Lecture Notes in Computer Science 1912

Edited by G. Goos, J. Hartmanis and J. van Leeuwen

3
Berlin
Heidelberg
New York
Barcelona
Hong Kong
London
Milan
Paris
Singapore
Tokyo
Yuri Gurevich Philipp W. Kutter
Martin Odersky Lothar Thiele (Eds.)

Abstract State Machines

Theory and Applications
International Workshop, ASM 2000
Monte Verità, Switzerland, March 19-24, 2000
Proceedings

13
Series Editors

Gerhard Goos, Karlsruhe University, Germany

Juris Hartmanis, Cornell University, NY, USA
Jan van Leeuwen, Utrecht University, The Netherlands

Volume Editors

Yuri Gurevich
Microsoft Research
One Microsoft Way, Redmond, WA 98052, USA
E-mail: [email protected]
Philipp W. Kutter
ETH Zürich, Institute TIK
Gloriastr. 35, 8092 Zürich, Switzerland
E-mail: [email protected]
Martin Odersky
EPFL Lausanne, DI
IN-Ecublens, 1015 Lausanne, Switzerland
E-mail: [email protected]
Lothar Thiele
ETH Zürich, Institute TIK
Gloriastr. 35, 8092 Zürich, Switzerland
E-mail: [email protected]
Cataloging-in-Publication Data applied for

Die Deutsche Bibliothek - CIP-Einheitsaufnahme

Abstract state machines : theory and applications ; international
workshop ; proceedings / ASM 2000, Monte Verità, Switzerland, March
19 - 24, 2000. Yuri Gurevich . . . (ed.). - Berlin ; Heidelberg ; New York ;
Barcelona ; Hong Kong ; London ; Milan ; Paris ; Singapore ; Tokyo :
Springer, 2000
(Lecture notes in computer science ; Vol. 1912)
ISBN 3-540-67959-6

CR Subject Classification (1998): F.3, D.2, F.4.1, D.3, H.2.3

ISSN 0302-9743
ISBN 3-540-67959-6 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.
Springer-Verlag Berlin Heidelberg New York
a member of BertelsmannSpringer Science+Business Media GmbH

c Springer-Verlag Berlin Heidelberg 2000
Printed in Germany
Typesetting: Camera-ready by author, data conversion by PTP-Berlin, Stefan Sossna
Printed on acid-free paper SPIN: 10722735 06/3142 543210
 Preface

The ASM 2000 workshop was held in the conference center of the Swiss Federal
Institute of Technology (ETH) at Monte Verità, Canton Ticino, March 19-24,
2000.
The ASM formalism was proposed together with the thesis that it is suitable
to model arbitrary computer systems on arbitrary abstraction levels. ASMs have
been successfully used to analyze and specify various hardware and software
systems including numerous computer languages.
The aim of the workshop was to bring together domain-experts, using ASMs
as a practical specification method, and theorists working with ASMs and related
methods. In addition the workshop served as a forum on theoretical and practical
topics that relate to ASMs in a broad sense. Three tutorials including hands-on
experience with tools were organized by U. Glässer and G. del Castillo (on the
topic “Specifying Concurrent Systems with ASMs”), H. Rüss and N. Shankar
(on the topic “A Tutorial Introduction to PVS”), M. Anlauff, P.W. Kutter, and
A. Pierantonio (on the topic “Developing Domain Specific Languages”).
In response to the organization committee’s call for papers, 30 papers were
submitted, each of which was independently reviewed by four members of the
program committee. This volume presents a selection of 12 of the refereed papers
and two reports on industrial ASM application at Siemens AG and Microsoft
Research, together with contributions based on the invited talks given by A.
Blass (University of Michigan), E. Börger (University of Pisa), G. Goos (Univer-
sity of Karlsruhe), M. Odersky (Swiss Federal Institute of Technology (EPFL),
Lausanne), W. Reisig (Humboldt University Berlin), and N. Shankar (SRI In-
ternational). The introduction written by E. Börger gives an overview on ASM
research from the beginning to the present.
On behalf of the program committee, we would like to express our apprecia-
tion to the six lecturers who accepted our invitation to speak, to the tutorial
organizers, to all the authors who submitted papers to ASM 2000 and to the
Centro Stefano Franscini (CSF) at Monte Verità and the local organizers who
made the workshop possible.

June 2000 Y.Gurevich, P.W.Kutter, M.Odersky, and L. Thiele

 Organization

ASM 2000 took place from March 19th to 24th at the conference center Monte
Verità of the Swiss Federal Institute of Technology (ETH).

Organization Committee

Yuri Gurevich (Microsoft Research, Redmond, WA, USA)

Philipp W. Kutter (ETH, Zürich, Switzerland)
Martin Odersky (EPFL, Lausanne, Switzerland)
Lothar Thiele (ETH, Zürich, Switzerland)

Program Committee

Andreas Blass (University of Michigan, USA)

Egon Börger (University of Pisa, Italy)
Uwe Glässer (HNI Paderborn, Germany)
Carla P. Gomes (Cornel University/Rome Labs, USA)
Georg Gottlob (TU Wien, Austria)
Erich Grädel (University of Aachen, Germany)
Irene Guessarian (LIAFA/University Paris6, France)
Yuri Gurevich (co-chair, Microsoft Research, USA)
Jim Huggins (Kettering University, USA)
Stefan Jähnichen (GMD FIRST Berlin, Germany)
Hans Langmaack (University of Kiehl, Germany)
Larry Moss (Indiana University, USA)
Peter Mosses (BRICS, Denmark)
Martin Odersky (co-chair, EPFL, Switzerland)
Alfonso Pierantonio (University of L’Aquila, Italy)
Arnd Poetzsch-Heffter (FernUni Hagen, Germany)
Elvinia Riccobene (University of Catania, Italy)
Dean Rosenzweig (University of Zagreb, Croatia)
Harald Ruess (SRI International, USA)
Daniel Schweizer (UBS Zürich, Switzerland)
Anatol Slissenko (University Paris 12, France)
Lothar Thiele (co-chair, ETH Zürich, Switzerland)
Richard Waldinger (SRI International, USA)
Alexandre V. Zamulin (Russian Academy of Science, Russia)
Wolf Zimmermann (University of Karlsruhe, Germany)
VIII Organization

Local Organizers
Bernhard Bron (ETH Zürich, Switzerland)
Samarjit Chakraborty (ETH Zürich, Switzerland)
Christoph Denzler (ETH Zürich, Switzerland)
Felix Etter (ETH Zürich, Switzerland)
Monica Fricker (ETH Zürich, Switzerland)
Philipp W. Kutter (ETH Zürich, Switzerland)
Milan Tadjan (ETH Zürich, Switzerland)

Sponsoring Institutions
Swiss Federal Institute of Technology, Zürich, Switzerland

Microsoft Research, Redmond, WA, USA

Swiss National Science Foundation (SNF)

Blue Capital Inc., Zürich, Switzerland

 Table of Contents

Introduction
Abstract State Machines at the Cusp of the Millennium . . . . . . . . . . . . . . . . . 1
E. Börger (Univ. of Pisa)

Mathematical Foundations
Abstract State Machines and Pure Mathematics . . . . . . . . . . . . . . . . . . . . . . . 9
A. Blass (Univ. of Michigan)

Abstract State Machines and Computationally Complete Query Languages 22

A. Blass (Univ. of Michigan), Y. Gurevich (Microsoft Research and
Univ. of Michigan), and J. Van den Bussche (Limburg Univ.)

On Verification of Refinements of Timed Distributed Algorithms . . . . . . . . . 34

J.Cohen (Univ. Paris-12) and A. Slissenko (Univ. Paris-12 and
Russian Academy of Science)

Abstract State Machine Languages

Objects + Views = Components? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
M. Odersky (EPFL Lausanne)

Xasm – An Extensible, Component-Based ASM Language . . . . . . . . . . . . . . 69

M. Anlauff (GMD First, Berlin)

Generic Facilities in Object-Oriented ASMs . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

A. V. Zamulin (Sibirian Branch of Russian Acad. of Sci.)

Distribution and Concurrency

Towards an ASM Thesis for Unconventional Algorithms . . . . . . . . . . . . . . . . 112
W. Reisig (Humboldt-Univ. Berlin)

Partially Ordered Runs: A Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

Y. Gurevich (Microsoft Research) and D. Rosenzweig (Univ. of Zagreb)

Investigating Java Concurrency Using Abstract State Machines . . . . . . . . . . 151

Y. Gurevich (Microsoft Research), W. Schulte (Microsoft Research),
and C. Wallace (Univ. of Delaware)

Compilers and Semantics

Verifying Compilers and ASMs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
G. Goos and W. Zimmermann (Univ. Karlsruhe)
X Table of Contents

An ASM Dynamic Semantics for Standard ML . . . . . . . . . . . . . . . . . . . . . . . . 203

S.C. Cater and J.K. Huggins (Kettering Univ.)

Modeling the Dynamics of UML State Machines . . . . . . . . . . . . . . . . . . . . . . . 223

E. Börger (Univ. of Pisa), A. Cavarra, and E. Riccobene
(Univ. of Catania)
On the Formal Semantics of SDL-2000: A Compilation Approach Based
on an Abstract SDL Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
R. Eschbach (Univ. of Kaiserslautern), U. Glässer (Univ. of
Paderborn), R. Gotzhein (Univ. of Kaiserslautern), and A. Prinz
(Univ. of Berlin)
Description and Simulation of Microprocessor Instruction Sets . . . . . . . . . . . 266
J. Teich (Univ. of Paderborn), P.W. Kutter (FIT Zürich), and
R. Weper (Univ. of Paderborn)

Automatic Verification and Model Checking

Symbolic Analysis of Transition Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
N. Shankar (SRI International)
Encoding Abstract State Machines in PVS . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
A. Gargantini (Politecnico di Milano) and E. Riccobene (Universita di
Catania)
Model Checking Abstract State Machines and Beyond . . . . . . . . . . . . . . . . . 323
M. Spielmann (RWTH Aachen)
Towards a Methodology for Model Checking ASM: Lessons Learned from
the FLASH Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
K. Winter (GMD FIRST)

Industrial Applications
Report on a Practical Application of ASMs in Software Design . . . . . . . . . . 361
E. Börger (Univ. of Pisa), P. Päppinghaus, and J. Schmid (Siemens
AG, Munich)
Using Abstract State Machines at Microsoft: A Case Study . . . . . . . . . . . . . . 367
M. Barnett, E. Börger, Y. Gurevich, W. Schulte, and M. Veanes
(Microsoft Research)

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381

 Abstract State Machines
at the Cusp of the Millenium

Egon Börger

Università di Pisa, Dipartimento di Informatica, I-56125 Pisa, Italy

[email protected] (Visiting Microsoft Research, Redmond)

Abstract. The ASM’2000 Workshop marks for the ASM method the
transition from its adolescence to the maturation period. The goals which
have been achieved open new frontiers and put us into the position to
embark on new challenges.

