Lecture Notes in Computer Science Edited by G. Goos, J. Hartmanis and J. van Leeuwen
1503
3 Berlin Heidelberg New Yor...

Author:
Giorgio Levi

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Lecture Notes in Computer Science Edited by G. Goos, J. Hartmanis and J. van Leeuwen

1503

3 Berlin Heidelberg New York Barcelona Budapest Hong Kong London Milan Paris Singapore Tokyo

Giorgio Levi (Ed.)

Static Analysis 5th International Symposium, SAS’98 Pisa, Italy, September 14-16, 1998 Proceedings

13

Series Editors Gerhard Goos, Karlsruhe University, Germany Juris Hartmanis, Cornell University, NY, USA Jan van Leeuwen, Utrecht University, The Netherlands

Volume Editor Giorgio Levi Universit`a di Pisa, Dipartimento di Informatica Corso Italia, 40, I-56125 Pisa, Italy E-mail: [email protected]

Cataloging-in-Publication data applied for Die Deutsche Bibliothek - CIP-Einheitsaufnahme Static analysis : 5th international symposium ; proceedings / SAS ’98, Pisa, Italy, September 14-16, 1998. Giorgio Levi (ed.). - Berlin ; Heidelberg ; New York ; Barcelona ; Budapest ; Hong Kong ; London ; Milan ; Paris ; Singapore ; Tokyo : Springer, 1998 (Lecture notes in computer science ; Vol. 1503) ISBN 3-540-65014-8

CR Subject Classification (1991): D.1, D.2.8, D.3.2-3, F.3.1-2, F.4.2 ISSN 0302-9743 ISBN 3-540-65014-8 Springer-Verlag Berlin Heidelberg New York This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting, reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer-Verlag. Violations are liable for prosecution under the German Copyright Law. c Springer-Verlag Berlin Heidelberg 1998 Printed in Germany Typesetting: Camera-ready by author SPIN 10638978 06/3142 – 5 4 3 2 1 0

Printed on acid-free paper

Foreword This volume contains the proceedings of the 1998 international symposium on static analysis (SAS’98) which was held in Pisa (Italy), on September 14-16, 1998 and was part of a federated conference with ALP-PLILP’98 and several workshops. SAS’98 is the annual conference and forum for researchers in all aspects of static analysis. It follows to SAS’94, SAS’95, SAS’96 and SAS’97 which were held respectively in Namur (Belgium), Glasgow (UK), Aachen (Germany) and Paris (France), and the international workshops WSA’92 held in Bordeaux (France) and WSA’93 held in Padova (Italy). In response to the call for papers, 48 papers were submitted. All papers were reviewed by at least three reviewers and the program committee met in Pisa to select 20 papers based on the referee reports. There was a consensus at the meeting that the technical papers were of very high quality. In addition to the submitted papers, SAS’98 had a number of outstanding invited speakers. Roberto Giacobazzi, Peter Lee, Amir Pnueli, Dave Schmidt, Scott Smolka, and Bernhard Steﬀen accepted our invitation to give invited talks or tutorials. Some of the papers (or abstracts) based on these talks are also included in this volume. SAS’98 has been fortunate to rely on a number of individuals and organizations. I want to thank all the program committee members and referees, for their hard work in producing the reviews and for such a smooth and enjoyable program committee meeting. Special thanks go to the conference chairman, Maurizio Gabbrielli, and to my students in Pisa who helped me a lot. More special thanks go to Vladimiro Sassone, who made available to SAS’98 his excellent system for handling submissions and reviews on the web, and to Ernesto Lastres and Ren`e Moreno who were my “system managers”. The use of this system made my life of program chairman much easier and I strongly recommend it to future program chairpersons. SAS’98 was sponsored by Universit` a di Pisa, Compulog Network, Consiglio Nazionale delle Ricerche, CNR-Gruppo Nazionale di Informatica Matematica, Comune di Pisa, and Unione Industriali di Pisa.

Giorgio Levi July 1998

Program Committee of SAS’98 Alex Aiken (Berkeley, USA) Maurice Bruynooghe (Leuven, Belgium) Michael Codish (Ben Gurion, Israel) Agostino Cortesi (Venezia, Italy) Radhia Cousot (Polytechnique Paris, France) Alain Deutsch (INRIA, France) Laurie Hendren (McGill, Canada) Fritz Henglein (DIKU, Denmark) Thomas Jensen (IRISA/CNRS, France) Alan Mycroft (Cambridge, UK) Flemming Nielson (Aarhus, Denmark) Thomas Reps (Wisconsin, USA) Dave Schmidt (Kansas State, USA) Mary Lou Soffa (Pittsburgh, USA) Harald Søndergaard (Melbourne, Australia) Bernhard Steffen (Passau, Germany)

Conference Chairman of SAS-ALP-PLILP’98 Maurizio Gabbrielli (Pisa, Italy)

List of Referees Torben Amtoft Andrew Appel Roberto Bagnara Dante Baldan Bruno Blanchet Rastislav Bodik Andrew Bromage Nicoletta Cocco Thomas Conway Patrick Cousot Bart Demoen Danny De Schreye Alessandra Di Pierro Manuel Faehndrich Gilberto Fil´e Jeff Foster Pascal Fradet Maurizio Gabbrielli Etienne Gagnon Roberto Giacobazzi Robert Gluck Eric Goubault Susanne Graf James Harland Nevin Heintze Frank Huch Jesper Jorgensen Andy King Jens Knoop Laura Lafave Vitaly Lagoon Martin Leucker Michael Leuschel Daniel Le Metayer Henning Makholm Elena Marchiori Kim Marriott Bern Martens Markus Mueller-Olm Anne Mulkers Hanne Riis Nielson Thomas Noll Dino Pedreschi Francesco Ranzato Jakob Rehof Olivier Ridoux Eva Rose Sabina Rossi Oliver Ruething David Sands Fausto Spoto Peter Stuckey Zhendong Su Cohavit Taboch Simon Taylor Peter Thiemann Arnaud Venet Zhe Yang Phillip Yelland

Contents

Data-Flow Analysis Bidirectional Data Flow Analysis in Code Motion: Myth and Reality . . . . . Oliver R¨ uthing

1

On a New Method for Dataflow Analysis of Java Virtual Machine Subroutines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 Masami Hagiya, Akihiko Tozawa Enabling Sparse Constant Propagation of Array Elements via Array SSA Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 Kathleen Knobe, Vivek Sarkar Assessing the Effects of Flow-Sensitivity on Pointer Alias Analyses . . . . . . . 57 Michael Hind, Anthony Pioli

Logic Programming Analysis of Normal Logic Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82 Fran¸cois Fages, Roberta Gori The Correctness of Set-Sharing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 Patricia M. Hill, Roberto Bagnara, Enea Zaffanella Deriving Analysers by Folding/Unfolding of Natural Semantics and a Case Study: Slicing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 Val´erie Gouranton

Concurrency A Symbolic Semantics for Abstract Model Checking . . . . . . . . . . . . . . . . . . . . 134 Francesca Levi Automatic Determination of Communication Topologies in Mobile Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152 Arnaud Venet Constructing Specific SOS Semantics for Concurrency via Abstract Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168 Chiara Bodei, Pierpaolo Degano, Corrado Priami

VIII

Contents

Abstract Domains A First-Order Language for Expressing Aliasing and Type Properties of Logic Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184 Paolo Volpe Refining Static Analyses by Trace-Based Partinioning Using Control Flow . 200 Maria Handjieva, Stanislav Tzolovski Building Complete Abstract Interpretations in a Linear Logic-Based Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215 Roberto Giacobazzi, Francesco Ranzato, Francesca Scozzari

Partial Evaluation On the Power of Homeomorphic Embedding for Online Termination . . . . . . 230 Michael Leuschel Analysis of Imperative Programs through Analysis of Constraint Logic Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246 Julio C. Peralta, John P. Gallagher, H¨ useyin Saˇglam Improving Control in Functional Logic Program Specialization . . . . . . . . . . . 262 E. Albert, M. Alpuente, M. Falaschi, P. Juli´ an, G. Vidal

Type Inference Directional Type Inference for Logic Programs . . . . . . . . . . . . . . . . . . . . . . . . . 278 Witold Charatonik, Andreas Podelski Finite Subtype Inference with Explicit Polymorphism . . . . . . . . . . . . . . . . . . . 295 Dominic Duggan

Optimization Sparse Jacobian Computation in Automatic Differentiation by Static Program Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311 M. Tadjouddine, F. Eyssette, C. Faure A New Solution to the Hidden Copy Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 327 Deepak Goyal, Robert Paige

Tutorials A Tutorial on Domain Theory in Abstract Interpretation . . . . . . . . . . . . . . . 349 Roberto Giacobazzi

Contents

IX

Program Analysis as Model Checking of Abstract Interpretations . . . . . . . . 351 David Schmidt, Bernhard Steffen

Invited Talks Certifying, Optimizing Compilation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381 Peter Lee Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383

Bidirectional Data Flow Analysis in Code Motion: Myth and Reality Oliver R¨ uthing Department of Computer Science, University of Dortmund, Germany [email protected]

Abstract. Bidirectional data flow analysis has become the standard technique for solving bit-vector based code motion problems in the presence of critical edges. Unfortunately, bidirectional analyses have turned out to be conceptually and computationally harder than their unidirectional counterparts. In this paper we show that code motion in the presence of critical edges can be achieved without bidirectional data flow analyses. This is demonstrated by means of an adaption of our algorithm for lazy code motion [15], which is developed from a fresh, specification oriented view. Besides revealing a better conceptual understanding of the phenomena caused by critical edges, this also settles the foundation for a new and efficient hybrid iteration strategy that intermixes conventional round-robin iteration with the exhaustive iteration on critical subparts.

1

Motivation

In data flow analysis equation systems involving bidirectional dependencies, i. e. dependencies from predecessor nodes as well as from successor nodes, are a well-known source for various kinds of difficulties. First, bidirectional equation systems are conceptually hard to understand. Mainly, this is caused by the lack of a corresponding operational specification like it is given by the the meet over all path (MOP) solution of a uni-directional data flow problem. Furthermore, Khedker and Dhamdhere recently proved that the costs for solving bidirectional data flow analysis problems may be significantly worse than for solving their unidirectional counterparts. This particularly holds for the only practically relevant class of bidirectional analyses, bit-vector based code motion problems. In fact, all known bidirectional problems are of this kind. Even more specifically, they are more or less variations of Morel’s and Renvoise’s pioneering algorithm for the elimination of partial redundancies [18,12,13,8,1,2,3,4,9]. Independently different researchers documented that bidirectionality is only required in programs that have critical edges [7,15], i. e. edges in a flow graph that directly lead from branch nodes to join nodes (see Fig. 1a for illustration). Ideally, critical edges can be completely eliminated by inserting empty synthetic nodes as depicted in Fig. 1b. In this example, the additional placement point enables the code motion transformation shown in Fig. 1c which eliminates the partial redundant G. Levi (Ed.): SAS’98, LNCS 1503, pp. 1–16, 1998. c Springer-Verlag Berlin Heidelberg 1998

2

O. R¨ uthing

a)

b)

1 x := a+b

2

c)

1 x := a+b

2

1 h := a+b x := h

2 h := a+b

3 y := a+b

3 y := a+b

3 y := h

Fig. 1. a) Critical Edge b) Edge splitting c) Transformational gain through edge splitting

computation on the path through node 1 and 3.1 However, in practice splitting of critical edges is sometimes avoided since it may cause additional unconditional jumps or decrease potential for pipelined execution.2 In this paper we investigate a new approach to code motion in the presence of critical edges. This is demonstrated by presenting a “critical” variant of our algorithm for lazy code motion [15]. However, the principal ideas straightforwardly carry over to all related code motion algorithms that employ bidirectional data flow analyses. Our algorithm is developed from a rigorous, specification oriented view. This particularly allows us to separate between different concerns. While safety in code motion is naturally associated with forward and backward oriented propagation of information, the presence of critical edges requires to impose additional homogeneity properties which can be expressed in terms of a side propagation of information. Actually, this clear separation allows us to avoid the usage of bidirectional dependencies in our specification. With regard to the variant of lazy code motion the contribution of this paper is threefold: – On a conceptual level we give a unidirectional specification of the problem. This particularly induces the first MOP characterization of code motion in the presence of critical edges. – We present a novel hybrid iteration strategy that separates the information flow along critical edges from the information flow along the uncritical ones. While the latter one is accomplished by an outer schedule proceeding in standard round-robin discipline the critical information flow is treated exhaustively by an inner schedule. – Almost as a by-product we obtain the first lifetime optimal algorithm for partial redundancy elimination in the presence of critical edges.

1 2

This is not possible in Fig. 1a, since hoisting a + b to node 2 introduces a new value on the rightmost path. Sometimes critical edges are not split only in situations that may harm the final code generation.

Bidirectional Data Flow Analysis in Code Motion: Myth and Reality

1.1

3

Related Work

As Khedker and Dhamdhere [14] and more recently Masticola et al. [17] noticed, critical edges do not add to the worst-case time complexity of iterative data flow analyses being based on a workset approach. However, this result cannot be generalized to bit-vector analyses where the iteration order has to be organized in a way such that structural properties of the flow graph are exploited in order to take maximum benefit of bit-wise parallel updates through efficient bit-vector operations. Hecht and Ullman [10] proved an upper bound on the number of roundrobin iterations that are necessary for stablization of monotone, unidirectional bit-vector problems. When proceeding in reverse postorder (or postorder for backward problems) d+2round-robin iterations are sufficient where dis the depth of the flow graph, i. e. the maximum number of backedges on an acyclic program path. Recently, Dhamdhere and Khedker [5,14] generalized this result towards bidirectional problems. However, a major drawback of their setting is that it is pinned to round-robin iterations. Unfortunately, such a schedule does not fit well to situations where information is side-propagated along critical edges. In this light, it is not astonishing that their results on the convergence speed of bidirectional bit-vector analyses are quite disappointing. They replace the depth dof a flow graph by its width w, which is the number of non-conform edge traversals on an information flow path.3 Unfortunately, the width is not a structural property of the flow graph, but varies with the problem under consideration, and unlike dwhich is 0 for acyclic programs is not even bounded in this case. Actually, the notion of width does not match to the intuition associated with the name, as even “slim” programs may have a large width. An intuitive reason for this behaviour is given in Fig. 2a which shows a program fragment with a number of critical edges. Let us consider that information flow in this example follows the equation Info(n) =

X m∈pred(n)

Info(m) +

X

Info(n0 )

n0 ∈succ(m)

which means that the information at node nis set to true if the information at a predecessor or the information of any “sibling” of nis true. We can easily see that the width of a flow graph with such a fragment directly depends on the number of critical edges, and therefore possibly grows linearly with the “length” of the program. It should be noted that such programs are by 3

Informatively, an information flow path is a sequence of backwards or forwards directed edges along which a change of information can be propagated. A forward traversal along a forward edge or a backward traversal along a backward edge are conform with a round-robin schedule proceeding (forwards) in reverse postorder. The other two kind of traversals are non-conform. Complemental notions apply to round-robin iterations proceeding in postorder.

4

O. R¨ uthing

a)

11

12

7 13

8

9

3 10

4

5

1 2

b)

6

Fig. 2. a) Acyclic program path responsible for the width of a program with reverse postordering of the nodes b) Slow information propagation in round-robin iterations

no means pathological and thus the linear growth of the width is not unlikely for real-life programs. In fact, considering the reverse postorder of nodes as given in Fig. 2a the large width is actually reflected in a poor behaviour of a round-robin iteration. Fig. 2b shows how the information slowly propagates along the obvious “path” displayed in this example being stopped in each round-robin iteration at a non-conform (critical) edge.4 Dhamdhere and Patil [6] proposed an elimination method for bidirectional problems that is as efficient as in the unidirectional case. However, it is restricted to a quite pathological class of problems, namely weakly bidirectional bit-vector problems and, as usual for elimination methods, is mainly designed for reducible control flow. Finally, our hybrid approach shares with the hybrid iteration strategy of [11] that it mixes a round-robin schedule with exhaustive subiterations. However, their subiterations are within strongly connected components and their approach is solely designed to speed up unidirectional iterations.

2

Preliminaries

We consider programs in terms of directed flow graphs = (N, E, s, e) with node set N , edge set E and unique start and end nodes s and e, respectively. Nodes n, m, . . . ∈ N represent (elementary) statements and are assumed to lie on a path from sto e. Finally, predecessors and successors of a node n ∈ N are denoted by pred(n)and succ(n), respectively, and P[n, m] stands for the set of finite paths between node n and m. 4

Shaded circles indicate the flow of informations along the “path”.

Bidirectional Data Flow Analysis in Code Motion: Myth and Reality

5

Local Predicates. As usual our reasoning is based on an arbitrary but fixed expression ϕ that is the running object for code movement. With each node of the flow graph two local predicates are associated. Comp(n): ϕ is computed at n, i. e. ϕ is part of the right-hand side expression associated with n. Transp(n): nis transparent for ϕ, i. e. none of ϕ’s variables is modified at n. Global Predicates. Based on these local predicates global program properties are specified. Usually, global predicates are associated with both entries and exits of nodes. In order to keep the presentation simple we assume that every node is split into an entry node and an empty exit node which inherit the set of predecessors and successors from the original node, respectively, and which are assumed to be connected by an edge leading from the entry node to the exit node. This step allows to restrict our reasoning to entry predicates. It should be noted, however, that this step is solely conceptual and does not eliminate any critical edge. In this paper partial redundancy elimination (PRE), or code motion (CM) as a synonym, stands for program transformations that 1. insert some instances of initialisation statements hϕ := ϕ at program points, where hϕ is a temporary variable that is exclusively assigned to ϕ and 2. replaces some original occurrences of ϕ by a usage of hϕ . In order to guarantee that the semantics of the argument program is preserved, we require that a code motion transformation must be admissible. Intuitively, this means that every insertion of a computation is safe, i. e. on no program path the computation of a new value is introduced at initialization sites, and that every substitution of an original occurrence of ϕ by hϕ is correct, i. e. hϕ always represents the same value as ϕ at use sites. This requires that hϕ is properly initialized on every program path leading to some use site in a way such that no modification occurs afterwards.5

3

Code Motion in the Absence of Critical Edges

Before presenting our new approach to PRE in the presence of critical edges we shall first briefly recall the basic steps of lazy code motion [15] as a typical representative of an algorithm that relies on the absence of critical edges. Lazy code motion was the first algorithm for partial redundancy elimination that succeeded in removing partial redundancies as good as possible, while avoiding any unnecessary register pressure. This was mainly achieved by a rigorous redesign of Morel’s and Renvoise’s algorithm. Starting from a specification oriented view the key points was a hierarchical separation between the primary and the secondary concern of partial redundancy elimination, namely to minimize the number of 5

For a formal definition see [16].

6

O. R¨ uthing

computations and to avoid unnecessary register pressure, respectively. This hierarchical organization is reflected in a two-step design of the algorithm: lazy code motion rests on busy code motion. In following we briefly summarize the details of these transformations. 3.1

Busy Code Motion

Busy code motion (BCM) [15,16,9] places initializations as early as possible while replacing all original occurrences of ϕ. This is achieved by determining the earliest program points, where an initialization is safe. Technically, the range of safe program points can be determined by separately computing down-safe and up-safe program points. Both are given through the greatest solutions of two uni-directional data flow analyses, respectively.6

DnSafe(n) = (n 6= e) · (Comp(n) + Transp(n) ·

Y

DnSafe(m))

m∈succ(n)

Y

UpSafe(n) = (n 6= s) · Transp(m) ·

(Comp(m) + UpSafe(m))

m∈pred(n) def

Safe (n) = UpSafe(n) + DnSafe(n) def

Earliest (n) = Safe (n) · ((n = s) +

X

Safe (m))

m∈pred(n)

Despite of its surprising simplicity, BCM already reaches computational optimality, i. e. programs resulting from this transformation have at most as many ϕ-occurrences on every path from sto eas any other result of an admissible code motion transformation (cp. [15,16]). 3.2

Lazy Code Motion

In addition to BCM, lazy code motion (LCM) takes the lifetimes of temporaries into account. This is accomplished by placing initialisations as late as possible but as early as necessary, where the latter requirement means “necessary in order to reach computational optimality”. Technically, this is achieved by determining the latest program points where a BCM-initialisation might be delayed to, which leads to one additional uni-directional data flow analysis.7 6 7

As common “·”,“+” and overlining stand for logical conjunction, disjunction and negation, respectively. In [15,16] an additional analysis is employed determining isolated program points, i. e. program points where initialisations are only used immediately afterwards. This aspects, however, can independently be treated by means of a postprocess. For the sake of simplicity we skip the isolation analysis in this paper.

Bidirectional Data Flow Analysis in Code Motion: Myth and Reality

Y

Delayed (n) = Earliest (n) + (n 6= s) ·

7

Delayed (m) · Comp(m)

m∈pred(n) def

Latest (n) = Delayed (n) · ( Comp(n) +

X

Delayed (m) )

m∈succ(n)

LCM is computationally optimal as well as lifetime optimal, i. e. the temporary associated with ϕhas a lifetime range that is included in the lifetime range of any other program resulting from a computationally optimal code motion transformation (cp. [15,16]). The lifetime range of a temporary comprises all nodes whose exits occur in between an initialisation site and a use site such that no other initialisations are situated in between.8

4

Code Motion in the Presence of Critical Edges

In this section our new approach to PRE in the presence of critical edges is elaborated in full details. First we shall investigate the principal differences to the setting presented in Sect. 3.2. As opposed to flow graphs without critical edges there are usually no computationally optimal representatives. In fact, Fig. 3 shows two admissible, but computationally incomparable transformations that cannot be improved any further. The first one is simply given by the identical transformation of the program in Fig. 3a, the result of the second one is displayed in Fig. 3b. Each of the resulting programs has exactly one computation on the path that is emphasised in the dark shade of grey, while having two computations on the path being emphasised in the light shade of grey, respectively. Thus there is no computationally optimal code motion transformation with respect to the original program in Fig. 3a. a)

b)

1

a+b

a+b 1 h := h

2 4

3 5

a+b

c)

6

a+b

2 h := a+b 4

3 5

h

a+b 1 h := h

6

a+b

3 5

h

2 h := a+b 4 6 h := a+b h

Fig. 3. a & b) Incomparable admissible program transformations c) Program degradation through a naive adaption of busy expression motion 8

In [20] we show that this optimality result is only adequate for flat universes of expressions. If both composite expressions and their subexpressions are moved, then the notion of lifetime optimality changes and a significantly more sophisticated technique has to be applied. Nonetheless, LCM still provides a basic ingredient of this approach.

8

O. R¨ uthing

This problem can be overcome by restricting the range of program transformations to those that are profitable, which means those that actually improve their argument programs. Note that this requirement excludes Fig. 3b as a reasonable code motion transformation. Obviously, profitability does not provide a further restriction for flow graphs without critical edges, where computationally optimal code motion transformations are granted to exist. In the presence of critical edges, however, this additional constraint is necessary in order to yield computationally optimal results at all. 4.1

Busy Code Motion

In this section we will develop a counterpart to BCM in the presence of critical edges. After briefly sketching the difficulties that prohibit a straightforward adaption of the uncritical solution, a correct approach is systematically developed from a specification that incorporates the special role of critical edges. Unfortunately, BCM as presented in Sect. 3.1 cannot straightforwardly be applied to flow graphs with critical edges. This is because such a naive adaption may include non-profitable transformations as it is illustrated in Fig. 3c, where the marked range of down-safe program points would yield earliest initialisation points at nodes 1, 2 and 6.9 Homogeneous Propagation of Down-Safety. The key for a useful critical variant of BCM is to impose an additional homogeneity requirement on downsafety that ensures that the information propagates either to all or to none of its predecessors, which grants that earliest program points become a proper upper borderline of the region of safe program points. In fact, in the absence of critical edges down-safety has the following homogeneity property: ∀ n ∈ N. DnSafe(n) ⇒ (∀ m ∈ pred(n). Safe (m) ∨ ∀ m ∈ pred(n). ¬DnSafe(m))

Note that the first term of the disjunction uses safety rather than down-safety, since propagation of down-safety needs not to be considered for predecessors that are up-safe anyhow.10 Now this propery has to be forced explicitly. For instance, in Fig. 3c node 6 as well as node 3 are down-safe, while node 4 is not. Therefore, let us consider the following notion of homogeneous down-safety: Definition 1 (Homogeneous Down-Safety). A predicate HDnSafe on the nodes of N is a homogeneous down-safety predicate iff for any n ∈ N 1. HDnSafe is conform with down-safety: HDnSafe (n) ⇒ (n 6= e) ∧ (Comp(n) ∨ Transp(n) ∧ ∀ m ∈ succ(n). HDnSafe (n)) 9

10

Modifying this example by removing the computation of a + bfrom node 1, would even result in a transformation that does not not improve any path while strictly impairing some. In the absence of critical edges this makes no difference to ∀ m ∈ pred(n). DnSafe(m).

Bidirectional Data Flow Analysis in Code Motion: Myth and Reality

9

2. HDnSafe is homogeneous: HDnSafe (n) ⇒ (∀ m ∈ pred(n). (HDnSafe (m) ∨ UpSafe(m)) ∨ ∀ m ∈ pred(n). ¬HDnSafe (m))

Obviously, homogeneous down-safety predicates are closed under “union”.11 Thus there exists a unique largest homogeneous down-safety predicate DnSafeHom , which gives rise to a homogeneous version of safety, too: def

∀ n ∈ N. SafeHom (n) = DnSafeHom (n) ∨ UpSafe(n) It should be noted that this definition is developed from a pure specification oriented reasoning and can be seen as a first rigorous characterization of downsafety in the presence of critical edges: down safety is described by a backward directed data flow problem which is restricted by additional homogeneity constraints. This is in contrast to other algorithms, where bidirectional equation systems are postulated in an ad-hoc fashion without any separation of their functional components. Earliest program points are defined as in the uncritical case, but with the difference of using the homogeneous version of down-safety in place of the usual one. def

EarliestHom (n) = DnSafeHom (n) · ((n 6= s) +

X

SafeHom (m))

m∈pred(n)

The earliest program points serve as insertion points of BCM for flow graphs with critical edges (CBCM). With a similar argumentation as for BCM is easy to prove that CBCM is indeed computationally optimal, however, only relatively to the profitable transformations.

Computing CBCM: The Data Flow Analyses. In this part we present how the specifying solution of CBCMcan be translated into appropriate data flow analyses determining the range of homogeneously safe program points. We will discuss three alternative approaches: (1) A “classical” one via bidirectional analyses, (2) a new non-standard approach that transforms the problem into one with purely unidirectional equations and (3) a hybrid approach that separates backwards flow from side propagation. The Bidirectional Approach The specification of Definition 1 can straightforwardly be transfered into a bidirectional equation system for down-system. 11

This means the predicate defined by the pointwise conjunction of the predicate values.

10

O. R¨ uthing

UpSafe(n) = (n 6= s) · Transp(m) ·

Y

(Comp(m) + UpSafe(m))

m∈pred(n)

DnSafeHom (n) = (n 6= e) · (Comp(n) + Transp(n) · Y Y DnSafeHom (m) · (UpSafe(n0 ) + DnSafeHom (n0 ))) m∈succ(n)

n0 ∈pred(m)

def

SafeHom (n) = UpSafe(n) + DnSafeHom (n)

Unfortunately, the above bidirectional data flow problem shares the problems sketched in Fig. 2 when subjected to a round-robin iteration strategy. In fact, violation of homogeneous safety follows exactly the same definition pattern as Info does in this example.12 Hence slow propagation of down-safety would be also apparent in CBCM. The Unidirectional Approach It is easy to see that in the bidirectional equation system there is no “true” Q forward propagation of down-safety information as the scope of the term n0 ∈pred(m) (UpSafe(n0 ) + DnSafeHom (n0 ))is restricted in its context. Rather this can be seen as a “side propagation” of down-safety information along zig-zag paths. For a technical description let us define the set of zig-zag successors zsucc(n)of a node n ∈ N as the smallest set of nodes satisfying (see Fig. 4 for illustration): 1. succ(n) ⊆ zsucc(n) 2. ∀m ∈ zsucc(n). succ(pred(m)) ⊆ zsucc(n) In our example zig-zag propgation of non-down-safety is further stopped at nodes where up-safety can be established. Hence we introduce a parameterized notion of zsucc(n)which is defined for M ⊆ N by: 1. succ(n) ⊆ zsucc M (n) 2. ∀m ∈ zsucc M (n). succ(pred(m) \ M ) ⊆ zsucc M (n) def

With XUS = rewritten as:

{m ∈ N | UpSafe(m)}the equation for down-safety can be

DnSafeHom (n) = n 6= e · (Comp(n) + Transp(n) ·

Y

DnSafeHom (m))

m∈zsucc XUS (n)

Note that this equation system can be seen as a unidirectional one that operates on a flow graph that is enriched by shortcut edges drawn between nodes and their zig-zag successors (see Fig. 4b). 12

Actually, here non-down-safety is propagated in a dual fashion.

Bidirectional Data Flow Analysis in Code Motion: Myth and Reality b)

a) 2

1

4

11

5

6

2

1

3

4

7

5

succ(1 )

3

6

7

zsucc(1)

Fig. 4. (a) Program fragment with a nest of critical edges (b) Zig-zag successors and virtual shortcut edges of node 1

However, we do not actually recommend to perform such a transformation, as this would require to introduce an unnecessarily large number of additional edges. For a zig-zag chain of k critical edges as shown in Fig. 4a the number of shortcut edges is of order k 2 . Although, long zig-zag chains of critical edges can be expected to be rare in practice, we will show that information propagation can be organized without such blow-up in the number of edges. However, the important contribution of the unidirectional approach is that it provides the first meet over all paths characterization of PRE in the presence of critical edges.13 The Hybrid Approach: The hybrid approach rather addresses the organization of the iteration process than the equation system itself. As we have learned, bidirectional problems like our formulation of homogeneous down-safety do not fit together with a round-robin schedule based upon postorder traversals. The hybrid approach modifies the conventional round-robin schedule by integrating zig-zag propagation of information. This is achieved by clustering the nodes in a flow graph in a way such that side propagation of information can take benefit of much potential for simultaneous work. The overall schedule of the approach can be sketched as follows: Preprocess: Collapsing of nodes according to side flow of information Outer Schedule: Process the collapsed nodes in postorder until stabilization is reached performing an Inner Schedule 1. For each node within the collapsed one perform information propagation along its outgoing uncritical edges 2. Perform exhaustive information propagation along the outgoing critical edges within the collapsed node In the following we will go into the details of this process. • The preprocess: Clustering of nodes groups together nodes of N according to the following equivalence relation: def

n ≡ m ⇔ zsucc(n) = zsucc(m) 13

Actually, only the notion of paths has to be extended towards paths across shortcut edges.

12

O. R¨ uthing

It should be noted that Gcan be decomposed into its equivalence classes easily by tracing zig-zag paths of critical edges originating at an unprocessed node. For instance, starting with node 1 in Fig. 4a we obtain the equivalence class {1, 2, 3}by following the critical edges. Clearly, this process can be managed in order O(e)where edenotes the number of edges in E. All nodes of an equivalence classes are collapsed into a single node that inherits all incoming and outgoing edges of its members (see Fig. 5 for illustration). a)

b)

Fig. 5. (a) Equivalent nodes (b) Collapsing equivalent nodes

• The outer schedule: The flow graph G0 that results from the collapsing preprocess is used in order to determine the round-robin schedule which drives information backwards. It should be noted that the depth of G0 may differ from the depth of the original flow graph Gin both directions: the depth may increase or decrease by collapsing. This is illustrated in Fig. 6. While collapsing nodes in Part a) decreases the depth, since the indicated path is no longer acyclic, collapsing in Part b) allows to construct a longer acyclic path as indicated by the dashed line connecting two independent acyclic paths.

a)

b)

Fig. 6. (a) Decrease of depth due to collapsing of nodes (b) Increase of depth due to collapsing of nodes

• The inner schedule: The first step of the inner schedule is quite trivial. Considering a node n ∈ N within the collapsed node under consideration and an

Bidirectional Data Flow Analysis in Code Motion: Myth and Reality

13

uncritical edge (n, m) ∈ Ethe value of DnSafeHom (n)is changed to false if and only if Transp(n) · DnSafeHom (m) holds.14 The second step of the inner schedule is the central innovation in our approach and has to be elaborated with some care. Within any collapsed node side propagation of down-safety information along critical edges is done exhaustively by using a subiteration process. Information propagation here means that for nodes n, min the collapsed node under consideration with m ∈ pred(succ(n))the value of DnSafeHom (m)is changed to false if and only if DnSafeHom (n) · UpSafe(n)holds. The crucial point, however, is to organize the information flow along the critical edges. The situation is easy if the zig-zag paths are acyclically shaped as displayed in Fig. 7a or Fig. 7b. In this case the equivalence class can be represented as a tree, which can already be built while preprocessing this class. Following the topological order of the tree, information can be propagated completely by a bottom-up traversal (from the leaves to the root) followed by a top-down traversal (from the root to the leaves). Unfortunately, in general there may be cycles of critical edges as shown in Fig. 7c and Fig. 7d. Hence a problem of the same difficulty as in the backward propagation of information shows up in the side-propagation step. However, separating both problems is useful as we expect nested cycles of critical edges to be a phenomenon that is extremely rare in practice. Nonetheless, to cope with them is quite straightforward. As in the acyclic case, the equivalence class can be represented as a tree with some additional non-tree edges establishing cycles. The only difference to the non-cyclic case is that the tree traversals have to be iterated more than once until the process gets stable. To estimate the number of traversal we borrow the arguments from conventional unidirectional analysis. Denoting the non-tree edges within the tree-like representation of an equivalence class as critical backedges the number of iterations is bound by dc , where dc is the maximum number of critical back edges along an acyclic path in any component representation. Complexity of the Hybrid Approach: All together the iteration of homogeneous down-safety in the hybrid approach requires to apply the outer schedule until stabilization. Since the inner schedule propagates the information completely within each collapsed node the overall effort can be estimated by (d0 + 2)(eu + 2(dc + 2)ec ) bit-vector steps, where eu and ec denote the number of uncritical and critical edges, respectively, d’ is the depth of the collapsed flow graph G0 and dc the critical depth as defined before. It is commonly argued that the depth of a flow graph is a reasonably small constant in practice. We already discussed that dc is at least as likely to be a small constant, too. Hence the algorithm is expected to behave linear in efor real-life programs. In particular, we succeed in giving the first linear worst-case estimation for acyclic programs as in our introductory example of Fig. 2. 14

Note that there are no uncritical edges directly connecting different nodes of an equivalence class.

14

O. R¨ uthing a)

c)

d)

b)

Fig. 7. Shapes of equivalent nodes: (a) chain of critical edges, (b) tree of critical edges, (c) cycle of critical edges and (d) structure with nested cycles of critical edges

4.2

Lazy Code Motion

Similar to the situation in Sect. 4.1 also the relevant analyses of LCM as defined in Sect. 3.2 cannot naively be adapted to flow graphs with critical edges. Again the reason for this behaviour lies in a homogeneity defect, but now with respect to delayability. In fact, for flow graphs without critical edges we have Delayed (n) ⇒ ( ∀ m ∈ succ(n). Delayed (m) ∨ ∀ m ∈ succ(n). ¬Delayed (m) ) This property may now be violated. Hence one has to force homogeneity explicitly in order to yield an appropriate critical variant of lazy code motion. Therefore, let us consider the following notion of homogeneous delayability. Definition 2 (Homogeneous Delayability). A predicate HDelayed on N is a homogeneous delayability predicate iff for any n ∈ N 1. HDelayed (n)is conform with delayability:

HDelayed (n) ⇒ EarliestHom (n) ∨ ((n 6= s) ∧ ∀ m ∈ pred(n). HDelayed (m) ∧ ¬Comp(m))

2. HDelayed (n)is homogeneous: Delayed (n) ⇒ ( ∀ m ∈ succ(n). HDelayed (m) ∨ ∀ m ∈ succ(n). ¬HDelayed (m))

Obviously, homogeneous delayability predicates are closed under “union”. Thus there exists a unique largest homogeneous delayability predicate DelayedHom . This gives rise to a new version of latestness characterizing the insertion points of lazy code motion for flow graphs with critical edges (CLCM). def

LatestHom (n) ⇔ DelayedHom (n) ∧ (Comp(n) ∨ ∃ m ∈ succ(n). ¬DelayedHom (m))

Bidirectional Data Flow Analysis in Code Motion: Myth and Reality

15

Using the same definitions for lifetime optimality as in Sect. 3.2 we succeed in proving lifetime optimality of CLCM. Computing CLCM. In analogy to Sect. 11 the delayability property can either be coded into a bidirectional equation system or, more interestingly, again be expressed using a unidirectional formulation:

DelayedHom (n) = EarliestHom (n) ∨ ((n 6= s) ∧ ∀ m ∈ zpred (n). DelayedHom (m) ∧ ¬Comp(m)) This definition is based on zig-zag predecessor, which are defined completely along the lines of zig-zag successors. However, in contrast to down-safety zpred needs not to be parameterized this time. Using this characterization the same techniques for hybrid iteration can be used as in Sect. 11.

5

Conclusion

We presented an adaption of lazy code motion to flow graphs with critical edges as a model how to cope with bidirectional dependencies in code motion. On the conceptual level we isolated homogeneity requirements as the source for bidirectional dependencies. This led to a new hybrid iteration strategy which is almost as fast as its unidirectional counterparts. This dramatically improves all known estimations for bidirectional bit-vector methods. Nonetheless, we still recommended to eliminate critical edges as far as possible, since critical edges are also responsible for problems of a different flavour [19]. However, any implementation of code motion that has to cope with critical edges will definitely benefit from the ideas presented in this paper.

References 1. D. M. Dhamdhere. A fast algorithm for code movement optimization. ACM SIGPLAN Notices, 23(10):172 – 180, 1988. 2. D. M. Dhamdhere. A new algorithm for composite hoisting and strength reduction optimisation (+ Corrigendum). International Journal of Computer Mathematics, 27:1 – 14 (+ 31 – 32), 1989. 3. D. M. Dhamdhere. A usually linear algorithm for register assignment using edge placement of load and store instructions. Journal of Computer Languages, 15(2):83 – 94, 1990. 4. D. M. Dhamdhere. Practical adaptation of the global optimization algorithm of Morel and Renvoise. ACM Transactions on Programming Languages and Systems, 13(2):291 – 294, 1991. Technical Correspondence. 5. D. M. Dhamdhere and U. P. Khedker. Complexity of bidirectional data flow analysis. In Conf. Record of the 20th ACM Symposium on the Principles of Programming Languages, pages 397–409, Charleston, SC, January 1993.

16

O. R¨ uthing

6. D. M. Dhamdhere and H. Patil. An elimination algorithm for bidirectional data flow problems using edge placement. ACM Transactions on Programming Languages and Systems, 15(2):312 – 336, April 1993. 7. D. M. Dhamdhere, B. K. Rosen, and F. K. Zadeck. How to analyze large programs efficiently and informatively. In Proc. ACM SIGPLAN Conference on Programming Language Design and Implementation’92, volume 27,7 of ACM SIGPLAN Notices, pages 212 – 223, San Francisco, CA, June 1992. 8. K.-H. Drechsler and M. P. Stadel. A solution to a problem with Morel and Renvoise’s “Global optimization by suppression of partial redundancies”. ACM Transactions on Programming Languages and Systems, 10(4):635 – 640, 1988. Technical Correspondence. 9. K.-H. Drechsler and M. P. Stadel. A variation of Knoop, R¨ uthing and Steffen’s lazy code motion. ACM SIGPLAN Notices, 28(5):29 – 38, 1993. 10. M. S. Hecht and J. D. Ullman. A simple algorithm for global data flow analysis problems. SIAM Journal on Computing, 4(4):519 – 532, 1977. 11. S. Horwitz, A. Demers, and T. Teitelbaum. An efficient general iterative algorithm for data flow analysis. Acta Informatica, 24:679 – 694, 1987. 12. S. M. Joshi and D. M. Dhamdhere. A composite hoisting-strength reduction transformation for global program optimization – part I. International Journal of Computer Mathematics, 11:21 – 41, 1982. 13. S. M. Joshi and D. M. Dhamdhere. A composite hoisting-strength reduction transformation for global program optimization – part II. International Journal of Computer Mathematics, 11:111 – 126, 1982. 14. U. P. Khedker and D. M. Dhamdhere. A generalized theory of bit vector data flow analysis. ACM Transactions on Programming Languages and Systems, 16(5):1472 – 1511, September 1994. 15. J. Knoop, O. R¨ uthing, and B. Steffen. Lazy code motion. In Proc. ACM SIGPLAN Conference on Programming Language Design and Implementation’92, volume 27,7 of ACM SIGPLAN Notices, pages 224 – 234, San Francisco, CA, June 1992. 16. J. Knoop, O. R¨ uthing, and B. Steffen. Optimal code motion: Theory and practice. ACM Transactions on Programming Languages and Systems, 16(4):1117– 1155, 1994. 17. P. M. Masticola, T. J. Marlowe, and B. G. Ryder. Lattice frameworks for multisource and bidirectional data flow problems. ACM Transactions on Programming Languages and Systems, 17(5):777 – 802, 1995. 18. E. Morel and C. Renvoise. Global optimization by suppression of partial redundancies. Communications of the ACM, 22(2):96 – 103, 1979. 19. O. R¨ uthing. Interacting Code Motion Transformations. Their Impact and their complexity. PhD thesis, Institut f¨ ur Informatik und Praktische Mathematik, Christian-Albrechts-Universit¨ at Kiel, Germany, 1997. Available as http://sunshine.cs.uni-dortmund.de/˜ruething/diss.ps.gz. 20. O. R¨ uthing. Optimal code motion in the presence of large expressions. In Proc. Internatinal Conference on Computer Languages (ICCL’98), Chicago, IL., 1998. IEEE.

On a New Method for Dataflow Analysis of Java Virtual Machine Subroutines Masami Hagiya and Akihiko Tozawa Department of Information Science, Graduate School of Science, University of Tokyo {hagiya,miles}@is.s.u-tokyo.ac.jp

Abstract. The bytecode verifier of the Java Virtual Machine, which statically checks the type safety of Java bytecode, is the basis of the security model of Java and guarantees the safety of mobile code sent from an untrusted remote host. However, the type system for Java bytecode has some technical problems, one of which is in the handling of subroutines. Based on the work of Stata and Abadi and that of Qian, this paper presents yet another type system for Java Virtual Machine subroutines. Our type system includes types of the form last(x). A value whose type is last(x) is the same as that of the x-th variable of the caller of the subroutine. In addition, we represent the type of a return address by the form return(n), which means returning to the n-th outer caller. By virtue of these types, we can analyze instructions purely in terms of type, and as a result the correctness proof of bytecode verification becomes extremely simple. Moreover, for some programs, our method is more powerful than existing ones. In particular, our method has no restrictions on the number of entries and exits of a subroutine.

1

Introduction

One contribution of Java is its bytecode verifier, which statically checks the type safety of bytecode for the JVM (Java Virtual Machine) prior to execution. Thanks to the bytecode verifier, bytecode sent from an untrusted remote host can be executed without the danger of causing type errors and destroying the entire security model of Java, even when the source code is not available. Verifying the type safety of bytecode (or native code) seems to be a new research area that is not only of technical interest but also of practical importance, due to the availability of remote binary code in web browsers and other applications. Much effort has been put into guaranteeing the security of Java programs, including the type safety of bytecode: – Security model for Java applets: The security model for Java applets is said to consist of three prongs: the bytecode verifier, the applet class loader and the security manager [9]. In this model, the bytecode verifier plays the most fundamental role, on which the other two prongs are based. If the bytecode verifier is cheated, the other two also become ineffective. G. Levi (Ed.): SAS’98, LNCS 1503, pp. 17–32, 1998. c Springer-Verlag Berlin Heidelberg 1998

18

M. Hagiya and A. Tozawa

– Type safety of source code: The type safety of Java programs has been proved, either formally (using a theorem proving assistant), or rigorously (but not formally) [3,4,17,12]. This means that a program that has passed static type checking is guaranteed to cause no type errors while it is running. – Safety of class loading: Java allows classes to be loaded lazily, i.e., only when classes are actually accessed. In order to support various loading disciplines, Java allows programmers to define their own class loaders. This has opened one of the security holes in Java [14]. To avoid such a security hole, Dean formalized part of the class loader functionality and formally proved its correctness using PVS [2]. Goldberg’s main concern is also class loading, though bytecode verification is addressed [5]. – Type safety of bytecode: The bytecode verifier statically checks the type safety of Java bytecode. If the bytecode verifier accepts incorrectly typed bytecode, it will break the entire security model of Java. It guarantees that no type error occurs at each instruction by performing dataflow analysis on the bytecode. Besides the researches into different aspects of security mentioned above, there are also some on-going projects that are developing more secure network programming environments [7,15]. This paper concerns bytecode verification. Since this is the basis of the entire security model of Java, it is desirable to rigorously prove that any bytecode program that has passed bytecode verification will never cause a runtime type error. In order to be able to show the correctness of the bytecode verifier, one has to begin by formally specifying the operational semantics of the virtual machine (e.g., [1]), and then give the formal specification of the bytecode verifier, based on its informal specification written in English [8]. Qian rigorously defined the operational semantics of a subset of the JVM and formulated the bytecode verifier as a type system [13]. He then succeeded in proving the correctness of the bytecode verifier, though not completely. Bytecode verification of the JVM has some technical challenges. One is that of handling object initialization, as objects created but not yet initialized may open a security hole. In Qian’s work, much attention is paid to the handling of object initialization. Another is that of handling the polymorphism of subroutines. This paper addresses this issue. Inside a JVM subroutine, which is the result of compiling a finally clause in a try statement, local variables may have values of different types, depending on the caller of the subroutine. This is a kind of polymorphism. To investigate how to analyze JVM subroutines, Stata and Abadi defined a type system for a small subset of the JVM and proved its correctness with respect to the operational semantics of the subset [16]. Qian’s system is similar to that of Stata and Abadi in its handling of subroutines [13]. Both systems faithfully follow the specification of the bytecode verifier, and make use of information as to which variables are accessed or modified in a subroutine. Those variables that are not accessed or modified are simply ignored during analysis of the subroutine.

On a New Method for Dataflow Analysis

19

This paper takes a different approach. We introduce types of the form last(x). A value whose type is last(x) is the same as that of the x-th variable of the caller of the subroutine. In addition, we represent the type of a return address by the form return(n), which means returning to the n-th outer caller. Our approach has the following advantages. – By virtue of the last and return types, we can analyze instructions purely in terms of types. As a result, the proof of the correctness of bytecode verification becomes extremely simple, and we do not need a separate analysis on variable access or modification. – For some programs (unfortunately, not those produced by the Java compiler), our method is more powerful than existing ones. In particular, our method has no restrictions on the number of entries and exits of a subroutine. Stata and Abadi enforce a stack-like behavior on subroutine calls with their analysis, which does not account for out-of-order returns from subroutines, though they are actually produced by the Java compiler. Due to these advantages, we hope that our method can be modified and applied to the analysis of other kinds of bytecode or native code [10,11]. This paper is organized as follows: In the next section, we explain JVM subroutines in more detail and our approach to bytecode verification. In Sect. 3, a subset of the JVM, similar to that of Stata and Abadi, is defined. In Sect. 4, our method for analysis of bytecode is described and its correctness is shown. In Sect. 5, issues of implementation are briefly discussed. Section 6 offers concluding remarks.

2

Analysis of Subroutines

The JVM is a classical virtual machine consisting of – a program counter, – an array for storing the values of local variables, – a stack for placing arguments and results of operators, called the operand stack, – a heap for storing method code and object bodies, and – a stack for frames, each of which is allocated for each invocation of a method and consists of the program counter, the local variables, and the operand stack. To allow for checking the type safety of bytecode, it prepares different instructions for the same operation depending on the type of the operands. For example, it has the instruction istore for storing integers and the instruction fstore for storing floating-point numbers. JVM subroutines are used mainly for compiling the finally clauses of try statements of Java. Notice that subroutine calls are completely different from method calls. A subroutine is locally defined inside a method and, unlike a method, is not allowed to call itself recursively.

20

M. Hagiya and A. Tozawa

In this paper, we define a virtual machine based on JVML0, which was formulated by State and Abadi for investigating how to analyze JVM subroutines. Subroutines are called by instructions of the following form. jsr(L) An instruction of this form pushes the address of its next instruction (i.e., the return address) onto the operand stack and jumps to the subroutine L. Subroutines are usually defined as follows. L : store(x) .. . ret(x)

store(x) pops a value from the operand stack and stores it in the x-th local variable. ret(x) is an instruction for jumping to the address stored in the x-th local variable. Note that, in contrast to the JVM, our virtual machine has only one instruction for the store operation. Subroutines in the JVM have made bytecode verification more difficult for the following reasons. – The return address of ret(x) can only be determined after the values of the local variables have been analyzed by the bytecode verifier. On the other hand, return addresses affect the control flow and the analysis of local variables. – In some situations, a subroutine does not return to the immediate caller, but returns to an outer caller, such as the caller of the caller. – Inside a subroutine, local variables may have values of different types, depending on the caller of the subroutine. In this paper, we introduce types of the form last(x) in order to address the last problem. A value having this type must have the same value as that of the x-th local variable in the caller of a subroutine. As an example, let us consider the following program. const0 store(1) 2 : jsr(7) constNULL store(1) 5 : jsr(7) halt

7 : store(0) load(1) store(2) ret(0)

Subroutine 7 is called from two callers (2 and 5). The return address is stored in variable 0 (the 0-th local variable). The value of variable 1 is an integer when the subroutine is called from caller 2, and is an object pointer when called from caller 5. This is a typical case in which a subroutine is polymorphic. In this

On a New Method for Dataflow Analysis

21

program, the value of variable 1 is copied to variable 2. As the JVM has different store instructions depending on the type of the operand, the above program is impossible under the JVM. However, even if the JVM allowed the untyped store instruction, the bytecode verifier of the JVM would signal an error for the above program, because it always assigns a unique type for any variable accessed in a subroutine. According to the specification of the JVM [8], – For each instruction and each jsr needed to reach that instruction, a bit vector is maintained of all local variables accessed or modified since the execution of the jsr instruction. – For any local variable for which the bit vector (constructed above) indicates that the subroutine has accessed or modified, use the type of the local variable at the time of the ret. – For other local variables, use the type of the local variable before the jsr instruction. The work of Stata and Abadi and that of Qian faithfully follow this specification. Local variables that are accessed or modified in each subroutine are recorded. Those variables that are not accessed or modified are simply ignored during the subsequent analysis of the subroutine. The method proposed in this paper assigns a type of the form last(1) to variable 1 in subroutine 7. This means that it includes a value passed from variable 1 of the caller. This type is propagated through instructions in the subroutine. In particular, by the instruction store(2), the type of variable 2 becomes last(1). This information is then used when the control returns from the subroutine to the caller. In this way, the polymorphism of local variables in a subroutine is expressed by types of the form last(x). In our method, return addresses have types of the form return(n). A type of the form return(n) means to return to the n-th outer caller. For example, the address returning to the immediate caller has type the return(1), while the address returning to the caller of the caller has type the return(2). In Stata and Abadi’s work (and similarly in Qian’s work), return addresses have types of the form (ret-from L), where L is the address of the subroutine, from which the callers of the subroutines are obtained. In our analysis, the callers are obtained from the set of histories assigned to each instruction (cf. Sect. 4.3).

3

Virtual Machine

In this section we formalize a subset of the JVM, which resembles that of Stata and Abadi [16]. Differences are mainly for the examples by which we want to show the power of our framework.

22

3.1

M. Hagiya and A. Tozawa

Values

A value is a return address or an integer or an object pointer. We can easily add other kinds of values, such as that of floating point number. In the following formal treatment of the operational semantics of the virtual machine, a return address has the constructor retaddr, an integer the constructor intval, and an object pointer the constructor objval. They all take an integer as an argument. 3.2

Instructions

A bytecode program is a list of instructions. An instruction takes one of the following formats. jsr(L) load(x) const0 inc(x) if0(L) halt

(L: subroutine address) (x: variable index) (x: variable index) (L: branch address)

ret(x) (x: variable index) store(x) (x: variable index) constNULL ifNULL(L)

(L: branch address)

Each mnemonic is considered as a constructor of instructions. Some of the mnemonics takes a nonnegative integer x or L as an operand. 3.3

Operational Semantics

The virtual machine consists of – the program, which is a list of instructions and denoted by P , – the program counter, which is an index to P , – the local variables, where the list of values of the local variables is denoted by f , and – the operand stack, denoted by s. Let us use the notation l[i] for extracting the i-th element of list l, where the first element of l has the index 0. The i-th instruction of the program P is denoted by P [i]. The value of the x-th local variable is denoted by f [x]. The p-th element of the operand stack s is denoted by s[p], where s[0] denotes the top element of s. As in the work by Stata and Abadi, the operational semantics of the virtual machine is defined as a transition relation between triples of the form hi, f, si, where i is the program counter, i.e., the index to the program P , f the value list of the local variables, and s the operand stack. While the length of s may change during execution of the virtual machine, the length of f , i.e., the number of local variables is unchanged. The program P , of course, never changes during execution. The transition relation is defined as follows.

On a New Method for Dataflow Analysis

23

– If P [i] = jsr(L), then hi, f, si → hL, f, retaddr(i + 1)::si. The return address retaddr(i + 1) is pushed onto the operand stack. The operator :: is the cons operator for lists. – If P [i] = ret(x) and f [x] = retaddr(j + 1), then hi, f, si → hj + 1, f, si. – If P [i] = load(x), then hi, f, si → hi + 1, f, f [x]::si. – If P [i] = store(x), then hi, f, v::si → hi + 1, f [x 7→v], si. The notation f [x 7→v] means a list whose element is the same as that of f except for the x-th element, which is set to v. – If P [i] = const0, then hi, f, si → hi + 1, intval(0)::si. – If P [i] = constNULL, then hi, f, si → hi + 1, objval(0)::si. – If P [i] = inc(x) and f [x] = intval(k), then hi, f, si → hi + 1, f [x 7→intval(k + 1)], si. – If P [i] = if0(L), then hi, f, intval(0)::si → hL, f, si. If P [i] = if0(L) and k 6= 0, then hi, f, intval(k)::si → hi + 1, f, si. – If P [i] = ifNULL(L), then hi, f, objval(0)::si → hL, f, si. If P [i] = ifNULL(L) and k 6= 0, then hi, f, objval(k)::si → hi + 1, f, si. The transition relation → is considered as the least relation satisfying the above conditions. The relation is defined so that when a type error occurs, no transition is defined. This means that to show the type safety of bytecode is to show that a transition sequence stops only at the halt instruction. For proving the correctness of our bytecode analysis, we also need another version of the operational semantics that maintains invocation histories of subroutines. This semantics corresponds to the structured dynamic semantics of Stata and Abadi. The transition relation is now defined for quadruples of the form hi, f, s, hi, where the last component h is an invocation history of subroutines. It is a list of addresses of callers of subroutines. This component is only changed by the jsr and ret instructions. – If P [i] = jsr(L), then hi, f, s, hi → hL, f, retaddr(i + 1)::s, i::hi. Note that the address i of the caller of the subroutine is pushed onto the invocation history. – If P [i] = ret(x), f [x] = retaddr(j + 1) and h = h0 @[j]@h00 , where j does not appear in h0 , then hi, f, s, hi → hj + 1, f, s, h00 i. The operator @ is the append operator for lists. For other instructions, the invocation histories before and after transition are the same. As for the two transition relations, we immediately have the following proposition. Proposition 1: If hi, f, s, hi → hi0 , f 0 , s0 , h0 i, then hi, f, si → hi0 , f 0 , s0 i.

4 4.1

Analysis Types

Types in our analysis are among the following syntactic entities:

24

M. Hagiya and A. Tozawa

>, ⊥ (top and bottom) return(n) (n: caller level)

INT, OBJ, · · · (basic types) last(x) (x: variable index)

A type is >, ⊥, a basic type, a return type, or a last type. In this paper, we assume as basic types INT, the type of integers, and OBJ, the type of object pointers. It is easy to add other basic types, such as that of floating point numbers. return types and last types are only meaningful inside a subroutine. A return type is the type of a return address. For positive integer n, return(n) denotes the type of the address for returning to the n-th outer caller. For example, return(1) denotes the type of the address for returning to the direct caller of the subroutine, and return(2) the type of the address for returning to the caller of the caller. A last type means that a value is passed from the caller of the subroutine. For nonnegative integer x, last(x) denotes the type of a value that was stored in the x-th local variable of the caller. A value can have this type only when it is exactly the same as the value of the x-th local variable when the subroutine was called. 4.2

Order among Types

We define the order among types as follows. > > INT > ⊥ > > return(n) > ⊥

> > OBJ > ⊥ > > last(x) > ⊥

Since we do not distinguish object pointers by their classes in this paper, the order is flat, with > and ⊥ as the top and bottom elements. This order is extended to lists of types. For type lists t1 and t2 , t1 > t2 holds if and only if t1 and t2 are of the same length and t1 [i] > t2 [i] holds for any i ranging over the indices for the lists. 4.3

Target of Analysis

The target of our bytecode analysis is to obtain the following pieces of information for the i-th instruction of the given program P . Fi

Si

Hi

Fi is a type list. Fi [x] describes the type of f [x], i.e., the value of the x-th local variable of the virtual machine. Si is a also type list. Each element of Si describes the type of the corresponding element of the operand stack of the virtual machine. Both Fi and Si describe the types of the components of the virtual machine just before the i-th instruction is executed. Hi is a set of invocation histories for the i-th instruction. F , S and H should follow a rule that is defined for each kind of P [i]. The rule says that certain conditions must be satisfied before and after the execution of P [i].

On a New Method for Dataflow Analysis

25

Rule for jsr) If h ∈ Hi and P [i] = jsr(L), then the following conditions must be satisfied. – 0 ≤ L < |P |. – For each variable index y, either FL [y] ≥ return(n + 1) (if Fi [y] = return(n)), and FL [y] ≥ Fi [y] (if Fi [y] is neither return nor last) or FL [y] ≥ last(y) (even if Fi [y] is last). – |SL | = |Si | + 1, where |l| denotes the length of list l. – SL [0] ≥ return(1). – For each index p, where 0 ≤ p < |Si |, Si [p] is not last, SL [p + 1] ≥ Si [p] (if Si [p] is not return), and SL [p + 1] ≥ return(n + 1) (if Si [p] = return(n)). – i does not appear in h. (Recursion is not allowed.) – i::h ∈ HL . Note that when Fi [y] is not last, FL [y] cannot be determined uniquely. We must make a nondeterministic choice between return(n + 1) and Fi [y]. See Sect. 5 for more discussions on the implementation of the analysis. The following figures show the two possibilities for typing local variables inside a subroutine. In this example, it is assumed that there is only one caller (2) of subroutine 7. The column [Fi [0], Fi [1], Fi [2]] shows the types of local variables before each instruction is executed. At subroutine 7, it is set to [>, INT, >] or [l(0), l(1), l(2)]. There are more possibilities. For example, one could also set it to [>, >, >], but this possibility is subsumed by the first. i 0 1 2 3 7 8 9 10

instruction const0 store(1) jsr(7) constNULL ··· store(0) load(1) store(2) ret(0)

[Fi [0], Fi [1], Fi [2]] [>, >, >] [>, >, >] [>, INT, >] [>, INT, INT]

Si [] [INT] [] []

Hi {[]} {[]} {[]} {[]}

[Fi [0], Fi [1], Fi [2]] [>, >, >] [>, >, >] [>, INT, >] [>, INT, INT]

Si [] [INT] [] []

Hi {[]} {[]} {[]} {[]}

[>, INT, >] [r(1), INT, >] [r(1), INT, >] [r(1), INT, INT]

[r(1)] [] [INT] []

{[2]} {[2]} {[2]} {[2]}

[l(0), l(1), l(2)] [r(1), l(1), l(2)] [r(1), l(1), l(2)] [r(1), l(1), l(1)]

[r(1)] [] [l(1)] []

{[2]} {[2]} {[2]} {[2]}

Rule for ret) If h ∈ Hi and P [i] = ret(x), then the following conditions must be satisfied. – Fi [x] = return(n).

26

M. Hagiya and A. Tozawa

– h = h0 @[j]@h00 , where |h0 | = n − 1. – 0 ≤ j + 1 < |P |. – For each variable index y, Fj+1 [y] ≥ follow last(n, h, Fi [y]). – Sj+1 ≥ follow last(n, h, Si ). – h00 ∈ Hj+1 . follow last is a function for extracting the type of a variable in a caller of a subroutine according to an invocation history. For nonnegative integer n, invocation history h and type t, follow last(n, h, t) is defined as follows. follow last(0, h, t) = t follow last(n + 1, i::h, return(m)) = if m > n + 1 then return(m − n − 1) else > follow last(n + 1, i::h, last(x)) = follow last(n, h, Fi [x]) follow last(n + 1, i::h, t) = t (otherwise) follow last is extended to type lists, i.e., follow last(n, h, t) is also defined when t is a type list. Rule for load) If h ∈ Hi and P [i] = load(x), then the following conditions must be satisfied. – 0 ≤ i + 1 < |P |. Fi+1 ≥ Fi . Si+1 ≥ Fi [x]::Si . h ∈ Hi+1 . Rule for store) If h ∈ Hi and P [i] = store(x), then the following conditions must be satisfied. – 0 ≤ i + 1 < |P |. Si = t::t. Fi+1 ≥ Fi [x 7→t]. Si+1 ≥ t. h ∈ Hi+1 . Rule for const0) If h ∈ Hi and P [i] = const0, then the following conditions must be satisfied. – 0 ≤ i + 1 < |P |. Fi+1 ≥ Fi . Si+1 ≥ INT ::Si . h ∈ Hi+1 . The rule for constNULL is similar. Rule for inc) If h ∈ Hi and P [i] = inc(x), then the following conditions must be satisfied. – 0 ≤ i + 1 < |P |. Fi [x] = INT. Fi+1 ≥ Fi . Si+1 ≥ Si . h ∈ Hi+1 . Rule for if0) If h ∈ Hi and P [i] = if0(L), then the following conditions must be satisfied. – 0 ≤ L < |P |. 0 ≤ i + 1 < |P |. – Si = INT ::t. FL ≥ Fi . Fi+1 ≥ Fi . SL ≥ t. Si+1 ≥ t. – h ∈ HL . h ∈ Hi+1 . The rule for ifNULL is similar. Rule for halt) There is no rule for halt.

On a New Method for Dataflow Analysis

4.4

27

Correctness of Analysis

In order to state the correctness of our analysis, we first introduce the following relation. hv, hi : t v is a value and h is an invocation history. t is a type. By hv, hi : t, we mean that the value v belongs to the type t provided that v appears with the invocation history h. Following is the definition of this relation. – – – – –

hv, hi : >. hintval(k), hi : INT. hobjval(k), hi : OBJ. If h[n − 1] = j, then hretaddr(j + 1), hi : return(n). If hv, hi : Fi [x], then hv, i::hi : last(x).

This definition is also inductive, i.e., hv, hi : t holds if and only if it can be derived only by the above rules. We have two lemmas. Lemma 1: If hv, hi : t and t0 ≥ t, then hv, hi : t0 . Lemma 2: Let h0 be a prefix of h of length n and h00 be its corresponding suffix, i.e., h = h0 @h00 and |h0 | = n. If hv, hi : t, then hv, h00 i : follow last(n, h, t). We say that the quadruple hi, f, s, hi is sound with respect to hF, S, Hi and write hi, f, s, hi : hF, S, Hi, if the following conditions are satisfied. – – – – –

0 ≤ i < |P |. For each variable index y, hf [y], hi : Fi [y]. For each index p for s, hs[p], hi : Si [p]. h ∈ Hi . h does not have duplication, i.e., no element of h occurs more than once in h.

We have the following correctness theorem. It says that if F , S and H follow the rule for each instruction of P , then the soundness is preserved under the transition of quadruples. This means that if the initial quadruple is sound, then quadruples that appear during execution of the virtual machine are always sound. Theorem (correctness of analysis): Assume that F , S and H follow the rule for each instruction of P . If hi, f, s, hi : hF, S, Hi and hi, f, s, hi → hi0 , f 0 , s0 , h0 i, then hi0 , f 0 , s0 , h0 i : hF, S, Hi. The theorem is proved by the case analysis on the kind of P [i]. In this short paper, we only examine the case when P [i] = ret(x). Assume that hi, f, s, hi : hF, S, Hi and hi, f, s, hi → hi0 , f 0 , s0 , h0 i. Since F , S and H follow the rule for ret, the following facts hold. (i) Fi [x] = return(n).

28

M. Hagiya and A. Tozawa

(ii) h = h1 @[j]@h2 , where |h1 | = n − 1. (iii) 0 ≤ j + 1 < |P |. (iv) For each variable index y, Fj+1 [y] ≥ follow last(n, h, Fi [y]). (v) Sj+1 ≥ follow last(n, h, Si ). (vi) h2 ∈ Hj+1 . By (i) and the soundness of hi, f, s, hi, hf [x], hi : return(n). Therefore, by (ii), f [x] = retaddr(j+1) and i0 = j+1. Moreover, since h does not have duplication, h1 does not contain j. This implies that h0 = h2 . We also have that f 0 = f and s0 = s. Let us check the conditions for the soundness of hi0 , f 0 , s0 , h0 i = hj+1, f, s, h2 i. – By (iii), 0 ≤ i0 < |P |. – By (iv), Fi0 [y] ≥ follow last(n, h, Fi [y]). By the soundness of hi, f, s, hi, hf [y], hi : Fi [y]. By Lemma 2, hf [y], h0 i : follow last(n, h, Fi [y]). Therefore, by Lemma 1, hf [y], h0 i : Fi0 [y]. – Similarly, by (v), we have that hs[p], h0 i : Si0 [p]. – By (vi) and since h0 = h2 , h0 ∈ Hi0 . – Finally, since h does not have duplication, h0 does not have duplication, either. Proposition 2: If hi, f, s, hi : hF, S, Hi and hi, f, si → hi0 , f 0 , s0 i, then there exists some h0 such that hi, f, s, hi → hi0 , f 0 , s0 , h0 i. The only case that must be examined is that of ret. Note that h0 is uniquely determined. The above proposition guarantees that if F , S and H follow the rule for each instruction and the initial quadruple hi, f, s, hi is sound, then the transition sequence starting from the triple hi, f, si can always be lifted to a sequence starting from hi, f, s, hi. This means that the semantics for triples and that for quadruples coincide when F , S and H follow the rule for each instruction. A similar lemma is stated in [16], which establishes a correspondence between their stackless semantics and their structured semantics. Lemma 3: If hi, f, s, hi : hF, S, Hi, then hi, f, s, hi has the next state unless P [i] = halt. The following final theorem, derived from the above lemma and the previous theorem, guarantees the type safety of bytecode. This corresponds to Theorem 1 (Soundness) in [16]. Theorem (type safety): If F , S and H follow the rule for each instruction and the initial quadruple hi, f, s, hi is sound, then a transition sequence stops only at the halt instruction.

On a New Method for Dataflow Analysis

4.5

29

Example

Let us abbreviate last(x) and return(n) by l(x) and r(n), respectively. Following is the result of analyzing the example in Sect. 2. i 0 1 2 3 4 5 6 7 8 9 10

instruction const0 store(1) jsr(7) constNULL store(1) jsr(7) halt store(0) load(1) store(2) ret(0)

[Fi [0], Fi [1], Fi [2]] [>, >, >] [>, >, >] [>, INT, >] [>, INT, INT] [>, INT, INT] [>, OBJ, INT] [>, OBJ, OBJ] [l(0), l(1), l(2)] [r(1), l(1), l(2)] [r(1), l(1), l(2)] [r(1), l(1), l(1)]

Si [] [INT] [] [] [OBJ] [] [] [r(1)] [] [l(1)] []

Hi {[]} {[]} {[]} {[]} {[]} {[]} {[]} {[2], [5]} {[2], [5]} {[2], [5]} {[2], [5]}

The rule for ret(0) at 10 is satisfied because, for [2] ∈ H10 , – F10 [0] = return(1). [2] = []@[2]@[]. 0 ≤ 2+1 = 3 < 11. – F3 [0] = > ≥ follow last(1, [2], F10 [0]) = follow last(1, [2], return(1)) = >. – F3 [1] = INT ≥ follow last(1, [2], F10 [1]) = follow last(1, [2], last(1)) = F2 [1] = INT. – F3 [2] = INT ≥ follow last(1, [2], F10 [2]) = follow last(1, [2], last(1)) = F2 [1] = INT. – S3 = []. [] ∈ {[]} = H3 , and similarly for [5] ∈ H10 . 4.6

Returning to an Outer Caller

When P [i] = ret(x) returns to the caller of the caller, for example, Fi [x] must be equal to the type return(2). In this case, Hi should consist of histories of length at least 2. If [j1 , j2 , j3 , · · ·] is in Hi , P [i] returns to i0 = j2 +1 and Fi0 should satisfy Fi0 [y] ≥ follow last(2, [j1 , j2 , j3 , · · ·], Fi [y]). If Fi [y] = last(y) and Fj1 = last(y), for example, then follow last(2, [j1 , j2 , j3 , · · ·], Fi [y]) = Fj2 [y]. This is how information at j2 (which is a jsr) is propagated to i0 = j2 +1. If Fi [y] is not last, information at i is propagated.

30

5

M. Hagiya and A. Tozawa

Implementation

A dataflow analysis is usually implemented by an iterative algorithm. For each instruction, we check if the rule for the instruction is satisfied by F , S and H. If not, we update F , S and H accordingly, and check the next instruction that is affected by the update. There are two problems for implementing our analysis by such an iterative algorithm. Firstly, the rule for jsr does not uniquely determine FL [y] when Fi [y] is not last. We have two choices: one is to set FL [y] = last(y), and the other is to set FL [y] = Fi [y] (or FL [y] = return(n + 1) if Fi [y] = return(n)). In our current implementation, we first set FL [y] = last(y) and proceed the analysis. If the analysis fails at some point because FL [y] = last(y), we take the alternative and redo the analysis from L. (We need not completely abandon the work after we set FL [y] = last(y).) The second problem is that by a na¨ıve iterative algorithm, a subroutine is analyzed each time it is called. In the worst case this may require an exponential number of steps with respect to the length of the program. This problem can be avoided by representing invocation histories by a node in the call graph of the program, which is a graph whose nodes are addresses of subroutines and whose (directed) edges are labeled with addresses of jsr instructions. Since the JVM does not allow recursion, it is a connected acyclic graph with a unique root node representing the initial address. A path from the root to a node in the graph corresponds to an invocation history by concatenating the labels of edges in the path. Each node in the graph then represents the set of all the invocation histories from the root to the node. Now, instead of keeping a set of invocation histories (i.e., Hi ), we can keep a set of nodes in the graph. From a program in the following (left), the call graph in the right is constructed. The node L3 represents the set {[c, b, a], [c, b0 , a0 ], [c0 , b, a], [c0 , b0 , a0 ]} of invocation histories. a : jsr(L1 ) ... a0 : jsr(L01 ) ... L1 : . . . b : jsr(L2 ) ... L01 : . . . b0 : jsr(L2 ) ... L2 : . . . c : jsr(L3 ) ... c0 : jsr(L3 ) ... L3 : . . .

L1

L3

HH HH Y c0

@ I @a @ @

b

c

L2

@ I @0 [email protected] @

a0 L01

On a New Method for Dataflow Analysis

31

If histories are represented by nodes in the call graph, then the values that Hi can take are bounded by the set of all the nodes in the call graph. This means that Hi can only be updated for the number of times equal to the number of nodes. The number of overall updates is, therefore, limited by n2 , where n is the number of instructions in the program. In order to achieve a polinomial complexity of the entire analysis, however, the nondeterministic choice in the handling of jsr instructions must be restricted.

6

Concluding Remark

Since we introduced types of the form last(x), it has become possible to assign types to polymorphic subroutines that move a value from a variable to another. Our analysis is towards the real polymorphism of subroutines in binary code, because we do not simply ignore unaccessed variables. For some programs, our method is more powerful than existing ones. In particular, we impose no restrictions on the number of entries and exits of a subroutine. It is also important that the proof of the correctness of bytecode verification becomes extremely simple. We only formalized a very small subset of the JVM. We believe that the framework of the paper can be extended to the full language. The extension is almost straight forward. In particular, adding new kinds of types seems to cause no difficulty. The correct handling of exceptions and that of object initialization are problematic but are not impossible [13,5]. The resulting bytecode verifier is expected to be more powerful than the existing one, so the Java compiler will gain more freedom in bytecode generation. It is also interesting whether the framework can be applied to the analysis of other kinds of bytecode or native code. Handling recursive calls is the key to such applications. In order to allow recursion, we must be able to represent histories of an indefinite length by a kind of regular expression. Stacks generated by recursive calls should also be represented by regular expressions. All this is left as future work. A dataflow analysis, in general, assigns an abstract value xi to the i-th instruction so that a certain predicate P (xi , σ) always holds for any state σ that reaches the i-th instruction. To this end, for any transition σ → σ 0 , where σ 0 is at i0 , one must show that P (xi , σ) implies P (xi0 , σ 0 ). Since σ corresponds to hi, f, s, hi in our analysis, xi seems to correspond to hFi , Si , Hi i. However, the predicate hi, f, s, hi : hF, S, Hi, which should correspond to P (xi , σ), does not only refer to hFi , Si , Hi i. When Fi [y] is last, it also refers to Fh[0] . This means that in terms of last types, our analysis relates values assigned to different instructions. This makes the analysis powerful while keeping the overall data structure for the analysis compact. The representation of invocation histories by a node in the call graph is also for making the data structure small and efficient. By this representation, the number of updates of Hi is limited by the size of the call graph, and an iterative algorithm is expected to stop in polynomial time with respect to the program size. This kind of complexity analysis should be made rigorous in the future.

32

M. Hagiya and A. Tozawa

Acknowledgments. The author would like to thank Zhenyu Qian and Mart´ın Abadi for their comments on the earlier draft of this paper. He also thanks anonymous referees whose comments greatly improved the paper.

References 1. Richard M. Cohen: The Defensive Java Virtual Machine Specification, Version Alpha 1 Release, DRAFT VERSION, 1997. http://www.cli.com/software/djvm/html-0.5/djvm-report.html 2. Drew Dean: The Security of Static Typing with Dynamic Linking, Fourth ACM Conference on Computer and Communication Security, 1997, pp.18–27. http://www.cs.princeton.edu/sip/pub/ccs4.html 3. Sophia Drossopoulou and Susan Eisenbach: Java is Type Safe — Probably, ECOOP’97 — Object-Oriented Programming, Lecture Notes in Computer Science, Vol.1241, 1997, pp.389–418. http://outoften.doc.ic.ac.uk/projects/slurp/papers.html\#ecoop 4. Sophia Drossopoulou, Susan Eisenbach and Sarfraz Khurshid: Is the Java Type System Sound? Proceedings of the Fourth International Workshop on Foundations of Object-Oriented Languages, 1997. http://outoften.doc.ic.ac.uk/projects/slurp/papers.html\#tapos 5. Allen Goldberg: A Specification of Java Loading and Bytecode Verification, 1997. http://www.kestrel.edu/˜goldberg/ 6. James Gosling, Bill Joy and Guy Steele: The Java TM Language Specification, Addison-Weslay, 1996. 7. Kimera: http://kimera.cs.washington.edu/ 8. Tim Lindholm and Frank Yellin: The Java TM Virtual Machine Specification, Addison-Weslay, 1997. 9. Gary McGraw and Edward W. Felten: Java Security: Hostile Applets, Holes and Antidotes, John Wiley and Sons, 1996. 10. George C. Necula: Proof-Carrying Code, the Proceedings of the 24th Annual SIGPLAN-SIGACT Symposium on Principles of Programming Languages, 1997, pp.106–117. 11. George C. Necula, Peter Lee: The Design and Implementation of a Certifying Compiler, submitted to PLDI’98. 12. Tobias Nipkow and David von Oheimb: Javalight is Type-Safe — Definitely, Proceedings of the 25th Annual SIGPLAN-SIGACT Symposium on Principles of Programming Languages, 1998, pp.161–170. 13. Zhenyu Qian: A Formal Specification of JavaTM Virtual Machine Instructions, 1997. http://www.informatik.uni-bremen.de/˜qian/abs-fsjvm.html 14. Vijay Saraswat: Java is not type-safe, 1997. http://www.research.att.com/˜vj/bug.html 15. Secure Internet Programming: http://www.cs.princeton/edu/sip/ 16. Raymie Stata and Mart´ın Abadi: A Type System for Java Bytecode Subroutines, Proceedings of the 25th Annual SIGPLAN-SIGACT Symposium on Principles of Programming Languages, 1998, pp.149–160. 17. Don Syme: Proving Java Type Soundness, 1997. http://www.cl.cam.ac.uk/users/drs1004/java.ps

Enabling Sparse Constant Propagation of Array Elements via Array SSA Form Vivek Sarkar1 and Kathleen Knobe2 1

IBM Thomas J. Watson Research Center P.O. Box 704, Yorktown Heights, NY 10598, USA [email protected] 2 Compaq Cambridge Research Laboratory One Kendall Square, Building 700, Cambridge, MA 02139, USA [email protected]

Abstract. We present a new static analysis technique based on Array SSA form [6]. Compared to traditional SSA form, the key enhancement in Array SSA form is that it deals with arrays at the element level instead of as monolithic objects. In addition, Array SSA form improves the φ function used for merging scalar or array variables in traditional SSA form. The computation of a φ function in traditional SSA form depends on the program’s control flow in addition to the arguments of the φ function. Our improved φ function (referred to as a Φ function) includes the relevant control flow information explicitly as arguments through auxiliary variables that are called @ variables. The @ variables and Φ functions were originally introduced as run-time computations in Array SSA form. In this paper, we use the elementlevel Φ functions in Array SSA form for enhanced static analysis. We use Array SSA form to extend past algorithms for Sparse Constant propagation (SC) and Sparse Conditional Constant propagation (SCC) by enabling constant propagation through array elements. In addition, our formulation of array constant propagation as a set of data flow equations enables integration with other analysis algorithms that are based on data flow equations. Keywords: static single assignment (SSA) form, constant propagation, conditional constant propagation, Array SSA form, unreachable code elimination.

G. Levi (Ed.): SAS’98, LNCS 1503, pp. 33–56, 1998. c Springer-Verlag Berlin Heidelberg 1998

34

1

V. Sarkar and K. Knobe

Introduction

The problems of constant propagation and conditional constant propagation (a combination of constant propagation and unreachable code elimination) have been studied for several years. However, past algorithms limited their attention to constant propagation of scalar variables only. In this paper, we introduce efficient new algorithms that perform constant propagation and conditional constant propagation through both scalar and array references. One motivation for constant propagation of array variables is in optimization of scientific programs in which certain array elements can be identified as constant. For example, the SPEC95fp [3] benchmark 107.mgrid contains an array variable A that is initialized to four constant-valued elements as shown in Fig. 1. A significant amount of the time in this application is spent in the triply nested loop shown at the bottom of Fig. 1. Since constant propagation can determine that A(2) equals zero in the loop, an effective optimization is to eliminate the entire multiplicand of A(2) in the loop nest. Doing so eliminates 11 of the 23 floating-point additions in the loop nest thus leading to a significant speedup. Another motivation is in analysis and optimization of field accesses of structure variables or objects in object-oriented languages such as Java and C++. A structure can be viewed as a fixed-size array, and a read/write operation of a structure field can be viewed as a read/write operation of an array element through a subscript that is a compile-time constant. This approach is more compact than an approach in which each field of a structure is modeled as a separate scalar variable. This technique for modeling structures as arrays directly extends to nested arrays and structures. For example, an array of rank n of some structure type can be modeled as an array of rank n + 1. Therefore, the constant propagation algorithms presented in this paper can be efficiently applied to structure variables and to arrays of structures. Extending these algorithms to analyze programs containing pointer aliasing is a subject for future research, however. The best known algorithms for sparse constant propagation of scalar variables [8,2] are based on static single assignment (SSA) form [4]. However, traditional SSA form views arrays as monolithic objects, which is an inadequate view for analyzing and optimizing programs that contain reads and writes of individual array elements. In past work, we introduced Array SSA form [6] to address this deficiency. The primary application of Array SSA form in [6] was to enable parallelization of loops not previously parallelizable by making Array SSA form manifest at run-time. In this paper, we use Array SSA form as a basis for static analysis, which means that the Array SSA form structures can be removed after the program properties of interest have been discovered. Array SSA form has two distinct advantages over traditional SSA form. First, the φ operator in traditional SSA form is not a pure function and returns different values for the same arguments depending on the control flow path that was taken. In contrast, the corresponding Φ operator in Array SSA form includes @ variables as extra arguments to capture the control information required i.e., x3 := φ(x2 , x1 ) in traditional SSA form becomes x3 := Φ(x2 , @x2 , x1 , @x1 ) in

Enabling Sparse Constant Propagation of Array Elements

35

! Initialization of array A REAL*8 A(0:3) . . . A(0) = -8.0D0/3.0D0 A(1) = 0.0D0 A(2) = 1.0D0/6.0D0 A(3) = 1.0D0/12.0D0 . . . ! Computation loop in subroutine RESID() do i3 = 2, n-1 do i2 = 2, n-1 do i1 = 2, n-1 R(i1,i2,i3)=V(i1,i2,i3) -A(0)*( U(i1, i2, i3 ) ) -A(1)*( U(i1-1,i2, i3 ) + U(i1+1,i2, i3 ) + U(i1, i2-1,i3 ) + U(i1, i2+1,i3 ) + U(i1, i2, i3-1) + U(i1, i2, i3+1) ) -A(2)*( U(i1-1,i2-1,i3 ) + U(i1+1,i2-1,i3 ) + U(i1-1,i2+1,i3 ) + U(i1+1,i2+1,i3 ) + U(i1, i2-1,i3-1) + U(i1, i2+1,i3-1) + U(i1, i2-1,i3+1) + U(i1, i2+1,i3+1) + U(i1-1,i2, i3-1) + U(i1-1,i2, i3+1) + U(i1+1,i2, i3-1) + U(i1+1,i2, i3+1) ) -A(3)*( U(i1-1,i2-1,i3-1) + U(i1+1,i2-1,i3-1) + U(i1-1,i2+1,i3-1) + U(i1+1,i2+1,i3-1) + U(i1-1,i2-1,i3+1) + U(i1+1,i2-1,i3+1) + U(i1-1,i2+1,i3+1) + U(i1+1,i2+1,i3+1) ) end do end do end do Fig. 1. Code fragments from the SPEC95fp 107.mgrid benchmark

Array SSA form. Second, Array SSA form operates on arrays at the element level rather than as monolithic objects. In particular, a Φ operator in Array SSA form, A3 := Φ(A2 , @A2 , A1 , @A1 ), represents an element-level merge of A2 and A1 . Both advantages of Array SSA form are significant for static analysis. The fact that Array SSA form operates at the element-level facilitates transfer of statically derived information across references to array elements. The fact that the Φ is a known pure function facilitates optimization and simplification of the Φ operations. For convenience, we assume that all array operations in the input program are expressed as reads and writes of individual array elements. The extension to more complex array operations (e.g., as in Fortran 90 array language) is

36

V. Sarkar and K. Knobe

straightforward, and omitted for the sake of brevity. Also, for simplicity, we will restrict constant propagation of array variables to cases in which both the subscript and the value of an array definition are constant e.g., our algorithm mights recognize that a definition A[k] := i is really A[2] := 99 and propagate this constant into a use of A[2]. The algorithms presented in this paper will not consider a definition to be constant if its subscript has a non-constant value e.g., A[m] := 99 where m is not a constant. Performing constant propagation for such references (e.g., propagating 99 into a use of A[m] when legal to do so) is a subject of future work. The rest of the paper is organized as follows. Section 2 reviews the Array SSA form introduced in [6]. Section 3 presents our extension to the Sparse Constant propagation (SC) algorithm from [8] that enables constant propagation through array elements. It describes how lattice values can be computed for array variables and Φ functions. Section 4 presents our extension to the Sparse Conditional Constant propagation (SCC) algorithm from [8] that enables constant propagation through array elements in conjunction with unreachable code elimination. Section 5 discusses related work, and Sect. 6 contains our conclusions.

2

Array SSA Form

In this section, we describe the Array SSA form introduced in [6]. The goal of Array SSA form is to provide the same benefits for arrays that traditional SSA provides for scalars but, as we will see, it has advantages over traditional SSA form for scalars as well. We first describe its use for scalar variables and then its use for array variables. The salient properties of traditional SSA form are as follows: 1. Each definition is assigned a unique name. 2. At certain points in the program, new names are generated which combine the results from several definitions. This combining is performed by a φ function which determines which of several values to use, based on the flow path traversed. 3. Each use refers to exactly one name generated from either of the two rules above. For example, traditional SSA form converts the code in Fig. 2 to that in Fig. 3. The S3 := φ(S1 , S2 ) statement defines S3 as a new name that represents the merge of definitions S1 and S2 . It is important to note that the φ function in traditional SSA form is not a pure function of S1 and S2 because its value depends on the path taken through the if statement. Notice that this path is unknown until runtime and may vary with each dynamic execution of this code. In contrast to traditional SSA form, the semantics of a Φ function is defined to be a pure function in our Array SSA form. This is accomplished by introducing @ variables (pronounced “at variables”), and by rewriting a φ function in traditional SSA form such as φ(S1 , S2 ) as a new kind of Φ function,

Enabling Sparse Constant Propagation of Array Elements

37

Φ(S1 , @S1 , S2 , @S2 ). For each static definition Sk , its @ variable @Sk identifies the most recent “time” at which Sk was modified by this definition. For an acyclic control flow graph, a static definition Sk may execute either zero times or one time. These two cases can be simply encoded as @Sk = false and @Sk = true to indicate whether or not definition Sk was executed. For a control flow graph with cycles (loops), a static definition Sk may execute an arbitrary number of times. In general, we need more detailed information for the @Sk = true case so as to distinguish among different dynamic execution instances of static definition Sk . Therefore, @Sk is set to contain the dynamic iteration vector at which the static definition Sk was last executed. The iteration vector of a static definition Sk identifies a single iteration in the iteration space of the set of loops that enclose the definition. Let n be the number of loops that enclose a given definition. For convenience, we treat the outermost region of acyclic control flow in a procedure as a dummy outermost loop with a single iteration. Therefore n ≥ 1 for each definition. A single point in the iteration space is specified by the iteration vector i = (i1 , . . . , in ), which is an n-tuple of iteration numbers one for each enclosing loop. We do not require that the surrounding loops be structured counted loops (i.e., like Fortran do loops) or that the surrounding loops be tightly nested. Our only assumption is that all loops are single-entry, or equivalently, that the control flow graph is reducible [5, 1]. For single-entry loops, we know that each def executes at most once in a given iteration of its surrounding loops. All structured loops (e.g., do, while, repeat-until) are single-entry even when they contain multiple exits; also, most unstructured loops (built out of goto statements) found in real programs are single-entry as well. A multiple-entry loop can be transformed into multiple single-entry loops by node splitting [5,1]. Array SSA form can be used either at run-time as discussed in [6] or for static analysis, as in the constant propagation algorithms presented in this paper. In this section, we explain the meaning of @ variables as if they are computed at run-time. We assume that all @ variables, @Sk , are initialized to the empty vector, @Sk := ( ), at the start of program execution. For each real (non-Φ) definition, Sk , we assume that a statement of the form @Sk := i is inserted immediately after definition Sk 1 , where i is the current iteration vector for all loops that surround Sk . Each Φ definition also has an associated @ variable. Its semantics will be defined shortly. All @ variables are initialized to the empty vector because the empty vector is the identity element for a lexicographic max operation i.e., max(( ), i) = i, for any @ variable value i. As a simple example, Fig. 4 shows the Array SSA form for the program in Fig. 2. Note that @ variables @S1 and @S2 are explicit arguments of the Φ function. In this example of acyclic code, there are only two possible values 1

It may appear that the @ variables do not satisfy the static single assignment property because each @Sk variable has two static definitions, one in the initialization and one at the real definition of Sk . However, the initialization def is executed only once at the start of program execution and can be treated as a special-case initial value rather than as a separate definition.

38

V. Sarkar and K. Knobe

for each @ variable — the empty vector ( ) and the unit vector (1) — which correspond to false and true respectively. Figure 5 shows an example for-loop and its conversion to Array SSA form. Because of the presence of a loop, the set of possible values for an @ variable becomes unbounded e.g., we may have @S1 = (100) on exit from the loop. However, @S1 and @S2 are still explicit arguments of the Φ function, and their iteration vector values are necessary for evaluating the Φ function at run-time. The semantics of a Φ function can now be specified by a conditional expression that is a pure function of the arguments of the Φ. For example, the semantics of the Φ function, S3 := Φ(S2 , @S2 , S1 , @S1 ) in Fig. 5, can be expressed as a conditional expression as follows (where denotes a lexicographic greater-thanor-equal comparison of iteration vectors): if @S2 @S1 then S2 S3 = else S1 end if Following each Φ def, S3 := Φ(S2 , @S2 , S1 , @S1 ), there is the definition of the associated @ variable, @S3 = max(@S2 , @S1 ), where max represents a lexicographic maximum operation of iteration vector values @S2 , @S1 . Consider, for example, a condition C in Fig. 5 that checks if the value of i is even. In this case, definition S1 is executed in every iteration and definition S2 is executed only in iterations 2, 4, 6, . . . . For this “even-value” branch condition, the final values of @S1 and @S2 are both equal to (100) if m = 100. Since these values satisfy the condition @S2 @S1 , the conditional expression will yield S 3 = S2 . Consider another execution of the for-loop in Fig. 5 in which condition C evaluates to false in each iteration of the for loop. For this execution, the final values of @S2 and @S1 will be the empty vector ( ) and (100) respectively. Therefore, S2 ≺ S1 , and the conditional expression for the Φ function will yield S3 = S1 for this execution. The above description outlines how @ variables and Φ functions can be computed at run-time. However, if Array SSA form is used for static analysis, then no run-time overhead is incurred due to the @ variables and Φ functions. Instead, the @ variables and Φ functions are inserted in the compiler intermediate representation prior to analysis, and then removed after the program properties of interest have been discovered by static analysis. We now describe Array SSA form for array variables. Figure 6 shows an example program with an array variable, and the conversion of the program to Array SSA form as defined in [6]. The key differences between Array SSA form for array variables and Array SSA form for scalar variables are as follows: 1. Array-valued @ variables: The @ variable is an array of the same shape as the array variable with which it is associated, and each element of an @ array is initialized to the empty vector. For example, the statement @A1 [k1 ] := (1) is inserted after

Enabling Sparse Constant Propagation of Array Elements

if (C) then S := . . . else S := . . . end if

Fig. 2. Control Flow with Scalar Definitions

if (C) then S1 := . . . else S2 := . . . end if S3 := φ(S1 , S2 )

Fig. 3. Traditional SSA form

@S1 := ( ) @S2 := ( ) if (C) then S1 := . . . @S1 := (1) else S2 := . . . @S2 := (1) end if S3 = Φ(S1 , @S1 , S2 , @S2 )

Fig. 4. After conversion of Fig. 2 to Array SSA form

39

40

V. Sarkar and K. Knobe

Example for-loop: S := . . . for i := 1 to m do S := . . . if (C) then S := . . . end if end for

After conversion to Array SSA form: @S1 := ( ) ; @S2 := ( ) S := . . . @S := (1) for i := 1 to m do S0 := Φ(S3 , @S3 , S, @S) @S0 := max(@S3 , @S) S1 := . . . @S1 := (i) if (C) then S2 := . . . @S2 := (i) end if S3 := Φ(S2 , @S2 , S1 , @S1 ) @S3 := max(@S2 , @S2 ) end for

Fig. 5. A for-loop and its conversion to Array SSA form

Enabling Sparse Constant Propagation of Array Elements

41

Example program with array variables: n1:

n2:

n3:

A[∗] := initial value of A i := 1 C := i < n if C then k := 2 ∗ i A[k] := i print A[k] endif print A[2]

After conversion to Array SSA form: n1:

n2:

n4:

@i := ( ) ; @C := ( ) ; @k := ( ) ; @A0 [∗] := ( ) ; @A1 [∗] := ( ) A0 [∗] := initial value of A @A0 [∗] := (1) i := 1 @i := (1) C := i < n @C := (1) if C then k := 2 ∗ i @k := (1) A1 [k] := i @A1 [k] := (1) A2 := dΦ(A1 , @A1 , A0 , @A0 ) @A2 := max(@A1 , @A0 ) print A2 [k] endif A3 := Φ(A2 , @A2 , A0 , @A0 ) @A3 := max(@A2 , @A0 ) print A3 [2]

Fig. 6. Example program with an array variable, and its conversion to Array SSA form

42

V. Sarkar and K. Knobe

statement A1 [k1 ] := i in Fig. 6. In general, @A1 can record a separate iteration vector for each element that is assigned by definition A1 . This initialization is only required for @ arrays corresponding to real (non-Φ) definitions. No initialization is required for an @ array for a Φ definition (such as @A2 and @A3 in Fig. 6) because its value is completely determined by other @ arrays. 2. Array-valued Φ functions: A Φ function for array variables returns an array value. For example, consider the Φ definition A3 := Φ(A2 , @A2 , A0 , @A0 ) in Fig. 6 which represents a merge of arrays A2 and A0 . The semantics of the Φ function is specified by the following conditional expression for each element, A3 [j]: if @A2 [j] @A0 [j] then A2 [j] A3 [j] = else A0 [j] end if Note that this conditional expression uses a lexicographic comparison () of @ values just as in the scalar case. 3. Definition Φ’s: The traditional placement of the Φ is at control merge points. We refer to this as a control Φ. A special new kind of Φ function is inserted immediately after each original program definition of an array variable that does not completely kill the array value. This definition Φ merges the value of the element modified in the definition with the values available immediately prior to the definition. Definition Φ’s did not need to be inserted for definitions of scalar variables because a scalar definition completely kills the old value of the variable. We will use the notation dΦ when we want to distinguish a definition Φ function from a control Φ function. For example, consider definition A1 in Fig. 6. The dΦ function, A2 := dΦ(A1 , @A1 , A0 , @A0 ) is inserted immediately after the def of A1 to represent an element-by-element merge of A1 and A0 . Any subsequent use of the original program variable A (before an intervening def) will now refer to A2 instead of A1 . The semantics of the dΦ function is specified by the following conditional expression for each element, A2 [j]: if @A1 [j] @A0 [j] then A1 [j] A2 [j] = else A0 [j] end if Note that this conditional expression is identical in structure to the conditional expression for the control Φ function in item 2 above.

3

Sparse Constant Propagation for Scalars and Array Elements

We now present our extension to the Sparse Constant propagation (SC) algorithm from [8] that enables constant propagation through array elements.

Enabling Sparse Constant Propagation of Array Elements

43

Section 3.1 contains our definitions of lattice elements for scalar and array variables; the main extension in this section is our modeling of lattice elements for array variables. Section 3.2 outlines our sparse constant propagation algorithm; the main extension in this section is the use of the definition Φ operator in Array SSA form to perform constant propagation through array elements. 3.1

Lattice Values for Scalar and Array Variables

Recall that a lattice consists of: – L, a set of lattice elements. A lattice element for a program variable v is written as L(v), and denotes set(L(v)) = a set of possible values for variable v. – > (“top”) and ⊥ (“bottom”), two distinguished elements of L. – A meet (or join) operator, u, such that for any lattice element e, e u > = e and e u ⊥ = ⊥. – A w operator such that e w f if and only if e u f = f , and a A operator such that e A f if and only if e w f and e 6= f . The height H of lattice L is the length of the largest sequence of lattice elements e1 , e2 , . . . , eH such that ei A ei+1 for all 1 ≤ i < H. We use the same approach as in [8] for modeling lattice elements for scalar variables. Given a scalar variable S, the value of L(S) in our framework can be >, Constant or ⊥ . When the value is Constant we also maintain the value of the constant. The sets denoted by these lattice elements are set(>) = { }, set(Constant) = {Constant}, and set(⊥) = U S , where U S is the universal set of values for variable S. We now describe how lattice elements for array variables are represented in A our framework. Let U A ind and U elem be the universal set of index values and the universal set of array element values respectively for an array variable A in Array SSA form. For an array variable, the set denoted by lattice element L(A) A is a subset of U A ind × U elem i.e., a set of index-element pairs. Since we restrict constant propagation of array variables to cases in which both the subscript and the value of an array definition are constant, there are only three kinds of lattice elements of interest: 1. L(A) = > ⇒ set(L(A)) = { } This “top” case means that the possible values of A have yet to be determined i.e., the set of possible index-element pairs that have been identified thus far for A is the empty set, { }. 2. L(A) = h(i1 , e1 ), (i2 , e2 ), . . . i A ⇒ set(L(A)) = {(i1 , e1 ), (i2 , e2 ), . . . } ∪ (U A ind − {i1 , i2 , . . . }) × U elem In general, the lattice value for this “constant” case is represented by a finite ordered list of index-element pairs, h(i1 , e1 ), (i2 , e2 ), . . . i where i1 , e1 , i2 , e2 , . . . are all constant. The list is sorted in ascending order of the index values, i1 , i2 , . . . , and all the index values assumed to be distinct.

44

V. Sarkar and K. Knobe

. . .

< (1,100), (2,101) >

. . .

< (1,100) >

. . .

. . .

< (2,101) >

< (2,101), (3,102) >

. . .

< (3,102) >

. . .

. . .

Fig. 7. Lattice elements of array values with maximum list size Z = 2

The meaning of this “constant” lattice value is that the current stage of analysis has determined some finite number of constant index-element pairs for array variable A, such that A[i1 ] = e1 , A[i2 ] = e2 , . . . . All other elements of A are assumed to be non-constant. These properties are captured by set(L(A)) defined above as the set denoted by lattice value L(A). For the sake of efficiency, we will restrict these constant lattice values to ordered lists that are bounded in size by a small constant, Z ≥ 1 e.g., if Z = 5 then all constant lattice values will have ≤ 5 index-element pairs. Doing so ensures that the height of the lattice for array values is at most (Z + 2). If any data flow equation yields a lattice value with P > Z pairs, then this size constraint is obeyed by conservatively dropping any (P − Z) index-element pairs from the ordered list. Note that the lattice value for a real (non-Φ) definition, will contain at most one index-element pair, since we assumed that an array assignment only modifies a single element. Ordered lists with size > 1 can only appear as the output of Φ functions. A 3. L(A) = ⊥ ⇒ set(L(A)) = U A ind × U elem This “bottom” case means that, according to the approximation in the current stage of analysis, array A may take on any value from the universal set of index-element pairs. Note that L(A) = ⊥ is equivalent to an empty ordered list, L(A) = h i.

The lattice ordering (A) for these elements is determined by the subset relationship among the sets that they denote. The lattice structure for the Z = 2 case is shown in Fig. 7. This lattice has four levels. The second level (just below >) contains all possible ordered lists that contain exactly two constant index-element pairs. The third level (just above ⊥) contains all possible ordered lists that contain a single constant index-element pair. The lattice ordering is determined by the subset relationship among the sets denoted by lattice elements. For example, consider two lattice elements L1 = h(1, 100), (2, 101)i

Enabling Sparse Constant Propagation of Array Elements

45

and L2 = h(2, 101)i. The sets denoted by these lattice elements are: set(L1 ) = {(1, 100), (2, 101)} ∪ (U ind − {1, 2}) × U elem set(L2 ) = {(2, 101)} ∪ (U ind − {2}) × U elem Therefore, set(L1 ) is a proper subset of set(L2 ) and we have L1 A L2 i.e., L1 is above L2 in the lattice in Fig. 7. Finally, the meet operator (u) for two lattice elements, L1 and L2 , for array variables is defined in Fig. 8 where L1 ∩ L2 denotes an intersection of ordered lists L1 and L2 . L3 L1 L1 L1

= L1 u L2 L2 = > L2 = h(i1 , e1 ), . . . i L2 = ⊥ => > L2 ⊥ = h(i01 , e01 ), . . . i L1 L1 ∩ L 2 ⊥ =⊥ ⊥ ⊥ ⊥

Fig. 8. Lattice computation for the meet operator, L3 = L1 u L2

3.2

The Algorithm

Recall that the @ variables defined in Sect. 2 were necessary for defining the full execution semantics of Array SSA form. For example, the semantics of a Φ operator, A2 := Φ(A1 , @A1 , A0 , @A0 ), is defined by the following conditional expression: if @A1 [j] @A0 [j] then A1 [j] A2 [j] = else A0 [j] end if The sparse constant propagation algorithm presented in this section is a static analysis that is based on conservative assumptions about runtime behavior. Let us first consider the case when the above Φ operator is a control Φ. Since algorithm in this section does not perform conditional constant propagation, the lattice computation of a control Φ can be simply defined as L(A2 ) = L(Φ(A1 , @A1 , A0 , @A0 ) = L(A1 ) u L(A0 ) i.e., as a join of the lattice values L(A1 ) and L(A0 ). Therefore, the lattice computation for A2 does not depend on @ variables @A1 and @A0 for a control Φ operator. Now, consider the case when the above Φ operator is a definition Φ. The lattice computation for a definition Φ is shown in Fig. 9. Since A1 corresponds to a definition of a single array element, the ordered list for L(A1 ) can contain at most one pair. The insert operation in Fig. 9 is assumed to return a new ordered list obtained by inserting (i0 , e0 ) into h(i1 , e1 ), . . . i with the following adjustments if needed:

46

V. Sarkar and K. Knobe

– If there exists an index-element pair (ij , ej ) in h(i1 , e1 ), . . . i such that i0 = ij , then the insert operation just replaces (ij , ej ) by (i0 , e0 ). – If the insert operation causes the size of the list to exceed the threshold size Z, then one of the pairs is dropped from the output list so as to satisfy the size constraint. Interestingly, we again do not need @ variables for the lattice computation in Fig. 9. This is because the ordered list representation for array lattice values already contains all the subscript information of interest, and overlaps with the information that would have been provided by @ variables. L(A2 ) L(A0 ) = > L(A0 ) = h(i1 , e1 ), . . . i L(A0 ) = ⊥ L(A1 ) = > > > > L(A1 ) = h(i0 , e0 )i > insert((i0 , e0 ), h(i1 , e1 ), . . . i) h(i0 , e0 )i L(A1 ) = ⊥ ⊥ ⊥ ⊥ Fig. 9. Lattice computation for L(A2 ) = LdΦ (L(A1 ), L(A0 ))

Therefore, we do not need to analyze @ variables for the sparse constant propagation algorithm described in this section, because of our ordered list representation of lattice values for array variables. Instead, we can use a partial Array SSA form which is simply Array SSA form with all definitions and uses of @ variables removed, and with φ operators instead of Φ operators. If only constant propagation is being performed, then it would be more efficient to only build the partial Array SSA form. However, if other optimizations are being performed that use Array SSA form, then we can build full Array SSA form and simply ignore the @ variables for this particular analysis. Our running example is shown in Fig. 10. The partial Array SSA form for this example is shown in Fig. 11. The partial Array SSA form does not contain any @ variables since @ variables are not necessary for the level of analysis performed by the constant propagation algorithms in this paper. The data flow equations for this example are shown in Fig. 12. Each assignment in the Array SSA form results in one data flow equation. The numbering S1 through S8 indicates the correspondence. The argument to these equations are simply the current lattice values of the variables. The lattice operations are specific to the operations within the statement. Figures 13, 9, and 14 show the lattice computations for an assignment to an array element (as in S3 and S5), definition φ (as in S4 and S6), a reference to an array element (as in the RHS of S3 and S5). The lattice computation for a φ assignment (as in S7) A3 = φ(A2 , A1 ) is L(A3 ) = Lφ (L(A2 ), L(A1 )) = L(A1 ) u L(A2 ) where u is shown in Fig. 8. Notice that we also include lattice computation for specific arithmetic computations such as the multiply in S3 and S5. This allows for constant computation as well as constant propagation. Tables for these arithmetic computations are straightforward and are not shown.

Enabling Sparse Constant Propagation of Array Elements

47

The simple data flow algorithm is shown in Fig. 15. Our example includes the propagation of values from a node to two successors (Y) and to a node from two predecessors (D). Equations S1 and S2 are evaluated because they are associated with the entry block. Element 3 of Y2 is known to have the value 99 at this stage. As a result of the modification of Y2 both S3 and S5 are inserted into the worklist (they reference the lattice value of Y2 ). S3 uses the propagated constant 99 to compute and propagate the constant 198 to element 1 of D1 and then, after evaluation of S4, to element 1 of D2 . Any subsequent references to D in the then block of the source become references D2 in the Array SSA form and are known to have a constant value at element 1. Depending on the order of computations via the worklist, we may then compute either D3 and then D4 or, because D2 has been modified, we may compute D5 . If we compute D5 at this point, it appears to have a constant value. Subsequent evaluations D3 and D4 cause D5 to be reevaluated and lowered from constant value to ⊥ because the value along one path is not constant. Notice that in this case, the reevaluation of D5 could have been avoided by choosing an optimal ordering of processing. Processing of programs with cyclic control flow is no more complex but may involve recomputation that can not be removed by optimal reordering. In particular, the loop entry is a control flow merge point since control may enter from the top or come from the loop body. It will contain a φ which combines the value entering from the top with that returning after the loop. The lattice values for such a node may require multiple evaluations. Also notice that in this example, if I in S5 is known to have the value 3, it will be recoded as a constant element. Upon evaluation of S7, the intersection of the sets associated with D2 and D4 will not be empty and element 1 of D5 will be recorded as a constant.

Y [3] := 99 if C then D[1] := Y [3] ∗ 2 else D[1] := Y [I] ∗ 2 endif Z := D[1]

Fig. 10. Sparse Constant Propagation Example

48

V. Sarkar and K. Knobe

S1: S2: S3: S4: S5: S6: S7: S8:

Y0 and D0 in effect here. ... Y1 [3] := 99 Y2 := φ(Y1 , Y0 ) if C then D1 [1] := Y2 [3] ∗ 2 D2 := φ(D1 , D0 ) else D3 [1] := Y2 [I] ∗ 2 D4 := φ(D3 , D0 ) endif D5 := φ(D2 , D4 ) Z := D5 [1]

Fig. 11. Array SSA form for the Sparse Constant Propagation Example S1: S2: S3: S4: S5: S6: S7: S8:

L(Y1 ) L(Y2 ) L(D1 ) L(D2 ) L(D3 ) L(D4 ) L(D5 ) L(Z)

= = = = = = = =

< (3, 99) > Ldφ (L(Y1 ), L(Y0 )) Ld[ ] (L∗ (L(Y2 [3]), 2)) Ldφ (L(D1 ), L(D0 )) Ld[ ] (L∗ (L(Y2 [I]), 2)) Ldφ (L(D3 ), L(D0 )) Lφ (L(D2 ), L(D4 )) L(D5 [1])

Fig. 12. Data Flow Equations for the Sparse Constant Propagation Example L(A1 ) L(i) = > L(i) = Constant L(i) = ⊥ L(k) = > > > ⊥ L(k) = Constant > h(L(k), L(i))i ⊥ L(k) = ⊥ ⊥ ⊥ ⊥ Fig. 13. Lattice computation for array definition operator, L(A1 ) = Ld[ ] (L(k), L(i)) L(A[k]) L(k) = > L(k) = Constant L(k) = ⊥ L(A) = > > > ⊥ L(A) = h(i1 , e1 ), . . . i > ej , if ∃ (ij , ej ) ∈ L(A) with ij = L(k) ⊥ ⊥, otherwise L(A) = ⊥ ⊥ ⊥ ⊥ Fig. 14. Lattice computation for array reference operator, L(A[k]) = L[ ] (L(A), L(k))

Enabling Sparse Constant Propagation of Array Elements

49

Initialization: L(v) ← > for all local variables, v. insert(Ev , work list) for each equation Ev defining v such that v is assigned to in the entry block. Body: while (work list != empty) Ev ← remove(work list) reevaluate(Ev ) insert(Ev0 0 , work list) for each equation Ev0 0 that uses Ev end while

Fig. 15. Algorithm for Sparse Constant propagation (SC) for array and scalar variables

4 4.1

Sparse Conditional Constant Propagation Lattice Values of Executable Flags for Nodes and Edges

As in the SCC algorithm in [8], our array conditional constant propagation algorithm maintains executable flags associated with each node and each edge in the CFG. Flag Xni indicates whether node ni may be executed, and Xei indicates whether edge ei may be traversed. The lattice value of an execution flag is either no or maybe, corresponding to unreachable code and reachable code respectively. The lattice value for an execution flag is initialized to no, and can be lowered to maybe in the course of the constant propagation algorithm. In practice, control dependence identities can be used to reduce the number of executable flag variables in the data flow equations e.g., a single flag can be used for all CFG nodes that are control equivalent. For the sake of simplicity, we ignore such optimizations in this paper. The executable flag of a node is computed from the executable flags of its incoming edges. The executable flag of an edge is computed from the executable flag of its source node and knowledge of the branch condition variable used to determine the execution path from that node. These executable flag mappings are summarized in Fig. 16 for a node n with two incoming edges, e1 and e2, and two outgoing edges, e3 and e4. The first function table in Fig. 16 defines the join operator u on executable flags such that Xn = Xe1 u Xe2 . We introduce a true operator LT and a false operator LF on lattice values such that Xe3 = LT (Xn , L(C)) and Xe4 = LF (Xn , L(C)). Complete function tables for the LT and LF operators are also shown in Fig. 16. Note that all three function tables are monotonic with respect to their inputs. Other cases for mapping a node Xn value to the Xe values of its outgoing edges can be defined similarly. If n has exactly one outgoing edge e, then Xe =

50

V. Sarkar and K. Knobe

Definition of join operator, Xn = Xe1 u Xe2 : Xn Xe2 = no Xe2 = maybe Xe1 = no no maybe Xe1 = maybe maybe maybe

X e1

X n

X e2

C?

TRUE

FALSE

X e3

X e4

Definition of true operator for branch condition C, Xe3 = LT (Xn , L(C)): Xe3 L(C) = > L(C) = true L(C) = false L(C) = ⊥ Xn = no no no no no Xn = maybe no maybe no maybe

Definition of false operator for branch condition C, Xe4 = LF (Xn , L(C)): Xe4 L(C) = > L(C) = true L(C) = false L(C) = ⊥ Xn = no no no no no Xn = maybe no no maybe maybe Fig. 16. Executable flag mappings for join operator (u), true operator (LT ), and false operator (LF )

L(k3 ) Xe2 = no Xe2 = maybe Xe1 = no > L(k2 ) X e1 = maybe L(k1 ) L(k1 ) u L(k2 ) Fig. 17. k3 := Φ(k1 , Xe1 , k2 , Xe2 ), where execution flags Xe1 and Xe2 control the selection of k1 and k2 respectively

Enabling Sparse Constant Propagation of Array Elements

51

Xn . If node n has more than two outgoing edges then the mapping for each edge is similar to the LT and LF operators. Recall that the LΦ function for a control Φ was defined in Sect. 3.2 by the meet function, LΦ (L(A2 ), L(A1 )) = L(A1 ) u L(A2 ). For the extended SCC algorithm described in this section, we use the definition of LΦ shown in Fig. 17. This definition uses executable flags Xe1 and Xe2 , where e1 and e2 are the incoming control flow edges for the Φ function. Thus, the lattice values of the executable flags Xe1 and Xe2 are used as compile-time approximations of @ variables @k1 and @k2 . 4.2

Sparse Conditional Constant Propagation Algorithm

We introduce our algorithm by first explaining how it works for acyclic scalar code. Consider the example program shown in Fig. 18. The basic blocks are labeled n1, n2, n3 and n4. Edges e12, e13, e24 and e34 connect nodes in the obvious way. The control flow following block n1 depends on the value of variable n. The first step is to transform the example in Fig. 18 to partial Array SSA form (with no @ variables) as shown in Fig. 19. Note that since k had multiple assignments in the original program, a φ function is required to compute k3 as a function of k1 and k2 . The second step is to use the partial Array SSA form to create a set of data flow equations on lattice values for use by our constant propagation algorithm. The conversion to equations is performed as follows. There is one equation created for each assignment statement in the program. There is one equation created for each node in the CFG. There is one equation created for each edge in the CFG. The equations for the assignments in our example are shown in Fig. 20. The equations for the nodes and edges in our example are found in Fig. 21 and 22 respectively. The lattice operations L< , L∗ , and Lmax use specific knowledge of their operation as well as the lattice values of their operands to compute resulting lattice operations. For example, L∗ (⊥, L(0)) results in L(0) because the result of multiplying 0 by any number is 0. Next, we employ a work list algorithm, shown in Fig. 23, that reevaluates the equations until there are no further changes. A solution to the data flow equations identifies lattice values for each variable in the Array SSA form of the program, and for each node executable flag and edge executable flag in the CFG. Reevaluation of an equation associated with an assignment may cause equations associated with other assignments to be inserted on the work list. If the value appears in a conditional expression, it may cause one of the equations associated with edges to be inserted on the work list. Reevaluation of an edge’s executable flag may cause an equation for a destination node’s executable flag to be inserted on the work list. If reevaluation of a node’s executable flag indicates that the node may be evaluated, then the equations associated with assignments within that node to be added to the work list. When the algorithm terminates, the lattice values for variables identify the constants in the program, and the lattice values for executable flags of nodes identify unreachable code.

52

V. Sarkar and K. Knobe n1: n2: n3: n4:

i := 1 C := i < n if C then k := 2 ∗ i else k := 2 ∗ n endif print k

n1 e12

e13

n2

n3

e24

e34 n4

Fig. 18. Acyclic Scalar Example

n1: n2: n3: n4:

i := 1 C := i < n if C then k1 := 2 ∗ i else k2 := 2 ∗ n endif k3 := φ(k1 , k2 ) print k3

Fig. 19. Partial Array SSA form for the Acyclic Scalar Example

L(i) = L(1) L(C) = L< (L(i), L(n)) L(k1 ) = L∗ (L(2), L(i)) L(k2 ) = L∗ (L(2),L(n)) L(k3 ) = LΦ (L(k1 ), Xe24 , L(k2 ), Xe34 )

Fig. 20. Equations for Assignments

Xn1 Xn2 Xn3 Xn4

= true = Xe12 = Xe13 = Xe24 u Xe34

Fig. 21. Equations for Nodes

Enabling Sparse Constant Propagation of Array Elements Xe12 Xe13 Xe24 Xe34

53

= LT (Xn1 , L(C)) = LF (Xn1 , L(C)) = Xn2 = Xn3

Fig. 22. Equations for Edges Initialization: L(v) ← > for all local variables, v. Xn ← maybe where Xn is the executable flag for the entry node. Xn ← no where Xn is the executable flag for any node other than the entry node. Xe ← no where Xe is the executable flag for any edge. insert(Ev , work list) for each equation Ev defining v such that v is assigned to in the entry block. Body: while (work list != empty) Ev ← remove(work list) reevaluate(Ev ) insert(Ev0 0 , work list) for each equation Ev0 0 that uses Ev if Ev defines the executable flag for some node n then insert(Ev0 0 , work list) for each equation Ev0 0 defining v 0 such that v 0 is assigned to in block n. end if end while

Fig. 23. Sparse Conditional Constant propagation (SCC) algorithm for scalar and array variables

Even though we assumed an acyclic CFG in the above discussion, the algorithm in Fig. 23 can be used unchanged for performing constant propagation analysis on a CFG that may have cycles. The only difference is that the CFG may now contain back edges. Each back edge will be evaluated when its source node is modified. The evaluation of this back edge may result in the reevaluation of its target node. As in past work, it is easy to show that the algorithm must take at most O(Q) time, where Q is the number of data flow equations, assuming that the maximum arity of a function is constant and the maximum height of the lattice is constant. As an example with array variables, Fig. 24 lists the data flow equations for the assignment statements in the Array SSA program in Fig. 6 (the data flow equations for nodes and edges follow the CFG structure as in Figs. 21 and 22). Given the definition of lattice elements for array variables from Sect. 3.1, the conditional constant propagation algorithm in Fig. 23 can also be used unchanged for array variables.

54

V. Sarkar and K. Knobe

L(A0 ) = ⊥ L(i) = L(1) L(C) = L< (L(i), L(n)) L(k) = L∗ (L(2), L(i)) L(A1 ) = Ld[ ] (L(k), L(i)) L(A2 ) = Ldφ (L(A1 ), L(A0 )) L(A3 ) = LΦ (L(A2 ), Xe24 , L(A0 ), Xe14 ) Fig. 24. Equations for Assignments from Fig. 6

5

Related Work

Static single assignment (SSA) form for scalar variables has been a significant advance. It has simplified the design of some optimizations and has made other optimizations more effective. The popularity of SSA form surged after an efficient algorithm for computing SSA form was made available [4]. SSA form is now a standard representation used in modern optimizing compilers in both industry and academia. However, it has been widely recognized that SSA form is much less effective for array variables than for scalar variables. The approach recommended in [4] is to treat an entire array like a single scalar variable in SSA form. The most serious limitation of this approach is that it lacks precise data flow information on a per-element basis. Array SSA form addresses this limitation by providing Φ functions that can combine array values on a per-element basis. The constant propagation algorithm described in this paper can propagate lattice values through Φ functions in Array SSA form, just like any other operation/function in the input program. The problem of conditional constant propagation for scalar variables has been studied for several years. Wegbreit [7] provided a general algorithm for solving data flow equations; his algorithm can be used to perform conditional constant propagation and more general combinations of program analyses. However, his algorithm was too slow to be practical for use on large programs. Wegman and Zadeck [8] introduced a Sparse Conditional Constant (SCC) propagation algorithm that is as precise as the conditional constant propagation obtained by Wegbreit’s algorithm, but runs faster than Wegbreit’s algorithm by a speedup factor that is at least O(V ), where V is the number of variables in the program. The improved efficiency of the SCC algorithm made it practical to perform conditional constant propagation on large programs, even in the context of industry-strength product compilers. The main limitation of the SCC algorithm is a conceptual one — the algorithm operates on two “worklists” (one containing edges in the SSA graph and another containing edges from the control flow

Enabling Sparse Constant Propagation of Array Elements

55

graph) rather than on data flow equations. The lack of data flow equations makes it hard to combine the algorithm in [8] with other program analyses. The problem of combining different program analyses based on scalar SSA form has been addressed by Click and Cooper in [2], where they present a framework for combining constant propagation, unreachable-code elimination, and value numbering that explicitly uses data flow equations. Of the conditional constant propagation algorithms mentioned above, our work is most closely related to that of [2] with two significant differences. First, our algorithm performs conditional constant propagation through both scalar and array references, while the algorithm in [2] is limited only to scalar variables. Second, the framework in [2] uses control flow predicates instead of execution flags. It wasn’t clear from the description in [2] how their framework deals with predicates that are logical combinations of multiple branch conditions; it appears that they must either allow the possibility of an arbitrary size predicate expression appearing in a data flow equation (which would increase the worst-case execution time complexity of their algorithm) or they must sacrifice precision by working with an approximation of the predicate expression.

6

Conclusions

We have presented a new sparse conditional constant propagation algorithm for scalar and array references based on Array SSA form [6]. Array SSA form has two advantages: It is designed to support analysis of arrays at the element level and it employs a new Φ function that is a pure function of its operands, and can be manipulated by the compiler just like any other operator in the input program. The original sparse conditional constant propagation algorithm in [8] dealt with control flow and data flow separately by maintaining two distinct work lists. Our algorithm uses a single set of data flow equations and is therefore conceptually simpler. In addition to being simpler, the algorithm presented in this paper is more powerful than its predecessors in that it handles constant propagation through array elements. It is also more effective because its use of data flow equations allows it to be totally integrated with other data flow algorithms, thus making it easier to combine other analyses with conditional constant propagation.

References 1. A.V. Aho, R. Sethi, and J.D. Ullman. Compilers: Principles, Techniques, and Tools. Addison-Wesley, 1986. 2. Cliff Click and Keith D. Cooper. Combining Analyses, Combining Optimizations. ACM Transactions on Programming Languages and Systems, 17(2):181–196, March 1995. 3. The Standard Performance Evaluation Corporation. SPEC CPU95 Benchmarks. http://open.specbench.org/osg/cpu95/, 1997.

56

V. Sarkar and K. Knobe

4. Ron Cytron, Jeanne Ferrante, Barry K. Rosen, Mark N. Wegman, and F. Kenneth Zadeck. Efficiently Computing Static Single Assignment Form and the Control Dependence Graph. ACM Transactions on Programming Languages and Systems, 13(4):451–490, October 1991. 5. Matthew S. Hecht. Flow Analysis of Computer Programs. Elsevier North-Holland, Inc., 1977. 6. Kathleen Knobe and Vivek Sarkar. Array SSA form and its use in Parallelization. Conf. Rec. Twenty-fifth ACM Symposium on Principles of Programming Languages, San Diego, California, January 1998. 7. B. Wegbreit. Property Extraction in Well-Founded Property Sets. IEEE Transactions on Software Engineering, 1:270–285, 1975. 8. Mark N. Wegman and F. Kenneth Zadeck. Constant Propagation with Conditional Branches. ACM Transactions on Programming Languages and Systems, 13(2):181– 210, April 1991.

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses Michael Hind and Anthony Pioli State University of New York at New Paltz IBM Thomas J. Watson Research Center {hind,pioli}@mcs.newpaltz.edu

Abstract. This paper describes an empirical comparison of four contextinsensitive pointer alias analysis algorithms that use varying degrees of ﬂow-sensitivity: a ﬂow-insensitive algorithm that tracks variables whose addresses were taken and stored; a ﬂow-insensitive algorithm that computes a solution for each function; a variant of this algorithm that uses precomputed kill information; and a ﬂow-sensitive algorithm. In addition to contrasting the precision and eﬃciency of these analyses, we describe implementation techniques and quantify their analysis-time speed-up.

1

Introduction

To eﬀectively analyze programs written in languages that make extensive use of pointers, such as C, C++, or Java (in the form of references), knowledge of pointer behavior is required. Without such knowledge, conservative assumptions about pointer values must be made, resulting in less precise data ﬂow information, which can adversely aﬀect the eﬀectiveness of analyses and tools that depend on this information. A pointer alias analysis is a compile-time analysis that, for each program point, attempts to determine what a pointer can point to. As such an analysis is, in general, undecidable [25, 34], approximation methods have been developed. These algorithms provide trade-oﬀs between the eﬃciency of the analysis and the precision of the computed solution. The goal of this work is to quantify how the use of ﬂow-sensitivity aﬀects precision and eﬃciency. Although several researchers have provided empirical results of their techniques, comparisons among algorithms can be diﬃcult because of diﬀering program representations, benchmark suites, and metrics. By holding these factors constant, we can focus more on the eﬃcacy of the algorithms and less on the manner in which the results were obtained. The contributions of this paper are the following: – empirical results that measure the precision and eﬃciency of four pointer alias analysis algorithms with varying degrees of ﬂow-sensitivity;

This work was supported in part by the National Science Foundation under grant CCR-9633010, by IBM Research, and by SUNY at New Paltz Research and Creative Project Awards.

G. Levi (Ed.): SAS’98, LNCS 1503, pp. 57–81, 1998. c Springer-Verlag Berlin Heidelberg 1998

58

M. Hind and A. Pioli

– empirical evidence of how various implementation enhancements signiﬁcantly improved analysis time of the ﬂow-sensitive analysis. In addition to the use of ﬂow-sensitivity, other factors that aﬀect the cost/precision trade-oﬀs of pointer alias analyses include the use of context-sensitivity and the manner in which aggregates (arrays and structs) and the heap are modeled. Our experiments hold these factors constant so that the results only reﬂect the usage of ﬂow-sensitivity. Section 2 highlights the four algorithms and their implementations. Section 3 describes the empirical study of the four algorithms, analyzes the results, and contrasts them with related results from other researchers. Section 4 overviews some of the performance-improving enhancements made in the implementation and quantiﬁes their analysis-time speed-up. Section 5 describes other related work. Section 6 states conclusions.

2

Analyses and Implementation

One manner of classifying interprocedural data ﬂow analyses is whether they consider control ﬂow information during the analysis. A flow-sensitive analysis considers control ﬂow information of a procedure during its analysis of the procedure. A flow-insensitive analysis does not consider control ﬂow information during its analysis, and thus can be more eﬃcient, but less precise. (See [31] for a full discussion of these deﬁnitions.) The algorithms we consider, listed in order of increasing precision, are AT: a ﬂow-insensitive algorithm that computes one solution set for the entire program that contains all named objects whose address has been taken and stored, FI: a ﬂow-insensitive algorithm [4, 5] that computes a solution set for every function, FIK: a ﬂow-insensitive algorithm [4, 5] that computes a solution set for every function, but attempts to improve precision by using precomputed (ﬂowsensitive) kill information, FS: a ﬂow-sensitive algorithm [8, 5] that computes a solution set for every program point. The following sections provide further information about these analyses and their implementation. 2.1

Algorithms

The program is represented as a program call (multi-) graph (PCG), in which a node corresponds to a function, and a directed edge represents a call to the target function.1 Each function body is represented by a control ﬂow graph (CFG). This 1

Potential calls can occur due to function pointers and virtual methods, in which the called function is not known until run time.

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

59

graph is used to build a simpliﬁed sparse evaluation graph (SEG) [9], discussed in Section 4. The address-taken analysis (AT) computes its solution by making a single pass over all functions in the program,2 adding to a global set all variables whose addresses have been assigned to another variable. These include actual parameters whose addresses are stored in the corresponding formal. Examples are statements such as “p = &a;”, “q = new ...;”, and “foo(&a);”, but not simple expression statements such as “&a;” because the address was not stored. AT is eﬃcient because it is linear in the size of the program and uses a single solution set, but it can be very imprecise. It is provided as a base case for comparison to the other three algorithms presented in this paper. The general manner in which the other three analyses compute their solutions is the same. A nested ﬁxed point computation is used in which the outer nest corresponds to computing solutions for each function in the PCG. Each such function computation triggers the computation of a local solution for all program points that are distinguished in the particular analysis. For the ﬂowsensitive (FS) analysis, the local solution corresponds to each CFG node in the function. For the other two ﬂow-insensitive analyses (FI, FIK), the local solution corresponds to one set that conservatively represents what can hold anywhere in the function. This general framework is presented as an iterative algorithm in Fig. 1 and is further described in [5]. An extension to handle virtual methods is described in [6]. Section 4 reports improvements due to the use of a worklistbased implementation.

S1 : S2 : S3 : S4 : S5 : S6 : S7 : S8 : S9 : S10 : S11 :

build the initial PCG foreach procedure, p, in the PCG, loop initialize interprocedural alias sets of p to {} end loop repeat foreach procedure, p, in the PCG, loop using the interprocedural alias sets (for entry of p and call sites in p), compute the intraprocedural alias sets of p using the intraprocedural alias sets of p, update the interprocedural alias sets representing the eﬀect of p on each procedure that calls or is called by p end loop using new function pointer aliases, update the PCG, initializing interprocedural alias sets of new functions to {} until the interprocedural alias sets and the PCG converge

Fig. 1. High-level description of general algorithm [5] 2

As the PCG is not used, this can include functions that are not called.

60

M. Hind and A. Pioli

The FS, FI, and FIK analyses utilize the compact representation [8, 5] to represent alias relations. This representation shares the property of the points-to representation [14], in that it captures the “edge” information of alias relations. For example, if variable a points to b, which in turn points to c, the compact representation records only the following alias set: {∗a, b, ∗b, c}, from which it can be inferred that ∗∗ a, c and ∗∗ a, ∗b are also aliases.3 All analyses are context-insensitive; they merge information ﬂowing from diﬀerent calls to the same function, and may suﬀer from the unrealizable path problem [27], i.e., they potentially propagate back to the wrong caller the aliases of the called function. (Sections 3.2 and 4.4 discuss this potential imprecision.) Context-sensitive analyses [14, 48] do not suﬀer from this problem, but may increase time/space costs. As in [22, 7], all analyses considered here represent the (possibly many) objects allocated at calls to new or malloc by creating a named object based on the CFG node number of the allocation statement. These objects are referred to as heapn , where n is the CFG node number of the allocation statement. These names are unique throughout the entire program. More precise heap modeling schemes [29, 22, 19, 7, 8, 11, 15, 16, 39, 40] can improve precision, but may also increase time/space costs. Quantifying the eﬀects of using context-sensitivity and various heap models is beyond the scope of this work. Consider the simple program in Fig. 2. The AT analysis computes only one set of objects, which it assumes all pointers may point to. This set will contain ﬁve objects {heapS1 , heapS3 , heapS4 , heapS6 , and heapS7 }, all of which will appear to be referenced at S8.

S1: S2: S3:

T ∗p, ∗q; void main(void) { p = new T; f(); p = new T; }

S4: S5: S6: S7:

void f(void) { p = new T; g(); p = new T; q = new T; }

S8:

void g(void) { T t = ∗p; }

Fig. 2. Example program

The FI analysis does not use any intraprocedural control ﬂow information. Instead it conservatively computes what can hold anywhere within a function, i.e., for each function, f , it uses only one alias set, Holdsf , to represent what may hold at any CFG node in f . Thus, in Fig. 2 the FI analysis assumes that ∗p, heapS1 and ∗p, heapS3 can ﬂow into f. This results in Holdsg = Holdsf = {∗p, heapS1 , ∗p, heapS3 , ∗p, heapS4 , ∗p, heapS6 , ∗q, heapS7 }, which re3

See [30, 5] for a discussion of precision trade-oﬀs between this representation and an explicit representation, which would contain all four alias pairs.

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

61

sults in four objects, heapS1 , heapS3 , heapS4 , and heapS6 potentially being referenced at S8.4 The FIK analysis attempts to improve the precision of the ﬂow-insensitive analysis by precomputing kill information for pointers, and then uses this information during the ﬂow-insensitive analysis at call sites.5 For example, the precomputation will determine that all alias relations involving ∗p on entry to f will be killed before the call to g at S5. Thus, Holdsg will contain only the alias relations that are generated in f and propagated to g, i.e., Holdsg = {∗p, heapS4 , ∗p, heapS6 ∗q, heapS7 }. This results in two objects, heapS4 and heapS6 , potentially being referenced at S8. The FS analysis associates an alias set before (Inn ) and after (Outn ) every CFG node, n. For example, OutS1 = {∗p, heapS1 } because ∗p and heapS1 refer to the same storage. At the entry to function g, the FS analysis will compute InS8 = {∗p, heapS4 }, which is the precise solution for this simple example. This example illustrates the theoretical precision levels of the four analyses, from FS (most precise) to AT (least precise). The AT analysis is our most eﬃcient analysis because it is linear and only uses one set. The FI analysis is more eﬃcient than the FIK analysis because it neither precomputes kill information nor uses it during the analysis. One would expect the FS analysis to be the least eﬃcient because it needs to distinguish solutions for every point in the program. Thus, a theoretical spectrum exists in terms of precision and eﬃciency with the AT analysis on the less precise/more eﬃcient side, the FS analysis on the more precise/less eﬃcient side, and the FI and FIK analyses in the middle. Not studied in this paper are other ﬂow-insensitive analyses [45, 42] that use one alias set for the whole program and limit the number of alias relations by sometimes grouping distinct variables into one named object. These analysis fall in between AT and FI in the theoretical precision spectrum. 2.2

Implementation

The analyses have been implemented in the NPIC system, an experimental program analysis system written in C++. The system uses multiple and virtual inheritance to provide an extensible framework for data ﬂow analyses [21, 33]. A prototype version of the IBM VisualAge C++ compiler [43, 32] is used as the front end. The abstract syntax tree constructed by the front end is transformed into a PCG and a CFG for each function, which serve as input to the alias analyses. No CFG is built for library functions. We model a call to a library function based on its semantics, thereby providing the beneﬁts of context-sensitive analysis of such calls. Library calls that cannot aﬀect the value of a pointer are treated as the identity transfer function. 4 5

Although a ﬁnal intraprocedural ﬂow-sensitive pass can be used to improve precision [5], this pass has not been implemented. Kill information is computed in a single ﬂow-sensitive prepass of each CFG. For each call site, c, we compute two sets, the set of pointers that are deﬁnitely killed on all paths from entry to c and the set of pointers that are deﬁnitely killed on all paths from c to exit [4, 5]. Only the ﬁrst set is used in our example.

62

M. Hind and A. Pioli

The FS, FI, and FIK analyses are implemented using worklists. (Section 4.2 discusses an earlier iterative implementation.) These three analyses incorporate function pointer analysis into the pointer alias analysis as described in [4, 5]. Currently, array elements and ﬁeld components are not distinguished, and setjmp/longjmp6 statements are not supported. The implementation also assumes that pointer values will only exist in pointer variables, and that pointer arithmetic does not result in the pointer going beyond array boundaries. As stated in Section 2.1, heap objects are named based on their allocation site. To model the values passed as argc and argv to the main function, a dummy main function was added, which called the benchmark’s main function, simulating the eﬀects of argc and argv. This function also initialized the iob array, used for standard I/O. The added function is similar to the one added by Ruf [36, 38] and Landi et al. [28, 26]. Initializations of global variables are automatically modeled as assignment statements in the dummy main function.

3

Results

Our benchmark suite contains 21 C programs, 18 provided by other researchers [28, 14, 36] and 3 from the SPEC CINT92 [3] and CINT95 [44] benchmarks.7 Table 1 describes characteristics of the suite. The third column contains the number of lines in the source and header ﬁles reported by the Unix utility wc. The fourth column reports the number of user-deﬁned functions (nodes in the PCG), which include the dummy main function. The next two columns give the number of call sites, distinguished between user and library function calls. The next two columns report cumulative statistics for all CFG nodes and edges. These ﬁgures include nodes and edges created by the initialization of globals. The following column computes the ratio of CFG edges to nodes. The next column reports the percentage of CFG nodes that are considered pointer-assignment nodes. The current analysis treats an assignment as a pointer-assignment if the variable involved in the pointer expression on the left side of the assignment is declared to be a pointer.8 The last two columns report the number of recursive functions (functions that are in PCG cycles) and heap allocation sites in each program. The last row of the table reports the average edge/node ratio and the 6 7

8

Although one program in our benchmark suite, anagram, does syntactically contain a call to longjmp, the code is unreachable. Some programs had to be syntactically modiﬁed to satisfy C++’s stricter type checking semantics. A few program names are diﬀerent than those reported by Ruf [36]. The SPEC CINT92 program 052.alvinn was named backprop in Todd Austin’s benchmark suite [2]. Ruf referred to ks as part, and ft as span [38]. This is more conservative than considering statements in which the left side expression is a pointer. Thus, statements such as “p->field = ...” are treated as pointer assignments no matter how the type of field is declared. A more accurate categorization would not aﬀect the precision of the analysis, but could improve the eﬃciency by reducing the number of nodes considered during the analysis as discussed in Section 4.1.

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

63

average pointer-assignment node percentage, both of which are computed by averaging the corresponding values over the 21 benchmarks. Table 1. Static characteristics of benchmark suite Name allroots 052.alvinn 01.qbsort 15.trie 04.bisect 17.bintr anagram lex315 ks 05.eks 08.main 09.vor loader 129.compress ft football compiler assembler yacr2 simulator 099.go Average

3.1

Source Landi SPEC92 McCat McCat McCat McCat Austin Landi Austin McCat McCat McCat Landi SPEC95 Austin Landi Landi Landi Austin Landi SPEC95

Call Sites LOC Funcs User Lib 227 7 19 35 272 9 8 13 325 8 9 25 358 13 19 21 463 9 11 18 496 17 27 28 650 16 22 38 733 17 102 52 782 14 17 67 1,202 30 62 49 1,206 41 68 53 1,406 52 174 28 1,539 30 79 102 1,934 25 35 28 2,156 38 63 55 2,354 58 257 274 2,360 40 349 107 3,446 52 247 243 3,979 59 158 169 4,639 111 447 226 29,637 373 2,054 22

Nodes 157 223 173 167 175 196 332 527 513 678 684 876 687 596 732 2,710 1,723 1,544 2,030 2,686 16,823

CFG Edges 167 243 187 181 189 205 365 615 576 732 700 917 773 652 808 3,164 2,090 1,738 2,328 3,123 20,176

Edges N odes

1.064 1.090 1.081 1.084 1.080 1.046 1.099 1.167 1.123 1.080 1.023 1.047 1.125 1.094 1.104 1.168 1.213 1.126 1.147 1.163 1.199 1.111

Ptr-Asgn Rec Allocation Nodes Funcs Sites 2.6% 2 1 11.2% 0 0 24.9% 1 5 23.4% 3 5 9.7% 0 2 9.7% 5 1 9.0% 1 2 2.5% 0 3 26.9% 0 5 3.8% 0 3 24.3% 3 8 27.5% 5 8 8.3% 2 7 6.4% 0 0 19.5% 0 5 1.8% 1 0 1.1% 14 0 8.0% 0 16 5.4% 5 26 2.7% 0 4 .2% 1 0 10.9%

Description of Experiment

This section presents precision and eﬃciency results. For each benchmark and each analysis, we report the analysis time, the maximum memory used, and the average number of objects the analysis computes a dereferenced pointer can point to. The precision results for the FIK analysis are exactly the same as the FI analysis for all benchmarks. Thus, we do not explicitly include this analysis in our precision data. The experiment was performed on a 132MHz IBM RS/6000 PowerPC 604 with 96MB RAM and 104MB paging space, running AIX 4.1.5. The executable was built with IBM’s xlC compiler using the “-O3” option. The analysis time is reported in seconds and does not include the time to build the PCG and CFGs, but does include any analysis-speciﬁc preprocessing, such as building the SEG from the CFG. This information is displayed in the top left chart in Fig. 3. The top right chart of this ﬁgure reports the high-water mark of memory usage during the analysis process. This memory usage includes the intermediate representation, the alias information, and statistics-related data. This information was obtained by using the “ps v” command under AIX 4.1.5. To collect precision information, the system traverses the representation visiting each expression containing a pointer dereference and, using the computed

64

M. Hind and A. Pioli Analysis Time (Secs)

Max Memory Usage (MB)

allroots

allroots

052.alvinn

052.alvinn

01.qbsort

01.qbsort

15.trie

15.trie

04.bisect

04.bisect

17.bintr

17.bintr

anagram

anagram

lex315

lex315

ks

ks

05.eks

05.eks

08.main

08.main 10.3 15.3 11.9

09.vor loader 129.compress

09.vor loader 129.compress

ft

ft

football

football

compiler

compiler

assembler yacr2 simulator 099.go 0

1

2

3

Flow Sensitive

4

5

21.0 19.0

052.alvinn

25.0

01.qbsort

13.0

15.trie

20.0

04.bisect

28.0

17.bintr

34.0

anagram

20.0

lex315

54.0

ks

16.0

05.eks

32.0

08.main

29.0

09.vor

97.0

loader

27.0

129.compress

38.0

ft

205.0

football

112.0

compiler

195.0

assembler

106.0

yacr2

239.0

simulator 099.go 2

Flow Sensitive

099.go 0

Flow Insensitive/Kill

allroots

1

yacr2 simulator

6

Reads Precision (Num Objects)

0

assembler

8.6 9.7 10.5 10.6 8.6 10.1 9.4 14.3 9.8

3

4

180.0 17.0 17.0

5

5

10

Flow Insensitive

32.6 20

15

Address Taken

Writes Precision (Num Objects) 21.0

allroots

19.0

052.alvinn

25.0

01.qbsort

13.0

15.trie

20.0

04.bisect

28.0

17.bintr

34.0

anagram

20.0

lex315

54.0

ks

16.0

05.eks

32.0

08.main

29.0

09.vor

97.0

loader

27.0

129.compress

38.0

ft

205.0

football compiler (No Derefs)

195.0

assembler

106.0

yacr2

239.0

simulator 099.go 0

Flow Insensitive

1

2

3

Address Taken

Fig. 3. Analysis time, memory usage, and precision results

4

180.0 13.6 13.6

5

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

65

alias information, reports how many named objects are aliased to the pointer expression. We report the average number of such dereferences for both reads and writes. Further precision information is provided in [21]. A pointer expression with multiple dereferences, such as ∗ ∗ ∗p, is counted as multiple dereference expressions, one for each dereference. The intermediate dereferences (∗p and ∗∗p) are counted as reads. The last dereference (∗∗∗p) is counted as a read or write depending on the context of the expression. Statements such as (∗p)++ and ∗p += increment are treated as both a read and a write of ∗p. We consider a pointer to be dereferenced if the variable is declared as a pointer or an array formal parameter, and one or more of the “∗”, “->”, or “[ ]” operators are used with that variable. Formal parameter arrays are included because their corresponding actual parameter(s) could be a pointer. We do not count the use of the “[ ]” operator on arrays that are not formal parameters because the resulting “pointer” (the array name) is constant, and therefore, counting it may skew results. Figure 4 classiﬁes the type of pointer dereferenced averaged over all programs. Information for each benchmark is given in [21]. 100 80

46.6

Percent

61.1 60

18.9

40

Global

38.9 20

34.5

22.8 0

Reads

Formal Local

Writes

Fig. 4. Classiﬁcation of dereferenced pointer types for all programs

The manner in which the heap is modeled must be considered in evaluating precision results. For example, a model that uses several names for objects in the heap may seem less precise when compared to a model that uses fewer names [36]. Similarly, analyses that represent invisible objects (objects not lexically visible in the current procedure) aliased to a formal parameteras a single object9 may report fewer objects. Our analyses do not use this technique. Assuming a correct input program, each pointer dereference should correspond to at least one object at run time, and thus one serves as a lower bound for the average. Although a precision result close to one demonstrates the analysis is precise (modulo heap and invisible object naming), a larger number could reﬂect an imprecise algorithm, a limitation of static analysis, or a pointer dereference that corresponds to diﬀerent memory locations over the program’s execution. 9

This modeling technique can increase the possibility of strong updates [7].

66

M. Hind and A. Pioli

Flow-Sensitive — Reads allroots 052.alvinn 01.qbsort 15.trie 04.bisect 17.bintr anagram lex315 ks 05.eks 08.main 09.vor loader 129.compress ft football compiler assembler yacr2 simulator 099.go

Flow-Sensitive — Writes

1.6 1.0 1.8 1.0 1.0 1.0 1.1 2.0 1.7 1.0 3.6 1.6 1.4 1.1 2.4 1.5 1.0 2.1 1.0 1.9 17.0* 0

1

2

3

4

allroots 052.alvinn 01.qbsort 15.trie 04.bisect 17.bintr anagram lex315 ks 05.eks 08.main 09.vor loader 129.compress ft football compiler 0.0 assembler yacr2 simulator 099.go

5

0

Flow-Insensitive — Reads allroots 052.alvinn 01.qbsort 15.trie 04.bisect 17.bintr anagram lex315 ks 05.eks 08.main 09.vor loader 129.compress ft football compiler assembler yacr2 simulator 099.go

1.0 1.8 1.0 1.0 1.0 1.1 2.0 1.8 1.0 4.6 1.7 1.7 1.1 2.5 1.6 1.0 2.2 1.0 1.9 17.0*

Local

1

2

3

4

Nonvisible Local

1.0 1.5 1.0 1.1 1.0 1.0 1.8 1.6 1.2 2.6 1.3 2.2 1.1 1.7 1.7 2.2 1.1 2.0 13.6* 1

2

3

4

5

Flow-Insensitive — Writes

1.6

0

1.0

1.0 allroots 1.0 052.alvinn 1.5 01.qbsort 1.1 15.trie 1.1 04.bisect 1.0 17.bintr 1.0 anagram 1.8 lex315 1.9 ks 1.2 05.eks 3.3 08.main 1.4 09.vor 2.4 loader 1.1 129.compress 1.8 ft 2.8 football compiler (No Derefs) 2.4 assembler 1.1 yacr2 2.1 simulator 099.go

5

0

Global

1

2

Formal Parameter

3

13.6* 4

5

Heap

∗ The bars for 099.go are truncated. The object type breakdown for both FS and FI is Nonvisible Formal Local Local Global Parameter Heap Reads 0.1 9.7 7.2 0.01 0 Writes 0.0 8.1 5.4 0.01 0

Fig. 5. Breakdown of average object type pointed to by a dereferenced pointer

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

67

The bottom two charts of Fig. 3 provide a graphical layout of precision information for reads and writes. Fig. 5 reﬁnes this information for the FS and FI analyses by providing a breakdown of the type of object pointed to. Fig. 6 provides similar information averaged over all programs. Charts E and F of this ﬁgure report the percentage of dereferenced pointers that resolve to exactly one object in our model. If the object is a named variable, as opposed to a heap object, the pointer dereference could be replaced with the variable. This information is expanded upon in [21]. 3.2

Discussion

As expected, the results from the analysis speed chart of Fig. 3 indicate that the AT analysis is eﬃcient; it takes less than .4 seconds on all programs. The FI analysis can be over twice as fast as the FS analysis, and was faster than the FS analysis in all but one program. The precision of the AT analysis leaves much to be desired. Fig. 6 reports on average 111.9 objects for reads and 96.68 objects for writes were in the AT set.10 As one would expect this set to increase with the size of the program, the precision for this analysis will worsen with larger programs. The results also indicate that the FIK analysis is not beneﬁcial. On our benchmark suite it is never more precise than the FI analysis, and on some occasions requires more analysis time than the FS analysis. One explanation for the precision result may be that an alias relation created to simulate a reference parameter, in which the formal points to the actual, typically is not killed in the called routine, i.e., the formal parameter is not modiﬁed, but rather is used to access the passed actual. Thus, programs containing these alias relations will not beneﬁt from the precomputed kill information. One surprising result is the overall precision of the FI analysis. In 12 of the 21 benchmarks the FI analysis is as precise as the FS analysis. This seems to suggest that the added precision obtained by the FS analysis in considering control ﬂow within a function is not signiﬁcant for those benchmarks, at least where pointers are dereferenced. We oﬀer two possible explanations: 1. Pointer variables are often not assigned more than one distinguished object within the same function. Thus, distinguishing program points within a function, a key diﬀerence between the FS and FI analyses, does not often result in an increase in precision. We have seen exceptions to this in the function InitLists of the ks benchmark and in the function InsertPoint in the 08.main benchmark. In both cases the same pointer is reused in two list-traversing loops. 2. It seems that a large number of alias relations are created at call sites because of actual/formal parameter bindings. The lack of a substantial precision diﬀerence between our FS and FI analyses may be because both algorithms rely on the same (context-insensitive) mapping mechanism at call sites. 10

The numbers diﬀer because they are weighted with the number of reads and writes through a pointer in each program.

68

M. Hind and A. Pioli Total Precision Reads 111.9

2.5

9.1 .9

80 82.5 40

0

2.12 FS

2.29 FI

18.9 .6

AT

2.29

2.12

Avg Num Objects

Avg Num Objects

120

2.0

1.13

1.00 1.5

.07

.07

1.0

.65

.61 0.5

.44 0

.44 0

0.0 FS

(A)

FI

(B) Total Precision Writes

96.68

80 60

2.5

9.0 1.5

Avg Num Objects

Avg Num Objects

100

62.7

40 20 0

2.27 FS

2.40 FI

AT

22.9 .5

2.40

2.27

2.0

1.00

.92

1.5 1.0

.06 .45

0.5

.84

.06 .50

.84

0

0.0

FS

(C)

0 FI

(D)

Total Percentage of Resolved Dereferences to One Object Reads

100

80 68

63

60 37

33

40 20 FS

Pct Resolved

Pct Resolved

80

0

1

1

26 4 0

24 4 0

FI

69

0 AT

Nonvisible Local

66

60 40 20 0

(E)

Local

Writes

100

FS

48

46

1 10 10 0

1 9 10 0

FI

0 AT

(F)

Global

Formal Parameter

Heap

Fig. 6. Charts A – D provide the average precision over all benchmarks for reads and writes. (Charts B and D do not include the AT analysis to allow the diﬀerence between the FS and FI analyses to be visible.) Charts E and F report the percentage of dereferenced pointers that resolve to one object in our model.

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

69

Considering charts B and D of Fig. 3, it seems that FI is as precise as FS for pointers directed to nonvisible locals and formal parameters. Therefore, FS, if employed at all, should focus on pointers directed to the heap and global variables. The precision results for 099.go merit discussion. An average of 17.03 and 13.64 objects are returned for reads and writes, respectively, with a maximum of 100. This program contains six small list-processing functions (using an arraybased “cursor” implementation) that accept a pointer to the head of a list as a parameter. One of these functions, addlist, is called 404 times, passing the address of 100 diﬀerent actuals for the list header, resulting in 100 aliases for the formal parameter. However, because the lifetime of the formal is limited to this function (it does not call any other function), these relations are not propagated to any other function. Therefore, these relations do not suﬀer the eﬀects of the unrealizable path problem mentioned in Section 2.1. Another conclusion from the results is that analysis time is not only a function of program size; it also depends on the amount of alias relation propagation along the PCG and SEGs. For example, 099.go, despite being our largest program and having a pointer aliased with 100 objects, is analyzed at one of the fastest rates (3,037 LOC/second, 1,724 CFG nodes/second) because these relations are not propagated throughout the program. As suggested by Shapiro and Horwitz [41] and Diwan [12], a more precise and time-consuming alias analysis may not be as ineﬃcient as it may appear because the time required to obtain increased precision may reduce the time required by subsequent analyses that utilize mod-use information, and thus pointer alias information, as their input. As the previous paragraph suggests, this can also be true about pointer alias analysis itself, which also utilizes pointer alias information during its analysis. 3.3

Comparison with Other Results

Landi et al. [28] report precision results for the computation of the MOD problem using a ﬂow-sensitive pointer alias algorithm with limited context-sensitive information. Among the metrics they report is the number of “thru-deref” assigns, which corresponds to the “write” metrics reported in Fig. 3. However, since their results include compiler-introduced temporaries in their “thru-deref” count [26], a direct comparison is not possible. Stocks et al. [46] use the same metric without including temporaries for the ﬂow-sensitive context-sensitive analysis of Landi and Ryder [27]. They report the average number of objects ranges from 1.0 to 2.0 on the eight common benchmarks. On these benchmarks our ﬂow-sensitive context-insensitive analysis ranges from 1.0 to 2.22. Two possible explanations for the slightly less precise results are 1) their algorithm uses some context-sensitivity; 2) the underlying representation is not identical, and thus pointer dereferences may not be counted in the same manner in all cases. For example, statements such as cfree(TP) located in allroots are treated as modifying the structure deallocated, and thus as a pointer dereference [26]. In fact, on the three programs in which our analysis

70

M. Hind and A. Pioli

reports the same, or close to the same, number of “writes” as “thru-derefs” (allroots, lex315, simulator), our precision is identical to that reported in [46]. The relative precision of the ﬂow-insensitive analysis compared to the ﬂowsensitive analysis is in contrast to the study of Stocks et al. [46], which compares the ﬂow-sensitive analysis of Landi and Ryder [27] with a ﬂow-insensitive analysis described in [49]. For the eight common benchmarks, our ﬂow-insensitive algorithm ranges from 1.0 to 2.81 objects on average for a write dereference, compared to 1.0 to approximately 6.3 for the ﬂow-insensitive analysis they studied. The analysis described in [49] shares the property of Steensgaard’s analysis [45] in that it groups all objects pointed-to by a variable into an equivalence class. Although this can lead to spurious alias relations not present in the FI analysis, it does allow for an almost linear time algorithm, which has been shown to be fast in practice [45, 42, 50]. Emami et al. [14] report precision results for a ﬂow-sensitive context-sensitive algorithm. Their results range from 1.0 to 1.77 for all indirect accesses using a heap naming scheme that represents all heap objects with one name. Because we were unable to obtain the benchmarks from their suite, a direct comparison with our results is not possible. Ruf [36] reports both read and write totals for a ﬂow-sensitive contextinsensitive analysis. However, unlike our analysis he counts use of the “[ ]” operator on arrays that are not formal parameters as a dereference [38]. Since such an array will always point to the same place, the average number of objects is improved.11 For the 11 benchmarks in common,12 Ruf reports an overall read and write average of 1.33 and 1.38, respectively. To facilitate comparisons, we have also counted in this manner. The results for the common benchmarks are averages of 1.35 and 1.47 for the FS analysis and 1.41 and 1.54 for the FI analysis. We attribute the slight diﬀerences in the FS analysis to the diﬀerence in representations. As Ruf [36] states, “the VDG intermediate representation often coalesces series of structure or array operations into a single memory write.” This coalescing can skew results in either direction. Shapiro and Horwitz [41] present an empirical comparison of four ﬂowinsensitive algorithms. The ﬁrst algorithm is the same as the AT algorithm. The remaining three algorithms [45, 42, 1] can be less precise and more eﬃcient than the algorithms studied in this paper.13 The authors measure the precision 11

12

13

The best illustration of this is in 099.go, which has a large number of array references, but a low number of pointer dereferences. In this program, the average changed from 17.03 to 1.13 for reads and from 13.64 to 1.48 for writes when all uses of the “[ ]” operator were counted. Although [36] reports results for ft (under the name span), our version of the benchmark is substantially larger than the one Ruf analyzed, and thus is not comparable. In theory, the FI analysis can be more precise than Andersen’s algorithm [1] because it considers function scopes, at the cost of using more than one alias set. However, both algorithms are likely to oﬀer similar precision in practice because the distinguishing case is probably uncommon.

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

71

of these analyses by implementing three dataﬂow analyses and an interprocedural slicing algorithm. In addition to these alias analysis clients, the authors also report the direct precision of the alias analysis algorithms in terms of the total number of points-to relations. We agree with [14, 36] that a more meaningful metric is to measure where the points-to information is used, such as where a pointer is dereferenced. They conclude 1) a more precise ﬂow-insensitive analysis generally leads to increased precision by the subsequent analyses that use this information with varying magnitudes; 2) metrics measuring the alias analysis precision tend to be good predictors on the precision of subsequent analyses that use alias information; and 3) more precise ﬂow-insensitive analysis can also improve the eﬃciency of subsequent analyses that use this information.

4

Eﬃciency Improvements

This section describes some of the performance-improving enhancements made in the implementation and quantiﬁes their eﬀects on analysis-time speed-up for the ﬂow-sensitive algorithm. Although novelty is not claimed, the eﬃcacy of each technique is shown. #1

P(char* f)

a=5.0;

Entry {,,} p #2 while(...)

#7 g = &b;

#3 l = &a;

#4

if(...)

Exit

p

{,}

#5 b=10.0;

#6 P(g);

Fig. 7. Example CFG (left) and procedure P (right). Note that g is a global, a,b,l are locals of the CFG, and f is a formal parameter of P. Possible topological numbers (ignoring the back edge 7 → 2) appear next to each node.

4.1

Sharing Alias Sets

As described in Section 2.1, the ﬂow-sensitive analysis computes solutions before and after every node in the CFG. However, this can result in storing redundant

72

M. Hind and A. Pioli

information. For example, node #1 in Fig. 7 does not aﬀect any pointer value; therefore its Out set will always equal its In set. Likewise, a node whose predecessor(s) Out set(s) have the same value will have an In set equal to this value. For example, all nodes in Fig. 7 except #1 and #2. Thus, for all nodes, except #1 and #2, alias sets can be shared. We use the term shared in a literal way — if the In set is shared with the Out set, they are the same object during the analysis. This is done with a shallow copy of the alias set data structures. We precompute these sharing sites in a separate forward pass over the CFG before performing the alias analysis. Each node that has its own set, which we call a deep set, is dubbed a “SEG” (Sparse Evaluation Graph) node.14 Such nodes have a list of “SEG” predecessors and “SEG” successors that are used during the analysis. The alias set allocation strategy for a node, n, is summarized as follows: Outp , if ∀p, q ∈ P red(n), p and q share Out sets Inn = a deep set, otherwise Outn =

Inn , if n’s transfer function is the identity function a deep set, otherwise

Our current implementation treats all call nodes as SEG nodes. Table 2 reports the number of alias sets before and after this optimization was applied as well as the percentage reduction. In addition to saving storage (on average 73.78% fewer alias sets were allocated), this method saves visits during the actual analysis to nodes that can not aﬀect pointer aliasing. Although we allocate fewer alias sets, the real beneﬁt of this technique is seen during the analysis. Since we do not visit and propagate alias relations between extraneous CFG nodes, we simply have fewer alias relations being stored and copied from one CFG node to the next. This aﬀects both the analysis’s time and space use in a signiﬁcant way. Table 3 shows the eﬀects of the eﬃciency improving ideas on the ﬂow-sensitive analysis. Run times were collected as in Section 3, using the C function clock(), which gives the CPU time, not the elapsed time. This metric was chosen because it eliminates the eﬀects of system load, amount of RAM vs. paging space, etc. The elapsed time for each program was approximately 2.5 times the CPU time. The headings for each column are read vertically — for example, the second column shows the analysis time in seconds with no sharing of alias sets as well as no other enhancements described in this section. The third column reports the analysis time and resulting speed-up using this sharing technique. The eﬀectiveness of this technique is mostly related to the percentage of CFG nodes that aﬀect a pointer and percentage of call nodes — the higher the percentages the lower the potential beneﬁt.15 For all our benchmarks, these averages were 10.9% 14 15

The method described is a conservative approximation to the SEG of [9] and is similar to [35]. The percentage of merge nodes plays a smaller role.

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

73

Table 2. Eﬀectiveness of sharing alias sets Num. Alias Sets Benchmark No Sharing Sharing Pct Saved allroots 328 87 73.48% 052.alvinn 464 89 80.82% 01.qbsort 362 111 69.34% 15.trie 360 125 65.28% 04.bisect 352 78 77.84% 17.bintr 350 124 64.57% anagram 696 159 77.16% lex315 1088 260 76.10% ks 1054 317 69.92% 05.eks 694 170 75.50% 08.main 1166 399 65.78% 09.vor 1842 633 65.64% loader 1434 382 73.36% 129.compress 1152 195 83.07% ft 1140 380 66.67% football 5536 962 82.62% compiler 3486 699 79.95% assembler 3184 911 71.39% yacr2 3960 782 80.25% simulator 5180 1203 76.78% 099.go — 4966 — Average 73.78%

and 20.8% respectively. (09.vor had high averages in these categories and has the lowest eﬀectiveness for this technique while yacr2 had lower than average percentages and resulted in a higher speed-up.) Over all benchmarks, this technique results in a signiﬁcant speed-up, 2.80 on average. For our largest benchmark, 099.go, the analysis did not run (due to insuﬃcient memory) until we applied this optimization. 4.2

Worklists

The initial implementation of the analyses used an iterative algorithm for simplicity. After correctness was veriﬁed, a worklist-based implementation was used to improve eﬃciency. Two types of worklists were used: SEG node worklists and function worklists. Each function has a worklist of SEG nodes. The PCG has two worklists of functions: “current” and “next.” (The motivation for using two worklists will be described in Section 4.3.) We use nested “while not empty” loops with the worklists — the outer loop visiting functions and the inner loop visiting SEG nodes. The worklist of functions initially contains all functions. On the ﬁrst visit to a function, we initialize the function’s SEG node worklist with all SEG nodes in that function. If a SEG node’s Out set changes, all its SEG successors are placed on its function’s SEG node worklist. If a function’s entry set changes, it is placed on the “next” function worklist. If the exit set of a function changes, all calling functions are placed on the “next” function worklist and the calling call node(s) are placed on their respective function’s SEG node worklist. The analysis runs until all worklists are empty.

74

M. Hind and A. Pioli

Table 3. Flow-sensitive analysis run time in seconds. Numbers in parentheses are the speed-up from the previous version to the left. Benchmark Name allroots (227 LOC) 052.alvinn (272 LOC) 01.qbsort (325 LOC) 15.trie (358 LOC) 04.bisect (463 LOC) 17.bintr (495 LOC) anagram (650 LOC) lex315 (733 LOC) ks (782 LOC) 05.eks (1202 LOC) 08.main (1206 LOC) 09.vor (1406 LOC) loader (1539 LOC) 129.compress (1934 LOC) ft (2156 LOC) football (2354 LOC) compiler (2360 LOC) assembler (3446 LOC) yacr2 (3979 LOC) simulator (4639 LOC) 099.go (29637 LOC) Averages

←−Sharing −→ ←−Worklists −→ Overall Unsorted Sorted Speed-up ←−No Forward Bind Filter −→ Filter 1.41 0.54 0.21 0.24 0.10 (2.61) (2.57) (0.88) (2.40) 14.10 1.03 0.39 0.09 0.10 0.10 (2.64) (4.33) (0.90) (1.00) 10.30 8.22 4.16 1.81 1.19 0.75 (1.98) (2.30) (1.52) (1.59) 10.96 4.59 2.26 0.83 0.72 0.45 (2.03) (2.72) (1.15) (1.60) 10.20 1.47 0.45 0.16 0.13 0.14 (3.27) (2.81) (1.23) (0.93) 10.50 2.20 1.19 0.33 0.32 0.32 (1.85) (3.61) (1.03) (1.00) 6.88 6.67 2.16 0.59 0.52 0.48 (3.09) (3.66) (1.13) (1.08) 13.90 5.03 2.09 0.73 0.70 0.41 (2.41) (2.86) (1.04) (1.71) 12.27 20.48 9.40 3.32 2.55 1.49 (2.18) (2.83) (1.30) (1.71) 13.74 9.21 2.87 0.92 0.87 0.67 (3.21) (3.12) (1.06) (1.30) 13.75 96.37 44.80 18.30 12.39 3.16 (2.15) (2.45) (1.48) (3.92) 30.50 217.19 113.71 38.70 32.96 11.92 (1.91) (2.94) (1.17) (2.77) 18.22 176.49 58.37 26.71 21.20 3.81 (3.02) (2.19) (1.26) (5.56) 46.32 6.00 1.82 0.56 0.53 0.41 (3.30) (3.25) (1.06) (1.29) 14.63 80.94 37.07 14.04 11.25 5.09 (2.18) (2.64) (1.25) (2.21) 15.90 275.63 61.30 26.87 22.75 2.70 (4.50) (2.28) (1.18) (8.43) 102.09 39.31 12.87 4.19 4.11 3.70 (3.05) (3.07) (1.02) (1.11) 10.62 668.25 240.46 119.31 97.26 9.69 (2.78) (2.02) (1.23) (10.04) 68.96 377.11 86.50 26.76 25.93 10.64 (4.36) (3.23) (1.03) (2.44) 35.44 511.49 146.35 82.41 61.84 10.12 (3.49) (1.78) (1.33) (6.11) 50.54 — 228.18 74.01 65.23 9.83 — (3.08) (1.13) (6.64) 23.21 2.80 2.84 1.16 3.09 25.38

No Sharing ←−Iterating −→

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

75

Column 4 of Table 3 shows the improvement of the worklist-based implementation over the iterating version, both of which use shared alias sets. The result was an average speed-up of 2.84 over the iterating version, which produced an average speed-up of 2.78 over the nonshared iterating version. 4.3

Sorted Worklists

Using an iterating analysis of a forward data-ﬂow problem, it is natural to process the nodes in topological order. The next enhancement was to use priority queues (based on a topological order16 ) for the SEG and function worklists. Consider the case of a loop body that generates aliasing information. It would be best to process the loop body before moving on to the loop exit. Topological order alone does not give this property — we may process the loop exit before the loop body. (This would occur in Fig. 7 using the given node numbers as the topological order.) However, during the construction of the CFG, nodes for loop bodies are created before those nodes after the loop body. (Thus, node #7 would be created before node #3.) Since nodes are assigned numbers as they are created (which is performed in a topological order, except in the presence of gotos), using the nodes creation number as a priority ensures that loop bodies are processed ﬁrst. A result of using a single priority-based worklist of functions was that calling functions were visited before called functions. However, unlike SEG nodes, aliases can be propagated in both directions along a PCG edge. Thus, an optimal order of function visits is not apparent. In our benchmark suite, using a single prioritybased worklist for functions provided only a marginal improvement over the iterating version. To increase eﬃciency, we use two function worklists — “current” and “next.” While visiting functions on the “current” worklist, we place functions on the “next” worklist. This has the eﬀect that once a set of functions is on the worklist, the visiting order is ﬁxed in a topological order. When the “current” worklist is empty, we swap the “current” and “next” worklists and continue the analysis. The ﬁfth column of Table 3 reports the analysis time using sorted worklists for both SEG nodes and functions and the previous enhancements. This enhancement resulted in an average speed-up of 1.16 over the previous version, which used nonpriority-based worklists and shared alias sets. 4.4

ForwardBind Filtering

ForwardBind() is a function in our analysis that propagates alias relations from call nodes to the entry set of called functions. If needed, it creates alias relations for the formals from the actuals in the called function’s entry set and then unions in the call node’s In set with the called function’s entry set. Consider those alias relations that cannot be changed or used in the called function (for example, the 16

Topological order on a cyclic graph can be obtained by performing a depth-ﬁrst search of the graph and then removing edges classiﬁed as back edges [24].

76

M. Hind and A. Pioli

relation ∗l, a in Fig. 7), but are still propagated through the called function until they reach the exit set of the function. These relations are then propagated back to the call node’s Out set. In eﬀect, the called function acts as an identity transfer function for those relations that are not relevant in the called function. Although correct, this is ineﬃcient. Our enhancement is not to propagate from the call node those alias relations that cannot be reached in a called function. To compute the set of alias relations that are not reachable in the called function, F , we ﬁrst call ForwardBind(), which propagates alias relations into the entry set of F , EntryF , as described above. We then view all alias relations in EntryF as a directed graph and remove from this graph all vertices (distinguished objects), and associated edges (alias relations), that cannot be reached from a global or a formal of F . These removed edges (alias relations) are simply unioned into the call node’s Out set directly. This can help limit the propagation eﬀects of the unrealizable path problem. The last column of Table 3 reports the eﬀectiveness of this optimization. It provided the most dramatic speed-up, an average of 3.09 over the previous implementation, which used all other enhancements. Some programs, in particular football and assembler, had a much higher than average speed-up resulting from this ﬁltering. These two programs both shared some common characteristics: they have a single function that has both an unusually high amount of pointer aﬀecting statements and a very high number of called functions. Figure 8 shows the eﬀects of these optimizations in a dramatic way for the loader benchmark.17 We collected the data presented in this graph by repeatedly executing the “ps v” command (under AIX 4.1.5) while each of the ﬁve FS analyses was running. We recorded the SIZE column, which gives the virtual size of the data section of the process; this will capture all heap allocated memory usage during the analysis. The x-axis of the chart is simply marked oﬀ in samples (a sample is a single call to “ps v”). As the majority of our heap allocations are used to represent alias relations, the resulting memory usage can be interpreted as the number of alias relations stored by the analysis while running. The “No Sharing” version shows a characteristic curve; it grows quickly early on, but then levels oﬀ as the analysis reaches its ﬁxed point. The diﬀerence between the No Sharing and Sharing versions shows how the large reduction in the number of alias sets reduced the number of alias relations that were stored. The cumulative eﬀect of all ﬁve optimizations was an average speed-up of 25.38. This illustrates the beneﬁts that can be obtained by limiting the propagation of extraneous alias relations and the number of visits to functions and nodes.

5

Other Related Work

This section describes other related work not mentioned in Section 3.3. 17

The other benchmarks had similarly shaped graphs.

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

77

loader

25

No Sharing Sharing Worklists Priority Queues Filtering

20

MB

15

10

5

0

0

500

1000

1500

Samples

2000

2500

3000

3500

Fig. 8. Memory usage over time for the loader benchmark.

Ruf [36] presents an empirical study of two algorithms: a ﬂow-sensitive algorithm similar to the one we have implemented, and a context-sensitive version of the same algorithm. His results showed that the context-sensitive algorithm did not improve precision for pointers where they are dereferenced, but cautioned that this may be a characteristic of the benchmark suite analyzed. Wilson and Lam [48, 47] present an algorithm for performing context-sensitive analysis that avoids redundant analyses of functions for similar calling contexts. Ghiya and Hendren [17] present empirical data showing how points-to and connection analyses can improve traditional transformations, array dependence testing, and program understanding. Ruf [37] describes a program partitioning technique that is used for a ﬂowsensitive points-to analysis, achieving a storage savings of 1.3–7.2 over existing methods. Diwan et al. [13] provide static and dynamic measurements of the eﬀectiveness of three ﬂow-insensitive analyses for a type-safe language (Modula3). With the exception of AT, all three algorithms are less precise than the versions we have studied. Zhang et al. [50] report the eﬀectiveness of applying diﬀerent pointer aliasing algorithms to diﬀerent parts of a program. Hasti and Horwitz [18] present a pessimistic algorithm that attempts to increase the precision of a ﬂow-insensitive analysis by iterating over a ﬂowinsensitive analysis and an SSA [10] construction. No empirical results are reported. Horwitz [23] deﬁnes precise ﬂow-insensitive alias analysis and proves that, even in the absence of dynamic memory allocation, computing it is NPhard.

6

Conclusions

This work has described an empirical study of four pointer alias analysis algorithms that use varying degrees of ﬂow-sensitivity. We have found that

78

M. Hind and A. Pioli

– the address-taken analysis, although eﬃcient, is unlikely to provide suﬃcient precision; – the ﬂow-insensitive analysis with kill is not beneﬁcial; – the precision of the ﬂow-insensitive analysis is identical to that of the ﬂowsensitive analysis in 12 of 21 programs in our benchmark suite; – most published implementations of ﬂow-sensitive pointer analysis have equivalent precision. Although the ﬂow-sensitive analysis eﬃciently analyzed a program on the order of 30K LOCs, further benchmarks are needed to see if this property generalizes. We have also empirically demonstrated how various implementation strategies result in signiﬁcant analysis-time speed-up.

7

Acknowledgements

We thank Michael Burke, Michael Karasick, and Lee Nackman for their support of this work. We also thank Todd Austin, Bill Landi, and Rakesh Ghiya for making their benchmarks available. Bill Landi, Laurie Hendren, Erik Ruf, Barbara Ryder, and Bob Wilson provided useful details concerning their implementations. Michael Burke, Paul Carini, and Jong-Deok Choi provided useful discussions regarding the algorithms described in [8]. Discussions with Manuel F¨ ahndrich led to reporting intermediate read dereferences, which were not considered in [20]. Harini Srinivasan designed and implemented the initial control ﬂow graph builder, an important early system component. NPIC group members Robert Culley, Lynne Delesky, Lap Chung Lam, Giampaolo Lauria, Mark Nicosia, Joseph Perillo, Keith Sanders, Truong Vu, and Ming Wu assisted with the implementation and testing of the system. David Bacon helped with the initial design of the program call graph representation. Michael Burke, Jong-Deok Choi, John Field, G. Ramalingam, Harini Srinivasan, Laureen Treacy, and the anonymous referees provided useful comments on earlier drafts of this paper.

References [1] Lars Ole Andersen. Program Analysis and Specialization for the C Programming Language. PhD thesis, DIKU, University of Copenhagen, May 1994. Available at ftp.diku.dk/pub/diku/semantics/papers/D-203.dvi.Z. 70, 70 [2] Todd Austin. Pointer-intensive benchmark suite, version 1.1. http://www.cs.wisc.edu/˜austin/ptr-dist.html, 1995. 62 [3] Subra Balan and Walter Bays. Spec announces new benchmark suites cint92 and cfp92. Technical report, Systems Performance Evaluation Cooperative, March 1992. SPEC Newsletter 4(1). 62 [4] Michael Burke, Paul Carini, Jong-Deok Choi, and Michael Hind. Flow-insensitive interprocedural alias analysis in the presence of pointers. In K. Pingali, U. Banerjee, D. Gelernter, A. Nicolau, and D. Padua, editors, Lecture Notes in Computer

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

[5]

[6]

[7]

[8]

[9]

[10]

[11]

[12] [13]

[14]

[15]

[16]

[17]

[18]

79

Science, 892, pages 234–250. Springer-Verlag, 1995. Proceedings from the 7th Workshop on Languages and Compilers for Parallel Computing. Extended version published as Research Report RC 19546, IBM T. J. Watson Research Center, September 1994. 58, 58, 61, 62 Michael Burke, Paul Carini, Jong-Deok Choi, and Michael Hind. Interprocedural pointer alias analysis. Research Report RC 21055, IBM T. J. Watson Research Center, December 1997. 58, 58, 58, 59, 59, 60, 60, 61, 61, 62 Paul Carini, Michael Hind, and Harini Srinivasan. Flow-sensitive interprocedural type analysis for C++. Research Report RC 20267, IBM T. J. Watson Research Center, November 1995. 59 David R. Chase, Mark Wegman, and F. Kenneth Zadeck. Analysis of pointers and structures. In SIGPLAN ’90 Conference on Programming Language Design and Implementation, pages 296–310, June 1990. SIGPLAN Notices 25(6). 60, 60, 65 Jong-Deok Choi, Michael Burke, and Paul Carini. Eﬃcient ﬂow-sensitive interprocedural computation of pointer-induced aliases and side eﬀects. In 20th Annual ACM SIGACT-SIGPLAN Symposium on the Principles of Programming Languages, pages 232–245, January 1993. 58, 60, 60, 78 Jong-Deok Choi, Ron Cytron, and Jeanne Ferrante. Automatic construction of sparse data ﬂow evaluation graphs. In 18th Annual ACM Symposium on the Principles of Programming Languages, pages 55–66, January 1991. 59, 72 Ron Cytron, Jeanne Ferrante, Barry K. Rosen, Mark N. Wegman, and F. Kenneth Zadeck. Eﬃciently computing static single assignment form and the control dependence graph. ACM Transactions on Programming Languages and Systems, pages 451–490, October 1991. 77 Alain Deutsch. Interprocedural may-alias analysis for pointers: Beyond k-limiting. In SIGPLAN ’94 Conference on Programming Language Design and Implementation, pages 230–241, June 1994. SIGPLAN Notices, 29(6). 60 Amer Diwan. Personal communication, August 1997. 69 Amer Diwan, Kathryn S. McKinley, and J. Elliot B. Moss. Type-based alias analysis. In SIGPLAN ’98 Conference on Programming Language Design and Implementation, pages 106–117, June 1998. SIGPLAN Notices, 33(5). 77 Maryam Emami, Rakesh Ghiya, and Laurie J. Hendren. Context-sensitive interprocedural points-to analysis in the presence of function pointers. In SIGPLAN ’94 Conference on Programming Language Design and Implementation, pages 242–256, June 1994. SIGPLAN Notices, 29(6). 60, 60, 62, 70, 71 Rakesh Ghiya and Laurie J. Hendren. Connection analysis: A practical interprocedural heap analysis for C. In Lecture Notes in Computer Science, 1033, pages 515–533, August 1995. Proceedings from the 8th Workshop on Languages and Compilers for Parallel Computing. 60 Rakesh Ghiya and Laurie J. Hendren. Is it a tree, a dag or a cyclic graph? A shape analysis for heap-directed pointers in C. In 23rd Annual ACM SIGACTSIGPLAN Symposium on the Principles of Programming Languages, pages 1–15, January 1996. 60 Rakesh Ghiya and Laurie J. Hendren. Putting pointer analysis to work. In 25th Annual ACM SIGACT-SIGPLAN Symposium on the Principles of Programming Languages, pages 121–133, January 1998. 77 Rebecca Hasti and Susan Horwitz. Using static single assignment form to improve ﬂow-insensitive pointer analysis. In SIGPLAN ’98 Conference on Programming Language Design and Implementation, pages 97–105, June 1998. SIGPLAN Notices, 33(5). 77

80

M. Hind and A. Pioli

[19] Laurie J. Hendren and Alexandru Nicolau. Parallelizing programs with recursive data structures. IEEE Transactions on Parallel and Distributed Systems, 1(1):35– 47, January 1990. 60 [20] Michael Hind and Anthony Pioli. An empirical comparison of interprocedural pointer alias analyses. Research Report RC 21058, IBM T. J. Watson Research Center, December 1997. 78 [21] Michael Hind and Anthony Pioli. Assessing the eﬀects of the ﬂow-sensitivity on pointer alias analyses (extended version). Technical Report 98-104, SUNY at New Paltz, June 1998. 61, 65, 65, 67 [22] S. Horwitz, P. Pfeiﬀer, and T. Reps. Dependence analysis for pointer variables. In SIGPLAN ’89 Conference on Programming Language Design and Implementation, pages 28–40, June 1989. SIGPLAN Notices 24(6). 60, 60 [23] Susan Horwitz. Precise ﬂow-insensitive may-alias analysis is NP-Hard. ACM Transactions on Programming Languages and Systems, 19(1):1–6, January 1997. 77 [24] John B. Kam and Jeﬀrey D. Ullman. Global data ﬂow analysis and iterative algorithms. Journal of the ACM, 23(1):158–171, January 1976. 75 [25] William Landi. Undecidability of static analysis. ACM Letters on Programming Languages and Systems, 1(4):323–337, December 1992. 57 [26] William Landi. Personal communication, October 1997. 62, 69, 69 [27] William Landi and Barbara Ryder. A safe approximate algorithm for interprocedural pointer aliasing. In SIGPLAN ’92 Conference on Programming Language Design and Implementation, pages 235–248, June 1992. SIGPLAN Notices 27(6). 60, 69, 70 [28] William Landi, Barbara Ryder, and Sean Zhang. Interprocedural modiﬁcation side eﬀect analysis with pointer aliasing. In SIGPLAN ’93 Conference on Programming Language Design and Implementation, pages 56–67, June 1993. SIGPLAN Notices 28(6). 62, 62, 69 [29] James R. Larus and Paul N. Hilﬁnger. Detecting conﬂicts between structure accesses. In SIGPLAN ’88 Conference on Programming Language Design and Implementation, pages 21–34, 1988. SIGPLAN Notices, 23(7). 60 [30] Thomas Marlowe, William Landi, Barbara Ryder, Jong-Deok Choi, Michael Burke, and Paul Carini. Pointer-induced aliasing: A clariﬁcation. SIGPLAN Notices, 28(9):67–70, September 1993. 60 [31] Thomas J. Marlowe, Barbara G. Ryder, and Michael G. Burke. Deﬁning ﬂow sensitivity in data ﬂow problems. Technical Report RC 20138, IBM T. J. Watson Research Center, July 1995. 58 [32] Lee R. Nackman. Codestore and incremental C++. Dr. Dobbs Journal, pages 92–95, December 1997. 61 [33] Anthony Pioli. Conditional pointer aliasing and constant propagation. Master’s thesis, SUNY at New Paltz, 1998. In preparation. 61 [34] G. Ramalingam. The undecidability of aliasing. ACM Transactions on Programming Languages and Systems, 16(6):1467–1471, November 1994. 57 [35] G. Ramalingam. On sparse evaluation representations. In Pascal Van Hentenryck, editor, Lecture Notes in Computer Science, 1302, pages 1–15. Springer-Verlag, 1997. Proceedings from the 4th International Static Analysis Symposium. 72 [36] Erik Ruf. Context-insensitive alias analysis reconsidered. In SIGPLAN ’95 Conference on Programming Language Design and Implementation, pages 13–22, June 1995. SIGPLAN Notices, 30(6). 62, 62, 62, 65, 70, 70, 70, 71, 77

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

81

[37] Erik Ruf. Partitioning dataﬂow analyses using types. In 24th Annual ACM SIGACT-SIGPLAN Symposium on the Principles of Programming Languages, pages 15–26, January 1997. 77 [38] Erik Ruf. Personal communication, October 1997. 62, 62, 70 [39] Mooly Sagiv, Thomas Reps, and Reinhard Wilhelm. Solving shape-analysis problems in languages with destructive updating. In 23rd Annual ACM SIGACTSIGPLAN Symposium on the Principles of Programming Languages, pages 16–31, January 1996. 60 [40] Mooly Sagiv, Thomas Reps, and Reinhard Wilhelm. Solving shape-analysis problems in languages with destructive updating. ACM Transactions on Programming Languages and Systems, 20(1):1–50, January 1998. 60 [41] Marc Shapiro and Susan Horwitz. The eﬀects of the precision of pointer analysis. In Pascal Van Hentenryck, editor, Lecture Notes in Computer Science, 1302, pages 16–34. Springer-Verlag, 1997. Proceedings from the 4th International Static Analysis Symposium. 69, 70 [42] Marc Shapiro and Susan Horwitz. Fast and accurate ﬂow-insensitive point-to analysis. In 24th Annual ACM SIGACT-SIGPLAN Symposium on the Principles of Programming Languages, pages 1–14, January 1997. 61, 70, 70 [43] D. Soroker, M. Karasick, J. Barton, and D. Streeter. Extension mechanisms in Montana. In 8th IEEE Israeli Conference on Software and Systems, pages 119– 128, June 1997. 61 [44] SPEC. SPEC CPU95, Version 1.0. Standard Performance Evaluation Corporation, http://www.specbench/org, August 1995. 62 [45] Bjarne Steensgaard. Points-to analysis in almost linear time. In 23rd Annual ACM SIGACT-SIGPLAN Symposium on the Principles of Programming Languages, pages 32–41, January 1996. 61, 70, 70, 70 [46] Philip A. Stocks, Barbara G. Ryder, William A. Landi, and Sean Zhang. Comparing ﬂow and context sensitivity on the modiﬁcations-side-eﬀects problem. In International Symposium on Software Testing and Analysis, pages 21–31, March 1998. 69, 70, 70 [47] Robert P. Wilson. Eﬃcient Context-Sensitive Pointer Analysis for C Programs. PhD thesis, Stanford University, December 1997. 77 [48] Robert P. Wilson and Monica S. Lam. Eﬃcient context-sensitive pointer analysis for C programs. In SIGPLAN ’95 Conference on Programming Language Design and Implementation, pages 1–12, June 1995. SIGPLAN Notices, 30(6). 60, 77 [49] Sean Zhang, Barbara G. Ryder, and William Landi. Program decomposition for pointer aliasing: A step toward practical analyses. In 4th Symposium on the Foundations of Software Engineering, pages 81–92, October 1996. 70, 70 [50] Sean Zhang, Barbara G. Ryder, and William Landi. Experiments with combined analysis for pointer aliasing. In ACM SIGPLAN-SIGSOFT Workshop on Program Analysis for Software Tools and Engineering, pages 11–18, June 1998. 70, 77

Analysis of Normal Logic Programs Fran¸cois Fages and Roberta Gori LIENS CNRS, Ecole Normale Sup´erieure, 45 rue d’Ulm, 75005 Paris, France, [email protected] Dipartimento di Informatica, Universit` a di Pisa, Corso Italia 40, 56125 Pisa, Italy [email protected] Ph.: +39-50-887248 Fax: +39-50-887226

Abstract. In this paper we present a dataflow analysis method for normal logic programs interpreted with negation as failure or constructive negation. We apply our method to a well known analysis for logic programs: the depth(k) analysis for approximating the set of computed answers. The analysis is correct w.r.t. SLDNF resolution and optimal w.r.t. constructive negation. Keywords: Abstract interpretation, static analysis, logic programming, constructive negation.