1 The Start at the End of the Twentieth Century

We went a long way since the Spring of 1987 when Yuri Gurevich visited Pisa
and, in a series of lectures on the fundamental problem of semantics of program-
ming languages, presented the world première of the concept of ASMs (then
called dynamic/evolving structures/algebras). He gave the main motivation: re-
consider Turing’s thesis in the light of the problem of semantics of programs. He
illustrated his ideas with examples, in particular specifications of Turing machi-
nes, stack machines and some Pascal programs. He gave also proofs of simple
properties of these programs. This material appeared a year later in [22]. It was
preceded by the first appearance of the ASM Thesis, in embryo in a 1984 tech-
nical report [20], and fully spelled out in a notice presented on May 13 of 1985
to the American Mathematical Society [21]. It was accompanied by the first
real-world application, namely the dynamic semantics of MODULA-2 [26], and
shortly afterwards followed by the ASM treatment of concurrency used to de-
fine the semantics of OCCAM [27], which was presented by Gurevich in another
series of lectures in Pisa in May 1990. Since then the concept of Abstract State
Machines essentially remained stable [23,24]1 and triggered hundreds of publica-
tions in various domains including finite model theory, complexity theory and
numerous areas of applied computer science, in particular programming langu-
ages, database query languages, protocols, architectures and embedded control
software [1].
The first attempts to put the bold ASM thesis to the test were focussed
on the problem of the dynamics of programming languages known to us, and
we came from a purely theoretical background and had no practical, let alone
1
The initially present construct to shrink domains, which was motivated by concerns
about resource bounds, was abandoned because it belongs to garbage collection
rather than to high-level specification. Some technical variation was later introduced
concerning the treatment of non determinism and of inconsistent update sets.

Y. Gurevich et al. (Eds.): ASM 2000, LNCS 1912, pp. 1–8, 2000.
c Springer-Verlag Berlin Heidelberg 2000

2 E. Börger

industrial, experience. What came out of that is a practical method which ex-
ploited ASMs for the development of a full-fledged refinement approach which
enabled us to rigorously define and analyse the dynamic semantics of real-life
programming languages and their implementation on virtual or real machines.
By now, the covered programming paradigms include the paradigms of all the
major modern programming languages. The method of constructing ground mo-
dels, described in [5] where the name primary model (rather than ground model)
was used, proved to be mature and was chosen for standardization purposes by
the International Standards Organization, see [8,9,11,12,3], and by the Interna-
tional Telecommunication Union, as reported in [19] and in these Proceedings
[18].
At the next step, the method was tried out for the specification and verifi-
cation of machine architectures and protocols. Eventually this was followed by
applications to software engineering. Here one starts by constructing a ground
model for a proposed (or, in the case of reverse engineering, for an existing)
software system. Through the entire design process, one refines and/or coarsens
the models linking the high level models in a traceable and inspectable way to
executable code; see the survey in [6].
The key for the surprisingly fast success of the method lies in (a) the two
constituents of the notion of ASM, namely being abstract (Abstract State) and
operational (Abstract Machine) (see the section Abstract Machines + Abstract
State = ASM in [6] for the historical reconstruction of the confluence of these
two concepts2 ) and (b) in the systematic way it offers for practical software
development to separate different concerns.
The abstract character of ASMs allows one, on one side, to tailor the mo-
dels to the needs or purposes of the design, and, on the other side, to make their
rigorous analysis feasible. The latter is due to the freedom to use those proof
methods which are appropriate for the present discourse. In other words, the ab-
straction mechanism, built into the notion of ASM, permits one to make real the
old dream of well documented and controllable hierarchical system development.
Use ASMs to do the following:

– Make the faithfulness of the models, with respect to the design intentions,
checkable by direct inspection (falsifiable in the Popperian sense). This holds
in particular in requirements engineering for the faithfulness of the ground
model with respect to the informally given requirements. The faithfulness
becomes checkable by the application domain expert once an ASM model is
there (see [16]).
– Link, by hierarchies of stepwise refinements, the high-level definition in a
transparent way to its implementation. Here each refinement step is supposed
to reflect design decisions one wants to document for future use, e.g. for
maintenance purposes or for changes by extensions and modifications.
2
Notice that the ASM abstract machine is different from the familiar abstract machi-
nes. It has built-in parallelism which allows one to abstract from irrelevant sequen-
tialization.
 Abstract State Machines at the Cusp of the Millenium 3

– Make the (mathematical, possibly machine checked) justification of the cor-

rectness of a complex design feasible. ASMs offer the necessary rigorous
framework for the analysis of run-time properties, at the appropriate level
of abstraction, which allows one to prove that the implementation conforms
to the high-level specification.

The operational character of ASMs provides a sufficient basis for turning

the definitions into executable models. These can be used for high-level validation
of user scenarios prior to coding or of test beds by means of mental or machine
driven experiments. This luckily breaks with the still widely held traditional
view that specifications should be, or are by definition, non-executable.
The separation of different concerns is incorporated into the techniques
provided by the ASM method for constructing different system views and linking
them into hierarchies of system levels. Here are the major software development
concerns we systematically separate and recombine, making divide and conquer
the main methodical principle, and it led us very far.

– The separation of orthogonal design decisions is the most important one in

practice. It is made possible by the most general abstraction and refinement
capabilities of ASMs, and it is motivated by the necessity to keep the design
space open as long as possible and to structure it, for “design for change”
and for modular development.
– The separation of design from analysis corrects the long standing tradition
of identifying “rigorous” with “formalized in logic”. This tradition is one of
the reasons why the so called formal methods have not really had a strong
impact on practical software development.
– The separation, within the analysis, of experimental validation from mathe-
matical verification is possible with ASMs because once the ASM models
have been made executable, they can be used for simulation. The simulation
of a higher-level model can be performed prior to writing the final code. It
can also enhance testing the code, so that the correspondence of the imple-
mentation to the abstract specification can be checked.
– The separation of different degrees of detail within verification allows one to
adapt the justification of a design to the current development stage. ASMs
give the means to differentiate justifying a design to domain experts from
justification in terms of mechanical reasoning systems. Further, one should
distinguish between interactive logical systems and fully automated tools
like model checkers or automatic theorem provers.
– The separation of mathematical concerns within verification. For example,
one may want to split the proof of a property P for a complex system S into
three steps, namely:
• prove P for an abstract model M of S under an appropriate assumption
A,
• refine M to S so that S implements M correctly,
• prove that S satisfies the assumption A.
4 E. Börger

Experience shows that it is not only easier to prove properties for complex
systems in this way, but this splitting of proof obligations often is the only
known way to show that a run-time system works the way it is supposed
to. A characteristic example are the ASM-based proofs of the correctness of
compilation schemes [13,10,17] and the implementation of those schemes by
provably correct real-life compilers, a problem which is addressed in these
Proceedings. By the way, collecting such justificational evidence yields an
interesting byproduct: a detailed analysis of the design itself.

Most of these lines of research and principles were present, in embryo, already
at the first international ASM workshop held as early as 1994, as part of the IFIP
World Computer Congress in Hamburg, Germany [30]. Although the applications
which appeared there were limited by our academic experience and were largely
motivated by an effort to critically test the ASM thesis, it is not incidental that
right from the beginning, by making good use of the freedom of abstraction
offered by the ASM concept, we were naturally led to “separate and combine”
design levels, design and analysis, verification and validation, and degrees of
detail in verification. Concerning validation, let us note that the first tools for
making ASMs executable date back to 1990; see the Prolog-based compiler for
ASMs Angelika Kappel developed in [29], to execute my ASMs for PROLOG
[8], and the Michigan interpreter mentioned in [23]3 .
Only 5 years later, at the ASM workshop which was held as part of the
International Formal Methods Conference FM’99 in Toulouse, France4 , one can
observe (see again the survey [6] for details) that

– the theory of ASMs is richly developed,

– the applications include industrially successful standardization and software
engineering projects, some of which become publicly visible for the first time
in these Proceedings [15,4],
– there is a proliferation of different tools5 , most of them developed in acade-
mia, for both the experimental validation and the machine supported veri-
fication of ASMs, providing execution mechanisms for ASMs (via interpre-
tation or compilation) and links to verification systems like PVS, KIV and
model checkers.
3
Here is what Jim Huggins wrote to me on June 4, 2000, about the details of the
history: “Yuri taught a course in programming language design at Michigan during
the Fall of 1990. Of course, he introduced ASMs in the class. A very sharp young
undergraduate named Ben Harrison wrote a bare-bones interpreter for ASMs in
LISP in a weekend and distributed it to the class. Yuri was impressed enough with
Harrison that he hired him to write a full-fledged ASM interpreter, this time in C.
Ben built the core of the interpreter in May-June 1991. At that point the interpreter
was handed over to me, and I worked on it for 3 years or so, finishing up the rest
of the unfinished business at that time. In 1994 development was handed over to
Raghu Mani, who worked on it for a couple of years...”. Let me add that Raghu’s
task was to upgrade the interpreter for one-thread ASMs to multi-agent ASMs.
4
In between, ASM workshops had been held in Paderborn (May 1996), Cannes (June
1997, June 1998) and Magdeburg (September 1998).
 Abstract State Machines at the Cusp of the Millenium 5

All of these themes are reflected in the rich program of ASM’2000. These
Proceedings, which constitute the first book entirely devoted to ASMs, docu-
ment what has been achieved in the first decade after the formulation of the
ASM concept. We see confirmation of our conviction expressed already at the
first ASM workshop in 1994, namely that (paraphrased) “the extraordinary po-
tential of the ASM method will change drastically the industrial future of formal
specifications” [5, pg.393].