1

Introduction

Important results have been achieved for static analysis using the theory of abstract interpretation [6]. Abstract interpretation is a general theory for specifying and validating program analysis. A key point in abstract interpretation is the choice of a reference semantics from which one can abstract the properties of interest. While it is always possible to use the operational semantics, it is possible to get rid of useless details, by choosing a more abstract semantics as reference semantics. In the case of definite logic programs, much work has been done in this sense. Choosing the most abstract logical least model semantics limits the analysis to type inference properties, that approximate the ground success set. Non-ground model semantics have thus been developed, under the name of S-semantics [2], and proved useful for a wide variety of goal-independent analysis ranging from groundness, to sharing, call patterns, etc. All the intermediate fixpoint semantics between the most abstract logical one and the most concrete operational one, form in fact a hierarchy related by abstract interpretation, in which one can define a notion of the best reference semantics [12] for a given analysis. On the other hand, less work has been done on the analysis of normal logic programs, although the finite failure principle, and hence SLDNF resolution, are standard practice. The most significant paper on the analysis of normal G. Levi (Ed.): SAS’98, LNCS 1503, pp. 82–98, 1998. c Springer-Verlag Berlin Heidelberg 1998

Analysis of Normal Logic Programs

83

logic programs, using the theory of abstract interpretation, is the one by Marriott and Søndergaard [19], which proposes a framework based on Fitting’s least three-valued model semantics [11]. Since this reference semantics is a ground semantics, the main application of this framework is type analysis. Marriott and Søndergaard already pointed out that a choice of a different reference semantics could lead to an improved analysis. Fitting’s least three-valued model semantics is, in fact, an abstraction (a non recursively enumerable one, yet easier to define) of Kunen’s three-valued logical semantics [14] which is more faithful to SLDNF resolution [15] and complete w.r.t. constructive negation. These are exactly the directions along which we try to improve the results of Marriott and Søndergaard. We consider the inference rule of constructive negation, which provides normal logic programs with a sound and complete [21] operational semantics w.r.t. Kunen’s logical semantics [14]. We propose an analysis method for normal logic programs interpreted with constructive negation, based on the generalized S-semantics given in [9] and on the hierarchy described in [10]. We present here an analysis based on the depth(k) domain which approximates the computed answers obtained by constructive negation and therefore the three-valued consequences of the program completion and CET (Clark’s equational theory). Other well known analyses for logic programs can be extended to normal logic programs. For example, starting from a suitable version of Clark’s semantics a groundness analysis was defined which is correct and optimal w.r.t. constructive negation. Here, for lack of space, we present only the depth(k) analysis. We show that it is correct and also optimal w.r.t. constructive negation. Finally we give an example which shows that in the case of type inference properties our semantics yields a result which is more precise than the one obtained by Marriott and Søndergaard. From the technical point of view, the contribution of the paper is the definition of a normal form for first order constraints on the Herbrand Universe, which is suitable for analysis. In fact the normal form allows us to define an abstraction function which is a congruence w.r.t. the equivalence on constraints induced by the Clark’s equality theory. The paper is organized as follows. In Sect. 2 we introduce some preliminary notions on constructive negation. In Sect. 3 we define a normal form on the concrete domain of constraints in order to deal, with equivalence classes of constraints w.r.t. the Clark’s equational theory. Section 4 defines the abstract domain and abstract operator and show its correctness and optimality (under suitable assumptions on the depth of the cut) w.r.t. the concrete one. Finally, Subsect. 4.5 shows an example.

2

Preliminaries

The reader is assumed to be familiar with the terminology of and the basic results in the semantics of logic programs [1,17] and with the theory of abstract interpretation as presented in [6,7].

84

2.1

F. Fages and R. Gori

Normal Logic Programs and Constructive Negation

We consider the equational version of normal logic programs, where a normal program is a finite set of clauses of the form A ← c|L1 , ..., Ln , where n ≥ 0, A is an atom, called the head, c is a conjunction of equalities, and L1 , ..., Ln are literals. The local variables of a program clause are the free variables in the clause which do not occur in the head. With V ar(A) we intend the free variables in the atom A. In order to deal with constructive negation, we need to consider the domain C of full first-order equality constraints on the structure H of the Herbrand domain. Assuming an infinite number of function symbols in the alphabet, Clark’s equational theory (CET) provides a complete decidable theory for the constraint language [18,14], i.e. 1. (soundness) H |= CET , 2. (completeness) for any constraint c, either CET |= ∃c or CET |= ¬∃c. A constraint is in prenex form if all its quantifiers are in the head. The set of free variables in a constraint c is denoted by V ar(c). For a constraint c, we shall use the notation ∃c (resp. ∀c) to represent the constraint ∃X c (resp. ∀X c) where X = V ar(c). A constrained atom is a pair c|A where c is an H-solvable constraint such that V ar(c) ⊆ V ar(A). The set of constrained atoms is denoted by B. A constrained interpretation is a subset of B. A three-valued or partial constrained interpretation is a pair of constrained interpretations < I + , I − >, representing the constrained atoms which are true and false respectively (note that because of our interest in abstract interpretations we do not impose any consistency condition between I + and I − ). Constructive negation is a rule of inference introduced by Chan for normal logic programs in [3], which provides normal logic programs with a sound and complete [21] operational semantics w.r.t. Kunen’s logical semantics [14]. In logic programming, Kunen’s semantics is simply the set of three-valued consequences of the program’s completion and the theory CET . The S-semantics of definite logic programs [2] has been generalized to normal logic programs in [9] for a version of constructive negation, called constructive negation by pruning. The idea of the fixpoint operator, which captures the set of computed answer constraints, is to consider a non-ground finitary (hence continuous) version of Fitting’s operator. Here we give a definition of the operator TPBD which is parametric w.r.t. the domain BD of constrained atoms and the operations on constraints on the domain D. Definition 1. Let P be a normal logic program. TPBD is an operator over P(BD )× P(BD ) defined by

Analysis of Normal Logic Programs

85

TPBD (I)+ = {c|p(X) ∈ BD : there exists a clause in P with local variables Y , C = p(X) ← d|A1 , ..., Am , ¬Am+1 , ..., ¬An . c1 |A1 , ..., cm |Am ∈ I + , cm+1 |Am+1 , ..., cn |An ∈ I − such that c = ∃Y (d ∧c1 ∧ . . . ∧cn )} TPBD (I)− = {c|p(X) ∈ BD : for each clause defining p in P with loc. var. Yk , Ck = p(X) ← dk |Ak,1 , ..., Ak,mk , βk . there exist ek,1 |Ak,1 , ..., ek,mk |Ak,mk ∈ I − , nk ≥ mk , ek,mk+1 |Ak,mk+1 , ..., ek,nk |Ak,nk ∈ I + , where for mk+1 ≤ j ≤ nk , ¬Ak,j occurs in βk , V such that c = k ∀Yk (¬ dk ∨ ek,1 . . . ∨ ek,nk )}. where the operations ∃, ∀, ¬, ∨, ∧, are the corresponding operations on the constraint domain of D. In the case of a normal logic program, the operator TPB defines a generalized S-semantics which is fully abstract w.r.t. the computed answer constraints with constructive negation by pruning [9]. By soundness it approximates also the set of computed answer constraints under the SLDNF resolution rule, or under the Prolog strategy. In [10] we have shown that this operator defines a hierarchy of reference semantics related by abstract interpretation, that extends the hierarchy defined by Giacobazzi for definite logic programs [12]. Here we show the use of the hierarchy for the static analysis of normal logic programs.

3

Normal Forms in CET

Unlike the semantics in Marriott and Søndergaard’s framework, our reference semantics is a non ground semantics and has to deal with first-order equality constraints. The first problem that arises is to define a normal form for such constraints on the Herbrand domain, so that abstraction functions on constrained atoms can be defined. In general, in fact, given a theory th, we are interested in working with equivalence classes of constraints w.r.t. the equivalence of the constraints in th. Namely c is equivalent to c0 if th |= c ↔ c0 . Therefore we need the abstraction function on the concrete constraint domain to be a congruence. This is a necessary property since it permits to be independent from the syntactic form of the constraints. Dealing with normal logic programs, we need to achieve this property in CET. We thus need to introduce a normal form for first-order equality constraints, in a similar way to what has been done for the analysis of definite programs where the normal form is the unification solved form [16]. Here we shall define a new notion of “false-simplified” normal forms, making use of Colmerauer’s solved forms for inequalities [4], Maher’s transformations for first-order constraints [18] and an extended disjunctive normal form [13]. First let us motivate the need of a “false-simplified” form. Let us call a basic constraint an equality or an inequality between a variable and a term. The abstraction function will be defined inductively on the basic constraints, and it

86

F. Fages and R. Gori

will sometimes (e.g. for groundness analysis) abstract to true some inequalities. Consider, for example, the following constraint d = ∀X(Y = b ∧ X 6= f (a)). d is H-equivalent to f alse. If the abstraction of X 6= f (a) is true then the abstraction of d will be the abstraction of Y = b, which cannot be H-equivalent to the abstraction of f alse. Therefore we need to define a normal form where the constraints which are H-equivalent to f alse, are eliminated. Definition 2. Consider a constraint d in prenex disjunctive form, d = ∆(∨i Ai ), where ∆ is a sequence of quantified variables and ∨i Ai is a finite disjunction. d is in a f alse-simplified form if, either there does not exist a proper subset I of the i0 s such that H |= ∆(∨i Ai ) ↔ ∆(∨i∈I Ai ), or such an I exists and there exists also a subset K of I, such that ∨j6∈I Aj is H-equivalent to ∨k∈K Ak . The latter condition assures that we really eliminate constraints that are Hequivalent to f alse and that are not just redundant in the constraint. Now the existence of a false-simplified form for any constraint can be proved simply with the following: Algorithm 3 Input: a constraint in prenex disjunctive form d = ∆(∨i Ai ). Let us call U the set of the indices i’s in d = ∆(∨i Ai ). 1. Let I and J be the partition of U such that i ∈ I if H |= ∃∆(Ai ), otherwise i ∈ J. 2. Repeat I := I ∪S as long as there exists an S ⊆ J such that H |= ∃∆(∨i∈S Ai ) and for all j ∈ S H 6|= ∃∆(∨i∈(S\{j}) Ai ). 3. Let S ⊆ J\I be any minimal set such that H |= ∃∆(∨s∈S As ∨i∈I Ai ) and H |= ∆(∨s∈S As ∨i∈I Ai ) ↔ d, do I := I ∪ S, 4. Output: ∆(∨i∈I Ai ). The idea of the algorithm is to find a subset of the conjunctions Ai ’s (those with i ∈ I) such that ∆(∨i∈I Ai ) is in false-simplified form and it is H-equivalent to ∆(∨i Ai ). In the first step we select the Ai ’s such that ∆(Ai ) is H-satisfiable. In this case, in fact, Ai cannot be H-equivalent to f alse and it can be put in the set I. In the second step from the remaining Ai ’s we select the set of Ai ’s such that their ∆ quantified disjunction is H-satisfiable, since we check that all the Ai ’s are necessary for the quantified disjunction to be H-satisfiable, the considered Ai ’s can not be H-equivalent to f alse. At the end of this process, if the resulting constraint is H-equivalent to the input constraint, we stop. Otherwise, we add a minimum number of the not yet selected Ai ’s such the ∆(∨i Ai ) for the selected i’s is H-equivalent to the input constraint. Since we add a minimum number of not yet selected Ai ’s, we are sure that the resulting constraint is in falsesimplified form. Example 4 shows how the algorithm 3 works on two examples. Example 4. 1. Input: c1 = ∀T (A1 ∨ A2 ∨ A3 ∨ A4 ), A1 = (T = f (H) ∧ Y = a), A2 = (T 6= f (a) ∧ Y = b), A3 = (Y 6= g(H, T )), A4 = (T 6= a ∧ Y = a). I1 = {3}. I2 = {3, 1, 2}. I3 = {3, 1, 2}( since H |= ∀T (∨i∈I2 Ai ) ↔ c1 ). Output: ∀T (A1 ∨ A2 ∨ A3 ).

Analysis of Normal Logic Programs

87

2. Input: c2 = ∀T (A01 ∨ A02 ∨ A03 ), A01 = (T 6= f (U ) ∧ T 6= f (V )), A02 = (T = H), A03 = (U 6= V ∧ T = f (a)). I1 = {}. I2 = {1, 2}. I3 = {1, 2, 3}( since H 6|= ∀T (∨i∈I2 A0i ) ↔ c2 ).1 Output: ∀T (A01 ∨ A02 ∨ A03 ). Theorem 5. For any input constraint c = ∆(∨i Ai ), algorithm 3 terminates and computes a false-simplified form logically equivalent to c. Note that all the false-simplified forms of a constraint c are H-equivalent. Now the intuitive idea for a normal form is the following. We put a constraint in prenex form and we compute the disjunctive form of its quantifier free part. We make equality and inequality constraints interact in every conjunction of the disjunctive form and then we compute the false-simplified form for the resulting constraint. The problem is that if we consider a standard disjunctive normal form, we would not be able to see explicitly all the relations between constraints in disjunctions. Consider, for example, the constraint (X = f (H) ∨ (H 6= f (a)). This constraint is equivalent, therefore H-equivalent, to the constraint ((X = f (f (a))∧H = f (a))∨H 6= f (a)). Note that the equality H = f (a) is not explicit in the first disjunction. Since the abstraction function will act on the terms of the disjunction independently, this could cause a problem. This is why we will use a well known extended disjunctive form defined for Boolean algebra and applied, in our case, to the algebra of quantifier free constraints. In the next theorem with Bi we denote basic equality or inequality constraints (X = t or X 6= t). For any Bi let Bi f alse = ¬Bi and Bi true = Bi . Theorem 6. [13] For every Boolean formula φ on basic equality or inequality constraints B1 , . . . , B Wn , φ ↔ ψ where ψ = ( (a1 ,... ,an )∈{f alse,true}n φ(a1 , . . . , an ) ∧ B1a1 ∧ . . . ∧ Bnan ). Note that ψ is a formula in disjunctive form. ψ has in fact a particular disjunctive form where each conjunction contains all the basic constraints (possibly negated) of φ. This is why, this form is able to capture all the possible relations between the different terms of a disjunction. We will call the formula ψ the extended disjunctive normal form (dnf ) of φ. The next example shows how the extended disjunctive normal form works on a constraint c1 . Example 7. c1 = (X = f (H) ∨ H 6= f (a)). dnf (c1 ) = ((X = f (f (a)) ∧ H = f (a))∨ (X = f (H) ∧ H 6= f (a)) ∨ (X 6= f (H) ∧ H 6= f (a))). Note that although c1 , dnf (c1 ) and ((X = f (f (a)) ∧ H = f (a)) ∨ H 6= f (a)) are H-equivalent, dnf (c1 ) is the most “complete” in the sense that it shows syntactically all the relations between constraints in disjunctions. 1

U = b, H = f (b), V = a, in fact, is an assignment (for the free variables of c2 ), which is a solution of c2 but is not a solution of ∀T (∨i∈I2 A0i ).

88

F. Fages and R. Gori

We can now define the normal form, Res(c), of a first-order equality constraint c, as the result of the following steps: 1. Put the constraint c in prenex form obtaining ∆(c1 ), where ∆ is a sequence of quantified variables and c1 is the quantifier free part of c. 2. Compute dnf (c1 ) = ∨(Ai ), 3. Simplify each conjunction Ai obtaining A0i = ResConj(Ai ), 4. Return a f alse-simplified form of the constraint ∆(∨A0i ). where the procedure for simplifying each conjunction is based on Maher’s canonical form [18] and Colmerauer’s simplification algorithm for inequalities [4]. The procedure performs the following steps, ResConj(A) 1. compute a unification solved form for the equalities in the conjunction A 2. for each equality X = t in A, substitute X by t at each occurrence of X in the inequalities of conjunction A. 3. simplify the inequalities by applying the following rules, a) replace f (t1 , . . . , tn ) 6= f (s1 , . . . , sn ) by t1 6= s1 ∨ . . . ∨ tn 6= sn . b) replace f (t1 , . . . , tn ) 6= g(s1 , . . . , sn ) by true. c) replace t = 6 x by x 6= t if t is not a variable. obtaining A0 , 4. if A0 is a conjunction then return A0 . 5. otherwise compute dnf (A) = ∨(Ai ) and return ∨ResConj(Ai ). It is worth noting that the previous algorithm terminates since each constraint contains a finite number of inequalities. Example 8 shows how the procedure Res computes the normal form of some constraints. Example 8. 1. c = (X = f (Y ) ∧ (Y = a ∨ Y = f (a)) ∧ ∀U X 6= f (f (U ))). = ∀U ( X = f (Y ) ∧ (Y = a ∨ Y = f (a)) ∧ X 6= f (f (U ))). c1 = ∀U ( (X = f (Y ) ∧ Y 6= a ∧ Y = f (a) ∧ X 6= f (f (U )))∨ c2 (X = f (Y ) ∧ Y = a ∧ Y 6= f (a) ∧ X 6= f (f (U )))∨ (X = f (Y ) ∧ Y = a ∧ Y = f (a) ∧ X 6= f (f (U )))). c3.1 = ∀U ( (X = f (f (a)) ∧ Y 6= a ∧ Y = f (a) ∧ X 6= f (f (U )))∨ (X = f (a) ∧ Y = a ∧ Y 6= f (a) ∧ X 6= f (f (U ))). c3.2 = ∀U ( (X = f (f (a)) ∧ f (a) 6= a ∧ Y = f (a) ∧ f (f (a)) 6= f (f (U )))∨ (X = f (a) ∧ Y = a ∧ a 6= f (a) ∧ f (a) 6= f (f (U ))). c3.3 = ∀U ( (X = f (f (a)) ∧ Y = f (a) ∧ a 6= U )∨ (X = f (a) ∧ Y = a ∧ a 6= f (U ))). c4 = (X = f (a) ∧ Y = a).

Analysis of Normal Logic Programs

89

2. c¯ = (X = f (Z, S) ∧ U = (f (H), H) ∧ S = a ∧ X 6= U ). = c1 = c¯. c2 c3.1 = (X = f (Z, a) ∧ U = f (f (H), H) ∧ S = a ∧ X 6= U ). c3.2 = (X = f (Z, a) ∧ U = f (f (H), H) ∧ S = a ∧ f (Z, a) 6= f (f (H), H)). c3.3 = (X = f (Z, a) ∧ U = f (f (H), H) ∧ S = a ∧ (Z 6= f (H) ∨ H 6= a). c3.4 = A1 ∨ A2 ∨ A3 . A1 = (X = f (Z, a) ∧ U = f (f (H), H) ∧ S = a ∧ Z = f (H) ∧ H 6= a). A2 = (X = f (Z, a) ∧ U = f (f (H), H) ∧ S = a ∧ Z 6= f (H) ∧ H 6= a). A3 = (X = f (Z, a) ∧ U = f (f (H), H) ∧ S = a ∧ Z 6= f (H) ∧ H = a). ResConj(A1 ) = A¯1 ResConj(A2 ) = A2 ResConj(A3 ) = A¯3 .

c4

A¯1 = (X = f (f (H), a) ∧ U = f (f (H), H) ∧ S = a ∧ Z = f (H) ∧ H 6= a). A¯3 = (X = f (Z, a) ∧ U = f (f (a), a) ∧ S = a ∧ Z 6= f (a) ∧ H = a). = A¯1 ∨ A2 ∨ A¯3 .

Note that all the steps in ResConj and Res preserve the H-equivalence, the third step of ResConj is Colmerauer’s simplification algorithm for inequalities [4], the first and second transformations are the usual ones for CET formulas [18], while the second step of Res is the extended disjunctive normal transformation [13]. Hence we get: Proposition 9. H |= φ ↔ Res(φ). Our concrete constraints domain N C will be the subset of constraints in C which are in normal form. The concrete operations on N C will be thus defined using the normal form: Definition 10. Let c1 , c2 ∈ N C, c1 ∨c c2 = Res(c1 ∨ c2 ) c1 ∧c c2 = Res(c1 ∧ c2 ) c ¬ c1 = Res(¬c1 ) ∃c X c1 = ∃X c1

∀c X c1 = ∀X c1

We denote by B the set of constrained atoms with constraints in N C, and by (I, ⊆) the complete lattice of (not necessarily consistent) partial constrained interpretations formed over B.

4

Depth(k) Analysis for Constructive Negation

The idea of depth(k) analysis was first introduced in [20]. The domain of depth(k) analysis was then used in order to approximate the ground success and failure sets for normal programs in [19]. We follow the formalization of [5] for positive logic programs. We want to approximate an infinite set of computed answer constraints by means of a constraint depth(k) cut, i.e. constraints where the equalities and inequalities are between variables and terms which have a depth not greater than k.

90

F. Fages and R. Gori

Our concrete domain is the complete lattice of partial constrained interpretations (I, ⊆) of the previous section. Since our aim is to approximate the computed answer constraints, the fixpoint semantics we choose in the hierarchy [10] is the one which generalizes the S-semantics to normal logic programs, the TPBD operator (cf def. 1). The version we consider here is the one defined on the domain B with the concrete operations in N C, ∧c , ∨c , ¬c , ∃c , ∀c , (the TPB operator). 4.1

The Abstract Domain

Terms are cut by replacing each-subterm rooted at depth greater than k by a new fresh variable taken from a set W , (disjoint from V ). The depth(k) terms represent each term obtained by instantiating the variables of W with terms built over V . Consider the depth function || : T erm → T erm such that 1 if t is a constant or a variable |t| = max{|t1 |, . . . , |tn |} + 1 if t = f (t1 , . . . , tn ) and a given positive integer k. The abstract term αk (t) is the term obtained from the concrete one by substituting a fresh variable (belonging to W ) to each subterm t0 in t, such that |t| − |t0 | = k. Consider now the abstract basic constraints c | c = (X = t) |t| ≤ k or ABC = c = (X 0 6= t0 ) |t0 | ≤ k, and V ar(t0 ) ∩ W = ∅ Note that V ar(t0 ) ∩ W = ∅ expresses the fact that inequalities between variables and cut terms are not allowed. The domain of abstract constraints is defined as follows, Definition 11. AN C =

c | c is a constraint in normal form built with the logical connectives ∨, ∧, ∀ and ∃ on ABC

The concepts of abstract constrained atoms and partial abstract interpretations are defined as expected. Definition 12. An abstract constrained atom is a pair c|A such that c ∈ AN C and c is a H − solvable constraint, A is an atom and V ar(c) ⊆ V ar(A). With B a we intend the set of abstract constrained atoms. The abstract domain is the set of partial interpretations on abstract constrained atoms. A partial abstract constrained interpretation for a program, is a pair of + − set of abstract constrained atoms, I a =< I a , I a >, not necessary consistent. We consider I a = {I a | I a is a partial interpretation}. With respect to the case of definite logic programs [5], we need to define a different order on the abstract constraint domain.

Analysis of Normal Logic Programs

91

This is because the result ca of an abstract and operation on the abstract constraint domain will be an approximation of the result c of the concrete and operation on the concrete constraint domain, in the sense that ca will be “more general” than the abstraction of c (where here “more general” means “is implied under H”) . This motivates the definition of the following relation on the abstract constraint domain. Definition 13. Let c, c0 ∈ AN C. c a c0 if H |= c → c0 . We consider the order ≤a induced by the preorder a , namely the order obtained considering the classes modulo the equivalence induced by a . We define the downward closure of a pair of sets w.r.t. the ≤a order, Definition 14. Consider a pair of sets of constrained atoms B. By ↓ B we denote the downward closure of < B + , B − >. c|A ∈↓ B + if there exists c0 |A ∈ B + and c ≤a c0 , c|A ∈↓ B − if there exists c0 |A ∈ B − and c ≤a c0 . Intuitively, a set of constrained atoms I is less or equal than J, if ↓ I ⊆↓ J. Definition 15. Consider I, J ∈ I a . + + I a J a ↔ for all c|A ∈ I a ∃c0 |A ∈ J a such that c ≤a c0 and − − for all c|A ∈ I a ∃c0 |A ∈ J a such that c ≤a c0 It is immediate to see that defines a preorder. We consider the order ≤ induced by the preorder , namely the order obtained by considering the classes Ia modulo the equivalence induced by . Then our abstract domain is (Ia , ≤). Since the operations on the equivalence classes are independent on the choice of the representative, we denote the class of an interpretation I a by I a itself. In the rest of the paper, we often abuse notation by denoting by I a the equivalence class of I a or the interpretation I a itself. 4.2

The Abstraction Function

Let us now define the abstraction function. To this aim we first define the function αc on constraints. The main idea is to define αc on the basic constraints as follows: an equality X = t is abstracted to X = αk (t), while an inequality X 6= t is abstracted to X 6= t if |t| ≤ k and to true otherwise. We denote by ∆(c) the constraint c0 in normal form and by ∆ the sequence of quantified variables of c0 , where c is the quantifier-free part of c0 . Definition 16. The depth(k) constraint abstraction function is the function αc : N C → AN C: αc (∆(c)) = ∆, ∆0 αc (c) where ∆0 = ∃Y1 , ∃Y2 , .., and Yi ∈ (W ∩ V ar(αc (c))) αc (X = t) = (X = αk (t)), αc (f alse) = f alse, αc (true) = true, αc (X 6= t) = true if |t| > k, αc (X 6= t) = (X 6= t) if |t| ≤ k, αc (A ∧ B) = αc (A) ∧ αc (B), αc (A ∨ B) = αc (A) ∨ αc (B).

92

F. Fages and R. Gori

Note that the first definition means that all the new variables introduced by the cut terms have to be considered existentially quantified. Example 17 shows an application of αc . Example 17. c = ∀U ((H = f (f (T )) ∧ T 6= f (f (U )) ∧ X = f (U ))∨ (H = f (f (T )) ∧ T 6= f (X) ∧ X 6= f (U ))), k = 2. αc (c) = αc ( ∀U (

(H = f (f (T )) ∧ T 6= f (f (U )) ∧ X = f (U ))∨ (H = f (f (T )) ∧ T 6= f (X) ∧ X 6= f (U )))) = ∀U ( (αc (H = f (f (T ))) ∧ αc (T 6= f (f (U ))) ∧ αc (X = f (U )))∨ (αc (H = f (f (T ))) ∧ αc (T 6= f (X)) ∧ αc (X 6= f (U ))) = ∀U, ∃Q1 , Q2 ((H = f (Q1 ) ∧ true ∧ X = f (U ))∨ (H = f (Q2 ) ∧ T 6= f (X) ∧ X 6= f (U ))) (Q1 , Q2 ∈ W ).

The abstraction function α is defined by applying αc to every constraint of the constrained atoms in the concrete interpretation. Definition 18. Let α : I → Ia : α =< α+ , α− > α+ (I) = {c|A | c0 |A ∈ I + and αc (c0 ) = c}, α− (I) = {c|A | c0 |A ∈ I − and αc (c0 ) = c}. As a consequence the function γ on (equivalence classes of) sets of abstract constraints is automatically determined as follows: Definition 19. Let γ : Ia → I: γ(I a ) = ∪{I | α(I) ≤ I a } = + ∪{I | ∀c|A ∈ α+ (I) ∃c0 |A ∈ I a such that c ≤a c0 and − ∀c|A ∈ α− (I) ∃c0 |A ∈ I a such that c ≤a c0 } = ∪{I | ↓ α(I) ⊆↓ I a } = ∪{I |α(I) ⊆↓ I a } Lemma 20. α is additive. Theorem 21. < α, γ > is a Galois insertion of (I, ⊆) into (Ia , ≤). 4.3

αc is a Congruence w.r.t. the H-Equivalence

As we have already pointed out in Sect. 3, we want to work with H-equivalence classes of constraints and, for this purpose, we need to be sure that the above defined function αc on N C is a congruence w.r.t. the H-equivalence. This means that if two constraints c, c0 ∈ N C are H-equivalent, then also αc (c) and αc (c0 ) have to be H-equivalent. In order to understand whether two constraints are H-equivalent, it is useful to state the following result. Lemma 22. Consider the inequality X 6= t. There exist no arbitrary quantified t1 , . . . , tn , where ti 6= t, such that X 6= t is H-equivalent to ∧i X 6= ti .

Analysis of Normal Logic Programs

93

This is a consequence of the fact that we consider the models of the theory CET without the DCA axiom. The previous result, together with the fact that constraints are in falsesimplified form, allows us to claim that αc is a congruence. Theorem 23. Let c, c0 ∈ N C. If H |= c ↔ c0 then H |= αc (c) ↔ αc (c0 ). 4.4

The Abstract Fixpoint Operator

We now define the abstract operations that will replace the concrete ones in the definition of the fixpoint abstract operator. We show that the abstract operations are a correct approximation of the concrete operations. The definition of the abstract and operation is not immediate. The example 24 is meant to give some intuition on some problems that may arise. Example 24. Consider the following two constraints: c1 = (X = f (Z, f (H)) ∧ S = f (a)) c2 = (U 6= X ∧ Y 6= f (S)) and k = 2. Consider αc (c1 ) = ∃Q(X = f (Z, Q) ∧ S = f (a)) αc (c2 ) = (U 6= X ∧ Y 6= f (S)). If we now consider the normalized form of αc (c1 )∧αc (c2 ) the resulting constraint is ∃Q(U 6= f (Z, Q) ∧ Y 6= f (f (a)) ∧ X = f (Z, Q) ∧ S = f (a)), which is not an abstract constraint according to definition 11. The problem is that the normalized form of the logical and operation on two abstract constraints is not in general an abstract constraint (the depth of the terms involved in equalities and inequalities can be greater than k and it can contain inequalities between variables and cut terms). This is the reason why we need to define a new M operator, on the normalized forms of abstract constraints. The M operator must cut terms deeper than k and replace by true all the inequalities which contain a cut term. Intuitively this is because X 6= t, where V ar(t) ∩ W 6= ∅, represents, on the concrete domain, an inequality between a variable and a term longer than k. On the abstract domain, such inequalities are abstracted to the constant true. Definition 25. Let M : N C → AN C M(∆(c)) = ∆, ∆0 M(c) where ∆0 = ∃Y1 , ∃Y2 , .., where Yi ∈ (W ∩V ar(M(c))). M(X = t) = (X = αk (t)) M(X 6= t) = (X 6= t) if |t| ≤ k and V ar(t) ∩ W = ∅ M(X 6= t) = (true) if |t| > k or V ar(t) ∩ W 6= ∅ M(A ∧ B) = αc (A) ∧ αc (B), M(A ∨ B) = αc (A) ∨ αc (B) As expected, the M operator is similar to the αc operator. The only difference is that M replace by true all the inequalities between variables and cut terms. Since AN C is a subset of N C, the Res form is defined also on the abstract constraints domain. Definition 26. Let c1 , c2 ∈ AN C ˜ c2 = M(Res(c1 ∧ c2 )), ˜ c2 = M(Res(c1 ∨ c2 )), c1 ∨ c1 ∧ ˜ c1 = ∃X c1 , ¬ ˜ c1 = M(Res(¬c1 )), ∃X

˜ c1 = ∀X c1 , ∀X

94

F. Fages and R. Gori

It is worth noting that the procedure Res on the abstract domain needs to perform the logical and on abstract constraints. This means that most of the observations that can be done on the behavior of the abstract and operation, concern also the abstract or and not operations. Example 27 illustrates the relation between the abstract and operation and the abstraction of the concrete and operation. For a sake of simplicity, since in this case it does not affect the result, we write the constraint c1 in the more compact standard disjunctive form rather than of in extended disjunctive form. Example 27. c1 = ∀K((Y = a ∧ U 6= f (f (K))) ∨ Z = a), c2 = (U = f (f (a))). Consider k = 1. αc (c1 ) = (Y = a ∨ Z = a), αc (c2 ) = ∃V U = f (V ). ˜ αc (c2 ) = ∃V ((Y = a ∧ U = f (V )) ∨ (Z = a ∧ U = f (V ))). αc (c1 )∧ αc (Res(c1 ∧ c2 )) = ∃V (Z = a ∧ U = f (V )). ˜ αc (c2 ) H |= αc (Res(c1 ∧ c2 )) → αc (c1 )∧ As already pointed out, the abstract and gives a more general constraint than the abstraction of the one computed by the concrete and and this is the reason why we have defined an approximation order based on implication (under H) between constraints. In order to show that the abstract operations are correct, we prove a stronger property. Theorem 28. Let c1 , c2 ∈ N C. ˜ αc (c2 ) ≥a αc (c1 ∧c c2 ), ˜ αc (c2 ) ≥a αc (c1 ∨c c2 ), αc (c1 )∨ αc (c1 )∧ ˜ αc (c1 ) = αc (∀c x c1 ). ˜ αc (c1 ) = αc (∃c x c1 ), ∀x ∃x As shown by example 29, the correctness property does not hold for the version of abstract “not” which we have defined, if we consider general constraints. Example 29. Consider c1 = (X 6= f (f (a))) and k = 1. αc (¬c (c1 )) = ∃Y X = f (Y ) which does not implies ¬ ˜ (αc (c1 )) = f alse. Since the not operator is used by the abstract fixpoint operator on “simpler “ constraints (the program constraints) only, we are interested in its behavior on conjunctions of equalities between variables and terms only. For this kind of constraints the following result holds. V ˜ αc (c1 ) ≥a αc (¬c (c1 )). Lemma 30. If c1 = ( i (Xi = ti )) ∈ N C, then ¬ Now that we have defined the abstract constraints domain and the abstract operations, we can define the abstract fixpoint operator. Definition 31. Let α(P ) be the program obtained by replacing every constraint c in a clause of P by αc (c). a Ba The abstract fixpoint operator: Ia → Ia is defined as follows, TPB (I a ) = Tα(P ) (↓ ˜ ∀, ˜ ¬ ˜, ∧ ˜ on AN C × AN C. ˜ on AN C and ∨ I a ), where the operations are ∃, a

By definition, TPB is a congruence respect to the equivalence classes of the a abstract domain. Note also that TPB is monotonic on the (Ia , ≤), because I ≤ J implies ↓ I ⊆↓ J.

Analysis of Normal Logic Programs

95

a

Lemma 32. TPB is monotonic on the (Ia , ≤). The proof that the abstract operator is correct w.r.t. the concrete one, is based on the correctness of the abstract operations on the abstract constraints domain. a

a

Theorem 33. α(TPB (γ(I a ))) ≤ TPB (I a ). Then α(lf p(TPB )) ≤ TPB (I a ). Consider now a k greater than the maximal depth of the terms involved in the constraints of the clauses in the program P . In this case the abstract operator is also optimal. a

Theorem 34. TPB (I a ) ≤ α(TPB (γ(I a ))). Let us finally discuss termination properties of the dataflow analyses presented in this section. First note that the set of not equivalent (w.r.t. H) set of constraints belonging to AN C is finite. Lemma 35. Assume that the signature of the program has a finite number of function and predicate symbols. Our depth(k) abstraction is ascending chain finite. 4.5

An Example

We now show how the depth-k analysis works on an example.The program of figure 1 computes the union of two sets represented as lists. We denote the a a equivalence class of TPB by TPB itself. All the computed constraints for the predicate ¬member are shown, while concerning the predicate ¬union, for a sake of simplicity, we choose to show just a small subset of the computed answer constraints (written in the more compact standard disjunctive form). Therefore, the concretization of the set of answer constraints for ¬union that we present in figure 1, contains some answer constraints computed by the concrete semantics but not all of them. As expected the set of answer constraints, computed by the abstract fixpoint operator, is an approximation of the answer constraints, computed by the concrete operator, for the predicates member, union and ¬member. For example, for the predicate ¬member(X, Y ), we compute the answer ∀L(Y 6= [X, L]) which correctly approximates the concrete answer ∀L, H, H1 , L1 (Y 6= [X, L] ∧ Y 6= [H, H1 , L1 ]). While the constraint answer ∃X∀H1 , L1 ∃Z1 , Z2 (A = [X, Z1 ] ∧ C = [X, Z2 ]∧B 6= [H1 , L1 ]) for union(A, B, C), approximates the concrete constraint A = [X, X], C = [X, X, K], B = K and B is not a list, computed by the concrete semantics. Note, in fact, that, if the second argument is not a list, the predicate member finitely fails. Let us now consider Marriott and Søndergaard’s abstraction for the program P , with a language where the only constant is a (this assumption does not affect the result). Concerning the predicate union with the empty list as first argument, their abstraction computes the following atoms union([ ], a, a), union([ ], [ ], [ ]), union([ ], [a], [a]), union([ ], [a, Z1 ], [a, Z2 ]), while we obtain the more precise answer (A = [ ] ∧ B = C)|union(A, B, C).

96

F. Fages and R. Gori

P : union(A, B, C) : −A = [ ], B = C. union(A, B, C) : −A = [X, L], C = [X, K], ¬member(X, B), union(L, B, K). union(A, B, C) : −A = [X, L], member(X, B), union(L, B, C). member(X, Y ) : −Y = [X, L]. member(X, Y ) : −Y = [H, L], member(X, L). Consider now a depth-2 analysis with Zi ∈ W . a+

∃L(

TPB Y = [X, L]

)|member(X, Y ).

∃H, Z1 (

Y = [H, Z1 ]

)|member(X, Y ).

A=[]∧B =C

|union(A, B, C).

∃X, L(

A = [X] ∧ B = [X, L] ∧ B = C

)|union(A, B, C).

∃X, Y, Z1 (

A = [X] ∧ B = [Y, Z1 ] ∧ B = C

)|union(A, B, C).

∃X, H, L, Z1 (

A = [X, Z1 ] ∧ B = [X, L] ∧ B = C

)|union(A, B, C).

∃X, H, Z1 , Z2 (

A = [X, Z1 ] ∧ B = [H, Z2 ] ∧ B = C

)|union(A, B, C).

∃X, K∀H, L( A = [X] ∧ C = [X, K] ∧ B 6= H [ , L] ∧ B = K )|union(A, B, C). ∃X∀H, L∃Z1 , Z2 (

A = [X, Z1 ] ∧ C = [X, Z2 ] ∧ B 6= H [ , L]

)|union(A, B, C).

∃X, K∀L( A = [X] ∧ C = [X, K] ∧ B 6= X [ , L] ∧ B = K )|union(A, B, C). ∃X∀L∃Z1 , Z2 (

A = [X, Z1 ] ∧ C = [X, Z2 ] ∧ B 6= X [ , L]

A subset of TPB

a−

)|union(A, B, C).

(complete for the predicate member)

∀H, L(

Y 6= H [ , L]

)|member(X, Y ).

∀L(

Y 6= X [ , L]

)|member(X, Y ).

∀X, L, KX1 , L1

((A 6= [ ]∧ A 6= X [ , L])∨ (B 6=C ∧ A 6= X [ , L]∨ (B 6=C ∧ C 6= X [ , K]) ∧ A 6= X [ 1 , L1 ]) )|union(A, B, C).

∀X, L, KX1 , L1 , H, L2

((A 6= [ ]∧ A 6= X [ , L])∨ (B 6=C ∧ A 6= X [ , L]∨ (B 6=C ∧ C 6= X [ , K]) ∧ A 6= X [ 1 , L1 ])∨ (B 6=C ∧ C 6= X [ , L] ∧ B 6= H [ , L2 ])∨ (A 6= [ ]∧ C 6= X [ , L] ∧ B 6= H [ , L2 ]) .. .

Fig. 1. Example 1

|)union(A, B, C)

Analysis of Normal Logic Programs

97

The atom union([ ], [a, Z1 ], [a, Z2 ]), in fact, correctly approximates the predicates deeper than k which have a successful behavior, but it has lost the relation between B and C. As a consequence all the other ground atoms for union computed using the atom union([ ], [a, Z1 ], [a, Z2 ]), are less precise than the ground instances of the atoms computed by our non-ground abstract semantics.

5

Conclusion

Starting from the hierarchy of semantics defined in [10], our aim was to show that well known analysis for logic programs could be extended to normal logic programs. Based on the framework of abstract interpretation [7,8], we have presented a depth(k) analysis which is able to approximate the answer set of normal logic programs. It is worth noting that our depth(k) analysis, can be easily generalized to constraint logic programs defined on H, whose program constraints can be conjunctions of equalities and inequalities. In order to deal with constructive negation, in fact, most of the results presented in this paper hold for first order equality constraints. The only exception is lemma 30 (and consequently theorem 33 and theorem 34), which is true only for conjunctions of equalities. But a more complex definition of the abstract not operator can be defined and proven correct on conjunctions of equalities and inequalities constraints. This alternative definition is, however, less precise than the one defined here. As a consequence theorem 33, where the abstract fixpoint operator uses the new abstract not operator, still holds for such “extended” logic programs, while it is not the case for theorem 34.

References 1. K. R. Apt. Introduction to Logic Programming. In J. van Leeuwen, editor, Handbook of Theoretical Computer Science, volume B: Formal Models and Semantics, pages 495–574. Elsevier and The MIT Press, 1990. 2. A. Bossi, M. Gabbrielli, G. Levi, and M. Martelli. The s-semantics approach: Theory and applications. Journal of Logic Programming, 19–20:149–197, 1994. 3. D. Chan. Constructive Negation Based on the Completed Database. In R. A. Kowalski and K. A. Bowen, editors, Proc. Fifth Int’l Conf. on Logic Programming, pages 111–125. The MIT Press, 1988. 4. A. Colmerauer. Equations and inequations on finite and infinite trees. In Proceedings of the International Conference on Fifth Generation Computer System, pages 85–99, 1984. 5. M. Comini. Abstract Interpretation framework for Semantics and Diagnosis of Logic Programs. PhD thesis, Dipartimento di Informatica, Universit` a di Pisa, 1998. 6. P. Cousot and R. Cousot. Abstract Interpretation: A Unified Lattice Model for Static Analysis of Programs by Construction or Approximation of Fixpoints. In Proc. Fourth ACM Symp. Principles of Programming Languages, pages 238–252, 1977.

98

F. Fages and R. Gori

7. P. Cousot and R. Cousot. Systematic Design of Program Analysis Frameworks. In Proc. Sixth ACM Symp. Principles of Programming Languages, pages 269–282, 1979. 8. P. Cousot and R. Cousot. Abstract Interpretation and Applications to Logic Programs. Journal of Logic Programming, 13(2 & 3):103–179, 1992. 9. F. Fages. Constructive negation by pruning. Journal of Logic Programming, 32(2):85–118, 1997. 10. F. Fages and R. Gori. A hierarchy of semantics for normal constraint logic programs. In M.Hanus M.Rodriguez-Artalejo, editor, Proc. Fifth Int’l Conf. on Algebraic and Logic Programming, volume 1139 of Lecture Notes in Computer Science, pages 77–91. Springer-Verlag, 1996. 11. M. Fitting. A Kripke-Kleene semantics for logic programs. Journal of Logic Programming, 2:295–312, 1985. 12. R. Giacobazzi. “Optimal” collecting semantics for analysis in a hierarchy of logic program semantics. In C. Puech and R. Reischuk, editors, Proc. 13th International Symposium on Theoretical Aspects of Computer Science (STACS’96), volume 1046 of Lecture Notes in Computer Science, pages 503–514. Springer-Verlag, 1996. 13. S. Koppelberg. Handbook of Boolean Algebras (Vol.I). Elsevier Science Publisher B.V.(North Holland), 1989. 14. K. Kunen. Negation in logic programming. Journal of Logic Programming, 4:289– 308, 1987. 15. K. Kunen. Signed Data Dependencies in Logic Programs. Journal of Logic Programming, 7(3):231–245, 1989. 16. J.-L. Lassez, M. J. Maher, and K. Marriott. Unification Revisited. In J. Minker, editor, Foundations of Deductive Databases and Logic Programming, pages 587– 625. Morgan Kaufmann, Los Altos, Ca., 1988. 17. J. W. Lloyd. Foundations of Logic Programming. Springer-Verlag, 1987. Second edition. 18. M.J. Maher. Complete axiomatizations of the algebra of finite, rational and infinite trees. In Third Symp. on Logic in Computer Science, pages 348–357, 1988. 19. K. Marriott and H. Sondergaard. Bottom-up Dataflow Analysis of Normal Logic Programs. Journal of Logic Programming, 13(2 & 3):181–204, 1992. 20. T. Sato and H. Tamaki. Enumeration of Success Patterns in Logic Programs. Theoretical Computer Science, 34:227–240, 1984. 21. P. Stuckey. Negation and constraint logic programming. Information and Computation, 118(1):12–33, 1995.

The Correctness of Set-Sharing Patricia M. Hill1 , Roberto Bagnara2? , and Enea Zaffanella3 1

School of Computer Studies, University of Leeds, Leeds, LS2 9JT, United Kingdom [email protected] 2 Dipartimento di Matematica, Universit` a degli Studi di Parma, Italy. [email protected] 3 Servizio IX Automazione, Universit` a degli Studi di Modena, Italy. [email protected]

Abstract. It is important that practical data flow analysers are backed by reliably proven theoretical results. Abstract interpretation provides a sound mathematical framework and necessary generic properties for an abstract domain to be well-defined and sound with respect to the concrete semantics. In logic programming, the abstract domain Sharing is a standard choice for sharing analysis for both practical work and further theoretical study. In spite of this, we found that there were no satisfactory proofs for the key properties of commutativity and idempotence that are essential for Sharing to be well-defined and that published statements of the safeness property assumed the occur-check. This paper provides a generalisation of the abstraction function for Sharing that can be applied to any language, with or without the occur-check. The results for safeness, idempotence and commutativity for abstract unification using this abstraction function are given. Keywords: abstract interpretation, logic programming, occur-check, rational trees, set-sharing.

1

Introduction

Today, talking about sharing analysis for logic programs is almost the same as talking about the set-sharing domain Sharing of Jacobs and Langen [8, 9]. Researchers are primarily concerned with extending the domain with linearity, freeness, depth-k abstract substitutions and so on [2, 4, 12, 13, 16]. Key properties such as commutativity and soundness of this domain and its associated abstract operations are normally assumed to hold. The main reason for this is that [9] not only includes a proof of the soundness but also refers the reader to the thesis of Langen [14] for proofs of commutativity and idempotence. In abstract interpretation, the concrete semantics of a program is approximated by an abstract semantics. In particular, the concrete domain is replaced ?

Much of this work was supported by EPSRC grant GR/L19515.

G. Levi (Ed.): SAS’98, LNCS 1503, pp. 99−114, 1998. Springer-Verlag Berlin Heidelberg 1998

100

P.M. Hill, R. Bagnara, and E. Zaffanella

by an abstract domain and each elementary operation on the concrete domain is replaced by a corresponding abstract operation on the abstract domain. Thus, assuming the global abstract procedure mimics the concrete execution procedure, each operation on elements in the abstract domain must produce an approximation of the corresponding operation on corresponding elements in the concrete domain. The key operation in a logic programming derivation is unification (unify) and the corresponding operation for an abstract domain is aunify. An important step in standard unification algorithms is the occur-check that avoids the generation of infinite data structures. However, in computational terms, it is expensive and it is well known that Prolog implementations by default omit this check. Although standard unification algorithms that include the occur-check produce a substitution that is idempotent, the resulting substitution when the occur-check is omitted, may not be idempotent. In spite of this, most theoretical work on data-flow analysis of logic programming assume the result of unify is always idempotent. In particular both [9] and [14] assume in their proofs of soundness that the concrete substitutions are idempotent. Thus their results do not apply to the analysis of all Prolog programs. If two terms in the concrete domain are unifiable, then unify computes the most general unifier (mgu). Up to renaming of variables, an mgu is unique. Moreover a substitution is defined as a set of bindings or equations between variables and other terms. Thus, for the concrete domain, the order and multiplicity of elements are irrelevant in both the computation and semantics of unify. It is therefore useful that the abstraction of the unification procedure should be unaffected by the order and multiplicity in which it abstracts the bindings that are present in the substitution. Furthermore, from a practical perspective, it is useful if the global abstract procedure can proceed in a different order to the concrete one without affecting the accuracy of the analysis results. Hence, it is extremely desirable that aunify is also commutative and idempotent. However, as discussed later in this paper, only a weak form of idempotence has ever been proved while the only previous proof of commutativity [14] is seriously flawed. As sharing is normally combined with linearity and freeness domains that are not idempotent or commutative, [2, 12] it may be asked why these properties are important for sharing. In answer to this, we observe that the order and multiplicity in which the bindings in a substitution are analysed affects the accuracy of the linearity and freeness domains. It is therefore a real advantage to be able to ignore these aspects as far as the sharing domain is concerned. This paper provides a generalisation of the abstraction function for Sharing that can be applied to any language, with or without the occur-check. The results for safeness, idempotence and commutativity for abstract unification using this abstraction function are given. Detailed proofs of the results stated in this paper are available in [7]. In the next section, the notation and definitions needed for equality and substitutions in the concrete domain are given. In Section 3, we introduce a new concept called variable-idempotence that generalises idempotence to allow for rational trees. In Section 4, we recall the definition of Sharing and define its

The Correctness of Set-Sharing

101

abstraction function, generalised to allow for non-idempotent substitutions. We conclude in Section 5.

2 2.1

Equations and Substitutions Notation

For a set S, # S is the cardinality of S, ℘(S) is the powerset of S, whereas ℘f (S) is the set of all the finite subsets of S. The symbol Vars denotes a denumerable set of variables, whereas TVars denotes the set of first-order terms over Vars for some given set of function symbols. The set of variables occurring in a syntactic object o is denoted by vars(o). 2.2

Substitutions

If x ∈ Vars and s ∈ TVars , then x 7→s is called a binding. A substitution is a total function σ : Vars → TVars that is the identity almost everywhere; in other words, the domain of σ, def dom(σ) = x ∈ Vars σ(x) 6= x is finite. If t ∈ TVars , we write tσ to denote σ(t). Substitutions by the set of are denoted their bindings, thus σ is identified with the set x 7→ σ(x) x ∈ dom(σ) . The composition of substitutions is defined in the usual way. Thus τ ◦ σ is the substitution such that, for all terms t, (τ ◦ σ)(t) = τ (σ(t)). A substitution is said circular if it has the form {x1 7→x2 , . . . , xn−1 7→xn , xn 7→x1 }. A substitution is in rational solved form if it has no circular subset. The set of all substitutions in rational solved form is denoted by Subst. 2.3

Equations

An equation is of the form s = t where s, t ∈ TVars . Eqs denotes the set of all equations. We are concerned in this paper to keep the results on sharing as general as possible. In particular, we do not want to restrict ourselves to a specific equality theory. Thus we allow for any equality theory T over TVars that includes the basic axioms denoted by the following schemata. s = s, s = t ⇐⇒ t = s, r = s ∧ s = t =⇒ r = t, f (s1 , . . . , sn ) = f (t1 , . . . , tn ) ⇐⇒ s1 = t1 , . . . , sn = tn .

(1) (2) (3) (4)

Of course, T can include other axioms. For example, it is usual in logic programming and most implementations of Prolog to assume an equality theory

102

P.M. Hill, R. Bagnara, and E. Zaffanella

based on syntactic identity and characterised by the axiom schemata given by Clark [3]. This consists of the basic axioms together with the following: ¬f (s1 , . . . , sn ) = g(t1 , . . . , tm ) ∀z ∈ Vars ∀t ∈ (TVars \ Vars) : z ∈ vars(t) =⇒ ¬(z = t).

(5) (6)

The identity axioms characterised by the schemata 5 ensure the equality theory is Herbrand and depends only on the syntax. Equality theory for a non-Herbrand domain replaces these axioms by ones that depend instead on the semantics of the domain. Axioms characterised by the schemata 6 are called the occur-check axioms and are an essential part of the standard unification procedure in SLDresolution. An alternative approach used in some implementations of Prolog, does not require the occur-check axioms. This approach is based on the theory of rational trees [5, 6]. It assumes the basic axioms and the identity axioms together with a set of uniqueness axioms [10, 11]. These state that each equation in rational solved form uniquely defines a set of trees. Thus, an equation z = t where z ∈ vars(t) and t ∈ (TVars \ Vars) denotes the axiom (expressed in terms of the usual first-order quantifiers [15]): ∀x ∈ Vars : z = t ∧ (x = t{z 7→x} =⇒ z = x) . The basic axioms defined by schemata 1, 2, 3, and 4, which are all that are required for the results in this paper, are included in both these theories. A substitution σ may be regarded as a set of equations { x = t | x 7→t ∈ σ }. A set of equations e ∈ ℘f (Eqs) is unifiable if there is σ ∈ Subst such that T ` (σ =⇒ e). σ is called a unifier for e. σ is said to be a relevant unifier of e if vars(σ) ⊆ vars(e). That is, σ does not introduce any new variables. σ is a most general unifier for e if, for every unifier σ 0 of e, T ` (σ 0 =⇒ σ). An mgu, if it exists, is unique up to the renaming of variables. In this paper, mgu(e) always denotes a relevant unifier of e.

3

Variable-Idempotence

It is usual in papers on sharing analysis to assume that all the substitutions are idempotent. Note that a substitution σ is idempotent if, for all t ∈ TVars , tσσ = tσ. However, the sharing domain is just concerned with the variables. So, to allow for substitutions representing rational trees, we generalise idempotence to variable-idempotence. Definition 1. A substitution σ is variable-idempotent if ∀t ∈ TVars : vars(tσσ) = vars(tσ). The set of all variable-idempotent substitutions is denoted by VSubst.

The Correctness of Set-Sharing

103

It is convenient to use the following alternative characterisation of variableidempotence: A substitution σ is variable-idempotent if and only if, ∀(x 7→t) ∈ σ : vars(tσ) = vars(t). Thus any substitution consisting of a single binding is variable-idempotent. Moreover, all idempotent substitutions are also variable-idempotent. Example 1. The substitution x 7→ f (x) is not idempotent but is variable idempotent. Also, x 7→f (y), y 7→z is not idempotent or variable-idempotent but is equivalent (with respect to some equality theory T ) to x 7→f (z), y 7→z , which is idempotent. S

We define the transformation 7−→ ⊆Subst × Subst, called S-transformation, as follows: (x 7→t) ∈ σ

(y 7→s) ∈ σ x 6= y σ 7−→ σ \ {y 7→s} ∪ {y 7→s[x/t]} S

Any substitution σ can be transformed to a variable-idempotent substitution σ 0 for σ by a finite sequence of S-transformations. Furthermore, if the substitutions σ and σ 0 are regarded as equations, then they are equivalent with respect to any equality theory that includes the basic equality axioms. These two statements are direct consequences of Lemmas 1 and 2, respectively. Lemma 1. Let T be an equality theory that satisfies the basic equality axioms. S Suppose that σ and σ 0 are substitutions such that σ 7−→σ 0 . Then, regarding σ 0 0 and σ as sets of equations, T ` (σ ⇐⇒ σ ). Proof. Suppose that (x 7→ t), (y 7→ s) ∈ σ where x 6= y and suppose also σ 0 = σ \ {y 7→s} ∪ {y 7→s[x/t]}. We first show by induction on the depth of the term s that x = t =⇒ s = s[x/t]. Suppose s has depth 1. If s is x, then s[x/t] = t and the result is trivial. If s is a variable distinct from x or a constant, then s[x/t] = s and the result follows from equality Axiom 1. Suppose now that s = f (s1 , . . . , sn ) and the result holds for all terms of depth less than that of s. Then, by the inductive hypothesis, for each i = 1, . . . , n, x = t =⇒ si = si [x/t]. Hence, by Axiom 4, x = t =⇒ f (s1 , . . . , sn ) = f s1 [x/t], . . . , sn [x/t] and hence x = t =⇒ f (s1 , . . . , sn ) = f (s1 , . . . , sn )[x/t].

104

P.M. Hill, R. Bagnara, and E. Zaffanella

Thus, combining this result with Axiom 3, we have {x = t, y = s} =⇒ x = t, y = s, s = s[x/t] =⇒ x = t, y = s[x/t] . Similarly, combining this result with Axioms 2 and 3, x = t, y = s[x/t] =⇒ x = t, y = s[x/t], s = s[x/t] =⇒ {x = t, y = s}. t u Note is necessary. For example, suppose that the conditionx 6= y in Lemma1 σ = x 7→f (x) and σ 0 = x 7→f (f (x)) . Then we do not have σ 0 =⇒ σ. Lemma 2. Suppose that, for each j = 0, . . . , n: σj = {x1 7→t1,j , . . . , xn 7→tn,j }, where tj,j = tj,j−1 and if j > 0, for each i = 1, . . . , n, where i 6= j, ti,j = ti,j−1 [xj /tj,j−1 ]. Then, for each j = 0, . . . , n, νj = {x1 7→t1,j , . . . , xj 7→tj,j } is variable-idempotent and, if j > 0, σj can be obtained from σj−1 by a sequence of S-transformations. Proof. The proof is by induction on j. Since ν0 is empty, the base case when j = 0 is trivial. Suppose, therefore that 1 ≤ j ≤ n and the hypothesis holds for νj−1 and σj−1 . By the definition of νj , we have νj = {xj 7→tj,j−1 } ◦ νj−1 . Consider an arbitrary i, 1 ≤ i ≤ j. We will show that vars(ti,j νj ) = vars(ti,j ). Suppose first that i = j. Then since tj,j = tj,j−1 , tj,j−1 = tj,0 νj−1 and, by the inductive hypothesis, vars(tj,0 νj−1 νj−1 ) = vars(tj,0 νj−1 ), we have vars(tj,j νj ) = vars tj,0 νj−1 νj−1 {xj 7→tj,j } = vars tj,0 νj−1 {xj 7→tj,j } = vars tj,j {xj 7→tj,j } = vars(tj,j ). Suppose now that i 6= j. Then,

vars(ti,j ) = vars ti,j−1 {xj 7→tj,j−1 } .

and, by the inductive hypothesis, vars(ti,j−1 νj−1 ) = vars(ti,j−1 ). If xj ∈ / vars(ti,j−1 ), then vars(ti,j νj−1 ) = vars ti,j−1 {xj 7→tj,j−1 }νj−1 = vars(ti,j−1 νj−1 ) = vars(ti,j ).

The Correctness of Set-Sharing

105

On the other hand, if xj ∈ vars(ti,j−1 ), then vars(ti,j νj−1 ) = vars ti,j−1 {xj 7→tj,j−1 }νj−1 = vars(ti,j−1 νj−1 ) \ {xj } ∪ vars(tj,j−1 νj−1 ) = vars(ti,j−1 ) \ {xj } ∪ vars(tj,j−1 ) = vars ti,j−1 {xj 7→tj,j−1 } = vars(ti,j ). Thus, in both cases, vars(ti,j νj ) = vars ti,j νj−1 {xj 7→tj,j−1 } = vars ti,j {xj 7→tj,j−1 }

= vars(ti,j−1 {xj 7→tj,j−1 }{xj 7→tj,j−1 } .

However, a substitution consisting of a single binding is variable-idempotent. Thus vars(ti,j νj ) = vars ti,j−1 {xj 7→tj,j−1 } = vars(ti,j ).

Therefore, for each i = 1, . . . , j, vars(ti,j νj ) = vars(ti,j ). It then follows (using the alternative characterisation of variable-idempotence) that νj is variableidempotent. t u Example 2. Let σ0 = x1 7→f (x2 ), x2 7→g(x3 , x4 ), x3 7→x1 . Then σ1 = x1 7→f (x2 ), x2 7→g(x3 , x4 ), x3 7→f (x2 ) , σ2 = x1 7→f (g(x3 , x4 )), x2 7→g(x3 , x4 ), x3 7→f (g(x3 , x4 )) , σ3 = x1 7→f (g(f (g(x3 , x4 )), x4 )), x2 7→g(f (g(x3 , x4 )), x4 ), x3 7→f (g(x3 , x4 )) . Note that σ3 is variable-idempotent and that T ` σ0 ⇐⇒ σ3 .

4 4.1

Set-Sharing The Sharing Domain

The Sharing domain is due to Jacobs and Langen [8]. However, we use the definition as presented in [1].

106

P.M. Hill, R. Bagnara, and E. Zaffanella

Definition 2. (The set-sharing lattice.) Let def SG = S ∈ ℘f (Vars) S 6= ∅ def

and let SH = ℘(SG). The set-sharing lattice is given by the set def

SS =

(sh, U ) sh ∈ SH , U ∈ ℘f (Vars), ∀S ∈ sh : S ⊆ U ∪ {⊥, >}

ordered by SS defined as follows, for each d, (sh 1 , U1 ), (sh 2 , U2 ) ∈ SS : ⊥ SS d, d SS >, (sh 1 , U1 ) SS (sh 2 , U2 )

⇐⇒

(U1 = U2 ) ∧ (sh 1 ⊆ sh 2 ).

It is straightforward to see that every subset of SS has a least upper bound with respect to SS . Hence SS is a complete lattice.1 An element sh of SH abstracts the property of sharing in a substitution σ. That is, if σ is idempotent, two variables x, y must be in the same set in sh if some variable, say v occurs in both xσ and yσ. In fact, this is also true for variable-idempotent substitutions although it is shown below that this needs to be generalised for substitutions that are not variable-idempotent. Thus, the definition of the abstraction function α for sharing, requires an ancillary definition for the notion of occurrence. Definition 3. (Occurrence.) For each n ∈ N, occi : Subst × Vars → ℘f (Vars) is defined for each σ ∈ Subst and each v ∈ Vars: def

occ0 (σ, v) = {v},

if v = vσ;

def

occ0 (σ, v) = ∅, def occn (σ, v) = y ∈ Vars x ∈ vars(yσ) ∩ occn−1 (σ, v) ,

if v 6= vσ; if n > 0.

It follows that, for fixed values of σ and v, occn (σ, v) is monotonic and extensive with respect to the index n. Hence, as the range of occn (σ, v) is restricted to the finite set of variables in σ, there is an ` = `(σ, v) ∈ N such that occ` (σ, v) = occn (σ, v)) for all n ≥ `. Let def

occ!(σ, v) = occ` (σ, v). Note that if σ is variable-idempotent, then occ!(σ, v) = occ1 (σ, v). Note also that if v 6= vσ, then occ!(σ, v) = ∅. Previous definitions for an occurrence operator such as that for sg in [8] have all been for idempotent substitutions. However, when σ is an idempotent substitution, occ!(σ, v) and sg(σ, v) are the same for all v ∈ Vars. We base the definition of abstraction on the occurrence operator, occ!. 1

Notice that the only reason we have > ∈ SS is in order to turn SS into a lattice rather than a CPO.

The Correctness of Set-Sharing

107

Definition 4. (Abstraction.) The concrete domain Subst is related to SS by means of the abstraction function α : ℘(Subst) × ℘f (Vars) → SS . For each Σ ∈ ℘(Subst) and each U ∈ ℘f (Vars), G def α(Σ, U ) = α(σ, U ), σ∈Σ

where α : Subst × ℘f (Vars) → SS is defined, for each σ ∈ Subst and each U ∈ ℘f (Vars), by def

α(σ, U ) =

occ!(σ, v) ∩ U v ∈ Vars \ {∅}, U .

The following result states that the abstraction for a substitution σ is the same as the abstraction for a variable-idempotent substitution for σ. Lemma 3. Let σ be a substitution, σ 0 a substitution obtained from σ by a sequence of S-transformations, U a set of variables and v ∈ Vars. Then v = vσ ⇐⇒ v = vσ 0 ,

occ!(σ, v) = occ!(σ 0 , v),

and α(σ, U ) = α(σ 0 , U ).

Proof. Suppose first that σ 0 is obtained from σ by a single S-transformation. Thus we can assume that x 7→t and y 7→s are in σ where x ∈ vars(s) and that σ 0 = σ \ {y 7→s} ∪ y 7→s[x/t] . It follows that, since σ is in rational solved form, σ has no circular subset and hence v = vσ ⇐⇒ v = vσ 0 . Thus, if v 6= vσ, then we have v 6= vσ 0 and occ!(σ, v) = occ!(σ 0 , v) = ∅. We now assume that v = vσ = vσ 0 and prove that occm (σ, v) ⊆ occ!(σ 0 , v). The proof is by induction on m. By Definition 3, occ0 (σ, v) = occ0 (σ 0 , v) = {v}, so that the result holds for m = 0. Suppose then that m > 0 and that vm ∈ occm (σ, v). By Definition 3, there exists vm−1 ∈ vars(vm σ) where vm−1 ∈ 0 occm−1 (σ, v). Hence, by the inductive hypothesis, vm−1 ∈ occ!(σ , v). If vm−1 ∈ vars(vm σ 0 ), then, by Definition 3, vm ∈ occ!(σ 0 , v) . On the other hand, if vm−1 ∈ / vars(vm σ 0 ), then vm = y, vm−1 = x, and x ∈ vars(s) (so that vars(t) ⊆ vars(s[x/t])). However, by hypothesis, v = vσ, so that x 6= v and m > 1. Thus, by Definition 3, there exists vm−2 ∈ vars(t) such that vm−2 ∈ occm−2 (σ, v). By the inductive hypothesis, vm−2 ∈ occ!(σ 0 , v). Since y 7→ s[x/t] ∈ σ 0 , and vm−2 ∈ vars(s[x/t]), vm−2 ∈ vars(yσ 0 ). Thus, by Definition 3, y ∈ occ!(σ 0 , v). Conversely, we now prove that, for all m, occm (σ 0 , v) ⊆ occ!(σ, v). The proof is again by induction on m. As in the previous case, occ0 (σ 0 , v) = occ0 (σ, v) = {v}, so that the result holds for m = 0. Suppose then that m > 0 and

108

P.M. Hill, R. Bagnara, and E. Zaffanella

that vm ∈ occm (σ 0 , v). By Definition 3, there exists vm−1 ∈ vars(vm σ 0 ) where vm−1 ∈ occm−1 (σ 0 , v). Hence, by the inductive hypothesis, vm−1 ∈ occ!(σ, v). If vm ∈ occ(σ, vm−1 ), then, by Definition 3, vm ∈ occ!(σ, v). On the other hand, if vm−1 ∈ / vars(vm σ), then vm = y, vm−1 ∈ vars(t) and x ∈ vars(s). Thus, as y 7→s ∈ σ, y ∈ vars(xσ). However, since x 7→t ∈ σ, vm−1 ∈ vars(xσ) so that, by Definition 3, x ∈ occ!(σ, v). Thus, again by Definition 3, y ∈ occ!(σ, v). Thus, if σ 0 is obtained from σ by a single S-transformation, we have the required results: v = vσ ⇐⇒ v = vσ 0 , occ!(σ, v) = occ!(σ 0 , v), and α(σ, U ) = α(σ 0 , U ). Suppose now that there is a sequence σ = σ1 , . . . , σn = σ 0 such that, for i = 2, . . . , n, σi is obtained from σi−1 by a single S-step. If n = 1, then σ = σ 0 . If n > 1, we have by the first part of the proof that, for each i = 2, . . . , n, v = vσi−1 ⇐⇒ v = vσi , occ!(σi−1 , v) = occ!(σi , v), and α(σi−1 , U ) = α(σi , U ), and hence the required results. t u Example 3. Consider again Example 2. Then occ1 (σ0 , x4 ) = {x2 , x4 }, occ2 (σ0 , x4 ) = {x1 , x2 , x4 }, occ3 (σ0 , x4 ) = {x1 , x2 , x3 , x4 } = occ!(σ0 , x4 ), and occ1 (σ3 , x4 ) = {x1 , x2 , x3 , x4 } = occ!(σ3 , x4 ). Thus, if V = {x1 , x2 , x3 , x4 },

α(σ0 , V ) = α(σ3 , V ) = {x1 , x2 , x3 , x4 } .

4.2

Abstract Operations for Sharing Sets

We are concerned in this paper in establishing results for the abstract operation aunify which is defined for arbitrary sets of equations. However, by building the definition of aunify in three steps via the definitions of amgu (for sharing sets) and Amgu (for sharing domains) and stating corresponding results for each of them, we provide an outline for the overall method of proof for the aunify results. Details of all proofs are available in [7]. In order to define the abstract operation amgu we need some ancillary definitions. Definition 5. (Auxiliary functions.) The closure under union function (also called star-union), (·)? : SH → SH , is, for each sh ∈ SH , def sh ? = S ∈ SG ∃n ≥ 1 . ∃T1 , . . . , Tn ∈ sh . S = T1 ∪ · · · ∪ Tn . For each sh ∈ SH and each T ∈ ℘f (Vars), the extraction of the relevant component of sh with respect to T is encoded by the function rel : ℘f (Vars)×SH → SH defined as def

rel(T, sh) = { S ∈ sh | S ∩ T 6= ∅ }.

The Correctness of Set-Sharing

109

For each sh 1 , sh 2 ∈ SH , the binary union function bin : SH × SH → SH is given by def

bin(sh 1 , sh 2 ) = { S1 ∪ S2 | S1 ∈ sh 1 , S2 ∈ sh 2 }. The function proj : SH × ℘f (Vars) → SH projects an element of SH onto a set of variables of interest: if sh ∈ SH and V ∈ ℘f (Vars), then def

proj(sh, V ) = { S ∩ V | S ∈ sh, S ∩ V 6= ∅ }. Definition 6. (amgu.) The function amgu captures the effects of a binding x 7→ t on an SH element. Let x be a variable and t a term. Let also sh ∈ SH and def A = rel {x}, sh ,

def B = rel vars(t), sh . Then def amgu(sh, x 7→t) = sh \ (A ∪ B) ∪ bin(A? , B ? ). Then we have the following soundness result for amgu. Lemma 4. Let (sh, U ) ∈ SS and {x 7→t}, σ, ν ∈ Subst such that ν is a relevant unifier of {xσ = tσ} and vars(x), vars(t), vars(σ) ⊆ U . Then α(σ, U ) SS (sh, U ) =⇒ α(ν ◦ σ, U ) SS (amgu(sh, x 7→t), U ). To prove this, observe that, by Lemma 2, if σ is not variable-idempotent, it can be transformed to a variable-idempotent substitution σ 0 . Hence, by Lemma 3, α(σ, U ) = α(σ 0 , U ). Therefore, the proof, which is given in [7], deals primarily with the case when σ is variable-idempotent. Since a relevant unifier of e is a relevant unifier of any other set e0 equivalent to e wrt to the equality theory T , this lemma shows that it is safe for the analyser to perform part or all of the concrete unification algorithm before computing amgu. The following lemmas, proved in [7], show that amgu is commutative and idempotent. Lemma 5. Let sh ∈ SH and {x 7→r} ∈ Subst. Then amgu(sh, x 7→r) = amgu amgu(sh, x 7→r), x 7→r . Lemma 6. Let sh ∈ SH and {x 7→r}, {y 7→t} ∈ Subst. Then amgu amgu(sh, x 7→r), y 7→t = amgu amgu(sh, y 7→t), x 7→r .

110

P.M. Hill, R. Bagnara, and E. Zaffanella

4.3

Abstract Operations for Sharing Domains

The definitions and results of Subsection 4.2 can be lifted to apply to sharing domains. Definition 7. (Amgu.) The operation Amgu : SS × Subst → SS extends the SS description it takes as an argument, to the set of variables occurring in the binding it is given as the second argument. Then it applies amgu: Amgu (sh, U ), x 7→t def = amgu sh ∪ {u} u ∈ vars(x 7→t) \ U , x 7→t , U ∪ vars(x 7→t) . The results for amgu can easily be extended to apply to Amgu. Definition 8. (aunify.) The function aunify : SS ×Eqs → SS generalises Amgu to a set of equations e: If (sh, U ) ∈ SS , x is a variable, r is a term, s = f (s1 , . . . , sn ) and t = f (t1 , . . . , tn ) are non-variable terms, and s = t denote the set of equations {s1 = t1 , . . . , sn = tn }, then def

aunify((sh, U ), ∅) = (sh, U ), if e ∈ ℘f (Eqs) is unifiable, def aunify (sh, U ), e ∪ {x = r} = aunify Amgu(sh, U ), x 7→r), e \ {x = r} , def aunify (sh, U ), e ∪ {s = x} = aunify (sh, U ), (e \ {s = x}) ∪ {x = s} , def aunify (sh, U ), e ∪ {s = t} = aunify (sh, U ), (e \ {s = t}) ∪ s = t , and, if e is not unifiable, def

aunify((sh, U ), e) = ⊥. For the distinguished elements ⊥ and > of SS def aunify ⊥, e = ⊥,

def aunify >, e = >.

As a consequence of this and the generalisation of Lemmas 4, 5 and 6 to Amgu, we have the following soundness, commutativity and idempotence results required for aunify to be sound and well-defined. As before, the proofs of these results are in [7]. Theorem 1. Let (sh, U ) ∈ SS , σ, ν ∈ Subst, and e ∈ ℘f (Eqs) be such that vars(σ) ⊆ U and ν is a relevant unifier of e. Then α(σ, U ) SS (sh, U ) =⇒ α(ν ◦ σ, U ) SS aunify((sh, U ), e).

The Correctness of Set-Sharing

111

Theorem 2. Let (sh, U ) ∈ SS and e ∈ ℘f (Eqs). Then aunify (sh, U ), e = aunify aunify (sh, U ), e , e . Theorem 3. Let (sh, U ) ∈ SS and e1 , e2 ∈ ℘f (Eqs). Then aunify aunify (sh, U ), e1 , e2 = aunify aunify (sh, U ), e2 , e1 .

5

Discussion

The SS domain which was first defined by Langen [14] and published by Jacobs and Langen [8] is an important domain for sharing analysis. In this paper, we have provided a framework for analysing non-idempotent substitutions and presented results for soundness, idempotence and commutativity of aunify. In fact, most researchers concerned with analysing sharing and related properties using the SS domain, assume these properties hold. Why therefore are the results in this paper necessary? Let us consider each of the above properties one at a time. 5.1

Soundness

We have shown that, for any substitution σ over a set of variables U , the abstraction α(σ, U ) = (sh, U ) is unique (Lemma 3) and the aunify operation is sound (Theorem 1). Note that, in Theorem 1, there are no restrictions on σ; it can be non-idempotent, possibly including cyclic bindings (that is, bindings where the domain variable occurs in its co-domain). Thus this result is widely applicable. Previous results on sharing have assumed that substitutions are idempotent. This is true if equality is syntactic identity and the implementation uses a unification algorithm based on that of Robinson [17] which includes the occur-check. With such algorithms, the resulting unifier is both unique and idempotent. Unfortunately, this is not what is implemented by most Prolog systems. In particular, if the algorithm is as described in [11] and used in Prolog III [5], then the resulting unifier is in rational solved form. This algorithm does not generate idempotent or even variable-idempotent substitutions even when the occur-check would never have succeeded. However, it has been shown that the substitution obtained in this way uniquely defines a system of rational trees [5]. Thus our results show that its abstraction using α, as defined in this paper, is also unique and that aunify is sound. Alternatively, if, as in most commercial Prolog systems, the unification algorithm is based on the Martelli-Montanari algorithm, but omits the occur check step, then the resulting substitution may not be idempotent. Consider the following example. Suppose we are given as input the equation p(z, f (x, y)) = p(f (z, y), z) with an initial substitution that is empty. We apply the steps in Martelli-Montanari procedure but without the occur-check:

112

P.M. Hill, R. Bagnara, and E. Zaffanella

1 2 3 4 5 6

equations p(z, f (x, y)) = p(f (z, y), z) z = f (z, y), f (x, y) = z f (x, y) = f (z, y) x = z, y = y y=y ∅

substitution ∅ ∅ {z 7→f (z, y)} {z 7→f (z, y)} {z 7→f (z, y), x 7→z} {z 7→f (z, y), x 7→z}

Note that we have used three kinds of steps here. In lines 1 and 3, neither argument of the selected equation is a variable. In this case, the outer nonvariable symbols (when, as in this example, they are the same) are removed and new equations are formed between the corresponding arguments. In lines 2 and 4, the selected equation has the form v = t, where v is a variable and t is not identical to v, then every occurrence of v is replaced by t in all the remaining equations and the range of the substitution. v 7→t is then added to the substitution. In line 5, the identity is removed. Let σ = {z 7→f (z, y), x 7→z}, be the computed substitution. Then, we have vars(xσ) = vars(z) = {z}, vars(xσ 2 ) = vars(f (z, y)) = {y, z}. Hence σ is not variable-idempotent. We conjecture that the resulting substitution is still unique (up to variable renaming). In this case our results can be applied so that its abstraction using α, as defined in this paper, is also unique and aunify is sound. 5.2

Idempotence

Definition 8 defines aunify inductively over a set of equations, so that it is important for this definition that aunify is both idempotent and commutative. The only previous result concerning the idempotence of aunify is given in thesis of Langen [14, Theorem 32]. However, the definition of aunify in [14] includes the renaming and projection operations and, in this case, only a weak form of idempotence holds. In fact, for the basic aunify operation as defined here and without projection and renaming, idempotence has never before been proven. 5.3

Commutativity

In the thesis of Langen the “proof” of commutativity of amguhas a number of omissions and errors [14, Lemma 30]. We highlight here, one error which we were unable to correct in the context of the given proof. To make it easier to compare, we adapt our notation and, define amge only in the case that a is a variable: def

amge(a, b, sh) = amgu(sh, a 7→b).

The Correctness of Set-Sharing

113

To prove the lemma, it has to show that: amge(a2 , b2 amge(a1 , b1 , sh)) = amge(a1 , b1 , amge(a2 , b2 , sh)). holds when a1 and a2 are variables. This corresponds to “the second base case” of the proof. We use Langen’s terminology: – A set of variables X is at a term t iff var(t) ∩ X 6= ∅. – A set of variables X is at i iff X is at ai or bi . – A union X ∪i Y is of Type i iff X is at ai and Y is at bi . def

def

Let lhs = amge(a2 , b2 , amge(a1 , b1 , S)), and rhs = amge(a1 , b1 , amge(a2 , b2 , S)). def

Let also Z ∈ lhs and T = aunify(a1 , b1 , S). Consider the case when Z = X ∪2 Y where X ∈ rel(a2 , T ), Y ∈ rel(b2 , T ), X = U ∪1 V where U ∈ rel(a1 , sh), V ∈ rel(b1 , sh) and U ∩ (vars(a2 ) ∪ vars(b2 )) = ∅ (that is, U is not at 2). Then the following quote [14, page 53, line 23] applies: In this case (U ∪1 V ) ∪2 Y = U ∪1 (V ∪2 Y ). By the inductive assumption V ∪2 Y is in the rhs and therefore so is Z. We give a counter-example to the statement “V ∪2 Y is in the rhs”. Suppose a1 , b1 , a2 , b2 are variables. We let each of a1 , b1 , a2 , b2 denote both the actual variable and the singleton set containing that variable. Suppose sh = {a1 , b1 a2 , b2 }. Then, from the definition of amge, lhs = {a1 b1 a2 b2 },

rhs = {a1 b1 a2 b2 },

T = {a1 b1 a2 , b2 }.

Let Z = a1 b1 a2 b2 , X = a1 b1 a2 , Y = b2 , U = a1 , V = b1 a2 . All the above conditions. However V ∪2 Y = b1 a2 b2 and this is not in {a1 b1 a2 b2 }.

References 1. R. Bagnara, P. M. Hill, and E. Zaffanella. Set-sharing is redundant for pair-sharing. In P. Van Hentenryck, editor, Static Analysis: Proceedings of the 4th International Symposium, volume 1302 of Lecture Notes in Computer Science, pages 53–67, Paris, France, 1997. Springer-Verlag, Berlin. 2. M. Bruynooghe and M. Codish. Freeness, sharing, linearity and correctness — All at once. In P. Cousot, M. Falaschi, G. Fil´e, and A. Rauzy, editors, Static Analysis, Proceedings of the Third International Workshop, volume 724 of Lecture Notes in Computer Science, pages 153–164, Padova, Italy, 1993. Springer-Verlag, Berlin. An extended version is available as Technical Report CW 179, Department of Computer Science, K.U. Leuven, September 1993. 3. K. L. Clark. Negation as failure. In H. Gallaire and J. Minker, editors, Logic and Databases, pages 293–322, Toulouse, France, 1978. Plenum Press.

114

P.M. Hill, R. Bagnara, and E. Zaffanella

4. M. Codish, D. Dams, G. Fil´e, and M. Bruynooghe. Freeness analysis for logic programs-and correctness? In D. S. Warren, editor, Logic Programming: Proceedings of the Tenth International Conference on Logic Programming, MIT Press Series in Logic Programming, pages 116–131, Budapest, Hungary, 1993. The MIT Press. An extended version is available as Technical Report CW 161, Department of Computer Science, K.U. Leuven, December 1992. 5. A. Colmerauer. Prolog and Infinite Trees. In K. L. Clark and S. ˚ A. T¨ arnlund, editors, Logic Programming, APIC Studies in Data Processing, volume 16, pages 231–251. Academic Press, New York, 1982. 6. A. Colmerauer. Equations and inequations on finite and infinite trees. In Proceedings of the International Conference on Fifth Generation Computer Systems (FGCS’84), pages 85–99, Tokyo, Japan, 1984. ICOT. 7. P. M. Hill, R. Bagnara, and E. Zaffanella. The correctness of set-sharing. Technical Report 98.03, School of Computer Studies, University of Leeds, 1998. 8. D. Jacobs and A. Langen. Accurate and efficient approximation of variable aliasing in logic programs. In E. L. Lusk and R. A. Overbeek, editors, Logic Programming: Proceedings of the North American Conference, MIT Press Series in Logic Programming, pages 154–165, Cleveland, Ohio, USA, 1989. The MIT Press. 9. D. Jacobs and A. Langen. Static analysis of logic programs for independent AND parallelism. Journal of Logic Programming, 13(2&3):291–314, 1992. 10. J. Jaffar, J-L. Lassez, and M. J. Maher. Prolog-II as an instance of the logic programming scheme. In M. Wirsing, editor, Formal Descriptions of Programming Concepts III, pages 275–299. North Holland, 1987. 11. T. Keisu. Tree Constraints. PhD thesis, The Royal Institute of Technology, Stockholm, Sweden, May 1994. Also available in the SICS Dissertation Series: SICS/D– 16–SE. 12. A. King. A synergistic analysis for sharing and groundness which traces linearity. In D. Sannella, editor, Proceedings of the Fifth European Symposium on Programming, volume 788 of Lecture Notes in Computer Science, pages 363–378, Edinburgh, UK, 1994. Springer-Verlag, Berlin. 13. A. King and P. Soper. Depth-k sharing and freeness. In P. Van Hentenryck, editor, Logic Programming: Proceedings of the Eleventh International Conference on Logic Programming, MIT Press Series in Logic Programming, pages 553–568, Santa Margherita Ligure, Italy, 1994. The MIT Press. 14. A. Langen. Static Analysis for Independent And-Parallelism in Logic Programs. PhD thesis, Computer Science Department, University of Southern California, 1990. Printed as Report TR 91-05. 15. M. J. Maher. Complete axiomatizations of the algebras of finite, rational and infinite trees. In Proceedings, Third Annual Symposium on Logic in Computer Science, pages 348–357, Edinburgh, Scotland, 1988. IEEE Computer Society. 16. K. Muthukumar and M. Hermenegildo. Compile-time derivation of variable dependency using abstract interpretation. Journal of Logic Programming, 13(2&3):315– 347, 1992. 17. J. A. Robinson. A machine-oriented logic based on the resolution principle. Journal of the ACM, 12(1):23–41, 1965.

Deriving Analysers by Folding/Unfolding of Natural Semantics and a Case Study: Slicing Val´erie Gouranton IRISA/INRIA, IFSIC Campus universitaire de Beaulieu, 35042 Rennes Cedex, France tel : 33 2 99 84 74 85, fax : 33 2 99 84 71 71, [email protected] http://www.irisa.fr/lande/

Abstract. We consider speciﬁcations of analysers expressed as compositions of two functions: a semantic function, which returns a natural semantics derivation tree, and a property deﬁned by recurrence on derivation trees. A recursive deﬁnition of a dynamic analyser can be obtained by fold/unfold program transformation combined with deforestation. A static analyser can then be derived by abstract interpretation of the dynamic analyser. We apply our framework to the derivation of a dynamic backward slicing analysis for a logic programming language.

1

Introduction

A large amount of work has been devoted to program analysis during the last two decades, both on the practical side and on the theoretical issues. However, most of the program analysers that have been implemented or reported in the literature so far are concerned with one speciﬁc property, one speciﬁc language and one speciﬁc service (dynamic or static). A few generic tools have been proposed but they are generally restricted to one class of properties or languages, or limited in their level of abstraction. We believe that there is a strong need for environments supporting the design of program analysers and that more eﬀort should be put on the software engineering of analysers. We present a framework for designing analysers from operational speciﬁcations by program transformation (folding/unfolding). The analysis speciﬁcation has two components: a semantics of the programming language and a deﬁnition of the property. The advantage of this two-fold speciﬁcation is that the deﬁnition of the property can be kept separate from the semantics of the programming language. Ideally, properties can be speciﬁed in terms of the derivation tree of the operational semantics. Speciﬁc analysers can then be obtained systematically by instantiating semantics of the programming language. We focus here on slicing of logic programs. The general approach is detailed in [12]. Natural semantics [7,14] are a good starting point for the deﬁnition of analyses because they are both structural (compositional) and intensional. They are structural because the semantics of a phrase in the programming language G. Levi (Ed.): SAS’98, LNCS 1503, pp. 115–133, 1998. c Springer-Verlag Berlin Heidelberg 1998

116

V. Gouranton

is derived from the semantics of subphrases; they are intensional because the derivation tree that is associated with a phrase in the programming language contains the intermediate results (the semantics of subphrases). These qualities are signiﬁcant in the context of program analysis because compositionality leads to more tractable proof techniques and intensionality makes it easier to establish the connection between the result of the analysis and its intended use. Our semantics is deﬁned formally as a function taking a term and an evaluation context and returning a derivation tree. The property itself is a function from derivation trees to a suitable abstract domain. The composition of these two functions deﬁnes a dynamic a posteriori analysis. It represents a function which initially calculates the trace of a complete execution (a derivation tree) of a program before extracting the required property. Program transformations via extended folding/unfolding techniques and simpliﬁcation rules allow to obtain a recursive deﬁnition of the dynamic analyser (which does not call the property function). This function is in fact a dynamic on the fly analyser in the sense that it calculates the required property progressively during program execution. The following diagram shows the general organisation: Context × Term Q

Semantics /

QQQ QAnalyser QQQ QQQ QQ(

Tree

Property

Result

The key points of the approach proposed here are the following: – The derivation is achieved in a systematic way by using functional transformations: unfolding and folding. – It is applicable to a wide variety of languages and properties because it is based on natural semantics deﬁnitions. As mentioned before, some of the analyses that we want to specify are dynamic and others are static. There is no real reason why these two categories of analyses should be seen as belonging to diﬀerent worlds. In the paper we focus on dynamic analysis, considering that static analysis can be obtained in a second stage as an abstract interpretation of the dynamic analysis as presented in [10]. We outline this derivation in the conclusion. Note that our approach introduces a clear separation between the speciﬁcation of an analysis (deﬁned as a property on semantics derivation trees) and the algorithm that implements it. We illustrate the framework by the formal derivation of a slicing analysis for a logic programming language. The diﬀerent stages of the derivation are detailed in the following sections. Section 2 introduces contexts, terms, derivation trees and the semantics function. The abstract domain and the property function are presented in section 3. The transformation of the composition of the two speciﬁcation functions (the semantics and the property) into a dynamic on the fly analyser is described in section 4. Related work, conclusion and avenues for further research are discussed in section 5 and section 6 respectively.

Deriving Analysers by Folding/Unfolding of Natural Semantics

2

117

Natural Semantics

The natural semantics of a language is a set of axioms and inference rules that deﬁne a relation between a context, a term in the programming language and a result. A natural semantics derivation tree has the form: Proof-Tree = [RN]

Proof-Tree1

... Proof-Treen STT

where RN is the name of the rule used to derive STT. The conclusion STT is a statement, that is to say a triple consisting of a context, a term and a result. Let C be the set of contexts, T the type of terms of the language and PT the type of derivation trees, we have: PT = STT × (list PT) × RN STT = C × T × NF T = PP × I

Derivation trees are made of a statement (the conclusion), a list of derivation trees (the premises) and the name of rule applied to derive the conclusion. We assume that a term is a pair of a program point and an expression. STT denotes the type of statements, RN rule names, NF normal forms (program results), PP program points and I expressions. The Semantics of a Logic Programming Language We assume a program P rog which is a collection of predicate deﬁnitions of the form [Pk (x1 , ..., xn ) = Bk ]. The body Bk of a predicate is in normal form and it contains only variables from {x1 , ..., xn }. Normal forms are ﬁrst order formulae (also called “goal formulae” in [16]) built up from predicate applications using only the connectives “and”, “or”, and “there exists”. Their syntax is deﬁned by: I ::= Op(x1 , x2 , x3 ) | x = t | U1 ∧ U2 | U1 ∨ U2 | ∃x.U1 | Pk (y1 , . . . , yn ) where Op stands for basic predicates1 and Pk for user-deﬁned predicates. Ui are terms of type T. We assume that each variable x occurring in a term ∃x.U1 is unique. In a program, each subterm in this syntax is associated with a program point (using pairs). As an illustration of this syntax, Figure 1 presents a small program in a logic programming syntax and shows its translation into normal form. Program points are represented by πi . Note that some program points are omitted for the sake of readability. The program deﬁnes two predicates P and Q. The main predicate is Q. The recursive predicate P computes the length n of the list l of integers, the sum sum of the elements of the list, the maximun max and the minimum min of the list l. The average av of the list is computed by the predicate Q via P (the value of sum obtained by P is divided by the length n of the list). 1

We consider only ternary basic predicates here, but other arities are treated in the same way.

118

V. Gouranton Deﬁnition of the program in a logic programming syntax P (nil, 1, 0, 0, 0) P ((x, nil), 1, x, x, x) P ((x, xs), n, sum, max, min) = (π1 , P (xs, n , sum , max , min )) (π2 , Ad(n , 1, n)) (π3 , Ad(sum , x, sum)) (π4 , M ax(max , x, max)) (π5 , M in(min , x, min)) Q(l, av, max, min) = (π6 , P (l, n, sum, max, min)) (π7 , Div(sum, n, av)) Normal form of the program

P (l, n, sum, max, min) = ((l = nil) ∧ (n = 1) ∧ (sum = 0) ∧ (max = 0) ∧ (min = 0))∨ (∃x. (l = (x, nil)) ∧ (n = 1) ∧ (sum = x) ∧ (max = x) ∧ (min = x))∨ (∃x. ∃xs. ∃n . ∃sum . ∃max . ∃min . l = (x, xs)∧ (π1 , P (xs, n , sum , max , min ))∧ (π2 , Ad(n , 1, n))∧ (π3 , Ad(sum , x, sum))∧ (π4 , M ax(max, x, max))∧ (π5 , M in(min , x, min))) Q(l, av, max, min) = (∃n. ∃sum. (π6 , P (l, n, sum, max, min))∧ (π7 , Div(sum, n, av))) Fig. 1. A simple logic program

Following [15], we assume an inﬁnite set of program variables Pvar and an inﬁnite set of renaming variables Rvar. Terms and substitutions are constructed using program variables and renaming variables. We distinguish two kinds of substitutions: program variable substitutions (Subst) whose domain and co-domain are subsets of Pvar and Rvar respectively, and renaming variable substitutions (Rsubst) whose domain and co-domain are subsets of Rvar: Subst = Pvar → Rterm Rsubst = Rvar → Rterm

where Rterm represents a term constructed with renaming variables Rvar. By convention, we use θ ∈ Subst for a program variable substitution and σ ∈ Rsubst for a renaming variable substitution. The deﬁnition of substitution composition is modiﬁed to take account the role held by renaming variables. The modiﬁcation occurs when θ ∈ Subst and σ ∈ Rsubst, we have σ ◦ θ ∈ Subst deﬁned by: dom(σ ◦ θ) = dom(θ) (σ ◦ θ)(x) = σ(θ(x)) for all x ∈ dom(θ)

Deriving Analysers by Folding/Unfolding of Natural Semantics

119

The domain of contexts for this language is deﬁned by C = Tree(Subst) where Tree(H) is the type of binary trees with leaves of type H. We deﬁne contexts as binary trees of substitutions to take into account the non deterministic nature of the language. So, we gather in one derivation the computation of all the substitutions of a program. A particular control strategy for the implementation of the language corresponds to a particular ordering of the leaves of substitutions trees. For instance, the list of results of the usual depth-ﬁrst evaluation strategy of Prolog is precisely the leaves of the substitution tree produced by our semantics ordered from left to right. We write N (T1 , T2 ) for a tree with subtrees T1 and T2 .

C (π, Op(x1 , x2 , x3 )) → op(C, x1 , x2 , x3 )

[Op] [Eq]

[∧]

C U1 → R1 R1 U2 → R2 C (π, U1 ∧ U2 ) → R2 [∃]

[Call]

C (π, x = t) → unif (C, x, t)

[∨]

C U1 → R1 C U2 → R2 C (π, U1 ∨ U2 ) → union (C, R1 , R2 )

Add (C, x, rx) U1 → R1 C (π, ∃x.U1 ) → Drop (R1 , x)

Renk (C) Bk → R1 C (π, Pk (y1 , ..., yn )) → Extk (C, R1 )

rx ∈ Rvar

fresh variable

with [Pk (x1 , ..., xn ) = Bk ] ∈ P rog

F (N (T1 , T2 ), x1 , . . . , xn ) = N (F (T1 , x1 , . . . , xn ), F (T2 , x1 , . . . , xn )) F (θ, x1 , . . . , xn ) = F (θ, x1 , . . . , xn ) op(θ, x1 , x2 , x3 ) = let σ = [(θ(x1 ) op θ(x2 ))/θ(x3 )] in σ ◦ θ if θ(x1 ) and θ(x2 ) are ground and θ(x3 ) ∈ Rvar,

⊥ otherwise

unif(θ, x, t) = let σ = mgu(θ(x), θ(t)) in σ ◦ θ if θ(x) and θ(t) can be uniﬁed , ⊥ otherwise union(N (T1 , T2 ), N (U1 , U2 ), N (V1 , V2 )) = N (union(T1 , U1 , V1 ), union(T2 , U2 , V2 )) union(θ, U, V ) = N (U, V ) Add(θ, pv, rv) = θ[rv/pv] with v = pv ⇒ θ[rv/pv](v) = θ(v) and θ[rv/pv](pv) = rv Drop(θ, pv) = θ/pv with v = pv ⇒ θ/pv (v) = θ(v) and θ/pv (pv) =⊥ Renk (θ) = [θ(yi )/xi ] Extk (θ, θ ) = σ ◦ θ with θ = σ ◦ [θ(yi )/xi ] Fig. 2. Natural semantics of a logic programming language

120

V. Gouranton

The natural semantics of a simple logic programming language using the usual inference rule presentation is presented in Figure 2. The normal forms calculated by the rules are contexts. In the ﬁgure, F (T ) denotes the application of a function F to all the substitutions of a tree T and its result is also a tree. The function op represents the interpretation of operator Op. The substitution unif (θ, x, t) of Subst is deﬁned for the uniﬁcation of x and t via θ (rule Eq). The rule ∧ is not surprising, the ﬁrst formula U1 of the conjonction is evaluated and the result R1 is taken as the new context for the evaluation of the second formula U2 of the conjonction; the result R2 is the ﬁnal result. For the rule ∨, the subtrees corresponding to the sub-formulae of the disjonction are evaluated independently. The function union(T1 , T2 , T3 ) is needed to build a new substitution tree joining the trees T2 and T3 produced by two subgoals. Its ﬁrst argument is the initial substitution, which is used to identify the points where the joins have to be introduced (these points are the leaves of T1 ). The argument θ as the initial substitution can be ignored because substitutions are added to contexts, generating new contexts. The rule ∃ uses two functions Add and Drop. Add is used to add a program variable in a substitution (the new program variable is attached to a free renaming variable) and Drop removes a variable from a substitution. For the rule Call, two deﬁnitions of substitutions are needed. Renk (C) creates a new substitution to execute the body of a clause (it amounts to a variable renaming) because the body Bk of a clause contains formal parameters xi and C contains program variables yi . Extk (C, R1 ) propagates the result of a predicate in the calling substitutions because C contains variables yi and R1 contains formal parameters xi . From the deﬁnition of Renk , we see that the body Bk of a predicate is evaluated in an environment deﬁning exactly the formal parameters of the predicate Pk . The formal deﬁnitions of the functions introduced informally before are presented in the bottom of Figure 2. In order to make formal manipulations easier, we express the construction of natural semantics derivation trees in a functional framework. The semantic function S is a partial function of type: C × T → PT

The important issue about the type of the semantic function is that it returns the whole natural semantics derivation tree, rather than just the result of the program. This choice makes it easier to deﬁne intensional analyses. The fact that we describe the semantics in a functional framework does not prevent us from dealing with non deterministic languages, as we show for a logic programming language. This is because we can use NF and C to represent sets of possible results and contexts. We use the notation X.ty to denote the ﬁeld of type TY of X. For example, we will make intensive use of the following expressions in the rest of the paper:

Deriving Analysers by Folding/Unfolding of Natural Semantics

type P T .stt P T .lpt P T .rn P T .stt.c P T .stt.t.i P T .stt.t.pp P T .stt.nf

121

meaning

STT conclusion of P T (list PT) premisses of P T RN name of the rule used at the root of P T C context of the conclusion sequent of P T I term of the conclusion sequent of P T PP program point of the conclusion sequent of P T NF normal form of the conclusion sequent of P T

The semantics in functional form is presented in Figure 3. The semantics function of Figure 3 takes two arguments (the context C and the term T ) and it returns a derivation tree. The derivation tree contains the conclusion (C, T, F k (C, R, E)) of type STT, where F k is the result of the program in functional form, the list of subtrees and the name k of the rule used to derive the conclusion. The body of the function is a list of cases selected by pattern matching on the form of the term. The function is deﬁned by recurrence on the term. The set of deﬁnitions Prog is used as an implicit parameter of the semantics.

S (C, T ) = case T of (π, Op (x1 , x2 , x3 )) : ((C, T, op(C, x1 , x2 , x3 )), nil, Op) (π, Eq(x, t)) : ((C, T, unif (C, x, t)), nil, Eq) (π, And (U1 , U2 )) : let P T1 = S (C, U1 ) R1 = P T1 .stt.nf P T2 = S (R1 , U2 ) R2 = P T2 .stt.nf in ((C, T, R2 ), [P T1 , P T2 ], ∧) (π, Or (U1 , U2 )) : let P T1 = S (C, U1 ) R1 = P T1 .stt.nf P T2 = S (C, U2 ) R2 = P T2 .stt.nf in ((C, T, union (C, R1 , R2 )), [P T1 , P T2 ], ∨) (π, Exists (x, U1 )) : let P T1 = S (Add (C, x, rx), U1 ) R1 = P T1 .stt.nf in ((C, T, Drop (R1 , x)), [P T1 ], ∃) (π, Call (Pk (y1 , . . . , yn ))) : let P T1 = S (Renk (C), Bk ) R1 = P T1 .stt.nf in ((C, T, Extk (C, R1 )), [P T1 ], Call) with [Pk (x1 , ..., xn ) = Bk ] ∈ Prog Fig. 3. The semantics function of a logic programming language

122

3

V. Gouranton

Specification of a Slicing Property

Slicing2 a program consists in constructing a reduced version of the program (called a program slice) containing only those statements that aﬀect a given set of variables at given program points (this set is called the slicing criterion). In program debugging, slicing makes it possible for a software engineer to focus on the relevant parts of the code. Slicing is also useful for testing, program understanding and in maintenance activities. Because of this diversity of applications, diﬀerent variations on the notion of slicing have been proposed, as well as a number of methods to compute slices. First, a program slice can either be executable or not. Producing an executable slice makes it possible to apply further treatments to the result of the analysis. Another important distinction is between static and dynamic slicing. In the ﬁrst case, the slice is computed without any assumption on the inputs, whereas the latter relies on some speciﬁc input data. Slicing algorithms can also be distinguished by their direction. Backward slicing identiﬁes the statements of a program that may have some impact on the criterion whereas forward slicing returns the statements which may be inﬂuenced by the criterion. In this paper, we consider dynamic backward slicing with executable slices. Static slicing algorithms can be derived by abstract interpretation of dynamic slicing analysers ; this construction is sketched in the conclusion. We can describe forward slicing analysers in a similar way but slicing analyses producing non executable slices do not ﬁt well into our framework since the speciﬁcation of the analysis is a relation between the semantics of the original program and the semantics of the slice as presented in [10] . Slicing was originally proposed by Weiser for imperative languages [29] and its application to logic programming [23] and functional programming [18] have been studied recently. In fact, the concept of slicing itself is very general: it is not tied to one speciﬁc style of programming3 and it can lead to dynamic as well as static analysers [25]. A slicing analysis for a logic programming language (with programs in normal form) according to a program point and a set of variables of interest consists in keeping only the sub-goals of disjunctions of each clause (a clause deﬁnes a predicate) being able to aﬀect the value of the variables of interest. If all subgoals of a formula of the disjunction are dropped, then this formula is dropped. If all formulae of the disjunction of goals are dropped, then the clause is dropped. In the opposite case, the head of the clause deﬁning the predicate is kept. Let us take the program in normal form of Figure 1 to illustrate dynamic backward slicing. We assume that we are interested only in the value of the variable av at the program point π7 . The pair {(π7 , av)} is called the slicing criterion. The dynamic slice of the program is extracted for one particular input. For instance, if we execute the predicate Q with nil as the initial value of l, we get: 2 3

More precisely “backward slicing”. Even if the details of the resulting analyses are of course.

Deriving Analysers by Folding/Unfolding of Natural Semantics

123

P (nil, 1, 0, 0, 0) Q(l, av, max, min) = (π6 , P (l, n, sum, max, min)) (π7 , Div(sum, n, av))

The predicate P is not recursively called and the ﬁrst disjunctive part is satisﬁed, the third clause of P is never executed. The deﬁnition of the predicate Q is kept because all its clauses are useful to compute the variable av. If we consider the execution of the program with (2, (3, nil)) as initial value of l, we recursively call the predicate P , we get: P ((x, nil), 1, x, x, x) P ((x, xs), n, sum, max, min) = (π1 , P (xs, n , sum , max , min )) (π2 , Ad(n , 1, n)) (π3 , Ad(sum , x, sum)) Q(l, av, max, min) = (π6 , P (l, n, sum, max, min)) (π7 , Div(sum, n, av))

Only a part of the third clause of the predicate P is kept, the program points π4 and π5 are dropped because they are not useful in computing av (they are needed to compute the values for max and min ). Assuming a set of pairs (πi , vi ), where πi is a program point and vi a variable, a backward slicing analysis produces the slice computing for each point πi the same values as the initial program for the variable vi . In our framework, a property is expressed by a function which takes at least an argument being the co-domain of the semantics function (a derivation tree of type PT) and the result of the property is an abstract domain. The slicing property takes an additional argument to represent the slicing criterion (of type PP → P(Pvar)) and the type of the result is P(PP) because slices are represented by sets of program points. The slicing criterion is represented in our approach by the mapping from program points to relevant variables. Because of the slicing property, we need extra information. We introduce a set of variables of interest according to a program point (this set represents the value of variable that must be preserved for computing the corresponding term). The initial value of the set is ∅. The property propagates this information of type P(Pvar) and ﬁnally the type of the property is: αsl : PT × (PP → P(Pvar)) × P(Pvar) → P(PP) × P(Pvar) The slicing property αsl for the logic programming language is presented in Figure 4. The property takes as arguments a derivation tree P T , plus two additional parameters RV ∈ PP → P(Pvar) and D ∈ P(Pvar). The second argument RV (for Relevant Variables) is the slicing criterion mentioned above. A program point π associated with a non-empty set RV (π) is called an observation point. The third argument D of the property represents the set of variables whose

124

V. Gouranton

values must be preserved in the output context4 (normal form) of the term, i.e. the set of variables that must be preserved in the result Ri of the evaluation of the derivation tree P Ti . In the initial call, D is the empty set. The function αsl is called recursively on the intermediate derivation trees (P Ti ) of the natural semantics and sets of observation variables. The result of the property is a pair (S, N ) with S ∈ P(PP) and N ∈ P(Pvar). S is the set of program points of the term T that must be kept in the slice and N is the set of variables whose value must be preserved in the input context5 . A program point must be kept in the slice if it can inﬂuence an observation point or the value of a variable of D in the output context. The same condition applies to decide which variables must be preserved in the input context. If the program point can be removed from the slice, the result of the property is (∅, D), which means that no program point is added to the slice and the variables whose values must be preserved in the input context are the variables that are necessary in the output context. Otherwise, the ﬁrst component of the result of the property is ∪ Si ∪ {π} because π has to be added to the program points collected in the i

subterms of T . The second component N of the result is the set of variables whose value must be preserved in the input context C. It contains at least the set D and the variables RV (π) of slicing criterion, thus we factorise that by setting D = D ∪ RV (π) in beginning of the slicing deﬁnition. We assume that the deﬁnitions of Si and Ni are not mutually recursive. The deﬁnition of the sets of observation variables (third argument of αsl ) do not use Nj , j > i. Note that this is a characteristic feature of a backward analysis. In Figure 4, the relation Indep(C, D1 , D2 ) is used to ensure that two sets of variables D1 and D2 are independent, which is the case when they do not share any renaming variables (in any substitution of the context C). The relation Indep appears in the ﬁrst two cases as a necessary condition to exclude the term from the slice. If the relation holds, then the (renaming variable) substitution resulting from the evaluation of the term cannot have any impact on the variables of D. The relation UF(C, x, t) is satisﬁed if the uniﬁcation of x and t cannot fail for any substitution of C. It is a prerequisite for excluding Eq(x, t) from the slice because a failure is recorded in the substitution tree as the ⊥ substitution6 ; as a consequence, it has an impact on all the variables. This condition was not included in the Op case, assuming that the logic programming language is equipped with mode annotations ensuring that operators are always called with their ﬁrst two arguments ground and the last one free7 . In both the Op and the Eq cases, the set of necessary variables (at the input of the program point) is D added to all the program variables of the term: the set {x1 , x2 , x3 } for 4 5 6 7

For a forward property this argument would characterise the input context rather than the output context. For a forward property this argument would characterise the output context rather than the input context. Note that ⊥ is an absorbing element for the semantics of the language. For instance op(⊥, x1 , x2 , x3 ) =⊥ and unif (⊥, x, t) =⊥. Otherwise an extra condition based on UF can be added as in the Eq case.

Deriving Analysers by Folding/Unfolding of Natural Semantics

125

Op (x1 , x2 , x3 ) and the set of program variables occurring in t increased with x for the rule Eq (x, t). The formal deﬁnitions of Indep and U F are presented in

the bottom of Figure 4. For the rule And, both branches are processed in turn (the second branch ﬁrst since our property is computed in a backward direction). The property is ﬁrst called with P T2 and D and the result is (S2 , N2 ); then the property is computed with P T1 and N2 , we have (S1 , N1 ) as the result. The program point π can be removed from the slice when both S1 and S2 are empty sets. When the program point is kept, the result of the operator And is then (S1 ∪ S2 ∪ {π}, N1 ) because the information about program points of both branches is kept and the set N1 represents the variables must be preserved in the input context since we consider a backward direction. The treatment of Or is diﬀerent: the term is systematically kept in the slice because it always inﬂuences the values of all the variables (through the introduction of subtrees in the derivation tree). Both branches are computed independently and the result gathers the information of these two branches. The rules for Exists and Call are not surprising. We assume that the variable x in Exists(x, U1 ) is unique in a normalised program; so x can be removed from the set of necessary variables yielded by the analysis of U1 (hence N1 − {x}). In the rule for Call, ﬁrst the derivation tree corresponding to the predicate Pk is computed with the set {xi | ¬Indep(C, D , {yi })} of variables to be preserved (i.e. the formal parameters xi of Pk bounded to arguments yi which are not independent from the set D ). The test in the rule for Call is similar to the test in the Op case. We could make more sophisticated choices to avoid including all the variables y1 ..., yn in the set of the necessary variables.

4

Derivation of the Dynamic on the Fly Analyser

We have presented in section 2 the semantics function S and the property αsl in functional form in section 3. The general organisation is described by the following diagram: S / PT C × TE EE ν EE a EE αsl E" Da The composition of the property αsl and the semantics S is a function of type C × T → Da , where Da is the domain of abstract values, the result of the analysis. This function computes successively the derivation tree related to a program, then the property of interest for this tree. It corresponds to a dynamic analysis a posteriori that inspects the trace produced after the program execution. It is interesting to formally describe dynamic analysers, because they are useful for instrumentation or debugging. We could also prefer dynamic analyses which, calculate their result on the fly i.e. during program execution. Their advantage is that they do not have to memorise traces before analysing them.

126

V. Gouranton

αsl (P T, RV, D) = let π = P T .stt.t.pp C = P T .stt.c D = D ∪ RV (π) in case (P T .lpt, P T .rn) of (nil, Op) : let Op (x1 , x2 , x3 ) = P T .stt.t.i in if RV (π) = ∅ and Indep(C, D, {x3 }) then (∅, D) else ({π}, D ∪ {x1 , x2 , x3 }) (nil, Eq) : let Eq (x, t) = P T .stt.t.i in if RV (π) = ∅ and UF(C, x, t) and Indep(C, D, P v(t) ∪ {x}) then (∅, D) else ({π}, D ∪ P v(t) ∪ {x}) ([P T1 , P T2 ], ∧) : let (S2 , N2 ) = αsl (P T2 , RV, D ) (S1 , N1 ) = αsl (P T1 , RV, N2 ) in if RV (π) = ∅ and S1 ∪ S2 = ∅ then (∅, D) else ( S1 ∪ S2 ∪ {π}, N1 ) ([P T1 , P T2 ], ∨) : let (S2 , N2 ) = αsl (P T2 , RV, D ) (S1 , N1 ) = αsl (P T1 , RV, D ) in ( S1 ∪ S2 ∪ {π}, N1 ∪ N2 ) (P T1 , ∃) : let Exists (x, U1 ) = P T .stt.t.i (S1 , N1 ) = αsl (P T1 , RV, D ) in if RV (π) = ∅ and S1 = ∅ then (∅, D) else (S1 ∪ {π}, N1 − {x}) (P T1 , Call) : let Call(Pk (y1 , ..., yn )) = P T .stt.t.i (S1 , N1 ) = αsl (P T1 , RV, {xi | ¬Indep(C, D , {yi })}) in if RV (π) = ∅ and S1 ∪ N1 = ∅ and Indep(C, D, {y1 , ..., yn }) then (∅, D) else (S1 ∪ {π}, D ∪ {y1 , ..., yn }) U F (C, x, t) = ∀θ ∈ C. θ =⊥⇒ ∃σ = mgu(θ(x), θ(t)) P v(t) = set of program variables occurring in t Rv(rt) = set of renaming variables occurring in rt Indep(C, D1 , D2 ) = ∀θ ∈ C. θ =⊥⇒ {Rv(θ(x)) | x ∈ D1 } ∩ {Rv(θ(x)) | x ∈ D2 } = ∅ Fig. 4. Slicing property

Deriving Analysers by Folding/Unfolding of Natural Semantics

127

The derivation of a dynamic on the fly analyser from a dynamic analyser a posteriori presents similarities with a well-known program transformation within the framework of functional programming. The program transformation is called deforestation [28] and its purpose is to eliminate the intermediate data structures induced by the composition of recursive functions. Here, the intermediate structure is the derivation tree of the natural semantics. We use folding and unfolding transformations to carry out deforestation. The three principal operations are the following: – unfoldings: we set νa (C, T ) = αsl (S (C, T )) and replace in the expression the calls to the recursive functions αsl and S by their deﬁnition. – applications of laws on the operators of the language (like the conditional ones, the expressions case and let ). – foldings which consist in replacing the occurrences of αsl (S (C , T )) from calls to νa (C , T ). The goal of these transformations is to remove all the calls to the property extraction function αsl , to obtain a closed deﬁnition of νa (C, T ). The function obtained is then a dynamic on the fly analyser since it does not build the intermediate derivation trees any more. The partial correction of the transformation by folding/unfolding is obvious. The total correction is not assured in general because some inopportune foldings can introduce cases of non-termination. The Improvement Theorem in [20] can be extented to a method (the extended improved unfold-fold method) presented in [21] which makes it possible to show the total correction of the method proposed in this paper.

Dynamic Slicing Analyser The deﬁnition of the dynamic slicing analyser for the logic programming language is the following: SLd (C, T, RV, D) = αsl (S (C, T ), RV, D) First, we use an unfolding technique applied to the semantics and the property functions. We present in [11] the transformation rules used for the derivation of the dynamic on the fly analyser by unfolding. Figure 5 presents these unfoldings for two rules (the other cases are straightforward). To obtain a dynamic on the fly analyser, we must apply folding steps that allows us to remove the calls of the function αsl . Figure 6 presents the result of these foldings. The fact that SLd itself calls S shows that it is a dynamic analysis.

128

V. Gouranton

SLd (C, T, RV, D) = let D = D ∪ RV (T.pp) in case T of (π, Op (x1 , x2 , x3 )) : if RV (π) = ∅ and Indep(C, D, {x3 }) then (∅, D) else ({π}, D ∪ {x1 , x2 , x3 }) (π, And (U1 , U2 )) : let P T1 = S (C, U1 ) R1 = P T1 .stt.nf (S2 , N2 ) = αsl (S (R1 , U2 ), RV, D ) (S1 , N1 ) = αsl (S (C, U1 ), RV, N2 ) in if RV (π) = ∅ and S1 ∪ S2 = ∅ then (∅, D) else ( S1 ∪ S2 ∪ {π}, N1 ) Fig. 5. Unfoldings of semantics and property functions

SLd (C, T, RV, D) = let D = D ∪ RV (T.pp) in case T of (π, Op (x1 , x2 , x3 )) : if RV (π) = ∅ and Indep(C, D, {x3 }) then (∅, D) else ({π}, D ∪ {x1 , x2 , x3 }) (π, And (U1 , U2 )) : let P T1 = S (C, U1 ) R1 = P T1 .stt.nf (S2 , N2 ) = SLd (R1 , U2 , RV, D ) (S1 , N1 ) = SLd (C, U1 , RV, N2 ) in if RV (π) = ∅ and S1 ∪ S2 = ∅ then (∅, D) else ( S1 ∪ S2 ∪ {π}, N1 ) Fig. 6. Dynamic (on the fly) slicing analysis

Deriving Analysers by Folding/Unfolding of Natural Semantics

5

129

Related Work

The fold/unfold transformation framework used here is based on seminal work by Burstall and Darlington [2,6]. The application of the technique to the derivation of programs has also been investigated in [5], which presents the synthesis of several sorting algorithms. The initial speciﬁcation is expressed in terms of sets and predicate logic constructs. Our transformations are also reminiscent of the deforestation technique [3,9,28]: in both cases the goal is to transform a composition of recursive functions into a single recursive deﬁnition. Generic frameworks for program analysis have been proposed in the context of logic programming languages [15] and data ﬂow analysis [26,30]. They rely on abstract interpretations of denotational semantics [15,26] or interpreters [30] and genericity is achieved by parameterising the abstract domains and choosing appropriate abstract functions. The implementation details of the analysis algorithm can be factorised. While these tools may attain a higher degree of mechanisation than our framework, they do not oﬀer to the user the same level of abstraction: they take as input the specification of an abstract interpreter rather than the specification of a property. Despite this diﬀerence of point of view, all these works are obviously inspired by the same goals. The framework introduced in [24] is closer to the spirit of the work presented in this paper but the technique itself is quite diﬀerent. Programs are represented as models in a modal logic and a data ﬂow analysis can be speciﬁed as a property in the logic. An eﬃcient data ﬂow analyser can be generated by partially evaluating a speciﬁc model checker with respect to the specifying modal formula. In comparison with this work, our framework trades mechanisation against generality: it is not limited to data ﬂow analyses but the derivation process by fold/unfold transformations is not fully automatic. Few papers have been devoted to the semantics of program slicing so far. A relationship between the behaviour of the original program and the behaviour of the slice is proved in [19]. The semantics of the language is expressed in terms of program dependence graphs; thus the programs are ﬁrst analysed in order to extract their dependences. This approach is well suited to the treatment of imperative languages. Formal deﬁnitions and a classiﬁcation of diﬀerent notions of slicing are provided in [27]. The main distinctions are backward vs forward analysers, executable vs non executable slices, and dynamic vs static analysers. Their deﬁnitions are based on denotational semantics and they focus on the speciﬁcations of the analyses. In [8] a description of a family of slicing algorithms generalising the notions of dynamic and static slice to that of a constrained slice is presented. Genericity with respect to the programming language is achieved through a translation into an intermediate representation called pim. Programs are represented as directed acyclic graphs whose semantics is deﬁned in terms of rewriting rules. Slicing is carried out using term graph rewriting with a technique for tracing dynamic dependence relations. It should be noted that a richer notion of slicing has been proposed for logic programming languages, which returns not only the set of program points that must be kept in the slice, but also the necessary variables at each program point [23]. This increased precision can also

130

V. Gouranton

be expressed in our framework, but we preferred to present the simpler version here for the sake of size and readability. By collecting the following information ( ∪ Si + {(π, D ∪ N )}, N ) i

we can modify straightforwardly each rule in order to get the same precision as [23].

6

Conclusion

We have presented a method to derive dynamic analysers by program transformation (folding/unfolding). A dynamic analyser is expressed as composition of a semantics and a property functions. The analyser is called a posteriori, it is a function computing ﬁrst a complete program execution trace (derivation tree) and then extracting the property of interest. A recursive deﬁnition of an analyser can be obtained by program transformation. This function is a dynamic on the fly analyser that computes the property during program execution. We have focussed on dynamic analysis in the body of paper. Our generic dynamic analyser is deﬁned in a strongly typed functional language8. As a consequence, we can rely on previous results on logical relations and abstract interpretation [1,4] in order to systematically construct static analysers from the dynamic analysers. The ﬁrst task is to provide abstract domains for the static slicing analyser and the corresponding abstraction functions. We recall that the type of the dynamic analyser is C × T × (PP → P(Pvar))× P(Pvar) → P(PP)× P(Pvar). Since PP → P(Pvar), P(Pvar) and P(PP) are already abstract domains associated with the dynamic analysis, only C needs to be abstracted9 . The next stage to derive a correct static analyser is to ﬁnd appropriate abstractions for the constants and operators occurring in the deﬁnition of the analyser. It is shown in [1] that the correctness of the abstract interpretation of the constants and operators of the language entails the correctness of the abstract interpretation of the whole language. The correctness of the abstract interpretation means that the results of the dynamic analysis and the static analysis are related if their arguments are. In fact, it is possible to deﬁne the most precise abstraction for each constant and operator of the language [1]. The basic idea to ﬁnd the best abstraction opa (v1a , . . . , vna ) of an operator op is to deﬁne it as the least upper bound of the abstractions of all the results of op applied to arguments vi belonging to the concretisation sets of the arguments of the via . The technique sketched here provides a systematic way to construct a correct abstract interpretation, and thus to derive a static analyser from a dynamic analyser [10,12]. By deriving static analysers as abstractions of dynamic analysers, we can see the dynamic 8 9

Note that the typing mentioned here has nothing to do with the language in which the analysed programs are written, this language itself can perfectly well be untyped. Of course, as usual in abstract interpretation, PP → P(Pvar), P(Pvar) and P(PP) can also be abstracted if further approximations are needed, but we do not consider this issue here.

Deriving Analysers by Folding/Unfolding of Natural Semantics

131

analyser either as an intermediate stage in the derivation of a static analyser (playing a role similar to a collecting semantics) or as the ﬁnal product of the derivation. The theory of abstract interpretation [4] provides a strong formal basis for static program analysis. The work described here does not provide an alternative formal underpinning for program analysis. Its goal is rather to put forward a derivation approach for the design of analysers from high level speciﬁcations. Our framework is applicable to a wide variety of languages, properties and type of service (dynamic or static). We have proposed in the body of the paper a formal deﬁnition of a dynamic slicing analyser for a logic programming language. To our knowledge, this deﬁnition is the ﬁrst one to be formal, so the beneﬁt of our approach is striking in this case. In [10], we present the derivation of dynamic and static analysers for a strictness analysis of a higher-order functional language and a live variable analysis for an imperative language. We have also applied this work for a globalisation analysis of a higher-order functional language and a generic sharing analysis. Pushing our approach ever further we arrive at a natural semantics format and a format for slicing, as presented in [12]. We have shown the correctness of the slicing property format. These formats can be instantiated for several programming languages (imperative language, logic programming language and functional language). The slicing property for the logic programming that we have presented here is an instantiation of the slicing format. As mentioned in the introduction, we wanted to establish the connection between the result of the analysis and its intended use. Analyses are generally performed to check assumptions about the behaviour of the program at speciﬁc points of its execution or to enable program optimisations. In both cases the intention of the analysis can be expressed in terms of a transformation and a relation as presented in [10,12]. The transformation depends on the result of the analysis and the relation establishes a correspondence between the semantics of the original program and the transformed program. For example, in the case of a program analysis for compiler optimisation the transformation expresses the optimisation that is allowed by the information provided by the analysis and the relation is the equality between the ﬁnal results (or outputs) of the original and the transformed program. It is not always the case that the relation is the equality: a counter-example is slicing analysis described in this paper (because the new program is required to behave like the original one only with respect to speciﬁc program points and variables). We have formally deﬁned and proved in [12] a property for the intention of a slicing analysis but space considerations prevent us from presenting the intentional property for slicing. There is a main aspect in which the work described here may seem limited: we have used only natural semantics and terminating programs. Structural Operational Semantics (SOS) are more precise than natural semantics and they are required for a proper treatment of non-determinism, non-termination and parallelism [17]. In fact, the natural semantics introduced in section 2 can be

132

V. Gouranton

replaced by SOS without diﬃculty10 and the dynamic analyses can be deﬁned in the very same way. The extra diﬃculty introduced by SOS is the fact that they create new program fragments which makes it necessary to abstract over the syntax of the language to derive a static analyser. This problem is discussed in [22]. We can also adapt our natural semantics to SOS by using the technique presented in [13]. To achieve this goal, the classical inductive interpretation of natural semantics has to be extended with coinduction mechanisms and rules must be deﬁned to express divergence.

Acknowledgements I would like to thank Pascal Fradet, Thomas Jensen, Daniel Le M´etayer and Ronan Gaugne for their useful comments. Thanks are also due to the anonymous referees SAS for their criticisms on an earlier draft.

References 1. S. Abramsky. Abstract interpretation, logical relations and Kan extensions. Journal of Logic and Computation, 1:5–40, 1990. 130, 130, 130 2. R. M. Burstall and J. Darlington. A transformation system for developing recursive programs. Journal of the ACM, 24:44–67, 1977. 129 3. W.N. Chin. Automatic methods for program transformation. PhD thesis, Imperial College, 1990. 129 4. P. Cousot and R. Cousot. Abstract interpretation: a uniﬁed lattice model for static analysis of programs by construction or approximation of ﬁxpoints. In POPL, pages 238–252, 1977. 130, 131 5. J. Darlington. A synthesis of several sorting algorithms. Acta Informatica, 11:1–30, 1978. 129 6. J. Darlington and R. M. Burstall. A system which automatically improves programs. Acta Informatica, 6:41–60, 1976. 129 7. T. Despeyroux. Typol: a formalism to implement natural semantics. Technical Report 94, INRIA, France, 1988. 115 8. J. Field, G. Ramalingam, and F. Tip. Parametric program slicing. In POPL, pages 379–392, 1995. 129 9. A. Gill, J. Launchbury, and S.L. Peyton Jones. A short cut to deforestation. In FPCA, pages 223–232, 1993. 129 10. V. Gouranton. D´erivation d’analyseurs dynamiques et statiques a ` partir de sp´ecifications op´ erationnelles. PhD thesis, Universit´e de Rennes, France, 1997. 116, 122, 130, 131, 131 11. V. Gouranton. Deriving analysers by folding/unfolding of natural semantics and a case study: slicing. Technical Report 3413, INRIA, France, 1998. 127 12. V. Gouranton and D. Le M´etayer. Dynamic slicing: a generic analysis based on a natural semantics format. Technical Report 3375, INRIA, France, 1998. 115, 130, 131, 131, 131 10

In order to deal with SOS, we basically need to change the type NF in STT and to introduce a global loop in the semantics since a SOS rule represents a single evaluation step.

Deriving Analysers by Folding/Unfolding of Natural Semantics

133

13. H. Ibraheem and D. A. Schmidt. Adapting big-step semantics to small-step style: coinductive interpretations and ”higher-order” derivations. In Second Workshop on Higher-Order Techniques in Operational Semantics (HOOTS2), 1997. 132 14. G. Kahn. Natural semantics. In STACS 87, number 247 in Lecture Notes in Computer Science, pages 22–39. Springer-Verlag, 1987. 115 15. B. Le Charlier, K. Musumbu, and P. Van Hentenryck. A generic interpretation algorithm and its complexity analysis. In ICLP, pages 64–78, 1991. 118, 129, 129 16. D. Miller and G. Nadathur. Higher-order logic programming. In ILPC, volume 225 of LNCS, pages 448–462, 1986. 117 17. H. Riis Nielson and F. Nielson. Semantics With Applications. John Wiley & Sons, 1992. 131 18. T. Reps and T. Turnidge. Program specialization via program slicing. In International Seminar on Partial Evaluation, 1996. 122 19. T. Reps and W. Yang. The semantics of program slicing. Technical Report 777, University of Wisconsin, Madison, 1988. 129 20. D. Sands. Proving the correctness of recursion-based automatic program transformations. In TAPSOFT. Springer-Verlag, 1995. 127 21. D. Sands. Total correctness by local improvement in the transformation of functional programs. TOPLAS, 18:175–234, 1996. 127 22. D.A. Schmidt. Abstract interpretation of small-step semantics. In 5th LOMAPS Workshop on Analysis and Verification of Multiple-Agent Languages. LNCS, 1996. 132 23. S. Schoenig and M. Ducass´e. A backward slicing algorithm for prolog. In SAS, number 1145 in LNCS, pages 317–331. Springer-Verlag, 1996. 122, 129, 130 24. B. Steﬀen. Generating data ﬂow analysis algorithms from modal speciﬁcations. Science of Computer Science, 21(2):115–139, 1993. 129 25. F. Tip. A survey of program slicing techniques. Journal of Programming Languages, 3:121–189, 1995. 122 26. G.A. Venkatesh. A framework for construction and evaluation of high-level speciﬁcations for program analysis techniques. In PLDI, volume 24, pages 1–12, 1989. 129, 129 27. G.A. Venkatesh. The semantics approach to program slicing. In PLDI, pages 107–119, 1991. 129 28. P. Wadler. Deforestation: transforming programs to eliminate trees. Theoretical Computer Science, 73:231–248, 1990. 127, 129 29. M. Weiser. Program slicing. IEEE Transactions on Software Engineering, 4:352– 357, 1984. 122 30. K. Yi and W. L. Harrison III. Automatic generation and management of interprocedural program analyses. In POPL, pages 246–259, 1993. 129, 129

A Symbolic Semantics for Abstract Model Checking? Francesca Levi Dipartimento di Informatica, Universit´ a di Pisa and ´ LIX, Ecole Polytechnique [email protected]

Abstract. We present a finite symbolic semantics of value-passing concurrent processes, that can be suitably interpreted over abstract values to compute a lower approximate semantics of full µ-calculus. The main feature of the semantics is that classical branching is replaced by explicit relations of non-deterministic and alternative choices among transitions. A combination of safe upper and lower approximations of the basic operators of the logic is used to handle negation. The relations of non-deterministic and alternative choices turn out to be very useful for the dual approximations of the existential next modality. Key words: Model checking, µ-calculus, abstract interpretation.

1

Introduction

Model Checking is a very successful technique for the automatic verification of temporal properties of reactive and concurrent systems, but it is only applicable to finite-state systems. Over the past few years, abstract interpretation has been widely applied to handle large as well as infinite systems with model checking [3,1,9,11,13,4,8,17,10,15]. Abstract interpretation [6,7] was originally conceived in the framework of data-flow analysis for designing approximate semantics of programs and relies on the idea of obtaining an approximate semantics from the standard one by substituting the concrete domain of computation and its basic operations with an abstract domain and corresponding abstract operations. The typical approach consists of constructing an abstract model over a chosen set of abstract states that can be used in model checking instead of the concrete one. To this aim the abstract model has to be safe, namely the formulas satisfied by the abstract model have to hold in the concrete one. For branching time logics the definition of a safe abstract transition relation among abstract states presents some basic difficulties and a single abstract transition relation cannot preserve both the existential and the universal next modality. Several authors [11,10,4] propose to adopt two different abstract transition relations: a ?

This work has been partially supported by the HCM project ABILE (ERBCHRXCT940624).

G. Levi (Ed.): SAS’98, LNCS 1503, pp. 134–151, 1998. c Springer-Verlag Berlin Heidelberg 1998

A Symbolic Semantics for Abstract Model Checking

135

free transition relation for computing the universal next modality, and a constrained transition relation for computing the existential next modality. Safeness of the free transition relation is ensured, if every concrete transition induces a free transition among the corresponding abstract states. In contrast, a constrained transition between abstract states is safe only if all the corresponding concrete transitions exist. Given this notion of safeness it turns out to be very difficult to effectively compute a sufficiently precise safe abstract model without constructing the concrete one. In this paper we propose a method for applying abstract interpretation to the µ-calculus model checking of value-passing concurrent processes. The main contribution is the definition of a symbolic semantics of processes in the style of [12], whose main feature is that explicit relations of non-determinism and alternative choice among transitions replace classical branching. Moreover, a finite graph for regular processes is achieved by avoiding the infinite paths of [12] due to parameterized recursion. Model checking of µ-calculus can be suitably performed by interpreting the obtained symbolic graph over concrete environments assigning concrete values to variables. However, since processes are capable of exchanging values taken from a typically infinite set, the fixpoint computation of µ-calculus semantics is not effectively computable. We define a technique to compute a lower approximation of the µ-calculus semantics by interpreting the symbolic graph over abstract environments on a given (finite) set of abstract values. Safeness of the lower approximation ensures indeed the preservation of any property. Following [13] the lower approximation is achieved by combining dual safe upper and lower abstract functions corresponding to all logical connectives except negation. The critical case is undoubtedly that of the next modality, where safe constrained and free transition relations among abstract processes have to be considered. We show that explicit non-deterministic and alternative choices between transitions allow us to avoid some typical problems due to abstract branching so that a more precise lower approximation of the next modality in particular is achieved with respect to previous proposals [10]. Finally, we discuss the basic problems that typically lead to miss optimality in the approximations of the next modalities. The paper is organized as follows. Section 2 presents value-passing concurrent processes, µ-calculus and concrete model checking. Section 3 summarizes the basic concepts of abstract interpretation. The symbolic graph is described in Sect. 4 and the corresponding model checking algorithm is shown in Sect. 5. Section 6 presents abstract model checking and Sect. 7 discuss optimality of abstract model checking.

2

Concrete Model Checking

We consider a value-passing version of CCS. Let V al a set of values (possibly infinite), Chan a set of channels and V ar a set of variables. Moreover, let Bexp and V exp be sets of boolean and values expressions. Processes P roc are generated

136

F. Levi

by the following grammar p ::= nil | x | a.p | be 5 p1 , p2 | p1 × p2 | p1 + p2 | p \ L | P (e1 , . . . , en ) where a ∈ {c!e, c?x, τ | c ∈ Chan, e ∈ V exp, x ∈ V ar}, L ⊆ Chan, and be ∈ Bexp is a boolean expression and ei ∈ V exp are expressions over values. Process c?x.p will receive a value v on the channel c and then behaves as p[v/x], where p[v/x] denotes the standard substitution of v for all free occurrences of x. Process c!e.p will send the value of the expression e and then behaves as p. The operator + represents choice, while × represents parallel composition. Process be 5 p1 , p2 behaves as p1 if the value of be is true, and as p2 otherwise. The operator \L is the standard restriction for a set of channels L. Finally, P (x1 , . . . , xn ) is a process constant, which has an associated definition P (x1 , . . . , xn ) ≡ p. We assume the usual definitions of free variables f v(p) and bound variables bv(p) of processes. A process p is closed iff f v(p) = ∅. In the following, we denote by capital letters T, P . . . open processes. For recursive processes P (x1 , . . . , xn ) ≡ p we assume f v(p) ⊆ {x1 , . . . , xn } and we assume the body to be guarded. The concrete semantics of processes is defined in Taa ble 1 of the appendix as a labelled transition system LT S(p) = (P ∗ , 7→) with actions a ∈ Act = {τ, c?v, c!v | c ∈ Chan, v ∈ V al}. Two semantic functions Sv : V exp → V al and Sb : Bexp → {tt, f f } for the evaluation of expressions are used. For a ∈ Act, we define chan(τ ) = ∅, chan(c?v) = chan(c!v) = {c}. Moreover, we denote by a ¯ the symmetric action of a, namely c!v for c?v and c?v for c!v. Note that, if the set of values is infinite, the labelled transition system is infinite and infinitely branching. For expressing temporal properties of processes we consider a simple extension of propositional µ-calculus [14]. Let Act be a set of actions and V AR be a set of logical variables. Formulas are inductively defined as follows. A ::= X | A ∧ A |< K > A | ¬A | µX.A where X ∈ V AR is a logical variable and K ∈ {τ, c?V, c!V | c ∈ Chan, V ⊆ V al}. We assume that each variable X occurs positively in formulas. The modality < K > subsumes the classical existential next modality < a > ranging over actions a ∈ Act and corresponds to ∨a∈K < a >. The dual universal modality is [K] ≡ ¬ < K > ¬. The operator µX.A denotes the least fixpoint and the dual operator of greatest fixpoint is equivalent to νX.A ≡ ¬µX.A[¬X/X]. Note that formulas with K infinite are needed for subsuming the logics CT L and CT L∗ , since Act can be infinite. For W instance, the classical W liveness property ∀F A can be expressed as µX.A ∨ [τ ]X c∈Chan [c?V al]X c∈Chan [c!sV al]X. Traditional global model checking corresponds to compute the semantics of a the formula k A k on the concrete labelled transition system LT S(p) = (P ∗ , 7→). ∗ ∗ Let δ : V AR → P(P ) be a valuation assigning subsets of P to logical variables. The semantics of an open formula A with respect to δ is defined as: k X kδ = δ(X) k A0 ∧ A1 kδ =k A0 kδ ∩ k A1 kδ k< K > A kδ = ∪a∈K k< a >k (k A kδ )

k ¬A kδ = P \ k A kδ k µX.A kδ = µV.(k A kδ[V /X] )

A Symbolic Semantics for Abstract Model Checking

137

where δ[V /X] stands for the valuation δ 0 , which agrees with δ except that δ 0 (V ) = δ(X). The next modality function k< a >k: P(P ∗ ) → P(P ∗ ) is given by k< a >k a (S) = {p ∈ P | ∃p 7→p0 , p0 ∈ S}. Therefore, p ∈k A k (p |= A) iff p ∈k A kδ0 , where δ0 is the empty evaluation.

3

Abstract Interpretation Theory

In this section we briefly recall the basic ideas of abstract interpretation, we refer the reader to [6,7] for more details. The theory of abstract interpretation provides a systematic method to design approximate semantics of programs by replacing the concrete domain of computation with a simpler abstract domain. The relation between the concrete and the abstract domain is precisely stated into a formal framework. Let (C, ≤) and (A, ≤# ) be two posets, where orderings ≤ and ≤# correspond to precision. A pair of functions (α, γ), where α : C → A (abstraction) and γ : A → C (concretization) is called a Galois connection iff ∀c ∈ C, ∀a ∈ A, α(c) ≤# a ⇔ c ≤ γ(c). These requirements can also be captured by saying that α is extensive (c ≤ γ(α(c))), γ is reductive (α(γ(a)) ≤# a), α and γ are total and monotonic. If α(γ(a)) = a, then (α, γ) is called a Galois insertion. Intuitively, the condition c ≤ γ(α(c)) ensures the loss of information of abstraction to be safe. On the other hand, condition α(γ(a)) ≤# a ensures that the concretization process introduce no loss of information. Let S(P ) be the semantics of a program P computed as the least fixpoint of a semantic function F over the concrete domain (C, ≤). The goal is that of computing an approximate semantics S # (P ) over (A, ≤# ), that is safe α(S(P )) ≤# S # (P ). The main result is that a safe approximate semantics S # (P ) can be computed as the least fixpoint of a safe approximate semantic function F # over (A, ≤# ), such that α(F (c)) ≤# F # (α(c)), for each c ∈ C. Moreover, it has been shown there there exists always a best approximate semantic function F # (optimal), when α(F (c)) = F # (α(c)). In the Galois insertion case this is equivalent to α(F (γ(a)) = F # (a), for each a ∈ A.

4

The Symbolic Graph

In this section we define the symbolic graph of processes. Symbolic semantics as introduced in [12] relies on the idea of using symbolic actions instead of concrete actions. For instance, transitions modeling input are represented by a single c?x transition c?x.p 7→ p. To handle conditional open processes symbolic actions (c,θ)

depend in addition on boolean guards. Thus, a symbolic transition T 7→ T 0 represents all concrete transitions with action a corresponding to symbolic action θ for the assignments to the free variables of T such that guard c is satisfied. Our definition of the symbolic graph differs in some aspects from the classical one. First, classical branching is replaced by explicit relations of non-deterministic (⊕) and alternative choices (⊗) among transitions. Moreover, a method to avoid

138

F. Levi

infinite applications of the standard rule for recursion is proposed. This way a finitely branching and finite graph for regular processes is obtained. We use the standard notions of substitutions and environments. A substitution is a partial function σ : V ar → V Exp and a simple substitution is an injective σ : V ar → V ar. We denote by Sub the set of simple substitutions. For a substitution σ we denote by tar(σ) and dom(σ) its target and source, respectively. An environment is a total function ρ : V ar → V al and Env is the set of environments. For an environment ρ we denote by ρ[x → v] the environment that agrees with ρ except that the value assigned to variable x is v. Let T be an open process and x ∈ V ar. We say that x is fresh in T iff x 6∈f v(T )∪bv(T ). Moreover, we say that a term T is free for a simple substitution σ with dom(σ) ⊆ f v(T ), iff for each x ∈ f v(tar(σ)), x 6∈ bv(T ) ∪ (f v(T ) \ tar(σ)). Let T be an open process and ρ an environment. We denote by (T, ρ) the closed process T ρ, that is obtained by substituting ρ(x) to variable x for each x ∈ f v(T ). Let C be the set of constraints with c ::= be | e = e | e ∈ V | true | ¬c | c ∧ c, where e ∈ V exp, be ∈ Bexp and V ⊆ V al. We denote by vars(c) the variables occurring in constraint c. For c ∈ C we consider the semantic function k c k: P(Env) → P(Env) obtained in the trivial way by k be k= {ρ ∈ Env | Sb (beρ) = tt}, k e1 = e2 k= {ρ ∈ Env | Sv (e1 ρ) = Sv (e2 ρ)} and k e ∈ V k= {ρ ∈ Env | Sv (eρ) ∈ V }. We use the notation ρ |= c for ρ ∈k c k. Let the symbolic actions be SymAct = {c?x, c!e, τ | c ∈ Chan, x ∈ V ar, e ∈ V exp} with bv(c?x) = {x}, bv(c!e) = bv(τ ) = ∅ and f v(c!e) = vars(e), f v(τ ) = f v(c?x) = ∅. Transitions of the symbolic semantics are T for each i ∈ {1, n} (i ∈ {1, . . . , n})

⊗i∈{1,n} Θi

−→

⊗i∈{1,n} Ωi , such that

1. Θi = (ci , ⊕ji ∈{1,ni } θi,ji ) with θi,ji ∈ SymAct ∪ {∗} and ci ∈ C; 2. Ωi = ⊕ji ∈{1,ni } Ti,ji where Ti,ji are open processes; 3. vars(ci ) ⊆ f v(T ), f v(θi,ji ) ⊆ f v(T ) and f v(Ti,ji ) ⊆ f v(T ) ∪ bv(θi,ji ). All possible behaviors of closed processes obtained from T are represented by ⊗i∈{1,n} Θi

−→ ⊗i∈{1,n} Ωi , where alternative choices are related a single transition T by ⊗ and non-deterministic choices by ⊕. The idea is that for each environment ρ there exists a unique alternative ci that is satisfied and symbolic actions θi,ji with corresponding processes Ti,ji represent the concrete transitions of (T, ρ). Let us explain informally the construction of transitions. The complete semantic rules are shown in Table 2 of the appendix. (true,a⊕∗)

−→ T ⊕ Transitions of a basic process a.T are obtained by a rule a.T a.T , where the special action ∗ denotes idle action and is used in the parallel composition rule. The non-deterministic choice of + is reflected by the composition of transitions with ⊕. The rule for choice is as follows {Ti T1 + T2

⊗j ∈{1,n } Θi,ji i i

−→

⊗ji ∈{1,ni } Ωi,ji }i∈{1,2}

⊗i∈{1,2},j ∈{1,n } Θj+ ,j i i 1 2

−→

⊗i∈{1,2},j∈{1,ni } Ωj+1 ,j2

+

A Symbolic Semantics for Abstract Model Checking

139

where Θj+1 ,j2 is guarded by constraint c1,j1 ∧c2,j2 and where all non-deterministic choices of Θi,ji = (ci,ji , ⊕hi ∈{1,kji } θi,ji ,hi ) for i ∈ {1, 2} are merged by ⊕. The resulting processes are analogously combined by Ωj+1 ,j2 . In fact, for each environment that satisfies both guards both actions of T1 and of T2 can be chosen. Example 1. For instance, a process T = c!x.T1 + a!x.T2 is modeled by c!x.T1 + a!x.T2

(true,c!x⊕a!x⊕∗)

−→

T1 ⊕ T2 ⊕ T.

The alternative choices of a conditional process are related by ⊗ through the following rule {Ti

⊗j ∈{1,n } Θi,ji i i

−→

be 5 T1 , T2

⊗ji ∈{1,ni } Ωi,ji }i∈{1,2}

¯ i,j ⊗i∈{1,2},j ∈{1,n } Θ i i i

−→

5

¯i,j ⊗i∈{1,2},j∈{1,ni } Ω i

¯ 1,j is equivalent to Θ1,j where the guard is additionally constrained by where Θ 1 1 ¯ 2,j is equivalent to Θ2,j where the guard is additionally constrained be, while Θ 2 2 by ¬be. Intuitively, transitions of be 5 T1 , T2 are either transitions of T1 , if be is satisfied, or transitions of T2 , if be is not satisfied. Θ ⊗Θ

1 2 Example 2. For T = x > o 5 c!x.T1 + a!x.T2 , b!x.nil we have T −→ Ω1 ⊗ Ω 2 , where Θ1 = (x > 0, c!x ⊕ a!x ⊕ ∗), Θ2 = (x ≤ 0, b!x ⊕ ∗), Ω1 = T1 ⊕ T2 ⊕ T and Ω2 = nil ⊕ T . In fact, for each environment ρ either x > 0 or x ≤ 0 is true and the process (T, ρ) is able to perform either both c!ρ(x) and a!ρ(x) or b!ρ(x), respectively.

The rule of parallel composition is quite complex. Since a single transition must represent all behaviors combined by the relations ⊕ and ⊗, a single rule performs at the same time synchronization and interleaving. The rule is defined as follows {Ti T1 × T2

⊗j ∈{1,n } Θi,ji i i

−→

⊗ji ∈{1,ni } Ωi,ji }i∈{1,2}

⊗i∈{1,2},j ∈{1,n } Θj× ,j i i 1 2

−→

×

⊗i∈{1,2},ji ∈{1,ni } Ωj×1 ,j2

Θj×1 ,j2

where is constrained by c1,j1 ∧ c2,j2 and its actions are all possible combinations of actions θ1,j1 ,h1 and θ2,j2 ,h2 for Θi,ji = (ci,ji , ⊕hi ∈{1,kji } θi,ji ,hi ) for i ∈ {1, 2}. The combination of processes corresponding to the combination of actions is realized by Ωj×1 ,j2 . Example 3. Consider an open process T1 ×T2 with T1 = x > 05c?x.T, a?x.T and T2 = c!y + 1.T + a!y − 1.T . We have T1 (true,c!y+1⊕a!y−1⊕∗)

(x>0,c?x⊕∗)⊗(x≤0,a?x⊕∗)

−→

T ⊕ T 1 ⊗ T ⊕ T1

and T2 −→ T ⊕ T ⊕ T2 . The transition resulting from parallel Θ1 ⊗Θ2 composition is T1 × T2 −→ Ω1 ⊗ Ω2 , where Θ1 = (x > 0, τ ⊕ c?z ⊕ c!y + 1 ⊕ a!y − 1 ⊕ ∗), Θ2 = (x ≤ 0, τ ⊕ a?z ⊕ c!y + 1 ⊕ a!y − 1 ⊕ ∗), Ω1 = T [y + 1/x] × T ⊕ T [z/x] × T2 ⊕ T1 × T ⊕ T1 × T ⊕ T1 × T2 and Ω2 = T [y − 1/x] × T ⊕ T [z/x] × T2 ⊕ T1 × T ⊕ T1 × T ⊕ T1 × T2 . Two alternative choices corresponding to constraints x > 0 and x ≤ 0 arise from composition of guard x > 0 with true and of guard

140

F. Levi

x ≤ 0 with true, respectively. For guard x > 0 the non-deterministic choices are obtained by the composition of actions c?x and ∗ with c!y +1, a!y −1 and ∗, while for guard x ≤ 0 they are obtained by the composition of actions a?x and ∗ with c!y + 1, a!y − 1 and ∗. Therefore, τ actions are obtained from the synchronization of actions c?x and c!y + 1 and of a?x with a!y − 1, respectively. The others arise from interleaving by composition with ∗. The corresponding processes are obtained in the obvious way. For instance, the process corresponding to τ with guard x > 0 is T [y +1/x]×T , since the value of y −1 is received by T1 . Note that in the interleaving case with a receive action the variable x must be renamed to z to avoid clash of variables with the free variables of T2 . Recursive processes are handled by the classical rule, where formal parameters are substituted by actual parameters. T [¯ e/¯ x] P (¯ e)

⊗i∈{1,n} Θi

−→

⊗i∈{1,n} Θi

−→

⊗i∈{1,n} Ωi

x) ≡ T rec P (¯

⊗i∈{1,n} Ωi

Unfortunately, the application of rule rec leads to an infinite graph also for regular processes, where there is no parallel composition inside the scope of recursion. The semantics of [12] suffers of the same problem. Example 4. Let us consider the process P (x) ≡ c!x.(P (x + 1) + P (x − 1)). Since the recursive process is unfolded infinitely times with a different argument an infinite graph arise. P (x)

(true,c!x⊕∗)

−→

P (x + 1) + P (x − 1) ⊕ P (x)

P (x + 1) + P (x − 1) P (x + 1) + P (x − 1) ...

(true,c!x+1⊕∗)

−→

P (x + 1 + 1) + P (x + 1 − 1) ⊕ P (x + 1)

−→

P (x − 1 + 1) + P (x − 1 − 1) ⊕ P (x − 1)

(true,c!x−1⊕∗)

This problem can be solved by replacing in the graph transitions of P (¯ e) by transitions of a process P (¯ x) for fresh variables x ¯. The semantics of P (¯ e) can naturally be obtained by the semantics of P (¯ x) by instantiating parameters x ¯ to the actual values corresponding to the evaluation of the expressions e¯. Let general processes GP have the following syntax GP ::= nil | a.T | P (¯ x) | GP1 + GP2 | be 5 GP1 , GP2 | GP1 × GP2 | GP \ L where x ¯ is a tuple of distinct variables and T ∈ P roc is a process. Note that a.T is a general process for any process T , since there are no current recursive calls. For general processes GP ∈ GP, the recursion variables rv(GP ) are defined as rv(a.T ) = rv(nil) = ∅, rv(P (¯ x)) = {¯ x}, rv(GP1 +GP2 ) = rv(GP1 × GP2 ) = rv(be 5 GP1 , GP2 ) = rv(GP1 ) ∪ rv(GP2 ) and rv(GP \ L) = rv(GP ). Our aim is that of finding a general process GP that can be used instead of T in the graph. For this purpose we define most general terms.

A Symbolic Semantics for Abstract Model Checking

141

Definition 5. Let T be a process. We define most general terms Π(T ): – if T = nil or T = a.T1 , Π(T ) = {T }; – for GPi ∈ Π(Ti ) such that, rv(GP1 ) ∩ f v(GP2 ) = ∅, rv(GP1 ) ∩ bv(GP2 ) = ∅ and vice-versa, then GP1 + GP2 ∈ Π(T1 + T2 ), GP1 × GP2 ∈ Π(T1 × T2 ) and be 5 GP1 , GP2 ∈ Π(be 5 T1 , T2 ); – if P (¯ z ) ≡ T and T is free for [¯ x/¯ z ], then P (¯ x) ∈ Π(P (¯ e)). Most general terms are general processes obtained by introducing fresh and distinct variables in current recursive calls. The following property is satisfied. Proposition 6. Let T be a process. For each GP ∈ Π(T ), there exists a substitution σ with dom(σ) = rv(GP ) such that GP σ = T . The substitution σ assigns actual parameters of T to formal parameters of GP . For ρ ∈ Env, let genT,GP : Env → Env, such that ρ(x) = genT,GP (ρ)(x), for x 6∈rv(GP ), and ρ(x) = Sv (σ(x)ρ), otherwise. In the environment genGP,T (ρ) formal parameters of GP are instantiated to the values of the actual parameters provided by σ. The result is that the closed process (GP, genT,GP (ρ)) is equivalent (bisimilar) to (T, ρ). For p1 , p2 ∈ P roc be processes, we say that p1 ≡ p2 a a iff for each a ∈ Act, for each p1 −→ p01 there exists p2 −→ p02 and p01 ≡ p02 and vice-versa. Proposition 7. Let T be a process and ρ ∈ Env. For each GP ∈ Π(T ), (GP, genT,GP (ρ)) ≡ (T, ρ). By proposition 2 the symbolic graph where transitions of T are replaced by transitions of GP correctly models the behavior of processes. Definition 8. Let T be a process. We define SG(T ) = (GP ∗ , T ∗ , ∗

∗

GP ⊆ GP, T ⊆ P roc and transitions are GP GP ∈ GP ∗ , where

⊗i∈{1,n} Θi

−→

⊗i∈{1,n} Θi

−→

) with

⊗i∈{1,n} Ωi for each

1. for each T 0 ∈ T ∗ there exists GP 0 ∈ GP ∗ ∩ Π(T 0 ); ⊗i∈{1,n} Θi

−→ ⊗i∈{1,n} Ωi , where Ωi = ⊕ji ∈{1,ni } Ti,ji , 2. T ∈ T ∗ and for each GP Ti,ji ∈ T ∗ , for i ∈ {1, n}, ji ∈ {1, ni }. Example 9. Consider for instance the process P (1), where P (x) is defined in example 4. Since P (z) + P (w) ∈ Π(P (x + 1) + P (x − 1)) then SG(P (1)) is finite: P (x)

(true,c!x⊕∗)

−→

P (z) + P (w)

P (x + 1) + P (x − 1) ⊕ P (x)

(true,c!z⊕c!w⊕∗)

−→

P (z + 1) + P (z − 1) ⊕ P (w + 1) + P (w − 1) ⊕ P (z) + P (w)

This graph correctly describes the behavior of P (1). For instance, the concrete c!1

c!2

computation P (1) 7→P (1 + 1) + P (1 − 1) 7→P (1 + 1 + 1) + P (1 + 1 − 1) . . . c!1

c!2

is simulated by (P (x), ρ1 ) −→ (P (z) + P (w), ρ2 ) −→ . . ., where ρ1 (x) = 1 and ρ2 (z) = 2, ρ2 (w) = 0. Environment ρ2 = genT,GP (ρ1 ) assigns to parameters z and w the result of the evaluation of expressions x + 1 and x − 1 with respect to ρ1 . Note that in P (z) + P (w) distinct fresh variables are used to model all recursive calls P (e1 ) + P (e2 ), where e1 may be different from e2 .

142

F. Levi

The equivalence is formally stated by the following theorem. Let K ∈ {τ, c!V, c?V } and θ ∈ SymAct. We define the constraint θ ∈ K ∈ C as θ ∈ K ≡ true, for K = θ = τ , θ ∈ K ≡ true, for θ = c?x and K = c?V , θ ∈ K ≡ e ∈ V , for θ = c!e and K = c!V , and θ ∈ K ≡ f alse, otherwise. For K = a we denote by a = θ the constraint θ ∈ a. ⊗i∈{1,n} Θi

−→ ⊗i∈{1,n} Ωi , where Theorem 10. 1. For each symbolic transition GP Θi = (ci , ⊗ji ∈{1,ni } θi,ji ) and Ωi = ⊗ji ∈{1,ni } Ti,ji and for each ρ ∈ Env, there exists one and only one i ∈ {1, n}, such that ρ |= ci and, for each c?v

ji ∈ {1, ni }, there exists a ∈ Act with ρ |= a = θi,ji such that, (GP, ρ) 7→ a (Ti,ji , ρ[x → v]) for θi,ji = c?x and v ∈ V al, and (T, ρ) 7→ (Ti,ji , ρ), for θi,ji ∈ {τ, c!e}; 2. for each p ∈ P ∗ there exists GP ∈ GP ∗ and ρ ∈ Env such that GP ρ ≡ p and, ⊗i∈{1,n} Θi

a

−→ for each p 7→p0 , there exist i ∈ {1, n} and ji ∈ {1, ni }, where GP ⊗i∈{1,n} Ωi with Θi = (ci , ⊗ji ∈{1,ni } θi,ji ) and Ωi = ⊗ji ∈{1,ni } Ti,ji , and ρ |= ci ∧ a = θi,ji and p0 ≡ (Ti,ji , ρ[x → v]), for a = c?v and θi,ji = c?x, and p0 ≡ (Ti,ji , ρ), for a ∈ {τ, c!v}. By theorem 1 the concrete semantics is safely represented by the symbolic one and in addition concrete non-deterministic and alternative choices are exactly composed by ⊕ and ⊗ in symbolic transitions. Since transitions are restricted to general processes, infinite applications of rec with different arguments are avoided and the symbolic graph is finite for regular processes. Theorem 11. Let p ∈ P roc be a regular process. SG(p) is a finite graph up to renaming.

5

Symbolic Model Checking

In this section we show that model checking can be realized by interpreting the symbolic graph over concrete environments. Let (P(D), ⊆), with D = {(T, ρ) | ⊗i∈{1,n} Θi

T ∈ GP ∗ and ρ ∈ Env} for SG(p) = (GP ∗ , T ∗ , −→ ). The semantics of a formula A is defined as in Sect. 2 by replacing evaluations with symbolic evaluations δ : V AR → P(D) and by taking the following function for the next modality. For K ∈ {τ, c!V, c?V } and S ∈ P(D), ⊗i∈{1,n} Θi

– k< K >kS (S) = {(T, ρ) | T −→ ⊗i∈{1,n} Ωi , with Θi = (ci , ⊕ji ∈{1,ni } θi,ji ) and Ωi = ⊕ji ∈{1,ni } Ti,ji , and there exist i ∈ {1, n} and ji ∈ {1, ni }, such that ρ |= ci ∧ θi,ji ∈ K and (Ti,j , ρ) ∈≡ S, for θi,ji ∈ {τ, c!e}, and (Ti,ji , ρ[x → v]) ∈≡ S, for some v ∈ V for K = c?V and θi,ji = c?x}. where (T, ρ) ∈≡ S if (GP, genT,GP (ρ)) ∈ S for some GP ∈ Π(T ). The definition of k< a >kS is based on the observation that process (T, ρ) is able to perform an action a, if the environment satisfies a guard ci for which there exists a symbolic action θi,ji corresponding to a. Moreover, the resulting process must be in S. However, since S is a set of general processes and Ti,ji is not a general process we look for an equivalent process (GP, genTi,ji ,GP (ρ)) ∈ S.

A Symbolic Semantics for Abstract Model Checking

143

Theorem 12. For each closed formula A, k A kS ∩P ∗ =k A k.

6

Abstract Model Checking

In this section we define the approximate semantics of the collecting semantics k A kS on the abstract domain obtained by replacing concrete environments with abstract environments on an abstract values domain. Let (αv , γv ) be a Galois insertion between the complete lattices (P(V al), ⊆) and (P(V al# ), ⊆) of concrete and abstract values. We consider the set of abstract environments Env # = {ρ# | ρ# : V ar → V al# } and the set of abstract processes ⊗i∈{1,n} Θi

D# = {(T, ρ# ) | T ∈ GP ∗ and ρ# ∈ Env # } for SG(p) = (GP ∗ , T ∗ , −→ ). Let γe : P(Env # ) → P(Env) and γ : P(D# ) → P(D) be the obvious functions induced by γv . Our purpose is that of computing a lower approximation of the semantics k A kl , such that γ(k A kl ) ⊆k A kS . Given the relation between concrete and abstract domain an approximate semantics can be obtained by replacing concrete functions with corresponding safe abstract functions. However, the operator of negation is not monotonic. Kelb [13] argues that a lower approximation of the full logic with negation can be obtained by combining duals approximations for formulas without negation k A kl and k A ku . In fact, k ¬A kl = D# \ k A ku is a safe lower approximation of k ¬A kS , since k A kS ⊆ γ(k A ku ). Analogously, k ¬A ku = D# \ k A kl is a safe upper approximation of k ¬A kS . Thus, the problem is reduced to the definition of safe dual approximations for all the logical operators except negation. The safeness of the dual approximations is formally defined with respect to the following framework. Proposition 13. There exist αl , αu : P(D) → P(D# ) such that (αu , γ) is a Galois insertion between (P(D), ⊆) and (P(D# ), ⊆) and (αl , γ) is a Galois insertion between (P(D), ⊇) and (P(D# ), ⊇). We show that both the upper and the lower approximation can be computed over the symbolic graph with respect to abstract environments. We exploit the following constraints based on relations ⊕ and ⊗. The symbol ∨ denotes with an abuse of notation the equivalent constraint. ⊗i∈{1,n} Θi

−→ ⊗i∈{1,n} Ωi , with Θi = (ci , ⊗ji ∈{1,ni } θi,ji ) and Definition 14. Let T Ωi = ⊗ji ∈{1,ni } Ti,ji , be a symbolic transition. For each K ∈ {τ, c!V, c?V } and i ∈ {1, n}, ji ∈ {1, ni } and I = ∪i∈{1,n} Ii with Ii ⊆ {1, ni } we define 1. F REEK,i,jiV≡ ci ∧ θi,ji ∈ K; W 2. CONK,I ≡ i∈{1,n} ¬ci ∨ ( ji ∈Ii θi,ji ∈ K). The abstract semantics is based on safe lower and upper approximations of the previous constraints and on dual approximations of the relation ∈≡ . Note that for each c ∈ C, k c ku and k c kl are safe approximations iff k c k⊆ γe (k c ku ) and γe (k c kl ) ⊆k c k, respectively. Moreover, ∈u≡ is a safe upper approximation,

144

F. Levi

if ρ ∈ γe (ρ# ) such that (T, ρ) ∈≡ γ(S # ) implies (T, ρ# ) ∈u≡ S # , while ∈l≡ is a safe lower approximation if (T, ρ# ) ∈l≡ S # only if, for all ρ ∈ γe (ρ# ), (T, ρ) ∈≡ γ(S # ). Let δ # : V AR → P(D# ) be an abstract evaluation. We define the abstract upper and lower semantics of an open formula A with respect to δ # as follows. Lower Approximation k X klδ# = δ # (X) k A0 ∧ A1 klδ# =k A0 klδ# ∩ k A1 klδ# k< K > A klδ# =k< K >kl (k A klδ# )

k µX.A klδ# = µV.(k A klδ# [V /X] ) k ¬A klδ# = D# \ (k A kuδ# )

where, for S # ∈ P(D# ), ⊗i∈{1,n} Θi

−→ ⊗i∈{1,n} Ωi , with Θi = (ci , ⊗ji ∈{1,ni } – k< K >kl (S # ) = {(T, ρ# ) | T θi,ji ) and Ωi = ⊗ji ∈{1,ni } Ti,ji , and there exists I = ∪i∈{1,n} Ii with Ii ⊆ {1, ni }, such that I is minimal and ρ# ∈k CONK,I kl , and, for each i ∈ {1, n} and ji ∈ Ii , (Ti,ji , ρ# ) ∈l≡ S # , for θi,ji ∈ {c!e, τ }, and (Ti,ji , ρ# [x → αv (v)]) ∈l≡ S # , for v ∈ V , K = c?V and θi,ji = c?x}. Upper Approximation k X kuδ# = δ # (X) k A0 ∧ A1 kuδ# =k A0 kuδ# ∩ k A1 kuδ# k< K > A kuδ# =k< K >ku (k A kuδ# )

k µX.A kuδ# = µV.(k A kuδ# [V /X] ) k ¬A kuδ# = D# \ (k A klδ# )

where, for S # ∈ P(D# ), ⊗i∈{1,n} Θi

−→ ⊗i∈{1,n} Ωi with Θi = (ci , ⊕ji ∈{1,ni } – k< K >ku (S # ) = {(T, ρ# ) | T θi,ji ) and Ωi = ⊕ji ∈{1,ni } Ti,ji , and there exist i ∈ {1, n} and ji ∈ {1, ni } such that ρ# ∈k F REEK,i,ji ku and (Ti,ji , ρ# ) ∈u≡ S # , for θi,ji ∈ {τ, c!e}, and (Ti,ji , ρ# [x → αv (v)]) ∈u≡ S # , for v ∈ V , K = c?V and θi,ji = c?x }. Let us explain the dual approximations of the next modality. 1. The upper approximation is safe, if (T, ρ) ∈k< a >kS (γ(S # )) implies (T, ρ# ) ∈k< a >ku (S # ) for ρ ∈ γe (ρ# ). In other words, safeness requires to consider at least the abstract transitions corresponding to the concrete a-transitions (free transition relation [10,4]). If constraint F REEa,i,ji ≡ ci ∧ θi,ji = a is safely upper approximate the previous condition is guaranteed, since ρ |= ci ∧ θi,ji = a implies ρ# ∈k ci ∧ θi,ji = a ku . 2. The definition of the lower approximation is quite complex. In this case safeness is ensured, only if (T, ρ# ) ∈k< a >kl (S # ) implies (T, ρ) ∈k< a >kS (γ(S # )), for each ρ ∈ γe (ρ# ). In other words, safeness requires to consider only the abstract transitions, for which all corresponding concrete a-transition exist (constrained transition relation [10,4]). Let us consider the

A Symbolic Semantics for Abstract Model Checking

145

V W constraint CONa,I ≡ i∈{1,n} (ci ⊃ ( ji ∈Ii (a = θi,ji ))) and its lower approximation. If ρ# ∈k CONa,I kl , then for each ρ ∈ γe (ρ# ), if ρ |= ci there exists an action θi,ji with ji ∈ Ii corresponding to a. By theorem 1 for each ρ there exists exactly one ci such that ρ |= ci so that, for each ρ there exists indeed a transition with action a. Note that by definition of the next modality safeness for both approximations requires in addition safeness of the dual approximations of ∈≡ to check the resulting processes to be in S # . Example 15. Consider a process T = x > 0 5 c?x.T1 , a?x.T2 , whose behavior is (x>0,c?x⊕∗)⊗(x≤0,a?x⊕∗)

described by T −→ T1 ⊕T ⊗T2 ⊕T . Let the values abstraction be αv (n) = • and let ρ# be the abstract environment with ρ# (x) = •. For a formula A ≡< c?n > true we obtain an upper approximation {(T, ρ# )} = k A ku . In fact, there exists ρ ∈ γe (ρ# ) such that ρ |= x > 0 ∧ (c?x = c?n) ≡ x > 0 ∧ true so that ρ# ∈k x > 0 ∧ true ku . Analogously, {(T, ρ# )} =k< a?n > true ku . The abstract operator is safe, since the existence of both concrete c?n

a?n

transitions (T, ρ1 ) 7→ (T1 [n/x], ρ1 ) and (T, ρ2 ) 7→ (T2 [n/x], ρ2 ) for ρ1 , ρ2 ∈ γe (ρ# ) is captured. Note that in the abstract process non-determinism among the two actions c?n and a?n arise even if in the concrete case these are two alternative choices. In contrast, let us consider the lower approximation for formula A. Since there a?n c?n exists ρ1 , ρ2 ∈ γe (ρ# ) such that (T, ρ1 ) 67→ and (T, ρ2 ) 67→ the lower approximation is safe if and only if (T, ρ# ) 6∈kA kl . Since k x > 0 kl = ∅, k x ≤ 0 kl = ∅ and c?n = a?x ≡ f alse, ρ# 6∈k(x ≥ 0 ∨ true) ∧ (x < 0 ∨ f alse) kl . Therefore, we have both k< c?n > true kl = ∅ and k< a?n > true kl = ∅. The abstract operator is able to observe that there is no real non-determinism between a?n and c?n, while this is an alternative choice as expressed by ⊗ and there are concrete processes for both alternatives. On the other hand, consider the values abstraction with αv (n) = P os, if n > 0, and αv (n) = N eg, otherwise. There are two abstract environments ρ# 1 , # # # with ρ# (x) = P os, and ρ , with ρ (x) = N eg. Since for each ρ ∈ γ(ρ 1 2 2 1 ), # l ρ ∈k c k for c = (x > 0 ⊃ true) ∧ (x ≤ 0 ⊃ f alse), then ρ1 ∈k c k is safe. With this evaluation of constraint we can safely obtain k< c?n > true kl = {(T, ρ# 1 )}. Since, for each concrete environment of ρ# only the alternative x > 0 is possible, 1 it is safe to conclude that all concrete processes indeed perform c?n. Lemma 16. Let k c kl and k c ku be safe lower and upper approximations for c ∈ C. Moreover, let ∈l≡ and ∈u≡ be safe lower and upper approximations of ∈≡ . For each S # ∈ P(D# ) and K ∈ {τ, c?V, c!V }, αu (k< K >kS (γ(S # ))) ⊆k< K >ku (S # ) and k< K >kl (S # ) ⊆ αl (k< K >kS )(γ(S # )). By lemma 1 the lower approximation of the full logic is safe. Theorem 17. Let k c kl and k c ku be safe lower and upper approximations for c ∈ C. Moreover, let ∈l≡ and ∈u≡ be safe lower and upper approximations of ∈≡ . For each closed µ-calculus formula A, k A kS ⊇ γ(k A kl ).

146

7

F. Levi

About Optimality

We have defined a method for constructing safe dual approximations of the next modality, that exploit dual approximations of constraints and of relation ∈≡ . It this setting precision of abstract model checking depends on precision of these approximations. An interesting problem is that of finding conditions on the approximations of constraints and of ∈≡ , that guarantee optimality of next modalities. It turns out that optimality of the upper approximation of constraint F REEK,i,ji and of the lower approximation of constraint CONK,I is sufficient, whenever SP(p) contains only general processes. ⊗i∈{1,n} Θi

Lemma 18. Let p be a closed process such that SP(p) = (GP ∗ , T ∗ , −→ ) where GP ∗ = T ∗ . Let k c ku and k c kl be optimal upper and lower approximations for c ∈ C. For each S # ∈ P(D# ) and K ∈ {τ, c?V, c!V }, αu (k< K >kS (γ(S # ))) ⊇k< K >ku (S # ) and αl (k< K >kS (γ(S # ))) ⊆k< K >kl (S # ). In the upper approximation case this result is quite obvious, since by optimality ρ# ∈k F REEa,i,ji ku implies the existence of an environment such that ρ ∈k F REEa,i,ji k, namely the existence of a corresponding concrete transition. The lower approximation is optimal, whenever there exists a constrained transition if and only if for each concrete process there exists a corresponding concrete a transition. Suppose that for each concrete process (T, ρ) 7→(T 0 , ρ0 ). For each ρ there exists a choice such that constraint ci is satisfied and a non-deterministic choice such that ρ |= θi,ji = a. Optimality is guaranteed, since theorem 1 ensures in addition that for each ρ all others alternatives cj are not true. Therefore, for each environment the constraint CONa,I is indeed satisfied so that by optimality the corresponding abstract transition is certainly considered. In contrast, if SP(p) contains non-general processes lemma 2 is no longer valid. Problems arise from the abstract evaluation of parameters included in the definition of ∈≡ . It is sufficient to consider the case of ρ# , such that ∃ρ1 , ρ2 ∈ γe (ρ# ) such that ρ1 |= F REEa,i,ji and (Ti,ji , ρ2 ) ∈≡ γ(S # ), but ρ1 6= ρ2 . Depending on the domain of values and expressions specific solutions must be studied for approximating constraints and ∈≡ . We suggest a general strategy. Safe dual approximations of constraints can be found on the basis of dual approximations of basic constraints be, e = e and e ∈ V and by combining lower and upper approximations in the obvious way. Definition 19. For c ∈ C, let k ¬c ku = P(Env # )\ k ¬c kl , k ¬c kl = P(Env # )\ k ¬c ku , k c1 ∧ c2 kl = ∩i∈{1,2} k ci kl and k c1 ∧ c2 ku = ∩i∈{1,2} k ci ku . If basic constraints are safely approximate these approximations are obviously safe. Unfortunately, they are in general non-optimal, since αu does not preserve ∩, while αl does not preserve ∪. Moreover, we have to compute dual approximations of ∈≡ , that realizes pa# # rameters evaluation. Let gen# T,GP : Env → P(Env ) be a safe approximation # # of genT,GP , where genT,GP (ρ) ∈ γe (gen# T,GP (ρ )) for each ρ ∈ γe (ρ ). This function can be suitably used to define approximations of ∈≡ .

A Symbolic Semantics for Abstract Model Checking

147

Definition 20. Let gen# T,GP be a safe approximation. We define # # # # 1. (T, ρ# ) ∈u≡ S # iff there exists ρ# 1 ∈ genT,GP (ρ ) such that (GP, ρ1 ) ∈ S ;

# # # # 2. (T, ρ# ) ∈l≡ S # iff for each ρ# 1 ∈ genT,GP (ρ ), (GP, ρ1 ) ∈ S .

If gen# T,GP is safe, then the previous approximations of ∈≡ are safe. The difficulties for computing optimal dual approximations are obvious. However, it is important to stress the essential role of ⊗ and ⊕ for limiting the loss of information the lower approximation. The relations ⊗ and ⊕ allows to improve the precision of the lower approximation with respect to the “trivial” definition, where a lower approximation of constraint F REEK,i,ji is considered. Example 21. Consider the process T1 × T2 with T1 = x > 0 5 c?x.T, a?x.T and T2 = c!y+1.T +a!y−1.T of example 3. With respect to the abstraction αv (n) = •, for each n ∈ N at, we trivially obtain ρ# ∈k c kl for c = (x > 0 ⊃ (τ = τ )) ∧ (x ≤ 0 ⊃ (τ = τ )) and ρ# (x) = •, since τ = τ ≡ true. Therefore, (T1 × T2 , ρ# ) ∈k< τ > true kl is established. The lower approximation of constraint c captures that there are processes for both alternatives, but in both cases the action τ τ can be performed. In other words, a constrained transition (T1 × T2 , ρ# ) 7→. . . is constructed. If we consider the trivial definition of lower approximation, we would not able to prove it, since neither ρ# ∈k (x > 0 ∧ (τ = τ )) kl nor ρ# ∈k (x ≤ 0 ∧ (τ = τ )) kl , even if the evaluation of constraints is optimal. Thus, this trivial method succeeds only if the alternative choice is the same for each concrete processes. In contrast, due to ⊗ and ⊕ a weaker condition can be considered and a more precise result is achieved. The proposed method could not give the same results on a classical symbolic semantics, where branching represents both non-deterministic and alternative choices, since it exploits the existence of exactly one alternative choice for each process and the representation of all non-deterministic choices for each possibility.

8

Related Works

The combination of abstract interpretation and model checking has been the topic of intensive research in the last few years. Much of the work concerned the definition of safe abstract model, namely of safe abstract transition relations, that preserves universal properties [2,3,1,17] and both universal and existential properties [11,10,13,4,15]. [11,10,4] propose the use of constrained and free transitions to handle branching modalities and [10] tackles also the problem of effectively computing for very simple programs an abstract model. The proposed method suffers of the problems of the trivial definition shown in example 7. A slight different approach is the one of Kelb [13], that investigates conditions for the safeness of abstract µ-calculus semantics instead that for the safeness of abstract models. In order to handle non-monotonic negation the combination of dual approximations is suggested. However, the problem of computing safe dual approximations of the next modality even over a given abstract model is not

148

F. Levi

addressed. In the framework of value-passing concurrent processes [5] proposes an abstract labelled transition system semantics for abstract closed processes obtained by an abstraction on the values domain. It is not obvious which class of temporal properties is preserved by the abstract model. In [15] it has been introduced the idea of representing the relation of non-determinism and of alternative choice among actions in order to compute more precise safe and constrained transition relations in the framework of closed abstract processes rather than in a symbolic approach with environments. Schmidt [17] shows a methodology for computing a finite approximate semantics of value-passing CCS by finitely approximating the semantics over abstract environments as a regular tree. Such a semantics is adequate for the verification of universal properties only. As far as concern the symbolic semantics, it is worth mentioning that other approaches have been proposed for representing regular processes by finite graphs [16].

9

Conclusions

In this paper we have applied abstract interpretation to the verification of µcalculus properties of value-passing concurrent processes. The main contribution is the definition of a finite symbolic semantics of processes, that differs from the classical one [12] in some aspects. First, classical branching of transitions is replaced by explicit relations of alternative and non-deterministic choices among transitions. Moreover, infinite branches are avoided by representing current recursive calls by means of general processes. The concrete semantics of µ-calculus can suitably be computed by interpreting the symbolic graph over concrete environments, but due to infinite values it is not effectively computable. We have proposed a technique to compute a lower approximate semantics on the symbolic graph by replacing concrete environments with abstract environments. Following the approach of [13] for explicitly treating negation the lower approximate semantics has been obtained by combining lower and upper approximations for each operator of the logic except negation. The relations of non-deterministic and alternative choices turn out to be very useful to approximate the next modality. The lower approximation in particular results undoubtedly more precise than previous proposals [11,10]. With respect to the classical approach to abstract model checking the proposed method does not rely on the construction of a safe abstract model, but on the computation of safe approximations of the “model checking” functions over the symbolic graph. This approach has several advantages. In order to prove a property it would be typically necessary to subsequently refine the chosen value abstraction by adding more information. This way the construction of the new abstract model is avoided. Moreover, this approach fits well in the traditional abstract interpretation framework and allows us to reason about safety and precision without introducing ad-hoc conditions as for instance the approximation ordering between abstract models of [4,10]. Recently, Schmidt [18] following the ideas of [19] has pointed out the very close connection among abstract model checking and data-flow analysis. These results suggest that the methods from

A Symbolic Semantics for Abstract Model Checking

149

one area can be usefully used in the other. There are many directions to continue this research. For instance, it seems interesting to study whether the refinement operators [7] that have been designed to systematically construct new more precise domain can be applied to the abstract model checking framework. Acknowledgments. I would like to thank Radhia and Patrick Cousot and Dave Schmidt for their helpful suggestions on the presentation of this work.

References 1. S. Bensalem, A. Bouajjani, C. Loiseaux, and J. Sifakis. Property preserving simulations. In Proceedings of CAV 92, volume 663 of Lecture Notes in Computer Science, pages 260–263. Springer-Verlag, Berlin, 1992. 2. E.M. Clarke, O. Grumberg, and D.E. Long. Model Checking and Abstraction. In Proc. 19th Annual ACM Symp. on Principles of Programming Languages, pages 343–354. ACM Press, 1992. 3. E.M. Clarke, O. Grumberg, and D.E. Long. Model checking and abstraction. ACM Transactions on Programming Languages and Systems, 5(16):1512–1542, 1994. 4. R. Cleaveland, P. Iyer, and D. Yankelevic. Optimality in Abstractions of Model Checking. In Proceedings of SAS 95, volume 983 of Lecture Notes in Computer Science, pages 51–63. Springer-Verlag, Berlin, 1995. 5. R. Cleaveland and J. Riely. Testing based abstractions for value-based systems. In Proceedings of CONCUR 94, volume 836 of Lecture Notes in Computer Science, pages 417–432. Springer-Verlag, Berlin, 1994. 6. P. Cousot and R. Cousot. Abstract Interpretation: A Unified Lattice Model for Static Analysis of Programs by Construction or Approximation of Fixpoints. In Proc. Fourth ACM Symp. Principles of Programming Languages, pages 238–252, 1977. 7. P. Cousot and R. Cousot. Systematic Design of Program Analysis Frameworks. In Proc. Sixth ACM Symp. Principles of Programming Languages, pages 269–282, 1979. 8. D. Dams. Abstract Interpretation and Partition Refinement for Model Checking. PhD thesis, Eindhoven university of Technology, 1996. 9. D. Dams, R. Gerth, and O. Grumberg. Generation of reduced models for checking fragments of CT L. In Proceedings of CAV 93, volume 697 of Lecture Notes in Computer Science, pages 479–490. Springer-Verlag, Berlin, 1993. 10. D. Dams, R. Gerth, and O. Grumberg. Abstract Interpretation of Reactive Systems. ACM Transactions on Programming Languages and Systems, 19(2):253–291, 1997. 11. D. Dams, O. Grumberg, and R. Gerth. Abstract interpretation of reactive systems: Abstractions preserving ∀CT L∗ , ∃CT L∗ and CT L∗ . In Proceedings of the Working Conference on Programming Concepts, Methods and Calculi (PROCOMET), 1994. 12. M. Hennessy and H. Lin. Symbolic bisimulations. Theoretical Computer Science, 138:353–389, 1995. 13. P. Kelb. Model Checking and Abstraction: A framework preserving both truth and failure information. Technical report, OFFIS, Oldenburg, Germany, 1994. 14. D. Kozen. Results on the Propositional mu-Calculus. Theoretical Computer Science, 27:333–354, 1983.

150

F. Levi

15. F. Levi. Abstract model checking of value-passing processes. In A. Bossi, editor, International Workshop on Verification, Model Checking and Abstract Interpretation, 1997. http://www.dsi.unive.it/ bossi/VMCAI.html. 16. H. Lin. Symbolic Transition Graph with Assignment. In Proc. of CONCUR 96, volume 1119 of Lecture Notes in Computer Science, pages 50–65. Springer-Verlag, Berlin, 1996. 17. D.A. Schmidt. Abstract Interpretation of Small-Step Semantics. In Proc. of the LOMAPS Workshop on “Analysis and Verification of Multiple-Agent Languages”, volume 1192 of Lecture Notes in Computer Science, pages 76–99, 1996. 18. D.A. Schmidt. Data Flow Analysis is Model Checking of Abstract Interpretation. In Proc. of the Annual ACM Symp. on Principles of Programming Languages, pages 38–48. ACM Press, 1998. 19. B. Steffen. Data Flow Analysis as Model Checking. In A. Meyer, editor, Proceedings of Theoretical Aspects of Computer Software (TACS 91), volume 526 of Lecture Notes in Computer Science, pages 346–364. Springer-Verlag, Berlin, 1991.

A

Appendix Table 1. The concrete semantics τ

c!v τ.p 7→p c?x.p c?v 7→p[v/x] v ∈ V al c!e.p 7→p Sv (e) = v a

a

pi 7→p0i

+

a

p1 + p2 7→p0i a

a

p1 7→p01 a

p1 × p2 7→p01 × p2

p1 × p2 7→p01 × p02

sync

T [¯ e/¯ x] 7→p a

P (¯ e) 7→p

a

be 5 p1 , p2 7→p02

int2 a

x) ≡ T rec P (¯

p1 7→p01 a

be 5 p1 , p2 7→p01

51 Sb (be) = tt

a

a

p2 7→p02

a

p1 × p2 7→p1 × p02

a

a ¯

p1 7→p01 p2 7→p02 τ

int1

p2 7→p02

52 Sb (be) = f f

p 7→p0 a

p \ L 7→p0 \ L

\ chan(a) ∩ L = ∅

The semantic rules for the symbolic semantics are based on some operations over actions. Let Θi,ji = (ci,ji , θi,ji ,1 ⊕ . . . ⊕ θi,ji ,kji ) and Ωi,ji = Ti,ji ,1 ⊕ . . . ⊕ Ti,ji ,kji , for i ∈ {1, 2}, ji ∈ {1, ni }, we define – Θj+1 ,j2 = (c1,j1 ∧ c2,j2 , ⊕h1 ∈{K1 } θ1,j1 ,h1 ⊕h2 ∈{1,K2 } θ1,j2 ,h2 ⊕ {∗}) and Ωj+1 ,j2 = ⊕h1 ∈{K1 } T1,j1 ⊕h2 ∈{1,K2 } T1,j2 ⊕T1 +T2 , for Ki = {hi ∈ {1, kji } | θi,ji ,hi 6= ∗}; ¯ 2,j = (¬be ∧ c2,j , θ2,j ,1 ⊕ ¯ 1,j = (be ∧ c1,j , θ1,j ,1 ⊕ . . . ⊕ θ1,j ,k ) and Θ – Θ 1 1 1 1 j1 2 2 2 ¯ ¯ . . . ⊕ θ2,j2 ,kj2 ), and Ωi,ji = ⊕hi ∈{1,kji } Ti,ji ,hi with T¯i,ji ,hi = be 5 T1 , T2 if θi,ji ,hi = ∗ and T¯i,ji ,hi = Ti,ji ,hi otherwise; – Θj×1 ,j2 = (c1,j1 ∧ c2,j2 , ⊕hi ∈{1,kji } θ1,j1 ,h1 × θ2,j2 ,h2 ) and Ωj×1 ,j2 = ⊕hi ∈{1,kji } 0 0 T1,j × T2,j where 1 ,h1 2 ,h2

A Symbolic Semantics for Abstract Model Checking

151

1. if θ1,j1 ,h1 ∈ {τ, c!e} and θ2,j2 ,h2 = ∗, then θ1,j1 ,h1 × θ2,j2 ,h2 = θ1,j1 ,h1 , 0 0 and T2,j = T2 (interleaving); T1,j1 ,h1 = T1,j 1 ,h1 2 ,h2 2. if θ1,j1 ,h1 = c?x and θ2,j2 ,h2 = ∗ then θ1,j1 ,h1 × θ2,j2 ,h2 = c?z, such 0 = that T1,j1 ,h1 is free for [z/x] and T2 is free for z. Moreover, T1,j 1 ,h1 0 T1,j1 ,h1 [z/x] and T2,j2 ,h2 = T2 (interleaving); 0 = T1 and 3. if θ1,j1 ,h1 = θ2,j2 ,j2 = ∗, then θ1,h1 ,h1 × θ2,j2 ,h2 = ∗ and T1,j 1 ,h1 0 T2,j2 ,h2 = T2 (idle action); 0 = 4. θ1,j1 ,h1 = c!e and θ2,j2 ,h2 = c?x, then θ1,j1 ,h1 × θ2,j2 ,h2 = τ and T1,j 1 ,h1 0 T1,j1 ,h1 and T2,j2 ,h2 = T2,j2 ,h2 [e/x] (synchronization); 5. symmetric rules; \ – for Θi = (ci , θi,1 ⊕. . .⊕θi,ni ) and Ωi = Ti,1 ⊕. . .⊕Ti,ni , Θi = (ci , ⊕ji ∈Ki θi,ji ) \ and Ki = {ji ∈ {1, ni } | chan(θi,ji ) ∩ L = ∅ and Ωi = ⊕ji ∈Ki Ti,ji \ L.

Table 2. The symbolic semantics (true,∗)

nil −→ nil {Ti T1 + T2

a.T

⊗j ∈{1,n } Θi,ji i i

−→

{Ti

P (¯ e)

T ⊕ a.T a ∈ {τ, c!e, c?x}

⊗ji ∈{1,ni } Ωi,ji }i∈{1,2}

−→

be 5 T1 , T2

T [¯ e/¯ x]

−→

⊗i∈{1,2},j ∈{1,n } Θj+ ,j i i 1 2

{Ti

T1 × T2

(true,a⊕∗)

⊗i∈{1,2},j∈{1,ni } Ωj+1 ,j2

⊗j ∈{1,n } Θi,ji i i

−→

⊗ji ∈{1,ni } Ωi,ji }i∈{1,2}

¯ 1,j ⊗ ¯ ⊗j ∈{1,n } Θ Θ 1 j2 ∈{1,n2 } 2,j2 1 1

−→

⊗j ∈{1,n } Θi,ji i i

−→

−→

−→

⊗i∈{1,n} Θi

−→

⊗i∈{1,n} Ωi

⊗i∈{1,n} Ωi

5

¯1,j1 ⊗j ∈{1,n } Ω ¯2,j2 ⊗j1 ∈{1,n1 } Ω 2 2

⊗ji ∈{1,ni } Ωi,ji }i∈{1,2}

⊗i∈{1,2},j ∈{1,n } Θj× ,j i i 1 2

⊗i∈{1,n} Θi

+

×

⊗i∈{1,2},ji ∈{1,ni } Ωj×1 ,j2

x) ≡ T rec P (¯

T

⊗i∈{1,n} Θi

−→

⊗i∈{1,n} Ωi \

T \L

⊗i∈{1,n} Θi

−→

\ ⊗i∈{1,n} Ωi

\

Automatic Determination of Communication Topologies in Mobile Systems Arnaud Venet ´ LIX, Ecole Polytechnique, 91128 Palaiseau, France. [email protected] http://lix.polytechnique.fr/˜venet

Abstract. The interconnection structure of mobile systems is very difficult to predict, since communication between component agents may carry information which dynamically changes that structure. In this paper we design an automatic analysis for statically determining all potential links between the agents of a mobile system specified in the πcalculus. For this purpose, we use a nonstandard semantics of the πcalculus which allows us to describe precisely the linkage of agents. The analysis algorithm is then derived by abstract interpretation of this semantics. Key words: π-calculus, nonstandard semantics, abstract interpretation.

1

Introduction

We are interested in analyzing the evolution of the interconnection structure, or communication topology, in a mobile system of processes, abstracting away all computational aspects but communication. Therefore, we can restrict our study to the π-calculus [18,19], which is a widely accepted formalism for describing communication in mobile systems. Whereas the communication topology of systems written in csp [13] or ccs [17] can be directly extracted from the text of the specification, a semantic analysis is required in the π-calculus, because communication links may be dynamically created between agents. In the absence of automatic analysis tools this makes the design and debugging of mobile systems very difficult tasks (see [11] for a detailed case study). In this paper we propose a semantic analysis of the π-calculus based on Abstract Interpretation [3,5] for automatically inferring approximate but sound descriptions of communication topologies in mobile systems. In a previous work [23] we have presented an analysis of the π-calculus which relies on a nonstandard concrete semantics. In that model recursively defined agents are identified by the sequence of replication unfoldings from which they stem, whereas the interconnection structure is given by an equivalence relation on the agent communication ports. That semantics is inspired of a representation of sharing in recursive data structures [14] which has been applied to alias analysis [10]. However, the equivalence relation does not capture an important piece of information for debugging and verification purposes: the instance of the G. Levi (Ed.): SAS’98, LNCS 1503, pp. 152–167, 1998. c Springer-Verlag Berlin Heidelberg 1998

Automatic Determination of Communication Topologies

153

channel that establishes a link between two agents. In this paper we redesign our previous analysis in order to take this information into account, while still preserving a comparable level of accuracy. Surprisingly enough, whereas our original analysis was rather complicated, involving heavy operations like transitive closure of binary relations, the refined one is tremendously simpler and only requires very basic primitives. The paper is organized as follows. In Sect. 2 we introduce our representation of mobile systems in the π-calculus. Section 3 describes the nonstandard semantics of mobile systems, which makes instances of recursively defined agents and channels explicit. The abstract interpretation gathering information on communication topologies is constructed in Sect. 4. In Sect. 5 we design a computable analysis which is able to infer accurate descriptions of unbounded and nonuniform communication topologies. Related work is discussed in Sect. 6.

2

Mobile Systems in the π-Calculus

We consider the asynchronous version of the polyadic π-calculus which was introduced by Turner [21] as a semantic basis of the pict programming language. This restricted version has simpler communication primitives and a more operational flavour than the full π-calculus, while still ensuring a high expressive power1 . Let N be a countable set of channel names. The syntax of processes is given by the following grammar: P ::= c![x1 , . . . , xn ] | c?[x1 , . . . , xn ].P | ∗c?[x1 , . . . , xn ].P | (P | P ) | (νx)P

Message Input guard Guarded replication Parallel composition Channel creation

where c, x, x1 , . . . , xn are channel names. Input guard and channel creation act as name binders, i.e. in the process c?[x1 , . . . , xn ].P (resp. (νx)P ) the occurrences of x1 , . . . , xn (resp. x) in P are considered bound. Usual rules about scoping, α-conversion and substitution apply. We denote by fn(P ) the set of free names of P , i.e. those names which are not in the scope of a binder. Following the cham style [1], the standard semantics of the π-calculus is given by a structural congruence and a reduction relation on processes. The congruence relation “≡” satisfies the following rules: (i) (ii) (iii) (iv) (v)

P ≡ Q whenever P and Q are α-equivalent. P | Q ≡ Q | P. P | (Q | R) ≡ (P | Q) | R. (νx)(νy)P ≡ (νy)(νx)P . (νx)P | Q ≡ (νx)(P | Q) if x 6∈fn(Q).

The reduction relation is defined in Fig. 1, where P {x1 /y1 , . . . , xn /yn } denotes the result of substituting every name xi for the name yi in P . This may involve α-conversion to avoid capturing one of the xi ’s. 1

We can encode the lazy λ-calculus for example [2].

154

A. Venet

c![x1 , . . . , xn ] | c?[y1 , . . . , yn ].P → P {x1 /y1 , . . . , xn /yn } c![x1 , . . . , xn ] | ∗c?[y1 , . . . , yn ].P → P {x1 /y1 , . . . , xn /yn } | ∗c?[y1 , . . . , yn ].P P → P0 (νx)P → P 0

P0 ≡ P

P → Q Q ≡ Q0 P 0 → Q0

P → P0 P | Q → P0 | Q

Fig. 1. Reduction relation in the standard semantics.

We now have to define what we mean by a “mobile system in the π-calculus”. We cannot simply allow a mobile system to be described by any process S. In fact, we are unable to design the nonstandard semantics, and hence the analysis, if we do not require S to be closed, i.e. fn(S) = ∅. In other words, we must consider the system in a whole. In order to make the semantic constructions of the following sections simpler, we add further constraints to the structure of mobile systems, which are inspired of Milner’s definition of friendly systems [18]. We denote by x a tuple (x1 , . . . , xn ) of channel names. We say that a process is a thread if it is made of a message possibly preceded by some input guards: c1 ?[x1 ] . . . cn ?[xn ].c![x]. We call resource a replicated process of the following form: ∗c?[x].(νy)(T1 | · · · | Tn ) where all the Ti ’s are threads. A mobile system S is then defined as: S ≡ (νc)(R1 | · · · | Rn | T0 ) where the Ri ’s are resources and T0 is a message, the initial thread, which originates all communications in the system. The names in c are called the initial channels. Therefore, all threads and channels created by a mobile system are fetched from its resources. Example 1. We model a system S which sets up a ring of communicating processes, where each component agent may only communicate with its left and right neighbours. The system generating a ring of arbitrary size is defined as follows: S ≡ (νmake)(νmon)(νleft 0 )(R1 | R2 | make![left 0 ]) where

R1 ≡ ∗make?[left].(νright)(mon![left, right] | make![right])

is the resource that adds a new component to the chain of processes and R2 ≡ ∗make?[left].mon![left, left 0 ] is the resource that closes the ring. The name “mon” should be seen as a reference to a hidden resource (for example a C program) which monitors the behaviour

Automatic Determination of Communication Topologies

155

of a ring component2 . For the sake of clarity, we denote by c the free names of all agents in the system at every stage of its evolution. Then, a ring with four components can be generated by S in four steps as follows: S → (νc)(R1 | R2 | mon![left 0 , right 1 ] | make![right 1 ]) → (νc)(R1 | R2 | mon![left 0 , right 1 ] | mon![right 1 , right 2 ] | make![right 2 ]) → (νc)(R1 | R2 | mon![left 0 , right 1 ] | mon![right 1 , right 2 ] | mon![right 2 , right 3 ] | make![right 3 ]) → (νc)(R1 | R2 | mon![left 0 , right 1 ] | mon![right 1 , right 2 ] | mon![right 2 , right 3 ] | mon![right 3 , left 0 ]) The right i ’s represent the successive instances of the channel right created at each request to R1 . As illustrated by the above example, the configuration of a mobile system at any stage of its evolution has the following particular form: (νc)(R1 | · · · | Rm | T1 | · · · | Tn ) | {z } | {z } Resources Threads where R1 , . . . , Rm are the resources originally present in the system. Intuitively, every thread Ti or channel ci present in the configuration could be unambiguously identified with the instant of its creation in the history of computations. Unfortunately, this precious information is not captured by the standard semantics. The process of α-conversion in particular, which is inherent to the definition of the semantics, destroys the identity of channels. The purpose of the next section is to introduce a refined semantics of the π-calculus which restores this information.

3

Nonstandard Semantics of Mobile Systems

Let S be a mobile system described in the π-calculus. In order to identify the threads and channels created by the system, we must be able to locate the syntactic components of S from which they stem. This is the role of the following notations. We denote by R(S) the number of resources in S. We assume that every resource is assigned a unique number r in {1, . . . , R(S)}. For any such r, we denote by Sr the corresponding resource in S and by T(r) the number of threads spawned by the resource. We similarly assign a unique number t in {1, . . . , T(r)} to every thread in Sr and we denote by Sr,t this thread, which has the following form: Sr,t = c1 ?[x1 ] . . . cA(r,t)−1 ?[xA(r,t)−1 ].cA(r,t) ![xA(r,t) ] 2

Recall that we only take communications into account, abstracting away all other computational aspects like the details of the monitoring procedure here.

156

A. Venet

where A(r, t) is the number of input/output actions in Sr,t . Note that A(r, t) is always nonzero because a thread contains at least one message. For 1 6 n 6 A(r, t), we denote by act(r, t, n) the n-th input/output action of Sr,t , and by Sr,t @ n the subterm cn ?[xn ] . . . cA(r,t) ![xA(r,t) ] of Sr,t starting at the n-th input/output action act(r, t, n). By convention, the initial thread is assigned the resource number 0. Finally, given a resource number r such that Sr ≡ ∗c?[x].(νy1 ) . . . (νyn )(T1 | · · · | TT(r) ) we put guard(r) = c?[x] and C(r) = {y1 , . . . , yn }. Example 2. We consider the system of Example 1 which contains two resources R1 and R2 . We put S1 = R1 and S2 = R2 . Thus we have: – guard(1) = make?[left], C(1) = {right}, T(1) = 2, S1,1 = mon![left, right], S1,2 = make![right], A(1, 1) = A(1, 2) = 1. – guard(2) = make?[left], C(2) = ∅, T(2) = 1, S2,1 = mon![left, left 0 ], A(2, 1) = 1. The initial thread is S0,1 = make![left 0 ].

A configuration of S in the nonstandard semantics is a finite set of thread instances. We do not need to represent resources since they are statically given and accessible by all the threads in any state of the system. A thread instance is a tuple hr, t, n, id , Ei where 1 6 r 6 R(S), 1 6 t 6 T(r), 1 6 n 6 A(r, t), id is a thread identifier and E is an environment. The thread identifier is the history of resource requests which led to the creation of the thread, starting from the initial one. A resource request being nothing more than a message to a replicated process, it can be identified with the thread that released this message. If we denote by Thr(S) = {(r, t) | 1 6 r 6 R(S), 1 6 t 6 T(r)} the set of all threads originally present in the system, then id ∈ Thr(S)∗ . The empty sequence ε is the identifier of the initial thread. The environment E maps every free name x ∈ fn(Sr,t @ n) of the thread instance to a channel instance. A channel instance is a tuple (r0 , y, id 0 ) where 1 6 r0 6 R(S), y ∈ C(r0 ) and id 0 is a channel identifier. Instances of initial channels are represented similarly except that they are assigned the resource number 0. The channel identifier is the history of resource requests that led to the creation of the instance of channel y by resource r0 . Therefore, channel identifiers and thread identifiers are represented identically, i.e. id 0 ∈ Thr(S)∗ . Similarly, the identifier of an initial channel is the empty sequence ε. We assume that there is no overlapping of scopes in the mobile system, i.e. we forbid terms like x?[y].y?[y].y![z] or ∗c?[y, z].(νy)z![y]. This assumption can always be satisfied by appropriate α-conversion. The transition relation “.” of the nonstandard operational semantics is defined in Fig. 2 and Fig. 3. It should be clear that without the hypothesis on name scoping, the definition of the resulting environments is ambiguous. The two transition rules correspond to the

Automatic Determination of Communication Topologies

157

two kinds of operations that may arise in a mobile system: resource fetching and communication between threads. The initial configuration C0 of the nonstandard semantics is given by C0 = {h0, 1, 1, ε, E0 i} where the environment E0 maps any z ∈ fn(S0,1 @ 1) to the instance (0, z, ε) of the corresponding initial channel. Example 3. Using the notations of Example 2, a ring with four components is described in the nonstandard semantics as follows: + * mon 7→(0, mon, ε), , ε), left 7→(0, left , 1, 1, 1, (0, 1), 0 right 7→(1, right, (0, 1)) * + mon 7→(0, mon, ε), left 7→(1, right, (0, 1)) 1, 1, 1, (0, 1).(1, 2), , right 7→(1, right, (0, 1).(1, 2)) + * mon 7→(0, mon, ε), 2 left 7→(1, right, (0, 1).(1, 2)) , , 1, 1, 1, (0, 1).(1, 2) 2 right 7→(1, right, (0, 1).(1, 2) ) + * mon 7→(0, mon, ε), 2 3 ) left 7→(1, right, (0, 1).(1, 2) , 2, 1, 1, (0, 1).(1, 2) left 0 7→(0, left 0 , ε) The generative process of the ring is now made entirely explicit thanks to the information carried by thread and channel identifiers (compare with the corresponding computation in the standard semantics given in example 1). Both semantics can be shown to be equivalent by defining the translation of a nonstandard configuration C of S to a term π(C) of the π-calculus and by applying a bisimulation argument. We may assume without loss of generality that the π-terms we will generate are built upon the set of names N 0 defined as the disjoint union of N with the set of all channel instances (r, y, id ) that could be created by S. We identify an environment with a substitution over N 0 . We denote by P σ the result of applying a substitution σ to a π-term P . Then, for any nonstandard configuration C = {hr1 , t1 , n1 , id 1 , E1 i, . . . , hrk , tk , nk , id k , Ek i} we define the translation π(C) as follows: 0 π(C) = (νc)(S1 E10 | · · · | SR(S) ER(S) | (Sr1 ,t1 @ n1 )E1 | · · · | (Srk ,tk @ nk )Ek )

where, for 1 6 r 6 R(S), Er0 is the environment which maps any z ∈ fn(Sr ) to (0, z, ε). The channels in c are all those which have a free occurrence in some agent of the top-level parallel composition of π(C). ∗

∗

Theorem 4. If C0 . C and C . C 0 , then π(C) → π(C 0 ). If C0 . C and π(C) → P , then there exists C 0 such that C . C 0 and P ≡ π(C 0 ).

158

A. Venet

If there are µ ∈ C and 1 6 r0 6 R(S) such that: – – – –

µ = hr, t, n, id , Ei act(r, t, n) = x![x1 , . . . , xk ] guard(r0 ) = c?[y1 , . . . , yk ] E(x) = (0, c, ε)

then

C . (C − {µ}) ∪ {hr0 , t0 , 1, id .(r, t), Et0 i | 1 6 t0 6 T(r0 )}

where, for all 1 6 t0 6 T(r0 ) and z ∈ fn(Sr0 ,t0 @ 1) if z = yi , 1 6 i 6 k E(xi ) if z ∈ C(r0 ) Et0 (z) = (r0 , z, id .(r, t)) (0, z, ε) otherwise, i.e. if z is an initial channel Fig. 2. Resource fetching.

A nonstandard configuration describes the communication topology of a mobile system at a particular moment of its evolution in terms of the resources initially present in the system. Therefore, the nonstandard semantics is a good basis for deriving an analysis, since the resulting information can be used to determine the role of each syntactic component of the system in the evolution of its interconnection structure. Constructing a computable abstraction of this semantics is the purpose of the next section.

4

Abstract Interpretation of Mobile Systems

We denote by C the set of all possible nonstandard configurations for a system ∗ S. We are actually interested in the set S = {C ∈ C | C0 . C} of configurations of the system which are accessible from the initial one by a finite sequence of computations. This is the collecting semantics of S [9], which can be expressed as the least fixpoint of the ∪-complete endomorphism IF on the complete lattice (℘(C), ⊆, ∪, ∅, ∩, C) defined as follows: IF(X) = {C0 } ∪ {C | ∃C 0 ∈ X : C 0 . C} Following the methodology of Abstract Interpretation [3,5], we construct a lattice (C ] , v, t, ⊥, u, >), the abstract domain, allowing us to give finite descriptions of infinite sets of configurations. This domain is related to ℘(C) via a monotone map γ : (C ] , v) → (℘(C), ⊆), the concretization function. Then we derive an abstract counterpart IF] : C ] → C ] of IF which must be sound with respect to γ, that is: IF ◦ γ ⊆ γ ◦ IF] . The abstract domain C ] is based upon a global abstraction of all environments in a nonstandard configuration. We assume that we are provided with a lattice (Id]2 , v2 , t2 , ⊥2 , u2 , >2 ) and a monotone map γ2 : (Id]2 , v2 ) →

Automatic Determination of Communication Topologies

159

If there are µ, ρ ∈ C such that: – – – – –

µ = hr, t, n, id , Ei ρ = hr0 , t0 , n0 , id 0 , E 0 i act(r, t, n) = x![x1 , . . . , xk ] act(r0 , t0 , n0 ) = y?[y1 , . . . , yk ] E(x) = E 0 (y)

then

C . (C − {µ, ρ}) ∪ {hr0 , t0 , n0 + 1, id 0 , E 00 i}

where, for all z ∈ fn(Sr0 ,t0 @ n0 + 1) E(xi ) E 00 (z) = E 0 (z)

if z = yi , 1 6 i 6 k otherwise

Fig. 3. Communication between threads.

(℘(Thr(S)∗ × Thr(S)∗ ), ⊆). This lattice is left as a parameter of our abstraction. It will be instantiated in the next section when we will set up an effective analysis. Let Chan(S) be the set {(r, t, n, x, r0 , y) | (r, t) ∈ Thr(S), n ∈ A(r, t), x ∈ fn(Sr,t @ n), 1 6 r0 6 R(S), y ∈ C(r0 )} of all possible syntactic relations between a free name in a thread and a channel created by a resource of the system. The abstract domain C ] is then defined as C ] = Chan(S) → Id]2 , the lattice operations being the pointwise extensions of those in Id]2 . Given an abstract configuration C ] , γ(C ] ) is the set of nonstandard configurations C such that, for any hr, t, n, id , Ei ∈ C and any x ∈ fn(Sr,t @ n), the following condition is satisfied: E(x) = (r0 , y, id 0 ) ⇒ (id , id 0 ) ∈ γ2 (C ] (r, t, n, x, r0 , y)) Monotonicity of γ is readily checked. For the sake of readability, we will denote an abstract configuration C ] by its graph {hκ1 7→id ]1 i, . . . , hκn 7→id ]n i}, where the κi ’s are in Chan(S) and each id ]j is in Id]2 . We may safely omit to write pairs of the form hκ 7→ ⊥2 i. The abstract semantics is given by a transition relation on abstract configurations. In the relation C1] C2] , the configuration C2] represents the modification to the communication topology of C1] induced by an abstract computation. Therefore, the function IF] is given by: IF] (C ] ) = C0] t C ] t

G

]

{C | C ]

]

C }

where C0] is the initial abstract configuration defined as {h0, 1, 1, x, 0, xi 7→ε2 | x ∈ fn(S0,1 @ 1)}, ε2 being a distinguished element of Id]2 such that (ε, ε) ∈ is defined in Fig. 4 and Fig. 5 by using the γ2 (ε2 ). The transition relation

160

A. Venet

If there are h(r, t, n, x, 0, c) 7→id ] i ∈ C ] and 1 6 r0 6 R(S) such that: – act(r, t, n) = x![x1 , . . . , xk ] – guard(r0 ) = c?[y1 , . . . , yk ] – id ] 6= ⊥2 then

C]

{hκ 7→id ]t0 ,z,r00 ,w i | κ = (r0 , t0 , 1, z, r00 , w), 1 6 t0 6 T(r0 )}

where, for all 1 6 t0 6 T(r0 ) push(r,t) (C ] (r, t, n, xi , r00 , w)) ⊥2 id ]t0 ,z,r00 ,w = dpush(r,t) (C ] (r, t, n, x, 0, c)) ⊥ 2 spush(C ] (r, t, n, x, 0, c))

if if if if if

z z z z z

= yi , 1 6 i 6 k ∈ C(r0 ) and (r00 , w) 6= (r0 , z) ∈ C(r0 ) and (r00 , w) = (r0 , z) is initial and (r00 , w) 6= (0, z) is initial and (r00 , w) = (0, z)

Fig. 4. Abstract resource fetching.

following abstract primitives: a “push” operation pushτ : Id]2 → Id]2 , a “double push” dpushτ : Id]2 → Id]2 and a “single push” spushτ : Id]2 → Id]2 defined for any τ ∈ Thr(S), a “synchronization” operator sync : Id]2 × Id]2 → Id]2 and a “swapping” operation swap : Id]2 → Id]2 . These operations depend on the choice of Id]2 , however they must satisfy some soundness conditions: – For any τ ∈ Thr(S) and id ] ∈ Id]2 , {(id .τ, id 0 ) | (id , id 0 ) ∈ γ2 (id ] )} ⊆ γ2 (pushτ (id ] )). – For any τ ∈ Thr(S) and id ] ∈ Id]2 , {(id .τ, id .τ ) | (id , id 0 ) ∈ γ2 (id ] )} ⊆ γ2 (dpushτ (id ] )). – For any τ ∈ Thr(S) and id ] ∈ Id]2 , {(id .τ, ε) | (id , id 0 ) ∈ γ2 (id ] )} ⊆ γ2 (spushτ (id ] )). – For any id ] ∈ Id]2 , {(id 0 , id ) | (id , id 0 ) ∈ γ2 (id ] )} ⊆ γ2 (swap(id ] )). – For any id ]1 , id ]2 ∈ Id]2 , {(id 1 , id 2 ) | ∃id 0 ∈ Thr(S)∗ : (id 1 , id 0 ) ∈ γ2 (id ]1 ) ∧ (id 2 , id 0 ) ∈ γ2 (id ]2 )} ⊆ γ2 (sync(id ]1 , id ]2 )). Intuitively, the pushτ operation concatenates τ to the first component of every pair of identifiers. The dpushτ and spushτ operations act similarly, except that dpushτ duplicates the first component into the second one and spushτ sets the second component to ε. The swap operation permutes the components of every pair of identifiers. The sync operation extracts pairs of identifiers corresponding to redexes, i.e. pairs of agents linked to the same instance of a channel. ]

Proposition 5. If C ∈ γ(C ] ) and C . C 0 , then there exists C in C ] , such that ] ] C and C 0 ∈ γ(C ] t C ). C] The soundness of IF] is then a simple consequence of the previous result. It is remarkable that the soundness of the whole semantics depends only on very

Automatic Determination of Communication Topologies

161

If there are hκ1 7→id ]1 i, hκ2 7→id ]2 i ∈ C ] such that: – – – – –

κ1 = (r, t, n, x, r 00 , u) κ2 = (r0 , t0 , n0 , y, r 00 , u) act(r, t, n) = x![x1 , . . . , xk ] act(r0 , t0 , n0 ) = y?[y1 , . . . , yk ] id ]s = sync(id ]1 , id ]2 ) 6= ⊥2

then

C]

{hκ 7→id ]z,r000 ,w i | κ = (r0 , t0 , n0 + 1, z, r000 , w)}

]

where, if we put id i,r000 ,w = C ] (r, n, t, xi , r000 , w) for 1 6 i 6 k ( ] swap(sync(swap(id i,r000 ,w ), swap(id ]s ))) if z = yi , 1 6 i 6 k ] id z,r000 ,w = C ] (r0 , t0 , n0 , z, r 000 , w) otherwise Fig. 5. Abstract communication between threads.

simple conditions over some primitive operations. This means that we only need to construct the domain Id]2 and instantiate those operations to obtain a computable and sound abstract semantics. This is the goal that we will achieve in the next section.

5

Design of a Computable Analysis

The collecting semantics S is the least fixpoint of IF. Therefore, by Kleene’s theorem, it is the limit of the following increasing iteration sequence:

S0 = ∅ Sn+1 = IF(Sn )

Following [3,9] we will compute a sound approximation S ] of S by mimicking this iteration, using IF] instead of IF. Since the resulting computation may not terminate, we use a widening operator to enforce convergence in finitely many steps. A widening operator ∇ : C ] × C ] → C ] must satisfy the following conditions: – For any C1] , C2] ∈ C ] , C1] t C2] v C1] ∇ C2] .

]

– For any sequence (Cn] )n>0 , the sequence (C n )n>0 defined as: (

is ultimately stationary.

]

C 0 = C0] ] ] ] C n+1 = C n ∇ Cn+1

162

A. Venet

Note that we can construct a widening on C ] from an existing widening ∇2 on Id]2 by pointwise application of ∇2 . We define the approximate iteration sequence (S ] n )n>0 as follows: ] S 0 = ⊥ S ] n+1 = S ] n ∇ IF] (S ] n ) ] S n+1 = S ] n

¬(IF] (S ] n ) v S ] n ) IF] (S ] n ) v S ] n

if if

Convergence is ensured by the following result: Theorem 6 [9]. The sequence (S ] n )n>0 is ultimately stationary and its limit S ] satisfies S ⊆ γ(S ] ). Moreover, if N > 0 is such that S ] N +1 = S ] N , then for all n > N , S ] n = S ] N . This provides us with an algorithm for automatically computing a sound approximation of S. It now remains to instantiate the domain Id]2 and the associated abstract primitives. We will design two abstractions of ℘(Thr(S)∗ × Thr(S)∗ ), each one capturing a particular kind of information. Our first abstraction captures sequencing information and is based upon an approximation of thread and channel identifiers by regular languages. Let (Reg, ⊆, ∪, ∅, ∩, Thr(S)∗ ) be the lattice of regular languages over the alphabet Thr(S) ordered by set inclusion. We define Id]reg as the product lattice Reg × Reg. The concretization γreg is given by γreg (L1 , L2 ) = L1 × L2 . The associated abstract primitives are defined as follows: pushreg τ (L1 , L2 ) = (L1 .τ, L2 ) dpushreg τ (L1 , L2 ) = (L1 .τ, L1 .τ ) spushreg τ (L1 , L2 ) = (L1 .τ, ε) swapreg (L1 , L2 ) = (L2 , L1 ) (L1 , L01 ) – syncreg ((L1 , L2 ), (L01 , L02 )) = (∅, ∅)

– – – –

if L2 ∩ L02 6= ∅ otherwise

The soundness conditions are easily checked. The element εreg 2 is given by (ε, ε). Since Id]reg may have infinite strictly increasing chains, we must define a widening operator ∇reg 2 . It is sufficient to construct a widening ∇reg on Reg and to apply it componentwise to elements of Id]reg . A simple choice for L1 ∇reg L2 consists of quotienting the minimal automaton of L1 ∪ L2 such that any letter of the alphabet may occur at most once in the automaton. The resulting automaton is minimal, and there are finitely many such automata, which ensures the stabilization property. The second approximation captures counting relations between the components of a tuple of thread or channel identifiers. This will allow us to give nonuniform descriptions of recursively defined communication topologies. Suppose that we are given an infinite set of variables V. We assign two distinct variables xτ and yτ to each element τ of Thr(S). Now we consider a finite system K of linear equality constraints over the variables V with coefficients in Q. If we denote

Automatic Determination of Communication Topologies

163

by |id |τ the number of occurrences of τ in the sequence id , the concretization γnum (K) of K is the set of all pairs (id , id 0 ) such that the following variable assignment: {xτ 7→ i|d |τ , yτ 7→ i|d 0 |τ | τ ∈ Thr(S)} is a solution of K. The domain of finite systems of linear equality constraints over V ordered by inclusion of solution sets can be turned into a lattice Id]num . This domain has been originally introduced by Karr [15]. We refer the reader to the original paper for a detailed algorithmic description of lattice operations. We could use more sophisticated domains of computable numerical constraints such as linear inequalities [6] or linear congruences [12], but the underlying algorithmics is much more involved. Nevertheless, giving a rigorous construction of the abstract primitives on Id]num would be still very technical. Therefore, for the sake of readability, we only outline the definition of these primitives: (K) is the system of constraints K in which we have replaced every – pushnum τ occurrence of the variable xτ by the expression xτ − 1. – If K is a system of linear equality constraints, we denote by Kx the system in which we have removed all constraints involving a variable yτ . Then (K) is the system pushnum (Kx ) with the additional constraints dpushnum τ τ xτ 0 = yτ 0 , for any τ 0 ∈ Thr(S). – Similarly, spushnum (K) is the system pushnum (Kx ) with the additional τ τ constraints yτ 0 = 0, for any τ 0 ∈ Thr(S). – swapnum (K) is the system in which we have replaced each occurrence of xτ by yτ and vice-versa. – Let K1 and K2 be two systems of linear equality constraints. For any τ ∈ Thr(S), let x0τ and yτ0 be fresh variables of V. Let K20 be the system K2 in which we have substituted each occurrence of xτ (resp. yτ ) by x0τ (resp. yτ0 ). We construct the system K1,2 as the union of K1 and K20 together with the additional constraints yτ = yτ0 , for any τ ∈ Thr(S). Then, we define syncnum (K1 , K2 ) as the system K1,2 in which we have removed all constraints involving a variable yτ or yτ0 , each remaining variable x0τ being renamed in yτ . Note that a normalization pass (namely a Gauss reduction) has to be performed is on the system after or during each of these operations. The element εnum 2 given by the system of constraints {xτ = 0, yτ = 0 | τ ∈ Thr(S)}. Since we only consider systems defined over the finite set of variables {xτ , yτ | τ ∈ Thr(S)}, we cannot have infinite strictly increasing chains [15]. Therefore, we can use the join operation t num as a widening. Example 7. We consider the product of domains Id]reg and Id]num and we run the analysis on the system of Example 1 with the notations of Example 2. For the sake of readability, at each step we only write the elements of the abstract configuration that differ from the previous iteration. Moreover, we do not figure trivial constraints of the form xτ = 0 whenever they can be deduced from the Id]reg component.

164

A. Venet

First iteration.

h0, 1, 1, make, 0, makei 7→ h(ε, ε), >num i, h0, 1, 1, left 0 , 0, left 0 i 7→ h(ε, ε), >num i

Second iteration. h1, 1, 1, mon, 0, moni 7→ h((0, 1), ε), x(0,1) = 1i, h1, 1, 1, left, 0, left i 7→ h ( (0, 1), ε), x = 1i, (0,1) 0 = y = 1i, h1, 1, 1, right, 1, righti 7→ h ( (0, 1), (0, 1)), x (0,1) (0,1) h1, 2, 1, make, 0, makei 7→ h((0, 1), ε), x(0,1) = 1i, h1, 2, 1, right, 1, righti 7→ h((0, 1), (0, 1)), x(0,1) = y(0,1) = 1i, h2, 1, 1, mon, 0, moni 7→ h((0, 1), ε), x(0,1) = 1i, h2, 1, 1, left , 0, left i 7→ h ( (0, 1), ε), x = 1i, (0,1) 0 0 h2, 1, 1, left, 0, left 0 i 7→ h((0, 1), ε), x(0,1) = 1i Third iteration. h1, 1, 1, mon, 0, moni 7→h((0, 1).(ε + (1, 2)), ε),x(0,1) = 1i, x(0,1) = y(0,1) = 1 , h1, 1, 1, left, 0, righti 7→ ((0, 1).(1, 2), (0, 1)), = 1 x (1,2) * ((0, 1).(ε + (1, 2)), (0, 1).(ε + (1, 2))), + = y = 1 x , h1, 1, 1, right, 1, righti 7→ (0,1) (0,1) x = y (1,2) (1,2) h1, 2, 1, make, 0, makei 7→ h((0, 1).(ε + (1, 2)), ε), x(0,1) = 1i, * ((0, 1).(ε + (1, 2)), (0, 1).(ε + (1, 2))), + = y = 1 x h1, 2, 1, right, 1, righti 7→ , (0,1) (0,1) x = y (1,2) (1,2) = 1i, h2, 1, 1, mon, 0, moni 7→ h ( (0, 1).(ε + (1, 2)), ε), x (0,1) h2, 1, 1, left , 0, left i 7→ h ( (0, 1).(ε + (1, 2))), ε), x = 1i, (0,1) 0 0 x(0,1) = y(0,1) = 1 h2, 1, 1, left, 0, righti 7→ ((0, 1).(1, 2), (0, 1)), x(1,2) = 1 Fourth iteration. h1, 1, 1, mon, 0, moni 7→ h((0, 1).(1, 2)∗ , ε), x(0,1) = 1i, * ((0, 1).(1, 2)∗ , (0, 1).(ε + (1, 2))), + x = y = 1 , h1, 1, 1, left, 0, righti 7→ (0,1) (0,1) = y + 1 x (1,2) (1,2) * + ∗ ∗ ((0, 1).(1, 2) , (0, 1).(1, 2) ), = y = 1 x h1, 1, 1, right, 1, righti 7→ , (0,1) (0,1) x = y (1,2) (1,2) h1, 2, 1, make, 0, makei 7→ h((0, 1).(1, 2)∗ , ε), x(0,1) = 1i, * ((0, 1).(1, 2)∗ , (0, 1).(1, 2)∗ ), + x(0,1) = y(0,1) = 1 , h1, 2, 1, right, 1, righti 7→ = y x (1,2) (1,2) ∗ h2, 1, 1, mon, 0, moni 7→ h ( (0, 1).(1, 2) , ε), x = 1i, (0,1) ∗ , 0, left i 7→ h ( (0, 1).(1, 2) , ε), x = 1i, h2, 1, 1, left (0,1) 0 0 * + ∗ ((0, 1).(1, 2) , (0, 1).(ε + (1, 2))), h2, 1, 1, left, 0, righti 7→ = y = 1 x (0,1) (0,1) x(1,2) = y(1,2) + 1

Automatic Determination of Communication Topologies

165

Fifth iteration. * ((0, 1).(1, 2)∗ , (0, 1).(1, 2)∗ ), + x = y = 1 , h1, 1, 1, left, 0, righti 7→ (0,1) (0,1) x(1,2) = y(1,2) + 1 * ((0, 1).(1, 2)∗ + , (0, 1).(1, 2)∗ ), = y = 1 x h2, 1, 1, left, 0, righti 7→ (0,1) (0,1) x(1,2) = y(1,2) + 1 At the sixth iteration step we find the same configuration. Therefore, following Theorem 6, we know that the limit has been reached. Putting all previous computations together, we obtain: h0, 1, 1, make, 0, makei 7→ h(ε, ε), >num i, h0, 1, 1, left 0 , 0, left 0 i 7→ h(ε, ε), >num i, ∗ , ε), x = 1i, h1, 1, 1, mon, 0, moni 7→ h ( (0, 1).(1, 2) (0,1) h1, 1, 1, left, 0, left i 7→ h ( (0, 1), ε), x = 1i, (0,1) 0 + * ∗ ∗ , (0, 1).(1, 2) ), ((0, 1).(1, 2) x = y = 1 , h1, 1, 1, left, 0, righti 7→ (0,1) (0,1) = y + 1 x (1,2) (1,2) * + ∗ ∗ ((0, 1).(1, 2) , (0, 1).(1, 2) ), = y = 1 x h1, 1, 1, right, 1, righti 7→ , (0,1) (0,1) x = y (1,2) (1,2) ] S = ∗ h1, 2, 1, make, 0, makei 7→ h((0, 1).(1, 2) , ε), x(0,1) = 1i, * ((0, 1).(1, 2)∗ ∗ + , (0, 1).(1, 2) ), h1, 2, 1, right, 1, righti 7→ = y = 1 x , (0,1) (0,1) x = y (1,2) (1,2) ∗ h2, 1, 1, mon, 0, moni 7→ h((0, 1).(1, 2) , ε), x(0,1) = 1i, ∗ , 0, left i 7→ h ( (0, 1).(1, 2) , ε), x = 1i, h2, 1, 1, left (0,1) 0 0 h2, 1, 1, left, 0, left i 7→ h ( (0, 1), ε), x = 1i, (0,1) 0 + * ∗ ∗ , (0, 1).(1, 2) ), ((0, 1).(1, 2) x = y = 1 h2, 1, 1, left, 0, righti 7→ (0,1) (0,1) x(1,2) = y(1,2) + 1 This is a very accurate description of the communication topology of the ring. In particular, we are able to distinguish between instances of recursively defined agents and channels.

6

Conclusion

We have described a parametric analysis framework for automatically inferring communication topologies of mobile systems specified in the π-calculus. We have instantiated this framework to obtain an effective analysis which is able to give finite descriptions of unbounded communication topologies that distinguish between instances of recursively defined components. To our knowledge this is the only existing analysis of mobile systems (excluding [23]) which can produce

166

A. Venet

results of that level of accuracy without any strong restriction on the base language. Previous works addressed the issue of communication analysis in csp [4, 16] or cml [20,7,8]. In the latter papers, the analysis techniques heavily rely on cml type information and cannot be applied to more general untyped languages like the π-calculus. In order to keep the presentation clear within a limited space, we had to make some simplifying assumptions that can be relaxed in many ways, for example by using more expressive abstract domains to denote relations between thread and channel identifiers, like cofibered domains [22,24], by refining the abstract semantics to take more information into account, like the number of instances of a channel or a thread, or by considering a richer version of the π-calculus with guarded choice, matching and nested replications. Finally, in view of the encodings of classical language constructs (data structures, references, control structures) in the π-calculus, it would be interesting to study the possiblity of using a static analysis of the π-calculus as a universal analysis back-end for high-level languages. Acknowledgements. I wish to thank Radhia Cousot, Patrick Cousot, Ian Mackie and the anonymous referees for useful comments on a first version of this paper.

References 1. G. Berry and G. Boudol. The chemical abstract machine. Theoretical Computer Science, 96:217–248, 1992. 2. G. Boudol. Asynchrony and the π-calculus. Technical Report 1702, INRIA, 1992. 3. P. Cousot and R. Cousot. Abstract interpretation: a unified lattice model for static analysis of programs by construction or approximation of fixpoints. In Conference Record of the Fourth ACM Symposium on Principles of Programming Languages, pages 238–252, 1977. 4. P. Cousot and R. Cousot. Semantic analysis of communicating sequential processes. In Seventh International Colloquium on Automata, Languages and Programming, volume 85 of LNCS, 1980. 5. P. Cousot and R. Cousot. Abstract interpretation frameworks. Journal of logic and computation, 2(4):511–547, August 1992. 6. P. Cousot and N. Halbwachs. Automatic discovery of linear restraints among variables of a program. In Proceedings of the Fifth Conference on Principles of Programming Languages. ACM Press, 1978. 7. C. Colby. Analyzing the communication topology of concurrent programs. In Symposium on Partial Evaluation and Program Manipulation, 1995. 8. C. Colby. Determining storage properties of sequential and concurrent programs with assignment and structured data. In Proceedings of the Second International Static Analysis Symposium, volume 983 of Lecture Notes in Computer Science, pages 64–81. Springer-Verlag, 1995. 9. P. Cousot. Semantic foundations of program analysis. In S.S. Muchnick and N.D. Jones, editors, Program Flow Analysis: Theory and Applications, chapter 10, pages 303–342. Prentice-Hall, Inc., Englewood Cliffs, 1981.

Automatic Determination of Communication Topologies

167

10. A. Deutsch. A storeless model of aliasing and its abstraction using finite representations of right-regular equivalence relations. In Proceedings of the 1992 International Conference on Computer Languages, pages 2–13. IEEE Computer Society Press, 1992. 11. P. Degano, C. Priami, L. Leth, and B. Thomsen. Analysis of facile programs: A case study. In Proceedings of the Fifth LOMAPS Workshop on Analysis and Verification of Multiple-Agent Languages, volume 1192 of Lecture Notes in Computer Science, pages 345–369. Springer-Verlag, 1996. 12. P. Granger. Static analysis of linear congruence equalities among variables of a program. In TAPSOFT’91, volume 493. Lecture Notes in Computer Science, 1991. 13. C.A.R. Hoare. Communicating Sequential Processes. Prentice Hall, 1985. 14. H.B.M Jonkers. Abstract storage structures. In De Bakker and Van Vliet, editors, Algorithmic languages, pages 321–343. IFIP, 1981. 15. M. Karr. Affine relationships among variables of a program. Acta Informatica, pages 133–151, 1976. 16. N. Mercouroff. An algorithm for analysing communicating processes. In Mathematical Foundations of Programming Semantics, volume 598 of LNCS, 1991. 17. R. Milner. Communication and Concurrency. Prentice Hall, 1989. 18. R. Milner. The polyadic π-calculus: a tutorial. In Proceedings of the International Summer School on Logic and Algebra of Specification. Springer-Verlag, 1991. 19. R. Milner, J. Parrow, and D. Walker. A calculus of mobile processes. Information and Computation, 100:1–77, 1992. 20. H. R. Nielson and F. Nielson. Higher-order concurrent programs with finite communication topology. In 21st ACM Symposium on Principles of Programming Languages, 1994. 21. D. N. Turner. The Polymorphic Pi-Calculus: Theory and Implementation. PhD thesis, Edinburgh University, 1995. 22. A. Venet. Abstract cofibered domains: Application to the alias analysis of untyped programs. In Proceedings of the Third International Static Analysis Symposium SAS’96, volume 1145 of Lecture Notes in Computer Science, pages 366–382. Springer-Verlag, 1996. 23. A. Venet. Abstract interpretation of the π-calculus. In Proceedings of the Fifth LOMAPS Workshop on Analysis and Verification of High-Level Concurrent Languages, volume 1192 of Lecture Notes in Computer Science, pages 51–75. SpringerVerlag, 1996. 24. A. Venet. Automatic analysis of pointer aliasing for untyped programs. Science of Computer Programming, 1999. To appear.

Constructing speci c SOS semantics for concurrency via abstract interpretation EXTENDED ABSTRACT

Chiara Bodei1 , Pierpaolo Degano,1 Corrado Priami2 1

Dipartimento di Informatica, Universita di Pisa Corso Italia 40, I-56100 Pisa, Italy fchiara,[email protected]

Istituto Policattedra, Universita di Verona Ca' Vignal 2, Strada Le Grazie 1, I-37134 Verona, Italy 2

[email protected]

Abstract. Most of the SOS semantics for concurrent systems can be

derived by abstracting on the inference rules of a concrete transition system, namely the proved transition system. Besides the standard interleaving semantics we mechanically derive the causal transition system for CCS , whose de nition is particularly dicult and paradigmatic. Its rules are shown to coincide with those presented in the literature. Also, the tree of its computations coincide with that obtained by abstracting the computations of the proved transition system. Keywords. Concurrency, abstract interpretation, SOS semantics, causality, non-interleaving descriptions.

1 Introduction

We apply the theory of abstract interpretation [10] to calculi for concurrency. We carry out our investigation in a pure setting and consider the basic, foundational calculus CCS [26]. However, we are con dent that more powerful calculi, e.g. those for mobile processes, like the -calculus [27], require only some technical adjustments to the machinery proposed here. The operational semantics of CCS is a transition system, de ned following the SOS approach. More in detail, a transition for a process P is deduced applying a set of inference rules, driven by the syntax of P . The original transition system describes the evolution of processes in an interleaving style, and a great deal of work has been devoted to de ne new transition systems expressing also other aspects of process evolutions, both qualitative and quantitative. Among the former, particularly relevant has been the study of transition systems that express the causality between transitions (see [7,14,24,28,6] to mention only a few of a long list of references). Another well-studied non-interleaving qualitative aspect concerns the description of the localities where processes are placed (among the other proposals, see [8,1,2,29]). Quantitative descriptions include transition systems that express temporal aspects [22,30,19,21,16], probabilistic aspects [34,25], and stochastic G. Levi (Ed.): SAS’98, LNCS 1503, pp. 168−183, 1998. Springer-Verlag Berlin Heidelberg 1998

Constructing Specific SOS Semantics for Concurrency via Abstract Interpretation

169

ones [20,23,3,9,31]. Besides its interest in se, an SOS semantics for causality is relevant because it is paradigmatic for others qualitative, non-interleaving description of concurrent systems, as well as for some of the quantitative ones. In particular, many non-interleaving SOS de nitions have rules in a form similar to those of the causal transition system of [14]. Our starting point is the proved transition system of CCS [17], PTS for short. This very concrete representation of systems has been used to describe qualitative and quantitative non-interleaving aspects of concurrent computations [32,33,4]. The transitions are labelled by encodings of their proofs, that encode also most of the aspects brie y mentioned above. The rst and the last author used abstract interpretation in [5] to extract the causal computations from the proved computations of the PTS, and sketched how other kinds of computations could be derived, as well. Our main goal here is to push further the use of abstract interpretation in order to mechanically obtain from the PTS the SOS inference rules for most of the non-interleaving transition systems mentioned above. We argue that small step operational semantics in SOS style for process calculi can be organized in a hierarchy by abstract interpretation, along the lines of abstraction of rule based speci cations presented in [11{13]. We work out in full detail the derivation of the causal transition system for CCS , possibly the most dicult non-interleaving transition system. We de ne a (family of) abstraction functions that yield CTS, a variant of the original causal transition system of [14]. As a consequence, the computations of CTS coincide with the abstraction of the computations of PTS, as de ned in [5]. The task is easier with Milner's original transition system [26], that we also derive through abstractions from both the proved and the causal transition systems. These abstractions constitute a fragment of a hierarchy of transition systems. Our approach is constructive, because the causal SOS rules are formally derived by abstraction of the proved ones and the same happens for the interleaving ones. This is a rst step towards giving evidence of the power of formal methods over empirical ones. In the conclusions, we brie y sketch the simple modi cations needed to cope with the locality transition system, that we omit here for lack of space. We also give hints on the derivation of temporal transition systems.

2 The concrete domain We recall the pure CCS [26] that we use here as a test-bed. As usual, we denote the countable set of atomic actions by , the set of co-actions by and the invisible actions with . Then A = [ [ f g. We also assume a set of agent identi ers with typical element A. Processes (denoted by P; Q; R; : : : 2 P ) are built from actions and agents according to the syntax

P ::= 0 j :P j P + P j P jP j (a)P j A where 2 A. Hereafter, we omit the trailing 0.

170

C. Bodei, P. Degano, and C. Priami

The pre x is the rst atomic action that the process :P can perform. Summation denotes nondeterministic choice. The operator j describes parallel composition of processes. The restriction operator (a) prevents P from performing a. A is the invocation of the agent identi er A, that has a unique de ning equation of the form A = P. Hereafter, we assume that (P = P ; +; 0) is a commutative monoid. As in [17], the parallel operator j is neither associative nor commutative, and this is the only dierence with the original CCS de nition, apart from the omission of the relabelling operator, which is irrelevant to our present study. The transition system of CCS is in Tab. 1. P ,! P ide : ; A = P A ,! P

act : :P ,! P

0

par0 : par1 :

0

P ,! P P jQ ,! P jQ 0

sum :

0

P ,! P QjP ,! QjP 0

res :

0

P ,! P P + Q ,! P 0

P ,! P ; 62 fa; ag (a)P ,! (a)P 0

0

a a P ,! P ; Q ,! Q com : P jQ ,! P jQ 0

0

Table 1.

0

0

0

Transition system for CCS .

As anticipated in the introduction, we enrich the labels of the standard interleaving transition system of CCS in the style of [7,15]. It is thus possible to derive dierent semantic models for CCS by extracting new kinds of labels from the enriched ones (see [17,18]). We start with the de nition of the enriched labels and of a function (`) that takes them to the corresponding action labels. De nition 1. Let # 2 fjj0; jj1g. Then proof terms (with metavariable 2 ) are de ned by the following syntax ::= # j #hjj0#00 ; jj1#11 i with 0 = a if and only if 1 = a, or vice versa. Function ` is de ned as `(#a) = a and `(#hjj0#0a; jj1#1a0 i) = . The proved transition system PTS for CCS is in Tab. 2, where the symmetric rule for communication (Com1 ) is omitted. These rules are our concrete domain.

Constructing Specific SOS Semantics for Concurrency via Abstract Interpretation

Act : :P ,! P

Ide :

P ,! P ;A = P A ,! P 0

0

Par0 :

P ,! P 0 P jQ ,! P jQ

Sum :

Par1 :

P ,! P 1 QjP ,! QjP

Res :

0

jj

0

0

jj

0

171

P ,! P P + Q ,! P 0

0

P ,! P ; `() 62 fa; ag (a)P ,! (a)P 0

0

#a #a P ,! P ; Q ,! Q Com0 : 0 #a; 1 # a P jQ ,! P jQ 0

0

hjj

Table 2.

jj

0

0

i

0

0

Proved transition system for CCS .

The proved transition system diers from the one in Tab. 1 in the rules for parallel composition and communication. Rule Par0 (Par1) adds to the label a tag jj0 (jj1) to record that the left (right) component is moving. The rule Com0 has in its conclusion a pair instead of a to record the components which interacted. The standard interleaving semantics [26] is obtained from the proved one by relabelling each transition through function ` in Def. 1. The relabelling is an abstraction from the proved transition system to the interleaving transition system semantics. This result is made precise in Subsection 6.2.

3 The abstract domain Causal CCS (CCCS) is a version of CCS introduced in [14] to give full account to causality relations between transitions. Processes are extended with a family of operators K ::, with K 2 }fin (IN). The intended meaning of K in K :: P is to encode the causes of the process P, i.e. the transition occurrences enabling P. The CCCS processes, called extended processes and denoted by t; t ; : : : 2 T , are de ned according to the syntax 0

t ::= K :: P j tjt j ( a)t: We assume that the operators K :: distribute over the restriction (a) and the parallel operator j, i.e. we assume on extended processes the least congruence that satis es the following clauses K

:: (r)t (r)K :: t

K :: (tjt ) (K :: t)j(K :: t ) 0

0

172

C. Bodei, P. Degano, and C. Priami

Also we let K :: K :: P = (K [ K ) :: P ; and we write K(t) for the set of causes in t, k for the singleton fkg, and K fK =kg for (K nk) [ K . Moreover, we let (K :: P )fK =kg = K fK =kg :: P and we homomorphically extend this replacement of a set of causes K for a single cause k onto extended processes t. Causal transitions have the form t 7,K! t , where is the standard action, ;k k 2 IN is the unique name of the transition and K represents the set of the transitions that enable k. The causal transition system in Tab. 3 is a variant of the original one of [14] (see also [24]), where the function f is1 k if 6= f(K; ) = K; ;; ; otherwise : 0

0

0

0

0

0

0

We write 2 (f(K; )) for the second component of f(K; ) and, by abuse of notation, 2 (f(K; )) 2= K, assuming it true if 2 (f(K; )) = ;. As usual, we omit in Tab. 3 the rule Kcom1 , symmetric to Kcom0 . ! Kact : K :: :P f 7, (f (K;)) [ K :: P; 2 (f (K;)) 62 K (K;) 2

Kide :

! t f 7, t (K;)

! t f 7,K; t 00

(

)

! t f 7, t (K;)

0

0

; t =t

Kres :

00

0

! (a)t f (7,K; (a)t )

! t f 7, t (K;)

0

Kpar0 :

(

)

0

00

; 2 (f (K;)) 62 K(t ) Ksum :

! t f 7, t (K;)

0

Kpar1 :

! t + t f 7, t (K;) 00

0

t0 K7,a0!;k t0 ; t1 K7,a1!;k t1 0

! t jt f 7, t jt (K;) 00

! t f 7, t (K;)

0

00

! tjt f 7,K; t jt 00

; 62 fa; ag

0

; 2 (f (K;)) 62 K(t ) Kcom0 : 0

0

t0 jt1 7,! t0 fK1 =kgjt1 fK0 =kg ; Table 3. Causal transition system for CCCS . 00

0

0

0

; ;

4 Rule-based inductive de nitions We rst recall some de nitions and results from [11]. Then, we instantiate them in our setting. 1

This function is introduced to avoid duplicating the rules: one set to handle transitions and the other for the visible transitions, as done in [24].

Constructing Specific SOS Semantics for Concurrency via Abstract Interpretation

173

De nition 2. An inductive de nition is a quadruple F = hU; ; ?; ti such that: { the universe U is a set; { is the set of rule instances Pc , with P U and c 2 U ; { ? 2 U is the basis; { t 2 }(}(U )) ! }(U ), is the (partial) join operator and the induced ordering v is x v y i x t y = y and is a partial order on }(U ). De nition 3. The operator induced by F = hU; ; ?; ti is : }(U ) ! }(U ) such that

P 2 g c The inductive de nition F is monotonic whenever is monotonic and ? v (?). Moreover F is on a cpo (resp. on a complete lattice) i h}(U ); v; ?; ti (resp. h}(U ); v; ?; >; t; ui ) is a cpo (resp. a complete lattice). (X ) = fc 2 U j9P X :

De nition 4. Given F = hU; ; ?; ti and F = hU ; ; ? ; t i, we write F ()

F for h}(U ); v; ?; ti ()

h}(U ); v ; ? ; t i, whenever h; i is a Galois 0

0

0

0

0

0

0

0

0

0

connection, i.e. if and only if { h}(U ); vi and h}(U ); v i are posets; { 2 }(U ) ! }(U ) and 2 }(U ) ! }(U ); { 8x 2 }(U ); 8y 2 }(U ) : [(x) v y] if and only if [x v (y)]. 0

0

0

0

0

0

The following proposition (corresponding to Prop.53 in [11]), will be useful later. Proposition 5. Let F = hU; ; ?; ti be such that 1. F is a monotonic inductive de nition on the cpo h}(U ); v; ?; ti, 2. h}(U ); v ; t i be a partial order, 3. 2 }(U ) ! }(U ) be a complete t- and [-morphism such that 8 Pc 2 ; 8X U : (P ) (X ) implies 9 Pc 2 : P X and (fcg) (fc g): 0

0

0

0

0

0

0

0

De ne = f (cP ) j Pc 2 ; with c 2 (c)g and ? = (?). Then F = hU ; ; ? ; t i is a well formed inductive de nition (i.e. a monotonic or extensive de nition on a cpo) such that : F ! F . If moreover h}(U ); v; ?; >; t; ui is a complete lattice then F () F , where

2 }(U ) ! }(U ) is (Y ) = tfX 2 U j (X ) v Y g. 0

0

0

0

0

0

0

0

0

0

0

0

More precisely, under the conditions of the above proposition we have the following (for the notation used and more technical details see Def. 46 and Proposition 53 in [11]). The abstraction function preserves least xed points lfp , [=] because F ()

F implies that (lfp ( )) = lfp ( ). In the following we use a special form of inductive de nition only, namely the positive one. These de nitions are well-formed, hence they have always the xed point (;). 0

v

1 [

v0

0

174

C. Bodei, P. Degano, and C. Priami

De nition 6. The inductive de nition FP = hUP ; P ; ?; ti is positive, and is written as hU; i, whenever ? is the emptyset ; and t is the set union [. The set of all the possible transitions involving CCS processes can be generated by a rule-based inductive de nition. In this way, instead of reasoning about the classical schemata of rules and axioms, we consider a set of rule instances, where axioms have empty premises. Some of the rules in Tables 2 and 3 have side-conditions, that can also be seen as additional premises. In fact, each rule with a side-condition stands for a set of rules, therefore we can safely use a rule-based inductive de nition.

4.1 The concrete domain FP

Here we give a rule-based inductive de nition of the proved transitions of CCS. De nition 7. The inductive de nition FP = hUP ; P i is the positive inductive de nition, where: { UP = fPP ,! P jP; P 2 P ; 2 g is the universe, and { P = f c j P UP ; c 2 UP and Pc is an instance of Act; Ide; Res; Sum; Par0; P ar1; Com0; Com1 g is the set of rule instances. The following properties of FP are easy to establish. 0

0

Proposition 8. The positive inductive de nition FP is monotonic and it is well-formed. Proof. We only need to prove that the operator P is monotonic and that

; P (;). Let X Y UP . Since c 2 P (X) implies 9P X : Pc 2 P and X Y , then c 2 P (Y ). Finally, ; P (;) holds because P (;) is a set. Proposition 9. h}(UP ); ; ;; UP ; [; \i is a complete lattice.

4.2 The abstract domain FC

The set of all the possible causal transitions involving CCCS processes can be generated by a rule-based positive inductive de nition, as well. De nition 10. FC = hUC ; C i is the positive inductive de nition, where: ! t j t; t 2 T ; 2 Act; K 2 }fin (IN)g is the universe; { UC = ft f 7,K; { C = f Pc j P UC ; c 2 UC and Pc is an instance ofKact; Kide; Kres; Ksum; Kpar ; Kpar ; Kcom ; Kcom g is the set of rule instances. Just as for FP , we have the following. Proposition 11. The positive inductive de nition FC is monotonic and well0

(

0

formed.

0

)

1

0

1

Proposition 12. h}(UC ); ; ;; UC ; [; \i is a complete lattice.

Constructing Specific SOS Semantics for Concurrency via Abstract Interpretation

175

5 The abstract interpretation function

To abstract inductive de nitions, we use an abstraction of subsets of the universe, according to Proposition 5. Our abstraction is based on a family of abstraction functions, which act on the components of the rule schemata and which are parameterized on a class S de ned according to the syntax S ::= K j (S jS ):

We write S i for the ith (from the left) set K 2 }fin (IN ) of (the frontier of the tree) S . For instance if S = (K j(K jK ))j(H jH ), then S 4 = H . The family of functions S is such that 0

K (P ) = K :: P

00

0

S0 S1 (P j Q) = S0 (P ) j S1 (Q): j

Moreover we assume that { K :: H (P ) = K H (P ); { K :: S0 S1 (P j Q) = K :: S0 (P ) j K :: S1 (Q); { (S0 j S1 )fH=kg = S0 fH=kg j S1 fH=kg. As an example of how an abstraction function works, consider the S introduced above and the process R = (P j(P jP ))j(a:QjQ ). Then [

j

0

00

0

S (R) = (K (K K )) (P j(P jP )) j (H H ) (a:QjQ ) = K (P ) j (K (P ) j K (P )) j (H (a:Q) j H (Q )) = (K :: P j (K :: P j K :: P )) j (H :: a:Q j H :: Q ) = tR . j

0j

0

00

0

0

0

00

0

00

j

0

0

00

00

0

0

00

0

0

Now, we can de ne the abstraction function pc that maps proved transitions in causal transitions. It will be used to obtain the causal transition system by abstracting the rules of the proved one. De nition 13. Let pc : }(UP ) ! }(UC ) be de ned as # pc(X ) = fpc(P ,! P ) = S (P ) ,! S (P ) j P ,! P 2 Xg [ ; a ;# a fpc(P (#a;# ,!a ) P ) = S0 (P ) ,! S1 (P ) j P # # ,! P 2 X; ; j j p p i i 9i; j #a : S1 = S1 = S0 [ S0 and 8p 6= i; j : S1 = S0 g [ a #a p fpc(P ,! P ) = S0 (P ) ,! K ;k S1 (P ) j P ,! P 2 X; 8p : k 62 S0 and 9i; K : S0i = K; S1i = K [ fkg; 8j 6= i : S1j = S0j g To see our de nition at work, consider again the S and R introduced above, 1 0a and the following proved transition R ,! R = (P j(P jP ))j(QjQ ): Then, its a abstraction is tR 7,H! t , for h 62 K [ K [ K [ H [ H and ;h R S = (K j(K jK ))j(H [ fhgjH ), where 0

0

0

; ;

0

0

0

h

0

0

0

00

00

i

0

; ;

0

0

0

jj jj

0

0

0

0

00

0

0

00

0

0

00

0

176

C. Bodei, P. Degano, and C. Priami

tR = S (R ) = (K :: P j (K :: P j K :: P )) j (H [ fhg :: Q j H :: Q ). 0

0

0

0

00

00

0

0

It is easy to deduce this causal transition with the rules in Tab. 3. The interested reader may wish to abstract step by step its whole proved derivation. We wish to ! (f (K;)) :: K (P ); 2 (f (K;)) 62 K r (Act) = K (:P ) f 7, (K;) 2

r (Ide) =

) 7,l(! S (P ) f (K;` (f (K;`())) :: S (P ) ( )) 2 ) 7,`(! S (A) f (K;` (P ) ( )) S ) 7,`(! S (P ) f (K;` (P ) ( )) S 0

0

,7 ` ! S (a)P ) f K;` S ((a)P ) ( )

(

r (Sum) =

) 7,`(! S (P ) f (K;` (P ) ( )) S 0

0

) 7,`(! S (P + Q) f (K;` (P ) ( )) S 0

0

r (Par0 ) =

) 7,`(! S0 (P ) f (K;` (P ) ( )) S0 0

0

S0 S1 (P j Q 7,! S0 S1 (P j Q) `( ) ) f (K;`( ))

j

r (Par1 ) =

; `() 62 fa; ag

0

0

( ))

; S (A) = S (P )

0

0

r (Res) =

0

0

0

0j

) 7,`(! S1 (Q) f (K;` (Q ) ( )) S1 0

0

S0 S1 (P j Q 7,! S0 S1 (P j Q ) `( ) ) f (K;`( ))

j

r (Com0) =

; 8i : 2 (f (K;`())) 62 S1i

j

; 8i : 2 (f (K;`())) 62 S0i

0

0

S0 (P0 ) f (7,Ka! (P0 ); S1 (P1) f (7,Ka! (P1 ) 0 ;a) S0 0 ;a) S1 0

0

S0 S1 (P0 j P1) 7,! S0 ; j

Table 4.

; ;

0

K1 =k g

0f

j

P j P1) 0

S10 fK0 =k g ( 0

The abstraction interpretation

0

0

r of rule instances.

use Proposition 5 to establish a Galois connection between the de nitions of the proved and the causal transition systems. We already know that the abstract domain is a complete lattice (Proposition 12). We are left to show that our abstraction function pc enjoys the properties required by item 3 of Proposition 5. The function pc a complete [-morphism, i.e. 8X UP : [X 2 UP implies [ pc (X ) = pc ([X ):

Proposition 14.

Constructing Specific SOS Semantics for Concurrency via Abstract Interpretation

177

Proof. It is easy to prove that the function pc is monotone. We now prove the

double inclusion and (note that induced ordering is the same for both domains). We start with . By de nition of [, 8X UP ; pc (X ) [pc(X ) and X [X . Then, by monotonicity of pc , it is pc (X ) pc ([X ): Since [ is the least upper bound, [ pc (X ) pc([X ): We now prove . Let F K be the function introduced in Def. 25 that deletes the causes from an extended process, restoring a pure CCS process. i By hypothesis [pc (X ) = fti K7,i! t j i 2 I g [ ftj 7,! tj j j 2 J g UC . Then, ;ki i ; 0

0

0

; ;

p by construction, there exists a set Y = fPp ,! Pp j F K (tp ) = Pp ; F K (tp ) = p Pp; `(p ) = p j p 2 I [ J g, such that c (Y ) = [pc(X ). First we prove 8Y~ X : Y~ Y . Per absurdum, let S X such that S Y . Then, by monotonicity of pc , pc (S ) pc (Y ) = [pc (X ), against the hypothesis of [pc(X ) to be the least upper bound. Then, 8Y~ X : Y~ [X , as well. From this and from Y~ Y , [X Y , because [X is the least upper bound. By monotonicity of pc , we conclude pc ([X ) pc (Y ) = [pc (X ). 0

0

0

Proposition 15. The function pc is such that 8 Pc 2 P ; 8X UP : pc(P ) pc (X ) implies 9 Pc 2 P : P X and pc (fcg) pc (fc g): Proof. Since pc (P ) = 6 pc (P ), whenever P 6= P , pc (P ) pc (X ) implies P X . Hence, we only need to put P = P . 0

0

0

0

0

0

0

To de ne the abstract interpretation of the operator , wep introduce in Tab. 4 the auxiliary family of functions r . Note that r ( Pc ) = cc(P ) with c 2 pc (c) (cf. the de nition of in Proposition 5). We can now establish the required Galois connection, that preserves least xed points. 0

0

0

Theorem 16. Let = fr ( Pc ) j Pc 2 P g and ; = r (;). The positive inducc p tive de nition F = fpc (U ); g is well formed and such that F () F , i.e.

cp hpc ; pc i is a Galois connection, where pc : }(UC ) ! }(UP ) is X 2 UP j pc (X ) Y g if Y 2 pc (UP ) c

p (Y ) = [f UP if Y 2= pc (UP ) 0

0

0

[=]

P

0

Proof. By Proposition 5, whose hypotheses are satis ed because of Proposi-

tions 8, 9, 11, 12, 14, 15.

Although pc (UP ) UC , we have established exactly what we would like to obtain. In fact, it is immediate to prove the following property, because the rules in Tables 3 and 4 can be easily put in bijection.

Proposition 17. r (P ) = C .

178

C. Bodei, P. Degano, and C. Priami

The above correspondence suces to establish that the computations of the causal transition system obtained via abstraction pc coincide with the abstraction of the proved computations as de ned in [5]. Actually, a stronger result holds: the trees of computations coincide, so the non deterministic aspects of computations are preserved, as well. Lack of space prevents us from recalling the relevant de nitions and results of [5]. We simply introduce some notation. Let P tree(P ) be the tree de ned as the set of the proved computations starting from P , ordered by pre x. Similarly, the tree of causal computations is Ktree(P ), with transitions derived by the rules in Tab. 3; and nally let pcP tree(P ) be the tree of computations with transitions deduced according to pc(P ). We denote by h; i the Galois connection of Def. 17 in [5], where maps the tree of proved computations P tree(P ) to the tree of causal computations Ktree(P ). Now, we can state the following. Theorem 18. For all P 2 P , we have that pcP tree(P ) = Ktree(P ). Proof. The bijection between Tab. 3 and Tab. 4 suces, because (P tree(P )) = Ktree(P ) by Theorem 18 of [5].

6 A fragment of a hierarchy

pc pc

C

pi

P ci

pi

ci Fig. 1.

I

The hierarchy of semantics.

We now de ne a simple hierarchy of transition systems involving the proved, the causal and the standard interleaving semantics of CCS . The hierarchy in Fig. 1 is established by studying three abstraction functions: from the proved to the causal model and to the interleaving model, and from the causal to the interleaving one. This fragment of a hierarchy is formally constructed, and in the conclusions we sketch how other models can be derived by further constructive abstractions of the proved model.

6.1 The abstract domain FI

We start with the de nition of the abstract domain made of the standard interleaving transition system of CCS [26]. The set of all the possible transitions involving CCS processes can be generated by the following positive rule-based inductive de nition.

Constructing Specific SOS Semantics for Concurrency via Abstract Interpretation

179

De nition 19. FI = hUI ; I i is the positive inductive de nition, where: { UI = fP ,! P j P; P 2 P ; 2 Actg is the universe; P { I = f c j P UI ; c 2 UI and Pc is an instance of act; ide; res; sum; par ; par ; com ; com g is the set of rule instances. Just as for FC , we have the following. Proposition 20. The positive inductive de nition FI is monotonic and well0

0

0

1

0

1

formed.

Proposition 21. h}(UI ); ; ;; UI ; [; \i is a complete lattice. 6.2 The abstraction from proved to interleaving transition systems We make now precise the observation made at the end of Section 2, by de ning a Galois connection between FP and FI . So one can easily recover the standard interleaving transition system from the proved one.

De nition 22. Let pi : }(UP ) ! }(UI ) be de ned as ` pi(X) = fpi(P ,! P ) = P ,! P j P ,! P 2 X g; where ` is as in Def. 1. ( )

0

0

0

Also in this case is possible to reformulate the results relative to the Galois p connection pand the consequent correspondences. Note again that r ( Pc ) = ic(P ) with c 2 i (c). 00

0

0

Theorem 23. Let =p fr ( Pc ) j Pc 2 P g and ; = r (;). The positive inductive de nition F = fi (UP ); g is well formed and is such that p i FP () F I , i.e. hpi ; pi i is a Galois connection, where pi : }(UI ) ! }(UP ) is

pi 00

00

[=]

00

00

00

0

X 2 UP j pi (X) Y g if Y 2 pip (UP )

pi (Y ) = [f UP if Y 2= i (UP )

The following proposition says that the induced operator above is indeed I , and justi es the use of ,! in Def. 22. 00

Proposition 24. r (P ) = I . 00

6.3 The abstraction from causal to interleaving transition systems Our next abstraction is the function ci that maps causal transitions in standard interleaving transitions. It will be used to obtain the interleaving transition system by abstracting the rules of the causal one. The steps are the same as above.

180

C. Bodei, P. Degano, and C. Priami

De nition 25. Let ci : }(UC ) ! }(UI ) be de ned as ci(X ) = fci(t f ,! K; t ) = F K (t) ,! F K (t ) j t f ,! K; t 2 X g: (

)

0

0

(

)

0

where F K deletes the causes from extended processes, de ned as F K (K :: t) = F K (t), F K (tjt ) = F K (t)jF K (t ), F K ((a)t) = (a)F K (t). 0

0

Theorem 26. Let = fr ( Pc ) j Pc 2 C g and ; = r (;). The positive inductive de nition F = fci (UC ); g is well formed and is such that ci FC () F I , i.e. hci ; ci i is a Galois connection, where ci : }(UI ) ! }(UC ) is

ci 000

000

000

[=]

000

000

000

ci (Y ) =

[fX 2 UC j ci (X ) Y g if Y 2 ci (UC ) UC if Y 2= ci (UC )

Note that F K (t) is a standard CCS process and the following proposition justi es the presence of ,! in Def. 25.

Proposition 27. r (C ) = I . 000

6.4 The hierarchy The commutativity of the diagram in Fig. 1 is now immediate.

Theorem 28. ci(pc(P )) = pi(P ) = I . The above theorem suces to show a correspondence between the trees of computations in the various models. The trees of computations obtained by abstracting proved computations to interleaving ones coincide with the trees of the computations based on the standard interleaving semantics. The same happens abstracting causal computations. Therefore, the non deterministic aspects of agent behaviour are fully respected. This supports our claim that formal constructive methods can be used to mechanically derive, clarify and compare various dierent models for concurrency. Let Stree(P ) be the synchronization tree de ned as the set of the interleaving CCS computations starting from P , ordered by pre x [26]. Also, let pi P tree(P ) and ci Ktree(P ) be the trees of computations with transitions deduced with p i (P ) and with ci (C ). The following is immediate.

Theorem 29. For all P 2 P , piP tree(P ) = Stree(P ) = ciKtree(P ). 7 Conclusions

We have constructed the causal and the interleaving transition systems for CCS by abstracting on the rules of the proved transition system. Lack of space prevents us from giving more evidence that quite similar constructions are possible, and give most of the interleaving and the non-interleaving SOS de nitions

Constructing Specific SOS Semantics for Concurrency via Abstract Interpretation

181

presented in the literature for concurrency calculi. Their designers aimed at describing dierent qualitative and quantitative aspects of process evolutions, and a considerable eort has been spent in their actual SOS de nitions. Our proposal may help in mechanically deriving these SOS de nitions, keeping in mind only the eect that a single transition has on the state of the whole system of concurrent processes. Consider for instance the eect that a visible action a has on the S in Def. 13: only the causes K of the process performing a are aected | this obvious fact, together with exchange of causes in communications, are sucient to obtain the rules for the causal transition system. Despite space limitation, we would like to give some hints of three more abstraction functions, thus showing that our construction is in a sense paradigmatic. Consider the location transition system of [8]. It was designed to keep track of the sites (taken from a given set Location) where the visible actions actually occur, and it assumes that the sites of invisible ones cannot be detected. Having this in mind, it is easy to de ne the location abstraction functions pl , mimicking Def. 13, as follows. pl(X) = fpl(P ,! P ) = L (P) ,! L (P ) j P ,! P 2 X; `() = g [ a #a #a fpl(P ,! P ) = L (P) ,! lu L (P ) j P ,! P 2 X; p i 8p : l 62 L0 and 9i : L0 = u; Li1 = lu; and 8j 6= k : Lj1 = Lj0 g; where l 2 Location; u 2 Location and L is either u or LjL (i.e. L is a tree of strings of locations, rather than a tree of sets of natural numbers), and L is a family of abstraction functions de ned just as S . Indeed, the only dierence between Def. 13 and the one sketched above is that communications do not exchange causes, a typical feature of causality, not shared by location models. It is easy then to establish a Galois connection hpl ; pl i and results similar to Theorems 16 and 18. At this point, the reader familiar with the local/global transition system of [24] can easily obtain it by de ning an abstraction function pl=g that combines pc and pl and that enjoys all the properties these abstraction functions have. As a third example, we consider a SOS semantics that expresses quantitative aspects of concurrent computations. Consider the timed transition system of [19], whose transitions have labels of the form ha; ni ? w, where n is the time required to complete the standard action pa at location w. We can easily modify pl to obtain the abstraction function t that renders the considered timed transition systems. Assume as given a function time(a) that assigns a duration to actions (we only consider visible actions, because the calculus of [19] has no 's, and we slightly change their notion of location). This abstraction function is as follows. #a #a ptp(X) = fpt(P ,! P 0) = L (P ) ha;n ,!i?w jL (Pj0) j P ,! P 0 2 X; 8p : l 62 i i L0 and 9i : L0 = u; L1 = lu; and 8j 6= k : L1 = L0 ; n = time(a)g: We have not yet considered transition systems expressing other qualitative aspects of concurrent processes. However, we are very con dent that most of them 0

0

0

0

0

0

0

0

182

C. Bodei, P. Degano, and C. Priami

can be cast in our framework, and that abstract interpretation is a exible tool for constructing speci c SOS semantics for concurrency calculi, including those for mobility, like the -calculus [27].

Acknowledgment. The authors wish to thank the referees for their precise remarks and helpful comments.

References 1. L. Aceto. A static view of localities. Formal Aspects of Computing, 1992. 2. R.M. Amadio and S. Prasad. Localities and failures. In Proceedings of FSTTCS'94, pages 205{216. Springer-Verlag, 1994. 3. M. Bernardo, L. Donatiello, and R. Gorrieri. Integrating performance and functional analysis of concurrent systems with EMPA. Technical Report UBLCS-95-14, University of Bologna, Laboratory for Computer Science, 1995. 4. C. Bodei, P. Degano, and C. Priami. Mobile processes with a distributed environment. In Proceedings of ICALP'96, LNCS 1099, pages 490{501. Springer-Verlag, 1996. To appear in TCS. 5. C. Bodei and C. Priami. True concurrency via abstract interpretation. In Proceedings of SAS'97, LNCS 1302, pages 202{216, 1997. 6. M. Boreale and D. Sangiorgi. A fully abstract semantics of causality in the calculus. In Proceedings of STACS'95, LNCS. Springer Verlag, 1995. 7. G. Boudol and I. Castellani. A non-interleaving semantics for CCS based on proved transitions. Fundamenta Informaticae, XI(4):433{452, 1988. 8. G. Boudol, I. Castellani, M. Hennessy, and A. Kiehn. A theory of processes with localities. Theoretical Computer Science, 114, 1993. 9. P. Buchholz. On a markovian process algebra. Technical report, Informatik IV, University of Dortmund, 1994. 10. P. Cousot and R. Cousot. Abstract interpretation frameworks. Journal of Logic and Computation, 2(4):511{547, 1992. 11. P. Cousot and R. Cousot. Inductive de nitions, semantics and abstract interpretation. In Proceedings of POPL'92, pages 83{94, 1992. 12. P. Cousot and R. Cousot. Higher-order abstract interpretation (and application to comportment analysis generalizing strictness, termination, projection and PER analysis of functional languages). In Procs. ICCL'94, IEEE, pages 95{112, 1994. 13. P. Cousot and R. Cousot. Compositional and inductive semantic de nitions in xpoint equational, constraint, closure-condition, rule-based and game-theoretic form. In Proceedings of CAV'95, LNCS 939, pages 293{308, 1995. 14. Ph. Darondeau and P. Degano. Causal trees. In Proceedings of ICALP'89, LNCS 372, pages 234{248. Springer-Verlag, 1989. 15. P. Degano, R. De Nicola, and U. Montanari. Partial ordering derivations for CCS. In Proceedings of FCT, LNCS 199, pages 520{533. Springer-Verlag, 1985. 16. P. Degano, J.-V. Loddo, and C. Priami. Mobile processes with local clocks. In Proceedings of Workshop on Analysis and Veri cation of Multiple-Agent Languages, Stockholm, Sweden, 1996. 17. P. Degano and C. Priami. Proved trees. In Proceedings of ICALP'92, LNCS 623, pages 629{640. Springer-Verlag, 1992. 18. P. Degano and C. Priami. Non interleaving semantics for mobile processes. Extended abstract. In Proceedings of ICALP'95, LNCS 944, pages 660{671. SpringerVerlag, 1995. To appear in TCS.

Constructing Specific SOS Semantics for Concurrency via Abstract Interpretation

183

19. R. Gorrieri, M. Roccetti, and E. Stancapiano. A theory of processes with durational actions. Theoretical Computer Science, (140), 1994. 20. N. Gotz, U. Herzog, and M. Rettelbach. TIPP- a language for timed processes and performance evaluation. Technical Report 4/92, IMMD VII, University of Erlangen-Nurnberg, 1992. 21. E. Goubault. Durations for truly concurrent transitions. In Proceedings of ESOP96, LNCS 1058, pages 173{187, 1995. 22. M. Hennessy and T. Regan. A temporal process algebra. Technical Report 2/90, University of Sussex, 1990. 23. J. Hillston. The nature of synchronization. In U. Herzog and M. Rettelbach, editors, Proceedings of PAPM'94, University of Erlangen, 1994. 24. A. Kiehn. Comparing causality and locality based equivalences. Acta Informatica, 31(8):697{718, 1994. 25. K.G. Larsen and A. Skou. Compositional veri cation of probabilistic processes. In Proceedings of CONCUR'92, volume 630 of LNCS. Springer-Verlag, 1992. 26. R. Milner. Communication and Concurrency. Prentice-Hall, London, 1989. 27. R. Milner, J. Parrow, and D. Walker. A calculus of mobile processes (I and II). Information and Computation, 100(1):1{77, 1992. 28. U. Montanari and M. Pistore. Concurrent semantics for the -calculus. In Electronic Notes in Computer Science, number 1. Elsevier, 1995. 29. U. Montanari and D. Yankelevich. A parametric approach to localities. In Proceedings of ICALP'92, LNCS 623, pages 617{628. Springer-Verlag, 1992. 30. X. Nicollin and J. Sifakis. An overview and synthesis on timed process algebras. In Real Time: Theory in Practice, LNCS 600, pages 526{548. Springer-Verlag, 1991. 31. C. Priami. Stochastic -calculus. The Computer Journal, 38(6):578{589, 1995. 32. C. Priami. Enhanced Operational Semantics for Concurrency. PhD thesis, Dipartimento di Informatica, Universita di Pisa, March 1996. Available as Tech. Rep. TD-08/96. 33. C. Priami. Interleaving-based partial odering semantics. In Proceedings of Italian Conference on Theoretical Computer Science, pages 264{278, Ravello, November 1995, 1996. World Scienti c. 34. R.J. van Glabbeek, S.A. Smolka, B. Steen, and C.M.N. Tofts. Reactive, generative and strati ed models of probabilistic processes. Information and Computation, 1995.

A First-Order Language for Expressing Aliasing and Type Properties of Logic Programs Paolo Volpe Dipartimento di Informatica Universit` a di Pisa Corso Italia 40, 56125 Pisa, Italy e-mail: [email protected] tel: +39-50-887248. fax: +39-50-887226.

Abstract. In this paper we study a first-order language that allows to express and prove properties reagarding the sharing of variables between non-ground terms and their types. The class of true formulas is proven to be decidable through a procedure of elimination of quantifiers and the language, with its proof procedure, is shown to have interesting applications in validation and debugging of logic programs. An interesting parallel is pointed out between the language of aliasing properties and the first order theories of Boolean algebras. Keywords: Verification of logic programs, languages of specification, first-order logic.

1

Introduction

In many approaches to the verification of properties of logic programs, a formal language is required that allows to express the properties of programs one is interested in. In the methods proposed in [13][1][12][16], an assertional language is assumed to verify properties of arguments of predicates of the programs. Some verification conditions are provided that imply the partial correctness of the programs with respect to various aspects of the computations. For example the method proposed in [12] and [6] allows to prove properties of the correct answers of the programs, while in [16] a method is provided to prove properties of the computed answers. The methods proposed in [13] and [1] allow to prove, in addition, that predicates verify given specifications at call time. In this paper we study a language that allows to express an interesting class of properties of non-ground terms, that is the data on which logic program operate. The language is sufficiently expressive to capture sharing, freeness and types of non-ground terms. Two or more terms are said to share, when they have at least one variable in common, while a term is free when it is a simple variable. For types, we will refer to term properties like being a list, a tree, a list of ground terms, etc. Fragments of this language have already been studied in [18] and [2] and shown to be decidable, but in this paper we show that the full first-order theory is decidable. This allows to use its full expressive power in existing proof methods G. Levi (Ed.): SAS’98, LNCS 1503, pp. 184−199, 1998. Springer-Verlag Berlin Heidelberg 1998

A First-Order Language for Expressing Aliasing and Type Properties

185

and algorithmically decide whether the verification conditions are true or not. Indeed for many methods the verification conditions are expressed by formulas of the language. If the verification condition is true, we show the partial correctness of logic programs with respect to a property belonging to the class of aliasing properties and type assertions. If the verification condition is false (and we stress that this can be checked in finite time), obviously this does not mean that the program has necessarily an error. Anyway a “warning” can be raisen up, signalling a possible wrong situation. The proof procedure shown in the paper can be easily enriched so as to provide a counterexample in this case. This allows the user to have more information about the warning and to decide whether to change the program (the counterexample is actually a wrong computation of the program) or to refine the specification (the verification condition is false because the specification is too “loose” and impossible computations are considered). We want to emphasize the independent importance of the proof of decidability of the language. It is based on the method of elimination of quantifiers and points out an interesting set of formulas, which can be viewed as expressing constraints on the cardinality of the sets of variables that can occur in terms. Our proof is based on the parallel between the satisfiability of formulas of our language and the satisfiability of such cardinality constraints, which can be proven decidable as a consequence of the decidability of the theory of Boolean algebras. We think that such class of constraints, which has a quite simple representation and operations of composition and cylindrification, can be of interest in program analysis. For example, well known abstract domains such as POS [9] and Sharing[14] can be naturally viewed as subdomain of the class of cardinality constraints (see also [19]), with their composition and the cylindrification operator obtained as instances of the general ones. The paper is organized as follows. In Section (2) we lay down the basic terminology. In Section (3) we define the class of types from which type assertions are built. In Section (4) we define formally the language, parametrically with respect to a family of regular types, and then we prove in Section (5) that such a language is decidable. Finally in Section (6) we give examples of application in the context of inductive proof methods.

2

Preliminaries

Throughout the paper, we assume familiarity with standard notions of logic programming and mathematical logic [17, 15]. A first order language L = hΣ, Π, Vi is based on a set Σ of function symbols, a set Π of predicate symbols of assigned arities and a set V of variables. The set Terms(Σ, V) of all terms of L with variables in V is defined as usual. The set Terms(Σ, ∅) ⊆ Terms(Σ, V) is the subset of ground terms (i.e. not containing variables). We write f , g for function symbols, p, q for predicate symbols, X, Y for variables, X for tuples of distinct variables, t, s for terms, t,s for tuples of terms. A (predicate or function) symbol f with arity n will be denoted by f (n) . Atomic formulas (also called atoms) are formulas like p(t1 , . . . , tn ), with

186

P. Volpe

p(n) ∈ Π and t1 , . . . , tn ∈ Terms(Σ, V). The set of formulas of L is the smallest set F containing the atomic formulas and such that if φ and ψ are in F then ¬φ, φ ∧ ψ and ∃Xφ, with X ∈ V, are in F. We will use the notation φ ∨ ψ as a shorthand for ¬(¬φ ∧ ¬ψ), φ ⇒ ψ for ¬φ ∨ ψ, φ ⇔ ψ for (φ ⇒ ψ) ∧ (ψ ⇒ φ) and ∀Xφ for ¬(∃X¬φ). We assume also the constants true and false to be in F. Given a syntactic object O of L, Vars(O) denotes the set of free variables (not bound by any quantifier) of V in O. Substitutions are defined as mapping θ : V → Terms(Σ, V), which differ from the identity just on a finite subset of V. They can be extended homomorphically to functions on terms. A preorder can be introduced on Terms(Σ, V). Given t, s ∈ Terms(Σ, V), t ≤ s iff there exists substitution η such that tη = s. The induced equivalence on terms is called variance. An interpretation I = hD, Λ, Γ i of L consists of a non-empty set D, the domain; a set of functions Λf : Dn → D for each function symbol f (n) ∈ Σ; a family of subsets Γp ⊆ Dn for each predicate symbol p(n) ∈ Π. A variable assignment (also called state) σ : V → D maps each variable into an element of D. It can be lifted homomorphically to a function, still denoted by σ, which maps terms in Terms(Σ, V) to elements of D. An atom p(t1 , . . . , tn ) is true in I under the state σ, written I |=σ p(t1 , . . . , tn ), iff (σ(t1 ), . . . , σ(tn )) ∈ Γp . The truth of each formula of L under the state σ, written I |=σ ϕ, is defined, as usual, by induction on the structure of ϕ. A formula ϕ is true in I, that is I |= ϕ, if and only if for each state σ : V → D, I |=σ ϕ. The set Th L (I) is defined as the set of all formulas of L true in I, that is as the set {ϕ | ϕ formula of L and I |= ϕ}. In this paper we will be mainly interested in the non ground term interpretations H = hTerms(Σ, V), Λ, Γ i, with the set Terms(Σ, V) as domain and with Λf (t1 , . . . , tn ) = f (t1 , . . . , tn ) for each f (n) ∈ Σ and tuple t1 , . . . , tn of Terms(Σ, V). In practice every (ground) term is interpreted by itself. We have chosen not to distinguish between the variables in formulas and the variables in the model. Indeed their roles are quite different and in practice no ambiguity arises. The preorder on Terms(Σ, V) induces a preorder on the states σ : V → Terms(Σ, V) . The induced equivalence between states is still called variance.

3

Regular term grammars

To specify families of types we will consider regular term grammars. There is a large amount of papers on regular types. They have proved them to be a good trade-off between expressibility and decidability. In fact they are strictly more expressive than regular languages, but strictly contained in context-free languages (which have an undecidable subset relation). Our main references are the papers of Dart and Zobel [10, 11] and Boye and Maluszynski [3, 2]. A regular term grammar is a tuple G = (Σ, V, T, R), where Σ is a set of function symbols, V is an infinite denumerable set of variables, T is a finite set of type symbols, including var and any, and R is a finite set of rules l → r where – l ∈ (T\{var , any})

A First-Order Language for Expressing Aliasing and Type Properties

187

– r ∈ Terms(Σ, T) For every T ∈ T\{var , any}, we define Def G (T ) (also denoted by Def R (T )) as the set {r | T → r ∈ R}. Def G (var ) is defined as the set of variable V, while Def G (any) as the set Terms(Σ, V). We use the notation T1 →G T2 if T2 is obtained from T1 by replacing a symbol T ∈ T by a term in Def G (T ). Let → →G be the transitive and reflexive closure of →G . Given the type symbol T ∈ T, we define the set of terms [T ]G , the type T , as the set {s ∈ Terms(Σ, V) | T → →G s}. Notice that [var ]G = V and [any]G = Terms(Σ, V). We assume function symbols in Σ to contain at least a constant and a function of arity 2. We will often omit the subscript when the grammar is clear from the context. Example 1. Let us see some examples (taken in part from [2]) of regular types (k ) (k ) and of the grammars that define them. Let us suppose that Σ = {f1 1 , . . . , fn n }, (ki ) where each function symbol fi has arity ki . The set of ground and instantiated terms can be defined as: (0)

inst → f1 .. .

(0)

(0)

inst → fi−1 (k ) inst → fi i (any, . . . , any) .. .

ground → f1 .. .

ground → fi−1 (k ) ground → fi i (ground , . . . , ground ) .. . (kn )

ground → fn

(ground , . . . , ground )

(0)

(kn )

inst → fn

(any, . . . , any)

The type of lists, the lists of instantiated terms, the ground lists and the list of variables can be defined as : list → [ ] ilist → [ ] glist → [ ] vlist → [ ] list → [any | list] ilist → [inst | ilist] glist → [ground | glist] vlist → [var | vlist] Notice that if the type symbol var is available, the type any can be defined by any → var (0) any → f1 .. . (0)

any → fi−1

(ki )

any → fi .. .

(any, . . . , any)

(kn )

any → fn

(any, . . . , any)

The only reason to retain it, is that in closed grammars the type var in no longer available, but we still want to define non-ground types. Regular term grammars enjoy several remarkable properties. The following lemmas can be shown, by slightly generalizing results and algorithms given in [10] and in [2]. Theorem 1 ([10]). Given a regular term grammar G and a symbol type T , the set [T ]G is decidable.

188

P. Volpe

Lemma 1 ([10]). The emptiness and the subset relation of regular types is decidable. Indeed, given a grammar G and two type symbols T and S, it is possible to extend G into G0 , with a new symbol S ∩ T and new rules in such a way that [S ∩ T ]G0 = [S]G ∩ [T ]G . Lemma 2 ([10]). There exists an algorithm that computes the intersection of regular types In this paper we will be mainly concerned with closed discriminative regular grammars in normal form. Definition 1. A regular term grammar G = (Σ, V, T, R) is in normal form if each rule have the form T → var or T → f (T1 , . . . , Tn ) with f (n) ∈ Σ , T ∈ T\{var , any}and T1 , . . . , Tn ∈ T. It can be easily shown that each type can be defined by a grammar in normal form. Definition 2. A regular term grammar G is discriminative if it is in normal form and, for each type symbol T , the top functors in Def G (T ) are pairwise distinct. Definition 3. A regular term grammar is closed if it is in normal form and, for each type symbol T , the symbol var does not occur in any element of Def G (T ). Notice that most of the types used in logic programming allow closed and discriminative term grammars. For example, all the grammars introduced in (1), but for vlist, are discriminative and closed. It can be easily shown that regular types defined by a closed grammar are indeed closed under substitution. Theorem 2 ([2]). Given a closed regular term grammar G and a symbol type T , the set [T ]G is closed under substitution. The types that can be defined by a discriminative and closed grammar will be referred to as simple types. Notice that if S and T are simple types, the intersection type S ∩ T is still simple. An operation on types which we will need in the following is the difference of types. Given a discriminative and closed regular term grammar G and two type symbols T and S, we want to extend it to G0 with a new symbol T /S and new rules in such a way that [T \S]G0 = [T ]G \ [S]G , if it is not the case that [T ]G ⊆ [S]G . In general the grammar G0 need not to be nor closed nor discriminative in general. Example 2. Consider the difference type any/inst. It can easily be checked that the set [any] / [inst] is equal to the set of variables V. Anyway there is no closed grammar for such set of terms.

A First-Order Language for Expressing Aliasing and Type Properties

189

We will provide an algorithm which computes the difference T /S, assuming that type S is simple. The algorithm for the general case can be defined but it is more complex and we do not need such a generality. In fact we just need to compute differences types like T1 ∩ · · · ∩Tn \S1 \ · · · \ Sk , where T1 , . . . , Tn and S1 , . . . , Sk are simple types. Difference Algorithm Input. Two type symbols T and S and the set R of rules defining T and S, with S simple. Output. A pair (T \S, S), where T \S is defined by the rules in S, if S 6= ∅; otherwise the difference is empty ([T ] ⊆ [S]). Method. The algorithm is defined by the following recursive function. A set I of difference symbols T \S is used to ensure termination. The type symbol any is supposed to be unfolded as in example (1). difference(T, S, R) = difference(T, S, R, ∅) difference(T, S, R, I) = – If T ⊆ S then return (T \S, ∅); – If the symbol T \S is in I then retun (T \S, R); – Otherwise, let Def R (T ) = {r1 , . . . , rk }. For each i ∈ {1, . . . , k}, let Hi be defined as follows: • if ri = var or (ri = fi (T1 , . . . , Tni ) and the functor fi does not occur in Def R (S)) then let Hi = {(T \S) → ri }; • If ri = fi (T1 , . . . , Tni ) and fi (S1 , . . . , Sni ) ∈ Def R (S) then let (Tj \Sj , Sj ) = S difference(Tj , Sj , R, I ∪ {T \S}), for each j = {1, . . . , ni }, and let Hi = Sj 6=∅ {(T \S) → fi (T1 , . . . , Tj \Sj , . . . , Tni )} ∪ Sj Sk Return (T \S, R ∪ i=1 Hi ) Lemma 3. Let T and S be two type terms defined by the rules of R, S simple. Then difference(T, S, R) terminates and returns a pair (T \S, S) such that [T \S]S = [T ]R \ [S]R . To carry on the elimination of quantifiers in the next section, we need to know the cardinalities of the sets of variables which may occur in a term of a given type. Definition 4. Given a type symbol T defined by a grammar G, the var-cardinality of T , written |T |, is defined as the set {|Vars(t)| | t ∈ [T ]G }. In other terms, k ∈ |T | if and only if there exists t ∈ [T ] such that |Vars(t)| = k. We can prove in a straightforward way the following lemmas. Lemma 4. Given a simple type T , then |T | is equal to {0} or to ω. It is decidable which is the case.

190

P. Volpe

Proof. The relation T ⊆ ground ,with ground as defined in example (1), is decidable by lemma (1). If it is the case then |T | = {0}, otherwise |T | = ω, since T is substitution closed, the signature Σ contains at least a function symbol of arity 2 and a constant, and the set of variables is infinite denumerable. For difference types things can be more complex. Example 3. Consider again the difference type any\inst. It can be easily seen that |any\inst| is the set {1}. Anyway the behaviour of difference types T1 ∩ · · · ∩Tn \S1 \ · · · \ Sk , with T1 , . . . , Tn and S1 , . . . , Sk simple types, is sufficiently regular so as to show the following theorem. Theorem 3. The var-cardinality of the type T1 ∩ · · · ∩Tn \S1 \ · · · \ Sk , with T1 , . . . , Tn and S1 , . . . , Sk simple types, is equal to S ∪ [k, ω], where S is a finite set of natural number and [k, ω], with k = 1, . . . , ω is the set of natural numbers greater or equal to k . The proof provides an effective procedure to compute the var-cardinality of a difference type.

4

A language of properties

In this section we introduce a language that allows to express properties of terms used in static analysis and verification of logic programs: these include groundness, freeness, sharing, type assertions. The language is parametric with respect to a family of types defined through a regular term grammar. It is an extension of the language proposed by Marchiori in [18]. We assume a regular term grammar G = (Σ, V, T, R), discriminative and closed, describing the family of types we are interested in. As before, the set of function symbols Σ is assumed to contain at least a constant and a function of arity 2, V is assumed to be a denumerable set of variables. We define then a first-order language LG = hΣ, Π, Vi, starting from the regular term grammar G. The set of predicate symbol Π consists of the predicates var (1) , share (n) , for (1) each natural n, and a unary predicate pT for each symbol type T ∈ T/{any}. Often we will write T (t) for the atomic formula pT (t). For example, pground (t) will be often written as ground (t). We will omit the subscript G in LG , when no confusion arises. Like in [18], we give the semantics of formulas L by considering the nonground Herbrand interpretation H = hTerms(Σ, V), Λ, Γ i. We define the interpretation Γ for predicate symbols directly through the truth relation |=σ . Given a state σ, the relation |=σ is defined on atoms as follows. – H |=σ var (t) iff σ(t) ∈ V; T n – H |=σ share(t1 , . . . , tn ) iff i=1 Vars(σ(ti )) 6= ∅; – H |=σ pT (t) iff σ(t) ∈ [T ]G , with T ∈ T/{any}.

A First-Order Language for Expressing Aliasing and Type Properties

191

The semantics of the other formulas of L can be derived as usual. We will often write |= ϕ (resp. |=σ ϕ) for H |= ϕ (resp. H |=σ ϕ). Example 4. Let us see examples of the expressive power of LG . – The formula ∀V var (V ) ⇒ ¬share(V, X) asserts the groundness of X; – the formula list(X)∧∃V var (V )∧share(V, X)∧(∀W var (W )∧share(W, X) ⇒ share(V, W )) says that X is a list in which exactly one variable occurs; – ∀V var (V ) ∧ share(V, Y ) ⇒ share(V, X) says that each variable in Y is also in X; – (∀V var (V ) ∧ share(V, Y ) ⇒ share(V, X)) ∧ ground (X) ⇒ ground (Y ) asserts that if ∀V var (V ) ∧ share(V, Y ) ⇒ share(V, X) (i.e. Vars(Y ) ⊆ Vars(X)) and ground (X) (i.e. Vars(X) = ∅) then ground (Y ) (i.e. Vars(Y ) = ∅). Notice that properties expressible in the language LG are invariant with respect to the name of variables. That is, intuitively if property ϕ is true about t ∈ Terms(Σ, V) and s is a variant of t then ϕ is true of s, too. More formally the following lemma can be shown. Lemma 5. For each ϕ, formula of L, if |=σ ϕ and σ 0 is a variant of σ then |=σ0 ϕ. An important class of formulas of LG , which are often considered in analysis and verification, is the class of monotone formulas, that is the formulas ϕ such that |=σ ϕ and σ ≤ σ 0 implies |=σ0 ϕ. For example, ground (X) and ¬var (X) are monotone properties, while var (X) and share(X, Y ) are not. Since the grammar G is closed, by lemma (2), each atom pT1 (X) is monotone. An interesting subclass of monotone properties are the dependences like ∀V var (V ) ∧ share(V, Y ) ⇒ share(V, X). This kind of formulas are used in logic programming analysis since they allow to relate the values to which two or more arguments of a predicate can be instantiated. As shown by the example, the formula ∀V var (V ) ∧ share(V, Y ) ⇒ share(V, X), could be read informally as saying that if X is instantiated to a ground value then also Y is. The class of monotone properties is closed with respect to the connectives ∧,∨ and the quantifiers. We think it would be very interesting to give a complete syntactical characterization of these properties inside L.

5

A proof procedure for L

We are interested in characterizing the set of formulas Th L (H), that is the formulas of the language L which are true in the interpretation H. It is known that the existential fragment of the language L without the type predicates is decidable. In fact in [18], it is proposed a proof procedure to decide the validity of formulas ∃(ϕ1 ∧ · · · ∧ ϕn ) where each atom ϕi is an atom var (t), ground (t), share(t1 , . . . , tn ), or its negation. It is also known that the implication between regular types is decidable. In fact in [3, 2] a procedure is proposed to decide the validity of implications like ∀(pT1 (t1 )∧· · ·∧pTn (tn ) ⇒ pT (t)). It is not

192

P. Volpe

clear whether putting them together and considering the full first order theory, such as in L, the language is still decidable. In this section we will show that this is indeed the case and it is not such a trivial extension of those previous results. To show that Th L (H) is recursive we will use the method of elimination of quantifiers [5, 15]. We single out a set Ω, the elimination set, of fomulas of L and show that each formula of L is equivalent in H to a boolean combination of formulas of Ω. Once proven the decidability of formulas in Ω, we end up with a complete decision procedure for formulas of L. In the following, we use the abbreviation ∃≥k var (V ) φ to say that there exist at least k distinct variables which verify formula φ. It is defined by induction on k. ∃≥1 var (V ) φ is ∃V var (V ) ∧ φ ∃≥k+1 var (V ) φ is ∃V var (V ) ∧ φ ∧ (∃≥k var (W ) φ [V \W ] ∧ ¬share(V, W )), where the formula φ [V \W ] is obtained by φ by replacing all occurrence of V with W . To carry on the elimination of quantifiers we will need a particular class of formulas, which we call cardinality constraints. Definition 5. Let B be the class of boolean terms on V, that is the terms built from the signature ({∩(2) , ∪(2) , ¬(1) , 0(0) , 1(0) }, V). Fixed a natural k and a boolean term t, a simple cardinality constraint αk (t) is defined as the formula ∃≥k var (V ) Ψt (V ), where Ψt (V ) is defined inductively on the sintax of t. – – – – –

Ψ0 (V ) = false and Ψ1 (V ) = true; ΨX (V ) = share(V, X), with X ∈ V; Ψt1 ∩t2 (V ) = Ψt1 (V ) ∧ Ψt2 (V ); Ψt1 ∪t2 (V ) = Ψt1 (V ) ∨ Ψt2 (V ); Ψ¬t (V ) = ¬Ψt (V ).

A simple cardinality constraint αk (t) asserts the membership of at least k elements to the combination of variables in t, seen as subsets of V. Often the term ¬t will be written as t. Example 5. Consider the simple cardinality constraint α3 (X ∩ Y ∩ Z), that is, the formula ∃≥3 var (V ) share(V, X) ∧ share(V, Y ) ∧ ¬share(V, Z). This formula is true in H under the state σ, if there exist at least three variables sharing with σ(X) and σ(Y ) and not with σ(Z), that is if the cardinality of Vars(σ(X)) ∩ Vars(σ(Y )) ∩ Vars(σ(Z)), is at least equal to 3. The following lemma allows us to work with simple cardinality constraints just as if they were assertions on set of variables. Lemma 6. Let σ : V → Terms(Σ, V). Let t∗σ be obtained by the boolean term t by replacing each occurence of a variable X with Vars(σ(X)), for each X ∈ V. Then |=σ αk (t) if and only if the cardinality of t∗σ is at least equal to k. Proof. It can be straightforwardly proved by induction on t and k, noting that, if W ∈ V, then |=σ[V \W ] Ψt (V ) if and only if W ∈ t∗σ

A First-Order Language for Expressing Aliasing and Type Properties

193

Lemmas (7) and (8) follow. Lemma 7. Let σ, σ 0 : V → Terms(Σ, V) two states such that, for each X ∈ V, Vars(σ(X)) = Vars(σ 0 (X)). Then |=σ αk (t) if and only if |=σ0 αk (t). Lemma 8. Let t1 and t2 be two terms equivalent as boolean terms. Then |= αk (t1 ) ⇔ αk (t2 ). We will use the formula α=k (t), that is t contains exactly k elements, as an abbreviation for the formula αk (t) ∧ ¬αk+1 (t). Let the elimination set Ω be composed by the atomic formulas var (X), pT1 (X), . . . , pTn (X), where X ∈ V, and by the set of simple cardinality costraints {αk (t) | k ≥ 1, t is a boolean term}. The idea is to exploit the striking similarity of the simple cardinality constraints in our language with formulas of the first-order theory of the powerset of V seen as a Boolean algebra. For such a theory the decidability has been shown by Skolem in 1919 just through an argument based on elimination of quantifier (see [15] for a slightly more general account). The main idea is to reduce satisfiability of a formula in L to satisfiability of a conjunction of cardinality constraints. Definition 6. A conjunction of simple cardinality constraints αk (t), possibly negated, is a cardinality constraint. The elimination of quantifiers can be carried on for cardinality constraints in a long but straightforward way. The proof is adapted from the proof of elimination of quantifiers of the theory of Boolean algebras in [15]. Theorem 4. Let ψ(X) be a cardinality constraint. Then ∃X ψ(X) is equivalent to a disjunction of cardinality constraints. Anyway, in general, in a formula of L, other formulas than cardinality constraints may occur. We will show then some results that allow to eliminate formulas different from cardinality constraints under existential quantifiers. Definition 7. A formula is flat if it does not contain any functor. Notice that each formula αk (t) is flat. Indeed we can consider only flat formulas as shown by next lemma. Lemma 9. Every formula ϕ of L is equivalent in H to a flat formula. Proof. The following equivalences in H, already appeared in [18] and [2], can be easily checked. – var (f (X)) ⇔ false for each functor f ∈ Σ; Wk – share(t1 , . . . , f (s1 , . . . , sk ), . . . , tn ) ⇔ i=1 share(t1 , . . . , si , . . . , tn ); – pT (f (s1 , . . . , sn )) ⇔ pT1 (s1 )∧· · ·∧pTn (sn ) for each f (T1 , . . . , Tk ) ∈ Def G (T ) (remember that G is discriminative). – pT (f (s1 , . . . , sn )) ⇔ false if f (T1 , . . . , Tk ) 6∈Def G (T ) for any T1 , . . . , Tk .

194

P. Volpe

Definition 8. A type formula is a conjunction of atomic flat formulas pT (X), X ∈ V, possibly negated. Type formulas can be eliminated under existential quantifier and substituted by cardinality constraints. Lemma 10. Let ψ(X) be a cardinality constraint and φ(X) = pT1 (X) ∧ · · · ∧ pTn (X) ∧ ¬pS1 (X) ∧ · · · ∧ ¬pSk (X) a type formula. Let S ∪ [k, ω] be the varcardinality of T1 ∩ · · · ∩Tn \S1 \ · · · \ Sk . Then the following are valid equivalences in H. – (∃X φ(X) ∧ ψ(X)) ⇔ false W if S is empty and k = ω; – (∃X φ(X) ∧ ψ(X)) ⇔ ( h∈S ∃X α=h (X) ∧ ψ(X)) ∨ (∃X αk (X) ∧ ψ(X)). Proof. It is a consequence of lemmas (3) and (7) and of [T1 ] ∩ · · · ∩ [Tn ] ∩ [S1 ]∩ · · · ∩[Sk ] = [T1 ∩ · · · ∩Tn \S1 \ · · · \ Sk ]. In the same way the formulas var (X) and ¬var (X) can be eliminated. Lemma 11. Let ψ(X) be a cardinality constraint. Then (∃X ¬var (X) ∧ ψ(X)) ⇔ (∃X ψ(X)) and (∃X var (X) ∧ ψ(X)) ⇔ (∃X α=1 (X) ∧ ψ(X)). We can prove then the following theorem, which is at the base of the procedure of elimination of quantifiers. Theorem 5. For each formula Ψ (X), conjunction of formulas of Ω, possibly negated, there exists Φ, a boolean composition of formulas of Ω, such that |= (∃X Ψ (X)) ⇔ Φ. Proof. Let Ψ (X) be a conjunction of formulas of Ω, possibly negated. We can write Ψ (X) = Ψ1 (X) ∧ Ψ2 (X), with Ψ1 (X) the conjunction of all simple cardinality constraints in Ψ (X) and Ψ2 (X) the conjunction of the remaining formulas. We can assume that var (X) or ¬var (X) appears in Ψ2 (X). In the first case we can assume that no other formula appears, that is Ψ2 (X) = var (X). In fact if pT (X) appears in Ψ2 (X), we have that Ψ2 (X) ⇔ false. Moreover we have the equivalence var (X) ∧ ¬pT (X) ⇔ var (X). We have then Ψ (X) = Ψ1 (X) ∧ var (X), with Ψ1 (X) a cardinality constraint. By lemma (11) we have that ∃X Ψ1 (X) ∧ var (X) ⇔ ∃X Ψ1 (X) ∧ α=1 (X). In case ¬var (X) appears in Ψ2 (X) and no other formula occurs in Ψ2 (X), we have Ψ (X) = Ψ1 (X) ∧ ¬var (X) and (∃X Ψ1 (X) ∧ ¬var (X)) ⇔ ∃X Ψ1 (X). Otherwise we may assume that Ψ2 (X) has the general form ¬var (X) ∧ pT1 (X) ∧ · · · ∧ pTn (X) ∧ ¬pS1 (X) ∧ · · · ∧ ¬pSk (X), where n + k ≥ 1, which is equivalent to pinst (X) ∧ pT1 (X) ∧ · · · ∧ pTn (X) ∧ ¬pS1 (X) ∧ · · · ∧ ¬pSk (X). We have then Ψ (X) = Ψ1 (X) ∧ (pinst (X) ∧ pT1 (X) ∧ · · · ∧ pTn (X) ∧ ¬pS1 (X) ∧ · · · ∧ ¬pSk (X)) with Ψ1 (X) a cardinality constraint. By lemma (10) we know that ∃X Ψ (X), is equivalent to false or to a disjunction of formulas ∃X ∆i (X), where each ∆i (X) is a cardinality constraint. In any case, we have that ∃X Ψ (X) is equivalent to a disjunction of formulas

A First-Order Language for Expressing Aliasing and Type Properties

195

∃X Φ(X), with Φ(X) a cardinality constraint. We can apply then lemma (4) to each disjunct and obtain a boolean formula, which is a combination of formulas of Ω equivalent to ∃X Ψ (X). Notice that the procedure given in the proof of the previous theorem is effective. At this point it is easy to show that for each first-order formula of L it can be computed a Boolean composition of formulas of Ω equivalent to it on H. Theorem 6. Every formula Ψ of L is equivalent in H to a Boolean composition of formulas of Ω. Proof. Notice that since share(X1 , . . . , Xn ) and α1 (X1 ∩ · · · ∩Xn ) are equivalent in H, we can assume that no formula with predicate share occurs in Ψ . The proof is by structural induction on Ψ . Every atom of L is equivalent to a combination of atomic formulas of Ω (see proof of lemma (9)). If Ψ is a Boolean composition of formulas, then, by induction, each one of these is equivalent to a Boolean combination of formulas of Ω, and thus Ψ is equivalent to a Boolean combination of formulas of Ω. If Ψ is equal to ∃X Λ, then by induction Λ is equivalent to Λ∗ , boolean combination of formulas of Ω. We put Λ∗ in disjunctive normal form. The existential quantifier distributes over the disjunction and Ψ is equivalent to a disjunction of formulas ∃X Ψi with each Ψi conjunction of atoms, possibly negated, of Ω. By the previous theorem, each one of these is equivalent to a boolean formula of Ω. We have proven that each first-order formula of L is equivalent to a boolean combination of formulas of Ω. With the next theorem we state explicitly the decidability of truth of formulas in Ω. Theorem V 7. There exists an algorithm which decides the satisfiability of formulas ϕi , with each ϕi a formula of Ω, possibly negated. V V Proof. The formula is satisfiable in H if ∃ ϕi is true on H, where ∃ ϕi is V the V existantial closure of ϕi . We can apply the elimination of quantifiers to ∃ ϕi . Since the resulting formula has no variable, it must be equivalent to true or false. V Alternatively, an algorithm can be provided that, in case ∃ ϕi is satisfiable, V find V a substitution for the variables of ϕi . It can be obtained by first reducing ∃ ϕi to a disjunction of cardinality constraints by lemmas (10) and (11). An algorithm proposed in [4] to decide validity of a subclass of the language of set theory, can be instantiated to our case to solve cardinality constraints. Once V found such solutions, we have, for each variable X Vof ϕi , the set of variables SX of the term to which X must be mapped for ϕi to be satisfied. At this point a term can be builtVfor each such X, containing the variables in SX and verifying the formulas in ϕi , that is verifying the assertion var (X), ¬var (X) or pT1 (X) ∧ · · · ∧ pTn (X) ∧ ¬pS1 (X) ∧ · · · ∧ ¬pSk (X). In the first two cases the procedure is obvious. In the third case we use the rules that define regular type T1 ∩ · · · ∩Tn \S1 \ · · · \ Sk .

196

P. Volpe

Now combination V Wfor Ψ a general WV W V of formulas of Ω, we have that ∀Ψ ⇔ ∀ ϕi ⇔ ¬∃ ¬ϕi ⇔ ¬ ∃( ¬ϕi ), with each ϕi a formula of VΩ, possibly negated. Then ∀Ψ is true if and only if none of the conjunction ¬ϕ V i is satisfiable. Notice that in case Ψ is not true that means that a conjunction ¬ϕi is satisfiable and, by the previous theorem, a counterexample can be provided.

6

Applications

The logic we have studied is expressive enough to capture many interesting properties of arguments of predicates of logic programs. This suggests to employ L as an assertional language to be used in inductive proof methods. We will show how the verification conditions of many such methods can be entirely expressed by a formula of L. Morover since L is equipped with an algorithmic proof procedure, such verification conditions can effectively be decided. Throughout the section we assume hΣ, Λ, Vi as the signature for logic programs, with Λ the set of predicate symbols, distinct from the predicates of L. An assertion Θ of L is said a specification for predicate p(n) ∈ Λ, if Vars(Θ) ⊆ {X1 , . . . , Xn }. Informally variable Xi refers to the i-th argument of p. An atom p(t1 , . . . , tn ) will be said to satisfy the assertion Θ, written p(t1 , . . . , tn ) |= Θ, iff H |=σ[X1 ,... ,Xn \t1 ,... ,tn ] Θ. The notation Θ [X\t] denotes the formula Θ in which the variables (X1 , . . . , Xn ) are simultaneously substituted by terms (t1 , . . . , tn ). Let us consider first the correct answers of a program. We associate a specification Θp to each predicate p ∈ Λ. Program P is success-correct with respect θ

to {Θp }p∈Pred iff ∀p(t) ∈ Atoms p(t) 2 implies p(t)θ |= Θp . A sufficient condition for correctness can be stated as follows. A program P is success-correct with respect to {Θp }p∈Pred if for each clause p(t) ← p1 (t1 ), . . . , pn (tn ) it is true that H |=

n ^

Θpi [Xi \ti ] ⇒ Θp [X\t] .

(1)

i=1

The method indeed has been proposed in [6] and [12]. In our case, condition (1) can be decided using the procedure of Section (5). If the formula (1) is proved to be true for each clause, then the program is partial correct. Obviously if the formula is false this does not imply that the clause is necessarily wrong. Anyway it could be considered a warning that something wrong can happen. To this aim the alternative algorithm proposed in the proof of Theorem (7), could be very useful, since it would allow to provide counterexamples in such cases. The user then would have more information to decide whether the warning can give raise to a real error or simply the specification is too loose and behaviours are considered that can never occur in practice. If we are interested to check the Input/Output behaviour of logic programs, we can proceed as follows. To each predicate p ∈ Λ, is associated a property

A First-Order Language for Expressing Aliasing and Type Properties

197

pre p _ post p , where pre p and post p are specifications for p. Now a program P is I/O-correct with respect to the properties {pre p _ post p }p∈Pred if p(t) |= pre p and p(t)

θ

2 implies p(t)θ |= post p .

Now in case each formula pre p is monotone, a sufficient condition for P to be I/O-correct with respect to the properties {pre p _ post p }p∈Pred is that H |= (

n ^ i=1

(pre pi [Xi \ti ] ⇒ post pi [Xi \ti ]) ∧ pre p [X\t]) ⇒ post p [X\t] .

This can be shown to correspond to a particular case of the previous method [16]. Anyway, we still have a formula of L that can be decided algorithmically. Finally, if we want to check the call correctness of predicates we can consider methods like those proposed in [13][1]. Like in the previous case, to each predicate p ∈ Λ, is associated a property pre p _ post p , with pre p and post p specifications for p. In this case, anyway, the pre-condition is used also as a specification for the argument of a predicate at call-time. In fact a program P is call-correct with respect to the properties {pre p _ post p }p∈Pred if p(t) |= pre p and p(t)

θ

2 implies p(t)θ |= post p

and ∗

p(t) |= pre p and p(t) → hq(s), Gi implies q(s) |= pre q . We are assuming a leftmost selection rule for SLD-derivations. In [1] it has been shown that, in the case pre p and post p are monotone for each p, then a sufficient condition for P to be call-correct with respect to {pre p _ post p }p∈Pred is that for each clause p(t) ← p1 (t1 ), . . . , pn (tn ), it is true that for each 1 ≤ k ≤ n + 1 H |= (pre p [X\t] ∧

k−1 ^ i=1

post pi [Xi \ti ] ⇒ pre pk [Xk \tk ])

where pre pn+1 [Xn+1 \tn+1 ] ≡ post p [X\t]. In this case we have a finer control on the possible run-time behaviour. In fact a warning in this case can be raised because there may be a computation that calls a predicate with arguments which violate the specification. Again if a counterexample is provided the user may decide whether the specification is too loose or an actual error has been discovered. We want to stress the restriction in previous methods to monotone properties. The reason for not considering all the expressible properties is that in those cases the verification condititions become much more complex and the mgu’s have to be considered explicitly (see [16]). While for monotone assertions, the verification conditions are expressible as formulas of L, this is no more the case if more general properties are considered. A solution we are working on, is to enrich L so as to express formally the mgu’s. Anyway the class of properties that can be mechanically checked is still quite large, including type assertions, groundness, dependencies.

198

7

P. Volpe

Conclusions

In this paper we have studied a language that allows to express and decide properties of finite terms on a given signature, that is the data on which logic programs operate. The language is able to express aliasing properties such as the sharing or freeness and is enriched with type assertions. Using formulas of L, we are able to express many properties of program and prove them in inductive proof methods. Moreover since the logic is decidable we can mechanically check the corresponding verification condition and raise a warning if the condition is not verified. In such cases counterexamples can be built, which can be helpful for the user who is carrying on the verification. The set of true formulas is proved to be decidable through a procedure of elimination of quantifiers. This points out an interesting class of formulas, which express cardinality constraints on the set of variables that can occur in a term. This can give an interesting insight on domains used for the analysis of logic program. In fact many of the domains used for aliasing analysis can be seen as fragments of the domain of formulas of cardinality constraints. For example the element (X ∧ Y ) → (Z ∨ W ) of POS [9], can be represented as (Vars(Z) ⊆ Vars(X) ∪ Vars(Y )) ∨ (Vars(W ) ⊆ Vars(X) ∪ Vars(Y )) , which is equivalent to (|Vars(Z) ∩ Vars(X) ∩ Vars(Y )| = 0) ∨ (|Vars(W ) ∩ (Vars(X) ∩ Vars(Y ))| = 0), that is the constraint α=0 (Z ∩ X ∩ Y ) ∨ α=0 (W ∩ X ∩ Y ) of L. Another example is Sharing [14]. In fact the element {∅, {X}, {Y, Z}} can be represented as α=0 (X ∩ Y ) ∧ α=0 (X ∩ Z) ∧ α=0 (Y ∩ Z) ∧ α=0 (Z ∩ Y ). We recall that cardinality constraints are equipped with an operation of conjunction and cylindrification (existential quantifier). We are thinking about ways of augmenting the expressive power of the logic L, obviously while retaining the decidability. We are currently investigating two possibilities. The first consists in adding a modal operator 2 defined as follows |=σ 2ϕ

iff for each σ 0 ≥ σ |=σ0 ϕ.

Such modality would allow, first of all, to characterize monotone properties of language L. In fact monotone properties would correspond to the formulas Ψ such that Ψ ⇔ 2Ψ . Morover we could express arbitrary dependences between properties, like 2(list(X) ⇒ list(Y )), whose informal meaning is that every state that instantiate X to a list will also bind Y to a list. Another extension is to consider Hoare-like triples {Φ}[[t, s]]{Ψ }, whose meaning is: if Φ is true under the state σ and θ = mgu(σ(t), σ(s)), then Ψ is true under the state σ◦θ. They have been considered in [7] and [8], where a formal calculus has been provided for a language of assertions different from L. These formulas would allow to express formally the verification of general proof methods, like those proposed in [13] and in [16], for the whole class of formulas of L. At the moment, it is still not known if, given a decidable logic such as L, it is possible to decide the validity of such triples.

A First-Order Language for Expressing Aliasing and Type Properties

199

References 1. A. Bossi and N. Cocco. Verifying Correctness of Logic Programs. In J. Diaz and F. Orejas, editors, Proc. TAPSOFT’89, pages 96–110, 1989. 2. J. Boye. Directional Types in Logic Programming. PhD thesis, University of Link¨ oping, Computer Science Department, 1997. 3. J. Boye and J. Maluszynski. Directional Types and the Annotation Method. Journal of Logic Programming, 33(3):179–220, 1997. 4. D. Cantone, E. G. Omodeo, and A. Policriti. The Automation of Syllogistic II. Optimization and Complexity Issues. Journal of Automated Reasoning, 6(2):173– 187, 1990. 5. C. C. Chang and H. J. Kreisler. Model Theory. Elsevier Science Publ., 1990. Third edition. 6. K. L. Clark. Predicate logic as a computational formalism. Res. Report DOC 79/59, Imperial College, Dept. of Computing, London, 1979. 7. L. Colussi and E. Marchiori. Proving Correctness of Logic Programs Using Axiomatic Semantics. In Proc. of the Eight International Conference on Logic Programming, pages 629–644. The MIT Press, Cambridge, Mass., 1991. 8. L. Colussi and E. Marchiori. Unification as Predicate Transformer. In Proc. of the Joint International Conference and Symposium on Logic Programming, pages 67–85. The MIT Press, Cambridge, Mass., 1992. 9. A. Cortesi, G. Fil`e, and W. Winsborough. Prop revisited: Propositional Formula as Abstract Domain for Groundness Analysis. In Proc. Sixth IEEE Symp. on Logic In Computer Science, pages 322–327. IEEE Computer Society Press, 1991. 10. P. Dart and J. Zobel. Efficient run-time type checking of typed logic program. Journal of Logic Programming, 14(1-2):31–70, 1992. 11. P. Dart and J. Zobel. A regular type language for logic programs. In F. Pfenning, editor, Types in logic programming, pages 157–187. The MIT Press, Cambridge, Mass., 1992. 12. P. Deransart. Proof Methods of Declarative Properties of Definite Programs. Theoretical Computer Science, 118(2):99–166, 1993. 13. W. Drabent and J. Maluszynski. Inductive Assertion Method for Logic Programs. Theoretical Computer Science, 59(1):133–155, 1988. 14. D. Jacobs and A. Langen. Accurate and Efficient Approximation of Variable Aliasing in Logic Programs. In E. Lusk and R. Overbeek, editors, Proc. North American Conf. on Logic Programming’89, pages 154–165. The MIT Press, Cambridge, Mass., 1989. 15. G. Kreisel and J. L. Krivine. Elements of Mathematical Logic (Model Theory). North-Holland, Amsterdam, 1967. 16. G. Levi and P. Volpe. Derivation of Proof Methods by Abstract Interpretation. (Submitted). Available at http://www.di.unipi.it/∼volpep/papers.html, 1998. 17. J. W. Lloyd. Foundations of Logic Programming. Springer-Verlag, Berlin, 1987. Second edition. 18. E. Marchiori. A Logic for Variable Aliasing in Logic Programs. In G. Levi and M. Rodriguez-Artalejo, editors, Proceedings of the 4th International Conference on Algebraic and Logic Programming (ALP’94), number 850 in LNCS, pages 287–304. Springer Verlag, 1994. 19. E. Marchiori. Design of Abstract Domains using First-order Logic. In M. Hanus and M. Rodriguez-Artalejo, editors, Proceedings of the 5th International Conference on Algebraic and Logic Programming (ALP’96), number 1139 in LNCS, pages 209– 223. Springer Verlag, 1996.

Refining Static Analyses by Trace-Based Partitioning Using Control Flow Maria Handjieva and Stanislav Tzolovski LIX, Ecole Polytechnique, France {handjiev,stivy}@lix.polytechnique.fr

Abstract. This paper presents a systematic method of building a more precise static analysis from a given one. The key idea is to lift an abstract domain to the finite sets of its labeled abstract properties. The labels are designed to gather information about the history of control flow and to obtain a finite partitioning of the program execution traces. The abstract operations of the lifted domain are derived from those of the original one. This is a particular instance of the reduced cardinal power introduced by P. and R. Cousot, where the base is the set of labels approximating the control history and the exponent is an abstract domain. The method is applied to the domain of convex polyhedra and to the domain of linear congruences. Key words: abstract interpretation, reduced cardinal power, trace semantics.

1

Introduction

An essential part of static analysis by abstract interpretation [4,5] is to build a machine-representable abstract domain expressing interesting properties about a computer program. Many abstract domains have been developed for many purposes [3,7,9,11]. The most significant analyses dealing with numerical variables are constant propagation, analysis using intervals, linear equalities, linear inequalities, linear congruences. We are interested in refining static analyses based on a given domain. For non-distributive abstract interpretations (such as convex polyhedra), the disjunctive completion [6] is an operator that systematically produces new and more precise abstract domains from simpler ones. This was illustrated in [8] for the domain PROP for ground-dependence analysis of logic programs. We propose a method for lifting an abstract domain to the finite sets of its labeled abstract properties. The labels are designed to gather information about the history of control flow and to obtain a finite partitioning of the execution traces. Each partition (the abstract properties with the same label) is approximated with an upper bound of its elements. The abstract operations of the new domain are derived from those of the original one. We apply this method to the domain of convex polyhedra and to the domain of linear congruences. G. Levi (Ed.): SAS’98, LNCS 1503, pp. 200–214, 1998. c Springer-Verlag Berlin Heidelberg 1998

Refining Static Analyses by Trace-Based Partitioning Using Control Flow

201

A Motivated Example. A static analysis of program variables using convex polyhedra is proposed in [7]. We slightly change the illustrated example from [7] as follows: {U0 } {U1 } {U2 } {U3 } {U4 } {U5 } {U6 } {U7 } {U8 } {U9 } {U10 } {U11 } {U12 } {U13 }

[c1 ] i = 2; [c2 ] j = 0; [c3 ] if (Cond1 ) then [c4 ] k = 1; else [c5 ] k = −1; endif [c6 ] while (Cond2 ) do [c7 ] if (Cond3 ) then [c8 ] i = i + 4 ∗ k; else [c9 ] i = i + 2 ∗ k; [c10 ] j = j + k; endif enddo; [c11 ] exit;

For each statement Stat of the program we have two abstract properties: Ui and Uo called input (before execution of Stat) and output (after execution of Stat) abstract properties respectively. We denote the label of statement i as [ci ] and write the label before the statement. If we apply the analysis based on convex polyhedra to the above program we will obtain the convex polyhedron U13 = (−1 ≤ k ≤ 1). The result is too abstract, because the analysis cannot find that the value of variable k in the loop is either 1 or −1. The intuition is to represent the restraints between variables with two convex polyhedra (one for k = −1 and another for k = 1). In this case the result is U13 = {(i − 2j ≤ 2, j ≤ 0, k = −1), (i − 2j ≥ 2, j ≥ 0, k = 1)}, which is obviously more precise than the previously obtained one. Observe in this example, that this is the reduced cardinal power [5] with boolean base and polyhedral exponent. Therefore we need an abstract domain, which treats finite sets of convex polyhedra. Because the domain is not finite and does not satisfy the ascending chain condition we need also a technique that guarantees the termination of the analysis. Widening operators are an appropriate technique for this purpose. Moreover, we can use the same idea for other abstract domain, hence a method that systematically lifts a given abstract domain to the set of abstract properties is needed. The rest of this paper is structured as follows. In the next section we introduce the set of labeled abstract properties corresponding to a a given abstract domain. Starting from a concrete trace semantics and passing through a collecting semantics we design an abstract semantics using a concretization function and a widening. In Sects. 3 and 4 we apply our method to the convex polyhedra domain and to the domain of linear congruences respectively. We conclude the paper in Sect. 5.

202

2

M. Handjieva and S. Tzolovski

Trace-Based Partitioning Using Control Flow

In order to simplify the presentation all formal reasoning is done using a simple programming language with assignments, “if-then-else” constructs and “while” loops. Our framework takes two parameters. The first one is an abstract domain: (D, vD , uD , tD , ⊥D , >D , αD , γD , ∇D , assign D , True D , False D ) where D is the set of abstract properties, vD is the abstract ordering, uD and tD are the upper and lower bound respectively, ⊥D and >D are the minimal and maximal element respectively, αD is the abstraction function (if any) and γD is the concretization function, ∇D is a widening operator (if any). We add three specific functions: assign D ∈ (D × ASSIGN) → D that treat the assignments (from ASSIGN), True D ∈ (D × COND) → D and False D ∈ (D × COND) → D that treat the conditions in the program. For example, in the domain of closed intervals (denoted by I) if the value of variable x is approximated with the interval [1, 5] then: assign I ([1, 5], x = x + 1) = [2, 6] False I ([1, 5], x ≥ 3) = [1, 3] True I ([1, 5], x ≥ 3) = [3, 5], An “if-then-else” or a “while” statement is called a test node. Each test node has a unique number. Let T be the set of test node numbers and t be the number of test nodes (t = |T |). Each test node can be analyzed in different ways. The analysis of an “if” statement could return a set of one abstract property, which contains information about both the “true” and “false” branches or it could return a set of two abstract properties one for each branch. In the case of a “while” statement the analysis could obtain a set of k + 1 elements: one when the control does not pass through the loop; one - when the control passes exactly once through the loop, etc.; one - when the control passes k or more times through the loop. So, the second parameter of our framework is a tuple of t elements, denoted by ∆ and called test approximation parameter, that contains information on how to approximate each test node. 2.1

Concrete Semantics

The concrete semantics of a program P at point z is the set of all execution traces from the entry point to point z. Sz [[P ]] = {[c1 , M1 ] → [c2 , M2 ] → . . . → [cz , Mz ], . . . } where ci is the label of statement i and Mi is the memory state (a vector of values) before the execution of statement i. 2.2

Collecting Semantics

The starting point of an abstract interpretation is a collecting semantics for the programming language. Our collecting semantics is a set of configurations. Each

Refining Static Analyses by Trace-Based Partitioning Using Control Flow

203

configuration is a pair of a string of labels and a vector of values. In what follows configurations are always surrounded with left and right angles. For each trace in the concrete semantics Sz [[P ]] there is a configuration in the collecting semantics that characterizes the control history of this trace and its current memory state. Cz [[P ]] = {hc1 c2 . . . cz , Mz i , hc01 c02 . . . c0z , Mz0 i , . . . , hc001 . . . c00z , Mz00 i} The concretization γcs and abstraction αsc functions, that describe the link between concrete and collecting semantics, are defined recursively as follows: γcs (C) =

[

γt (hc1 · · · cz , Mz i)

hc1 ···cz ,Mz i∈C

γt (hc1 · · · cz , Mz i) = [c1 , M1 ] → γt (hc2 · · · cz , Mz i), γt (hcz , Mz i) = [cz , Mz ]

M1 is any memory state

To define the abstraction function we use the following notations: all letters in bold are strings, is the empty string and “·” is the string concatenation. αsc (T ) = {αt (, t) | t ∈ T } αt (c, [c, M ] → t) = αt (c · c, t) αt (c, [c, M ]) = hc · c, M i S S Note that the pair (αsc , γcs ) is a Galois connection, since αsc ( Ti ) = αsc (Ti ). i

2.3

i

Set of Labeled Abstract Property Domain

The second step is the abstraction of the collecting semantics. Our abstract domain is a finite set of labeled abstract properties from a given abstract domain. An abstract label is a string from the alphabet A = {b, tj , fj , vj }, for j ∈ T , such that it begins with b and there are no more b. The letter b describes the path from the entry node to the first test node. The letter tj (resp. fj ) describes the part of the “true” (resp. “false”) branch of test node j (from j to the next test node). The abstract letter vj represents: – all paths that pass a finite number of times through j, if j is a “while” loop; – the two paths that pass through j, if j is an “if-then-else” statement. The concrete control flow graph contains one node for each statement and a node for each junction point (“endif”). The abstract control flow graph contains: one node for each test point j, labeled with j, an entry, an exit node and one node for each junction point. The junction nodes are added for simplification. For example, the concrete control flow graph and the abstract control flow graph of the program from Sect. 1 are illustrated in Fig. 1.

204

M. Handjieva and S. Tzolovski

1

c4 2

c1 c2 c3

b t1

2

c7

f2

t2 c9

c8

f1

c5

c6 3

1

c11

t3

3

f3

c10

Fig. 1. Concrete and abstract control flow graphs

The meaning of the abstract letters (from alphabet A) are defined with the help of the concretization function γl . For the program from Sect. 1 it is: γl (b) = {c1 c2 c3 } γl (t1 ) = {c4 c6 } γl (t2 ) = {c7 }

γl (f1 ) = {c5 c6 } γl (f2 ) = {c11 }

γl (t3 ) = {c8 c6 }

γl (f2 ) = {c9 c10 c6 }

The intuition behind vj is expressed using function δ, recursively defined below. j vj ≡ δ P

tj

δ P

= (tj · δ(P ))∗ j

tj

δP

fj

= (tj · δ(P ))∗ · fj

tj

j

fj

R = tj · δ(P ) | fj · δ(R)

P δ = δ(P ) · δ(R) R

δ( | ) =

where P and R are part of the reducible abstract control flow graph with exactly one entry edge and exactly one exit edge. In this way for each abstract letter vj we obtain a regular expression that characterizes it. For example, v2 ≡ (t2 · (t3 | f3 ))∗ = {, t2 t3 , t2 f3 , t2 t3 t2 t3 , t2 t3 t2 f3 , . . . } The meaning of an abstract letter vj is formally given by function γl over regular expressions recursively defined as follows: γl (X · Y ) = {x · y | x ∈ γl (X), y ∈ γl (Y )} γl (X ∗ ) = {x · · x} | x ∈ γl (X), n ∈ IN} | ·{z n

γl (X | Y ) = {x | x ∈ γl (X) ∨ x ∈ γl (Y )}

Refining Static Analyses by Trace-Based Partitioning Using Control Flow

205

where X and Y are regular expressions. So γl (vj ) is γl (δ(Gj )), where Gj is the subgraph of the abstract control flow graph with nodes from the strongly connected components with head j. Function γL gives the meaning of an abstract label a (which is a string of abstract letters a = a1 . . . an ) : γL (a) = γL (a1 . . . an ) = {s1 . . . sn | si ∈ γl (ai ), i = 1, . . . , n} The abstract labels are ordered by: a a0 iff γL (a) ⊆ γL (a0 ) For example, bt1 t2 t3 bt1 v2 , but bt1 f2 6bt1 v2 . An abstract configuration is an element of the reduced cardinal power [5] with sets of abstract labels as atomic base and abstract domain D as exponent which represent as a pair of an abstract label and an abstract property, denoted as ha, Di, where D ∈ D. Our abstract domain consists of sets of abstract configurations with an additional condition: there are not two elements with comparable labels in each set. This additional condition guarantees correct trace partitioning. 2.4

Abstract Operations

The last step consists in computing the least fixpoint of the abstract semantics. In this subsection we define the operations on the abstract domain needed when computing this least fixpoint (described in the next subsection). Approximation ordering. We introduce the following ordering so that we can compare two sets of labeled abstract properties. U v U 0 iff ∀ ha, Di ∈ U, ∃ ha0 , D0 i ∈ U 0 , a a0 ∧ D vD D0 Upper bound. The intuition behind designing the upper bound operation is to obtain either a set union over the paths with non-comparable labels or merge over the paths with comparable labels. In this way we cannot merge paths from different partitioning. ha, Di ∈ U, ha0k , Dk0 i ∈ U 0 , 0 0 0 ha, D tD D1 tD · · · tD Dn i 0 U tU = ak a, k = 1, . . . , n ha0 , D0 i ∈ U 0 , hak , Dk i ∈ U, 0 0 ∪ ha , D tD D1 tD · · · tD Dm i ak a0 , k = 1, . . . , m ∪ {ha, Di | ha, Di ∈ U, 6 ∃ ha0 , D0 i ∈ U 0 , (a a0 ∨ a0 a)} ∪ {ha0 , D0 i | ha0 , D0 i ∈ U 0 , 6 ∃ ha, Di ∈ U, (a a0 ∨ a0 a)} Widening operator. A convergence technique has been proposed in [4] that uses so called widening operator to transform infinitely iterative computation into finite but approximate one. Let (L, v, ⊥, >, t) be a poset with maximal element >. The following definition is due to Cousot [2].

206

M. Handjieva and S. Tzolovski

Definition 1 A widening operator ∇ ∈ (IN → (L × L → L)) is such that1 : 1. ∀i > 0, ∀x, y ∈ L, (x t y) v (x∇i y); 2. For any ascending chain y 0 v y 1 v · · · v y n v · · · of elements of L, the ascending chain x0 = y 0 , x1 = x0 ∇1 y 1 , . . . , xn = xn−1 ∇n y n , . . . is eventually stable, i.e. there exists k ≥ 0 such that for i > k, xi = xk . The widening on our abstract domain is defined as follows: U ∇U 0 = {ha, D∇D (D tD D1 tD . . .tD Dn )i | ha, Di ∈ U, hai , Di i ∈ U 0 , ai a} ∪ {ha, Di | ha, Di ∈ U ∧ 6 ∃ ha0 , D0 i ∈ U 0 , a0 a} It is designed only for sets of labeled abstract properties U and U 0 such that: ∀ ha0 , D0 i ∈ U 0 , ∃ ha, Di ∈ U, a0 a This condition is not a restriction because, the widening operator ∇ is always applied between two successive iterations of a “while” loop and we analyze only structured programs. Proposition 1 The operator ∇ defined above is a widening. Proof: It is clear that U t U 0 v U ∇U 0 . On the other hand, the number of abstract properties in U ∇U 0 is equal to the number of abstract properties in U (|U | = |U ∇U 0 |). Moreover for each abstract label a we cannot have infinite many abstract properties D1 vD D2 vD . . . from successive iterations, because of the application of the widening ∇D . Therefore the operator ∇ cannot iterate infinitely without convergence. 2 2.5

Abstract Semantics

The abstract semantics of a program P at point z is a set of abstract configurations ha, Di, where a is an abstract label and D ∈ D is an abstract property. Az [[P ]] = {hai , Di i | k = 1, . . . , n} In our approach, the link between the abstract and collecting semantics is given by a monotone concretization function Γ defined below. [ Γ (U ) = γ(ha, Di) ha,Di∈U

γ(ha, Di) = {hs, M i | s ∈ γL (a), M ∈ γD (D)} Assignments. Consider the following assignment: {U } 1

x = Expr

We denote ∇n the operator ∇(n), n ∈ IN.

{U 0 }

Refining Static Analyses by Trace-Based Partitioning Using Control Flow

207

Its output set of labeled abstract properties is given by: U 0 = assign(U, x = Expr ) = {ha, D0 i | ha, Di ∈ U, D0 = assign D (D, x = Expr )} If-then-else constructs. We now give the abstract meaning of an “if-then-else” statement, where BlockThen and BlockElse are blocks of statements: {U } {U1 } {U2 }

if (Cond) then BlockThen else BlockElse endif

{U11 } {U22 } {U 0 }

There are two possible ways to approximate the information after “if-then-else” statement i. If ki = 1 (ki is the i-th element of the test approximation parameter) then we set lt = ti and lf = fi , otherwise (when ki = 0) we set lt = lf = vi . U1 = {ha · lt , D0 i | ha, Di ∈ U, D0 = True D (D, Cond), D0 6= ⊥D } U2 = {ha · lf , D0 i | ha, Di ∈ U, D0 = False D (D, Cond), D0 6= ⊥D } U 0 = U11 t U22 Note that when ki = 1 there is not a merge of the “true” and “false” paths, because the upper bound is simply the set union. Which means that there is no loss of information in this case. Contrarily to common choice of abstracting the merge of these two paths by an upper bound on the domain, which is an upper approximation for non-distributive abstract interpretation frameworks. For example, in the convex polyhedra domain, where the union of two polyhedra does not necessarily correspond to the convex polyhedron, it is approximated by the convex hull. While loops. Consider the following loop statement: {U } {U1 }

while (Cond) do Block enddo

{U2 } {U 0 }

In order to find the fixpoint of the abstract semantics we use additional sets of labeled abstract property V i which is always valid before the i-th iteration. when i = 0 U,i U2 , when 0 < i < kj i V = kj {ha · v , Di | ha, Di ∈ U }, when i = kj j 2 i−1 V ∇U2i when i > kj U1i = {ha · tj , D0 i | ha, Di ∈ V i−1 , D0 = True D (D, Cond), D0 6= ⊥D } [ U0 = {ha · fj , D0 i | ha, Di ∈ V i , D0 = False D (D, Cond), D0 6= ⊥D } i=0,kj −1

∪{ha · fj , D0 i | ha, Di ∈ V kj +y , D0 = False D (D, Cond), D0 6= ⊥D } where y is such that the local fixpoint (for the loop j) is reached, which means the smallest y such that V kj +y = V kj +y+1 .

208

M. Handjieva and S. Tzolovski

2.6

Complexity and Precision

The complexity of our analysis depends on its two parameters - the abstract domain and the test approximation parameter. For each program Q, having its test approximation parameter ∆Q we can calculate the maximal number of labeled abstract properties in the abstract semantics (which is in fact the number of partitions). Let σ be a function which for a given abstract control flow subgraph (with exactly one entry and exactly one exit edge) calculates this number according to ∆Q = (k1 , k2 , . . . , kt ) . We recursively define σ as follows: j σP σ P

tj

fj

tj

j

, u = u + σ(P, u) + · · · + σ(P, . . . , σ( P, u) . . . ) | {z }

fj R

, u =

kj

u, if kj = 0 σ(P, u) + σ(R, u), if kj = 1

δ ( | , u) = u

! σ

P R

,u

= σ(R, σ(P, u))

where j is a test node, P and R are part of the abstract control flow graph with exactly one entry edge and exactly one exit edge and u is a number of partitions at the entry edge. Therefore the maximal number of elements in the abstract semantics at the and of program Q is σ(Q, 1). Obviously the analysis based on the lifted abstract domain is more costly than the analysis based on the original one. For example, the maximal number of elements in the set abstract properties at Pkof2 labeled point 13 of the program from Sect. 1 is 2k1 i=0 2k3 i , where k1 , k3 ∈ {0, 1} and k2 ∈ IN. The precision of our analysis also depends on the test approximation parameter. We can argue that our analysis based on the lifted domain provides more accurate results than or the same results as the analysis based on abstract domain D. Proposition 2 The analysis based on finite sets of labeled abstract properties of D is more precise than the analysis based on D. In the worst case of our analysis, when ∆ = [0, . . . , 0] the result is a set of one abstract property, which is exactly the same as the result of the static analysis based on D. We can design a family of static analyses changing the test approximation parameter as making different compromises between complexity and precision.

3 3.1

Finite Sets of Labeled Convex Polyhedra Convex Polyhedra Domain

Static analysis of linear inequalities among variables of a program have been studied by P.Cousot and N.Halbwachs in [7]. To set up some notations we briefly

Refining Static Analyses by Trace-Based Partitioning Using Control Flow

209

present the definitions related to convex polyhedra [12]. The convex hull of a set X of vectors is the smallest convex set containing X and it is denoted by convexHull (X), so: convexHull (X) = {λ1 x1 + · · · + λk xk | k ≥ 1; x1 , . . . , xk ∈ X; λ1 , . . . , λk ∈ IR; λ1 , . . . , λk ≥ 0; λ1 + · · · + λk = 1} A set P of vectors in IRn is called a convex polyhedron if P = {x | Ax ≤ b} for some matrix A and vector b. In other words, the polyhedron is the intersection of finitely many affine half-spaces. An affine half-space is a set {x | ax ≤ b} for some nonzero row vector a and some number b. A set of vectors is a convex polytope if it is the convex hull of finitely many vectors. The convex hull of two convex polyhedra is the smallest convex set defined as follows: chull (P1 , P2 ) = convexHull ({X | X ∈ P1 ∨ X ∈ P2 }) Upper Bound. The upper bound of two convex polyhedra is their convex hull. So, we have: P1 tp P2 = chull (P1 , P2 ) and P1 tp · · · tp Pn = chull (P1 , . . . , Pn ). Affine Transformation. The affine transformation τp transforms a polyhedron P into another polyhedron P 0 according to a given affine map x → Ax + b. Polyhedron P 0 is specified as follows: P 0 = {x0 | x ∈ P, x0 = Ax + b} Widening Operator. The widening operator of convex polyhedra ∇p (defined in [7] and optimized in [10]) is based on the following heuristic: the widening of two polyhedra is a new polyhedron obtained by removing from the system of the first polyhedron all inequalities which are not satisfied by the second polyhedron. 3.2

Specific Functions

In the following, specific functions assign p , True p and False p that deal with convex polyhedra are presented. Assignments. Let x = Expr be an assignment. For a given polyhedron P , function assign p returns a polyhedron P 0 in the following manner: (a) If Expr is nonlinear expression we make a conservative approximation, assuming that any value can be assigned to x. Therefore, variable x is eliminated from P by projecting P along the x dimension. P n then function (b) If Expr is linear expression, i.e. Expr ≡ ( i=1 ai xi + b) P n the affine map x → i=1 ai xi + b. assign p is an affine transformation defined byP n In the case of invertible assignments xk = i=1 ai xi + b (when ak 6= 0) the system of restraints of each polyhedron can be computed directly [7]. In the case of non-invertible assignments (when ak = 0) we firstPproject polyhedron P along n the xk dimension and then add the restraint xk = i=1 ai xi + b. Conditions. According to condition Cond we have the following three cases: (a) condition Cond is nonlinear. We ignore the test by setting both functions True p and False p to behave as the identity function on convex polyhedra: True p (P, Cond) = False p (P, Cond) = P ; (b) condition Cond is a linear equality, i.e. Cond ≡ (aX = b). Let H be the hyperplane {X ∈ Rn | aX = b}. If polyhedron P is included in H then

210

M. Handjieva and S. Tzolovski

T ruep (P, Cond) = P and F alsep (P, Cond) = ⊥p . Otherwise T ruep (P, Cond) = P ∩ H and F alsep (P, Cond) = P ; (c) the condition Cond is a linear inequality, i.e. Cond ≡ (aX ≤ b). Let H1 be the closed half-space {X ∈ Rn | aX ≤ b} and let H2 be the closed halfspace {X ∈ Rn | aX ≥ b}. In this case we set True p (P, Cond) = P ∩ H1 and False p (P, Cond) = P ∩ H2 . 3.3

Example

We illustrate the application of the method described in Sect. 2 using convex polyhedra domain on the motivated example (see Sect. 1). The program is simple enough to allow hand computation and simple geometrical representation. All tests involve nonlinear conditions, which are not taken into account by the analysis. The test approximation parameter is set up to be ∆ = (1, 0, 1). Which means that the first test node (the “if” with label [c3 ]) will be treated exactly (because k1 = 1), the second test point (the “while” loop) will be approximated at the beginning (because k2 = 0) etc. Each set of labeled polyhedra Ui , (i = 0, . . . , 13) is initially empty (Ui0 = ∅). The superscript shows the number of steps the analysis passes through point i. At the entry node we set up U01 = IRn , where n is the number of variables involved in the analysis. U01 = IR3 U11 = assign(assign(U01 , i = 2), j = 0) = {hb, (i = 2, j = 0)i} U21 = {hbt1 , (i = 2, j = 0)i} U31 = assign(U21 , k = 1) = {hbt1 , (i = 2, j = 0, k = 1)i} U41 = {hbf1 , (i = 2, j = 0)i} U51 = assign(U41 , k = −1) = {hbf1 , (i = 2, j = 0, k = −1)i} U61 = U31 t U51 = {hbt1 , (i = 2, j = 0, k = 1)i , hbf1 , (i = 2, j = 0, k = −1)i} Now, we are going to analyze the “while” loop. V 0 = {hbt1 v2 , (i = 2, j = 0, k = 1)i , hbf1 v2 , (i = 2, j = 0, k = −1)i} U71 = {hbt1 v2 t2 , (i = 2, j = 0, k = 1)i , hbf1 v2 t2 , (i = 2, j = 0, k = −1)i} U81 = {hbt1 v2 t2 t3 , (i = 2, j = 0, k = 1)i , hbf1 v2 t2 t3 , (i = 2, j = 0, k = −1)i} U91 = assign(U81 , i = i + 4 ∗ k) = = {hbt1 v2 t2 t3 , (i = 6, j = 0, k = 1)i , hbf1 v2 t2 t3 , (i = −2, j = 0, k = −1)i} 1 U10 = {hbt1 v2 t2 f3 , (i = 2, j = 0, k = 1)i , hbf1 v2 t2 f3 , (i = 2, j = 0, k = −1)i} 1 1 U11 = assign(assign(U10 , i = i + 2 ∗ k), j = j + k) = = {hbt1 v2 t2 f3 , (i = 4, j = 1, k = 1)i , hbf1 v2 t2 f3 , (i = 0, j = −1, k = −1)i} 1 1 U12 = U91 t U11 = hbt1 v2 t2 t3 , (i = 6, j = 0, k = 1)i , hbf1 v2 t2 t3 , (i = −2, j = 0, k = −1)i = hbt1 v2 t2 f3 , (i = 4, j = 1, k = 1)i , hbf1 v2 t2 f3 , (i = 0, j = −1, k = −1)i

Refining Static Analyses by Trace-Based Partitioning Using Control Flow

211

If we unroll the loop once again we will obtain the following result: 1 V 1 = V 0 ∇U12 = hbt1 v2 , (i − 2j ≥ 2, i + 2j ≤ 6, j ≥ 0, k = 1)i = hbf1 v2 , (i − 2j ≤ 2, i + 2j ≥ −2, j ≤ 0, k = −1)i

The above widening behaves as an upper bound. It returns a set of two polyhedra. The first one is the convex hull of the polyhedra with labels bt1 v2 , bt1 v2 t2 t3 and bt1 v2 t2 f3 (because bt1 v2 t2 t3 bt1 v2 , bt1 v2 t2 f3 bt1 v2 ) and the second one is the convex hull of the polyhedra with labels bf1 v2 , bf1 v2 t2 t3 and bf1 v2 t2 f3 . hbt1 v2 t2 t3 , (i − 2j ≥ 6, i + 2j ≤ 10, j ≥ 0, k = 1)i hbf1 v2 t2 t3 , (i − 2j ≤ −2, i + 2j ≥ −6, j ≤ 0, k = −1)i 2 U12 = hbt1 v2 t2 f3 , (i − 2j ≥ 2, i + 2j ≤ 10, j ≥ 1, k = 1)i hbf1 v2 t2 f3 , (i − 2j ≤ 2, i + 2j ≥ −6, j ≤ −1, k = −1)i At this point the widening operator takes place: hbt1 v2 , (i − 2j ≥ 2, j ≥ 0, k = 1)i , 2 1 2 V = V ∇U12 = hbf1 v2 , (i − 2j ≤ 2, j ≤ 0, k = −1)i Finally, this is the fixpoint, therefore we find: hbt1 v2 f2 , (i − 2j ≥ 2, j ≥ 0, k = 1)i , 1 U13 = hbf1 v2 f2 , (i − 2j ≤ 2, j ≤ 0, k = −1)i 1 2 = V 2 , U12 ) are shown on Fig. 2. Some of these results (V 1 , U13

j Plane k=1

3

V

2

2

1 0 1

2

3

4

5

6

7

8

9

10

i

j -6

-5

-4

-3

-2

-1

1

2

i -1

V

1

2 U12

V

2

-2

Plane k=-1

Fig. 2. The set of polyhedra for the motivated example

212

4 4.1

M. Handjieva and S. Tzolovski

Finite Sets of Labeled Linear Congruences Linear Congruence Domain

The static analysis of linear congruence equalities among program variables was studied by P.Granger in [9]. We briefly recall some notions related to linear congruences. Let G be an abelian group and H be a subgroup of G. Any subset of G of the form: a + H = {x ∈ G | ∃h ∈ H, x = a + h} where a is an element of G is called coset of H in G. The element a is said to be a representative of the coset and H the unique modulo of the coset. For any element b of a + H, we have a + H = b + H. The set of all cosets in the abelian group G with the singleton {∅} is called the congruence lattice of G and is denoted as C(G) (or C when G is Z). The set inclusion order can be simplified by the following formula: (a1 + H1 ) ⊆ (a2 + H2 ) ≡ (a1 − a2 ∈ H2 ) ∧ (H1 ⊆ H2 ) Upper bound. The upper bound operations t in C(G) is (a1 + H1 ) t (a2 + H2 ) = a1 + Z(a1 − a2 ) + H1 + H2 where Za denotes the subgroup generated by a i.e. {. . . , −a − a, −a, 0, a, a + a, . . . }. In what follows we use Zn as a group G. Affine transformation. We use r(H) to denote the rank of H and k(u) the set {X ∈ Zn | u(X) = (0, . . . , 0)}. Let τD be an affine transformation on Zn , u be its linear part and a ∈ Zn , (ei )1≤i≤k be a basis of the subgroup H of Zn . The coset a + H is transformed by τD in the following manner: τD (a + H) =τD (a + Ze1 + · · · + Zek ) = τD (a) + Zu(e1 ) + · · · + Z(ek ) =τD (a) + Zu(H) where the rank of u(H) is equal to k − r(k(u) ∩ (H)). We have to mention that the congruence lattice satisfies the ascending chain condition, that guarantees the systematic termination of the analysis. This can be deduced from the fact that the complete lattice of all subgroups of Zn satisfies the ascending chain condition. Therefore no widening operators are needed for this domain. 4.2

Lifting the Domain of Linear Congruences

The domain of sets of labeled congruences is obtained by lifting the domain of linear congruences using the method described in Sect. 2. The abstract semantics for an imperative programming language with assignments, “if-then-else” and “while” statements is close to this in Subsect. 2.5. Only the “while” statement is treated in a different manner, because no widening is needed for this analysis. Therefore, the widening operator is simply replaced by the upper bound operator.

Refining Static Analyses by Trace-Based Partitioning Using Control Flow

4.3

213

Example

Let us now illustrate with help of the motivated example the analysis based on finite sets of labeled linear congruences. As in Subsect. 3.3 ∆ = (1, 0, 1). Before the while loop at point 6 of the program we have: * + * 2 2 + 2 b, U61 = bt1 , 0 , bf1 , 0 U11 = 0 1 −1 The first pass of the analysis through the “while” loop gives: * + * + 6 −2 0 0 bt v t t , , bf v t t , 1 2 2 3 1 2 2 3 1 −1 1 U12 = * + * + 4 0 1 −1 v t f , , bf v t f , bt 1 2 2 3 1 2 2 3 1 −1 The second pass produces the following set of labeled congruences: * + * + 2 42 2 42 1 = bt1 v2 , 0 + 0 1 Z2 , bf1 v2 , 0 + 0 1 Z2 V 2 = V 1 ∇U12 1 00 −1 00 2 2 2 = U11 = U12 =V2 U72 = U82 = U92 = U10

The analysis converges in the next iteration step (the third one), because V 3 = V 2 . The interpretation of the obtained result is the two sets - {(i, j, k) ∈ Z3 | i − 2j ≡ 2 [4], k ≡ 1[0]} and {(i, j, k) ∈ Z3 | i − 2j ≡ 2 [4], k ≡ −1[0]}. This is a more precise result than the following one obtained by the analysis based on linear congruences {(i, j, k) ∈ Z3 | i − 2j ≡ 2 [4], k ≡ 1 [2]}.

5

Conclusions and Further Work

We have shown how to lift an abstract domain to the finite sets of its labeled abstract properties using control flow information. The significance of this method is that it always produces a static analysis which performs as well as or better than the original one. We have shown how this method can be used with two currently available abstract domains - convex polyhedra and linear congruences. The ideas are in fact general enough to capture other abstract properties as interval congruences, congruence properties on rational numbers etc. The main contributions of this work are: – the design of a method that lifts an abstract domain to the finite sets of its labeled abstract properties as a particular instance of the reduced cardinal power [5], where the base is the set of labels approximating the control history and the exponent is the original domain;

214

M. Handjieva and S. Tzolovski

– the application of this method to the convex polyhedra domain and to the domain of linear congruences; An implementation is currently ongoing and it will allow us to investigate how the algorithms perform in practice. Acknowledgements. We would like to thank Patrick Cousot for pointing out the connection with the reduced cardinal power. We are grateful to our colleagues from LIX for helpful discussions.

References 1. F. Bourdoncle. Abstract interpretation by dynamic partitioning. Journal of Functional Programming, 2(4) (1992) 407-435. 2. P. Cousot. Semantic Foundations of Program Analysis. In Muchnick and Jones Eds. Program Flow Analysis, Theory and Applications, pp. 303-343, Prentice-Hall, 1981. 3. P. Cousot and R. Cousot. Static determination of dynamic properties of programs, In Proceedings of the 2nd Int. Symposium on Programming, pp. 106-130, 1976. 4. P. Cousot and R. Cousot. Abstract interpretation: a unified lattice model for static analysis of programs by construction or approximation of fixpoints. In Proceedings of the 4th ACM Symposium on Principles of Programming Languages, pp. 238-252, 1977. 5. P. Cousot, and R. Cousot, Systematic Design of Program Analysis frameworks. In Proceedings of the 6th ACM Symposium on Principles of Programming Languages, pp. 269-282, 1979. 6. P. Cousot, and R. Cousot, Abstract Interpretation and Application to Logic Programs. In Journal of Logic Programming, pp. 103–179, 1992. 7. P. Cousot and N. Halbwachs. Automatic discovery of linear restraints among variables of a program. In Proceedings of the 5th Annual ACM Symposium on Principles of Programming Languages, pp. 84-97, 1978. 8. G. Fil´e and F. Ranzato, Improving abstract interpretations by systematic lifting to the powerset. In Proceedings of the International Logic Programming Symposium, Ithaca, NY, pages 655-669. The MIT Press, 1994. 9. P. Granger, Static analysis of linear congruence equalities among variables of a program, In Proceedings of the International Joint Conference on Theory and Practice of Software Development, pp. 169-192, number 493 in LNCS, 1991. 10. N. Halbwachs and Y.-E. Proy and P. Raymond, Verification of linear hybrid systems by means of convex approximations, In Proceedings of the International Static Analysis Symposium, pp. 223-237, number 864 in LNCS, 1994. 11. F. Masdupuy. Array operations abstractions using semantics analysis of trapezoid congruences. In Proceedings of the International Conference on Supercomputing, Washington, 1992. 12. A. Schrijver, Theory of Linear and Integer Programming, John Wiley & Sons, 1986.

Building Complete Abstract Interpretations in a Linear Logic-Based Setting Roberto Giacobazzi† †

‡

Francesco Ranzato‡,

Francesca Scozzari†

Dipartimento di Informatica, Universit` a di Pisa Corso Italia 40, 56125 Pisa, Italy {giaco,scozzari}@di.unipi.it

Dipartimento di Matematica Pura ed Applicata, Universit` a di Padova Via Belzoni 7, 35131 Padova, Italy [email protected]

Abstract. Completeness is an important, but rather uncommon, property of abstract interpretations, ensuring that abstract computations are as precise as possible w.r.t. concrete ones. It turns out that completeness for an abstract interpretation depends only on its underlying abstract domains, and therefore it is an abstract domain property. Recently, the ﬁrst two authors proved that for a given abstract domain A, in all significant cases, there exists the most abstract domain, called least complete extension of A, which includes A and induces a complete abstract interpretation. In addition to the standard formulation, we introduce and study a novel and particularly interesting type of completeness, called observation completeness. Standard and observation completeness are here considered in the context of quantales, i.e. models of linear logic, as concrete interpretations. In this setting, we prove that various kinds of least complete and observationally complete extensions exist and, more importantly, we show that such complete extensions can all be explicitly characterized by elegant linear logic-based formulations. As an application, we determine the least complete extension of a generic abstract domain w.r.t. a standard bottom-up semantics for logic programs observing computed answer substitutions. This general result is then instantiated to the relevant case of groundness analysis.

1

Introduction

It is widely held that the ideal goal of any semantics design method is to ﬁnd sound and complete representations for some properties of concrete (actual) computations. Abstract interpretation is one such methodology, where soundness is always required, while completeness more rarely holds. Completeness issues in abstract interpretation have been studied since the Cousot and Cousot seminal paper [5]. The intuition is that a complete abstract interpretation induces an

The work of Francesco Ranzato has been supported by an individual post-doctoral grant from Universit` a di Padova, Italy.

G. Levi (Ed.): SAS’98, LNCS 1503, pp. 215–229, 1998. c Springer-Verlag Berlin Heidelberg 1998

216

R. Giacobazzi, F. Ranzato, and F. Scozzari

abstract semantics which is as precise as possible relatively to its underlying abstract domains and to the concrete interpretation of reference. The paradigmatic rule of signs is a typical and simple example of an abstract interpretation which is (sound and) complete for integer multiplication but merely sound for integer addition [13]. Although in static program analysis decidability issues often force to sacriﬁce completeness for achieving termination and/or eﬃciency, examples of complete abstract interpretations are common in other ﬁelds of application. For instance, several complete abstractions of algebraic polynomial systems have been studied by Cousot and Cousot in [7] and many complete abstract interpretations can be found in comparative program semantics [3,6,9]. Moreover, being completeness a notion relative to the concrete semantics of reference, complete abstract interpretations which are more concrete than a certain, possibly approximated and decidable, property of interest, yield an absolute upper bound for the precision that one can achieve in computing that property. Thus, complete abstract interpretations may play a useful rˆ ole in static program analysis as well. These argumentations probably stimulated the recent trend of research on completeness in abstract interpretation [2,10,11,13,15,17,18]. One key feature of completeness in abstract interpretation is that this property uniquely depends upon the abstraction function. Let us denote by LC the so-called lattice of abstract interpretations of a concrete domain C (cf. [4,5]), where, for all A, B ∈ LC , A B means that A is more precise (i.e. concrete) than B. Let us consider the simple case of an abstract interpretation f : A → A of a concrete semantic function f : C → C, where the abstract domain A ∈ LC is related to C by an adjoint pair of abstraction and concretization maps α : C → A and γ : A → C. Then, f is (sound and) complete if α ◦ f = f ◦ α. It is easily seen that if f is complete then the best correct approximation f b of f in A, i.e. def f b = α ◦ f ◦ γ : A → A, is complete as well, and, in this case, f indeed coincides with f b (cf. [10]). Thus, given an abstract domain A ∈ LC , one can deﬁne a complete abstract semantic function f : A → A over A if and only if f b : A → A is complete. This simple observation makes completeness an abstract domain property, namely a characteristic of the abstract domain. It is then clear that a key problem consists in devising systematic and constructive methodologies for transforming abstract domains in such a way that completeness is achieved and the resulting complete abstract domains are as close as possible to the initial (noncomplete) ones. This problem has been ﬁrst raised in a predicate-based approach to abstract interpretation by Mycroft [13, Section 3.2], who gave a methodology for deriving the most concrete domain, called “canonical abstract interpretation”, which is complete and included in a given domain of properties. More recently, the ﬁrst two authors proved in [10] that when the concrete semantic function f is continuous, any domain A can always be extended into the most abstract domain which includes A and is complete for f — the so-called least complete extension of A. Analogously, [10] solved the aforementioned problem raised by Mycroft, by showing that, for any given domain A, the most concrete domain which is complete and included in A — the so-called complete kernel

Building Complete Abstract Interpretations in a Linear Logic-Based Setting

217

of A — exists for any monotone semantic function. Very recently, we improved considerably such results, by providing explicit constructive characterizations for least complete extensions and complete kernels of abstract domains [11]. In this paper, we are concerned with completeness problems arising when concrete semantic binary1 operations of type C1 × C2 → C are assumed to give rise to a generalized form of quantales, called typed quantales. Quantales are well-known algebraic structures which turn out to be models of intuitionistic linear logic [16,19]. In this logical setting, we provide elegant linear logic-based solutions to a number of interesting completeness problems for abstract interpretations. Such solutions ﬁnd relevant applications in static program analysis and comparative semantics, for instance in logic programming, where uniﬁcation — the prime computational step of any logic program semantics — turns out to be a binary operation in a quantale of substitutions, or in data structure analysis, considering binary data constructors such as cons for lists. More in detail, a typed quantale C, C1 , C2 , ⊗ consists of three complete lattices C, C1 and C2 , and of an operation ⊗ : C1 × C2 → C which is additive (i.e., preserves lub’s) on both arguments. When C = C1 = C2 , typed quantales boil down to standard quantales C, ⊗. The main feature of (typed) quantales is that they support a notion of left and right linear implication between domain’s objects: Given a ∈ C1 and b ∈ C, there exists a unique greatest object a b ∈ C2 which, when combined by ⊗ with a, gives a result less than or equal to b. In other terms, the b holds. Analogous left following right modus ponens law a ⊗ x ≤ b ⇔ x ≤ a implicational objects exist for a correspoding left modus ponens law. When solving completeness problems in a setting where concrete interpretations are typed quantales, implicational domain objects allow to elegantly characterize complete abstract domains in a variety of situations. Our ﬁrst result provides a characterization based on linear implications between domain’s objects of the least complete extension of any abstract domain of any quantale. Then, we consider the following completeness problem over typed quantales: Given a typed quantale C, C1 , C2 , ⊗, a ﬁxed abstraction A ∈ LC , with corresponding abstraction map αA : C → A, and a pair of abstract domains A1 , A2 ∈ LC1 × LC2 , does there exist the most abstract pair of domains A1 , A2 ∈ LC1 × LC2 , with corresponding abstraction maps αAi : Ci → Ai (i = 1, 2), such that A1 , A2 L C1 ×L C2 A1 , A2 and αA ( ⊗ ) = αA (αA1 ( ) ⊗ αA2 ( ))? Here, the observation domain A is ﬁxed, and we are thus looking for the most abstract pair of domains in LC1 × LC2 which is more concrete than an initial pair A1 , A2 and simultaneously induces a complete abstract interpretation w.r.t. ⊗. This is termed an observation completeness problem. Again, solutions to this observation completeness problem are built in terms of linear implications between domains. To illustrate the practical scope of our results, we ﬁrst consider a simple example in data structure analysis involving abstract domains for lists. In particular, our results are applied in order to solve various observation completeness problems concerning abstract domains useful for detecting irredundant lists of 1

Clearly, a generalization to n-ary operations would be straightforward.

218

R. Giacobazzi, F. Ranzato, and F. Scozzari

objects. Then, in the context of logic program semantics, we consider an immediate consequences operator TP deﬁned in terms of uniﬁcation and union of sets of idempotent substitutions, and characterizing computed answer substitutions in a s-semantics style (cf. [1,8]). As usual, uniﬁcation turns out to be the key operation to take into account in order to build least complete extensions of abstract domains. Sets of idempotent substitutions and uniﬁcation give rise to a unital commutative quantale: Given an abstract domain A, we show how the least complete extension of A w.r.t. this quantale naturally induces the least complete extension of A w.r.t. TP functions. This permits to give explicitly, in terms of linear implications, the least complete extension, for any TP , of a generic domain abstracting sets of substitutions. As a remarkable instance of our construction, we characterize the least complete extension of the plain groundness domain w.r.t. computed answer substitutions s-semantics.

2

Basic Notions

The lattice of abstract interpretations. In standard Cousot and Cousot’s abstract interpretation theory, abstract domains can be equivalently speciﬁed either by Galois connections (GCs), i.e. adjunctions, or by upper closure operators (uco’s) [5]. In the ﬁrst case, the concrete and abstract domains C and A (both assumed to be complete lattices) are related by a pair of adjoint functions of a GC (α, C, A, γ). Also, it is generally assumed that (α, C, A, γ) is a Galois insertion (GI), i.e. α is onto or, equivalently, γ is 1-1. In the second case instead, an abstract domain is speciﬁed as a uco on the concrete domain C, i.e. a monotone, idempotent and extensive operator on C. These two approaches are equivalent, modulo isomorphic representation of domain’s objects. Given a complete lattice C, it is well known that the set uco(C) of all uco’s on C, endowed with the pointwise ordering , is a complete lattice uco(C), , , , λx. C , id (id denotes the identity function). Let us also recall that each ρ ∈ uco(C) is uniquely determined by the set of its ﬁxpoints, which is its image, i.e. ρ(C) = {x ∈ C | ρ(x) = x}, and that ρ η iﬀ η(C) ⊆ ρ(C). Moreover, a subset X ⊆ C is the set of ﬁxdef points of a uco on C iﬀ X is meet-closed, i.e. X = (X) = { C Y | Y ⊆ X} (note that C = C ∈ (X)). Often, we will identify closures with their sets of ﬁxpoints. This does not give rise to ambiguity, since one can distinguish their use as functions or sets according to the context. In view of the equivalence above, throughout the paper, uco(C), will play the rˆ ole of the lattice LC of abstract interpretations of C [4,5], i.e. the complete lattice of all possible abstract domains of the concrete domain C. For an abstract domain A ∈ LC , ρA ∈ uco(C) will denote the corresponding uco on C, and if A is speciﬁed by def a GI (α, C, A, γ) then ρA = γ ◦ α. The ordering on uco(C) corresponds to the standard order used to compare abstract domains with regard to their precision: A1 is more precise than A2 (i.e., A1 is more concrete than A2 or A2 is more abstract than A1 ) iﬀ A1 A2 in uco(C). Lub and glb on uco(C) have therefore the following reading as operators on domains. Suppose {Ai }i∈I ⊆ uco(C): (i) i∈I Ai is the most concrete among the domains which are abstractions of all

Building Complete Abstract Interpretations in a Linear Logic-Based Setting

219

the Ai ’s, i.e. it is their least (w.r.t. ) common abstraction; (ii) i∈I Ai is the most abstract among the domains (abstracting C) which are more concrete than every Ai ; this domain is also known as reduced product of all the Ai ’s. Quantales and linear logic. Quantales originated as algebraic foundations of the so-called quantum logic. They have been successively considered for the latticetheoretic semantics of Girard’s linear logic (see [16] for an exhaustive treatment of quantales). We introduce a mild generalization of the notion of quantale, which, up to knowledge, appears to be new. A typed quantale is a multisorted and ⊗ : C1 ×C2 → C algebra C, C1 , C2 , ⊗, where C, C1 , C2are complete lattices is a function such that ( i xi ) ⊗ c2 = i (xi ⊗ c2 ) and c1 ⊗ ( i xi ) = i (c1 ⊗ xi ). In other terms, a typed quantale is a 3-sorted algebra endowed with a “product” ⊗ which distributes over arbitrary lub’s on both sides. Thus, for any c1 ∈ C1 and c2 ∈ C2 , both functions c1 ⊗ and ⊗ c2 have right adjoints denoted, resp., by c1 and c2 . Hence, for all c ∈ C, c1 ⊗ c2 ≤ c ⇔ c2 ≤ c1 c, and, dually, c1 ⊗ c2 ≤ c ⇔ c1 ≤ c c2 . Two functions : C1 × C → C2 and : C × C2 → C1 can be therefore deﬁned as follows: def def c1 c= {z ∈ C2 | c1 ⊗ z ≤ c}; c c2 = {y ∈ C1 | y ⊗ c2 ≤ c}.

Any typed quantale C, C1 , C2 , ⊗ enjoys the following main properties: For all c ∈ C, {xi }i∈I ⊆ C, c1 ∈ C1 , {yi }i∈I ⊆ C1 , c2 ∈ C2 and {zi }i∈I ⊆ C2 : (i) (iii) (v)

c) ≤ c c1 ⊗ (c1 c1 ( xi ) = (c1 (

i∈I

i∈I

yi )

(ii) xi )

(iv)

i∈I

c=

(c c2 ) ⊗ c2 ≤ c ( xi ) c2 = (xi

i∈I

(yi

c)

i∈I

(vi)

c

c ) 2

i∈I

( z ) = (c z ) i

i∈I

i

i∈I

When C = C1 = C2 and ⊗ is associative, a typed quantale is called quantale. It is well known that quantales turn out to be models of noncommutative intuitionistic linear logic [16,19]. A quantale C, ⊗ is called commutative when ⊗ is comb=b a. mutative, and this is equivalent to require that, for all a, b ∈ C, a Also, a commutative quantale C, ⊗ is called unital if there exists an object 1 ∈ C such that 1 ⊗ a = a = a ⊗ 1, for all a ∈ C. For a quantale C, ⊗, the following additional properties hold for all a, b, c ∈ C:

(vii) (viii) (ix)

3

b) = (a c) b b (a c) = (a ⊗ b) c (c b) a = c (a ⊗ b)

a

(c

Completeness Problems in Abstract Interpretations

Let C, C1 , C2 , ⊗ be a concrete interpretation, i.e. C, C1 and C2 are concrete semantic domains provided with a semantic operation ⊗ : C1 × C2 → C. When

220

R. Giacobazzi, F. Ranzato, and F. Scozzari

C = C1 = C2 , we adopt the simpler notation C, ⊗. Given the abstractions A1 ∈ LC1 , A2 ∈ LC2 and A ∈ LC , let us recall [5] that the best correct approxdef imation ⊗b : A1 × A2 → A of ⊗ is deﬁned as ⊗b = αC,A ◦ ⊗ ◦ γA1 ,C1 , γA2 ,C2 . It has been shown in [10] that completeness for an abstract interpretation is a property depending only on the underlying abstract domains. In our setting, this means that an abstract interpretation A, A1 , A2 , ⊗ , with ⊗ : A1 × A2 → A, is complete for C, C1 , C2 , ⊗ iﬀ A, A1 , A2 , ⊗b is complete and ⊗ = ⊗b . In other terms, the best correct approximation induced by the underlying abstract domains determines the property of being complete. Hence, we ﬁnd more convenient, elegant and, of course, completely equivalent, to reason on completeness by using abstract domains speciﬁed by the closure operator approach. Full completeness problems. Within the closure operator approach, given a concrete interpretation C, ⊗, an abstraction A ∈ LC is complete when the equation ρA ◦⊗◦ρA , ρA = ρA ◦⊗ holds. Giacobazzi and Ranzato [10] stated the following full completeness problem, here specialized to binary semantic operations: Given an abstract domain A ∈ LC , does the following system with variable ρ admit a most abstract solution? ρA (1) ρ ◦ ⊗ ◦ ρ, ρ = ρ ◦ ⊗ Hence, a solution to the above full completeness problem is the necessarily unique (up to domain isomorphism) most abstract domain which includes A and induces a complete abstract interpretation. Following [10], such most abstract solution to System (1) is called the least complete extension of A w.r.t. ⊗. It is shown in [10] that if ⊗ is continuous in both arguments, then least complete extensions of any A exist. Full completeness problems clearly make sense only for concrete interpretations of type C, ⊗. When generic concrete interpretations of type C, C1 , C2 , ⊗ are considered, diﬀerent abstractions are involved in a completeness equation, and therefore various completeness problems arise by ﬁxing some of these abstractions. In the following, we introduce three such completeness problems, which turn out to be of particular interest. Such completeness problems still depend only on best correct approximations of the concrete operation, and therefore the corresponding completeness notions are again abstract domain properties. Observation completeness problems. Let C, C1 , C2 , ⊗ be a concrete interpretation. An observation domain is any abstraction of the range C of ⊗. Observation completeness problems arise when in a completeness equation an observation domain is ﬁxed. Hence, let us consider a ﬁxed observation domain A ∈ LC . The observation completeness problem for a pair A1 , A2 ∈ LC1 × LC2 admits solution when there exists the most abstract solution in LC1 × LC2 of the following system: η, µ A1 , A2 (2) ρA ◦ ⊗ ◦ η, µ = ρA ◦ ⊗

Building Complete Abstract Interpretations in a Linear Logic-Based Setting

221

Let us remark that, by using adjunctions, the observation completeness equation ρA ◦ ⊗ ◦ η, µ = ρA ◦ ⊗ is equivalent to require that for all x ∈ C1 and y ∈ C2 , αC1 ,η (x) ⊗b αC2 ,µ (y) = αC,A (x ⊗ y). When in addition to the observation domain we also ﬁx one (or more, if we would deal with n-ary operations) of the abstractions of the argument domains of ⊗, we obtain yet diﬀerent completeness problems. In the left observation completeness problem, A ∈ LC and A2 ∈ LC2 are ﬁxed, and the solution to this problem for a given A1 ∈ LC1 exists when the following system with variable η admits a most abstract solution in LC1 : η A1 (3) ρA ◦ ⊗ ◦ η, ρA2 = ρA ◦ ⊗ ◦ id, ρA2 Of course, right observation completeness problems are analogously formulated. It turns out that a left observation completeness equation ρA ◦ ⊗ ◦ η, ρA2 = ρA ◦ ⊗ ◦ id, ρA2 formulated in terms of GIs amounts to require that for any x ∈ C1 and y ∈ A2 , αC1 ,η (x) ⊗b y = αC,A (x ⊗ γA2 ,C2 (y)). Hence, a left observation completeness equation ρA ◦ ⊗ ◦ η, ρA2 = ρA ◦ ⊗ ◦ id, ρA2 states that completeness for the abstractions A ∈ LC , η ∈ LC1 and A2 ∈ LC2 holds when the semantic operation ⊗ acts over C1 × A2 . Examples of observation completeness problems will be considered and solved later in Section 5.