2 New Frontiers

The experience accumulated with the ASM concept, and with the method which
has been developed for its use, did change the way we think about high-level
software design and analysis. Now we have to actualize this vision to make
it work for established software development disciplines, at a large scale. The
achievements of the last decade open new frontiers and put us into the position
to face the new challenges.
Through the extensive ASM modeling, validation and verification work of
the past decade, the ASM thesis was experimentally confirmed. But this year
brought us a theoretical explanation of the observed phenomenon, namely via
a proof [25] that the sequential version of the thesis follows from three funda-
mental system theory axioms. Once established, the thesis allows one to draw
conclusions of practical importance, as is illustrated by an example in these
Proceedings [7]: the thesis guarantees that there is no loss of generality in sub-
stituting the fundamental but vague UML concepts of action and activity by the
mathematically rigorous concepts of ASM step and ASM run. It seems that in
UML [31,2] the meanings of action/activity were intentionally left unspecified,
namely to leave the space of possible implementations as open as possible. But
this was achieved at the price of making it difficult to control the implications
the concepts have in the context of the event-driven run-to-completion scheme,
in particular concerning the possibly numerous and nested exit/entry actions,
coming through interrupts, and concerning the launch and abortion of internal
activities.
On the practical side we have to take advantage of the experience, acquired
with building tools for executing ASMs, to develop an entire tool environment
which is also industrially satisfactory. It has to support the different activities
of defining, transforming (by refinements and by code generation) and analysing
ASM models (by testing, via visualization supported simulation, and by verifi-
cation). The tool environment has to enable us to capture the design knowledge
in a rigorous, electronically available and reusable way, and to achieve this goal
it must be integrated into established design flows and their tool environments.
The integration potential of ASMs, as a universal model of computation which
5
When I was working on this introduction, a message from Prof. Igor Soloviev, of
St.Petersburg State University, arrived: “One of my students, Andrew Usov, has
written an ASM interpreter. It is implemented as an java-applet designed to run
under an internet-browser (it has been tested under IE 5.0 and NS 4.7).”
6 E. Börger

is well established by the accumulated experimental evidence and by the theo-

retical explanation we have by now for the ASM thesis, is helpful to capture
the overall behavior of a complex system by combined use of whatever rigorous
descriptions are appropriate and mandatory in established design approaches
(static, dynamic, functional, state-based, object-oriented, etc.).
This is a difficult and probably long way to go, “a ridge walk between freedom
and discipline, creativity and pattern oriented design, generality and specializa-
tion, expressability and limitations by tool support” [6]. But it is worth the effort.
Here are some among other challenging problems where I see a large potential
for fruitful exploitation of the ASM method.
– If we succeed to construct paradigmatic and parameterized ASM compo-
nents and to extract (de)composition techniques that can be made available
in libraries, the “codeless” form of ASM programming will help porting ap-
plication programs from one platform or language to another and can lead
to fruitful applications for plug-and-play software technology.
– If we succeed to exploit ASMs for defining and implementing methods for
generating test suites from high-level specifications, this will turn a dark and
at present overwhelming part of software development into an intellectually
challenging and methodologically well supported task of enormous practical
value. Indeed using ASMs one can solve the crucial and essentially creative
part of test case selection, given that this selection is driven typically by
appplication domain expert knowledge and thus can be formulated using the
ASM ground model. Similarly the ground model supports solving the oracle
problem of testing: the expected output, which has to be compared with
the execution output, can be defined using the ground model specification
(which is independent of the programming language where the system will
be encoded)6 .
– If we exploit ASMs to enhance current (mostly signature oriented) software
architecture description techniques by adding to the structural definitions
also relevant semantical content, we will solve a widely felt need for building
reliably reconfigurable conceptual and module interconnection architectures
[28].
– If we succeed to exploit the atomic transaction nature of the notion of ASM-
step to model practically useful patterns for communication and synchroniza-
tion of multi-agent ASMs, typically based on shared memory or on message
passing, then we will contribute to solve a crucial problem of distributed
computing.
On the theoretical side a model and a proof theory of ASMs are needed. We
need definitions which capture and enhance the practical refinement schemes we
have used with success for ASMs, together with useful proof principles which
can be built into state-of-the-art mechanical verification systems (for some steps
in this direction see the contributions to PVS and model checking in these Pro-
ceedings and [32,14]). The proof theory we need should alleviate the verification
effort encountered in practical applications, namely by offering structuring and
layering of proof obligations which avoid the bottleneck of a priori fixed levels
6
A similar remark applies also to static testing (code inspection) where one has to
formulate the properties to be checked.
 Abstract State Machines at the Cusp of the Millenium 7

of overwhelming proof details. We need to find the right way to exploit the no-
tion of monitored (real-valued) function for connecting the discrete ASM world
to the continuous world of control theory. We need to help in building models
for mobile computing. We badly need to extract the inherent object oriented
features of ASMs, which are visible in the concept of ASM agents and of their
state, to make them explicitly and syntactically available, adapted to established
object-oriented programming techniques.
These Proceedings contain numerous contributions where the mentioned is-
sues are raised and thus constitute a good point of departure to help solving the
challenging problems which are waiting for us.

References
1. Abstract State Machines. http://www.eecs.umich.edu/gasm/.
2. Rational Software Corporation, Unified Modeling Language UML, version 1.3,
1999.
3. ISO/IEC 13211-1. Prolog-Part 1: General Core. In Information Technology-
Programming Languages. International Standards Organization, 1995.
4. M. Barnett, E. Börger, Y. Gurevich, W. Schulte, and M. Veanes. Using ASMs at
Microsoft: A Case Study. In This volume.
5. E. Börger. Logic Programming: The Evolving Algebra Approach. In B. Pehrson
and I. Simon, editor, IFIP 13th World Computer Congress, number I (Techno-
logy/Foundations), pages 391–395. Elsevier, 1994.
6. E. Börger. High Level System Design and Analysis using Abstract State Machines.
In D. Hutter and W. Stephan and P. Traverso and M. Ullmann, editor, Current
Trends in Applied Formal Methods (FM-Trends 98), number 1641 in LNCS, pages
1–43. Springer-Verlag, 1999.
7. E. Börger, A. Cavarra, and E. Riccobene. Modeling the Dynamics of UML State
Machines. In This volume.
8. E. Börger and K. Dässler. PROLOG. DIN Papers for Discussion. Report 58,
ISO/IEC JTC1 SC22 WG17, April 1990.
9. E. Börger and B. Demoen. The view on database updates in Standard Prolog: a
proposal and a rationale. Report 74, ISO/IEC JTC1 SC22 WG17 , February 1991.
10. E. Börger and I. Durdanovic. Correctness of Compiling Occam to Transputer
Code. Computer Journal, (39(1)):52–92, 1996.
11. E. Börger and D. Rosenzweig. An Analysis of Prolog Database Views and Their
Uniform Implementation. In K. Dässler and R. Scowen, editor, Prolog. Paris
Papers–2, number 80, pages 87–130. National Physical Laboratory, Middlesex, July
1991.
12. E. Börger and D. Rosenzweig. The Mathematics of Set Predicates in Prolog. In K.
Dässler and R. Scowen, editor, Prolog. Copenhagen Papers–2, number 105, pages
33–42. National Physical Laboratory, Middlesex, 1993.
13. E. Börger and D. Rosenzweig. The WAM–Definition and Compiler Correctness.
In Ch. Beierle and L. Plümer, editors, Logic Programming: Formal Methods and
Practical Applications, pages 20–90. Elsevier Science B.V./North–Holland, 1995.
14. E. Börger and J. Schmid. Composition and Submachine Concepts for Sequential
ASMs. In P. Clote and H. Schwichtenberg, editor, Gurevich Festschrift CSL 2000,
LNCS. Springer-Verlag, 2000. (In print).
15. E. Börger, J. Schmid, and P. Päppinghaus. Report on a Practical Application of
ASMs in Software Design. In This volume.
8 E. Börger