4

Quantales and Solutions to Completeness Problems

When concrete interpretations are quantales and typed quantales, solutions to the above completeness problems exist and can be characterized explicitly and elegantly in terms of linear implications. Let C, C1 , C2 , ⊗ be a typed quantale playing the rˆ ole of concrete interpretation. In this setting, solutions to completeness problems will be characterized by exploiting two basic domain transformers : uco(C1 ) × uco(C) → uco(C2 ) and : uco(C) × uco(C2 ) → uco(C1 ), deﬁned by lifting left and right linear and to abstract domains as follows: For any A1 ∈ LC1 , A2 ∈ implications LC2 , A ∈ LC :

({a a ∈ C A A = ({a a ∈ C A1

def

A=

1

2

| a1 ∈ A1 , a ∈ A});

1

| a ∈ A, a2 ∈ A2 }).

def

2

2

Hence, A1 A is deﬁned to be the most abstract domain in LC2 containing all the linear implications from A1 to A. From the logic properties of linear implication recalled in Section 2, it is not too hard to derive the following useful distributivity laws for and over reduced product of abstract domains.

Proposition 1. For all {Bi }i∈I ⊆ LC , A1

(

i∈I

Bi ) =

i∈I (A1

Bi )

and

(

i∈I

Bi )

A = 2

i∈I (Bi

A ). 2

222

R. Giacobazzi, F. Ranzato, and F. Scozzari

Solutions to full completeness problems. Solutions to full completeness problems exist and are characterized in terms of linear implications among domain’s objects as stated by the following result. Theorem 2. A(C of System (1).

A)(A

C)((C A) C) is the most abstract solution

Solutions to observation completeness problems. Let us ﬁrst consider left observation completeness problems. A left observation completeness equation can be characterized as follows. Theorem 3. ρA ◦ ⊗ ◦ η, ρA2 = ρA ◦ ⊗ ◦ id, ρA2 ⇔ η A

A . 2

Thus, as an immediate consequence, left observation completeness problems admit the following solutions. Corollary 4. A1 (A

A ) is the most abstract solution of System (3). 2

Of course, dual results can be given for right observation completeness problems. In this case, the most abstract solution therefore is A2 (A1 A). The above results for left and right observation completeness turn out to be useful for solving observation completeness problems. In fact, an observation completeness equation is characterized as an independent combination of left and right observation completeness equations as follows. Theorem 5. ρA ◦ ⊗ ◦ η, µ = ρA ◦ ⊗ ⇔ η, µ A

C , C 2

1

A.

As a straight consequence, we get the following result. Corollary 6. A1 (A System (2). 4.1

C ), A 2

2

(C1

A) is the most abstract solution of

The Case of Unital Commutative Quantales

When we deal with unital and commutative quantales — i.e. models of intuitionistic linear logic [16,19] — the above solutions to full completeness problems given by Theorem 2 can be signiﬁcantly simpliﬁed by exploiting the logical properties of linear implication. In a unital commutative quantale C, ⊗, the following additional properties hold: For any a, b, c ∈ C: – a (b – c ≤ (c

c) = b a) a

(a

c) – 1 a=a – ((c a) a)

a=c

a

Therefore, from these properties it is not hard to check that for all a ∈ C, λc.(c a) a ∈ uco(C). This turns out to be the key observation in order to give a more compact form to solutions of full completeness problems on unital commutative quantales. Moreover, the objects of such solutions enjoy a clean logical characterization in terms of linear implications as speciﬁed by the third point of the next result.

Building Complete Abstract Interpretations in a Linear Logic-Based Setting

223

Theorem 7. Let C, ⊗ be a unital commutative quantale.

1. C A is the most abstract solution of System (1); 2. If B ∈ LC is such that B C A, then C A = B A; 3. For all c ∈ C, c ∈ C A ⇔ c = a∈A (c a) a.

5

An Application in Data Structure Completeness

In this section, we consider some examples of completeness problems in abstract interpretation of list data structures. For the sake of practicality, we consider lists of natural numbers, even if our discussion holds more in general for lists of objects of arbitrary type. Consider the structure ℘(list( )), ℘( ), ℘(list( )), ::, where list( ) is the set of all ﬁnite lists of natural numbers, ℘(list( )) and ℘( ) are complete lattices w.r.t. set-inclusion, and :: : ℘( ) × ℘(list( )) → ℘(list( )) is a “collecting” version of concatenation deﬁned as follows: def

N :: L = {[n|l] | n ∈ N, l ∈ L}, where :: L = N :: = . It is clear that this structure is a typed quantale. We say that a list is irredundant if it does not contain two occurrences of the same object. An abstract domain ρ ∈ uco(℘(list( ))) for detecting irredundant def lists can be deﬁned by ρ = { list(), Irr }, where Irr ⊆ list( ) is the set of irredundant lists over . We consider ρ as an observation domain and we look for the most abstract solution X, Y ∈ uco(℘( )) × uco(℘(list( ))) to the following observation completeness problem: ρ ◦ :: ◦ ρX , ρY = ρ ◦ :: Here, we dropped the ﬁrst constraint of the generic System (2), since it amounts to the trivial constraint ρX , ρY { }, { list() } which is always satisﬁed. By Corollary 6, the most abstract solution of this observation completeness equation exists. This will be the most abstract pair of domains X, Y for which abstract concatenation in X and Y results to be complete when observing irredundancy as represented by ρ. By Corollary 6, the solution X, Y is as follows. X = { } (ρ

℘(list()))

℘(list( )) = ({L M | L ∈ ρ, M ∈ ℘(list( ))}) (since, for all M , list( ) M = list( )) = ({Irr M | M ∈ ℘(list( ))}) (since, for all n ∈ , Irr {[n]} = {n})

=ρ

= ℘( )

224

R. Giacobazzi, F. Ranzato, and F. Scozzari

Y = { list( ) } (℘( ) = ℘( )

ρ)

= ({N

ρ

= ({N

(since, for all N , N

L | N ∈ ℘( ), L ∈ ρ}) Irr | N ∈ ℘( )})

list( ) = list( ))

(since, by (v) in Section 2,

i (Ni

Irr) = ( i Ni )

Irr)

= {N Irr | N ∈ ℘( )} = N ⊆ {L ∈ ℘(list( )) | l ∈ L ⇔ (l ∈ Irr and ∀n ∈ N. n is not in l)} Thus, in order to be complete for concatenation when observing irredundancy, X must coincide with the concrete domain ℘( ), while it suﬃces that Y contains all the sets of irreduntant lists which do not contain some set of numbers. Note that Y coincides with the set of all the sets of irredundant lists closed by permutation of their objects, and this is a strict abstraction of the concrete domain ℘(list( )). Let us now consider the standard abstract domain η ∈ uco(℘( )) for pardef ity analysis given by η = { , even, odd, }. We consider the following left and right observation completeness problems: We look respectively for the the most abstract domains X ∈ uco(℘( )) and Y ∈ uco(℘(list( ))) such that: (i) ρ ◦ :: ◦ ρX , ρ = ρ ◦ :: id, ρ

(ii) ρ ◦ :: ◦ η, ρY = ρ ◦ :: η, id

Here again, there are no upper bound constraints for X and Y . By Corollary 4 (and its dual), we get the following solutions:

X = { } (ρ ρ) =ρ ρ = ({ list( ) list( ), list( ) = { , },

Irr, Irr list(), Irr Irr })

Y = { list() } (η ρ) =η ρ Irr, even Irr, odd Irr, = ({ = { { [] }, Irreven , Irrodd , list( ) },

Irr })

where Irreven = {l ∈ list( ) | l ∈ Irr and l does not contain even numbers} and def Irrodd = {l ∈ list( ) | l ∈ Irr and l does not contain odd numbers}. Thus, for problem (i), in order to get completeness, it is enough to check whether a given set of numbers is empty or not, while, for problem (ii), we only need to consider sets of irreduntant lists which do not contain either even or odd numbers. def

6

Complete Semantics for Logic Program Analysis

In this section, we determine the least complete extension of any logic program property w.r.t. a bottom-up semantics characterizing computed answer substitutions, which turns out to be equivalent to the well-known s-semantics [8]. This

Building Complete Abstract Interpretations in a Linear Logic-Based Setting

225

complete semantics can be thought of as a logic program semantics which is “optimal” (i.e. neither too concrete nor too abstract) for characterizing a given property of interest. The problem of determining optimal semantics for program analysis was raised in [9]. We show that such optimal semantics can be obtained by solving a full completeness problem relatively to the operation of uniﬁcation on sets of substitutions. Completeness for an abstract domain, in general, depends both on the considered concrete domains and on the semantic operation deﬁned on them. However, the following important result shows that, under certain conditions, completeness is instead independent from the choice of the concrete semantics. Theorem 8. Let α ∈ LC and f, g : C → C such that f g α ◦ f . Then, α is complete for f iﬀ α is complete for g. We will exploit this result for computing our least complete extension of abstract domains w.r.t. s-semantics. The idea is that of considering a “simpliﬁed” and more concrete version, denoted by TP , of the immediate consequences operator TPs of s-semantics, which does not take into account variable renaming. This will simplify a lot the technical development of this section. Hence, s-semantics results to be an abstract interpretation of our TP semantics: If r denotes the closure under variable renaming of sets of substitutions, then we have that TPs = r ◦ TP . Then, leaving out the details, if α denotes the least complete extension, relatively to TP , of any domain A abstracting computed answer substitutions, since TP r ◦ TP = TPs α ◦ TP holds, we can apply Theorem 8, from which we get that α not only is complete for TPs , but actually α turns out to be the least complete extension of A relatively to TPs . 6.1

Notation

Let V be an inﬁnite, recursively enumerable (r.e.) set of variables, Σ be a set of function symbols and Π be a set of predicate symbols, deﬁning a r.e. ﬁrst-order language L. Term denotes the set of terms of L. If s is any syntactic object and σ and θ are substitutions, then vars(s) denotes the set of variables occurring def def in s, dom(σ) = {v ∈ V | σ(v) = v}, rng(σ) = ∪ {vars(σ(x)) | x ∈ dom(σ)}, sσ denotes the application of σ to s, and σ ◦ θ denotes the standard composition of θ and σ (i.e., σ ◦ θ = λx.(θ(x))σ). The set of idempotent substitutions modulo renaming (i.e., θ ∼ σ iﬀ there exists β and δ such that θ = σ ◦ β and σ = θ ◦ δ) on L is denoted by Sub. Objects in Sub are partially ordered by instantiation, denoted by . By adding to Sub an extra object τ as least element, one gets a complete lattice Subτ , , ∨, ∧, , τ , where ∨ is least general anti-instance, ∧ is standard uniﬁcation and is the empty substitution (see [14] for more details). def ¯ | p ∈ Π}, where X ¯ is The set of most general atoms is given by GAtom = {p(X) a tuple of distinct variables. We consider logic programs in normalized form, that ¯ 1 ), . . . , qn (X ¯ n ), where all the tuples of ¯ : −c, q1 (X is, a generic Horn clause is p(X) variables are distinct and c ∈ Sub is an idempotent substitution binding variables to terms.

226

6.2

R. Giacobazzi, F. Ranzato, and F. Scozzari

TP -Completeness

Our basic semantic structure is the unital commutative quantale ℘(Sub), ⊗, where ℘(Sub), ⊆ is a complete lattice, ⊗ : ℘(Sub) × ℘(Sub) → ℘(Sub) is the obvious lifting of uniﬁcation ∧ to sets of substitutions, i.e. it is deﬁned def by: X ⊗ Y = {x ∧ y | x ∈ X, y ∈ Y }, and {} ∈ ℘(Sub) is the unit of ⊗. It is immediate to check that ℘(Sub), ⊗ actually is a unital commutative quantale [16, Example 10, p. 18]. In the following, we will slightly abuse notation by applying the operation ⊗ also to substitutions. As mentioned above, we consider a bottom-up semantics based on an immediate consequences operator TP which is more concrete than the standard operator TPs of s-semantics. In fact, TP is deﬁned using only the operations of uniﬁcation ⊗ and union of sets of idempotent substitutions. The s-semantics operator TPs can be recovered from TP by a simple step of abstract interpretation considering the closure of sets of substitutions over variables renaming. We consider a concrete semantic domain CInt of functions — as usual, called interpretations — which map most general atoms to sets of substitutions: def CInt = GAtom → ℘(Sub), which ordered pointwise is trivially a complete lattice. Often, we will ﬁnd convenient to denote an interpretation I ∈ CInt by the ¯ I(p(X)) ¯ set {p(X), | p ∈ Π}. Then, for any program P , the immediate consequences operator TP : CInt → CInt is deﬁned as follows: For any I ∈ CInt, ¯ (c ⊗ ( I(qi (X¯i ))) ⊗ {Y¯ = X}), TP (I)(p(Y¯ )) = C

P

i=1..n

¯ : −c, q1 (X¯1 ), . . . , qn (X¯n ). Here, C where C = p(X) P denotes that the clause ¯i )) is intended modulo C of P is renamed apart with fresh variables, and I(qi (X renaming. If ρ ∈ uco(℘(Sub)) is any abstraction on sets of substitutions, a corresponding abstraction on interpretations ρ ∈ uco(CInt) which acts accordingly to ρ can def ¯ ρ(c) | p(X), ¯ c ∈ I}. be deﬁned as follows: For any I ∈ CInt, ρ (I) = {p(X), Note that · is monotone, i.e., for all ρ, η ∈ uco(℘(Sub)), if ρ η then ρ η .

Given a basic abstract domain of properties of interest π ∈ uco(℘(Sub), ⊆), our goal is therefore to ﬁnd the most abstract domain which contains π (more precisely, π ) and is complete for any TP operator. Our strategy consists in characterizing the least complete extension of π for the basic operations involved in the deﬁnition of TP , and then to show that, under reasonable hypotheses, this domain turns out to be the right one. Since every abstract domain is trivially always complete for lub’s, it turns out that union of sets of substitutions is troubleless, and therefore it is enough to concentrate on uniﬁcation ⊗. Indeed, the following result shows that completeness for ⊗ implies completeness for any TP .

Theorem 9. Let ρ ∈ uco(℘(Sub)). If ρ is complete for ⊗ then ρ is complete for any TP .

Building Complete Abstract Interpretations in a Linear Logic-Based Setting

227

By Theorem 7, given π ∈ uco(℘(Sub)), the least complete extension of π w.r.t. the quantale operation ⊗ exists and is the domain ℘(Sub) π of linear implications from ℘(Sub) to π. This complete domain ℘(Sub) π results to be the right one whenever the abstract domain π satisﬁes the following weak decidability property. Definition 10. π ∈ uco(℘(Sub)) is decidable if any S ∈ π is a r.e. set.

¾

It should be clear that decidability is a reasonable requirement for most abstract domains used in program analysis: In fact, for such a decidable abstraction (of sets) of substitutions, an eﬀective procedure for checking whether a substitution belongs to (is approximated by) an abstract object is available. As announced, the following key result shows that least complete extensions of decidable abstract domains w.r.t. uniﬁcation ⊗ actually are least complete extensions for TP operators as well.

Theorem 11. Let π ∈ uco(℘(Sub)) be decidable. Then, ℘(Sub) least complete extension of π for any TP .

π is the

As we discussed just after Theorem 8, it turns out that, for any decidable π ∈ uco(℘(Sub)), ℘(Sub) π actually is the least complete extension of π for any immediate consequences operator TPs of s-semantics. In fact, if r ∈ uco(℘(Sub)) is the closure under renaming of sets of substitions, namely variables in the range of substitutions are renamed in each possible way, then TPs = r ◦ TP ; moreover, ℘(Sub) π clearly induces a semantic transformer less precise than TPs , i.e. s TP ℘(Sub) π ◦ TP . Hence, by Theorems 8 and 11, we get the following desired consequence.

Corollary 12. Let π ∈ uco(℘(Sub)) be decidable, such that TPs π ◦ TP for any P . Then, ℘(Sub) π is the least complete extension of π for any TPs .

6.3

Complete Semantics for Groundness Analysis

Groundness analysis is arguably one of the most important analysis for logicbased programming languages. Groundness analysis aims to statically detect whether variables will be bound to ground terms in successful derivations. By instantiating the results above, we are able to characterize the least complete extension of the basic abstract domain representing plain groundness information w.r.t. any immediate consequences operator of s-semantics. The resulting semantics can be therefore interpreted as the “optimal” semantics for groundness analysis. If V ⊆ V is a ﬁnite set of variables of interest, the simplest abstract domain for def representing plain groundness information of variables in V is GV = ℘(V ), ⊇, as ﬁrst put forward by Jones and Søndergaard [12]. The intuition is that each W ∈ GV represents the set of substitutions which ground every variable in W . GV is related to the concrete domain ℘(Sub) by the following concretization map: def For each W ∈ G, γG V (W ) = {θ ∈ Sub | ∀v ∈ W. vars(θ(v)) = }. As usual, we

228

R. Giacobazzi, F. Ranzato, and F. Scozzari

shall abuse notation by denoting with GV ∈ uco(℘(Sub)) the corresponding isomorphic image γG V (GV ) in ℘(Sub). A variable independent abstract domain G ∈ uco(CInt) for representing plain groundness information can be therefore deﬁned as follows: def

¯ GX¯ (c) | p(X), ¯ c ∈ I}. G(I) = {p(X), It is easy to check that G actually is a uco on the concrete semantic domain CInt. Furthermore, G is clearly decidable and TPs G ◦ TP . Thus, as an easy consequence of Corollary 12, we can prove the following result, where V ⊂f V means that V is a ﬁnite subset of V. Theorem 13. any TPs .

V ⊂f V

℘(Sub) G is the least complete extension of G for V

References 1. A. Bossi, M. Gabbrielli, G. Levi, and M. Martelli. The s-semantics approach: theory and applications. J. Logic Program., 19-20:149–197, 1994. 218 2. P. Cousot. Completeness in abstract interpretation (Invited Talk). In Proc. 1995 Joint Italian-Spanish Conference on Declarative Programming, pages 37–38, 1995. 216 3. P. Cousot. Constructive design of a hierarchy of semantics of a transition system by abstract interpretation (Invited Paper). In Proc. of the 13th Int. Symp. on Math. Found. of Programming Semantics (MFPS’97), vol. 6 of Electronic Notes in Theor. Comput. Sci., 1997. 216 4. P. Cousot and R. Cousot. Abstract interpretation: a uniﬁed lattice model for static analysis of programs by construction or approximation of ﬁxpoints. In Proc. 4th ACM POPL, pages 238–252, 1977. 216, 218 5. P. Cousot and R. Cousot. Systematic design of program analysis frameworks. In Proc. 6th ACM POPL, pages 269–282, 1979. 215, 216, 218, 218, 220 6. P. Cousot and R. Cousot. Inductive deﬁnitions, semantics and abstract interpretation. In Proc. 19th ACM POPL, pages 83–94, 1992. 216 7. P. Cousot and R. Cousot. Abstract interpretation of algebraic polynomial systems. In Proc. 6th Int. Conf. on Algebraic Methodology and Software Technology (AMAST’97), LNCS 1349, pages 138–154, 1997. 216 8. M. Falaschi, G. Levi, M. Martelli, and C. Palamidessi. Declarative modeling of the operational behavior of logic languages. Theor. Comput. Sci., 69(3):289–318, 1989. 218, 224 9. R. Giacobazzi. “Optimal” collecting semantics for analysis in a hierarchy of logic program semantics. In Proc. 13th Int. Symp. on Theor. Aspects of Comput. Sci. (STACS’96), LNCS 1046, pages 503–514, 1996. 216, 225 10. R. Giacobazzi and F. Ranzato. Completeness in abstract interpretation: a domain perspective. In Proc. 6th Int. Conf. on Algebraic Methodology and Software Technology (AMAST’97), LNCS 1349, pages 231–245, 1997. 216, 216, 216, 216, 220, 220, 220, 220 11. R. Giacobazzi, F. Ranzato and F. Scozzari. Complete abstract interpretations made constructive. In Proc. 23rd Int. Symp. on Math. Found. of Comp. Sci. (MFCS’98), LNCS, 1998. 216, 217

Building Complete Abstract Interpretations in a Linear Logic-Based Setting

229

12. N.D. Jones and H. Søndergaard. A semantics-based framework for the abstract interpretation of Prolog. In S. Abramsky and C. Hankin, editors, Abstract Interpretation of Declarative Languages, pages 123–142. Ellis Horwood Ltd, 1987. 227 13. A. Mycroft. Completeness and predicate-based abstract interpretation. In Proc. ACM Conf. on Partial Evaluation and Program Manipulation (PEPM’93), pages 179–185, 1993. 216, 216, 216 14. C. Palamidessi. Algebraic properties of idempotent substitutions. In Proc. 17th Int. Colloq. on Automata, Languages and Programming (ICALP’90), LNCS 443, pages 386–399, 1990. 225 15. U.S. Reddy and S.N. Kamin. On the power of abstract interpretation. Computer Languages, 19(2):79–89, 1993. 216 16. K.I Rosenthal. Quantales and their Applications. Longman Scientiﬁc & Technical, 1990. 217, 219, 219, 222, 226 17. R.C. Sekar, P. Mishra, and I.V. Ramakrishnan. On the power and limitation of strictness analysis. J. ACM, 44(3):505–525, 1997. 216 18. B. Steﬀen. Optimal data ﬂow analysis via observational equivalence. In Proc. 14th Int. Symp. on Math. Found. of Comp. Sci. (MFCS’89), LNCS 379, pages 492–502, 1989. 216 19. D. Yetter. Quantales and (noncommutative) linear logic. J. Symbolic Logic, 55(1):41–64, 1990. 217, 219, 222

On the Power of Homeomorphic Embedding for Online Termination Michael Leuschel? Department of Computer Science, K.U. Leuven, Belgium Department of Electronics and Computer Science, University of Southampton, UK DIKU, University of Copenhagen, Denmark [email protected] Abstract. Recently well-quasi orders in general, and homeomorphic embedding in particular, have gained popularity to ensure the termination of program analysis, specialisation and transformation techniques. In this paper we investigate and clarify for the first time, both intuitively and formally, the advantages of such an approach over one using wellfounded orders. Notably we show that the homeomorphic embedding relation is strictly more powerful than a large class of involved well-founded approaches.

1

Introduction

The problem of ensuring termination arises in many areas of computer science and a lot of work has been devoted to proving termination of term rewriting systems (e.g. [7,8,9,37] and references therein) or of logic programs (e.g. [6,38] and references therein). It is also an important issue within all areas of program analysis, specialisation and transformation: one usually strives for methods which are guaranteed to terminate. One can basically distinguish between two kinds of techniques for ensuring termination: • static techniques, which prove or ensure termination of a program or process beforehand (i.e. off-line) without any kind of execution, and • online (or dynamic) techniques, which ensure termination of a process during its execution. (The process itself can of course be, e.g., performing a static analysis.) Static approaches have less information at their disposal but do not require runtime intervention (which might be impossible). Which of the two approaches is taken depends entirely on the application area. For instance, static termination analysis of logic programs [6,38] falls within the former context, while termination of program specialisation, transformation or analysis is often ensured in an online manner. This paper is primarily aimed at studying and improving online termination techniques. Let us examine the case of partial deduction [29,10,23] — an automatic technique for specialising logic programs. Henceforth we suppose some familiarity with basic notions in logic programming [3,28]. ?

Part of the work was done while the author was Post-doctoral Fellow of the Fund for Scientific Research - Flanders Belgium (FWO) and visiting DIKU, University of Copenhagen.

G. Levi (Ed.): SAS’98, LNCS 1503, pp. 230–245, 1998. c Springer-Verlag Berlin Heidelberg 1998

On the Power of Homeomorphic Embedding for Online Termination

231

Partial deduction based upon the Lloyd and Shepherdson framework [29] generates (possibly incomplete) SLDNF-trees for a set A of atoms. The specialised program is extracted from these trees by producing one clause for every nonfailing branch. The resolution steps within the SLDNF-trees — often referred to as unfolding steps — are those that have been performed beforehand, justifying the hope that the specialised program is more efficient. Now, to ensure termination of partial deduction two issues arise [10,34]. One is called the local termination problem, corresponding to the fact that each generated SLDNF-tree should be finite. The other is called the global termination problem, meaning that the set A should contain only a finite number of atoms. A similar classification can be done for most other program specialisation techniques (cf. e.g. [26]). Below we mainly use local termination to illustrate our concepts. (As shown in [34] the atoms in A can be structured into a global tree and methods similar to the one for local termination can be used to ensure global termination.) However, the discussions and contributions of the present paper are also (immediately) applicable in the context of analysis, specialisation and transformation techniques in general, especially when applied to computational paradigms, such as logic programming, constrained logic programming, conditional term rewriting, functional programming and functional & logic programming. For instance, abstract interpretation techniques usually analyse a set of abstract calls to which new call patterns are continuously added. One thus faces a problem very similar to global termination of partial deduction. Depth Bounds. One, albeit ad-hoc, way to solve the local termination problem is to simply impose an arbitrary depth bound. Such a depth bound is of course not motivated by any property, structural or otherwise, of the program or goal under consideration. In the context of local termination, the depth bound will therefore typically lead either to too little or too much unfolding. Determinacy. Another approach, often used in partial evaluation of functional programs [17] is to (only) expand a tree while it is determinate (i.e. it only has one non-failing branch). However, this approach can be very restrictive and in itself does not guarantee termination, as there can be infinitely failing determinate computations at specialisation time. Well-Founded Orders. Luckily, more refined approaches to ensure local termination exist. The first non-ad-hoc methods [5,33,32,31] in logic and [40,47] functional programming were based on well-founded orders, inspired by their usefulness in the context of static termination analysis. These techniques ensure termination, while at the same time allowing unfolding related to the structural aspect of the program and goal to be specialised, e.g., permitting the consumption of static input within the atoms of A.

232

M. Leuschel

Definition 1. (wfo) A (strict) partial order >S on a set S is an anti-reflexive, anti-symmetric and transitive binary relation on S × S. A sequence of elements s1 , s2 , . . . in S is called admissible wrt >S iff si > si+1 , for all i ≥ 1. We call >S a well-founded order (wfo) iff there is no infinite admissible sequence wrt >S To ensure local termination, one has to find a sensible well-founded order on atoms and then only allow SLDNF-trees in which the sequence of selected atoms is admissible wrt the well-founded order. If an atom that we want to select is not strictly smaller than its ancestors, we either have to select another atom or stop unfolding altogether. Example 1. Let P be the reverse program using an accumulator: rev ([], Acc, Acc) ← rev ([H|T ], Acc, Res) ← rev (T, [H|Acc], Res) A simple well-founded order on atoms of the form rev (t1 , t2 , t3 ) might be based on comparing the termsize (i.e., the number of function and constant symbols) of the first argument. We then define the wfo on atoms by: rev (t1 , t2 , t3 ) > rev (s1 , s2 , s3 ) iff term size(t2 ) > term size(s2 ). Based on that wfo, the goal ← rev ([a, b|T ], [], R) can be unfolded into the goal ← rev ([b|T ], [a], R) and further into ← rev (T, [b, a], R) because the termsize of the first argument strictly decreases at each step (even though the overall termsize does not decrease). However, ← rev (T, [b, a], R) cannot be further unfolded into ← rev (T 0 , [H 0 , b, a], R) because there is no such strict decrease. Much more elaborate techniques, which e.g. split the expressions into classes, use lexicographical ordering on subsequences of the arguments and even continuously refine the orders during the unfolding process, exist and we refer the reader to [5,33,32,31] for precise details. These works also present some further refinements on how to apply wfo’s, especially in the context of partial deduction. For instance, instead of requiring a decrease wrt every ancestor, one can only request a decrease wrt the covering ancestors, i.e. one only compares with the ancestor atoms from which the current atom descends (via resolution). Other refinements consist in allowing the wfo’s not only to depend upon the selected atom but on the context as well [32] or to ignore calls to non-recursive predicates. [32] also discusses a way to relax the condition of a “strict decrease” when refining a wfo. (Most of these refinements can also be applied to other approaches, notably the one we will present in the next section.) However, it has been felt by several researchers that well-founded orders are sometimes too rigid or (conceptually) too complex in an online setting. Recently, well-quasi orders have therefore gained popularity to ensure online termination of program manipulation techniques [4,41,42,25,26,11,18,1,20,46]. Unfortunately, this move to well-quasi orders has never been formally justified nor has the relation to well-founded approaches been investigated. We strive to do so in this paper and will actually prove that a rather simple well-quasi approach is strictly more powerful than a large class of involved well-founded approaches.

On the Power of Homeomorphic Embedding for Online Termination

2

233

Well-Quasi Orders and Homeomorphic Embedding

Formally, well-quasi orders can be defined as follows. Definition 2. (quasi order) A quasi order ≥S on a set S is a reflexive and transitive binary relation on S × S. Henceforth, we will use symbols like (possibly annotated by some subscript) to refer to strict partial orders and ≤, ≥ to refer to quasi orders. We will use either “directionality” as is convenient in the context. We also define an expression to be either a term (built-up from variables and function symbols of arity ≥ 0) or an atom (a predicate symbol applied to a, possibly empty, sequence of terms), and then treat predicate symbols as functors, but suppose that no confusion between function and predicate symbols can arise (i.e. predicate and function symbols are distinct). Definition 3. (wbr,wqo) Let ≤S be a binary relation on S × S. A sequence of elements s1 , s2 , . . . in S is called admissible wrt ≤S iff there are no i < j such that si ≤S sj . We say that ≤S is a well-binary relation (wbr) on S iff there are no infinite admissible sequences wrt ≤S . If ≤S is a quasi order on S then we also say that ≤S is a well-quasi order (wqo) on S. Observe that, in contrast to wfo’s, non-comparable elements are allowed within admissible sequences. An admissible sequence is sometimes called bad while a non-admissible one is called good. A well-binary relation is then such that all infinite sequences are good. There are several other equivalent definitions of well-binary relations and well-quasi orders. Higman [14] used an alternate definition of well-quasi orders in terms of the “finite basis property” (or “finite generating set” in [19]). A different (but also equivalent) definition of a wqo is: A quasi-order ≤V is a wqo iff for all quasi-orders V which contain ≤V (i.e. v≤V v 0 ⇒ vV v 0 ) the corresponding strict partial order ≺V is a wfo. This property has been exploited in the context of static termination analysis to dynamically construct well-founded orders from well-quasi ones and led to the initial use of wqo’s in the offline setting [7,8]. The use of well-quasi orders in an online setting has only emerged recently (it is mentioned, e.g., in [4] but also [41]) and has never been compared to well-founded approaches. There has been some comparison between wfo’s and wqo’s in the offline setting, e.g., in [37] it is argued that (for “simply terminating” rewrite systems) approaches based upon quasi-orders are less interesting than ones based upon a partial orders. In this paper we will show that the situation is somewhat reversed in an online setting. Furthermore, in the online setting, transitivity of a wqo is not really interesting and one can therefore drop this requirement, leading to the use of wbr’s. [24] contains some useful wbr’s which are not wqo’s. An interesting wqo is the homeomorphic embedding relation , which derives from results by Higman [14] and Kruskal [19]. It has been used in the context of term rewriting systems in [7,8], and adapted for use in supercompilation [45] in [42]. Its usefulness as a stop criterion for partial evaluation is also discussed and

234

M. Leuschel

advocated in [30]. Some complexity results can be found in [44] and [13] (also summarised in [30]). The following is the definition from [42], which adapts the pure homeomorphic embedding from [8] by adding a rudimentary treatment of variables. Definition 4. () The (pure) homeomorphic embedding relation on expressions is defined inductively as follows (i.e. is the least relation satisfying the rules): 1. X Y for all variables X, Y 2. s f (t1 , . . . , tn ) if s ti for some i 3. f (s1 , . . . , sn ) f (t1 , . . . , tn ) if ∀i ∈ {1, . . . , n} : si ti . The second rule is sometimes called the diving rule, and the third rule is sometimes called the coupling rule. When s t we also say that s is embedded in t or t is embedding s. By s t we denote that s t and t 6s. Example 2. The intuition behind the above definition is that A B iff A can be obtained from B by “striking out” certain parts, or said another way, the structure of A reappears within B. For instance we have p(a) p(f (a)) and indeed p(a) can be obtained from p(f (a)) by “striking out” the f . Observe that the “striking out” corresponds to the application of the diving rule 2 and that we even have p(a) p(f (a)). We also have, e.g., that: X X, p(X) p(f (Y )), p(X, X) p(X, Y ) and p(X, Y ) p(X, X). Proposition 1. The relation is a wqo on the set of expressions over a finite alphabet. For a complete proof, reusing Higman’s and Kruskal’s results [14,19] in a very straightforward manner, see, e.g., [23]. Extensions to infinite alphabets and improved treatment of variables can be found in [24]. To ensure, e.g., local termination of partial deduction, we have to ensure that the constructed SLDNF-trees are such that the selected atoms do not embed any of their ancestors (when using a well-founded order as in Example 1, we had to require a strict decrease at every step). If an atom that we want to select embeds one of its ancestors, we either have to select another atom or stop unfolding altogether. For example, based on , the goal ← rev ([a, b|T ], [], R) of Example 1 can be unfolded into ← rev ([b|T ], [a], R) and further into ← rev (T, [b, a], R) as rev ([a, b|T ], [], R) 6rev ([b|T ], [a], R), rev ([a, b|T ], [], R) 6rev (T, [b, a], R) and rev ([b|T ], [a], R) 6 rev (T, [b, a], R). However, ← rev (T, [b, a], R) cannot be further unfolded into ← rev (T 0 , [H 0 , b, a], R) as rev (T, [b, a], R) rev (T 0 , [H 0 , b, a], R). Observe that, in contrast to Example 1, we did not have to choose how to measure which arguments. We further elaborate on the inherent flexibility of in the next section. The homeomorphic embedding relation is also useful for handling structures other than expressions. It has, e.g., been successfully applied in [25,23,26] to detect (potentially) non-terminating sequences of characteristic trees. Also, seems to have the desired property that very often only “real” loops are detected and that they are detected at the earliest possible moment (see [30]).

On the Power of Homeomorphic Embedding for Online Termination

3

235

Comparing wbr’s and wfo’s

3.1

General Comparison

It follows from Definitions 1 and 3 that if ≤V is a wqo then t3 . Let, e.g., s < t denote that s is strictly more general than t. Then < is a wfo (see below) but p(X, X, a) 6> p(X, Z, b) and p(X, Z, b) 6> p(X, Y, a) even though p(X, X, a) > p(X, Y, a). Let us now examine the power of one particular wqo, the earlier defined . 3.2

Homeomorphic Embedding and Monotonic wfo’s

The homeomorphic embedding relation is very flexible. It will for example, when applied to the sequence of covering ancestors, permit the full unfolding of most terminating Datalog programs, the quicksort or even the mergesort program when the list to be sorted is known (the latter poses problems to some static termination analysis methods [38,27]; for some experiments see Appendix A). Also, the produce-consume example from [31] requires rather involved techniques (considering the context) to be solved by wfo’s. Again, this example poses no problem to (cf. Appendix A). The homeomorphic embedding is also very powerful in the context of metaprogramming. Notably, it has the ability to “penetrate” layers of (non-ground) meta-encodings (cf. also Appendix A). For instance, will admit the following sequences (where Example 1 is progressively wrapped into “vanilla” metainterpreters counting resolution steps and keeping track of the selected predicates respectively): Sequence rev ([a, b|T ], [], R) ; rev ([b|T ], [a], R) solve(rev ([a, b|T ], [], R), 0) ; solve(rev ([b|T ], [a], R), s(0)) solve 0 (solve(rev ([a, b|T ], [], R), 0), []) ; solve 0 (solve(rev ([b|T ], [a], R), s(0)), [rev])

Again, this is very difficult for wfo’s and requires refined and involved techniques (of which to our knowledge no implementation exists). We have intuitively demonstrated the usefulness of and that it is often more flexible than wfo’s. But can we prove some “hard” results? It turns out that we can and we now establish that — in the online setting — is strictly more generous than a large class of refined wfo’s.

236

M. Leuschel

Definition 5. A well-founded order ≺ on expressions is said to be monotonic iff the following rules hold: 1. X 6Y for all variables X, Y , 2. s 6f (t1 , . . . , tn ) whenever f is a function symbol and s 6ti for some i and 3. f (s1 , . . . , sn ) 6f (t1 , . . . , tn ) whenever ∀i ∈ {1, . . . , n} : si 6fi . Note the similarity of structure with the definition of (but, contrary to , 6does not have to be the least relation satisfying the rules). This similarity of structure will later enable us to prove that any sequence admissible wrt ≺ must also be admissible wrt (by showing that st ⇒ s 6t). Also observe that point 2 need not hold for predicate symbols and that point 3 implies that c 6c for all constant and proposition symbols c. Finally, there is a subtle difference between monotonic wfo’s as of Definition 5 and wfo’s which possess the replacement property (such orders are called rewrite orders in [37] and monotonic in [7]). More on that below. Similarly, we say that a norm k.k (i.e. a mapping from expressions to IN ) is said to be monotonic iff the associated wfo ≺k.k is monotonic (t1 ≺k.k t2 iff kt1 k < kt2 k). For instance the termsize norm (see below) is trivially monotonic. More generally, any semi-linear norm of the following form is monotonic: Proposition 2.P Let the norm k.k : Expr →IN be defined by: n • ktk = cf + i=1 cf,i kti k if t = f (t1 , . . . , tn ), n ≥ 0 • ktk = cv otherwise (i.e. t is a variable) Then k.k is monotonic if all coefficients cv , cf , cf,i are ≥ 0 and cf,i ≥ 1 for all function symbols f of arity ≥ 1 (but not necessarily for all predicate symbols). Proof. As < on IN is total we have that s 6> t is equivalent to s ≤ t. The proof proceeds by induction on the structure of the expressions and examines every rule of Definition 5 separately: 1. X ≤ Y for all variables X, Y this trivially holds as we use the same constant cv for all variables. 2. s ≤ f (t1 , . . . , tn ) whenever s ≤ ti for some i This holds trivially if all coefficients are ≥ 0 and if cf,i ≥ 1. This is verified, as the rule only applies if f is a function symbol. 3. f (s1 , . . . , sn ) ≤ f (t1 , . . . , tn ) whenever ∀i ∈ {1, . . . , n} : si ≤ fi This holds trivially, independently of whether f is a function or predicate symbol, as all coefficients are positive (and the same coefficient is applied to si and ti ).

By taking cv = 0 and cf,i = cf = 1 for all f we get the termsize norm k.kts , which by the above proposition is thus monotonic. Also, by taking cv = 1 and cf,i = cf = 1 for all f we also get a monotonic norm, counting symbols. Finally, a linear norm can always be obtained [38] by setting cv = 0, cf,i = 1 and cf ∈ IN for all f . Thus, as another corollary of the above, any linear norm is monotonic. Proposition 3. Let k.k1 , . . . , k.kk be monotonic norms satisfying Proposition 2. Then the lexicographical ordering ≺lex defined by s ≺lex t iff ∃i ∈ {1, . . . , k} such that kski < ktki and ∀j ∈ {1, . . . , i − 1}: kskj = ktkj is a monotonic wfo.

On the Power of Homeomorphic Embedding for Online Termination

237

Proof. By standard results (see, e.g., [8]) we know that ≺lex is a wfo (as < is a wfo on IN ). We will prove that ≺lex satisfies all the rules of Definition 5. 1. First, rule 1 is easy as kXki = kY ki for all i and variables X, Y and therefore we never have X ≺lex Y . 2. Before examining the other rules, let us note that s 6lex t is equivalent to saying that either a) ∀j ∈ {1, . . . , k} kskj = ktkj or b) there exists an j ∈ {1, . . . , k} such that kskj < ktkj and ∀l ∈ {1, . . . , j − 1}: kskl = ktkl . Let us now examine rule 2 of Definition 5. We have to prove that whenever s 6lex ti the conclusion of the rule holds. Let us first examine case a) for s 6lex ti . We have kskj = kti kj and thus we know that kskj ≤ kf (t1 , . . . , tn )kj by monotonicity of k.kj (as < on IN is total we have that s 6> t is equivalent to s ≤ t). As this holds for all k.kj we cannot have sj lex f (t1 , . . . , tn ). Let us now examine the second case b) for s 6lex ti . Let j ∈ {1, . . . , k} such that kskj < kti kj and ∀l ∈ {1, . . . , j − 1}: kskl = kti kl . For all l we can deduce as above that kskl ≤ kf (t1 , . . . , tn )kl . However, we still have to prove that kskj < kf (t1 , . . . , tn )kj . By monotonicity of k.kj we only know that kskj ≤ kf (t1 , . . . , tn )kj . But we can also apply monotonicity of k.kj to deduce that kti kj ≤ kf (t1 , . . . , tn )kj and hence we can infer the desired property (as kskj < kti kj ). 3. Now, for rule 3 we have to prove that whenever si 6lex ti for all i ∈ {1, . . . , n} the conclusion of the rule holds. There are again two cases. a) We can have ksi kj = kti kj for all i, j. By monotonicity of each k.kj we know that kf (s1 , . . . , sn )kj ≤ kf (t1 , . . . , tn )kj for all j ∈ {1, . . . , k}. Hence, we cannot have f (s1 , . . . , sn ) lex f (t1 , . . . , tn ). b) In the other case we know that there must be a value j 0 ∈ {1, . . . , k} such that for some i: ksi kj 0 < kti kj 0 and ∀l ∈ {1, . . . , j 0 − 1}: ksi kl = kti kl . I.e., by letting j denote the minimum value j 0 for which this holds, we know that for some i: ksi kj < kti kj and for all i0 : ∀l ∈ {1, . . . , j}: ksi0 kl ≤ kti0 kl . By monotonicity of each k.kl we can therefore deduce that ∀l ∈ {1, . . . , j}: kf (s1 , . . . , sn )kl ≤ kf (t1 , . . . , tn )kl . We can also deduce by monotonicity of k.kj that kf (s1 , . . . , sn )kj ≤ kf (t1 , . . . , tn )kj . We can even deduce that kf (s1 , . . . , sn )kj ≤ kf (t1 , . . . , ti−1 , si , ti+1 , . . . , tn )kj ≤ kf (t1 , . . . , tn )kj . Now, we just have to prove that: kf (t1 , . . . , ti−1 , si , ti+1 , . . . , tn )kj < kf (t1 , . . . , tn )kj in order to affirm that kf (s1 , . . . , sn )kj 6lex kf (t1 , . . . , tn )kj . This does not hold for all monotonic norms, but as we know that k.kj satisfies Proposition 2, this can be deduced by the fact that the coefficient cf,i in k.kj must be ≥ 1.

It is important that the norms k.k1 , . . . , k.kk satisfy Proposition 2. Otherwise, a counterexample would be as follows. Let kak1 = 1, kbk1 = 2 and kf (a)k1 = kf (b)k1 = 5. Also let kak2 = 2, kbk2 = 1 and kf (a)k2 = 3, kf (b)k2 = 2. Now we have a ≺lex b, i.e. a 6lex b, but also f (a) lex f (b) and condition 3 of monotonicity for ≺lex is violated. One could make Proposition 3 slightly more general, but the current version is sufficient to show the desired result, namely that most of the wfo’s used in online practice are actually monotonic. For example almost all of the refined wfo’s defined in [5,33,32,31] are monotonic: • Definitions 3.4 of [5], 3.2 of [33] and 2.14 of [32] all sum up the number of function symbols (i.e. termsize) of a subset of the argument positions of

238

M. Leuschel

atoms. These wfo’s are therefore immediately covered by Proposition 2. The algorithms only differ in the way of choosing the positions to measure. The early algorithms simply measure the input positions, while the later ones dynamically refine the argument positions to be measured (but which are still measured using the termsize norm). • Definitions 3.2 of [32] as well as 8.2.2 of [31] use the lexicographical order on the termsizes of some selected argument positions. These wfo’s are therefore monotonic as a corollary to Propositions 2 and 3. The only non-monotonic wfo in that collection of articles is the one defined specifically for metainterpreters in Definition 3.4 of [5] (also in Sect. 8.6 of [31]) which uses selector functions to focus on subterms to be measured. We will return to this approach below. Also, as already mentioned, some of the techniques in [32,31] (in Sects. 3.4 and 8.2.4 respectively) do not require the whole sequence to be admissible wrt a unique wfo, i.e. one can split up a sequence into a (finite) number of subsequences and apply different (monotonic) wfo’s on these subsequences. Similar refinements can also be developed for wqo’s and the formal study of these refinements are (thus) not the main focus of the paper. Before showing that is strictly more powerful than the union of all monotonic wfo’s, we adapt the class of simplification orderings from term rewriting systems. It will turn out that the power of this class is also subsumed by . Definition 6. A simplification ordering is a wfo ≺ on expressions which satisfies 1. s ≺ t ⇒ f (t1 , . . . , s, . . . , tn ) ≺ f (t1 , . . . , t, . . . , tn ) (replacement property), 2. t ≺ f (t1 , . . . , t, . . . , tn ) (subterm property) and 3. s ≺ t ⇒ sσ ≺ tγ for all variable only substitutions σ and γ (invariance under variable replacement). The third rule of the above definition is new wrt term-rewriting systems and implies that all variables must be treated like a unique new constant. It turns out that a lot powerful wfo’s are simplification orderings [7,37]: recursive path ordering, Knuth-Bendix ordering or lexicographic path ordering, to name just a few. However, not all wfo’s of Proposition 2 are simplification orderings: e.g., for cf = 0, ca = 1 we have kak = kf (a)k and the subterm property does not hold (for the associated wfo). In addition, Proposition 2 allows a special treatment for predicates. On the other hand, there are wfo’s which are simplification orderings but are not monotonic according to Definition 5. Proposition 4. Let ≺ be a wfo on expressions. Then any admissible sequence wrt ≺ is also an admissible sequence wrt if ≺ is a) monotonic or if it is b) a simplification ordering. Proof. First, let us observe that for a given wfo ≺ on expressions, any admissible sequence wrt ≺ is also an admissible sequence wrt iff s t ⇒ s 6t. Indeed (⇒), whenever s t then s 6t, and this trivially implies (by transitivity of ≺) that any sequence not admissible wrt cannot be strictly descending wrt ≺. On the other hand

On the Power of Homeomorphic Embedding for Online Termination

239

(⇐), let us assume that for some s and t s t but s t. This means that the sequence s, t is admissible wrt but not wrt and we have a contradiction. a) The proof that for a monotonic wfo ≺ we have st ⇒ s 6t is by straightforward induction on the structure of s and t. The only “tricky” aspect is that the second rule for monotonicity only holds if f is a function symbol. But if f is a predicate symbol, then s ti cannot hold because we supposed that predicate and function symbols are distinct. b) If ≺ is a simplification ordering then we can apply Lemma 3.3 of [37] to deduce that ≺ is the superset of the strict part of (i.e., ≺⊇ ). Let us examine the two possibilities for s t. First, we can have s t. In that case we can deduce s ≺ t and thus s 6t. Second, we can have s t and t s. In that case s and t are identical, except for the variables. If we now take the substitution σ which assigns all variables in s and t to a unique variable we have sσ = tσ, i.e., sσ 6tσ. This means that s t cannot hold (because is invariant under variable replacement).

This means that the admissible sequences wrt are a superset of the union of all admissible sequences wrt simplification orderings and monotonic wfo’s. In other words, no matter how much refinement we put into an approach based upon monotonic wfo’s or upon simplification orderings we can only expect to approach in the limit. But by a simple example we can even dispel that hope. Example 3. Take the sequence δ = f (a), f (b), b, a. This sequence is admissible wrt as f (a) 6f (b), f (a) 6b, f (a) 6a, f (b) 6b, f (b) 6a and a 6b. However, there is no monotonic wfo ≺ which admits this sequence. More precisely, to admit δ we must have f (a) f (b) as well as b a, i.e. a 6b. But this violates rule 3 of Definition 5 and ≺ cannot be monotonic. This also violates rule 1 of Definition 6 and ≺ cannot be a simplification ordering. These new results relating to monotonic wfo’s shed light on ’s usefulness in the context of ensuring online termination. But of course the admissible sequences wrt are not a superset of the union of all admissible sequences wrt any wfo.1 For instance the list-length norm k.kllen is not monotonic, and indeed we have for t1 = [1, 2, 3] and t2 = [[1, 2, 3], 4] that kt1 kllen = 3 > kt2 kllen = 2 although t1 t2 . So there are sequences admissible wrt list-length but not wrt . The reason is that k.kllen in particular and nonmonotonic wfo’s in general can completely ignore certain parts of the term, while will always inspect that part. E.g., if we have s f (. . . t . . .) and ignores the subterm t then it will also be true that s f (. . . s . . .) while s f (. . . s . . .),2 i.e. the sequence s, f (. . . s . . .) is admissible wrt but not wrt . For that same reason the wfo’s for metainterpreters defined in Definition 3.4 of [5] mentioned above are not monotonic, as they are allowed to completely focus on subterms, fully ignoring other subterms. However, automation of that technique is not addressed in [5]. E.g., for this wfo one cannot immediately apply 1 2

Otherwise could not be a wqo, as all finite sequences without repetitions are admissible wrt some wfo (map last element to 1, second last element to 2, . . . ). Observe that if f is a predicate symbols then f (. . . s . . .) is not a valid expression, which enabled us to ignore arguments to predicates in e.g. Proposition 2.

240

M. Leuschel

the idea of continually refining the measured subterms, because otherwise one might simply plunge deeper and deeper into the terms and termination would not be ensured. A step towards an automatic implementation is presented in Sect. 8.6 of [31] and it will require further work to formally compare it with wqo-based approaches and whether the ability to completely ignore certain parts of an expression can be beneficial for practical programs. But, as we have seen earlier, alone is already very flexible for metainterpreters, even more so when combined with characteristic trees [26] (see also [46]). Of course, for any wfo (monotonic or not) one can devise a wbr (cf. Lemma 1) which has the same admissible sequences. Still there are some feats that are easily attained, even by using , but which cannot be achieved by a wfo approach (monotonic or not). Take the sequences S1 = p([], [a]), p([a], []) and S2 = p([a], []), p([], [a]). Both of these sequences are admissible wrt This illustrates the flexibility of using well-quasi orders compared to well-founded ones in an online setting, as there exists no wfo (monotonic or not) which will admit both these sequences. It however also illustrates why, when using a wqo in that way, one has to compare with every predecessor state of a process. Otherwise one can get infinite derivations of the form p([a], []) → p([], [a]) → p([a], []) →. . . .3 Short Note on Offline Termination. This example also shows why (or well-quasi orders in general) cannot be used directly for static termination analysis. Let us explain what we mean. Take, e.g., a program containing the clauses C1 = p([a], []) ← p([], [a]) and C2 = p([], [a]) ← p([a], []). Then, in both cases the body is not embedding the head, but still the combination of the two clauses leads to a non-terminating program. However, can be used to construct wellfounded orders for static termination analysis. Take the clause C1 . The head and the body are incomparable according to . So, we can simply extend by stating that p([a], []) p([], a) (thus making the head strictly larger than the body atom). As already mentioned, for any extension ≤ of a wqo we have that < is a wfo. Thus we know that the program just consisting of C1 is terminating. If we now analyse C2 we have that, according to the extended wqo, the body is strictly larger than the head and (luckily) we cannot prove termination (i.e. there is no way of extending so that for both C1 and C2 the head is strictly larger than the body).

4

Discussion and Conclusion

Of course is not the ultimate relation for ensuring online termination. On its own in the context of local control of partial deduction, will sometimes allow 3

When using a wfo one has to compare only to the closest predecessor [32], because of the transitivity of the order and the strict decrease enforced at each step. However, wfo’s are usually extended to incorporate variant checking and then require inspecting every predecessor anyway (though only when there is no strict weight decrease, see, e.g., [31,32]).

On the Power of Homeomorphic Embedding for Online Termination

241

too much unfolding than desirable for efficiency concerns (i.e. more unfolding does not always imply a better specialised program) and we refer to the solutions developed in, e.g., [26,18]. For some applications, remains too restrictive. In particular, it does not always deal satisfactorily with fluctuating structure (arising, e.g., for certain meta-interpretation tasks) [46]. The use of characteristic trees [23,26] remedies this problem to some extent, but not totally. A further step towards a solution is presented in [46]. In that light, it might be of interest to study whether the extensions of the homeomorphic embedding relation proposed in [39] and [21] (in the context of static termination analysis of term rewrite systems) can be useful in an online setting. As shown in [24] the treatment of variables of is rather rudimentary and several ways to remedy this problem are presented. In summary, we have shed new light on the relation between wqo’s and wfo’s and have formally shown (for the first time) why wqo’s are more interesting than wfo’s for ensuring termination in an online setting (such as program specialisation or analysis). We have illustrated the inherent flexibility of and proved that, despite its simplicity, it is strictly more generous than the class of monotonic wfo’s. As all the wfo’s used for automatic online termination (so far) are actually monotonic, this formally establishes the interest of in that context. Acknowledgements. I would like to thank Danny De Schreye, Robert Gl¨uck, Jesper Jørgensen, Bern Martens, Maurizio Proietti, Jacques Riche and Morten Heine Sørensen for all the discussions, comments and joint research which led to this paper. Anonymous referees as well as Bern Martens provided extremely useful feedback on this paper.

References 1. M. Alpuente, M. Falaschi, P. Juli´ an, and G. Vidal. Spezialisation of lazy functional logic programs. In Proceedings PEPM’97, pages 151–162, Amsterdam, The Netherlands, 1997. ACM Press. 2. M. Alpuente, M. Falaschi, and G. Vidal. Narrowing-driven partial evaluation of functional logic programs. In H. Riis Nielson, editor, Proceedings ESOP’96, LNCS 1058, pages 45–61. Springer-Verlag, 1996. 3. K. R. Apt. Introduction to logic programming. In J. van Leeuwen, editor, Handbook of Theoretical Computer Science, chapter 10, pages 495–574. North-Holland Amsterdam, 1990. 4. R. Bol. Loop checking in partial deduction. The Journal of Logic Programming, 16(1&2):25–46, 1993. 5. M. Bruynooghe, D. De Schreye, and B. Martens. A general criterion for avoiding infinite unfolding during partial deduction. New Generation Computing, 11(1):47– 79, 1992. 6. D. De Schreye and S. Decorte. Termination of logic programs: The never ending story. The Journal of Logic Programming, 19 & 20:199–260, May 1994. 7. N. Dershowitz. Termination of rewriting. Journal of Symbolic Computation, 3:69– 116, 1987.

242

M. Leuschel

8. N. Dershowitz and J.-P. Jouannaud. Rewrite systems. In J. van Leeuwen, editor, Handbook of Theoretical Computer Science, Vol. B, pages 243–320. Elsevier, MIT Press, 1990. 9. N. Dershowitz and Z. Manna. Proving termination with multiset orderings. Communications of the ACM, 22(8):465–476, 1979. 10. J. Gallagher. Tutorial on specialisation of logic programs. In Proceedings PEPM’93, pages 88–98. ACM Press, 1993. 11. R. Gl¨ uck, J. Jørgensen, B. Martens, and M. H. Sørensen. Controlling conjunctive partial deduction of definite logic programs. In H. Kuchen and S. Swierstra, editors, Proceedings PLILP’96, LNCS 1140, pages 152–166, Aachen, Germany, September 1996. Springer-Verlag. 12. R. Gl¨ uck and M. H. Sørensen. A roadmap to supercompilation. In O. Danvy, R. Gl¨ uck, and P. Thiemann, editors, Proceedings of the 1996 Dagstuhl Seminar on Partial Evaluation, LNCS 1110, pages 137–160, Schloß Dagstuhl, 1996. SpringerVerlag. 13. J. Gustedt. Algorithmic Aspects of Ordered Structures. PhD thesis, Technische Universit¨ at Berlin, 1992. 14. G. Higman. Ordering by divisibility in abstract algebras. Proceedings of the London Mathematical Society, 2:326–336, 1952. 15. G. Huet. Confluent reductions: Abstract properties and applications to term rewriting systems. Journal of the ACM, 27(4):797–821, 1980. 16. N. D. Jones. An introduction to partial evaluation. ACM Computing Surveys, 28(3):480–503, September 1996. 17. N. D. Jones, C. K. Gomard, and P. Sestoft. Partial Evaluation and Automatic Program Generation. Prentice Hall, 1993. 18. J. Jørgensen, M. Leuschel, and B. Martens. Conjunctive partial deduction in practice. In J. Gallagher, editor, Proceedings LOPSTR’96, LNCS 1207, pages 59– 82, Stockholm, Sweden, August 1996. Springer-Verlag. 19. J. B. Kruskal. Well-quasi ordering, the tree theorem, and Vazsonyi’s conjecture. Transactions of the American Mathematical Society, 95:210–225, 1960. 20. L. Lafave and J. Gallagher. Constraint-based partial evaluation of rewriting-based functional logic programs. In N. Fuchs, editor, Proceedings LOPSTR’97, LNCS, Leuven, Belgium, July 1997. Springer-Verlag. 21. P. Lescanne. Well rewrite orderings and well quasi-orderings. Technical Report No 1385, INRIA-Lorraine, France, January 1991. 22. M. Leuschel. The ecce partial deduction system and the dppd library of benchmarks. 1996–1998. Obtainable via http://www.cs.kuleuven.ac.be/˜dtai. 23. M. Leuschel. Advanced Techniques for Logic Program Specialisation. PhD thesis, K.U. Leuven, May 1997. Accessible via http://www.cs.kuleuven.ac.be/˜michael. 24. M. Leuschel. Homeomorphic Embedding for Online Termination. Technical Report, Department of Electronics and Computer Science, University of Southampton, 1998. 25. M. Leuschel and B. Martens. Global control for partial deduction through characteristic atoms and global trees. In O. Danvy, R. Gl¨ uck, and P. Thiemann, editors, Proceedings of the 1996 Dagstuhl Seminar on Partial Evaluation, LNCS 1110, pages 263–283, Schloß Dagstuhl, 1996. Springer-Verlag. 26. M. Leuschel, B. Martens, and D. De Schreye. Controlling generalisation and polyvariance in partial deduction of normal logic programs. ACM Transactions on Programming Languages and Systems, 20(1):208–258, January 1998.

On the Power of Homeomorphic Embedding for Online Termination

243

27. N. Lindenstrauss, Y. Sagiv, and A. Serebrenik. Unfolding the mystery of mergesort. In N. Fuchs, editor, Proceedings LOPSTR’97, LNCS, Leuven, Belgium, July 1997. Springer-Verlag. 28. J. W. Lloyd. Foundations of Logic Programming. Springer-Verlag, 1987. 29. J. W. Lloyd and J. C. Shepherdson. Partial evaluation in logic programming. The Journal of Logic Programming, 11(3& 4):217–242, 1991. ´ 30. R. Marlet. Vers une Formalisation de l’Evaluation Partielle. PhD thesis, Universit´e de Nice - Sophia Antipolis, December 1994. 31. B. Martens. On the Semantics of Meta-Programming and the Control of Partial Deduction in Logic Programming. PhD thesis, K.U. Leuven, February 1994. 32. B. Martens and D. De Schreye. Automatic finite unfolding using well-founded measures. The Journal of Logic Programming, 28(2):89–146, August 1996. 33. B. Martens, D. De Schreye, and T. Horv´ ath. Sound and complete partial deduction with unfolding based on well-founded measures. Theoretical Computer Science, 122(1–2):97–117, 1994. 34. B. Martens and J. Gallagher. Ensuring global termination of partial deduction while allowing flexible polyvariance. In L. Sterling, editor, Proceedings ICLP’95, pages 597–613, Kanagawa, Japan, June 1995. MIT Press. 35. J. Martin. Sonic partial deduction. Technical Report, Department of Electronics and Computer Science, University of Southampton, 1998. 36. P.-A. Melli`es. On a duality between Kruskal and Dershowitz theorems. In K. G. Larsen, editor, Proceedings of ICALP’98, LNCS, Aalborg, Denmark, 1998. Springer-Verlag. 37. A. Middeldorp and H. Zantema. Simple termination of rewrite systems. Theoretical Computer Science, 175(1):127–158, 1997. 38. L. Pl¨ umer. Termination Proofs for Logic Programs. LNCS 446. Springer-Verlag, 1990. 39. L. Puel. Using unavoidable set of trees to generalize Kruskal’s theorem. Journal of Symbolic Computation, 8:335–382, 1989. 40. E. Ruf. Topics in Online Partial Evaluation. PhD thesis, Stanford University, March 1993. 41. D. Sahlin. Mixtus: An automatic partial evaluator for full Prolog. New Generation Computing, 12(1):7–51, 1993. 42. M. H. Sørensen and R. Gl¨ uck. An algorithm of generalization in positive supercompilation. In J. W. Lloyd, editor, Proceedings ILPS’95, pages 465–479, Portland, USA, December 1995. MIT Press. 43. M. H. Sørensen, R. Gl¨ uck and N.D. Jones. A positive supercompiler. Journal of Functional Programming, 1996. 44. J. Stillman. Computational Problems in Equational Theorem Proving. PhD thesis, State University of New York at Albany, 1988. 45. V. F. Turchin. The concept of a supercompiler. ACM Transactions on Programming Languages and Systems, 8(3):292–325, 1986. 46. W. Vanhoof and B. Martens. To parse or not to parse. In N. Fuchs, editor, Proceedings LOPSTR’97, LNCS, Leuven, Belgium, July 1997. Springer-Verlag. 47. D. Weise, R. Conybeare, E. Ruf, and S. Seligman. Automatic online partial evaluation. In Proceedings of the Conference on Functional Programming Languages and Computer Architectures, LNCS 523, pages 165–191, Harvard University, 1991. Springer-Verlag.

244

A

M. Leuschel

Small Experiments with the ecce System

The purpose of this appendix is to illustrate the flexibility which the homeomorphic embedding relation provides straight “out of the box” (other more intricate well-quasi orders, like the one used by Mixtus [41], can handle some of the examples below as well). For that we experiment with the ecce partial deduction system [22] using an unfolding rule based on which allows the selection of either determinate literals or left-most literals within a goal, given that no covering ancestor [5] is embedded (via ). To ease readability, the specialised programs are sometimes presented in unrenamed form. First, let us take the mergesort program, which is somewhat problematic for a lot of static termination analysis methods [38,27]. mergesort([],[]). mergesort([X],[X]). mergesort([X,Y|Xs],Ys) :split([X,Y|Xs],X1s,X2s), mergesort(X1s,Y1s),mergesort(X2s,Y2s), merge(Y1s,Y2s,Ys). split([],[],[]). split([X|Xs],[X|Ys],Zs) :- split(Xs,Zs,Ys). merge([],Xs,Xs). merge(Xs,[],Xs). merge([X|Xs],[Y|Ys],[X|Zs]) :- X =< Y, merge(Xs,[Y|Ys],Zs). merge([X|Xs],[Y|Ys],[Y|Zs]) :- X>Y, merge([X|Xs],Ys,Zs).

The partial evaluation query: ?- mergesort([3,1,2],X). As the following resulting specialised program shows, homeomorphic embedding allowed the full unfolding of mergesort: mergesort([3,1,2],[1,2,3]). It took ecce less than 0.5 s on a Sparc Classic to produce the above program (including post-processing and writing to file). We can even achieve this same feat even if we interpose one or more levels of metainterpretation! Take a vanilla solve metainterpreter with mergesort at the object-level: solve([]). solve([H|T]) :claus(H,Bdy),solve(Bdy),solve(T). claus(mergesort([],[]), []). claus(mergesort([X],[X]), []).

On the Power of Homeomorphic Embedding for Online Termination

245

claus(mergesort([X,Y|Xs],Ys), [split([X,Y|Xs],X1s,X2s), mergesort(X1s,Y1s),mergesort(X2s,Y2s), merge(Y1s,Y2s,Ys) ]). claus(split([],[],[]), []). claus(split([X|Xs],[X|Ys],Zs) , [ split(Xs,Zs,Ys) ]). claus(merge([],Xs,Xs), []). claus(merge(Xs,[],Xs), []). claus(merge([X|Xs],[Y|Ys],[X|Zs]) , [ X =< Y, merge(Xs,[Y|Ys],Zs) ]). claus(merge([X|Xs],[Y|Ys],[Y|Zs]) , [ X>Y, merge([X|Xs],Ys,Zs)]). claus(’=’(X,Y),[]) :- X > Y. mergesort_test(X) :- solve([mergesort([3,1,2],X)]).

The partial evaluation query: ?- mergesort_test(X). Again homeomorphic embedding allowed the full unfolding: mergesort_test([1,2,3]). It took ecce 2.86 s on a Sparc Classic to produce the above program (including post-processing and writing to file). The following example is taken from [31]. produce([],[]). produce([X|Xs],[X|Ys]) :- produce(Xs,Ys). consume([]). consume([X|Xs]) :- consume(Xs).

The partial evaluation query: ?- produce([1,2|X],Y), consume(Y). To solve it in the setting of unfolding based upon wfo’s one needs both partition based measure functions as well as taking the context into account. The same adequate unfolding can simply be achieved by based determinate unfolding, as illustrated by the following specialised program derived by ecce (default settings): produce_conj__1([],[1,2]). produce_conj__1([X1|X2],[1,2,X1|X3]) :produce_conj__2(X2,X3). produce_conj__2([],[]). produce_conj__2([X1|X2],[X1|X3]) :produce_conj__2(X2,X3).

Analysis of Imperative Programs through Analysis of Constraint Logic Programs Julio C. Peralta, John P. Gallagher, and H¨ useyin Saˇ glam University of Bristol Dept. of Computer Science Merchant Venturers Building Woodland Rd., Bristol, U.K. BS8 1UB fax: +44-117-9545208 e-mail: {jperalta, john}@cs.bris.ac.uk, [email protected]

Abstract. In this paper a method is proposed for carrying out analysis of imperative programs. We achieve this by writing down the language semantics as a declarative program (a constraint logic program, in the approach shown here). We propose an effective style of writing operational semantics suitable for analysis which we call one-state small-step semantics. Through controlled partial evaluation we are able to generate residual programs where the relationship between imperative statements and predicates is straightforward. Then we use a static analyser for constraint logic programs on the residual program. The analysis results are interpreted through program points associating predicates in the partially evaluated interpreter to statements in its corresponding imperative program. We used an analyser that allows us to determine linear equality, inequality and disequality relations among the variables of a program without user-provided inductive assertions or human interaction. The proposed method intends to serve as a framework for the analysis of programs in any imperative language. The tools required are a partial evaluator and a static analyser for the declarative language.

Keywords: Partial Evaluation, Constraint Logic Programming, Operational Semantics, Imperative Program Analysis.

1

Introduction

Program semantics has long been used as a formal basis for program manipulation. By this we mean that the formal semantics of a programming language is written down in some mathematical framework, which is then used to establish program properties such as termination, correctness with respect to specifications, or the correctness of program transformations. In the case of imperative languages the gap between semantics and programs is greater than in the case of declarative languages. Perhaps for this reason, semantics-based tools for declarative languages, such as abstract interpreters and partial evaluators, have been the subject of more intense research than similar tools for imperative languages. G. Levi (Ed.): SAS’98, LNCS 1503, pp. 246–261, 1998. c Springer-Verlag Berlin Heidelberg 1998

Analysis of Imperative Programs through Analysis

247

The aim of this paper is to show how to transfer results achieved in declarative languages to imperative languages. The approach is to implement the semantics of imperative languages carefully in a declarative language for which analysis and transformation tools exist. There exist several kinds of semantics for imperative languages. The choice of which one is better suited for the particular application is a subject of ongoing research [3]. We shall see that for the purposes of this paper a variant of operational semantics will suffice. Logic programming appears to be an elegant paradigm for imperative language semantics implementation. When written carefully, the semantics can be regarded as an interpreter for the imperative language. Appropriate techniques for implementing semantics are discussed in Sect. 2. Partial evaluation of the interpreter with respect to an imperative program yields an equivalent declarative program. By so doing we open up the possibility of applying well-developed techniques for analysis and transformation of constraint logic programs to imperative programs as well. Nevertheless, it is not clear how to relate the results of such analysis and/or transformation back to the original imperative program. In order to obtain a correspondence between the imperative program and its corresponding declarative program some tuning of the partial evaluator is needed. Otherwise, the partial evaluator may remove important information needed to relate imperative statements and variables with their declarative counterpart. Such tuning involves selecting among the predicates of the semanticsbased interpreter those we want to be defined in the residual program. Hence, we choose predicates from the semantics-based interpreter that relate directly to the meaning of the statements in the imperative program to be partially evaluated. As a result we get one predicate for each statement of the imperative program, thus highlighting the correspondence between imperative statements and predicates in the residual program. In this paper we propose a method that intends to serve as a framework for the analysis of programs in any imperative language, by writing down its semantics as a declarative program (a constraint logic program, in the approach shown here). The tools required are a partial evaluator and a static analyser for the declarative language. Section 2 considers the overall problem of encoding semantics in a logic programming language, in a form suitable for analysis. Section 3 provides some remarks on the implementation of the operational semantics for a small imperative language 1 . In Sect. 4 we show the partial evaluation process and give an example. Section 5 illustrates how to relate the results of the analysis back to its imperative program in a systematic way. Finally, in Sects. 6 and 7 we discuss related work, and state our final remarks and some trends for future work.

1

Currently we are writing the semantics for Java in a similar style

248

2

J.C. Peralta, J.P. Gallagher, and H. Saˇ glam

Analysis through Semantics

Before describing our experiments in detail let us consider some critical points concerning representing semantics in a form suitable for analysis. There are several styles of semantics for imperative languages to be found in textbooks. These may all be translated more or less directly into declarative programming languages, but it is necessary to consider carefully the choice of semantics and the style in which the semantics is represented. The choice is influenced by two questions: firstly, what kind of analysis is to be performed on the imperative program, and secondly, how can the complexity of the analysis be minimised? 2.1

Big-Step and Small-Step Semantics

The usual distinction between different kinds of semantics is between the compositional and operational styles. However for our purpose, the most relevant division is between big-step and small-step semantics. Note that operational semantics can be either big-step (natural semantics) or small-step (structural operational semantics). Let us represent program statements by S and computation states by E. Big-step semantics is modelled using a relation of the form bigstep(S, E1 , E2 ), which means that the statement S transforms the initial state E1 to final state E2 . The effect of each program construct is defined independently; compound statements are modelled by composing together the effects of its parts. On the other hand, small-step semantics is typically modelled by a transition relation of the form smallstep(hS1 , E1 i, hS2 , E2 i). This states that execution of statement S1 in state E1 is followed by the execution of statement S2 in state E2 . Small-step semantics models a computation as a sequence of computation states, and the effect of a program construct is defined as its effect on the computation state in the context of the program in which it occurs. Big steps can be derived from small steps by defining a special terminating statement, say halt, and expressing big-step relations as a sequence of small steps. bigstep(S, E1 , E2 ) ← smallstep(hS, E1 i, hhalt, E2 i). bigstep(S, E1 , E3 ) ← smallstep(hS, E1 i, hS1 , E2 i), bigstep(S1 , E2 , E3 ). For the purposes of program analysis, the two styles of semantics have significant differences. Analysis of the bigstep relation allows direct comparison of the initial and final states of a computation. As shown above, big steps are derivable from small steps but the analysis becomes more complex. If the purpose of analysis is to derive relationships between initial and final states, then big-step semantics would be recommended. On the other hand, the smallstep relation represents directly the relation between one computation state and the next, information which would be awkward (though not impossible) to extract from the big-step semantics. Small-step semantics is more appropriate for analyses where local information about states is required, such as the relationship of variables within the same state, or between successive states.

Analysis of Imperative Programs through Analysis

2.2

249

One-State and Two-State Semantics

There is another option for representing small-step semantics, which leads to programs significantly simpler to analyse. Replace the smallstep(hS1 , E1 i, hS2 , E2 i) relation by a clause exec(S1 , E1 ) ← exec(S2 , E2 ). The meaning of the predicate exec(S, E) is that there exists a terminating computation of the statement S starting in the state E. We also have to add a special terminal statement called halt which is placed at the end of every program, and a statement exec(halt, E) ← true. We call this style of small-step semantics a one-state semantics since the relation exec represents only one state, in contrast to two-state semantics given by the bigstep and smallstep relations. As an aside, the one-state and two-state styles of representation follow a pattern identified by Kowalski in [10], when discussing graph-searching algorithms in logic programming. Kowalski noted that there were two ways to formalise the task of searching for a path from node a to node z in a directed graph. One way is to represent the graph as a set of facts of the form go(X, Y ) representing arcs from node X to node Y . A relation path(X, Y ) could then be recursively defined on the arcs. The search task is to solve the goal ← path(a, z). Alternatively, an arc could be represented as a clause of form go(Y ) ← go(X). In this case, to perform the task of searching for a path from node a to node z, the fact go(a) ← true is added to the program, and computation is performed by solving the goal ← go(z). There is no need for a recursively defined path relation, since the underlying logic of the implication relation fulfils the need. One-state semantics corresponds to the use of the relation go(X) while twostate semantics corresponds to using the relation go(X, Y ). The “start state” corresponds to the clause exec(halt, E) ← true. Our experiments show that the one-state semantics is considerably simpler to analyse than the two-state semantics. 2.3

Analysis of the Semantics

It may be asked whether one-state semantics is expressive enough, since no output state can be computed. That is, given a program P and initial state E, the computation is simulated by running the goal ← exec(P, E) and the final state is not observable. This is certainly inadequate if we are using our semantics to simulate the full effect of computations. However, during program analysis of ← exec(P, E) more things are observable than during normal execution of the same goal. In particular we can use a program analysis algorithm that gives information about calls to different program subgoals. In one-state semantics with a relation exec(S, E), the analysis can determine (an approximation of) every instance of exec(S, E) that arises in the computation. We can even derive information about the relation of successive states since the analysis can derive information about instances of a clause exec(S1 , E1 ) ← exec(S2 , E2 ). In the experiments to be described below, we start from a structural operational semantics in a textbook style, and derive a one-state small-step semantics.

250

3

J.C. Peralta, J.P. Gallagher, and H. Saˇ glam

A Semantics-Based Interpreter

As already mentioned a desirable property of structural operational semantics is the way it reflects every change in the computation state. Here we present briefly a way of systematically translating formal operational semantics (adapted from [16]) into a constraint logic program. We shall first provide some elements of the syntax of the imperative language and the metavariables used in the semantics descriptions. Assume we have an imperative language L with assignments, arithmetic expressions, while statements, if-then-else conditionals, empty statement, statement composition2 , and boolean expressions. Let S be a statement, a be an arithmetic expression, b a boolean expression, e a variable environment (mapping variables to their value), and x a program variable. All these variables may occur subscripted. A pair hS, ei is a configuration. Also a state is a special terminal configuration. The operational semantics below give meaning to programs by defining an appropriate transition relation holding between configurations. 3.1

Structural Operational Semantics

Using structural operational semantics [16] a transition relation ⇒ defines the relationship between successive configurations. There are different kinds of transition corresponding to different kinds of statement. Accordingly, the meaning of empty statement, assignment statement, if-then-else statement, while-do statement and composition of statements are: hskip, ei ⇒ e hx := a, ei ⇒ e[x 7→ A[[a]]e] hif b then S1 else S2 , ei ⇒ hS1 , ei if B[[b]]e = tt hif b then S1 else S2 , ei ⇒ hS2 , ei if B[[b]]e = ff hwhile b do S, ei ⇒ hif b then (S; while b do S) else skip, ei hS1 ; S2 , ei ⇒ hS10 ; S2 , e0 i if hS1 , ei ⇒ hS10 , e0 i hS1 ; S2 , ei ⇒ hS2 , e0 i if hS1 , ei ⇒ e0

(1) (2) (3) (4) (5) (6) (7)

where A[[·]] is the semantics function for arithmetic expressions and B[[·]] is the semantics function for boolean expressions. Intuitively, the assignment axiom schema above says that in a state e, x := a is executed to yield a state e[x 7→ A[[a]]e] which is as e except that x has the value A[[a]]e. Moreover, transition 6 expresses that if S1 is not a primitive statement of the language then execution won’t proceed to S2 until the rest of S1 , S10 , has been fully executed. Transition 7 considers the case when execution of S1 has been completed thus yielding state e0 , hence execution of S2 starts from this new state. We may specialise transitions 6 and 7 by unfolding their conditions with respect to transitions 1, 2, 3, 4, and 5 above to transitions 9, 10, 11, 12, and 13 2

The statement composition operator ‘;’ is assumed to be right associative

Analysis of Imperative Programs through Analysis

251

below. Transition 8 below is obtained by unfolding the condition of transition 6 with respect to itself and applying the associativity property of the ‘;’ operator. We assume that cases 2, 3, 4 and 5 do not occur since all programs are terminated by skip. Hence the new semantics: hskip, ei h(S1 ; S2 ); S3 , ei hskip; S2 , ei hx := a; S2 , ei

⇒ e ⇒ hS1 ; (S2 ; S3 ), ei ⇒ hS2 , ei

⇒ hS2 , e[x 7→ A[[a]]e]i h(if b then S1 else S2 ); S3 , ei ⇒ hS1 ; S3 , ei if hB[[b]]e = tti h(if b then S1 else S2 ); S3 , ei ⇒ hS2 ; S3 , ei if hB[[b]]e = ff i h(while b do S); S2 , ei ⇒

(8) (9) (10) (11) (12)

(13) hif b then (S; while b do S else skip); S2 , ei

Next, we express the semantics as a constraint logic program below. It is worth noting that this representation aids the analysis phase by carrying a single environment instead of double environment (that is, the one-state small-step semantics discussed in Sect. 2). exec(skip,E) <exec(compose(compose(S1,S2),S3),E) <exec(compose(S1,compose(S2,S3)),E) exec(compose(skip,S2),E) =t+1,b>=j+1} tm := k[j]; k[j] := k[j+1]; k[j+1] := tm; t := j; {n>=b,j>=1,t>=0,b>=j+1,j>=t} j := j+1 if (t == 0) then {n>=b,t=0,j>=1,b>=j,b<j+1} b := -1; else {n>=b,t>=0,j>=t+1,b>=j,b<j+1} b := t;

These results are the same as those obtained in [2].

6

Related Work

The first practical results on imperative languages for deriving linear equality or disequality relations among the variables of a program is due to Cousot and Halbwachs [2]. Their system was implemented in Pascal. The model execution used is based on flow-charts and an approximation method based on convex polyhedra. Incidentally the analyser used on the experiments here reported uses a similar approximation method integrated with other constraint solvers [20]. Later on in [3] the author poses the possibility of deriving different static analysers parameterised by the language semantics. In a similar way [8] show how to

Analysis of Imperative Programs through Analysis

259

obtain a static analyser for a non strict functional language. Such a static analyser is derived by successive refinements of the original language specification, natural operational semantics. The possible analyses obtained by the analyser derived with this method depend on the program property sought. This program property should be provided in advance. It appears that this technique has been applied to obtain some classical compiler analyses of programs in the sense of [1]. A good source of related work on implementation/derivation of static analysers from operational semantics for different programming languages is [8]. In [21] the authors describe a technique based on the style of abstract interpretation to statically estimate the ranges of variables throughout a program. Their implementation has been realised in the context of an optimising/parallelising compiler for C. Again, this is an example of using a variant of operational semantics to describe the abstract interpreters for static analysis of imperative programs. In [13] the author lays out the theory of abstract interpretation using twolevel semantics. Two-level semantics had been previously used in [15] to describe code generation. A summary of both can be found in [14]. Using denotational definitions the semantics of Pascal-like languages is given making explicit the distinction between compile time computations and run time computations, hence the two levels of the metalanguage. For program analysis, an appropriate interpretation of the run time metalanguage aids the analysis by giving a nonstandard semantics to run time constructs. By contrast in the present work, the semantic definitions are given in a standard way, and the translation is carried out by the partial evaluator where the distinction between compile time and run time computations is accounted for. Another sort of problem reduction for analysis is provided in [17]. The authors convert the problem of identifying dependences among program elements to a problem of identifying may-aliases. The transformation output is a program in the same language as the input program where may-aliases in the transformed program yield information directly translatable to control flow dependences between statements in the original program. In a similar way the authors claim that control flow dependences in the transformed program have a direct reading as may-aliases in the corresponding program. Presumably the their method and ours could be combined to obtain other problem reduction results. In [19] it is shown how to use logic programs to aid the analysis of imperative programs with pointers. The formalism is shown for the case of the pointer equality problem in Pascal. During analysis a set of assertions, represented as unary clauses, is updated according to the meaning of the program statement evaluated, the update operation designated and a set of consistency rules. The update operations resemble operations in deductive databases. The semantics of the imperative program is not explicitly represented as a logic program as in our approach, but in both approaches logic programs are used to express program properties.

260

7

J.C. Peralta, J.P. Gallagher, and H. Saˇ glam

Final Remarks

We have developed a language-independent method for analysing imperative programs. The method is based on encoding the semantics of an imperative language in a logic programming language for which there are advanced tools available for program analysis and transformation. This allows us to transfer the results of research on analysis of logic programs to the analysis of imperative programs. The emphasis of our work is to find practical and efficient techniques for achieving this aim. A key aspect is to write the semantics in a way that is amenable to analysis. We identified the one-state small-step semantics as a suitable style. The problem of relating the results of analysis of a logic program to the source code of the original imperative program was also solved. A representation of the imperative program was constructed in which program points were represented by special terms in the logic program. The partial evaluation algorithm was modified to exploit these terms, and thus produce a residual logic program whose structure mirrored that of the imperative program. Thus results of analysis of the logic programs could be related directly to the imperative code. The correctness of our results follow from correctness of the partial evaluator and correctness of the analyser. Both correctness proofs may be done independently of the imperative language to be analysed, which we claim is one of the contributions of the present work. Future Work We have performed some promising experiments on a simple language, as shown in this paper, but our aim is to analyse programs in a mainstream imperative language. Currently we are well advanced in writing the operational semantics for a significant subset of Java. We aim to enhance our current analysis tools by handling non-linear arithmetic constraints, and boolean constraints. Moreover, we aim to increase the flexibility of analysis by using pre-interpretations to express properties and abstract compilation to encode them as logic programs [18]. The use of the same method to perform other analyses, such as contextsensitive or alias analyses remains to be explored.

References 1. Alfred V. Aho, Ravi Sethi, and Jeffrey D. Ullman. Compilers: Principles, Techniques, and Tools. Addison Wesley, 1986. 2. P. Cousot and N. Halbwachs. Automatic discovery of linear restraints amog variables of a program. In Proceedings of the Conference Record of the 5th ACM Symposium on Principles of Programming Languages, pages 84–97, Albuquerque, New Mexico, 1978. 3. Patrick Cousot. Abstract interpretation based static analysis parametrized by semantics. In Proceedings of the Fourth International Symposium on Static Analysis, SAS’97, pages 388–394, Paris, France, 1997. LNCS 1302, Springer Verlag.

Analysis of Imperative Programs through Analysis

261

4. J.P. Gallagher. Tutorial on specialisation of logic programs. In Proceedings of the ACM SIGPLAN Symposium on Partial Evaluation and Semantics-Based Program Manipulation, pages 88–98, Copenhagen, Denmark, 1993. ACM Press. 5. J. Gallagher and M. Bruynooghe. Some low-level source transformations for logic programs. In Proceedings of Meta90 Workshop on Meta Programming in Logic. Katholieke Universiteit Leuven, Belgium, 1990. 6. J. Gallagher and M. Bruynooghe. The derivation of an algorithm for program specialisation. New Generation Computing, 9(3&4):305–333, 1991. 7. J. Gallagher and D.A. de Waal. Fast and precise regular approximation of logic programs. In P. Van Hentenryck, editor, Proceedings of the International Conference on Logic Programming (ICLP’94), Santa Margherita Ligure, Italy. MIT Press, 1994. 8. Val´erie Gouranton and Daniel Le M´etayer. Formal development of static program analysers. In Proceedings of the 8th Israeli Conference on Computer Systems and Software Engineering, pages 101–110, Israel, 1997. 9. Donald E. Knuth. The Art of Computer Programming, volume 3 of Sorting and Searching. Addison-Wesley Publishing Company, 1973. 10. Robert Kowalski. Logic for Problem Solving. North Holland, 1979. 11. Michael Leuschel. Advanced Techniques for Logic Program Specialisation. PhD thesis, Katholieke Universiteit Leuven, Department of Computer Science, Leuven, Belgium, May 1997. 12. J. W. Lloyd and J. C. Shepherdson. Partial evaluation in logic programming. The Journal of Logic Programming, 11(3&4):217–242, 1991. 13. Flemming Nielson. Two-level semantics and abstract interpretation. Theoretical Computer Science, (69):117–242, 1989. 14. Flemming Nielson. Semantics-directed program analysis: A toolmaker’s perspective. In Third International Symposium, SAS’96. Springer Verlag, LNCS 1145, 1996. 15. Flemming Nielson and Hanne Riis Nielson. Two-level semantics and code generation. Theoretical Computer Science, (56):59–133, 1988. 16. Hanne Riis Nielson and Flemming Nielson. Semantics with Applications. John Wiley and Sons, 1992. 17. John L. Ross and Mooly Sagiv. Building a bridge between pointer aliases and program dependences. In European Symposium On Programming, Lisbon, Portugal, 1998. 18. H¨ useyin Sa˘ glam. A Toolkit for Static Analysis of Constraint Logic Programs. PhD thesis, Bristol University, Department of Computer Science, Bristol, U.K., March 1998. 19. Mooly Sagiv, Nissim Francez, Michael Rodeh, and Reinhard Wilhelm. A logicbased approach to program flow analysis. 1998. Submitted to Acta Informatica. 20. H. Sa˘ glam and J. Gallagher. Constrained regular approximation of logic programs. In N. Fuchs, editor, Logic Program Synthesis and Transformation (LOPSTR’97). Springer-Verlag, Lecture Notes in Computer Science, 1998. in press. 21. Clark Verbrugge, Phong Co, and Laurie Hendren. Generalized constant propagation a study of C. Lecture Notes in Computer Science, Compiler Construction, (1060):74–90, 1996.

Improving Control in Functional Logic Program Specialization? E. Albert1 , M. Alpuente1 , M. Falaschi2 , P. Juli´ an3 , and G. Vidal1 1 2

DSIC, U. Polit´ecnica de Valencia, Camino de Vera s/n, 46022 Valencia, Spain. {ealbert,alpuente,gvidal}@dsic.upv.es Dip. di Mat. e Informatica, U. Udine, Via delle Scienze 206, 33100 Udine, Italy. [email protected] 3 Dep. de Inform´ atica, Ronda de Calatrava s/n, 13071 Ciudad Real, Spain. [email protected]

Abstract. We have recently defined a framework for Narrowing-driven Partial Evaluation (NPE) of functional logic programs. This method is as powerful as partial deduction of logic programs and positive supercompilation of functional programs. Although it is possible to treat complex terms containing primitive functions (e.g. conjunctions or equations) in the NPE framework, its basic control mechanisms do not allow for effective polygenetic specialization of these complex expressions. We introduce a sophisticated unfolding rule endowed with a dynamic narrowing strategy which permits flexible scheduling of the elements (in conjunctions) which are reduced during specialization. We also present a novel abstraction operator which extends some partitioning techniques defined in the framework of conjunctive partial deduction. We provide experimental results obtained from an implementation using the Indy system which demonstrate that the control refinements produce better specializations.

1

Introduction

Functional logic programming languages allow us to integrate some of the best features of the classical declarative paradigms, namely functional and logic programming. Lazy, efficient, functional computations are combined with the expressivity of logic variables, which allows for function inversion as well as logical search. The operational semantics of functional logic languages is usually based on (some form of) narrowing, which is a unification-based, parameter-passing mechanism which extends functional evaluation through goal solving capabilities as in logic programming (see [14] for a survey). In order to avoid unnecessary computations and to compute with infinite data structures, most recent work has concentrated on lazy narrowing strategies [11,15,23,25]. The aim of partial evaluation (PE) is to specialize a given program w.r.t. part of its input data (hence, also called program specialization). PE techniques ?

This work has been partially supported by CICYT TIC 95-0433-C03-03, by HCM project CONSOLE, and by Acci´ on Integrada HA1997-0073.

G. Levi (Ed.): SAS’98, LNCS 1503, pp. 262–277, 1998. c Springer-Verlag Berlin Heidelberg 1998

Improving Control in Functional Logic Program Specialization

263

have been widely applied to the optimization of functional (see [16] and references therein) and logic programs [10,18,22,27]. Unfortunately, these techniques generally cannot be easily transferred to functional logic languages, since logical variables in function calls place specific demands that have to be tackled in order to achieve effective specialization. Narrowing-driven PE [4] (NPE) provides a general scheme for the specialization of functional logic languages. The method is formalized within the theoretical framework established in [22,24] for the PE of logic programs (also known as partial deduction, PD). However, a number of concepts have been generalized for dealing with features such as nested function calls, eager and lazy evaluation strategies and the standard optimization based on deterministically reducing functions. Control issues are managed by using standard techniques as in [24, 28]. The NPE method of [4] distinguishes a local and a global levels of control. At the local level, (finite) narrowing trees for (nested) function calls are constructed. At the global level, the calls extracted from the leaves of the local trees are considered for the next iteration of the algorithm, after a proper abstraction (generalization) that guarantees that only a finite number of calls is specialized. A close, automatic approach is that of positive supercompilation (PS) [29], whose basic transformation operation is driving, a unification-based transformation mechanism which is similar to (lazy) narrowing. A different PE method for a rewriting-based, functional logic language is considered in [19]. Classical PD computes partial evaluations for separate atoms independently. Recently, [12,21] have introduced a technique for the partial deduction of conjunctions of atoms. This technique achieves a number of program optimizations such as (some form of) tupling and deforestation which are usually obtained through more expensive fold/unfold transformations, which are difficult to automate and which cannot be obtained through classical PD. The NPE method of [4] is able to produce polygenetic specializations [13], i.e. it is able to extract specialized definitions which combine several function definitions of the original program. That means that NPE has the same potential for specialization as conjunctive PD or PS within the considered paradigm (a detailed comparison can be found in [5,6]). This is because the generic method of [4] may allow one to deal with equations and conjunctions during specialization by simply considering the equality and conjunction operators as primitive function symbols of the language. Unfortunately, the use of primitive functions may encumber the nature of the specialization problems and it often turns out that some form of tupling (as defined in [27] for logic programs) is required for specializing expressions which contain conjunctive calls. The NPE algorithm of [4] does not incorporate a specific treatment for such primitive symbols, which depletes many opportunities for reaching the closedness condition and forces the method to dramatically generalize calls, thus giving up the intended specialization (see Example 1). Inspired by the challenging results of conjunctive PD in [12], this paper extends [4,3] by formulating and experimentally testing concrete NPE control options that effectively handle primitive function symbols in lazy functional logic languages.

264

E. Albert et al.

Some of the original contributions of our paper are as follows: i) we introduce a well-balanced dynamic unfolding rule and a novel abstraction operator that do not depend on the narrowing strategy and which highly improve the specialization of the NPE method; ii) these options allow us to tune the specialization algorithm to handle conjunctions (and other expressions containing primitive functions, such as conditionals and strict equalities) in a natural way, which provides for polygenetic specialization without any ad-hoc artifice; and iii) our method is applicable to modern functional logic languages with a lazy narrowing semantics such as Babel [25], Curry [15] and Toy [8], thus giving a specialization method which subsumes both lazy functional and conventional logic program specialization. We demonstrate the quality of these improvements by specializing some examples which were not handled well by classical NPE. The control strategies have been tested in the NPE system Indy [2]. The structure of the paper is as follows. Section 2 contains basic definitions. Section 3 extends the NPE algorithm of [3] to care for the appropriate handling of primitive function symbols. The algorithm is still generic in that no concrete control options are settled (i.e., there is no commitment to any concrete unfolding rule or abstraction operator). In Sect. 4, the concrete control options are described by formalizing some appropriate unfolding and abstraction operators. We illustrate the usefulness of our approach through some simple examples. Preliminary performance results, given in Sect. 5, show the practical importance of the proposed strategies. Finally, Sect. 6 concludes the paper. An extended version of this paper containing more details and proofs can be found in [1].

2

Preliminaries

We briefly summarize some well-known results about rewrite systems and functional logic programs [9,14]. The definitions below are given in the homogeneous case. The extension to many-sorted signatures is straightforward [26]. Throughout this paper, X denotes a countably infinite set of variables and F denotes a set of function symbols (also called the signature), each of which has a fixed associated arity. We assume that the signature F is partitioned into two sets F = C ∪ D with C ∩ D = Ø. Symbols in C are called constructors and symbols in D are called defined functions. T (F, X ) denotes the set of terms or expressions built from F and X . T (F) denotes the set of ground terms, while T (C, X ) denotes the set of constructor terms. If t 6∈ X, then Head(t) is the function symbol heading term t, also called the root symbol of t. A pattern is a term of the form f (d1 , . . . , dn ) where f /n ∈ D and d1 , . . . , dn are constructor terms. Var(s) is the set of variables occurring in the syntactic object s. A substitution is a mapping from X to T (F, X ) s.t. its domain Dom(σ) = {x ∈ X | xσ 6≡ x} is finite. We frequently identify a substitution σ with the set {x 7→ xσ | x ∈ Dom(σ)}. We denote the identity substitution by id. We consider the usual preorder on substitutions ≤: θ is more general than σ (in symbols θ ≤ σ) iff ∃γ. σ ≡ θγ. The restriction σ|V of a substitution σ to a set

Improving Control in Functional Logic Program Specialization

265

V of variables is defined by σ|V = xσ if x ∈ V and σ|V = x if x 6∈V . We write σ =V θ iff σ|V = θ|V , and σ ≤V θ iff ∃γ. σγ =V θ. A term t is more general than s (or s is an instance of t), in symbols t ≤ s, if ∃σ. tσ ≡ s. A unifier of a pair of terms {t1 , t2 } is a substitution σ such that t1 σ ≡ t2 σ. A unifier σ is called most general unifier (mgu) if σ ≤ σ 0 for every other unifier σ 0 . A generalization of a set of terms {t1 , . . . , tn } is a pair ht, {θ1 , . . . , θn }i such that tθi = ti , i = 1, . . . , n. A generalization ht, Θi is the most specific generalization (msg) if t0 ≤ t for every other generalization ht0 , Θ0 i. Positions of a term t are represented by sequences of natural numbers used to address subterms of t. They are ordered by the prefix ordering: p ≤ q if there is w such that p.w = q, where p.w denotes the concatenation of sequences p and w. We let Λ denote the empty sequence. Pos(t) and FPos(t) denote, respectively, the set of positions and the set of nonvariable positions of the term t. t|p is the subterm of t at position p. t[s]p is the term t with the subterm at position p replaced with s. We find it useful to simplify our description by limiting the discussion to unconditional term rewriting systems. A rewrite rule is pair l → r with l, r ∈ T (F, X ), l 6∈ X, and Var(r) ⊆ Var(l). l and r are called the left-hand side (lhs) and right-hand side (rhs) of the rewrite rule, respectively. A term rewriting system (TRS) R is a finite set of rewrite rules. A rewrite step is an application of a rewrite rule to a term, i.e. t →p,l→r s if there exists a position p ∈ Pos(t), a rewrite rule l → r, and a substitution σ with t|p = lσ and s = t[rσ]p . We say that t|p is a redex (reducible expression) of t. A term t is reducible to term s if t →∗ s. A term t is irreducible or in normal form if there is no term s with t → s. A TRS is terminating if there are no infinite sequences of the form t1 → t2 → . . . A TRS is called confluent if, whenever a term s reduces to two terms t1 and t2 , both t1 and t2 reduce to the same term. Since we do not require terminating rules, normal forms may not exist. Functional Logic Programming The operational semantics of functional logic programs is usually based on (some variant) of narrowing. Essentially, narrowing consists of computing an appropriate substitution so that when applied to the current term, it becomes reducible, and then reducing it [14]. In this section, we briefly introduce a functional logic language whose syntax and demand-driven reduction mechanism is essentially equivalent to that of (a subset of) Babel [23,25], Toy [8], and Curry [15], which has been recently proposed to become a standard in the area. A TRS R is constructor-based (CB) if for each rule l → r ∈ R the lhs l is a pattern. A CB TRS R is weakly-orthogonal if R is left-linear (i.e., for each rule l → r ∈ R, the lhs l does not contain multiple occurrences of the same variable) and R contains only trivial overlaps (i.e., if l → r and l0 → r0 are variants of distinct rules in R and σ is a unifier for l and l0 , then rσ ≡ r0 σ). It is well-known that weakly-orthogonal TRS’s are confluent. We henceforth consider CB weakly-orthogonal TRS’s as programs. For this class of programs, a term t is a head normal form if t is a variable or Head(t) ∈ C.

266

E. Albert et al.

The signature F is augmented with a set of primitive function symbols P = {≈, ∧, ⇒} in order to handle complex expressions containing equations s ≈ t, conjunctions b1 ∧b2 , and conditional (guarded) terms b ⇒ t, i.e. F = C∪D∪P. We usually treat the symbols in P as infix operators. We assume that the following predefined rules belong to any given program: c ≈ c → true % c/0 ∈ C c(x1 , . . . , xn ) ≈ c(y1 , . . . , yn ) → (x1 ≈ y1 ) ∧ . . . ∧ (xn ≈ yn ) % c/n ∈ C true ∧ x → x x ∧ true → x (true ⇒ x) → x These rules are weakly-orthogonal and define the validity of an equation as a strict equality between terms, which is common in functional languages when computations may not terminate [11,25]. Note that, although the basic computation model only supports unconditional rules, it is still adequate to support logic programs since conditional rewrite rules l → r ⇐ C can be encompassed by guarded unconditional rules l → (C ⇒ r) by using the conditional primitive ‘⇒’ as in Babel [25]. For reasons of simplicity, we assume the associativity of ‘∧’ and assume that ‘≈’ binds more than ‘∧’ and ‘∧’ binds more than ‘⇒’. We consider that programs are executed by lazy narrowing, which allows us to deal with nonterminating functions [23,25]. Roughly speaking, laziness means that a given expression is only narrowed at inner positions if they are demanded (by the pattern in the lhs of some rule) and this contributes to a later narrowing step at an outer position. Formally, given a program R, we define the one-step narrowing relation as follows. A term s narrows to t in R, in symbols s ;p,l→r,σ t (or simply s ;σ t), iff there exists a position p ∈ ϕ(s), a (standardized apart) rule l → r ∈ R, and a substitution σ such that σ = mgu({s|p , l}) and t = (s[r]p )σ. The selection strategy ϕ(t) is responsible for computing the set of demanded positions of a given term t. A formal definition of this strategy in terms of an inference system is shown in [1]. Lazy narrowing is strong complete w.r.t. constructor substitutions in CB, weakly-orthogonal TRS’s [25, 14]. This means that the interpreter is free to disregard from ϕ(t) all components of each conjunction which may occur in t except one, even if all arguments are demanded by the predefined rules of ‘∧’ (that is, completeness holds for all scheduling policies). A formal definition can be found in [3]. If s0 ;σ1 s1 ;σ2 . . . ;σn sn (in symbols, s0 ;∗σ sn , σ = σ1 σ2 . . . σn ), we speak of a lazy narrowing derivation for the goal s0 with (partial) result sn . A lazy narrowing derivation s ;∗σ t is successful iff t ∈ T (C ∪ X ), where σ|Var(s) is the computed answer substitution.

3

The Generalized Specialization Algorithm

In this section, we generalize some basic concepts and techniques for the NPE of (lazy) functional logic programs (as presented in [3]). These extended notions will prove to be extremely useful for formulating new unfolding and abstraction operators which are well-suited to cope with primitive function symbols. In the original NPE framework, no distinction is made between primitive and defined function symbols during specialization. For instance, a conjunction b1 ∧ b2 is

Improving Control in Functional Logic Program Specialization

267

sorted bits(x : xs) ∧ 1 ≤ x

(( (((( AA ( ( ( (

true ∧ 1 ≤ x

sorted bits(x0 : xs0 ) ∧ x ≤ x0 ∧ 1 ≤ x

... Fig. 1. Incomplete narrowing tree for sorted bits(x : xs) ∧ 1 ≤ x

considered as a single entity when checking whether it is covered by the set of specialized calls. This commonly implies a drastic generalization of the involved calls, which causes losing all specialization, as the following example illustrates. Example 1. Let us consider the program excerpt: sorted bits(x : [ ]) → true sorted bits(x1 : x2 : xs) → sorted bits(x2 : xs) ∧ x1 ≤ x2 0 ≤ 0 → true 0 ≤ 1 → true 1 ≤ 1 → true and the call sorted bits(x : xs) ∧ 1 ≤ x. The lazy narrowing tree depicted1 in Fig. 1 is built up by using the nonembedding unfolding rule of [3], which expands derivations while new redexes are not “greater” (with the homeomorphic embedding ordering, see e.g. [4,28]) than previous, comparable redexes in the branch (i.e., redexes with the same outermost function symbol). From this tree, we can identify two main weaknesses of the plain NPE algorithm: – The rightmost branch stops because the leftmost redex sorted bits(x0 : xs0 ) “embeds” the previous redex sorted bits(x : xs), even if no reductions have been performed on the other elements of the conjunction. – At the global level, since the call sorted bits(x0 : xs0 ) ∧ x ≤ x0 ∧ 1 ≤ x in the leaf of the tree embeds (but does not cover) the specialized call sorted bits(x : xs) ∧ 1 ≤ x (and they are comparable), the msg sorted bits(x : xs) ∧ z is computed, which gives up the intended specialization. The first drawback pointed out in this example motivates the definition of more sophisticated unfolding rules which are able to achieve a balanced evaluation of the given expression by narrowing appropriate redexes (e.g., by using some kind of dynamic scheduling strategy which takes into account the ancestors narrowed in the same branch). The second drawback suggests the definition of a more flexible abstraction operator which is able to automatically split complex terms before attempting folding or generalization. In the following, we refine the framework of [3] in order to overcome these problems. The PE of a term is obtained by constructing a (partial) narrowing tree and then extracting the resultants associated to the root-to-leaf paths of the tree. Definition 1 (resultant). Let s be a term and R be a program. Given a lazy narrowing derivation s ;∗σ t, its associated resultant is the rewrite rule sσ → t. 1

We assume a fixed left-to-right selection of components within conjunctions and underline the selected redex at each step.

268

E. Albert et al.

Definition 2 (partial evaluation). Let R be a program and s be a term. Let τ be a finite (possibly incomplete) narrowing tree for s in R such that no goal in the tree is narrowed beyond its head normal form. Let {t1 , . . . , tn } be the terms in the leaves of τ . Then, the set of resultants for the narrowing sequences {s ;+ σi ti | i = 1, . . . , n} is called a partial evaluation of s in R. The partial evaluation of a set of terms S in R is defined as the union of the partial evaluations for the terms in S. Intuitively, the reason for requiring that the PE of a term s do not surpass its head normal form is that, at runtime, the evaluation of a (nested) call C[s]p containing the partially evaluated term s at some position p might not demand evaluating s beyond its head normal form. Since the “contexts” C[ ] in which s will appear are not known at PE time, we avoid to interfering with the “lazy nature” of computations in the specialized program by imposing this condition. A recursive closedness condition is formalized by inductively checking that the different calls in the rules are sufficiently covered by the specialized functions. Definition 3 (closedness). Let S be a finite set of terms. A term t is S-closed if closed(S, t) holds,where the predicate closed is defined inductively as follows: if t ∈ X true closed(S, t ) ∧ . . . ∧ closed(S, t ) if t ≡ c(t1 , . . . , tn ) 1 n ^ closed(S, t) ⇔ + 0 ∃s ∈ S . sθ = t ∧ closed(S, t ) if t ≡ f (t1 , . . . , tn ) +

x/t0 ∈θ

where c ∈ C, f ∈ (D ∪ P), and S = S ∪ {p(x, y) | p ∈ P}. We say that a set of terms T is S-closed, written closed(S, T ), if closed(S, t) holds for all t ∈ T , and we say that a program R is S-closed if closed(S, Rcalls ) holds. Here we denote by Rcalls the set of terms in the rhs’s of the rules in R. The novelty w.r.t. [4,3] is that a complex expression headed by a primitive function symbol, such as a conjunction, is proved closed w.r.t. S either by checking that it is an instance of a call in S (followed by an inductive test of the subterms), or by splitting it into two conjuncts and then trying to match with “simpler” terms in S (which happens when matching is first attempted w.r.t. one of the ‘flat’ calls p(x, y) in S + ). This extension is safe since the rules which define the primitive functions are automatically added to each program. The way in which a concrete PE is made is given by an unfolding rule (which decides how to stop the construction of lazy narrowing trees) and an abstraction operator (which ensures the finiteness of the set of specialized calls). Definition 4 (unfolding rule [3]). An unfolding rule U is a mapping which, when given a program R and a term s, returns a concrete PE for s in R (a set of resultants). By U (S, R) we denote the union of U (s, R) for all s ∈ S. Definition 5 (abstraction operator). Given a finite set of terms T and a set of terms S, an abstraction operator is a function which returns a finite set of terms abstract(S, T ) such that: i) if s ∈ abstract(S, T ), then there exists t ∈ (S ∪ T ) such that t|p = sθ for some position p and substitution θ; ii) for all t ∈ (S ∪ T ), t is closed w.r.t. the set of terms in abstract(S, T ).

Improving Control in Functional Logic Program Specialization

269

Roughly speaking, the first condition guarantees that the abstraction operator does not introduce new function symbols, while the second condition ensures that the resulting set of terms “covers” the calls previously specialized and that closedness is preserved throughout successive abstractions. The following basic algorithm for NPE is parameterized by the unfolding rule U and the abstraction operator abstract in the style of [10]. Algorithm 1.

Input: a program R and a set of terms T Output: a set of terms S Initialization: i := 0; T0 := T Repeat

1. R0 := U (Ti , R); 2. Ti+1 := abstract(Ti , R0calls ); 3. i := i + 1; Until Ti = Ti−1 (modulo renaming) Return S := Ti

The output of the NPE algorithm, given a program R, is not a PE, but a set of terms S from which the partial evaluations U (S, R) are automatically derived. Note that, whenever the specialized call is not a pattern, lhs’s of resultants are not patterns either and hence resultants are not (CB) program rules. In [3], we introduced a post-processing renaming which is useful for producing CB rules. Roughly speaking, we construct an “independent renaming” S 0 of S as follows: for each term s in S, we define its independent renaming s0 = fs (x1 , . . . , xn ), where x1 , . . . , xn are the distinct variables in s in the order of their first occurrence and fs is a new fresh function symbol. Then, we fold each call t in the rules of U (S, R) by replacing the old call t by a call to the corresponding term t0 in S 0 (details can be found in [3]). After the algorithm terminates, the specialized program is obtained by applying this post-processing renaming to U (S, R). The (partial) correctness of the NPE algorithm is stated as follows. Theorem 2. Given a program R and a term t, if Algorithm 1 terminates by computing the set of terms S, then R0 and t are S-closed, where R0 = U (S, R). The correctness of the generic algorithm is stated in the following theorem, which generalizes Theorem 4.5 of [3]. Theorem 3. Let R be a program, t a term, and S a finite set of terms. Let R0 be a PE of R w.r.t. S such that R0 and t are S-closed. Let S 0 be an independent renaming of S, and t00 (resp. R00 ) be a renaming of t (resp. R0 ) w.r.t. S 0 . Then t computes in R the result d with computed answer θ iff t00 computes in R00 the result d with computed answer θ0 and θ0 ≤Var(t) θ.

4

Improving Control of NPE

In Sect. 4.1 we improve control in functional logic specialization by fixing an unfolding strategy which is specifically designed for “conjunctive specialization”.

270

E. Albert et al. app(x, y) ≈ w ∧ app(w, z) ≈ r

((aa (((( aa ( ( ( ( a

.. .

x0 : app(xs0 , y) ≈ w ∧ app(w, z) ≈ r

x0 ≈ w0 ∧ app(xs0 , y) ≈ ws0 ∧ app(w0 : ws0 , z) ≈ r true ∧ app(xs0 , y) ≈ ws0 ∧ app(w0 : ws0 , z) ≈ r app(xs0 , y) ≈ ws0 ∧ app(w0 : ws0 , z) ≈ r Fig. 2. Na¨ıve local control for app(x, y) ≈ w ∧ app(w, z) ≈ r

As for global control, a specific treatment of the primitive function symbols ‘≈’, ‘∧’ and ‘⇒’ is introduced in Sect. 4.2 which produces more effective and powerful, polygenetic specializations, as compared to classical NPE. 4.1

Local Control

The unfolding rule of [3] simply exploits the redexes selected by the lazy narrowing strategy ϕ (using a static selection rule which determines the next conjunct to be reduced) whenever none of them embed a previous (comparable) redex. The following example reveals that this strategy is not elaborated enough for specializing calls which may contain primitive symbols like conjunctions. Example 2. Consider the well-known program append: app([ ], y) → y app(x : xs, y) → x : app(xs, y) with the input goal “app(x, y) ≈ w ∧ app(w, z) ≈ r”. Using the nonembedding unfolding rule of [3], we obtain the tree depicted2 in Fig. 2 (using a fixed leftto-right selection rule for conjunctions). From this local tree, no appropriate specialized definition for the initial goal can be obtained, since the leaf cannot be folded into the input call in the root of the tree and generalization is required (which causes losing all specialization, as in Example 1). We note that the NPE method [4,3] succeeds with this example when the specialized call is written as a nested expression app(app(x, y), z). This is because exploiting the nesting capability of the functional syntax allows us to transform the original tupling3 problem illustrated by Example 2 into a simpler, deforestation problem, which is easily solved by the original NPE method. 2 3

We adopt the standard optimization which makes use of the built-in unification to solve strict equalities s ≈ t (provided that only constructor bindings are produced). Here we refer to tupling of logic programs, which subsumes both deforestation and tupling of functional programs [27].

Improving Control in Functional Logic Program Specialization

271

Now we introduce a dynamic lazy unfolding rule which attempts to achieve a fair evaluation of the complete input term, rather than a deeper evaluation of some given subterm. This novel concrete unfolding rule dynamically selects the positions to be reduced within a given conjunction, by exploiting some dependency information between redexes gathered along the derivation. Definition 6 (dependent positions). Let D ≡ (s ;p,l→r,σ t) be a narrowing step. The set of dependent positions of a position q of s by D, written q\\D, is: {q.u | u ∈ FPos(r) ∧ Head(r|u ) 6∈ C} if q = p if q 6≤p q\\D = {q} {p.u0 .v | r|u0 = x} if q = p.u.v and l|u = x ∈ X This notion can be naturally lifted to narrowing derivations. The notion of dependency for terms stems directly from the corresponding notion for positions. Note that the above definition is a strict extension of the standard notion of descendant in functional programming. Intuitively, the descendants of the subterm s|q are computed as follows: underline the root symbol of s|q and perform the narrowing step s ; t. Then, every subterm of t with and underlined root symbol is a descendant of s|q [17]. Intuitively, a position q 0 of t depends on a position q of s (by D) if q 0 is a descendant of q (second and third cases), or if the position q 0 has been introduced by the rhs of the rule applied in the reduction of the former position q and it addresses a subterm headed by a defined function symbol (first case). Note that this notion is an extension of the standard PD concept of (covering) ancestor to the functional logic framework. By abuse, we also say that the term addressed by q is an ancestor of the term addressed by q 0 in D. If s is an ancestor of t and Head(s) = Head(t), we say that s is a comparable ancestor of t in D. Now we formalize the way in which the dynamic selection is performed. Definition 7 (dynamic narrowing selection strategy). Let D ≡ (t0 ; t1 ; . . . ; tn ), n ≥ 0 be a lazy narrowing derivation. We define the dynamic selection rule ϕdynamic as: ϕdynamic (tn , D) = select(tn , Λ, D), where the auxiliary function select is: select(t, p, D) = if p ∈ ϕ(t) then if dependency clash(t|p , D) then {⊥} else {p} else case t|p of x ∈ V: Ø s1 ∧ s2 : let Oi = select(t, p.i, D), i ∈ {1, 2}, in [if ∃i. (⊥ 6∈Oi ∧ Oi 6≡Ø) then Oi else if (O1 ≡ O2 ≡ Ø) then Ø else {⊥}] otherwise: let t|p = S f (s1 , . . . , sn ) and n Oargs = i=1 select(t, p.i, D) in [if ⊥ ∈ Oargs then {⊥} else Oargs ] where dependency clash(t, D) is a generic Boolean function that looks at the ancestors of t in D to determine whether there is a risk of nontermination.

272

E. Albert et al. app(x, y) ≈ w ∧ app(w, z) ≈ r

(((aa ((((( aa ( ( ( ( ( a

.. .

x0 : app(xs0 , y) ≈ w ∧ app(w, z) ≈ r

x0 ≈ w0 ∧ app(xs0 , y) ≈ ws0 ∧ app(w0 : ws0 , z) ≈ r x0 ≈ w0 ∧ app(xs0 , y) ≈ ws0 ∧ w0 : app(ws0 , z) ≈ r x0 ≈ w0 ∧ app(xs0 , y) ≈ ws0 ∧ w0 ≈ r0 ∧ app(ws0 , z) ≈ rs0 Fig. 3. Improved local control for app(x, y) ≈ w ∧ app(w, z) ≈ r

For the sake of simplicity, in the remainder of this section we consider that dependency clash(t, D) holds whenever there is a comparable ancestor of the selected redex t in D. Another approach, that we investigate in the experiments, is to additionally test homeomorphic embedding on comparable ancestors. Roughly speaking, the dynamic selection strategy recurs over the structure of the goal and determines the set of positions to be unfolded by a don’t-care selection within each conjunction of exactly one of the components (among those that do not incur into a dependency clash). We introduce a dynamic unfolding rule Udynamic (t, R) which simply expands lazy narrowing trees according to the dynamic lazy narrowing strategy ϕdynamic . The “mark” ⊥ of Definition 7 is used as a whistle to warn us that the derivation must be cut off since it runs into a dependency clash, i.e., each branch D of the tree is stopped whenever ϕdynamic (t, D) = {⊥} or the term t (of the leaf) is in head normal form. Example 3. Consider again the program and goal of Example 2. Using the dynamic unfolding rule Udynamic , we get the tree depicted in Fig. 3. From this tree, an optimal (recursive) specialized definition for the initial call can be derived, provided there is a suitable splitting mechanism to extract, from the leaf of the tree, an appropriate subconjunction such as app(xs0 , y) ≈ ws0 ∧app(ws0 , z) ≈ rs0 , which is covered by the initial call in the root of the tree (see Example 4). 4.2

Global Control

In the presence of primitive functions like ‘∧’ or ‘≈’, using an abstraction operator which respects the structure of the terms (as in [3]) is not very effective, since the generalization of two conjunctions (resp. equations) might be a term of the form x ≈ y ∧ z (resp. x ≈ y) in most cases. The drastical solution of decomposing the term into subterms containing just one function call can avoid the problem, but has the negative consequence of losing nearly all specialization. In this section, we introduce a more concerned abstraction operator which is inspired by the partitioning techniques of conjunctive PD [12,21], and which uses the homeomorphic embedding relation E as defined in [28].

Improving Control in Functional Logic Program Specialization

273

The following notion of best matching terms, which is aimed at avoiding loss of specialization due to generalization, is a proper generalization of the notion of best matching conjunction in [12]. Definition 8 (best matching terms). Let S = {s1 , . . . , sn } be a set of terms, t a term, and consider the set of terms W = {wi | hwi , {θi1 , θi2 }i = msg({si , t}), i = 1, . . . , n}. The best matching terms BM T (S, t) for t in S are those terms sj ∈ S such that the corresponding wj in W is a minimally general element. The notion of BMT is used in the abstraction process at two stages: i) when selecting the more appropriate term in S which covers a new call t, and ii) when determining whether a call t headed by a primitive function symbol could be (safely) added to the current set of specialized calls or should be split. Definition 9 (concrete abstraction operator). Let S and T be sets of terms. We define abstractE (S, T ) inductively as follows: abstractE (S, T ) = S if T ≡ Ø or T ≡ {t}, t ∈ X abstractE (. . . abstractE (S, t1 ), . . . , tn ) if T ≡ {t1 , . . . , tn }, n > 0 abstractE (S, {t1 , . . . , tn }) if T ≡ {t}, t ≡ c(t1 , . . . , tn ), c ∈ C 0 abs def (S, T , t) if T ≡ {t}, Head(t) ∈ D if T ≡ {t}, Head(t) ∈ P abs prim(S, T 0 , t) where T 0 = {s ∈ S | Head(s) = Head(t) ∧ s E t}. The functions abs def and abs prim are defined as follows: abs def (S, Ø, t) = abs prim(S, Ø, t) = S ∪ {t} abs def (S, T, t) = abstractE (S \ {s}, {w} ∪ Ran(θ1 ) ∪ Ran(θ2 )) if hw, {θ1 , θ2 }i = msg({s, t}), with s ∈ BM T (T, t) abs def (S, T, t) if ∃s ∈ BM T (T, t) s.t. def (t) = def (s) abs prim(S, T, t) = ∧ abstractE (S, T, {t1 , t2 }) otherwise, where t = p(t1 , t2 ) where def (t) denotes a sequence with the defined function symbols of t in lexico∧ graphical order, and = is equality up to reordering of elements in a conjunction. Essentially, the way in which the abstraction operator proceeds is simple. We distinguish the cases when the considered term i) is a variable, ii) is headed by a constructor symbol, iii) by a defined function symbol, or iv) by a primitive function symbol. The actions that the abstraction operator takes, respectively, are: i) to ignore it, ii) to recursively inspect the subterms, iii) to generalize the given term w.r.t. some of its best matching terms (recursively inspecting the msg w and the subterms of θ1 , θ2 not covered by the generalization), and iv) the same as in iii), but considering the possibility of splitting the given expression before generalizing it when def (t) 6= def (s) (which essentially states that some defined function symbols would be lost due to the application of msg). The function abstractE is an abstraction operator in the sense of Definition 5 [1]. The following result establishes the termination of the global specialization process Theorem 4. Algorithm 1 terminates for the unfolding rule Udynamic and the abstraction operator abstractE .

274

E. Albert et al.

Our final example witnesses that abstractE behaves well w.r.t. Example 3. Example 4. Consider again the tree depicted in Fig. 3. By applying Algorithm 1, the following call to abstractE is undertaken: abstractE ({app(x, y) ≈ w ∧ app(w, z) ≈ r}, {x0 ≈ w0 ∧ app(xs0 , y) ≈ ws0 ∧ w0 ≈ r0 ∧ app(ws0 , z) ≈ rs0 }). Following Definition 9, by two recursive calls to abs prim, we get: {app(x, y) ≈ w ∧ app(w, z) ≈ r, x0 ≈ w0 }. By considering the independent renaming dapp(x, y, w, z, r) for the specialized call app(x, y) ≈ w ∧ app(w, z) ≈ r, the method derives a (recursive) rule of the form: dapp(x : xs, y, w : ws, z, r : rs) → x ≈ w ∧ w ≈ r ∧ dapp(xs, y, ws, z, rs), which embodies the intended optimal specialization for this example.

5

Experiments

The refinements presented so far have been incorporated into the NPE prototype implementation system Indy (Integrated Narrowing-Driven specialization system [2]). Indy is written in SICStus Prolog v3.6 and is publicly available [2]. In order to assess the practicality of our approach, we have benchmarked the speed and specialization achieved by the extended implementation. The benchmarks used for the analysis are: applast, which appends an element at the end of a given list and returns the last element of the resulting list; double app, see Example 2; double flip, which flips a tree structure twice, then returning the original tree back; fibonacci, fibonacci’s function; heads&legs, which computes the number of heads and legs of a given number of birds and cats; match-app, the extremely na¨ıve string pattern matcher based on using append; match-kmp, a semi-na¨ıve string pattern matcher; maxlength, which returns the maximum and the length of a list; palindrome, a program to check whether a given list is a palindrome; and sorted bits, see Example 1. Some of the examples are typical PD benchmarks (see [20]) adapted to a functional logic syntax, while others come from the literature of functional program transformations, such as PS [29], fold/unfold transformations [7], and deforestation [30]. We have considered the following settings to test the benchmarks: – Evaluation strategy: All benchmarks were executed by lazy narrowing. – Unfolding rule: We have tested three different alternatives: (1) emb goal: it expands derivations while new goals do not embed a previous comparable goal in the same branch; (2) emb redex: the concrete unfolding rule of Sect. 4.1 which implements the dependency clash test using homeomorphic embedding on comparable ancestors of selected redexes to ensure finiteness (note that it differs from emb goal in that emb redex implements dynamic scheduling on conjunctions and that homeomorphic embedding is checked on simple redexes rather than on whole goals); (3) comp redex: the unfolding rule of Sect. 4.1 which uses the simpler definition of dependency clash based on comparable ancestors of selected redexes as a whistle. – Abstraction operator: Abstraction is always done as explained in Def. 9.

Improving Control in Functional Logic Program Specialization

275

Table 1. Benchmark results Benchmarks applast double app double flip fibonacci heads&legs match-app match-kmp maxlength palindrome sorted bits Average TMix average

Original Rw RT 10 90 8 106 8 62 5 119 8 176 8 201 12 120 14 94 10 119 8 110 9.1 119.7

emb goal Rw Speedup 13 1.32 39 1.63 26 1.51 11 1.19 24 4.63 12 1.25 14 3.43 51 1.17 19 1.25 16 1.15 22.5 1.85 1881

emb redex Rw Speedup 28 2.20 61 1.28 17 1.55 7 1.08 22 2.41 20 2.75 14 3.64 20 1.27 10 1.35 31 2.89 23 2.04 7441

comp redex Rw Speedup 13 1.10 15 3.12 17 1.55 7 1.08 21 2.48 23 2.79 13 3.43 18 1.25 10 1.35 10 2.68 14.7 2.08 5788

Table 1 summarizes our benchmark results. The first two columns measure the number of rewrite rules (Rw) and the absolute runtimes (RT) for each original program. The other columns show the number of rewrite rules and the speedups achieved for the specialized programs obtained by using the three considered unfolding rules. The row at the bottom of the table (TMix) indicates the average specialization time for each considered unfolding rule. Times are expressed in milliseconds and are the average of 10 executions. Speedups were computed by running the original and specialized programs under the publicly available lazy functional logic language Toy [8]. Runtime input goals were chosen to give a reasonably long overall time. The complete code for benchmarks, the specialized calls, and the partially evaluated programs can be found in [1]. The figures in Table 1 demonstrate that the control refinements that we have incorporated into the Indy system provide satisfactory speedups on all benchmarks (which is very encouraging, given the fact that no partial input data were provided in any example, except for match-app, match-kmp, and sorted bits). On the other hand, our extensions are conservative in the sense that there is no penalty w.r.t. the specialization achieved by the original system on nonconjunctive goals (although some specialization times are slightly higher due to the more complex processing being done). Let us note that, from the speedup results in Table 1, it can appear that there is no significant difference between the strategies emb redex and comp redex. However, when we also consider the specialization times (TMix) and the size of the specialized programs (Rw), we find out that comp redex has a better overall behaviour. A detailed comparison between the considered unfolding strategies can be found in [1].

6

Discussion

In functional logic languages, expressions can be written by exploiting the nesting capability of the functional syntax, as in app(app(x, y), z) ≈ r, but in many cases

276

E. Albert et al.

it can be appropriate (or necessary) to decompose nested expressions as in logic programming, and write app(x, y) ≈ w ∧ app(w, z) ≈ r (for instance, if some test such as sorted bits(w) on the intermediate list w were necessary). The original Indy system behaves well on programs written with the “pure” functional syntax [5]. However, Indy is not able to produce good specialization on the benchmarks of Table 1 when they are written as conjunctions of subgoals. For this we could not achieve some of the standard, difficult transformations such as tupling [7,27] within the classical NPE framework. As opposed to the classical PD framework (in which only folding on single atoms can be done), the NPE algorithm is able to perform folding on complex expressions (containing an arbitrary number of function calls). This does not suffice to achieve tupling in practice, since complex expressions are often generalized and specialization is lost. We have shown that the NPE general framework can be supplied with appropriate control options to specialize complex expressions containing primitive functions, thus providing a powerful polygenetic specialization framework with no ad-hoc setting. To the best of our knowledge, this is the first practical framework for the specialization of modern functional logic languages with partitioning techniques and dynamic scheduling. As future research, there is room for further improvement in performance by introducing more powerful abstraction operators based on better analyses to determine the optimal way to split expressions (trying not to endanger the communication of data structures with shared variables), and by considering in practice the problem of controlling particular algebraic laws for primitive symbols.

References 1. E. Albert, M. Alpuente, M. Falaschi, P. Juli´ an, and G. Vidal. Improving Control in Functional Logic Program Specialization. Technical Report DSIC-II/2/97, UPV, 1998. Available from URL: http://www.dsic.upv.es/users/elp/papers.html. 2. E. Albert, M. Alpuente, M. Falaschi, and G. Vidal. Indy User’s Manual. Technical Report, available from http://www.dsic.upv.es/users/elp/papers.html. 3. M. Alpuente, M. Falaschi, P. Juli´ an, and G. Vidal. Specialization of Lazy Functional Logic Programs. In Proc. of PEPM’97, volume 32(12) of Sigplan Notices, pages 151–162, New York, 1997. ACM Press. 4. M. Alpuente, M. Falaschi, and G. Vidal. Narrowing-driven Partial Evaluation of Functional Logic Programs. In H. Riis Nielson, editor, Proc. of the 6th European Symp. on Programming, ESOP’96, pages 45–61. Springer LNCS 1058, 1996. 5. M. Alpuente, M. Falaschi, and G. Vidal. Partial Evaluation of Functional Logic Programs. ACM TOPLAS, 1998. To appear. 6. M. Alpuente, M. Falaschi, and G. Vidal. A Unifying View of Functional and Logic Program Specialization. ACM Computing Surveys, 1998. To appear. 7. R.M. Burstall and J. Darlington. A Transformation System for Developing Recursive Programs. Journal of the ACM, 24(1):44–67, 1977. 8. R. Caballero-Rold´ an, F.J. L´ opez-Fraguas, and J. S´ anchez-Hern´ andez. User’s manual for Toy. Technical Report SIP-5797, UCM, Madrid (Spain), April 1997. 9. N. Dershowitz and J.-P. Jouannaud. Rewrite Systems. In J. van Leeuwen, editor, Handbook of Theoretical Computer Science, volume B: Formal Models and Semantics, pages 243–320. Elsevier, Amsterdam, 1990.

Improving Control in Functional Logic Program Specialization

277

10. J. Gallagher. Tutorial on Specialisation of Logic Programs. In Proc. of PEPM’93, pages 88–98. ACM, New York, 1993. 11. E. Giovannetti, G. Levi, C. Moiso, and C. Palamidessi. Kernel Leaf: A Logic plus Functional Language. J. of Computer and System Sciences, 42:363–377, 1991. 12. R. Gl¨ uck, J. Jørgensen, B. Martens, and M.H. Sørensen. Controlling Conjunctive Partial Deduction of Definite Logic Programs. In Proc. of PLILP’96, pages 152– 166. Springer LNCS 1140, 1996. 13. R. Gl¨ uck and M.H. Sørensen. A Roadmap to Metacomputation by Supercompilation. In Partial Evaluation, Int’l Seminar, Dagstuhl Castle, Germany, pages 137–160. Springer LNCS 1110, February 1996. 14. M. Hanus. The Integration of Functions into Logic Programming: From Theory to Practice. Journal of Logic Programming, 19&20:583–628, 1994. 15. M. Hanus, H. Kuchen, and J.J. Moreno-Navarro. Curry: A Truly Functional Logic Language. In Proc. ILPS’95 Workshop on Visions for the Future of Logic Programming, pages 95–107, 1995. 16. N.D. Jones, C.K. Gomard, and P. Sestoft. Partial Evaluation and Automatic Program Generation. Prentice-Hall, Englewood Cliffs, NJ, 1993. 17. J.W. Klop and A. Middeldorp. Sequentiality in Orthogonal Term Rewriting Systems. Journal of Symbolic Computation, pages 161–195, 1991. 18. J. Komorowski. An Introduction to Partial Deduction. In A. Pettorossi, editor, Meta-Programming in Logic, pages 49–69. Springer LNCS 649, 1992. 19. L. Lafave and J.P. Gallagher. Partial Evaluation of Functional Logic Programs in Rewriting-based Languages. Technical Report CSTR-97-001, Department of Computer Science, University of Bristol, Bristol, England, March 1997. 20. M. Leuschel. The ecce partial deduction system and the dppd library of benchmarks. Tech. Rep., accessible via http://www.cs.kuleuven.ac.be/˜lpai, 1998. 21. M. Leuschel, D. De Schreye, and A. de Waal. A Conceptual Embedding of Folding into Partial Deduction: Towards a Maximal Integration. In M. Maher, editor, Proc. of JICSLP’96, pages 319–332. The MIT Press, Cambridge, MA, 1996. 22. J.W. Lloyd and J.C. Shepherdson. Partial Evaluation in Logic Programming. Journal of Logic Programming, 11:217–242, 1991. 23. R. Loogen, F. L´ opez-Fraguas, and M. Rodr´ıguez-Artalejo. A Demand Driven Computation Strategy for Lazy Narrowing. In J. Penjam and M. Bruynooghe, editors, Proc. of PLILP’93, pages 184–200. Springer LNCS 714, 1993. 24. B. Martens and J. Gallagher. Ensuring Global Termination of Partial Deduction while Allowing Flexible Polyvariance. In L. Sterling, editor, Proc. of ICLP’95, pages 597–611. MIT Press, 1995. 25. J.J. Moreno-Navarro and M. Rodr´ıguez-Artalejo. Logic programming with functions and predicates: the language Babel. J. Logic Program., 12(3):191–224, 1992. 26. P. Padawitz. Computing in Horn Clause Theories, volume 16 of EATCS Monographs on Theoretical Computer Science. Springer-Verlag, Berlin, 1988. 27. A. Pettorossi and M. Proietti. Rules and Strategies for Transforming Functional and Logic Programs. ACM Computing Surveys, 28(2):360–414, 1996. 28. M.H. Sørensen and R. Gl¨ uck. An Algorithm of Generalization in Positive Supercompilation. In Proc. of ILPS’95, pages 465–479. The MIT Press, 1995. 29. M.H. Sørensen, R. Gl¨ uck, and N.D. Jones. A Positive Supercompiler. Journal of Functional Programming, 6(6):811–838, 1996. 30. P.L. Wadler. Deforestation: Transforming programs to eliminate trees. Theoretical Computer Science, 73:231–248, 1990.

Directional Type Inference for Logic Programs Witold Charatonik? and Andreas Podelski Max-Planck-Institut f¨ ur Informatik Im Stadtwald, D-66123 Saarbr¨ ucken {witold;podelski}@mpi-sb.mpg.de

Abstract. We follow the set-based approach to directional types proposed by Aiken and Lakshman [1]. Their type checking algorithm works via set constraint solving and is sound and complete for given discriminative types. We characterize directional types in model-theoretic terms. We present an algorithm for inferring directional types. The directional type that we derive from a logic program P is uniformly at least as precise as any discriminative directional type of P, i.e., any directional type out of the class for which the type checking algorithm of Aiken and Lakshman is sound and complete. We improve their algorithm as well as their lower bound and thereby settle the complexity (Dexptime-complete) of the corresponding problem.

1

Introduction

Directional types form a type system for logic programs which is based on the view of a predicate as a directional procedure which, when applied to a tuple of input terms, generates a tuple of output terms. There is a rich literature on types and directional types for which we can give only some entry points. Directional types occur as predicate profiles in [28], as mode dependencies in [8], and simply as types in [4,2,3]. Our use of the terminology “directional type” stems from [1]. A directional type for a program P assigns input types Ip and output types Op to each predicate p of P. A program can have many directional types. For example, consider the predicate append defined by append([], L, L). append([X|Xs], Y, [X|Z])←append(Xs, Y, Z). We can give this predicate the directional type (list, list, >) → (list, list, list), where list denotes the set of all lists and > is the set of all terms, but also (>, >, list) → (list, list, list), as well as the least precise type (>, >, >) → (>, >, >). A predicate defined by a single fact p(X) has a directional type τ → τ for all types τ . In [1], Aiken and Lakshman present an algorithm for automatic type checking of logic programs wrt. given directional types. The algorithm runs in ?

On leave from University of Wroclaw, Poland. Partially supported by Polish KBN grant 8T11C02913.

G. Levi (Ed.): SAS’98, LNCS 1503, pp. 278–294, 1998. c Springer-Verlag Berlin Heidelberg 1998

Directional Type Inference for Logic Programs

279

NEXPTIME; they show that the problem is DEXPTIME-hard in general and PSPACE-hard for discriminative types. The algorithm works via set constraint solving; its correctness relies on a connection between the well-typedness conditions and the set constraints to which they are translated. The connection is such that the type check is sound and complete for discriminative types (it is still sound for general types). Our results. In this paper, we answer two questions left open in [1]. First, we give an algorithm for inferring directional types. Second, we establish the DEXPTIME-completeness of the problem of directional type checking wrt. discriminative types. To fix just one (the “right” one) directional type for a given logic program, we assume that the program comes with a query which, wlog., consists of one atom main(t). Clearly, the choice of the type makes sense only if the input type Imain for the query predicate main contains at least the expected set of input terms for main. Ideally, among all those directional types T that satisfy this condition, we would like to infer the uniformly (i.e., for the input types and output types of all predicates) most precise one. The uniformly most precise directional type Tmin (P) of a program P together with a specification of the query input terms does exist, as we will show. It is, however, not effectively computable in general. This is naturally the place where abstraction comes in. We can compute a directional type Tsb (P), a regular approximation of Tmin (P) which is defined through the set-based abstraction a` la Heintze and Jaffar [24]. There is no objective criterion to evaluate the quality of the approximation of a non-regular set by a regular one in the sense that the most precise approximation does not exist; this fact applies also to our type inference procedure. We can show, however, that Tsb (P) is uniformly more precise than any discriminative directional type of P, i.e., any directional type out of the class for which the type check of Aiken and Lakshman is sound and complete. The above comparison is interesting for intrinsic reasons and it indicates that our type inference procedure produces “good” directional types. We exploit it furthermore in order to derive a type checking algorithm whose complexity improves upon the one of the original algorithm in [1]. A simple refinement of the arguments given in [1] suffices to make the lower bound more precise. We thus settle the complexity (DEXPTIME-complete) of the problem of directional type checking of logic programs wrt. discriminative types. Technically, our results are based on several basic properties of three kinds of abstraction (and their interrelation): the set-based abstraction (obtained by Cartesian approximation and related to our inference procedure), the set-valued abstraction (obtained by replacing membership constraints with set inclusions and related to the type-checking procedure for arbitrary types), and path closure abstraction (related to the type-checking procedure for discriminative types). These properties, that we collect in Sect. 2, are of general interest; in particular, the abstraction by path closure keeps reappearing (see, e.g., [27,26,31,19,20]). Furthermore, we establish that the directional types of a program P are exactly the models of an associated logic program PI nOut . In fact, PI nOut is a kind of

280

W. Charatonik and A. Podelski

“magic set transformation” (see, e.g., [20]) of P. We obtain our results (and the soundness and completeness results in [1]) by combining the model-theoretic characterization of directional types with the properties of abstractions established in Sect. 2. In fact, by having factored out general properties of abstractions from the aspects proper to directional types, we have maybe given a new view of the results in [1]. Other related work. Rouzaud and Nguyen-Phong [28] describe a type system where types are sets of non-ground terms and express directionality, but these sets must be tuple-distributive. Our types need not be tuple-distributive. In [14], Codish and Demoen infer type dependencies for logic programs. Their techniques (abstract compilation) are quite different from ours and the derived dependencies express all possible input-output relationships. Probably the work of Heintze and Jaffar is the one that is most closely related to ours. This is not only due to the fact that we use set-based analysis [24] to approximate the type program. Some of their papers [23,25] contain examples where they compute for each predicate p a pair of sets Callp and Retp , which can be viewed as (ground) directional type Callp → Retp . We are not aware, however, of a general, formal treatment of directional types inference in their work. Also the work of Gallagher and de Waal [20,21] is closely related to ours. They use a kind of “magic set transformation” to compute query and answer predicates for each predicate in the given program, which is essentially the same as our type program. Then they approximate the success set of the new predicates with ground, regular and tupledistributive sets, which they do not call types. Boye in [5] (see also [6,7]) presents a procedure that infers directional types for logic programs. The procedure is not fully automatic (sometimes it requires an interaction with the user), it requires the set of possible types to be finite, and no complexity analysis is given. In our approach, any regular set of terms is an admissible type; our procedure is fully automatic in the presence of a query for the program (or lower bounds for input types) and it runs in single-exponential time. We refer to [1] for comparison with still other type systems. Most of those interpret types as sets of ground terms, while we interpret types as sets of non-ground terms. Most type systems do not express the directionality of predicates.

2

Abstractions

In this section we discuss three kinds of abstraction of a given program. The motivation for this discussion is the following. In the next section we will define a type program for a given program P. The models of this program are directional types of P. The least model of the set-based abstraction of the type program correspond to the type that we infer. The models of the set-valued abstraction correspond to the directional types that pass the type-check from [1]. The models of the path-closure abstraction correspond to discriminative directional types of P. Later we will use the relations between these abstractions to compare different directional types of a given program.

Directional Type Inference for Logic Programs

281

Preliminaries. We follow the notation and terminology of [1] unless specified otherwise. For example, we use the symbols p, q, p0 , . . . for predicates (instead of f, g, f0 , . . . as in [1]) and write p(t) for predicate atoms (instead of f t). Wlog., all predicates are unary. Terms t are of the form x or f (t1 , . . . , tn ). We may write t[x1 , . . . , xm ] for t if x1 , . . . , xm are the variable occurrences in t (we distinguish between multiple occurrences of the same variable), and t[t1 , . . . , tm ] for the term obtained from t by substituting tj for the occurrence xj . A logic program P is a set of Horn clauses, i.e., implications of the form p0 (t0 ) ← p1 (t1 ), . . . , pn (tn ). A program comes with a query (“the main loop”) which, wlog., is specified by one atom main(t). The interpretation of programs, which is defined as usual, may be viewed as a mapping from predicate symbols to sets of trees. Programs may be viewed as formulas whose free variables are set-valued (and are referred to via predicate symbols), hence as a large class of set constraints. We will not use (positive) set expressions and set constraints as in [1] but, instead, logic programs, for specifying sets of trees as well as for abstraction. Positive set expressions, which denote sets of trees, can be readily translated into equivalent alternating tree automata, which again form the special case of logic programs whose clauses are all of the form p(f (x1 , . . . , xn )) ← p11 (x1 ), . . . , p1m1 (x1 ), . . . , pn1 (xn ), . . . , pnmn (xn ) where x1 , . . . , xn are pairwise different. A non-deterministic tree automaton is the special case where m1 = . . . = mn = 1, i.e., a logic program whose clauses are all of the form p(f (x1 , . . . , xn )) ← p1 (x1 ), . . . , pn (xn ). A set of trees is regular if it can be denoted by a predicate p in the least model of a non-deterministic tree automaton. A uniform program [19] consists of Horn clauses in one of the following two forms. (In a linear term, each variable occurs at most once.) – p(t) ← p1 (x1 ), . . . , pk (xm ) where the term t is linear. – q(x) ← p1 (t1 ), . . . , pm (tm ) where t1 , . . . , tm are any terms over Σ(Var). A uniform program can be transformed (in single-exponential time) into an equivalent non-deterministic tree automaton [19,18,10]. Set-based abstraction. We use set-based analysis in the sense of [24] but in the formulation using logic programs as in [19,18,10] (instead of set constraints as in [24] and [1]). Definition 1 (P # , the set-based abstraction of P). The uniform program P # is obtained from a program P by translating every clause p(t) ← body, whose head term t contains the n variables x1 , . . . , xn , into the (n + 1) clauses 1 mn 1 p(t˜) ← p1 (x11 ), . . . , p1 (xm 1 ), . . . , pn (xn ), . . . , pn (xn )

pi (xi ) ← body

(for i = 1, . . . , n)

where t˜ is obtained from t by replacing the mi different occurrences of variables xi i by different renamings x1i , . . . , xm i , and p1 , . . . , pn are new predicate names.

282

W. Charatonik and A. Podelski

The least model of the program P # expresses the set-based abstraction of P, which is defined in [24] as the least fixpoint of the operator τP . The operator τP is defined via set-based substitutions in [24]; it can also be defined by (for a subset M of the Herbrand base) τP (M ) = {p0 (t0 [x1 θ1 , . . . , xm θm ]) | p0 (t0 ) ← p1 (t1 ), . . . , pn (tn ) ∈ P, θ1 , . . . , θm ground substitutions, p1 (t1 θ1 ), . . . , pn (tn θ1 ) ∈ M .. .

p1 (t1 θm ), . . . , pn (tn θm ) ∈ M }

where x1 , . . . , xm are the variable occurrences in t0 (we distinguish between multiple occurrences of the same variable). As noted in [19], the logical consequence operator associated with P # is equal to the set-based consequence operator; i.e., TP # = τP . We recall that the logical consequence operator associated with the program P is defined by TP (M ) = {p0 (t0 [x1 θ, . . . , xm θ] | p0 (t0 ) ← p1 (t1 ), . . . , pn (tn ) ∈ P, θ ground substitution, p1 (t1 θ), . . . , pn (tn θ) ∈ M }. The set-based abstraction can be formalized in the abstract interpretation framework [16] by the application of the Cartesian approximation C to the semanticsdefining fixpoint operator TP ; i.e., TP # = C(TP ) (see [17]; roughly, C maps a set of tuples to the smallest Cartesian product containing it). Thus, we have TP # = τP = C(TP ). We note the following fact, keeping symmetry with Remarks 2 and 3 on the two other abstractions of TP that we will introduce. Given two set-valued functions F and F 0 , we write F ≤ F 0 if F is smaller than F 0 wrt. to pointwise subset inclusion, i.e., F (x) ⊆ F 0 (x) for all x. Remark 1 (Set-based approximation). The direct-consequence operator associated with P # approximates the one associated with P; i.e., TP ≤ T P # . Proof. Obvious by definition.

2

The following statement will be used for the soundness of our type inference algorithm (Theorem 3). Its converse does, of course, not hold in general (the least models of P may be strictly smaller than the least models of P # ). Proposition 1. Each model M of the set-based abstraction P # of a program P is also a model of P.

Directional Type Inference for Logic Programs

283

Proof. A ground instance of a clause of P is also a ground instance of the 2 corresponding clause of P # . Set-valued abstraction. The second abstraction that we consider is also defined via a program transformation: an atom p(t) is simply replaced by an inclusion t ⊆ p. The transformed program is interpreted over the domain of sets of trees; i.e., the valuations are mappings θ : Var → 2TΣ . These mappings are extended canonically from variables x to terms t; i.e., tθ is a set of trees. We repeat that an interpretation M, i.e., a subset of the Herbrand base, maps predicates p to sets of trees pM = {t ∈ TΣ | p(t) ∈ M}. The inclusion t ⊆ p holds in M under the valuation θ if tθ is a subset of pM . Definition 2 (P ⊆ , the set-valued abstraction of P). Given a program P, its set-valued program abstraction is a program P ⊆ that is interpreted over sets of trees (instead of trees). It is obtained by replacing membership with subset inclusion; i.e., it contains, for each clause p0 (t0 ) ← p1 (t1 ), . . . , pn (tn ) in P, the implication (1) t0 ⊆ p0 ← t1 ⊆ p1 , . . . , tn ⊆ pn . The models of the program P ⊆ are, as one expects, interpretations M (subsets of the Herbrand base) such that all implications all valid in M. An implications of the form (1) is valid in M if for all valuations θ : Var → 2TΣ , if ti θ is a subset of pM i for i = 1, . . . , n then also for i = 0. The models of P ⊆ are the fixpoints of TP ⊆ , the direct consequence operator associated with P ⊆ , which is defined in a way analogous to TP (using set-valued substitutions instead of tree-valued substitutions). Hence, we will be able to use the following remark when we compare models of P # with models of P ⊆ (Proposition 4). Remark 2 (Set-valued approximation). The direct-consequence operator associated with P ⊆ approximates the one associated with P; i.e., TP ≤ TP ⊆ . Proof. If, for some subset M of the Herbrand base, TP (M ) contains the ground atom p0 (t0 ) because M contains the ground atoms p1 (t1 ), . . . , pn (tn ) and p0 (t0 ) ← p1 (t1 ), . . . , pn (tn ) is a ground instance of some clause of P, then the singleton {ti } is a subset of the denotation of pi under M (for i = 1, . . . , n) and, hence, the singleton {t0 } is a subset of the denotation of p0 under TP ⊆ (M ). 2 The next statement underlies the soundness of the type checking procedure of [1], which essentially checks if a given interpretation is a model of the program PI⊆nOut defined in the next section (cf. Theorem 9 in [1]). It says that being a model wrt. P ⊆ is a sufficient condition for being a model of the program P. (The model property of a regular set wrt. P ⊆ is nothing else than an entailment relation between set constraints. The entailment is equivalent to satisfiability of negative constraints and can be tested in NEXPTIME [12]).

284

W. Charatonik and A. Podelski

Proposition 2. Each model M of the set-valued abstraction P ⊆ of a program P is also a model of P. Proof. If c ≡ p(t0 ) ← p1 (t1 ), . . . , pn (tn ) is a ground instance of a clause of P, then {t0 } ⊆ p ← {t1 } ⊆ p1 , . . . , {tn } ⊆ pn is a ground instance of the corresponding implication of P ⊆ (which holds in M if M is a model of P ⊆ , and, thus, c also holds). 2 The converse of the statement above does not hold in general. Take, for example, the program P defined by the clause p(f (x, x)) ← q(f (x, x)) and the four facts q(f (a, a)), q(f (a, b)), q(f (b, a)), q(f (b, b)). Then P ⊆ consists of the implication f (x, x) ⊆ p ← f (x, x) ⊆ q and the four inclusions f (a, a) ⊆ q, f (a, b) ⊆ q, f (b, a) ⊆ q, f (b, b) ⊆ q. Then M = {p(f (a, a))), p(f (b, b)), q(f (a, a)), q(f (a, b)), q(f (b, a)), q(f (b, b))} is a model of P but M is not a model of P ⊆ . (This example transfers, in the essence, Example 1 in [1] from the setting of directional types to a general setting.) The converse of the statement in Proposition 2 does, however, hold in the special case of path closed models (Proposition 3) which we will introduce below. The least fixpoint of TP ⊆ is in general not regular. To see this note that it is equal to the least fixpoint of TP if, for example, P is the length program. This example also shows that the least model of P # is in general not contained in every model of P ⊆ . This is the case, however, in the special case where the model of P ⊆ is path closed (Proposition 4). Path closed models. The following notion of a path-closed set originates from [22]. It is equivalent to other notions occurring in the literature: tupledistributive [27,28], discriminative [1], or deterministic. Definition 3 (Path-closed). A [regular] set of ground terms is called pathclosed if it can be defined by a [finite] deterministic top-down tree automaton. A deterministic finite tree automaton translates to a logic program which does not contain two different clauses with the same head (modulo variable renaming), e.g., p(f (x1 , . . . , xn )) ← p1 (x1 ), . . . , pn (xn ) and p(f (x1 , . . . , xn )) ← p01 (x1 ), . . . , p0n (xn ). A discriminative set expression as defined in [1] translates to a deterministic finite tree automaton, and vice versa. That is, discriminative set expressions denote exactly path-closed regular sets. It is argued in [1] that discriminative set expressions are quite expressive and are used to express commonly used data structures. Note that lists, for example, can be defined by the program with the two clauses l ist(cons(x, y)) ← list(y) and l ist(nil). The following fact is the fundamental property of path closed sets in the context of set constraints (see also Theorem 12 and Lemma 14 in [1]). It will be directly used in Proposition 3. For comparison, take the constraint f (x, y) ⊆ f (a, a) ∪ f (b, b); here, {f (a, a), f (b, b)} is a set that is not path closed, and the union of the (set-valued) solutions θ1 : x, y 7→ {a} and θ2 : x, y 7→ {b} is not a solution. Also, take the constraint f (x, y) ⊆ ∅; here, the union of the solutions over possibly empty sets θ1 : x 7→ {a}, y 7→ ∅and θ2 : x 7→ ∅, y 7→ {a} is not a solution.

Directional Type Inference for Logic Programs

285

Lemma 1. Solutions of conjunctions of inclusions t ⊆ e between terms t interpreted over nonempty sets and expressions e denoting path closed sets of trees are closed under union; i.e., if S is a set of solutions, then Θ defined S by Θ(x) = {θ(x) | θ ∈ S} is again a solution. Proof. The statement follows from the fact (shown, e.g., in [13]) that in the interpretation over nonempty sets, if the upper bounds e denote path closed sets then inclusions of the form f (x1 , . . . , xn ) ⊆ e are equivalent to the conjunction −1 −1 x1 ⊆ f(1) (e) ∧ . . . ∧ xn ⊆ f(n) (e) −1 (e) denotes the set {ti | f (t1 , . . . , tn ) ∈ e}. where f(i)

2

The following statement underlies the completeness of the type checking procedure of [1] for discriminative directional types (see Theorem 12, Lemma 14 and Theorem 15 in [1]). Proposition 3. Each path closed model M of a program P is also a model of P ⊆ , its set-valued abstraction. Proof. Assume that the clause p0 (t0 ) ← p1 (t1 ), . . . , pn (tn ) is valid in M (i.e., it holds under all ground substitutions σ : Var 7→TΣ , under the interpretation of the predicates p0 , p1 , . . . , pn by M), and that θ is a substitution mapping variables to nonempty sets such that t1 θ ⊆ p1 , . . . , tn θ ⊆ pn holds in M. We have to show that also t0 θ ⊆ p0 holds in M. The assumption yields that, for every ground substitution σ : Var 7→TΣ such that σ(x) ∈ θ(x) for all x ∈ Var, t1 σ ∈ p1 , . . . , tn σ ∈ pn holds in M. Thus, also t0 σ ∈ p0 holds in M. Since we have that – t0 σ ∈ p0 is equivalent to t0 σ ¯ ⊆ p0 where σ ¯ is the set substitution defined by σ ¯ (x) = {σ(x)}, – θ is the union of all σ ¯ such that σ(x) ∈ θ(x) for all x ∈ Var, – solutions of t0 ⊆ p0 , where p0 is interpreted by M as a path closed set, are closed under union (Lemma 1), the inclusion t0 ⊆ p0 also holds in M under the substitution θ.

2

Path closure abstraction. The path closure P C of a set M of trees is the smallest path closed set containing M . We consider a third abstraction of the operator TP by composing the path closure P C with TP . We note that we do not know whether the least fixpoint of the operator P C ◦ TP is always a regular set. The following comparison of two of the three abstractions that we have defined so far will be used in the proof of Proposition 4.

286

W. Charatonik and A. Podelski

Remark 3 (Set-based approximation and path closure). The path closure abstraction of the direct-consequence operator of P approximates also its set-based abstraction; i.e., TP # ≤ (P C ◦ TP ). Proof. We use the equality TP # = τP . If p0 (t0 [x1 θ1 , . . . , xm θm ]) ∈ τP (M ) because p1 (t1 θj ), . . . , pn (tn θj ) ∈ M , then p0 (t0 θj ) ∈ TP (M ), for j = 1, . . . , m. But then we have p0 (t0 [x1 θ1 , . . . , xm θm ]) ∈ (P C ◦ TP )(M ). 2 We will use the following statement later in order to compare the directional types obtained by our type inference procedure with the subclass of discriminative directional types for which the type check in [1] is sound and complete. (Note that the path closure of a model of a program is in general not itself a model, and that the least model of P # is in general not contained in the path closure of P). Proposition 4. The least model of P # , the set-based abstraction of a program P, is contained in every path closed model of P ⊆ , the set-valued abstraction of a program P. Proof. By Remark 3, TP # ⊆ (P C ◦ TP ) and thus, by Remark 2, TP # ⊆ (P C ◦ TP ⊆ ). If M is a path-closed model of P ⊆ , then it is also a fixpoint of P C ◦ TP , and hence it contains the least fixpoint of TP # , i.e., the least model of P # . 2

3

Directional Types and Type Programs

A type T is a set of terms t closed under substitution [2]. A ground type is a set of ground terms (i.e., trees), and thus a special case of a type. A term t has type T , in symbols t : T , if t ∈ T . A type judgment is an implication t1 : T1 ∧ . . . ∧ tn : Tn → t0 : T0 that holds under all term substitutions θ : Var → TΣ (Var). Definition 4 (Directional type of a program [8,1]). A directional type of a program P is a family T = (Ip → Op )p∈Pred assigning to each predicate p of P an input type Ip and an output type Op such that, for each clause p0 (t0 ) ← p1 (t1 ), . . . , pn (tn ) of P, the following type judgments hold. t0 : Ip0 → t1 : Ip1 t0 : Ip0 ∧ t1 : Op1 → t2 : Ip2 .. .

t0 : Ip0 ∧ t1 : Op1 ∧ . . . ∧ tn−1 : Opn−1 → tn : Ipn t0 : Ip0 ∧ t1 : Op1 ∧ . . . ∧ tn : Opn → t0 : Op0 We then also say that P is well-typed wrt. T . A program together with its query main(t) is well-typed wrt. T if furthermore the query argument t is well-typed wrt. the input type for main (i.e, the type judgment t : Imain holds).

Directional Type Inference for Logic Programs

287

Definition 5 (Ordering on directional types). We define that T = (Ip → Op )p∈Pred is uniformly more precise than T 0 = (Ip0 → Op0 )p∈Pred if Ip ⊆ Ip0 and Op ⊆ Op0 for all predicates p. The least precise directional type for which any program (possibly together with a query) is well-typed is T> = (> → >)p∈Pred assigning the set of all terms to each input and output type. In the absence of a query, the most precise one is T⊥ = (⊥ → ⊥)p∈Pred assigning the empty set to each input and output type. This changes if, for example, a query of the form main is present (see Sect. 4). Definition 6 (S at(T ), the type of terms satisfying T [1]). Given the ground type T , the set S at(T ) of terms satisfying T is the type S at(T ) = {t ∈ TΣ (Var) | θ(t) ∈ T for all ground substitutions θ : Var → TΣ }. Definition 7 (Discriminative type). A directional type is called discriminative if it is of the form (S at(Ip ) → S at(Op ))p∈Pred , where the sets Ip , Op are path-closed. Remark 4. The clause p0 (t0 ) ← p1 (t1 ), . . . , pn (tn ) is valid in some model M if and only if the type judgment t0 : S at(p0 ) ← t1 : S at(p1 ) ∧ . . . ∧ tn : S at(pn ) holds in M (i.e., under the interpretation of p0 , p1 , . . . , pn by M). Proof. Membership of the application of substitutions to terms in sets of the form S at(E) is defined by the application of ground substitutions to the terms in E. 2 A directional type of the form T = (S at(Ip ) → S at(Op ))p∈Pred , for ground types Ip , Op ⊆ TΣ , satisfies a type judgment if and only if the corresponding directional ground type Tg = (Ip → Op )p∈Pred does. We will next transform the well-typedness condition in Definition 4 into a logic program by replacing t : Ip with the atom pI n (t) and t : Op with pOut (t). Definition 8 (PI nOut , the type program for P). Given a program P, the corresponding type program PI nOut defines an in-predicate pI n and an outpredicate pOut for each predicate p of P. Namely, for every clause p0 (t0 ) ← p1 (t1 ), . . . , pn (tn ) in P, PI nOut contains the n clauses defining in-predicates corresponding to each atom in the body of the clause, pI1n (t1 ) ← pI0n (t0 ) pI2n (t2 ) ← pI0n (t0 ), pOut (t1 ) 1 .. .

(t1 ), . . . , pOut pInn (tn ) ← pI0n (t0 ), pOut 1 n−1 (tn−1 )

and the clause defining the out-predicate corresponding to the head of the clause, (t0 ) ← pI0n (t0 ), pOut (t1 ), . . . , pOut pOut 0 1 n (tn ).

288

W. Charatonik and A. Podelski

If the program P comes together with a query main(t), we add the clause mainI n (t) ← true to PI nOut . The next statement extends naturally to a characterization of the well-typedness of a program together with a query. Theorem 1 (Types and models of type programs). The program P is well-typed wrt. the directional type T = (S at(Ip ) → S at(Op ))p∈Pred (with ground types Ip , Op ) if and only if the subset of the Herbrand base corresponding to T , MT = {pI n (t) | t ∈ Ip } ∪ {pOut (t) | t ∈ Op }, is a model of the type program PI nOut . Proof. The validity of the well-typing conditions under ground substitutions is exactly the logical validity of the clauses of PI nOut in the model MT . The statement then follows by Remark 4. 2 We next define two abstractions of type programs. We will use PI#nOut for type inference (Sect. 4) and PI⊆nOut for type checking (Sect. 5). Given a directional type T , the interpretation of PI⊆nOut by the corresponding subset MT is the set constraint condition which is used in [1] to replace the well-typedness condition in Definition 4. Definition 9 (Abstractions of type programs). The set-based type program PI#nOut is the set-based abstraction of PI nOut , and the set-valued type program PI⊆nOut is the set-valued abstraction of PI nOut ; i.e., PI#nOut = (PI nOut )# , PI⊆nOut = (PI nOut )⊆ . The following direct consequence of Theorem 1 and Propositions 2 and 3 restates the soundness and the conditional completeness of the type check in [1]. Theorem 2 ([Discriminative] types and models of set-valued type programs). A directional type of the form T = (S at(Ip ) → S at(Op ))p∈Pred is a directional type of the program P if the corresponding subset MT of the Herbrand base is a model of PI⊆nOut , the set-valued abstraction of the type program of P. For discriminative directional types T , the converse also holds. Proof. The first part follows from Proposition 2 together with Theorem 1, the second from from Proposition 3 together with Theorem 1. 2

Directional Type Inference for Logic Programs

4

289

Directional Type Inference

We consider three different scenarios in which we may want to infer directional types from a program. (1) The program comes together with a query consisting of one atom main without arguments (and there is a clause main ← p1 (t1 ), . . . , pn (tn ) calling the actual query). In this case, we are interested in inferring the most precise directional type T such that P together with the query main is well-typed. According to the model-theoretic characterization of the well-typedness of a program together with a query (Theorem 1), we must add the clause mainI n ← true to PI nOut . This means that T⊥ = (⊥ → ⊥)p∈Pred is generally not a directional type of a program together with the query main and, hence, it is nontrivial to infer a precise one. (2) The program comes together with a query consisting of one atom main(t) and a lower bound Mmain for the input type of main is given (the user hereby encodes which input terms to the query predicates are expected). In this case, we are interested in inferring the most precise directional type T = (Ip → Op )p∈Pred for P such that the input type for main lies above the lower bound for main, i.e., such that Mmain ⊆ Imain . For example, take the program defining the predicate r everse(x, y) r everse([], []). r everse([X|Xs], Y )←r everse(Xs, Z), append(Z, [X], Y, ). together with the definition of append and with the query r everse(x, y). If the expected input terms are lists (for x) and non-instantiated variables (for y), i.e., the lower bound specified is Mr ev = (list, >), then the type inferred by our algorithm for r everse is (list, >) → (list, list) and the type inferred by our algorithm for append is (list, [>], >) → (list, [>], list), where [>] is the type of all singleton lists. (3) Lower bounds Mp for the input types Ip of all predicates p of P are given. This may be done explicitly for some p and implicitly, with Mp = ∅, for the others. In this case, we are interested in inferring the most precise directional type T = (Ip → Op )p∈Pred for P such that the input types for p lie above the given lower bounds, i.e., such that Mp ⊆ Ip for all p. In a setting with program modules, for example, the lower bounds that are explicitly specified may be the query goals of exported predicates. The scenario (3) resembles the one imagined by Aiken and Lakshman in the conclusion of [1]. Note that in our setting, however, the sets Mp need not already be input types. For example, if the inputs for x and y in append(x, y, z) are expected to be lists of even length only, i.e., the lower bound for Iapp is given by Mapp = (evenlist, evenlist, >), then we would infer the set of all lists as the input type for x, i.e., Iapp = (list, evenlist, >) (note that there are recursive calls to append with lists of odd length). We obtain (1) and (2)as the special case of (3) where the lower bounds for all input types but main are given as the empty set. Hence, it is sufficient to formulate the type inference only for the case (3).

290

W. Charatonik and A. Podelski

Definition 10 (Inference of the set-based directional type Tsb (P, (Mp )p∈Pred )). Given a program P and a family of lower bounds Mp for the input types of the predicates p of the program, we infer the set-based directional type Tsb (P, (Mp )p∈Pred ) = (S at(Ip ) → S at(Op ))p∈Pred where Ip and Op are the denotations of the predicates pI n and pOut in the least model of the program PI#nOut ∪ {pI n (x) ← Mp (x) | p ∈ Pred}.

(2)

The definition of type inference above leaves open in which formalism the lower bounds are specified and how the inferred directional types are presented to the user. There is a wide variety of formalisms that coincide in the expressive power of regular sets of trees and that can be effectively translated one into another and, hence, for which our type inference yields an effective procedure. More concretely, we propose to represent the lower bounds Mp through logic programs in restricted syntax (see Sect. 2) that corresponds directly to alternating tree automata (and also to the positive set expressions considered in [1]; an even more restricted syntax corresponds to non-deterministic tree automata and to positive set expressions without intersection). We attach these logic programs to the program in (2) as definitions of the sets Mp . Then, we can apply one of the known algorithms (see, e.g., [19,18,11,10]) in order to compute, in single-exponential time (in the size of the program P and the programs for the sets Mp ), a logic program that corresponds to a non-deterministic tree automaton and that is equivalent to the program in (2) wrt. the least model and, thus, represents Tsb (P, (Mp )p∈Pred ). The representation of sets by a non-deterministic tree automaton is a good representation in the sense that, for example, the test of emptiness can be done in linear time. If, in Definition 10, we replace PI#nOut with PI nOut , then we obtain the uniformly most precise directional type Tmin (P) wrt. given lower bounds for input types. In general, we cannot effectively compute Tmin (P) (e.g., test emptiness of its input and output types). We repeat that the following comparison of the set-based directional type of P with discriminative directional types of P is interesting because these form the subclass of directional types for which the type check in [1] is sound and complete. Theorem 3 (Soundness and Quality of Type Inference). The program P is well-typed wrt. the set-based directional type Tsb (P, (Mp )p∈Pred ). Moreover, this type is uniformly more precise than every discriminative directional type of P whose family of input types contains (Mp )p∈Pred . Proof. The two statements are direct consequences of Propositions 1 and 4, respectively, together with Theorem 1. 2

Directional Type Inference for Logic Programs

5

291

Complexity of Directional Type Checking

Theorem 4. The complexity of directional type checking of logic programs for discriminative types is DEXPTIME-complete. Proof. The DEXPTIME-hardness follows from refining the argument in the proof of Theorem 18 in [1] with the two facts that (1) discriminative types are denoted by path closed sets, and (2) testing the non-emptiness of a sequence of n tree automata is DEXPTIME-hard even if the automata are restricted to deterministic ones (which recognize exactly path closed sets) [29]. We obtain a type checking algorithm in single-exponential time as follows. The input is the program P and the discriminative directional type T = (S at(Ip ) → S at(Op )) where all types Ip and Op are given by ground set expressions (a special case of which are regular tree expressions or non-deterministic tree automata). Ground set expressions correspond to alternating tree automata which are self-dual; i.e., fp representing the complement by we can obtain ground set expressions Iep and O a syntactic transformation in linear time. We can translate ground expressions into logic programs defining predicates pIp and pOp such that they denote Ip and e Op wrt. the least-model semantics (and similarly predicates pIe and pO f for Ip p

p

fp ). and O We now use one of the well-known single-exponential time algorithms [19,18, 11,10] to compute (the non-deterministic tree automaton representing) the least model of the set-based abstraction P # of a logic program P. We apply such an algorithm to the program PI#nOut (T ) that we obtain from PI#nOut by adding the clauses pI n (x) ← pIp (x) and pOut (x) ← pOp (x) and the logic programs defining pIp and pOp . We use the result in order to test whether the denotations of pI n and pOut under the least model of the program PI nOut (T ) are exactly Ip and Op . This holds if and only if T is a directional type of P (otherwise, we have a proper inclusion for at least one p; see the correctness proof below). The test works by testing whether the intersection of (the non-deterministic tree automaton representing) the denotations of pI n and pOut under the least model of the program PI nOut (T ) with the complements pIe and pO f is empty. This can be p

p

done in one pass by taking the conjunction of PI nOut (T ) with the logic programs defining pIe and pO f and testing the emptiness of new predicates defined as the p

p

intersection of pIp and pIe (and pOp and pO fp ). p The correctness of the algorithm follows with Theorem 1 and Proposition 1 and 4. In detail: Given a discriminative directional type T and a program P, we have that P is well-typed wrt. T iff the corresponding subset MT of the Herbrand base is a (path closed) model of PI nOut by Theorem 1. If this is the case then MT is equal to the least model M0 of PI nOut (T ), since MT ⊆ M0 by definition of PI nOut (T ), and MT ⊇ M0 holds by Proposition 4 (note that MT

292

W. Charatonik and A. Podelski

is a path closed model of PI nOut and of the additional clauses translating T and, thus, contains the least one). On the other hand, if MT is equal to the least model of PI#nOut (T ), then T is a directional type by Proposition 1 and Theorem 1. 2 We note that the procedure above yields a semi-test (in the same sense as the one in [1]) for well-typedness wrt. the class of general directional types, since the equivalence between the inferred type and the given one is a sufficient (but generally not necessary) condition for well-typedness wrt. the given type. We repeat that it implements a full test wrt. discriminative directional types.

6

Future Work

One of the obvious directions for future work is an implementation. We already have a working prototype implementation on top of the saturation-based theorem prover SPASS [30]. The first results are very promising; due to specific theorem-proving techniques like powerful redundancy criteria, one obtains a decision procedure for the emptiness test that is quite efficient on our examples. We ran our prototype on the benchmark programs from [9]. We were able to detect emptiness of the main predicate for six of the twelve programs, while a similar method of Gallagher succeeded in three cases. We believe that by using the tree-automata techniques suggested in [18] together with automata minimization (and conversion to the syntax of ground set expressions), we can further improve the efficiency of our implementation and the readability of the output. Acknowledgments. We thank David McAllester for turning us on to magic sets and thereby to directional types. We thank Harald Ganzinger for useful comments.

References 1. A. Aiken and T. K. Lakshman. Directional type checking of logic programs. In B. L. Charlier, editor, 1st International Symposium on Static Analysis, volume 864 of Lecture Notes in Computer Science, pages 43–60, Namur, Belgium, Sept. 1994. Springer Verlag. 2. K. R. Apt. Declarative programming in Prolog. In D. Miller, editor, Logic Programming - Proceedings of the 1993 International Symposium, pages 12–35, Vancouver, Canada, 1993. The MIT Press. 3. K. R. Apt. Program verification and Prolog. In E. B¨ orger, editor, Specification and Validation methods for Programming languages and systems, pages 55–95. Oxford University Press, 1995. 4. K. R. Apt and S. Etalle. On the unification free Prolog programs. In A. M. Borzyszkowski and S. Sokolowski, editors, Mathematical Foundations of Computer Science 1993, 18th International Symposium, volume 711 of lncs, pages 1–19, Gdansk, Poland, 30 Aug.– 3 Sept. 1993. Springer.

Directional Type Inference for Logic Programs

293

5. J. Boye. Directional Types in Logic Programming. PhD thesis, Department of Computer and Information Science, Link¨ oping University, 1996. 6. J. Boye and J. Maluszynski. Two aspects of directional types. In L. Sterling, editor, Proceedings of the 12th International Conference on Logic Programming, pages 747–764, Cambridge, June13–18 1995. MIT Press. 7. J. Boye and J. Maluszynski. Directional types and the annotation method. Journal of Logic Programming, 33(3):179–220, Dec. 1997. 8. F. Bronsard, T. K. Lakshman, and U. S. Reddy. A framework of directionality for proving termination of logic programs. In K. Apt, editor, Proceedings of the Joint International Conference and Symposium on Logic Programming, pages 321–335, Washington, USA, 1992. The MIT Press. 9. M. Bruynooghe, H. Vandecasteele, A. de Waal, and M. Denecker. Detecting unsolvable queries for definite logic programs. In Proceedings of the Symposium on Programming Language Implementation and Logic Programming (PLILP’98), LNCS. Springer-Verlag, 1998. to appear. 10. W. Charatonik, D. McAllester, D. Niwi´ nski, A. Podelski, and I. Walukiewicz. The Horn mu-calculus. To appear in Vaughan Pratt, editor, Proceedings of the 13th IEEE Annual Symposium on Logic in Computer Science. 11. W. Charatonik, D. McAllester, and A. Podelski. Computing the least and the greatest model of the set-based abstraction of logic programs. Presented at the Dagstuhl Workshop on Tree Automata, October 1997. 12. W. Charatonik and L. Pacholski. Negative set constraints with equality. In Ninth Annual IEEE Symposium on Logic in Computer Science, pages 128–136, 1994. 13. W. Charatonik and A. Podelski. Set constraints for greatest models. Technical Report MPI-I-97-2-004, Max-Planck-Institut f¨ ur Informatik, April 1997. www.mpisb.mpg.de/˜podelski/papers/greatest.html. 14. M. Codish and B. Demoen. Deriving polymorphic type dependencies for logic programs using multiple incarnations of prop. In B. L. Charlier, editor, Proceedings of the First International Static Analysis Symposium, Lecture Notes in Computer Science 864, pages 281–296. Springer Verlag, 1994. 15. M. Comini, G. Levi, M. C. Meo, and G. Vitiello. Proving properties of logic programs by abstract diagnosis. In M. Dam, editor, Analysis and Verification of Multiple-Agent Languages, volume 1192 of LNCS, pages 22–50. Springer-Verlag, June 1996. 16. P. Cousot and R. Cousot. Inductive definitions, semantics and abstract interpretation. In Proc. POPL ’92, pages 83–94. ACM Press, 1992. 17. P. Cousot and R. Cousot. Formal language, grammar and set-constraint-based program analysis by abstract interpretation. In Record of FPCA ’95 - Conference on Functional Programming and Computer Architecture, pages 170–181, La Jolla, California, USA, 25-28 June 1995. SIGPLAN/SIGARCH/WG2.8, ACM Press, New York, USA. 18. P. Devienne, J.-M. Talbot, and S. Tison. Set-based analysis for logic programming and tree automata. In Proceedings of the Static Analysis Symposium, SAS’97, volume 1302 of LNCS, pages 127–140. Springer-Verlag, 1997. 19. T. Fr¨ uhwirth, E. Shapiro, M. Vardi, and E. Yardeni. Logic programs as types for logic programs. In Sixth Annual IEEE Symposium on Logic in Computer Science, pages 300–309, July 1991. 20. J. Gallagher and D. A. de Waal. Regular approximations of logic programs and their uses. Technical Report CSTR-92-06, Department of Computer Science, University of Bristol, 1992.

294

W. Charatonik and A. Podelski

21. J. Gallagher and D. A. de Waal. Fast and precise regular approximations of logic programs. In P. V. Hentenryck, editor, Proceedings of the Eleventh International Conference on Logic Programming, pages 599–613, Santa Margherita Ligure, Italy, 1994. The MIT Press. 22. F. G´ecseg and M. Steinby. Tree Automata. Akademiai Kiado, 1984. 23. N. Heintze. Practical aspects of set based analysis. In K. Apt, editor, Proceedings of the Joint International Conference and Symposium on Logic Programming, pages 765–779, Washington, USA, 1992. The MIT Press. 24. N. Heintze and J. Jaffar. A finite presentation theorem for approximating logic programs. In Seventeenth Annual ACM Symposium on Principles of Programming Languages, pages 197–209, January 1990. 25. N. Heintze and J. Jaffar. Set constraints and set-based analysis. In Proceedings of the Workshop on Principles and Practice of Constraint Programming, LNCS 874, pages 281–298. Springer-Verlag, 1994. 26. G. Janssens and M. Bruynooghe. Deriving descriptions of possible values of program variables. Journal of Logic Programming, 13(2-3):205–258, 1992. 27. P. Mishra. Towards a theory of types in Prolog. In IEEE International Symposium on Logic Programming, pages 289–298, 1984. 28. Y. Rouzaud and L. Nguyen-Phuong. Integrating modes and subtypes into a Prolog type-checker. In K. Apt, editor, Proceedings of the Joint International Conference and Symposium on Logic Programming, pages 85–97, Washington, USA, 1992. The MIT Press. 29. H. Seidl. Haskell overloading is DEXPTIME-complete. Information Processing Letters, 52:57–60, 1994. 30. C. Weidenbach. Spass version 0.49. Journal of Automated Reasoning, 18(2):247– 252, 1997. 31. E. Yardeni and E. Shapiro. A type system for logic programs. In E. Shapiro, editor, Concurrent Prolog, volume 2, chapter 28, pages 211–244. The MIT Press, 1987.

Finite Subtype Inference with Explicit Polymorphism Dominic Duggan Department of Computer Science, Stevens Institute of Technology, Castle Point on the Hudson, Hoboken, New Jersey 07030. [email protected]

Abstract. Finite subtype inference occupies a middle ground between HindleyMilner type inference (as in ML) and subtype inference with recursively constrained types. It refers to subtype inference where only finite types are allowed as solutions. This approach avoids some open problems with general subtype inference, and has practical motivation where recursively constrained types are not appropriate. This paper presents algorithms for finite subtype inference, including checking for entailment of inferred types against explicitly declared polymorphic types. This resolves for finite types a problem that is still open for recursively constrained types. Some motivation for this work, particularly for finite types and explicit polymorphism, is in providing subtype inference for first-class container objects with polymorphic methods.

1

Introduction

Type inference is the process of statically type-checking a program where some or all of the type information has been omitted from the program text. ML and Haskell are examples of programming languages where type inference has been a spectacular success. The particular flavor of type inference used by ML and Haskell is HindleyMilner type inference [14]. The type-checker accumulates equality constraints via a tree walk of the abstract syntax tree, and then uses a unification algorithm to compute a (most general) unifying substitution for these constraints. More recently attention has been focused on subtype inference [2,3,17,7,19,22]. With this work, the type-checker accumulates subtype constraints while traversing the abstract syntax tree, and then applies a constraint solver to check these constraints for consistency. Pottier [19] and Smith and Trifonov [22] have considered the problem of entailment in these type systems, which is important for example in interface matching. Subtype inference continues to be an important avenue of research, particularly in simplifying inferred types to make them practically useful. Hindley-Milner type inference and subtype inference represent two extremes in the type inference continuum: G. Levi (Ed.): SAS’98, LNCS 1503, pp. 295–310, 1998. c Springer-Verlag Berlin Heidelberg 1998

296

D. Duggan

Hindley-Milner Finite Subtype Subtype Equality Subtyping Subtyping Finite types Finite types Infinite types Inferred monotypes Inferred monotypes Inferred monotypes Inferred polytypes Inferred polytypes Inferred polytypes Specified polytypes∗ Between these two extremes, there is an intermediate point: finite subtype inference. While this alternative allows subtyping and type subsumption, it does not assume that types are potentially infinite trees (as with the most recent work on subtype inference). Why should we consider subtype inference with finite types? It is worth recalling why ML for example does not allow circular types (types as potentially infinite trees). The problem was pointed out by Solomon [21]: the problem of deciding the equality of parameterized recursive types is equivalent to the problem of deciding the equality of deterministic context-free languages (DCFLs), which is still after several decades an open problem. This problem is avoided in ML type inference by making the folding and unfolding of recursive types explicit (using data constructors and pattern-matching, respectively), so that circular types are not needed. A motivation for infinite types in subtype inference is to support objects with recursive interfaces. However the problem discovered by Solomon also holds for recursive interfaces for container objects. Consider for example a set object with interface: set(α) = {map : ∀β.(α → β) → set(β), product : ∀β.set(β) → set(α ∗ β), power : ∀β.unit → set(set(α)) All of the methods in this interface are examples of non-regular recursion in the object interface. In a companion paper [5], we have developed an object design for ML-like languages that avoids this problem. The approach there is to again make the folding and unfolding of recursive object interfaces explicit, in object creation and method invocation, respectively. This work introduces the possibility that circular types for recursive object interfaces, while useful for simple objects, may not be so useful for container objects. Early work on subtype inference considered atomic subtyping with finite types [15,12,9]. However there are non-trivial differences between finite subtype inference with atomic subtyping and with record containment. For example even checking for finiteness is markedly different, as elaborated upon in Sect. 5. Another design point is whether polymorphic types should be inferred or specified. All of the work so far assumes that polymorphic types are inferred. The disadvantage of these approaches is that the inferred types are large and complex, diminishing their practical usefulness, despite recent work on simplifying inferred types [7,19,22]. One way to avoid this problem is to require that the programmer provide explicit interfaces for polymorphic functions. This approach introduces fresh technical complications of its own. In Hindley-Milner type inference, mixed-prefix unification has been used to control scoping of type variables with explicit polymorphic type declarations (an idea originally used by Leroy and Mauny [11], and subsequently rediscovered by Odersky and L¨aufer

Finite Subtype Inference with Explicit Polymorphism

297

[16]). In this paper we extend subtype inference with constraint-solving under a mixed prefix, in order to support subtype inference with explicit polymorphism. Explicit polymorphism also derives motivation from our work on container objects with recursive interfaces [5]. We avoid the problems with first-class polymorphism in Hindley-Milner type inference, by requiring explicit type specifications on polymorphic methods. This is similar to the use of universal types to incorporate impredicativity into Hindley-Milner type inference [20,16,10], but tied to the object system instead of to datatypes (again because we are concerned with container objects with polymorphic methods). Even if we are not concerned with polymorphic methods, explicit polymorphism is required if we wish to provide a type-checking algorithm for the Abadi and Cardelli object type system, for example [1]. In that type system, the type rule for method update is structural [1, Sect. 16.2], meaning that the new method definition must be parametric in the type of self (the type of self is a type parameter constrained by the object interface). Explicit polymorphism requires that it be possible to check for entailment of inferred types from declared types. For infinite types this is problematic. Although incomplete algorithms have been published [19,22], the decidability of entailment remains open [22]. In this paper we demonstrate that entailment is decidable for finite subtyping, giving further motivation for our approach. Sect. 2 introduces our type system. We do not overburden the paper with any details of the object system mentioned earlier [5], but present a familiar ML-like language with record-based subtyping and explicit polymorphism. Sect. 3 provides the type inference algorithm. Sect. 4 provides algorithms for checking consistency and entailment; these algorithms must be defined mutually recursively. Sect. 5 considers the check for finite solutions; perhaps surprisingly, this check must be done after consistency and entailment checking. Sect. 6 considers the use of mixed prefix constraint-solving to check that there are well-scoped solutions. Finally Sect. 8 considers further related work and provides conclusions.

2 Type System The mini-language we consider is a language with functions, pairs and records. Subtyping is based on containment between record types. This is extended contravariantly to function types and covariantly to product types. Polymorphic types allow quantified type variables to be constrained by upper bounds. We use ∀αn

1503

3 Berlin Heidelberg New York Barcelona Budapest Hong Kong London Milan Paris Singapore Tokyo

Giorgio Levi (Ed.)

Static Analysis 5th International Symposium, SAS’98 Pisa, Italy, September 14-16, 1998 Proceedings

13

Series Editors Gerhard Goos, Karlsruhe University, Germany Juris Hartmanis, Cornell University, NY, USA Jan van Leeuwen, Utrecht University, The Netherlands

Volume Editor Giorgio Levi Universit`a di Pisa, Dipartimento di Informatica Corso Italia, 40, I-56125 Pisa, Italy E-mail: [email protected]

Cataloging-in-Publication data applied for Die Deutsche Bibliothek - CIP-Einheitsaufnahme Static analysis : 5th international symposium ; proceedings / SAS ’98, Pisa, Italy, September 14-16, 1998. Giorgio Levi (ed.). - Berlin ; Heidelberg ; New York ; Barcelona ; Budapest ; Hong Kong ; London ; Milan ; Paris ; Singapore ; Tokyo : Springer, 1998 (Lecture notes in computer science ; Vol. 1503) ISBN 3-540-65014-8

CR Subject Classification (1991): D.1, D.2.8, D.3.2-3, F.3.1-2, F.4.2 ISSN 0302-9743 ISBN 3-540-65014-8 Springer-Verlag Berlin Heidelberg New York This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting, reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer-Verlag. Violations are liable for prosecution under the German Copyright Law. c Springer-Verlag Berlin Heidelberg 1998 Printed in Germany Typesetting: Camera-ready by author SPIN 10638978 06/3142 – 5 4 3 2 1 0

Printed on acid-free paper

Foreword This volume contains the proceedings of the 1998 international symposium on static analysis (SAS’98) which was held in Pisa (Italy), on September 14-16, 1998 and was part of a federated conference with ALP-PLILP’98 and several workshops. SAS’98 is the annual conference and forum for researchers in all aspects of static analysis. It follows to SAS’94, SAS’95, SAS’96 and SAS’97 which were held respectively in Namur (Belgium), Glasgow (UK), Aachen (Germany) and Paris (France), and the international workshops WSA’92 held in Bordeaux (France) and WSA’93 held in Padova (Italy). In response to the call for papers, 48 papers were submitted. All papers were reviewed by at least three reviewers and the program committee met in Pisa to select 20 papers based on the referee reports. There was a consensus at the meeting that the technical papers were of very high quality. In addition to the submitted papers, SAS’98 had a number of outstanding invited speakers. Roberto Giacobazzi, Peter Lee, Amir Pnueli, Dave Schmidt, Scott Smolka, and Bernhard Steﬀen accepted our invitation to give invited talks or tutorials. Some of the papers (or abstracts) based on these talks are also included in this volume. SAS’98 has been fortunate to rely on a number of individuals and organizations. I want to thank all the program committee members and referees, for their hard work in producing the reviews and for such a smooth and enjoyable program committee meeting. Special thanks go to the conference chairman, Maurizio Gabbrielli, and to my students in Pisa who helped me a lot. More special thanks go to Vladimiro Sassone, who made available to SAS’98 his excellent system for handling submissions and reviews on the web, and to Ernesto Lastres and Ren`e Moreno who were my “system managers”. The use of this system made my life of program chairman much easier and I strongly recommend it to future program chairpersons. SAS’98 was sponsored by Universit` a di Pisa, Compulog Network, Consiglio Nazionale delle Ricerche, CNR-Gruppo Nazionale di Informatica Matematica, Comune di Pisa, and Unione Industriali di Pisa.

Giorgio Levi July 1998

Program Committee of SAS’98 Alex Aiken (Berkeley, USA) Maurice Bruynooghe (Leuven, Belgium) Michael Codish (Ben Gurion, Israel) Agostino Cortesi (Venezia, Italy) Radhia Cousot (Polytechnique Paris, France) Alain Deutsch (INRIA, France) Laurie Hendren (McGill, Canada) Fritz Henglein (DIKU, Denmark) Thomas Jensen (IRISA/CNRS, France) Alan Mycroft (Cambridge, UK) Flemming Nielson (Aarhus, Denmark) Thomas Reps (Wisconsin, USA) Dave Schmidt (Kansas State, USA) Mary Lou Soffa (Pittsburgh, USA) Harald Søndergaard (Melbourne, Australia) Bernhard Steffen (Passau, Germany)

Conference Chairman of SAS-ALP-PLILP’98 Maurizio Gabbrielli (Pisa, Italy)

List of Referees Torben Amtoft Andrew Appel Roberto Bagnara Dante Baldan Bruno Blanchet Rastislav Bodik Andrew Bromage Nicoletta Cocco Thomas Conway Patrick Cousot Bart Demoen Danny De Schreye Alessandra Di Pierro Manuel Faehndrich Gilberto Fil´e Jeff Foster Pascal Fradet Maurizio Gabbrielli Etienne Gagnon Roberto Giacobazzi Robert Gluck Eric Goubault Susanne Graf James Harland Nevin Heintze Frank Huch Jesper Jorgensen Andy King Jens Knoop Laura Lafave Vitaly Lagoon Martin Leucker Michael Leuschel Daniel Le Metayer Henning Makholm Elena Marchiori Kim Marriott Bern Martens Markus Mueller-Olm Anne Mulkers Hanne Riis Nielson Thomas Noll Dino Pedreschi Francesco Ranzato Jakob Rehof Olivier Ridoux Eva Rose Sabina Rossi Oliver Ruething David Sands Fausto Spoto Peter Stuckey Zhendong Su Cohavit Taboch Simon Taylor Peter Thiemann Arnaud Venet Zhe Yang Phillip Yelland

Contents

Data-Flow Analysis Bidirectional Data Flow Analysis in Code Motion: Myth and Reality . . . . . Oliver R¨ uthing

1

On a New Method for Dataflow Analysis of Java Virtual Machine Subroutines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 Masami Hagiya, Akihiko Tozawa Enabling Sparse Constant Propagation of Array Elements via Array SSA Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 Kathleen Knobe, Vivek Sarkar Assessing the Effects of Flow-Sensitivity on Pointer Alias Analyses . . . . . . . 57 Michael Hind, Anthony Pioli

Logic Programming Analysis of Normal Logic Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82 Fran¸cois Fages, Roberta Gori The Correctness of Set-Sharing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 Patricia M. Hill, Roberto Bagnara, Enea Zaffanella Deriving Analysers by Folding/Unfolding of Natural Semantics and a Case Study: Slicing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 Val´erie Gouranton

Concurrency A Symbolic Semantics for Abstract Model Checking . . . . . . . . . . . . . . . . . . . . 134 Francesca Levi Automatic Determination of Communication Topologies in Mobile Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152 Arnaud Venet Constructing Specific SOS Semantics for Concurrency via Abstract Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168 Chiara Bodei, Pierpaolo Degano, Corrado Priami

VIII

Contents

Abstract Domains A First-Order Language for Expressing Aliasing and Type Properties of Logic Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184 Paolo Volpe Refining Static Analyses by Trace-Based Partinioning Using Control Flow . 200 Maria Handjieva, Stanislav Tzolovski Building Complete Abstract Interpretations in a Linear Logic-Based Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215 Roberto Giacobazzi, Francesco Ranzato, Francesca Scozzari

Partial Evaluation On the Power of Homeomorphic Embedding for Online Termination . . . . . . 230 Michael Leuschel Analysis of Imperative Programs through Analysis of Constraint Logic Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246 Julio C. Peralta, John P. Gallagher, H¨ useyin Saˇglam Improving Control in Functional Logic Program Specialization . . . . . . . . . . . 262 E. Albert, M. Alpuente, M. Falaschi, P. Juli´ an, G. Vidal

Type Inference Directional Type Inference for Logic Programs . . . . . . . . . . . . . . . . . . . . . . . . . 278 Witold Charatonik, Andreas Podelski Finite Subtype Inference with Explicit Polymorphism . . . . . . . . . . . . . . . . . . . 295 Dominic Duggan

Optimization Sparse Jacobian Computation in Automatic Differentiation by Static Program Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311 M. Tadjouddine, F. Eyssette, C. Faure A New Solution to the Hidden Copy Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 327 Deepak Goyal, Robert Paige

Tutorials A Tutorial on Domain Theory in Abstract Interpretation . . . . . . . . . . . . . . . 349 Roberto Giacobazzi

Contents

IX

Program Analysis as Model Checking of Abstract Interpretations . . . . . . . . 351 David Schmidt, Bernhard Steffen

Invited Talks Certifying, Optimizing Compilation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381 Peter Lee Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383

Bidirectional Data Flow Analysis in Code Motion: Myth and Reality Oliver R¨ uthing Department of Computer Science, University of Dortmund, Germany [email protected]

Abstract. Bidirectional data flow analysis has become the standard technique for solving bit-vector based code motion problems in the presence of critical edges. Unfortunately, bidirectional analyses have turned out to be conceptually and computationally harder than their unidirectional counterparts. In this paper we show that code motion in the presence of critical edges can be achieved without bidirectional data flow analyses. This is demonstrated by means of an adaption of our algorithm for lazy code motion [15], which is developed from a fresh, specification oriented view. Besides revealing a better conceptual understanding of the phenomena caused by critical edges, this also settles the foundation for a new and efficient hybrid iteration strategy that intermixes conventional round-robin iteration with the exhaustive iteration on critical subparts.

1

Motivation

In data flow analysis equation systems involving bidirectional dependencies, i. e. dependencies from predecessor nodes as well as from successor nodes, are a well-known source for various kinds of difficulties. First, bidirectional equation systems are conceptually hard to understand. Mainly, this is caused by the lack of a corresponding operational specification like it is given by the the meet over all path (MOP) solution of a uni-directional data flow problem. Furthermore, Khedker and Dhamdhere recently proved that the costs for solving bidirectional data flow analysis problems may be significantly worse than for solving their unidirectional counterparts. This particularly holds for the only practically relevant class of bidirectional analyses, bit-vector based code motion problems. In fact, all known bidirectional problems are of this kind. Even more specifically, they are more or less variations of Morel’s and Renvoise’s pioneering algorithm for the elimination of partial redundancies [18,12,13,8,1,2,3,4,9]. Independently different researchers documented that bidirectionality is only required in programs that have critical edges [7,15], i. e. edges in a flow graph that directly lead from branch nodes to join nodes (see Fig. 1a for illustration). Ideally, critical edges can be completely eliminated by inserting empty synthetic nodes as depicted in Fig. 1b. In this example, the additional placement point enables the code motion transformation shown in Fig. 1c which eliminates the partial redundant G. Levi (Ed.): SAS’98, LNCS 1503, pp. 1–16, 1998. c Springer-Verlag Berlin Heidelberg 1998

2

O. R¨ uthing

a)

b)

1 x := a+b

2

c)

1 x := a+b

2

1 h := a+b x := h

2 h := a+b

3 y := a+b

3 y := a+b

3 y := h

Fig. 1. a) Critical Edge b) Edge splitting c) Transformational gain through edge splitting

computation on the path through node 1 and 3.1 However, in practice splitting of critical edges is sometimes avoided since it may cause additional unconditional jumps or decrease potential for pipelined execution.2 In this paper we investigate a new approach to code motion in the presence of critical edges. This is demonstrated by presenting a “critical” variant of our algorithm for lazy code motion [15]. However, the principal ideas straightforwardly carry over to all related code motion algorithms that employ bidirectional data flow analyses. Our algorithm is developed from a rigorous, specification oriented view. This particularly allows us to separate between different concerns. While safety in code motion is naturally associated with forward and backward oriented propagation of information, the presence of critical edges requires to impose additional homogeneity properties which can be expressed in terms of a side propagation of information. Actually, this clear separation allows us to avoid the usage of bidirectional dependencies in our specification. With regard to the variant of lazy code motion the contribution of this paper is threefold: – On a conceptual level we give a unidirectional specification of the problem. This particularly induces the first MOP characterization of code motion in the presence of critical edges. – We present a novel hybrid iteration strategy that separates the information flow along critical edges from the information flow along the uncritical ones. While the latter one is accomplished by an outer schedule proceeding in standard round-robin discipline the critical information flow is treated exhaustively by an inner schedule. – Almost as a by-product we obtain the first lifetime optimal algorithm for partial redundancy elimination in the presence of critical edges.

1 2

This is not possible in Fig. 1a, since hoisting a + b to node 2 introduces a new value on the rightmost path. Sometimes critical edges are not split only in situations that may harm the final code generation.

Bidirectional Data Flow Analysis in Code Motion: Myth and Reality

1.1

3

Related Work

As Khedker and Dhamdhere [14] and more recently Masticola et al. [17] noticed, critical edges do not add to the worst-case time complexity of iterative data flow analyses being based on a workset approach. However, this result cannot be generalized to bit-vector analyses where the iteration order has to be organized in a way such that structural properties of the flow graph are exploited in order to take maximum benefit of bit-wise parallel updates through efficient bit-vector operations. Hecht and Ullman [10] proved an upper bound on the number of roundrobin iterations that are necessary for stablization of monotone, unidirectional bit-vector problems. When proceeding in reverse postorder (or postorder for backward problems) d+2round-robin iterations are sufficient where dis the depth of the flow graph, i. e. the maximum number of backedges on an acyclic program path. Recently, Dhamdhere and Khedker [5,14] generalized this result towards bidirectional problems. However, a major drawback of their setting is that it is pinned to round-robin iterations. Unfortunately, such a schedule does not fit well to situations where information is side-propagated along critical edges. In this light, it is not astonishing that their results on the convergence speed of bidirectional bit-vector analyses are quite disappointing. They replace the depth dof a flow graph by its width w, which is the number of non-conform edge traversals on an information flow path.3 Unfortunately, the width is not a structural property of the flow graph, but varies with the problem under consideration, and unlike dwhich is 0 for acyclic programs is not even bounded in this case. Actually, the notion of width does not match to the intuition associated with the name, as even “slim” programs may have a large width. An intuitive reason for this behaviour is given in Fig. 2a which shows a program fragment with a number of critical edges. Let us consider that information flow in this example follows the equation Info(n) =

X m∈pred(n)

Info(m) +

X

Info(n0 )

n0 ∈succ(m)

which means that the information at node nis set to true if the information at a predecessor or the information of any “sibling” of nis true. We can easily see that the width of a flow graph with such a fragment directly depends on the number of critical edges, and therefore possibly grows linearly with the “length” of the program. It should be noted that such programs are by 3

Informatively, an information flow path is a sequence of backwards or forwards directed edges along which a change of information can be propagated. A forward traversal along a forward edge or a backward traversal along a backward edge are conform with a round-robin schedule proceeding (forwards) in reverse postorder. The other two kind of traversals are non-conform. Complemental notions apply to round-robin iterations proceeding in postorder.

4

O. R¨ uthing

a)

11

12

7 13

8

9

3 10

4

5

1 2

b)

6

Fig. 2. a) Acyclic program path responsible for the width of a program with reverse postordering of the nodes b) Slow information propagation in round-robin iterations

no means pathological and thus the linear growth of the width is not unlikely for real-life programs. In fact, considering the reverse postorder of nodes as given in Fig. 2a the large width is actually reflected in a poor behaviour of a round-robin iteration. Fig. 2b shows how the information slowly propagates along the obvious “path” displayed in this example being stopped in each round-robin iteration at a non-conform (critical) edge.4 Dhamdhere and Patil [6] proposed an elimination method for bidirectional problems that is as efficient as in the unidirectional case. However, it is restricted to a quite pathological class of problems, namely weakly bidirectional bit-vector problems and, as usual for elimination methods, is mainly designed for reducible control flow. Finally, our hybrid approach shares with the hybrid iteration strategy of [11] that it mixes a round-robin schedule with exhaustive subiterations. However, their subiterations are within strongly connected components and their approach is solely designed to speed up unidirectional iterations.

2

Preliminaries

We consider programs in terms of directed flow graphs = (N, E, s, e) with node set N , edge set E and unique start and end nodes s and e, respectively. Nodes n, m, . . . ∈ N represent (elementary) statements and are assumed to lie on a path from sto e. Finally, predecessors and successors of a node n ∈ N are denoted by pred(n)and succ(n), respectively, and P[n, m] stands for the set of finite paths between node n and m. 4

Shaded circles indicate the flow of informations along the “path”.

Bidirectional Data Flow Analysis in Code Motion: Myth and Reality

5

Local Predicates. As usual our reasoning is based on an arbitrary but fixed expression ϕ that is the running object for code movement. With each node of the flow graph two local predicates are associated. Comp(n): ϕ is computed at n, i. e. ϕ is part of the right-hand side expression associated with n. Transp(n): nis transparent for ϕ, i. e. none of ϕ’s variables is modified at n. Global Predicates. Based on these local predicates global program properties are specified. Usually, global predicates are associated with both entries and exits of nodes. In order to keep the presentation simple we assume that every node is split into an entry node and an empty exit node which inherit the set of predecessors and successors from the original node, respectively, and which are assumed to be connected by an edge leading from the entry node to the exit node. This step allows to restrict our reasoning to entry predicates. It should be noted, however, that this step is solely conceptual and does not eliminate any critical edge. In this paper partial redundancy elimination (PRE), or code motion (CM) as a synonym, stands for program transformations that 1. insert some instances of initialisation statements hϕ := ϕ at program points, where hϕ is a temporary variable that is exclusively assigned to ϕ and 2. replaces some original occurrences of ϕ by a usage of hϕ . In order to guarantee that the semantics of the argument program is preserved, we require that a code motion transformation must be admissible. Intuitively, this means that every insertion of a computation is safe, i. e. on no program path the computation of a new value is introduced at initialization sites, and that every substitution of an original occurrence of ϕ by hϕ is correct, i. e. hϕ always represents the same value as ϕ at use sites. This requires that hϕ is properly initialized on every program path leading to some use site in a way such that no modification occurs afterwards.5

3

Code Motion in the Absence of Critical Edges

Before presenting our new approach to PRE in the presence of critical edges we shall first briefly recall the basic steps of lazy code motion [15] as a typical representative of an algorithm that relies on the absence of critical edges. Lazy code motion was the first algorithm for partial redundancy elimination that succeeded in removing partial redundancies as good as possible, while avoiding any unnecessary register pressure. This was mainly achieved by a rigorous redesign of Morel’s and Renvoise’s algorithm. Starting from a specification oriented view the key points was a hierarchical separation between the primary and the secondary concern of partial redundancy elimination, namely to minimize the number of 5

For a formal definition see [16].

6

O. R¨ uthing

computations and to avoid unnecessary register pressure, respectively. This hierarchical organization is reflected in a two-step design of the algorithm: lazy code motion rests on busy code motion. In following we briefly summarize the details of these transformations. 3.1

Busy Code Motion

Busy code motion (BCM) [15,16,9] places initializations as early as possible while replacing all original occurrences of ϕ. This is achieved by determining the earliest program points, where an initialization is safe. Technically, the range of safe program points can be determined by separately computing down-safe and up-safe program points. Both are given through the greatest solutions of two uni-directional data flow analyses, respectively.6

DnSafe(n) = (n 6= e) · (Comp(n) + Transp(n) ·

Y

DnSafe(m))

m∈succ(n)

Y

UpSafe(n) = (n 6= s) · Transp(m) ·

(Comp(m) + UpSafe(m))

m∈pred(n) def

Safe (n) = UpSafe(n) + DnSafe(n) def

Earliest (n) = Safe (n) · ((n = s) +

X

Safe (m))

m∈pred(n)

Despite of its surprising simplicity, BCM already reaches computational optimality, i. e. programs resulting from this transformation have at most as many ϕ-occurrences on every path from sto eas any other result of an admissible code motion transformation (cp. [15,16]). 3.2

Lazy Code Motion

In addition to BCM, lazy code motion (LCM) takes the lifetimes of temporaries into account. This is accomplished by placing initialisations as late as possible but as early as necessary, where the latter requirement means “necessary in order to reach computational optimality”. Technically, this is achieved by determining the latest program points where a BCM-initialisation might be delayed to, which leads to one additional uni-directional data flow analysis.7 6 7

As common “·”,“+” and overlining stand for logical conjunction, disjunction and negation, respectively. In [15,16] an additional analysis is employed determining isolated program points, i. e. program points where initialisations are only used immediately afterwards. This aspects, however, can independently be treated by means of a postprocess. For the sake of simplicity we skip the isolation analysis in this paper.

Bidirectional Data Flow Analysis in Code Motion: Myth and Reality

Y

Delayed (n) = Earliest (n) + (n 6= s) ·

7

Delayed (m) · Comp(m)

m∈pred(n) def

Latest (n) = Delayed (n) · ( Comp(n) +

X

Delayed (m) )

m∈succ(n)

LCM is computationally optimal as well as lifetime optimal, i. e. the temporary associated with ϕhas a lifetime range that is included in the lifetime range of any other program resulting from a computationally optimal code motion transformation (cp. [15,16]). The lifetime range of a temporary comprises all nodes whose exits occur in between an initialisation site and a use site such that no other initialisations are situated in between.8

4

Code Motion in the Presence of Critical Edges

In this section our new approach to PRE in the presence of critical edges is elaborated in full details. First we shall investigate the principal differences to the setting presented in Sect. 3.2. As opposed to flow graphs without critical edges there are usually no computationally optimal representatives. In fact, Fig. 3 shows two admissible, but computationally incomparable transformations that cannot be improved any further. The first one is simply given by the identical transformation of the program in Fig. 3a, the result of the second one is displayed in Fig. 3b. Each of the resulting programs has exactly one computation on the path that is emphasised in the dark shade of grey, while having two computations on the path being emphasised in the light shade of grey, respectively. Thus there is no computationally optimal code motion transformation with respect to the original program in Fig. 3a. a)

b)

1

a+b

a+b 1 h := h

2 4

3 5

a+b

c)

6

a+b

2 h := a+b 4

3 5

h

a+b 1 h := h

6

a+b

3 5

h

2 h := a+b 4 6 h := a+b h

Fig. 3. a & b) Incomparable admissible program transformations c) Program degradation through a naive adaption of busy expression motion 8

In [20] we show that this optimality result is only adequate for flat universes of expressions. If both composite expressions and their subexpressions are moved, then the notion of lifetime optimality changes and a significantly more sophisticated technique has to be applied. Nonetheless, LCM still provides a basic ingredient of this approach.

8

O. R¨ uthing

This problem can be overcome by restricting the range of program transformations to those that are profitable, which means those that actually improve their argument programs. Note that this requirement excludes Fig. 3b as a reasonable code motion transformation. Obviously, profitability does not provide a further restriction for flow graphs without critical edges, where computationally optimal code motion transformations are granted to exist. In the presence of critical edges, however, this additional constraint is necessary in order to yield computationally optimal results at all. 4.1

Busy Code Motion

In this section we will develop a counterpart to BCM in the presence of critical edges. After briefly sketching the difficulties that prohibit a straightforward adaption of the uncritical solution, a correct approach is systematically developed from a specification that incorporates the special role of critical edges. Unfortunately, BCM as presented in Sect. 3.1 cannot straightforwardly be applied to flow graphs with critical edges. This is because such a naive adaption may include non-profitable transformations as it is illustrated in Fig. 3c, where the marked range of down-safe program points would yield earliest initialisation points at nodes 1, 2 and 6.9 Homogeneous Propagation of Down-Safety. The key for a useful critical variant of BCM is to impose an additional homogeneity requirement on downsafety that ensures that the information propagates either to all or to none of its predecessors, which grants that earliest program points become a proper upper borderline of the region of safe program points. In fact, in the absence of critical edges down-safety has the following homogeneity property: ∀ n ∈ N. DnSafe(n) ⇒ (∀ m ∈ pred(n). Safe (m) ∨ ∀ m ∈ pred(n). ¬DnSafe(m))

Note that the first term of the disjunction uses safety rather than down-safety, since propagation of down-safety needs not to be considered for predecessors that are up-safe anyhow.10 Now this propery has to be forced explicitly. For instance, in Fig. 3c node 6 as well as node 3 are down-safe, while node 4 is not. Therefore, let us consider the following notion of homogeneous down-safety: Definition 1 (Homogeneous Down-Safety). A predicate HDnSafe on the nodes of N is a homogeneous down-safety predicate iff for any n ∈ N 1. HDnSafe is conform with down-safety: HDnSafe (n) ⇒ (n 6= e) ∧ (Comp(n) ∨ Transp(n) ∧ ∀ m ∈ succ(n). HDnSafe (n)) 9

10

Modifying this example by removing the computation of a + bfrom node 1, would even result in a transformation that does not not improve any path while strictly impairing some. In the absence of critical edges this makes no difference to ∀ m ∈ pred(n). DnSafe(m).

Bidirectional Data Flow Analysis in Code Motion: Myth and Reality

9

2. HDnSafe is homogeneous: HDnSafe (n) ⇒ (∀ m ∈ pred(n). (HDnSafe (m) ∨ UpSafe(m)) ∨ ∀ m ∈ pred(n). ¬HDnSafe (m))

Obviously, homogeneous down-safety predicates are closed under “union”.11 Thus there exists a unique largest homogeneous down-safety predicate DnSafeHom , which gives rise to a homogeneous version of safety, too: def

∀ n ∈ N. SafeHom (n) = DnSafeHom (n) ∨ UpSafe(n) It should be noted that this definition is developed from a pure specification oriented reasoning and can be seen as a first rigorous characterization of downsafety in the presence of critical edges: down safety is described by a backward directed data flow problem which is restricted by additional homogeneity constraints. This is in contrast to other algorithms, where bidirectional equation systems are postulated in an ad-hoc fashion without any separation of their functional components. Earliest program points are defined as in the uncritical case, but with the difference of using the homogeneous version of down-safety in place of the usual one. def

EarliestHom (n) = DnSafeHom (n) · ((n 6= s) +

X

SafeHom (m))

m∈pred(n)

The earliest program points serve as insertion points of BCM for flow graphs with critical edges (CBCM). With a similar argumentation as for BCM is easy to prove that CBCM is indeed computationally optimal, however, only relatively to the profitable transformations.

Computing CBCM: The Data Flow Analyses. In this part we present how the specifying solution of CBCMcan be translated into appropriate data flow analyses determining the range of homogeneously safe program points. We will discuss three alternative approaches: (1) A “classical” one via bidirectional analyses, (2) a new non-standard approach that transforms the problem into one with purely unidirectional equations and (3) a hybrid approach that separates backwards flow from side propagation. The Bidirectional Approach The specification of Definition 1 can straightforwardly be transfered into a bidirectional equation system for down-system. 11

This means the predicate defined by the pointwise conjunction of the predicate values.

10

O. R¨ uthing

UpSafe(n) = (n 6= s) · Transp(m) ·

Y

(Comp(m) + UpSafe(m))

m∈pred(n)

DnSafeHom (n) = (n 6= e) · (Comp(n) + Transp(n) · Y Y DnSafeHom (m) · (UpSafe(n0 ) + DnSafeHom (n0 ))) m∈succ(n)

n0 ∈pred(m)

def

SafeHom (n) = UpSafe(n) + DnSafeHom (n)

Unfortunately, the above bidirectional data flow problem shares the problems sketched in Fig. 2 when subjected to a round-robin iteration strategy. In fact, violation of homogeneous safety follows exactly the same definition pattern as Info does in this example.12 Hence slow propagation of down-safety would be also apparent in CBCM. The Unidirectional Approach It is easy to see that in the bidirectional equation system there is no “true” Q forward propagation of down-safety information as the scope of the term n0 ∈pred(m) (UpSafe(n0 ) + DnSafeHom (n0 ))is restricted in its context. Rather this can be seen as a “side propagation” of down-safety information along zig-zag paths. For a technical description let us define the set of zig-zag successors zsucc(n)of a node n ∈ N as the smallest set of nodes satisfying (see Fig. 4 for illustration): 1. succ(n) ⊆ zsucc(n) 2. ∀m ∈ zsucc(n). succ(pred(m)) ⊆ zsucc(n) In our example zig-zag propgation of non-down-safety is further stopped at nodes where up-safety can be established. Hence we introduce a parameterized notion of zsucc(n)which is defined for M ⊆ N by: 1. succ(n) ⊆ zsucc M (n) 2. ∀m ∈ zsucc M (n). succ(pred(m) \ M ) ⊆ zsucc M (n) def

With XUS = rewritten as:

{m ∈ N | UpSafe(m)}the equation for down-safety can be

DnSafeHom (n) = n 6= e · (Comp(n) + Transp(n) ·

Y

DnSafeHom (m))

m∈zsucc XUS (n)

Note that this equation system can be seen as a unidirectional one that operates on a flow graph that is enriched by shortcut edges drawn between nodes and their zig-zag successors (see Fig. 4b). 12

Actually, here non-down-safety is propagated in a dual fashion.

Bidirectional Data Flow Analysis in Code Motion: Myth and Reality b)

a) 2

1

4

11

5

6

2

1

3

4

7

5

succ(1 )

3

6

7

zsucc(1)

Fig. 4. (a) Program fragment with a nest of critical edges (b) Zig-zag successors and virtual shortcut edges of node 1

However, we do not actually recommend to perform such a transformation, as this would require to introduce an unnecessarily large number of additional edges. For a zig-zag chain of k critical edges as shown in Fig. 4a the number of shortcut edges is of order k 2 . Although, long zig-zag chains of critical edges can be expected to be rare in practice, we will show that information propagation can be organized without such blow-up in the number of edges. However, the important contribution of the unidirectional approach is that it provides the first meet over all paths characterization of PRE in the presence of critical edges.13 The Hybrid Approach: The hybrid approach rather addresses the organization of the iteration process than the equation system itself. As we have learned, bidirectional problems like our formulation of homogeneous down-safety do not fit together with a round-robin schedule based upon postorder traversals. The hybrid approach modifies the conventional round-robin schedule by integrating zig-zag propagation of information. This is achieved by clustering the nodes in a flow graph in a way such that side propagation of information can take benefit of much potential for simultaneous work. The overall schedule of the approach can be sketched as follows: Preprocess: Collapsing of nodes according to side flow of information Outer Schedule: Process the collapsed nodes in postorder until stabilization is reached performing an Inner Schedule 1. For each node within the collapsed one perform information propagation along its outgoing uncritical edges 2. Perform exhaustive information propagation along the outgoing critical edges within the collapsed node In the following we will go into the details of this process. • The preprocess: Clustering of nodes groups together nodes of N according to the following equivalence relation: def

n ≡ m ⇔ zsucc(n) = zsucc(m) 13

Actually, only the notion of paths has to be extended towards paths across shortcut edges.

12

O. R¨ uthing

It should be noted that Gcan be decomposed into its equivalence classes easily by tracing zig-zag paths of critical edges originating at an unprocessed node. For instance, starting with node 1 in Fig. 4a we obtain the equivalence class {1, 2, 3}by following the critical edges. Clearly, this process can be managed in order O(e)where edenotes the number of edges in E. All nodes of an equivalence classes are collapsed into a single node that inherits all incoming and outgoing edges of its members (see Fig. 5 for illustration). a)

b)

Fig. 5. (a) Equivalent nodes (b) Collapsing equivalent nodes

• The outer schedule: The flow graph G0 that results from the collapsing preprocess is used in order to determine the round-robin schedule which drives information backwards. It should be noted that the depth of G0 may differ from the depth of the original flow graph Gin both directions: the depth may increase or decrease by collapsing. This is illustrated in Fig. 6. While collapsing nodes in Part a) decreases the depth, since the indicated path is no longer acyclic, collapsing in Part b) allows to construct a longer acyclic path as indicated by the dashed line connecting two independent acyclic paths.

a)

b)

Fig. 6. (a) Decrease of depth due to collapsing of nodes (b) Increase of depth due to collapsing of nodes

• The inner schedule: The first step of the inner schedule is quite trivial. Considering a node n ∈ N within the collapsed node under consideration and an

Bidirectional Data Flow Analysis in Code Motion: Myth and Reality

13

uncritical edge (n, m) ∈ Ethe value of DnSafeHom (n)is changed to false if and only if Transp(n) · DnSafeHom (m) holds.14 The second step of the inner schedule is the central innovation in our approach and has to be elaborated with some care. Within any collapsed node side propagation of down-safety information along critical edges is done exhaustively by using a subiteration process. Information propagation here means that for nodes n, min the collapsed node under consideration with m ∈ pred(succ(n))the value of DnSafeHom (m)is changed to false if and only if DnSafeHom (n) · UpSafe(n)holds. The crucial point, however, is to organize the information flow along the critical edges. The situation is easy if the zig-zag paths are acyclically shaped as displayed in Fig. 7a or Fig. 7b. In this case the equivalence class can be represented as a tree, which can already be built while preprocessing this class. Following the topological order of the tree, information can be propagated completely by a bottom-up traversal (from the leaves to the root) followed by a top-down traversal (from the root to the leaves). Unfortunately, in general there may be cycles of critical edges as shown in Fig. 7c and Fig. 7d. Hence a problem of the same difficulty as in the backward propagation of information shows up in the side-propagation step. However, separating both problems is useful as we expect nested cycles of critical edges to be a phenomenon that is extremely rare in practice. Nonetheless, to cope with them is quite straightforward. As in the acyclic case, the equivalence class can be represented as a tree with some additional non-tree edges establishing cycles. The only difference to the non-cyclic case is that the tree traversals have to be iterated more than once until the process gets stable. To estimate the number of traversal we borrow the arguments from conventional unidirectional analysis. Denoting the non-tree edges within the tree-like representation of an equivalence class as critical backedges the number of iterations is bound by dc , where dc is the maximum number of critical back edges along an acyclic path in any component representation. Complexity of the Hybrid Approach: All together the iteration of homogeneous down-safety in the hybrid approach requires to apply the outer schedule until stabilization. Since the inner schedule propagates the information completely within each collapsed node the overall effort can be estimated by (d0 + 2)(eu + 2(dc + 2)ec ) bit-vector steps, where eu and ec denote the number of uncritical and critical edges, respectively, d’ is the depth of the collapsed flow graph G0 and dc the critical depth as defined before. It is commonly argued that the depth of a flow graph is a reasonably small constant in practice. We already discussed that dc is at least as likely to be a small constant, too. Hence the algorithm is expected to behave linear in efor real-life programs. In particular, we succeed in giving the first linear worst-case estimation for acyclic programs as in our introductory example of Fig. 2. 14

Note that there are no uncritical edges directly connecting different nodes of an equivalence class.

14

O. R¨ uthing a)

c)

d)

b)

Fig. 7. Shapes of equivalent nodes: (a) chain of critical edges, (b) tree of critical edges, (c) cycle of critical edges and (d) structure with nested cycles of critical edges

4.2

Lazy Code Motion

Similar to the situation in Sect. 4.1 also the relevant analyses of LCM as defined in Sect. 3.2 cannot naively be adapted to flow graphs with critical edges. Again the reason for this behaviour lies in a homogeneity defect, but now with respect to delayability. In fact, for flow graphs without critical edges we have Delayed (n) ⇒ ( ∀ m ∈ succ(n). Delayed (m) ∨ ∀ m ∈ succ(n). ¬Delayed (m) ) This property may now be violated. Hence one has to force homogeneity explicitly in order to yield an appropriate critical variant of lazy code motion. Therefore, let us consider the following notion of homogeneous delayability. Definition 2 (Homogeneous Delayability). A predicate HDelayed on N is a homogeneous delayability predicate iff for any n ∈ N 1. HDelayed (n)is conform with delayability:

HDelayed (n) ⇒ EarliestHom (n) ∨ ((n 6= s) ∧ ∀ m ∈ pred(n). HDelayed (m) ∧ ¬Comp(m))

2. HDelayed (n)is homogeneous: Delayed (n) ⇒ ( ∀ m ∈ succ(n). HDelayed (m) ∨ ∀ m ∈ succ(n). ¬HDelayed (m))

Obviously, homogeneous delayability predicates are closed under “union”. Thus there exists a unique largest homogeneous delayability predicate DelayedHom . This gives rise to a new version of latestness characterizing the insertion points of lazy code motion for flow graphs with critical edges (CLCM). def

LatestHom (n) ⇔ DelayedHom (n) ∧ (Comp(n) ∨ ∃ m ∈ succ(n). ¬DelayedHom (m))

Bidirectional Data Flow Analysis in Code Motion: Myth and Reality

15

Using the same definitions for lifetime optimality as in Sect. 3.2 we succeed in proving lifetime optimality of CLCM. Computing CLCM. In analogy to Sect. 11 the delayability property can either be coded into a bidirectional equation system or, more interestingly, again be expressed using a unidirectional formulation:

DelayedHom (n) = EarliestHom (n) ∨ ((n 6= s) ∧ ∀ m ∈ zpred (n). DelayedHom (m) ∧ ¬Comp(m)) This definition is based on zig-zag predecessor, which are defined completely along the lines of zig-zag successors. However, in contrast to down-safety zpred needs not to be parameterized this time. Using this characterization the same techniques for hybrid iteration can be used as in Sect. 11.

5

Conclusion

We presented an adaption of lazy code motion to flow graphs with critical edges as a model how to cope with bidirectional dependencies in code motion. On the conceptual level we isolated homogeneity requirements as the source for bidirectional dependencies. This led to a new hybrid iteration strategy which is almost as fast as its unidirectional counterparts. This dramatically improves all known estimations for bidirectional bit-vector methods. Nonetheless, we still recommended to eliminate critical edges as far as possible, since critical edges are also responsible for problems of a different flavour [19]. However, any implementation of code motion that has to cope with critical edges will definitely benefit from the ideas presented in this paper.

References 1. D. M. Dhamdhere. A fast algorithm for code movement optimization. ACM SIGPLAN Notices, 23(10):172 – 180, 1988. 2. D. M. Dhamdhere. A new algorithm for composite hoisting and strength reduction optimisation (+ Corrigendum). International Journal of Computer Mathematics, 27:1 – 14 (+ 31 – 32), 1989. 3. D. M. Dhamdhere. A usually linear algorithm for register assignment using edge placement of load and store instructions. Journal of Computer Languages, 15(2):83 – 94, 1990. 4. D. M. Dhamdhere. Practical adaptation of the global optimization algorithm of Morel and Renvoise. ACM Transactions on Programming Languages and Systems, 13(2):291 – 294, 1991. Technical Correspondence. 5. D. M. Dhamdhere and U. P. Khedker. Complexity of bidirectional data flow analysis. In Conf. Record of the 20th ACM Symposium on the Principles of Programming Languages, pages 397–409, Charleston, SC, January 1993.

16

O. R¨ uthing

6. D. M. Dhamdhere and H. Patil. An elimination algorithm for bidirectional data flow problems using edge placement. ACM Transactions on Programming Languages and Systems, 15(2):312 – 336, April 1993. 7. D. M. Dhamdhere, B. K. Rosen, and F. K. Zadeck. How to analyze large programs efficiently and informatively. In Proc. ACM SIGPLAN Conference on Programming Language Design and Implementation’92, volume 27,7 of ACM SIGPLAN Notices, pages 212 – 223, San Francisco, CA, June 1992. 8. K.-H. Drechsler and M. P. Stadel. A solution to a problem with Morel and Renvoise’s “Global optimization by suppression of partial redundancies”. ACM Transactions on Programming Languages and Systems, 10(4):635 – 640, 1988. Technical Correspondence. 9. K.-H. Drechsler and M. P. Stadel. A variation of Knoop, R¨ uthing and Steffen’s lazy code motion. ACM SIGPLAN Notices, 28(5):29 – 38, 1993. 10. M. S. Hecht and J. D. Ullman. A simple algorithm for global data flow analysis problems. SIAM Journal on Computing, 4(4):519 – 532, 1977. 11. S. Horwitz, A. Demers, and T. Teitelbaum. An efficient general iterative algorithm for data flow analysis. Acta Informatica, 24:679 – 694, 1987. 12. S. M. Joshi and D. M. Dhamdhere. A composite hoisting-strength reduction transformation for global program optimization – part I. International Journal of Computer Mathematics, 11:21 – 41, 1982. 13. S. M. Joshi and D. M. Dhamdhere. A composite hoisting-strength reduction transformation for global program optimization – part II. International Journal of Computer Mathematics, 11:111 – 126, 1982. 14. U. P. Khedker and D. M. Dhamdhere. A generalized theory of bit vector data flow analysis. ACM Transactions on Programming Languages and Systems, 16(5):1472 – 1511, September 1994. 15. J. Knoop, O. R¨ uthing, and B. Steffen. Lazy code motion. In Proc. ACM SIGPLAN Conference on Programming Language Design and Implementation’92, volume 27,7 of ACM SIGPLAN Notices, pages 224 – 234, San Francisco, CA, June 1992. 16. J. Knoop, O. R¨ uthing, and B. Steffen. Optimal code motion: Theory and practice. ACM Transactions on Programming Languages and Systems, 16(4):1117– 1155, 1994. 17. P. M. Masticola, T. J. Marlowe, and B. G. Ryder. Lattice frameworks for multisource and bidirectional data flow problems. ACM Transactions on Programming Languages and Systems, 17(5):777 – 802, 1995. 18. E. Morel and C. Renvoise. Global optimization by suppression of partial redundancies. Communications of the ACM, 22(2):96 – 103, 1979. 19. O. R¨ uthing. Interacting Code Motion Transformations. Their Impact and their complexity. PhD thesis, Institut f¨ ur Informatik und Praktische Mathematik, Christian-Albrechts-Universit¨ at Kiel, Germany, 1997. Available as http://sunshine.cs.uni-dortmund.de/˜ruething/diss.ps.gz. 20. O. R¨ uthing. Optimal code motion in the presence of large expressions. In Proc. Internatinal Conference on Computer Languages (ICCL’98), Chicago, IL., 1998. IEEE.

On a New Method for Dataflow Analysis of Java Virtual Machine Subroutines Masami Hagiya and Akihiko Tozawa Department of Information Science, Graduate School of Science, University of Tokyo {hagiya,miles}@is.s.u-tokyo.ac.jp

Abstract. The bytecode verifier of the Java Virtual Machine, which statically checks the type safety of Java bytecode, is the basis of the security model of Java and guarantees the safety of mobile code sent from an untrusted remote host. However, the type system for Java bytecode has some technical problems, one of which is in the handling of subroutines. Based on the work of Stata and Abadi and that of Qian, this paper presents yet another type system for Java Virtual Machine subroutines. Our type system includes types of the form last(x). A value whose type is last(x) is the same as that of the x-th variable of the caller of the subroutine. In addition, we represent the type of a return address by the form return(n), which means returning to the n-th outer caller. By virtue of these types, we can analyze instructions purely in terms of type, and as a result the correctness proof of bytecode verification becomes extremely simple. Moreover, for some programs, our method is more powerful than existing ones. In particular, our method has no restrictions on the number of entries and exits of a subroutine.

1

Introduction

One contribution of Java is its bytecode verifier, which statically checks the type safety of bytecode for the JVM (Java Virtual Machine) prior to execution. Thanks to the bytecode verifier, bytecode sent from an untrusted remote host can be executed without the danger of causing type errors and destroying the entire security model of Java, even when the source code is not available. Verifying the type safety of bytecode (or native code) seems to be a new research area that is not only of technical interest but also of practical importance, due to the availability of remote binary code in web browsers and other applications. Much effort has been put into guaranteeing the security of Java programs, including the type safety of bytecode: – Security model for Java applets: The security model for Java applets is said to consist of three prongs: the bytecode verifier, the applet class loader and the security manager [9]. In this model, the bytecode verifier plays the most fundamental role, on which the other two prongs are based. If the bytecode verifier is cheated, the other two also become ineffective. G. Levi (Ed.): SAS’98, LNCS 1503, pp. 17–32, 1998. c Springer-Verlag Berlin Heidelberg 1998

18

M. Hagiya and A. Tozawa

– Type safety of source code: The type safety of Java programs has been proved, either formally (using a theorem proving assistant), or rigorously (but not formally) [3,4,17,12]. This means that a program that has passed static type checking is guaranteed to cause no type errors while it is running. – Safety of class loading: Java allows classes to be loaded lazily, i.e., only when classes are actually accessed. In order to support various loading disciplines, Java allows programmers to define their own class loaders. This has opened one of the security holes in Java [14]. To avoid such a security hole, Dean formalized part of the class loader functionality and formally proved its correctness using PVS [2]. Goldberg’s main concern is also class loading, though bytecode verification is addressed [5]. – Type safety of bytecode: The bytecode verifier statically checks the type safety of Java bytecode. If the bytecode verifier accepts incorrectly typed bytecode, it will break the entire security model of Java. It guarantees that no type error occurs at each instruction by performing dataflow analysis on the bytecode. Besides the researches into different aspects of security mentioned above, there are also some on-going projects that are developing more secure network programming environments [7,15]. This paper concerns bytecode verification. Since this is the basis of the entire security model of Java, it is desirable to rigorously prove that any bytecode program that has passed bytecode verification will never cause a runtime type error. In order to be able to show the correctness of the bytecode verifier, one has to begin by formally specifying the operational semantics of the virtual machine (e.g., [1]), and then give the formal specification of the bytecode verifier, based on its informal specification written in English [8]. Qian rigorously defined the operational semantics of a subset of the JVM and formulated the bytecode verifier as a type system [13]. He then succeeded in proving the correctness of the bytecode verifier, though not completely. Bytecode verification of the JVM has some technical challenges. One is that of handling object initialization, as objects created but not yet initialized may open a security hole. In Qian’s work, much attention is paid to the handling of object initialization. Another is that of handling the polymorphism of subroutines. This paper addresses this issue. Inside a JVM subroutine, which is the result of compiling a finally clause in a try statement, local variables may have values of different types, depending on the caller of the subroutine. This is a kind of polymorphism. To investigate how to analyze JVM subroutines, Stata and Abadi defined a type system for a small subset of the JVM and proved its correctness with respect to the operational semantics of the subset [16]. Qian’s system is similar to that of Stata and Abadi in its handling of subroutines [13]. Both systems faithfully follow the specification of the bytecode verifier, and make use of information as to which variables are accessed or modified in a subroutine. Those variables that are not accessed or modified are simply ignored during analysis of the subroutine.

On a New Method for Dataflow Analysis

19

This paper takes a different approach. We introduce types of the form last(x). A value whose type is last(x) is the same as that of the x-th variable of the caller of the subroutine. In addition, we represent the type of a return address by the form return(n), which means returning to the n-th outer caller. Our approach has the following advantages. – By virtue of the last and return types, we can analyze instructions purely in terms of types. As a result, the proof of the correctness of bytecode verification becomes extremely simple, and we do not need a separate analysis on variable access or modification. – For some programs (unfortunately, not those produced by the Java compiler), our method is more powerful than existing ones. In particular, our method has no restrictions on the number of entries and exits of a subroutine. Stata and Abadi enforce a stack-like behavior on subroutine calls with their analysis, which does not account for out-of-order returns from subroutines, though they are actually produced by the Java compiler. Due to these advantages, we hope that our method can be modified and applied to the analysis of other kinds of bytecode or native code [10,11]. This paper is organized as follows: In the next section, we explain JVM subroutines in more detail and our approach to bytecode verification. In Sect. 3, a subset of the JVM, similar to that of Stata and Abadi, is defined. In Sect. 4, our method for analysis of bytecode is described and its correctness is shown. In Sect. 5, issues of implementation are briefly discussed. Section 6 offers concluding remarks.

2

Analysis of Subroutines

The JVM is a classical virtual machine consisting of – a program counter, – an array for storing the values of local variables, – a stack for placing arguments and results of operators, called the operand stack, – a heap for storing method code and object bodies, and – a stack for frames, each of which is allocated for each invocation of a method and consists of the program counter, the local variables, and the operand stack. To allow for checking the type safety of bytecode, it prepares different instructions for the same operation depending on the type of the operands. For example, it has the instruction istore for storing integers and the instruction fstore for storing floating-point numbers. JVM subroutines are used mainly for compiling the finally clauses of try statements of Java. Notice that subroutine calls are completely different from method calls. A subroutine is locally defined inside a method and, unlike a method, is not allowed to call itself recursively.

20

M. Hagiya and A. Tozawa

In this paper, we define a virtual machine based on JVML0, which was formulated by State and Abadi for investigating how to analyze JVM subroutines. Subroutines are called by instructions of the following form. jsr(L) An instruction of this form pushes the address of its next instruction (i.e., the return address) onto the operand stack and jumps to the subroutine L. Subroutines are usually defined as follows. L : store(x) .. . ret(x)

store(x) pops a value from the operand stack and stores it in the x-th local variable. ret(x) is an instruction for jumping to the address stored in the x-th local variable. Note that, in contrast to the JVM, our virtual machine has only one instruction for the store operation. Subroutines in the JVM have made bytecode verification more difficult for the following reasons. – The return address of ret(x) can only be determined after the values of the local variables have been analyzed by the bytecode verifier. On the other hand, return addresses affect the control flow and the analysis of local variables. – In some situations, a subroutine does not return to the immediate caller, but returns to an outer caller, such as the caller of the caller. – Inside a subroutine, local variables may have values of different types, depending on the caller of the subroutine. In this paper, we introduce types of the form last(x) in order to address the last problem. A value having this type must have the same value as that of the x-th local variable in the caller of a subroutine. As an example, let us consider the following program. const0 store(1) 2 : jsr(7) constNULL store(1) 5 : jsr(7) halt

7 : store(0) load(1) store(2) ret(0)

Subroutine 7 is called from two callers (2 and 5). The return address is stored in variable 0 (the 0-th local variable). The value of variable 1 is an integer when the subroutine is called from caller 2, and is an object pointer when called from caller 5. This is a typical case in which a subroutine is polymorphic. In this

On a New Method for Dataflow Analysis

21

program, the value of variable 1 is copied to variable 2. As the JVM has different store instructions depending on the type of the operand, the above program is impossible under the JVM. However, even if the JVM allowed the untyped store instruction, the bytecode verifier of the JVM would signal an error for the above program, because it always assigns a unique type for any variable accessed in a subroutine. According to the specification of the JVM [8], – For each instruction and each jsr needed to reach that instruction, a bit vector is maintained of all local variables accessed or modified since the execution of the jsr instruction. – For any local variable for which the bit vector (constructed above) indicates that the subroutine has accessed or modified, use the type of the local variable at the time of the ret. – For other local variables, use the type of the local variable before the jsr instruction. The work of Stata and Abadi and that of Qian faithfully follow this specification. Local variables that are accessed or modified in each subroutine are recorded. Those variables that are not accessed or modified are simply ignored during the subsequent analysis of the subroutine. The method proposed in this paper assigns a type of the form last(1) to variable 1 in subroutine 7. This means that it includes a value passed from variable 1 of the caller. This type is propagated through instructions in the subroutine. In particular, by the instruction store(2), the type of variable 2 becomes last(1). This information is then used when the control returns from the subroutine to the caller. In this way, the polymorphism of local variables in a subroutine is expressed by types of the form last(x). In our method, return addresses have types of the form return(n). A type of the form return(n) means to return to the n-th outer caller. For example, the address returning to the immediate caller has type the return(1), while the address returning to the caller of the caller has type the return(2). In Stata and Abadi’s work (and similarly in Qian’s work), return addresses have types of the form (ret-from L), where L is the address of the subroutine, from which the callers of the subroutines are obtained. In our analysis, the callers are obtained from the set of histories assigned to each instruction (cf. Sect. 4.3).

3

Virtual Machine

In this section we formalize a subset of the JVM, which resembles that of Stata and Abadi [16]. Differences are mainly for the examples by which we want to show the power of our framework.

22

3.1

M. Hagiya and A. Tozawa

Values

A value is a return address or an integer or an object pointer. We can easily add other kinds of values, such as that of floating point number. In the following formal treatment of the operational semantics of the virtual machine, a return address has the constructor retaddr, an integer the constructor intval, and an object pointer the constructor objval. They all take an integer as an argument. 3.2

Instructions

A bytecode program is a list of instructions. An instruction takes one of the following formats. jsr(L) load(x) const0 inc(x) if0(L) halt

(L: subroutine address) (x: variable index) (x: variable index) (L: branch address)

ret(x) (x: variable index) store(x) (x: variable index) constNULL ifNULL(L)

(L: branch address)

Each mnemonic is considered as a constructor of instructions. Some of the mnemonics takes a nonnegative integer x or L as an operand. 3.3

Operational Semantics

The virtual machine consists of – the program, which is a list of instructions and denoted by P , – the program counter, which is an index to P , – the local variables, where the list of values of the local variables is denoted by f , and – the operand stack, denoted by s. Let us use the notation l[i] for extracting the i-th element of list l, where the first element of l has the index 0. The i-th instruction of the program P is denoted by P [i]. The value of the x-th local variable is denoted by f [x]. The p-th element of the operand stack s is denoted by s[p], where s[0] denotes the top element of s. As in the work by Stata and Abadi, the operational semantics of the virtual machine is defined as a transition relation between triples of the form hi, f, si, where i is the program counter, i.e., the index to the program P , f the value list of the local variables, and s the operand stack. While the length of s may change during execution of the virtual machine, the length of f , i.e., the number of local variables is unchanged. The program P , of course, never changes during execution. The transition relation is defined as follows.

On a New Method for Dataflow Analysis

23

– If P [i] = jsr(L), then hi, f, si → hL, f, retaddr(i + 1)::si. The return address retaddr(i + 1) is pushed onto the operand stack. The operator :: is the cons operator for lists. – If P [i] = ret(x) and f [x] = retaddr(j + 1), then hi, f, si → hj + 1, f, si. – If P [i] = load(x), then hi, f, si → hi + 1, f, f [x]::si. – If P [i] = store(x), then hi, f, v::si → hi + 1, f [x 7→v], si. The notation f [x 7→v] means a list whose element is the same as that of f except for the x-th element, which is set to v. – If P [i] = const0, then hi, f, si → hi + 1, intval(0)::si. – If P [i] = constNULL, then hi, f, si → hi + 1, objval(0)::si. – If P [i] = inc(x) and f [x] = intval(k), then hi, f, si → hi + 1, f [x 7→intval(k + 1)], si. – If P [i] = if0(L), then hi, f, intval(0)::si → hL, f, si. If P [i] = if0(L) and k 6= 0, then hi, f, intval(k)::si → hi + 1, f, si. – If P [i] = ifNULL(L), then hi, f, objval(0)::si → hL, f, si. If P [i] = ifNULL(L) and k 6= 0, then hi, f, objval(k)::si → hi + 1, f, si. The transition relation → is considered as the least relation satisfying the above conditions. The relation is defined so that when a type error occurs, no transition is defined. This means that to show the type safety of bytecode is to show that a transition sequence stops only at the halt instruction. For proving the correctness of our bytecode analysis, we also need another version of the operational semantics that maintains invocation histories of subroutines. This semantics corresponds to the structured dynamic semantics of Stata and Abadi. The transition relation is now defined for quadruples of the form hi, f, s, hi, where the last component h is an invocation history of subroutines. It is a list of addresses of callers of subroutines. This component is only changed by the jsr and ret instructions. – If P [i] = jsr(L), then hi, f, s, hi → hL, f, retaddr(i + 1)::s, i::hi. Note that the address i of the caller of the subroutine is pushed onto the invocation history. – If P [i] = ret(x), f [x] = retaddr(j + 1) and h = h0 @[j]@h00 , where j does not appear in h0 , then hi, f, s, hi → hj + 1, f, s, h00 i. The operator @ is the append operator for lists. For other instructions, the invocation histories before and after transition are the same. As for the two transition relations, we immediately have the following proposition. Proposition 1: If hi, f, s, hi → hi0 , f 0 , s0 , h0 i, then hi, f, si → hi0 , f 0 , s0 i.

4 4.1

Analysis Types

Types in our analysis are among the following syntactic entities:

24

M. Hagiya and A. Tozawa

>, ⊥ (top and bottom) return(n) (n: caller level)

INT, OBJ, · · · (basic types) last(x) (x: variable index)

A type is >, ⊥, a basic type, a return type, or a last type. In this paper, we assume as basic types INT, the type of integers, and OBJ, the type of object pointers. It is easy to add other basic types, such as that of floating point numbers. return types and last types are only meaningful inside a subroutine. A return type is the type of a return address. For positive integer n, return(n) denotes the type of the address for returning to the n-th outer caller. For example, return(1) denotes the type of the address for returning to the direct caller of the subroutine, and return(2) the type of the address for returning to the caller of the caller. A last type means that a value is passed from the caller of the subroutine. For nonnegative integer x, last(x) denotes the type of a value that was stored in the x-th local variable of the caller. A value can have this type only when it is exactly the same as the value of the x-th local variable when the subroutine was called. 4.2

Order among Types

We define the order among types as follows. > > INT > ⊥ > > return(n) > ⊥

> > OBJ > ⊥ > > last(x) > ⊥

Since we do not distinguish object pointers by their classes in this paper, the order is flat, with > and ⊥ as the top and bottom elements. This order is extended to lists of types. For type lists t1 and t2 , t1 > t2 holds if and only if t1 and t2 are of the same length and t1 [i] > t2 [i] holds for any i ranging over the indices for the lists. 4.3

Target of Analysis

The target of our bytecode analysis is to obtain the following pieces of information for the i-th instruction of the given program P . Fi

Si

Hi

Fi is a type list. Fi [x] describes the type of f [x], i.e., the value of the x-th local variable of the virtual machine. Si is a also type list. Each element of Si describes the type of the corresponding element of the operand stack of the virtual machine. Both Fi and Si describe the types of the components of the virtual machine just before the i-th instruction is executed. Hi is a set of invocation histories for the i-th instruction. F , S and H should follow a rule that is defined for each kind of P [i]. The rule says that certain conditions must be satisfied before and after the execution of P [i].

On a New Method for Dataflow Analysis

25

Rule for jsr) If h ∈ Hi and P [i] = jsr(L), then the following conditions must be satisfied. – 0 ≤ L < |P |. – For each variable index y, either FL [y] ≥ return(n + 1) (if Fi [y] = return(n)), and FL [y] ≥ Fi [y] (if Fi [y] is neither return nor last) or FL [y] ≥ last(y) (even if Fi [y] is last). – |SL | = |Si | + 1, where |l| denotes the length of list l. – SL [0] ≥ return(1). – For each index p, where 0 ≤ p < |Si |, Si [p] is not last, SL [p + 1] ≥ Si [p] (if Si [p] is not return), and SL [p + 1] ≥ return(n + 1) (if Si [p] = return(n)). – i does not appear in h. (Recursion is not allowed.) – i::h ∈ HL . Note that when Fi [y] is not last, FL [y] cannot be determined uniquely. We must make a nondeterministic choice between return(n + 1) and Fi [y]. See Sect. 5 for more discussions on the implementation of the analysis. The following figures show the two possibilities for typing local variables inside a subroutine. In this example, it is assumed that there is only one caller (2) of subroutine 7. The column [Fi [0], Fi [1], Fi [2]] shows the types of local variables before each instruction is executed. At subroutine 7, it is set to [>, INT, >] or [l(0), l(1), l(2)]. There are more possibilities. For example, one could also set it to [>, >, >], but this possibility is subsumed by the first. i 0 1 2 3 7 8 9 10

instruction const0 store(1) jsr(7) constNULL ··· store(0) load(1) store(2) ret(0)

[Fi [0], Fi [1], Fi [2]] [>, >, >] [>, >, >] [>, INT, >] [>, INT, INT]

Si [] [INT] [] []

Hi {[]} {[]} {[]} {[]}

[Fi [0], Fi [1], Fi [2]] [>, >, >] [>, >, >] [>, INT, >] [>, INT, INT]

Si [] [INT] [] []

Hi {[]} {[]} {[]} {[]}

[>, INT, >] [r(1), INT, >] [r(1), INT, >] [r(1), INT, INT]

[r(1)] [] [INT] []

{[2]} {[2]} {[2]} {[2]}

[l(0), l(1), l(2)] [r(1), l(1), l(2)] [r(1), l(1), l(2)] [r(1), l(1), l(1)]

[r(1)] [] [l(1)] []

{[2]} {[2]} {[2]} {[2]}

Rule for ret) If h ∈ Hi and P [i] = ret(x), then the following conditions must be satisfied. – Fi [x] = return(n).

26

M. Hagiya and A. Tozawa

– h = h0 @[j]@h00 , where |h0 | = n − 1. – 0 ≤ j + 1 < |P |. – For each variable index y, Fj+1 [y] ≥ follow last(n, h, Fi [y]). – Sj+1 ≥ follow last(n, h, Si ). – h00 ∈ Hj+1 . follow last is a function for extracting the type of a variable in a caller of a subroutine according to an invocation history. For nonnegative integer n, invocation history h and type t, follow last(n, h, t) is defined as follows. follow last(0, h, t) = t follow last(n + 1, i::h, return(m)) = if m > n + 1 then return(m − n − 1) else > follow last(n + 1, i::h, last(x)) = follow last(n, h, Fi [x]) follow last(n + 1, i::h, t) = t (otherwise) follow last is extended to type lists, i.e., follow last(n, h, t) is also defined when t is a type list. Rule for load) If h ∈ Hi and P [i] = load(x), then the following conditions must be satisfied. – 0 ≤ i + 1 < |P |. Fi+1 ≥ Fi . Si+1 ≥ Fi [x]::Si . h ∈ Hi+1 . Rule for store) If h ∈ Hi and P [i] = store(x), then the following conditions must be satisfied. – 0 ≤ i + 1 < |P |. Si = t::t. Fi+1 ≥ Fi [x 7→t]. Si+1 ≥ t. h ∈ Hi+1 . Rule for const0) If h ∈ Hi and P [i] = const0, then the following conditions must be satisfied. – 0 ≤ i + 1 < |P |. Fi+1 ≥ Fi . Si+1 ≥ INT ::Si . h ∈ Hi+1 . The rule for constNULL is similar. Rule for inc) If h ∈ Hi and P [i] = inc(x), then the following conditions must be satisfied. – 0 ≤ i + 1 < |P |. Fi [x] = INT. Fi+1 ≥ Fi . Si+1 ≥ Si . h ∈ Hi+1 . Rule for if0) If h ∈ Hi and P [i] = if0(L), then the following conditions must be satisfied. – 0 ≤ L < |P |. 0 ≤ i + 1 < |P |. – Si = INT ::t. FL ≥ Fi . Fi+1 ≥ Fi . SL ≥ t. Si+1 ≥ t. – h ∈ HL . h ∈ Hi+1 . The rule for ifNULL is similar. Rule for halt) There is no rule for halt.

On a New Method for Dataflow Analysis

4.4

27

Correctness of Analysis

In order to state the correctness of our analysis, we first introduce the following relation. hv, hi : t v is a value and h is an invocation history. t is a type. By hv, hi : t, we mean that the value v belongs to the type t provided that v appears with the invocation history h. Following is the definition of this relation. – – – – –

hv, hi : >. hintval(k), hi : INT. hobjval(k), hi : OBJ. If h[n − 1] = j, then hretaddr(j + 1), hi : return(n). If hv, hi : Fi [x], then hv, i::hi : last(x).

This definition is also inductive, i.e., hv, hi : t holds if and only if it can be derived only by the above rules. We have two lemmas. Lemma 1: If hv, hi : t and t0 ≥ t, then hv, hi : t0 . Lemma 2: Let h0 be a prefix of h of length n and h00 be its corresponding suffix, i.e., h = h0 @h00 and |h0 | = n. If hv, hi : t, then hv, h00 i : follow last(n, h, t). We say that the quadruple hi, f, s, hi is sound with respect to hF, S, Hi and write hi, f, s, hi : hF, S, Hi, if the following conditions are satisfied. – – – – –

0 ≤ i < |P |. For each variable index y, hf [y], hi : Fi [y]. For each index p for s, hs[p], hi : Si [p]. h ∈ Hi . h does not have duplication, i.e., no element of h occurs more than once in h.

We have the following correctness theorem. It says that if F , S and H follow the rule for each instruction of P , then the soundness is preserved under the transition of quadruples. This means that if the initial quadruple is sound, then quadruples that appear during execution of the virtual machine are always sound. Theorem (correctness of analysis): Assume that F , S and H follow the rule for each instruction of P . If hi, f, s, hi : hF, S, Hi and hi, f, s, hi → hi0 , f 0 , s0 , h0 i, then hi0 , f 0 , s0 , h0 i : hF, S, Hi. The theorem is proved by the case analysis on the kind of P [i]. In this short paper, we only examine the case when P [i] = ret(x). Assume that hi, f, s, hi : hF, S, Hi and hi, f, s, hi → hi0 , f 0 , s0 , h0 i. Since F , S and H follow the rule for ret, the following facts hold. (i) Fi [x] = return(n).

28

M. Hagiya and A. Tozawa

(ii) h = h1 @[j]@h2 , where |h1 | = n − 1. (iii) 0 ≤ j + 1 < |P |. (iv) For each variable index y, Fj+1 [y] ≥ follow last(n, h, Fi [y]). (v) Sj+1 ≥ follow last(n, h, Si ). (vi) h2 ∈ Hj+1 . By (i) and the soundness of hi, f, s, hi, hf [x], hi : return(n). Therefore, by (ii), f [x] = retaddr(j+1) and i0 = j+1. Moreover, since h does not have duplication, h1 does not contain j. This implies that h0 = h2 . We also have that f 0 = f and s0 = s. Let us check the conditions for the soundness of hi0 , f 0 , s0 , h0 i = hj+1, f, s, h2 i. – By (iii), 0 ≤ i0 < |P |. – By (iv), Fi0 [y] ≥ follow last(n, h, Fi [y]). By the soundness of hi, f, s, hi, hf [y], hi : Fi [y]. By Lemma 2, hf [y], h0 i : follow last(n, h, Fi [y]). Therefore, by Lemma 1, hf [y], h0 i : Fi0 [y]. – Similarly, by (v), we have that hs[p], h0 i : Si0 [p]. – By (vi) and since h0 = h2 , h0 ∈ Hi0 . – Finally, since h does not have duplication, h0 does not have duplication, either. Proposition 2: If hi, f, s, hi : hF, S, Hi and hi, f, si → hi0 , f 0 , s0 i, then there exists some h0 such that hi, f, s, hi → hi0 , f 0 , s0 , h0 i. The only case that must be examined is that of ret. Note that h0 is uniquely determined. The above proposition guarantees that if F , S and H follow the rule for each instruction and the initial quadruple hi, f, s, hi is sound, then the transition sequence starting from the triple hi, f, si can always be lifted to a sequence starting from hi, f, s, hi. This means that the semantics for triples and that for quadruples coincide when F , S and H follow the rule for each instruction. A similar lemma is stated in [16], which establishes a correspondence between their stackless semantics and their structured semantics. Lemma 3: If hi, f, s, hi : hF, S, Hi, then hi, f, s, hi has the next state unless P [i] = halt. The following final theorem, derived from the above lemma and the previous theorem, guarantees the type safety of bytecode. This corresponds to Theorem 1 (Soundness) in [16]. Theorem (type safety): If F , S and H follow the rule for each instruction and the initial quadruple hi, f, s, hi is sound, then a transition sequence stops only at the halt instruction.

On a New Method for Dataflow Analysis

4.5

29

Example

Let us abbreviate last(x) and return(n) by l(x) and r(n), respectively. Following is the result of analyzing the example in Sect. 2. i 0 1 2 3 4 5 6 7 8 9 10

instruction const0 store(1) jsr(7) constNULL store(1) jsr(7) halt store(0) load(1) store(2) ret(0)

[Fi [0], Fi [1], Fi [2]] [>, >, >] [>, >, >] [>, INT, >] [>, INT, INT] [>, INT, INT] [>, OBJ, INT] [>, OBJ, OBJ] [l(0), l(1), l(2)] [r(1), l(1), l(2)] [r(1), l(1), l(2)] [r(1), l(1), l(1)]

Si [] [INT] [] [] [OBJ] [] [] [r(1)] [] [l(1)] []

Hi {[]} {[]} {[]} {[]} {[]} {[]} {[]} {[2], [5]} {[2], [5]} {[2], [5]} {[2], [5]}

The rule for ret(0) at 10 is satisfied because, for [2] ∈ H10 , – F10 [0] = return(1). [2] = []@[2]@[]. 0 ≤ 2+1 = 3 < 11. – F3 [0] = > ≥ follow last(1, [2], F10 [0]) = follow last(1, [2], return(1)) = >. – F3 [1] = INT ≥ follow last(1, [2], F10 [1]) = follow last(1, [2], last(1)) = F2 [1] = INT. – F3 [2] = INT ≥ follow last(1, [2], F10 [2]) = follow last(1, [2], last(1)) = F2 [1] = INT. – S3 = []. [] ∈ {[]} = H3 , and similarly for [5] ∈ H10 . 4.6

Returning to an Outer Caller

When P [i] = ret(x) returns to the caller of the caller, for example, Fi [x] must be equal to the type return(2). In this case, Hi should consist of histories of length at least 2. If [j1 , j2 , j3 , · · ·] is in Hi , P [i] returns to i0 = j2 +1 and Fi0 should satisfy Fi0 [y] ≥ follow last(2, [j1 , j2 , j3 , · · ·], Fi [y]). If Fi [y] = last(y) and Fj1 = last(y), for example, then follow last(2, [j1 , j2 , j3 , · · ·], Fi [y]) = Fj2 [y]. This is how information at j2 (which is a jsr) is propagated to i0 = j2 +1. If Fi [y] is not last, information at i is propagated.

30

5

M. Hagiya and A. Tozawa

Implementation

A dataflow analysis is usually implemented by an iterative algorithm. For each instruction, we check if the rule for the instruction is satisfied by F , S and H. If not, we update F , S and H accordingly, and check the next instruction that is affected by the update. There are two problems for implementing our analysis by such an iterative algorithm. Firstly, the rule for jsr does not uniquely determine FL [y] when Fi [y] is not last. We have two choices: one is to set FL [y] = last(y), and the other is to set FL [y] = Fi [y] (or FL [y] = return(n + 1) if Fi [y] = return(n)). In our current implementation, we first set FL [y] = last(y) and proceed the analysis. If the analysis fails at some point because FL [y] = last(y), we take the alternative and redo the analysis from L. (We need not completely abandon the work after we set FL [y] = last(y).) The second problem is that by a na¨ıve iterative algorithm, a subroutine is analyzed each time it is called. In the worst case this may require an exponential number of steps with respect to the length of the program. This problem can be avoided by representing invocation histories by a node in the call graph of the program, which is a graph whose nodes are addresses of subroutines and whose (directed) edges are labeled with addresses of jsr instructions. Since the JVM does not allow recursion, it is a connected acyclic graph with a unique root node representing the initial address. A path from the root to a node in the graph corresponds to an invocation history by concatenating the labels of edges in the path. Each node in the graph then represents the set of all the invocation histories from the root to the node. Now, instead of keeping a set of invocation histories (i.e., Hi ), we can keep a set of nodes in the graph. From a program in the following (left), the call graph in the right is constructed. The node L3 represents the set {[c, b, a], [c, b0 , a0 ], [c0 , b, a], [c0 , b0 , a0 ]} of invocation histories. a : jsr(L1 ) ... a0 : jsr(L01 ) ... L1 : . . . b : jsr(L2 ) ... L01 : . . . b0 : jsr(L2 ) ... L2 : . . . c : jsr(L3 ) ... c0 : jsr(L3 ) ... L3 : . . .

L1

L3

HH HH Y c0

@ I @a @ @

b

c

L2

@ I @0 [email protected] @

a0 L01

On a New Method for Dataflow Analysis

31

If histories are represented by nodes in the call graph, then the values that Hi can take are bounded by the set of all the nodes in the call graph. This means that Hi can only be updated for the number of times equal to the number of nodes. The number of overall updates is, therefore, limited by n2 , where n is the number of instructions in the program. In order to achieve a polinomial complexity of the entire analysis, however, the nondeterministic choice in the handling of jsr instructions must be restricted.

6

Concluding Remark

Since we introduced types of the form last(x), it has become possible to assign types to polymorphic subroutines that move a value from a variable to another. Our analysis is towards the real polymorphism of subroutines in binary code, because we do not simply ignore unaccessed variables. For some programs, our method is more powerful than existing ones. In particular, we impose no restrictions on the number of entries and exits of a subroutine. It is also important that the proof of the correctness of bytecode verification becomes extremely simple. We only formalized a very small subset of the JVM. We believe that the framework of the paper can be extended to the full language. The extension is almost straight forward. In particular, adding new kinds of types seems to cause no difficulty. The correct handling of exceptions and that of object initialization are problematic but are not impossible [13,5]. The resulting bytecode verifier is expected to be more powerful than the existing one, so the Java compiler will gain more freedom in bytecode generation. It is also interesting whether the framework can be applied to the analysis of other kinds of bytecode or native code. Handling recursive calls is the key to such applications. In order to allow recursion, we must be able to represent histories of an indefinite length by a kind of regular expression. Stacks generated by recursive calls should also be represented by regular expressions. All this is left as future work. A dataflow analysis, in general, assigns an abstract value xi to the i-th instruction so that a certain predicate P (xi , σ) always holds for any state σ that reaches the i-th instruction. To this end, for any transition σ → σ 0 , where σ 0 is at i0 , one must show that P (xi , σ) implies P (xi0 , σ 0 ). Since σ corresponds to hi, f, s, hi in our analysis, xi seems to correspond to hFi , Si , Hi i. However, the predicate hi, f, s, hi : hF, S, Hi, which should correspond to P (xi , σ), does not only refer to hFi , Si , Hi i. When Fi [y] is last, it also refers to Fh[0] . This means that in terms of last types, our analysis relates values assigned to different instructions. This makes the analysis powerful while keeping the overall data structure for the analysis compact. The representation of invocation histories by a node in the call graph is also for making the data structure small and efficient. By this representation, the number of updates of Hi is limited by the size of the call graph, and an iterative algorithm is expected to stop in polynomial time with respect to the program size. This kind of complexity analysis should be made rigorous in the future.

32

M. Hagiya and A. Tozawa

Acknowledgments. The author would like to thank Zhenyu Qian and Mart´ın Abadi for their comments on the earlier draft of this paper. He also thanks anonymous referees whose comments greatly improved the paper.

References 1. Richard M. Cohen: The Defensive Java Virtual Machine Specification, Version Alpha 1 Release, DRAFT VERSION, 1997. http://www.cli.com/software/djvm/html-0.5/djvm-report.html 2. Drew Dean: The Security of Static Typing with Dynamic Linking, Fourth ACM Conference on Computer and Communication Security, 1997, pp.18–27. http://www.cs.princeton.edu/sip/pub/ccs4.html 3. Sophia Drossopoulou and Susan Eisenbach: Java is Type Safe — Probably, ECOOP’97 — Object-Oriented Programming, Lecture Notes in Computer Science, Vol.1241, 1997, pp.389–418. http://outoften.doc.ic.ac.uk/projects/slurp/papers.html\#ecoop 4. Sophia Drossopoulou, Susan Eisenbach and Sarfraz Khurshid: Is the Java Type System Sound? Proceedings of the Fourth International Workshop on Foundations of Object-Oriented Languages, 1997. http://outoften.doc.ic.ac.uk/projects/slurp/papers.html\#tapos 5. Allen Goldberg: A Specification of Java Loading and Bytecode Verification, 1997. http://www.kestrel.edu/˜goldberg/ 6. James Gosling, Bill Joy and Guy Steele: The Java TM Language Specification, Addison-Weslay, 1996. 7. Kimera: http://kimera.cs.washington.edu/ 8. Tim Lindholm and Frank Yellin: The Java TM Virtual Machine Specification, Addison-Weslay, 1997. 9. Gary McGraw and Edward W. Felten: Java Security: Hostile Applets, Holes and Antidotes, John Wiley and Sons, 1996. 10. George C. Necula: Proof-Carrying Code, the Proceedings of the 24th Annual SIGPLAN-SIGACT Symposium on Principles of Programming Languages, 1997, pp.106–117. 11. George C. Necula, Peter Lee: The Design and Implementation of a Certifying Compiler, submitted to PLDI’98. 12. Tobias Nipkow and David von Oheimb: Javalight is Type-Safe — Definitely, Proceedings of the 25th Annual SIGPLAN-SIGACT Symposium on Principles of Programming Languages, 1998, pp.161–170. 13. Zhenyu Qian: A Formal Specification of JavaTM Virtual Machine Instructions, 1997. http://www.informatik.uni-bremen.de/˜qian/abs-fsjvm.html 14. Vijay Saraswat: Java is not type-safe, 1997. http://www.research.att.com/˜vj/bug.html 15. Secure Internet Programming: http://www.cs.princeton/edu/sip/ 16. Raymie Stata and Mart´ın Abadi: A Type System for Java Bytecode Subroutines, Proceedings of the 25th Annual SIGPLAN-SIGACT Symposium on Principles of Programming Languages, 1998, pp.149–160. 17. Don Syme: Proving Java Type Soundness, 1997. http://www.cl.cam.ac.uk/users/drs1004/java.ps

Enabling Sparse Constant Propagation of Array Elements via Array SSA Form Vivek Sarkar1 and Kathleen Knobe2 1

IBM Thomas J. Watson Research Center P.O. Box 704, Yorktown Heights, NY 10598, USA [email protected] 2 Compaq Cambridge Research Laboratory One Kendall Square, Building 700, Cambridge, MA 02139, USA [email protected]

Abstract. We present a new static analysis technique based on Array SSA form [6]. Compared to traditional SSA form, the key enhancement in Array SSA form is that it deals with arrays at the element level instead of as monolithic objects. In addition, Array SSA form improves the φ function used for merging scalar or array variables in traditional SSA form. The computation of a φ function in traditional SSA form depends on the program’s control flow in addition to the arguments of the φ function. Our improved φ function (referred to as a Φ function) includes the relevant control flow information explicitly as arguments through auxiliary variables that are called @ variables. The @ variables and Φ functions were originally introduced as run-time computations in Array SSA form. In this paper, we use the elementlevel Φ functions in Array SSA form for enhanced static analysis. We use Array SSA form to extend past algorithms for Sparse Constant propagation (SC) and Sparse Conditional Constant propagation (SCC) by enabling constant propagation through array elements. In addition, our formulation of array constant propagation as a set of data flow equations enables integration with other analysis algorithms that are based on data flow equations. Keywords: static single assignment (SSA) form, constant propagation, conditional constant propagation, Array SSA form, unreachable code elimination.

G. Levi (Ed.): SAS’98, LNCS 1503, pp. 33–56, 1998. c Springer-Verlag Berlin Heidelberg 1998

34

1

V. Sarkar and K. Knobe

Introduction

The problems of constant propagation and conditional constant propagation (a combination of constant propagation and unreachable code elimination) have been studied for several years. However, past algorithms limited their attention to constant propagation of scalar variables only. In this paper, we introduce efficient new algorithms that perform constant propagation and conditional constant propagation through both scalar and array references. One motivation for constant propagation of array variables is in optimization of scientific programs in which certain array elements can be identified as constant. For example, the SPEC95fp [3] benchmark 107.mgrid contains an array variable A that is initialized to four constant-valued elements as shown in Fig. 1. A significant amount of the time in this application is spent in the triply nested loop shown at the bottom of Fig. 1. Since constant propagation can determine that A(2) equals zero in the loop, an effective optimization is to eliminate the entire multiplicand of A(2) in the loop nest. Doing so eliminates 11 of the 23 floating-point additions in the loop nest thus leading to a significant speedup. Another motivation is in analysis and optimization of field accesses of structure variables or objects in object-oriented languages such as Java and C++. A structure can be viewed as a fixed-size array, and a read/write operation of a structure field can be viewed as a read/write operation of an array element through a subscript that is a compile-time constant. This approach is more compact than an approach in which each field of a structure is modeled as a separate scalar variable. This technique for modeling structures as arrays directly extends to nested arrays and structures. For example, an array of rank n of some structure type can be modeled as an array of rank n + 1. Therefore, the constant propagation algorithms presented in this paper can be efficiently applied to structure variables and to arrays of structures. Extending these algorithms to analyze programs containing pointer aliasing is a subject for future research, however. The best known algorithms for sparse constant propagation of scalar variables [8,2] are based on static single assignment (SSA) form [4]. However, traditional SSA form views arrays as monolithic objects, which is an inadequate view for analyzing and optimizing programs that contain reads and writes of individual array elements. In past work, we introduced Array SSA form [6] to address this deficiency. The primary application of Array SSA form in [6] was to enable parallelization of loops not previously parallelizable by making Array SSA form manifest at run-time. In this paper, we use Array SSA form as a basis for static analysis, which means that the Array SSA form structures can be removed after the program properties of interest have been discovered. Array SSA form has two distinct advantages over traditional SSA form. First, the φ operator in traditional SSA form is not a pure function and returns different values for the same arguments depending on the control flow path that was taken. In contrast, the corresponding Φ operator in Array SSA form includes @ variables as extra arguments to capture the control information required i.e., x3 := φ(x2 , x1 ) in traditional SSA form becomes x3 := Φ(x2 , @x2 , x1 , @x1 ) in

Enabling Sparse Constant Propagation of Array Elements

35

! Initialization of array A REAL*8 A(0:3) . . . A(0) = -8.0D0/3.0D0 A(1) = 0.0D0 A(2) = 1.0D0/6.0D0 A(3) = 1.0D0/12.0D0 . . . ! Computation loop in subroutine RESID() do i3 = 2, n-1 do i2 = 2, n-1 do i1 = 2, n-1 R(i1,i2,i3)=V(i1,i2,i3) -A(0)*( U(i1, i2, i3 ) ) -A(1)*( U(i1-1,i2, i3 ) + U(i1+1,i2, i3 ) + U(i1, i2-1,i3 ) + U(i1, i2+1,i3 ) + U(i1, i2, i3-1) + U(i1, i2, i3+1) ) -A(2)*( U(i1-1,i2-1,i3 ) + U(i1+1,i2-1,i3 ) + U(i1-1,i2+1,i3 ) + U(i1+1,i2+1,i3 ) + U(i1, i2-1,i3-1) + U(i1, i2+1,i3-1) + U(i1, i2-1,i3+1) + U(i1, i2+1,i3+1) + U(i1-1,i2, i3-1) + U(i1-1,i2, i3+1) + U(i1+1,i2, i3-1) + U(i1+1,i2, i3+1) ) -A(3)*( U(i1-1,i2-1,i3-1) + U(i1+1,i2-1,i3-1) + U(i1-1,i2+1,i3-1) + U(i1+1,i2+1,i3-1) + U(i1-1,i2-1,i3+1) + U(i1+1,i2-1,i3+1) + U(i1-1,i2+1,i3+1) + U(i1+1,i2+1,i3+1) ) end do end do end do Fig. 1. Code fragments from the SPEC95fp 107.mgrid benchmark

Array SSA form. Second, Array SSA form operates on arrays at the element level rather than as monolithic objects. In particular, a Φ operator in Array SSA form, A3 := Φ(A2 , @A2 , A1 , @A1 ), represents an element-level merge of A2 and A1 . Both advantages of Array SSA form are significant for static analysis. The fact that Array SSA form operates at the element-level facilitates transfer of statically derived information across references to array elements. The fact that the Φ is a known pure function facilitates optimization and simplification of the Φ operations. For convenience, we assume that all array operations in the input program are expressed as reads and writes of individual array elements. The extension to more complex array operations (e.g., as in Fortran 90 array language) is

36

V. Sarkar and K. Knobe

straightforward, and omitted for the sake of brevity. Also, for simplicity, we will restrict constant propagation of array variables to cases in which both the subscript and the value of an array definition are constant e.g., our algorithm mights recognize that a definition A[k] := i is really A[2] := 99 and propagate this constant into a use of A[2]. The algorithms presented in this paper will not consider a definition to be constant if its subscript has a non-constant value e.g., A[m] := 99 where m is not a constant. Performing constant propagation for such references (e.g., propagating 99 into a use of A[m] when legal to do so) is a subject of future work. The rest of the paper is organized as follows. Section 2 reviews the Array SSA form introduced in [6]. Section 3 presents our extension to the Sparse Constant propagation (SC) algorithm from [8] that enables constant propagation through array elements. It describes how lattice values can be computed for array variables and Φ functions. Section 4 presents our extension to the Sparse Conditional Constant propagation (SCC) algorithm from [8] that enables constant propagation through array elements in conjunction with unreachable code elimination. Section 5 discusses related work, and Sect. 6 contains our conclusions.

2

Array SSA Form

In this section, we describe the Array SSA form introduced in [6]. The goal of Array SSA form is to provide the same benefits for arrays that traditional SSA provides for scalars but, as we will see, it has advantages over traditional SSA form for scalars as well. We first describe its use for scalar variables and then its use for array variables. The salient properties of traditional SSA form are as follows: 1. Each definition is assigned a unique name. 2. At certain points in the program, new names are generated which combine the results from several definitions. This combining is performed by a φ function which determines which of several values to use, based on the flow path traversed. 3. Each use refers to exactly one name generated from either of the two rules above. For example, traditional SSA form converts the code in Fig. 2 to that in Fig. 3. The S3 := φ(S1 , S2 ) statement defines S3 as a new name that represents the merge of definitions S1 and S2 . It is important to note that the φ function in traditional SSA form is not a pure function of S1 and S2 because its value depends on the path taken through the if statement. Notice that this path is unknown until runtime and may vary with each dynamic execution of this code. In contrast to traditional SSA form, the semantics of a Φ function is defined to be a pure function in our Array SSA form. This is accomplished by introducing @ variables (pronounced “at variables”), and by rewriting a φ function in traditional SSA form such as φ(S1 , S2 ) as a new kind of Φ function,

Enabling Sparse Constant Propagation of Array Elements

37

Φ(S1 , @S1 , S2 , @S2 ). For each static definition Sk , its @ variable @Sk identifies the most recent “time” at which Sk was modified by this definition. For an acyclic control flow graph, a static definition Sk may execute either zero times or one time. These two cases can be simply encoded as @Sk = false and @Sk = true to indicate whether or not definition Sk was executed. For a control flow graph with cycles (loops), a static definition Sk may execute an arbitrary number of times. In general, we need more detailed information for the @Sk = true case so as to distinguish among different dynamic execution instances of static definition Sk . Therefore, @Sk is set to contain the dynamic iteration vector at which the static definition Sk was last executed. The iteration vector of a static definition Sk identifies a single iteration in the iteration space of the set of loops that enclose the definition. Let n be the number of loops that enclose a given definition. For convenience, we treat the outermost region of acyclic control flow in a procedure as a dummy outermost loop with a single iteration. Therefore n ≥ 1 for each definition. A single point in the iteration space is specified by the iteration vector i = (i1 , . . . , in ), which is an n-tuple of iteration numbers one for each enclosing loop. We do not require that the surrounding loops be structured counted loops (i.e., like Fortran do loops) or that the surrounding loops be tightly nested. Our only assumption is that all loops are single-entry, or equivalently, that the control flow graph is reducible [5, 1]. For single-entry loops, we know that each def executes at most once in a given iteration of its surrounding loops. All structured loops (e.g., do, while, repeat-until) are single-entry even when they contain multiple exits; also, most unstructured loops (built out of goto statements) found in real programs are single-entry as well. A multiple-entry loop can be transformed into multiple single-entry loops by node splitting [5,1]. Array SSA form can be used either at run-time as discussed in [6] or for static analysis, as in the constant propagation algorithms presented in this paper. In this section, we explain the meaning of @ variables as if they are computed at run-time. We assume that all @ variables, @Sk , are initialized to the empty vector, @Sk := ( ), at the start of program execution. For each real (non-Φ) definition, Sk , we assume that a statement of the form @Sk := i is inserted immediately after definition Sk 1 , where i is the current iteration vector for all loops that surround Sk . Each Φ definition also has an associated @ variable. Its semantics will be defined shortly. All @ variables are initialized to the empty vector because the empty vector is the identity element for a lexicographic max operation i.e., max(( ), i) = i, for any @ variable value i. As a simple example, Fig. 4 shows the Array SSA form for the program in Fig. 2. Note that @ variables @S1 and @S2 are explicit arguments of the Φ function. In this example of acyclic code, there are only two possible values 1

It may appear that the @ variables do not satisfy the static single assignment property because each @Sk variable has two static definitions, one in the initialization and one at the real definition of Sk . However, the initialization def is executed only once at the start of program execution and can be treated as a special-case initial value rather than as a separate definition.

38

V. Sarkar and K. Knobe

for each @ variable — the empty vector ( ) and the unit vector (1) — which correspond to false and true respectively. Figure 5 shows an example for-loop and its conversion to Array SSA form. Because of the presence of a loop, the set of possible values for an @ variable becomes unbounded e.g., we may have @S1 = (100) on exit from the loop. However, @S1 and @S2 are still explicit arguments of the Φ function, and their iteration vector values are necessary for evaluating the Φ function at run-time. The semantics of a Φ function can now be specified by a conditional expression that is a pure function of the arguments of the Φ. For example, the semantics of the Φ function, S3 := Φ(S2 , @S2 , S1 , @S1 ) in Fig. 5, can be expressed as a conditional expression as follows (where denotes a lexicographic greater-thanor-equal comparison of iteration vectors): if @S2 @S1 then S2 S3 = else S1 end if Following each Φ def, S3 := Φ(S2 , @S2 , S1 , @S1 ), there is the definition of the associated @ variable, @S3 = max(@S2 , @S1 ), where max represents a lexicographic maximum operation of iteration vector values @S2 , @S1 . Consider, for example, a condition C in Fig. 5 that checks if the value of i is even. In this case, definition S1 is executed in every iteration and definition S2 is executed only in iterations 2, 4, 6, . . . . For this “even-value” branch condition, the final values of @S1 and @S2 are both equal to (100) if m = 100. Since these values satisfy the condition @S2 @S1 , the conditional expression will yield S 3 = S2 . Consider another execution of the for-loop in Fig. 5 in which condition C evaluates to false in each iteration of the for loop. For this execution, the final values of @S2 and @S1 will be the empty vector ( ) and (100) respectively. Therefore, S2 ≺ S1 , and the conditional expression for the Φ function will yield S3 = S1 for this execution. The above description outlines how @ variables and Φ functions can be computed at run-time. However, if Array SSA form is used for static analysis, then no run-time overhead is incurred due to the @ variables and Φ functions. Instead, the @ variables and Φ functions are inserted in the compiler intermediate representation prior to analysis, and then removed after the program properties of interest have been discovered by static analysis. We now describe Array SSA form for array variables. Figure 6 shows an example program with an array variable, and the conversion of the program to Array SSA form as defined in [6]. The key differences between Array SSA form for array variables and Array SSA form for scalar variables are as follows: 1. Array-valued @ variables: The @ variable is an array of the same shape as the array variable with which it is associated, and each element of an @ array is initialized to the empty vector. For example, the statement @A1 [k1 ] := (1) is inserted after

Enabling Sparse Constant Propagation of Array Elements

if (C) then S := . . . else S := . . . end if

Fig. 2. Control Flow with Scalar Definitions

if (C) then S1 := . . . else S2 := . . . end if S3 := φ(S1 , S2 )

Fig. 3. Traditional SSA form

@S1 := ( ) @S2 := ( ) if (C) then S1 := . . . @S1 := (1) else S2 := . . . @S2 := (1) end if S3 = Φ(S1 , @S1 , S2 , @S2 )

Fig. 4. After conversion of Fig. 2 to Array SSA form

39

40

V. Sarkar and K. Knobe

Example for-loop: S := . . . for i := 1 to m do S := . . . if (C) then S := . . . end if end for

After conversion to Array SSA form: @S1 := ( ) ; @S2 := ( ) S := . . . @S := (1) for i := 1 to m do S0 := Φ(S3 , @S3 , S, @S) @S0 := max(@S3 , @S) S1 := . . . @S1 := (i) if (C) then S2 := . . . @S2 := (i) end if S3 := Φ(S2 , @S2 , S1 , @S1 ) @S3 := max(@S2 , @S2 ) end for

Fig. 5. A for-loop and its conversion to Array SSA form

Enabling Sparse Constant Propagation of Array Elements

41

Example program with array variables: n1:

n2:

n3:

A[∗] := initial value of A i := 1 C := i < n if C then k := 2 ∗ i A[k] := i print A[k] endif print A[2]

After conversion to Array SSA form: n1:

n2:

n4:

@i := ( ) ; @C := ( ) ; @k := ( ) ; @A0 [∗] := ( ) ; @A1 [∗] := ( ) A0 [∗] := initial value of A @A0 [∗] := (1) i := 1 @i := (1) C := i < n @C := (1) if C then k := 2 ∗ i @k := (1) A1 [k] := i @A1 [k] := (1) A2 := dΦ(A1 , @A1 , A0 , @A0 ) @A2 := max(@A1 , @A0 ) print A2 [k] endif A3 := Φ(A2 , @A2 , A0 , @A0 ) @A3 := max(@A2 , @A0 ) print A3 [2]

Fig. 6. Example program with an array variable, and its conversion to Array SSA form

42

V. Sarkar and K. Knobe

statement A1 [k1 ] := i in Fig. 6. In general, @A1 can record a separate iteration vector for each element that is assigned by definition A1 . This initialization is only required for @ arrays corresponding to real (non-Φ) definitions. No initialization is required for an @ array for a Φ definition (such as @A2 and @A3 in Fig. 6) because its value is completely determined by other @ arrays. 2. Array-valued Φ functions: A Φ function for array variables returns an array value. For example, consider the Φ definition A3 := Φ(A2 , @A2 , A0 , @A0 ) in Fig. 6 which represents a merge of arrays A2 and A0 . The semantics of the Φ function is specified by the following conditional expression for each element, A3 [j]: if @A2 [j] @A0 [j] then A2 [j] A3 [j] = else A0 [j] end if Note that this conditional expression uses a lexicographic comparison () of @ values just as in the scalar case. 3. Definition Φ’s: The traditional placement of the Φ is at control merge points. We refer to this as a control Φ. A special new kind of Φ function is inserted immediately after each original program definition of an array variable that does not completely kill the array value. This definition Φ merges the value of the element modified in the definition with the values available immediately prior to the definition. Definition Φ’s did not need to be inserted for definitions of scalar variables because a scalar definition completely kills the old value of the variable. We will use the notation dΦ when we want to distinguish a definition Φ function from a control Φ function. For example, consider definition A1 in Fig. 6. The dΦ function, A2 := dΦ(A1 , @A1 , A0 , @A0 ) is inserted immediately after the def of A1 to represent an element-by-element merge of A1 and A0 . Any subsequent use of the original program variable A (before an intervening def) will now refer to A2 instead of A1 . The semantics of the dΦ function is specified by the following conditional expression for each element, A2 [j]: if @A1 [j] @A0 [j] then A1 [j] A2 [j] = else A0 [j] end if Note that this conditional expression is identical in structure to the conditional expression for the control Φ function in item 2 above.

3

Sparse Constant Propagation for Scalars and Array Elements

We now present our extension to the Sparse Constant propagation (SC) algorithm from [8] that enables constant propagation through array elements.

Enabling Sparse Constant Propagation of Array Elements

43

Section 3.1 contains our definitions of lattice elements for scalar and array variables; the main extension in this section is our modeling of lattice elements for array variables. Section 3.2 outlines our sparse constant propagation algorithm; the main extension in this section is the use of the definition Φ operator in Array SSA form to perform constant propagation through array elements. 3.1

Lattice Values for Scalar and Array Variables

Recall that a lattice consists of: – L, a set of lattice elements. A lattice element for a program variable v is written as L(v), and denotes set(L(v)) = a set of possible values for variable v. – > (“top”) and ⊥ (“bottom”), two distinguished elements of L. – A meet (or join) operator, u, such that for any lattice element e, e u > = e and e u ⊥ = ⊥. – A w operator such that e w f if and only if e u f = f , and a A operator such that e A f if and only if e w f and e 6= f . The height H of lattice L is the length of the largest sequence of lattice elements e1 , e2 , . . . , eH such that ei A ei+1 for all 1 ≤ i < H. We use the same approach as in [8] for modeling lattice elements for scalar variables. Given a scalar variable S, the value of L(S) in our framework can be >, Constant or ⊥ . When the value is Constant we also maintain the value of the constant. The sets denoted by these lattice elements are set(>) = { }, set(Constant) = {Constant}, and set(⊥) = U S , where U S is the universal set of values for variable S. We now describe how lattice elements for array variables are represented in A our framework. Let U A ind and U elem be the universal set of index values and the universal set of array element values respectively for an array variable A in Array SSA form. For an array variable, the set denoted by lattice element L(A) A is a subset of U A ind × U elem i.e., a set of index-element pairs. Since we restrict constant propagation of array variables to cases in which both the subscript and the value of an array definition are constant, there are only three kinds of lattice elements of interest: 1. L(A) = > ⇒ set(L(A)) = { } This “top” case means that the possible values of A have yet to be determined i.e., the set of possible index-element pairs that have been identified thus far for A is the empty set, { }. 2. L(A) = h(i1 , e1 ), (i2 , e2 ), . . . i A ⇒ set(L(A)) = {(i1 , e1 ), (i2 , e2 ), . . . } ∪ (U A ind − {i1 , i2 , . . . }) × U elem In general, the lattice value for this “constant” case is represented by a finite ordered list of index-element pairs, h(i1 , e1 ), (i2 , e2 ), . . . i where i1 , e1 , i2 , e2 , . . . are all constant. The list is sorted in ascending order of the index values, i1 , i2 , . . . , and all the index values assumed to be distinct.

44

V. Sarkar and K. Knobe

. . .

< (1,100), (2,101) >

. . .

< (1,100) >

. . .

. . .

< (2,101) >

< (2,101), (3,102) >

. . .

< (3,102) >

. . .

. . .

Fig. 7. Lattice elements of array values with maximum list size Z = 2

The meaning of this “constant” lattice value is that the current stage of analysis has determined some finite number of constant index-element pairs for array variable A, such that A[i1 ] = e1 , A[i2 ] = e2 , . . . . All other elements of A are assumed to be non-constant. These properties are captured by set(L(A)) defined above as the set denoted by lattice value L(A). For the sake of efficiency, we will restrict these constant lattice values to ordered lists that are bounded in size by a small constant, Z ≥ 1 e.g., if Z = 5 then all constant lattice values will have ≤ 5 index-element pairs. Doing so ensures that the height of the lattice for array values is at most (Z + 2). If any data flow equation yields a lattice value with P > Z pairs, then this size constraint is obeyed by conservatively dropping any (P − Z) index-element pairs from the ordered list. Note that the lattice value for a real (non-Φ) definition, will contain at most one index-element pair, since we assumed that an array assignment only modifies a single element. Ordered lists with size > 1 can only appear as the output of Φ functions. A 3. L(A) = ⊥ ⇒ set(L(A)) = U A ind × U elem This “bottom” case means that, according to the approximation in the current stage of analysis, array A may take on any value from the universal set of index-element pairs. Note that L(A) = ⊥ is equivalent to an empty ordered list, L(A) = h i.

The lattice ordering (A) for these elements is determined by the subset relationship among the sets that they denote. The lattice structure for the Z = 2 case is shown in Fig. 7. This lattice has four levels. The second level (just below >) contains all possible ordered lists that contain exactly two constant index-element pairs. The third level (just above ⊥) contains all possible ordered lists that contain a single constant index-element pair. The lattice ordering is determined by the subset relationship among the sets denoted by lattice elements. For example, consider two lattice elements L1 = h(1, 100), (2, 101)i

Enabling Sparse Constant Propagation of Array Elements

45

and L2 = h(2, 101)i. The sets denoted by these lattice elements are: set(L1 ) = {(1, 100), (2, 101)} ∪ (U ind − {1, 2}) × U elem set(L2 ) = {(2, 101)} ∪ (U ind − {2}) × U elem Therefore, set(L1 ) is a proper subset of set(L2 ) and we have L1 A L2 i.e., L1 is above L2 in the lattice in Fig. 7. Finally, the meet operator (u) for two lattice elements, L1 and L2 , for array variables is defined in Fig. 8 where L1 ∩ L2 denotes an intersection of ordered lists L1 and L2 . L3 L1 L1 L1

= L1 u L2 L2 = > L2 = h(i1 , e1 ), . . . i L2 = ⊥ => > L2 ⊥ = h(i01 , e01 ), . . . i L1 L1 ∩ L 2 ⊥ =⊥ ⊥ ⊥ ⊥

Fig. 8. Lattice computation for the meet operator, L3 = L1 u L2

3.2

The Algorithm

Recall that the @ variables defined in Sect. 2 were necessary for defining the full execution semantics of Array SSA form. For example, the semantics of a Φ operator, A2 := Φ(A1 , @A1 , A0 , @A0 ), is defined by the following conditional expression: if @A1 [j] @A0 [j] then A1 [j] A2 [j] = else A0 [j] end if The sparse constant propagation algorithm presented in this section is a static analysis that is based on conservative assumptions about runtime behavior. Let us first consider the case when the above Φ operator is a control Φ. Since algorithm in this section does not perform conditional constant propagation, the lattice computation of a control Φ can be simply defined as L(A2 ) = L(Φ(A1 , @A1 , A0 , @A0 ) = L(A1 ) u L(A0 ) i.e., as a join of the lattice values L(A1 ) and L(A0 ). Therefore, the lattice computation for A2 does not depend on @ variables @A1 and @A0 for a control Φ operator. Now, consider the case when the above Φ operator is a definition Φ. The lattice computation for a definition Φ is shown in Fig. 9. Since A1 corresponds to a definition of a single array element, the ordered list for L(A1 ) can contain at most one pair. The insert operation in Fig. 9 is assumed to return a new ordered list obtained by inserting (i0 , e0 ) into h(i1 , e1 ), . . . i with the following adjustments if needed:

46

V. Sarkar and K. Knobe

– If there exists an index-element pair (ij , ej ) in h(i1 , e1 ), . . . i such that i0 = ij , then the insert operation just replaces (ij , ej ) by (i0 , e0 ). – If the insert operation causes the size of the list to exceed the threshold size Z, then one of the pairs is dropped from the output list so as to satisfy the size constraint. Interestingly, we again do not need @ variables for the lattice computation in Fig. 9. This is because the ordered list representation for array lattice values already contains all the subscript information of interest, and overlaps with the information that would have been provided by @ variables. L(A2 ) L(A0 ) = > L(A0 ) = h(i1 , e1 ), . . . i L(A0 ) = ⊥ L(A1 ) = > > > > L(A1 ) = h(i0 , e0 )i > insert((i0 , e0 ), h(i1 , e1 ), . . . i) h(i0 , e0 )i L(A1 ) = ⊥ ⊥ ⊥ ⊥ Fig. 9. Lattice computation for L(A2 ) = LdΦ (L(A1 ), L(A0 ))

Therefore, we do not need to analyze @ variables for the sparse constant propagation algorithm described in this section, because of our ordered list representation of lattice values for array variables. Instead, we can use a partial Array SSA form which is simply Array SSA form with all definitions and uses of @ variables removed, and with φ operators instead of Φ operators. If only constant propagation is being performed, then it would be more efficient to only build the partial Array SSA form. However, if other optimizations are being performed that use Array SSA form, then we can build full Array SSA form and simply ignore the @ variables for this particular analysis. Our running example is shown in Fig. 10. The partial Array SSA form for this example is shown in Fig. 11. The partial Array SSA form does not contain any @ variables since @ variables are not necessary for the level of analysis performed by the constant propagation algorithms in this paper. The data flow equations for this example are shown in Fig. 12. Each assignment in the Array SSA form results in one data flow equation. The numbering S1 through S8 indicates the correspondence. The argument to these equations are simply the current lattice values of the variables. The lattice operations are specific to the operations within the statement. Figures 13, 9, and 14 show the lattice computations for an assignment to an array element (as in S3 and S5), definition φ (as in S4 and S6), a reference to an array element (as in the RHS of S3 and S5). The lattice computation for a φ assignment (as in S7) A3 = φ(A2 , A1 ) is L(A3 ) = Lφ (L(A2 ), L(A1 )) = L(A1 ) u L(A2 ) where u is shown in Fig. 8. Notice that we also include lattice computation for specific arithmetic computations such as the multiply in S3 and S5. This allows for constant computation as well as constant propagation. Tables for these arithmetic computations are straightforward and are not shown.

Enabling Sparse Constant Propagation of Array Elements

47

The simple data flow algorithm is shown in Fig. 15. Our example includes the propagation of values from a node to two successors (Y) and to a node from two predecessors (D). Equations S1 and S2 are evaluated because they are associated with the entry block. Element 3 of Y2 is known to have the value 99 at this stage. As a result of the modification of Y2 both S3 and S5 are inserted into the worklist (they reference the lattice value of Y2 ). S3 uses the propagated constant 99 to compute and propagate the constant 198 to element 1 of D1 and then, after evaluation of S4, to element 1 of D2 . Any subsequent references to D in the then block of the source become references D2 in the Array SSA form and are known to have a constant value at element 1. Depending on the order of computations via the worklist, we may then compute either D3 and then D4 or, because D2 has been modified, we may compute D5 . If we compute D5 at this point, it appears to have a constant value. Subsequent evaluations D3 and D4 cause D5 to be reevaluated and lowered from constant value to ⊥ because the value along one path is not constant. Notice that in this case, the reevaluation of D5 could have been avoided by choosing an optimal ordering of processing. Processing of programs with cyclic control flow is no more complex but may involve recomputation that can not be removed by optimal reordering. In particular, the loop entry is a control flow merge point since control may enter from the top or come from the loop body. It will contain a φ which combines the value entering from the top with that returning after the loop. The lattice values for such a node may require multiple evaluations. Also notice that in this example, if I in S5 is known to have the value 3, it will be recoded as a constant element. Upon evaluation of S7, the intersection of the sets associated with D2 and D4 will not be empty and element 1 of D5 will be recorded as a constant.

Y [3] := 99 if C then D[1] := Y [3] ∗ 2 else D[1] := Y [I] ∗ 2 endif Z := D[1]

Fig. 10. Sparse Constant Propagation Example

48

V. Sarkar and K. Knobe

S1: S2: S3: S4: S5: S6: S7: S8:

Y0 and D0 in effect here. ... Y1 [3] := 99 Y2 := φ(Y1 , Y0 ) if C then D1 [1] := Y2 [3] ∗ 2 D2 := φ(D1 , D0 ) else D3 [1] := Y2 [I] ∗ 2 D4 := φ(D3 , D0 ) endif D5 := φ(D2 , D4 ) Z := D5 [1]

Fig. 11. Array SSA form for the Sparse Constant Propagation Example S1: S2: S3: S4: S5: S6: S7: S8:

L(Y1 ) L(Y2 ) L(D1 ) L(D2 ) L(D3 ) L(D4 ) L(D5 ) L(Z)

= = = = = = = =

< (3, 99) > Ldφ (L(Y1 ), L(Y0 )) Ld[ ] (L∗ (L(Y2 [3]), 2)) Ldφ (L(D1 ), L(D0 )) Ld[ ] (L∗ (L(Y2 [I]), 2)) Ldφ (L(D3 ), L(D0 )) Lφ (L(D2 ), L(D4 )) L(D5 [1])

Fig. 12. Data Flow Equations for the Sparse Constant Propagation Example L(A1 ) L(i) = > L(i) = Constant L(i) = ⊥ L(k) = > > > ⊥ L(k) = Constant > h(L(k), L(i))i ⊥ L(k) = ⊥ ⊥ ⊥ ⊥ Fig. 13. Lattice computation for array definition operator, L(A1 ) = Ld[ ] (L(k), L(i)) L(A[k]) L(k) = > L(k) = Constant L(k) = ⊥ L(A) = > > > ⊥ L(A) = h(i1 , e1 ), . . . i > ej , if ∃ (ij , ej ) ∈ L(A) with ij = L(k) ⊥ ⊥, otherwise L(A) = ⊥ ⊥ ⊥ ⊥ Fig. 14. Lattice computation for array reference operator, L(A[k]) = L[ ] (L(A), L(k))

Enabling Sparse Constant Propagation of Array Elements

49

Initialization: L(v) ← > for all local variables, v. insert(Ev , work list) for each equation Ev defining v such that v is assigned to in the entry block. Body: while (work list != empty) Ev ← remove(work list) reevaluate(Ev ) insert(Ev0 0 , work list) for each equation Ev0 0 that uses Ev end while

Fig. 15. Algorithm for Sparse Constant propagation (SC) for array and scalar variables

4 4.1

Sparse Conditional Constant Propagation Lattice Values of Executable Flags for Nodes and Edges

As in the SCC algorithm in [8], our array conditional constant propagation algorithm maintains executable flags associated with each node and each edge in the CFG. Flag Xni indicates whether node ni may be executed, and Xei indicates whether edge ei may be traversed. The lattice value of an execution flag is either no or maybe, corresponding to unreachable code and reachable code respectively. The lattice value for an execution flag is initialized to no, and can be lowered to maybe in the course of the constant propagation algorithm. In practice, control dependence identities can be used to reduce the number of executable flag variables in the data flow equations e.g., a single flag can be used for all CFG nodes that are control equivalent. For the sake of simplicity, we ignore such optimizations in this paper. The executable flag of a node is computed from the executable flags of its incoming edges. The executable flag of an edge is computed from the executable flag of its source node and knowledge of the branch condition variable used to determine the execution path from that node. These executable flag mappings are summarized in Fig. 16 for a node n with two incoming edges, e1 and e2, and two outgoing edges, e3 and e4. The first function table in Fig. 16 defines the join operator u on executable flags such that Xn = Xe1 u Xe2 . We introduce a true operator LT and a false operator LF on lattice values such that Xe3 = LT (Xn , L(C)) and Xe4 = LF (Xn , L(C)). Complete function tables for the LT and LF operators are also shown in Fig. 16. Note that all three function tables are monotonic with respect to their inputs. Other cases for mapping a node Xn value to the Xe values of its outgoing edges can be defined similarly. If n has exactly one outgoing edge e, then Xe =

50

V. Sarkar and K. Knobe

Definition of join operator, Xn = Xe1 u Xe2 : Xn Xe2 = no Xe2 = maybe Xe1 = no no maybe Xe1 = maybe maybe maybe

X e1

X n

X e2

C?

TRUE

FALSE

X e3

X e4

Definition of true operator for branch condition C, Xe3 = LT (Xn , L(C)): Xe3 L(C) = > L(C) = true L(C) = false L(C) = ⊥ Xn = no no no no no Xn = maybe no maybe no maybe

Definition of false operator for branch condition C, Xe4 = LF (Xn , L(C)): Xe4 L(C) = > L(C) = true L(C) = false L(C) = ⊥ Xn = no no no no no Xn = maybe no no maybe maybe Fig. 16. Executable flag mappings for join operator (u), true operator (LT ), and false operator (LF )

L(k3 ) Xe2 = no Xe2 = maybe Xe1 = no > L(k2 ) X e1 = maybe L(k1 ) L(k1 ) u L(k2 ) Fig. 17. k3 := Φ(k1 , Xe1 , k2 , Xe2 ), where execution flags Xe1 and Xe2 control the selection of k1 and k2 respectively

Enabling Sparse Constant Propagation of Array Elements

51

Xn . If node n has more than two outgoing edges then the mapping for each edge is similar to the LT and LF operators. Recall that the LΦ function for a control Φ was defined in Sect. 3.2 by the meet function, LΦ (L(A2 ), L(A1 )) = L(A1 ) u L(A2 ). For the extended SCC algorithm described in this section, we use the definition of LΦ shown in Fig. 17. This definition uses executable flags Xe1 and Xe2 , where e1 and e2 are the incoming control flow edges for the Φ function. Thus, the lattice values of the executable flags Xe1 and Xe2 are used as compile-time approximations of @ variables @k1 and @k2 . 4.2

Sparse Conditional Constant Propagation Algorithm

We introduce our algorithm by first explaining how it works for acyclic scalar code. Consider the example program shown in Fig. 18. The basic blocks are labeled n1, n2, n3 and n4. Edges e12, e13, e24 and e34 connect nodes in the obvious way. The control flow following block n1 depends on the value of variable n. The first step is to transform the example in Fig. 18 to partial Array SSA form (with no @ variables) as shown in Fig. 19. Note that since k had multiple assignments in the original program, a φ function is required to compute k3 as a function of k1 and k2 . The second step is to use the partial Array SSA form to create a set of data flow equations on lattice values for use by our constant propagation algorithm. The conversion to equations is performed as follows. There is one equation created for each assignment statement in the program. There is one equation created for each node in the CFG. There is one equation created for each edge in the CFG. The equations for the assignments in our example are shown in Fig. 20. The equations for the nodes and edges in our example are found in Fig. 21 and 22 respectively. The lattice operations L< , L∗ , and Lmax use specific knowledge of their operation as well as the lattice values of their operands to compute resulting lattice operations. For example, L∗ (⊥, L(0)) results in L(0) because the result of multiplying 0 by any number is 0. Next, we employ a work list algorithm, shown in Fig. 23, that reevaluates the equations until there are no further changes. A solution to the data flow equations identifies lattice values for each variable in the Array SSA form of the program, and for each node executable flag and edge executable flag in the CFG. Reevaluation of an equation associated with an assignment may cause equations associated with other assignments to be inserted on the work list. If the value appears in a conditional expression, it may cause one of the equations associated with edges to be inserted on the work list. Reevaluation of an edge’s executable flag may cause an equation for a destination node’s executable flag to be inserted on the work list. If reevaluation of a node’s executable flag indicates that the node may be evaluated, then the equations associated with assignments within that node to be added to the work list. When the algorithm terminates, the lattice values for variables identify the constants in the program, and the lattice values for executable flags of nodes identify unreachable code.

52

V. Sarkar and K. Knobe n1: n2: n3: n4:

i := 1 C := i < n if C then k := 2 ∗ i else k := 2 ∗ n endif print k

n1 e12

e13

n2

n3

e24

e34 n4

Fig. 18. Acyclic Scalar Example

n1: n2: n3: n4:

i := 1 C := i < n if C then k1 := 2 ∗ i else k2 := 2 ∗ n endif k3 := φ(k1 , k2 ) print k3

Fig. 19. Partial Array SSA form for the Acyclic Scalar Example

L(i) = L(1) L(C) = L< (L(i), L(n)) L(k1 ) = L∗ (L(2), L(i)) L(k2 ) = L∗ (L(2),L(n)) L(k3 ) = LΦ (L(k1 ), Xe24 , L(k2 ), Xe34 )

Fig. 20. Equations for Assignments

Xn1 Xn2 Xn3 Xn4

= true = Xe12 = Xe13 = Xe24 u Xe34

Fig. 21. Equations for Nodes

Enabling Sparse Constant Propagation of Array Elements Xe12 Xe13 Xe24 Xe34

53

= LT (Xn1 , L(C)) = LF (Xn1 , L(C)) = Xn2 = Xn3

Fig. 22. Equations for Edges Initialization: L(v) ← > for all local variables, v. Xn ← maybe where Xn is the executable flag for the entry node. Xn ← no where Xn is the executable flag for any node other than the entry node. Xe ← no where Xe is the executable flag for any edge. insert(Ev , work list) for each equation Ev defining v such that v is assigned to in the entry block. Body: while (work list != empty) Ev ← remove(work list) reevaluate(Ev ) insert(Ev0 0 , work list) for each equation Ev0 0 that uses Ev if Ev defines the executable flag for some node n then insert(Ev0 0 , work list) for each equation Ev0 0 defining v 0 such that v 0 is assigned to in block n. end if end while

Fig. 23. Sparse Conditional Constant propagation (SCC) algorithm for scalar and array variables

Even though we assumed an acyclic CFG in the above discussion, the algorithm in Fig. 23 can be used unchanged for performing constant propagation analysis on a CFG that may have cycles. The only difference is that the CFG may now contain back edges. Each back edge will be evaluated when its source node is modified. The evaluation of this back edge may result in the reevaluation of its target node. As in past work, it is easy to show that the algorithm must take at most O(Q) time, where Q is the number of data flow equations, assuming that the maximum arity of a function is constant and the maximum height of the lattice is constant. As an example with array variables, Fig. 24 lists the data flow equations for the assignment statements in the Array SSA program in Fig. 6 (the data flow equations for nodes and edges follow the CFG structure as in Figs. 21 and 22). Given the definition of lattice elements for array variables from Sect. 3.1, the conditional constant propagation algorithm in Fig. 23 can also be used unchanged for array variables.

54

V. Sarkar and K. Knobe

L(A0 ) = ⊥ L(i) = L(1) L(C) = L< (L(i), L(n)) L(k) = L∗ (L(2), L(i)) L(A1 ) = Ld[ ] (L(k), L(i)) L(A2 ) = Ldφ (L(A1 ), L(A0 )) L(A3 ) = LΦ (L(A2 ), Xe24 , L(A0 ), Xe14 ) Fig. 24. Equations for Assignments from Fig. 6

5

Related Work

Static single assignment (SSA) form for scalar variables has been a significant advance. It has simplified the design of some optimizations and has made other optimizations more effective. The popularity of SSA form surged after an efficient algorithm for computing SSA form was made available [4]. SSA form is now a standard representation used in modern optimizing compilers in both industry and academia. However, it has been widely recognized that SSA form is much less effective for array variables than for scalar variables. The approach recommended in [4] is to treat an entire array like a single scalar variable in SSA form. The most serious limitation of this approach is that it lacks precise data flow information on a per-element basis. Array SSA form addresses this limitation by providing Φ functions that can combine array values on a per-element basis. The constant propagation algorithm described in this paper can propagate lattice values through Φ functions in Array SSA form, just like any other operation/function in the input program. The problem of conditional constant propagation for scalar variables has been studied for several years. Wegbreit [7] provided a general algorithm for solving data flow equations; his algorithm can be used to perform conditional constant propagation and more general combinations of program analyses. However, his algorithm was too slow to be practical for use on large programs. Wegman and Zadeck [8] introduced a Sparse Conditional Constant (SCC) propagation algorithm that is as precise as the conditional constant propagation obtained by Wegbreit’s algorithm, but runs faster than Wegbreit’s algorithm by a speedup factor that is at least O(V ), where V is the number of variables in the program. The improved efficiency of the SCC algorithm made it practical to perform conditional constant propagation on large programs, even in the context of industry-strength product compilers. The main limitation of the SCC algorithm is a conceptual one — the algorithm operates on two “worklists” (one containing edges in the SSA graph and another containing edges from the control flow

Enabling Sparse Constant Propagation of Array Elements

55

graph) rather than on data flow equations. The lack of data flow equations makes it hard to combine the algorithm in [8] with other program analyses. The problem of combining different program analyses based on scalar SSA form has been addressed by Click and Cooper in [2], where they present a framework for combining constant propagation, unreachable-code elimination, and value numbering that explicitly uses data flow equations. Of the conditional constant propagation algorithms mentioned above, our work is most closely related to that of [2] with two significant differences. First, our algorithm performs conditional constant propagation through both scalar and array references, while the algorithm in [2] is limited only to scalar variables. Second, the framework in [2] uses control flow predicates instead of execution flags. It wasn’t clear from the description in [2] how their framework deals with predicates that are logical combinations of multiple branch conditions; it appears that they must either allow the possibility of an arbitrary size predicate expression appearing in a data flow equation (which would increase the worst-case execution time complexity of their algorithm) or they must sacrifice precision by working with an approximation of the predicate expression.

6

Conclusions

We have presented a new sparse conditional constant propagation algorithm for scalar and array references based on Array SSA form [6]. Array SSA form has two advantages: It is designed to support analysis of arrays at the element level and it employs a new Φ function that is a pure function of its operands, and can be manipulated by the compiler just like any other operator in the input program. The original sparse conditional constant propagation algorithm in [8] dealt with control flow and data flow separately by maintaining two distinct work lists. Our algorithm uses a single set of data flow equations and is therefore conceptually simpler. In addition to being simpler, the algorithm presented in this paper is more powerful than its predecessors in that it handles constant propagation through array elements. It is also more effective because its use of data flow equations allows it to be totally integrated with other data flow algorithms, thus making it easier to combine other analyses with conditional constant propagation.

References 1. A.V. Aho, R. Sethi, and J.D. Ullman. Compilers: Principles, Techniques, and Tools. Addison-Wesley, 1986. 2. Cliff Click and Keith D. Cooper. Combining Analyses, Combining Optimizations. ACM Transactions on Programming Languages and Systems, 17(2):181–196, March 1995. 3. The Standard Performance Evaluation Corporation. SPEC CPU95 Benchmarks. http://open.specbench.org/osg/cpu95/, 1997.

56

V. Sarkar and K. Knobe

4. Ron Cytron, Jeanne Ferrante, Barry K. Rosen, Mark N. Wegman, and F. Kenneth Zadeck. Efficiently Computing Static Single Assignment Form and the Control Dependence Graph. ACM Transactions on Programming Languages and Systems, 13(4):451–490, October 1991. 5. Matthew S. Hecht. Flow Analysis of Computer Programs. Elsevier North-Holland, Inc., 1977. 6. Kathleen Knobe and Vivek Sarkar. Array SSA form and its use in Parallelization. Conf. Rec. Twenty-fifth ACM Symposium on Principles of Programming Languages, San Diego, California, January 1998. 7. B. Wegbreit. Property Extraction in Well-Founded Property Sets. IEEE Transactions on Software Engineering, 1:270–285, 1975. 8. Mark N. Wegman and F. Kenneth Zadeck. Constant Propagation with Conditional Branches. ACM Transactions on Programming Languages and Systems, 13(2):181– 210, April 1991.

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses Michael Hind and Anthony Pioli State University of New York at New Paltz IBM Thomas J. Watson Research Center {hind,pioli}@mcs.newpaltz.edu

Abstract. This paper describes an empirical comparison of four contextinsensitive pointer alias analysis algorithms that use varying degrees of ﬂow-sensitivity: a ﬂow-insensitive algorithm that tracks variables whose addresses were taken and stored; a ﬂow-insensitive algorithm that computes a solution for each function; a variant of this algorithm that uses precomputed kill information; and a ﬂow-sensitive algorithm. In addition to contrasting the precision and eﬃciency of these analyses, we describe implementation techniques and quantify their analysis-time speed-up.

1

Introduction

To eﬀectively analyze programs written in languages that make extensive use of pointers, such as C, C++, or Java (in the form of references), knowledge of pointer behavior is required. Without such knowledge, conservative assumptions about pointer values must be made, resulting in less precise data ﬂow information, which can adversely aﬀect the eﬀectiveness of analyses and tools that depend on this information. A pointer alias analysis is a compile-time analysis that, for each program point, attempts to determine what a pointer can point to. As such an analysis is, in general, undecidable [25, 34], approximation methods have been developed. These algorithms provide trade-oﬀs between the eﬃciency of the analysis and the precision of the computed solution. The goal of this work is to quantify how the use of ﬂow-sensitivity aﬀects precision and eﬃciency. Although several researchers have provided empirical results of their techniques, comparisons among algorithms can be diﬃcult because of diﬀering program representations, benchmark suites, and metrics. By holding these factors constant, we can focus more on the eﬃcacy of the algorithms and less on the manner in which the results were obtained. The contributions of this paper are the following: – empirical results that measure the precision and eﬃciency of four pointer alias analysis algorithms with varying degrees of ﬂow-sensitivity;

This work was supported in part by the National Science Foundation under grant CCR-9633010, by IBM Research, and by SUNY at New Paltz Research and Creative Project Awards.

G. Levi (Ed.): SAS’98, LNCS 1503, pp. 57–81, 1998. c Springer-Verlag Berlin Heidelberg 1998

58

M. Hind and A. Pioli

– empirical evidence of how various implementation enhancements signiﬁcantly improved analysis time of the ﬂow-sensitive analysis. In addition to the use of ﬂow-sensitivity, other factors that aﬀect the cost/precision trade-oﬀs of pointer alias analyses include the use of context-sensitivity and the manner in which aggregates (arrays and structs) and the heap are modeled. Our experiments hold these factors constant so that the results only reﬂect the usage of ﬂow-sensitivity. Section 2 highlights the four algorithms and their implementations. Section 3 describes the empirical study of the four algorithms, analyzes the results, and contrasts them with related results from other researchers. Section 4 overviews some of the performance-improving enhancements made in the implementation and quantiﬁes their analysis-time speed-up. Section 5 describes other related work. Section 6 states conclusions.

2

Analyses and Implementation

One manner of classifying interprocedural data ﬂow analyses is whether they consider control ﬂow information during the analysis. A flow-sensitive analysis considers control ﬂow information of a procedure during its analysis of the procedure. A flow-insensitive analysis does not consider control ﬂow information during its analysis, and thus can be more eﬃcient, but less precise. (See [31] for a full discussion of these deﬁnitions.) The algorithms we consider, listed in order of increasing precision, are AT: a ﬂow-insensitive algorithm that computes one solution set for the entire program that contains all named objects whose address has been taken and stored, FI: a ﬂow-insensitive algorithm [4, 5] that computes a solution set for every function, FIK: a ﬂow-insensitive algorithm [4, 5] that computes a solution set for every function, but attempts to improve precision by using precomputed (ﬂowsensitive) kill information, FS: a ﬂow-sensitive algorithm [8, 5] that computes a solution set for every program point. The following sections provide further information about these analyses and their implementation. 2.1

Algorithms

The program is represented as a program call (multi-) graph (PCG), in which a node corresponds to a function, and a directed edge represents a call to the target function.1 Each function body is represented by a control ﬂow graph (CFG). This 1

Potential calls can occur due to function pointers and virtual methods, in which the called function is not known until run time.

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

59

graph is used to build a simpliﬁed sparse evaluation graph (SEG) [9], discussed in Section 4. The address-taken analysis (AT) computes its solution by making a single pass over all functions in the program,2 adding to a global set all variables whose addresses have been assigned to another variable. These include actual parameters whose addresses are stored in the corresponding formal. Examples are statements such as “p = &a;”, “q = new ...;”, and “foo(&a);”, but not simple expression statements such as “&a;” because the address was not stored. AT is eﬃcient because it is linear in the size of the program and uses a single solution set, but it can be very imprecise. It is provided as a base case for comparison to the other three algorithms presented in this paper. The general manner in which the other three analyses compute their solutions is the same. A nested ﬁxed point computation is used in which the outer nest corresponds to computing solutions for each function in the PCG. Each such function computation triggers the computation of a local solution for all program points that are distinguished in the particular analysis. For the ﬂowsensitive (FS) analysis, the local solution corresponds to each CFG node in the function. For the other two ﬂow-insensitive analyses (FI, FIK), the local solution corresponds to one set that conservatively represents what can hold anywhere in the function. This general framework is presented as an iterative algorithm in Fig. 1 and is further described in [5]. An extension to handle virtual methods is described in [6]. Section 4 reports improvements due to the use of a worklistbased implementation.

S1 : S2 : S3 : S4 : S5 : S6 : S7 : S8 : S9 : S10 : S11 :

build the initial PCG foreach procedure, p, in the PCG, loop initialize interprocedural alias sets of p to {} end loop repeat foreach procedure, p, in the PCG, loop using the interprocedural alias sets (for entry of p and call sites in p), compute the intraprocedural alias sets of p using the intraprocedural alias sets of p, update the interprocedural alias sets representing the eﬀect of p on each procedure that calls or is called by p end loop using new function pointer aliases, update the PCG, initializing interprocedural alias sets of new functions to {} until the interprocedural alias sets and the PCG converge

Fig. 1. High-level description of general algorithm [5] 2

As the PCG is not used, this can include functions that are not called.

60

M. Hind and A. Pioli

The FS, FI, and FIK analyses utilize the compact representation [8, 5] to represent alias relations. This representation shares the property of the points-to representation [14], in that it captures the “edge” information of alias relations. For example, if variable a points to b, which in turn points to c, the compact representation records only the following alias set: {∗a, b, ∗b, c}, from which it can be inferred that ∗∗ a, c and ∗∗ a, ∗b are also aliases.3 All analyses are context-insensitive; they merge information ﬂowing from diﬀerent calls to the same function, and may suﬀer from the unrealizable path problem [27], i.e., they potentially propagate back to the wrong caller the aliases of the called function. (Sections 3.2 and 4.4 discuss this potential imprecision.) Context-sensitive analyses [14, 48] do not suﬀer from this problem, but may increase time/space costs. As in [22, 7], all analyses considered here represent the (possibly many) objects allocated at calls to new or malloc by creating a named object based on the CFG node number of the allocation statement. These objects are referred to as heapn , where n is the CFG node number of the allocation statement. These names are unique throughout the entire program. More precise heap modeling schemes [29, 22, 19, 7, 8, 11, 15, 16, 39, 40] can improve precision, but may also increase time/space costs. Quantifying the eﬀects of using context-sensitivity and various heap models is beyond the scope of this work. Consider the simple program in Fig. 2. The AT analysis computes only one set of objects, which it assumes all pointers may point to. This set will contain ﬁve objects {heapS1 , heapS3 , heapS4 , heapS6 , and heapS7 }, all of which will appear to be referenced at S8.

S1: S2: S3:

T ∗p, ∗q; void main(void) { p = new T; f(); p = new T; }

S4: S5: S6: S7:

void f(void) { p = new T; g(); p = new T; q = new T; }

S8:

void g(void) { T t = ∗p; }

Fig. 2. Example program

The FI analysis does not use any intraprocedural control ﬂow information. Instead it conservatively computes what can hold anywhere within a function, i.e., for each function, f , it uses only one alias set, Holdsf , to represent what may hold at any CFG node in f . Thus, in Fig. 2 the FI analysis assumes that ∗p, heapS1 and ∗p, heapS3 can ﬂow into f. This results in Holdsg = Holdsf = {∗p, heapS1 , ∗p, heapS3 , ∗p, heapS4 , ∗p, heapS6 , ∗q, heapS7 }, which re3

See [30, 5] for a discussion of precision trade-oﬀs between this representation and an explicit representation, which would contain all four alias pairs.

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

61

sults in four objects, heapS1 , heapS3 , heapS4 , and heapS6 potentially being referenced at S8.4 The FIK analysis attempts to improve the precision of the ﬂow-insensitive analysis by precomputing kill information for pointers, and then uses this information during the ﬂow-insensitive analysis at call sites.5 For example, the precomputation will determine that all alias relations involving ∗p on entry to f will be killed before the call to g at S5. Thus, Holdsg will contain only the alias relations that are generated in f and propagated to g, i.e., Holdsg = {∗p, heapS4 , ∗p, heapS6 ∗q, heapS7 }. This results in two objects, heapS4 and heapS6 , potentially being referenced at S8. The FS analysis associates an alias set before (Inn ) and after (Outn ) every CFG node, n. For example, OutS1 = {∗p, heapS1 } because ∗p and heapS1 refer to the same storage. At the entry to function g, the FS analysis will compute InS8 = {∗p, heapS4 }, which is the precise solution for this simple example. This example illustrates the theoretical precision levels of the four analyses, from FS (most precise) to AT (least precise). The AT analysis is our most eﬃcient analysis because it is linear and only uses one set. The FI analysis is more eﬃcient than the FIK analysis because it neither precomputes kill information nor uses it during the analysis. One would expect the FS analysis to be the least eﬃcient because it needs to distinguish solutions for every point in the program. Thus, a theoretical spectrum exists in terms of precision and eﬃciency with the AT analysis on the less precise/more eﬃcient side, the FS analysis on the more precise/less eﬃcient side, and the FI and FIK analyses in the middle. Not studied in this paper are other ﬂow-insensitive analyses [45, 42] that use one alias set for the whole program and limit the number of alias relations by sometimes grouping distinct variables into one named object. These analysis fall in between AT and FI in the theoretical precision spectrum. 2.2

Implementation

The analyses have been implemented in the NPIC system, an experimental program analysis system written in C++. The system uses multiple and virtual inheritance to provide an extensible framework for data ﬂow analyses [21, 33]. A prototype version of the IBM VisualAge C++ compiler [43, 32] is used as the front end. The abstract syntax tree constructed by the front end is transformed into a PCG and a CFG for each function, which serve as input to the alias analyses. No CFG is built for library functions. We model a call to a library function based on its semantics, thereby providing the beneﬁts of context-sensitive analysis of such calls. Library calls that cannot aﬀect the value of a pointer are treated as the identity transfer function. 4 5

Although a ﬁnal intraprocedural ﬂow-sensitive pass can be used to improve precision [5], this pass has not been implemented. Kill information is computed in a single ﬂow-sensitive prepass of each CFG. For each call site, c, we compute two sets, the set of pointers that are deﬁnitely killed on all paths from entry to c and the set of pointers that are deﬁnitely killed on all paths from c to exit [4, 5]. Only the ﬁrst set is used in our example.

62

M. Hind and A. Pioli

The FS, FI, and FIK analyses are implemented using worklists. (Section 4.2 discusses an earlier iterative implementation.) These three analyses incorporate function pointer analysis into the pointer alias analysis as described in [4, 5]. Currently, array elements and ﬁeld components are not distinguished, and setjmp/longjmp6 statements are not supported. The implementation also assumes that pointer values will only exist in pointer variables, and that pointer arithmetic does not result in the pointer going beyond array boundaries. As stated in Section 2.1, heap objects are named based on their allocation site. To model the values passed as argc and argv to the main function, a dummy main function was added, which called the benchmark’s main function, simulating the eﬀects of argc and argv. This function also initialized the iob array, used for standard I/O. The added function is similar to the one added by Ruf [36, 38] and Landi et al. [28, 26]. Initializations of global variables are automatically modeled as assignment statements in the dummy main function.

3

Results

Our benchmark suite contains 21 C programs, 18 provided by other researchers [28, 14, 36] and 3 from the SPEC CINT92 [3] and CINT95 [44] benchmarks.7 Table 1 describes characteristics of the suite. The third column contains the number of lines in the source and header ﬁles reported by the Unix utility wc. The fourth column reports the number of user-deﬁned functions (nodes in the PCG), which include the dummy main function. The next two columns give the number of call sites, distinguished between user and library function calls. The next two columns report cumulative statistics for all CFG nodes and edges. These ﬁgures include nodes and edges created by the initialization of globals. The following column computes the ratio of CFG edges to nodes. The next column reports the percentage of CFG nodes that are considered pointer-assignment nodes. The current analysis treats an assignment as a pointer-assignment if the variable involved in the pointer expression on the left side of the assignment is declared to be a pointer.8 The last two columns report the number of recursive functions (functions that are in PCG cycles) and heap allocation sites in each program. The last row of the table reports the average edge/node ratio and the 6 7

8

Although one program in our benchmark suite, anagram, does syntactically contain a call to longjmp, the code is unreachable. Some programs had to be syntactically modiﬁed to satisfy C++’s stricter type checking semantics. A few program names are diﬀerent than those reported by Ruf [36]. The SPEC CINT92 program 052.alvinn was named backprop in Todd Austin’s benchmark suite [2]. Ruf referred to ks as part, and ft as span [38]. This is more conservative than considering statements in which the left side expression is a pointer. Thus, statements such as “p->field = ...” are treated as pointer assignments no matter how the type of field is declared. A more accurate categorization would not aﬀect the precision of the analysis, but could improve the eﬃciency by reducing the number of nodes considered during the analysis as discussed in Section 4.1.

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

63

average pointer-assignment node percentage, both of which are computed by averaging the corresponding values over the 21 benchmarks. Table 1. Static characteristics of benchmark suite Name allroots 052.alvinn 01.qbsort 15.trie 04.bisect 17.bintr anagram lex315 ks 05.eks 08.main 09.vor loader 129.compress ft football compiler assembler yacr2 simulator 099.go Average

3.1

Source Landi SPEC92 McCat McCat McCat McCat Austin Landi Austin McCat McCat McCat Landi SPEC95 Austin Landi Landi Landi Austin Landi SPEC95

Call Sites LOC Funcs User Lib 227 7 19 35 272 9 8 13 325 8 9 25 358 13 19 21 463 9 11 18 496 17 27 28 650 16 22 38 733 17 102 52 782 14 17 67 1,202 30 62 49 1,206 41 68 53 1,406 52 174 28 1,539 30 79 102 1,934 25 35 28 2,156 38 63 55 2,354 58 257 274 2,360 40 349 107 3,446 52 247 243 3,979 59 158 169 4,639 111 447 226 29,637 373 2,054 22

Nodes 157 223 173 167 175 196 332 527 513 678 684 876 687 596 732 2,710 1,723 1,544 2,030 2,686 16,823

CFG Edges 167 243 187 181 189 205 365 615 576 732 700 917 773 652 808 3,164 2,090 1,738 2,328 3,123 20,176

Edges N odes

1.064 1.090 1.081 1.084 1.080 1.046 1.099 1.167 1.123 1.080 1.023 1.047 1.125 1.094 1.104 1.168 1.213 1.126 1.147 1.163 1.199 1.111

Ptr-Asgn Rec Allocation Nodes Funcs Sites 2.6% 2 1 11.2% 0 0 24.9% 1 5 23.4% 3 5 9.7% 0 2 9.7% 5 1 9.0% 1 2 2.5% 0 3 26.9% 0 5 3.8% 0 3 24.3% 3 8 27.5% 5 8 8.3% 2 7 6.4% 0 0 19.5% 0 5 1.8% 1 0 1.1% 14 0 8.0% 0 16 5.4% 5 26 2.7% 0 4 .2% 1 0 10.9%

Description of Experiment

This section presents precision and eﬃciency results. For each benchmark and each analysis, we report the analysis time, the maximum memory used, and the average number of objects the analysis computes a dereferenced pointer can point to. The precision results for the FIK analysis are exactly the same as the FI analysis for all benchmarks. Thus, we do not explicitly include this analysis in our precision data. The experiment was performed on a 132MHz IBM RS/6000 PowerPC 604 with 96MB RAM and 104MB paging space, running AIX 4.1.5. The executable was built with IBM’s xlC compiler using the “-O3” option. The analysis time is reported in seconds and does not include the time to build the PCG and CFGs, but does include any analysis-speciﬁc preprocessing, such as building the SEG from the CFG. This information is displayed in the top left chart in Fig. 3. The top right chart of this ﬁgure reports the high-water mark of memory usage during the analysis process. This memory usage includes the intermediate representation, the alias information, and statistics-related data. This information was obtained by using the “ps v” command under AIX 4.1.5. To collect precision information, the system traverses the representation visiting each expression containing a pointer dereference and, using the computed

64

M. Hind and A. Pioli Analysis Time (Secs)

Max Memory Usage (MB)

allroots

allroots

052.alvinn

052.alvinn

01.qbsort

01.qbsort

15.trie

15.trie

04.bisect

04.bisect

17.bintr

17.bintr

anagram

anagram

lex315

lex315

ks

ks

05.eks

05.eks

08.main

08.main 10.3 15.3 11.9

09.vor loader 129.compress

09.vor loader 129.compress

ft

ft

football

football

compiler

compiler

assembler yacr2 simulator 099.go 0

1

2

3

Flow Sensitive

4

5

21.0 19.0

052.alvinn

25.0

01.qbsort

13.0

15.trie

20.0

04.bisect

28.0

17.bintr

34.0

anagram

20.0

lex315

54.0

ks

16.0

05.eks

32.0

08.main

29.0

09.vor

97.0

loader

27.0

129.compress

38.0

ft

205.0

football

112.0

compiler

195.0

assembler

106.0

yacr2

239.0

simulator 099.go 2

Flow Sensitive

099.go 0

Flow Insensitive/Kill

allroots

1

yacr2 simulator

6

Reads Precision (Num Objects)

0

assembler

8.6 9.7 10.5 10.6 8.6 10.1 9.4 14.3 9.8

3

4

180.0 17.0 17.0

5

5

10

Flow Insensitive

32.6 20

15

Address Taken

Writes Precision (Num Objects) 21.0

allroots

19.0

052.alvinn

25.0

01.qbsort

13.0

15.trie

20.0

04.bisect

28.0

17.bintr

34.0

anagram

20.0

lex315

54.0

ks

16.0

05.eks

32.0

08.main

29.0

09.vor

97.0

loader

27.0

129.compress

38.0

ft

205.0

football compiler (No Derefs)

195.0

assembler

106.0

yacr2

239.0

simulator 099.go 0

Flow Insensitive

1

2

3

Address Taken

Fig. 3. Analysis time, memory usage, and precision results

4

180.0 13.6 13.6

5

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

65

alias information, reports how many named objects are aliased to the pointer expression. We report the average number of such dereferences for both reads and writes. Further precision information is provided in [21]. A pointer expression with multiple dereferences, such as ∗ ∗ ∗p, is counted as multiple dereference expressions, one for each dereference. The intermediate dereferences (∗p and ∗∗p) are counted as reads. The last dereference (∗∗∗p) is counted as a read or write depending on the context of the expression. Statements such as (∗p)++ and ∗p += increment are treated as both a read and a write of ∗p. We consider a pointer to be dereferenced if the variable is declared as a pointer or an array formal parameter, and one or more of the “∗”, “->”, or “[ ]” operators are used with that variable. Formal parameter arrays are included because their corresponding actual parameter(s) could be a pointer. We do not count the use of the “[ ]” operator on arrays that are not formal parameters because the resulting “pointer” (the array name) is constant, and therefore, counting it may skew results. Figure 4 classiﬁes the type of pointer dereferenced averaged over all programs. Information for each benchmark is given in [21]. 100 80

46.6

Percent

61.1 60

18.9

40

Global

38.9 20

34.5

22.8 0

Reads

Formal Local

Writes

Fig. 4. Classiﬁcation of dereferenced pointer types for all programs

The manner in which the heap is modeled must be considered in evaluating precision results. For example, a model that uses several names for objects in the heap may seem less precise when compared to a model that uses fewer names [36]. Similarly, analyses that represent invisible objects (objects not lexically visible in the current procedure) aliased to a formal parameteras a single object9 may report fewer objects. Our analyses do not use this technique. Assuming a correct input program, each pointer dereference should correspond to at least one object at run time, and thus one serves as a lower bound for the average. Although a precision result close to one demonstrates the analysis is precise (modulo heap and invisible object naming), a larger number could reﬂect an imprecise algorithm, a limitation of static analysis, or a pointer dereference that corresponds to diﬀerent memory locations over the program’s execution. 9

This modeling technique can increase the possibility of strong updates [7].

66

M. Hind and A. Pioli

Flow-Sensitive — Reads allroots 052.alvinn 01.qbsort 15.trie 04.bisect 17.bintr anagram lex315 ks 05.eks 08.main 09.vor loader 129.compress ft football compiler assembler yacr2 simulator 099.go

Flow-Sensitive — Writes

1.6 1.0 1.8 1.0 1.0 1.0 1.1 2.0 1.7 1.0 3.6 1.6 1.4 1.1 2.4 1.5 1.0 2.1 1.0 1.9 17.0* 0

1

2

3

4

allroots 052.alvinn 01.qbsort 15.trie 04.bisect 17.bintr anagram lex315 ks 05.eks 08.main 09.vor loader 129.compress ft football compiler 0.0 assembler yacr2 simulator 099.go

5

0

Flow-Insensitive — Reads allroots 052.alvinn 01.qbsort 15.trie 04.bisect 17.bintr anagram lex315 ks 05.eks 08.main 09.vor loader 129.compress ft football compiler assembler yacr2 simulator 099.go

1.0 1.8 1.0 1.0 1.0 1.1 2.0 1.8 1.0 4.6 1.7 1.7 1.1 2.5 1.6 1.0 2.2 1.0 1.9 17.0*

Local

1

2

3

4

Nonvisible Local

1.0 1.5 1.0 1.1 1.0 1.0 1.8 1.6 1.2 2.6 1.3 2.2 1.1 1.7 1.7 2.2 1.1 2.0 13.6* 1

2

3

4

5

Flow-Insensitive — Writes

1.6

0

1.0

1.0 allroots 1.0 052.alvinn 1.5 01.qbsort 1.1 15.trie 1.1 04.bisect 1.0 17.bintr 1.0 anagram 1.8 lex315 1.9 ks 1.2 05.eks 3.3 08.main 1.4 09.vor 2.4 loader 1.1 129.compress 1.8 ft 2.8 football compiler (No Derefs) 2.4 assembler 1.1 yacr2 2.1 simulator 099.go

5

0

Global

1

2

Formal Parameter

3

13.6* 4

5

Heap

∗ The bars for 099.go are truncated. The object type breakdown for both FS and FI is Nonvisible Formal Local Local Global Parameter Heap Reads 0.1 9.7 7.2 0.01 0 Writes 0.0 8.1 5.4 0.01 0

Fig. 5. Breakdown of average object type pointed to by a dereferenced pointer

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

67

The bottom two charts of Fig. 3 provide a graphical layout of precision information for reads and writes. Fig. 5 reﬁnes this information for the FS and FI analyses by providing a breakdown of the type of object pointed to. Fig. 6 provides similar information averaged over all programs. Charts E and F of this ﬁgure report the percentage of dereferenced pointers that resolve to exactly one object in our model. If the object is a named variable, as opposed to a heap object, the pointer dereference could be replaced with the variable. This information is expanded upon in [21]. 3.2

Discussion

As expected, the results from the analysis speed chart of Fig. 3 indicate that the AT analysis is eﬃcient; it takes less than .4 seconds on all programs. The FI analysis can be over twice as fast as the FS analysis, and was faster than the FS analysis in all but one program. The precision of the AT analysis leaves much to be desired. Fig. 6 reports on average 111.9 objects for reads and 96.68 objects for writes were in the AT set.10 As one would expect this set to increase with the size of the program, the precision for this analysis will worsen with larger programs. The results also indicate that the FIK analysis is not beneﬁcial. On our benchmark suite it is never more precise than the FI analysis, and on some occasions requires more analysis time than the FS analysis. One explanation for the precision result may be that an alias relation created to simulate a reference parameter, in which the formal points to the actual, typically is not killed in the called routine, i.e., the formal parameter is not modiﬁed, but rather is used to access the passed actual. Thus, programs containing these alias relations will not beneﬁt from the precomputed kill information. One surprising result is the overall precision of the FI analysis. In 12 of the 21 benchmarks the FI analysis is as precise as the FS analysis. This seems to suggest that the added precision obtained by the FS analysis in considering control ﬂow within a function is not signiﬁcant for those benchmarks, at least where pointers are dereferenced. We oﬀer two possible explanations: 1. Pointer variables are often not assigned more than one distinguished object within the same function. Thus, distinguishing program points within a function, a key diﬀerence between the FS and FI analyses, does not often result in an increase in precision. We have seen exceptions to this in the function InitLists of the ks benchmark and in the function InsertPoint in the 08.main benchmark. In both cases the same pointer is reused in two list-traversing loops. 2. It seems that a large number of alias relations are created at call sites because of actual/formal parameter bindings. The lack of a substantial precision diﬀerence between our FS and FI analyses may be because both algorithms rely on the same (context-insensitive) mapping mechanism at call sites. 10

The numbers diﬀer because they are weighted with the number of reads and writes through a pointer in each program.

68

M. Hind and A. Pioli Total Precision Reads 111.9

2.5

9.1 .9

80 82.5 40

0

2.12 FS

2.29 FI

18.9 .6

AT

2.29

2.12

Avg Num Objects

Avg Num Objects

120

2.0

1.13

1.00 1.5

.07

.07

1.0

.65

.61 0.5

.44 0

.44 0

0.0 FS

(A)

FI

(B) Total Precision Writes

96.68

80 60

2.5

9.0 1.5

Avg Num Objects

Avg Num Objects

100

62.7

40 20 0

2.27 FS

2.40 FI

AT

22.9 .5

2.40

2.27

2.0

1.00

.92

1.5 1.0

.06 .45

0.5

.84

.06 .50

.84

0

0.0

FS

(C)

0 FI

(D)

Total Percentage of Resolved Dereferences to One Object Reads

100

80 68

63

60 37

33

40 20 FS

Pct Resolved

Pct Resolved

80

0

1

1

26 4 0

24 4 0

FI

69

0 AT

Nonvisible Local

66

60 40 20 0

(E)

Local

Writes

100

FS

48

46

1 10 10 0

1 9 10 0

FI

0 AT

(F)

Global

Formal Parameter

Heap

Fig. 6. Charts A – D provide the average precision over all benchmarks for reads and writes. (Charts B and D do not include the AT analysis to allow the diﬀerence between the FS and FI analyses to be visible.) Charts E and F report the percentage of dereferenced pointers that resolve to one object in our model.

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

69

Considering charts B and D of Fig. 3, it seems that FI is as precise as FS for pointers directed to nonvisible locals and formal parameters. Therefore, FS, if employed at all, should focus on pointers directed to the heap and global variables. The precision results for 099.go merit discussion. An average of 17.03 and 13.64 objects are returned for reads and writes, respectively, with a maximum of 100. This program contains six small list-processing functions (using an arraybased “cursor” implementation) that accept a pointer to the head of a list as a parameter. One of these functions, addlist, is called 404 times, passing the address of 100 diﬀerent actuals for the list header, resulting in 100 aliases for the formal parameter. However, because the lifetime of the formal is limited to this function (it does not call any other function), these relations are not propagated to any other function. Therefore, these relations do not suﬀer the eﬀects of the unrealizable path problem mentioned in Section 2.1. Another conclusion from the results is that analysis time is not only a function of program size; it also depends on the amount of alias relation propagation along the PCG and SEGs. For example, 099.go, despite being our largest program and having a pointer aliased with 100 objects, is analyzed at one of the fastest rates (3,037 LOC/second, 1,724 CFG nodes/second) because these relations are not propagated throughout the program. As suggested by Shapiro and Horwitz [41] and Diwan [12], a more precise and time-consuming alias analysis may not be as ineﬃcient as it may appear because the time required to obtain increased precision may reduce the time required by subsequent analyses that utilize mod-use information, and thus pointer alias information, as their input. As the previous paragraph suggests, this can also be true about pointer alias analysis itself, which also utilizes pointer alias information during its analysis. 3.3

Comparison with Other Results

Landi et al. [28] report precision results for the computation of the MOD problem using a ﬂow-sensitive pointer alias algorithm with limited context-sensitive information. Among the metrics they report is the number of “thru-deref” assigns, which corresponds to the “write” metrics reported in Fig. 3. However, since their results include compiler-introduced temporaries in their “thru-deref” count [26], a direct comparison is not possible. Stocks et al. [46] use the same metric without including temporaries for the ﬂow-sensitive context-sensitive analysis of Landi and Ryder [27]. They report the average number of objects ranges from 1.0 to 2.0 on the eight common benchmarks. On these benchmarks our ﬂow-sensitive context-insensitive analysis ranges from 1.0 to 2.22. Two possible explanations for the slightly less precise results are 1) their algorithm uses some context-sensitivity; 2) the underlying representation is not identical, and thus pointer dereferences may not be counted in the same manner in all cases. For example, statements such as cfree(TP) located in allroots are treated as modifying the structure deallocated, and thus as a pointer dereference [26]. In fact, on the three programs in which our analysis

70

M. Hind and A. Pioli

reports the same, or close to the same, number of “writes” as “thru-derefs” (allroots, lex315, simulator), our precision is identical to that reported in [46]. The relative precision of the ﬂow-insensitive analysis compared to the ﬂowsensitive analysis is in contrast to the study of Stocks et al. [46], which compares the ﬂow-sensitive analysis of Landi and Ryder [27] with a ﬂow-insensitive analysis described in [49]. For the eight common benchmarks, our ﬂow-insensitive algorithm ranges from 1.0 to 2.81 objects on average for a write dereference, compared to 1.0 to approximately 6.3 for the ﬂow-insensitive analysis they studied. The analysis described in [49] shares the property of Steensgaard’s analysis [45] in that it groups all objects pointed-to by a variable into an equivalence class. Although this can lead to spurious alias relations not present in the FI analysis, it does allow for an almost linear time algorithm, which has been shown to be fast in practice [45, 42, 50]. Emami et al. [14] report precision results for a ﬂow-sensitive context-sensitive algorithm. Their results range from 1.0 to 1.77 for all indirect accesses using a heap naming scheme that represents all heap objects with one name. Because we were unable to obtain the benchmarks from their suite, a direct comparison with our results is not possible. Ruf [36] reports both read and write totals for a ﬂow-sensitive contextinsensitive analysis. However, unlike our analysis he counts use of the “[ ]” operator on arrays that are not formal parameters as a dereference [38]. Since such an array will always point to the same place, the average number of objects is improved.11 For the 11 benchmarks in common,12 Ruf reports an overall read and write average of 1.33 and 1.38, respectively. To facilitate comparisons, we have also counted in this manner. The results for the common benchmarks are averages of 1.35 and 1.47 for the FS analysis and 1.41 and 1.54 for the FI analysis. We attribute the slight diﬀerences in the FS analysis to the diﬀerence in representations. As Ruf [36] states, “the VDG intermediate representation often coalesces series of structure or array operations into a single memory write.” This coalescing can skew results in either direction. Shapiro and Horwitz [41] present an empirical comparison of four ﬂowinsensitive algorithms. The ﬁrst algorithm is the same as the AT algorithm. The remaining three algorithms [45, 42, 1] can be less precise and more eﬃcient than the algorithms studied in this paper.13 The authors measure the precision 11

12

13

The best illustration of this is in 099.go, which has a large number of array references, but a low number of pointer dereferences. In this program, the average changed from 17.03 to 1.13 for reads and from 13.64 to 1.48 for writes when all uses of the “[ ]” operator were counted. Although [36] reports results for ft (under the name span), our version of the benchmark is substantially larger than the one Ruf analyzed, and thus is not comparable. In theory, the FI analysis can be more precise than Andersen’s algorithm [1] because it considers function scopes, at the cost of using more than one alias set. However, both algorithms are likely to oﬀer similar precision in practice because the distinguishing case is probably uncommon.

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

71

of these analyses by implementing three dataﬂow analyses and an interprocedural slicing algorithm. In addition to these alias analysis clients, the authors also report the direct precision of the alias analysis algorithms in terms of the total number of points-to relations. We agree with [14, 36] that a more meaningful metric is to measure where the points-to information is used, such as where a pointer is dereferenced. They conclude 1) a more precise ﬂow-insensitive analysis generally leads to increased precision by the subsequent analyses that use this information with varying magnitudes; 2) metrics measuring the alias analysis precision tend to be good predictors on the precision of subsequent analyses that use alias information; and 3) more precise ﬂow-insensitive analysis can also improve the eﬃciency of subsequent analyses that use this information.

4

Eﬃciency Improvements

This section describes some of the performance-improving enhancements made in the implementation and quantiﬁes their eﬀects on analysis-time speed-up for the ﬂow-sensitive algorithm. Although novelty is not claimed, the eﬃcacy of each technique is shown. #1

P(char* f)

a=5.0;

Entry {,,} p #2 while(...)

#7 g = &b;

#3 l = &a;

#4

if(...)

Exit

p

{,}

#5 b=10.0;

#6 P(g);

Fig. 7. Example CFG (left) and procedure P (right). Note that g is a global, a,b,l are locals of the CFG, and f is a formal parameter of P. Possible topological numbers (ignoring the back edge 7 → 2) appear next to each node.

4.1

Sharing Alias Sets

As described in Section 2.1, the ﬂow-sensitive analysis computes solutions before and after every node in the CFG. However, this can result in storing redundant

72

M. Hind and A. Pioli

information. For example, node #1 in Fig. 7 does not aﬀect any pointer value; therefore its Out set will always equal its In set. Likewise, a node whose predecessor(s) Out set(s) have the same value will have an In set equal to this value. For example, all nodes in Fig. 7 except #1 and #2. Thus, for all nodes, except #1 and #2, alias sets can be shared. We use the term shared in a literal way — if the In set is shared with the Out set, they are the same object during the analysis. This is done with a shallow copy of the alias set data structures. We precompute these sharing sites in a separate forward pass over the CFG before performing the alias analysis. Each node that has its own set, which we call a deep set, is dubbed a “SEG” (Sparse Evaluation Graph) node.14 Such nodes have a list of “SEG” predecessors and “SEG” successors that are used during the analysis. The alias set allocation strategy for a node, n, is summarized as follows: Outp , if ∀p, q ∈ P red(n), p and q share Out sets Inn = a deep set, otherwise Outn =

Inn , if n’s transfer function is the identity function a deep set, otherwise

Our current implementation treats all call nodes as SEG nodes. Table 2 reports the number of alias sets before and after this optimization was applied as well as the percentage reduction. In addition to saving storage (on average 73.78% fewer alias sets were allocated), this method saves visits during the actual analysis to nodes that can not aﬀect pointer aliasing. Although we allocate fewer alias sets, the real beneﬁt of this technique is seen during the analysis. Since we do not visit and propagate alias relations between extraneous CFG nodes, we simply have fewer alias relations being stored and copied from one CFG node to the next. This aﬀects both the analysis’s time and space use in a signiﬁcant way. Table 3 shows the eﬀects of the eﬃciency improving ideas on the ﬂow-sensitive analysis. Run times were collected as in Section 3, using the C function clock(), which gives the CPU time, not the elapsed time. This metric was chosen because it eliminates the eﬀects of system load, amount of RAM vs. paging space, etc. The elapsed time for each program was approximately 2.5 times the CPU time. The headings for each column are read vertically — for example, the second column shows the analysis time in seconds with no sharing of alias sets as well as no other enhancements described in this section. The third column reports the analysis time and resulting speed-up using this sharing technique. The eﬀectiveness of this technique is mostly related to the percentage of CFG nodes that aﬀect a pointer and percentage of call nodes — the higher the percentages the lower the potential beneﬁt.15 For all our benchmarks, these averages were 10.9% 14 15

The method described is a conservative approximation to the SEG of [9] and is similar to [35]. The percentage of merge nodes plays a smaller role.

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

73

Table 2. Eﬀectiveness of sharing alias sets Num. Alias Sets Benchmark No Sharing Sharing Pct Saved allroots 328 87 73.48% 052.alvinn 464 89 80.82% 01.qbsort 362 111 69.34% 15.trie 360 125 65.28% 04.bisect 352 78 77.84% 17.bintr 350 124 64.57% anagram 696 159 77.16% lex315 1088 260 76.10% ks 1054 317 69.92% 05.eks 694 170 75.50% 08.main 1166 399 65.78% 09.vor 1842 633 65.64% loader 1434 382 73.36% 129.compress 1152 195 83.07% ft 1140 380 66.67% football 5536 962 82.62% compiler 3486 699 79.95% assembler 3184 911 71.39% yacr2 3960 782 80.25% simulator 5180 1203 76.78% 099.go — 4966 — Average 73.78%

and 20.8% respectively. (09.vor had high averages in these categories and has the lowest eﬀectiveness for this technique while yacr2 had lower than average percentages and resulted in a higher speed-up.) Over all benchmarks, this technique results in a signiﬁcant speed-up, 2.80 on average. For our largest benchmark, 099.go, the analysis did not run (due to insuﬃcient memory) until we applied this optimization. 4.2

Worklists

The initial implementation of the analyses used an iterative algorithm for simplicity. After correctness was veriﬁed, a worklist-based implementation was used to improve eﬃciency. Two types of worklists were used: SEG node worklists and function worklists. Each function has a worklist of SEG nodes. The PCG has two worklists of functions: “current” and “next.” (The motivation for using two worklists will be described in Section 4.3.) We use nested “while not empty” loops with the worklists — the outer loop visiting functions and the inner loop visiting SEG nodes. The worklist of functions initially contains all functions. On the ﬁrst visit to a function, we initialize the function’s SEG node worklist with all SEG nodes in that function. If a SEG node’s Out set changes, all its SEG successors are placed on its function’s SEG node worklist. If a function’s entry set changes, it is placed on the “next” function worklist. If the exit set of a function changes, all calling functions are placed on the “next” function worklist and the calling call node(s) are placed on their respective function’s SEG node worklist. The analysis runs until all worklists are empty.

74

M. Hind and A. Pioli

Table 3. Flow-sensitive analysis run time in seconds. Numbers in parentheses are the speed-up from the previous version to the left. Benchmark Name allroots (227 LOC) 052.alvinn (272 LOC) 01.qbsort (325 LOC) 15.trie (358 LOC) 04.bisect (463 LOC) 17.bintr (495 LOC) anagram (650 LOC) lex315 (733 LOC) ks (782 LOC) 05.eks (1202 LOC) 08.main (1206 LOC) 09.vor (1406 LOC) loader (1539 LOC) 129.compress (1934 LOC) ft (2156 LOC) football (2354 LOC) compiler (2360 LOC) assembler (3446 LOC) yacr2 (3979 LOC) simulator (4639 LOC) 099.go (29637 LOC) Averages

←−Sharing −→ ←−Worklists −→ Overall Unsorted Sorted Speed-up ←−No Forward Bind Filter −→ Filter 1.41 0.54 0.21 0.24 0.10 (2.61) (2.57) (0.88) (2.40) 14.10 1.03 0.39 0.09 0.10 0.10 (2.64) (4.33) (0.90) (1.00) 10.30 8.22 4.16 1.81 1.19 0.75 (1.98) (2.30) (1.52) (1.59) 10.96 4.59 2.26 0.83 0.72 0.45 (2.03) (2.72) (1.15) (1.60) 10.20 1.47 0.45 0.16 0.13 0.14 (3.27) (2.81) (1.23) (0.93) 10.50 2.20 1.19 0.33 0.32 0.32 (1.85) (3.61) (1.03) (1.00) 6.88 6.67 2.16 0.59 0.52 0.48 (3.09) (3.66) (1.13) (1.08) 13.90 5.03 2.09 0.73 0.70 0.41 (2.41) (2.86) (1.04) (1.71) 12.27 20.48 9.40 3.32 2.55 1.49 (2.18) (2.83) (1.30) (1.71) 13.74 9.21 2.87 0.92 0.87 0.67 (3.21) (3.12) (1.06) (1.30) 13.75 96.37 44.80 18.30 12.39 3.16 (2.15) (2.45) (1.48) (3.92) 30.50 217.19 113.71 38.70 32.96 11.92 (1.91) (2.94) (1.17) (2.77) 18.22 176.49 58.37 26.71 21.20 3.81 (3.02) (2.19) (1.26) (5.56) 46.32 6.00 1.82 0.56 0.53 0.41 (3.30) (3.25) (1.06) (1.29) 14.63 80.94 37.07 14.04 11.25 5.09 (2.18) (2.64) (1.25) (2.21) 15.90 275.63 61.30 26.87 22.75 2.70 (4.50) (2.28) (1.18) (8.43) 102.09 39.31 12.87 4.19 4.11 3.70 (3.05) (3.07) (1.02) (1.11) 10.62 668.25 240.46 119.31 97.26 9.69 (2.78) (2.02) (1.23) (10.04) 68.96 377.11 86.50 26.76 25.93 10.64 (4.36) (3.23) (1.03) (2.44) 35.44 511.49 146.35 82.41 61.84 10.12 (3.49) (1.78) (1.33) (6.11) 50.54 — 228.18 74.01 65.23 9.83 — (3.08) (1.13) (6.64) 23.21 2.80 2.84 1.16 3.09 25.38

No Sharing ←−Iterating −→

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

75

Column 4 of Table 3 shows the improvement of the worklist-based implementation over the iterating version, both of which use shared alias sets. The result was an average speed-up of 2.84 over the iterating version, which produced an average speed-up of 2.78 over the nonshared iterating version. 4.3

Sorted Worklists

Using an iterating analysis of a forward data-ﬂow problem, it is natural to process the nodes in topological order. The next enhancement was to use priority queues (based on a topological order16 ) for the SEG and function worklists. Consider the case of a loop body that generates aliasing information. It would be best to process the loop body before moving on to the loop exit. Topological order alone does not give this property — we may process the loop exit before the loop body. (This would occur in Fig. 7 using the given node numbers as the topological order.) However, during the construction of the CFG, nodes for loop bodies are created before those nodes after the loop body. (Thus, node #7 would be created before node #3.) Since nodes are assigned numbers as they are created (which is performed in a topological order, except in the presence of gotos), using the nodes creation number as a priority ensures that loop bodies are processed ﬁrst. A result of using a single priority-based worklist of functions was that calling functions were visited before called functions. However, unlike SEG nodes, aliases can be propagated in both directions along a PCG edge. Thus, an optimal order of function visits is not apparent. In our benchmark suite, using a single prioritybased worklist for functions provided only a marginal improvement over the iterating version. To increase eﬃciency, we use two function worklists — “current” and “next.” While visiting functions on the “current” worklist, we place functions on the “next” worklist. This has the eﬀect that once a set of functions is on the worklist, the visiting order is ﬁxed in a topological order. When the “current” worklist is empty, we swap the “current” and “next” worklists and continue the analysis. The ﬁfth column of Table 3 reports the analysis time using sorted worklists for both SEG nodes and functions and the previous enhancements. This enhancement resulted in an average speed-up of 1.16 over the previous version, which used nonpriority-based worklists and shared alias sets. 4.4

ForwardBind Filtering

ForwardBind() is a function in our analysis that propagates alias relations from call nodes to the entry set of called functions. If needed, it creates alias relations for the formals from the actuals in the called function’s entry set and then unions in the call node’s In set with the called function’s entry set. Consider those alias relations that cannot be changed or used in the called function (for example, the 16

Topological order on a cyclic graph can be obtained by performing a depth-ﬁrst search of the graph and then removing edges classiﬁed as back edges [24].

76

M. Hind and A. Pioli

relation ∗l, a in Fig. 7), but are still propagated through the called function until they reach the exit set of the function. These relations are then propagated back to the call node’s Out set. In eﬀect, the called function acts as an identity transfer function for those relations that are not relevant in the called function. Although correct, this is ineﬃcient. Our enhancement is not to propagate from the call node those alias relations that cannot be reached in a called function. To compute the set of alias relations that are not reachable in the called function, F , we ﬁrst call ForwardBind(), which propagates alias relations into the entry set of F , EntryF , as described above. We then view all alias relations in EntryF as a directed graph and remove from this graph all vertices (distinguished objects), and associated edges (alias relations), that cannot be reached from a global or a formal of F . These removed edges (alias relations) are simply unioned into the call node’s Out set directly. This can help limit the propagation eﬀects of the unrealizable path problem. The last column of Table 3 reports the eﬀectiveness of this optimization. It provided the most dramatic speed-up, an average of 3.09 over the previous implementation, which used all other enhancements. Some programs, in particular football and assembler, had a much higher than average speed-up resulting from this ﬁltering. These two programs both shared some common characteristics: they have a single function that has both an unusually high amount of pointer aﬀecting statements and a very high number of called functions. Figure 8 shows the eﬀects of these optimizations in a dramatic way for the loader benchmark.17 We collected the data presented in this graph by repeatedly executing the “ps v” command (under AIX 4.1.5) while each of the ﬁve FS analyses was running. We recorded the SIZE column, which gives the virtual size of the data section of the process; this will capture all heap allocated memory usage during the analysis. The x-axis of the chart is simply marked oﬀ in samples (a sample is a single call to “ps v”). As the majority of our heap allocations are used to represent alias relations, the resulting memory usage can be interpreted as the number of alias relations stored by the analysis while running. The “No Sharing” version shows a characteristic curve; it grows quickly early on, but then levels oﬀ as the analysis reaches its ﬁxed point. The diﬀerence between the No Sharing and Sharing versions shows how the large reduction in the number of alias sets reduced the number of alias relations that were stored. The cumulative eﬀect of all ﬁve optimizations was an average speed-up of 25.38. This illustrates the beneﬁts that can be obtained by limiting the propagation of extraneous alias relations and the number of visits to functions and nodes.

5

Other Related Work

This section describes other related work not mentioned in Section 3.3. 17

The other benchmarks had similarly shaped graphs.

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

77

loader

25

No Sharing Sharing Worklists Priority Queues Filtering

20

MB

15

10

5

0

0

500

1000

1500

Samples

2000

2500

3000

3500

Fig. 8. Memory usage over time for the loader benchmark.

Ruf [36] presents an empirical study of two algorithms: a ﬂow-sensitive algorithm similar to the one we have implemented, and a context-sensitive version of the same algorithm. His results showed that the context-sensitive algorithm did not improve precision for pointers where they are dereferenced, but cautioned that this may be a characteristic of the benchmark suite analyzed. Wilson and Lam [48, 47] present an algorithm for performing context-sensitive analysis that avoids redundant analyses of functions for similar calling contexts. Ghiya and Hendren [17] present empirical data showing how points-to and connection analyses can improve traditional transformations, array dependence testing, and program understanding. Ruf [37] describes a program partitioning technique that is used for a ﬂowsensitive points-to analysis, achieving a storage savings of 1.3–7.2 over existing methods. Diwan et al. [13] provide static and dynamic measurements of the eﬀectiveness of three ﬂow-insensitive analyses for a type-safe language (Modula3). With the exception of AT, all three algorithms are less precise than the versions we have studied. Zhang et al. [50] report the eﬀectiveness of applying diﬀerent pointer aliasing algorithms to diﬀerent parts of a program. Hasti and Horwitz [18] present a pessimistic algorithm that attempts to increase the precision of a ﬂow-insensitive analysis by iterating over a ﬂowinsensitive analysis and an SSA [10] construction. No empirical results are reported. Horwitz [23] deﬁnes precise ﬂow-insensitive alias analysis and proves that, even in the absence of dynamic memory allocation, computing it is NPhard.

6

Conclusions

This work has described an empirical study of four pointer alias analysis algorithms that use varying degrees of ﬂow-sensitivity. We have found that

78

M. Hind and A. Pioli

– the address-taken analysis, although eﬃcient, is unlikely to provide suﬃcient precision; – the ﬂow-insensitive analysis with kill is not beneﬁcial; – the precision of the ﬂow-insensitive analysis is identical to that of the ﬂowsensitive analysis in 12 of 21 programs in our benchmark suite; – most published implementations of ﬂow-sensitive pointer analysis have equivalent precision. Although the ﬂow-sensitive analysis eﬃciently analyzed a program on the order of 30K LOCs, further benchmarks are needed to see if this property generalizes. We have also empirically demonstrated how various implementation strategies result in signiﬁcant analysis-time speed-up.

7

Acknowledgements

We thank Michael Burke, Michael Karasick, and Lee Nackman for their support of this work. We also thank Todd Austin, Bill Landi, and Rakesh Ghiya for making their benchmarks available. Bill Landi, Laurie Hendren, Erik Ruf, Barbara Ryder, and Bob Wilson provided useful details concerning their implementations. Michael Burke, Paul Carini, and Jong-Deok Choi provided useful discussions regarding the algorithms described in [8]. Discussions with Manuel F¨ ahndrich led to reporting intermediate read dereferences, which were not considered in [20]. Harini Srinivasan designed and implemented the initial control ﬂow graph builder, an important early system component. NPIC group members Robert Culley, Lynne Delesky, Lap Chung Lam, Giampaolo Lauria, Mark Nicosia, Joseph Perillo, Keith Sanders, Truong Vu, and Ming Wu assisted with the implementation and testing of the system. David Bacon helped with the initial design of the program call graph representation. Michael Burke, Jong-Deok Choi, John Field, G. Ramalingam, Harini Srinivasan, Laureen Treacy, and the anonymous referees provided useful comments on earlier drafts of this paper.

References [1] Lars Ole Andersen. Program Analysis and Specialization for the C Programming Language. PhD thesis, DIKU, University of Copenhagen, May 1994. Available at ftp.diku.dk/pub/diku/semantics/papers/D-203.dvi.Z. 70, 70 [2] Todd Austin. Pointer-intensive benchmark suite, version 1.1. http://www.cs.wisc.edu/˜austin/ptr-dist.html, 1995. 62 [3] Subra Balan and Walter Bays. Spec announces new benchmark suites cint92 and cfp92. Technical report, Systems Performance Evaluation Cooperative, March 1992. SPEC Newsletter 4(1). 62 [4] Michael Burke, Paul Carini, Jong-Deok Choi, and Michael Hind. Flow-insensitive interprocedural alias analysis in the presence of pointers. In K. Pingali, U. Banerjee, D. Gelernter, A. Nicolau, and D. Padua, editors, Lecture Notes in Computer

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

[5]

[6]

[7]

[8]

[9]

[10]

[11]

[12] [13]

[14]

[15]

[16]

[17]

[18]

79

Science, 892, pages 234–250. Springer-Verlag, 1995. Proceedings from the 7th Workshop on Languages and Compilers for Parallel Computing. Extended version published as Research Report RC 19546, IBM T. J. Watson Research Center, September 1994. 58, 58, 61, 62 Michael Burke, Paul Carini, Jong-Deok Choi, and Michael Hind. Interprocedural pointer alias analysis. Research Report RC 21055, IBM T. J. Watson Research Center, December 1997. 58, 58, 58, 59, 59, 60, 60, 61, 61, 62 Paul Carini, Michael Hind, and Harini Srinivasan. Flow-sensitive interprocedural type analysis for C++. Research Report RC 20267, IBM T. J. Watson Research Center, November 1995. 59 David R. Chase, Mark Wegman, and F. Kenneth Zadeck. Analysis of pointers and structures. In SIGPLAN ’90 Conference on Programming Language Design and Implementation, pages 296–310, June 1990. SIGPLAN Notices 25(6). 60, 60, 65 Jong-Deok Choi, Michael Burke, and Paul Carini. Eﬃcient ﬂow-sensitive interprocedural computation of pointer-induced aliases and side eﬀects. In 20th Annual ACM SIGACT-SIGPLAN Symposium on the Principles of Programming Languages, pages 232–245, January 1993. 58, 60, 60, 78 Jong-Deok Choi, Ron Cytron, and Jeanne Ferrante. Automatic construction of sparse data ﬂow evaluation graphs. In 18th Annual ACM Symposium on the Principles of Programming Languages, pages 55–66, January 1991. 59, 72 Ron Cytron, Jeanne Ferrante, Barry K. Rosen, Mark N. Wegman, and F. Kenneth Zadeck. Eﬃciently computing static single assignment form and the control dependence graph. ACM Transactions on Programming Languages and Systems, pages 451–490, October 1991. 77 Alain Deutsch. Interprocedural may-alias analysis for pointers: Beyond k-limiting. In SIGPLAN ’94 Conference on Programming Language Design and Implementation, pages 230–241, June 1994. SIGPLAN Notices, 29(6). 60 Amer Diwan. Personal communication, August 1997. 69 Amer Diwan, Kathryn S. McKinley, and J. Elliot B. Moss. Type-based alias analysis. In SIGPLAN ’98 Conference on Programming Language Design and Implementation, pages 106–117, June 1998. SIGPLAN Notices, 33(5). 77 Maryam Emami, Rakesh Ghiya, and Laurie J. Hendren. Context-sensitive interprocedural points-to analysis in the presence of function pointers. In SIGPLAN ’94 Conference on Programming Language Design and Implementation, pages 242–256, June 1994. SIGPLAN Notices, 29(6). 60, 60, 62, 70, 71 Rakesh Ghiya and Laurie J. Hendren. Connection analysis: A practical interprocedural heap analysis for C. In Lecture Notes in Computer Science, 1033, pages 515–533, August 1995. Proceedings from the 8th Workshop on Languages and Compilers for Parallel Computing. 60 Rakesh Ghiya and Laurie J. Hendren. Is it a tree, a dag or a cyclic graph? A shape analysis for heap-directed pointers in C. In 23rd Annual ACM SIGACTSIGPLAN Symposium on the Principles of Programming Languages, pages 1–15, January 1996. 60 Rakesh Ghiya and Laurie J. Hendren. Putting pointer analysis to work. In 25th Annual ACM SIGACT-SIGPLAN Symposium on the Principles of Programming Languages, pages 121–133, January 1998. 77 Rebecca Hasti and Susan Horwitz. Using static single assignment form to improve ﬂow-insensitive pointer analysis. In SIGPLAN ’98 Conference on Programming Language Design and Implementation, pages 97–105, June 1998. SIGPLAN Notices, 33(5). 77

80

M. Hind and A. Pioli

[19] Laurie J. Hendren and Alexandru Nicolau. Parallelizing programs with recursive data structures. IEEE Transactions on Parallel and Distributed Systems, 1(1):35– 47, January 1990. 60 [20] Michael Hind and Anthony Pioli. An empirical comparison of interprocedural pointer alias analyses. Research Report RC 21058, IBM T. J. Watson Research Center, December 1997. 78 [21] Michael Hind and Anthony Pioli. Assessing the eﬀects of the ﬂow-sensitivity on pointer alias analyses (extended version). Technical Report 98-104, SUNY at New Paltz, June 1998. 61, 65, 65, 67 [22] S. Horwitz, P. Pfeiﬀer, and T. Reps. Dependence analysis for pointer variables. In SIGPLAN ’89 Conference on Programming Language Design and Implementation, pages 28–40, June 1989. SIGPLAN Notices 24(6). 60, 60 [23] Susan Horwitz. Precise ﬂow-insensitive may-alias analysis is NP-Hard. ACM Transactions on Programming Languages and Systems, 19(1):1–6, January 1997. 77 [24] John B. Kam and Jeﬀrey D. Ullman. Global data ﬂow analysis and iterative algorithms. Journal of the ACM, 23(1):158–171, January 1976. 75 [25] William Landi. Undecidability of static analysis. ACM Letters on Programming Languages and Systems, 1(4):323–337, December 1992. 57 [26] William Landi. Personal communication, October 1997. 62, 69, 69 [27] William Landi and Barbara Ryder. A safe approximate algorithm for interprocedural pointer aliasing. In SIGPLAN ’92 Conference on Programming Language Design and Implementation, pages 235–248, June 1992. SIGPLAN Notices 27(6). 60, 69, 70 [28] William Landi, Barbara Ryder, and Sean Zhang. Interprocedural modiﬁcation side eﬀect analysis with pointer aliasing. In SIGPLAN ’93 Conference on Programming Language Design and Implementation, pages 56–67, June 1993. SIGPLAN Notices 28(6). 62, 62, 69 [29] James R. Larus and Paul N. Hilﬁnger. Detecting conﬂicts between structure accesses. In SIGPLAN ’88 Conference on Programming Language Design and Implementation, pages 21–34, 1988. SIGPLAN Notices, 23(7). 60 [30] Thomas Marlowe, William Landi, Barbara Ryder, Jong-Deok Choi, Michael Burke, and Paul Carini. Pointer-induced aliasing: A clariﬁcation. SIGPLAN Notices, 28(9):67–70, September 1993. 60 [31] Thomas J. Marlowe, Barbara G. Ryder, and Michael G. Burke. Deﬁning ﬂow sensitivity in data ﬂow problems. Technical Report RC 20138, IBM T. J. Watson Research Center, July 1995. 58 [32] Lee R. Nackman. Codestore and incremental C++. Dr. Dobbs Journal, pages 92–95, December 1997. 61 [33] Anthony Pioli. Conditional pointer aliasing and constant propagation. Master’s thesis, SUNY at New Paltz, 1998. In preparation. 61 [34] G. Ramalingam. The undecidability of aliasing. ACM Transactions on Programming Languages and Systems, 16(6):1467–1471, November 1994. 57 [35] G. Ramalingam. On sparse evaluation representations. In Pascal Van Hentenryck, editor, Lecture Notes in Computer Science, 1302, pages 1–15. Springer-Verlag, 1997. Proceedings from the 4th International Static Analysis Symposium. 72 [36] Erik Ruf. Context-insensitive alias analysis reconsidered. In SIGPLAN ’95 Conference on Programming Language Design and Implementation, pages 13–22, June 1995. SIGPLAN Notices, 30(6). 62, 62, 62, 65, 70, 70, 70, 71, 77

Assessing the Eﬀects of Flow-Sensitivity on Pointer Alias Analyses

81

[37] Erik Ruf. Partitioning dataﬂow analyses using types. In 24th Annual ACM SIGACT-SIGPLAN Symposium on the Principles of Programming Languages, pages 15–26, January 1997. 77 [38] Erik Ruf. Personal communication, October 1997. 62, 62, 70 [39] Mooly Sagiv, Thomas Reps, and Reinhard Wilhelm. Solving shape-analysis problems in languages with destructive updating. In 23rd Annual ACM SIGACTSIGPLAN Symposium on the Principles of Programming Languages, pages 16–31, January 1996. 60 [40] Mooly Sagiv, Thomas Reps, and Reinhard Wilhelm. Solving shape-analysis problems in languages with destructive updating. ACM Transactions on Programming Languages and Systems, 20(1):1–50, January 1998. 60 [41] Marc Shapiro and Susan Horwitz. The eﬀects of the precision of pointer analysis. In Pascal Van Hentenryck, editor, Lecture Notes in Computer Science, 1302, pages 16–34. Springer-Verlag, 1997. Proceedings from the 4th International Static Analysis Symposium. 69, 70 [42] Marc Shapiro and Susan Horwitz. Fast and accurate ﬂow-insensitive point-to analysis. In 24th Annual ACM SIGACT-SIGPLAN Symposium on the Principles of Programming Languages, pages 1–14, January 1997. 61, 70, 70 [43] D. Soroker, M. Karasick, J. Barton, and D. Streeter. Extension mechanisms in Montana. In 8th IEEE Israeli Conference on Software and Systems, pages 119– 128, June 1997. 61 [44] SPEC. SPEC CPU95, Version 1.0. Standard Performance Evaluation Corporation, http://www.specbench/org, August 1995. 62 [45] Bjarne Steensgaard. Points-to analysis in almost linear time. In 23rd Annual ACM SIGACT-SIGPLAN Symposium on the Principles of Programming Languages, pages 32–41, January 1996. 61, 70, 70, 70 [46] Philip A. Stocks, Barbara G. Ryder, William A. Landi, and Sean Zhang. Comparing ﬂow and context sensitivity on the modiﬁcations-side-eﬀects problem. In International Symposium on Software Testing and Analysis, pages 21–31, March 1998. 69, 70, 70 [47] Robert P. Wilson. Eﬃcient Context-Sensitive Pointer Analysis for C Programs. PhD thesis, Stanford University, December 1997. 77 [48] Robert P. Wilson and Monica S. Lam. Eﬃcient context-sensitive pointer analysis for C programs. In SIGPLAN ’95 Conference on Programming Language Design and Implementation, pages 1–12, June 1995. SIGPLAN Notices, 30(6). 60, 77 [49] Sean Zhang, Barbara G. Ryder, and William Landi. Program decomposition for pointer aliasing: A step toward practical analyses. In 4th Symposium on the Foundations of Software Engineering, pages 81–92, October 1996. 70, 70 [50] Sean Zhang, Barbara G. Ryder, and William Landi. Experiments with combined analysis for pointer aliasing. In ACM SIGPLAN-SIGSOFT Workshop on Program Analysis for Software Tools and Engineering, pages 11–18, June 1998. 70, 77

Analysis of Normal Logic Programs Fran¸cois Fages and Roberta Gori LIENS CNRS, Ecole Normale Sup´erieure, 45 rue d’Ulm, 75005 Paris, France, [email protected] Dipartimento di Informatica, Universit` a di Pisa, Corso Italia 40, 56125 Pisa, Italy [email protected] Ph.: +39-50-887248 Fax: +39-50-887226

Abstract. In this paper we present a dataflow analysis method for normal logic programs interpreted with negation as failure or constructive negation. We apply our method to a well known analysis for logic programs: the depth(k) analysis for approximating the set of computed answers. The analysis is correct w.r.t. SLDNF resolution and optimal w.r.t. constructive negation. Keywords: Abstract interpretation, static analysis, logic programming, constructive negation.