16. E. Börger, J. Schmid, and E. Riccobene. Capturing Requirements by Abstract

State Machines: The Light Control Case Study. J. Universal Computer Science,
2000. Special Requirement Engineering Issue, to appear.
17. E. Börger, J. Schmid, R. Stärk, and W. Schulte. Java and the Java Virtual Ma-
chine. Springer-Verlag, 2000. to appear.
18. R. Eschbach, U. Glässer, R. Gotzhein, and A. Prinz. On the Formal Semantics of
SDL-2000: a Compilation Approach Based on an Abstract SDL Machine. In This
volume.
19. U. Glässer, R. Gotzhein, and A. Prinz. Towards a New Formal SDL Semantics
Based on Abstract State Machines. In R. Dssouli and G.v. Bochmann and Y.Lahav
, editor, SDL’99 - The Next Millenium (Proc. of the 9th SDL FORUM). Elsevier
Science B.V., 1999.
20. Y. Gurevich. Reconsidering Turing’s Thesis: Toward More Realistic Semantics of
Programs. Technical Report CRL-TR-36-84, University of Michigan, Computing
Research Lab, 1984.
21. Y. Gurevich. A New Thesis. Notices of the American Mathematical Society, page
317, 1985. abstract 85T-68-203, received May 13.
22. Y. Gurevich. Logic and the Challenge of Computer Science. In E. Börger, editor,
Current Trends in Theoretical Computer Science, pages 1–57. Computer Science
Press, 1988.
23. Y. Gurevich. Evolving Algebras: An Attempt to Discover Semantics. In G. Rozen-
berg and A. Salomaa, editors, Current Trends in Theoretical Computer Science,
pages 225–234. World Scientific, 1993. A reprint of the article in the Bulletin of
the European Association for Theoretical Computer Science, Number 35 (1991),
pp.71-82.
24. Y. Gurevich. Evolving Algebras 1993: Lipari Guide. In E. Börger, editor, Specifi-
cation and Validation Methods, pages 9–36. Oxford University Press, 1995.
25. Y. Gurevich. Sequential Abstract State Machines Capture Sequential Algorithms.
ACM Transactions on Computational Logic, 1, 2000.
26. Y. Gurevich and J. Morris. Algebraic operational semantics and modula-2. In
Börger, E. et al., editor, CSL’87, 1st Workshop on Computer Science Logic, number
329 in LNCS, pages 81–101. Springer-Verlag, 1988.
27. Y. Gurevich and L. A. Moss. Algebraic operational semantics and occam. In
Börger, E. et al., editor, CSL’89, 3d Workshop on Computer Science Logic, number
440 in LNCS, pages 176–192. Springer-Verlag, 1990.
28. C. Hofmeister, R.L. Nord, and D. Soni. Applied Software Architecture. Addison
Wesley, 1999.
29. A.M. Kappel. Implementation of Dynamic Algebras with an Application to Prolog.
Master’s thesis, CS Dept., University of Dortmund, Germany, November 1990.
An extended abstract ”Executable Specifications based on Dynamic Algebras”
appeared in A. Voronkov (ed.): Logic Programming and Automated Reasoning,
volume 698 of LNAI, Springer, 1993, pages 229-240.
30. B. Pehrson and I. Simon. IFIP 13th World Computer Congress. Vol.I: Techno-
logy/Foundations. Elsevier, 1994.
31. J. Rumbaugh, I. Jacobson, and G. Booch. The Unified Modeling Language Refe-
rence Manual. Addison Wesley, 1999.
32. G. Schellhorn. Verifikation abstrakter Zustandsmaschinen. Phd thesis, CS
Dept., University of Ulm, Germany, 1999. For an English version consult
www.informatik.uni-ulm.de/pm/kiv/papers/verif-asms-english.ps.gz.
Abstract State Machines and Pure Mathematics

Andreas Blass ?

University of Michigan, USA

[email protected]

Abstract. We discuss connections, similarities, and differences between

the concepts and issues arising in the study of abstract state machines
and those arising in pure mathematics, particularly in set theory and
logic. Among the topics from pure mathematics are the foundational
role of set theory, permutation models of set theory without the axiom
of choice, and interpretations (between theories or vocabularies) regarded
as transformations acting on structures.

1 Introduction

The purpose of this paper is to describe some connections between the theory
of abstract state machines (ASM’s) and concepts from pure mathematics. These
connections are of three general sorts.
First, there are direct uses of mathematical concepts in ASM’s. A well known
instance of this is the use of structures, in the sense of first-order logic, in the
very definition of ASM’s. A less well known instance is the use of interpretations,
also in the sense of first-order logic, to describe transitions of ASM’s as well as
certain sorts of simulations.
Second, there are modifications of mathematical concepts, adapting them
to the purposes of computation theory and ASM’s. As an example of this, we
discuss the ASM analog of the set-theoretic concept of permutation model.
Finally, there are analogies between ASM’s and some aspects of pure ma-
thematics. In this connection, we discuss the multifaceted philosophical issue
of “universality”: Do ASM’s provide a universal framework for descriptions of
algorithms in the same sense that set theory provides a universal framework for
mathematical proofs? In this connection, we also discuss the value of explicit
formalization and the role of definitions. We comment briefly on possible alter-
native “foundations” based on, for example, ordered lists or multisets instead of
sets.
We also discuss the issue of “objects whose identity doesn’t matter,” which
arises in connection with the import rules of ASM’s and also in various contexts
in pure mathematics.
?
Preparation of this paper was partially supported by a grant from Microsoft Corpo-
ration. The opinions expressed here are, however, entirely my own.

Y. Gurevich et al. (Eds.): ASM 2000, LNCS 1912, pp. 9–21, 2000.
c Springer-Verlag Berlin Heidelberg 2000

10 A. Blass

Acknowledgements. I thank the organizers of the Monte Verità conference for

inviting me to present the talk on which this paper is based. I also thank Yuri
Gurevich for his suggestions for improving this paper and for many helpful and
informative discussions about ASM’s and other aspects of computer science.

2 Structures and Interpretations

From the very beginning, the concept of abstract state machines has been closely
linked to pure mathematics. Indeed, one of the central ideas is to model states of
computations as structures in the sense of mathematical logic. Another central
idea is that an ASM program describes one step in a computation, not the entire
iteration. Because of this, the programs themselves correspond to a familiar
concept from first-order logic, that of interpretation.
Interpretations of one vocabulary or theory in another were introduced in
[17] for the purpose of deducing undecidability of many theories from essential
undecidability of one theory. See also [15, Section 4.7], whose formulation we
follow here. In general, an interpretation of a vocabulary Υ in a theory T (with
vocabulary ΥT ) consists of
– a unary predicate U in ΥT ,
– for each function symbol F of Υ a function symbol FI of ΥT , and
– for each predicate symbol P of Υ a predicate symbol PI of ΥT (equality does
not count as a predicate symbol here but as a logical symbol),
such that it is provable in T that U is nonempty and closed under all the FI ’s.
Thus, given a model M for T , we obtain a structure for Υ by taking as its base
set the extent in M of U and as its functions and predicates the (restrictions of
the) interpretations in M of the corresponding symbols of ΥT . The concept of
interpretation is often extended by saying “interpretation in T ” when one really
means “interpretation in an extension by definitions of T .” Then U , FI , and
PI need not be primitive symbols of the vocabulary ΥT but could be given by
formulas in that vocabulary. (For historical accuracy, I should mention that the
definition of “interpretation” in [17] already included extensions by definitions
but did not include the unary predicate U ; the latter was treated separately
under the name of “relativization.”)
Consider the special case of an interpretation where the theory T has no
axioms (so the interpretation can be applied to arbitrary ΥT -structures), where
Υ = ΥT , and where U is identically true (so applying the interpretation doesn’t
change the vocabulary or the base set of a structure). The transformations of
Υ -structures obtainable by such transformations are just those obtainable by
executing one step of an ASM with vocabulary Υ , provided the ASM doesn’t
create new elements and is non-distributed. (In the distributed case, the interpre-
tations should be restricted so that the steps they describe can each be executed
by a single agent.)
From this point of view, one could say that an ASM is “just a special sort
of interpretation,” but the word “just” here is unjust. Though formally the
 Abstract State Machines and Pure Mathematics 11

same, ASM’s are intended to be used quite differently from interpretations: An

ASM is to be applied iteratively (for a number of steps not usually known in
advance). For interpretations in general, there is a notion of composition, but
only interpretations from a vocabulary (or theory) to itself can be repeated
arbitrarily often.
It may be noteworthy that, in the correspondence between ASM’s and inter-
pretations, the sequential ASM’s correspond to quantifier-free interpretations.
Thus, the same restriction is quite natural from both the computational and the
logical points of view.
Can interpretations contribute anything to our understanding of ASM’s? In
a sense, they already did, at least if one considers interpretations in a generalized
sense that has become common in model theory. There (following an idea that
I believe was introduced by Shelah), one more or less automatically enlarges
structures to include, in addition to their actual elements,
– tuples of elements and
– equivalence classes with respect to definable equivalence relations.
See for example the M eq construction in [13, page 10]. Then interpretations can
use these additional elements; thus for example the standard definition of the
integers in terms of natural numbers (as equivalence classes of pairs) amounts to
an interpretation. Some of the central ideas and results of [2] first appeared, in
notes of Shelah, in the context of interpretations of just this sort. (They weren’t
explicitly called interpretations, but the influence of M eq was visible.) Later,
this material was clarified by expressing it in terms of ASM’s. The distinctive
set-theoretic context of [2] was motivated, at least in part, by the desire for a
natural framework in which tuples and equivalence classes are easily handled.
Might interpretations contribute something more to ASM theory? What ab-
out interpretations between different vocabularies, which are therefore not itera-
ble? Perhaps these will provide a useful point of view for such operations as
setting up an initial state for an ASM. Often the input to a computation is only
part of what should be present in the initial state of a computation. Adding the
extra material to convert the input into the initial state is, in some cases, an
interpretation in the generalized (M eq ) sense discussed above. In other cases,
this preparation is more complicated, for example adding a whole set-theoretic
superstructure as in [2]. It is not clear whether the notion of interpretation can
be usefully stretched to cover such cases as well.
In addition to setting up the initial state for a computation, interpretati-
ons might be appropriate for extracting the “answer” from the final state of a
computation. More generally, one can imagine many sorts of interfaces — not
only input/output but more general interfaces between computations or between
parts of a computation — as described by interpretations.
It also seems that interpretations may provide a good way to describe the
connection between two ASM’s that “compute the same thing in different ways.”
The idea here is that certain states, the “pose” states, in a run of the one machine
are obtainable from those of the other machine by a uniform interpretation.
(Between successive pose states, each machine may go through states that have
12 A. Blass

no strict counterpart in the other machine, states that depend on the details of
how the machine does its work.) The interpretation tells how what one machine
does is to be seen from the perspective of the other.
Finally, since interpretations are usually (in mathematical logic) regarded
primarily as syntactic transformations rather than semantical ones, I should
add a few words about the syntactic aspect. An interpretation of a vocabulary
Υ in a theory T provides a translation of terms and formulas from Υ to the
vocabulary of T . The translation consists of replacing each function symbol F
or predicate symbol P in the given term or formula with FI or PI and restricting
all quantifiers to UI . Then the translation ϕI of a sentence ϕ is true in a model
M of T if and only if the sentence ϕ itself is true in the Υ -structure MI obtained
from M via the interpretation.
Thinking of interpretations primarily on the semantic level, transforming mo-
dels M of T into Υ -structures MI , we can describe the syntactic transformation
ϕ 7→ ϕI as producing weakest preconditions. That is, ϕI expresses, in a structure
M , exactly what is needed in order that MI satisfy ϕ. Here is a toy example to
indicate what goes on in general.
Consider the update rule c := d. The corresponding interpretation has PI =
P and FI = F for all predicate and function symbols other than c, and cI = d
(and UI is the identically true predicate). If M is a structure, then MI for this
interpretation is the same as M except that it gives c the value that d had in
M , i.e., it is the sequel of M with respect to the rule c := d. Now consider a
formula ϕ, the simplest relevant example being P (c) for a unary predicate P .
Its translation ϕI under our interpretation is obtained simply by replacing every
occurrence of c by d, so in our simple example P (c)I is P (d). And this ϕI is
exactly what M must satisfy in order to guarantee that MI satisfies ϕ.