1

Introduction

Important results have been achieved for static analysis using the theory of abstract interpretation [6]. Abstract interpretation is a general theory for specifying and validating program analysis. A key point in abstract interpretation is the choice of a reference semantics from which one can abstract the properties of interest. While it is always possible to use the operational semantics, it is possible to get rid of useless details, by choosing a more abstract semantics as reference semantics. In the case of definite logic programs, much work has been done in this sense. Choosing the most abstract logical least model semantics limits the analysis to type inference properties, that approximate the ground success set. Non-ground model semantics have thus been developed, under the name of S-semantics [2], and proved useful for a wide variety of goal-independent analysis ranging from groundness, to sharing, call patterns, etc. All the intermediate fixpoint semantics between the most abstract logical one and the most concrete operational one, form in fact a hierarchy related by abstract interpretation, in which one can define a notion of the best reference semantics [12] for a given analysis. On the other hand, less work has been done on the analysis of normal logic programs, although the finite failure principle, and hence SLDNF resolution, are standard practice. The most significant paper on the analysis of normal G. Levi (Ed.): SAS’98, LNCS 1503, pp. 82–98, 1998. c Springer-Verlag Berlin Heidelberg 1998

Analysis of Normal Logic Programs

83

logic programs, using the theory of abstract interpretation, is the one by Marriott and Søndergaard [19], which proposes a framework based on Fitting’s least three-valued model semantics [11]. Since this reference semantics is a ground semantics, the main application of this framework is type analysis. Marriott and Søndergaard already pointed out that a choice of a different reference semantics could lead to an improved analysis. Fitting’s least three-valued model semantics is, in fact, an abstraction (a non recursively enumerable one, yet easier to define) of Kunen’s three-valued logical semantics [14] which is more faithful to SLDNF resolution [15] and complete w.r.t. constructive negation. These are exactly the directions along which we try to improve the results of Marriott and Søndergaard. We consider the inference rule of constructive negation, which provides normal logic programs with a sound and complete [21] operational semantics w.r.t. Kunen’s logical semantics [14]. We propose an analysis method for normal logic programs interpreted with constructive negation, based on the generalized S-semantics given in [9] and on the hierarchy described in [10]. We present here an analysis based on the depth(k) domain which approximates the computed answers obtained by constructive negation and therefore the three-valued consequences of the program completion and CET (Clark’s equational theory). Other well known analyses for logic programs can be extended to normal logic programs. For example, starting from a suitable version of Clark’s semantics a groundness analysis was defined which is correct and optimal w.r.t. constructive negation. Here, for lack of space, we present only the depth(k) analysis. We show that it is correct and also optimal w.r.t. constructive negation. Finally we give an example which shows that in the case of type inference properties our semantics yields a result which is more precise than the one obtained by Marriott and Søndergaard. From the technical point of view, the contribution of the paper is the definition of a normal form for first order constraints on the Herbrand Universe, which is suitable for analysis. In fact the normal form allows us to define an abstraction function which is a congruence w.r.t. the equivalence on constraints induced by the Clark’s equality theory. The paper is organized as follows. In Sect. 2 we introduce some preliminary notions on constructive negation. In Sect. 3 we define a normal form on the concrete domain of constraints in order to deal, with equivalence classes of constraints w.r.t. the Clark’s equational theory. Section 4 defines the abstract domain and abstract operator and show its correctness and optimality (under suitable assumptions on the depth of the cut) w.r.t. the concrete one. Finally, Subsect. 4.5 shows an example.

2

Preliminaries

The reader is assumed to be familiar with the terminology of and the basic results in the semantics of logic programs [1,17] and with the theory of abstract interpretation as presented in [6,7].

84

2.1

F. Fages and R. Gori

Normal Logic Programs and Constructive Negation

We consider the equational version of normal logic programs, where a normal program is a finite set of clauses of the form A ← c|L1 , ..., Ln , where n ≥ 0, A is an atom, called the head, c is a conjunction of equalities, and L1 , ..., Ln are literals. The local variables of a program clause are the free variables in the clause which do not occur in the head. With V ar(A) we intend the free variables in the atom A. In order to deal with constructive negation, we need to consider the domain C of full first-order equality constraints on the structure H of the Herbrand domain. Assuming an infinite number of function symbols in the alphabet, Clark’s equational theory (CET) provides a complete decidable theory for the constraint language [18,14], i.e. 1. (soundness) H |= CET , 2. (completeness) for any constraint c, either CET |= ∃c or CET |= ¬∃c. A constraint is in prenex form if all its quantifiers are in the head. The set of free variables in a constraint c is denoted by V ar(c). For a constraint c, we shall use the notation ∃c (resp. ∀c) to represent the constraint ∃X c (resp. ∀X c) where X = V ar(c). A constrained atom is a pair c|A where c is an H-solvable constraint such that V ar(c) ⊆ V ar(A). The set of constrained atoms is denoted by B. A constrained interpretation is a subset of B. A three-valued or partial constrained interpretation is a pair of constrained interpretations < I + , I − >, representing the constrained atoms which are true and false respectively (note that because of our interest in abstract interpretations we do not impose any consistency condition between I + and I − ). Constructive negation is a rule of inference introduced by Chan for normal logic programs in [3], which provides normal logic programs with a sound and complete [21] operational semantics w.r.t. Kunen’s logical semantics [14]. In logic programming, Kunen’s semantics is simply the set of three-valued consequences of the program’s completion and the theory CET . The S-semantics of definite logic programs [2] has been generalized to normal logic programs in [9] for a version of constructive negation, called constructive negation by pruning. The idea of the fixpoint operator, which captures the set of computed answer constraints, is to consider a non-ground finitary (hence continuous) version of Fitting’s operator. Here we give a definition of the operator TPBD which is parametric w.r.t. the domain BD of constrained atoms and the operations on constraints on the domain D. Definition 1. Let P be a normal logic program. TPBD is an operator over P(BD )× P(BD ) defined by

Analysis of Normal Logic Programs

85

TPBD (I)+ = {c|p(X) ∈ BD : there exists a clause in P with local variables Y , C = p(X) ← d|A1 , ..., Am , ¬Am+1 , ..., ¬An . c1 |A1 , ..., cm |Am ∈ I + , cm+1 |Am+1 , ..., cn |An ∈ I − such that c = ∃Y (d ∧c1 ∧ . . . ∧cn )} TPBD (I)− = {c|p(X) ∈ BD : for each clause defining p in P with loc. var. Yk , Ck = p(X) ← dk |Ak,1 , ..., Ak,mk , βk . there exist ek,1 |Ak,1 , ..., ek,mk |Ak,mk ∈ I − , nk ≥ mk , ek,mk+1 |Ak,mk+1 , ..., ek,nk |Ak,nk ∈ I + , where for mk+1 ≤ j ≤ nk , ¬Ak,j occurs in βk , V such that c = k ∀Yk (¬ dk ∨ ek,1 . . . ∨ ek,nk )}. where the operations ∃, ∀, ¬, ∨, ∧, are the corresponding operations on the constraint domain of D. In the case of a normal logic program, the operator TPB defines a generalized S-semantics which is fully abstract w.r.t. the computed answer constraints with constructive negation by pruning [9]. By soundness it approximates also the set of computed answer constraints under the SLDNF resolution rule, or under the Prolog strategy. In [10] we have shown that this operator defines a hierarchy of reference semantics related by abstract interpretation, that extends the hierarchy defined by Giacobazzi for definite logic programs [12]. Here we show the use of the hierarchy for the static analysis of normal logic programs.

3

Normal Forms in CET

Unlike the semantics in Marriott and Søndergaard’s framework, our reference semantics is a non ground semantics and has to deal with first-order equality constraints. The first problem that arises is to define a normal form for such constraints on the Herbrand domain, so that abstraction functions on constrained atoms can be defined. In general, in fact, given a theory th, we are interested in working with equivalence classes of constraints w.r.t. the equivalence of the constraints in th. Namely c is equivalent to c0 if th |= c ↔ c0 . Therefore we need the abstraction function on the concrete constraint domain to be a congruence. This is a necessary property since it permits to be independent from the syntactic form of the constraints. Dealing with normal logic programs, we need to achieve this property in CET. We thus need to introduce a normal form for first-order equality constraints, in a similar way to what has been done for the analysis of definite programs where the normal form is the unification solved form [16]. Here we shall define a new notion of “false-simplified” normal forms, making use of Colmerauer’s solved forms for inequalities [4], Maher’s transformations for first-order constraints [18] and an extended disjunctive normal form [13]. First let us motivate the need of a “false-simplified” form. Let us call a basic constraint an equality or an inequality between a variable and a term. The abstraction function will be defined inductively on the basic constraints, and it

86

F. Fages and R. Gori

will sometimes (e.g. for groundness analysis) abstract to true some inequalities. Consider, for example, the following constraint d = ∀X(Y = b ∧ X 6= f (a)). d is H-equivalent to f alse. If the abstraction of X 6= f (a) is true then the abstraction of d will be the abstraction of Y = b, which cannot be H-equivalent to the abstraction of f alse. Therefore we need to define a normal form where the constraints which are H-equivalent to f alse, are eliminated. Definition 2. Consider a constraint d in prenex disjunctive form, d = ∆(∨i Ai ), where ∆ is a sequence of quantified variables and ∨i Ai is a finite disjunction. d is in a f alse-simplified form if, either there does not exist a proper subset I of the i0 s such that H |= ∆(∨i Ai ) ↔ ∆(∨i∈I Ai ), or such an I exists and there exists also a subset K of I, such that ∨j6∈I Aj is H-equivalent to ∨k∈K Ak . The latter condition assures that we really eliminate constraints that are Hequivalent to f alse and that are not just redundant in the constraint. Now the existence of a false-simplified form for any constraint can be proved simply with the following: Algorithm 3 Input: a constraint in prenex disjunctive form d = ∆(∨i Ai ). Let us call U the set of the indices i’s in d = ∆(∨i Ai ). 1. Let I and J be the partition of U such that i ∈ I if H |= ∃∆(Ai ), otherwise i ∈ J. 2. Repeat I := I ∪S as long as there exists an S ⊆ J such that H |= ∃∆(∨i∈S Ai ) and for all j ∈ S H 6|= ∃∆(∨i∈(S\{j}) Ai ). 3. Let S ⊆ J\I be any minimal set such that H |= ∃∆(∨s∈S As ∨i∈I Ai ) and H |= ∆(∨s∈S As ∨i∈I Ai ) ↔ d, do I := I ∪ S, 4. Output: ∆(∨i∈I Ai ). The idea of the algorithm is to find a subset of the conjunctions Ai ’s (those with i ∈ I) such that ∆(∨i∈I Ai ) is in false-simplified form and it is H-equivalent to ∆(∨i Ai ). In the first step we select the Ai ’s such that ∆(Ai ) is H-satisfiable. In this case, in fact, Ai cannot be H-equivalent to f alse and it can be put in the set I. In the second step from the remaining Ai ’s we select the set of Ai ’s such that their ∆ quantified disjunction is H-satisfiable, since we check that all the Ai ’s are necessary for the quantified disjunction to be H-satisfiable, the considered Ai ’s can not be H-equivalent to f alse. At the end of this process, if the resulting constraint is H-equivalent to the input constraint, we stop. Otherwise, we add a minimum number of the not yet selected Ai ’s such the ∆(∨i Ai ) for the selected i’s is H-equivalent to the input constraint. Since we add a minimum number of not yet selected Ai ’s, we are sure that the resulting constraint is in falsesimplified form. Example 4 shows how the algorithm 3 works on two examples. Example 4. 1. Input: c1 = ∀T (A1 ∨ A2 ∨ A3 ∨ A4 ), A1 = (T = f (H) ∧ Y = a), A2 = (T 6= f (a) ∧ Y = b), A3 = (Y 6= g(H, T )), A4 = (T 6= a ∧ Y = a). I1 = {3}. I2 = {3, 1, 2}. I3 = {3, 1, 2}( since H |= ∀T (∨i∈I2 Ai ) ↔ c1 ). Output: ∀T (A1 ∨ A2 ∨ A3 ).

Analysis of Normal Logic Programs

87

2. Input: c2 = ∀T (A01 ∨ A02 ∨ A03 ), A01 = (T 6= f (U ) ∧ T 6= f (V )), A02 = (T = H), A03 = (U 6= V ∧ T = f (a)). I1 = {}. I2 = {1, 2}. I3 = {1, 2, 3}( since H 6|= ∀T (∨i∈I2 A0i ) ↔ c2 ).1 Output: ∀T (A01 ∨ A02 ∨ A03 ). Theorem 5. For any input constraint c = ∆(∨i Ai ), algorithm 3 terminates and computes a false-simplified form logically equivalent to c. Note that all the false-simplified forms of a constraint c are H-equivalent. Now the intuitive idea for a normal form is the following. We put a constraint in prenex form and we compute the disjunctive form of its quantifier free part. We make equality and inequality constraints interact in every conjunction of the disjunctive form and then we compute the false-simplified form for the resulting constraint. The problem is that if we consider a standard disjunctive normal form, we would not be able to see explicitly all the relations between constraints in disjunctions. Consider, for example, the constraint (X = f (H) ∨ (H 6= f (a)). This constraint is equivalent, therefore H-equivalent, to the constraint ((X = f (f (a))∧H = f (a))∨H 6= f (a)). Note that the equality H = f (a) is not explicit in the first disjunction. Since the abstraction function will act on the terms of the disjunction independently, this could cause a problem. This is why we will use a well known extended disjunctive form defined for Boolean algebra and applied, in our case, to the algebra of quantifier free constraints. In the next theorem with Bi we denote basic equality or inequality constraints (X = t or X 6= t). For any Bi let Bi f alse = ¬Bi and Bi true = Bi . Theorem 6. [13] For every Boolean formula φ on basic equality or inequality constraints B1 , . . . , B Wn , φ ↔ ψ where ψ = ( (a1 ,... ,an )∈{f alse,true}n φ(a1 , . . . , an ) ∧ B1a1 ∧ . . . ∧ Bnan ). Note that ψ is a formula in disjunctive form. ψ has in fact a particular disjunctive form where each conjunction contains all the basic constraints (possibly negated) of φ. This is why, this form is able to capture all the possible relations between the different terms of a disjunction. We will call the formula ψ the extended disjunctive normal form (dnf ) of φ. The next example shows how the extended disjunctive normal form works on a constraint c1 . Example 7. c1 = (X = f (H) ∨ H 6= f (a)). dnf (c1 ) = ((X = f (f (a)) ∧ H = f (a))∨ (X = f (H) ∧ H 6= f (a)) ∨ (X 6= f (H) ∧ H 6= f (a))). Note that although c1 , dnf (c1 ) and ((X = f (f (a)) ∧ H = f (a)) ∨ H 6= f (a)) are H-equivalent, dnf (c1 ) is the most “complete” in the sense that it shows syntactically all the relations between constraints in disjunctions. 1

U = b, H = f (b), V = a, in fact, is an assignment (for the free variables of c2 ), which is a solution of c2 but is not a solution of ∀T (∨i∈I2 A0i ).

88

F. Fages and R. Gori

We can now define the normal form, Res(c), of a first-order equality constraint c, as the result of the following steps: 1. Put the constraint c in prenex form obtaining ∆(c1 ), where ∆ is a sequence of quantified variables and c1 is the quantifier free part of c. 2. Compute dnf (c1 ) = ∨(Ai ), 3. Simplify each conjunction Ai obtaining A0i = ResConj(Ai ), 4. Return a f alse-simplified form of the constraint ∆(∨A0i ). where the procedure for simplifying each conjunction is based on Maher’s canonical form [18] and Colmerauer’s simplification algorithm for inequalities [4]. The procedure performs the following steps, ResConj(A) 1. compute a unification solved form for the equalities in the conjunction A 2. for each equality X = t in A, substitute X by t at each occurrence of X in the inequalities of conjunction A. 3. simplify the inequalities by applying the following rules, a) replace f (t1 , . . . , tn ) 6= f (s1 , . . . , sn ) by t1 6= s1 ∨ . . . ∨ tn 6= sn . b) replace f (t1 , . . . , tn ) 6= g(s1 , . . . , sn ) by true. c) replace t = 6 x by x 6= t if t is not a variable. obtaining A0 , 4. if A0 is a conjunction then return A0 . 5. otherwise compute dnf (A) = ∨(Ai ) and return ∨ResConj(Ai ). It is worth noting that the previous algorithm terminates since each constraint contains a finite number of inequalities. Example 8 shows how the procedure Res computes the normal form of some constraints. Example 8. 1. c = (X = f (Y ) ∧ (Y = a ∨ Y = f (a)) ∧ ∀U X 6= f (f (U ))). = ∀U ( X = f (Y ) ∧ (Y = a ∨ Y = f (a)) ∧ X 6= f (f (U ))). c1 = ∀U ( (X = f (Y ) ∧ Y 6= a ∧ Y = f (a) ∧ X 6= f (f (U )))∨ c2 (X = f (Y ) ∧ Y = a ∧ Y 6= f (a) ∧ X 6= f (f (U )))∨ (X = f (Y ) ∧ Y = a ∧ Y = f (a) ∧ X 6= f (f (U )))). c3.1 = ∀U ( (X = f (f (a)) ∧ Y 6= a ∧ Y = f (a) ∧ X 6= f (f (U )))∨ (X = f (a) ∧ Y = a ∧ Y 6= f (a) ∧ X 6= f (f (U ))). c3.2 = ∀U ( (X = f (f (a)) ∧ f (a) 6= a ∧ Y = f (a) ∧ f (f (a)) 6= f (f (U )))∨ (X = f (a) ∧ Y = a ∧ a 6= f (a) ∧ f (a) 6= f (f (U ))). c3.3 = ∀U ( (X = f (f (a)) ∧ Y = f (a) ∧ a 6= U )∨ (X = f (a) ∧ Y = a ∧ a 6= f (U ))). c4 = (X = f (a) ∧ Y = a).

Analysis of Normal Logic Programs

89

2. c¯ = (X = f (Z, S) ∧ U = (f (H), H) ∧ S = a ∧ X 6= U ). = c1 = c¯. c2 c3.1 = (X = f (Z, a) ∧ U = f (f (H), H) ∧ S = a ∧ X 6= U ). c3.2 = (X = f (Z, a) ∧ U = f (f (H), H) ∧ S = a ∧ f (Z, a) 6= f (f (H), H)). c3.3 = (X = f (Z, a) ∧ U = f (f (H), H) ∧ S = a ∧ (Z 6= f (H) ∨ H 6= a). c3.4 = A1 ∨ A2 ∨ A3 . A1 = (X = f (Z, a) ∧ U = f (f (H), H) ∧ S = a ∧ Z = f (H) ∧ H 6= a). A2 = (X = f (Z, a) ∧ U = f (f (H), H) ∧ S = a ∧ Z 6= f (H) ∧ H 6= a). A3 = (X = f (Z, a) ∧ U = f (f (H), H) ∧ S = a ∧ Z 6= f (H) ∧ H = a). ResConj(A1 ) = A¯1 ResConj(A2 ) = A2 ResConj(A3 ) = A¯3 .

c4

A¯1 = (X = f (f (H), a) ∧ U = f (f (H), H) ∧ S = a ∧ Z = f (H) ∧ H 6= a). A¯3 = (X = f (Z, a) ∧ U = f (f (a), a) ∧ S = a ∧ Z 6= f (a) ∧ H = a). = A¯1 ∨ A2 ∨ A¯3 .

Note that all the steps in ResConj and Res preserve the H-equivalence, the third step of ResConj is Colmerauer’s simplification algorithm for inequalities [4], the first and second transformations are the usual ones for CET formulas [18], while the second step of Res is the extended disjunctive normal transformation [13]. Hence we get: Proposition 9. H |= φ ↔ Res(φ). Our concrete constraints domain N C will be the subset of constraints in C which are in normal form. The concrete operations on N C will be thus defined using the normal form: Definition 10. Let c1 , c2 ∈ N C, c1 ∨c c2 = Res(c1 ∨ c2 ) c1 ∧c c2 = Res(c1 ∧ c2 ) c ¬ c1 = Res(¬c1 ) ∃c X c1 = ∃X c1

∀c X c1 = ∀X c1

We denote by B the set of constrained atoms with constraints in N C, and by (I, ⊆) the complete lattice of (not necessarily consistent) partial constrained interpretations formed over B.

4

Depth(k) Analysis for Constructive Negation

The idea of depth(k) analysis was first introduced in [20]. The domain of depth(k) analysis was then used in order to approximate the ground success and failure sets for normal programs in [19]. We follow the formalization of [5] for positive logic programs. We want to approximate an infinite set of computed answer constraints by means of a constraint depth(k) cut, i.e. constraints where the equalities and inequalities are between variables and terms which have a depth not greater than k.

90

F. Fages and R. Gori

Our concrete domain is the complete lattice of partial constrained interpretations (I, ⊆) of the previous section. Since our aim is to approximate the computed answer constraints, the fixpoint semantics we choose in the hierarchy [10] is the one which generalizes the S-semantics to normal logic programs, the TPBD operator (cf def. 1). The version we consider here is the one defined on the domain B with the concrete operations in N C, ∧c , ∨c , ¬c , ∃c , ∀c , (the TPB operator). 4.1

The Abstract Domain

Terms are cut by replacing each-subterm rooted at depth greater than k by a new fresh variable taken from a set W , (disjoint from V ). The depth(k) terms represent each term obtained by instantiating the variables of W with terms built over V . Consider the depth function || : T erm → T erm such that 1 if t is a constant or a variable |t| = max{|t1 |, . . . , |tn |} + 1 if t = f (t1 , . . . , tn ) and a given positive integer k. The abstract term αk (t) is the term obtained from the concrete one by substituting a fresh variable (belonging to W ) to each subterm t0 in t, such that |t| − |t0 | = k. Consider now the abstract basic constraints c | c = (X = t) |t| ≤ k or ABC = c = (X 0 6= t0 ) |t0 | ≤ k, and V ar(t0 ) ∩ W = ∅ Note that V ar(t0 ) ∩ W = ∅ expresses the fact that inequalities between variables and cut terms are not allowed. The domain of abstract constraints is defined as follows, Definition 11. AN C =

c | c is a constraint in normal form built with the logical connectives ∨, ∧, ∀ and ∃ on ABC

The concepts of abstract constrained atoms and partial abstract interpretations are defined as expected. Definition 12. An abstract constrained atom is a pair c|A such that c ∈ AN C and c is a H − solvable constraint, A is an atom and V ar(c) ⊆ V ar(A). With B a we intend the set of abstract constrained atoms. The abstract domain is the set of partial interpretations on abstract constrained atoms. A partial abstract constrained interpretation for a program, is a pair of + − set of abstract constrained atoms, I a =< I a , I a >, not necessary consistent. We consider I a = {I a | I a is a partial interpretation}. With respect to the case of definite logic programs [5], we need to define a different order on the abstract constraint domain.

Analysis of Normal Logic Programs

91

This is because the result ca of an abstract and operation on the abstract constraint domain will be an approximation of the result c of the concrete and operation on the concrete constraint domain, in the sense that ca will be “more general” than the abstraction of c (where here “more general” means “is implied under H”) . This motivates the definition of the following relation on the abstract constraint domain. Definition 13. Let c, c0 ∈ AN C. c a c0 if H |= c → c0 . We consider the order ≤a induced by the preorder a , namely the order obtained considering the classes modulo the equivalence induced by a . We define the downward closure of a pair of sets w.r.t. the ≤a order, Definition 14. Consider a pair of sets of constrained atoms B. By ↓ B we denote the downward closure of < B + , B − >. c|A ∈↓ B + if there exists c0 |A ∈ B + and c ≤a c0 , c|A ∈↓ B − if there exists c0 |A ∈ B − and c ≤a c0 . Intuitively, a set of constrained atoms I is less or equal than J, if ↓ I ⊆↓ J. Definition 15. Consider I, J ∈ I a . + + I a J a ↔ for all c|A ∈ I a ∃c0 |A ∈ J a such that c ≤a c0 and − − for all c|A ∈ I a ∃c0 |A ∈ J a such that c ≤a c0 It is immediate to see that defines a preorder. We consider the order ≤ induced by the preorder , namely the order obtained by considering the classes Ia modulo the equivalence induced by . Then our abstract domain is (Ia , ≤). Since the operations on the equivalence classes are independent on the choice of the representative, we denote the class of an interpretation I a by I a itself. In the rest of the paper, we often abuse notation by denoting by I a the equivalence class of I a or the interpretation I a itself. 4.2

The Abstraction Function

Let us now define the abstraction function. To this aim we first define the function αc on constraints. The main idea is to define αc on the basic constraints as follows: an equality X = t is abstracted to X = αk (t), while an inequality X 6= t is abstracted to X 6= t if |t| ≤ k and to true otherwise. We denote by ∆(c) the constraint c0 in normal form and by ∆ the sequence of quantified variables of c0 , where c is the quantifier-free part of c0 . Definition 16. The depth(k) constraint abstraction function is the function αc : N C → AN C: αc (∆(c)) = ∆, ∆0 αc (c) where ∆0 = ∃Y1 , ∃Y2 , .., and Yi ∈ (W ∩ V ar(αc (c))) αc (X = t) = (X = αk (t)), αc (f alse) = f alse, αc (true) = true, αc (X 6= t) = true if |t| > k, αc (X 6= t) = (X 6= t) if |t| ≤ k, αc (A ∧ B) = αc (A) ∧ αc (B), αc (A ∨ B) = αc (A) ∨ αc (B).

92

F. Fages and R. Gori

Note that the first definition means that all the new variables introduced by the cut terms have to be considered existentially quantified. Example 17 shows an application of αc . Example 17. c = ∀U ((H = f (f (T )) ∧ T 6= f (f (U )) ∧ X = f (U ))∨ (H = f (f (T )) ∧ T 6= f (X) ∧ X 6= f (U ))), k = 2. αc (c) = αc ( ∀U (

(H = f (f (T )) ∧ T 6= f (f (U )) ∧ X = f (U ))∨ (H = f (f (T )) ∧ T 6= f (X) ∧ X 6= f (U )))) = ∀U ( (αc (H = f (f (T ))) ∧ αc (T 6= f (f (U ))) ∧ αc (X = f (U )))∨ (αc (H = f (f (T ))) ∧ αc (T 6= f (X)) ∧ αc (X 6= f (U ))) = ∀U, ∃Q1 , Q2 ((H = f (Q1 ) ∧ true ∧ X = f (U ))∨ (H = f (Q2 ) ∧ T 6= f (X) ∧ X 6= f (U ))) (Q1 , Q2 ∈ W ).

The abstraction function α is defined by applying αc to every constraint of the constrained atoms in the concrete interpretation. Definition 18. Let α : I → Ia : α =< α+ , α− > α+ (I) = {c|A | c0 |A ∈ I + and αc (c0 ) = c}, α− (I) = {c|A | c0 |A ∈ I − and αc (c0 ) = c}. As a consequence the function γ on (equivalence classes of) sets of abstract constraints is automatically determined as follows: Definition 19. Let γ : Ia → I: γ(I a ) = ∪{I | α(I) ≤ I a } = + ∪{I | ∀c|A ∈ α+ (I) ∃c0 |A ∈ I a such that c ≤a c0 and − ∀c|A ∈ α− (I) ∃c0 |A ∈ I a such that c ≤a c0 } = ∪{I | ↓ α(I) ⊆↓ I a } = ∪{I |α(I) ⊆↓ I a } Lemma 20. α is additive. Theorem 21. < α, γ > is a Galois insertion of (I, ⊆) into (Ia , ≤). 4.3

αc is a Congruence w.r.t. the H-Equivalence

As we have already pointed out in Sect. 3, we want to work with H-equivalence classes of constraints and, for this purpose, we need to be sure that the above defined function αc on N C is a congruence w.r.t. the H-equivalence. This means that if two constraints c, c0 ∈ N C are H-equivalent, then also αc (c) and αc (c0 ) have to be H-equivalent. In order to understand whether two constraints are H-equivalent, it is useful to state the following result. Lemma 22. Consider the inequality X 6= t. There exist no arbitrary quantified t1 , . . . , tn , where ti 6= t, such that X 6= t is H-equivalent to ∧i X 6= ti .

Analysis of Normal Logic Programs

93

This is a consequence of the fact that we consider the models of the theory CET without the DCA axiom. The previous result, together with the fact that constraints are in falsesimplified form, allows us to claim that αc is a congruence. Theorem 23. Let c, c0 ∈ N C. If H |= c ↔ c0 then H |= αc (c) ↔ αc (c0 ). 4.4

The Abstract Fixpoint Operator

We now define the abstract operations that will replace the concrete ones in the definition of the fixpoint abstract operator. We show that the abstract operations are a correct approximation of the concrete operations. The definition of the abstract and operation is not immediate. The example 24 is meant to give some intuition on some problems that may arise. Example 24. Consider the following two constraints: c1 = (X = f (Z, f (H)) ∧ S = f (a)) c2 = (U 6= X ∧ Y 6= f (S)) and k = 2. Consider αc (c1 ) = ∃Q(X = f (Z, Q) ∧ S = f (a)) αc (c2 ) = (U 6= X ∧ Y 6= f (S)). If we now consider the normalized form of αc (c1 )∧αc (c2 ) the resulting constraint is ∃Q(U 6= f (Z, Q) ∧ Y 6= f (f (a)) ∧ X = f (Z, Q) ∧ S = f (a)), which is not an abstract constraint according to definition 11. The problem is that the normalized form of the logical and operation on two abstract constraints is not in general an abstract constraint (the depth of the terms involved in equalities and inequalities can be greater than k and it can contain inequalities between variables and cut terms). This is the reason why we need to define a new M operator, on the normalized forms of abstract constraints. The M operator must cut terms deeper than k and replace by true all the inequalities which contain a cut term. Intuitively this is because X 6= t, where V ar(t) ∩ W 6= ∅, represents, on the concrete domain, an inequality between a variable and a term longer than k. On the abstract domain, such inequalities are abstracted to the constant true. Definition 25. Let M : N C → AN C M(∆(c)) = ∆, ∆0 M(c) where ∆0 = ∃Y1 , ∃Y2 , .., where Yi ∈ (W ∩V ar(M(c))). M(X = t) = (X = αk (t)) M(X 6= t) = (X 6= t) if |t| ≤ k and V ar(t) ∩ W = ∅ M(X 6= t) = (true) if |t| > k or V ar(t) ∩ W 6= ∅ M(A ∧ B) = αc (A) ∧ αc (B), M(A ∨ B) = αc (A) ∨ αc (B) As expected, the M operator is similar to the αc operator. The only difference is that M replace by true all the inequalities between variables and cut terms. Since AN C is a subset of N C, the Res form is defined also on the abstract constraints domain. Definition 26. Let c1 , c2 ∈ AN C ˜ c2 = M(Res(c1 ∧ c2 )), ˜ c2 = M(Res(c1 ∨ c2 )), c1 ∨ c1 ∧ ˜ c1 = ∃X c1 , ¬ ˜ c1 = M(Res(¬c1 )), ∃X

˜ c1 = ∀X c1 , ∀X

94

F. Fages and R. Gori

It is worth noting that the procedure Res on the abstract domain needs to perform the logical and on abstract constraints. This means that most of the observations that can be done on the behavior of the abstract and operation, concern also the abstract or and not operations. Example 27 illustrates the relation between the abstract and operation and the abstraction of the concrete and operation. For a sake of simplicity, since in this case it does not affect the result, we write the constraint c1 in the more compact standard disjunctive form rather than of in extended disjunctive form. Example 27. c1 = ∀K((Y = a ∧ U 6= f (f (K))) ∨ Z = a), c2 = (U = f (f (a))). Consider k = 1. αc (c1 ) = (Y = a ∨ Z = a), αc (c2 ) = ∃V U = f (V ). ˜ αc (c2 ) = ∃V ((Y = a ∧ U = f (V )) ∨ (Z = a ∧ U = f (V ))). αc (c1 )∧ αc (Res(c1 ∧ c2 )) = ∃V (Z = a ∧ U = f (V )). ˜ αc (c2 ) H |= αc (Res(c1 ∧ c2 )) → αc (c1 )∧ As already pointed out, the abstract and gives a more general constraint than the abstraction of the one computed by the concrete and and this is the reason why we have defined an approximation order based on implication (under H) between constraints. In order to show that the abstract operations are correct, we prove a stronger property. Theorem 28. Let c1 , c2 ∈ N C. ˜ αc (c2 ) ≥a αc (c1 ∧c c2 ), ˜ αc (c2 ) ≥a αc (c1 ∨c c2 ), αc (c1 )∨ αc (c1 )∧ ˜ αc (c1 ) = αc (∀c x c1 ). ˜ αc (c1 ) = αc (∃c x c1 ), ∀x ∃x As shown by example 29, the correctness property does not hold for the version of abstract “not” which we have defined, if we consider general constraints. Example 29. Consider c1 = (X 6= f (f (a))) and k = 1. αc (¬c (c1 )) = ∃Y X = f (Y ) which does not implies ¬ ˜ (αc (c1 )) = f alse. Since the not operator is used by the abstract fixpoint operator on “simpler “ constraints (the program constraints) only, we are interested in its behavior on conjunctions of equalities between variables and terms only. For this kind of constraints the following result holds. V ˜ αc (c1 ) ≥a αc (¬c (c1 )). Lemma 30. If c1 = ( i (Xi = ti )) ∈ N C, then ¬ Now that we have defined the abstract constraints domain and the abstract operations, we can define the abstract fixpoint operator. Definition 31. Let α(P ) be the program obtained by replacing every constraint c in a clause of P by αc (c). a Ba The abstract fixpoint operator: Ia → Ia is defined as follows, TPB (I a ) = Tα(P ) (↓ ˜ ∀, ˜ ¬ ˜, ∧ ˜ on AN C × AN C. ˜ on AN C and ∨ I a ), where the operations are ∃, a

By definition, TPB is a congruence respect to the equivalence classes of the a abstract domain. Note also that TPB is monotonic on the (Ia , ≤), because I ≤ J implies ↓ I ⊆↓ J.

Analysis of Normal Logic Programs

95

a

Lemma 32. TPB is monotonic on the (Ia , ≤). The proof that the abstract operator is correct w.r.t. the concrete one, is based on the correctness of the abstract operations on the abstract constraints domain. a

a

Theorem 33. α(TPB (γ(I a ))) ≤ TPB (I a ). Then α(lf p(TPB )) ≤ TPB (I a ). Consider now a k greater than the maximal depth of the terms involved in the constraints of the clauses in the program P . In this case the abstract operator is also optimal. a

Theorem 34. TPB (I a ) ≤ α(TPB (γ(I a ))). Let us finally discuss termination properties of the dataflow analyses presented in this section. First note that the set of not equivalent (w.r.t. H) set of constraints belonging to AN C is finite. Lemma 35. Assume that the signature of the program has a finite number of function and predicate symbols. Our depth(k) abstraction is ascending chain finite. 4.5

An Example

We now show how the depth-k analysis works on an example.The program of figure 1 computes the union of two sets represented as lists. We denote the a a equivalence class of TPB by TPB itself. All the computed constraints for the predicate ¬member are shown, while concerning the predicate ¬union, for a sake of simplicity, we choose to show just a small subset of the computed answer constraints (written in the more compact standard disjunctive form). Therefore, the concretization of the set of answer constraints for ¬union that we present in figure 1, contains some answer constraints computed by the concrete semantics but not all of them. As expected the set of answer constraints, computed by the abstract fixpoint operator, is an approximation of the answer constraints, computed by the concrete operator, for the predicates member, union and ¬member. For example, for the predicate ¬member(X, Y ), we compute the answer ∀L(Y 6= [X, L]) which correctly approximates the concrete answer ∀L, H, H1 , L1 (Y 6= [X, L] ∧ Y 6= [H, H1 , L1 ]). While the constraint answer ∃X∀H1 , L1 ∃Z1 , Z2 (A = [X, Z1 ] ∧ C = [X, Z2 ]∧B 6= [H1 , L1 ]) for union(A, B, C), approximates the concrete constraint A = [X, X], C = [X, X, K], B = K and B is not a list, computed by the concrete semantics. Note, in fact, that, if the second argument is not a list, the predicate member finitely fails. Let us now consider Marriott and Søndergaard’s abstraction for the program P , with a language where the only constant is a (this assumption does not affect the result). Concerning the predicate union with the empty list as first argument, their abstraction computes the following atoms union([ ], a, a), union([ ], [ ], [ ]), union([ ], [a], [a]), union([ ], [a, Z1 ], [a, Z2 ]), while we obtain the more precise answer (A = [ ] ∧ B = C)|union(A, B, C).

96

F. Fages and R. Gori

P : union(A, B, C) : −A = [ ], B = C. union(A, B, C) : −A = [X, L], C = [X, K], ¬member(X, B), union(L, B, K). union(A, B, C) : −A = [X, L], member(X, B), union(L, B, C). member(X, Y ) : −Y = [X, L]. member(X, Y ) : −Y = [H, L], member(X, L). Consider now a depth-2 analysis with Zi ∈ W . a+

∃L(

TPB Y = [X, L]

)|member(X, Y ).

∃H, Z1 (

Y = [H, Z1 ]

)|member(X, Y ).

A=[]∧B =C

|union(A, B, C).

∃X, L(

A = [X] ∧ B = [X, L] ∧ B = C

)|union(A, B, C).

∃X, Y, Z1 (

A = [X] ∧ B = [Y, Z1 ] ∧ B = C

)|union(A, B, C).

∃X, H, L, Z1 (

A = [X, Z1 ] ∧ B = [X, L] ∧ B = C

)|union(A, B, C).

∃X, H, Z1 , Z2 (

A = [X, Z1 ] ∧ B = [H, Z2 ] ∧ B = C

)|union(A, B, C).

∃X, K∀H, L( A = [X] ∧ C = [X, K] ∧ B 6= H [ , L] ∧ B = K )|union(A, B, C). ∃X∀H, L∃Z1 , Z2 (

A = [X, Z1 ] ∧ C = [X, Z2 ] ∧ B 6= H [ , L]

)|union(A, B, C).

∃X, K∀L( A = [X] ∧ C = [X, K] ∧ B 6= X [ , L] ∧ B = K )|union(A, B, C). ∃X∀L∃Z1 , Z2 (

A = [X, Z1 ] ∧ C = [X, Z2 ] ∧ B 6= X [ , L]

A subset of TPB

a−

)|union(A, B, C).

(complete for the predicate member)

∀H, L(

Y 6= H [ , L]

)|member(X, Y ).

∀L(

Y 6= X [ , L]

)|member(X, Y ).

∀X, L, KX1 , L1

((A 6= [ ]∧ A 6= X [ , L])∨ (B 6=C ∧ A 6= X [ , L]∨ (B 6=C ∧ C 6= X [ , K]) ∧ A 6= X [ 1 , L1 ]) )|union(A, B, C).

∀X, L, KX1 , L1 , H, L2

((A 6= [ ]∧ A 6= X [ , L])∨ (B 6=C ∧ A 6= X [ , L]∨ (B 6=C ∧ C 6= X [ , K]) ∧ A 6= X [ 1 , L1 ])∨ (B 6=C ∧ C 6= X [ , L] ∧ B 6= H [ , L2 ])∨ (A 6= [ ]∧ C 6= X [ , L] ∧ B 6= H [ , L2 ]) .. .

Fig. 1. Example 1

|)union(A, B, C)

Analysis of Normal Logic Programs

97

The atom union([ ], [a, Z1 ], [a, Z2 ]), in fact, correctly approximates the predicates deeper than k which have a successful behavior, but it has lost the relation between B and C. As a consequence all the other ground atoms for union computed using the atom union([ ], [a, Z1 ], [a, Z2 ]), are less precise than the ground instances of the atoms computed by our non-ground abstract semantics.

5

Conclusion

Starting from the hierarchy of semantics defined in [10], our aim was to show that well known analysis for logic programs could be extended to normal logic programs. Based on the framework of abstract interpretation [7,8], we have presented a depth(k) analysis which is able to approximate the answer set of normal logic programs. It is worth noting that our depth(k) analysis, can be easily generalized to constraint logic programs defined on H, whose program constraints can be conjunctions of equalities and inequalities. In order to deal with constructive negation, in fact, most of the results presented in this paper hold for first order equality constraints. The only exception is lemma 30 (and consequently theorem 33 and theorem 34), which is true only for conjunctions of equalities. But a more complex definition of the abstract not operator can be defined and proven correct on conjunctions of equalities and inequalities constraints. This alternative definition is, however, less precise than the one defined here. As a consequence theorem 33, where the abstract fixpoint operator uses the new abstract not operator, still holds for such “extended” logic programs, while it is not the case for theorem 34.

References 1. K. R. Apt. Introduction to Logic Programming. In J. van Leeuwen, editor, Handbook of Theoretical Computer Science, volume B: Formal Models and Semantics, pages 495–574. Elsevier and The MIT Press, 1990. 2. A. Bossi, M. Gabbrielli, G. Levi, and M. Martelli. The s-semantics approach: Theory and applications. Journal of Logic Programming, 19–20:149–197, 1994. 3. D. Chan. Constructive Negation Based on the Completed Database. In R. A. Kowalski and K. A. Bowen, editors, Proc. Fifth Int’l Conf. on Logic Programming, pages 111–125. The MIT Press, 1988. 4. A. Colmerauer. Equations and inequations on finite and infinite trees. In Proceedings of the International Conference on Fifth Generation Computer System, pages 85–99, 1984. 5. M. Comini. Abstract Interpretation framework for Semantics and Diagnosis of Logic Programs. PhD thesis, Dipartimento di Informatica, Universit` a di Pisa, 1998. 6. P. Cousot and R. Cousot. Abstract Interpretation: A Unified Lattice Model for Static Analysis of Programs by Construction or Approximation of Fixpoints. In Proc. Fourth ACM Symp. Principles of Programming Languages, pages 238–252, 1977.

98

F. Fages and R. Gori

7. P. Cousot and R. Cousot. Systematic Design of Program Analysis Frameworks. In Proc. Sixth ACM Symp. Principles of Programming Languages, pages 269–282, 1979. 8. P. Cousot and R. Cousot. Abstract Interpretation and Applications to Logic Programs. Journal of Logic Programming, 13(2 & 3):103–179, 1992. 9. F. Fages. Constructive negation by pruning. Journal of Logic Programming, 32(2):85–118, 1997. 10. F. Fages and R. Gori. A hierarchy of semantics for normal constraint logic programs. In M.Hanus M.Rodriguez-Artalejo, editor, Proc. Fifth Int’l Conf. on Algebraic and Logic Programming, volume 1139 of Lecture Notes in Computer Science, pages 77–91. Springer-Verlag, 1996. 11. M. Fitting. A Kripke-Kleene semantics for logic programs. Journal of Logic Programming, 2:295–312, 1985. 12. R. Giacobazzi. “Optimal” collecting semantics for analysis in a hierarchy of logic program semantics. In C. Puech and R. Reischuk, editors, Proc. 13th International Symposium on Theoretical Aspects of Computer Science (STACS’96), volume 1046 of Lecture Notes in Computer Science, pages 503–514. Springer-Verlag, 1996. 13. S. Koppelberg. Handbook of Boolean Algebras (Vol.I). Elsevier Science Publisher B.V.(North Holland), 1989. 14. K. Kunen. Negation in logic programming. Journal of Logic Programming, 4:289– 308, 1987. 15. K. Kunen. Signed Data Dependencies in Logic Programs. Journal of Logic Programming, 7(3):231–245, 1989. 16. J.-L. Lassez, M. J. Maher, and K. Marriott. Unification Revisited. In J. Minker, editor, Foundations of Deductive Databases and Logic Programming, pages 587– 625. Morgan Kaufmann, Los Altos, Ca., 1988. 17. J. W. Lloyd. Foundations of Logic Programming. Springer-Verlag, 1987. Second edition. 18. M.J. Maher. Complete axiomatizations of the algebra of finite, rational and infinite trees. In Third Symp. on Logic in Computer Science, pages 348–357, 1988. 19. K. Marriott and H. Sondergaard. Bottom-up Dataflow Analysis of Normal Logic Programs. Journal of Logic Programming, 13(2 & 3):181–204, 1992. 20. T. Sato and H. Tamaki. Enumeration of Success Patterns in Logic Programs. Theoretical Computer Science, 34:227–240, 1984. 21. P. Stuckey. Negation and constraint logic programming. Information and Computation, 118(1):12–33, 1995.

The Correctness of Set-Sharing Patricia M. Hill1 , Roberto Bagnara2? , and Enea Zaffanella3 1

School of Computer Studies, University of Leeds, Leeds, LS2 9JT, United Kingdom [email protected] 2 Dipartimento di Matematica, Universit` a degli Studi di Parma, Italy. [email protected] 3 Servizio IX Automazione, Universit` a degli Studi di Modena, Italy. [email protected]

Abstract. It is important that practical data flow analysers are backed by reliably proven theoretical results. Abstract interpretation provides a sound mathematical framework and necessary generic properties for an abstract domain to be well-defined and sound with respect to the concrete semantics. In logic programming, the abstract domain Sharing is a standard choice for sharing analysis for both practical work and further theoretical study. In spite of this, we found that there were no satisfactory proofs for the key properties of commutativity and idempotence that are essential for Sharing to be well-defined and that published statements of the safeness property assumed the occur-check. This paper provides a generalisation of the abstraction function for Sharing that can be applied to any language, with or without the occur-check. The results for safeness, idempotence and commutativity for abstract unification using this abstraction function are given. Keywords: abstract interpretation, logic programming, occur-check, rational trees, set-sharing.

1

Introduction

Today, talking about sharing analysis for logic programs is almost the same as talking about the set-sharing domain Sharing of Jacobs and Langen [8, 9]. Researchers are primarily concerned with extending the domain with linearity, freeness, depth-k abstract substitutions and so on [2, 4, 12, 13, 16]. Key properties such as commutativity and soundness of this domain and its associated abstract operations are normally assumed to hold. The main reason for this is that [9] not only includes a proof of the soundness but also refers the reader to the thesis of Langen [14] for proofs of commutativity and idempotence. In abstract interpretation, the concrete semantics of a program is approximated by an abstract semantics. In particular, the concrete domain is replaced ?

Much of this work was supported by EPSRC grant GR/L19515.

G. Levi (Ed.): SAS’98, LNCS 1503, pp. 99−114, 1998. Springer-Verlag Berlin Heidelberg 1998

100

P.M. Hill, R. Bagnara, and E. Zaffanella

by an abstract domain and each elementary operation on the concrete domain is replaced by a corresponding abstract operation on the abstract domain. Thus, assuming the global abstract procedure mimics the concrete execution procedure, each operation on elements in the abstract domain must produce an approximation of the corresponding operation on corresponding elements in the concrete domain. The key operation in a logic programming derivation is unification (unify) and the corresponding operation for an abstract domain is aunify. An important step in standard unification algorithms is the occur-check that avoids the generation of infinite data structures. However, in computational terms, it is expensive and it is well known that Prolog implementations by default omit this check. Although standard unification algorithms that include the occur-check produce a substitution that is idempotent, the resulting substitution when the occur-check is omitted, may not be idempotent. In spite of this, most theoretical work on data-flow analysis of logic programming assume the result of unify is always idempotent. In particular both [9] and [14] assume in their proofs of soundness that the concrete substitutions are idempotent. Thus their results do not apply to the analysis of all Prolog programs. If two terms in the concrete domain are unifiable, then unify computes the most general unifier (mgu). Up to renaming of variables, an mgu is unique. Moreover a substitution is defined as a set of bindings or equations between variables and other terms. Thus, for the concrete domain, the order and multiplicity of elements are irrelevant in both the computation and semantics of unify. It is therefore useful that the abstraction of the unification procedure should be unaffected by the order and multiplicity in which it abstracts the bindings that are present in the substitution. Furthermore, from a practical perspective, it is useful if the global abstract procedure can proceed in a different order to the concrete one without affecting the accuracy of the analysis results. Hence, it is extremely desirable that aunify is also commutative and idempotent. However, as discussed later in this paper, only a weak form of idempotence has ever been proved while the only previous proof of commutativity [14] is seriously flawed. As sharing is normally combined with linearity and freeness domains that are not idempotent or commutative, [2, 12] it may be asked why these properties are important for sharing. In answer to this, we observe that the order and multiplicity in which the bindings in a substitution are analysed affects the accuracy of the linearity and freeness domains. It is therefore a real advantage to be able to ignore these aspects as far as the sharing domain is concerned. This paper provides a generalisation of the abstraction function for Sharing that can be applied to any language, with or without the occur-check. The results for safeness, idempotence and commutativity for abstract unification using this abstraction function are given. Detailed proofs of the results stated in this paper are available in [7]. In the next section, the notation and definitions needed for equality and substitutions in the concrete domain are given. In Section 3, we introduce a new concept called variable-idempotence that generalises idempotence to allow for rational trees. In Section 4, we recall the definition of Sharing and define its

The Correctness of Set-Sharing

101

abstraction function, generalised to allow for non-idempotent substitutions. We conclude in Section 5.

2 2.1

Equations and Substitutions Notation

For a set S, # S is the cardinality of S, ℘(S) is the powerset of S, whereas ℘f (S) is the set of all the finite subsets of S. The symbol Vars denotes a denumerable set of variables, whereas TVars denotes the set of first-order terms over Vars for some given set of function symbols. The set of variables occurring in a syntactic object o is denoted by vars(o). 2.2

Substitutions

If x ∈ Vars and s ∈ TVars , then x 7→s is called a binding. A substitution is a total function σ : Vars → TVars that is the identity almost everywhere; in other words, the domain of σ, def dom(σ) = x ∈ Vars σ(x) 6= x is finite. If t ∈ TVars , we write tσ to denote σ(t). Substitutions by the set of are denoted their bindings, thus σ is identified with the set x 7→ σ(x) x ∈ dom(σ) . The composition of substitutions is defined in the usual way. Thus τ ◦ σ is the substitution such that, for all terms t, (τ ◦ σ)(t) = τ (σ(t)). A substitution is said circular if it has the form {x1 7→x2 , . . . , xn−1 7→xn , xn 7→x1 }. A substitution is in rational solved form if it has no circular subset. The set of all substitutions in rational solved form is denoted by Subst. 2.3

Equations

An equation is of the form s = t where s, t ∈ TVars . Eqs denotes the set of all equations. We are concerned in this paper to keep the results on sharing as general as possible. In particular, we do not want to restrict ourselves to a specific equality theory. Thus we allow for any equality theory T over TVars that includes the basic axioms denoted by the following schemata. s = s, s = t ⇐⇒ t = s, r = s ∧ s = t =⇒ r = t, f (s1 , . . . , sn ) = f (t1 , . . . , tn ) ⇐⇒ s1 = t1 , . . . , sn = tn .

(1) (2) (3) (4)

Of course, T can include other axioms. For example, it is usual in logic programming and most implementations of Prolog to assume an equality theory

102

P.M. Hill, R. Bagnara, and E. Zaffanella

based on syntactic identity and characterised by the axiom schemata given by Clark [3]. This consists of the basic axioms together with the following: ¬f (s1 , . . . , sn ) = g(t1 , . . . , tm ) ∀z ∈ Vars ∀t ∈ (TVars \ Vars) : z ∈ vars(t) =⇒ ¬(z = t).

(5) (6)

The identity axioms characterised by the schemata 5 ensure the equality theory is Herbrand and depends only on the syntax. Equality theory for a non-Herbrand domain replaces these axioms by ones that depend instead on the semantics of the domain. Axioms characterised by the schemata 6 are called the occur-check axioms and are an essential part of the standard unification procedure in SLDresolution. An alternative approach used in some implementations of Prolog, does not require the occur-check axioms. This approach is based on the theory of rational trees [5, 6]. It assumes the basic axioms and the identity axioms together with a set of uniqueness axioms [10, 11]. These state that each equation in rational solved form uniquely defines a set of trees. Thus, an equation z = t where z ∈ vars(t) and t ∈ (TVars \ Vars) denotes the axiom (expressed in terms of the usual first-order quantifiers [15]): ∀x ∈ Vars : z = t ∧ (x = t{z 7→x} =⇒ z = x) . The basic axioms defined by schemata 1, 2, 3, and 4, which are all that are required for the results in this paper, are included in both these theories. A substitution σ may be regarded as a set of equations { x = t | x 7→t ∈ σ }. A set of equations e ∈ ℘f (Eqs) is unifiable if there is σ ∈ Subst such that T ` (σ =⇒ e). σ is called a unifier for e. σ is said to be a relevant unifier of e if vars(σ) ⊆ vars(e). That is, σ does not introduce any new variables. σ is a most general unifier for e if, for every unifier σ 0 of e, T ` (σ 0 =⇒ σ). An mgu, if it exists, is unique up to the renaming of variables. In this paper, mgu(e) always denotes a relevant unifier of e.

3

Variable-Idempotence

It is usual in papers on sharing analysis to assume that all the substitutions are idempotent. Note that a substitution σ is idempotent if, for all t ∈ TVars , tσσ = tσ. However, the sharing domain is just concerned with the variables. So, to allow for substitutions representing rational trees, we generalise idempotence to variable-idempotence. Definition 1. A substitution σ is variable-idempotent if ∀t ∈ TVars : vars(tσσ) = vars(tσ). The set of all variable-idempotent substitutions is denoted by VSubst.

The Correctness of Set-Sharing

103

It is convenient to use the following alternative characterisation of variableidempotence: A substitution σ is variable-idempotent if and only if, ∀(x 7→t) ∈ σ : vars(tσ) = vars(t). Thus any substitution consisting of a single binding is variable-idempotent. Moreover, all idempotent substitutions are also variable-idempotent. Example 1. The substitution x 7→ f (x) is not idempotent but is variable idempotent. Also, x 7→f (y), y 7→z is not idempotent or variable-idempotent but is equivalent (with respect to some equality theory T ) to x 7→f (z), y 7→z , which is idempotent. S

We define the transformation 7−→ ⊆Subst × Subst, called S-transformation, as follows: (x 7→t) ∈ σ

(y 7→s) ∈ σ x 6= y σ 7−→ σ \ {y 7→s} ∪ {y 7→s[x/t]} S

Any substitution σ can be transformed to a variable-idempotent substitution σ 0 for σ by a finite sequence of S-transformations. Furthermore, if the substitutions σ and σ 0 are regarded as equations, then they are equivalent with respect to any equality theory that includes the basic equality axioms. These two statements are direct consequences of Lemmas 1 and 2, respectively. Lemma 1. Let T be an equality theory that satisfies the basic equality axioms. S Suppose that σ and σ 0 are substitutions such that σ 7−→σ 0 . Then, regarding σ 0 0 and σ as sets of equations, T ` (σ ⇐⇒ σ ). Proof. Suppose that (x 7→ t), (y 7→ s) ∈ σ where x 6= y and suppose also σ 0 = σ \ {y 7→s} ∪ {y 7→s[x/t]}. We first show by induction on the depth of the term s that x = t =⇒ s = s[x/t]. Suppose s has depth 1. If s is x, then s[x/t] = t and the result is trivial. If s is a variable distinct from x or a constant, then s[x/t] = s and the result follows from equality Axiom 1. Suppose now that s = f (s1 , . . . , sn ) and the result holds for all terms of depth less than that of s. Then, by the inductive hypothesis, for each i = 1, . . . , n, x = t =⇒ si = si [x/t]. Hence, by Axiom 4, x = t =⇒ f (s1 , . . . , sn ) = f s1 [x/t], . . . , sn [x/t] and hence x = t =⇒ f (s1 , . . . , sn ) = f (s1 , . . . , sn )[x/t].

104

P.M. Hill, R. Bagnara, and E. Zaffanella

Thus, combining this result with Axiom 3, we have {x = t, y = s} =⇒ x = t, y = s, s = s[x/t] =⇒ x = t, y = s[x/t] . Similarly, combining this result with Axioms 2 and 3, x = t, y = s[x/t] =⇒ x = t, y = s[x/t], s = s[x/t] =⇒ {x = t, y = s}. t u Note is necessary. For example, suppose that the conditionx 6= y in Lemma1 σ = x 7→f (x) and σ 0 = x 7→f (f (x)) . Then we do not have σ 0 =⇒ σ. Lemma 2. Suppose that, for each j = 0, . . . , n: σj = {x1 7→t1,j , . . . , xn 7→tn,j }, where tj,j = tj,j−1 and if j > 0, for each i = 1, . . . , n, where i 6= j, ti,j = ti,j−1 [xj /tj,j−1 ]. Then, for each j = 0, . . . , n, νj = {x1 7→t1,j , . . . , xj 7→tj,j } is variable-idempotent and, if j > 0, σj can be obtained from σj−1 by a sequence of S-transformations. Proof. The proof is by induction on j. Since ν0 is empty, the base case when j = 0 is trivial. Suppose, therefore that 1 ≤ j ≤ n and the hypothesis holds for νj−1 and σj−1 . By the definition of νj , we have νj = {xj 7→tj,j−1 } ◦ νj−1 . Consider an arbitrary i, 1 ≤ i ≤ j. We will show that vars(ti,j νj ) = vars(ti,j ). Suppose first that i = j. Then since tj,j = tj,j−1 , tj,j−1 = tj,0 νj−1 and, by the inductive hypothesis, vars(tj,0 νj−1 νj−1 ) = vars(tj,0 νj−1 ), we have vars(tj,j νj ) = vars tj,0 νj−1 νj−1 {xj 7→tj,j } = vars tj,0 νj−1 {xj 7→tj,j } = vars tj,j {xj 7→tj,j } = vars(tj,j ). Suppose now that i 6= j. Then,

vars(ti,j ) = vars ti,j−1 {xj 7→tj,j−1 } .

and, by the inductive hypothesis, vars(ti,j−1 νj−1 ) = vars(ti,j−1 ). If xj ∈ / vars(ti,j−1 ), then vars(ti,j νj−1 ) = vars ti,j−1 {xj 7→tj,j−1 }νj−1 = vars(ti,j−1 νj−1 ) = vars(ti,j ).

The Correctness of Set-Sharing

105

On the other hand, if xj ∈ vars(ti,j−1 ), then vars(ti,j νj−1 ) = vars ti,j−1 {xj 7→tj,j−1 }νj−1 = vars(ti,j−1 νj−1 ) \ {xj } ∪ vars(tj,j−1 νj−1 ) = vars(ti,j−1 ) \ {xj } ∪ vars(tj,j−1 ) = vars ti,j−1 {xj 7→tj,j−1 } = vars(ti,j ). Thus, in both cases, vars(ti,j νj ) = vars ti,j νj−1 {xj 7→tj,j−1 } = vars ti,j {xj 7→tj,j−1 }

= vars(ti,j−1 {xj 7→tj,j−1 }{xj 7→tj,j−1 } .

However, a substitution consisting of a single binding is variable-idempotent. Thus vars(ti,j νj ) = vars ti,j−1 {xj 7→tj,j−1 } = vars(ti,j ).

Therefore, for each i = 1, . . . , j, vars(ti,j νj ) = vars(ti,j ). It then follows (using the alternative characterisation of variable-idempotence) that νj is variableidempotent. t u Example 2. Let σ0 = x1 7→f (x2 ), x2 7→g(x3 , x4 ), x3 7→x1 . Then σ1 = x1 7→f (x2 ), x2 7→g(x3 , x4 ), x3 7→f (x2 ) , σ2 = x1 7→f (g(x3 , x4 )), x2 7→g(x3 , x4 ), x3 7→f (g(x3 , x4 )) , σ3 = x1 7→f (g(f (g(x3 , x4 )), x4 )), x2 7→g(f (g(x3 , x4 )), x4 ), x3 7→f (g(x3 , x4 )) . Note that σ3 is variable-idempotent and that T ` σ0 ⇐⇒ σ3 .

4 4.1

Set-Sharing The Sharing Domain

The Sharing domain is due to Jacobs and Langen [8]. However, we use the definition as presented in [1].

106

P.M. Hill, R. Bagnara, and E. Zaffanella

Definition 2. (The set-sharing lattice.) Let def SG = S ∈ ℘f (Vars) S 6= ∅ def

and let SH = ℘(SG). The set-sharing lattice is given by the set def

SS =

(sh, U ) sh ∈ SH , U ∈ ℘f (Vars), ∀S ∈ sh : S ⊆ U ∪ {⊥, >}

ordered by SS defined as follows, for each d, (sh 1 , U1 ), (sh 2 , U2 ) ∈ SS : ⊥ SS d, d SS >, (sh 1 , U1 ) SS (sh 2 , U2 )

⇐⇒

(U1 = U2 ) ∧ (sh 1 ⊆ sh 2 ).

It is straightforward to see that every subset of SS has a least upper bound with respect to SS . Hence SS is a complete lattice.1 An element sh of SH abstracts the property of sharing in a substitution σ. That is, if σ is idempotent, two variables x, y must be in the same set in sh if some variable, say v occurs in both xσ and yσ. In fact, this is also true for variable-idempotent substitutions although it is shown below that this needs to be generalised for substitutions that are not variable-idempotent. Thus, the definition of the abstraction function α for sharing, requires an ancillary definition for the notion of occurrence. Definition 3. (Occurrence.) For each n ∈ N, occi : Subst × Vars → ℘f (Vars) is defined for each σ ∈ Subst and each v ∈ Vars: def

occ0 (σ, v) = {v},

if v = vσ;

def

occ0 (σ, v) = ∅, def occn (σ, v) = y ∈ Vars x ∈ vars(yσ) ∩ occn−1 (σ, v) ,

if v 6= vσ; if n > 0.

It follows that, for fixed values of σ and v, occn (σ, v) is monotonic and extensive with respect to the index n. Hence, as the range of occn (σ, v) is restricted to the finite set of variables in σ, there is an ` = `(σ, v) ∈ N such that occ` (σ, v) = occn (σ, v)) for all n ≥ `. Let def

occ!(σ, v) = occ` (σ, v). Note that if σ is variable-idempotent, then occ!(σ, v) = occ1 (σ, v). Note also that if v 6= vσ, then occ!(σ, v) = ∅. Previous definitions for an occurrence operator such as that for sg in [8] have all been for idempotent substitutions. However, when σ is an idempotent substitution, occ!(σ, v) and sg(σ, v) are the same for all v ∈ Vars. We base the definition of abstraction on the occurrence operator, occ!. 1

Notice that the only reason we have > ∈ SS is in order to turn SS into a lattice rather than a CPO.

The Correctness of Set-Sharing

107

Definition 4. (Abstraction.) The concrete domain Subst is related to SS by means of the abstraction function α : ℘(Subst) × ℘f (Vars) → SS . For each Σ ∈ ℘(Subst) and each U ∈ ℘f (Vars), G def α(Σ, U ) = α(σ, U ), σ∈Σ

where α : Subst × ℘f (Vars) → SS is defined, for each σ ∈ Subst and each U ∈ ℘f (Vars), by def

α(σ, U ) =

occ!(σ, v) ∩ U v ∈ Vars \ {∅}, U .

The following result states that the abstraction for a substitution σ is the same as the abstraction for a variable-idempotent substitution for σ. Lemma 3. Let σ be a substitution, σ 0 a substitution obtained from σ by a sequence of S-transformations, U a set of variables and v ∈ Vars. Then v = vσ ⇐⇒ v = vσ 0 ,

occ!(σ, v) = occ!(σ 0 , v),

and α(σ, U ) = α(σ 0 , U ).

Proof. Suppose first that σ 0 is obtained from σ by a single S-transformation. Thus we can assume that x 7→t and y 7→s are in σ where x ∈ vars(s) and that σ 0 = σ \ {y 7→s} ∪ y 7→s[x/t] . It follows that, since σ is in rational solved form, σ has no circular subset and hence v = vσ ⇐⇒ v = vσ 0 . Thus, if v 6= vσ, then we have v 6= vσ 0 and occ!(σ, v) = occ!(σ 0 , v) = ∅. We now assume that v = vσ = vσ 0 and prove that occm (σ, v) ⊆ occ!(σ 0 , v). The proof is by induction on m. By Definition 3, occ0 (σ, v) = occ0 (σ 0 , v) = {v}, so that the result holds for m = 0. Suppose then that m > 0 and that vm ∈ occm (σ, v). By Definition 3, there exists vm−1 ∈ vars(vm σ) where vm−1 ∈ 0 occm−1 (σ, v). Hence, by the inductive hypothesis, vm−1 ∈ occ!(σ , v). If vm−1 ∈ vars(vm σ 0 ), then, by Definition 3, vm ∈ occ!(σ 0 , v) . On the other hand, if vm−1 ∈ / vars(vm σ 0 ), then vm = y, vm−1 = x, and x ∈ vars(s) (so that vars(t) ⊆ vars(s[x/t])). However, by hypothesis, v = vσ, so that x 6= v and m > 1. Thus, by Definition 3, there exists vm−2 ∈ vars(t) such that vm−2 ∈ occm−2 (σ, v). By the inductive hypothesis, vm−2 ∈ occ!(σ 0 , v). Since y 7→ s[x/t] ∈ σ 0 , and vm−2 ∈ vars(s[x/t]), vm−2 ∈ vars(yσ 0 ). Thus, by Definition 3, y ∈ occ!(σ 0 , v). Conversely, we now prove that, for all m, occm (σ 0 , v) ⊆ occ!(σ, v). The proof is again by induction on m. As in the previous case, occ0 (σ 0 , v) = occ0 (σ, v) = {v}, so that the result holds for m = 0. Suppose then that m > 0 and

108

P.M. Hill, R. Bagnara, and E. Zaffanella

that vm ∈ occm (σ 0 , v). By Definition 3, there exists vm−1 ∈ vars(vm σ 0 ) where vm−1 ∈ occm−1 (σ 0 , v). Hence, by the inductive hypothesis, vm−1 ∈ occ!(σ, v). If vm ∈ occ(σ, vm−1 ), then, by Definition 3, vm ∈ occ!(σ, v). On the other hand, if vm−1 ∈ / vars(vm σ), then vm = y, vm−1 ∈ vars(t) and x ∈ vars(s). Thus, as y 7→s ∈ σ, y ∈ vars(xσ). However, since x 7→t ∈ σ, vm−1 ∈ vars(xσ) so that, by Definition 3, x ∈ occ!(σ, v). Thus, again by Definition 3, y ∈ occ!(σ, v). Thus, if σ 0 is obtained from σ by a single S-transformation, we have the required results: v = vσ ⇐⇒ v = vσ 0 , occ!(σ, v) = occ!(σ 0 , v), and α(σ, U ) = α(σ 0 , U ). Suppose now that there is a sequence σ = σ1 , . . . , σn = σ 0 such that, for i = 2, . . . , n, σi is obtained from σi−1 by a single S-step. If n = 1, then σ = σ 0 . If n > 1, we have by the first part of the proof that, for each i = 2, . . . , n, v = vσi−1 ⇐⇒ v = vσi , occ!(σi−1 , v) = occ!(σi , v), and α(σi−1 , U ) = α(σi , U ), and hence the required results. t u Example 3. Consider again Example 2. Then occ1 (σ0 , x4 ) = {x2 , x4 }, occ2 (σ0 , x4 ) = {x1 , x2 , x4 }, occ3 (σ0 , x4 ) = {x1 , x2 , x3 , x4 } = occ!(σ0 , x4 ), and occ1 (σ3 , x4 ) = {x1 , x2 , x3 , x4 } = occ!(σ3 , x4 ). Thus, if V = {x1 , x2 , x3 , x4 },

α(σ0 , V ) = α(σ3 , V ) = {x1 , x2 , x3 , x4 } .

4.2

Abstract Operations for Sharing Sets

We are concerned in this paper in establishing results for the abstract operation aunify which is defined for arbitrary sets of equations. However, by building the definition of aunify in three steps via the definitions of amgu (for sharing sets) and Amgu (for sharing domains) and stating corresponding results for each of them, we provide an outline for the overall method of proof for the aunify results. Details of all proofs are available in [7]. In order to define the abstract operation amgu we need some ancillary definitions. Definition 5. (Auxiliary functions.) The closure under union function (also called star-union), (·)? : SH → SH , is, for each sh ∈ SH , def sh ? = S ∈ SG ∃n ≥ 1 . ∃T1 , . . . , Tn ∈ sh . S = T1 ∪ · · · ∪ Tn . For each sh ∈ SH and each T ∈ ℘f (Vars), the extraction of the relevant component of sh with respect to T is encoded by the function rel : ℘f (Vars)×SH → SH defined as def

rel(T, sh) = { S ∈ sh | S ∩ T 6= ∅ }.

The Correctness of Set-Sharing

109

For each sh 1 , sh 2 ∈ SH , the binary union function bin : SH × SH → SH is given by def

bin(sh 1 , sh 2 ) = { S1 ∪ S2 | S1 ∈ sh 1 , S2 ∈ sh 2 }. The function proj : SH × ℘f (Vars) → SH projects an element of SH onto a set of variables of interest: if sh ∈ SH and V ∈ ℘f (Vars), then def

proj(sh, V ) = { S ∩ V | S ∈ sh, S ∩ V 6= ∅ }. Definition 6. (amgu.) The function amgu captures the effects of a binding x 7→ t on an SH element. Let x be a variable and t a term. Let also sh ∈ SH and def A = rel {x}, sh ,

def B = rel vars(t), sh . Then def amgu(sh, x 7→t) = sh \ (A ∪ B) ∪ bin(A? , B ? ). Then we have the following soundness result for amgu. Lemma 4. Let (sh, U ) ∈ SS and {x 7→t}, σ, ν ∈ Subst such that ν is a relevant unifier of {xσ = tσ} and vars(x), vars(t), vars(σ) ⊆ U . Then α(σ, U ) SS (sh, U ) =⇒ α(ν ◦ σ, U ) SS (amgu(sh, x 7→t), U ). To prove this, observe that, by Lemma 2, if σ is not variable-idempotent, it can be transformed to a variable-idempotent substitution σ 0 . Hence, by Lemma 3, α(σ, U ) = α(σ 0 , U ). Therefore, the proof, which is given in [7], deals primarily with the case when σ is variable-idempotent. Since a relevant unifier of e is a relevant unifier of any other set e0 equivalent to e wrt to the equality theory T , this lemma shows that it is safe for the analyser to perform part or all of the concrete unification algorithm before computing amgu. The following lemmas, proved in [7], show that amgu is commutative and idempotent. Lemma 5. Let sh ∈ SH and {x 7→r} ∈ Subst. Then amgu(sh, x 7→r) = amgu amgu(sh, x 7→r), x 7→r . Lemma 6. Let sh ∈ SH and {x 7→r}, {y 7→t} ∈ Subst. Then amgu amgu(sh, x 7→r), y 7→t = amgu amgu(sh, y 7→t), x 7→r .

110

P.M. Hill, R. Bagnara, and E. Zaffanella

4.3

Abstract Operations for Sharing Domains

The definitions and results of Subsection 4.2 can be lifted to apply to sharing domains. Definition 7. (Amgu.) The operation Amgu : SS × Subst → SS extends the SS description it takes as an argument, to the set of variables occurring in the binding it is given as the second argument. Then it applies amgu: Amgu (sh, U ), x 7→t def = amgu sh ∪ {u} u ∈ vars(x 7→t) \ U , x 7→t , U ∪ vars(x 7→t) . The results for amgu can easily be extended to apply to Amgu. Definition 8. (aunify.) The function aunify : SS ×Eqs → SS generalises Amgu to a set of equations e: If (sh, U ) ∈ SS , x is a variable, r is a term, s = f (s1 , . . . , sn ) and t = f (t1 , . . . , tn ) are non-variable terms, and s = t denote the set of equations {s1 = t1 , . . . , sn = tn }, then def

aunify((sh, U ), ∅) = (sh, U ), if e ∈ ℘f (Eqs) is unifiable, def aunify (sh, U ), e ∪ {x = r} = aunify Amgu(sh, U ), x 7→r), e \ {x = r} , def aunify (sh, U ), e ∪ {s = x} = aunify (sh, U ), (e \ {s = x}) ∪ {x = s} , def aunify (sh, U ), e ∪ {s = t} = aunify (sh, U ), (e \ {s = t}) ∪ s = t , and, if e is not unifiable, def

aunify((sh, U ), e) = ⊥. For the distinguished elements ⊥ and > of SS def aunify ⊥, e = ⊥,

def aunify >, e = >.

As a consequence of this and the generalisation of Lemmas 4, 5 and 6 to Amgu, we have the following soundness, commutativity and idempotence results required for aunify to be sound and well-defined. As before, the proofs of these results are in [7]. Theorem 1. Let (sh, U ) ∈ SS , σ, ν ∈ Subst, and e ∈ ℘f (Eqs) be such that vars(σ) ⊆ U and ν is a relevant unifier of e. Then α(σ, U ) SS (sh, U ) =⇒ α(ν ◦ σ, U ) SS aunify((sh, U ), e).

The Correctness of Set-Sharing

111

Theorem 2. Let (sh, U ) ∈ SS and e ∈ ℘f (Eqs). Then aunify (sh, U ), e = aunify aunify (sh, U ), e , e . Theorem 3. Let (sh, U ) ∈ SS and e1 , e2 ∈ ℘f (Eqs). Then aunify aunify (sh, U ), e1 , e2 = aunify aunify (sh, U ), e2 , e1 .

5

Discussion

The SS domain which was first defined by Langen [14] and published by Jacobs and Langen [8] is an important domain for sharing analysis. In this paper, we have provided a framework for analysing non-idempotent substitutions and presented results for soundness, idempotence and commutativity of aunify. In fact, most researchers concerned with analysing sharing and related properties using the SS domain, assume these properties hold. Why therefore are the results in this paper necessary? Let us consider each of the above properties one at a time. 5.1

Soundness

We have shown that, for any substitution σ over a set of variables U , the abstraction α(σ, U ) = (sh, U ) is unique (Lemma 3) and the aunify operation is sound (Theorem 1). Note that, in Theorem 1, there are no restrictions on σ; it can be non-idempotent, possibly including cyclic bindings (that is, bindings where the domain variable occurs in its co-domain). Thus this result is widely applicable. Previous results on sharing have assumed that substitutions are idempotent. This is true if equality is syntactic identity and the implementation uses a unification algorithm based on that of Robinson [17] which includes the occur-check. With such algorithms, the resulting unifier is both unique and idempotent. Unfortunately, this is not what is implemented by most Prolog systems. In particular, if the algorithm is as described in [11] and used in Prolog III [5], then the resulting unifier is in rational solved form. This algorithm does not generate idempotent or even variable-idempotent substitutions even when the occur-check would never have succeeded. However, it has been shown that the substitution obtained in this way uniquely defines a system of rational trees [5]. Thus our results show that its abstraction using α, as defined in this paper, is also unique and that aunify is sound. Alternatively, if, as in most commercial Prolog systems, the unification algorithm is based on the Martelli-Montanari algorithm, but omits the occur check step, then the resulting substitution may not be idempotent. Consider the following example. Suppose we are given as input the equation p(z, f (x, y)) = p(f (z, y), z) with an initial substitution that is empty. We apply the steps in Martelli-Montanari procedure but without the occur-check:

112

P.M. Hill, R. Bagnara, and E. Zaffanella

1 2 3 4 5 6

equations p(z, f (x, y)) = p(f (z, y), z) z = f (z, y), f (x, y) = z f (x, y) = f (z, y) x = z, y = y y=y ∅

substitution ∅ ∅ {z 7→f (z, y)} {z 7→f (z, y)} {z 7→f (z, y), x 7→z} {z 7→f (z, y), x 7→z}

Note that we have used three kinds of steps here. In lines 1 and 3, neither argument of the selected equation is a variable. In this case, the outer nonvariable symbols (when, as in this example, they are the same) are removed and new equations are formed between the corresponding arguments. In lines 2 and 4, the selected equation has the form v = t, where v is a variable and t is not identical to v, then every occurrence of v is replaced by t in all the remaining equations and the range of the substitution. v 7→t is then added to the substitution. In line 5, the identity is removed. Let σ = {z 7→f (z, y), x 7→z}, be the computed substitution. Then, we have vars(xσ) = vars(z) = {z}, vars(xσ 2 ) = vars(f (z, y)) = {y, z}. Hence σ is not variable-idempotent. We conjecture that the resulting substitution is still unique (up to variable renaming). In this case our results can be applied so that its abstraction using α, as defined in this paper, is also unique and aunify is sound. 5.2

Idempotence

Definition 8 defines aunify inductively over a set of equations, so that it is important for this definition that aunify is both idempotent and commutative. The only previous result concerning the idempotence of aunify is given in thesis of Langen [14, Theorem 32]. However, the definition of aunify in [14] includes the renaming and projection operations and, in this case, only a weak form of idempotence holds. In fact, for the basic aunify operation as defined here and without projection and renaming, idempotence has never before been proven. 5.3

Commutativity

In the thesis of Langen the “proof” of commutativity of amguhas a number of omissions and errors [14, Lemma 30]. We highlight here, one error which we were unable to correct in the context of the given proof. To make it easier to compare, we adapt our notation and, define amge only in the case that a is a variable: def

amge(a, b, sh) = amgu(sh, a 7→b).

The Correctness of Set-Sharing

113

To prove the lemma, it has to show that: amge(a2 , b2 amge(a1 , b1 , sh)) = amge(a1 , b1 , amge(a2 , b2 , sh)). holds when a1 and a2 are variables. This corresponds to “the second base case” of the proof. We use Langen’s terminology: – A set of variables X is at a term t iff var(t) ∩ X 6= ∅. – A set of variables X is at i iff X is at ai or bi . – A union X ∪i Y is of Type i iff X is at ai and Y is at bi . def

def

Let lhs = amge(a2 , b2 , amge(a1 , b1 , S)), and rhs = amge(a1 , b1 , amge(a2 , b2 , S)). def

Let also Z ∈ lhs and T = aunify(a1 , b1 , S). Consider the case when Z = X ∪2 Y where X ∈ rel(a2 , T ), Y ∈ rel(b2 , T ), X = U ∪1 V where U ∈ rel(a1 , sh), V ∈ rel(b1 , sh) and U ∩ (vars(a2 ) ∪ vars(b2 )) = ∅ (that is, U is not at 2). Then the following quote [14, page 53, line 23] applies: In this case (U ∪1 V ) ∪2 Y = U ∪1 (V ∪2 Y ). By the inductive assumption V ∪2 Y is in the rhs and therefore so is Z. We give a counter-example to the statement “V ∪2 Y is in the rhs”. Suppose a1 , b1 , a2 , b2 are variables. We let each of a1 , b1 , a2 , b2 denote both the actual variable and the singleton set containing that variable. Suppose sh = {a1 , b1 a2 , b2 }. Then, from the definition of amge, lhs = {a1 b1 a2 b2 },

rhs = {a1 b1 a2 b2 },

T = {a1 b1 a2 , b2 }.

Let Z = a1 b1 a2 b2 , X = a1 b1 a2 , Y = b2 , U = a1 , V = b1 a2 . All the above conditions. However V ∪2 Y = b1 a2 b2 and this is not in {a1 b1 a2 b2 }.

References 1. R. Bagnara, P. M. Hill, and E. Zaffanella. Set-sharing is redundant for pair-sharing. In P. Van Hentenryck, editor, Static Analysis: Proceedings of the 4th International Symposium, volume 1302 of Lecture Notes in Computer Science, pages 53–67, Paris, France, 1997. Springer-Verlag, Berlin. 2. M. Bruynooghe and M. Codish. Freeness, sharing, linearity and correctness — All at once. In P. Cousot, M. Falaschi, G. Fil´e, and A. Rauzy, editors, Static Analysis, Proceedings of the Third International Workshop, volume 724 of Lecture Notes in Computer Science, pages 153–164, Padova, Italy, 1993. Springer-Verlag, Berlin. An extended version is available as Technical Report CW 179, Department of Computer Science, K.U. Leuven, September 1993. 3. K. L. Clark. Negation as failure. In H. Gallaire and J. Minker, editors, Logic and Databases, pages 293–322, Toulouse, France, 1978. Plenum Press.

114

P.M. Hill, R. Bagnara, and E. Zaffanella

4. M. Codish, D. Dams, G. Fil´e, and M. Bruynooghe. Freeness analysis for logic programs-and correctness? In D. S. Warren, editor, Logic Programming: Proceedings of the Tenth International Conference on Logic Programming, MIT Press Series in Logic Programming, pages 116–131, Budapest, Hungary, 1993. The MIT Press. An extended version is available as Technical Report CW 161, Department of Computer Science, K.U. Leuven, December 1992. 5. A. Colmerauer. Prolog and Infinite Trees. In K. L. Clark and S. ˚ A. T¨ arnlund, editors, Logic Programming, APIC Studies in Data Processing, volume 16, pages 231–251. Academic Press, New York, 1982. 6. A. Colmerauer. Equations and inequations on finite and infinite trees. In Proceedings of the International Conference on Fifth Generation Computer Systems (FGCS’84), pages 85–99, Tokyo, Japan, 1984. ICOT. 7. P. M. Hill, R. Bagnara, and E. Zaffanella. The correctness of set-sharing. Technical Report 98.03, School of Computer Studies, University of Leeds, 1998. 8. D. Jacobs and A. Langen. Accurate and efficient approximation of variable aliasing in logic programs. In E. L. Lusk and R. A. Overbeek, editors, Logic Programming: Proceedings of the North American Conference, MIT Press Series in Logic Programming, pages 154–165, Cleveland, Ohio, USA, 1989. The MIT Press. 9. D. Jacobs and A. Langen. Static analysis of logic programs for independent AND parallelism. Journal of Logic Programming, 13(2&3):291–314, 1992. 10. J. Jaffar, J-L. Lassez, and M. J. Maher. Prolog-II as an instance of the logic programming scheme. In M. Wirsing, editor, Formal Descriptions of Programming Concepts III, pages 275–299. North Holland, 1987. 11. T. Keisu. Tree Constraints. PhD thesis, The Royal Institute of Technology, Stockholm, Sweden, May 1994. Also available in the SICS Dissertation Series: SICS/D– 16–SE. 12. A. King. A synergistic analysis for sharing and groundness which traces linearity. In D. Sannella, editor, Proceedings of the Fifth European Symposium on Programming, volume 788 of Lecture Notes in Computer Science, pages 363–378, Edinburgh, UK, 1994. Springer-Verlag, Berlin. 13. A. King and P. Soper. Depth-k sharing and freeness. In P. Van Hentenryck, editor, Logic Programming: Proceedings of the Eleventh International Conference on Logic Programming, MIT Press Series in Logic Programming, pages 553–568, Santa Margherita Ligure, Italy, 1994. The MIT Press. 14. A. Langen. Static Analysis for Independent And-Parallelism in Logic Programs. PhD thesis, Computer Science Department, University of Southern California, 1990. Printed as Report TR 91-05. 15. M. J. Maher. Complete axiomatizations of the algebras of finite, rational and infinite trees. In Proceedings, Third Annual Symposium on Logic in Computer Science, pages 348–357, Edinburgh, Scotland, 1988. IEEE Computer Society. 16. K. Muthukumar and M. Hermenegildo. Compile-time derivation of variable dependency using abstract interpretation. Journal of Logic Programming, 13(2&3):315– 347, 1992. 17. J. A. Robinson. A machine-oriented logic based on the resolution principle. Journal of the ACM, 12(1):23–41, 1965.

Deriving Analysers by Folding/Unfolding of Natural Semantics and a Case Study: Slicing Val´erie Gouranton IRISA/INRIA, IFSIC Campus universitaire de Beaulieu, 35042 Rennes Cedex, France tel : 33 2 99 84 74 85, fax : 33 2 99 84 71 71, [email protected] http://www.irisa.fr/lande/

Abstract. We consider speciﬁcations of analysers expressed as compositions of two functions: a semantic function, which returns a natural semantics derivation tree, and a property deﬁned by recurrence on derivation trees. A recursive deﬁnition of a dynamic analyser can be obtained by fold/unfold program transformation combined with deforestation. A static analyser can then be derived by abstract interpretation of the dynamic analyser. We apply our framework to the derivation of a dynamic backward slicing analysis for a logic programming language.

1

Introduction

A large amount of work has been devoted to program analysis during the last two decades, both on the practical side and on the theoretical issues. However, most of the program analysers that have been implemented or reported in the literature so far are concerned with one speciﬁc property, one speciﬁc language and one speciﬁc service (dynamic or static). A few generic tools have been proposed but they are generally restricted to one class of properties or languages, or limited in their level of abstraction. We believe that there is a strong need for environments supporting the design of program analysers and that more eﬀort should be put on the software engineering of analysers. We present a framework for designing analysers from operational speciﬁcations by program transformation (folding/unfolding). The analysis speciﬁcation has two components: a semantics of the programming language and a deﬁnition of the property. The advantage of this two-fold speciﬁcation is that the deﬁnition of the property can be kept separate from the semantics of the programming language. Ideally, properties can be speciﬁed in terms of the derivation tree of the operational semantics. Speciﬁc analysers can then be obtained systematically by instantiating semantics of the programming language. We focus here on slicing of logic programs. The general approach is detailed in [12]. Natural semantics [7,14] are a good starting point for the deﬁnition of analyses because they are both structural (compositional) and intensional. They are structural because the semantics of a phrase in the programming language G. Levi (Ed.): SAS’98, LNCS 1503, pp. 115–133, 1998. c Springer-Verlag Berlin Heidelberg 1998

116

V. Gouranton

is derived from the semantics of subphrases; they are intensional because the derivation tree that is associated with a phrase in the programming language contains the intermediate results (the semantics of subphrases). These qualities are signiﬁcant in the context of program analysis because compositionality leads to more tractable proof techniques and intensionality makes it easier to establish the connection between the result of the analysis and its intended use. Our semantics is deﬁned formally as a function taking a term and an evaluation context and returning a derivation tree. The property itself is a function from derivation trees to a suitable abstract domain. The composition of these two functions deﬁnes a dynamic a posteriori analysis. It represents a function which initially calculates the trace of a complete execution (a derivation tree) of a program before extracting the required property. Program transformations via extended folding/unfolding techniques and simpliﬁcation rules allow to obtain a recursive deﬁnition of the dynamic analyser (which does not call the property function). This function is in fact a dynamic on the fly analyser in the sense that it calculates the required property progressively during program execution. The following diagram shows the general organisation: Context × Term Q

Semantics /

QQQ QAnalyser QQQ QQQ QQ(

Tree

Property

Result

The key points of the approach proposed here are the following: – The derivation is achieved in a systematic way by using functional transformations: unfolding and folding. – It is applicable to a wide variety of languages and properties because it is based on natural semantics deﬁnitions. As mentioned before, some of the analyses that we want to specify are dynamic and others are static. There is no real reason why these two categories of analyses should be seen as belonging to diﬀerent worlds. In the paper we focus on dynamic analysis, considering that static analysis can be obtained in a second stage as an abstract interpretation of the dynamic analysis as presented in [10]. We outline this derivation in the conclusion. Note that our approach introduces a clear separation between the speciﬁcation of an analysis (deﬁned as a property on semantics derivation trees) and the algorithm that implements it. We illustrate the framework by the formal derivation of a slicing analysis for a logic programming language. The diﬀerent stages of the derivation are detailed in the following sections. Section 2 introduces contexts, terms, derivation trees and the semantics function. The abstract domain and the property function are presented in section 3. The transformation of the composition of the two speciﬁcation functions (the semantics and the property) into a dynamic on the fly analyser is described in section 4. Related work, conclusion and avenues for further research are discussed in section 5 and section 6 respectively.

Deriving Analysers by Folding/Unfolding of Natural Semantics

2

117

Natural Semantics

The natural semantics of a language is a set of axioms and inference rules that deﬁne a relation between a context, a term in the programming language and a result. A natural semantics derivation tree has the form: Proof-Tree = [RN]

Proof-Tree1

... Proof-Treen STT

where RN is the name of the rule used to derive STT. The conclusion STT is a statement, that is to say a triple consisting of a context, a term and a result. Let C be the set of contexts, T the type of terms of the language and PT the type of derivation trees, we have: PT = STT × (list PT) × RN STT = C × T × NF T = PP × I

Derivation trees are made of a statement (the conclusion), a list of derivation trees (the premises) and the name of rule applied to derive the conclusion. We assume that a term is a pair of a program point and an expression. STT denotes the type of statements, RN rule names, NF normal forms (program results), PP program points and I expressions. The Semantics of a Logic Programming Language We assume a program P rog which is a collection of predicate deﬁnitions of the form [Pk (x1 , ..., xn ) = Bk ]. The body Bk of a predicate is in normal form and it contains only variables from {x1 , ..., xn }. Normal forms are ﬁrst order formulae (also called “goal formulae” in [16]) built up from predicate applications using only the connectives “and”, “or”, and “there exists”. Their syntax is deﬁned by: I ::= Op(x1 , x2 , x3 ) | x = t | U1 ∧ U2 | U1 ∨ U2 | ∃x.U1 | Pk (y1 , . . . , yn ) where Op stands for basic predicates1 and Pk for user-deﬁned predicates. Ui are terms of type T. We assume that each variable x occurring in a term ∃x.U1 is unique. In a program, each subterm in this syntax is associated with a program point (using pairs). As an illustration of this syntax, Figure 1 presents a small program in a logic programming syntax and shows its translation into normal form. Program points are represented by πi . Note that some program points are omitted for the sake of readability. The program deﬁnes two predicates P and Q. The main predicate is Q. The recursive predicate P computes the length n of the list l of integers, the sum sum of the elements of the list, the maximun max and the minimum min of the list l. The average av of the list is computed by the predicate Q via P (the value of sum obtained by P is divided by the length n of the list). 1

We consider only ternary basic predicates here, but other arities are treated in the same way.

118

V. Gouranton Deﬁnition of the program in a logic programming syntax P (nil, 1, 0, 0, 0) P ((x, nil), 1, x, x, x) P ((x, xs), n, sum, max, min) = (π1 , P (xs, n , sum , max , min )) (π2 , Ad(n , 1, n)) (π3 , Ad(sum , x, sum)) (π4 , M ax(max , x, max)) (π5 , M in(min , x, min)) Q(l, av, max, min) = (π6 , P (l, n, sum, max, min)) (π7 , Div(sum, n, av)) Normal form of the program

P (l, n, sum, max, min) = ((l = nil) ∧ (n = 1) ∧ (sum = 0) ∧ (max = 0) ∧ (min = 0))∨ (∃x. (l = (x, nil)) ∧ (n = 1) ∧ (sum = x) ∧ (max = x) ∧ (min = x))∨ (∃x. ∃xs. ∃n . ∃sum . ∃max . ∃min . l = (x, xs)∧ (π1 , P (xs, n , sum , max , min ))∧ (π2 , Ad(n , 1, n))∧ (π3 , Ad(sum , x, sum))∧ (π4 , M ax(max, x, max))∧ (π5 , M in(min , x, min))) Q(l, av, max, min) = (∃n. ∃sum. (π6 , P (l, n, sum, max, min))∧ (π7 , Div(sum, n, av))) Fig. 1. A simple logic program

Following [15], we assume an inﬁnite set of program variables Pvar and an inﬁnite set of renaming variables Rvar. Terms and substitutions are constructed using program variables and renaming variables. We distinguish two kinds of substitutions: program variable substitutions (Subst) whose domain and co-domain are subsets of Pvar and Rvar respectively, and renaming variable substitutions (Rsubst) whose domain and co-domain are subsets of Rvar: Subst = Pvar → Rterm Rsubst = Rvar → Rterm

where Rterm represents a term constructed with renaming variables Rvar. By convention, we use θ ∈ Subst for a program variable substitution and σ ∈ Rsubst for a renaming variable substitution. The deﬁnition of substitution composition is modiﬁed to take account the role held by renaming variables. The modiﬁcation occurs when θ ∈ Subst and σ ∈ Rsubst, we have σ ◦ θ ∈ Subst deﬁned by: dom(σ ◦ θ) = dom(θ) (σ ◦ θ)(x) = σ(θ(x)) for all x ∈ dom(θ)

Deriving Analysers by Folding/Unfolding of Natural Semantics

119

The domain of contexts for this language is deﬁned by C = Tree(Subst) where Tree(H) is the type of binary trees with leaves of type H. We deﬁne contexts as binary trees of substitutions to take into account the non deterministic nature of the language. So, we gather in one derivation the computation of all the substitutions of a program. A particular control strategy for the implementation of the language corresponds to a particular ordering of the leaves of substitutions trees. For instance, the list of results of the usual depth-ﬁrst evaluation strategy of Prolog is precisely the leaves of the substitution tree produced by our semantics ordered from left to right. We write N (T1 , T2 ) for a tree with subtrees T1 and T2 .

C (π, Op(x1 , x2 , x3 )) → op(C, x1 , x2 , x3 )

[Op] [Eq]

[∧]

C U1 → R1 R1 U2 → R2 C (π, U1 ∧ U2 ) → R2 [∃]

[Call]

C (π, x = t) → unif (C, x, t)

[∨]

C U1 → R1 C U2 → R2 C (π, U1 ∨ U2 ) → union (C, R1 , R2 )

Add (C, x, rx) U1 → R1 C (π, ∃x.U1 ) → Drop (R1 , x)

Renk (C) Bk → R1 C (π, Pk (y1 , ..., yn )) → Extk (C, R1 )

rx ∈ Rvar

fresh variable

with [Pk (x1 , ..., xn ) = Bk ] ∈ P rog

F (N (T1 , T2 ), x1 , . . . , xn ) = N (F (T1 , x1 , . . . , xn ), F (T2 , x1 , . . . , xn )) F (θ, x1 , . . . , xn ) = F (θ, x1 , . . . , xn ) op(θ, x1 , x2 , x3 ) = let σ = [(θ(x1 ) op θ(x2 ))/θ(x3 )] in σ ◦ θ if θ(x1 ) and θ(x2 ) are ground and θ(x3 ) ∈ Rvar,

⊥ otherwise

unif(θ, x, t) = let σ = mgu(θ(x), θ(t)) in σ ◦ θ if θ(x) and θ(t) can be uniﬁed , ⊥ otherwise union(N (T1 , T2 ), N (U1 , U2 ), N (V1 , V2 )) = N (union(T1 , U1 , V1 ), union(T2 , U2 , V2 )) union(θ, U, V ) = N (U, V ) Add(θ, pv, rv) = θ[rv/pv] with v = pv ⇒ θ[rv/pv](v) = θ(v) and θ[rv/pv](pv) = rv Drop(θ, pv) = θ/pv with v = pv ⇒ θ/pv (v) = θ(v) and θ/pv (pv) =⊥ Renk (θ) = [θ(yi )/xi ] Extk (θ, θ ) = σ ◦ θ with θ = σ ◦ [θ(yi )/xi ] Fig. 2. Natural semantics of a logic programming language

120

V. Gouranton

The natural semantics of a simple logic programming language using the usual inference rule presentation is presented in Figure 2. The normal forms calculated by the rules are contexts. In the ﬁgure, F (T ) denotes the application of a function F to all the substitutions of a tree T and its result is also a tree. The function op represents the interpretation of operator Op. The substitution unif (θ, x, t) of Subst is deﬁned for the uniﬁcation of x and t via θ (rule Eq). The rule ∧ is not surprising, the ﬁrst formula U1 of the conjonction is evaluated and the result R1 is taken as the new context for the evaluation of the second formula U2 of the conjonction; the result R2 is the ﬁnal result. For the rule ∨, the subtrees corresponding to the sub-formulae of the disjonction are evaluated independently. The function union(T1 , T2 , T3 ) is needed to build a new substitution tree joining the trees T2 and T3 produced by two subgoals. Its ﬁrst argument is the initial substitution, which is used to identify the points where the joins have to be introduced (these points are the leaves of T1 ). The argument θ as the initial substitution can be ignored because substitutions are added to contexts, generating new contexts. The rule ∃ uses two functions Add and Drop. Add is used to add a program variable in a substitution (the new program variable is attached to a free renaming variable) and Drop removes a variable from a substitution. For the rule Call, two deﬁnitions of substitutions are needed. Renk (C) creates a new substitution to execute the body of a clause (it amounts to a variable renaming) because the body Bk of a clause contains formal parameters xi and C contains program variables yi . Extk (C, R1 ) propagates the result of a predicate in the calling substitutions because C contains variables yi and R1 contains formal parameters xi . From the deﬁnition of Renk , we see that the body Bk of a predicate is evaluated in an environment deﬁning exactly the formal parameters of the predicate Pk . The formal deﬁnitions of the functions introduced informally before are presented in the bottom of Figure 2. In order to make formal manipulations easier, we express the construction of natural semantics derivation trees in a functional framework. The semantic function S is a partial function of type: C × T → PT

The important issue about the type of the semantic function is that it returns the whole natural semantics derivation tree, rather than just the result of the program. This choice makes it easier to deﬁne intensional analyses. The fact that we describe the semantics in a functional framework does not prevent us from dealing with non deterministic languages, as we show for a logic programming language. This is because we can use NF and C to represent sets of possible results and contexts. We use the notation X.ty to denote the ﬁeld of type TY of X. For example, we will make intensive use of the following expressions in the rest of the paper:

Deriving Analysers by Folding/Unfolding of Natural Semantics

type P T .stt P T .lpt P T .rn P T .stt.c P T .stt.t.i P T .stt.t.pp P T .stt.nf

121

meaning

STT conclusion of P T (list PT) premisses of P T RN name of the rule used at the root of P T C context of the conclusion sequent of P T I term of the conclusion sequent of P T PP program point of the conclusion sequent of P T NF normal form of the conclusion sequent of P T

The semantics in functional form is presented in Figure 3. The semantics function of Figure 3 takes two arguments (the context C and the term T ) and it returns a derivation tree. The derivation tree contains the conclusion (C, T, F k (C, R, E)) of type STT, where F k is the result of the program in functional form, the list of subtrees and the name k of the rule used to derive the conclusion. The body of the function is a list of cases selected by pattern matching on the form of the term. The function is deﬁned by recurrence on the term. The set of deﬁnitions Prog is used as an implicit parameter of the semantics.

S (C, T ) = case T of (π, Op (x1 , x2 , x3 )) : ((C, T, op(C, x1 , x2 , x3 )), nil, Op) (π, Eq(x, t)) : ((C, T, unif (C, x, t)), nil, Eq) (π, And (U1 , U2 )) : let P T1 = S (C, U1 ) R1 = P T1 .stt.nf P T2 = S (R1 , U2 ) R2 = P T2 .stt.nf in ((C, T, R2 ), [P T1 , P T2 ], ∧) (π, Or (U1 , U2 )) : let P T1 = S (C, U1 ) R1 = P T1 .stt.nf P T2 = S (C, U2 ) R2 = P T2 .stt.nf in ((C, T, union (C, R1 , R2 )), [P T1 , P T2 ], ∨) (π, Exists (x, U1 )) : let P T1 = S (Add (C, x, rx), U1 ) R1 = P T1 .stt.nf in ((C, T, Drop (R1 , x)), [P T1 ], ∃) (π, Call (Pk (y1 , . . . , yn ))) : let P T1 = S (Renk (C), Bk ) R1 = P T1 .stt.nf in ((C, T, Extk (C, R1 )), [P T1 ], Call) with [Pk (x1 , ..., xn ) = Bk ] ∈ Prog Fig. 3. The semantics function of a logic programming language

122

3

V. Gouranton

Specification of a Slicing Property

Slicing2 a program consists in constructing a reduced version of the program (called a program slice) containing only those statements that aﬀect a given set of variables at given program points (this set is called the slicing criterion). In program debugging, slicing makes it possible for a software engineer to focus on the relevant parts of the code. Slicing is also useful for testing, program understanding and in maintenance activities. Because of this diversity of applications, diﬀerent variations on the notion of slicing have been proposed, as well as a number of methods to compute slices. First, a program slice can either be executable or not. Producing an executable slice makes it possible to apply further treatments to the result of the analysis. Another important distinction is between static and dynamic slicing. In the ﬁrst case, the slice is computed without any assumption on the inputs, whereas the latter relies on some speciﬁc input data. Slicing algorithms can also be distinguished by their direction. Backward slicing identiﬁes the statements of a program that may have some impact on the criterion whereas forward slicing returns the statements which may be inﬂuenced by the criterion. In this paper, we consider dynamic backward slicing with executable slices. Static slicing algorithms can be derived by abstract interpretation of dynamic slicing analysers ; this construction is sketched in the conclusion. We can describe forward slicing analysers in a similar way but slicing analyses producing non executable slices do not ﬁt well into our framework since the speciﬁcation of the analysis is a relation between the semantics of the original program and the semantics of the slice as presented in [10] . Slicing was originally proposed by Weiser for imperative languages [29] and its application to logic programming [23] and functional programming [18] have been studied recently. In fact, the concept of slicing itself is very general: it is not tied to one speciﬁc style of programming3 and it can lead to dynamic as well as static analysers [25]. A slicing analysis for a logic programming language (with programs in normal form) according to a program point and a set of variables of interest consists in keeping only the sub-goals of disjunctions of each clause (a clause deﬁnes a predicate) being able to aﬀect the value of the variables of interest. If all subgoals of a formula of the disjunction are dropped, then this formula is dropped. If all formulae of the disjunction of goals are dropped, then the clause is dropped. In the opposite case, the head of the clause deﬁning the predicate is kept. Let us take the program in normal form of Figure 1 to illustrate dynamic backward slicing. We assume that we are interested only in the value of the variable av at the program point π7 . The pair {(π7 , av)} is called the slicing criterion. The dynamic slice of the program is extracted for one particular input. For instance, if we execute the predicate Q with nil as the initial value of l, we get: 2 3

More precisely “backward slicing”. Even if the details of the resulting analyses are of course.

Deriving Analysers by Folding/Unfolding of Natural Semantics

123

P (nil, 1, 0, 0, 0) Q(l, av, max, min) = (π6 , P (l, n, sum, max, min)) (π7 , Div(sum, n, av))

The predicate P is not recursively called and the ﬁrst disjunctive part is satisﬁed, the third clause of P is never executed. The deﬁnition of the predicate Q is kept because all its clauses are useful to compute the variable av. If we consider the execution of the program with (2, (3, nil)) as initial value of l, we recursively call the predicate P , we get: P ((x, nil), 1, x, x, x) P ((x, xs), n, sum, max, min) = (π1 , P (xs, n , sum , max , min )) (π2 , Ad(n , 1, n)) (π3 , Ad(sum , x, sum)) Q(l, av, max, min) = (π6 , P (l, n, sum, max, min)) (π7 , Div(sum, n, av))

Only a part of the third clause of the predicate P is kept, the program points π4 and π5 are dropped because they are not useful in computing av (they are needed to compute the values for max and min ). Assuming a set of pairs (πi , vi ), where πi is a program point and vi a variable, a backward slicing analysis produces the slice computing for each point πi the same values as the initial program for the variable vi . In our framework, a property is expressed by a function which takes at least an argument being the co-domain of the semantics function (a derivation tree of type PT) and the result of the property is an abstract domain. The slicing property takes an additional argument to represent the slicing criterion (of type PP → P(Pvar)) and the type of the result is P(PP) because slices are represented by sets of program points. The slicing criterion is represented in our approach by the mapping from program points to relevant variables. Because of the slicing property, we need extra information. We introduce a set of variables of interest according to a program point (this set represents the value of variable that must be preserved for computing the corresponding term). The initial value of the set is ∅. The property propagates this information of type P(Pvar) and ﬁnally the type of the property is: αsl : PT × (PP → P(Pvar)) × P(Pvar) → P(PP) × P(Pvar) The slicing property αsl for the logic programming language is presented in Figure 4. The property takes as arguments a derivation tree P T , plus two additional parameters RV ∈ PP → P(Pvar) and D ∈ P(Pvar). The second argument RV (for Relevant Variables) is the slicing criterion mentioned above. A program point π associated with a non-empty set RV (π) is called an observation point. The third argument D of the property represents the set of variables whose

124

V. Gouranton

values must be preserved in the output context4 (normal form) of the term, i.e. the set of variables that must be preserved in the result Ri of the evaluation of the derivation tree P Ti . In the initial call, D is the empty set. The function αsl is called recursively on the intermediate derivation trees (P Ti ) of the natural semantics and sets of observation variables. The result of the property is a pair (S, N ) with S ∈ P(PP) and N ∈ P(Pvar). S is the set of program points of the term T that must be kept in the slice and N is the set of variables whose value must be preserved in the input context5 . A program point must be kept in the slice if it can inﬂuence an observation point or the value of a variable of D in the output context. The same condition applies to decide which variables must be preserved in the input context. If the program point can be removed from the slice, the result of the property is (∅, D), which means that no program point is added to the slice and the variables whose values must be preserved in the input context are the variables that are necessary in the output context. Otherwise, the ﬁrst component of the result of the property is ∪ Si ∪ {π} because π has to be added to the program points collected in the i

subterms of T . The second component N of the result is the set of variables whose value must be preserved in the input context C. It contains at least the set D and the variables RV (π) of slicing criterion, thus we factorise that by setting D = D ∪ RV (π) in beginning of the slicing deﬁnition. We assume that the deﬁnitions of Si and Ni are not mutually recursive. The deﬁnition of the sets of observation variables (third argument of αsl ) do not use Nj , j > i. Note that this is a characteristic feature of a backward analysis. In Figure 4, the relation Indep(C, D1 , D2 ) is used to ensure that two sets of variables D1 and D2 are independent, which is the case when they do not share any renaming variables (in any substitution of the context C). The relation Indep appears in the ﬁrst two cases as a necessary condition to exclude the term from the slice. If the relation holds, then the (renaming variable) substitution resulting from the evaluation of the term cannot have any impact on the variables of D. The relation UF(C, x, t) is satisﬁed if the uniﬁcation of x and t cannot fail for any substitution of C. It is a prerequisite for excluding Eq(x, t) from the slice because a failure is recorded in the substitution tree as the ⊥ substitution6 ; as a consequence, it has an impact on all the variables. This condition was not included in the Op case, assuming that the logic programming language is equipped with mode annotations ensuring that operators are always called with their ﬁrst two arguments ground and the last one free7 . In both the Op and the Eq cases, the set of necessary variables (at the input of the program point) is D added to all the program variables of the term: the set {x1 , x2 , x3 } for 4 5 6 7

For a forward property this argument would characterise the input context rather than the output context. For a forward property this argument would characterise the output context rather than the input context. Note that ⊥ is an absorbing element for the semantics of the language. For instance op(⊥, x1 , x2 , x3 ) =⊥ and unif (⊥, x, t) =⊥. Otherwise an extra condition based on UF can be added as in the Eq case.

Deriving Analysers by Folding/Unfolding of Natural Semantics

125

Op (x1 , x2 , x3 ) and the set of program variables occurring in t increased with x for the rule Eq (x, t). The formal deﬁnitions of Indep and U F are presented in

the bottom of Figure 4. For the rule And, both branches are processed in turn (the second branch ﬁrst since our property is computed in a backward direction). The property is ﬁrst called with P T2 and D and the result is (S2 , N2 ); then the property is computed with P T1 and N2 , we have (S1 , N1 ) as the result. The program point π can be removed from the slice when both S1 and S2 are empty sets. When the program point is kept, the result of the operator And is then (S1 ∪ S2 ∪ {π}, N1 ) because the information about program points of both branches is kept and the set N1 represents the variables must be preserved in the input context since we consider a backward direction. The treatment of Or is diﬀerent: the term is systematically kept in the slice because it always inﬂuences the values of all the variables (through the introduction of subtrees in the derivation tree). Both branches are computed independently and the result gathers the information of these two branches. The rules for Exists and Call are not surprising. We assume that the variable x in Exists(x, U1 ) is unique in a normalised program; so x can be removed from the set of necessary variables yielded by the analysis of U1 (hence N1 − {x}). In the rule for Call, ﬁrst the derivation tree corresponding to the predicate Pk is computed with the set {xi | ¬Indep(C, D , {yi })} of variables to be preserved (i.e. the formal parameters xi of Pk bounded to arguments yi which are not independent from the set D ). The test in the rule for Call is similar to the test in the Op case. We could make more sophisticated choices to avoid including all the variables y1 ..., yn in the set of the necessary variables.

4

Derivation of the Dynamic on the Fly Analyser

We have presented in section 2 the semantics function S and the property αsl in functional form in section 3. The general organisation is described by the following diagram: S / PT C × TE EE ν EE a EE αsl E" Da The composition of the property αsl and the semantics S is a function of type C × T → Da , where Da is the domain of abstract values, the result of the analysis. This function computes successively the derivation tree related to a program, then the property of interest for this tree. It corresponds to a dynamic analysis a posteriori that inspects the trace produced after the program execution. It is interesting to formally describe dynamic analysers, because they are useful for instrumentation or debugging. We could also prefer dynamic analyses which, calculate their result on the fly i.e. during program execution. Their advantage is that they do not have to memorise traces before analysing them.

126

V. Gouranton

αsl (P T, RV, D) = let π = P T .stt.t.pp C = P T .stt.c D = D ∪ RV (π) in case (P T .lpt, P T .rn) of (nil, Op) : let Op (x1 , x2 , x3 ) = P T .stt.t.i in if RV (π) = ∅ and Indep(C, D, {x3 }) then (∅, D) else ({π}, D ∪ {x1 , x2 , x3 }) (nil, Eq) : let Eq (x, t) = P T .stt.t.i in if RV (π) = ∅ and UF(C, x, t) and Indep(C, D, P v(t) ∪ {x}) then (∅, D) else ({π}, D ∪ P v(t) ∪ {x}) ([P T1 , P T2 ], ∧) : let (S2 , N2 ) = αsl (P T2 , RV, D ) (S1 , N1 ) = αsl (P T1 , RV, N2 ) in if RV (π) = ∅ and S1 ∪ S2 = ∅ then (∅, D) else ( S1 ∪ S2 ∪ {π}, N1 ) ([P T1 , P T2 ], ∨) : let (S2 , N2 ) = αsl (P T2 , RV, D ) (S1 , N1 ) = αsl (P T1 , RV, D ) in ( S1 ∪ S2 ∪ {π}, N1 ∪ N2 ) (P T1 , ∃) : let Exists (x, U1 ) = P T .stt.t.i (S1 , N1 ) = αsl (P T1 , RV, D ) in if RV (π) = ∅ and S1 = ∅ then (∅, D) else (S1 ∪ {π}, N1 − {x}) (P T1 , Call) : let Call(Pk (y1 , ..., yn )) = P T .stt.t.i (S1 , N1 ) = αsl (P T1 , RV, {xi | ¬Indep(C, D , {yi })}) in if RV (π) = ∅ and S1 ∪ N1 = ∅ and Indep(C, D, {y1 , ..., yn }) then (∅, D) else (S1 ∪ {π}, D ∪ {y1 , ..., yn }) U F (C, x, t) = ∀θ ∈ C. θ =⊥⇒ ∃σ = mgu(θ(x), θ(t)) P v(t) = set of program variables occurring in t Rv(rt) = set of renaming variables occurring in rt Indep(C, D1 , D2 ) = ∀θ ∈ C. θ =⊥⇒ {Rv(θ(x)) | x ∈ D1 } ∩ {Rv(θ(x)) | x ∈ D2 } = ∅ Fig. 4. Slicing property

Deriving Analysers by Folding/Unfolding of Natural Semantics

127

The derivation of a dynamic on the fly analyser from a dynamic analyser a posteriori presents similarities with a well-known program transformation within the framework of functional programming. The program transformation is called deforestation [28] and its purpose is to eliminate the intermediate data structures induced by the composition of recursive functions. Here, the intermediate structure is the derivation tree of the natural semantics. We use folding and unfolding transformations to carry out deforestation. The three principal operations are the following: – unfoldings: we set νa (C, T ) = αsl (S (C, T )) and replace in the expression the calls to the recursive functions αsl and S by their deﬁnition. – applications of laws on the operators of the language (like the conditional ones, the expressions case and let ). – foldings which consist in replacing the occurrences of αsl (S (C , T )) from calls to νa (C , T ). The goal of these transformations is to remove all the calls to the property extraction function αsl , to obtain a closed deﬁnition of νa (C, T ). The function obtained is then a dynamic on the fly analyser since it does not build the intermediate derivation trees any more. The partial correction of the transformation by folding/unfolding is obvious. The total correction is not assured in general because some inopportune foldings can introduce cases of non-termination. The Improvement Theorem in [20] can be extented to a method (the extended improved unfold-fold method) presented in [21] which makes it possible to show the total correction of the method proposed in this paper.

Dynamic Slicing Analyser The deﬁnition of the dynamic slicing analyser for the logic programming language is the following: SLd (C, T, RV, D) = αsl (S (C, T ), RV, D) First, we use an unfolding technique applied to the semantics and the property functions. We present in [11] the transformation rules used for the derivation of the dynamic on the fly analyser by unfolding. Figure 5 presents these unfoldings for two rules (the other cases are straightforward). To obtain a dynamic on the fly analyser, we must apply folding steps that allows us to remove the calls of the function αsl . Figure 6 presents the result of these foldings. The fact that SLd itself calls S shows that it is a dynamic analysis.

128

V. Gouranton

SLd (C, T, RV, D) = let D = D ∪ RV (T.pp) in case T of (π, Op (x1 , x2 , x3 )) : if RV (π) = ∅ and Indep(C, D, {x3 }) then (∅, D) else ({π}, D ∪ {x1 , x2 , x3 }) (π, And (U1 , U2 )) : let P T1 = S (C, U1 ) R1 = P T1 .stt.nf (S2 , N2 ) = αsl (S (R1 , U2 ), RV, D ) (S1 , N1 ) = αsl (S (C, U1 ), RV, N2 ) in if RV (π) = ∅ and S1 ∪ S2 = ∅ then (∅, D) else ( S1 ∪ S2 ∪ {π}, N1 ) Fig. 5. Unfoldings of semantics and property functions

SLd (C, T, RV, D) = let D = D ∪ RV (T.pp) in case T of (π, Op (x1 , x2 , x3 )) : if RV (π) = ∅ and Indep(C, D, {x3 }) then (∅, D) else ({π}, D ∪ {x1 , x2 , x3 }) (π, And (U1 , U2 )) : let P T1 = S (C, U1 ) R1 = P T1 .stt.nf (S2 , N2 ) = SLd (R1 , U2 , RV, D ) (S1 , N1 ) = SLd (C, U1 , RV, N2 ) in if RV (π) = ∅ and S1 ∪ S2 = ∅ then (∅, D) else ( S1 ∪ S2 ∪ {π}, N1 ) Fig. 6. Dynamic (on the fly) slicing analysis

Deriving Analysers by Folding/Unfolding of Natural Semantics

5

129

Related Work

The fold/unfold transformation framework used here is based on seminal work by Burstall and Darlington [2,6]. The application of the technique to the derivation of programs has also been investigated in [5], which presents the synthesis of several sorting algorithms. The initial speciﬁcation is expressed in terms of sets and predicate logic constructs. Our transformations are also reminiscent of the deforestation technique [3,9,28]: in both cases the goal is to transform a composition of recursive functions into a single recursive deﬁnition. Generic frameworks for program analysis have been proposed in the context of logic programming languages [15] and data ﬂow analysis [26,30]. They rely on abstract interpretations of denotational semantics [15,26] or interpreters [30] and genericity is achieved by parameterising the abstract domains and choosing appropriate abstract functions. The implementation details of the analysis algorithm can be factorised. While these tools may attain a higher degree of mechanisation than our framework, they do not oﬀer to the user the same level of abstraction: they take as input the specification of an abstract interpreter rather than the specification of a property. Despite this diﬀerence of point of view, all these works are obviously inspired by the same goals. The framework introduced in [24] is closer to the spirit of the work presented in this paper but the technique itself is quite diﬀerent. Programs are represented as models in a modal logic and a data ﬂow analysis can be speciﬁed as a property in the logic. An eﬃcient data ﬂow analyser can be generated by partially evaluating a speciﬁc model checker with respect to the specifying modal formula. In comparison with this work, our framework trades mechanisation against generality: it is not limited to data ﬂow analyses but the derivation process by fold/unfold transformations is not fully automatic. Few papers have been devoted to the semantics of program slicing so far. A relationship between the behaviour of the original program and the behaviour of the slice is proved in [19]. The semantics of the language is expressed in terms of program dependence graphs; thus the programs are ﬁrst analysed in order to extract their dependences. This approach is well suited to the treatment of imperative languages. Formal deﬁnitions and a classiﬁcation of diﬀerent notions of slicing are provided in [27]. The main distinctions are backward vs forward analysers, executable vs non executable slices, and dynamic vs static analysers. Their deﬁnitions are based on denotational semantics and they focus on the speciﬁcations of the analyses. In [8] a description of a family of slicing algorithms generalising the notions of dynamic and static slice to that of a constrained slice is presented. Genericity with respect to the programming language is achieved through a translation into an intermediate representation called pim. Programs are represented as directed acyclic graphs whose semantics is deﬁned in terms of rewriting rules. Slicing is carried out using term graph rewriting with a technique for tracing dynamic dependence relations. It should be noted that a richer notion of slicing has been proposed for logic programming languages, which returns not only the set of program points that must be kept in the slice, but also the necessary variables at each program point [23]. This increased precision can also

130

V. Gouranton

be expressed in our framework, but we preferred to present the simpler version here for the sake of size and readability. By collecting the following information ( ∪ Si + {(π, D ∪ N )}, N ) i

we can modify straightforwardly each rule in order to get the same precision as [23].

6

Conclusion

We have presented a method to derive dynamic analysers by program transformation (folding/unfolding). A dynamic analyser is expressed as composition of a semantics and a property functions. The analyser is called a posteriori, it is a function computing ﬁrst a complete program execution trace (derivation tree) and then extracting the property of interest. A recursive deﬁnition of an analyser can be obtained by program transformation. This function is a dynamic on the fly analyser that computes the property during program execution. We have focussed on dynamic analysis in the body of paper. Our generic dynamic analyser is deﬁned in a strongly typed functional language8. As a consequence, we can rely on previous results on logical relations and abstract interpretation [1,4] in order to systematically construct static analysers from the dynamic analysers. The ﬁrst task is to provide abstract domains for the static slicing analyser and the corresponding abstraction functions. We recall that the type of the dynamic analyser is C × T × (PP → P(Pvar))× P(Pvar) → P(PP)× P(Pvar). Since PP → P(Pvar), P(Pvar) and P(PP) are already abstract domains associated with the dynamic analysis, only C needs to be abstracted9 . The next stage to derive a correct static analyser is to ﬁnd appropriate abstractions for the constants and operators occurring in the deﬁnition of the analyser. It is shown in [1] that the correctness of the abstract interpretation of the constants and operators of the language entails the correctness of the abstract interpretation of the whole language. The correctness of the abstract interpretation means that the results of the dynamic analysis and the static analysis are related if their arguments are. In fact, it is possible to deﬁne the most precise abstraction for each constant and operator of the language [1]. The basic idea to ﬁnd the best abstraction opa (v1a , . . . , vna ) of an operator op is to deﬁne it as the least upper bound of the abstractions of all the results of op applied to arguments vi belonging to the concretisation sets of the arguments of the via . The technique sketched here provides a systematic way to construct a correct abstract interpretation, and thus to derive a static analyser from a dynamic analyser [10,12]. By deriving static analysers as abstractions of dynamic analysers, we can see the dynamic 8 9

Note that the typing mentioned here has nothing to do with the language in which the analysed programs are written, this language itself can perfectly well be untyped. Of course, as usual in abstract interpretation, PP → P(Pvar), P(Pvar) and P(PP) can also be abstracted if further approximations are needed, but we do not consider this issue here.

Deriving Analysers by Folding/Unfolding of Natural Semantics

131

analyser either as an intermediate stage in the derivation of a static analyser (playing a role similar to a collecting semantics) or as the ﬁnal product of the derivation. The theory of abstract interpretation [4] provides a strong formal basis for static program analysis. The work described here does not provide an alternative formal underpinning for program analysis. Its goal is rather to put forward a derivation approach for the design of analysers from high level speciﬁcations. Our framework is applicable to a wide variety of languages, properties and type of service (dynamic or static). We have proposed in the body of the paper a formal deﬁnition of a dynamic slicing analyser for a logic programming language. To our knowledge, this deﬁnition is the ﬁrst one to be formal, so the beneﬁt of our approach is striking in this case. In [10], we present the derivation of dynamic and static analysers for a strictness analysis of a higher-order functional language and a live variable analysis for an imperative language. We have also applied this work for a globalisation analysis of a higher-order functional language and a generic sharing analysis. Pushing our approach ever further we arrive at a natural semantics format and a format for slicing, as presented in [12]. We have shown the correctness of the slicing property format. These formats can be instantiated for several programming languages (imperative language, logic programming language and functional language). The slicing property for the logic programming that we have presented here is an instantiation of the slicing format. As mentioned in the introduction, we wanted to establish the connection between the result of the analysis and its intended use. Analyses are generally performed to check assumptions about the behaviour of the program at speciﬁc points of its execution or to enable program optimisations. In both cases the intention of the analysis can be expressed in terms of a transformation and a relation as presented in [10,12]. The transformation depends on the result of the analysis and the relation establishes a correspondence between the semantics of the original program and the transformed program. For example, in the case of a program analysis for compiler optimisation the transformation expresses the optimisation that is allowed by the information provided by the analysis and the relation is the equality between the ﬁnal results (or outputs) of the original and the transformed program. It is not always the case that the relation is the equality: a counter-example is slicing analysis described in this paper (because the new program is required to behave like the original one only with respect to speciﬁc program points and variables). We have formally deﬁned and proved in [12] a property for the intention of a slicing analysis but space considerations prevent us from presenting the intentional property for slicing. There is a main aspect in which the work described here may seem limited: we have used only natural semantics and terminating programs. Structural Operational Semantics (SOS) are more precise than natural semantics and they are required for a proper treatment of non-determinism, non-termination and parallelism [17]. In fact, the natural semantics introduced in section 2 can be

132

V. Gouranton

replaced by SOS without diﬃculty10 and the dynamic analyses can be deﬁned in the very same way. The extra diﬃculty introduced by SOS is the fact that they create new program fragments which makes it necessary to abstract over the syntax of the language to derive a static analyser. This problem is discussed in [22]. We can also adapt our natural semantics to SOS by using the technique presented in [13]. To achieve this goal, the classical inductive interpretation of natural semantics has to be extended with coinduction mechanisms and rules must be deﬁned to express divergence.

Acknowledgements I would like to thank Pascal Fradet, Thomas Jensen, Daniel Le M´etayer and Ronan Gaugne for their useful comments. Thanks are also due to the anonymous referees SAS for their criticisms on an earlier draft.

References 1. S. Abramsky. Abstract interpretation, logical relations and Kan extensions. Journal of Logic and Computation, 1:5–40, 1990. 130, 130, 130 2. R. M. Burstall and J. Darlington. A transformation system for developing recursive programs. Journal of the ACM, 24:44–67, 1977. 129 3. W.N. Chin. Automatic methods for program transformation. PhD thesis, Imperial College, 1990. 129 4. P. Cousot and R. Cousot. Abstract interpretation: a uniﬁed lattice model for static analysis of programs by construction or approximation of ﬁxpoints. In POPL, pages 238–252, 1977. 130, 131 5. J. Darlington. A synthesis of several sorting algorithms. Acta Informatica, 11:1–30, 1978. 129 6. J. Darlington and R. M. Burstall. A system which automatically improves programs. Acta Informatica, 6:41–60, 1976. 129 7. T. Despeyroux. Typol: a formalism to implement natural semantics. Technical Report 94, INRIA, France, 1988. 115 8. J. Field, G. Ramalingam, and F. Tip. Parametric program slicing. In POPL, pages 379–392, 1995. 129 9. A. Gill, J. Launchbury, and S.L. Peyton Jones. A short cut to deforestation. In FPCA, pages 223–232, 1993. 129 10. V. Gouranton. D´erivation d’analyseurs dynamiques et statiques a ` partir de sp´ecifications op´ erationnelles. PhD thesis, Universit´e de Rennes, France, 1997. 116, 122, 130, 131, 131 11. V. Gouranton. Deriving analysers by folding/unfolding of natural semantics and a case study: slicing. Technical Report 3413, INRIA, France, 1998. 127 12. V. Gouranton and D. Le M´etayer. Dynamic slicing: a generic analysis based on a natural semantics format. Technical Report 3375, INRIA, France, 1998. 115, 130, 131, 131, 131 10

In order to deal with SOS, we basically need to change the type NF in STT and to introduce a global loop in the semantics since a SOS rule represents a single evaluation step.

Deriving Analysers by Folding/Unfolding of Natural Semantics

133

13. H. Ibraheem and D. A. Schmidt. Adapting big-step semantics to small-step style: coinductive interpretations and ”higher-order” derivations. In Second Workshop on Higher-Order Techniques in Operational Semantics (HOOTS2), 1997. 132 14. G. Kahn. Natural semantics. In STACS 87, number 247 in Lecture Notes in Computer Science, pages 22–39. Springer-Verlag, 1987. 115 15. B. Le Charlier, K. Musumbu, and P. Van Hentenryck. A generic interpretation algorithm and its complexity analysis. In ICLP, pages 64–78, 1991. 118, 129, 129 16. D. Miller and G. Nadathur. Higher-order logic programming. In ILPC, volume 225 of LNCS, pages 448–462, 1986. 117 17. H. Riis Nielson and F. Nielson. Semantics With Applications. John Wiley & Sons, 1992. 131 18. T. Reps and T. Turnidge. Program specialization via program slicing. In International Seminar on Partial Evaluation, 1996. 122 19. T. Reps and W. Yang. The semantics of program slicing. Technical Report 777, University of Wisconsin, Madison, 1988. 129 20. D. Sands. Proving the correctness of recursion-based automatic program transformations. In TAPSOFT. Springer-Verlag, 1995. 127 21. D. Sands. Total correctness by local improvement in the transformation of functional programs. TOPLAS, 18:175–234, 1996. 127 22. D.A. Schmidt. Abstract interpretation of small-step semantics. In 5th LOMAPS Workshop on Analysis and Verification of Multiple-Agent Languages. LNCS, 1996. 132 23. S. Schoenig and M. Ducass´e. A backward slicing algorithm for prolog. In SAS, number 1145 in LNCS, pages 317–331. Springer-Verlag, 1996. 122, 129, 130 24. B. Steﬀen. Generating data ﬂow analysis algorithms from modal speciﬁcations. Science of Computer Science, 21(2):115–139, 1993. 129 25. F. Tip. A survey of program slicing techniques. Journal of Programming Languages, 3:121–189, 1995. 122 26. G.A. Venkatesh. A framework for construction and evaluation of high-level speciﬁcations for program analysis techniques. In PLDI, volume 24, pages 1–12, 1989. 129, 129 27. G.A. Venkatesh. The semantics approach to program slicing. In PLDI, pages 107–119, 1991. 129 28. P. Wadler. Deforestation: transforming programs to eliminate trees. Theoretical Computer Science, 73:231–248, 1990. 127, 129 29. M. Weiser. Program slicing. IEEE Transactions on Software Engineering, 4:352– 357, 1984. 122 30. K. Yi and W. L. Harrison III. Automatic generation and management of interprocedural program analyses. In POPL, pages 246–259, 1993. 129, 129

A Symbolic Semantics for Abstract Model Checking? Francesca Levi Dipartimento di Informatica, Universit´ a di Pisa and ´ LIX, Ecole Polytechnique [email protected]

Abstract. We present a finite symbolic semantics of value-passing concurrent processes, that can be suitably interpreted over abstract values to compute a lower approximate semantics of full µ-calculus. The main feature of the semantics is that classical branching is replaced by explicit relations of non-deterministic and alternative choices among transitions. A combination of safe upper and lower approximations of the basic operators of the logic is used to handle negation. The relations of non-deterministic and alternative choices turn out to be very useful for the dual approximations of the existential next modality. Key words: Model checking, µ-calculus, abstract interpretation.

1

Introduction

Model Checking is a very successful technique for the automatic verification of temporal properties of reactive and concurrent systems, but it is only applicable to finite-state systems. Over the past few years, abstract interpretation has been widely applied to handle large as well as infinite systems with model checking [3,1,9,11,13,4,8,17,10,15]. Abstract interpretation [6,7] was originally conceived in the framework of data-flow analysis for designing approximate semantics of programs and relies on the idea of obtaining an approximate semantics from the standard one by substituting the concrete domain of computation and its basic operations with an abstract domain and corresponding abstract operations. The typical approach consists of constructing an abstract model over a chosen set of abstract states that can be used in model checking instead of the concrete one. To this aim the abstract model has to be safe, namely the formulas satisfied by the abstract model have to hold in the concrete one. For branching time logics the definition of a safe abstract transition relation among abstract states presents some basic difficulties and a single abstract transition relation cannot preserve both the existential and the universal next modality. Several authors [11,10,4] propose to adopt two different abstract transition relations: a ?

This work has been partially supported by the HCM project ABILE (ERBCHRXCT940624).

G. Levi (Ed.): SAS’98, LNCS 1503, pp. 134–151, 1998. c Springer-Verlag Berlin Heidelberg 1998

A Symbolic Semantics for Abstract Model Checking

135

free transition relation for computing the universal next modality, and a constrained transition relation for computing the existential next modality. Safeness of the free transition relation is ensured, if every concrete transition induces a free transition among the corresponding abstract states. In contrast, a constrained transition between abstract states is safe only if all the corresponding concrete transitions exist. Given this notion of safeness it turns out to be very difficult to effectively compute a sufficiently precise safe abstract model without constructing the concrete one. In this paper we propose a method for applying abstract interpretation to the µ-calculus model checking of value-passing concurrent processes. The main contribution is the definition of a symbolic semantics of processes in the style of [12], whose main feature is that explicit relations of non-determinism and alternative choice among transitions replace classical branching. Moreover, a finite graph for regular processes is achieved by avoiding the infinite paths of [12] due to parameterized recursion. Model checking of µ-calculus can be suitably performed by interpreting the obtained symbolic graph over concrete environments assigning concrete values to variables. However, since processes are capable of exchanging values taken from a typically infinite set, the fixpoint computation of µ-calculus semantics is not effectively computable. We define a technique to compute a lower approximation of the µ-calculus semantics by interpreting the symbolic graph over abstract environments on a given (finite) set of abstract values. Safeness of the lower approximation ensures indeed the preservation of any property. Following [13] the lower approximation is achieved by combining dual safe upper and lower abstract functions corresponding to all logical connectives except negation. The critical case is undoubtedly that of the next modality, where safe constrained and free transition relations among abstract processes have to be considered. We show that explicit non-deterministic and alternative choices between transitions allow us to avoid some typical problems due to abstract branching so that a more precise lower approximation of the next modality in particular is achieved with respect to previous proposals [10]. Finally, we discuss the basic problems that typically lead to miss optimality in the approximations of the next modalities. The paper is organized as follows. Section 2 presents value-passing concurrent processes, µ-calculus and concrete model checking. Section 3 summarizes the basic concepts of abstract interpretation. The symbolic graph is described in Sect. 4 and the corresponding model checking algorithm is shown in Sect. 5. Section 6 presents abstract model checking and Sect. 7 discuss optimality of abstract model checking.

2

Concrete Model Checking

We consider a value-passing version of CCS. Let V al a set of values (possibly infinite), Chan a set of channels and V ar a set of variables. Moreover, let Bexp and V exp be sets of boolean and values expressions. Processes P roc are generated

136

F. Levi

by the following grammar p ::= nil | x | a.p | be 5 p1 , p2 | p1 × p2 | p1 + p2 | p \ L | P (e1 , . . . , en ) where a ∈ {c!e, c?x, τ | c ∈ Chan, e ∈ V exp, x ∈ V ar}, L ⊆ Chan, and be ∈ Bexp is a boolean expression and ei ∈ V exp are expressions over values. Process c?x.p will receive a value v on the channel c and then behaves as p[v/x], where p[v/x] denotes the standard substitution of v for all free occurrences of x. Process c!e.p will send the value of the expression e and then behaves as p. The operator + represents choice, while × represents parallel composition. Process be 5 p1 , p2 behaves as p1 if the value of be is true, and as p2 otherwise. The operator \L is the standard restriction for a set of channels L. Finally, P (x1 , . . . , xn ) is a process constant, which has an associated definition P (x1 , . . . , xn ) ≡ p. We assume the usual definitions of free variables f v(p) and bound variables bv(p) of processes. A process p is closed iff f v(p) = ∅. In the following, we denote by capital letters T, P . . . open processes. For recursive processes P (x1 , . . . , xn ) ≡ p we assume f v(p) ⊆ {x1 , . . . , xn } and we assume the body to be guarded. The concrete semantics of processes is defined in Taa ble 1 of the appendix as a labelled transition system LT S(p) = (P ∗ , 7→) with actions a ∈ Act = {τ, c?v, c!v | c ∈ Chan, v ∈ V al}. Two semantic functions Sv : V exp → V al and Sb : Bexp → {tt, f f } for the evaluation of expressions are used. For a ∈ Act, we define chan(τ ) = ∅, chan(c?v) = chan(c!v) = {c}. Moreover, we denote by a ¯ the symmetric action of a, namely c!v for c?v and c?v for c!v. Note that, if the set of values is infinite, the labelled transition system is infinite and infinitely branching. For expressing temporal properties of processes we consider a simple extension of propositional µ-calculus [14]. Let Act be a set of actions and V AR be a set of logical variables. Formulas are inductively defined as follows. A ::= X | A ∧ A |< K > A | ¬A | µX.A where X ∈ V AR is a logical variable and K ∈ {τ, c?V, c!V | c ∈ Chan, V ⊆ V al}. We assume that each variable X occurs positively in formulas. The modality < K > subsumes the classical existential next modality < a > ranging over actions a ∈ Act and corresponds to ∨a∈K < a >. The dual universal modality is [K] ≡ ¬ < K > ¬. The operator µX.A denotes the least fixpoint and the dual operator of greatest fixpoint is equivalent to νX.A ≡ ¬µX.A[¬X/X]. Note that formulas with K infinite are needed for subsuming the logics CT L and CT L∗ , since Act can be infinite. For W instance, the classical W liveness property ∀F A can be expressed as µX.A ∨ [τ ]X c∈Chan [c?V al]X c∈Chan [c!sV al]X. Traditional global model checking corresponds to compute the semantics of a the formula k A k on the concrete labelled transition system LT S(p) = (P ∗ , 7→). ∗ ∗ Let δ : V AR → P(P ) be a valuation assigning subsets of P to logical variables. The semantics of an open formula A with respect to δ is defined as: k X kδ = δ(X) k A0 ∧ A1 kδ =k A0 kδ ∩ k A1 kδ k< K > A kδ = ∪a∈K k< a >k (k A kδ )

k ¬A kδ = P \ k A kδ k µX.A kδ = µV.(k A kδ[V /X] )

A Symbolic Semantics for Abstract Model Checking

137

where δ[V /X] stands for the valuation δ 0 , which agrees with δ except that δ 0 (V ) = δ(X). The next modality function k< a >k: P(P ∗ ) → P(P ∗ ) is given by k< a >k a (S) = {p ∈ P | ∃p 7→p0 , p0 ∈ S}. Therefore, p ∈k A k (p |= A) iff p ∈k A kδ0 , where δ0 is the empty evaluation.

3

Abstract Interpretation Theory

In this section we briefly recall the basic ideas of abstract interpretation, we refer the reader to [6,7] for more details. The theory of abstract interpretation provides a systematic method to design approximate semantics of programs by replacing the concrete domain of computation with a simpler abstract domain. The relation between the concrete and the abstract domain is precisely stated into a formal framework. Let (C, ≤) and (A, ≤# ) be two posets, where orderings ≤ and ≤# correspond to precision. A pair of functions (α, γ), where α : C → A (abstraction) and γ : A → C (concretization) is called a Galois connection iff ∀c ∈ C, ∀a ∈ A, α(c) ≤# a ⇔ c ≤ γ(c). These requirements can also be captured by saying that α is extensive (c ≤ γ(α(c))), γ is reductive (α(γ(a)) ≤# a), α and γ are total and monotonic. If α(γ(a)) = a, then (α, γ) is called a Galois insertion. Intuitively, the condition c ≤ γ(α(c)) ensures the loss of information of abstraction to be safe. On the other hand, condition α(γ(a)) ≤# a ensures that the concretization process introduce no loss of information. Let S(P ) be the semantics of a program P computed as the least fixpoint of a semantic function F over the concrete domain (C, ≤). The goal is that of computing an approximate semantics S # (P ) over (A, ≤# ), that is safe α(S(P )) ≤# S # (P ). The main result is that a safe approximate semantics S # (P ) can be computed as the least fixpoint of a safe approximate semantic function F # over (A, ≤# ), such that α(F (c)) ≤# F # (α(c)), for each c ∈ C. Moreover, it has been shown there there exists always a best approximate semantic function F # (optimal), when α(F (c)) = F # (α(c)). In the Galois insertion case this is equivalent to α(F (γ(a)) = F # (a), for each a ∈ A.

4

The Symbolic Graph

In this section we define the symbolic graph of processes. Symbolic semantics as introduced in [12] relies on the idea of using symbolic actions instead of concrete actions. For instance, transitions modeling input are represented by a single c?x transition c?x.p 7→ p. To handle conditional open processes symbolic actions (c,θ)

depend in addition on boolean guards. Thus, a symbolic transition T 7→ T 0 represents all concrete transitions with action a corresponding to symbolic action θ for the assignments to the free variables of T such that guard c is satisfied. Our definition of the symbolic graph differs in some aspects from the classical one. First, classical branching is replaced by explicit relations of non-deterministic (⊕) and alternative choices (⊗) among transitions. Moreover, a method to avoid

138

F. Levi

infinite applications of the standard rule for recursion is proposed. This way a finitely branching and finite graph for regular processes is obtained. We use the standard notions of substitutions and environments. A substitution is a partial function σ : V ar → V Exp and a simple substitution is an injective σ : V ar → V ar. We denote by Sub the set of simple substitutions. For a substitution σ we denote by tar(σ) and dom(σ) its target and source, respectively. An environment is a total function ρ : V ar → V al and Env is the set of environments. For an environment ρ we denote by ρ[x → v] the environment that agrees with ρ except that the value assigned to variable x is v. Let T be an open process and x ∈ V ar. We say that x is fresh in T iff x 6∈f v(T )∪bv(T ). Moreover, we say that a term T is free for a simple substitution σ with dom(σ) ⊆ f v(T ), iff for each x ∈ f v(tar(σ)), x 6∈ bv(T ) ∪ (f v(T ) \ tar(σ)). Let T be an open process and ρ an environment. We denote by (T, ρ) the closed process T ρ, that is obtained by substituting ρ(x) to variable x for each x ∈ f v(T ). Let C be the set of constraints with c ::= be | e = e | e ∈ V | true | ¬c | c ∧ c, where e ∈ V exp, be ∈ Bexp and V ⊆ V al. We denote by vars(c) the variables occurring in constraint c. For c ∈ C we consider the semantic function k c k: P(Env) → P(Env) obtained in the trivial way by k be k= {ρ ∈ Env | Sb (beρ) = tt}, k e1 = e2 k= {ρ ∈ Env | Sv (e1 ρ) = Sv (e2 ρ)} and k e ∈ V k= {ρ ∈ Env | Sv (eρ) ∈ V }. We use the notation ρ |= c for ρ ∈k c k. Let the symbolic actions be SymAct = {c?x, c!e, τ | c ∈ Chan, x ∈ V ar, e ∈ V exp} with bv(c?x) = {x}, bv(c!e) = bv(τ ) = ∅ and f v(c!e) = vars(e), f v(τ ) = f v(c?x) = ∅. Transitions of the symbolic semantics are T for each i ∈ {1, n} (i ∈ {1, . . . , n})

⊗i∈{1,n} Θi

−→

⊗i∈{1,n} Ωi , such that

1. Θi = (ci , ⊕ji ∈{1,ni } θi,ji ) with θi,ji ∈ SymAct ∪ {∗} and ci ∈ C; 2. Ωi = ⊕ji ∈{1,ni } Ti,ji where Ti,ji are open processes; 3. vars(ci ) ⊆ f v(T ), f v(θi,ji ) ⊆ f v(T ) and f v(Ti,ji ) ⊆ f v(T ) ∪ bv(θi,ji ). All possible behaviors of closed processes obtained from T are represented by ⊗i∈{1,n} Θi

−→ ⊗i∈{1,n} Ωi , where alternative choices are related a single transition T by ⊗ and non-deterministic choices by ⊕. The idea is that for each environment ρ there exists a unique alternative ci that is satisfied and symbolic actions θi,ji with corresponding processes Ti,ji represent the concrete transitions of (T, ρ). Let us explain informally the construction of transitions. The complete semantic rules are shown in Table 2 of the appendix. (true,a⊕∗)

−→ T ⊕ Transitions of a basic process a.T are obtained by a rule a.T a.T , where the special action ∗ denotes idle action and is used in the parallel composition rule. The non-deterministic choice of + is reflected by the composition of transitions with ⊕. The rule for choice is as follows {Ti T1 + T2

⊗j ∈{1,n } Θi,ji i i

−→

⊗ji ∈{1,ni } Ωi,ji }i∈{1,2}

⊗i∈{1,2},j ∈{1,n } Θj+ ,j i i 1 2

−→

⊗i∈{1,2},j∈{1,ni } Ωj+1 ,j2

+

A Symbolic Semantics for Abstract Model Checking

139

where Θj+1 ,j2 is guarded by constraint c1,j1 ∧c2,j2 and where all non-deterministic choices of Θi,ji = (ci,ji , ⊕hi ∈{1,kji } θi,ji ,hi ) for i ∈ {1, 2} are merged by ⊕. The resulting processes are analogously combined by Ωj+1 ,j2 . In fact, for each environment that satisfies both guards both actions of T1 and of T2 can be chosen. Example 1. For instance, a process T = c!x.T1 + a!x.T2 is modeled by c!x.T1 + a!x.T2

(true,c!x⊕a!x⊕∗)

−→

T1 ⊕ T2 ⊕ T.

The alternative choices of a conditional process are related by ⊗ through the following rule {Ti

⊗j ∈{1,n } Θi,ji i i

−→

be 5 T1 , T2

⊗ji ∈{1,ni } Ωi,ji }i∈{1,2}

¯ i,j ⊗i∈{1,2},j ∈{1,n } Θ i i i

−→

5

¯i,j ⊗i∈{1,2},j∈{1,ni } Ω i

¯ 1,j is equivalent to Θ1,j where the guard is additionally constrained by where Θ 1 1 ¯ 2,j is equivalent to Θ2,j where the guard is additionally constrained be, while Θ 2 2 by ¬be. Intuitively, transitions of be 5 T1 , T2 are either transitions of T1 , if be is satisfied, or transitions of T2 , if be is not satisfied. Θ ⊗Θ

1 2 Example 2. For T = x > o 5 c!x.T1 + a!x.T2 , b!x.nil we have T −→ Ω1 ⊗ Ω 2 , where Θ1 = (x > 0, c!x ⊕ a!x ⊕ ∗), Θ2 = (x ≤ 0, b!x ⊕ ∗), Ω1 = T1 ⊕ T2 ⊕ T and Ω2 = nil ⊕ T . In fact, for each environment ρ either x > 0 or x ≤ 0 is true and the process (T, ρ) is able to perform either both c!ρ(x) and a!ρ(x) or b!ρ(x), respectively.

The rule of parallel composition is quite complex. Since a single transition must represent all behaviors combined by the relations ⊕ and ⊗, a single rule performs at the same time synchronization and interleaving. The rule is defined as follows {Ti T1 × T2

⊗j ∈{1,n } Θi,ji i i

−→

⊗ji ∈{1,ni } Ωi,ji }i∈{1,2}

⊗i∈{1,2},j ∈{1,n } Θj× ,j i i 1 2

−→

×

⊗i∈{1,2},ji ∈{1,ni } Ωj×1 ,j2

Θj×1 ,j2

where is constrained by c1,j1 ∧ c2,j2 and its actions are all possible combinations of actions θ1,j1 ,h1 and θ2,j2 ,h2 for Θi,ji = (ci,ji , ⊕hi ∈{1,kji } θi,ji ,hi ) for i ∈ {1, 2}. The combination of processes corresponding to the combination of actions is realized by Ωj×1 ,j2 . Example 3. Consider an open process T1 ×T2 with T1 = x > 05c?x.T, a?x.T and T2 = c!y + 1.T + a!y − 1.T . We have T1 (true,c!y+1⊕a!y−1⊕∗)

(x>0,c?x⊕∗)⊗(x≤0,a?x⊕∗)

−→

T ⊕ T 1 ⊗ T ⊕ T1

and T2 −→ T ⊕ T ⊕ T2 . The transition resulting from parallel Θ1 ⊗Θ2 composition is T1 × T2 −→ Ω1 ⊗ Ω2 , where Θ1 = (x > 0, τ ⊕ c?z ⊕ c!y + 1 ⊕ a!y − 1 ⊕ ∗), Θ2 = (x ≤ 0, τ ⊕ a?z ⊕ c!y + 1 ⊕ a!y − 1 ⊕ ∗), Ω1 = T [y + 1/x] × T ⊕ T [z/x] × T2 ⊕ T1 × T ⊕ T1 × T ⊕ T1 × T2 and Ω2 = T [y − 1/x] × T ⊕ T [z/x] × T2 ⊕ T1 × T ⊕ T1 × T ⊕ T1 × T2 . Two alternative choices corresponding to constraints x > 0 and x ≤ 0 arise from composition of guard x > 0 with true and of guard

140

F. Levi

x ≤ 0 with true, respectively. For guard x > 0 the non-deterministic choices are obtained by the composition of actions c?x and ∗ with c!y +1, a!y −1 and ∗, while for guard x ≤ 0 they are obtained by the composition of actions a?x and ∗ with c!y + 1, a!y − 1 and ∗. Therefore, τ actions are obtained from the synchronization of actions c?x and c!y + 1 and of a?x with a!y − 1, respectively. The others arise from interleaving by composition with ∗. The corresponding processes are obtained in the obvious way. For instance, the process corresponding to τ with guard x > 0 is T [y +1/x]×T , since the value of y −1 is received by T1 . Note that in the interleaving case with a receive action the variable x must be renamed to z to avoid clash of variables with the free variables of T2 . Recursive processes are handled by the classical rule, where formal parameters are substituted by actual parameters. T [¯ e/¯ x] P (¯ e)

⊗i∈{1,n} Θi

−→

⊗i∈{1,n} Θi

−→

⊗i∈{1,n} Ωi

x) ≡ T rec P (¯

⊗i∈{1,n} Ωi

Unfortunately, the application of rule rec leads to an infinite graph also for regular processes, where there is no parallel composition inside the scope of recursion. The semantics of [12] suffers of the same problem. Example 4. Let us consider the process P (x) ≡ c!x.(P (x + 1) + P (x − 1)). Since the recursive process is unfolded infinitely times with a different argument an infinite graph arise. P (x)

(true,c!x⊕∗)

−→

P (x + 1) + P (x − 1) ⊕ P (x)

P (x + 1) + P (x − 1) P (x + 1) + P (x − 1) ...

(true,c!x+1⊕∗)

−→

P (x + 1 + 1) + P (x + 1 − 1) ⊕ P (x + 1)

−→

P (x − 1 + 1) + P (x − 1 − 1) ⊕ P (x − 1)

(true,c!x−1⊕∗)

This problem can be solved by replacing in the graph transitions of P (¯ e) by transitions of a process P (¯ x) for fresh variables x ¯. The semantics of P (¯ e) can naturally be obtained by the semantics of P (¯ x) by instantiating parameters x ¯ to the actual values corresponding to the evaluation of the expressions e¯. Let general processes GP have the following syntax GP ::= nil | a.T | P (¯ x) | GP1 + GP2 | be 5 GP1 , GP2 | GP1 × GP2 | GP \ L where x ¯ is a tuple of distinct variables and T ∈ P roc is a process. Note that a.T is a general process for any process T , since there are no current recursive calls. For general processes GP ∈ GP, the recursion variables rv(GP ) are defined as rv(a.T ) = rv(nil) = ∅, rv(P (¯ x)) = {¯ x}, rv(GP1 +GP2 ) = rv(GP1 × GP2 ) = rv(be 5 GP1 , GP2 ) = rv(GP1 ) ∪ rv(GP2 ) and rv(GP \ L) = rv(GP ). Our aim is that of finding a general process GP that can be used instead of T in the graph. For this purpose we define most general terms.

A Symbolic Semantics for Abstract Model Checking

141

Definition 5. Let T be a process. We define most general terms Π(T ): – if T = nil or T = a.T1 , Π(T ) = {T }; – for GPi ∈ Π(Ti ) such that, rv(GP1 ) ∩ f v(GP2 ) = ∅, rv(GP1 ) ∩ bv(GP2 ) = ∅ and vice-versa, then GP1 + GP2 ∈ Π(T1 + T2 ), GP1 × GP2 ∈ Π(T1 × T2 ) and be 5 GP1 , GP2 ∈ Π(be 5 T1 , T2 ); – if P (¯ z ) ≡ T and T is free for [¯ x/¯ z ], then P (¯ x) ∈ Π(P (¯ e)). Most general terms are general processes obtained by introducing fresh and distinct variables in current recursive calls. The following property is satisfied. Proposition 6. Let T be a process. For each GP ∈ Π(T ), there exists a substitution σ with dom(σ) = rv(GP ) such that GP σ = T . The substitution σ assigns actual parameters of T to formal parameters of GP . For ρ ∈ Env, let genT,GP : Env → Env, such that ρ(x) = genT,GP (ρ)(x), for x 6∈rv(GP ), and ρ(x) = Sv (σ(x)ρ), otherwise. In the environment genGP,T (ρ) formal parameters of GP are instantiated to the values of the actual parameters provided by σ. The result is that the closed process (GP, genT,GP (ρ)) is equivalent (bisimilar) to (T, ρ). For p1 , p2 ∈ P roc be processes, we say that p1 ≡ p2 a a iff for each a ∈ Act, for each p1 −→ p01 there exists p2 −→ p02 and p01 ≡ p02 and vice-versa. Proposition 7. Let T be a process and ρ ∈ Env. For each GP ∈ Π(T ), (GP, genT,GP (ρ)) ≡ (T, ρ). By proposition 2 the symbolic graph where transitions of T are replaced by transitions of GP correctly models the behavior of processes. Definition 8. Let T be a process. We define SG(T ) = (GP ∗ , T ∗ , ∗

∗

GP ⊆ GP, T ⊆ P roc and transitions are GP GP ∈ GP ∗ , where

⊗i∈{1,n} Θi

−→

⊗i∈{1,n} Θi

−→

) with

⊗i∈{1,n} Ωi for each

1. for each T 0 ∈ T ∗ there exists GP 0 ∈ GP ∗ ∩ Π(T 0 ); ⊗i∈{1,n} Θi

−→ ⊗i∈{1,n} Ωi , where Ωi = ⊕ji ∈{1,ni } Ti,ji , 2. T ∈ T ∗ and for each GP Ti,ji ∈ T ∗ , for i ∈ {1, n}, ji ∈ {1, ni }. Example 9. Consider for instance the process P (1), where P (x) is defined in example 4. Since P (z) + P (w) ∈ Π(P (x + 1) + P (x − 1)) then SG(P (1)) is finite: P (x)

(true,c!x⊕∗)

−→

P (z) + P (w)

P (x + 1) + P (x − 1) ⊕ P (x)

(true,c!z⊕c!w⊕∗)

−→

P (z + 1) + P (z − 1) ⊕ P (w + 1) + P (w − 1) ⊕ P (z) + P (w)

This graph correctly describes the behavior of P (1). For instance, the concrete c!1

c!2

computation P (1) 7→P (1 + 1) + P (1 − 1) 7→P (1 + 1 + 1) + P (1 + 1 − 1) . . . c!1

c!2

is simulated by (P (x), ρ1 ) −→ (P (z) + P (w), ρ2 ) −→ . . ., where ρ1 (x) = 1 and ρ2 (z) = 2, ρ2 (w) = 0. Environment ρ2 = genT,GP (ρ1 ) assigns to parameters z and w the result of the evaluation of expressions x + 1 and x − 1 with respect to ρ1 . Note that in P (z) + P (w) distinct fresh variables are used to model all recursive calls P (e1 ) + P (e2 ), where e1 may be different from e2 .

142

F. Levi

The equivalence is formally stated by the following theorem. Let K ∈ {τ, c!V, c?V } and θ ∈ SymAct. We define the constraint θ ∈ K ∈ C as θ ∈ K ≡ true, for K = θ = τ , θ ∈ K ≡ true, for θ = c?x and K = c?V , θ ∈ K ≡ e ∈ V , for θ = c!e and K = c!V , and θ ∈ K ≡ f alse, otherwise. For K = a we denote by a = θ the constraint θ ∈ a. ⊗i∈{1,n} Θi

−→ ⊗i∈{1,n} Ωi , where Theorem 10. 1. For each symbolic transition GP Θi = (ci , ⊗ji ∈{1,ni } θi,ji ) and Ωi = ⊗ji ∈{1,ni } Ti,ji and for each ρ ∈ Env, there exists one and only one i ∈ {1, n}, such that ρ |= ci and, for each c?v

ji ∈ {1, ni }, there exists a ∈ Act with ρ |= a = θi,ji such that, (GP, ρ) 7→ a (Ti,ji , ρ[x → v]) for θi,ji = c?x and v ∈ V al, and (T, ρ) 7→ (Ti,ji , ρ), for θi,ji ∈ {τ, c!e}; 2. for each p ∈ P ∗ there exists GP ∈ GP ∗ and ρ ∈ Env such that GP ρ ≡ p and, ⊗i∈{1,n} Θi

a

−→ for each p 7→p0 , there exist i ∈ {1, n} and ji ∈ {1, ni }, where GP ⊗i∈{1,n} Ωi with Θi = (ci , ⊗ji ∈{1,ni } θi,ji ) and Ωi = ⊗ji ∈{1,ni } Ti,ji , and ρ |= ci ∧ a = θi,ji and p0 ≡ (Ti,ji , ρ[x → v]), for a = c?v and θi,ji = c?x, and p0 ≡ (Ti,ji , ρ), for a ∈ {τ, c!v}. By theorem 1 the concrete semantics is safely represented by the symbolic one and in addition concrete non-deterministic and alternative choices are exactly composed by ⊕ and ⊗ in symbolic transitions. Since transitions are restricted to general processes, infinite applications of rec with different arguments are avoided and the symbolic graph is finite for regular processes. Theorem 11. Let p ∈ P roc be a regular process. SG(p) is a finite graph up to renaming.

5

Symbolic Model Checking

In this section we show that model checking can be realized by interpreting the symbolic graph over concrete environments. Let (P(D), ⊆), with D = {(T, ρ) | ⊗i∈{1,n} Θi

T ∈ GP ∗ and ρ ∈ Env} for SG(p) = (GP ∗ , T ∗ , −→ ). The semantics of a formula A is defined as in Sect. 2 by replacing evaluations with symbolic evaluations δ : V AR → P(D) and by taking the following function for the next modality. For K ∈ {τ, c!V, c?V } and S ∈ P(D), ⊗i∈{1,n} Θi

– k< K >kS (S) = {(T, ρ) | T −→ ⊗i∈{1,n} Ωi , with Θi = (ci , ⊕ji ∈{1,ni } θi,ji ) and Ωi = ⊕ji ∈{1,ni } Ti,ji , and there exist i ∈ {1, n} and ji ∈ {1, ni }, such that ρ |= ci ∧ θi,ji ∈ K and (Ti,j , ρ) ∈≡ S, for θi,ji ∈ {τ, c!e}, and (Ti,ji , ρ[x → v]) ∈≡ S, for some v ∈ V for K = c?V and θi,ji = c?x}. where (T, ρ) ∈≡ S if (GP, genT,GP (ρ)) ∈ S for some GP ∈ Π(T ). The definition of k< a >kS is based on the observation that process (T, ρ) is able to perform an action a, if the environment satisfies a guard ci for which there exists a symbolic action θi,ji corresponding to a. Moreover, the resulting process must be in S. However, since S is a set of general processes and Ti,ji is not a general process we look for an equivalent process (GP, genTi,ji ,GP (ρ)) ∈ S.

A Symbolic Semantics for Abstract Model Checking

143

Theorem 12. For each closed formula A, k A kS ∩P ∗ =k A k.

6

Abstract Model Checking

In this section we define the approximate semantics of the collecting semantics k A kS on the abstract domain obtained by replacing concrete environments with abstract environments on an abstract values domain. Let (αv , γv ) be a Galois insertion between the complete lattices (P(V al), ⊆) and (P(V al# ), ⊆) of concrete and abstract values. We consider the set of abstract environments Env # = {ρ# | ρ# : V ar → V al# } and the set of abstract processes ⊗i∈{1,n} Θi

D# = {(T, ρ# ) | T ∈ GP ∗ and ρ# ∈ Env # } for SG(p) = (GP ∗ , T ∗ , −→ ). Let γe : P(Env # ) → P(Env) and γ : P(D# ) → P(D) be the obvious functions induced by γv . Our purpose is that of computing a lower approximation of the semantics k A kl , such that γ(k A kl ) ⊆k A kS . Given the relation between concrete and abstract domain an approximate semantics can be obtained by replacing concrete functions with corresponding safe abstract functions. However, the operator of negation is not monotonic. Kelb [13] argues that a lower approximation of the full logic with negation can be obtained by combining duals approximations for formulas without negation k A kl and k A ku . In fact, k ¬A kl = D# \ k A ku is a safe lower approximation of k ¬A kS , since k A kS ⊆ γ(k A ku ). Analogously, k ¬A ku = D# \ k A kl is a safe upper approximation of k ¬A kS . Thus, the problem is reduced to the definition of safe dual approximations for all the logical operators except negation. The safeness of the dual approximations is formally defined with respect to the following framework. Proposition 13. There exist αl , αu : P(D) → P(D# ) such that (αu , γ) is a Galois insertion between (P(D), ⊆) and (P(D# ), ⊆) and (αl , γ) is a Galois insertion between (P(D), ⊇) and (P(D# ), ⊇). We show that both the upper and the lower approximation can be computed over the symbolic graph with respect to abstract environments. We exploit the following constraints based on relations ⊕ and ⊗. The symbol ∨ denotes with an abuse of notation the equivalent constraint. ⊗i∈{1,n} Θi

−→ ⊗i∈{1,n} Ωi , with Θi = (ci , ⊗ji ∈{1,ni } θi,ji ) and Definition 14. Let T Ωi = ⊗ji ∈{1,ni } Ti,ji , be a symbolic transition. For each K ∈ {τ, c!V, c?V } and i ∈ {1, n}, ji ∈ {1, ni } and I = ∪i∈{1,n} Ii with Ii ⊆ {1, ni } we define 1. F REEK,i,jiV≡ ci ∧ θi,ji ∈ K; W 2. CONK,I ≡ i∈{1,n} ¬ci ∨ ( ji ∈Ii θi,ji ∈ K). The abstract semantics is based on safe lower and upper approximations of the previous constraints and on dual approximations of the relation ∈≡ . Note that for each c ∈ C, k c ku and k c kl are safe approximations iff k c k⊆ γe (k c ku ) and γe (k c kl ) ⊆k c k, respectively. Moreover, ∈u≡ is a safe upper approximation,

144

F. Levi

if ρ ∈ γe (ρ# ) such that (T, ρ) ∈≡ γ(S # ) implies (T, ρ# ) ∈u≡ S # , while ∈l≡ is a safe lower approximation if (T, ρ# ) ∈l≡ S # only if, for all ρ ∈ γe (ρ# ), (T, ρ) ∈≡ γ(S # ). Let δ # : V AR → P(D# ) be an abstract evaluation. We define the abstract upper and lower semantics of an open formula A with respect to δ # as follows. Lower Approximation k X klδ# = δ # (X) k A0 ∧ A1 klδ# =k A0 klδ# ∩ k A1 klδ# k< K > A klδ# =k< K >kl (k A klδ# )

k µX.A klδ# = µV.(k A klδ# [V /X] ) k ¬A klδ# = D# \ (k A kuδ# )

where, for S # ∈ P(D# ), ⊗i∈{1,n} Θi

−→ ⊗i∈{1,n} Ωi , with Θi = (ci , ⊗ji ∈{1,ni } – k< K >kl (S # ) = {(T, ρ# ) | T θi,ji ) and Ωi = ⊗ji ∈{1,ni } Ti,ji , and there exists I = ∪i∈{1,n} Ii with Ii ⊆ {1, ni }, such that I is minimal and ρ# ∈k CONK,I kl , and, for each i ∈ {1, n} and ji ∈ Ii , (Ti,ji , ρ# ) ∈l≡ S # , for θi,ji ∈ {c!e, τ }, and (Ti,ji , ρ# [x → αv (v)]) ∈l≡ S # , for v ∈ V , K = c?V and θi,ji = c?x}. Upper Approximation k X kuδ# = δ # (X) k A0 ∧ A1 kuδ# =k A0 kuδ# ∩ k A1 kuδ# k< K > A kuδ# =k< K >ku (k A kuδ# )

k µX.A kuδ# = µV.(k A kuδ# [V /X] ) k ¬A kuδ# = D# \ (k A klδ# )

where, for S # ∈ P(D# ), ⊗i∈{1,n} Θi

−→ ⊗i∈{1,n} Ωi with Θi = (ci , ⊕ji ∈{1,ni } – k< K >ku (S # ) = {(T, ρ# ) | T θi,ji ) and Ωi = ⊕ji ∈{1,ni } Ti,ji , and there exist i ∈ {1, n} and ji ∈ {1, ni } such that ρ# ∈k F REEK,i,ji ku and (Ti,ji , ρ# ) ∈u≡ S # , for θi,ji ∈ {τ, c!e}, and (Ti,ji , ρ# [x → αv (v)]) ∈u≡ S # , for v ∈ V , K = c?V and θi,ji = c?x }. Let us explain the dual approximations of the next modality. 1. The upper approximation is safe, if (T, ρ) ∈k< a >kS (γ(S # )) implies (T, ρ# ) ∈k< a >ku (S # ) for ρ ∈ γe (ρ# ). In other words, safeness requires to consider at least the abstract transitions corresponding to the concrete a-transitions (free transition relation [10,4]). If constraint F REEa,i,ji ≡ ci ∧ θi,ji = a is safely upper approximate the previous condition is guaranteed, since ρ |= ci ∧ θi,ji = a implies ρ# ∈k ci ∧ θi,ji = a ku . 2. The definition of the lower approximation is quite complex. In this case safeness is ensured, only if (T, ρ# ) ∈k< a >kl (S # ) implies (T, ρ) ∈k< a >kS (γ(S # )), for each ρ ∈ γe (ρ# ). In other words, safeness requires to consider only the abstract transitions, for which all corresponding concrete a-transition exist (constrained transition relation [10,4]). Let us consider the

A Symbolic Semantics for Abstract Model Checking

145

V W constraint CONa,I ≡ i∈{1,n} (ci ⊃ ( ji ∈Ii (a = θi,ji ))) and its lower approximation. If ρ# ∈k CONa,I kl , then for each ρ ∈ γe (ρ# ), if ρ |= ci there exists an action θi,ji with ji ∈ Ii corresponding to a. By theorem 1 for each ρ there exists exactly one ci such that ρ |= ci so that, for each ρ there exists indeed a transition with action a. Note that by definition of the next modality safeness for both approximations requires in addition safeness of the dual approximations of ∈≡ to check the resulting processes to be in S # . Example 15. Consider a process T = x > 0 5 c?x.T1 , a?x.T2 , whose behavior is (x>0,c?x⊕∗)⊗(x≤0,a?x⊕∗)

described by T −→ T1 ⊕T ⊗T2 ⊕T . Let the values abstraction be αv (n) = • and let ρ# be the abstract environment with ρ# (x) = •. For a formula A ≡< c?n > true we obtain an upper approximation {(T, ρ# )} = k A ku . In fact, there exists ρ ∈ γe (ρ# ) such that ρ |= x > 0 ∧ (c?x = c?n) ≡ x > 0 ∧ true so that ρ# ∈k x > 0 ∧ true ku . Analogously, {(T, ρ# )} =k< a?n > true ku . The abstract operator is safe, since the existence of both concrete c?n

a?n

transitions (T, ρ1 ) 7→ (T1 [n/x], ρ1 ) and (T, ρ2 ) 7→ (T2 [n/x], ρ2 ) for ρ1 , ρ2 ∈ γe (ρ# ) is captured. Note that in the abstract process non-determinism among the two actions c?n and a?n arise even if in the concrete case these are two alternative choices. In contrast, let us consider the lower approximation for formula A. Since there a?n c?n exists ρ1 , ρ2 ∈ γe (ρ# ) such that (T, ρ1 ) 67→ and (T, ρ2 ) 67→ the lower approximation is safe if and only if (T, ρ# ) 6∈kA kl . Since k x > 0 kl = ∅, k x ≤ 0 kl = ∅ and c?n = a?x ≡ f alse, ρ# 6∈k(x ≥ 0 ∨ true) ∧ (x < 0 ∨ f alse) kl . Therefore, we have both k< c?n > true kl = ∅ and k< a?n > true kl = ∅. The abstract operator is able to observe that there is no real non-determinism between a?n and c?n, while this is an alternative choice as expressed by ⊗ and there are concrete processes for both alternatives. On the other hand, consider the values abstraction with αv (n) = P os, if n > 0, and αv (n) = N eg, otherwise. There are two abstract environments ρ# 1 , # # # with ρ# (x) = P os, and ρ , with ρ (x) = N eg. Since for each ρ ∈ γ(ρ 1 2 2 1 ), # l ρ ∈k c k for c = (x > 0 ⊃ true) ∧ (x ≤ 0 ⊃ f alse), then ρ1 ∈k c k is safe. With this evaluation of constraint we can safely obtain k< c?n > true kl = {(T, ρ# 1 )}. Since, for each concrete environment of ρ# only the alternative x > 0 is possible, 1 it is safe to conclude that all concrete processes indeed perform c?n. Lemma 16. Let k c kl and k c ku be safe lower and upper approximations for c ∈ C. Moreover, let ∈l≡ and ∈u≡ be safe lower and upper approximations of ∈≡ . For each S # ∈ P(D# ) and K ∈ {τ, c?V, c!V }, αu (k< K >kS (γ(S # ))) ⊆k< K >ku (S # ) and k< K >kl (S # ) ⊆ αl (k< K >kS )(γ(S # )). By lemma 1 the lower approximation of the full logic is safe. Theorem 17. Let k c kl and k c ku be safe lower and upper approximations for c ∈ C. Moreover, let ∈l≡ and ∈u≡ be safe lower and upper approximations of ∈≡ . For each closed µ-calculus formula A, k A kS ⊇ γ(k A kl ).

146

7

F. Levi

About Optimality

We have defined a method for constructing safe dual approximations of the next modality, that exploit dual approximations of constraints and of relation ∈≡ . It this setting precision of abstract model checking depends on precision of these approximations. An interesting problem is that of finding conditions on the approximations of constraints and of ∈≡ , that guarantee optimality of next modalities. It turns out that optimality of the upper approximation of constraint F REEK,i,ji and of the lower approximation of constraint CONK,I is sufficient, whenever SP(p) contains only general processes. ⊗i∈{1,n} Θi

Lemma 18. Let p be a closed process such that SP(p) = (GP ∗ , T ∗ , −→ ) where GP ∗ = T ∗ . Let k c ku and k c kl be optimal upper and lower approximations for c ∈ C. For each S # ∈ P(D# ) and K ∈ {τ, c?V, c!V }, αu (k< K >kS (γ(S # ))) ⊇k< K >ku (S # ) and αl (k< K >kS (γ(S # ))) ⊆k< K >kl (S # ). In the upper approximation case this result is quite obvious, since by optimality ρ# ∈k F REEa,i,ji ku implies the existence of an environment such that ρ ∈k F REEa,i,ji k, namely the existence of a corresponding concrete transition. The lower approximation is optimal, whenever there exists a constrained transition if and only if for each concrete process there exists a corresponding concrete a transition. Suppose that for each concrete process (T, ρ) 7→(T 0 , ρ0 ). For each ρ there exists a choice such that constraint ci is satisfied and a non-deterministic choice such that ρ |= θi,ji = a. Optimality is guaranteed, since theorem 1 ensures in addition that for each ρ all others alternatives cj are not true. Therefore, for each environment the constraint CONa,I is indeed satisfied so that by optimality the corresponding abstract transition is certainly considered. In contrast, if SP(p) contains non-general processes lemma 2 is no longer valid. Problems arise from the abstract evaluation of parameters included in the definition of ∈≡ . It is sufficient to consider the case of ρ# , such that ∃ρ1 , ρ2 ∈ γe (ρ# ) such that ρ1 |= F REEa,i,ji and (Ti,ji , ρ2 ) ∈≡ γ(S # ), but ρ1 6= ρ2 . Depending on the domain of values and expressions specific solutions must be studied for approximating constraints and ∈≡ . We suggest a general strategy. Safe dual approximations of constraints can be found on the basis of dual approximations of basic constraints be, e = e and e ∈ V and by combining lower and upper approximations in the obvious way. Definition 19. For c ∈ C, let k ¬c ku = P(Env # )\ k ¬c kl , k ¬c kl = P(Env # )\ k ¬c ku , k c1 ∧ c2 kl = ∩i∈{1,2} k ci kl and k c1 ∧ c2 ku = ∩i∈{1,2} k ci ku . If basic constraints are safely approximate these approximations are obviously safe. Unfortunately, they are in general non-optimal, since αu does not preserve ∩, while αl does not preserve ∪. Moreover, we have to compute dual approximations of ∈≡ , that realizes pa# # rameters evaluation. Let gen# T,GP : Env → P(Env ) be a safe approximation # # of genT,GP , where genT,GP (ρ) ∈ γe (gen# T,GP (ρ )) for each ρ ∈ γe (ρ ). This function can be suitably used to define approximations of ∈≡ .

A Symbolic Semantics for Abstract Model Checking

147

Definition 20. Let gen# T,GP be a safe approximation. We define # # # # 1. (T, ρ# ) ∈u≡ S # iff there exists ρ# 1 ∈ genT,GP (ρ ) such that (GP, ρ1 ) ∈ S ;

# # # # 2. (T, ρ# ) ∈l≡ S # iff for each ρ# 1 ∈ genT,GP (ρ ), (GP, ρ1 ) ∈ S .

If gen# T,GP is safe, then the previous approximations of ∈≡ are safe. The difficulties for computing optimal dual approximations are obvious. However, it is important to stress the essential role of ⊗ and ⊕ for limiting the loss of information the lower approximation. The relations ⊗ and ⊕ allows to improve the precision of the lower approximation with respect to the “trivial” definition, where a lower approximation of constraint F REEK,i,ji is considered. Example 21. Consider the process T1 × T2 with T1 = x > 0 5 c?x.T, a?x.T and T2 = c!y+1.T +a!y−1.T of example 3. With respect to the abstraction αv (n) = •, for each n ∈ N at, we trivially obtain ρ# ∈k c kl for c = (x > 0 ⊃ (τ = τ )) ∧ (x ≤ 0 ⊃ (τ = τ )) and ρ# (x) = •, since τ = τ ≡ true. Therefore, (T1 × T2 , ρ# ) ∈k< τ > true kl is established. The lower approximation of constraint c captures that there are processes for both alternatives, but in both cases the action τ τ can be performed. In other words, a constrained transition (T1 × T2 , ρ# ) 7→. . . is constructed. If we consider the trivial definition of lower approximation, we would not able to prove it, since neither ρ# ∈k (x > 0 ∧ (τ = τ )) kl nor ρ# ∈k (x ≤ 0 ∧ (τ = τ )) kl , even if the evaluation of constraints is optimal. Thus, this trivial method succeeds only if the alternative choice is the same for each concrete processes. In contrast, due to ⊗ and ⊕ a weaker condition can be considered and a more precise result is achieved. The proposed method could not give the same results on a classical symbolic semantics, where branching represents both non-deterministic and alternative choices, since it exploits the existence of exactly one alternative choice for each process and the representation of all non-deterministic choices for each possibility.

8

Related Works

The combination of abstract interpretation and model checking has been the topic of intensive research in the last few years. Much of the work concerned the definition of safe abstract model, namely of safe abstract transition relations, that preserves universal properties [2,3,1,17] and both universal and existential properties [11,10,13,4,15]. [11,10,4] propose the use of constrained and free transitions to handle branching modalities and [10] tackles also the problem of effectively computing for very simple programs an abstract model. The proposed method suffers of the problems of the trivial definition shown in example 7. A slight different approach is the one of Kelb [13], that investigates conditions for the safeness of abstract µ-calculus semantics instead that for the safeness of abstract models. In order to handle non-monotonic negation the combination of dual approximations is suggested. However, the problem of computing safe dual approximations of the next modality even over a given abstract model is not

148

F. Levi

addressed. In the framework of value-passing concurrent processes [5] proposes an abstract labelled transition system semantics for abstract closed processes obtained by an abstraction on the values domain. It is not obvious which class of temporal properties is preserved by the abstract model. In [15] it has been introduced the idea of representing the relation of non-determinism and of alternative choice among actions in order to compute more precise safe and constrained transition relations in the framework of closed abstract processes rather than in a symbolic approach with environments. Schmidt [17] shows a methodology for computing a finite approximate semantics of value-passing CCS by finitely approximating the semantics over abstract environments as a regular tree. Such a semantics is adequate for the verification of universal properties only. As far as concern the symbolic semantics, it is worth mentioning that other approaches have been proposed for representing regular processes by finite graphs [16].

9

Conclusions

In this paper we have applied abstract interpretation to the verification of µcalculus properties of value-passing concurrent processes. The main contribution is the definition of a finite symbolic semantics of processes, that differs from the classical one [12] in some aspects. First, classical branching of transitions is replaced by explicit relations of alternative and non-deterministic choices among transitions. Moreover, infinite branches are avoided by representing current recursive calls by means of general processes. The concrete semantics of µ-calculus can suitably be computed by interpreting the symbolic graph over concrete environments, but due to infinite values it is not effectively computable. We have proposed a technique to compute a lower approximate semantics on the symbolic graph by replacing concrete environments with abstract environments. Following the approach of [13] for explicitly treating negation the lower approximate semantics has been obtained by combining lower and upper approximations for each operator of the logic except negation. The relations of non-deterministic and alternative choices turn out to be very useful to approximate the next modality. The lower approximation in particular results undoubtedly more precise than previous proposals [11,10]. With respect to the classical approach to abstract model checking the proposed method does not rely on the construction of a safe abstract model, but on the computation of safe approximations of the “model checking” functions over the symbolic graph. This approach has several advantages. In order to prove a property it would be typically necessary to subsequently refine the chosen value abstraction by adding more information. This way the construction of the new abstract model is avoided. Moreover, this approach fits well in the traditional abstract interpretation framework and allows us to reason about safety and precision without introducing ad-hoc conditions as for instance the approximation ordering between abstract models of [4,10]. Recently, Schmidt [18] following the ideas of [19] has pointed out the very close connection among abstract model checking and data-flow analysis. These results suggest that the methods from

A Symbolic Semantics for Abstract Model Checking

149

one area can be usefully used in the other. There are many directions to continue this research. For instance, it seems interesting to study whether the refinement operators [7] that have been designed to systematically construct new more precise domain can be applied to the abstract model checking framework. Acknowledgments. I would like to thank Radhia and Patrick Cousot and Dave Schmidt for their helpful suggestions on the presentation of this work.

References 1. S. Bensalem, A. Bouajjani, C. Loiseaux, and J. Sifakis. Property preserving simulations. In Proceedings of CAV 92, volume 663 of Lecture Notes in Computer Science, pages 260–263. Springer-Verlag, Berlin, 1992. 2. E.M. Clarke, O. Grumberg, and D.E. Long. Model Checking and Abstraction. In Proc. 19th Annual ACM Symp. on Principles of Programming Languages, pages 343–354. ACM Press, 1992. 3. E.M. Clarke, O. Grumberg, and D.E. Long. Model checking and abstraction. ACM Transactions on Programming Languages and Systems, 5(16):1512–1542, 1994. 4. R. Cleaveland, P. Iyer, and D. Yankelevic. Optimality in Abstractions of Model Checking. In Proceedings of SAS 95, volume 983 of Lecture Notes in Computer Science, pages 51–63. Springer-Verlag, Berlin, 1995. 5. R. Cleaveland and J. Riely. Testing based abstractions for value-based systems. In Proceedings of CONCUR 94, volume 836 of Lecture Notes in Computer Science, pages 417–432. Springer-Verlag, Berlin, 1994. 6. P. Cousot and R. Cousot. Abstract Interpretation: A Unified Lattice Model for Static Analysis of Programs by Construction or Approximation of Fixpoints. In Proc. Fourth ACM Symp. Principles of Programming Languages, pages 238–252, 1977. 7. P. Cousot and R. Cousot. Systematic Design of Program Analysis Frameworks. In Proc. Sixth ACM Symp. Principles of Programming Languages, pages 269–282, 1979. 8. D. Dams. Abstract Interpretation and Partition Refinement for Model Checking. PhD thesis, Eindhoven university of Technology, 1996. 9. D. Dams, R. Gerth, and O. Grumberg. Generation of reduced models for checking fragments of CT L. In Proceedings of CAV 93, volume 697 of Lecture Notes in Computer Science, pages 479–490. Springer-Verlag, Berlin, 1993. 10. D. Dams, R. Gerth, and O. Grumberg. Abstract Interpretation of Reactive Systems. ACM Transactions on Programming Languages and Systems, 19(2):253–291, 1997. 11. D. Dams, O. Grumberg, and R. Gerth. Abstract interpretation of reactive systems: Abstractions preserving ∀CT L∗ , ∃CT L∗ and CT L∗ . In Proceedings of the Working Conference on Programming Concepts, Methods and Calculi (PROCOMET), 1994. 12. M. Hennessy and H. Lin. Symbolic bisimulations. Theoretical Computer Science, 138:353–389, 1995. 13. P. Kelb. Model Checking and Abstraction: A framework preserving both truth and failure information. Technical report, OFFIS, Oldenburg, Germany, 1994. 14. D. Kozen. Results on the Propositional mu-Calculus. Theoretical Computer Science, 27:333–354, 1983.

150

F. Levi

15. F. Levi. Abstract model checking of value-passing processes. In A. Bossi, editor, International Workshop on Verification, Model Checking and Abstract Interpretation, 1997. http://www.dsi.unive.it/ bossi/VMCAI.html. 16. H. Lin. Symbolic Transition Graph with Assignment. In Proc. of CONCUR 96, volume 1119 of Lecture Notes in Computer Science, pages 50–65. Springer-Verlag, Berlin, 1996. 17. D.A. Schmidt. Abstract Interpretation of Small-Step Semantics. In Proc. of the LOMAPS Workshop on “Analysis and Verification of Multiple-Agent Languages”, volume 1192 of Lecture Notes in Computer Science, pages 76–99, 1996. 18. D.A. Schmidt. Data Flow Analysis is Model Checking of Abstract Interpretation. In Proc. of the Annual ACM Symp. on Principles of Programming Languages, pages 38–48. ACM Press, 1998. 19. B. Steffen. Data Flow Analysis as Model Checking. In A. Meyer, editor, Proceedings of Theoretical Aspects of Computer Software (TACS 91), volume 526 of Lecture Notes in Computer Science, pages 346–364. Springer-Verlag, Berlin, 1991.