3 Permutation Models

In this section, I’ll describe a technique for describing and analyzing the idea
of “not permitting arbitrary choices,” both in set theory and in computation
theory.
In 1922, Fraenkel [5] constructed models of set theory (with infinitely many
urelements, also called atoms) in which the axiom of choice is false. The central
idea is that all the atoms in a model of set theory “look alike”; more precisely, any
permutation of the atoms induces an automorphism of the whole set-theoretic
universe. Fraenkel formed the subuniverse of “sufficiently symmetric” sets, and
showed that it satisfies all the usual axioms of set theory except the axiom of
choice. (“Usual” is an anachronism here, since Fraenkel himself had just recently
introduced the replacement axiom and given a precise formulation of Zermelo’s
separation axiom. But the axiom system, known as Zermelo-Fraenkel set theory,
ZF, is certainly the usual one nowadays.) More precisely, Fraenkel’s model con-
sisted of those sets x which depend on only a finite subset F of the atoms, in the
sense that any permutation of the atoms fixing those in F will, when extended to
an automorphism of the whole universe, also fix x. (Subsequently, other authors
 Abstract State Machines and Pure Mathematics 13

studied other notions of “sufficiently symmetric” in order to obtain independence

results concerning weak forms of the axiom of choice, but these other notions
will not be relevant here.)

A modification of Fraenkel’s idea is at the heart of the main results of [2].

This paper introduced a rather liberal model of choiceless polynomial time com-
putation. The model is an ASM whose initial state is obtained by building a uni-
verse of (hereditarily) finite sets over the input structure regarded as consisting
of atoms; this makes many sorts of data structures available for the computa-
tion. In addition, parallelism is allowed both in do-for-all rules and in some
set-formation constructions. It is shown, however, that such apparently simple
things as the parity of the number of elements in an unstructured set cannot be
computed in polynomial time, even in this liberal model, as long as arbitrary
choices are prohibited. The proof of this theorem uses the hypothesis of choice-
lessness by deducing that, if the computation “uses” a set x, then it must also
use all sets π(x) obtainable from x by automorphisms of the set-theoretic struc-
ture, i.e., by permutations of the atoms of the input structure. The polynomial
time hypothesis then ensures that there cannot be too many of these π(x)’s. The
essential combinatorial step in the proof is a lemma allowing us to infer from
“not too many π(x)’s” to x being fixed by all permutations that fix a certain
number of atoms. In contrast to Fraenkel’s situation, “a certain number” cannot
be taken to be simply “finitely many,” as there are only finitely many atoms
altogether in the present situation. Instead, a quantitative estimate is needed,
which depends on the ASM program under consideration. Thus, the “infinite
vs. finite” dichotomy exploited by Fraenkel has been replaced by “large finite
(growing as the input structure grows) vs. small finite (depending only on the
ASM program).

A further refinement of Fraenkel’s idea occurs in [14], where Shelah proves a

zero-one law for properties of graphs computable in choiceless polynomial time.
For the purposes of permutation arguments, there is a crucial difference between
the unstructured (or the slightly more general “colored”) sets considered in [2]
and the graphs considered in [14]: Almost all finite graphs have no non-trivial
automorphisms. So at first it would seem that symmetry arguments cannot be
applied to computations taking random finite graphs as inputs. Shelah circum-
vented this problem by working not with automorphisms but with partial auto-
morphisms of graphs. This required a rather delicate development, with careful
attention to the sizes of the domains of these partial automorphisms, to ensure
that the behavior of a choiceless polynomial time computation is invariant (in a
suitable sense) under partial automorphisms of the input. For the (rather com-
plicated) details, we refer the reader to [14] or to an easier to read (we hope)
exposition that Yuri Gurevich and I are preparing.
14 A. Blass

4 ASM’s and Set Theory

At a 1993 meeting in Dagstuhl, after hearing several talks about formulating
various algorithms as ASM’s (then still called “evolving algebras”), I made a
comment along the following lines:
It seems to me that expressing algorithms in the ASM formalism is rather
like formalizing mathematical proofs in ZFC. At the beginning, one needs
a number of examples of such formalizations, but after a while it becomes
clear that any “reasonable” algorithm can be written as an ASM, just
as any “reasonable” proof can be formalized in ZFC. And after a while,
there is little point in actually carrying out the formalizations just in
order to verify formalizability (in either situation) unless and until a
genuinely problematic case arises.
In the case of sequential algorithms, it seems fair to say that this comment
has been confirmed by subsequent experience. Better yet, it is confirmed by the
main result of Gurevich [8], which asserts that any sequential algorithm can be
expressed by an ASM provided it satisfies certain very natural postulates.
I’ll comment later on the situation with non-sequential algorithms, but first
let me point out some differences between the roles of ASM’s and of ZFC even
in the sequential case.
The most obvious difference is that most of us have never seen a non-trivial
proof fully formalized in ZFC (nor would we want to see one), but we have
seen non-trivial algorithms fully formalized as ASM’s. ASM’s are, by design,
reasonably close to our mental images of algorithms; ZFC is similarly close to
mental images of only small parts of mathematics (primarily set theory, and not
even all of that). To write out the proof of, say, the commutative law for real
addition in the primitive language of set theory would be a very time-consuming
exercise for which I see no value at all.
In contrast, writing out explicit ASM programs for various algorithms is not
only feasible but worthwhile. It has, for example, led to the detection of errors or
ambiguities in specifications of programming languages and computer systems.
This distinction is reflected in a second difference between ASM and ZFC
formalizations: the ubiquity of definitions in the development of mathematics on
a set-theoretic basis. Such a development begins with axioms written in purely
set-theoretic notation, but it soon introduces other concepts by definition, and it
is not at all unusual for a concept to be many definitional layers removed from the
primitive set-theoretic notions. Think, for example of the real numbers, defined
as equivalence classes of Cauchy sequences of rational numbers. The definition
of R sits at the top of a rather high tower (or pile) of other definitions.
Something roughly analogous happens in ASM’s, in the idea of successive
refinement, working from a high-level description of an algorithm down to a
specific implementation. But the analogy is only a rough one for two reasons.
First, the tower that connects low-level and high-level versions of an algorithm
consists of ASM’s at every level. The tower connecting ZFC to, say, the theory
of partial differential equations, departs from ZFC formalization (in the strict
 Abstract State Machines and Pure Mathematics 15

sense) as soon as one gets above the bottom level; the higher levels are formalized,
if at all, in first-order theories obtained (formally) by adding definitions as new
axioms. Second, the refinement process is a serious object of study in the ASM
world, whereas in the set-theoretic world the process of definition is usually
swept under the rug — so much so that many logicians would be taken aback
by my comment a moment ago that differential equation theory is formalized
not in ZFC but in an extension by definitions. They would probably say “what’s
the difference?” and accuse me of splitting hairs. The only authors I know of
who have made a serious study of the role of definitions in set theory are Morse
[12], who carefully analyzes the permissible syntax (far more general than what
is usually considered), and Leśniewski [11] who actually uses definitions in a
non-conservative way, as an elegant formulation of existence axioms.
This suggests a question for ASM theory: Can one push the refinement pro-
cess as far into the background as set-theorists have pushed definitions (and if
not then how far can one push it)? That is, can one arrange things so that all
one has to write is a high-level specification and the sequence of “definitions” of
its functions in terms of lower levels? The idea would be that the corresponding
low-level ASM would either be produced automatically or would (as in the set-
theoretic situation) be entirely irrelevant. And of course the high-level ASM and
the definitions together should be significantly shorter than the low-level ASM
they describe.
A third difference between ASM’s and ZFC is the use of ASM’s for detecting
and correcting errors in intuitive specifications, programs, etc. As far as I know,
formalization in ZFC has not played a role in detection and correction of errors
in mathematical proofs. Errors certainly occur, and they are detected and cor-
rected. (For particularly embarrassing examples, see the footnotes on page 118
of [9] and the introduction of [16]. I emphasize that the authors of [9] and [16]
did not commit the errors but conveniently summarized them.) But the detec-
tion and correction seems to proceed quite independently of formalization. It is
based on intuitive understanding and is therefore not very systematic. There are,
of course, projects to formalize and systematically check, by computer, various
parts of mathematics. But these seem to be of more interest, at the moment,
to computer scientists than to typical mathematicians. The latter seem to be
generally quite willing to rely on intuitive understanding rather than formal
checking.
A fourth difference between the role of ASM’s in formalizing algorithms and
the role of ZFC in formalizing proofs is that in the latter context there is (to
the best of my knowledge) no analog of [8]. That is, there is no general criterion
guaranteeing that any proof, subject to some natural constraints, is formalizable
in ZFC. Mathematicians other than set theorists generally take on faith that
anything they would consider a correct proof can be formalized in ZFC. Set
theorists are aware of possible difficulties (with universe-sized sets), but we are
confident that we could recognize any such difficulty and determine whether
a proposed argument really goes beyond ZFC (without actually writing out a
formalization).
16 A. Blass

Having promised to comment on the situation for non-sequential algorithms,

let me say that the situation there is less clear than in the sequential case. There
is no analog of [8] (yet), and the world of parallel and distributed algorithms
seems much more difficult to survey completely than the world of sequential
algorithms. (Distinguishing between “parallel” and “distributed” in the sense of
[7], it seems that an analog of [8] is within reach for parallel algorithms but not
for distributed ones.)
A particular headache that has come up several times in my discussions
with Yuri Gurevich is the question of cumulative updates. As a simple example,
suppose we want to count the number of elements in some finite universe U (in
an ASM). The natural way to do this is to have a counter, initialized to zero,
which each of the relevant elements then increments by one. The problem is
to do this in parallel, without arbitrarily fixing an order in which the various
elements are to act. The rule
do for all v with U (v)
c := c + 1
enddo
only increments the counter by one, no matter how many elements are in U .
If we have an ordering of these elements, then we can let each one increment
the counter in turn. But without an ordering, a genuinely parallel version of the
algorithm seems to require something new.
The fundamental problem, though, isn’t whether this or that situation calls
for an extension of the framework but rather whether the framework can be
completed at all. I am inclined to be optimistic: A few additions should cover
all parallel algorithms. But I realize that my optimism may be the result of an
overly naı̈ve picture of the wild world of parallelism.