A

Appendix Table 1. The concrete semantics τ

c!v τ.p 7→p c?x.p c?v 7→p[v/x] v ∈ V al c!e.p 7→p Sv (e) = v a

a

pi 7→p0i

+

a

p1 + p2 7→p0i a

a

p1 7→p01 a

p1 × p2 7→p01 × p2

p1 × p2 7→p01 × p02

sync

T [¯ e/¯ x] 7→p a

P (¯ e) 7→p

a

be 5 p1 , p2 7→p02

int2 a

x) ≡ T rec P (¯

p1 7→p01 a

be 5 p1 , p2 7→p01

51 Sb (be) = tt

a

a

p2 7→p02

a

p1 × p2 7→p1 × p02

a

a ¯

p1 7→p01 p2 7→p02 τ

int1

p2 7→p02

52 Sb (be) = f f

p 7→p0 a

p \ L 7→p0 \ L

\ chan(a) ∩ L = ∅

The semantic rules for the symbolic semantics are based on some operations over actions. Let Θi,ji = (ci,ji , θi,ji ,1 ⊕ . . . ⊕ θi,ji ,kji ) and Ωi,ji = Ti,ji ,1 ⊕ . . . ⊕ Ti,ji ,kji , for i ∈ {1, 2}, ji ∈ {1, ni }, we define – Θj+1 ,j2 = (c1,j1 ∧ c2,j2 , ⊕h1 ∈{K1 } θ1,j1 ,h1 ⊕h2 ∈{1,K2 } θ1,j2 ,h2 ⊕ {∗}) and Ωj+1 ,j2 = ⊕h1 ∈{K1 } T1,j1 ⊕h2 ∈{1,K2 } T1,j2 ⊕T1 +T2 , for Ki = {hi ∈ {1, kji } | θi,ji ,hi 6= ∗}; ¯ 2,j = (¬be ∧ c2,j , θ2,j ,1 ⊕ ¯ 1,j = (be ∧ c1,j , θ1,j ,1 ⊕ . . . ⊕ θ1,j ,k ) and Θ – Θ 1 1 1 1 j1 2 2 2 ¯ ¯ . . . ⊕ θ2,j2 ,kj2 ), and Ωi,ji = ⊕hi ∈{1,kji } Ti,ji ,hi with T¯i,ji ,hi = be 5 T1 , T2 if θi,ji ,hi = ∗ and T¯i,ji ,hi = Ti,ji ,hi otherwise; – Θj×1 ,j2 = (c1,j1 ∧ c2,j2 , ⊕hi ∈{1,kji } θ1,j1 ,h1 × θ2,j2 ,h2 ) and Ωj×1 ,j2 = ⊕hi ∈{1,kji } 0 0 T1,j × T2,j where 1 ,h1 2 ,h2

A Symbolic Semantics for Abstract Model Checking

151

1. if θ1,j1 ,h1 ∈ {τ, c!e} and θ2,j2 ,h2 = ∗, then θ1,j1 ,h1 × θ2,j2 ,h2 = θ1,j1 ,h1 , 0 0 and T2,j = T2 (interleaving); T1,j1 ,h1 = T1,j 1 ,h1 2 ,h2 2. if θ1,j1 ,h1 = c?x and θ2,j2 ,h2 = ∗ then θ1,j1 ,h1 × θ2,j2 ,h2 = c?z, such 0 = that T1,j1 ,h1 is free for [z/x] and T2 is free for z. Moreover, T1,j 1 ,h1 0 T1,j1 ,h1 [z/x] and T2,j2 ,h2 = T2 (interleaving); 0 = T1 and 3. if θ1,j1 ,h1 = θ2,j2 ,j2 = ∗, then θ1,h1 ,h1 × θ2,j2 ,h2 = ∗ and T1,j 1 ,h1 0 T2,j2 ,h2 = T2 (idle action); 0 = 4. θ1,j1 ,h1 = c!e and θ2,j2 ,h2 = c?x, then θ1,j1 ,h1 × θ2,j2 ,h2 = τ and T1,j 1 ,h1 0 T1,j1 ,h1 and T2,j2 ,h2 = T2,j2 ,h2 [e/x] (synchronization); 5. symmetric rules; \ – for Θi = (ci , θi,1 ⊕. . .⊕θi,ni ) and Ωi = Ti,1 ⊕. . .⊕Ti,ni , Θi = (ci , ⊕ji ∈Ki θi,ji ) \ and Ki = {ji ∈ {1, ni } | chan(θi,ji ) ∩ L = ∅ and Ωi = ⊕ji ∈Ki Ti,ji \ L.

Table 2. The symbolic semantics (true,∗)

nil −→ nil {Ti T1 + T2

a.T

⊗j ∈{1,n } Θi,ji i i

−→

{Ti

P (¯ e)

T ⊕ a.T a ∈ {τ, c!e, c?x}

⊗ji ∈{1,ni } Ωi,ji }i∈{1,2}

−→

be 5 T1 , T2

T [¯ e/¯ x]

−→

⊗i∈{1,2},j ∈{1,n } Θj+ ,j i i 1 2

{Ti

T1 × T2

(true,a⊕∗)

⊗i∈{1,2},j∈{1,ni } Ωj+1 ,j2

⊗j ∈{1,n } Θi,ji i i

−→

⊗ji ∈{1,ni } Ωi,ji }i∈{1,2}

¯ 1,j ⊗ ¯ ⊗j ∈{1,n } Θ Θ 1 j2 ∈{1,n2 } 2,j2 1 1

−→

⊗j ∈{1,n } Θi,ji i i

−→

−→

−→

⊗i∈{1,n} Θi

−→

⊗i∈{1,n} Ωi

⊗i∈{1,n} Ωi

5

¯1,j1 ⊗j ∈{1,n } Ω ¯2,j2 ⊗j1 ∈{1,n1 } Ω 2 2

⊗ji ∈{1,ni } Ωi,ji }i∈{1,2}

⊗i∈{1,2},j ∈{1,n } Θj× ,j i i 1 2

⊗i∈{1,n} Θi

+

×

⊗i∈{1,2},ji ∈{1,ni } Ωj×1 ,j2

x) ≡ T rec P (¯

T

⊗i∈{1,n} Θi

−→

⊗i∈{1,n} Ωi \

T \L

⊗i∈{1,n} Θi

−→

\ ⊗i∈{1,n} Ωi

\

Automatic Determination of Communication Topologies in Mobile Systems Arnaud Venet ´ LIX, Ecole Polytechnique, 91128 Palaiseau, France. [email protected] http://lix.polytechnique.fr/˜venet

Abstract. The interconnection structure of mobile systems is very difficult to predict, since communication between component agents may carry information which dynamically changes that structure. In this paper we design an automatic analysis for statically determining all potential links between the agents of a mobile system specified in the πcalculus. For this purpose, we use a nonstandard semantics of the πcalculus which allows us to describe precisely the linkage of agents. The analysis algorithm is then derived by abstract interpretation of this semantics. Key words: π-calculus, nonstandard semantics, abstract interpretation.

1

Introduction

We are interested in analyzing the evolution of the interconnection structure, or communication topology, in a mobile system of processes, abstracting away all computational aspects but communication. Therefore, we can restrict our study to the π-calculus [18,19], which is a widely accepted formalism for describing communication in mobile systems. Whereas the communication topology of systems written in csp [13] or ccs [17] can be directly extracted from the text of the specification, a semantic analysis is required in the π-calculus, because communication links may be dynamically created between agents. In the absence of automatic analysis tools this makes the design and debugging of mobile systems very difficult tasks (see [11] for a detailed case study). In this paper we propose a semantic analysis of the π-calculus based on Abstract Interpretation [3,5] for automatically inferring approximate but sound descriptions of communication topologies in mobile systems. In a previous work [23] we have presented an analysis of the π-calculus which relies on a nonstandard concrete semantics. In that model recursively defined agents are identified by the sequence of replication unfoldings from which they stem, whereas the interconnection structure is given by an equivalence relation on the agent communication ports. That semantics is inspired of a representation of sharing in recursive data structures [14] which has been applied to alias analysis [10]. However, the equivalence relation does not capture an important piece of information for debugging and verification purposes: the instance of the G. Levi (Ed.): SAS’98, LNCS 1503, pp. 152–167, 1998. c Springer-Verlag Berlin Heidelberg 1998

Automatic Determination of Communication Topologies

153

channel that establishes a link between two agents. In this paper we redesign our previous analysis in order to take this information into account, while still preserving a comparable level of accuracy. Surprisingly enough, whereas our original analysis was rather complicated, involving heavy operations like transitive closure of binary relations, the refined one is tremendously simpler and only requires very basic primitives. The paper is organized as follows. In Sect. 2 we introduce our representation of mobile systems in the π-calculus. Section 3 describes the nonstandard semantics of mobile systems, which makes instances of recursively defined agents and channels explicit. The abstract interpretation gathering information on communication topologies is constructed in Sect. 4. In Sect. 5 we design a computable analysis which is able to infer accurate descriptions of unbounded and nonuniform communication topologies. Related work is discussed in Sect. 6.

2

Mobile Systems in the π-Calculus

We consider the asynchronous version of the polyadic π-calculus which was introduced by Turner [21] as a semantic basis of the pict programming language. This restricted version has simpler communication primitives and a more operational flavour than the full π-calculus, while still ensuring a high expressive power1 . Let N be a countable set of channel names. The syntax of processes is given by the following grammar: P ::= c![x1 , . . . , xn ] | c?[x1 , . . . , xn ].P | ∗c?[x1 , . . . , xn ].P | (P | P ) | (νx)P

Message Input guard Guarded replication Parallel composition Channel creation

where c, x, x1 , . . . , xn are channel names. Input guard and channel creation act as name binders, i.e. in the process c?[x1 , . . . , xn ].P (resp. (νx)P ) the occurrences of x1 , . . . , xn (resp. x) in P are considered bound. Usual rules about scoping, α-conversion and substitution apply. We denote by fn(P ) the set of free names of P , i.e. those names which are not in the scope of a binder. Following the cham style [1], the standard semantics of the π-calculus is given by a structural congruence and a reduction relation on processes. The congruence relation “≡” satisfies the following rules: (i) (ii) (iii) (iv) (v)

P ≡ Q whenever P and Q are α-equivalent. P | Q ≡ Q | P. P | (Q | R) ≡ (P | Q) | R. (νx)(νy)P ≡ (νy)(νx)P . (νx)P | Q ≡ (νx)(P | Q) if x 6∈fn(Q).

The reduction relation is defined in Fig. 1, where P {x1 /y1 , . . . , xn /yn } denotes the result of substituting every name xi for the name yi in P . This may involve α-conversion to avoid capturing one of the xi ’s. 1

We can encode the lazy λ-calculus for example [2].

154

A. Venet

c![x1 , . . . , xn ] | c?[y1 , . . . , yn ].P → P {x1 /y1 , . . . , xn /yn } c![x1 , . . . , xn ] | ∗c?[y1 , . . . , yn ].P → P {x1 /y1 , . . . , xn /yn } | ∗c?[y1 , . . . , yn ].P P → P0 (νx)P → P 0

P0 ≡ P

P → Q Q ≡ Q0 P 0 → Q0

P → P0 P | Q → P0 | Q

Fig. 1. Reduction relation in the standard semantics.

We now have to define what we mean by a “mobile system in the π-calculus”. We cannot simply allow a mobile system to be described by any process S. In fact, we are unable to design the nonstandard semantics, and hence the analysis, if we do not require S to be closed, i.e. fn(S) = ∅. In other words, we must consider the system in a whole. In order to make the semantic constructions of the following sections simpler, we add further constraints to the structure of mobile systems, which are inspired of Milner’s definition of friendly systems [18]. We denote by x a tuple (x1 , . . . , xn ) of channel names. We say that a process is a thread if it is made of a message possibly preceded by some input guards: c1 ?[x1 ] . . . cn ?[xn ].c![x]. We call resource a replicated process of the following form: ∗c?[x].(νy)(T1 | · · · | Tn ) where all the Ti ’s are threads. A mobile system S is then defined as: S ≡ (νc)(R1 | · · · | Rn | T0 ) where the Ri ’s are resources and T0 is a message, the initial thread, which originates all communications in the system. The names in c are called the initial channels. Therefore, all threads and channels created by a mobile system are fetched from its resources. Example 1. We model a system S which sets up a ring of communicating processes, where each component agent may only communicate with its left and right neighbours. The system generating a ring of arbitrary size is defined as follows: S ≡ (νmake)(νmon)(νleft 0 )(R1 | R2 | make![left 0 ]) where

R1 ≡ ∗make?[left].(νright)(mon![left, right] | make![right])

is the resource that adds a new component to the chain of processes and R2 ≡ ∗make?[left].mon![left, left 0 ] is the resource that closes the ring. The name “mon” should be seen as a reference to a hidden resource (for example a C program) which monitors the behaviour

Automatic Determination of Communication Topologies

155

of a ring component2 . For the sake of clarity, we denote by c the free names of all agents in the system at every stage of its evolution. Then, a ring with four components can be generated by S in four steps as follows: S → (νc)(R1 | R2 | mon![left 0 , right 1 ] | make![right 1 ]) → (νc)(R1 | R2 | mon![left 0 , right 1 ] | mon![right 1 , right 2 ] | make![right 2 ]) → (νc)(R1 | R2 | mon![left 0 , right 1 ] | mon![right 1 , right 2 ] | mon![right 2 , right 3 ] | make![right 3 ]) → (νc)(R1 | R2 | mon![left 0 , right 1 ] | mon![right 1 , right 2 ] | mon![right 2 , right 3 ] | mon![right 3 , left 0 ]) The right i ’s represent the successive instances of the channel right created at each request to R1 . As illustrated by the above example, the configuration of a mobile system at any stage of its evolution has the following particular form: (νc)(R1 | · · · | Rm | T1 | · · · | Tn ) | {z } | {z } Resources Threads where R1 , . . . , Rm are the resources originally present in the system. Intuitively, every thread Ti or channel ci present in the configuration could be unambiguously identified with the instant of its creation in the history of computations. Unfortunately, this precious information is not captured by the standard semantics. The process of α-conversion in particular, which is inherent to the definition of the semantics, destroys the identity of channels. The purpose of the next section is to introduce a refined semantics of the π-calculus which restores this information.

3

Nonstandard Semantics of Mobile Systems

Let S be a mobile system described in the π-calculus. In order to identify the threads and channels created by the system, we must be able to locate the syntactic components of S from which they stem. This is the role of the following notations. We denote by R(S) the number of resources in S. We assume that every resource is assigned a unique number r in {1, . . . , R(S)}. For any such r, we denote by Sr the corresponding resource in S and by T(r) the number of threads spawned by the resource. We similarly assign a unique number t in {1, . . . , T(r)} to every thread in Sr and we denote by Sr,t this thread, which has the following form: Sr,t = c1 ?[x1 ] . . . cA(r,t)−1 ?[xA(r,t)−1 ].cA(r,t) ![xA(r,t) ] 2

Recall that we only take communications into account, abstracting away all other computational aspects like the details of the monitoring procedure here.

156

A. Venet

where A(r, t) is the number of input/output actions in Sr,t . Note that A(r, t) is always nonzero because a thread contains at least one message. For 1 6 n 6 A(r, t), we denote by act(r, t, n) the n-th input/output action of Sr,t , and by Sr,t @ n the subterm cn ?[xn ] . . . cA(r,t) ![xA(r,t) ] of Sr,t starting at the n-th input/output action act(r, t, n). By convention, the initial thread is assigned the resource number 0. Finally, given a resource number r such that Sr ≡ ∗c?[x].(νy1 ) . . . (νyn )(T1 | · · · | TT(r) ) we put guard(r) = c?[x] and C(r) = {y1 , . . . , yn }. Example 2. We consider the system of Example 1 which contains two resources R1 and R2 . We put S1 = R1 and S2 = R2 . Thus we have: – guard(1) = make?[left], C(1) = {right}, T(1) = 2, S1,1 = mon![left, right], S1,2 = make![right], A(1, 1) = A(1, 2) = 1. – guard(2) = make?[left], C(2) = ∅, T(2) = 1, S2,1 = mon![left, left 0 ], A(2, 1) = 1. The initial thread is S0,1 = make![left 0 ].

A configuration of S in the nonstandard semantics is a finite set of thread instances. We do not need to represent resources since they are statically given and accessible by all the threads in any state of the system. A thread instance is a tuple hr, t, n, id , Ei where 1 6 r 6 R(S), 1 6 t 6 T(r), 1 6 n 6 A(r, t), id is a thread identifier and E is an environment. The thread identifier is the history of resource requests which led to the creation of the thread, starting from the initial one. A resource request being nothing more than a message to a replicated process, it can be identified with the thread that released this message. If we denote by Thr(S) = {(r, t) | 1 6 r 6 R(S), 1 6 t 6 T(r)} the set of all threads originally present in the system, then id ∈ Thr(S)∗ . The empty sequence ε is the identifier of the initial thread. The environment E maps every free name x ∈ fn(Sr,t @ n) of the thread instance to a channel instance. A channel instance is a tuple (r0 , y, id 0 ) where 1 6 r0 6 R(S), y ∈ C(r0 ) and id 0 is a channel identifier. Instances of initial channels are represented similarly except that they are assigned the resource number 0. The channel identifier is the history of resource requests that led to the creation of the instance of channel y by resource r0 . Therefore, channel identifiers and thread identifiers are represented identically, i.e. id 0 ∈ Thr(S)∗ . Similarly, the identifier of an initial channel is the empty sequence ε. We assume that there is no overlapping of scopes in the mobile system, i.e. we forbid terms like x?[y].y?[y].y![z] or ∗c?[y, z].(νy)z![y]. This assumption can always be satisfied by appropriate α-conversion. The transition relation “.” of the nonstandard operational semantics is defined in Fig. 2 and Fig. 3. It should be clear that without the hypothesis on name scoping, the definition of the resulting environments is ambiguous. The two transition rules correspond to the

Automatic Determination of Communication Topologies

157

two kinds of operations that may arise in a mobile system: resource fetching and communication between threads. The initial configuration C0 of the nonstandard semantics is given by C0 = {h0, 1, 1, ε, E0 i} where the environment E0 maps any z ∈ fn(S0,1 @ 1) to the instance (0, z, ε) of the corresponding initial channel. Example 3. Using the notations of Example 2, a ring with four components is described in the nonstandard semantics as follows: + * mon 7→(0, mon, ε), , ε), left 7→(0, left , 1, 1, 1, (0, 1), 0 right 7→(1, right, (0, 1)) * + mon 7→(0, mon, ε), left 7→(1, right, (0, 1)) 1, 1, 1, (0, 1).(1, 2), , right 7→(1, right, (0, 1).(1, 2)) + * mon 7→(0, mon, ε), 2 left 7→(1, right, (0, 1).(1, 2)) , , 1, 1, 1, (0, 1).(1, 2) 2 right 7→(1, right, (0, 1).(1, 2) ) + * mon 7→(0, mon, ε), 2 3 ) left 7→(1, right, (0, 1).(1, 2) , 2, 1, 1, (0, 1).(1, 2) left 0 7→(0, left 0 , ε) The generative process of the ring is now made entirely explicit thanks to the information carried by thread and channel identifiers (compare with the corresponding computation in the standard semantics given in example 1). Both semantics can be shown to be equivalent by defining the translation of a nonstandard configuration C of S to a term π(C) of the π-calculus and by applying a bisimulation argument. We may assume without loss of generality that the π-terms we will generate are built upon the set of names N 0 defined as the disjoint union of N with the set of all channel instances (r, y, id ) that could be created by S. We identify an environment with a substitution over N 0 . We denote by P σ the result of applying a substitution σ to a π-term P . Then, for any nonstandard configuration C = {hr1 , t1 , n1 , id 1 , E1 i, . . . , hrk , tk , nk , id k , Ek i} we define the translation π(C) as follows: 0 π(C) = (νc)(S1 E10 | · · · | SR(S) ER(S) | (Sr1 ,t1 @ n1 )E1 | · · · | (Srk ,tk @ nk )Ek )

where, for 1 6 r 6 R(S), Er0 is the environment which maps any z ∈ fn(Sr ) to (0, z, ε). The channels in c are all those which have a free occurrence in some agent of the top-level parallel composition of π(C). ∗

∗

Theorem 4. If C0 . C and C . C 0 , then π(C) → π(C 0 ). If C0 . C and π(C) → P , then there exists C 0 such that C . C 0 and P ≡ π(C 0 ).

158

A. Venet

If there are µ ∈ C and 1 6 r0 6 R(S) such that: – – – –

µ = hr, t, n, id , Ei act(r, t, n) = x![x1 , . . . , xk ] guard(r0 ) = c?[y1 , . . . , yk ] E(x) = (0, c, ε)

then

C . (C − {µ}) ∪ {hr0 , t0 , 1, id .(r, t), Et0 i | 1 6 t0 6 T(r0 )}

where, for all 1 6 t0 6 T(r0 ) and z ∈ fn(Sr0 ,t0 @ 1) if z = yi , 1 6 i 6 k E(xi ) if z ∈ C(r0 ) Et0 (z) = (r0 , z, id .(r, t)) (0, z, ε) otherwise, i.e. if z is an initial channel Fig. 2. Resource fetching.

A nonstandard configuration describes the communication topology of a mobile system at a particular moment of its evolution in terms of the resources initially present in the system. Therefore, the nonstandard semantics is a good basis for deriving an analysis, since the resulting information can be used to determine the role of each syntactic component of the system in the evolution of its interconnection structure. Constructing a computable abstraction of this semantics is the purpose of the next section.

4

Abstract Interpretation of Mobile Systems

We denote by C the set of all possible nonstandard configurations for a system ∗ S. We are actually interested in the set S = {C ∈ C | C0 . C} of configurations of the system which are accessible from the initial one by a finite sequence of computations. This is the collecting semantics of S [9], which can be expressed as the least fixpoint of the ∪-complete endomorphism IF on the complete lattice (℘(C), ⊆, ∪, ∅, ∩, C) defined as follows: IF(X) = {C0 } ∪ {C | ∃C 0 ∈ X : C 0 . C} Following the methodology of Abstract Interpretation [3,5], we construct a lattice (C ] , v, t, ⊥, u, >), the abstract domain, allowing us to give finite descriptions of infinite sets of configurations. This domain is related to ℘(C) via a monotone map γ : (C ] , v) → (℘(C), ⊆), the concretization function. Then we derive an abstract counterpart IF] : C ] → C ] of IF which must be sound with respect to γ, that is: IF ◦ γ ⊆ γ ◦ IF] . The abstract domain C ] is based upon a global abstraction of all environments in a nonstandard configuration. We assume that we are provided with a lattice (Id]2 , v2 , t2 , ⊥2 , u2 , >2 ) and a monotone map γ2 : (Id]2 , v2 ) →

Automatic Determination of Communication Topologies

159

If there are µ, ρ ∈ C such that: – – – – –

µ = hr, t, n, id , Ei ρ = hr0 , t0 , n0 , id 0 , E 0 i act(r, t, n) = x![x1 , . . . , xk ] act(r0 , t0 , n0 ) = y?[y1 , . . . , yk ] E(x) = E 0 (y)

then

C . (C − {µ, ρ}) ∪ {hr0 , t0 , n0 + 1, id 0 , E 00 i}

where, for all z ∈ fn(Sr0 ,t0 @ n0 + 1) E(xi ) E 00 (z) = E 0 (z)

if z = yi , 1 6 i 6 k otherwise

Fig. 3. Communication between threads.

(℘(Thr(S)∗ × Thr(S)∗ ), ⊆). This lattice is left as a parameter of our abstraction. It will be instantiated in the next section when we will set up an effective analysis. Let Chan(S) be the set {(r, t, n, x, r0 , y) | (r, t) ∈ Thr(S), n ∈ A(r, t), x ∈ fn(Sr,t @ n), 1 6 r0 6 R(S), y ∈ C(r0 )} of all possible syntactic relations between a free name in a thread and a channel created by a resource of the system. The abstract domain C ] is then defined as C ] = Chan(S) → Id]2 , the lattice operations being the pointwise extensions of those in Id]2 . Given an abstract configuration C ] , γ(C ] ) is the set of nonstandard configurations C such that, for any hr, t, n, id , Ei ∈ C and any x ∈ fn(Sr,t @ n), the following condition is satisfied: E(x) = (r0 , y, id 0 ) ⇒ (id , id 0 ) ∈ γ2 (C ] (r, t, n, x, r0 , y)) Monotonicity of γ is readily checked. For the sake of readability, we will denote an abstract configuration C ] by its graph {hκ1 7→id ]1 i, . . . , hκn 7→id ]n i}, where the κi ’s are in Chan(S) and each id ]j is in Id]2 . We may safely omit to write pairs of the form hκ 7→ ⊥2 i. The abstract semantics is given by a transition relation on abstract configurations. In the relation C1] C2] , the configuration C2] represents the modification to the communication topology of C1] induced by an abstract computation. Therefore, the function IF] is given by: IF] (C ] ) = C0] t C ] t

G

]

{C | C ]

]

C }

where C0] is the initial abstract configuration defined as {h0, 1, 1, x, 0, xi 7→ε2 | x ∈ fn(S0,1 @ 1)}, ε2 being a distinguished element of Id]2 such that (ε, ε) ∈ is defined in Fig. 4 and Fig. 5 by using the γ2 (ε2 ). The transition relation

160

A. Venet

If there are h(r, t, n, x, 0, c) 7→id ] i ∈ C ] and 1 6 r0 6 R(S) such that: – act(r, t, n) = x![x1 , . . . , xk ] – guard(r0 ) = c?[y1 , . . . , yk ] – id ] 6= ⊥2 then

C]

{hκ 7→id ]t0 ,z,r00 ,w i | κ = (r0 , t0 , 1, z, r00 , w), 1 6 t0 6 T(r0 )}

where, for all 1 6 t0 6 T(r0 ) push(r,t) (C ] (r, t, n, xi , r00 , w)) ⊥2 id ]t0 ,z,r00 ,w = dpush(r,t) (C ] (r, t, n, x, 0, c)) ⊥ 2 spush(C ] (r, t, n, x, 0, c))

if if if if if

z z z z z

= yi , 1 6 i 6 k ∈ C(r0 ) and (r00 , w) 6= (r0 , z) ∈ C(r0 ) and (r00 , w) = (r0 , z) is initial and (r00 , w) 6= (0, z) is initial and (r00 , w) = (0, z)

Fig. 4. Abstract resource fetching.

following abstract primitives: a “push” operation pushτ : Id]2 → Id]2 , a “double push” dpushτ : Id]2 → Id]2 and a “single push” spushτ : Id]2 → Id]2 defined for any τ ∈ Thr(S), a “synchronization” operator sync : Id]2 × Id]2 → Id]2 and a “swapping” operation swap : Id]2 → Id]2 . These operations depend on the choice of Id]2 , however they must satisfy some soundness conditions: – For any τ ∈ Thr(S) and id ] ∈ Id]2 , {(id .τ, id 0 ) | (id , id 0 ) ∈ γ2 (id ] )} ⊆ γ2 (pushτ (id ] )). – For any τ ∈ Thr(S) and id ] ∈ Id]2 , {(id .τ, id .τ ) | (id , id 0 ) ∈ γ2 (id ] )} ⊆ γ2 (dpushτ (id ] )). – For any τ ∈ Thr(S) and id ] ∈ Id]2 , {(id .τ, ε) | (id , id 0 ) ∈ γ2 (id ] )} ⊆ γ2 (spushτ (id ] )). – For any id ] ∈ Id]2 , {(id 0 , id ) | (id , id 0 ) ∈ γ2 (id ] )} ⊆ γ2 (swap(id ] )). – For any id ]1 , id ]2 ∈ Id]2 , {(id 1 , id 2 ) | ∃id 0 ∈ Thr(S)∗ : (id 1 , id 0 ) ∈ γ2 (id ]1 ) ∧ (id 2 , id 0 ) ∈ γ2 (id ]2 )} ⊆ γ2 (sync(id ]1 , id ]2 )). Intuitively, the pushτ operation concatenates τ to the first component of every pair of identifiers. The dpushτ and spushτ operations act similarly, except that dpushτ duplicates the first component into the second one and spushτ sets the second component to ε. The swap operation permutes the components of every pair of identifiers. The sync operation extracts pairs of identifiers corresponding to redexes, i.e. pairs of agents linked to the same instance of a channel. ]

Proposition 5. If C ∈ γ(C ] ) and C . C 0 , then there exists C in C ] , such that ] ] C and C 0 ∈ γ(C ] t C ). C] The soundness of IF] is then a simple consequence of the previous result. It is remarkable that the soundness of the whole semantics depends only on very

Automatic Determination of Communication Topologies

161

If there are hκ1 7→id ]1 i, hκ2 7→id ]2 i ∈ C ] such that: – – – – –

κ1 = (r, t, n, x, r 00 , u) κ2 = (r0 , t0 , n0 , y, r 00 , u) act(r, t, n) = x![x1 , . . . , xk ] act(r0 , t0 , n0 ) = y?[y1 , . . . , yk ] id ]s = sync(id ]1 , id ]2 ) 6= ⊥2

then

C]

{hκ 7→id ]z,r000 ,w i | κ = (r0 , t0 , n0 + 1, z, r000 , w)}

]

where, if we put id i,r000 ,w = C ] (r, n, t, xi , r000 , w) for 1 6 i 6 k ( ] swap(sync(swap(id i,r000 ,w ), swap(id ]s ))) if z = yi , 1 6 i 6 k ] id z,r000 ,w = C ] (r0 , t0 , n0 , z, r 000 , w) otherwise Fig. 5. Abstract communication between threads.

simple conditions over some primitive operations. This means that we only need to construct the domain Id]2 and instantiate those operations to obtain a computable and sound abstract semantics. This is the goal that we will achieve in the next section.

5

Design of a Computable Analysis

The collecting semantics S is the least fixpoint of IF. Therefore, by Kleene’s theorem, it is the limit of the following increasing iteration sequence:

S0 = ∅ Sn+1 = IF(Sn )

Following [3,9] we will compute a sound approximation S ] of S by mimicking this iteration, using IF] instead of IF. Since the resulting computation may not terminate, we use a widening operator to enforce convergence in finitely many steps. A widening operator ∇ : C ] × C ] → C ] must satisfy the following conditions: – For any C1] , C2] ∈ C ] , C1] t C2] v C1] ∇ C2] .

]

– For any sequence (Cn] )n>0 , the sequence (C n )n>0 defined as: (

is ultimately stationary.

]

C 0 = C0] ] ] ] C n+1 = C n ∇ Cn+1

162

A. Venet

Note that we can construct a widening on C ] from an existing widening ∇2 on Id]2 by pointwise application of ∇2 . We define the approximate iteration sequence (S ] n )n>0 as follows: ] S 0 = ⊥ S ] n+1 = S ] n ∇ IF] (S ] n ) ] S n+1 = S ] n

¬(IF] (S ] n ) v S ] n ) IF] (S ] n ) v S ] n

if if

Convergence is ensured by the following result: Theorem 6 [9]. The sequence (S ] n )n>0 is ultimately stationary and its limit S ] satisfies S ⊆ γ(S ] ). Moreover, if N > 0 is such that S ] N +1 = S ] N , then for all n > N , S ] n = S ] N . This provides us with an algorithm for automatically computing a sound approximation of S. It now remains to instantiate the domain Id]2 and the associated abstract primitives. We will design two abstractions of ℘(Thr(S)∗ × Thr(S)∗ ), each one capturing a particular kind of information. Our first abstraction captures sequencing information and is based upon an approximation of thread and channel identifiers by regular languages. Let (Reg, ⊆, ∪, ∅, ∩, Thr(S)∗ ) be the lattice of regular languages over the alphabet Thr(S) ordered by set inclusion. We define Id]reg as the product lattice Reg × Reg. The concretization γreg is given by γreg (L1 , L2 ) = L1 × L2 . The associated abstract primitives are defined as follows: pushreg τ (L1 , L2 ) = (L1 .τ, L2 ) dpushreg τ (L1 , L2 ) = (L1 .τ, L1 .τ ) spushreg τ (L1 , L2 ) = (L1 .τ, ε) swapreg (L1 , L2 ) = (L2 , L1 ) (L1 , L01 ) – syncreg ((L1 , L2 ), (L01 , L02 )) = (∅, ∅)

– – – –

if L2 ∩ L02 6= ∅ otherwise

The soundness conditions are easily checked. The element εreg 2 is given by (ε, ε). Since Id]reg may have infinite strictly increasing chains, we must define a widening operator ∇reg 2 . It is sufficient to construct a widening ∇reg on Reg and to apply it componentwise to elements of Id]reg . A simple choice for L1 ∇reg L2 consists of quotienting the minimal automaton of L1 ∪ L2 such that any letter of the alphabet may occur at most once in the automaton. The resulting automaton is minimal, and there are finitely many such automata, which ensures the stabilization property. The second approximation captures counting relations between the components of a tuple of thread or channel identifiers. This will allow us to give nonuniform descriptions of recursively defined communication topologies. Suppose that we are given an infinite set of variables V. We assign two distinct variables xτ and yτ to each element τ of Thr(S). Now we consider a finite system K of linear equality constraints over the variables V with coefficients in Q. If we denote

Automatic Determination of Communication Topologies

163

by |id |τ the number of occurrences of τ in the sequence id , the concretization γnum (K) of K is the set of all pairs (id , id 0 ) such that the following variable assignment: {xτ 7→ i|d |τ , yτ 7→ i|d 0 |τ | τ ∈ Thr(S)} is a solution of K. The domain of finite systems of linear equality constraints over V ordered by inclusion of solution sets can be turned into a lattice Id]num . This domain has been originally introduced by Karr [15]. We refer the reader to the original paper for a detailed algorithmic description of lattice operations. We could use more sophisticated domains of computable numerical constraints such as linear inequalities [6] or linear congruences [12], but the underlying algorithmics is much more involved. Nevertheless, giving a rigorous construction of the abstract primitives on Id]num would be still very technical. Therefore, for the sake of readability, we only outline the definition of these primitives: (K) is the system of constraints K in which we have replaced every – pushnum τ occurrence of the variable xτ by the expression xτ − 1. – If K is a system of linear equality constraints, we denote by Kx the system in which we have removed all constraints involving a variable yτ . Then (K) is the system pushnum (Kx ) with the additional constraints dpushnum τ τ xτ 0 = yτ 0 , for any τ 0 ∈ Thr(S). – Similarly, spushnum (K) is the system pushnum (Kx ) with the additional τ τ constraints yτ 0 = 0, for any τ 0 ∈ Thr(S). – swapnum (K) is the system in which we have replaced each occurrence of xτ by yτ and vice-versa. – Let K1 and K2 be two systems of linear equality constraints. For any τ ∈ Thr(S), let x0τ and yτ0 be fresh variables of V. Let K20 be the system K2 in which we have substituted each occurrence of xτ (resp. yτ ) by x0τ (resp. yτ0 ). We construct the system K1,2 as the union of K1 and K20 together with the additional constraints yτ = yτ0 , for any τ ∈ Thr(S). Then, we define syncnum (K1 , K2 ) as the system K1,2 in which we have removed all constraints involving a variable yτ or yτ0 , each remaining variable x0τ being renamed in yτ . Note that a normalization pass (namely a Gauss reduction) has to be performed is on the system after or during each of these operations. The element εnum 2 given by the system of constraints {xτ = 0, yτ = 0 | τ ∈ Thr(S)}. Since we only consider systems defined over the finite set of variables {xτ , yτ | τ ∈ Thr(S)}, we cannot have infinite strictly increasing chains [15]. Therefore, we can use the join operation t num as a widening. Example 7. We consider the product of domains Id]reg and Id]num and we run the analysis on the system of Example 1 with the notations of Example 2. For the sake of readability, at each step we only write the elements of the abstract configuration that differ from the previous iteration. Moreover, we do not figure trivial constraints of the form xτ = 0 whenever they can be deduced from the Id]reg component.

164

A. Venet

First iteration.

h0, 1, 1, make, 0, makei 7→ h(ε, ε), >num i, h0, 1, 1, left 0 , 0, left 0 i 7→ h(ε, ε), >num i

Second iteration. h1, 1, 1, mon, 0, moni 7→ h((0, 1), ε), x(0,1) = 1i, h1, 1, 1, left, 0, left i 7→ h ( (0, 1), ε), x = 1i, (0,1) 0 = y = 1i, h1, 1, 1, right, 1, righti 7→ h ( (0, 1), (0, 1)), x (0,1) (0,1) h1, 2, 1, make, 0, makei 7→ h((0, 1), ε), x(0,1) = 1i, h1, 2, 1, right, 1, righti 7→ h((0, 1), (0, 1)), x(0,1) = y(0,1) = 1i, h2, 1, 1, mon, 0, moni 7→ h((0, 1), ε), x(0,1) = 1i, h2, 1, 1, left , 0, left i 7→ h ( (0, 1), ε), x = 1i, (0,1) 0 0 h2, 1, 1, left, 0, left 0 i 7→ h((0, 1), ε), x(0,1) = 1i Third iteration. h1, 1, 1, mon, 0, moni 7→h((0, 1).(ε + (1, 2)), ε),x(0,1) = 1i, x(0,1) = y(0,1) = 1 , h1, 1, 1, left, 0, righti 7→ ((0, 1).(1, 2), (0, 1)), = 1 x (1,2) * ((0, 1).(ε + (1, 2)), (0, 1).(ε + (1, 2))), + = y = 1 x , h1, 1, 1, right, 1, righti 7→ (0,1) (0,1) x = y (1,2) (1,2) h1, 2, 1, make, 0, makei 7→ h((0, 1).(ε + (1, 2)), ε), x(0,1) = 1i, * ((0, 1).(ε + (1, 2)), (0, 1).(ε + (1, 2))), + = y = 1 x h1, 2, 1, right, 1, righti 7→ , (0,1) (0,1) x = y (1,2) (1,2) = 1i, h2, 1, 1, mon, 0, moni 7→ h ( (0, 1).(ε + (1, 2)), ε), x (0,1) h2, 1, 1, left , 0, left i 7→ h ( (0, 1).(ε + (1, 2))), ε), x = 1i, (0,1) 0 0 x(0,1) = y(0,1) = 1 h2, 1, 1, left, 0, righti 7→ ((0, 1).(1, 2), (0, 1)), x(1,2) = 1 Fourth iteration. h1, 1, 1, mon, 0, moni 7→ h((0, 1).(1, 2)∗ , ε), x(0,1) = 1i, * ((0, 1).(1, 2)∗ , (0, 1).(ε + (1, 2))), + x = y = 1 , h1, 1, 1, left, 0, righti 7→ (0,1) (0,1) = y + 1 x (1,2) (1,2) * + ∗ ∗ ((0, 1).(1, 2) , (0, 1).(1, 2) ), = y = 1 x h1, 1, 1, right, 1, righti 7→ , (0,1) (0,1) x = y (1,2) (1,2) h1, 2, 1, make, 0, makei 7→ h((0, 1).(1, 2)∗ , ε), x(0,1) = 1i, * ((0, 1).(1, 2)∗ , (0, 1).(1, 2)∗ ), + x(0,1) = y(0,1) = 1 , h1, 2, 1, right, 1, righti 7→ = y x (1,2) (1,2) ∗ h2, 1, 1, mon, 0, moni 7→ h ( (0, 1).(1, 2) , ε), x = 1i, (0,1) ∗ , 0, left i 7→ h ( (0, 1).(1, 2) , ε), x = 1i, h2, 1, 1, left (0,1) 0 0 * + ∗ ((0, 1).(1, 2) , (0, 1).(ε + (1, 2))), h2, 1, 1, left, 0, righti 7→ = y = 1 x (0,1) (0,1) x(1,2) = y(1,2) + 1

Automatic Determination of Communication Topologies

165

Fifth iteration. * ((0, 1).(1, 2)∗ , (0, 1).(1, 2)∗ ), + x = y = 1 , h1, 1, 1, left, 0, righti 7→ (0,1) (0,1) x(1,2) = y(1,2) + 1 * ((0, 1).(1, 2)∗ + , (0, 1).(1, 2)∗ ), = y = 1 x h2, 1, 1, left, 0, righti 7→ (0,1) (0,1) x(1,2) = y(1,2) + 1 At the sixth iteration step we find the same configuration. Therefore, following Theorem 6, we know that the limit has been reached. Putting all previous computations together, we obtain: h0, 1, 1, make, 0, makei 7→ h(ε, ε), >num i, h0, 1, 1, left 0 , 0, left 0 i 7→ h(ε, ε), >num i, ∗ , ε), x = 1i, h1, 1, 1, mon, 0, moni 7→ h ( (0, 1).(1, 2) (0,1) h1, 1, 1, left, 0, left i 7→ h ( (0, 1), ε), x = 1i, (0,1) 0 + * ∗ ∗ , (0, 1).(1, 2) ), ((0, 1).(1, 2) x = y = 1 , h1, 1, 1, left, 0, righti 7→ (0,1) (0,1) = y + 1 x (1,2) (1,2) * + ∗ ∗ ((0, 1).(1, 2) , (0, 1).(1, 2) ), = y = 1 x h1, 1, 1, right, 1, righti 7→ , (0,1) (0,1) x = y (1,2) (1,2) ] S = ∗ h1, 2, 1, make, 0, makei 7→ h((0, 1).(1, 2) , ε), x(0,1) = 1i, * ((0, 1).(1, 2)∗ ∗ + , (0, 1).(1, 2) ), h1, 2, 1, right, 1, righti 7→ = y = 1 x , (0,1) (0,1) x = y (1,2) (1,2) ∗ h2, 1, 1, mon, 0, moni 7→ h((0, 1).(1, 2) , ε), x(0,1) = 1i, ∗ , 0, left i 7→ h ( (0, 1).(1, 2) , ε), x = 1i, h2, 1, 1, left (0,1) 0 0 h2, 1, 1, left, 0, left i 7→ h ( (0, 1), ε), x = 1i, (0,1) 0 + * ∗ ∗ , (0, 1).(1, 2) ), ((0, 1).(1, 2) x = y = 1 h2, 1, 1, left, 0, righti 7→ (0,1) (0,1) x(1,2) = y(1,2) + 1 This is a very accurate description of the communication topology of the ring. In particular, we are able to distinguish between instances of recursively defined agents and channels.

6

Conclusion

We have described a parametric analysis framework for automatically inferring communication topologies of mobile systems specified in the π-calculus. We have instantiated this framework to obtain an effective analysis which is able to give finite descriptions of unbounded communication topologies that distinguish between instances of recursively defined components. To our knowledge this is the only existing analysis of mobile systems (excluding [23]) which can produce

166

A. Venet

results of that level of accuracy without any strong restriction on the base language. Previous works addressed the issue of communication analysis in csp [4, 16] or cml [20,7,8]. In the latter papers, the analysis techniques heavily rely on cml type information and cannot be applied to more general untyped languages like the π-calculus. In order to keep the presentation clear within a limited space, we had to make some simplifying assumptions that can be relaxed in many ways, for example by using more expressive abstract domains to denote relations between thread and channel identifiers, like cofibered domains [22,24], by refining the abstract semantics to take more information into account, like the number of instances of a channel or a thread, or by considering a richer version of the π-calculus with guarded choice, matching and nested replications. Finally, in view of the encodings of classical language constructs (data structures, references, control structures) in the π-calculus, it would be interesting to study the possiblity of using a static analysis of the π-calculus as a universal analysis back-end for high-level languages. Acknowledgements. I wish to thank Radhia Cousot, Patrick Cousot, Ian Mackie and the anonymous referees for useful comments on a first version of this paper.

References 1. G. Berry and G. Boudol. The chemical abstract machine. Theoretical Computer Science, 96:217–248, 1992. 2. G. Boudol. Asynchrony and the π-calculus. Technical Report 1702, INRIA, 1992. 3. P. Cousot and R. Cousot. Abstract interpretation: a unified lattice model for static analysis of programs by construction or approximation of fixpoints. In Conference Record of the Fourth ACM Symposium on Principles of Programming Languages, pages 238–252, 1977. 4. P. Cousot and R. Cousot. Semantic analysis of communicating sequential processes. In Seventh International Colloquium on Automata, Languages and Programming, volume 85 of LNCS, 1980. 5. P. Cousot and R. Cousot. Abstract interpretation frameworks. Journal of logic and computation, 2(4):511–547, August 1992. 6. P. Cousot and N. Halbwachs. Automatic discovery of linear restraints among variables of a program. In Proceedings of the Fifth Conference on Principles of Programming Languages. ACM Press, 1978. 7. C. Colby. Analyzing the communication topology of concurrent programs. In Symposium on Partial Evaluation and Program Manipulation, 1995. 8. C. Colby. Determining storage properties of sequential and concurrent programs with assignment and structured data. In Proceedings of the Second International Static Analysis Symposium, volume 983 of Lecture Notes in Computer Science, pages 64–81. Springer-Verlag, 1995. 9. P. Cousot. Semantic foundations of program analysis. In S.S. Muchnick and N.D. Jones, editors, Program Flow Analysis: Theory and Applications, chapter 10, pages 303–342. Prentice-Hall, Inc., Englewood Cliffs, 1981.

Automatic Determination of Communication Topologies

167

10. A. Deutsch. A storeless model of aliasing and its abstraction using finite representations of right-regular equivalence relations. In Proceedings of the 1992 International Conference on Computer Languages, pages 2–13. IEEE Computer Society Press, 1992. 11. P. Degano, C. Priami, L. Leth, and B. Thomsen. Analysis of facile programs: A case study. In Proceedings of the Fifth LOMAPS Workshop on Analysis and Verification of Multiple-Agent Languages, volume 1192 of Lecture Notes in Computer Science, pages 345–369. Springer-Verlag, 1996. 12. P. Granger. Static analysis of linear congruence equalities among variables of a program. In TAPSOFT’91, volume 493. Lecture Notes in Computer Science, 1991. 13. C.A.R. Hoare. Communicating Sequential Processes. Prentice Hall, 1985. 14. H.B.M Jonkers. Abstract storage structures. In De Bakker and Van Vliet, editors, Algorithmic languages, pages 321–343. IFIP, 1981. 15. M. Karr. Affine relationships among variables of a program. Acta Informatica, pages 133–151, 1976. 16. N. Mercouroff. An algorithm for analysing communicating processes. In Mathematical Foundations of Programming Semantics, volume 598 of LNCS, 1991. 17. R. Milner. Communication and Concurrency. Prentice Hall, 1989. 18. R. Milner. The polyadic π-calculus: a tutorial. In Proceedings of the International Summer School on Logic and Algebra of Specification. Springer-Verlag, 1991. 19. R. Milner, J. Parrow, and D. Walker. A calculus of mobile processes. Information and Computation, 100:1–77, 1992. 20. H. R. Nielson and F. Nielson. Higher-order concurrent programs with finite communication topology. In 21st ACM Symposium on Principles of Programming Languages, 1994. 21. D. N. Turner. The Polymorphic Pi-Calculus: Theory and Implementation. PhD thesis, Edinburgh University, 1995. 22. A. Venet. Abstract cofibered domains: Application to the alias analysis of untyped programs. In Proceedings of the Third International Static Analysis Symposium SAS’96, volume 1145 of Lecture Notes in Computer Science, pages 366–382. Springer-Verlag, 1996. 23. A. Venet. Abstract interpretation of the π-calculus. In Proceedings of the Fifth LOMAPS Workshop on Analysis and Verification of High-Level Concurrent Languages, volume 1192 of Lecture Notes in Computer Science, pages 51–75. SpringerVerlag, 1996. 24. A. Venet. Automatic analysis of pointer aliasing for untyped programs. Science of Computer Programming, 1999. To appear.

Constructing speci c SOS semantics for concurrency via abstract interpretation EXTENDED ABSTRACT

Chiara Bodei1 , Pierpaolo Degano,1 Corrado Priami2 1

Dipartimento di Informatica, Universita di Pisa Corso Italia 40, I-56100 Pisa, Italy fchiara,[email protected]

Istituto Policattedra, Universita di Verona Ca' Vignal 2, Strada Le Grazie 1, I-37134 Verona, Italy 2

[email protected]

Abstract. Most of the SOS semantics for concurrent systems can be

derived by abstracting on the inference rules of a concrete transition system, namely the proved transition system. Besides the standard interleaving semantics we mechanically derive the causal transition system for CCS , whose de nition is particularly dicult and paradigmatic. Its rules are shown to coincide with those presented in the literature. Also, the tree of its computations coincide with that obtained by abstracting the computations of the proved transition system. Keywords. Concurrency, abstract interpretation, SOS semantics, causality, non-interleaving descriptions.

1 Introduction

We apply the theory of abstract interpretation [10] to calculi for concurrency. We carry out our investigation in a pure setting and consider the basic, foundational calculus CCS [26]. However, we are con dent that more powerful calculi, e.g. those for mobile processes, like the -calculus [27], require only some technical adjustments to the machinery proposed here. The operational semantics of CCS is a transition system, de ned following the SOS approach. More in detail, a transition for a process P is deduced applying a set of inference rules, driven by the syntax of P . The original transition system describes the evolution of processes in an interleaving style, and a great deal of work has been devoted to de ne new transition systems expressing also other aspects of process evolutions, both qualitative and quantitative. Among the former, particularly relevant has been the study of transition systems that express the causality between transitions (see [7,14,24,28,6] to mention only a few of a long list of references). Another well-studied non-interleaving qualitative aspect concerns the description of the localities where processes are placed (among the other proposals, see [8,1,2,29]). Quantitative descriptions include transition systems that express temporal aspects [22,30,19,21,16], probabilistic aspects [34,25], and stochastic G. Levi (Ed.): SAS’98, LNCS 1503, pp. 168−183, 1998. Springer-Verlag Berlin Heidelberg 1998

Constructing Specific SOS Semantics for Concurrency via Abstract Interpretation

169

ones [20,23,3,9,31]. Besides its interest in se, an SOS semantics for causality is relevant because it is paradigmatic for others qualitative, non-interleaving description of concurrent systems, as well as for some of the quantitative ones. In particular, many non-interleaving SOS de nitions have rules in a form similar to those of the causal transition system of [14]. Our starting point is the proved transition system of CCS [17], PTS for short. This very concrete representation of systems has been used to describe qualitative and quantitative non-interleaving aspects of concurrent computations [32,33,4]. The transitions are labelled by encodings of their proofs, that encode also most of the aspects brie y mentioned above. The rst and the last author used abstract interpretation in [5] to extract the causal computations from the proved computations of the PTS, and sketched how other kinds of computations could be derived, as well. Our main goal here is to push further the use of abstract interpretation in order to mechanically obtain from the PTS the SOS inference rules for most of the non-interleaving transition systems mentioned above. We argue that small step operational semantics in SOS style for process calculi can be organized in a hierarchy by abstract interpretation, along the lines of abstraction of rule based speci cations presented in [11{13]. We work out in full detail the derivation of the causal transition system for CCS , possibly the most dicult non-interleaving transition system. We de ne a (family of) abstraction functions that yield CTS, a variant of the original causal transition system of [14]. As a consequence, the computations of CTS coincide with the abstraction of the computations of PTS, as de ned in [5]. The task is easier with Milner's original transition system [26], that we also derive through abstractions from both the proved and the causal transition systems. These abstractions constitute a fragment of a hierarchy of transition systems. Our approach is constructive, because the causal SOS rules are formally derived by abstraction of the proved ones and the same happens for the interleaving ones. This is a rst step towards giving evidence of the power of formal methods over empirical ones. In the conclusions, we brie y sketch the simple modi cations needed to cope with the locality transition system, that we omit here for lack of space. We also give hints on the derivation of temporal transition systems.

2 The concrete domain We recall the pure CCS [26] that we use here as a test-bed. As usual, we denote the countable set of atomic actions by , the set of co-actions by and the invisible actions with . Then A = [ [ f g. We also assume a set of agent identi ers with typical element A. Processes (denoted by P; Q; R; : : : 2 P ) are built from actions and agents according to the syntax

P ::= 0 j :P j P + P j P jP j (a)P j A where 2 A. Hereafter, we omit the trailing 0.

170

C. Bodei, P. Degano, and C. Priami

The pre x is the rst atomic action that the process :P can perform. Summation denotes nondeterministic choice. The operator j describes parallel composition of processes. The restriction operator (a) prevents P from performing a. A is the invocation of the agent identi er A, that has a unique de ning equation of the form A = P. Hereafter, we assume that (P = P ; +; 0) is a commutative monoid. As in [17], the parallel operator j is neither associative nor commutative, and this is the only dierence with the original CCS de nition, apart from the omission of the relabelling operator, which is irrelevant to our present study. The transition system of CCS is in Tab. 1. P ,! P ide : ; A = P A ,! P

act : :P ,! P

0

par0 : par1 :

0

P ,! P P jQ ,! P jQ 0

sum :

0

P ,! P QjP ,! QjP 0

res :

0

P ,! P P + Q ,! P 0

P ,! P ; 62 fa; ag (a)P ,! (a)P 0

0

a a P ,! P ; Q ,! Q com : P jQ ,! P jQ 0

0

Table 1.

0

0

0

Transition system for CCS .

As anticipated in the introduction, we enrich the labels of the standard interleaving transition system of CCS in the style of [7,15]. It is thus possible to derive dierent semantic models for CCS by extracting new kinds of labels from the enriched ones (see [17,18]). We start with the de nition of the enriched labels and of a function (`) that takes them to the corresponding action labels. De nition 1. Let # 2 fjj0; jj1g. Then proof terms (with metavariable 2 ) are de ned by the following syntax ::= # j #hjj0#00 ; jj1#11 i with 0 = a if and only if 1 = a, or vice versa. Function ` is de ned as `(#a) = a and `(#hjj0#0a; jj1#1a0 i) = . The proved transition system PTS for CCS is in Tab. 2, where the symmetric rule for communication (Com1 ) is omitted. These rules are our concrete domain.

Constructing Specific SOS Semantics for Concurrency via Abstract Interpretation

Act : :P ,! P

Ide :

P ,! P ;A = P A ,! P 0

0

Par0 :

P ,! P 0 P jQ ,! P jQ

Sum :

Par1 :

P ,! P 1 QjP ,! QjP

Res :

0

jj

0

0

jj

0

171

P ,! P P + Q ,! P 0

0

P ,! P ; `() 62 fa; ag (a)P ,! (a)P 0

0

#a #a P ,! P ; Q ,! Q Com0 : 0 #a; 1 # a P jQ ,! P jQ 0

0

hjj

Table 2.

jj

0

0

i

0

0

Proved transition system for CCS .

The proved transition system diers from the one in Tab. 1 in the rules for parallel composition and communication. Rule Par0 (Par1) adds to the label a tag jj0 (jj1) to record that the left (right) component is moving. The rule Com0 has in its conclusion a pair instead of a to record the components which interacted. The standard interleaving semantics [26] is obtained from the proved one by relabelling each transition through function ` in Def. 1. The relabelling is an abstraction from the proved transition system to the interleaving transition system semantics. This result is made precise in Subsection 6.2.

3 The abstract domain Causal CCS (CCCS) is a version of CCS introduced in [14] to give full account to causality relations between transitions. Processes are extended with a family of operators K ::, with K 2 }fin (IN). The intended meaning of K in K :: P is to encode the causes of the process P, i.e. the transition occurrences enabling P. The CCCS processes, called extended processes and denoted by t; t ; : : : 2 T , are de ned according to the syntax 0

t ::= K :: P j tjt j ( a)t: We assume that the operators K :: distribute over the restriction (a) and the parallel operator j, i.e. we assume on extended processes the least congruence that satis es the following clauses K

:: (r)t (r)K :: t

K :: (tjt ) (K :: t)j(K :: t ) 0

0

172

C. Bodei, P. Degano, and C. Priami

Also we let K :: K :: P = (K [ K ) :: P ; and we write K(t) for the set of causes in t, k for the singleton fkg, and K fK =kg for (K nk) [ K . Moreover, we let (K :: P )fK =kg = K fK =kg :: P and we homomorphically extend this replacement of a set of causes K for a single cause k onto extended processes t. Causal transitions have the form t 7,K! t , where is the standard action, ;k k 2 IN is the unique name of the transition and K represents the set of the transitions that enable k. The causal transition system in Tab. 3 is a variant of the original one of [14] (see also [24]), where the function f is1 k if 6= f(K; ) = K; ;; ; otherwise : 0

0

0

0

0

0

0

We write 2 (f(K; )) for the second component of f(K; ) and, by abuse of notation, 2 (f(K; )) 2= K, assuming it true if 2 (f(K; )) = ;. As usual, we omit in Tab. 3 the rule Kcom1 , symmetric to Kcom0 . ! Kact : K :: :P f 7, (f (K;)) [ K :: P; 2 (f (K;)) 62 K (K;) 2

Kide :

! t f 7, t (K;)

! t f 7,K; t 00

(

)

! t f 7, t (K;)

0

0

; t =t

Kres :

00

0

! (a)t f (7,K; (a)t )

! t f 7, t (K;)

0

Kpar0 :

(

)

0

00

; 2 (f (K;)) 62 K(t ) Ksum :

! t f 7, t (K;)

0

Kpar1 :

! t + t f 7, t (K;) 00

0

t0 K7,a0!;k t0 ; t1 K7,a1!;k t1 0

! t jt f 7, t jt (K;) 00

! t f 7, t (K;)

0

00

! tjt f 7,K; t jt 00

; 62 fa; ag

0

; 2 (f (K;)) 62 K(t ) Kcom0 : 0

0

t0 jt1 7,! t0 fK1 =kgjt1 fK0 =kg ; Table 3. Causal transition system for CCCS . 00

0

0

0

; ;

4 Rule-based inductive de nitions We rst recall some de nitions and results from [11]. Then, we instantiate them in our setting. 1

This function is introduced to avoid duplicating the rules: one set to handle transitions and the other for the visible transitions, as done in [24].

Constructing Specific SOS Semantics for Concurrency via Abstract Interpretation

173

De nition 2. An inductive de nition is a quadruple F = hU; ; ?; ti such that: { the universe U is a set; { is the set of rule instances Pc , with P U and c 2 U ; { ? 2 U is the basis; { t 2 }(}(U )) ! }(U ), is the (partial) join operator and the induced ordering v is x v y i x t y = y and is a partial order on }(U ). De nition 3. The operator induced by F = hU; ; ?; ti is : }(U ) ! }(U ) such that

P 2 g c The inductive de nition F is monotonic whenever is monotonic and ? v (?). Moreover F is on a cpo (resp. on a complete lattice) i h}(U ); v; ?; ti (resp. h}(U ); v; ?; >; t; ui ) is a cpo (resp. a complete lattice). (X ) = fc 2 U j9P X :

De nition 4. Given F = hU; ; ?; ti and F = hU ; ; ? ; t i, we write F ()

F for h}(U ); v; ?; ti ()

h}(U ); v ; ? ; t i, whenever h; i is a Galois 0

0

0

0

0

0

0

0

0

0

connection, i.e. if and only if { h}(U ); vi and h}(U ); v i are posets; { 2 }(U ) ! }(U ) and 2 }(U ) ! }(U ); { 8x 2 }(U ); 8y 2 }(U ) : [(x) v y] if and only if [x v (y)]. 0

0

0

0

0

0

The following proposition (corresponding to Prop.53 in [11]), will be useful later. Proposition 5. Let F = hU; ; ?; ti be such that 1. F is a monotonic inductive de nition on the cpo h}(U ); v; ?; ti, 2. h}(U ); v ; t i be a partial order, 3. 2 }(U ) ! }(U ) be a complete t- and [-morphism such that 8 Pc 2 ; 8X U : (P ) (X ) implies 9 Pc 2 : P X and (fcg) (fc g): 0

0

0

0

0

0

0

0

De ne = f (cP ) j Pc 2 ; with c 2 (c)g and ? = (?). Then F = hU ; ; ? ; t i is a well formed inductive de nition (i.e. a monotonic or extensive de nition on a cpo) such that : F ! F . If moreover h}(U ); v; ?; >; t; ui is a complete lattice then F () F , where

2 }(U ) ! }(U ) is (Y ) = tfX 2 U j (X ) v Y g. 0

0

0

0

0

0

0

0

0

0

0

0

More precisely, under the conditions of the above proposition we have the following (for the notation used and more technical details see Def. 46 and Proposition 53 in [11]). The abstraction function preserves least xed points lfp , [=] because F ()

F implies that (lfp ( )) = lfp ( ). In the following we use a special form of inductive de nition only, namely the positive one. These de nitions are well-formed, hence they have always the xed point (;). 0

v

1 [

v0

0

174

C. Bodei, P. Degano, and C. Priami

De nition 6. The inductive de nition FP = hUP ; P ; ?; ti is positive, and is written as hU; i, whenever ? is the emptyset ; and t is the set union [. The set of all the possible transitions involving CCS processes can be generated by a rule-based inductive de nition. In this way, instead of reasoning about the classical schemata of rules and axioms, we consider a set of rule instances, where axioms have empty premises. Some of the rules in Tables 2 and 3 have side-conditions, that can also be seen as additional premises. In fact, each rule with a side-condition stands for a set of rules, therefore we can safely use a rule-based inductive de nition.

4.1 The concrete domain FP

Here we give a rule-based inductive de nition of the proved transitions of CCS. De nition 7. The inductive de nition FP = hUP ; P i is the positive inductive de nition, where: { UP = fPP ,! P jP; P 2 P ; 2 g is the universe, and { P = f c j P UP ; c 2 UP and Pc is an instance of Act; Ide; Res; Sum; Par0; P ar1; Com0; Com1 g is the set of rule instances. The following properties of FP are easy to establish. 0

0

Proposition 8. The positive inductive de nition FP is monotonic and it is well-formed. Proof. We only need to prove that the operator P is monotonic and that

; P (;). Let X Y UP . Since c 2 P (X) implies 9P X : Pc 2 P and X Y , then c 2 P (Y ). Finally, ; P (;) holds because P (;) is a set. Proposition 9. h}(UP ); ; ;; UP ; [; \i is a complete lattice.

4.2 The abstract domain FC

The set of all the possible causal transitions involving CCCS processes can be generated by a rule-based positive inductive de nition, as well. De nition 10. FC = hUC ; C i is the positive inductive de nition, where: ! t j t; t 2 T ; 2 Act; K 2 }fin (IN)g is the universe; { UC = ft f 7,K; { C = f Pc j P UC ; c 2 UC and Pc is an instance ofKact; Kide; Kres; Ksum; Kpar ; Kpar ; Kcom ; Kcom g is the set of rule instances. Just as for FP , we have the following. Proposition 11. The positive inductive de nition FC is monotonic and well0

(

0

formed.

0

)

1

0

1

Proposition 12. h}(UC ); ; ;; UC ; [; \i is a complete lattice.

Constructing Specific SOS Semantics for Concurrency via Abstract Interpretation

175

5 The abstract interpretation function

To abstract inductive de nitions, we use an abstraction of subsets of the universe, according to Proposition 5. Our abstraction is based on a family of abstraction functions, which act on the components of the rule schemata and which are parameterized on a class S de ned according to the syntax S ::= K j (S jS ):

We write S i for the ith (from the left) set K 2 }fin (IN ) of (the frontier of the tree) S . For instance if S = (K j(K jK ))j(H jH ), then S 4 = H . The family of functions S is such that 0

K (P ) = K :: P

00

0

S0 S1 (P j Q) = S0 (P ) j S1 (Q): j

Moreover we assume that { K :: H (P ) = K H (P ); { K :: S0 S1 (P j Q) = K :: S0 (P ) j K :: S1 (Q); { (S0 j S1 )fH=kg = S0 fH=kg j S1 fH=kg. As an example of how an abstraction function works, consider the S introduced above and the process R = (P j(P jP ))j(a:QjQ ). Then [

j

0

00

0

S (R) = (K (K K )) (P j(P jP )) j (H H ) (a:QjQ ) = K (P ) j (K (P ) j K (P )) j (H (a:Q) j H (Q )) = (K :: P j (K :: P j K :: P )) j (H :: a:Q j H :: Q ) = tR . j

0j

0

00

0

0

0

00

0

00

j

0

0

00

00

0

0

00

0

0

Now, we can de ne the abstraction function pc that maps proved transitions in causal transitions. It will be used to obtain the causal transition system by abstracting the rules of the proved one. De nition 13. Let pc : }(UP ) ! }(UC ) be de ned as # pc(X ) = fpc(P ,! P ) = S (P ) ,! S (P ) j P ,! P 2 Xg [ ; a ;# a fpc(P (#a;# ,!a ) P ) = S0 (P ) ,! S1 (P ) j P # # ,! P 2 X; ; j j p p i i 9i; j #a : S1 = S1 = S0 [ S0 and 8p 6= i; j : S1 = S0 g [ a #a p fpc(P ,! P ) = S0 (P ) ,! K ;k S1 (P ) j P ,! P 2 X; 8p : k 62 S0 and 9i; K : S0i = K; S1i = K [ fkg; 8j 6= i : S1j = S0j g To see our de nition at work, consider again the S and R introduced above, 1 0a and the following proved transition R ,! R = (P j(P jP ))j(QjQ ): Then, its a abstraction is tR 7,H! t , for h 62 K [ K [ K [ H [ H and ;h R S = (K j(K jK ))j(H [ fhgjH ), where 0

0

0

; ;

0

0

0

h

0

0

0

00

00

i

0

; ;

0

0

0

jj jj

0

0

0

0

00

0

0

00

0

0

00

0

176

C. Bodei, P. Degano, and C. Priami

tR = S (R ) = (K :: P j (K :: P j K :: P )) j (H [ fhg :: Q j H :: Q ). 0

0

0

0

00

00

0

0

It is easy to deduce this causal transition with the rules in Tab. 3. The interested reader may wish to abstract step by step its whole proved derivation. We wish to ! (f (K;)) :: K (P ); 2 (f (K;)) 62 K r (Act) = K (:P ) f 7, (K;) 2

r (Ide) =

) 7,l(! S (P ) f (K;` (f (K;`())) :: S (P ) ( )) 2 ) 7,`(! S (A) f (K;` (P ) ( )) S ) 7,`(! S (P ) f (K;` (P ) ( )) S 0

0

,7 ` ! S (a)P ) f K;` S ((a)P ) ( )

(

r (Sum) =

) 7,`(! S (P ) f (K;` (P ) ( )) S 0

0

) 7,`(! S (P + Q) f (K;` (P ) ( )) S 0

0

r (Par0 ) =

) 7,`(! S0 (P ) f (K;` (P ) ( )) S0 0

0

S0 S1 (P j Q 7,! S0 S1 (P j Q) `( ) ) f (K;`( ))

j

r (Par1 ) =

; `() 62 fa; ag

0

0

( ))

; S (A) = S (P )

0

0

r (Res) =

0

0

0

0j

) 7,`(! S1 (Q) f (K;` (Q ) ( )) S1 0

0

S0 S1 (P j Q 7,! S0 S1 (P j Q ) `( ) ) f (K;`( ))

j

r (Com0) =

; 8i : 2 (f (K;`())) 62 S1i

j

; 8i : 2 (f (K;`())) 62 S0i

0

0

S0 (P0 ) f (7,Ka! (P0 ); S1 (P1) f (7,Ka! (P1 ) 0 ;a) S0 0 ;a) S1 0

0

S0 S1 (P0 j P1) 7,! S0 ; j

Table 4.

; ;

0

K1 =k g

0f

j

P j P1) 0

S10 fK0 =k g ( 0

The abstraction interpretation

0

0

r of rule instances.

use Proposition 5 to establish a Galois connection between the de nitions of the proved and the causal transition systems. We already know that the abstract domain is a complete lattice (Proposition 12). We are left to show that our abstraction function pc enjoys the properties required by item 3 of Proposition 5. The function pc a complete [-morphism, i.e. 8X UP : [X 2 UP implies [ pc (X ) = pc ([X ):

Proposition 14.

Constructing Specific SOS Semantics for Concurrency via Abstract Interpretation

177

Proof. It is easy to prove that the function pc is monotone. We now prove the

double inclusion and (note that induced ordering is the same for both domains). We start with . By de nition of [, 8X UP ; pc (X ) [pc(X ) and X [X . Then, by monotonicity of pc , it is pc (X ) pc ([X ): Since [ is the least upper bound, [ pc (X ) pc([X ): We now prove . Let F K be the function introduced in Def. 25 that deletes the causes from an extended process, restoring a pure CCS process. i By hypothesis [pc (X ) = fti K7,i! t j i 2 I g [ ftj 7,! tj j j 2 J g UC . Then, ;ki i ; 0

0

0

; ;

p by construction, there exists a set Y = fPp ,! Pp j F K (tp ) = Pp ; F K (tp ) = p Pp; `(p ) = p j p 2 I [ J g, such that c (Y ) = [pc(X ). First we prove 8Y~ X : Y~ Y . Per absurdum, let S X such that S Y . Then, by monotonicity of pc , pc (S ) pc (Y ) = [pc (X ), against the hypothesis of [pc(X ) to be the least upper bound. Then, 8Y~ X : Y~ [X , as well. From this and from Y~ Y , [X Y , because [X is the least upper bound. By monotonicity of pc , we conclude pc ([X ) pc (Y ) = [pc (X ). 0

0

0

Proposition 15. The function pc is such that 8 Pc 2 P ; 8X UP : pc(P ) pc (X ) implies 9 Pc 2 P : P X and pc (fcg) pc (fc g): Proof. Since pc (P ) = 6 pc (P ), whenever P 6= P , pc (P ) pc (X ) implies P X . Hence, we only need to put P = P . 0

0

0

0

0

0

0

To de ne the abstract interpretation of the operator , wep introduce in Tab. 4 the auxiliary family of functions r . Note that r ( Pc ) = cc(P ) with c 2 pc (c) (cf. the de nition of in Proposition 5). We can now establish the required Galois connection, that preserves least xed points. 0

0

0

Theorem 16. Let = fr ( Pc ) j Pc 2 P g and ; = r (;). The positive inducc p tive de nition F = fpc (U ); g is well formed and such that F () F , i.e.

cp hpc ; pc i is a Galois connection, where pc : }(UC ) ! }(UP ) is X 2 UP j pc (X ) Y g if Y 2 pc (UP ) c

p (Y ) = [f UP if Y 2= pc (UP ) 0

0

0

[=]

P

0

Proof. By Proposition 5, whose hypotheses are satis ed because of Proposi-

tions 8, 9, 11, 12, 14, 15.

Although pc (UP ) UC , we have established exactly what we would like to obtain. In fact, it is immediate to prove the following property, because the rules in Tables 3 and 4 can be easily put in bijection.

Proposition 17. r (P ) = C .

178

C. Bodei, P. Degano, and C. Priami

The above correspondence suces to establish that the computations of the causal transition system obtained via abstraction pc coincide with the abstraction of the proved computations as de ned in [5]. Actually, a stronger result holds: the trees of computations coincide, so the non deterministic aspects of computations are preserved, as well. Lack of space prevents us from recalling the relevant de nitions and results of [5]. We simply introduce some notation. Let P tree(P ) be the tree de ned as the set of the proved computations starting from P , ordered by pre x. Similarly, the tree of causal computations is Ktree(P ), with transitions derived by the rules in Tab. 3; and nally let pcP tree(P ) be the tree of computations with transitions deduced according to pc(P ). We denote by h; i the Galois connection of Def. 17 in [5], where maps the tree of proved computations P tree(P ) to the tree of causal computations Ktree(P ). Now, we can state the following. Theorem 18. For all P 2 P , we have that pcP tree(P ) = Ktree(P ). Proof. The bijection between Tab. 3 and Tab. 4 suces, because (P tree(P )) = Ktree(P ) by Theorem 18 of [5].

6 A fragment of a hierarchy

pc pc

C

pi

P ci

pi

ci Fig. 1.

I

The hierarchy of semantics.

We now de ne a simple hierarchy of transition systems involving the proved, the causal and the standard interleaving semantics of CCS . The hierarchy in Fig. 1 is established by studying three abstraction functions: from the proved to the causal model and to the interleaving model, and from the causal to the interleaving one. This fragment of a hierarchy is formally constructed, and in the conclusions we sketch how other models can be derived by further constructive abstractions of the proved model.

6.1 The abstract domain FI

We start with the de nition of the abstract domain made of the standard interleaving transition system of CCS [26]. The set of all the possible transitions involving CCS processes can be generated by the following positive rule-based inductive de nition.

Constructing Specific SOS Semantics for Concurrency via Abstract Interpretation

179

De nition 19. FI = hUI ; I i is the positive inductive de nition, where: { UI = fP ,! P j P; P 2 P ; 2 Actg is the universe; P { I = f c j P UI ; c 2 UI and Pc is an instance of act; ide; res; sum; par ; par ; com ; com g is the set of rule instances. Just as for FC , we have the following. Proposition 20. The positive inductive de nition FI is monotonic and well0

0

0

1

0

1

formed.

Proposition 21. h}(UI ); ; ;; UI ; [; \i is a complete lattice. 6.2 The abstraction from proved to interleaving transition systems We make now precise the observation made at the end of Section 2, by de ning a Galois connection between FP and FI . So one can easily recover the standard interleaving transition system from the proved one.

De nition 22. Let pi : }(UP ) ! }(UI ) be de ned as ` pi(X) = fpi(P ,! P ) = P ,! P j P ,! P 2 X g; where ` is as in Def. 1. ( )

0

0

0

Also in this case is possible to reformulate the results relative to the Galois p connection pand the consequent correspondences. Note again that r ( Pc ) = ic(P ) with c 2 i (c). 00

0

0

Theorem 23. Let =p fr ( Pc ) j Pc 2 P g and ; = r (;). The positive inductive de nition F = fi (UP ); g is well formed and is such that p i FP () F I , i.e. hpi ; pi i is a Galois connection, where pi : }(UI ) ! }(UP ) is

pi 00

00

[=]

00

00

00

0

X 2 UP j pi (X) Y g if Y 2 pip (UP )

pi (Y ) = [f UP if Y 2= i (UP )

The following proposition says that the induced operator above is indeed I , and justi es the use of ,! in Def. 22. 00

Proposition 24. r (P ) = I . 00

6.3 The abstraction from causal to interleaving transition systems Our next abstraction is the function ci that maps causal transitions in standard interleaving transitions. It will be used to obtain the interleaving transition system by abstracting the rules of the causal one. The steps are the same as above.

180

C. Bodei, P. Degano, and C. Priami

De nition 25. Let ci : }(UC ) ! }(UI ) be de ned as ci(X ) = fci(t f ,! K; t ) = F K (t) ,! F K (t ) j t f ,! K; t 2 X g: (

)

0

0

(

)

0

where F K deletes the causes from extended processes, de ned as F K (K :: t) = F K (t), F K (tjt ) = F K (t)jF K (t ), F K ((a)t) = (a)F K (t). 0

0

Theorem 26. Let = fr ( Pc ) j Pc 2 C g and ; = r (;). The positive inductive de nition F = fci (UC ); g is well formed and is such that ci FC () F I , i.e. hci ; ci i is a Galois connection, where ci : }(UI ) ! }(UC ) is

ci 000

000

000

[=]

000

000

000

ci (Y ) =

[fX 2 UC j ci (X ) Y g if Y 2 ci (UC ) UC if Y 2= ci (UC )

Note that F K (t) is a standard CCS process and the following proposition justi es the presence of ,! in Def. 25.

Proposition 27. r (C ) = I . 000

6.4 The hierarchy The commutativity of the diagram in Fig. 1 is now immediate.

Theorem 28. ci(pc(P )) = pi(P ) = I . The above theorem suces to show a correspondence between the trees of computations in the various models. The trees of computations obtained by abstracting proved computations to interleaving ones coincide with the trees of the computations based on the standard interleaving semantics. The same happens abstracting causal computations. Therefore, the non deterministic aspects of agent behaviour are fully respected. This supports our claim that formal constructive methods can be used to mechanically derive, clarify and compare various dierent models for concurrency. Let Stree(P ) be the synchronization tree de ned as the set of the interleaving CCS computations starting from P , ordered by pre x [26]. Also, let pi P tree(P ) and ci Ktree(P ) be the trees of computations with transitions deduced with p i (P ) and with ci (C ). The following is immediate.

Theorem 29. For all P 2 P , piP tree(P ) = Stree(P ) = ciKtree(P ). 7 Conclusions

We have constructed the causal and the interleaving transition systems for CCS by abstracting on the rules of the proved transition system. Lack of space prevents us from giving more evidence that quite similar constructions are possible, and give most of the interleaving and the non-interleaving SOS de nitions

Constructing Specific SOS Semantics for Concurrency via Abstract Interpretation

181

presented in the literature for concurrency calculi. Their designers aimed at describing dierent qualitative and quantitative aspects of process evolutions, and a considerable eort has been spent in their actual SOS de nitions. Our proposal may help in mechanically deriving these SOS de nitions, keeping in mind only the eect that a single transition has on the state of the whole system of concurrent processes. Consider for instance the eect that a visible action a has on the S in Def. 13: only the causes K of the process performing a are aected | this obvious fact, together with exchange of causes in communications, are sucient to obtain the rules for the causal transition system. Despite space limitation, we would like to give some hints of three more abstraction functions, thus showing that our construction is in a sense paradigmatic. Consider the location transition system of [8]. It was designed to keep track of the sites (taken from a given set Location) where the visible actions actually occur, and it assumes that the sites of invisible ones cannot be detected. Having this in mind, it is easy to de ne the location abstraction functions pl , mimicking Def. 13, as follows. pl(X) = fpl(P ,! P ) = L (P) ,! L (P ) j P ,! P 2 X; `() = g [ a #a #a fpl(P ,! P ) = L (P) ,! lu L (P ) j P ,! P 2 X; p i 8p : l 62 L0 and 9i : L0 = u; Li1 = lu; and 8j 6= k : Lj1 = Lj0 g; where l 2 Location; u 2 Location and L is either u or LjL (i.e. L is a tree of strings of locations, rather than a tree of sets of natural numbers), and L is a family of abstraction functions de ned just as S . Indeed, the only dierence between Def. 13 and the one sketched above is that communications do not exchange causes, a typical feature of causality, not shared by location models. It is easy then to establish a Galois connection hpl ; pl i and results similar to Theorems 16 and 18. At this point, the reader familiar with the local/global transition system of [24] can easily obtain it by de ning an abstraction function pl=g that combines pc and pl and that enjoys all the properties these abstraction functions have. As a third example, we consider a SOS semantics that expresses quantitative aspects of concurrent computations. Consider the timed transition system of [19], whose transitions have labels of the form ha; ni ? w, where n is the time required to complete the standard action pa at location w. We can easily modify pl to obtain the abstraction function t that renders the considered timed transition systems. Assume as given a function time(a) that assigns a duration to actions (we only consider visible actions, because the calculus of [19] has no 's, and we slightly change their notion of location). This abstraction function is as follows. #a #a ptp(X) = fpt(P ,! P 0) = L (P ) ha;n ,!i?w jL (Pj0) j P ,! P 0 2 X; 8p : l 62 i i L0 and 9i : L0 = u; L1 = lu; and 8j 6= k : L1 = L0 ; n = time(a)g: We have not yet considered transition systems expressing other qualitative aspects of concurrent processes. However, we are very con dent that most of them 0

0

0

0

0

0

0

0

182

C. Bodei, P. Degano, and C. Priami

can be cast in our framework, and that abstract interpretation is a exible tool for constructing speci c SOS semantics for concurrency calculi, including those for mobility, like the -calculus [27].

Acknowledgment. The authors wish to thank the referees for their precise remarks and helpful comments.

References 1. L. Aceto. A static view of localities. Formal Aspects of Computing, 1992. 2. R.M. Amadio and S. Prasad. Localities and failures. In Proceedings of FSTTCS'94, pages 205{216. Springer-Verlag, 1994. 3. M. Bernardo, L. Donatiello, and R. Gorrieri. Integrating performance and functional analysis of concurrent systems with EMPA. Technical Report UBLCS-95-14, University of Bologna, Laboratory for Computer Science, 1995. 4. C. Bodei, P. Degano, and C. Priami. Mobile processes with a distributed environment. In Proceedings of ICALP'96, LNCS 1099, pages 490{501. Springer-Verlag, 1996. To appear in TCS. 5. C. Bodei and C. Priami. True concurrency via abstract interpretation. In Proceedings of SAS'97, LNCS 1302, pages 202{216, 1997. 6. M. Boreale and D. Sangiorgi. A fully abstract semantics of causality in the calculus. In Proceedings of STACS'95, LNCS. Springer Verlag, 1995. 7. G. Boudol and I. Castellani. A non-interleaving semantics for CCS based on proved transitions. Fundamenta Informaticae, XI(4):433{452, 1988. 8. G. Boudol, I. Castellani, M. Hennessy, and A. Kiehn. A theory of processes with localities. Theoretical Computer Science, 114, 1993. 9. P. Buchholz. On a markovian process algebra. Technical report, Informatik IV, University of Dortmund, 1994. 10. P. Cousot and R. Cousot. Abstract interpretation frameworks. Journal of Logic and Computation, 2(4):511{547, 1992. 11. P. Cousot and R. Cousot. Inductive de nitions, semantics and abstract interpretation. In Proceedings of POPL'92, pages 83{94, 1992. 12. P. Cousot and R. Cousot. Higher-order abstract interpretation (and application to comportment analysis generalizing strictness, termination, projection and PER analysis of functional languages). In Procs. ICCL'94, IEEE, pages 95{112, 1994. 13. P. Cousot and R. Cousot. Compositional and inductive semantic de nitions in xpoint equational, constraint, closure-condition, rule-based and game-theoretic form. In Proceedings of CAV'95, LNCS 939, pages 293{308, 1995. 14. Ph. Darondeau and P. Degano. Causal trees. In Proceedings of ICALP'89, LNCS 372, pages 234{248. Springer-Verlag, 1989. 15. P. Degano, R. De Nicola, and U. Montanari. Partial ordering derivations for CCS. In Proceedings of FCT, LNCS 199, pages 520{533. Springer-Verlag, 1985. 16. P. Degano, J.-V. Loddo, and C. Priami. Mobile processes with local clocks. In Proceedings of Workshop on Analysis and Veri cation of Multiple-Agent Languages, Stockholm, Sweden, 1996. 17. P. Degano and C. Priami. Proved trees. In Proceedings of ICALP'92, LNCS 623, pages 629{640. Springer-Verlag, 1992. 18. P. Degano and C. Priami. Non interleaving semantics for mobile processes. Extended abstract. In Proceedings of ICALP'95, LNCS 944, pages 660{671. SpringerVerlag, 1995. To appear in TCS.

Constructing Specific SOS Semantics for Concurrency via Abstract Interpretation

183

19. R. Gorrieri, M. Roccetti, and E. Stancapiano. A theory of processes with durational actions. Theoretical Computer Science, (140), 1994. 20. N. Gotz, U. Herzog, and M. Rettelbach. TIPP- a language for timed processes and performance evaluation. Technical Report 4/92, IMMD VII, University of Erlangen-Nurnberg, 1992. 21. E. Goubault. Durations for truly concurrent transitions. In Proceedings of ESOP96, LNCS 1058, pages 173{187, 1995. 22. M. Hennessy and T. Regan. A temporal process algebra. Technical Report 2/90, University of Sussex, 1990. 23. J. Hillston. The nature of synchronization. In U. Herzog and M. Rettelbach, editors, Proceedings of PAPM'94, University of Erlangen, 1994. 24. A. Kiehn. Comparing causality and locality based equivalences. Acta Informatica, 31(8):697{718, 1994. 25. K.G. Larsen and A. Skou. Compositional veri cation of probabilistic processes. In Proceedings of CONCUR'92, volume 630 of LNCS. Springer-Verlag, 1992. 26. R. Milner. Communication and Concurrency. Prentice-Hall, London, 1989. 27. R. Milner, J. Parrow, and D. Walker. A calculus of mobile processes (I and II). Information and Computation, 100(1):1{77, 1992. 28. U. Montanari and M. Pistore. Concurrent semantics for the -calculus. In Electronic Notes in Computer Science, number 1. Elsevier, 1995. 29. U. Montanari and D. Yankelevich. A parametric approach to localities. In Proceedings of ICALP'92, LNCS 623, pages 617{628. Springer-Verlag, 1992. 30. X. Nicollin and J. Sifakis. An overview and synthesis on timed process algebras. In Real Time: Theory in Practice, LNCS 600, pages 526{548. Springer-Verlag, 1991. 31. C. Priami. Stochastic -calculus. The Computer Journal, 38(6):578{589, 1995. 32. C. Priami. Enhanced Operational Semantics for Concurrency. PhD thesis, Dipartimento di Informatica, Universita di Pisa, March 1996. Available as Tech. Rep. TD-08/96. 33. C. Priami. Interleaving-based partial odering semantics. In Proceedings of Italian Conference on Theoretical Computer Science, pages 264{278, Ravello, November 1995, 1996. World Scienti c. 34. R.J. van Glabbeek, S.A. Smolka, B. Steen, and C.M.N. Tofts. Reactive, generative and strati ed models of probabilistic processes. Information and Computation, 1995.

A First-Order Language for Expressing Aliasing and Type Properties of Logic Programs Paolo Volpe Dipartimento di Informatica Universit` a di Pisa Corso Italia 40, 56125 Pisa, Italy e-mail: [email protected] tel: +39-50-887248. fax: +39-50-887226.

Abstract. In this paper we study a first-order language that allows to express and prove properties reagarding the sharing of variables between non-ground terms and their types. The class of true formulas is proven to be decidable through a procedure of elimination of quantifiers and the language, with its proof procedure, is shown to have interesting applications in validation and debugging of logic programs. An interesting parallel is pointed out between the language of aliasing properties and the first order theories of Boolean algebras. Keywords: Verification of logic programs, languages of specification, first-order logic.

1

Introduction

In many approaches to the verification of properties of logic programs, a formal language is required that allows to express the properties of programs one is interested in. In the methods proposed in [13][1][12][16], an assertional language is assumed to verify properties of arguments of predicates of the programs. Some verification conditions are provided that imply the partial correctness of the programs with respect to various aspects of the computations. For example the method proposed in [12] and [6] allows to prove properties of the correct answers of the programs, while in [16] a method is provided to prove properties of the computed answers. The methods proposed in [13] and [1] allow to prove, in addition, that predicates verify given specifications at call time. In this paper we study a language that allows to express an interesting class of properties of non-ground terms, that is the data on which logic program operate. The language is sufficiently expressive to capture sharing, freeness and types of non-ground terms. Two or more terms are said to share, when they have at least one variable in common, while a term is free when it is a simple variable. For types, we will refer to term properties like being a list, a tree, a list of ground terms, etc. Fragments of this language have already been studied in [18] and [2] and shown to be decidable, but in this paper we show that the full first-order theory is decidable. This allows to use its full expressive power in existing proof methods G. Levi (Ed.): SAS’98, LNCS 1503, pp. 184−199, 1998. Springer-Verlag Berlin Heidelberg 1998

A First-Order Language for Expressing Aliasing and Type Properties

185

and algorithmically decide whether the verification conditions are true or not. Indeed for many methods the verification conditions are expressed by formulas of the language. If the verification condition is true, we show the partial correctness of logic programs with respect to a property belonging to the class of aliasing properties and type assertions. If the verification condition is false (and we stress that this can be checked in finite time), obviously this does not mean that the program has necessarily an error. Anyway a “warning” can be raisen up, signalling a possible wrong situation. The proof procedure shown in the paper can be easily enriched so as to provide a counterexample in this case. This allows the user to have more information about the warning and to decide whether to change the program (the counterexample is actually a wrong computation of the program) or to refine the specification (the verification condition is false because the specification is too “loose” and impossible computations are considered). We want to emphasize the independent importance of the proof of decidability of the language. It is based on the method of elimination of quantifiers and points out an interesting set of formulas, which can be viewed as expressing constraints on the cardinality of the sets of variables that can occur in terms. Our proof is based on the parallel between the satisfiability of formulas of our language and the satisfiability of such cardinality constraints, which can be proven decidable as a consequence of the decidability of the theory of Boolean algebras. We think that such class of constraints, which has a quite simple representation and operations of composition and cylindrification, can be of interest in program analysis. For example, well known abstract domains such as POS [9] and Sharing[14] can be naturally viewed as subdomain of the class of cardinality constraints (see also [19]), with their composition and the cylindrification operator obtained as instances of the general ones. The paper is organized as follows. In Section (2) we lay down the basic terminology. In Section (3) we define the class of types from which type assertions are built. In Section (4) we define formally the language, parametrically with respect to a family of regular types, and then we prove in Section (5) that such a language is decidable. Finally in Section (6) we give examples of application in the context of inductive proof methods.

2

Preliminaries

Throughout the paper, we assume familiarity with standard notions of logic programming and mathematical logic [17, 15]. A first order language L = hΣ, Π, Vi is based on a set Σ of function symbols, a set Π of predicate symbols of assigned arities and a set V of variables. The set Terms(Σ, V) of all terms of L with variables in V is defined as usual. The set Terms(Σ, ∅) ⊆ Terms(Σ, V) is the subset of ground terms (i.e. not containing variables). We write f , g for function symbols, p, q for predicate symbols, X, Y for variables, X for tuples of distinct variables, t, s for terms, t,s for tuples of terms. A (predicate or function) symbol f with arity n will be denoted by f (n) . Atomic formulas (also called atoms) are formulas like p(t1 , . . . , tn ), with

186

P. Volpe

p(n) ∈ Π and t1 , . . . , tn ∈ Terms(Σ, V). The set of formulas of L is the smallest set F containing the atomic formulas and such that if φ and ψ are in F then ¬φ, φ ∧ ψ and ∃Xφ, with X ∈ V, are in F. We will use the notation φ ∨ ψ as a shorthand for ¬(¬φ ∧ ¬ψ), φ ⇒ ψ for ¬φ ∨ ψ, φ ⇔ ψ for (φ ⇒ ψ) ∧ (ψ ⇒ φ) and ∀Xφ for ¬(∃X¬φ). We assume also the constants true and false to be in F. Given a syntactic object O of L, Vars(O) denotes the set of free variables (not bound by any quantifier) of V in O. Substitutions are defined as mapping θ : V → Terms(Σ, V), which differ from the identity just on a finite subset of V. They can be extended homomorphically to functions on terms. A preorder can be introduced on Terms(Σ, V). Given t, s ∈ Terms(Σ, V), t ≤ s iff there exists substitution η such that tη = s. The induced equivalence on terms is called variance. An interpretation I = hD, Λ, Γ i of L consists of a non-empty set D, the domain; a set of functions Λf : Dn → D for each function symbol f (n) ∈ Σ; a family of subsets Γp ⊆ Dn for each predicate symbol p(n) ∈ Π. A variable assignment (also called state) σ : V → D maps each variable into an element of D. It can be lifted homomorphically to a function, still denoted by σ, which maps terms in Terms(Σ, V) to elements of D. An atom p(t1 , . . . , tn ) is true in I under the state σ, written I |=σ p(t1 , . . . , tn ), iff (σ(t1 ), . . . , σ(tn )) ∈ Γp . The truth of each formula of L under the state σ, written I |=σ ϕ, is defined, as usual, by induction on the structure of ϕ. A formula ϕ is true in I, that is I |= ϕ, if and only if for each state σ : V → D, I |=σ ϕ. The set Th L (I) is defined as the set of all formulas of L true in I, that is as the set {ϕ | ϕ formula of L and I |= ϕ}. In this paper we will be mainly interested in the non ground term interpretations H = hTerms(Σ, V), Λ, Γ i, with the set Terms(Σ, V) as domain and with Λf (t1 , . . . , tn ) = f (t1 , . . . , tn ) for each f (n) ∈ Σ and tuple t1 , . . . , tn of Terms(Σ, V). In practice every (ground) term is interpreted by itself. We have chosen not to distinguish between the variables in formulas and the variables in the model. Indeed their roles are quite different and in practice no ambiguity arises. The preorder on Terms(Σ, V) induces a preorder on the states σ : V → Terms(Σ, V) . The induced equivalence between states is still called variance.

3

Regular term grammars

To specify families of types we will consider regular term grammars. There is a large amount of papers on regular types. They have proved them to be a good trade-off between expressibility and decidability. In fact they are strictly more expressive than regular languages, but strictly contained in context-free languages (which have an undecidable subset relation). Our main references are the papers of Dart and Zobel [10, 11] and Boye and Maluszynski [3, 2]. A regular term grammar is a tuple G = (Σ, V, T, R), where Σ is a set of function symbols, V is an infinite denumerable set of variables, T is a finite set of type symbols, including var and any, and R is a finite set of rules l → r where – l ∈ (T\{var , any})

A First-Order Language for Expressing Aliasing and Type Properties

187

– r ∈ Terms(Σ, T) For every T ∈ T\{var , any}, we define Def G (T ) (also denoted by Def R (T )) as the set {r | T → r ∈ R}. Def G (var ) is defined as the set of variable V, while Def G (any) as the set Terms(Σ, V). We use the notation T1 →G T2 if T2 is obtained from T1 by replacing a symbol T ∈ T by a term in Def G (T ). Let → →G be the transitive and reflexive closure of →G . Given the type symbol T ∈ T, we define the set of terms [T ]G , the type T , as the set {s ∈ Terms(Σ, V) | T → →G s}. Notice that [var ]G = V and [any]G = Terms(Σ, V). We assume function symbols in Σ to contain at least a constant and a function of arity 2. We will often omit the subscript when the grammar is clear from the context. Example 1. Let us see some examples (taken in part from [2]) of regular types (k ) (k ) and of the grammars that define them. Let us suppose that Σ = {f1 1 , . . . , fn n }, (ki ) where each function symbol fi has arity ki . The set of ground and instantiated terms can be defined as: (0)

inst → f1 .. .

(0)

(0)

inst → fi−1 (k ) inst → fi i (any, . . . , any) .. .

ground → f1 .. .

ground → fi−1 (k ) ground → fi i (ground , . . . , ground ) .. . (kn )

ground → fn

(ground , . . . , ground )

(0)

(kn )

inst → fn

(any, . . . , any)

The type of lists, the lists of instantiated terms, the ground lists and the list of variables can be defined as : list → [ ] ilist → [ ] glist → [ ] vlist → [ ] list → [any | list] ilist → [inst | ilist] glist → [ground | glist] vlist → [var | vlist] Notice that if the type symbol var is available, the type any can be defined by any → var (0) any → f1 .. . (0)

any → fi−1

(ki )

any → fi .. .

(any, . . . , any)

(kn )

any → fn

(any, . . . , any)

The only reason to retain it, is that in closed grammars the type var in no longer available, but we still want to define non-ground types. Regular term grammars enjoy several remarkable properties. The following lemmas can be shown, by slightly generalizing results and algorithms given in [10] and in [2]. Theorem 1 ([10]). Given a regular term grammar G and a symbol type T , the set [T ]G is decidable.

188

P. Volpe

Lemma 1 ([10]). The emptiness and the subset relation of regular types is decidable. Indeed, given a grammar G and two type symbols T and S, it is possible to extend G into G0 , with a new symbol S ∩ T and new rules in such a way that [S ∩ T ]G0 = [S]G ∩ [T ]G . Lemma 2 ([10]). There exists an algorithm that computes the intersection of regular types In this paper we will be mainly concerned with closed discriminative regular grammars in normal form. Definition 1. A regular term grammar G = (Σ, V, T, R) is in normal form if each rule have the form T → var or T → f (T1 , . . . , Tn ) with f (n) ∈ Σ , T ∈ T\{var , any}and T1 , . . . , Tn ∈ T. It can be easily shown that each type can be defined by a grammar in normal form. Definition 2. A regular term grammar G is discriminative if it is in normal form and, for each type symbol T , the top functors in Def G (T ) are pairwise distinct. Definition 3. A regular term grammar is closed if it is in normal form and, for each type symbol T , the symbol var does not occur in any element of Def G (T ). Notice that most of the types used in logic programming allow closed and discriminative term grammars. For example, all the grammars introduced in (1), but for vlist, are discriminative and closed. It can be easily shown that regular types defined by a closed grammar are indeed closed under substitution. Theorem 2 ([2]). Given a closed regular term grammar G and a symbol type T , the set [T ]G is closed under substitution. The types that can be defined by a discriminative and closed grammar will be referred to as simple types. Notice that if S and T are simple types, the intersection type S ∩ T is still simple. An operation on types which we will need in the following is the difference of types. Given a discriminative and closed regular term grammar G and two type symbols T and S, we want to extend it to G0 with a new symbol T /S and new rules in such a way that [T \S]G0 = [T ]G \ [S]G , if it is not the case that [T ]G ⊆ [S]G . In general the grammar G0 need not to be nor closed nor discriminative in general. Example 2. Consider the difference type any/inst. It can easily be checked that the set [any] / [inst] is equal to the set of variables V. Anyway there is no closed grammar for such set of terms.

A First-Order Language for Expressing Aliasing and Type Properties

189

We will provide an algorithm which computes the difference T /S, assuming that type S is simple. The algorithm for the general case can be defined but it is more complex and we do not need such a generality. In fact we just need to compute differences types like T1 ∩ · · · ∩Tn \S1 \ · · · \ Sk , where T1 , . . . , Tn and S1 , . . . , Sk are simple types. Difference Algorithm Input. Two type symbols T and S and the set R of rules defining T and S, with S simple. Output. A pair (T \S, S), where T \S is defined by the rules in S, if S 6= ∅; otherwise the difference is empty ([T ] ⊆ [S]). Method. The algorithm is defined by the following recursive function. A set I of difference symbols T \S is used to ensure termination. The type symbol any is supposed to be unfolded as in example (1). difference(T, S, R) = difference(T, S, R, ∅) difference(T, S, R, I) = – If T ⊆ S then return (T \S, ∅); – If the symbol T \S is in I then retun (T \S, R); – Otherwise, let Def R (T ) = {r1 , . . . , rk }. For each i ∈ {1, . . . , k}, let Hi be defined as follows: • if ri = var or (ri = fi (T1 , . . . , Tni ) and the functor fi does not occur in Def R (S)) then let Hi = {(T \S) → ri }; • If ri = fi (T1 , . . . , Tni ) and fi (S1 , . . . , Sni ) ∈ Def R (S) then let (Tj \Sj , Sj ) = S difference(Tj , Sj , R, I ∪ {T \S}), for each j = {1, . . . , ni }, and let Hi = Sj 6=∅ {(T \S) → fi (T1 , . . . , Tj \Sj , . . . , Tni )} ∪ Sj Sk Return (T \S, R ∪ i=1 Hi ) Lemma 3. Let T and S be two type terms defined by the rules of R, S simple. Then difference(T, S, R) terminates and returns a pair (T \S, S) such that [T \S]S = [T ]R \ [S]R . To carry on the elimination of quantifiers in the next section, we need to know the cardinalities of the sets of variables which may occur in a term of a given type. Definition 4. Given a type symbol T defined by a grammar G, the var-cardinality of T , written |T |, is defined as the set {|Vars(t)| | t ∈ [T ]G }. In other terms, k ∈ |T | if and only if there exists t ∈ [T ] such that |Vars(t)| = k. We can prove in a straightforward way the following lemmas. Lemma 4. Given a simple type T , then |T | is equal to {0} or to ω. It is decidable which is the case.

190

P. Volpe

Proof. The relation T ⊆ ground ,with ground as defined in example (1), is decidable by lemma (1). If it is the case then |T | = {0}, otherwise |T | = ω, since T is substitution closed, the signature Σ contains at least a function symbol of arity 2 and a constant, and the set of variables is infinite denumerable. For difference types things can be more complex. Example 3. Consider again the difference type any\inst. It can be easily seen that |any\inst| is the set {1}. Anyway the behaviour of difference types T1 ∩ · · · ∩Tn \S1 \ · · · \ Sk , with T1 , . . . , Tn and S1 , . . . , Sk simple types, is sufficiently regular so as to show the following theorem. Theorem 3. The var-cardinality of the type T1 ∩ · · · ∩Tn \S1 \ · · · \ Sk , with T1 , . . . , Tn and S1 , . . . , Sk simple types, is equal to S ∪ [k, ω], where S is a finite set of natural number and [k, ω], with k = 1, . . . , ω is the set of natural numbers greater or equal to k . The proof provides an effective procedure to compute the var-cardinality of a difference type.

4

A language of properties

In this section we introduce a language that allows to express properties of terms used in static analysis and verification of logic programs: these include groundness, freeness, sharing, type assertions. The language is parametric with respect to a family of types defined through a regular term grammar. It is an extension of the language proposed by Marchiori in [18]. We assume a regular term grammar G = (Σ, V, T, R), discriminative and closed, describing the family of types we are interested in. As before, the set of function symbols Σ is assumed to contain at least a constant and a function of arity 2, V is assumed to be a denumerable set of variables. We define then a first-order language LG = hΣ, Π, Vi, starting from the regular term grammar G. The set of predicate symbol Π consists of the predicates var (1) , share (n) , for (1) each natural n, and a unary predicate pT for each symbol type T ∈ T/{any}. Often we will write T (t) for the atomic formula pT (t). For example, pground (t) will be often written as ground (t). We will omit the subscript G in LG , when no confusion arises. Like in [18], we give the semantics of formulas L by considering the nonground Herbrand interpretation H = hTerms(Σ, V), Λ, Γ i. We define the interpretation Γ for predicate symbols directly through the truth relation |=σ . Given a state σ, the relation |=σ is defined on atoms as follows. – H |=σ var (t) iff σ(t) ∈ V; T n – H |=σ share(t1 , . . . , tn ) iff i=1 Vars(σ(ti )) 6= ∅; – H |=σ pT (t) iff σ(t) ∈ [T ]G , with T ∈ T/{any}.

A First-Order Language for Expressing Aliasing and Type Properties

191

The semantics of the other formulas of L can be derived as usual. We will often write |= ϕ (resp. |=σ ϕ) for H |= ϕ (resp. H |=σ ϕ). Example 4. Let us see examples of the expressive power of LG . – The formula ∀V var (V ) ⇒ ¬share(V, X) asserts the groundness of X; – the formula list(X)∧∃V var (V )∧share(V, X)∧(∀W var (W )∧share(W, X) ⇒ share(V, W )) says that X is a list in which exactly one variable occurs; – ∀V var (V ) ∧ share(V, Y ) ⇒ share(V, X) says that each variable in Y is also in X; – (∀V var (V ) ∧ share(V, Y ) ⇒ share(V, X)) ∧ ground (X) ⇒ ground (Y ) asserts that if ∀V var (V ) ∧ share(V, Y ) ⇒ share(V, X) (i.e. Vars(Y ) ⊆ Vars(X)) and ground (X) (i.e. Vars(X) = ∅) then ground (Y ) (i.e. Vars(Y ) = ∅). Notice that properties expressible in the language LG are invariant with respect to the name of variables. That is, intuitively if property ϕ is true about t ∈ Terms(Σ, V) and s is a variant of t then ϕ is true of s, too. More formally the following lemma can be shown. Lemma 5. For each ϕ, formula of L, if |=σ ϕ and σ 0 is a variant of σ then |=σ0 ϕ. An important class of formulas of LG , which are often considered in analysis and verification, is the class of monotone formulas, that is the formulas ϕ such that |=σ ϕ and σ ≤ σ 0 implies |=σ0 ϕ. For example, ground (X) and ¬var (X) are monotone properties, while var (X) and share(X, Y ) are not. Since the grammar G is closed, by lemma (2), each atom pT1 (X) is monotone. An interesting subclass of monotone properties are the dependences like ∀V var (V ) ∧ share(V, Y ) ⇒ share(V, X). This kind of formulas are used in logic programming analysis since they allow to relate the values to which two or more arguments of a predicate can be instantiated. As shown by the example, the formula ∀V var (V ) ∧ share(V, Y ) ⇒ share(V, X), could be read informally as saying that if X is instantiated to a ground value then also Y is. The class of monotone properties is closed with respect to the connectives ∧,∨ and the quantifiers. We think it would be very interesting to give a complete syntactical characterization of these properties inside L.

5

A proof procedure for L

We are interested in characterizing the set of formulas Th L (H), that is the formulas of the language L which are true in the interpretation H. It is known that the existential fragment of the language L without the type predicates is decidable. In fact in [18], it is proposed a proof procedure to decide the validity of formulas ∃(ϕ1 ∧ · · · ∧ ϕn ) where each atom ϕi is an atom var (t), ground (t), share(t1 , . . . , tn ), or its negation. It is also known that the implication between regular types is decidable. In fact in [3, 2] a procedure is proposed to decide the validity of implications like ∀(pT1 (t1 )∧· · ·∧pTn (tn ) ⇒ pT (t)). It is not

192

P. Volpe

clear whether putting them together and considering the full first order theory, such as in L, the language is still decidable. In this section we will show that this is indeed the case and it is not such a trivial extension of those previous results. To show that Th L (H) is recursive we will use the method of elimination of quantifiers [5, 15]. We single out a set Ω, the elimination set, of fomulas of L and show that each formula of L is equivalent in H to a boolean combination of formulas of Ω. Once proven the decidability of formulas in Ω, we end up with a complete decision procedure for formulas of L. In the following, we use the abbreviation ∃≥k var (V ) φ to say that there exist at least k distinct variables which verify formula φ. It is defined by induction on k. ∃≥1 var (V ) φ is ∃V var (V ) ∧ φ ∃≥k+1 var (V ) φ is ∃V var (V ) ∧ φ ∧ (∃≥k var (W ) φ [V \W ] ∧ ¬share(V, W )), where the formula φ [V \W ] is obtained by φ by replacing all occurrence of V with W . To carry on the elimination of quantifiers we will need a particular class of formulas, which we call cardinality constraints. Definition 5. Let B be the class of boolean terms on V, that is the terms built from the signature ({∩(2) , ∪(2) , ¬(1) , 0(0) , 1(0) }, V). Fixed a natural k and a boolean term t, a simple cardinality constraint αk (t) is defined as the formula ∃≥k var (V ) Ψt (V ), where Ψt (V ) is defined inductively on the sintax of t. – – – – –

Ψ0 (V ) = false and Ψ1 (V ) = true; ΨX (V ) = share(V, X), with X ∈ V; Ψt1 ∩t2 (V ) = Ψt1 (V ) ∧ Ψt2 (V ); Ψt1 ∪t2 (V ) = Ψt1 (V ) ∨ Ψt2 (V ); Ψ¬t (V ) = ¬Ψt (V ).

A simple cardinality constraint αk (t) asserts the membership of at least k elements to the combination of variables in t, seen as subsets of V. Often the term ¬t will be written as t. Example 5. Consider the simple cardinality constraint α3 (X ∩ Y ∩ Z), that is, the formula ∃≥3 var (V ) share(V, X) ∧ share(V, Y ) ∧ ¬share(V, Z). This formula is true in H under the state σ, if there exist at least three variables sharing with σ(X) and σ(Y ) and not with σ(Z), that is if the cardinality of Vars(σ(X)) ∩ Vars(σ(Y )) ∩ Vars(σ(Z)), is at least equal to 3. The following lemma allows us to work with simple cardinality constraints just as if they were assertions on set of variables. Lemma 6. Let σ : V → Terms(Σ, V). Let t∗σ be obtained by the boolean term t by replacing each occurence of a variable X with Vars(σ(X)), for each X ∈ V. Then |=σ αk (t) if and only if the cardinality of t∗σ is at least equal to k. Proof. It can be straightforwardly proved by induction on t and k, noting that, if W ∈ V, then |=σ[V \W ] Ψt (V ) if and only if W ∈ t∗σ

A First-Order Language for Expressing Aliasing and Type Properties

193

Lemmas (7) and (8) follow. Lemma 7. Let σ, σ 0 : V → Terms(Σ, V) two states such that, for each X ∈ V, Vars(σ(X)) = Vars(σ 0 (X)). Then |=σ αk (t) if and only if |=σ0 αk (t). Lemma 8. Let t1 and t2 be two terms equivalent as boolean terms. Then |= αk (t1 ) ⇔ αk (t2 ). We will use the formula α=k (t), that is t contains exactly k elements, as an abbreviation for the formula αk (t) ∧ ¬αk+1 (t). Let the elimination set Ω be composed by the atomic formulas var (X), pT1 (X), . . . , pTn (X), where X ∈ V, and by the set of simple cardinality costraints {αk (t) | k ≥ 1, t is a boolean term}. The idea is to exploit the striking similarity of the simple cardinality constraints in our language with formulas of the first-order theory of the powerset of V seen as a Boolean algebra. For such a theory the decidability has been shown by Skolem in 1919 just through an argument based on elimination of quantifier (see [15] for a slightly more general account). The main idea is to reduce satisfiability of a formula in L to satisfiability of a conjunction of cardinality constraints. Definition 6. A conjunction of simple cardinality constraints αk (t), possibly negated, is a cardinality constraint. The elimination of quantifiers can be carried on for cardinality constraints in a long but straightforward way. The proof is adapted from the proof of elimination of quantifiers of the theory of Boolean algebras in [15]. Theorem 4. Let ψ(X) be a cardinality constraint. Then ∃X ψ(X) is equivalent to a disjunction of cardinality constraints. Anyway, in general, in a formula of L, other formulas than cardinality constraints may occur. We will show then some results that allow to eliminate formulas different from cardinality constraints under existential quantifiers. Definition 7. A formula is flat if it does not contain any functor. Notice that each formula αk (t) is flat. Indeed we can consider only flat formulas as shown by next lemma. Lemma 9. Every formula ϕ of L is equivalent in H to a flat formula. Proof. The following equivalences in H, already appeared in [18] and [2], can be easily checked. – var (f (X)) ⇔ false for each functor f ∈ Σ; Wk – share(t1 , . . . , f (s1 , . . . , sk ), . . . , tn ) ⇔ i=1 share(t1 , . . . , si , . . . , tn ); – pT (f (s1 , . . . , sn )) ⇔ pT1 (s1 )∧· · ·∧pTn (sn ) for each f (T1 , . . . , Tk ) ∈ Def G (T ) (remember that G is discriminative). – pT (f (s1 , . . . , sn )) ⇔ false if f (T1 , . . . , Tk ) 6∈Def G (T ) for any T1 , . . . , Tk .

194

P. Volpe

Definition 8. A type formula is a conjunction of atomic flat formulas pT (X), X ∈ V, possibly negated. Type formulas can be eliminated under existential quantifier and substituted by cardinality constraints. Lemma 10. Let ψ(X) be a cardinality constraint and φ(X) = pT1 (X) ∧ · · · ∧ pTn (X) ∧ ¬pS1 (X) ∧ · · · ∧ ¬pSk (X) a type formula. Let S ∪ [k, ω] be the varcardinality of T1 ∩ · · · ∩Tn \S1 \ · · · \ Sk . Then the following are valid equivalences in H. – (∃X φ(X) ∧ ψ(X)) ⇔ false W if S is empty and k = ω; – (∃X φ(X) ∧ ψ(X)) ⇔ ( h∈S ∃X α=h (X) ∧ ψ(X)) ∨ (∃X αk (X) ∧ ψ(X)). Proof. It is a consequence of lemmas (3) and (7) and of [T1 ] ∩ · · · ∩ [Tn ] ∩ [S1 ]∩ · · · ∩[Sk ] = [T1 ∩ · · · ∩Tn \S1 \ · · · \ Sk ]. In the same way the formulas var (X) and ¬var (X) can be eliminated. Lemma 11. Let ψ(X) be a cardinality constraint. Then (∃X ¬var (X) ∧ ψ(X)) ⇔ (∃X ψ(X)) and (∃X var (X) ∧ ψ(X)) ⇔ (∃X α=1 (X) ∧ ψ(X)). We can prove then the following theorem, which is at the base of the procedure of elimination of quantifiers. Theorem 5. For each formula Ψ (X), conjunction of formulas of Ω, possibly negated, there exists Φ, a boolean composition of formulas of Ω, such that |= (∃X Ψ (X)) ⇔ Φ. Proof. Let Ψ (X) be a conjunction of formulas of Ω, possibly negated. We can write Ψ (X) = Ψ1 (X) ∧ Ψ2 (X), with Ψ1 (X) the conjunction of all simple cardinality constraints in Ψ (X) and Ψ2 (X) the conjunction of the remaining formulas. We can assume that var (X) or ¬var (X) appears in Ψ2 (X). In the first case we can assume that no other formula appears, that is Ψ2 (X) = var (X). In fact if pT (X) appears in Ψ2 (X), we have that Ψ2 (X) ⇔ false. Moreover we have the equivalence var (X) ∧ ¬pT (X) ⇔ var (X). We have then Ψ (X) = Ψ1 (X) ∧ var (X), with Ψ1 (X) a cardinality constraint. By lemma (11) we have that ∃X Ψ1 (X) ∧ var (X) ⇔ ∃X Ψ1 (X) ∧ α=1 (X). In case ¬var (X) appears in Ψ2 (X) and no other formula occurs in Ψ2 (X), we have Ψ (X) = Ψ1 (X) ∧ ¬var (X) and (∃X Ψ1 (X) ∧ ¬var (X)) ⇔ ∃X Ψ1 (X). Otherwise we may assume that Ψ2 (X) has the general form ¬var (X) ∧ pT1 (X) ∧ · · · ∧ pTn (X) ∧ ¬pS1 (X) ∧ · · · ∧ ¬pSk (X), where n + k ≥ 1, which is equivalent to pinst (X) ∧ pT1 (X) ∧ · · · ∧ pTn (X) ∧ ¬pS1 (X) ∧ · · · ∧ ¬pSk (X). We have then Ψ (X) = Ψ1 (X) ∧ (pinst (X) ∧ pT1 (X) ∧ · · · ∧ pTn (X) ∧ ¬pS1 (X) ∧ · · · ∧ ¬pSk (X)) with Ψ1 (X) a cardinality constraint. By lemma (10) we know that ∃X Ψ (X), is equivalent to false or to a disjunction of formulas ∃X ∆i (X), where each ∆i (X) is a cardinality constraint. In any case, we have that ∃X Ψ (X) is equivalent to a disjunction of formulas

A First-Order Language for Expressing Aliasing and Type Properties

195

∃X Φ(X), with Φ(X) a cardinality constraint. We can apply then lemma (4) to each disjunct and obtain a boolean formula, which is a combination of formulas of Ω equivalent to ∃X Ψ (X). Notice that the procedure given in the proof of the previous theorem is effective. At this point it is easy to show that for each first-order formula of L it can be computed a Boolean composition of formulas of Ω equivalent to it on H. Theorem 6. Every formula Ψ of L is equivalent in H to a Boolean composition of formulas of Ω. Proof. Notice that since share(X1 , . . . , Xn ) and α1 (X1 ∩ · · · ∩Xn ) are equivalent in H, we can assume that no formula with predicate share occurs in Ψ . The proof is by structural induction on Ψ . Every atom of L is equivalent to a combination of atomic formulas of Ω (see proof of lemma (9)). If Ψ is a Boolean composition of formulas, then, by induction, each one of these is equivalent to a Boolean combination of formulas of Ω, and thus Ψ is equivalent to a Boolean combination of formulas of Ω. If Ψ is equal to ∃X Λ, then by induction Λ is equivalent to Λ∗ , boolean combination of formulas of Ω. We put Λ∗ in disjunctive normal form. The existential quantifier distributes over the disjunction and Ψ is equivalent to a disjunction of formulas ∃X Ψi with each Ψi conjunction of atoms, possibly negated, of Ω. By the previous theorem, each one of these is equivalent to a boolean formula of Ω. We have proven that each first-order formula of L is equivalent to a boolean combination of formulas of Ω. With the next theorem we state explicitly the decidability of truth of formulas in Ω. Theorem V 7. There exists an algorithm which decides the satisfiability of formulas ϕi , with each ϕi a formula of Ω, possibly negated. V V Proof. The formula is satisfiable in H if ∃ ϕi is true on H, where ∃ ϕi is V the V existantial closure of ϕi . We can apply the elimination of quantifiers to ∃ ϕi . Since the resulting formula has no variable, it must be equivalent to true or false. V Alternatively, an algorithm can be provided that, in case ∃ ϕi is satisfiable, V find V a substitution for the variables of ϕi . It can be obtained by first reducing ∃ ϕi to a disjunction of cardinality constraints by lemmas (10) and (11). An algorithm proposed in [4] to decide validity of a subclass of the language of set theory, can be instantiated to our case to solve cardinality constraints. Once V found such solutions, we have, for each variable X Vof ϕi , the set of variables SX of the term to which X must be mapped for ϕi to be satisfied. At this point a term can be builtVfor each such X, containing the variables in SX and verifying the formulas in ϕi , that is verifying the assertion var (X), ¬var (X) or pT1 (X) ∧ · · · ∧ pTn (X) ∧ ¬pS1 (X) ∧ · · · ∧ ¬pSk (X). In the first two cases the procedure is obvious. In the third case we use the rules that define regular type T1 ∩ · · · ∩Tn \S1 \ · · · \ Sk .

196

P. Volpe

Now combination V Wfor Ψ a general WV W V of formulas of Ω, we have that ∀Ψ ⇔ ∀ ϕi ⇔ ¬∃ ¬ϕi ⇔ ¬ ∃( ¬ϕi ), with each ϕi a formula of VΩ, possibly negated. Then ∀Ψ is true if and only if none of the conjunction ¬ϕ V i is satisfiable. Notice that in case Ψ is not true that means that a conjunction ¬ϕi is satisfiable and, by the previous theorem, a counterexample can be provided.

6

Applications

The logic we have studied is expressive enough to capture many interesting properties of arguments of predicates of logic programs. This suggests to employ L as an assertional language to be used in inductive proof methods. We will show how the verification conditions of many such methods can be entirely expressed by a formula of L. Morover since L is equipped with an algorithmic proof procedure, such verification conditions can effectively be decided. Throughout the section we assume hΣ, Λ, Vi as the signature for logic programs, with Λ the set of predicate symbols, distinct from the predicates of L. An assertion Θ of L is said a specification for predicate p(n) ∈ Λ, if Vars(Θ) ⊆ {X1 , . . . , Xn }. Informally variable Xi refers to the i-th argument of p. An atom p(t1 , . . . , tn ) will be said to satisfy the assertion Θ, written p(t1 , . . . , tn ) |= Θ, iff H |=σ[X1 ,... ,Xn \t1 ,... ,tn ] Θ. The notation Θ [X\t] denotes the formula Θ in which the variables (X1 , . . . , Xn ) are simultaneously substituted by terms (t1 , . . . , tn ). Let us consider first the correct answers of a program. We associate a specification Θp to each predicate p ∈ Λ. Program P is success-correct with respect θ

to {Θp }p∈Pred iff ∀p(t) ∈ Atoms p(t) 2 implies p(t)θ |= Θp . A sufficient condition for correctness can be stated as follows. A program P is success-correct with respect to {Θp }p∈Pred if for each clause p(t) ← p1 (t1 ), . . . , pn (tn ) it is true that H |=

n ^

Θpi [Xi \ti ] ⇒ Θp [X\t] .

(1)

i=1

The method indeed has been proposed in [6] and [12]. In our case, condition (1) can be decided using the procedure of Section (5). If the formula (1) is proved to be true for each clause, then the program is partial correct. Obviously if the formula is false this does not imply that the clause is necessarily wrong. Anyway it could be considered a warning that something wrong can happen. To this aim the alternative algorithm proposed in the proof of Theorem (7), could be very useful, since it would allow to provide counterexamples in such cases. The user then would have more information to decide whether the warning can give raise to a real error or simply the specification is too loose and behaviours are considered that can never occur in practice. If we are interested to check the Input/Output behaviour of logic programs, we can proceed as follows. To each predicate p ∈ Λ, is associated a property

A First-Order Language for Expressing Aliasing and Type Properties

197

pre p _ post p , where pre p and post p are specifications for p. Now a program P is I/O-correct with respect to the properties {pre p _ post p }p∈Pred if p(t) |= pre p and p(t)

θ

2 implies p(t)θ |= post p .

Now in case each formula pre p is monotone, a sufficient condition for P to be I/O-correct with respect to the properties {pre p _ post p }p∈Pred is that H |= (

n ^ i=1

(pre pi [Xi \ti ] ⇒ post pi [Xi \ti ]) ∧ pre p [X\t]) ⇒ post p [X\t] .

This can be shown to correspond to a particular case of the previous method [16]. Anyway, we still have a formula of L that can be decided algorithmically. Finally, if we want to check the call correctness of predicates we can consider methods like those proposed in [13][1]. Like in the previous case, to each predicate p ∈ Λ, is associated a property pre p _ post p , with pre p and post p specifications for p. In this case, anyway, the pre-condition is used also as a specification for the argument of a predicate at call-time. In fact a program P is call-correct with respect to the properties {pre p _ post p }p∈Pred if p(t) |= pre p and p(t)

θ

2 implies p(t)θ |= post p

and ∗

p(t) |= pre p and p(t) → hq(s), Gi implies q(s) |= pre q . We are assuming a leftmost selection rule for SLD-derivations. In [1] it has been shown that, in the case pre p and post p are monotone for each p, then a sufficient condition for P to be call-correct with respect to {pre p _ post p }p∈Pred is that for each clause p(t) ← p1 (t1 ), . . . , pn (tn ), it is true that for each 1 ≤ k ≤ n + 1 H |= (pre p [X\t] ∧

k−1 ^ i=1

post pi [Xi \ti ] ⇒ pre pk [Xk \tk ])

where pre pn+1 [Xn+1 \tn+1 ] ≡ post p [X\t]. In this case we have a finer control on the possible run-time behaviour. In fact a warning in this case can be raised because there may be a computation that calls a predicate with arguments which violate the specification. Again if a counterexample is provided the user may decide whether the specification is too loose or an actual error has been discovered. We want to stress the restriction in previous methods to monotone properties. The reason for not considering all the expressible properties is that in those cases the verification condititions become much more complex and the mgu’s have to be considered explicitly (see [16]). While for monotone assertions, the verification conditions are expressible as formulas of L, this is no more the case if more general properties are considered. A solution we are working on, is to enrich L so as to express formally the mgu’s. Anyway the class of properties that can be mechanically checked is still quite large, including type assertions, groundness, dependencies.

198

7

P. Volpe

Conclusions

In this paper we have studied a language that allows to express and decide properties of finite terms on a given signature, that is the data on which logic programs operate. The language is able to express aliasing properties such as the sharing or freeness and is enriched with type assertions. Using formulas of L, we are able to express many properties of program and prove them in inductive proof methods. Moreover since the logic is decidable we can mechanically check the corresponding verification condition and raise a warning if the condition is not verified. In such cases counterexamples can be built, which can be helpful for the user who is carrying on the verification. The set of true formulas is proved to be decidable through a procedure of elimination of quantifiers. This points out an interesting class of formulas, which express cardinality constraints on the set of variables that can occur in a term. This can give an interesting insight on domains used for the analysis of logic program. In fact many of the domains used for aliasing analysis can be seen as fragments of the domain of formulas of cardinality constraints. For example the element (X ∧ Y ) → (Z ∨ W ) of POS [9], can be represented as (Vars(Z) ⊆ Vars(X) ∪ Vars(Y )) ∨ (Vars(W ) ⊆ Vars(X) ∪ Vars(Y )) , which is equivalent to (|Vars(Z) ∩ Vars(X) ∩ Vars(Y )| = 0) ∨ (|Vars(W ) ∩ (Vars(X) ∩ Vars(Y ))| = 0), that is the constraint α=0 (Z ∩ X ∩ Y ) ∨ α=0 (W ∩ X ∩ Y ) of L. Another example is Sharing [14]. In fact the element {∅, {X}, {Y, Z}} can be represented as α=0 (X ∩ Y ) ∧ α=0 (X ∩ Z) ∧ α=0 (Y ∩ Z) ∧ α=0 (Z ∩ Y ). We recall that cardinality constraints are equipped with an operation of conjunction and cylindrification (existential quantifier). We are thinking about ways of augmenting the expressive power of the logic L, obviously while retaining the decidability. We are currently investigating two possibilities. The first consists in adding a modal operator 2 defined as follows |=σ 2ϕ

iff for each σ 0 ≥ σ |=σ0 ϕ.

Such modality would allow, first of all, to characterize monotone properties of language L. In fact monotone properties would correspond to the formulas Ψ such that Ψ ⇔ 2Ψ . Morover we could express arbitrary dependences between properties, like 2(list(X) ⇒ list(Y )), whose informal meaning is that every state that instantiate X to a list will also bind Y to a list. Another extension is to consider Hoare-like triples {Φ}[[t, s]]{Ψ }, whose meaning is: if Φ is true under the state σ and θ = mgu(σ(t), σ(s)), then Ψ is true under the state σ◦θ. They have been considered in [7] and [8], where a formal calculus has been provided for a language of assertions different from L. These formulas would allow to express formally the verification of general proof methods, like those proposed in [13] and in [16], for the whole class of formulas of L. At the moment, it is still not known if, given a decidable logic such as L, it is possible to decide the validity of such triples.

A First-Order Language for Expressing Aliasing and Type Properties

199

References 1. A. Bossi and N. Cocco. Verifying Correctness of Logic Programs. In J. Diaz and F. Orejas, editors, Proc. TAPSOFT’89, pages 96–110, 1989. 2. J. Boye. Directional Types in Logic Programming. PhD thesis, University of Link¨ oping, Computer Science Department, 1997. 3. J. Boye and J. Maluszynski. Directional Types and the Annotation Method. Journal of Logic Programming, 33(3):179–220, 1997. 4. D. Cantone, E. G. Omodeo, and A. Policriti. The Automation of Syllogistic II. Optimization and Complexity Issues. Journal of Automated Reasoning, 6(2):173– 187, 1990. 5. C. C. Chang and H. J. Kreisler. Model Theory. Elsevier Science Publ., 1990. Third edition. 6. K. L. Clark. Predicate logic as a computational formalism. Res. Report DOC 79/59, Imperial College, Dept. of Computing, London, 1979. 7. L. Colussi and E. Marchiori. Proving Correctness of Logic Programs Using Axiomatic Semantics. In Proc. of the Eight International Conference on Logic Programming, pages 629–644. The MIT Press, Cambridge, Mass., 1991. 8. L. Colussi and E. Marchiori. Unification as Predicate Transformer. In Proc. of the Joint International Conference and Symposium on Logic Programming, pages 67–85. The MIT Press, Cambridge, Mass., 1992. 9. A. Cortesi, G. Fil`e, and W. Winsborough. Prop revisited: Propositional Formula as Abstract Domain for Groundness Analysis. In Proc. Sixth IEEE Symp. on Logic In Computer Science, pages 322–327. IEEE Computer Society Press, 1991. 10. P. Dart and J. Zobel. Efficient run-time type checking of typed logic program. Journal of Logic Programming, 14(1-2):31–70, 1992. 11. P. Dart and J. Zobel. A regular type language for logic programs. In F. Pfenning, editor, Types in logic programming, pages 157–187. The MIT Press, Cambridge, Mass., 1992. 12. P. Deransart. Proof Methods of Declarative Properties of Definite Programs. Theoretical Computer Science, 118(2):99–166, 1993. 13. W. Drabent and J. Maluszynski. Inductive Assertion Method for Logic Programs. Theoretical Computer Science, 59(1):133–155, 1988. 14. D. Jacobs and A. Langen. Accurate and Efficient Approximation of Variable Aliasing in Logic Programs. In E. Lusk and R. Overbeek, editors, Proc. North American Conf. on Logic Programming’89, pages 154–165. The MIT Press, Cambridge, Mass., 1989. 15. G. Kreisel and J. L. Krivine. Elements of Mathematical Logic (Model Theory). North-Holland, Amsterdam, 1967. 16. G. Levi and P. Volpe. Derivation of Proof Methods by Abstract Interpretation. (Submitted). Available at http://www.di.unipi.it/∼volpep/papers.html, 1998. 17. J. W. Lloyd. Foundations of Logic Programming. Springer-Verlag, Berlin, 1987. Second edition. 18. E. Marchiori. A Logic for Variable Aliasing in Logic Programs. In G. Levi and M. Rodriguez-Artalejo, editors, Proceedings of the 4th International Conference on Algebraic and Logic Programming (ALP’94), number 850 in LNCS, pages 287–304. Springer Verlag, 1994. 19. E. Marchiori. Design of Abstract Domains using First-order Logic. In M. Hanus and M. Rodriguez-Artalejo, editors, Proceedings of the 5th International Conference on Algebraic and Logic Programming (ALP’96), number 1139 in LNCS, pages 209– 223. Springer Verlag, 1996.

Refining Static Analyses by Trace-Based Partitioning Using Control Flow Maria Handjieva and Stanislav Tzolovski LIX, Ecole Polytechnique, France {handjiev,stivy}@lix.polytechnique.fr

Abstract. This paper presents a systematic method of building a more precise static analysis from a given one. The key idea is to lift an abstract domain to the finite sets of its labeled abstract properties. The labels are designed to gather information about the history of control flow and to obtain a finite partitioning of the program execution traces. The abstract operations of the lifted domain are derived from those of the original one. This is a particular instance of the reduced cardinal power introduced by P. and R. Cousot, where the base is the set of labels approximating the control history and the exponent is an abstract domain. The method is applied to the domain of convex polyhedra and to the domain of linear congruences. Key words: abstract interpretation, reduced cardinal power, trace semantics.

1

Introduction

An essential part of static analysis by abstract interpretation [4,5] is to build a machine-representable abstract domain expressing interesting properties about a computer program. Many abstract domains have been developed for many purposes [3,7,9,11]. The most significant analyses dealing with numerical variables are constant propagation, analysis using intervals, linear equalities, linear inequalities, linear congruences. We are interested in refining static analyses based on a given domain. For non-distributive abstract interpretations (such as convex polyhedra), the disjunctive completion [6] is an operator that systematically produces new and more precise abstract domains from simpler ones. This was illustrated in [8] for the domain PROP for ground-dependence analysis of logic programs. We propose a method for lifting an abstract domain to the finite sets of its labeled abstract properties. The labels are designed to gather information about the history of control flow and to obtain a finite partitioning of the execution traces. Each partition (the abstract properties with the same label) is approximated with an upper bound of its elements. The abstract operations of the new domain are derived from those of the original one. We apply this method to the domain of convex polyhedra and to the domain of linear congruences. G. Levi (Ed.): SAS’98, LNCS 1503, pp. 200–214, 1998. c Springer-Verlag Berlin Heidelberg 1998

Refining Static Analyses by Trace-Based Partitioning Using Control Flow

201

A Motivated Example. A static analysis of program variables using convex polyhedra is proposed in [7]. We slightly change the illustrated example from [7] as follows: {U0 } {U1 } {U2 } {U3 } {U4 } {U5 } {U6 } {U7 } {U8 } {U9 } {U10 } {U11 } {U12 } {U13 }

[c1 ] i = 2; [c2 ] j = 0; [c3 ] if (Cond1 ) then [c4 ] k = 1; else [c5 ] k = −1; endif [c6 ] while (Cond2 ) do [c7 ] if (Cond3 ) then [c8 ] i = i + 4 ∗ k; else [c9 ] i = i + 2 ∗ k; [c10 ] j = j + k; endif enddo; [c11 ] exit;

For each statement Stat of the program we have two abstract properties: Ui and Uo called input (before execution of Stat) and output (after execution of Stat) abstract properties respectively. We denote the label of statement i as [ci ] and write the label before the statement. If we apply the analysis based on convex polyhedra to the above program we will obtain the convex polyhedron U13 = (−1 ≤ k ≤ 1). The result is too abstract, because the analysis cannot find that the value of variable k in the loop is either 1 or −1. The intuition is to represent the restraints between variables with two convex polyhedra (one for k = −1 and another for k = 1). In this case the result is U13 = {(i − 2j ≤ 2, j ≤ 0, k = −1), (i − 2j ≥ 2, j ≥ 0, k = 1)}, which is obviously more precise than the previously obtained one. Observe in this example, that this is the reduced cardinal power [5] with boolean base and polyhedral exponent. Therefore we need an abstract domain, which treats finite sets of convex polyhedra. Because the domain is not finite and does not satisfy the ascending chain condition we need also a technique that guarantees the termination of the analysis. Widening operators are an appropriate technique for this purpose. Moreover, we can use the same idea for other abstract domain, hence a method that systematically lifts a given abstract domain to the set of abstract properties is needed. The rest of this paper is structured as follows. In the next section we introduce the set of labeled abstract properties corresponding to a a given abstract domain. Starting from a concrete trace semantics and passing through a collecting semantics we design an abstract semantics using a concretization function and a widening. In Sects. 3 and 4 we apply our method to the convex polyhedra domain and to the domain of linear congruences respectively. We conclude the paper in Sect. 5.

202

2

M. Handjieva and S. Tzolovski

Trace-Based Partitioning Using Control Flow

In order to simplify the presentation all formal reasoning is done using a simple programming language with assignments, “if-then-else” constructs and “while” loops. Our framework takes two parameters. The first one is an abstract domain: (D, vD , uD , tD , ⊥D , >D , αD , γD , ∇D , assign D , True D , False D ) where D is the set of abstract properties, vD is the abstract ordering, uD and tD are the upper and lower bound respectively, ⊥D and >D are the minimal and maximal element respectively, αD is the abstraction function (if any) and γD is the concretization function, ∇D is a widening operator (if any). We add three specific functions: assign D ∈ (D × ASSIGN) → D that treat the assignments (from ASSIGN), True D ∈ (D × COND) → D and False D ∈ (D × COND) → D that treat the conditions in the program. For example, in the domain of closed intervals (denoted by I) if the value of variable x is approximated with the interval [1, 5] then: assign I ([1, 5], x = x + 1) = [2, 6] False I ([1, 5], x ≥ 3) = [1, 3] True I ([1, 5], x ≥ 3) = [3, 5], An “if-then-else” or a “while” statement is called a test node. Each test node has a unique number. Let T be the set of test node numbers and t be the number of test nodes (t = |T |). Each test node can be analyzed in different ways. The analysis of an “if” statement could return a set of one abstract property, which contains information about both the “true” and “false” branches or it could return a set of two abstract properties one for each branch. In the case of a “while” statement the analysis could obtain a set of k + 1 elements: one when the control does not pass through the loop; one - when the control passes exactly once through the loop, etc.; one - when the control passes k or more times through the loop. So, the second parameter of our framework is a tuple of t elements, denoted by ∆ and called test approximation parameter, that contains information on how to approximate each test node. 2.1

Concrete Semantics

The concrete semantics of a program P at point z is the set of all execution traces from the entry point to point z. Sz [[P ]] = {[c1 , M1 ] → [c2 , M2 ] → . . . → [cz , Mz ], . . . } where ci is the label of statement i and Mi is the memory state (a vector of values) before the execution of statement i. 2.2

Collecting Semantics

The starting point of an abstract interpretation is a collecting semantics for the programming language. Our collecting semantics is a set of configurations. Each

Refining Static Analyses by Trace-Based Partitioning Using Control Flow

203

configuration is a pair of a string of labels and a vector of values. In what follows configurations are always surrounded with left and right angles. For each trace in the concrete semantics Sz [[P ]] there is a configuration in the collecting semantics that characterizes the control history of this trace and its current memory state. Cz [[P ]] = {hc1 c2 . . . cz , Mz i , hc01 c02 . . . c0z , Mz0 i , . . . , hc001 . . . c00z , Mz00 i} The concretization γcs and abstraction αsc functions, that describe the link between concrete and collecting semantics, are defined recursively as follows: γcs (C) =

[

γt (hc1 · · · cz , Mz i)

hc1 ···cz ,Mz i∈C

γt (hc1 · · · cz , Mz i) = [c1 , M1 ] → γt (hc2 · · · cz , Mz i), γt (hcz , Mz i) = [cz , Mz ]

M1 is any memory state

To define the abstraction function we use the following notations: all letters in bold are strings, is the empty string and “·” is the string concatenation. αsc (T ) = {αt (, t) | t ∈ T } αt (c, [c, M ] → t) = αt (c · c, t) αt (c, [c, M ]) = hc · c, M i S S Note that the pair (αsc , γcs ) is a Galois connection, since αsc ( Ti ) = αsc (Ti ). i

2.3

i

Set of Labeled Abstract Property Domain

The second step is the abstraction of the collecting semantics. Our abstract domain is a finite set of labeled abstract properties from a given abstract domain. An abstract label is a string from the alphabet A = {b, tj , fj , vj }, for j ∈ T , such that it begins with b and there are no more b. The letter b describes the path from the entry node to the first test node. The letter tj (resp. fj ) describes the part of the “true” (resp. “false”) branch of test node j (from j to the next test node). The abstract letter vj represents: – all paths that pass a finite number of times through j, if j is a “while” loop; – the two paths that pass through j, if j is an “if-then-else” statement. The concrete control flow graph contains one node for each statement and a node for each junction point (“endif”). The abstract control flow graph contains: one node for each test point j, labeled with j, an entry, an exit node and one node for each junction point. The junction nodes are added for simplification. For example, the concrete control flow graph and the abstract control flow graph of the program from Sect. 1 are illustrated in Fig. 1.

204

M. Handjieva and S. Tzolovski

1

c4 2

c1 c2 c3

b t1

2

c7

f2

t2 c9

c8

f1

c5

c6 3

1

c11

t3

3

f3

c10

Fig. 1. Concrete and abstract control flow graphs

The meaning of the abstract letters (from alphabet A) are defined with the help of the concretization function γl . For the program from Sect. 1 it is: γl (b) = {c1 c2 c3 } γl (t1 ) = {c4 c6 } γl (t2 ) = {c7 }

γl (f1 ) = {c5 c6 } γl (f2 ) = {c11 }

γl (t3 ) = {c8 c6 }

γl (f2 ) = {c9 c10 c6 }

The intuition behind vj is expressed using function δ, recursively defined below. j vj ≡ δ P

tj

δ P

= (tj · δ(P ))∗ j

tj

δP

fj

= (tj · δ(P ))∗ · fj

tj

j

fj

R = tj · δ(P ) | fj · δ(R)

P δ = δ(P ) · δ(R) R

δ( | ) =

where P and R are part of the reducible abstract control flow graph with exactly one entry edge and exactly one exit edge. In this way for each abstract letter vj we obtain a regular expression that characterizes it. For example, v2 ≡ (t2 · (t3 | f3 ))∗ = {, t2 t3 , t2 f3 , t2 t3 t2 t3 , t2 t3 t2 f3 , . . . } The meaning of an abstract letter vj is formally given by function γl over regular expressions recursively defined as follows: γl (X · Y ) = {x · y | x ∈ γl (X), y ∈ γl (Y )} γl (X ∗ ) = {x · · x} | x ∈ γl (X), n ∈ IN} | ·{z n

γl (X | Y ) = {x | x ∈ γl (X) ∨ x ∈ γl (Y )}

Refining Static Analyses by Trace-Based Partitioning Using Control Flow

205

where X and Y are regular expressions. So γl (vj ) is γl (δ(Gj )), where Gj is the subgraph of the abstract control flow graph with nodes from the strongly connected components with head j. Function γL gives the meaning of an abstract label a (which is a string of abstract letters a = a1 . . . an ) : γL (a) = γL (a1 . . . an ) = {s1 . . . sn | si ∈ γl (ai ), i = 1, . . . , n} The abstract labels are ordered by: a a0 iff γL (a) ⊆ γL (a0 ) For example, bt1 t2 t3 bt1 v2 , but bt1 f2 6bt1 v2 . An abstract configuration is an element of the reduced cardinal power [5] with sets of abstract labels as atomic base and abstract domain D as exponent which represent as a pair of an abstract label and an abstract property, denoted as ha, Di, where D ∈ D. Our abstract domain consists of sets of abstract configurations with an additional condition: there are not two elements with comparable labels in each set. This additional condition guarantees correct trace partitioning. 2.4

Abstract Operations

The last step consists in computing the least fixpoint of the abstract semantics. In this subsection we define the operations on the abstract domain needed when computing this least fixpoint (described in the next subsection). Approximation ordering. We introduce the following ordering so that we can compare two sets of labeled abstract properties. U v U 0 iff ∀ ha, Di ∈ U, ∃ ha0 , D0 i ∈ U 0 , a a0 ∧ D vD D0 Upper bound. The intuition behind designing the upper bound operation is to obtain either a set union over the paths with non-comparable labels or merge over the paths with comparable labels. In this way we cannot merge paths from different partitioning. ha, Di ∈ U, ha0k , Dk0 i ∈ U 0 , 0 0 0 ha, D tD D1 tD · · · tD Dn i 0 U tU = ak a, k = 1, . . . , n ha0 , D0 i ∈ U 0 , hak , Dk i ∈ U, 0 0 ∪ ha , D tD D1 tD · · · tD Dm i ak a0 , k = 1, . . . , m ∪ {ha, Di | ha, Di ∈ U, 6 ∃ ha0 , D0 i ∈ U 0 , (a a0 ∨ a0 a)} ∪ {ha0 , D0 i | ha0 , D0 i ∈ U 0 , 6 ∃ ha, Di ∈ U, (a a0 ∨ a0 a)} Widening operator. A convergence technique has been proposed in [4] that uses so called widening operator to transform infinitely iterative computation into finite but approximate one. Let (L, v, ⊥, >, t) be a poset with maximal element >. The following definition is due to Cousot [2].

206

M. Handjieva and S. Tzolovski

Definition 1 A widening operator ∇ ∈ (IN → (L × L → L)) is such that1 : 1. ∀i > 0, ∀x, y ∈ L, (x t y) v (x∇i y); 2. For any ascending chain y 0 v y 1 v · · · v y n v · · · of elements of L, the ascending chain x0 = y 0 , x1 = x0 ∇1 y 1 , . . . , xn = xn−1 ∇n y n , . . . is eventually stable, i.e. there exists k ≥ 0 such that for i > k, xi = xk . The widening on our abstract domain is defined as follows: U ∇U 0 = {ha, D∇D (D tD D1 tD . . .tD Dn )i | ha, Di ∈ U, hai , Di i ∈ U 0 , ai a} ∪ {ha, Di | ha, Di ∈ U ∧ 6 ∃ ha0 , D0 i ∈ U 0 , a0 a} It is designed only for sets of labeled abstract properties U and U 0 such that: ∀ ha0 , D0 i ∈ U 0 , ∃ ha, Di ∈ U, a0 a This condition is not a restriction because, the widening operator ∇ is always applied between two successive iterations of a “while” loop and we analyze only structured programs. Proposition 1 The operator ∇ defined above is a widening. Proof: It is clear that U t U 0 v U ∇U 0 . On the other hand, the number of abstract properties in U ∇U 0 is equal to the number of abstract properties in U (|U | = |U ∇U 0 |). Moreover for each abstract label a we cannot have infinite many abstract properties D1 vD D2 vD . . . from successive iterations, because of the application of the widening ∇D . Therefore the operator ∇ cannot iterate infinitely without convergence. 2 2.5

Abstract Semantics

The abstract semantics of a program P at point z is a set of abstract configurations ha, Di, where a is an abstract label and D ∈ D is an abstract property. Az [[P ]] = {hai , Di i | k = 1, . . . , n} In our approach, the link between the abstract and collecting semantics is given by a monotone concretization function Γ defined below. [ Γ (U ) = γ(ha, Di) ha,Di∈U

γ(ha, Di) = {hs, M i | s ∈ γL (a), M ∈ γD (D)} Assignments. Consider the following assignment: {U } 1

x = Expr

We denote ∇n the operator ∇(n), n ∈ IN.

{U 0 }

Refining Static Analyses by Trace-Based Partitioning Using Control Flow

207

Its output set of labeled abstract properties is given by: U 0 = assign(U, x = Expr ) = {ha, D0 i | ha, Di ∈ U, D0 = assign D (D, x = Expr )} If-then-else constructs. We now give the abstract meaning of an “if-then-else” statement, where BlockThen and BlockElse are blocks of statements: {U } {U1 } {U2 }

if (Cond) then BlockThen else BlockElse endif

{U11 } {U22 } {U 0 }

There are two possible ways to approximate the information after “if-then-else” statement i. If ki = 1 (ki is the i-th element of the test approximation parameter) then we set lt = ti and lf = fi , otherwise (when ki = 0) we set lt = lf = vi . U1 = {ha · lt , D0 i | ha, Di ∈ U, D0 = True D (D, Cond), D0 6= ⊥D } U2 = {ha · lf , D0 i | ha, Di ∈ U, D0 = False D (D, Cond), D0 6= ⊥D } U 0 = U11 t U22 Note that when ki = 1 there is not a merge of the “true” and “false” paths, because the upper bound is simply the set union. Which means that there is no loss of information in this case. Contrarily to common choice of abstracting the merge of these two paths by an upper bound on the domain, which is an upper approximation for non-distributive abstract interpretation frameworks. For example, in the convex polyhedra domain, where the union of two polyhedra does not necessarily correspond to the convex polyhedron, it is approximated by the convex hull. While loops. Consider the following loop statement: {U } {U1 }

while (Cond) do Block enddo

{U2 } {U 0 }

In order to find the fixpoint of the abstract semantics we use additional sets of labeled abstract property V i which is always valid before the i-th iteration. when i = 0 U,i U2 , when 0 < i < kj i V = kj {ha · v , Di | ha, Di ∈ U }, when i = kj j 2 i−1 V ∇U2i when i > kj U1i = {ha · tj , D0 i | ha, Di ∈ V i−1 , D0 = True D (D, Cond), D0 6= ⊥D } [ U0 = {ha · fj , D0 i | ha, Di ∈ V i , D0 = False D (D, Cond), D0 6= ⊥D } i=0,kj −1

∪{ha · fj , D0 i | ha, Di ∈ V kj +y , D0 = False D (D, Cond), D0 6= ⊥D } where y is such that the local fixpoint (for the loop j) is reached, which means the smallest y such that V kj +y = V kj +y+1 .

208

M. Handjieva and S. Tzolovski

2.6

Complexity and Precision

The complexity of our analysis depends on its two parameters - the abstract domain and the test approximation parameter. For each program Q, having its test approximation parameter ∆Q we can calculate the maximal number of labeled abstract properties in the abstract semantics (which is in fact the number of partitions). Let σ be a function which for a given abstract control flow subgraph (with exactly one entry and exactly one exit edge) calculates this number according to ∆Q = (k1 , k2 , . . . , kt ) . We recursively define σ as follows: j σP σ P

tj

fj

tj

j

, u = u + σ(P, u) + · · · + σ(P, . . . , σ( P, u) . . . ) | {z }

fj R

, u =

kj

u, if kj = 0 σ(P, u) + σ(R, u), if kj = 1

δ ( | , u) = u

! σ

P R

,u

= σ(R, σ(P, u))

where j is a test node, P and R are part of the abstract control flow graph with exactly one entry edge and exactly one exit edge and u is a number of partitions at the entry edge. Therefore the maximal number of elements in the abstract semantics at the and of program Q is σ(Q, 1). Obviously the analysis based on the lifted abstract domain is more costly than the analysis based on the original one. For example, the maximal number of elements in the set abstract properties at Pkof2 labeled point 13 of the program from Sect. 1 is 2k1 i=0 2k3 i , where k1 , k3 ∈ {0, 1} and k2 ∈ IN. The precision of our analysis also depends on the test approximation parameter. We can argue that our analysis based on the lifted domain provides more accurate results than or the same results as the analysis based on abstract domain D. Proposition 2 The analysis based on finite sets of labeled abstract properties of D is more precise than the analysis based on D. In the worst case of our analysis, when ∆ = [0, . . . , 0] the result is a set of one abstract property, which is exactly the same as the result of the static analysis based on D. We can design a family of static analyses changing the test approximation parameter as making different compromises between complexity and precision.

3 3.1

Finite Sets of Labeled Convex Polyhedra Convex Polyhedra Domain

Static analysis of linear inequalities among variables of a program have been studied by P.Cousot and N.Halbwachs in [7]. To set up some notations we briefly

Refining Static Analyses by Trace-Based Partitioning Using Control Flow

209

present the definitions related to convex polyhedra [12]. The convex hull of a set X of vectors is the smallest convex set containing X and it is denoted by convexHull (X), so: convexHull (X) = {λ1 x1 + · · · + λk xk | k ≥ 1; x1 , . . . , xk ∈ X; λ1 , . . . , λk ∈ IR; λ1 , . . . , λk ≥ 0; λ1 + · · · + λk = 1} A set P of vectors in IRn is called a convex polyhedron if P = {x | Ax ≤ b} for some matrix A and vector b. In other words, the polyhedron is the intersection of finitely many affine half-spaces. An affine half-space is a set {x | ax ≤ b} for some nonzero row vector a and some number b. A set of vectors is a convex polytope if it is the convex hull of finitely many vectors. The convex hull of two convex polyhedra is the smallest convex set defined as follows: chull (P1 , P2 ) = convexHull ({X | X ∈ P1 ∨ X ∈ P2 }) Upper Bound. The upper bound of two convex polyhedra is their convex hull. So, we have: P1 tp P2 = chull (P1 , P2 ) and P1 tp · · · tp Pn = chull (P1 , . . . , Pn ). Affine Transformation. The affine transformation τp transforms a polyhedron P into another polyhedron P 0 according to a given affine map x → Ax + b. Polyhedron P 0 is specified as follows: P 0 = {x0 | x ∈ P, x0 = Ax + b} Widening Operator. The widening operator of convex polyhedra ∇p (defined in [7] and optimized in [10]) is based on the following heuristic: the widening of two polyhedra is a new polyhedron obtained by removing from the system of the first polyhedron all inequalities which are not satisfied by the second polyhedron. 3.2

Specific Functions

In the following, specific functions assign p , True p and False p that deal with convex polyhedra are presented. Assignments. Let x = Expr be an assignment. For a given polyhedron P , function assign p returns a polyhedron P 0 in the following manner: (a) If Expr is nonlinear expression we make a conservative approximation, assuming that any value can be assigned to x. Therefore, variable x is eliminated from P by projecting P along the x dimension. P n then function (b) If Expr is linear expression, i.e. Expr ≡ ( i=1 ai xi + b) P n the affine map x → i=1 ai xi + b. assign p is an affine transformation defined byP n In the case of invertible assignments xk = i=1 ai xi + b (when ak 6= 0) the system of restraints of each polyhedron can be computed directly [7]. In the case of non-invertible assignments (when ak = 0) we firstPproject polyhedron P along n the xk dimension and then add the restraint xk = i=1 ai xi + b. Conditions. According to condition Cond we have the following three cases: (a) condition Cond is nonlinear. We ignore the test by setting both functions True p and False p to behave as the identity function on convex polyhedra: True p (P, Cond) = False p (P, Cond) = P ; (b) condition Cond is a linear equality, i.e. Cond ≡ (aX = b). Let H be the hyperplane {X ∈ Rn | aX = b}. If polyhedron P is included in H then

210

M. Handjieva and S. Tzolovski

T ruep (P, Cond) = P and F alsep (P, Cond) = ⊥p . Otherwise T ruep (P, Cond) = P ∩ H and F alsep (P, Cond) = P ; (c) the condition Cond is a linear inequality, i.e. Cond ≡ (aX ≤ b). Let H1 be the closed half-space {X ∈ Rn | aX ≤ b} and let H2 be the closed halfspace {X ∈ Rn | aX ≥ b}. In this case we set True p (P, Cond) = P ∩ H1 and False p (P, Cond) = P ∩ H2 . 3.3

Example

We illustrate the application of the method described in Sect. 2 using convex polyhedra domain on the motivated example (see Sect. 1). The program is simple enough to allow hand computation and simple geometrical representation. All tests involve nonlinear conditions, which are not taken into account by the analysis. The test approximation parameter is set up to be ∆ = (1, 0, 1). Which means that the first test node (the “if” with label [c3 ]) will be treated exactly (because k1 = 1), the second test point (the “while” loop) will be approximated at the beginning (because k2 = 0) etc. Each set of labeled polyhedra Ui , (i = 0, . . . , 13) is initially empty (Ui0 = ∅). The superscript shows the number of steps the analysis passes through point i. At the entry node we set up U01 = IRn , where n is the number of variables involved in the analysis. U01 = IR3 U11 = assign(assign(U01 , i = 2), j = 0) = {hb, (i = 2, j = 0)i} U21 = {hbt1 , (i = 2, j = 0)i} U31 = assign(U21 , k = 1) = {hbt1 , (i = 2, j = 0, k = 1)i} U41 = {hbf1 , (i = 2, j = 0)i} U51 = assign(U41 , k = −1) = {hbf1 , (i = 2, j = 0, k = −1)i} U61 = U31 t U51 = {hbt1 , (i = 2, j = 0, k = 1)i , hbf1 , (i = 2, j = 0, k = −1)i} Now, we are going to analyze the “while” loop. V 0 = {hbt1 v2 , (i = 2, j = 0, k = 1)i , hbf1 v2 , (i = 2, j = 0, k = −1)i} U71 = {hbt1 v2 t2 , (i = 2, j = 0, k = 1)i , hbf1 v2 t2 , (i = 2, j = 0, k = −1)i} U81 = {hbt1 v2 t2 t3 , (i = 2, j = 0, k = 1)i , hbf1 v2 t2 t3 , (i = 2, j = 0, k = −1)i} U91 = assign(U81 , i = i + 4 ∗ k) = = {hbt1 v2 t2 t3 , (i = 6, j = 0, k = 1)i , hbf1 v2 t2 t3 , (i = −2, j = 0, k = −1)i} 1 U10 = {hbt1 v2 t2 f3 , (i = 2, j = 0, k = 1)i , hbf1 v2 t2 f3 , (i = 2, j = 0, k = −1)i} 1 1 U11 = assign(assign(U10 , i = i + 2 ∗ k), j = j + k) = = {hbt1 v2 t2 f3 , (i = 4, j = 1, k = 1)i , hbf1 v2 t2 f3 , (i = 0, j = −1, k = −1)i} 1 1 U12 = U91 t U11 = hbt1 v2 t2 t3 , (i = 6, j = 0, k = 1)i , hbf1 v2 t2 t3 , (i = −2, j = 0, k = −1)i = hbt1 v2 t2 f3 , (i = 4, j = 1, k = 1)i , hbf1 v2 t2 f3 , (i = 0, j = −1, k = −1)i

Refining Static Analyses by Trace-Based Partitioning Using Control Flow

211

If we unroll the loop once again we will obtain the following result: 1 V 1 = V 0 ∇U12 = hbt1 v2 , (i − 2j ≥ 2, i + 2j ≤ 6, j ≥ 0, k = 1)i = hbf1 v2 , (i − 2j ≤ 2, i + 2j ≥ −2, j ≤ 0, k = −1)i

The above widening behaves as an upper bound. It returns a set of two polyhedra. The first one is the convex hull of the polyhedra with labels bt1 v2 , bt1 v2 t2 t3 and bt1 v2 t2 f3 (because bt1 v2 t2 t3 bt1 v2 , bt1 v2 t2 f3 bt1 v2 ) and the second one is the convex hull of the polyhedra with labels bf1 v2 , bf1 v2 t2 t3 and bf1 v2 t2 f3 . hbt1 v2 t2 t3 , (i − 2j ≥ 6, i + 2j ≤ 10, j ≥ 0, k = 1)i hbf1 v2 t2 t3 , (i − 2j ≤ −2, i + 2j ≥ −6, j ≤ 0, k = −1)i 2 U12 = hbt1 v2 t2 f3 , (i − 2j ≥ 2, i + 2j ≤ 10, j ≥ 1, k = 1)i hbf1 v2 t2 f3 , (i − 2j ≤ 2, i + 2j ≥ −6, j ≤ −1, k = −1)i At this point the widening operator takes place: hbt1 v2 , (i − 2j ≥ 2, j ≥ 0, k = 1)i , 2 1 2 V = V ∇U12 = hbf1 v2 , (i − 2j ≤ 2, j ≤ 0, k = −1)i Finally, this is the fixpoint, therefore we find: hbt1 v2 f2 , (i − 2j ≥ 2, j ≥ 0, k = 1)i , 1 U13 = hbf1 v2 f2 , (i − 2j ≤ 2, j ≤ 0, k = −1)i 1 2 = V 2 , U12 ) are shown on Fig. 2. Some of these results (V 1 , U13

j Plane k=1

3

V

2

2

1 0 1

2

3

4

5

6

7

8

9

10

i

j -6

-5

-4

-3

-2

-1

1

2

i -1

V

1

2 U12

V

2

-2

Plane k=-1

Fig. 2. The set of polyhedra for the motivated example

212

4 4.1

M. Handjieva and S. Tzolovski

Finite Sets of Labeled Linear Congruences Linear Congruence Domain

The static analysis of linear congruence equalities among program variables was studied by P.Granger in [9]. We briefly recall some notions related to linear congruences. Let G be an abelian group and H be a subgroup of G. Any subset of G of the form: a + H = {x ∈ G | ∃h ∈ H, x = a + h} where a is an element of G is called coset of H in G. The element a is said to be a representative of the coset and H the unique modulo of the coset. For any element b of a + H, we have a + H = b + H. The set of all cosets in the abelian group G with the singleton {∅} is called the congruence lattice of G and is denoted as C(G) (or C when G is Z). The set inclusion order can be simplified by the following formula: (a1 + H1 ) ⊆ (a2 + H2 ) ≡ (a1 − a2 ∈ H2 ) ∧ (H1 ⊆ H2 ) Upper bound. The upper bound operations t in C(G) is (a1 + H1 ) t (a2 + H2 ) = a1 + Z(a1 − a2 ) + H1 + H2 where Za denotes the subgroup generated by a i.e. {. . . , −a − a, −a, 0, a, a + a, . . . }. In what follows we use Zn as a group G. Affine transformation. We use r(H) to denote the rank of H and k(u) the set {X ∈ Zn | u(X) = (0, . . . , 0)}. Let τD be an affine transformation on Zn , u be its linear part and a ∈ Zn , (ei )1≤i≤k be a basis of the subgroup H of Zn . The coset a + H is transformed by τD in the following manner: τD (a + H) =τD (a + Ze1 + · · · + Zek ) = τD (a) + Zu(e1 ) + · · · + Z(ek ) =τD (a) + Zu(H) where the rank of u(H) is equal to k − r(k(u) ∩ (H)). We have to mention that the congruence lattice satisfies the ascending chain condition, that guarantees the systematic termination of the analysis. This can be deduced from the fact that the complete lattice of all subgroups of Zn satisfies the ascending chain condition. Therefore no widening operators are needed for this domain. 4.2

Lifting the Domain of Linear Congruences

The domain of sets of labeled congruences is obtained by lifting the domain of linear congruences using the method described in Sect. 2. The abstract semantics for an imperative programming language with assignments, “if-then-else” and “while” statements is close to this in Subsect. 2.5. Only the “while” statement is treated in a different manner, because no widening is needed for this analysis. Therefore, the widening operator is simply replaced by the upper bound operator.

Refining Static Analyses by Trace-Based Partitioning Using Control Flow

4.3

213

Example

Let us now illustrate with help of the motivated example the analysis based on finite sets of labeled linear congruences. As in Subsect. 3.3 ∆ = (1, 0, 1). Before the while loop at point 6 of the program we have: * + * 2 2 + 2 b, U61 = bt1 , 0 , bf1 , 0 U11 = 0 1 −1 The first pass of the analysis through the “while” loop gives: * + * + 6 −2 0 0 bt v t t , , bf v t t , 1 2 2 3 1 2 2 3 1 −1 1 U12 = * + * + 4 0 1 −1 v t f , , bf v t f , bt 1 2 2 3 1 2 2 3 1 −1 The second pass produces the following set of labeled congruences: * + * + 2 42 2 42 1 = bt1 v2 , 0 + 0 1 Z2 , bf1 v2 , 0 + 0 1 Z2 V 2 = V 1 ∇U12 1 00 −1 00 2 2 2 = U11 = U12 =V2 U72 = U82 = U92 = U10

The analysis converges in the next iteration step (the third one), because V 3 = V 2 . The interpretation of the obtained result is the two sets - {(i, j, k) ∈ Z3 | i − 2j ≡ 2 [4], k ≡ 1[0]} and {(i, j, k) ∈ Z3 | i − 2j ≡ 2 [4], k ≡ −1[0]}. This is a more precise result than the following one obtained by the analysis based on linear congruences {(i, j, k) ∈ Z3 | i − 2j ≡ 2 [4], k ≡ 1 [2]}.

5

Conclusions and Further Work

We have shown how to lift an abstract domain to the finite sets of its labeled abstract properties using control flow information. The significance of this method is that it always produces a static analysis which performs as well as or better than the original one. We have shown how this method can be used with two currently available abstract domains - convex polyhedra and linear congruences. The ideas are in fact general enough to capture other abstract properties as interval congruences, congruence properties on rational numbers etc. The main contributions of this work are: – the design of a method that lifts an abstract domain to the finite sets of its labeled abstract properties as a particular instance of the reduced cardinal power [5], where the base is the set of labels approximating the control history and the exponent is the original domain;

214

M. Handjieva and S. Tzolovski

– the application of this method to the convex polyhedra domain and to the domain of linear congruences; An implementation is currently ongoing and it will allow us to investigate how the algorithms perform in practice. Acknowledgements. We would like to thank Patrick Cousot for pointing out the connection with the reduced cardinal power. We are grateful to our colleagues from LIX for helpful discussions.

References 1. F. Bourdoncle. Abstract interpretation by dynamic partitioning. Journal of Functional Programming, 2(4) (1992) 407-435. 2. P. Cousot. Semantic Foundations of Program Analysis. In Muchnick and Jones Eds. Program Flow Analysis, Theory and Applications, pp. 303-343, Prentice-Hall, 1981. 3. P. Cousot and R. Cousot. Static determination of dynamic properties of programs, In Proceedings of the 2nd Int. Symposium on Programming, pp. 106-130, 1976. 4. P. Cousot and R. Cousot. Abstract interpretation: a unified lattice model for static analysis of programs by construction or approximation of fixpoints. In Proceedings of the 4th ACM Symposium on Principles of Programming Languages, pp. 238-252, 1977. 5. P. Cousot, and R. Cousot, Systematic Design of Program Analysis frameworks. In Proceedings of the 6th ACM Symposium on Principles of Programming Languages, pp. 269-282, 1979. 6. P. Cousot, and R. Cousot, Abstract Interpretation and Application to Logic Programs. In Journal of Logic Programming, pp. 103–179, 1992. 7. P. Cousot and N. Halbwachs. Automatic discovery of linear restraints among variables of a program. In Proceedings of the 5th Annual ACM Symposium on Principles of Programming Languages, pp. 84-97, 1978. 8. G. Fil´e and F. Ranzato, Improving abstract interpretations by systematic lifting to the powerset. In Proceedings of the International Logic Programming Symposium, Ithaca, NY, pages 655-669. The MIT Press, 1994. 9. P. Granger, Static analysis of linear congruence equalities among variables of a program, In Proceedings of the International Joint Conference on Theory and Practice of Software Development, pp. 169-192, number 493 in LNCS, 1991. 10. N. Halbwachs and Y.-E. Proy and P. Raymond, Verification of linear hybrid systems by means of convex approximations, In Proceedings of the International Static Analysis Symposium, pp. 223-237, number 864 in LNCS, 1994. 11. F. Masdupuy. Array operations abstractions using semantics analysis of trapezoid congruences. In Proceedings of the International Conference on Supercomputing, Washington, 1992. 12. A. Schrijver, Theory of Linear and Integer Programming, John Wiley & Sons, 1986.

Building Complete Abstract Interpretations in a Linear Logic-Based Setting Roberto Giacobazzi† †

‡

Francesco Ranzato‡,

Francesca Scozzari†

Dipartimento di Informatica, Universit` a di Pisa Corso Italia 40, 56125 Pisa, Italy {giaco,scozzari}@di.unipi.it

Dipartimento di Matematica Pura ed Applicata, Universit` a di Padova Via Belzoni 7, 35131 Padova, Italy [email protected]

Abstract. Completeness is an important, but rather uncommon, property of abstract interpretations, ensuring that abstract computations are as precise as possible w.r.t. concrete ones. It turns out that completeness for an abstract interpretation depends only on its underlying abstract domains, and therefore it is an abstract domain property. Recently, the ﬁrst two authors proved that for a given abstract domain A, in all significant cases, there exists the most abstract domain, called least complete extension of A, which includes A and induces a complete abstract interpretation. In addition to the standard formulation, we introduce and study a novel and particularly interesting type of completeness, called observation completeness. Standard and observation completeness are here considered in the context of quantales, i.e. models of linear logic, as concrete interpretations. In this setting, we prove that various kinds of least complete and observationally complete extensions exist and, more importantly, we show that such complete extensions can all be explicitly characterized by elegant linear logic-based formulations. As an application, we determine the least complete extension of a generic abstract domain w.r.t. a standard bottom-up semantics for logic programs observing computed answer substitutions. This general result is then instantiated to the relevant case of groundness analysis.

1

Introduction

It is widely held that the ideal goal of any semantics design method is to ﬁnd sound and complete representations for some properties of concrete (actual) computations. Abstract interpretation is one such methodology, where soundness is always required, while completeness more rarely holds. Completeness issues in abstract interpretation have been studied since the Cousot and Cousot seminal paper [5]. The intuition is that a complete abstract interpretation induces an

The work of Francesco Ranzato has been supported by an individual post-doctoral grant from Universit` a di Padova, Italy.

G. Levi (Ed.): SAS’98, LNCS 1503, pp. 215–229, 1998. c Springer-Verlag Berlin Heidelberg 1998

216

R. Giacobazzi, F. Ranzato, and F. Scozzari

abstract semantics which is as precise as possible relatively to its underlying abstract domains and to the concrete interpretation of reference. The paradigmatic rule of signs is a typical and simple example of an abstract interpretation which is (sound and) complete for integer multiplication but merely sound for integer addition [13]. Although in static program analysis decidability issues often force to sacriﬁce completeness for achieving termination and/or eﬃciency, examples of complete abstract interpretations are common in other ﬁelds of application. For instance, several complete abstractions of algebraic polynomial systems have been studied by Cousot and Cousot in [7] and many complete abstract interpretations can be found in comparative program semantics [3,6,9]. Moreover, being completeness a notion relative to the concrete semantics of reference, complete abstract interpretations which are more concrete than a certain, possibly approximated and decidable, property of interest, yield an absolute upper bound for the precision that one can achieve in computing that property. Thus, complete abstract interpretations may play a useful rˆ ole in static program analysis as well. These argumentations probably stimulated the recent trend of research on completeness in abstract interpretation [2,10,11,13,15,17,18]. One key feature of completeness in abstract interpretation is that this property uniquely depends upon the abstraction function. Let us denote by LC the so-called lattice of abstract interpretations of a concrete domain C (cf. [4,5]), where, for all A, B ∈ LC , A B means that A is more precise (i.e. concrete) than B. Let us consider the simple case of an abstract interpretation f : A → A of a concrete semantic function f : C → C, where the abstract domain A ∈ LC is related to C by an adjoint pair of abstraction and concretization maps α : C → A and γ : A → C. Then, f is (sound and) complete if α ◦ f = f ◦ α. It is easily seen that if f is complete then the best correct approximation f b of f in A, i.e. def f b = α ◦ f ◦ γ : A → A, is complete as well, and, in this case, f indeed coincides with f b (cf. [10]). Thus, given an abstract domain A ∈ LC , one can deﬁne a complete abstract semantic function f : A → A over A if and only if f b : A → A is complete. This simple observation makes completeness an abstract domain property, namely a characteristic of the abstract domain. It is then clear that a key problem consists in devising systematic and constructive methodologies for transforming abstract domains in such a way that completeness is achieved and the resulting complete abstract domains are as close as possible to the initial (noncomplete) ones. This problem has been ﬁrst raised in a predicate-based approach to abstract interpretation by Mycroft [13, Section 3.2], who gave a methodology for deriving the most concrete domain, called “canonical abstract interpretation”, which is complete and included in a given domain of properties. More recently, the ﬁrst two authors proved in [10] that when the concrete semantic function f is continuous, any domain A can always be extended into the most abstract domain which includes A and is complete for f — the so-called least complete extension of A. Analogously, [10] solved the aforementioned problem raised by Mycroft, by showing that, for any given domain A, the most concrete domain which is complete and included in A — the so-called complete kernel

Building Complete Abstract Interpretations in a Linear Logic-Based Setting

217

of A — exists for any monotone semantic function. Very recently, we improved considerably such results, by providing explicit constructive characterizations for least complete extensions and complete kernels of abstract domains [11]. In this paper, we are concerned with completeness problems arising when concrete semantic binary1 operations of type C1 × C2 → C are assumed to give rise to a generalized form of quantales, called typed quantales. Quantales are well-known algebraic structures which turn out to be models of intuitionistic linear logic [16,19]. In this logical setting, we provide elegant linear logic-based solutions to a number of interesting completeness problems for abstract interpretations. Such solutions ﬁnd relevant applications in static program analysis and comparative semantics, for instance in logic programming, where uniﬁcation — the prime computational step of any logic program semantics — turns out to be a binary operation in a quantale of substitutions, or in data structure analysis, considering binary data constructors such as cons for lists. More in detail, a typed quantale C, C1 , C2 , ⊗ consists of three complete lattices C, C1 and C2 , and of an operation ⊗ : C1 × C2 → C which is additive (i.e., preserves lub’s) on both arguments. When C = C1 = C2 , typed quantales boil down to standard quantales C, ⊗. The main feature of (typed) quantales is that they support a notion of left and right linear implication between domain’s objects: Given a ∈ C1 and b ∈ C, there exists a unique greatest object a b ∈ C2 which, when combined by ⊗ with a, gives a result less than or equal to b. In other terms, the b holds. Analogous left following right modus ponens law a ⊗ x ≤ b ⇔ x ≤ a implicational objects exist for a correspoding left modus ponens law. When solving completeness problems in a setting where concrete interpretations are typed quantales, implicational domain objects allow to elegantly characterize complete abstract domains in a variety of situations. Our ﬁrst result provides a characterization based on linear implications between domain’s objects of the least complete extension of any abstract domain of any quantale. Then, we consider the following completeness problem over typed quantales: Given a typed quantale C, C1 , C2 , ⊗, a ﬁxed abstraction A ∈ LC , with corresponding abstraction map αA : C → A, and a pair of abstract domains A1 , A2 ∈ LC1 × LC2 , does there exist the most abstract pair of domains A1 , A2 ∈ LC1 × LC2 , with corresponding abstraction maps αAi : Ci → Ai (i = 1, 2), such that A1 , A2 L C1 ×L C2 A1 , A2 and αA ( ⊗ ) = αA (αA1 ( ) ⊗ αA2 ( ))? Here, the observation domain A is ﬁxed, and we are thus looking for the most abstract pair of domains in LC1 × LC2 which is more concrete than an initial pair A1 , A2 and simultaneously induces a complete abstract interpretation w.r.t. ⊗. This is termed an observation completeness problem. Again, solutions to this observation completeness problem are built in terms of linear implications between domains. To illustrate the practical scope of our results, we ﬁrst consider a simple example in data structure analysis involving abstract domains for lists. In particular, our results are applied in order to solve various observation completeness problems concerning abstract domains useful for detecting irredundant lists of 1

Clearly, a generalization to n-ary operations would be straightforward.

218

R. Giacobazzi, F. Ranzato, and F. Scozzari

objects. Then, in the context of logic program semantics, we consider an immediate consequences operator TP deﬁned in terms of uniﬁcation and union of sets of idempotent substitutions, and characterizing computed answer substitutions in a s-semantics style (cf. [1,8]). As usual, uniﬁcation turns out to be the key operation to take into account in order to build least complete extensions of abstract domains. Sets of idempotent substitutions and uniﬁcation give rise to a unital commutative quantale: Given an abstract domain A, we show how the least complete extension of A w.r.t. this quantale naturally induces the least complete extension of A w.r.t. TP functions. This permits to give explicitly, in terms of linear implications, the least complete extension, for any TP , of a generic domain abstracting sets of substitutions. As a remarkable instance of our construction, we characterize the least complete extension of the plain groundness domain w.r.t. computed answer substitutions s-semantics.

2

Basic Notions

The lattice of abstract interpretations. In standard Cousot and Cousot’s abstract interpretation theory, abstract domains can be equivalently speciﬁed either by Galois connections (GCs), i.e. adjunctions, or by upper closure operators (uco’s) [5]. In the ﬁrst case, the concrete and abstract domains C and A (both assumed to be complete lattices) are related by a pair of adjoint functions of a GC (α, C, A, γ). Also, it is generally assumed that (α, C, A, γ) is a Galois insertion (GI), i.e. α is onto or, equivalently, γ is 1-1. In the second case instead, an abstract domain is speciﬁed as a uco on the concrete domain C, i.e. a monotone, idempotent and extensive operator on C. These two approaches are equivalent, modulo isomorphic representation of domain’s objects. Given a complete lattice C, it is well known that the set uco(C) of all uco’s on C, endowed with the pointwise ordering , is a complete lattice uco(C), , , , λx. C , id (id denotes the identity function). Let us also recall that each ρ ∈ uco(C) is uniquely determined by the set of its ﬁxpoints, which is its image, i.e. ρ(C) = {x ∈ C | ρ(x) = x}, and that ρ η iﬀ η(C) ⊆ ρ(C). Moreover, a subset X ⊆ C is the set of ﬁxdef points of a uco on C iﬀ X is meet-closed, i.e. X = (X) = { C Y | Y ⊆ X} (note that C = C ∈ (X)). Often, we will identify closures with their sets of ﬁxpoints. This does not give rise to ambiguity, since one can distinguish their use as functions or sets according to the context. In view of the equivalence above, throughout the paper, uco(C), will play the rˆ ole of the lattice LC of abstract interpretations of C [4,5], i.e. the complete lattice of all possible abstract domains of the concrete domain C. For an abstract domain A ∈ LC , ρA ∈ uco(C) will denote the corresponding uco on C, and if A is speciﬁed by def a GI (α, C, A, γ) then ρA = γ ◦ α. The ordering on uco(C) corresponds to the standard order used to compare abstract domains with regard to their precision: A1 is more precise than A2 (i.e., A1 is more concrete than A2 or A2 is more abstract than A1 ) iﬀ A1 A2 in uco(C). Lub and glb on uco(C) have therefore the following reading as operators on domains. Suppose {Ai }i∈I ⊆ uco(C): (i) i∈I Ai is the most concrete among the domains which are abstractions of all

Building Complete Abstract Interpretations in a Linear Logic-Based Setting

219

the Ai ’s, i.e. it is their least (w.r.t. ) common abstraction; (ii) i∈I Ai is the most abstract among the domains (abstracting C) which are more concrete than every Ai ; this domain is also known as reduced product of all the Ai ’s. Quantales and linear logic. Quantales originated as algebraic foundations of the so-called quantum logic. They have been successively considered for the latticetheoretic semantics of Girard’s linear logic (see [16] for an exhaustive treatment of quantales). We introduce a mild generalization of the notion of quantale, which, up to knowledge, appears to be new. A typed quantale is a multisorted and ⊗ : C1 ×C2 → C algebra C, C1 , C2 , ⊗, where C, C1 , C2are complete lattices is a function such that ( i xi ) ⊗ c2 = i (xi ⊗ c2 ) and c1 ⊗ ( i xi ) = i (c1 ⊗ xi ). In other terms, a typed quantale is a 3-sorted algebra endowed with a “product” ⊗ which distributes over arbitrary lub’s on both sides. Thus, for any c1 ∈ C1 and c2 ∈ C2 , both functions c1 ⊗ and ⊗ c2 have right adjoints denoted, resp., by c1 and c2 . Hence, for all c ∈ C, c1 ⊗ c2 ≤ c ⇔ c2 ≤ c1 c, and, dually, c1 ⊗ c2 ≤ c ⇔ c1 ≤ c c2 . Two functions : C1 × C → C2 and : C × C2 → C1 can be therefore deﬁned as follows: def def c1 c= {z ∈ C2 | c1 ⊗ z ≤ c}; c c2 = {y ∈ C1 | y ⊗ c2 ≤ c}.

Any typed quantale C, C1 , C2 , ⊗ enjoys the following main properties: For all c ∈ C, {xi }i∈I ⊆ C, c1 ∈ C1 , {yi }i∈I ⊆ C1 , c2 ∈ C2 and {zi }i∈I ⊆ C2 : (i) (iii) (v)

c) ≤ c c1 ⊗ (c1 c1 ( xi ) = (c1 (

i∈I

i∈I

yi )

(ii) xi )

(iv)

i∈I

c=

(c c2 ) ⊗ c2 ≤ c ( xi ) c2 = (xi

i∈I

(yi

c)

i∈I

(vi)

c

c ) 2

i∈I

( z ) = (c z ) i

i∈I

i

i∈I

When C = C1 = C2 and ⊗ is associative, a typed quantale is called quantale. It is well known that quantales turn out to be models of noncommutative intuitionistic linear logic [16,19]. A quantale C, ⊗ is called commutative when ⊗ is comb=b a. mutative, and this is equivalent to require that, for all a, b ∈ C, a Also, a commutative quantale C, ⊗ is called unital if there exists an object 1 ∈ C such that 1 ⊗ a = a = a ⊗ 1, for all a ∈ C. For a quantale C, ⊗, the following additional properties hold for all a, b, c ∈ C:

(vii) (viii) (ix)

3

b) = (a c) b b (a c) = (a ⊗ b) c (c b) a = c (a ⊗ b)

a

(c

Completeness Problems in Abstract Interpretations

Let C, C1 , C2 , ⊗ be a concrete interpretation, i.e. C, C1 and C2 are concrete semantic domains provided with a semantic operation ⊗ : C1 × C2 → C. When

220

R. Giacobazzi, F. Ranzato, and F. Scozzari

C = C1 = C2 , we adopt the simpler notation C, ⊗. Given the abstractions A1 ∈ LC1 , A2 ∈ LC2 and A ∈ LC , let us recall [5] that the best correct approxdef imation ⊗b : A1 × A2 → A of ⊗ is deﬁned as ⊗b = αC,A ◦ ⊗ ◦ γA1 ,C1 , γA2 ,C2 . It has been shown in [10] that completeness for an abstract interpretation is a property depending only on the underlying abstract domains. In our setting, this means that an abstract interpretation A, A1 , A2 , ⊗ , with ⊗ : A1 × A2 → A, is complete for C, C1 , C2 , ⊗ iﬀ A, A1 , A2 , ⊗b is complete and ⊗ = ⊗b . In other terms, the best correct approximation induced by the underlying abstract domains determines the property of being complete. Hence, we ﬁnd more convenient, elegant and, of course, completely equivalent, to reason on completeness by using abstract domains speciﬁed by the closure operator approach. Full completeness problems. Within the closure operator approach, given a concrete interpretation C, ⊗, an abstraction A ∈ LC is complete when the equation ρA ◦⊗◦ρA , ρA = ρA ◦⊗ holds. Giacobazzi and Ranzato [10] stated the following full completeness problem, here specialized to binary semantic operations: Given an abstract domain A ∈ LC , does the following system with variable ρ admit a most abstract solution? ρA (1) ρ ◦ ⊗ ◦ ρ, ρ = ρ ◦ ⊗ Hence, a solution to the above full completeness problem is the necessarily unique (up to domain isomorphism) most abstract domain which includes A and induces a complete abstract interpretation. Following [10], such most abstract solution to System (1) is called the least complete extension of A w.r.t. ⊗. It is shown in [10] that if ⊗ is continuous in both arguments, then least complete extensions of any A exist. Full completeness problems clearly make sense only for concrete interpretations of type C, ⊗. When generic concrete interpretations of type C, C1 , C2 , ⊗ are considered, diﬀerent abstractions are involved in a completeness equation, and therefore various completeness problems arise by ﬁxing some of these abstractions. In the following, we introduce three such completeness problems, which turn out to be of particular interest. Such completeness problems still depend only on best correct approximations of the concrete operation, and therefore the corresponding completeness notions are again abstract domain properties. Observation completeness problems. Let C, C1 , C2 , ⊗ be a concrete interpretation. An observation domain is any abstraction of the range C of ⊗. Observation completeness problems arise when in a completeness equation an observation domain is ﬁxed. Hence, let us consider a ﬁxed observation domain A ∈ LC . The observation completeness problem for a pair A1 , A2 ∈ LC1 × LC2 admits solution when there exists the most abstract solution in LC1 × LC2 of the following system: η, µ A1 , A2 (2) ρA ◦ ⊗ ◦ η, µ = ρA ◦ ⊗

Building Complete Abstract Interpretations in a Linear Logic-Based Setting

221

Let us remark that, by using adjunctions, the observation completeness equation ρA ◦ ⊗ ◦ η, µ = ρA ◦ ⊗ is equivalent to require that for all x ∈ C1 and y ∈ C2 , αC1 ,η (x) ⊗b αC2 ,µ (y) = αC,A (x ⊗ y). When in addition to the observation domain we also ﬁx one (or more, if we would deal with n-ary operations) of the abstractions of the argument domains of ⊗, we obtain yet diﬀerent completeness problems. In the left observation completeness problem, A ∈ LC and A2 ∈ LC2 are ﬁxed, and the solution to this problem for a given A1 ∈ LC1 exists when the following system with variable η admits a most abstract solution in LC1 : η A1 (3) ρA ◦ ⊗ ◦ η, ρA2 = ρA ◦ ⊗ ◦ id, ρA2 Of course, right observation completeness problems are analogously formulated. It turns out that a left observation completeness equation ρA ◦ ⊗ ◦ η, ρA2 = ρA ◦ ⊗ ◦ id, ρA2 formulated in terms of GIs amounts to require that for any x ∈ C1 and y ∈ A2 , αC1 ,η (x) ⊗b y = αC,A (x ⊗ γA2 ,C2 (y)). Hence, a left observation completeness equation ρA ◦ ⊗ ◦ η, ρA2 = ρA ◦ ⊗ ◦ id, ρA2 states that completeness for the abstractions A ∈ LC , η ∈ LC1 and A2 ∈ LC2 holds when the semantic operation ⊗ acts over C1 × A2 . Examples of observation completeness problems will be considered and solved later in Section 5.