5 How Basic Are Sets?

The title of this section is intended to refer to the computational (ASM) side
of the picture, not the pure mathematical (ZFC) side. Of course, sets are basic
in ZFC, but this is to some extent the result of an arbitrary choice. Indeed,
Cantor, the father of set theory, seems to have regarded sets as intrinsically
equipped with an ordering, so perhaps some sort of (generalized) lists would be
more fundamental.
On the computational side, too, there is a tendency to view interdefinable
concepts, like (finite) sets and lists, as having equal claim to being fundamental.
But it is shown in [3] that this tendency is not always appropriate. When we
do not allow arbitrary choices (or, equivalently, an ordering) but do allow paral-
lelism, and when we impose polynomial bounds on the total computation time
of all processors, then a set-theoretic environment as in [2] is strictly stronger
than one using tuples or lists instead. (The polynomial time bound is essential
here. Without it, parallelism could compensate for the absence of an ordering
by exploring all orderings.)
 Abstract State Machines and Pure Mathematics 17

This result suggests to me that a set-based environment is more appropriate

for this sort of computation than the more familiar list-based environments.
There is some tension here with the fact that lists can be more straightforwardly
implemented than sets. But the sort of computation under consideration, with
no ordering allowed, is already intrinsically removed from implementation, where
an ordering is implicitly given by the machine representation of elements.
These considerations raise a further question concerning choiceless polyno-
mial time computation: Might some other environment be even better than the
set-based one? Perhaps the considerations at the end of the preceding section,
about cumulative updates, suggest an answer, namely to use multisets and al-
low cumulative updating of multiplicities of membership in a multiset. In other
words, allow updates of the form “throw element x into multiset y,” with the
convention that, if several parallel processes execute this update (with the same
x and y), then the result is that the multiplicity of x’s membership in y is
increased by the number of these processes.
But can one do better yet? Might there even be entirely new data structures
that are particularly useful in the choiceless context?

6 Unidentified Objects

In this final section, I’d like to comment on an issue that comes up in many
contexts, throughout computer science and mathematics, but is almost always
ignored because it’s trivial. My main point is that, if it’s so trivial, we should be
able to give a clear explanation very close to our trivializing intuition.
The issue arises in the theory of ASM’s in connection with the importing or
creating of new elements. It doesn’t matter what the new element is, as long as
it’s new. So we refuse to worry about the exact choice of the new element. (If two
or more elements are to be imported simultaneously, then we make sure they’re
distinct, but beyond this avoidance of “clashes” we refuse to worry.) A mathe-
matician’s immediate reaction is that we work, not with a particular structure
in which a particular element has been imported, but with an isomorphism class
of structures, a different but isomorphic structure for each choice of imported
element. This works, but it doesn’t quite correspond to intuition; intuition is
still dealing with one structure, not a whole class of them. In some sense, intui-
tion deals with a “generic” member of the isomorphism class, but what exactly
(mathematically) is that?
The same issue and the same mathematical solution occur in connection with
the syntax of just about any formalized language. One uses bound variables,
but one doesn’t care what particular variable is used (as long as clashes are
avoided). It has become customary to annihilate the issue by saying “we work
modulo α-conversion,” meaning that expressions differing only in the choice of
bound variables are to be identified. This amounts to the “isomorphism class”
viewpoint of the preceding paragraph. Another approach was taken by Bourbaki
[4]. Their official syntax for logic and set theory (and thus for all of mathematics)
has no bound variables. In their places are occurrences of the symbol . To
18 A. Blass

indicate which occurrences correspond to the same variable, these are joined to
each other (and to the binding operator) by lines. This seems to me to be closer
to intuition, but rather unreadable. And it applies only to syntactic situations; I
don’t see how to adapt it to a semantic situation like ASM’s. (I find it somewhat
reassuring that the highly respected mathematicians of Bourbaki — or at least
some of them — found this issue to be worth thinking about to the extent of
producing a rather unorthodox solution.)
Another approach to this issue, in the syntactic context of bound variables,
is the use of de Bruijn indices. These indices link a variable-binder, such as
a quantifier, with the occurrences of the variables it binds, by specifying the
difference between them in quantifier depth. That is, in place of a Bourbaki box
with a line joining it to a quantifier, one would have a number indicating that
that the relevant quantifier is to be found so and so many levels less deep in the
parse tree. This notation strikes me as farther from intuition than Bourbaki’s
boxes but closer than isomorphism classes. In terms of human readability, it
seems no better than the boxes, but I understand computers read it quite well.
(Perhaps we humans just need to be wired differently.) I see no way to adapt this
approach to non-syntactic situations, like the choice of new elements created in
an ASM.
A variant of the isomorphism class approach is suggested by the topos-
theoretic view of generic or variable objects [10]. In the present context, this
amounts to regarding the isomorphism class not simply as a class but as an in-
dexed or parametrized family, the parameters being the individual choices. Thus,
for instance, an ASM that imports two elements would be viewed as a family
of ASM’s indexed by ordered pairs of distinct elements. The ASM indexed by
the pair (x, y) is the one in which first x and then y were imported. An elegant
framework and notation for this approach (applied to bound variables) is given
in [6].
The issue of unidentified objects arises in many other contexts. Indeed, one
can claim that the exact identity of mathematical objects never matters; all one
cares about is the structure relating them. For example, it never matters whether
real numbers are Dedekind cuts or equivalence classes of Cauchy sequences, as
long as they form a complete ordered field. The following quotation from [10,
page 119], though intended to emphasize the topos theorist’s view of sets in
contrast to the set theorist’s cumulative hierarchy, seems to accurately describe
how most mathematicians view sets.

An abstract set X has elements each of which has no internal struc-

ture whatsoever; X has no internal structure except for equality and
inequality of pairs of elements, and has no external properties save its
cardinality; still an abstract set is more refined (less abstract) than a
cardinal number in that it does have elements while a cardinal number
does not.

This comes very close to saying that the actual elements don’t matter as long
as there are some (and equality is determined). But I claim that we have no
 Abstract State Machines and Pure Mathematics 19

entirely satisfactory semantics for dealing with the concept of an arbitrary ob-
ject whose actual identity doesn’t matter (but whose distinctness from other
arbitrary objects may be important).
Notice that the problem I am raising is a semantical one. Syntactically, we
can handle objects whose identity doesn’t matter: just add constant symbols
to the language to serve as names for the desired objects. This process is well
understood in the context of first-order logic (see for example [15, Chapter 4]; it
plays a crucial role in the proof of Gödel’s completeness theorem). But passing
from syntax to semantics requires choosing elements to serve as the denotati-
ons of the new constant symbols, and that brings us right back to the original
problem: We need to choose an element but without caring which element it is.
Finally, let me mention that this problem threatens to become more acute if
one considers quantum computation or indeed any quantum phenomena. In the
quantum world, actual physical objects, like elementary particles, have identities
in only a very limited sense. It makes sense to talk about an electron and another
electron, but not to talk about this electron and that one — interchanging the
two electrons in some physical process yields not another process but another
channel for the same process, and two channels may interfere (constructively or
destructively).

A Appendix on Defaults and Environment

In this appendix, we collect some observations and questions about ASM’s whose
connection to pure mathematics is even more tenuous than in the last section.
Indeed, most of them are based on a contrast rather than a connection between
the computational and mathematical worlds.
A major difference between the viewpoints of ASM’s (together with most of
computer science) and pure mathematics is that in the former a dynamic aspect
is always present. An ASM doesn’t just sit there, it undergoes transitions —
as the old name “evolving algebra” emphasizes. Of course, pure mathematics
is also capable of dealing with dynamic situations, but this is always explicitly
emphasized, not part of a universal background as in ASM’s.
There is more to the dynamic aspect of ASM’s than just that their dynamic
functions can change their values. It is also important that dynamic functions
retain their previous values unless explicitly changed. Ironically, this apparently
simple default assumption, favoring persistence of values, can increase the logical
complexity of what an ASM does. For example, the action of a parallel rule
do for all v
R(v)
enddo
could be described in existential logic (or in the existential fixed point logic
advocated in [1]) provided its body R(v) could be so described. The parallel rule
executes an update if and only if there exists a value of v for which R(v) executes
that update. But the next state cannot be described existentially, because the
20 A. Blass

inaction of the rule, the persistence of dynamic functions when not updated,
requires a universal quantifier for its description.
The fact that parallel ASM’s can be limited to quantifier-free guards, using do
for all to simulate quantifiers, is ultimately due to the presence of an implicit
universal quantifier in the default requirement that values persist unless expli-
citly changed. (In the last two paragraphs, I’ve pretended for simplicity that the
ASM’s under consideration never attempt conflicting updates. The possibility
of clashes would introduce additional universal quantifiers, because an update is
executed if some R(v) would execute it (ignoring clashes) and there are no w1
and w2 for which R(w1 ) and R(w2 ) would execute conflicting updates.)
Another sort of default adds complexity not in the logic but in the computa-
tional aspects of ASM’s. This default is the assumption that, when new elements
are created (or imported from the reserve), they arrive with no structure: All
Boolean functions, except equality, produce false when one of their arguments
is a freshly created element; all non-Boolean functions produce undef under
these circumstances. Thus, for example, if P is a binary Boolean function and
x is a newly created element, then P (x, y) miraculously has the value false for
all y. If we ask how it got that value, there are several possible viewpoints. One
is that creating an element really means importing it from the reserve, and that
the appropriate default values were already there while x was in the reserve —
so the default value of P (x, y) for newly imported x is just a matter of the per-
sistence default discussed above. But of course this viewpoint requires that the
initial state of a computation include the appropriate default values for reserve
elements, an assumption that is appropriate only at a sufficiently high level of
abstraction. At a lower level, one would have to ask how this initialization is to
be performed. Another viewpoint is that the defaults are (at least implicitly) set
by the ASM at the time the new element is created. This amounts to a fairly
large scale parallel operation, not available in the sequential situation; it may be
the approach closest to what happens in actual computers when new memory
locations are allocated to a computation. A third viewpoint is that the setting of
the defaults is, like the choice of which element to import, the responsibility of
the environment. This seems to be the simplest approach, but it strikes me as a
bit unfair to the environment. Making the environment choose the new element
is reasonable, because this cannot be accomplished algorithmically (unless an
ordering or some similar structure is available); but setting the defaults could
be done algorithmically (in the case of parallel ASM’s) so the justification for
turning the job over to the environment seems to be only that it’s more work
than the ASM wants to do.
Let me close with a brief comment, related to the preceding only because it’s
about the environment. The environment of an ASM is used to model a variety of
things: the choice of elements to import (or create), the arbitrary choices involved
in non-determinism, input-output operations, and, in distributed computing,
all the agents other than the one under consideration. How similar are these,
really? There is of course one similarity, which caused them to all be called
“environment” in the first place: They are not part of the algorithm (ASM)
 Abstract State Machines and Pure Mathematics 21

under consideration, but they interact with it. Is there further similarity among
some (or all) of these aspects of the environment? Are there useful distinctions
to be made? (“Useful” means, at a minimum, more useful than just listing the
various items as I did above.)
One rather imprecise but perhaps useful distinction is obtained by singling
out those aspects of the environment’s activity that one would expect to be
included in a system for executing ASM programs. Such a system should not be
expected to provide input or to perform the actions of distributed agents other
than the one being simulated. But it could reasonably be expected to execute
import rules on its own, or at most with the help of the operating system under
which it runs.

References
1. Andreas Blass and Yuri Gurevich, “Existential fixed-point logic,” in Computa-
tion Theory and Logic, ed. by E. Börger, Lecture Notes in Computer Science 270,
Springer-Verlag (1987) 20–36.
2. Andreas Blass, Yuri Gurevich, and Saharon Shelah, “Choiceless polynomial time,”
Ann. Pure Appl. Logic, 100 (1999) 141–187.
3. Andreas Blass, Yuri Gurevich, and Jan Van den Bussche, “Abstract state machines
and computationally complete query languages,” this volume.
4. Nicolas Bourbaki, Elements of Mathematics. Theory of Sets, Hermann (1968).
5. Abraham Fraenkel, Der Begriff “definit” und die Unabhängigkeit des Aus-
wahlaxioms, Sitzungsberichte der Preussischen Akademie der Wissenschaften,
Physikalisch-Mathematische Klasse (1922) 253–257.
6. Murdoch J. Gabbay and Andrew M. Pitts, “A New Approach to Abstract Syn-
tax Involving Binders,” in Proceedings 14th Annual IEEE Symposium on Logic in
Computer Science, Trento, Italy, July 1999, IEEE Computer Society Press (1999)
214-224.
7. Yuri Gurevich, “Evolving Algebra 1993: Lipari Guide”, in Specification and Vali-
dation Methods, ed. by E. Boerger, Oxford University Press, 1995, 9–36.
8. Yuri Gurevich, “Sequential abstract state machines capture sequential algorithms,”
ACM Transactions on Computational Logic, to appear.
9. Thomas Jech, The Axiom of Choice, North-Holland (1973).
10. F. William Lawvere, “Variable quantities and variable structures in topoi,” in Al-
gebra, Topology, and Category Theory (A Collection of Papers in Honor of Samuel
Eilenberg), ed. by A. Heller and M. Tierney, Academic Press (1976) 101–131.
11. Eugene C. Luschei, The Logical Systems of Leśniewski, North-Holland (1962).
12. Anthony P. Morse, A Theory of Sets, Academic Press (1965).
13. Anand Pillay, Geometric Stability Theory, Oxford University Press (1996).
14. Saharon Shelah, Choiceless polynomial time logic: Inability to express, paper num-
ber 634, to appear.
15. Joseph Shoenfield, Mathematical Logic, Addison-Wesley (1967).
16. Edward E. Slaminka, “A Brouwer translation theorem for free homeomorphisms,”
Trans. Amer. Math. Soc., 306 (1988) 277–291.
17. Alfred Tarski, Andrzej Mostowski, and Abraham Robinson, Undecidable Theories,
North-Holland (1953).
 Abstract State Machines and Computationally
Complete Query Languages

Andreas Blass1 , Yuri Gurevich2 , and Jan Van den Bussche3

1
University of Michigan, USA
[email protected]
2
Microsoft Research and
University of Michigan, USA
[email protected]
3
Limburg University, Belgium
[email protected]

Abstract. Abstract state machines (ASMs) form a relatively new com-

putation model holding the promise that they can simulate any compu-
tational system in lockstep. In particular, an instance of the ASM model
has recently been introduced for computing queries to relational databa-
ses. This model, to which we refer as the BGS model, provides a powerful
query language in which all computable queries can be expressed. In this
paper, we show that when one is only interested in polynomial-time com-
putations, BGS is strictly more powerful than both QL and while new , two
well-known computationally complete query languages. We then show
that when a language such as while new is extended with a duplicate eli-
mination mechanism, polynomial-time simulations between the language
and BGS become possible.

1 Introduction
Abstract state machines (ASMs) were introduced as a new computation model,
accompanied by the “ASM thesis” stating that any algorithm, or more bro-
adly, any computational system, at any level of abstraction, can be simulated in
lockstep by an ASM [7,13,14,15]. Recently, Blass, Gurevich, and Shelah (BGS)
introduced an instance of the ASM model for expressing queries to relational
databases [8].
Roughly, a BGS program is a complex rule, changing the values of certain
dynamic functions at various arguments during the run of the program. Rules are
built up from elementary updates by conditionals and parallel composition. The
program is iterated until a halting condition is reached. A powerful sublanguage
of terms provides set-theoretic operations on arbitrarily nested sets over the
input data elements. Once “activated,” these sets are incorporated in the run
of the program, and can become arguments and values of dynamic functions.
While any computable query can be expressed in BGS, the actual motivation
of BGS to introduce their model was to study the complexity class denoted by
e
CPTIME, corresponding to BGS programs under a polynomial time restriction.

Y. Gurevich et al. (Eds.): ASM 2000, LNCS 1912, pp. 22–33, 2000.
c Springer-Verlag Berlin Heidelberg 2000

 Abstract State Machines and Computationally Complete Query Languages 23

Computationally complete query languages have been known in database

e
theory for some years now [1], and complexity classes similar to CPTIME, de-
noted by GEN-PTIME and GEN-PSPACE, were introduced by Abiteboul and
Vianu [6]. These classes can be defined in terms of the language while new .1 This
language is the extension of first-order logic with the following features: (1) as-
signment to relation variables; (2) sequential composition; (3) while-loops; and
(4) the introduction of new data elements in terms of tuples of existing ones. All
computable queries can be expressed in while new . The complexity classes GEN-
PSPACE and GEN-PTIME are obtained by putting polynomial space and time
restrictions on while new programs. Abiteboul and Vianu illustrated the effect of
such restrictions by showing that under a polynomial space restriction, while new
programs can no longer check the parity of the cardinality of a set.
The advent of the BGS model thus raises the natural question: how does
e
CPTIME compare to GEN-PTIME? We will show that CPTIME e is strictly
stronger than GEN-PTIME, in the sense that there are classes of structures
e
that can be separated in CPTIME but not in GEN-PSPACE (and hence neither
2
in GEN-PTIME). We also identify the reason for this inequality: while new only
has tuple-based invention: new data elements can only be introduced in terms of
tuples of existing ones. By repeated application of tuple-based invention one can
construct arbitrary lists. BGS, on the other hand, allowing the construction of
arbitrary sets, also has a form of set-based invention. In the absence of an order
on the data elements, it is impossible to simulate sets (which are unordered)
using lists (which are ordered) without introducing a lot of duplication.
Our result should be correctly compared to what is known from the theory of
object-creating query languages. It is already known [18] that set-based invention
cannot be expressed in while new . However, this is a statement about object-
creating queries where invention is not merely a tool to give more power to
query languages, but where we really want to see the new data elements in the
result of the query. When only considering standard domain-preserving, or even
just boolean queries, set-based invention seemed less relevant because for such
queries while new is already complete. Our results show that set-based invention
is still relevant, but we have to take complexity into account to see it.
When while new is extended with set-based invention, we show that the lan-
guage obtained, denoted by while sets
new , becomes polynomial-time equivalent with
BGS (in a sense that will be made precise). Our work is thus related to the
update language detTL for relational databases, introduced by Abiteboul and
Vianu [3,5]. Some of the spirit of the ASM model (of which BGS is an instance)
is clearly present in detTL, and the equivalence between detTL and while new

1
Abiteboul and Vianu used the name while invent in their paper [6], but use the name
while new in their book with Hull [1], so we use the latter name.
2
A program separates two classes K0 and K1 if it outputs ‘false’ on all structures in
K0 and ‘true’ on all structures in K1 .
24 A. Blass, Y. Gurevich, and J. Van den Bussche

seems to go without saying.3 New to our result are the programming with sets
and the added focus on polynomial time.
We conclude this introduction by mentioning some other related work. The
very first computationally complete query language was QL, introduced by
Chandra and Harel [9]. Because QL can be simulated in while new with only
a polynomial time overhead [6,19], our negative result concerning while new ap-
plies as well to QL. We also should note that the well-known object-creating
query language IQL, introduced by Abiteboul and Kanellakis [2], was set in a
complex-object data model with set values, where the distinction between tuples
and sets is blurred as one can always have a tuple with a set as a component. In-
deed, IQL is polynomial-time equivalent to while sets
new [19] and thus also to BGS.
Finally, we point out that interest in object creation in query languages has re-
cently resurged in the context of Web databases [12]. Current proposals in this
field introduce new data elements by constructing terms, and thus essentially
employ tuple-based invention.