4

Quantales and Solutions to Completeness Problems

When concrete interpretations are quantales and typed quantales, solutions to the above completeness problems exist and can be characterized explicitly and elegantly in terms of linear implications. Let C, C1 , C2 , ⊗ be a typed quantale playing the rˆ ole of concrete interpretation. In this setting, solutions to completeness problems will be characterized by exploiting two basic domain transformers : uco(C1 ) × uco(C) → uco(C2 ) and : uco(C) × uco(C2 ) → uco(C1 ), deﬁned by lifting left and right linear and to abstract domains as follows: For any A1 ∈ LC1 , A2 ∈ implications LC2 , A ∈ LC :

({a a ∈ C A A = ({a a ∈ C A1

def

A=

1

2

| a1 ∈ A1 , a ∈ A});

1

| a ∈ A, a2 ∈ A2 }).

def

2

2

Hence, A1 A is deﬁned to be the most abstract domain in LC2 containing all the linear implications from A1 to A. From the logic properties of linear implication recalled in Section 2, it is not too hard to derive the following useful distributivity laws for and over reduced product of abstract domains.

Proposition 1. For all {Bi }i∈I ⊆ LC , A1

(

i∈I

Bi ) =

i∈I (A1

Bi )

and

(

i∈I

Bi )

A = 2

i∈I (Bi

A ). 2

222

R. Giacobazzi, F. Ranzato, and F. Scozzari

Solutions to full completeness problems. Solutions to full completeness problems exist and are characterized in terms of linear implications among domain’s objects as stated by the following result. Theorem 2. A(C of System (1).

A)(A

C)((C A) C) is the most abstract solution

Solutions to observation completeness problems. Let us ﬁrst consider left observation completeness problems. A left observation completeness equation can be characterized as follows. Theorem 3. ρA ◦ ⊗ ◦ η, ρA2 = ρA ◦ ⊗ ◦ id, ρA2 ⇔ η A

A . 2

Thus, as an immediate consequence, left observation completeness problems admit the following solutions. Corollary 4. A1 (A

A ) is the most abstract solution of System (3). 2

Of course, dual results can be given for right observation completeness problems. In this case, the most abstract solution therefore is A2 (A1 A). The above results for left and right observation completeness turn out to be useful for solving observation completeness problems. In fact, an observation completeness equation is characterized as an independent combination of left and right observation completeness equations as follows. Theorem 5. ρA ◦ ⊗ ◦ η, µ = ρA ◦ ⊗ ⇔ η, µ A

C , C 2

1

A.

As a straight consequence, we get the following result. Corollary 6. A1 (A System (2). 4.1

C ), A 2

2

(C1

A) is the most abstract solution of

The Case of Unital Commutative Quantales

When we deal with unital and commutative quantales — i.e. models of intuitionistic linear logic [16,19] — the above solutions to full completeness problems given by Theorem 2 can be signiﬁcantly simpliﬁed by exploiting the logical properties of linear implication. In a unital commutative quantale C, ⊗, the following additional properties hold: For any a, b, c ∈ C: – a (b – c ≤ (c

c) = b a) a

(a

c) – 1 a=a – ((c a) a)

a=c

a

Therefore, from these properties it is not hard to check that for all a ∈ C, λc.(c a) a ∈ uco(C). This turns out to be the key observation in order to give a more compact form to solutions of full completeness problems on unital commutative quantales. Moreover, the objects of such solutions enjoy a clean logical characterization in terms of linear implications as speciﬁed by the third point of the next result.

Building Complete Abstract Interpretations in a Linear Logic-Based Setting

223

Theorem 7. Let C, ⊗ be a unital commutative quantale.

1. C A is the most abstract solution of System (1); 2. If B ∈ LC is such that B C A, then C A = B A; 3. For all c ∈ C, c ∈ C A ⇔ c = a∈A (c a) a.

5

An Application in Data Structure Completeness

In this section, we consider some examples of completeness problems in abstract interpretation of list data structures. For the sake of practicality, we consider lists of natural numbers, even if our discussion holds more in general for lists of objects of arbitrary type. Consider the structure ℘(list( )), ℘( ), ℘(list( )), ::, where list( ) is the set of all ﬁnite lists of natural numbers, ℘(list( )) and ℘( ) are complete lattices w.r.t. set-inclusion, and :: : ℘( ) × ℘(list( )) → ℘(list( )) is a “collecting” version of concatenation deﬁned as follows: def

N :: L = {[n|l] | n ∈ N, l ∈ L}, where :: L = N :: = . It is clear that this structure is a typed quantale. We say that a list is irredundant if it does not contain two occurrences of the same object. An abstract domain ρ ∈ uco(℘(list( ))) for detecting irredundant def lists can be deﬁned by ρ = { list(), Irr }, where Irr ⊆ list( ) is the set of irredundant lists over . We consider ρ as an observation domain and we look for the most abstract solution X, Y ∈ uco(℘( )) × uco(℘(list( ))) to the following observation completeness problem: ρ ◦ :: ◦ ρX , ρY = ρ ◦ :: Here, we dropped the ﬁrst constraint of the generic System (2), since it amounts to the trivial constraint ρX , ρY { }, { list() } which is always satisﬁed. By Corollary 6, the most abstract solution of this observation completeness equation exists. This will be the most abstract pair of domains X, Y for which abstract concatenation in X and Y results to be complete when observing irredundancy as represented by ρ. By Corollary 6, the solution X, Y is as follows. X = { } (ρ

℘(list()))

℘(list( )) = ({L M | L ∈ ρ, M ∈ ℘(list( ))}) (since, for all M , list( ) M = list( )) = ({Irr M | M ∈ ℘(list( ))}) (since, for all n ∈ , Irr {[n]} = {n})

=ρ

= ℘( )

224

R. Giacobazzi, F. Ranzato, and F. Scozzari

Y = { list( ) } (℘( ) = ℘( )

ρ)

= ({N

ρ

= ({N

(since, for all N , N

L | N ∈ ℘( ), L ∈ ρ}) Irr | N ∈ ℘( )})

list( ) = list( ))

(since, by (v) in Section 2,

i (Ni

Irr) = ( i Ni )

Irr)

= {N Irr | N ∈ ℘( )} = N ⊆ {L ∈ ℘(list( )) | l ∈ L ⇔ (l ∈ Irr and ∀n ∈ N. n is not in l)} Thus, in order to be complete for concatenation when observing irredundancy, X must coincide with the concrete domain ℘( ), while it suﬃces that Y contains all the sets of irreduntant lists which do not contain some set of numbers. Note that Y coincides with the set of all the sets of irredundant lists closed by permutation of their objects, and this is a strict abstraction of the concrete domain ℘(list( )). Let us now consider the standard abstract domain η ∈ uco(℘( )) for pardef ity analysis given by η = { , even, odd, }. We consider the following left and right observation completeness problems: We look respectively for the the most abstract domains X ∈ uco(℘( )) and Y ∈ uco(℘(list( ))) such that: (i) ρ ◦ :: ◦ ρX , ρ = ρ ◦ :: id, ρ

(ii) ρ ◦ :: ◦ η, ρY = ρ ◦ :: η, id

Here again, there are no upper bound constraints for X and Y . By Corollary 4 (and its dual), we get the following solutions:

X = { } (ρ ρ) =ρ ρ = ({ list( ) list( ), list( ) = { , },

Irr, Irr list(), Irr Irr })

Y = { list() } (η ρ) =η ρ Irr, even Irr, odd Irr, = ({ = { { [] }, Irreven , Irrodd , list( ) },

Irr })

where Irreven = {l ∈ list( ) | l ∈ Irr and l does not contain even numbers} and def Irrodd = {l ∈ list( ) | l ∈ Irr and l does not contain odd numbers}. Thus, for problem (i), in order to get completeness, it is enough to check whether a given set of numbers is empty or not, while, for problem (ii), we only need to consider sets of irreduntant lists which do not contain either even or odd numbers. def

6

Complete Semantics for Logic Program Analysis

In this section, we determine the least complete extension of any logic program property w.r.t. a bottom-up semantics characterizing computed answer substitutions, which turns out to be equivalent to the well-known s-semantics [8]. This

Building Complete Abstract Interpretations in a Linear Logic-Based Setting

225

complete semantics can be thought of as a logic program semantics which is “optimal” (i.e. neither too concrete nor too abstract) for characterizing a given property of interest. The problem of determining optimal semantics for program analysis was raised in [9]. We show that such optimal semantics can be obtained by solving a full completeness problem relatively to the operation of uniﬁcation on sets of substitutions. Completeness for an abstract domain, in general, depends both on the considered concrete domains and on the semantic operation deﬁned on them. However, the following important result shows that, under certain conditions, completeness is instead independent from the choice of the concrete semantics. Theorem 8. Let α ∈ LC and f, g : C → C such that f g α ◦ f . Then, α is complete for f iﬀ α is complete for g. We will exploit this result for computing our least complete extension of abstract domains w.r.t. s-semantics. The idea is that of considering a “simpliﬁed” and more concrete version, denoted by TP , of the immediate consequences operator TPs of s-semantics, which does not take into account variable renaming. This will simplify a lot the technical development of this section. Hence, s-semantics results to be an abstract interpretation of our TP semantics: If r denotes the closure under variable renaming of sets of substitutions, then we have that TPs = r ◦ TP . Then, leaving out the details, if α denotes the least complete extension, relatively to TP , of any domain A abstracting computed answer substitutions, since TP r ◦ TP = TPs α ◦ TP holds, we can apply Theorem 8, from which we get that α not only is complete for TPs , but actually α turns out to be the least complete extension of A relatively to TPs . 6.1

Notation

Let V be an inﬁnite, recursively enumerable (r.e.) set of variables, Σ be a set of function symbols and Π be a set of predicate symbols, deﬁning a r.e. ﬁrst-order language L. Term denotes the set of terms of L. If s is any syntactic object and σ and θ are substitutions, then vars(s) denotes the set of variables occurring def def in s, dom(σ) = {v ∈ V | σ(v) = v}, rng(σ) = ∪ {vars(σ(x)) | x ∈ dom(σ)}, sσ denotes the application of σ to s, and σ ◦ θ denotes the standard composition of θ and σ (i.e., σ ◦ θ = λx.(θ(x))σ). The set of idempotent substitutions modulo renaming (i.e., θ ∼ σ iﬀ there exists β and δ such that θ = σ ◦ β and σ = θ ◦ δ) on L is denoted by Sub. Objects in Sub are partially ordered by instantiation, denoted by . By adding to Sub an extra object τ as least element, one gets a complete lattice Subτ , , ∨, ∧, , τ , where ∨ is least general anti-instance, ∧ is standard uniﬁcation and is the empty substitution (see [14] for more details). def ¯ | p ∈ Π}, where X ¯ is The set of most general atoms is given by GAtom = {p(X) a tuple of distinct variables. We consider logic programs in normalized form, that ¯ 1 ), . . . , qn (X ¯ n ), where all the tuples of ¯ : −c, q1 (X is, a generic Horn clause is p(X) variables are distinct and c ∈ Sub is an idempotent substitution binding variables to terms.

226

6.2

R. Giacobazzi, F. Ranzato, and F. Scozzari

TP -Completeness

Our basic semantic structure is the unital commutative quantale ℘(Sub), ⊗, where ℘(Sub), ⊆ is a complete lattice, ⊗ : ℘(Sub) × ℘(Sub) → ℘(Sub) is the obvious lifting of uniﬁcation ∧ to sets of substitutions, i.e. it is deﬁned def by: X ⊗ Y = {x ∧ y | x ∈ X, y ∈ Y }, and {} ∈ ℘(Sub) is the unit of ⊗. It is immediate to check that ℘(Sub), ⊗ actually is a unital commutative quantale [16, Example 10, p. 18]. In the following, we will slightly abuse notation by applying the operation ⊗ also to substitutions. As mentioned above, we consider a bottom-up semantics based on an immediate consequences operator TP which is more concrete than the standard operator TPs of s-semantics. In fact, TP is deﬁned using only the operations of uniﬁcation ⊗ and union of sets of idempotent substitutions. The s-semantics operator TPs can be recovered from TP by a simple step of abstract interpretation considering the closure of sets of substitutions over variables renaming. We consider a concrete semantic domain CInt of functions — as usual, called interpretations — which map most general atoms to sets of substitutions: def CInt = GAtom → ℘(Sub), which ordered pointwise is trivially a complete lattice. Often, we will ﬁnd convenient to denote an interpretation I ∈ CInt by the ¯ I(p(X)) ¯ set {p(X), | p ∈ Π}. Then, for any program P , the immediate consequences operator TP : CInt → CInt is deﬁned as follows: For any I ∈ CInt, ¯ (c ⊗ ( I(qi (X¯i ))) ⊗ {Y¯ = X}), TP (I)(p(Y¯ )) = C

P

i=1..n

¯ : −c, q1 (X¯1 ), . . . , qn (X¯n ). Here, C where C = p(X) P denotes that the clause ¯i )) is intended modulo C of P is renamed apart with fresh variables, and I(qi (X renaming. If ρ ∈ uco(℘(Sub)) is any abstraction on sets of substitutions, a corresponding abstraction on interpretations ρ ∈ uco(CInt) which acts accordingly to ρ can def ¯ ρ(c) | p(X), ¯ c ∈ I}. be deﬁned as follows: For any I ∈ CInt, ρ (I) = {p(X), Note that · is monotone, i.e., for all ρ, η ∈ uco(℘(Sub)), if ρ η then ρ η .

Given a basic abstract domain of properties of interest π ∈ uco(℘(Sub), ⊆), our goal is therefore to ﬁnd the most abstract domain which contains π (more precisely, π ) and is complete for any TP operator. Our strategy consists in characterizing the least complete extension of π for the basic operations involved in the deﬁnition of TP , and then to show that, under reasonable hypotheses, this domain turns out to be the right one. Since every abstract domain is trivially always complete for lub’s, it turns out that union of sets of substitutions is troubleless, and therefore it is enough to concentrate on uniﬁcation ⊗. Indeed, the following result shows that completeness for ⊗ implies completeness for any TP .

Theorem 9. Let ρ ∈ uco(℘(Sub)). If ρ is complete for ⊗ then ρ is complete for any TP .

Building Complete Abstract Interpretations in a Linear Logic-Based Setting

227

By Theorem 7, given π ∈ uco(℘(Sub)), the least complete extension of π w.r.t. the quantale operation ⊗ exists and is the domain ℘(Sub) π of linear implications from ℘(Sub) to π. This complete domain ℘(Sub) π results to be the right one whenever the abstract domain π satisﬁes the following weak decidability property. Definition 10. π ∈ uco(℘(Sub)) is decidable if any S ∈ π is a r.e. set.

¾

It should be clear that decidability is a reasonable requirement for most abstract domains used in program analysis: In fact, for such a decidable abstraction (of sets) of substitutions, an eﬀective procedure for checking whether a substitution belongs to (is approximated by) an abstract object is available. As announced, the following key result shows that least complete extensions of decidable abstract domains w.r.t. uniﬁcation ⊗ actually are least complete extensions for TP operators as well.

Theorem 11. Let π ∈ uco(℘(Sub)) be decidable. Then, ℘(Sub) least complete extension of π for any TP .

π is the

As we discussed just after Theorem 8, it turns out that, for any decidable π ∈ uco(℘(Sub)), ℘(Sub) π actually is the least complete extension of π for any immediate consequences operator TPs of s-semantics. In fact, if r ∈ uco(℘(Sub)) is the closure under renaming of sets of substitions, namely variables in the range of substitutions are renamed in each possible way, then TPs = r ◦ TP ; moreover, ℘(Sub) π clearly induces a semantic transformer less precise than TPs , i.e. s TP ℘(Sub) π ◦ TP . Hence, by Theorems 8 and 11, we get the following desired consequence.

Corollary 12. Let π ∈ uco(℘(Sub)) be decidable, such that TPs π ◦ TP for any P . Then, ℘(Sub) π is the least complete extension of π for any TPs .

6.3

Complete Semantics for Groundness Analysis

Groundness analysis is arguably one of the most important analysis for logicbased programming languages. Groundness analysis aims to statically detect whether variables will be bound to ground terms in successful derivations. By instantiating the results above, we are able to characterize the least complete extension of the basic abstract domain representing plain groundness information w.r.t. any immediate consequences operator of s-semantics. The resulting semantics can be therefore interpreted as the “optimal” semantics for groundness analysis. If V ⊆ V is a ﬁnite set of variables of interest, the simplest abstract domain for def representing plain groundness information of variables in V is GV = ℘(V ), ⊇, as ﬁrst put forward by Jones and Søndergaard [12]. The intuition is that each W ∈ GV represents the set of substitutions which ground every variable in W . GV is related to the concrete domain ℘(Sub) by the following concretization map: def For each W ∈ G, γG V (W ) = {θ ∈ Sub | ∀v ∈ W. vars(θ(v)) = }. As usual, we

228

R. Giacobazzi, F. Ranzato, and F. Scozzari

shall abuse notation by denoting with GV ∈ uco(℘(Sub)) the corresponding isomorphic image γG V (GV ) in ℘(Sub). A variable independent abstract domain G ∈ uco(CInt) for representing plain groundness information can be therefore deﬁned as follows: def

¯ GX¯ (c) | p(X), ¯ c ∈ I}. G(I) = {p(X), It is easy to check that G actually is a uco on the concrete semantic domain CInt. Furthermore, G is clearly decidable and TPs G ◦ TP . Thus, as an easy consequence of Corollary 12, we can prove the following result, where V ⊂f V means that V is a ﬁnite subset of V. Theorem 13. any TPs .

V ⊂f V

℘(Sub) G is the least complete extension of G for V

References 1. A. Bossi, M. Gabbrielli, G. Levi, and M. Martelli. The s-semantics approach: theory and applications. J. Logic Program., 19-20:149–197, 1994. 218 2. P. Cousot. Completeness in abstract interpretation (Invited Talk). In Proc. 1995 Joint Italian-Spanish Conference on Declarative Programming, pages 37–38, 1995. 216 3. P. Cousot. Constructive design of a hierarchy of semantics of a transition system by abstract interpretation (Invited Paper). In Proc. of the 13th Int. Symp. on Math. Found. of Programming Semantics (MFPS’97), vol. 6 of Electronic Notes in Theor. Comput. Sci., 1997. 216 4. P. Cousot and R. Cousot. Abstract interpretation: a uniﬁed lattice model for static analysis of programs by construction or approximation of ﬁxpoints. In Proc. 4th ACM POPL, pages 238–252, 1977. 216, 218 5. P. Cousot and R. Cousot. Systematic design of program analysis frameworks. In Proc. 6th ACM POPL, pages 269–282, 1979. 215, 216, 218, 218, 220 6. P. Cousot and R. Cousot. Inductive deﬁnitions, semantics and abstract interpretation. In Proc. 19th ACM POPL, pages 83–94, 1992. 216 7. P. Cousot and R. Cousot. Abstract interpretation of algebraic polynomial systems. In Proc. 6th Int. Conf. on Algebraic Methodology and Software Technology (AMAST’97), LNCS 1349, pages 138–154, 1997. 216 8. M. Falaschi, G. Levi, M. Martelli, and C. Palamidessi. Declarative modeling of the operational behavior of logic languages. Theor. Comput. Sci., 69(3):289–318, 1989. 218, 224 9. R. Giacobazzi. “Optimal” collecting semantics for analysis in a hierarchy of logic program semantics. In Proc. 13th Int. Symp. on Theor. Aspects of Comput. Sci. (STACS’96), LNCS 1046, pages 503–514, 1996. 216, 225 10. R. Giacobazzi and F. Ranzato. Completeness in abstract interpretation: a domain perspective. In Proc. 6th Int. Conf. on Algebraic Methodology and Software Technology (AMAST’97), LNCS 1349, pages 231–245, 1997. 216, 216, 216, 216, 220, 220, 220, 220 11. R. Giacobazzi, F. Ranzato and F. Scozzari. Complete abstract interpretations made constructive. In Proc. 23rd Int. Symp. on Math. Found. of Comp. Sci. (MFCS’98), LNCS, 1998. 216, 217

Building Complete Abstract Interpretations in a Linear Logic-Based Setting

229

12. N.D. Jones and H. Søndergaard. A semantics-based framework for the abstract interpretation of Prolog. In S. Abramsky and C. Hankin, editors, Abstract Interpretation of Declarative Languages, pages 123–142. Ellis Horwood Ltd, 1987. 227 13. A. Mycroft. Completeness and predicate-based abstract interpretation. In Proc. ACM Conf. on Partial Evaluation and Program Manipulation (PEPM’93), pages 179–185, 1993. 216, 216, 216 14. C. Palamidessi. Algebraic properties of idempotent substitutions. In Proc. 17th Int. Colloq. on Automata, Languages and Programming (ICALP’90), LNCS 443, pages 386–399, 1990. 225 15. U.S. Reddy and S.N. Kamin. On the power of abstract interpretation. Computer Languages, 19(2):79–89, 1993. 216 16. K.I Rosenthal. Quantales and their Applications. Longman Scientiﬁc & Technical, 1990. 217, 219, 219, 222, 226 17. R.C. Sekar, P. Mishra, and I.V. Ramakrishnan. On the power and limitation of strictness analysis. J. ACM, 44(3):505–525, 1997. 216 18. B. Steﬀen. Optimal data ﬂow analysis via observational equivalence. In Proc. 14th Int. Symp. on Math. Found. of Comp. Sci. (MFCS’89), LNCS 379, pages 492–502, 1989. 216 19. D. Yetter. Quantales and (noncommutative) linear logic. J. Symbolic Logic, 55(1):41–64, 1990. 217, 219, 222

On the Power of Homeomorphic Embedding for Online Termination Michael Leuschel? Department of Computer Science, K.U. Leuven, Belgium Department of Electronics and Computer Science, University of Southampton, UK DIKU, University of Copenhagen, Denmark [email protected] Abstract. Recently well-quasi orders in general, and homeomorphic embedding in particular, have gained popularity to ensure the termination of program analysis, specialisation and transformation techniques. In this paper we investigate and clarify for the first time, both intuitively and formally, the advantages of such an approach over one using wellfounded orders. Notably we show that the homeomorphic embedding relation is strictly more powerful than a large class of involved well-founded approaches.

1

Introduction

The problem of ensuring termination arises in many areas of computer science and a lot of work has been devoted to proving termination of term rewriting systems (e.g. [7,8,9,37] and references therein) or of logic programs (e.g. [6,38] and references therein). It is also an important issue within all areas of program analysis, specialisation and transformation: one usually strives for methods which are guaranteed to terminate. One can basically distinguish between two kinds of techniques for ensuring termination: • static techniques, which prove or ensure termination of a program or process beforehand (i.e. off-line) without any kind of execution, and • online (or dynamic) techniques, which ensure termination of a process during its execution. (The process itself can of course be, e.g., performing a static analysis.) Static approaches have less information at their disposal but do not require runtime intervention (which might be impossible). Which of the two approaches is taken depends entirely on the application area. For instance, static termination analysis of logic programs [6,38] falls within the former context, while termination of program specialisation, transformation or analysis is often ensured in an online manner. This paper is primarily aimed at studying and improving online termination techniques. Let us examine the case of partial deduction [29,10,23] — an automatic technique for specialising logic programs. Henceforth we suppose some familiarity with basic notions in logic programming [3,28]. ?

Part of the work was done while the author was Post-doctoral Fellow of the Fund for Scientific Research - Flanders Belgium (FWO) and visiting DIKU, University of Copenhagen.

G. Levi (Ed.): SAS’98, LNCS 1503, pp. 230–245, 1998. c Springer-Verlag Berlin Heidelberg 1998

On the Power of Homeomorphic Embedding for Online Termination

231

Partial deduction based upon the Lloyd and Shepherdson framework [29] generates (possibly incomplete) SLDNF-trees for a set A of atoms. The specialised program is extracted from these trees by producing one clause for every nonfailing branch. The resolution steps within the SLDNF-trees — often referred to as unfolding steps — are those that have been performed beforehand, justifying the hope that the specialised program is more efficient. Now, to ensure termination of partial deduction two issues arise [10,34]. One is called the local termination problem, corresponding to the fact that each generated SLDNF-tree should be finite. The other is called the global termination problem, meaning that the set A should contain only a finite number of atoms. A similar classification can be done for most other program specialisation techniques (cf. e.g. [26]). Below we mainly use local termination to illustrate our concepts. (As shown in [34] the atoms in A can be structured into a global tree and methods similar to the one for local termination can be used to ensure global termination.) However, the discussions and contributions of the present paper are also (immediately) applicable in the context of analysis, specialisation and transformation techniques in general, especially when applied to computational paradigms, such as logic programming, constrained logic programming, conditional term rewriting, functional programming and functional & logic programming. For instance, abstract interpretation techniques usually analyse a set of abstract calls to which new call patterns are continuously added. One thus faces a problem very similar to global termination of partial deduction. Depth Bounds. One, albeit ad-hoc, way to solve the local termination problem is to simply impose an arbitrary depth bound. Such a depth bound is of course not motivated by any property, structural or otherwise, of the program or goal under consideration. In the context of local termination, the depth bound will therefore typically lead either to too little or too much unfolding. Determinacy. Another approach, often used in partial evaluation of functional programs [17] is to (only) expand a tree while it is determinate (i.e. it only has one non-failing branch). However, this approach can be very restrictive and in itself does not guarantee termination, as there can be infinitely failing determinate computations at specialisation time. Well-Founded Orders. Luckily, more refined approaches to ensure local termination exist. The first non-ad-hoc methods [5,33,32,31] in logic and [40,47] functional programming were based on well-founded orders, inspired by their usefulness in the context of static termination analysis. These techniques ensure termination, while at the same time allowing unfolding related to the structural aspect of the program and goal to be specialised, e.g., permitting the consumption of static input within the atoms of A.

232

M. Leuschel

Definition 1. (wfo) A (strict) partial order >S on a set S is an anti-reflexive, anti-symmetric and transitive binary relation on S × S. A sequence of elements s1 , s2 , . . . in S is called admissible wrt >S iff si > si+1 , for all i ≥ 1. We call >S a well-founded order (wfo) iff there is no infinite admissible sequence wrt >S To ensure local termination, one has to find a sensible well-founded order on atoms and then only allow SLDNF-trees in which the sequence of selected atoms is admissible wrt the well-founded order. If an atom that we want to select is not strictly smaller than its ancestors, we either have to select another atom or stop unfolding altogether. Example 1. Let P be the reverse program using an accumulator: rev ([], Acc, Acc) ← rev ([H|T ], Acc, Res) ← rev (T, [H|Acc], Res) A simple well-founded order on atoms of the form rev (t1 , t2 , t3 ) might be based on comparing the termsize (i.e., the number of function and constant symbols) of the first argument. We then define the wfo on atoms by: rev (t1 , t2 , t3 ) > rev (s1 , s2 , s3 ) iff term size(t2 ) > term size(s2 ). Based on that wfo, the goal ← rev ([a, b|T ], [], R) can be unfolded into the goal ← rev ([b|T ], [a], R) and further into ← rev (T, [b, a], R) because the termsize of the first argument strictly decreases at each step (even though the overall termsize does not decrease). However, ← rev (T, [b, a], R) cannot be further unfolded into ← rev (T 0 , [H 0 , b, a], R) because there is no such strict decrease. Much more elaborate techniques, which e.g. split the expressions into classes, use lexicographical ordering on subsequences of the arguments and even continuously refine the orders during the unfolding process, exist and we refer the reader to [5,33,32,31] for precise details. These works also present some further refinements on how to apply wfo’s, especially in the context of partial deduction. For instance, instead of requiring a decrease wrt every ancestor, one can only request a decrease wrt the covering ancestors, i.e. one only compares with the ancestor atoms from which the current atom descends (via resolution). Other refinements consist in allowing the wfo’s not only to depend upon the selected atom but on the context as well [32] or to ignore calls to non-recursive predicates. [32] also discusses a way to relax the condition of a “strict decrease” when refining a wfo. (Most of these refinements can also be applied to other approaches, notably the one we will present in the next section.) However, it has been felt by several researchers that well-founded orders are sometimes too rigid or (conceptually) too complex in an online setting. Recently, well-quasi orders have therefore gained popularity to ensure online termination of program manipulation techniques [4,41,42,25,26,11,18,1,20,46]. Unfortunately, this move to well-quasi orders has never been formally justified nor has the relation to well-founded approaches been investigated. We strive to do so in this paper and will actually prove that a rather simple well-quasi approach is strictly more powerful than a large class of involved well-founded approaches.

On the Power of Homeomorphic Embedding for Online Termination

2

233

Well-Quasi Orders and Homeomorphic Embedding

Formally, well-quasi orders can be defined as follows. Definition 2. (quasi order) A quasi order ≥S on a set S is a reflexive and transitive binary relation on S × S. Henceforth, we will use symbols like (possibly annotated by some subscript) to refer to strict partial orders and ≤, ≥ to refer to quasi orders. We will use either “directionality” as is convenient in the context. We also define an expression to be either a term (built-up from variables and function symbols of arity ≥ 0) or an atom (a predicate symbol applied to a, possibly empty, sequence of terms), and then treat predicate symbols as functors, but suppose that no confusion between function and predicate symbols can arise (i.e. predicate and function symbols are distinct). Definition 3. (wbr,wqo) Let ≤S be a binary relation on S × S. A sequence of elements s1 , s2 , . . . in S is called admissible wrt ≤S iff there are no i < j such that si ≤S sj . We say that ≤S is a well-binary relation (wbr) on S iff there are no infinite admissible sequences wrt ≤S . If ≤S is a quasi order on S then we also say that ≤S is a well-quasi order (wqo) on S. Observe that, in contrast to wfo’s, non-comparable elements are allowed within admissible sequences. An admissible sequence is sometimes called bad while a non-admissible one is called good. A well-binary relation is then such that all infinite sequences are good. There are several other equivalent definitions of well-binary relations and well-quasi orders. Higman [14] used an alternate definition of well-quasi orders in terms of the “finite basis property” (or “finite generating set” in [19]). A different (but also equivalent) definition of a wqo is: A quasi-order ≤V is a wqo iff for all quasi-orders V which contain ≤V (i.e. v≤V v 0 ⇒ vV v 0 ) the corresponding strict partial order ≺V is a wfo. This property has been exploited in the context of static termination analysis to dynamically construct well-founded orders from well-quasi ones and led to the initial use of wqo’s in the offline setting [7,8]. The use of well-quasi orders in an online setting has only emerged recently (it is mentioned, e.g., in [4] but also [41]) and has never been compared to well-founded approaches. There has been some comparison between wfo’s and wqo’s in the offline setting, e.g., in [37] it is argued that (for “simply terminating” rewrite systems) approaches based upon quasi-orders are less interesting than ones based upon a partial orders. In this paper we will show that the situation is somewhat reversed in an online setting. Furthermore, in the online setting, transitivity of a wqo is not really interesting and one can therefore drop this requirement, leading to the use of wbr’s. [24] contains some useful wbr’s which are not wqo’s. An interesting wqo is the homeomorphic embedding relation , which derives from results by Higman [14] and Kruskal [19]. It has been used in the context of term rewriting systems in [7,8], and adapted for use in supercompilation [45] in [42]. Its usefulness as a stop criterion for partial evaluation is also discussed and

234

M. Leuschel

advocated in [30]. Some complexity results can be found in [44] and [13] (also summarised in [30]). The following is the definition from [42], which adapts the pure homeomorphic embedding from [8] by adding a rudimentary treatment of variables. Definition 4. () The (pure) homeomorphic embedding relation on expressions is defined inductively as follows (i.e. is the least relation satisfying the rules): 1. X Y for all variables X, Y 2. s f (t1 , . . . , tn ) if s ti for some i 3. f (s1 , . . . , sn ) f (t1 , . . . , tn ) if ∀i ∈ {1, . . . , n} : si ti . The second rule is sometimes called the diving rule, and the third rule is sometimes called the coupling rule. When s t we also say that s is embedded in t or t is embedding s. By s t we denote that s t and t 6s. Example 2. The intuition behind the above definition is that A B iff A can be obtained from B by “striking out” certain parts, or said another way, the structure of A reappears within B. For instance we have p(a) p(f (a)) and indeed p(a) can be obtained from p(f (a)) by “striking out” the f . Observe that the “striking out” corresponds to the application of the diving rule 2 and that we even have p(a) p(f (a)). We also have, e.g., that: X X, p(X) p(f (Y )), p(X, X) p(X, Y ) and p(X, Y ) p(X, X). Proposition 1. The relation is a wqo on the set of expressions over a finite alphabet. For a complete proof, reusing Higman’s and Kruskal’s results [14,19] in a very straightforward manner, see, e.g., [23]. Extensions to infinite alphabets and improved treatment of variables can be found in [24]. To ensure, e.g., local termination of partial deduction, we have to ensure that the constructed SLDNF-trees are such that the selected atoms do not embed any of their ancestors (when using a well-founded order as in Example 1, we had to require a strict decrease at every step). If an atom that we want to select embeds one of its ancestors, we either have to select another atom or stop unfolding altogether. For example, based on , the goal ← rev ([a, b|T ], [], R) of Example 1 can be unfolded into ← rev ([b|T ], [a], R) and further into ← rev (T, [b, a], R) as rev ([a, b|T ], [], R) 6rev ([b|T ], [a], R), rev ([a, b|T ], [], R) 6rev (T, [b, a], R) and rev ([b|T ], [a], R) 6 rev (T, [b, a], R). However, ← rev (T, [b, a], R) cannot be further unfolded into ← rev (T 0 , [H 0 , b, a], R) as rev (T, [b, a], R) rev (T 0 , [H 0 , b, a], R). Observe that, in contrast to Example 1, we did not have to choose how to measure which arguments. We further elaborate on the inherent flexibility of in the next section. The homeomorphic embedding relation is also useful for handling structures other than expressions. It has, e.g., been successfully applied in [25,23,26] to detect (potentially) non-terminating sequences of characteristic trees. Also, seems to have the desired property that very often only “real” loops are detected and that they are detected at the earliest possible moment (see [30]).

On the Power of Homeomorphic Embedding for Online Termination

3

235

Comparing wbr’s and wfo’s

3.1

General Comparison

It follows from Definitions 1 and 3 that if ≤V is a wqo then t3 . Let, e.g., s < t denote that s is strictly more general than t. Then < is a wfo (see below) but p(X, X, a) 6> p(X, Z, b) and p(X, Z, b) 6> p(X, Y, a) even though p(X, X, a) > p(X, Y, a). Let us now examine the power of one particular wqo, the earlier defined . 3.2

Homeomorphic Embedding and Monotonic wfo’s

The homeomorphic embedding relation is very flexible. It will for example, when applied to the sequence of covering ancestors, permit the full unfolding of most terminating Datalog programs, the quicksort or even the mergesort program when the list to be sorted is known (the latter poses problems to some static termination analysis methods [38,27]; for some experiments see Appendix A). Also, the produce-consume example from [31] requires rather involved techniques (considering the context) to be solved by wfo’s. Again, this example poses no problem to (cf. Appendix A). The homeomorphic embedding is also very powerful in the context of metaprogramming. Notably, it has the ability to “penetrate” layers of (non-ground) meta-encodings (cf. also Appendix A). For instance, will admit the following sequences (where Example 1 is progressively wrapped into “vanilla” metainterpreters counting resolution steps and keeping track of the selected predicates respectively): Sequence rev ([a, b|T ], [], R) ; rev ([b|T ], [a], R) solve(rev ([a, b|T ], [], R), 0) ; solve(rev ([b|T ], [a], R), s(0)) solve 0 (solve(rev ([a, b|T ], [], R), 0), []) ; solve 0 (solve(rev ([b|T ], [a], R), s(0)), [rev])

Again, this is very difficult for wfo’s and requires refined and involved techniques (of which to our knowledge no implementation exists). We have intuitively demonstrated the usefulness of and that it is often more flexible than wfo’s. But can we prove some “hard” results? It turns out that we can and we now establish that — in the online setting — is strictly more generous than a large class of refined wfo’s.

236

M. Leuschel

Definition 5. A well-founded order ≺ on expressions is said to be monotonic iff the following rules hold: 1. X 6Y for all variables X, Y , 2. s 6f (t1 , . . . , tn ) whenever f is a function symbol and s 6ti for some i and 3. f (s1 , . . . , sn ) 6f (t1 , . . . , tn ) whenever ∀i ∈ {1, . . . , n} : si 6fi . Note the similarity of structure with the definition of (but, contrary to , 6does not have to be the least relation satisfying the rules). This similarity of structure will later enable us to prove that any sequence admissible wrt ≺ must also be admissible wrt (by showing that st ⇒ s 6t). Also observe that point 2 need not hold for predicate symbols and that point 3 implies that c 6c for all constant and proposition symbols c. Finally, there is a subtle difference between monotonic wfo’s as of Definition 5 and wfo’s which possess the replacement property (such orders are called rewrite orders in [37] and monotonic in [7]). More on that below. Similarly, we say that a norm k.k (i.e. a mapping from expressions to IN ) is said to be monotonic iff the associated wfo ≺k.k is monotonic (t1 ≺k.k t2 iff kt1 k < kt2 k). For instance the termsize norm (see below) is trivially monotonic. More generally, any semi-linear norm of the following form is monotonic: Proposition 2.P Let the norm k.k : Expr →IN be defined by: n • ktk = cf + i=1 cf,i kti k if t = f (t1 , . . . , tn ), n ≥ 0 • ktk = cv otherwise (i.e. t is a variable) Then k.k is monotonic if all coefficients cv , cf , cf,i are ≥ 0 and cf,i ≥ 1 for all function symbols f of arity ≥ 1 (but not necessarily for all predicate symbols). Proof. As < on IN is total we have that s 6> t is equivalent to s ≤ t. The proof proceeds by induction on the structure of the expressions and examines every rule of Definition 5 separately: 1. X ≤ Y for all variables X, Y this trivially holds as we use the same constant cv for all variables. 2. s ≤ f (t1 , . . . , tn ) whenever s ≤ ti for some i This holds trivially if all coefficients are ≥ 0 and if cf,i ≥ 1. This is verified, as the rule only applies if f is a function symbol. 3. f (s1 , . . . , sn ) ≤ f (t1 , . . . , tn ) whenever ∀i ∈ {1, . . . , n} : si ≤ fi This holds trivially, independently of whether f is a function or predicate symbol, as all coefficients are positive (and the same coefficient is applied to si and ti ).

By taking cv = 0 and cf,i = cf = 1 for all f we get the termsize norm k.kts , which by the above proposition is thus monotonic. Also, by taking cv = 1 and cf,i = cf = 1 for all f we also get a monotonic norm, counting symbols. Finally, a linear norm can always be obtained [38] by setting cv = 0, cf,i = 1 and cf ∈ IN for all f . Thus, as another corollary of the above, any linear norm is monotonic. Proposition 3. Let k.k1 , . . . , k.kk be monotonic norms satisfying Proposition 2. Then the lexicographical ordering ≺lex defined by s ≺lex t iff ∃i ∈ {1, . . . , k} such that kski < ktki and ∀j ∈ {1, . . . , i − 1}: kskj = ktkj is a monotonic wfo.

On the Power of Homeomorphic Embedding for Online Termination

237

Proof. By standard results (see, e.g., [8]) we know that ≺lex is a wfo (as < is a wfo on IN ). We will prove that ≺lex satisfies all the rules of Definition 5. 1. First, rule 1 is easy as kXki = kY ki for all i and variables X, Y and therefore we never have X ≺lex Y . 2. Before examining the other rules, let us note that s 6lex t is equivalent to saying that either a) ∀j ∈ {1, . . . , k} kskj = ktkj or b) there exists an j ∈ {1, . . . , k} such that kskj < ktkj and ∀l ∈ {1, . . . , j − 1}: kskl = ktkl . Let us now examine rule 2 of Definition 5. We have to prove that whenever s 6lex ti the conclusion of the rule holds. Let us first examine case a) for s 6lex ti . We have kskj = kti kj and thus we know that kskj ≤ kf (t1 , . . . , tn )kj by monotonicity of k.kj (as < on IN is total we have that s 6> t is equivalent to s ≤ t). As this holds for all k.kj we cannot have sj lex f (t1 , . . . , tn ). Let us now examine the second case b) for s 6lex ti . Let j ∈ {1, . . . , k} such that kskj < kti kj and ∀l ∈ {1, . . . , j − 1}: kskl = kti kl . For all l we can deduce as above that kskl ≤ kf (t1 , . . . , tn )kl . However, we still have to prove that kskj < kf (t1 , . . . , tn )kj . By monotonicity of k.kj we only know that kskj ≤ kf (t1 , . . . , tn )kj . But we can also apply monotonicity of k.kj to deduce that kti kj ≤ kf (t1 , . . . , tn )kj and hence we can infer the desired property (as kskj < kti kj ). 3. Now, for rule 3 we have to prove that whenever si 6lex ti for all i ∈ {1, . . . , n} the conclusion of the rule holds. There are again two cases. a) We can have ksi kj = kti kj for all i, j. By monotonicity of each k.kj we know that kf (s1 , . . . , sn )kj ≤ kf (t1 , . . . , tn )kj for all j ∈ {1, . . . , k}. Hence, we cannot have f (s1 , . . . , sn ) lex f (t1 , . . . , tn ). b) In the other case we know that there must be a value j 0 ∈ {1, . . . , k} such that for some i: ksi kj 0 < kti kj 0 and ∀l ∈ {1, . . . , j 0 − 1}: ksi kl = kti kl . I.e., by letting j denote the minimum value j 0 for which this holds, we know that for some i: ksi kj < kti kj and for all i0 : ∀l ∈ {1, . . . , j}: ksi0 kl ≤ kti0 kl . By monotonicity of each k.kl we can therefore deduce that ∀l ∈ {1, . . . , j}: kf (s1 , . . . , sn )kl ≤ kf (t1 , . . . , tn )kl . We can also deduce by monotonicity of k.kj that kf (s1 , . . . , sn )kj ≤ kf (t1 , . . . , tn )kj . We can even deduce that kf (s1 , . . . , sn )kj ≤ kf (t1 , . . . , ti−1 , si , ti+1 , . . . , tn )kj ≤ kf (t1 , . . . , tn )kj . Now, we just have to prove that: kf (t1 , . . . , ti−1 , si , ti+1 , . . . , tn )kj < kf (t1 , . . . , tn )kj in order to affirm that kf (s1 , . . . , sn )kj 6lex kf (t1 , . . . , tn )kj . This does not hold for all monotonic norms, but as we know that k.kj satisfies Proposition 2, this can be deduced by the fact that the coefficient cf,i in k.kj must be ≥ 1.

It is important that the norms k.k1 , . . . , k.kk satisfy Proposition 2. Otherwise, a counterexample would be as follows. Let kak1 = 1, kbk1 = 2 and kf (a)k1 = kf (b)k1 = 5. Also let kak2 = 2, kbk2 = 1 and kf (a)k2 = 3, kf (b)k2 = 2. Now we have a ≺lex b, i.e. a 6lex b, but also f (a) lex f (b) and condition 3 of monotonicity for ≺lex is violated. One could make Proposition 3 slightly more general, but the current version is sufficient to show the desired result, namely that most of the wfo’s used in online practice are actually monotonic. For example almost all of the refined wfo’s defined in [5,33,32,31] are monotonic: • Definitions 3.4 of [5], 3.2 of [33] and 2.14 of [32] all sum up the number of function symbols (i.e. termsize) of a subset of the argument positions of

238

M. Leuschel

atoms. These wfo’s are therefore immediately covered by Proposition 2. The algorithms only differ in the way of choosing the positions to measure. The early algorithms simply measure the input positions, while the later ones dynamically refine the argument positions to be measured (but which are still measured using the termsize norm). • Definitions 3.2 of [32] as well as 8.2.2 of [31] use the lexicographical order on the termsizes of some selected argument positions. These wfo’s are therefore monotonic as a corollary to Propositions 2 and 3. The only non-monotonic wfo in that collection of articles is the one defined specifically for metainterpreters in Definition 3.4 of [5] (also in Sect. 8.6 of [31]) which uses selector functions to focus on subterms to be measured. We will return to this approach below. Also, as already mentioned, some of the techniques in [32,31] (in Sects. 3.4 and 8.2.4 respectively) do not require the whole sequence to be admissible wrt a unique wfo, i.e. one can split up a sequence into a (finite) number of subsequences and apply different (monotonic) wfo’s on these subsequences. Similar refinements can also be developed for wqo’s and the formal study of these refinements are (thus) not the main focus of the paper. Before showing that is strictly more powerful than the union of all monotonic wfo’s, we adapt the class of simplification orderings from term rewriting systems. It will turn out that the power of this class is also subsumed by . Definition 6. A simplification ordering is a wfo ≺ on expressions which satisfies 1. s ≺ t ⇒ f (t1 , . . . , s, . . . , tn ) ≺ f (t1 , . . . , t, . . . , tn ) (replacement property), 2. t ≺ f (t1 , . . . , t, . . . , tn ) (subterm property) and 3. s ≺ t ⇒ sσ ≺ tγ for all variable only substitutions σ and γ (invariance under variable replacement). The third rule of the above definition is new wrt term-rewriting systems and implies that all variables must be treated like a unique new constant. It turns out that a lot powerful wfo’s are simplification orderings [7,37]: recursive path ordering, Knuth-Bendix ordering or lexicographic path ordering, to name just a few. However, not all wfo’s of Proposition 2 are simplification orderings: e.g., for cf = 0, ca = 1 we have kak = kf (a)k and the subterm property does not hold (for the associated wfo). In addition, Proposition 2 allows a special treatment for predicates. On the other hand, there are wfo’s which are simplification orderings but are not monotonic according to Definition 5. Proposition 4. Let ≺ be a wfo on expressions. Then any admissible sequence wrt ≺ is also an admissible sequence wrt if ≺ is a) monotonic or if it is b) a simplification ordering. Proof. First, let us observe that for a given wfo ≺ on expressions, any admissible sequence wrt ≺ is also an admissible sequence wrt iff s t ⇒ s 6t. Indeed (⇒), whenever s t then s 6t, and this trivially implies (by transitivity of ≺) that any sequence not admissible wrt cannot be strictly descending wrt ≺. On the other hand

On the Power of Homeomorphic Embedding for Online Termination

239

(⇐), let us assume that for some s and t s t but s t. This means that the sequence s, t is admissible wrt but not wrt and we have a contradiction. a) The proof that for a monotonic wfo ≺ we have st ⇒ s 6t is by straightforward induction on the structure of s and t. The only “tricky” aspect is that the second rule for monotonicity only holds if f is a function symbol. But if f is a predicate symbol, then s ti cannot hold because we supposed that predicate and function symbols are distinct. b) If ≺ is a simplification ordering then we can apply Lemma 3.3 of [37] to deduce that ≺ is the superset of the strict part of (i.e., ≺⊇ ). Let us examine the two possibilities for s t. First, we can have s t. In that case we can deduce s ≺ t and thus s 6t. Second, we can have s t and t s. In that case s and t are identical, except for the variables. If we now take the substitution σ which assigns all variables in s and t to a unique variable we have sσ = tσ, i.e., sσ 6tσ. This means that s t cannot hold (because is invariant under variable replacement).

This means that the admissible sequences wrt are a superset of the union of all admissible sequences wrt simplification orderings and monotonic wfo’s. In other words, no matter how much refinement we put into an approach based upon monotonic wfo’s or upon simplification orderings we can only expect to approach in the limit. But by a simple example we can even dispel that hope. Example 3. Take the sequence δ = f (a), f (b), b, a. This sequence is admissible wrt as f (a) 6f (b), f (a) 6b, f (a) 6a, f (b) 6b, f (b) 6a and a 6b. However, there is no monotonic wfo ≺ which admits this sequence. More precisely, to admit δ we must have f (a) f (b) as well as b a, i.e. a 6b. But this violates rule 3 of Definition 5 and ≺ cannot be monotonic. This also violates rule 1 of Definition 6 and ≺ cannot be a simplification ordering. These new results relating to monotonic wfo’s shed light on ’s usefulness in the context of ensuring online termination. But of course the admissible sequences wrt are not a superset of the union of all admissible sequences wrt any wfo.1 For instance the list-length norm k.kllen is not monotonic, and indeed we have for t1 = [1, 2, 3] and t2 = [[1, 2, 3], 4] that kt1 kllen = 3 > kt2 kllen = 2 although t1 t2 . So there are sequences admissible wrt list-length but not wrt . The reason is that k.kllen in particular and nonmonotonic wfo’s in general can completely ignore certain parts of the term, while will always inspect that part. E.g., if we have s f (. . . t . . .) and ignores the subterm t then it will also be true that s f (. . . s . . .) while s f (. . . s . . .),2 i.e. the sequence s, f (. . . s . . .) is admissible wrt but not wrt . For that same reason the wfo’s for metainterpreters defined in Definition 3.4 of [5] mentioned above are not monotonic, as they are allowed to completely focus on subterms, fully ignoring other subterms. However, automation of that technique is not addressed in [5]. E.g., for this wfo one cannot immediately apply 1 2

Otherwise could not be a wqo, as all finite sequences without repetitions are admissible wrt some wfo (map last element to 1, second last element to 2, . . . ). Observe that if f is a predicate symbols then f (. . . s . . .) is not a valid expression, which enabled us to ignore arguments to predicates in e.g. Proposition 2.

240

M. Leuschel

the idea of continually refining the measured subterms, because otherwise one might simply plunge deeper and deeper into the terms and termination would not be ensured. A step towards an automatic implementation is presented in Sect. 8.6 of [31] and it will require further work to formally compare it with wqo-based approaches and whether the ability to completely ignore certain parts of an expression can be beneficial for practical programs. But, as we have seen earlier, alone is already very flexible for metainterpreters, even more so when combined with characteristic trees [26] (see also [46]). Of course, for any wfo (monotonic or not) one can devise a wbr (cf. Lemma 1) which has the same admissible sequences. Still there are some feats that are easily attained, even by using , but which cannot be achieved by a wfo approach (monotonic or not). Take the sequences S1 = p([], [a]), p([a], []) and S2 = p([a], []), p([], [a]). Both of these sequences are admissible wrt This illustrates the flexibility of using well-quasi orders compared to well-founded ones in an online setting, as there exists no wfo (monotonic or not) which will admit both these sequences. It however also illustrates why, when using a wqo in that way, one has to compare with every predecessor state of a process. Otherwise one can get infinite derivations of the form p([a], []) → p([], [a]) → p([a], []) →. . . .3 Short Note on Offline Termination. This example also shows why (or well-quasi orders in general) cannot be used directly for static termination analysis. Let us explain what we mean. Take, e.g., a program containing the clauses C1 = p([a], []) ← p([], [a]) and C2 = p([], [a]) ← p([a], []). Then, in both cases the body is not embedding the head, but still the combination of the two clauses leads to a non-terminating program. However, can be used to construct wellfounded orders for static termination analysis. Take the clause C1 . The head and the body are incomparable according to . So, we can simply extend by stating that p([a], []) p([], a) (thus making the head strictly larger than the body atom). As already mentioned, for any extension ≤ of a wqo we have that < is a wfo. Thus we know that the program just consisting of C1 is terminating. If we now analyse C2 we have that, according to the extended wqo, the body is strictly larger than the head and (luckily) we cannot prove termination (i.e. there is no way of extending so that for both C1 and C2 the head is strictly larger than the body).

4

Discussion and Conclusion

Of course is not the ultimate relation for ensuring online termination. On its own in the context of local control of partial deduction, will sometimes allow 3

When using a wfo one has to compare only to the closest predecessor [32], because of the transitivity of the order and the strict decrease enforced at each step. However, wfo’s are usually extended to incorporate variant checking and then require inspecting every predecessor anyway (though only when there is no strict weight decrease, see, e.g., [31,32]).

On the Power of Homeomorphic Embedding for Online Termination

241

too much unfolding than desirable for efficiency concerns (i.e. more unfolding does not always imply a better specialised program) and we refer to the solutions developed in, e.g., [26,18]. For some applications, remains too restrictive. In particular, it does not always deal satisfactorily with fluctuating structure (arising, e.g., for certain meta-interpretation tasks) [46]. The use of characteristic trees [23,26] remedies this problem to some extent, but not totally. A further step towards a solution is presented in [46]. In that light, it might be of interest to study whether the extensions of the homeomorphic embedding relation proposed in [39] and [21] (in the context of static termination analysis of term rewrite systems) can be useful in an online setting. As shown in [24] the treatment of variables of is rather rudimentary and several ways to remedy this problem are presented. In summary, we have shed new light on the relation between wqo’s and wfo’s and have formally shown (for the first time) why wqo’s are more interesting than wfo’s for ensuring termination in an online setting (such as program specialisation or analysis). We have illustrated the inherent flexibility of and proved that, despite its simplicity, it is strictly more generous than the class of monotonic wfo’s. As all the wfo’s used for automatic online termination (so far) are actually monotonic, this formally establishes the interest of in that context. Acknowledgements. I would like to thank Danny De Schreye, Robert Gl¨uck, Jesper Jørgensen, Bern Martens, Maurizio Proietti, Jacques Riche and Morten Heine Sørensen for all the discussions, comments and joint research which led to this paper. Anonymous referees as well as Bern Martens provided extremely useful feedback on this paper.

References 1. M. Alpuente, M. Falaschi, P. Juli´ an, and G. Vidal. Spezialisation of lazy functional logic programs. In Proceedings PEPM’97, pages 151–162, Amsterdam, The Netherlands, 1997. ACM Press. 2. M. Alpuente, M. Falaschi, and G. Vidal. Narrowing-driven partial evaluation of functional logic programs. In H. Riis Nielson, editor, Proceedings ESOP’96, LNCS 1058, pages 45–61. Springer-Verlag, 1996. 3. K. R. Apt. Introduction to logic programming. In J. van Leeuwen, editor, Handbook of Theoretical Computer Science, chapter 10, pages 495–574. North-Holland Amsterdam, 1990. 4. R. Bol. Loop checking in partial deduction. The Journal of Logic Programming, 16(1&2):25–46, 1993. 5. M. Bruynooghe, D. De Schreye, and B. Martens. A general criterion for avoiding infinite unfolding during partial deduction. New Generation Computing, 11(1):47– 79, 1992. 6. D. De Schreye and S. Decorte. Termination of logic programs: The never ending story. The Journal of Logic Programming, 19 & 20:199–260, May 1994. 7. N. Dershowitz. Termination of rewriting. Journal of Symbolic Computation, 3:69– 116, 1987.

242

M. Leuschel

8. N. Dershowitz and J.-P. Jouannaud. Rewrite systems. In J. van Leeuwen, editor, Handbook of Theoretical Computer Science, Vol. B, pages 243–320. Elsevier, MIT Press, 1990. 9. N. Dershowitz and Z. Manna. Proving termination with multiset orderings. Communications of the ACM, 22(8):465–476, 1979. 10. J. Gallagher. Tutorial on specialisation of logic programs. In Proceedings PEPM’93, pages 88–98. ACM Press, 1993. 11. R. Gl¨ uck, J. Jørgensen, B. Martens, and M. H. Sørensen. Controlling conjunctive partial deduction of definite logic programs. In H. Kuchen and S. Swierstra, editors, Proceedings PLILP’96, LNCS 1140, pages 152–166, Aachen, Germany, September 1996. Springer-Verlag. 12. R. Gl¨ uck and M. H. Sørensen. A roadmap to supercompilation. In O. Danvy, R. Gl¨ uck, and P. Thiemann, editors, Proceedings of the 1996 Dagstuhl Seminar on Partial Evaluation, LNCS 1110, pages 137–160, Schloß Dagstuhl, 1996. SpringerVerlag. 13. J. Gustedt. Algorithmic Aspects of Ordered Structures. PhD thesis, Technische Universit¨ at Berlin, 1992. 14. G. Higman. Ordering by divisibility in abstract algebras. Proceedings of the London Mathematical Society, 2:326–336, 1952. 15. G. Huet. Confluent reductions: Abstract properties and applications to term rewriting systems. Journal of the ACM, 27(4):797–821, 1980. 16. N. D. Jones. An introduction to partial evaluation. ACM Computing Surveys, 28(3):480–503, September 1996. 17. N. D. Jones, C. K. Gomard, and P. Sestoft. Partial Evaluation and Automatic Program Generation. Prentice Hall, 1993. 18. J. Jørgensen, M. Leuschel, and B. Martens. Conjunctive partial deduction in practice. In J. Gallagher, editor, Proceedings LOPSTR’96, LNCS 1207, pages 59– 82, Stockholm, Sweden, August 1996. Springer-Verlag. 19. J. B. Kruskal. Well-quasi ordering, the tree theorem, and Vazsonyi’s conjecture. Transactions of the American Mathematical Society, 95:210–225, 1960. 20. L. Lafave and J. Gallagher. Constraint-based partial evaluation of rewriting-based functional logic programs. In N. Fuchs, editor, Proceedings LOPSTR’97, LNCS, Leuven, Belgium, July 1997. Springer-Verlag. 21. P. Lescanne. Well rewrite orderings and well quasi-orderings. Technical Report No 1385, INRIA-Lorraine, France, January 1991. 22. M. Leuschel. The ecce partial deduction system and the dppd library of benchmarks. 1996–1998. Obtainable via http://www.cs.kuleuven.ac.be/˜dtai. 23. M. Leuschel. Advanced Techniques for Logic Program Specialisation. PhD thesis, K.U. Leuven, May 1997. Accessible via http://www.cs.kuleuven.ac.be/˜michael. 24. M. Leuschel. Homeomorphic Embedding for Online Termination. Technical Report, Department of Electronics and Computer Science, University of Southampton, 1998. 25. M. Leuschel and B. Martens. Global control for partial deduction through characteristic atoms and global trees. In O. Danvy, R. Gl¨ uck, and P. Thiemann, editors, Proceedings of the 1996 Dagstuhl Seminar on Partial Evaluation, LNCS 1110, pages 263–283, Schloß Dagstuhl, 1996. Springer-Verlag. 26. M. Leuschel, B. Martens, and D. De Schreye. Controlling generalisation and polyvariance in partial deduction of normal logic programs. ACM Transactions on Programming Languages and Systems, 20(1):208–258, January 1998.

On the Power of Homeomorphic Embedding for Online Termination

243

27. N. Lindenstrauss, Y. Sagiv, and A. Serebrenik. Unfolding the mystery of mergesort. In N. Fuchs, editor, Proceedings LOPSTR’97, LNCS, Leuven, Belgium, July 1997. Springer-Verlag. 28. J. W. Lloyd. Foundations of Logic Programming. Springer-Verlag, 1987. 29. J. W. Lloyd and J. C. Shepherdson. Partial evaluation in logic programming. The Journal of Logic Programming, 11(3& 4):217–242, 1991. ´ 30. R. Marlet. Vers une Formalisation de l’Evaluation Partielle. PhD thesis, Universit´e de Nice - Sophia Antipolis, December 1994. 31. B. Martens. On the Semantics of Meta-Programming and the Control of Partial Deduction in Logic Programming. PhD thesis, K.U. Leuven, February 1994. 32. B. Martens and D. De Schreye. Automatic finite unfolding using well-founded measures. The Journal of Logic Programming, 28(2):89–146, August 1996. 33. B. Martens, D. De Schreye, and T. Horv´ ath. Sound and complete partial deduction with unfolding based on well-founded measures. Theoretical Computer Science, 122(1–2):97–117, 1994. 34. B. Martens and J. Gallagher. Ensuring global termination of partial deduction while allowing flexible polyvariance. In L. Sterling, editor, Proceedings ICLP’95, pages 597–613, Kanagawa, Japan, June 1995. MIT Press. 35. J. Martin. Sonic partial deduction. Technical Report, Department of Electronics and Computer Science, University of Southampton, 1998. 36. P.-A. Melli`es. On a duality between Kruskal and Dershowitz theorems. In K. G. Larsen, editor, Proceedings of ICALP’98, LNCS, Aalborg, Denmark, 1998. Springer-Verlag. 37. A. Middeldorp and H. Zantema. Simple termination of rewrite systems. Theoretical Computer Science, 175(1):127–158, 1997. 38. L. Pl¨ umer. Termination Proofs for Logic Programs. LNCS 446. Springer-Verlag, 1990. 39. L. Puel. Using unavoidable set of trees to generalize Kruskal’s theorem. Journal of Symbolic Computation, 8:335–382, 1989. 40. E. Ruf. Topics in Online Partial Evaluation. PhD thesis, Stanford University, March 1993. 41. D. Sahlin. Mixtus: An automatic partial evaluator for full Prolog. New Generation Computing, 12(1):7–51, 1993. 42. M. H. Sørensen and R. Gl¨ uck. An algorithm of generalization in positive supercompilation. In J. W. Lloyd, editor, Proceedings ILPS’95, pages 465–479, Portland, USA, December 1995. MIT Press. 43. M. H. Sørensen, R. Gl¨ uck and N.D. Jones. A positive supercompiler. Journal of Functional Programming, 1996. 44. J. Stillman. Computational Problems in Equational Theorem Proving. PhD thesis, State University of New York at Albany, 1988. 45. V. F. Turchin. The concept of a supercompiler. ACM Transactions on Programming Languages and Systems, 8(3):292–325, 1986. 46. W. Vanhoof and B. Martens. To parse or not to parse. In N. Fuchs, editor, Proceedings LOPSTR’97, LNCS, Leuven, Belgium, July 1997. Springer-Verlag. 47. D. Weise, R. Conybeare, E. Ruf, and S. Seligman. Automatic online partial evaluation. In Proceedings of the Conference on Functional Programming Languages and Computer Architectures, LNCS 523, pages 165–191, Harvard University, 1991. Springer-Verlag.

244

A

M. Leuschel

Small Experiments with the ecce System

The purpose of this appendix is to illustrate the flexibility which the homeomorphic embedding relation provides straight “out of the box” (other more intricate well-quasi orders, like the one used by Mixtus [41], can handle some of the examples below as well). For that we experiment with the ecce partial deduction system [22] using an unfolding rule based on which allows the selection of either determinate literals or left-most literals within a goal, given that no covering ancestor [5] is embedded (via ). To ease readability, the specialised programs are sometimes presented in unrenamed form. First, let us take the mergesort program, which is somewhat problematic for a lot of static termination analysis methods [38,27]. mergesort([],[]). mergesort([X],[X]). mergesort([X,Y|Xs],Ys) :split([X,Y|Xs],X1s,X2s), mergesort(X1s,Y1s),mergesort(X2s,Y2s), merge(Y1s,Y2s,Ys). split([],[],[]). split([X|Xs],[X|Ys],Zs) :- split(Xs,Zs,Ys). merge([],Xs,Xs). merge(Xs,[],Xs). merge([X|Xs],[Y|Ys],[X|Zs]) :- X =< Y, merge(Xs,[Y|Ys],Zs). merge([X|Xs],[Y|Ys],[Y|Zs]) :- X>Y, merge([X|Xs],Ys,Zs).

The partial evaluation query: ?- mergesort([3,1,2],X). As the following resulting specialised program shows, homeomorphic embedding allowed the full unfolding of mergesort: mergesort([3,1,2],[1,2,3]). It took ecce less than 0.5 s on a Sparc Classic to produce the above program (including post-processing and writing to file). We can even achieve this same feat even if we interpose one or more levels of metainterpretation! Take a vanilla solve metainterpreter with mergesort at the object-level: solve([]). solve([H|T]) :claus(H,Bdy),solve(Bdy),solve(T). claus(mergesort([],[]), []). claus(mergesort([X],[X]), []).

On the Power of Homeomorphic Embedding for Online Termination

245

claus(mergesort([X,Y|Xs],Ys), [split([X,Y|Xs],X1s,X2s), mergesort(X1s,Y1s),mergesort(X2s,Y2s), merge(Y1s,Y2s,Ys) ]). claus(split([],[],[]), []). claus(split([X|Xs],[X|Ys],Zs) , [ split(Xs,Zs,Ys) ]). claus(merge([],Xs,Xs), []). claus(merge(Xs,[],Xs), []). claus(merge([X|Xs],[Y|Ys],[X|Zs]) , [ X =< Y, merge(Xs,[Y|Ys],Zs) ]). claus(merge([X|Xs],[Y|Ys],[Y|Zs]) , [ X>Y, merge([X|Xs],Ys,Zs)]). claus(’=’(X,Y),[]) :- X > Y. mergesort_test(X) :- solve([mergesort([3,1,2],X)]).

The partial evaluation query: ?- mergesort_test(X). Again homeomorphic embedding allowed the full unfolding: mergesort_test([1,2,3]). It took ecce 2.86 s on a Sparc Classic to produce the above program (including post-processing and writing to file). The following example is taken from [31]. produce([],[]). produce([X|Xs],[X|Ys]) :- produce(Xs,Ys). consume([]). consume([X|Xs]) :- consume(Xs).

The partial evaluation query: ?- produce([1,2|X],Y), consume(Y). To solve it in the setting of unfolding based upon wfo’s one needs both partition based measure functions as well as taking the context into account. The same adequate unfolding can simply be achieved by based determinate unfolding, as illustrated by the following specialised program derived by ecce (default settings): produce_conj__1([],[1,2]). produce_conj__1([X1|X2],[1,2,X1|X3]) :produce_conj__2(X2,X3). produce_conj__2([],[]). produce_conj__2([X1|X2],[X1|X3]) :produce_conj__2(X2,X3).

Analysis of Imperative Programs through Analysis of Constraint Logic Programs Julio C. Peralta, John P. Gallagher, and H¨ useyin Saˇ glam University of Bristol Dept. of Computer Science Merchant Venturers Building Woodland Rd., Bristol, U.K. BS8 1UB fax: +44-117-9545208 e-mail: {jperalta, john}@cs.bris.ac.uk, [email protected]

Abstract. In this paper a method is proposed for carrying out analysis of imperative programs. We achieve this by writing down the language semantics as a declarative program (a constraint logic program, in the approach shown here). We propose an effective style of writing operational semantics suitable for analysis which we call one-state small-step semantics. Through controlled partial evaluation we are able to generate residual programs where the relationship between imperative statements and predicates is straightforward. Then we use a static analyser for constraint logic programs on the residual program. The analysis results are interpreted through program points associating predicates in the partially evaluated interpreter to statements in its corresponding imperative program. We used an analyser that allows us to determine linear equality, inequality and disequality relations among the variables of a program without user-provided inductive assertions or human interaction. The proposed method intends to serve as a framework for the analysis of programs in any imperative language. The tools required are a partial evaluator and a static analyser for the declarative language.

Keywords: Partial Evaluation, Constraint Logic Programming, Operational Semantics, Imperative Program Analysis.

1

Introduction

Program semantics has long been used as a formal basis for program manipulation. By this we mean that the formal semantics of a programming language is written down in some mathematical framework, which is then used to establish program properties such as termination, correctness with respect to specifications, or the correctness of program transformations. In the case of imperative languages the gap between semantics and programs is greater than in the case of declarative languages. Perhaps for this reason, semantics-based tools for declarative languages, such as abstract interpreters and partial evaluators, have been the subject of more intense research than similar tools for imperative languages. G. Levi (Ed.): SAS’98, LNCS 1503, pp. 246–261, 1998. c Springer-Verlag Berlin Heidelberg 1998

Analysis of Imperative Programs through Analysis

247

The aim of this paper is to show how to transfer results achieved in declarative languages to imperative languages. The approach is to implement the semantics of imperative languages carefully in a declarative language for which analysis and transformation tools exist. There exist several kinds of semantics for imperative languages. The choice of which one is better suited for the particular application is a subject of ongoing research [3]. We shall see that for the purposes of this paper a variant of operational semantics will suffice. Logic programming appears to be an elegant paradigm for imperative language semantics implementation. When written carefully, the semantics can be regarded as an interpreter for the imperative language. Appropriate techniques for implementing semantics are discussed in Sect. 2. Partial evaluation of the interpreter with respect to an imperative program yields an equivalent declarative program. By so doing we open up the possibility of applying well-developed techniques for analysis and transformation of constraint logic programs to imperative programs as well. Nevertheless, it is not clear how to relate the results of such analysis and/or transformation back to the original imperative program. In order to obtain a correspondence between the imperative program and its corresponding declarative program some tuning of the partial evaluator is needed. Otherwise, the partial evaluator may remove important information needed to relate imperative statements and variables with their declarative counterpart. Such tuning involves selecting among the predicates of the semanticsbased interpreter those we want to be defined in the residual program. Hence, we choose predicates from the semantics-based interpreter that relate directly to the meaning of the statements in the imperative program to be partially evaluated. As a result we get one predicate for each statement of the imperative program, thus highlighting the correspondence between imperative statements and predicates in the residual program. In this paper we propose a method that intends to serve as a framework for the analysis of programs in any imperative language, by writing down its semantics as a declarative program (a constraint logic program, in the approach shown here). The tools required are a partial evaluator and a static analyser for the declarative language. Section 2 considers the overall problem of encoding semantics in a logic programming language, in a form suitable for analysis. Section 3 provides some remarks on the implementation of the operational semantics for a small imperative language 1 . In Sect. 4 we show the partial evaluation process and give an example. Section 5 illustrates how to relate the results of the analysis back to its imperative program in a systematic way. Finally, in Sects. 6 and 7 we discuss related work, and state our final remarks and some trends for future work.

1

Currently we are writing the semantics for Java in a similar style

248

2

J.C. Peralta, J.P. Gallagher, and H. Saˇ glam

Analysis through Semantics

Before describing our experiments in detail let us consider some critical points concerning representing semantics in a form suitable for analysis. There are several styles of semantics for imperative languages to be found in textbooks. These may all be translated more or less directly into declarative programming languages, but it is necessary to consider carefully the choice of semantics and the style in which the semantics is represented. The choice is influenced by two questions: firstly, what kind of analysis is to be performed on the imperative program, and secondly, how can the complexity of the analysis be minimised? 2.1

Big-Step and Small-Step Semantics

The usual distinction between different kinds of semantics is between the compositional and operational styles. However for our purpose, the most relevant division is between big-step and small-step semantics. Note that operational semantics can be either big-step (natural semantics) or small-step (structural operational semantics). Let us represent program statements by S and computation states by E. Big-step semantics is modelled using a relation of the form bigstep(S, E1 , E2 ), which means that the statement S transforms the initial state E1 to final state E2 . The effect of each program construct is defined independently; compound statements are modelled by composing together the effects of its parts. On the other hand, small-step semantics is typically modelled by a transition relation of the form smallstep(hS1 , E1 i, hS2 , E2 i). This states that execution of statement S1 in state E1 is followed by the execution of statement S2 in state E2 . Small-step semantics models a computation as a sequence of computation states, and the effect of a program construct is defined as its effect on the computation state in the context of the program in which it occurs. Big steps can be derived from small steps by defining a special terminating statement, say halt, and expressing big-step relations as a sequence of small steps. bigstep(S, E1 , E2 ) ← smallstep(hS, E1 i, hhalt, E2 i). bigstep(S, E1 , E3 ) ← smallstep(hS, E1 i, hS1 , E2 i), bigstep(S1 , E2 , E3 ). For the purposes of program analysis, the two styles of semantics have significant differences. Analysis of the bigstep relation allows direct comparison of the initial and final states of a computation. As shown above, big steps are derivable from small steps but the analysis becomes more complex. If the purpose of analysis is to derive relationships between initial and final states, then big-step semantics would be recommended. On the other hand, the smallstep relation represents directly the relation between one computation state and the next, information which would be awkward (though not impossible) to extract from the big-step semantics. Small-step semantics is more appropriate for analyses where local information about states is required, such as the relationship of variables within the same state, or between successive states.

Analysis of Imperative Programs through Analysis

2.2

249

One-State and Two-State Semantics

There is another option for representing small-step semantics, which leads to programs significantly simpler to analyse. Replace the smallstep(hS1 , E1 i, hS2 , E2 i) relation by a clause exec(S1 , E1 ) ← exec(S2 , E2 ). The meaning of the predicate exec(S, E) is that there exists a terminating computation of the statement S starting in the state E. We also have to add a special terminal statement called halt which is placed at the end of every program, and a statement exec(halt, E) ← true. We call this style of small-step semantics a one-state semantics since the relation exec represents only one state, in contrast to two-state semantics given by the bigstep and smallstep relations. As an aside, the one-state and two-state styles of representation follow a pattern identified by Kowalski in [10], when discussing graph-searching algorithms in logic programming. Kowalski noted that there were two ways to formalise the task of searching for a path from node a to node z in a directed graph. One way is to represent the graph as a set of facts of the form go(X, Y ) representing arcs from node X to node Y . A relation path(X, Y ) could then be recursively defined on the arcs. The search task is to solve the goal ← path(a, z). Alternatively, an arc could be represented as a clause of form go(Y ) ← go(X). In this case, to perform the task of searching for a path from node a to node z, the fact go(a) ← true is added to the program, and computation is performed by solving the goal ← go(z). There is no need for a recursively defined path relation, since the underlying logic of the implication relation fulfils the need. One-state semantics corresponds to the use of the relation go(X) while twostate semantics corresponds to using the relation go(X, Y ). The “start state” corresponds to the clause exec(halt, E) ← true. Our experiments show that the one-state semantics is considerably simpler to analyse than the two-state semantics. 2.3

Analysis of the Semantics

It may be asked whether one-state semantics is expressive enough, since no output state can be computed. That is, given a program P and initial state E, the computation is simulated by running the goal ← exec(P, E) and the final state is not observable. This is certainly inadequate if we are using our semantics to simulate the full effect of computations. However, during program analysis of ← exec(P, E) more things are observable than during normal execution of the same goal. In particular we can use a program analysis algorithm that gives information about calls to different program subgoals. In one-state semantics with a relation exec(S, E), the analysis can determine (an approximation of) every instance of exec(S, E) that arises in the computation. We can even derive information about the relation of successive states since the analysis can derive information about instances of a clause exec(S1 , E1 ) ← exec(S2 , E2 ). In the experiments to be described below, we start from a structural operational semantics in a textbook style, and derive a one-state small-step semantics.

250

3

J.C. Peralta, J.P. Gallagher, and H. Saˇ glam

A Semantics-Based Interpreter

As already mentioned a desirable property of structural operational semantics is the way it reflects every change in the computation state. Here we present briefly a way of systematically translating formal operational semantics (adapted from [16]) into a constraint logic program. We shall first provide some elements of the syntax of the imperative language and the metavariables used in the semantics descriptions. Assume we have an imperative language L with assignments, arithmetic expressions, while statements, if-then-else conditionals, empty statement, statement composition2 , and boolean expressions. Let S be a statement, a be an arithmetic expression, b a boolean expression, e a variable environment (mapping variables to their value), and x a program variable. All these variables may occur subscripted. A pair hS, ei is a configuration. Also a state is a special terminal configuration. The operational semantics below give meaning to programs by defining an appropriate transition relation holding between configurations. 3.1

Structural Operational Semantics

Using structural operational semantics [16] a transition relation ⇒ defines the relationship between successive configurations. There are different kinds of transition corresponding to different kinds of statement. Accordingly, the meaning of empty statement, assignment statement, if-then-else statement, while-do statement and composition of statements are: hskip, ei ⇒ e hx := a, ei ⇒ e[x 7→ A[[a]]e] hif b then S1 else S2 , ei ⇒ hS1 , ei if B[[b]]e = tt hif b then S1 else S2 , ei ⇒ hS2 , ei if B[[b]]e = ff hwhile b do S, ei ⇒ hif b then (S; while b do S) else skip, ei hS1 ; S2 , ei ⇒ hS10 ; S2 , e0 i if hS1 , ei ⇒ hS10 , e0 i hS1 ; S2 , ei ⇒ hS2 , e0 i if hS1 , ei ⇒ e0

(1) (2) (3) (4) (5) (6) (7)

where A[[·]] is the semantics function for arithmetic expressions and B[[·]] is the semantics function for boolean expressions. Intuitively, the assignment axiom schema above says that in a state e, x := a is executed to yield a state e[x 7→ A[[a]]e] which is as e except that x has the value A[[a]]e. Moreover, transition 6 expresses that if S1 is not a primitive statement of the language then execution won’t proceed to S2 until the rest of S1 , S10 , has been fully executed. Transition 7 considers the case when execution of S1 has been completed thus yielding state e0 , hence execution of S2 starts from this new state. We may specialise transitions 6 and 7 by unfolding their conditions with respect to transitions 1, 2, 3, 4, and 5 above to transitions 9, 10, 11, 12, and 13 2

The statement composition operator ‘;’ is assumed to be right associative

Analysis of Imperative Programs through Analysis

251

below. Transition 8 below is obtained by unfolding the condition of transition 6 with respect to itself and applying the associativity property of the ‘;’ operator. We assume that cases 2, 3, 4 and 5 do not occur since all programs are terminated by skip. Hence the new semantics: hskip, ei h(S1 ; S2 ); S3 , ei hskip; S2 , ei hx := a; S2 , ei

⇒ e ⇒ hS1 ; (S2 ; S3 ), ei ⇒ hS2 , ei

⇒ hS2 , e[x 7→ A[[a]]e]i h(if b then S1 else S2 ); S3 , ei ⇒ hS1 ; S3 , ei if hB[[b]]e = tti h(if b then S1 else S2 ); S3 , ei ⇒ hS2 ; S3 , ei if hB[[b]]e = ff i h(while b do S); S2 , ei ⇒

(8) (9) (10) (11) (12)

(13) hif b then (S; while b do S else skip); S2 , ei

Next, we express the semantics as a constraint logic program below. It is worth noting that this representation aids the analysis phase by carrying a single environment instead of double environment (that is, the one-state small-step semantics discussed in Sect. 2). exec(skip,E) <exec(compose(compose(S1,S2),S3),E) <exec(compose(S1,compose(S2,S3)),E) exec(compose(skip,S2),E) =t+1,b>=j+1} tm := k[j]; k[j] := k[j+1]; k[j+1] := tm; t := j; {n>=b,j>=1,t>=0,b>=j+1,j>=t} j := j+1 if (t == 0) then {n>=b,t=0,j>=1,b>=j,b<j+1} b := -1; else {n>=b,t>=0,j>=t+1,b>=j,b<j+1} b := t;

These results are the same as those obtained in [2].

6

Related Work

The first practical results on imperative languages for deriving linear equality or disequality relations among the variables of a program is due to Cousot and Halbwachs [2]. Their system was implemented in Pascal. The model execution used is based on flow-charts and an approximation method based on convex polyhedra. Incidentally the analyser used on the experiments here reported uses a similar approximation method integrated with other constraint solvers [20]. Later on in [3] the author poses the possibility of deriving different static analysers parameterised by the language semantics. In a similar way [8] show how to

Analysis of Imperative Programs through Analysis

259

obtain a static analyser for a non strict functional language. Such a static analyser is derived by successive refinements of the original language specification, natural operational semantics. The possible analyses obtained by the analyser derived with this method depend on the program property sought. This program property should be provided in advance. It appears that this technique has been applied to obtain some classical compiler analyses of programs in the sense of [1]. A good source of related work on implementation/derivation of static analysers from operational semantics for different programming languages is [8]. In [21] the authors describe a technique based on the style of abstract interpretation to statically estimate the ranges of variables throughout a program. Their implementation has been realised in the context of an optimising/parallelising compiler for C. Again, this is an example of using a variant of operational semantics to describe the abstract interpreters for static analysis of imperative programs. In [13] the author lays out the theory of abstract interpretation using twolevel semantics. Two-level semantics had been previously used in [15] to describe code generation. A summary of both can be found in [14]. Using denotational definitions the semantics of Pascal-like languages is given making explicit the distinction between compile time computations and run time computations, hence the two levels of the metalanguage. For program analysis, an appropriate interpretation of the run time metalanguage aids the analysis by giving a nonstandard semantics to run time constructs. By contrast in the present work, the semantic definitions are given in a standard way, and the translation is carried out by the partial evaluator where the distinction between compile time and run time computations is accounted for. Another sort of problem reduction for analysis is provided in [17]. The authors convert the problem of identifying dependences among program elements to a problem of identifying may-aliases. The transformation output is a program in the same language as the input program where may-aliases in the transformed program yield information directly translatable to control flow dependences between statements in the original program. In a similar way the authors claim that control flow dependences in the transformed program have a direct reading as may-aliases in the corresponding program. Presumably the their method and ours could be combined to obtain other problem reduction results. In [19] it is shown how to use logic programs to aid the analysis of imperative programs with pointers. The formalism is shown for the case of the pointer equality problem in Pascal. During analysis a set of assertions, represented as unary clauses, is updated according to the meaning of the program statement evaluated, the update operation designated and a set of consistency rules. The update operations resemble operations in deductive databases. The semantics of the imperative program is not explicitly represented as a logic program as in our approach, but in both approaches logic programs are used to express program properties.

260

7

J.C. Peralta, J.P. Gallagher, and H. Saˇ glam

Final Remarks

We have developed a language-independent method for analysing imperative programs. The method is based on encoding the semantics of an imperative language in a logic programming language for which there are advanced tools available for program analysis and transformation. This allows us to transfer the results of research on analysis of logic programs to the analysis of imperative programs. The emphasis of our work is to find practical and efficient techniques for achieving this aim. A key aspect is to write the semantics in a way that is amenable to analysis. We identified the one-state small-step semantics as a suitable style. The problem of relating the results of analysis of a logic program to the source code of the original imperative program was also solved. A representation of the imperative program was constructed in which program points were represented by special terms in the logic program. The partial evaluation algorithm was modified to exploit these terms, and thus produce a residual logic program whose structure mirrored that of the imperative program. Thus results of analysis of the logic programs could be related directly to the imperative code. The correctness of our results follow from correctness of the partial evaluator and correctness of the analyser. Both correctness proofs may be done independently of the imperative language to be analysed, which we claim is one of the contributions of the present work. Future Work We have performed some promising experiments on a simple language, as shown in this paper, but our aim is to analyse programs in a mainstream imperative language. Currently we are well advanced in writing the operational semantics for a significant subset of Java. We aim to enhance our current analysis tools by handling non-linear arithmetic constraints, and boolean constraints. Moreover, we aim to increase the flexibility of analysis by using pre-interpretations to express properties and abstract compilation to encode them as logic programs [18]. The use of the same method to perform other analyses, such as contextsensitive or alias analyses remains to be explored.

References 1. Alfred V. Aho, Ravi Sethi, and Jeffrey D. Ullman. Compilers: Principles, Techniques, and Tools. Addison Wesley, 1986. 2. P. Cousot and N. Halbwachs. Automatic discovery of linear restraints amog variables of a program. In Proceedings of the Conference Record of the 5th ACM Symposium on Principles of Programming Languages, pages 84–97, Albuquerque, New Mexico, 1978. 3. Patrick Cousot. Abstract interpretation based static analysis parametrized by semantics. In Proceedings of the Fourth International Symposium on Static Analysis, SAS’97, pages 388–394, Paris, France, 1997. LNCS 1302, Springer Verlag.

Analysis of Imperative Programs through Analysis

261

4. J.P. Gallagher. Tutorial on specialisation of logic programs. In Proceedings of the ACM SIGPLAN Symposium on Partial Evaluation and Semantics-Based Program Manipulation, pages 88–98, Copenhagen, Denmark, 1993. ACM Press. 5. J. Gallagher and M. Bruynooghe. Some low-level source transformations for logic programs. In Proceedings of Meta90 Workshop on Meta Programming in Logic. Katholieke Universiteit Leuven, Belgium, 1990. 6. J. Gallagher and M. Bruynooghe. The derivation of an algorithm for program specialisation. New Generation Computing, 9(3&4):305–333, 1991. 7. J. Gallagher and D.A. de Waal. Fast and precise regular approximation of logic programs. In P. Van Hentenryck, editor, Proceedings of the International Conference on Logic Programming (ICLP’94), Santa Margherita Ligure, Italy. MIT Press, 1994. 8. Val´erie Gouranton and Daniel Le M´etayer. Formal development of static program analysers. In Proceedings of the 8th Israeli Conference on Computer Systems and Software Engineering, pages 101–110, Israel, 1997. 9. Donald E. Knuth. The Art of Computer Programming, volume 3 of Sorting and Searching. Addison-Wesley Publishing Company, 1973. 10. Robert Kowalski. Logic for Problem Solving. North Holland, 1979. 11. Michael Leuschel. Advanced Techniques for Logic Program Specialisation. PhD thesis, Katholieke Universiteit Leuven, Department of Computer Science, Leuven, Belgium, May 1997. 12. J. W. Lloyd and J. C. Shepherdson. Partial evaluation in logic programming. The Journal of Logic Programming, 11(3&4):217–242, 1991. 13. Flemming Nielson. Two-level semantics and abstract interpretation. Theoretical Computer Science, (69):117–242, 1989. 14. Flemming Nielson. Semantics-directed program analysis: A toolmaker’s perspective. In Third International Symposium, SAS’96. Springer Verlag, LNCS 1145, 1996. 15. Flemming Nielson and Hanne Riis Nielson. Two-level semantics and code generation. Theoretical Computer Science, (56):59–133, 1988. 16. Hanne Riis Nielson and Flemming Nielson. Semantics with Applications. John Wiley and Sons, 1992. 17. John L. Ross and Mooly Sagiv. Building a bridge between pointer aliases and program dependences. In European Symposium On Programming, Lisbon, Portugal, 1998. 18. H¨ useyin Sa˘ glam. A Toolkit for Static Analysis of Constraint Logic Programs. PhD thesis, Bristol University, Department of Computer Science, Bristol, U.K., March 1998. 19. Mooly Sagiv, Nissim Francez, Michael Rodeh, and Reinhard Wilhelm. A logicbased approach to program flow analysis. 1998. Submitted to Acta Informatica. 20. H. Sa˘ glam and J. Gallagher. Constrained regular approximation of logic programs. In N. Fuchs, editor, Logic Program Synthesis and Transformation (LOPSTR’97). Springer-Verlag, Lecture Notes in Computer Science, 1998. in press. 21. Clark Verbrugge, Phong Co, and Laurie Hendren. Generalized constant propagation a study of C. Lecture Notes in Computer Science, Compiler Construction, (1060):74–90, 1996.

Improving Control in Functional Logic Program Specialization? E. Albert1 , M. Alpuente1 , M. Falaschi2 , P. Juli´ an3 , and G. Vidal1 1 2

DSIC, U. Polit´ecnica de Valencia, Camino de Vera s/n, 46022 Valencia, Spain. {ealbert,alpuente,gvidal}@dsic.upv.es Dip. di Mat. e Informatica, U. Udine, Via delle Scienze 206, 33100 Udine, Italy. [email protected] 3 Dep. de Inform´ atica, Ronda de Calatrava s/n, 13071 Ciudad Real, Spain. [email protected]

Abstract. We have recently defined a framework for Narrowing-driven Partial Evaluation (NPE) of functional logic programs. This method is as powerful as partial deduction of logic programs and positive supercompilation of functional programs. Although it is possible to treat complex terms containing primitive functions (e.g. conjunctions or equations) in the NPE framework, its basic control mechanisms do not allow for effective polygenetic specialization of these complex expressions. We introduce a sophisticated unfolding rule endowed with a dynamic narrowing strategy which permits flexible scheduling of the elements (in conjunctions) which are reduced during specialization. We also present a novel abstraction operator which extends some partitioning techniques defined in the framework of conjunctive partial deduction. We provide experimental results obtained from an implementation using the Indy system which demonstrate that the control refinements produce better specializations.

1

Introduction

Functional logic programming languages allow us to integrate some of the best features of the classical declarative paradigms, namely functional and logic programming. Lazy, efficient, functional computations are combined with the expressivity of logic variables, which allows for function inversion as well as logical search. The operational semantics of functional logic languages is usually based on (some form of) narrowing, which is a unification-based, parameter-passing mechanism which extends functional evaluation through goal solving capabilities as in logic programming (see [14] for a survey). In order to avoid unnecessary computations and to compute with infinite data structures, most recent work has concentrated on lazy narrowing strategies [11,15,23,25]. The aim of partial evaluation (PE) is to specialize a given program w.r.t. part of its input data (hence, also called program specialization). PE techniques ?

This work has been partially supported by CICYT TIC 95-0433-C03-03, by HCM project CONSOLE, and by Acci´ on Integrada HA1997-0073.

G. Levi (Ed.): SAS’98, LNCS 1503, pp. 262–277, 1998. c Springer-Verlag Berlin Heidelberg 1998

Improving Control in Functional Logic Program Specialization

263

have been widely applied to the optimization of functional (see [16] and references therein) and logic programs [10,18,22,27]. Unfortunately, these techniques generally cannot be easily transferred to functional logic languages, since logical variables in function calls place specific demands that have to be tackled in order to achieve effective specialization. Narrowing-driven PE [4] (NPE) provides a general scheme for the specialization of functional logic languages. The method is formalized within the theoretical framework established in [22,24] for the PE of logic programs (also known as partial deduction, PD). However, a number of concepts have been generalized for dealing with features such as nested function calls, eager and lazy evaluation strategies and the standard optimization based on deterministically reducing functions. Control issues are managed by using standard techniques as in [24, 28]. The NPE method of [4] distinguishes a local and a global levels of control. At the local level, (finite) narrowing trees for (nested) function calls are constructed. At the global level, the calls extracted from the leaves of the local trees are considered for the next iteration of the algorithm, after a proper abstraction (generalization) that guarantees that only a finite number of calls is specialized. A close, automatic approach is that of positive supercompilation (PS) [29], whose basic transformation operation is driving, a unification-based transformation mechanism which is similar to (lazy) narrowing. A different PE method for a rewriting-based, functional logic language is considered in [19]. Classical PD computes partial evaluations for separate atoms independently. Recently, [12,21] have introduced a technique for the partial deduction of conjunctions of atoms. This technique achieves a number of program optimizations such as (some form of) tupling and deforestation which are usually obtained through more expensive fold/unfold transformations, which are difficult to automate and which cannot be obtained through classical PD. The NPE method of [4] is able to produce polygenetic specializations [13], i.e. it is able to extract specialized definitions which combine several function definitions of the original program. That means that NPE has the same potential for specialization as conjunctive PD or PS within the considered paradigm (a detailed comparison can be found in [5,6]). This is because the generic method of [4] may allow one to deal with equations and conjunctions during specialization by simply considering the equality and conjunction operators as primitive function symbols of the language. Unfortunately, the use of primitive functions may encumber the nature of the specialization problems and it often turns out that some form of tupling (as defined in [27] for logic programs) is required for specializing expressions which contain conjunctive calls. The NPE algorithm of [4] does not incorporate a specific treatment for such primitive symbols, which depletes many opportunities for reaching the closedness condition and forces the method to dramatically generalize calls, thus giving up the intended specialization (see Example 1). Inspired by the challenging results of conjunctive PD in [12], this paper extends [4,3] by formulating and experimentally testing concrete NPE control options that effectively handle primitive function symbols in lazy functional logic languages.

264

E. Albert et al.

Some of the original contributions of our paper are as follows: i) we introduce a well-balanced dynamic unfolding rule and a novel abstraction operator that do not depend on the narrowing strategy and which highly improve the specialization of the NPE method; ii) these options allow us to tune the specialization algorithm to handle conjunctions (and other expressions containing primitive functions, such as conditionals and strict equalities) in a natural way, which provides for polygenetic specialization without any ad-hoc artifice; and iii) our method is applicable to modern functional logic languages with a lazy narrowing semantics such as Babel [25], Curry [15] and Toy [8], thus giving a specialization method which subsumes both lazy functional and conventional logic program specialization. We demonstrate the quality of these improvements by specializing some examples which were not handled well by classical NPE. The control strategies have been tested in the NPE system Indy [2]. The structure of the paper is as follows. Section 2 contains basic definitions. Section 3 extends the NPE algorithm of [3] to care for the appropriate handling of primitive function symbols. The algorithm is still generic in that no concrete control options are settled (i.e., there is no commitment to any concrete unfolding rule or abstraction operator). In Sect. 4, the concrete control options are described by formalizing some appropriate unfolding and abstraction operators. We illustrate the usefulness of our approach through some simple examples. Preliminary performance results, given in Sect. 5, show the practical importance of the proposed strategies. Finally, Sect. 6 concludes the paper. An extended version of this paper containing more details and proofs can be found in [1].

2

Preliminaries

We briefly summarize some well-known results about rewrite systems and functional logic programs [9,14]. The definitions below are given in the homogeneous case. The extension to many-sorted signatures is straightforward [26]. Throughout this paper, X denotes a countably infinite set of variables and F denotes a set of function symbols (also called the signature), each of which has a fixed associated arity. We assume that the signature F is partitioned into two sets F = C ∪ D with C ∩ D = Ø. Symbols in C are called constructors and symbols in D are called defined functions. T (F, X ) denotes the set of terms or expressions built from F and X . T (F) denotes the set of ground terms, while T (C, X ) denotes the set of constructor terms. If t 6∈ X, then Head(t) is the function symbol heading term t, also called the root symbol of t. A pattern is a term of the form f (d1 , . . . , dn ) where f /n ∈ D and d1 , . . . , dn are constructor terms. Var(s) is the set of variables occurring in the syntactic object s. A substitution is a mapping from X to T (F, X ) s.t. its domain Dom(σ) = {x ∈ X | xσ 6≡ x} is finite. We frequently identify a substitution σ with the set {x 7→ xσ | x ∈ Dom(σ)}. We denote the identity substitution by id. We consider the usual preorder on substitutions ≤: θ is more general than σ (in symbols θ ≤ σ) iff ∃γ. σ ≡ θγ. The restriction σ|V of a substitution σ to a set

Improving Control in Functional Logic Program Specialization

265

V of variables is defined by σ|V = xσ if x ∈ V and σ|V = x if x 6∈V . We write σ =V θ iff σ|V = θ|V , and σ ≤V θ iff ∃γ. σγ =V θ. A term t is more general than s (or s is an instance of t), in symbols t ≤ s, if ∃σ. tσ ≡ s. A unifier of a pair of terms {t1 , t2 } is a substitution σ such that t1 σ ≡ t2 σ. A unifier σ is called most general unifier (mgu) if σ ≤ σ 0 for every other unifier σ 0 . A generalization of a set of terms {t1 , . . . , tn } is a pair ht, {θ1 , . . . , θn }i such that tθi = ti , i = 1, . . . , n. A generalization ht, Θi is the most specific generalization (msg) if t0 ≤ t for every other generalization ht0 , Θ0 i. Positions of a term t are represented by sequences of natural numbers used to address subterms of t. They are ordered by the prefix ordering: p ≤ q if there is w such that p.w = q, where p.w denotes the concatenation of sequences p and w. We let Λ denote the empty sequence. Pos(t) and FPos(t) denote, respectively, the set of positions and the set of nonvariable positions of the term t. t|p is the subterm of t at position p. t[s]p is the term t with the subterm at position p replaced with s. We find it useful to simplify our description by limiting the discussion to unconditional term rewriting systems. A rewrite rule is pair l → r with l, r ∈ T (F, X ), l 6∈ X, and Var(r) ⊆ Var(l). l and r are called the left-hand side (lhs) and right-hand side (rhs) of the rewrite rule, respectively. A term rewriting system (TRS) R is a finite set of rewrite rules. A rewrite step is an application of a rewrite rule to a term, i.e. t →p,l→r s if there exists a position p ∈ Pos(t), a rewrite rule l → r, and a substitution σ with t|p = lσ and s = t[rσ]p . We say that t|p is a redex (reducible expression) of t. A term t is reducible to term s if t →∗ s. A term t is irreducible or in normal form if there is no term s with t → s. A TRS is terminating if there are no infinite sequences of the form t1 → t2 → . . . A TRS is called confluent if, whenever a term s reduces to two terms t1 and t2 , both t1 and t2 reduce to the same term. Since we do not require terminating rules, normal forms may not exist. Functional Logic Programming The operational semantics of functional logic programs is usually based on (some variant) of narrowing. Essentially, narrowing consists of computing an appropriate substitution so that when applied to the current term, it becomes reducible, and then reducing it [14]. In this section, we briefly introduce a functional logic language whose syntax and demand-driven reduction mechanism is essentially equivalent to that of (a subset of) Babel [23,25], Toy [8], and Curry [15], which has been recently proposed to become a standard in the area. A TRS R is constructor-based (CB) if for each rule l → r ∈ R the lhs l is a pattern. A CB TRS R is weakly-orthogonal if R is left-linear (i.e., for each rule l → r ∈ R, the lhs l does not contain multiple occurrences of the same variable) and R contains only trivial overlaps (i.e., if l → r and l0 → r0 are variants of distinct rules in R and σ is a unifier for l and l0 , then rσ ≡ r0 σ). It is well-known that weakly-orthogonal TRS’s are confluent. We henceforth consider CB weakly-orthogonal TRS’s as programs. For this class of programs, a term t is a head normal form if t is a variable or Head(t) ∈ C.

266

E. Albert et al.

The signature F is augmented with a set of primitive function symbols P = {≈, ∧, ⇒} in order to handle complex expressions containing equations s ≈ t, conjunctions b1 ∧b2 , and conditional (guarded) terms b ⇒ t, i.e. F = C∪D∪P. We usually treat the symbols in P as infix operators. We assume that the following predefined rules belong to any given program: c ≈ c → true % c/0 ∈ C c(x1 , . . . , xn ) ≈ c(y1 , . . . , yn ) → (x1 ≈ y1 ) ∧ . . . ∧ (xn ≈ yn ) % c/n ∈ C true ∧ x → x x ∧ true → x (true ⇒ x) → x These rules are weakly-orthogonal and define the validity of an equation as a strict equality between terms, which is common in functional languages when computations may not terminate [11,25]. Note that, although the basic computation model only supports unconditional rules, it is still adequate to support logic programs since conditional rewrite rules l → r ⇐ C can be encompassed by guarded unconditional rules l → (C ⇒ r) by using the conditional primitive ‘⇒’ as in Babel [25]. For reasons of simplicity, we assume the associativity of ‘∧’ and assume that ‘≈’ binds more than ‘∧’ and ‘∧’ binds more than ‘⇒’. We consider that programs are executed by lazy narrowing, which allows us to deal with nonterminating functions [23,25]. Roughly speaking, laziness means that a given expression is only narrowed at inner positions if they are demanded (by the pattern in the lhs of some rule) and this contributes to a later narrowing step at an outer position. Formally, given a program R, we define the one-step narrowing relation as follows. A term s narrows to t in R, in symbols s ;p,l→r,σ t (or simply s ;σ t), iff there exists a position p ∈ ϕ(s), a (standardized apart) rule l → r ∈ R, and a substitution σ such that σ = mgu({s|p , l}) and t = (s[r]p )σ. The selection strategy ϕ(t) is responsible for computing the set of demanded positions of a given term t. A formal definition of this strategy in terms of an inference system is shown in [1]. Lazy narrowing is strong complete w.r.t. constructor substitutions in CB, weakly-orthogonal TRS’s [25, 14]. This means that the interpreter is free to disregard from ϕ(t) all components of each conjunction which may occur in t except one, even if all arguments are demanded by the predefined rules of ‘∧’ (that is, completeness holds for all scheduling policies). A formal definition can be found in [3]. If s0 ;σ1 s1 ;σ2 . . . ;σn sn (in symbols, s0 ;∗σ sn , σ = σ1 σ2 . . . σn ), we speak of a lazy narrowing derivation for the goal s0 with (partial) result sn . A lazy narrowing derivation s ;∗σ t is successful iff t ∈ T (C ∪ X ), where σ|Var(s) is the computed answer substitution.

3

The Generalized Specialization Algorithm

In this section, we generalize some basic concepts and techniques for the NPE of (lazy) functional logic programs (as presented in [3]). These extended notions will prove to be extremely useful for formulating new unfolding and abstraction operators which are well-suited to cope with primitive function symbols. In the original NPE framework, no distinction is made between primitive and defined function symbols during specialization. For instance, a conjunction b1 ∧ b2 is

Improving Control in Functional Logic Program Specialization

267

sorted bits(x : xs) ∧ 1 ≤ x

(( (((( AA ( ( ( (

true ∧ 1 ≤ x

sorted bits(x0 : xs0 ) ∧ x ≤ x0 ∧ 1 ≤ x

... Fig. 1. Incomplete narrowing tree for sorted bits(x : xs) ∧ 1 ≤ x

considered as a single entity when checking whether it is covered by the set of specialized calls. This commonly implies a drastic generalization of the involved calls, which causes losing all specialization, as the following example illustrates. Example 1. Let us consider the program excerpt: sorted bits(x : [ ]) → true sorted bits(x1 : x2 : xs) → sorted bits(x2 : xs) ∧ x1 ≤ x2 0 ≤ 0 → true 0 ≤ 1 → true 1 ≤ 1 → true and the call sorted bits(x : xs) ∧ 1 ≤ x. The lazy narrowing tree depicted1 in Fig. 1 is built up by using the nonembedding unfolding rule of [3], which expands derivations while new redexes are not “greater” (with the homeomorphic embedding ordering, see e.g. [4,28]) than previous, comparable redexes in the branch (i.e., redexes with the same outermost function symbol). From this tree, we can identify two main weaknesses of the plain NPE algorithm: – The rightmost branch stops because the leftmost redex sorted bits(x0 : xs0 ) “embeds” the previous redex sorted bits(x : xs), even if no reductions have been performed on the other elements of the conjunction. – At the global level, since the call sorted bits(x0 : xs0 ) ∧ x ≤ x0 ∧ 1 ≤ x in the leaf of the tree embeds (but does not cover) the specialized call sorted bits(x : xs) ∧ 1 ≤ x (and they are comparable), the msg sorted bits(x : xs) ∧ z is computed, which gives up the intended specialization. The first drawback pointed out in this example motivates the definition of more sophisticated unfolding rules which are able to achieve a balanced evaluation of the given expression by narrowing appropriate redexes (e.g., by using some kind of dynamic scheduling strategy which takes into account the ancestors narrowed in the same branch). The second drawback suggests the definition of a more flexible abstraction operator which is able to automatically split complex terms before attempting folding or generalization. In the following, we refine the framework of [3] in order to overcome these problems. The PE of a term is obtained by constructing a (partial) narrowing tree and then extracting the resultants associated to the root-to-leaf paths of the tree. Definition 1 (resultant). Let s be a term and R be a program. Given a lazy narrowing derivation s ;∗σ t, its associated resultant is the rewrite rule sσ → t. 1

We assume a fixed left-to-right selection of components within conjunctions and underline the selected redex at each step.

268

E. Albert et al.

Definition 2 (partial evaluation). Let R be a program and s be a term. Let τ be a finite (possibly incomplete) narrowing tree for s in R such that no goal in the tree is narrowed beyond its head normal form. Let {t1 , . . . , tn } be the terms in the leaves of τ . Then, the set of resultants for the narrowing sequences {s ;+ σi ti | i = 1, . . . , n} is called a partial evaluation of s in R. The partial evaluation of a set of terms S in R is defined as the union of the partial evaluations for the terms in S. Intuitively, the reason for requiring that the PE of a term s do not surpass its head normal form is that, at runtime, the evaluation of a (nested) call C[s]p containing the partially evaluated term s at some position p might not demand evaluating s beyond its head normal form. Since the “contexts” C[ ] in which s will appear are not known at PE time, we avoid to interfering with the “lazy nature” of computations in the specialized program by imposing this condition. A recursive closedness condition is formalized by inductively checking that the different calls in the rules are sufficiently covered by the specialized functions. Definition 3 (closedness). Let S be a finite set of terms. A term t is S-closed if closed(S, t) holds,where the predicate closed is defined inductively as follows: if t ∈ X true closed(S, t ) ∧ . . . ∧ closed(S, t ) if t ≡ c(t1 , . . . , tn ) 1 n ^ closed(S, t) ⇔ + 0 ∃s ∈ S . sθ = t ∧ closed(S, t ) if t ≡ f (t1 , . . . , tn ) +

x/t0 ∈θ

where c ∈ C, f ∈ (D ∪ P), and S = S ∪ {p(x, y) | p ∈ P}. We say that a set of terms T is S-closed, written closed(S, T ), if closed(S, t) holds for all t ∈ T , and we say that a program R is S-closed if closed(S, Rcalls ) holds. Here we denote by Rcalls the set of terms in the rhs’s of the rules in R. The novelty w.r.t. [4,3] is that a complex expression headed by a primitive function symbol, such as a conjunction, is proved closed w.r.t. S either by checking that it is an instance of a call in S (followed by an inductive test of the subterms), or by splitting it into two conjuncts and then trying to match with “simpler” terms in S (which happens when matching is first attempted w.r.t. one of the ‘flat’ calls p(x, y) in S + ). This extension is safe since the rules which define the primitive functions are automatically added to each program. The way in which a concrete PE is made is given by an unfolding rule (which decides how to stop the construction of lazy narrowing trees) and an abstraction operator (which ensures the finiteness of the set of specialized calls). Definition 4 (unfolding rule [3]). An unfolding rule U is a mapping which, when given a program R and a term s, returns a concrete PE for s in R (a set of resultants). By U (S, R) we denote the union of U (s, R) for all s ∈ S. Definition 5 (abstraction operator). Given a finite set of terms T and a set of terms S, an abstraction operator is a function which returns a finite set of terms abstract(S, T ) such that: i) if s ∈ abstract(S, T ), then there exists t ∈ (S ∪ T ) such that t|p = sθ for some position p and substitution θ; ii) for all t ∈ (S ∪ T ), t is closed w.r.t. the set of terms in abstract(S, T ).

Improving Control in Functional Logic Program Specialization

269

Roughly speaking, the first condition guarantees that the abstraction operator does not introduce new function symbols, while the second condition ensures that the resulting set of terms “covers” the calls previously specialized and that closedness is preserved throughout successive abstractions. The following basic algorithm for NPE is parameterized by the unfolding rule U and the abstraction operator abstract in the style of [10]. Algorithm 1.

Input: a program R and a set of terms T Output: a set of terms S Initialization: i := 0; T0 := T Repeat

1. R0 := U (Ti , R); 2. Ti+1 := abstract(Ti , R0calls ); 3. i := i + 1; Until Ti = Ti−1 (modulo renaming) Return S := Ti

The output of the NPE algorithm, given a program R, is not a PE, but a set of terms S from which the partial evaluations U (S, R) are automatically derived. Note that, whenever the specialized call is not a pattern, lhs’s of resultants are not patterns either and hence resultants are not (CB) program rules. In [3], we introduced a post-processing renaming which is useful for producing CB rules. Roughly speaking, we construct an “independent renaming” S 0 of S as follows: for each term s in S, we define its independent renaming s0 = fs (x1 , . . . , xn ), where x1 , . . . , xn are the distinct variables in s in the order of their first occurrence and fs is a new fresh function symbol. Then, we fold each call t in the rules of U (S, R) by replacing the old call t by a call to the corresponding term t0 in S 0 (details can be found in [3]). After the algorithm terminates, the specialized program is obtained by applying this post-processing renaming to U (S, R). The (partial) correctness of the NPE algorithm is stated as follows. Theorem 2. Given a program R and a term t, if Algorithm 1 terminates by computing the set of terms S, then R0 and t are S-closed, where R0 = U (S, R). The correctness of the generic algorithm is stated in the following theorem, which generalizes Theorem 4.5 of [3]. Theorem 3. Let R be a program, t a term, and S a finite set of terms. Let R0 be a PE of R w.r.t. S such that R0 and t are S-closed. Let S 0 be an independent renaming of S, and t00 (resp. R00 ) be a renaming of t (resp. R0 ) w.r.t. S 0 . Then t computes in R the result d with computed answer θ iff t00 computes in R00 the result d with computed answer θ0 and θ0 ≤Var(t) θ.

4

Improving Control of NPE

In Sect. 4.1 we improve control in functional logic specialization by fixing an unfolding strategy which is specifically designed for “conjunctive specialization”.

270

E. Albert et al. app(x, y) ≈ w ∧ app(w, z) ≈ r

((aa (((( aa ( ( ( ( a

.. .

x0 : app(xs0 , y) ≈ w ∧ app(w, z) ≈ r

x0 ≈ w0 ∧ app(xs0 , y) ≈ ws0 ∧ app(w0 : ws0 , z) ≈ r true ∧ app(xs0 , y) ≈ ws0 ∧ app(w0 : ws0 , z) ≈ r app(xs0 , y) ≈ ws0 ∧ app(w0 : ws0 , z) ≈ r Fig. 2. Na¨ıve local control for app(x, y) ≈ w ∧ app(w, z) ≈ r

As for global control, a specific treatment of the primitive function symbols ‘≈’, ‘∧’ and ‘⇒’ is introduced in Sect. 4.2 which produces more effective and powerful, polygenetic specializations, as compared to classical NPE. 4.1

Local Control

The unfolding rule of [3] simply exploits the redexes selected by the lazy narrowing strategy ϕ (using a static selection rule which determines the next conjunct to be reduced) whenever none of them embed a previous (comparable) redex. The following example reveals that this strategy is not elaborated enough for specializing calls which may contain primitive symbols like conjunctions. Example 2. Consider the well-known program append: app([ ], y) → y app(x : xs, y) → x : app(xs, y) with the input goal “app(x, y) ≈ w ∧ app(w, z) ≈ r”. Using the nonembedding unfolding rule of [3], we obtain the tree depicted2 in Fig. 2 (using a fixed leftto-right selection rule for conjunctions). From this local tree, no appropriate specialized definition for the initial goal can be obtained, since the leaf cannot be folded into the input call in the root of the tree and generalization is required (which causes losing all specialization, as in Example 1). We note that the NPE method [4,3] succeeds with this example when the specialized call is written as a nested expression app(app(x, y), z). This is because exploiting the nesting capability of the functional syntax allows us to transform the original tupling3 problem illustrated by Example 2 into a simpler, deforestation problem, which is easily solved by the original NPE method. 2 3

We adopt the standard optimization which makes use of the built-in unification to solve strict equalities s ≈ t (provided that only constructor bindings are produced). Here we refer to tupling of logic programs, which subsumes both deforestation and tupling of functional programs [27].

Improving Control in Functional Logic Program Specialization

271

Now we introduce a dynamic lazy unfolding rule which attempts to achieve a fair evaluation of the complete input term, rather than a deeper evaluation of some given subterm. This novel concrete unfolding rule dynamically selects the positions to be reduced within a given conjunction, by exploiting some dependency information between redexes gathered along the derivation. Definition 6 (dependent positions). Let D ≡ (s ;p,l→r,σ t) be a narrowing step. The set of dependent positions of a position q of s by D, written q\\D, is: {q.u | u ∈ FPos(r) ∧ Head(r|u ) 6∈ C} if q = p if q 6≤p q\\D = {q} {p.u0 .v | r|u0 = x} if q = p.u.v and l|u = x ∈ X This notion can be naturally lifted to narrowing derivations. The notion of dependency for terms stems directly from the corresponding notion for positions. Note that the above definition is a strict extension of the standard notion of descendant in functional programming. Intuitively, the descendants of the subterm s|q are computed as follows: underline the root symbol of s|q and perform the narrowing step s ; t. Then, every subterm of t with and underlined root symbol is a descendant of s|q [17]. Intuitively, a position q 0 of t depends on a position q of s (by D) if q 0 is a descendant of q (second and third cases), or if the position q 0 has been introduced by the rhs of the rule applied in the reduction of the former position q and it addresses a subterm headed by a defined function symbol (first case). Note that this notion is an extension of the standard PD concept of (covering) ancestor to the functional logic framework. By abuse, we also say that the term addressed by q is an ancestor of the term addressed by q 0 in D. If s is an ancestor of t and Head(s) = Head(t), we say that s is a comparable ancestor of t in D. Now we formalize the way in which the dynamic selection is performed. Definition 7 (dynamic narrowing selection strategy). Let D ≡ (t0 ; t1 ; . . . ; tn ), n ≥ 0 be a lazy narrowing derivation. We define the dynamic selection rule ϕdynamic as: ϕdynamic (tn , D) = select(tn , Λ, D), where the auxiliary function select is: select(t, p, D) = if p ∈ ϕ(t) then if dependency clash(t|p , D) then {⊥} else {p} else case t|p of x ∈ V: Ø s1 ∧ s2 : let Oi = select(t, p.i, D), i ∈ {1, 2}, in [if ∃i. (⊥ 6∈Oi ∧ Oi 6≡Ø) then Oi else if (O1 ≡ O2 ≡ Ø) then Ø else {⊥}] otherwise: let t|p = S f (s1 , . . . , sn ) and n Oargs = i=1 select(t, p.i, D) in [if ⊥ ∈ Oargs then {⊥} else Oargs ] where dependency clash(t, D) is a generic Boolean function that looks at the ancestors of t in D to determine whether there is a risk of nontermination.

272

E. Albert et al. app(x, y) ≈ w ∧ app(w, z) ≈ r

(((aa ((((( aa ( ( ( ( ( a

.. .

x0 : app(xs0 , y) ≈ w ∧ app(w, z) ≈ r

x0 ≈ w0 ∧ app(xs0 , y) ≈ ws0 ∧ app(w0 : ws0 , z) ≈ r x0 ≈ w0 ∧ app(xs0 , y) ≈ ws0 ∧ w0 : app(ws0 , z) ≈ r x0 ≈ w0 ∧ app(xs0 , y) ≈ ws0 ∧ w0 ≈ r0 ∧ app(ws0 , z) ≈ rs0 Fig. 3. Improved local control for app(x, y) ≈ w ∧ app(w, z) ≈ r

For the sake of simplicity, in the remainder of this section we consider that dependency clash(t, D) holds whenever there is a comparable ancestor of the selected redex t in D. Another approach, that we investigate in the experiments, is to additionally test homeomorphic embedding on comparable ancestors. Roughly speaking, the dynamic selection strategy recurs over the structure of the goal and determines the set of positions to be unfolded by a don’t-care selection within each conjunction of exactly one of the components (among those that do not incur into a dependency clash). We introduce a dynamic unfolding rule Udynamic (t, R) which simply expands lazy narrowing trees according to the dynamic lazy narrowing strategy ϕdynamic . The “mark” ⊥ of Definition 7 is used as a whistle to warn us that the derivation must be cut off since it runs into a dependency clash, i.e., each branch D of the tree is stopped whenever ϕdynamic (t, D) = {⊥} or the term t (of the leaf) is in head normal form. Example 3. Consider again the program and goal of Example 2. Using the dynamic unfolding rule Udynamic , we get the tree depicted in Fig. 3. From this tree, an optimal (recursive) specialized definition for the initial call can be derived, provided there is a suitable splitting mechanism to extract, from the leaf of the tree, an appropriate subconjunction such as app(xs0 , y) ≈ ws0 ∧app(ws0 , z) ≈ rs0 , which is covered by the initial call in the root of the tree (see Example 4). 4.2

Global Control

In the presence of primitive functions like ‘∧’ or ‘≈’, using an abstraction operator which respects the structure of the terms (as in [3]) is not very effective, since the generalization of two conjunctions (resp. equations) might be a term of the form x ≈ y ∧ z (resp. x ≈ y) in most cases. The drastical solution of decomposing the term into subterms containing just one function call can avoid the problem, but has the negative consequence of losing nearly all specialization. In this section, we introduce a more concerned abstraction operator which is inspired by the partitioning techniques of conjunctive PD [12,21], and which uses the homeomorphic embedding relation E as defined in [28].

Improving Control in Functional Logic Program Specialization

273

The following notion of best matching terms, which is aimed at avoiding loss of specialization due to generalization, is a proper generalization of the notion of best matching conjunction in [12]. Definition 8 (best matching terms). Let S = {s1 , . . . , sn } be a set of terms, t a term, and consider the set of terms W = {wi | hwi , {θi1 , θi2 }i = msg({si , t}), i = 1, . . . , n}. The best matching terms BM T (S, t) for t in S are those terms sj ∈ S such that the corresponding wj in W is a minimally general element. The notion of BMT is used in the abstraction process at two stages: i) when selecting the more appropriate term in S which covers a new call t, and ii) when determining whether a call t headed by a primitive function symbol could be (safely) added to the current set of specialized calls or should be split. Definition 9 (concrete abstraction operator). Let S and T be sets of terms. We define abstractE (S, T ) inductively as follows: abstractE (S, T ) = S if T ≡ Ø or T ≡ {t}, t ∈ X abstractE (. . . abstractE (S, t1 ), . . . , tn ) if T ≡ {t1 , . . . , tn }, n > 0 abstractE (S, {t1 , . . . , tn }) if T ≡ {t}, t ≡ c(t1 , . . . , tn ), c ∈ C 0 abs def (S, T , t) if T ≡ {t}, Head(t) ∈ D if T ≡ {t}, Head(t) ∈ P abs prim(S, T 0 , t) where T 0 = {s ∈ S | Head(s) = Head(t) ∧ s E t}. The functions abs def and abs prim are defined as follows: abs def (S, Ø, t) = abs prim(S, Ø, t) = S ∪ {t} abs def (S, T, t) = abstractE (S \ {s}, {w} ∪ Ran(θ1 ) ∪ Ran(θ2 )) if hw, {θ1 , θ2 }i = msg({s, t}), with s ∈ BM T (T, t) abs def (S, T, t) if ∃s ∈ BM T (T, t) s.t. def (t) = def (s) abs prim(S, T, t) = ∧ abstractE (S, T, {t1 , t2 }) otherwise, where t = p(t1 , t2 ) where def (t) denotes a sequence with the defined function symbols of t in lexico∧ graphical order, and = is equality up to reordering of elements in a conjunction. Essentially, the way in which the abstraction operator proceeds is simple. We distinguish the cases when the considered term i) is a variable, ii) is headed by a constructor symbol, iii) by a defined function symbol, or iv) by a primitive function symbol. The actions that the abstraction operator takes, respectively, are: i) to ignore it, ii) to recursively inspect the subterms, iii) to generalize the given term w.r.t. some of its best matching terms (recursively inspecting the msg w and the subterms of θ1 , θ2 not covered by the generalization), and iv) the same as in iii), but considering the possibility of splitting the given expression before generalizing it when def (t) 6= def (s) (which essentially states that some defined function symbols would be lost due to the application of msg). The function abstractE is an abstraction operator in the sense of Definition 5 [1]. The following result establishes the termination of the global specialization process Theorem 4. Algorithm 1 terminates for the unfolding rule Udynamic and the abstraction operator abstractE .

274

E. Albert et al.

Our final example witnesses that abstractE behaves well w.r.t. Example 3. Example 4. Consider again the tree depicted in Fig. 3. By applying Algorithm 1, the following call to abstractE is undertaken: abstractE ({app(x, y) ≈ w ∧ app(w, z) ≈ r}, {x0 ≈ w0 ∧ app(xs0 , y) ≈ ws0 ∧ w0 ≈ r0 ∧ app(ws0 , z) ≈ rs0 }). Following Definition 9, by two recursive calls to abs prim, we get: {app(x, y) ≈ w ∧ app(w, z) ≈ r, x0 ≈ w0 }. By considering the independent renaming dapp(x, y, w, z, r) for the specialized call app(x, y) ≈ w ∧ app(w, z) ≈ r, the method derives a (recursive) rule of the form: dapp(x : xs, y, w : ws, z, r : rs) → x ≈ w ∧ w ≈ r ∧ dapp(xs, y, ws, z, rs), which embodies the intended optimal specialization for this example.

5

Experiments

The refinements presented so far have been incorporated into the NPE prototype implementation system Indy (Integrated Narrowing-Driven specialization system [2]). Indy is written in SICStus Prolog v3.6 and is publicly available [2]. In order to assess the practicality of our approach, we have benchmarked the speed and specialization achieved by the extended implementation. The benchmarks used for the analysis are: applast, which appends an element at the end of a given list and returns the last element of the resulting list; double app, see Example 2; double flip, which flips a tree structure twice, then returning the original tree back; fibonacci, fibonacci’s function; heads&legs, which computes the number of heads and legs of a given number of birds and cats; match-app, the extremely na¨ıve string pattern matcher based on using append; match-kmp, a semi-na¨ıve string pattern matcher; maxlength, which returns the maximum and the length of a list; palindrome, a program to check whether a given list is a palindrome; and sorted bits, see Example 1. Some of the examples are typical PD benchmarks (see [20]) adapted to a functional logic syntax, while others come from the literature of functional program transformations, such as PS [29], fold/unfold transformations [7], and deforestation [30]. We have considered the following settings to test the benchmarks: – Evaluation strategy: All benchmarks were executed by lazy narrowing. – Unfolding rule: We have tested three different alternatives: (1) emb goal: it expands derivations while new goals do not embed a previous comparable goal in the same branch; (2) emb redex: the concrete unfolding rule of Sect. 4.1 which implements the dependency clash test using homeomorphic embedding on comparable ancestors of selected redexes to ensure finiteness (note that it differs from emb goal in that emb redex implements dynamic scheduling on conjunctions and that homeomorphic embedding is checked on simple redexes rather than on whole goals); (3) comp redex: the unfolding rule of Sect. 4.1 which uses the simpler definition of dependency clash based on comparable ancestors of selected redexes as a whistle. – Abstraction operator: Abstraction is always done as explained in Def. 9.

Improving Control in Functional Logic Program Specialization

275

Table 1. Benchmark results Benchmarks applast double app double flip fibonacci heads&legs match-app match-kmp maxlength palindrome sorted bits Average TMix average

Original Rw RT 10 90 8 106 8 62 5 119 8 176 8 201 12 120 14 94 10 119 8 110 9.1 119.7

emb goal Rw Speedup 13 1.32 39 1.63 26 1.51 11 1.19 24 4.63 12 1.25 14 3.43 51 1.17 19 1.25 16 1.15 22.5 1.85 1881

emb redex Rw Speedup 28 2.20 61 1.28 17 1.55 7 1.08 22 2.41 20 2.75 14 3.64 20 1.27 10 1.35 31 2.89 23 2.04 7441

comp redex Rw Speedup 13 1.10 15 3.12 17 1.55 7 1.08 21 2.48 23 2.79 13 3.43 18 1.25 10 1.35 10 2.68 14.7 2.08 5788

Table 1 summarizes our benchmark results. The first two columns measure the number of rewrite rules (Rw) and the absolute runtimes (RT) for each original program. The other columns show the number of rewrite rules and the speedups achieved for the specialized programs obtained by using the three considered unfolding rules. The row at the bottom of the table (TMix) indicates the average specialization time for each considered unfolding rule. Times are expressed in milliseconds and are the average of 10 executions. Speedups were computed by running the original and specialized programs under the publicly available lazy functional logic language Toy [8]. Runtime input goals were chosen to give a reasonably long overall time. The complete code for benchmarks, the specialized calls, and the partially evaluated programs can be found in [1]. The figures in Table 1 demonstrate that the control refinements that we have incorporated into the Indy system provide satisfactory speedups on all benchmarks (which is very encouraging, given the fact that no partial input data were provided in any example, except for match-app, match-kmp, and sorted bits). On the other hand, our extensions are conservative in the sense that there is no penalty w.r.t. the specialization achieved by the original system on nonconjunctive goals (although some specialization times are slightly higher due to the more complex processing being done). Let us note that, from the speedup results in Table 1, it can appear that there is no significant difference between the strategies emb redex and comp redex. However, when we also consider the specialization times (TMix) and the size of the specialized programs (Rw), we find out that comp redex has a better overall behaviour. A detailed comparison between the considered unfolding strategies can be found in [1].

6

Discussion

In functional logic languages, expressions can be written by exploiting the nesting capability of the functional syntax, as in app(app(x, y), z) ≈ r, but in many cases

276

E. Albert et al.

it can be appropriate (or necessary) to decompose nested expressions as in logic programming, and write app(x, y) ≈ w ∧ app(w, z) ≈ r (for instance, if some test such as sorted bits(w) on the intermediate list w were necessary). The original Indy system behaves well on programs written with the “pure” functional syntax [5]. However, Indy is not able to produce good specialization on the benchmarks of Table 1 when they are written as conjunctions of subgoals. For this we could not achieve some of the standard, difficult transformations such as tupling [7,27] within the classical NPE framework. As opposed to the classical PD framework (in which only folding on single atoms can be done), the NPE algorithm is able to perform folding on complex expressions (containing an arbitrary number of function calls). This does not suffice to achieve tupling in practice, since complex expressions are often generalized and specialization is lost. We have shown that the NPE general framework can be supplied with appropriate control options to specialize complex expressions containing primitive functions, thus providing a powerful polygenetic specialization framework with no ad-hoc setting. To the best of our knowledge, this is the first practical framework for the specialization of modern functional logic languages with partitioning techniques and dynamic scheduling. As future research, there is room for further improvement in performance by introducing more powerful abstraction operators based on better analyses to determine the optimal way to split expressions (trying not to endanger the communication of data structures with shared variables), and by considering in practice the problem of controlling particular algebraic laws for primitive symbols.

References 1. E. Albert, M. Alpuente, M. Falaschi, P. Juli´ an, and G. Vidal. Improving Control in Functional Logic Program Specialization. Technical Report DSIC-II/2/97, UPV, 1998. Available from URL: http://www.dsic.upv.es/users/elp/papers.html. 2. E. Albert, M. Alpuente, M. Falaschi, and G. Vidal. Indy User’s Manual. Technical Report, available from http://www.dsic.upv.es/users/elp/papers.html. 3. M. Alpuente, M. Falaschi, P. Juli´ an, and G. Vidal. Specialization of Lazy Functional Logic Programs. In Proc. of PEPM’97, volume 32(12) of Sigplan Notices, pages 151–162, New York, 1997. ACM Press. 4. M. Alpuente, M. Falaschi, and G. Vidal. Narrowing-driven Partial Evaluation of Functional Logic Programs. In H. Riis Nielson, editor, Proc. of the 6th European Symp. on Programming, ESOP’96, pages 45–61. Springer LNCS 1058, 1996. 5. M. Alpuente, M. Falaschi, and G. Vidal. Partial Evaluation of Functional Logic Programs. ACM TOPLAS, 1998. To appear. 6. M. Alpuente, M. Falaschi, and G. Vidal. A Unifying View of Functional and Logic Program Specialization. ACM Computing Surveys, 1998. To appear. 7. R.M. Burstall and J. Darlington. A Transformation System for Developing Recursive Programs. Journal of the ACM, 24(1):44–67, 1977. 8. R. Caballero-Rold´ an, F.J. L´ opez-Fraguas, and J. S´ anchez-Hern´ andez. User’s manual for Toy. Technical Report SIP-5797, UCM, Madrid (Spain), April 1997. 9. N. Dershowitz and J.-P. Jouannaud. Rewrite Systems. In J. van Leeuwen, editor, Handbook of Theoretical Computer Science, volume B: Formal Models and Semantics, pages 243–320. Elsevier, Amsterdam, 1990.

Improving Control in Functional Logic Program Specialization

277

10. J. Gallagher. Tutorial on Specialisation of Logic Programs. In Proc. of PEPM’93, pages 88–98. ACM, New York, 1993. 11. E. Giovannetti, G. Levi, C. Moiso, and C. Palamidessi. Kernel Leaf: A Logic plus Functional Language. J. of Computer and System Sciences, 42:363–377, 1991. 12. R. Gl¨ uck, J. Jørgensen, B. Martens, and M.H. Sørensen. Controlling Conjunctive Partial Deduction of Definite Logic Programs. In Proc. of PLILP’96, pages 152– 166. Springer LNCS 1140, 1996. 13. R. Gl¨ uck and M.H. Sørensen. A Roadmap to Metacomputation by Supercompilation. In Partial Evaluation, Int’l Seminar, Dagstuhl Castle, Germany, pages 137–160. Springer LNCS 1110, February 1996. 14. M. Hanus. The Integration of Functions into Logic Programming: From Theory to Practice. Journal of Logic Programming, 19&20:583–628, 1994. 15. M. Hanus, H. Kuchen, and J.J. Moreno-Navarro. Curry: A Truly Functional Logic Language. In Proc. ILPS’95 Workshop on Visions for the Future of Logic Programming, pages 95–107, 1995. 16. N.D. Jones, C.K. Gomard, and P. Sestoft. Partial Evaluation and Automatic Program Generation. Prentice-Hall, Englewood Cliffs, NJ, 1993. 17. J.W. Klop and A. Middeldorp. Sequentiality in Orthogonal Term Rewriting Systems. Journal of Symbolic Computation, pages 161–195, 1991. 18. J. Komorowski. An Introduction to Partial Deduction. In A. Pettorossi, editor, Meta-Programming in Logic, pages 49–69. Springer LNCS 649, 1992. 19. L. Lafave and J.P. Gallagher. Partial Evaluation of Functional Logic Programs in Rewriting-based Languages. Technical Report CSTR-97-001, Department of Computer Science, University of Bristol, Bristol, England, March 1997. 20. M. Leuschel. The ecce partial deduction system and the dppd library of benchmarks. Tech. Rep., accessible via http://www.cs.kuleuven.ac.be/˜lpai, 1998. 21. M. Leuschel, D. De Schreye, and A. de Waal. A Conceptual Embedding of Folding into Partial Deduction: Towards a Maximal Integration. In M. Maher, editor, Proc. of JICSLP’96, pages 319–332. The MIT Press, Cambridge, MA, 1996. 22. J.W. Lloyd and J.C. Shepherdson. Partial Evaluation in Logic Programming. Journal of Logic Programming, 11:217–242, 1991. 23. R. Loogen, F. L´ opez-Fraguas, and M. Rodr´ıguez-Artalejo. A Demand Driven Computation Strategy for Lazy Narrowing. In J. Penjam and M. Bruynooghe, editors, Proc. of PLILP’93, pages 184–200. Springer LNCS 714, 1993. 24. B. Martens and J. Gallagher. Ensuring Global Termination of Partial Deduction while Allowing Flexible Polyvariance. In L. Sterling, editor, Proc. of ICLP’95, pages 597–611. MIT Press, 1995. 25. J.J. Moreno-Navarro and M. Rodr´ıguez-Artalejo. Logic programming with functions and predicates: the language Babel. J. Logic Program., 12(3):191–224, 1992. 26. P. Padawitz. Computing in Horn Clause Theories, volume 16 of EATCS Monographs on Theoretical Computer Science. Springer-Verlag, Berlin, 1988. 27. A. Pettorossi and M. Proietti. Rules and Strategies for Transforming Functional and Logic Programs. ACM Computing Surveys, 28(2):360–414, 1996. 28. M.H. Sørensen and R. Gl¨ uck. An Algorithm of Generalization in Positive Supercompilation. In Proc. of ILPS’95, pages 465–479. The MIT Press, 1995. 29. M.H. Sørensen, R. Gl¨ uck, and N.D. Jones. A Positive Supercompiler. Journal of Functional Programming, 6(6):811–838, 1996. 30. P.L. Wadler. Deforestation: Transforming programs to eliminate trees. Theoretical Computer Science, 73:231–248, 1990.

Directional Type Inference for Logic Programs Witold Charatonik? and Andreas Podelski Max-Planck-Institut f¨ ur Informatik Im Stadtwald, D-66123 Saarbr¨ ucken {witold;podelski}@mpi-sb.mpg.de

Abstract. We follow the set-based approach to directional types proposed by Aiken and Lakshman [1]. Their type checking algorithm works via set constraint solving and is sound and complete for given discriminative types. We characterize directional types in model-theoretic terms. We present an algorithm for inferring directional types. The directional type that we derive from a logic program P is uniformly at least as precise as any discriminative directional type of P, i.e., any directional type out of the class for which the type checking algorithm of Aiken and Lakshman is sound and complete. We improve their algorithm as well as their lower bound and thereby settle the complexity (Dexptime-complete) of the corresponding problem.

1

Introduction

Directional types form a type system for logic programs which is based on the view of a predicate as a directional procedure which, when applied to a tuple of input terms, generates a tuple of output terms. There is a rich literature on types and directional types for which we can give only some entry points. Directional types occur as predicate profiles in [28], as mode dependencies in [8], and simply as types in [4,2,3]. Our use of the terminology “directional type” stems from [1]. A directional type for a program P assigns input types Ip and output types Op to each predicate p of P. A program can have many directional types. For example, consider the predicate append defined by append([], L, L). append([X|Xs], Y, [X|Z])←append(Xs, Y, Z). We can give this predicate the directional type (list, list, >) → (list, list, list), where list denotes the set of all lists and > is the set of all terms, but also (>, >, list) → (list, list, list), as well as the least precise type (>, >, >) → (>, >, >). A predicate defined by a single fact p(X) has a directional type τ → τ for all types τ . In [1], Aiken and Lakshman present an algorithm for automatic type checking of logic programs wrt. given directional types. The algorithm runs in ?

On leave from University of Wroclaw, Poland. Partially supported by Polish KBN grant 8T11C02913.

G. Levi (Ed.): SAS’98, LNCS 1503, pp. 278–294, 1998. c Springer-Verlag Berlin Heidelberg 1998

Directional Type Inference for Logic Programs

279

NEXPTIME; they show that the problem is DEXPTIME-hard in general and PSPACE-hard for discriminative types. The algorithm works via set constraint solving; its correctness relies on a connection between the well-typedness conditions and the set constraints to which they are translated. The connection is such that the type check is sound and complete for discriminative types (it is still sound for general types). Our results. In this paper, we answer two questions left open in [1]. First, we give an algorithm for inferring directional types. Second, we establish the DEXPTIME-completeness of the problem of directional type checking wrt. discriminative types. To fix just one (the “right” one) directional type for a given logic program, we assume that the program comes with a query which, wlog., consists of one atom main(t). Clearly, the choice of the type makes sense only if the input type Imain for the query predicate main contains at least the expected set of input terms for main. Ideally, among all those directional types T that satisfy this condition, we would like to infer the uniformly (i.e., for the input types and output types of all predicates) most precise one. The uniformly most precise directional type Tmin (P) of a program P together with a specification of the query input terms does exist, as we will show. It is, however, not effectively computable in general. This is naturally the place where abstraction comes in. We can compute a directional type Tsb (P), a regular approximation of Tmin (P) which is defined through the set-based abstraction a` la Heintze and Jaffar [24]. There is no objective criterion to evaluate the quality of the approximation of a non-regular set by a regular one in the sense that the most precise approximation does not exist; this fact applies also to our type inference procedure. We can show, however, that Tsb (P) is uniformly more precise than any discriminative directional type of P, i.e., any directional type out of the class for which the type check of Aiken and Lakshman is sound and complete. The above comparison is interesting for intrinsic reasons and it indicates that our type inference procedure produces “good” directional types. We exploit it furthermore in order to derive a type checking algorithm whose complexity improves upon the one of the original algorithm in [1]. A simple refinement of the arguments given in [1] suffices to make the lower bound more precise. We thus settle the complexity (DEXPTIME-complete) of the problem of directional type checking of logic programs wrt. discriminative types. Technically, our results are based on several basic properties of three kinds of abstraction (and their interrelation): the set-based abstraction (obtained by Cartesian approximation and related to our inference procedure), the set-valued abstraction (obtained by replacing membership constraints with set inclusions and related to the type-checking procedure for arbitrary types), and path closure abstraction (related to the type-checking procedure for discriminative types). These properties, that we collect in Sect. 2, are of general interest; in particular, the abstraction by path closure keeps reappearing (see, e.g., [27,26,31,19,20]). Furthermore, we establish that the directional types of a program P are exactly the models of an associated logic program PI nOut . In fact, PI nOut is a kind of

280

W. Charatonik and A. Podelski

“magic set transformation” (see, e.g., [20]) of P. We obtain our results (and the soundness and completeness results in [1]) by combining the model-theoretic characterization of directional types with the properties of abstractions established in Sect. 2. In fact, by having factored out general properties of abstractions from the aspects proper to directional types, we have maybe given a new view of the results in [1]. Other related work. Rouzaud and Nguyen-Phong [28] describe a type system where types are sets of non-ground terms and express directionality, but these sets must be tuple-distributive. Our types need not be tuple-distributive. In [14], Codish and Demoen infer type dependencies for logic programs. Their techniques (abstract compilation) are quite different from ours and the derived dependencies express all possible input-output relationships. Probably the work of Heintze and Jaffar is the one that is most closely related to ours. This is not only due to the fact that we use set-based analysis [24] to approximate the type program. Some of their papers [23,25] contain examples where they compute for each predicate p a pair of sets Callp and Retp , which can be viewed as (ground) directional type Callp → Retp . We are not aware, however, of a general, formal treatment of directional types inference in their work. Also the work of Gallagher and de Waal [20,21] is closely related to ours. They use a kind of “magic set transformation” to compute query and answer predicates for each predicate in the given program, which is essentially the same as our type program. Then they approximate the success set of the new predicates with ground, regular and tupledistributive sets, which they do not call types. Boye in [5] (see also [6,7]) presents a procedure that infers directional types for logic programs. The procedure is not fully automatic (sometimes it requires an interaction with the user), it requires the set of possible types to be finite, and no complexity analysis is given. In our approach, any regular set of terms is an admissible type; our procedure is fully automatic in the presence of a query for the program (or lower bounds for input types) and it runs in single-exponential time. We refer to [1] for comparison with still other type systems. Most of those interpret types as sets of ground terms, while we interpret types as sets of non-ground terms. Most type systems do not express the directionality of predicates.

2

Abstractions

In this section we discuss three kinds of abstraction of a given program. The motivation for this discussion is the following. In the next section we will define a type program for a given program P. The models of this program are directional types of P. The least model of the set-based abstraction of the type program correspond to the type that we infer. The models of the set-valued abstraction correspond to the directional types that pass the type-check from [1]. The models of the path-closure abstraction correspond to discriminative directional types of P. Later we will use the relations between these abstractions to compare different directional types of a given program.

Directional Type Inference for Logic Programs

281

Preliminaries. We follow the notation and terminology of [1] unless specified otherwise. For example, we use the symbols p, q, p0 , . . . for predicates (instead of f, g, f0 , . . . as in [1]) and write p(t) for predicate atoms (instead of f t). Wlog., all predicates are unary. Terms t are of the form x or f (t1 , . . . , tn ). We may write t[x1 , . . . , xm ] for t if x1 , . . . , xm are the variable occurrences in t (we distinguish between multiple occurrences of the same variable), and t[t1 , . . . , tm ] for the term obtained from t by substituting tj for the occurrence xj . A logic program P is a set of Horn clauses, i.e., implications of the form p0 (t0 ) ← p1 (t1 ), . . . , pn (tn ). A program comes with a query (“the main loop”) which, wlog., is specified by one atom main(t). The interpretation of programs, which is defined as usual, may be viewed as a mapping from predicate symbols to sets of trees. Programs may be viewed as formulas whose free variables are set-valued (and are referred to via predicate symbols), hence as a large class of set constraints. We will not use (positive) set expressions and set constraints as in [1] but, instead, logic programs, for specifying sets of trees as well as for abstraction. Positive set expressions, which denote sets of trees, can be readily translated into equivalent alternating tree automata, which again form the special case of logic programs whose clauses are all of the form p(f (x1 , . . . , xn )) ← p11 (x1 ), . . . , p1m1 (x1 ), . . . , pn1 (xn ), . . . , pnmn (xn ) where x1 , . . . , xn are pairwise different. A non-deterministic tree automaton is the special case where m1 = . . . = mn = 1, i.e., a logic program whose clauses are all of the form p(f (x1 , . . . , xn )) ← p1 (x1 ), . . . , pn (xn ). A set of trees is regular if it can be denoted by a predicate p in the least model of a non-deterministic tree automaton. A uniform program [19] consists of Horn clauses in one of the following two forms. (In a linear term, each variable occurs at most once.) – p(t) ← p1 (x1 ), . . . , pk (xm ) where the term t is linear. – q(x) ← p1 (t1 ), . . . , pm (tm ) where t1 , . . . , tm are any terms over Σ(Var). A uniform program can be transformed (in single-exponential time) into an equivalent non-deterministic tree automaton [19,18,10]. Set-based abstraction. We use set-based analysis in the sense of [24] but in the formulation using logic programs as in [19,18,10] (instead of set constraints as in [24] and [1]). Definition 1 (P # , the set-based abstraction of P). The uniform program P # is obtained from a program P by translating every clause p(t) ← body, whose head term t contains the n variables x1 , . . . , xn , into the (n + 1) clauses 1 mn 1 p(t˜) ← p1 (x11 ), . . . , p1 (xm 1 ), . . . , pn (xn ), . . . , pn (xn )

pi (xi ) ← body

(for i = 1, . . . , n)

where t˜ is obtained from t by replacing the mi different occurrences of variables xi i by different renamings x1i , . . . , xm i , and p1 , . . . , pn are new predicate names.

282

W. Charatonik and A. Podelski

The least model of the program P # expresses the set-based abstraction of P, which is defined in [24] as the least fixpoint of the operator τP . The operator τP is defined via set-based substitutions in [24]; it can also be defined by (for a subset M of the Herbrand base) τP (M ) = {p0 (t0 [x1 θ1 , . . . , xm θm ]) | p0 (t0 ) ← p1 (t1 ), . . . , pn (tn ) ∈ P, θ1 , . . . , θm ground substitutions, p1 (t1 θ1 ), . . . , pn (tn θ1 ) ∈ M .. .

p1 (t1 θm ), . . . , pn (tn θm ) ∈ M }

where x1 , . . . , xm are the variable occurrences in t0 (we distinguish between multiple occurrences of the same variable). As noted in [19], the logical consequence operator associated with P # is equal to the set-based consequence operator; i.e., TP # = τP . We recall that the logical consequence operator associated with the program P is defined by TP (M ) = {p0 (t0 [x1 θ, . . . , xm θ] | p0 (t0 ) ← p1 (t1 ), . . . , pn (tn ) ∈ P, θ ground substitution, p1 (t1 θ), . . . , pn (tn θ) ∈ M }. The set-based abstraction can be formalized in the abstract interpretation framework [16] by the application of the Cartesian approximation C to the semanticsdefining fixpoint operator TP ; i.e., TP # = C(TP ) (see [17]; roughly, C maps a set of tuples to the smallest Cartesian product containing it). Thus, we have TP # = τP = C(TP ). We note the following fact, keeping symmetry with Remarks 2 and 3 on the two other abstractions of TP that we will introduce. Given two set-valued functions F and F 0 , we write F ≤ F 0 if F is smaller than F 0 wrt. to pointwise subset inclusion, i.e., F (x) ⊆ F 0 (x) for all x. Remark 1 (Set-based approximation). The direct-consequence operator associated with P # approximates the one associated with P; i.e., TP ≤ T P # . Proof. Obvious by definition.

2

The following statement will be used for the soundness of our type inference algorithm (Theorem 3). Its converse does, of course, not hold in general (the least models of P may be strictly smaller than the least models of P # ). Proposition 1. Each model M of the set-based abstraction P # of a program P is also a model of P.

Directional Type Inference for Logic Programs

283

Proof. A ground instance of a clause of P is also a ground instance of the 2 corresponding clause of P # . Set-valued abstraction. The second abstraction that we consider is also defined via a program transformation: an atom p(t) is simply replaced by an inclusion t ⊆ p. The transformed program is interpreted over the domain of sets of trees; i.e., the valuations are mappings θ : Var → 2TΣ . These mappings are extended canonically from variables x to terms t; i.e., tθ is a set of trees. We repeat that an interpretation M, i.e., a subset of the Herbrand base, maps predicates p to sets of trees pM = {t ∈ TΣ | p(t) ∈ M}. The inclusion t ⊆ p holds in M under the valuation θ if tθ is a subset of pM . Definition 2 (P ⊆ , the set-valued abstraction of P). Given a program P, its set-valued program abstraction is a program P ⊆ that is interpreted over sets of trees (instead of trees). It is obtained by replacing membership with subset inclusion; i.e., it contains, for each clause p0 (t0 ) ← p1 (t1 ), . . . , pn (tn ) in P, the implication (1) t0 ⊆ p0 ← t1 ⊆ p1 , . . . , tn ⊆ pn . The models of the program P ⊆ are, as one expects, interpretations M (subsets of the Herbrand base) such that all implications all valid in M. An implications of the form (1) is valid in M if for all valuations θ : Var → 2TΣ , if ti θ is a subset of pM i for i = 1, . . . , n then also for i = 0. The models of P ⊆ are the fixpoints of TP ⊆ , the direct consequence operator associated with P ⊆ , which is defined in a way analogous to TP (using set-valued substitutions instead of tree-valued substitutions). Hence, we will be able to use the following remark when we compare models of P # with models of P ⊆ (Proposition 4). Remark 2 (Set-valued approximation). The direct-consequence operator associated with P ⊆ approximates the one associated with P; i.e., TP ≤ TP ⊆ . Proof. If, for some subset M of the Herbrand base, TP (M ) contains the ground atom p0 (t0 ) because M contains the ground atoms p1 (t1 ), . . . , pn (tn ) and p0 (t0 ) ← p1 (t1 ), . . . , pn (tn ) is a ground instance of some clause of P, then the singleton {ti } is a subset of the denotation of pi under M (for i = 1, . . . , n) and, hence, the singleton {t0 } is a subset of the denotation of p0 under TP ⊆ (M ). 2 The next statement underlies the soundness of the type checking procedure of [1], which essentially checks if a given interpretation is a model of the program PI⊆nOut defined in the next section (cf. Theorem 9 in [1]). It says that being a model wrt. P ⊆ is a sufficient condition for being a model of the program P. (The model property of a regular set wrt. P ⊆ is nothing else than an entailment relation between set constraints. The entailment is equivalent to satisfiability of negative constraints and can be tested in NEXPTIME [12]).

284

W. Charatonik and A. Podelski

Proposition 2. Each model M of the set-valued abstraction P ⊆ of a program P is also a model of P. Proof. If c ≡ p(t0 ) ← p1 (t1 ), . . . , pn (tn ) is a ground instance of a clause of P, then {t0 } ⊆ p ← {t1 } ⊆ p1 , . . . , {tn } ⊆ pn is a ground instance of the corresponding implication of P ⊆ (which holds in M if M is a model of P ⊆ , and, thus, c also holds). 2 The converse of the statement above does not hold in general. Take, for example, the program P defined by the clause p(f (x, x)) ← q(f (x, x)) and the four facts q(f (a, a)), q(f (a, b)), q(f (b, a)), q(f (b, b)). Then P ⊆ consists of the implication f (x, x) ⊆ p ← f (x, x) ⊆ q and the four inclusions f (a, a) ⊆ q, f (a, b) ⊆ q, f (b, a) ⊆ q, f (b, b) ⊆ q. Then M = {p(f (a, a))), p(f (b, b)), q(f (a, a)), q(f (a, b)), q(f (b, a)), q(f (b, b))} is a model of P but M is not a model of P ⊆ . (This example transfers, in the essence, Example 1 in [1] from the setting of directional types to a general setting.) The converse of the statement in Proposition 2 does, however, hold in the special case of path closed models (Proposition 3) which we will introduce below. The least fixpoint of TP ⊆ is in general not regular. To see this note that it is equal to the least fixpoint of TP if, for example, P is the length program. This example also shows that the least model of P # is in general not contained in every model of P ⊆ . This is the case, however, in the special case where the model of P ⊆ is path closed (Proposition 4). Path closed models. The following notion of a path-closed set originates from [22]. It is equivalent to other notions occurring in the literature: tupledistributive [27,28], discriminative [1], or deterministic. Definition 3 (Path-closed). A [regular] set of ground terms is called pathclosed if it can be defined by a [finite] deterministic top-down tree automaton. A deterministic finite tree automaton translates to a logic program which does not contain two different clauses with the same head (modulo variable renaming), e.g., p(f (x1 , . . . , xn )) ← p1 (x1 ), . . . , pn (xn ) and p(f (x1 , . . . , xn )) ← p01 (x1 ), . . . , p0n (xn ). A discriminative set expression as defined in [1] translates to a deterministic finite tree automaton, and vice versa. That is, discriminative set expressions denote exactly path-closed regular sets. It is argued in [1] that discriminative set expressions are quite expressive and are used to express commonly used data structures. Note that lists, for example, can be defined by the program with the two clauses l ist(cons(x, y)) ← list(y) and l ist(nil). The following fact is the fundamental property of path closed sets in the context of set constraints (see also Theorem 12 and Lemma 14 in [1]). It will be directly used in Proposition 3. For comparison, take the constraint f (x, y) ⊆ f (a, a) ∪ f (b, b); here, {f (a, a), f (b, b)} is a set that is not path closed, and the union of the (set-valued) solutions θ1 : x, y 7→ {a} and θ2 : x, y 7→ {b} is not a solution. Also, take the constraint f (x, y) ⊆ ∅; here, the union of the solutions over possibly empty sets θ1 : x 7→ {a}, y 7→ ∅and θ2 : x 7→ ∅, y 7→ {a} is not a solution.

Directional Type Inference for Logic Programs

285

Lemma 1. Solutions of conjunctions of inclusions t ⊆ e between terms t interpreted over nonempty sets and expressions e denoting path closed sets of trees are closed under union; i.e., if S is a set of solutions, then Θ defined S by Θ(x) = {θ(x) | θ ∈ S} is again a solution. Proof. The statement follows from the fact (shown, e.g., in [13]) that in the interpretation over nonempty sets, if the upper bounds e denote path closed sets then inclusions of the form f (x1 , . . . , xn ) ⊆ e are equivalent to the conjunction −1 −1 x1 ⊆ f(1) (e) ∧ . . . ∧ xn ⊆ f(n) (e) −1 (e) denotes the set {ti | f (t1 , . . . , tn ) ∈ e}. where f(i)

2

The following statement underlies the completeness of the type checking procedure of [1] for discriminative directional types (see Theorem 12, Lemma 14 and Theorem 15 in [1]). Proposition 3. Each path closed model M of a program P is also a model of P ⊆ , its set-valued abstraction. Proof. Assume that the clause p0 (t0 ) ← p1 (t1 ), . . . , pn (tn ) is valid in M (i.e., it holds under all ground substitutions σ : Var 7→TΣ , under the interpretation of the predicates p0 , p1 , . . . , pn by M), and that θ is a substitution mapping variables to nonempty sets such that t1 θ ⊆ p1 , . . . , tn θ ⊆ pn holds in M. We have to show that also t0 θ ⊆ p0 holds in M. The assumption yields that, for every ground substitution σ : Var 7→TΣ such that σ(x) ∈ θ(x) for all x ∈ Var, t1 σ ∈ p1 , . . . , tn σ ∈ pn holds in M. Thus, also t0 σ ∈ p0 holds in M. Since we have that – t0 σ ∈ p0 is equivalent to t0 σ ¯ ⊆ p0 where σ ¯ is the set substitution defined by σ ¯ (x) = {σ(x)}, – θ is the union of all σ ¯ such that σ(x) ∈ θ(x) for all x ∈ Var, – solutions of t0 ⊆ p0 , where p0 is interpreted by M as a path closed set, are closed under union (Lemma 1), the inclusion t0 ⊆ p0 also holds in M under the substitution θ.

2

Path closure abstraction. The path closure P C of a set M of trees is the smallest path closed set containing M . We consider a third abstraction of the operator TP by composing the path closure P C with TP . We note that we do not know whether the least fixpoint of the operator P C ◦ TP is always a regular set. The following comparison of two of the three abstractions that we have defined so far will be used in the proof of Proposition 4.

286

W. Charatonik and A. Podelski

Remark 3 (Set-based approximation and path closure). The path closure abstraction of the direct-consequence operator of P approximates also its set-based abstraction; i.e., TP # ≤ (P C ◦ TP ). Proof. We use the equality TP # = τP . If p0 (t0 [x1 θ1 , . . . , xm θm ]) ∈ τP (M ) because p1 (t1 θj ), . . . , pn (tn θj ) ∈ M , then p0 (t0 θj ) ∈ TP (M ), for j = 1, . . . , m. But then we have p0 (t0 [x1 θ1 , . . . , xm θm ]) ∈ (P C ◦ TP )(M ). 2 We will use the following statement later in order to compare the directional types obtained by our type inference procedure with the subclass of discriminative directional types for which the type check in [1] is sound and complete. (Note that the path closure of a model of a program is in general not itself a model, and that the least model of P # is in general not contained in the path closure of P). Proposition 4. The least model of P # , the set-based abstraction of a program P, is contained in every path closed model of P ⊆ , the set-valued abstraction of a program P. Proof. By Remark 3, TP # ⊆ (P C ◦ TP ) and thus, by Remark 2, TP # ⊆ (P C ◦ TP ⊆ ). If M is a path-closed model of P ⊆ , then it is also a fixpoint of P C ◦ TP , and hence it contains the least fixpoint of TP # , i.e., the least model of P # . 2

3

Directional Types and Type Programs

A type T is a set of terms t closed under substitution [2]. A ground type is a set of ground terms (i.e., trees), and thus a special case of a type. A term t has type T , in symbols t : T , if t ∈ T . A type judgment is an implication t1 : T1 ∧ . . . ∧ tn : Tn → t0 : T0 that holds under all term substitutions θ : Var → TΣ (Var). Definition 4 (Directional type of a program [8,1]). A directional type of a program P is a family T = (Ip → Op )p∈Pred assigning to each predicate p of P an input type Ip and an output type Op such that, for each clause p0 (t0 ) ← p1 (t1 ), . . . , pn (tn ) of P, the following type judgments hold. t0 : Ip0 → t1 : Ip1 t0 : Ip0 ∧ t1 : Op1 → t2 : Ip2 .. .

t0 : Ip0 ∧ t1 : Op1 ∧ . . . ∧ tn−1 : Opn−1 → tn : Ipn t0 : Ip0 ∧ t1 : Op1 ∧ . . . ∧ tn : Opn → t0 : Op0 We then also say that P is well-typed wrt. T . A program together with its query main(t) is well-typed wrt. T if furthermore the query argument t is well-typed wrt. the input type for main (i.e, the type judgment t : Imain holds).

Directional Type Inference for Logic Programs

287

Definition 5 (Ordering on directional types). We define that T = (Ip → Op )p∈Pred is uniformly more precise than T 0 = (Ip0 → Op0 )p∈Pred if Ip ⊆ Ip0 and Op ⊆ Op0 for all predicates p. The least precise directional type for which any program (possibly together with a query) is well-typed is T> = (> → >)p∈Pred assigning the set of all terms to each input and output type. In the absence of a query, the most precise one is T⊥ = (⊥ → ⊥)p∈Pred assigning the empty set to each input and output type. This changes if, for example, a query of the form main is present (see Sect. 4). Definition 6 (S at(T ), the type of terms satisfying T [1]). Given the ground type T , the set S at(T ) of terms satisfying T is the type S at(T ) = {t ∈ TΣ (Var) | θ(t) ∈ T for all ground substitutions θ : Var → TΣ }. Definition 7 (Discriminative type). A directional type is called discriminative if it is of the form (S at(Ip ) → S at(Op ))p∈Pred , where the sets Ip , Op are path-closed. Remark 4. The clause p0 (t0 ) ← p1 (t1 ), . . . , pn (tn ) is valid in some model M if and only if the type judgment t0 : S at(p0 ) ← t1 : S at(p1 ) ∧ . . . ∧ tn : S at(pn ) holds in M (i.e., under the interpretation of p0 , p1 , . . . , pn by M). Proof. Membership of the application of substitutions to terms in sets of the form S at(E) is defined by the application of ground substitutions to the terms in E. 2 A directional type of the form T = (S at(Ip ) → S at(Op ))p∈Pred , for ground types Ip , Op ⊆ TΣ , satisfies a type judgment if and only if the corresponding directional ground type Tg = (Ip → Op )p∈Pred does. We will next transform the well-typedness condition in Definition 4 into a logic program by replacing t : Ip with the atom pI n (t) and t : Op with pOut (t). Definition 8 (PI nOut , the type program for P). Given a program P, the corresponding type program PI nOut defines an in-predicate pI n and an outpredicate pOut for each predicate p of P. Namely, for every clause p0 (t0 ) ← p1 (t1 ), . . . , pn (tn ) in P, PI nOut contains the n clauses defining in-predicates corresponding to each atom in the body of the clause, pI1n (t1 ) ← pI0n (t0 ) pI2n (t2 ) ← pI0n (t0 ), pOut (t1 ) 1 .. .

(t1 ), . . . , pOut pInn (tn ) ← pI0n (t0 ), pOut 1 n−1 (tn−1 )

and the clause defining the out-predicate corresponding to the head of the clause, (t0 ) ← pI0n (t0 ), pOut (t1 ), . . . , pOut pOut 0 1 n (tn ).

288

W. Charatonik and A. Podelski

If the program P comes together with a query main(t), we add the clause mainI n (t) ← true to PI nOut . The next statement extends naturally to a characterization of the well-typedness of a program together with a query. Theorem 1 (Types and models of type programs). The program P is well-typed wrt. the directional type T = (S at(Ip ) → S at(Op ))p∈Pred (with ground types Ip , Op ) if and only if the subset of the Herbrand base corresponding to T , MT = {pI n (t) | t ∈ Ip } ∪ {pOut (t) | t ∈ Op }, is a model of the type program PI nOut . Proof. The validity of the well-typing conditions under ground substitutions is exactly the logical validity of the clauses of PI nOut in the model MT . The statement then follows by Remark 4. 2 We next define two abstractions of type programs. We will use PI#nOut for type inference (Sect. 4) and PI⊆nOut for type checking (Sect. 5). Given a directional type T , the interpretation of PI⊆nOut by the corresponding subset MT is the set constraint condition which is used in [1] to replace the well-typedness condition in Definition 4. Definition 9 (Abstractions of type programs). The set-based type program PI#nOut is the set-based abstraction of PI nOut , and the set-valued type program PI⊆nOut is the set-valued abstraction of PI nOut ; i.e., PI#nOut = (PI nOut )# , PI⊆nOut = (PI nOut )⊆ . The following direct consequence of Theorem 1 and Propositions 2 and 3 restates the soundness and the conditional completeness of the type check in [1]. Theorem 2 ([Discriminative] types and models of set-valued type programs). A directional type of the form T = (S at(Ip ) → S at(Op ))p∈Pred is a directional type of the program P if the corresponding subset MT of the Herbrand base is a model of PI⊆nOut , the set-valued abstraction of the type program of P. For discriminative directional types T , the converse also holds. Proof. The first part follows from Proposition 2 together with Theorem 1, the second from from Proposition 3 together with Theorem 1. 2

Directional Type Inference for Logic Programs

4

289

Directional Type Inference

We consider three different scenarios in which we may want to infer directional types from a program. (1) The program comes together with a query consisting of one atom main without arguments (and there is a clause main ← p1 (t1 ), . . . , pn (tn ) calling the actual query). In this case, we are interested in inferring the most precise directional type T such that P together with the query main is well-typed. According to the model-theoretic characterization of the well-typedness of a program together with a query (Theorem 1), we must add the clause mainI n ← true to PI nOut . This means that T⊥ = (⊥ → ⊥)p∈Pred is generally not a directional type of a program together with the query main and, hence, it is nontrivial to infer a precise one. (2) The program comes together with a query consisting of one atom main(t) and a lower bound Mmain for the input type of main is given (the user hereby encodes which input terms to the query predicates are expected). In this case, we are interested in inferring the most precise directional type T = (Ip → Op )p∈Pred for P such that the input type for main lies above the lower bound for main, i.e., such that Mmain ⊆ Imain . For example, take the program defining the predicate r everse(x, y) r everse([], []). r everse([X|Xs], Y )←r everse(Xs, Z), append(Z, [X], Y, ). together with the definition of append and with the query r everse(x, y). If the expected input terms are lists (for x) and non-instantiated variables (for y), i.e., the lower bound specified is Mr ev = (list, >), then the type inferred by our algorithm for r everse is (list, >) → (list, list) and the type inferred by our algorithm for append is (list, [>], >) → (list, [>], list), where [>] is the type of all singleton lists. (3) Lower bounds Mp for the input types Ip of all predicates p of P are given. This may be done explicitly for some p and implicitly, with Mp = ∅, for the others. In this case, we are interested in inferring the most precise directional type T = (Ip → Op )p∈Pred for P such that the input types for p lie above the given lower bounds, i.e., such that Mp ⊆ Ip for all p. In a setting with program modules, for example, the lower bounds that are explicitly specified may be the query goals of exported predicates. The scenario (3) resembles the one imagined by Aiken and Lakshman in the conclusion of [1]. Note that in our setting, however, the sets Mp need not already be input types. For example, if the inputs for x and y in append(x, y, z) are expected to be lists of even length only, i.e., the lower bound for Iapp is given by Mapp = (evenlist, evenlist, >), then we would infer the set of all lists as the input type for x, i.e., Iapp = (list, evenlist, >) (note that there are recursive calls to append with lists of odd length). We obtain (1) and (2)as the special case of (3) where the lower bounds for all input types but main are given as the empty set. Hence, it is sufficient to formulate the type inference only for the case (3).

290

W. Charatonik and A. Podelski

Definition 10 (Inference of the set-based directional type Tsb (P, (Mp )p∈Pred )). Given a program P and a family of lower bounds Mp for the input types of the predicates p of the program, we infer the set-based directional type Tsb (P, (Mp )p∈Pred ) = (S at(Ip ) → S at(Op ))p∈Pred where Ip and Op are the denotations of the predicates pI n and pOut in the least model of the program PI#nOut ∪ {pI n (x) ← Mp (x) | p ∈ Pred}.

(2)

The definition of type inference above leaves open in which formalism the lower bounds are specified and how the inferred directional types are presented to the user. There is a wide variety of formalisms that coincide in the expressive power of regular sets of trees and that can be effectively translated one into another and, hence, for which our type inference yields an effective procedure. More concretely, we propose to represent the lower bounds Mp through logic programs in restricted syntax (see Sect. 2) that corresponds directly to alternating tree automata (and also to the positive set expressions considered in [1]; an even more restricted syntax corresponds to non-deterministic tree automata and to positive set expressions without intersection). We attach these logic programs to the program in (2) as definitions of the sets Mp . Then, we can apply one of the known algorithms (see, e.g., [19,18,11,10]) in order to compute, in single-exponential time (in the size of the program P and the programs for the sets Mp ), a logic program that corresponds to a non-deterministic tree automaton and that is equivalent to the program in (2) wrt. the least model and, thus, represents Tsb (P, (Mp )p∈Pred ). The representation of sets by a non-deterministic tree automaton is a good representation in the sense that, for example, the test of emptiness can be done in linear time. If, in Definition 10, we replace PI#nOut with PI nOut , then we obtain the uniformly most precise directional type Tmin (P) wrt. given lower bounds for input types. In general, we cannot effectively compute Tmin (P) (e.g., test emptiness of its input and output types). We repeat that the following comparison of the set-based directional type of P with discriminative directional types of P is interesting because these form the subclass of directional types for which the type check in [1] is sound and complete. Theorem 3 (Soundness and Quality of Type Inference). The program P is well-typed wrt. the set-based directional type Tsb (P, (Mp )p∈Pred ). Moreover, this type is uniformly more precise than every discriminative directional type of P whose family of input types contains (Mp )p∈Pred . Proof. The two statements are direct consequences of Propositions 1 and 4, respectively, together with Theorem 1. 2

Directional Type Inference for Logic Programs

5

291

Complexity of Directional Type Checking

Theorem 4. The complexity of directional type checking of logic programs for discriminative types is DEXPTIME-complete. Proof. The DEXPTIME-hardness follows from refining the argument in the proof of Theorem 18 in [1] with the two facts that (1) discriminative types are denoted by path closed sets, and (2) testing the non-emptiness of a sequence of n tree automata is DEXPTIME-hard even if the automata are restricted to deterministic ones (which recognize exactly path closed sets) [29]. We obtain a type checking algorithm in single-exponential time as follows. The input is the program P and the discriminative directional type T = (S at(Ip ) → S at(Op )) where all types Ip and Op are given by ground set expressions (a special case of which are regular tree expressions or non-deterministic tree automata). Ground set expressions correspond to alternating tree automata which are self-dual; i.e., fp representing the complement by we can obtain ground set expressions Iep and O a syntactic transformation in linear time. We can translate ground expressions into logic programs defining predicates pIp and pOp such that they denote Ip and e Op wrt. the least-model semantics (and similarly predicates pIe and pO f for Ip p

p

fp ). and O We now use one of the well-known single-exponential time algorithms [19,18, 11,10] to compute (the non-deterministic tree automaton representing) the least model of the set-based abstraction P # of a logic program P. We apply such an algorithm to the program PI#nOut (T ) that we obtain from PI#nOut by adding the clauses pI n (x) ← pIp (x) and pOut (x) ← pOp (x) and the logic programs defining pIp and pOp . We use the result in order to test whether the denotations of pI n and pOut under the least model of the program PI nOut (T ) are exactly Ip and Op . This holds if and only if T is a directional type of P (otherwise, we have a proper inclusion for at least one p; see the correctness proof below). The test works by testing whether the intersection of (the non-deterministic tree automaton representing) the denotations of pI n and pOut under the least model of the program PI nOut (T ) with the complements pIe and pO f is empty. This can be p

p

done in one pass by taking the conjunction of PI nOut (T ) with the logic programs defining pIe and pO f and testing the emptiness of new predicates defined as the p

p

intersection of pIp and pIe (and pOp and pO fp ). p The correctness of the algorithm follows with Theorem 1 and Proposition 1 and 4. In detail: Given a discriminative directional type T and a program P, we have that P is well-typed wrt. T iff the corresponding subset MT of the Herbrand base is a (path closed) model of PI nOut by Theorem 1. If this is the case then MT is equal to the least model M0 of PI nOut (T ), since MT ⊆ M0 by definition of PI nOut (T ), and MT ⊇ M0 holds by Proposition 4 (note that MT

292

W. Charatonik and A. Podelski

is a path closed model of PI nOut and of the additional clauses translating T and, thus, contains the least one). On the other hand, if MT is equal to the least model of PI#nOut (T ), then T is a directional type by Proposition 1 and Theorem 1. 2 We note that the procedure above yields a semi-test (in the same sense as the one in [1]) for well-typedness wrt. the class of general directional types, since the equivalence between the inferred type and the given one is a sufficient (but generally not necessary) condition for well-typedness wrt. the given type. We repeat that it implements a full test wrt. discriminative directional types.

6

Future Work

One of the obvious directions for future work is an implementation. We already have a working prototype implementation on top of the saturation-based theorem prover SPASS [30]. The first results are very promising; due to specific theorem-proving techniques like powerful redundancy criteria, one obtains a decision procedure for the emptiness test that is quite efficient on our examples. We ran our prototype on the benchmark programs from [9]. We were able to detect emptiness of the main predicate for six of the twelve programs, while a similar method of Gallagher succeeded in three cases. We believe that by using the tree-automata techniques suggested in [18] together with automata minimization (and conversion to the syntax of ground set expressions), we can further improve the efficiency of our implementation and the readability of the output. Acknowledgments. We thank David McAllester for turning us on to magic sets and thereby to directional types. We thank Harald Ganzinger for useful comments.

References 1. A. Aiken and T. K. Lakshman. Directional type checking of logic programs. In B. L. Charlier, editor, 1st International Symposium on Static Analysis, volume 864 of Lecture Notes in Computer Science, pages 43–60, Namur, Belgium, Sept. 1994. Springer Verlag. 2. K. R. Apt. Declarative programming in Prolog. In D. Miller, editor, Logic Programming - Proceedings of the 1993 International Symposium, pages 12–35, Vancouver, Canada, 1993. The MIT Press. 3. K. R. Apt. Program verification and Prolog. In E. B¨ orger, editor, Specification and Validation methods for Programming languages and systems, pages 55–95. Oxford University Press, 1995. 4. K. R. Apt and S. Etalle. On the unification free Prolog programs. In A. M. Borzyszkowski and S. Sokolowski, editors, Mathematical Foundations of Computer Science 1993, 18th International Symposium, volume 711 of lncs, pages 1–19, Gdansk, Poland, 30 Aug.– 3 Sept. 1993. Springer.

Directional Type Inference for Logic Programs

293

5. J. Boye. Directional Types in Logic Programming. PhD thesis, Department of Computer and Information Science, Link¨ oping University, 1996. 6. J. Boye and J. Maluszynski. Two aspects of directional types. In L. Sterling, editor, Proceedings of the 12th International Conference on Logic Programming, pages 747–764, Cambridge, June13–18 1995. MIT Press. 7. J. Boye and J. Maluszynski. Directional types and the annotation method. Journal of Logic Programming, 33(3):179–220, Dec. 1997. 8. F. Bronsard, T. K. Lakshman, and U. S. Reddy. A framework of directionality for proving termination of logic programs. In K. Apt, editor, Proceedings of the Joint International Conference and Symposium on Logic Programming, pages 321–335, Washington, USA, 1992. The MIT Press. 9. M. Bruynooghe, H. Vandecasteele, A. de Waal, and M. Denecker. Detecting unsolvable queries for definite logic programs. In Proceedings of the Symposium on Programming Language Implementation and Logic Programming (PLILP’98), LNCS. Springer-Verlag, 1998. to appear. 10. W. Charatonik, D. McAllester, D. Niwi´ nski, A. Podelski, and I. Walukiewicz. The Horn mu-calculus. To appear in Vaughan Pratt, editor, Proceedings of the 13th IEEE Annual Symposium on Logic in Computer Science. 11. W. Charatonik, D. McAllester, and A. Podelski. Computing the least and the greatest model of the set-based abstraction of logic programs. Presented at the Dagstuhl Workshop on Tree Automata, October 1997. 12. W. Charatonik and L. Pacholski. Negative set constraints with equality. In Ninth Annual IEEE Symposium on Logic in Computer Science, pages 128–136, 1994. 13. W. Charatonik and A. Podelski. Set constraints for greatest models. Technical Report MPI-I-97-2-004, Max-Planck-Institut f¨ ur Informatik, April 1997. www.mpisb.mpg.de/˜podelski/papers/greatest.html. 14. M. Codish and B. Demoen. Deriving polymorphic type dependencies for logic programs using multiple incarnations of prop. In B. L. Charlier, editor, Proceedings of the First International Static Analysis Symposium, Lecture Notes in Computer Science 864, pages 281–296. Springer Verlag, 1994. 15. M. Comini, G. Levi, M. C. Meo, and G. Vitiello. Proving properties of logic programs by abstract diagnosis. In M. Dam, editor, Analysis and Verification of Multiple-Agent Languages, volume 1192 of LNCS, pages 22–50. Springer-Verlag, June 1996. 16. P. Cousot and R. Cousot. Inductive definitions, semantics and abstract interpretation. In Proc. POPL ’92, pages 83–94. ACM Press, 1992. 17. P. Cousot and R. Cousot. Formal language, grammar and set-constraint-based program analysis by abstract interpretation. In Record of FPCA ’95 - Conference on Functional Programming and Computer Architecture, pages 170–181, La Jolla, California, USA, 25-28 June 1995. SIGPLAN/SIGARCH/WG2.8, ACM Press, New York, USA. 18. P. Devienne, J.-M. Talbot, and S. Tison. Set-based analysis for logic programming and tree automata. In Proceedings of the Static Analysis Symposium, SAS’97, volume 1302 of LNCS, pages 127–140. Springer-Verlag, 1997. 19. T. Fr¨ uhwirth, E. Shapiro, M. Vardi, and E. Yardeni. Logic programs as types for logic programs. In Sixth Annual IEEE Symposium on Logic in Computer Science, pages 300–309, July 1991. 20. J. Gallagher and D. A. de Waal. Regular approximations of logic programs and their uses. Technical Report CSTR-92-06, Department of Computer Science, University of Bristol, 1992.

294

W. Charatonik and A. Podelski

21. J. Gallagher and D. A. de Waal. Fast and precise regular approximations of logic programs. In P. V. Hentenryck, editor, Proceedings of the Eleventh International Conference on Logic Programming, pages 599–613, Santa Margherita Ligure, Italy, 1994. The MIT Press. 22. F. G´ecseg and M. Steinby. Tree Automata. Akademiai Kiado, 1984. 23. N. Heintze. Practical aspects of set based analysis. In K. Apt, editor, Proceedings of the Joint International Conference and Symposium on Logic Programming, pages 765–779, Washington, USA, 1992. The MIT Press. 24. N. Heintze and J. Jaffar. A finite presentation theorem for approximating logic programs. In Seventeenth Annual ACM Symposium on Principles of Programming Languages, pages 197–209, January 1990. 25. N. Heintze and J. Jaffar. Set constraints and set-based analysis. In Proceedings of the Workshop on Principles and Practice of Constraint Programming, LNCS 874, pages 281–298. Springer-Verlag, 1994. 26. G. Janssens and M. Bruynooghe. Deriving descriptions of possible values of program variables. Journal of Logic Programming, 13(2-3):205–258, 1992. 27. P. Mishra. Towards a theory of types in Prolog. In IEEE International Symposium on Logic Programming, pages 289–298, 1984. 28. Y. Rouzaud and L. Nguyen-Phuong. Integrating modes and subtypes into a Prolog type-checker. In K. Apt, editor, Proceedings of the Joint International Conference and Symposium on Logic Programming, pages 85–97, Washington, USA, 1992. The MIT Press. 29. H. Seidl. Haskell overloading is DEXPTIME-complete. Information Processing Letters, 52:57–60, 1994. 30. C. Weidenbach. Spass version 0.49. Journal of Automated Reasoning, 18(2):247– 252, 1997. 31. E. Yardeni and E. Shapiro. A type system for logic programs. In E. Shapiro, editor, Concurrent Prolog, volume 2, chapter 28, pages 211–244. The MIT Press, 1987.

Finite Subtype Inference with Explicit Polymorphism Dominic Duggan Department of Computer Science, Stevens Institute of Technology, Castle Point on the Hudson, Hoboken, New Jersey 07030. [email protected]

Abstract. Finite subtype inference occupies a middle ground between HindleyMilner type inference (as in ML) and subtype inference with recursively constrained types. It refers to subtype inference where only finite types are allowed as solutions. This approach avoids some open problems with general subtype inference, and has practical motivation where recursively constrained types are not appropriate. This paper presents algorithms for finite subtype inference, including checking for entailment of inferred types against explicitly declared polymorphic types. This resolves for finite types a problem that is still open for recursively constrained types. Some motivation for this work, particularly for finite types and explicit polymorphism, is in providing subtype inference for first-class container objects with polymorphic methods.

1

Introduction

Type inference is the process of statically type-checking a program where some or all of the type information has been omitted from the program text. ML and Haskell are examples of programming languages where type inference has been a spectacular success. The particular flavor of type inference used by ML and Haskell is HindleyMilner type inference [14]. The type-checker accumulates equality constraints via a tree walk of the abstract syntax tree, and then uses a unification algorithm to compute a (most general) unifying substitution for these constraints. More recently attention has been focused on subtype inference [2,3,17,7,19,22]. With this work, the type-checker accumulates subtype constraints while traversing the abstract syntax tree, and then applies a constraint solver to check these constraints for consistency. Pottier [19] and Smith and Trifonov [22] have considered the problem of entailment in these type systems, which is important for example in interface matching. Subtype inference continues to be an important avenue of research, particularly in simplifying inferred types to make them practically useful. Hindley-Milner type inference and subtype inference represent two extremes in the type inference continuum: G. Levi (Ed.): SAS’98, LNCS 1503, pp. 295–310, 1998. c Springer-Verlag Berlin Heidelberg 1998

296

D. Duggan

Hindley-Milner Finite Subtype Subtype Equality Subtyping Subtyping Finite types Finite types Infinite types Inferred monotypes Inferred monotypes Inferred monotypes Inferred polytypes Inferred polytypes Inferred polytypes Specified polytypes∗ Between these two extremes, there is an intermediate point: finite subtype inference. While this alternative allows subtyping and type subsumption, it does not assume that types are potentially infinite trees (as with the most recent work on subtype inference). Why should we consider subtype inference with finite types? It is worth recalling why ML for example does not allow circular types (types as potentially infinite trees). The problem was pointed out by Solomon [21]: the problem of deciding the equality of parameterized recursive types is equivalent to the problem of deciding the equality of deterministic context-free languages (DCFLs), which is still after several decades an open problem. This problem is avoided in ML type inference by making the folding and unfolding of recursive types explicit (using data constructors and pattern-matching, respectively), so that circular types are not needed. A motivation for infinite types in subtype inference is to support objects with recursive interfaces. However the problem discovered by Solomon also holds for recursive interfaces for container objects. Consider for example a set object with interface: set(α) = {map : ∀β.(α → β) → set(β), product : ∀β.set(β) → set(α ∗ β), power : ∀β.unit → set(set(α)) All of the methods in this interface are examples of non-regular recursion in the object interface. In a companion paper [5], we have developed an object design for ML-like languages that avoids this problem. The approach there is to again make the folding and unfolding of recursive object interfaces explicit, in object creation and method invocation, respectively. This work introduces the possibility that circular types for recursive object interfaces, while useful for simple objects, may not be so useful for container objects. Early work on subtype inference considered atomic subtyping with finite types [15,12,9]. However there are non-trivial differences between finite subtype inference with atomic subtyping and with record containment. For example even checking for finiteness is markedly different, as elaborated upon in Sect. 5. Another design point is whether polymorphic types should be inferred or specified. All of the work so far assumes that polymorphic types are inferred. The disadvantage of these approaches is that the inferred types are large and complex, diminishing their practical usefulness, despite recent work on simplifying inferred types [7,19,22]. One way to avoid this problem is to require that the programmer provide explicit interfaces for polymorphic functions. This approach introduces fresh technical complications of its own. In Hindley-Milner type inference, mixed-prefix unification has been used to control scoping of type variables with explicit polymorphic type declarations (an idea originally used by Leroy and Mauny [11], and subsequently rediscovered by Odersky and L¨aufer

Finite Subtype Inference with Explicit Polymorphism

297

[16]). In this paper we extend subtype inference with constraint-solving under a mixed prefix, in order to support subtype inference with explicit polymorphism. Explicit polymorphism also derives motivation from our work on container objects with recursive interfaces [5]. We avoid the problems with first-class polymorphism in Hindley-Milner type inference, by requiring explicit type specifications on polymorphic methods. This is similar to the use of universal types to incorporate impredicativity into Hindley-Milner type inference [20,16,10], but tied to the object system instead of to datatypes (again because we are concerned with container objects with polymorphic methods). Even if we are not concerned with polymorphic methods, explicit polymorphism is required if we wish to provide a type-checking algorithm for the Abadi and Cardelli object type system, for example [1]. In that type system, the type rule for method update is structural [1, Sect. 16.2], meaning that the new method definition must be parametric in the type of self (the type of self is a type parameter constrained by the object interface). Explicit polymorphism requires that it be possible to check for entailment of inferred types from declared types. For infinite types this is problematic. Although incomplete algorithms have been published [19,22], the decidability of entailment remains open [22]. In this paper we demonstrate that entailment is decidable for finite subtyping, giving further motivation for our approach. Sect. 2 introduces our type system. We do not overburden the paper with any details of the object system mentioned earlier [5], but present a familiar ML-like language with record-based subtyping and explicit polymorphism. Sect. 3 provides the type inference algorithm. Sect. 4 provides algorithms for checking consistency and entailment; these algorithms must be defined mutually recursively. Sect. 5 considers the check for finite solutions; perhaps surprisingly, this check must be done after consistency and entailment checking. Sect. 6 considers the use of mixed prefix constraint-solving to check that there are well-scoped solutions. Finally Sect. 8 considers further related work and provides conclusions.

2 Type System The mini-language we consider is a language with functions, pairs and records. Subtyping is based on containment between record types. This is extended contravariantly to function types and covariantly to product types. Polymorphic types allow quantified type variables to be constrained by upper bounds. We use ∀αn

Our partners will collect data and use cookies for ad personalization and measurement. Learn how we and our ad partner Google, collect and use data. Agree & close