2 Preliminaries

A relational database scheme is modeled by a finite relational vocabulary, in

the sense of mathematical logic [11], i.e., a finite set of relation names with
associated arities. A relational database over a scheme Υ is modeled by a finite
structure B over Υ , i.e., a finite domain D and, for each relation name R ∈ Υ ,
a relation RB ⊆ Dr , where r is the arity of R. The reader is assumed to be
familiar with the syntax of first-order logic formulas over Υ , and the notion of
truth of a formula ϕ in a structure B.
We next briefly describe the languages while new , while sets
new , and BGS. For full
details we refer to the literature [1,8,19].

2.1 The Language while new

An FO statement is any expression of the form

X := {(x1 , . . . , xj ) | ϕ}

where X is a j-ary relation name, and ϕ(x1 , . . . , xj ) is a first-order formula. A

tuple-new statement is any expression of the form

Y := tup-new{(x1 , . . . , xj ) | ϕ}

where Y is a relation name of arity j + 1, and ϕ is as before.

Programs in the language while new are now defined as follows: FO statements
and tuple-new statements are programs; if Π1 and Π2 are programs, then so is
3
To witness, in their book with Hull [1], Abiteboul and Vianu refer to their paper on
detTL [5] as the original source for the language while new , although no language in
the style of while new is discussed in that paper.
 Abstract State Machines and Computationally Complete Query Languages 25

their composition Π1 ; Π2 ; and if Π is a program and ϕ is a first-order sentence,

then the while-loop while ϕ do Π od is a program.
Let Π be a program, let Υ be the vocabulary consisting of all the relation
names mentioned in Π, and let A be a finite Υ -structure. The result of applying
Π to A, denoted by Π(A), is the Υ -structure defined as follows:
– If Π is the FO statement X := {(x1 , . . . , xj ) | ϕ}, then Π(A) equals A
except for the interpretation of X, which is replaced by
{(a1 , . . . , aj ) ∈ Aj | A |= ϕ(a1 , . . . , aj )}. (∗)
– If Π is the tuple-new statement Y := tup-new{(x1 , . . . , xj ) | ϕ}, then
Π(A) equals A in the interpretation of every relation name other than Y .
The domain of Π(A) is that of A, extended with as many new elements as
there are tuples in the above set (∗). Let ι be an arbitrary bijection between
the set (∗) and these new elements. Then the interpretation of Y in Π(A) is
defined as
{(ā, ι(ā)) | A |= ϕ(ā)}.
– If Π is of the form Π1 ; Π2 then Π(A) equals Π2 (Π1 (A)).
– If Π is of the form while ϕ do Π̄ od, then Π(A) equals Π̄ n (A), where n is
the smallest natural number such that Π̄ n (A) 6|= ϕ. If such a number does
not exist, then Π(A) is undefined (the while-loop does not terminate).
By the semantics of tuple-new statements (second item), Π(A) is clearly
defined up to A-isomorphism only (isomorphisms that leave A pointwise fixed).
This is OK, because the particular choice of the newly invented domain elements
really does not matter to us. When doing a complexity analysis, we will assume
that the domain of A is an initial segment of the natural numbers, and that a
tuple-new statement simply extends this initial segment.
When Υ0 is a subset of Υ , and A is an Υ0 -structure, we can view A also as an
Υ -structure by setting A(X) empty for every relation name X in Υ not in Υ0 . In
this way we can also talk about Π(A). This convention formalizes the intuition
of initializing relation names not part of the vocabulary of the input structure
to the empty set. These relation names are used by the program as variables to
do its computation and to contain its final output.

2.2 The Language while

The sublanguage obtained from while new by disallowing tuple-new statements
is called while and has been extensively studied [1,10]. In finite model theory,
the language while is better known under the equivalent form of first-order logic
extended with the partial fixpoint operator [4].

2.3 The Language while sets

new

A set-new statement is an expression of the form

Y := set-new{(x, y) | ϕ},
26 A. Blass, Y. Gurevich, and J. Van den Bussche

where Y is a binary relation name, and ϕ(x, y) is a first-order formula.

The result Π(A) of applying this set-new statement Π to a structure A,
equals A in the interpretations of every relation name other than Y . In order
to define the domain of Π(A) and its interpretation of Y , consider the binary
relation
S = {(a, b) ∈ A2 | A |= ϕ(a, b)}.
We can view this relation as a set-valued function in the canonical way: for any
a in the first column of S, S(a) := {b | (a, b) ∈ S}.4 Now the domain of Π(A)
is that of A, extended with as many new elements as there are different sets in
the range of S. Let ι be an arbitrary bijection between the range of S and these
new elements. Then the interpretation of Y in Π(A) is defined as

{(a, ι(S(a))) | ∃b : S(a, b)}.

For example, the result of applying

Y := set-new{(x, y) | E(x, y)}

to the structure with domain {1, 2, 3} where E equals

{(1, 1), (1, 2), (2, 1), (2, 2), (3, 1), (3, 2), (3, 3)},

is the structure with domain {1, 2, 3, 4, 5} where Y equals

{(1, 4), (2, 4), (3, 5)}.

By adding set-new statements to the language while new , we obtain the lan-
guage while sets
new .

2.4 The BGS Model

BGS takes a functional point of view: computing means updating the values of
certain user-defined, named, “dynamic” functions at various arguments. Argu-
ments and values can be elements of the domain D of the input structure, as
well as hereditarily finite sets built over D during the execution of the program.
Formally, the set HF(D) of hereditarily finite sets over D is the smallest set
such that if x1 , . . . , xn are in D ∪ HF(D), then {x1 , . . . , xn } is in HF(D). Every
dynamic function name has an associated arity r, and thus has, at any stage
of the computation, an interpretation (which can be updated in later stages) as
a function from (D ∪ HF(D))r to D ∪ HF(D). The extent of such a function f
is the set {(x̄, f (x̄)) | x̄ ∈ (D ∪ HF(D))r and f (x̄) 6= ∅}. At any stage of the
computation, the extent of the interpretation of any dynamic function will be
finite.
A number of static functions, which cannot be updated, are predefined: The
relations of the input structure are given as boolean functions. The usual logical
4
In SQL terminology this corresponds to grouping by the first column.
 Abstract State Machines and Computationally Complete Query Languages 27

constants5 and functions (true, false, and, or, not, equality) are provided. Finally,
some set-theoretic constants and functions are provided: the empty set; the input
domain; set membership; set union; singleton extraction, and pairing. The input
domain is called ‘Atoms’. Union is unary, working on a set of sets.
Terms can now be built up from variables, constants, functions, and the set
constructor {t : v ∈ r : g}, where v is a variable that does not occur free in term
r but can occur free in term t and boolean term g. Variable v becomes bound
by the set constructor. The semantics is the obvious one of {t : v ∈ r and g}.
Finally, rules express transitions between states by updating the dynamic
functions. Elementary update rules are of the form f (t1 , . . . , tj ) := t0 , where f
is a dynamic function name (of arity j) and t1 , . . . , tj are terms. The semantics
is obvious. From elementary update rules more complex rules can be built by
conditionals and parallel composition. More specifically:

– If g is a boolean term and R1 and R2 are rules, then so is if g then R1 else

R2 endif, again with the obvious semantics.
– If v is a variable, r is a term in which v does not occur free, and R0 is a rule
in which v can occur free, then forall v ∈ r do R0 enddo is a rule in which
v becomes bound. The semantics is to perform R0 in parallel for all v ∈ r,
except if this yields conflicting updates in which case we do nothing.

A BGS program now is simply a rule without free variables. A program Π is

started in the initial state, where all dynamic functions have the empty extent,
and all static functions are initialized by the input structure I. In a run of the
program, successive states are computed, until the dynamic boolean constant
‘Halt’ (which is present in all programs) becomes true. The final state is then
the result Π(I). As with while new programs, a BGS program may not terminate
on some inputs.

2.5 Examples

An example of a while new program is shown in Figure 1, and an example of a

BGS program is shown in Figure 2. Both example programs work on directed
graphs, modeled as structures whose domain is the set of nodes and which have
a binary relation E holding the edges. Both programs compute, for all pairs
of nodes (x, y), all shortest paths from x to y. They do not follow exactly the
same algorithm; the while new program does a single-source single-target search
in parallel for all source-target pairs (x, y), while the BGS program does a single-
source all-targets search in parallel for all sources x.
In the while new program, a path x1 . . . xn is represented by invented values
p1 , . . . , pn such that the following relations, defined by the program, hold:
Path(x1 , xn , pi ) for i = 1, . . . , n; Ref (pi , xi ) for i = 1, . . . , n; and Child (pi , pi+1 )
for i = 1, . . . , n − 1. The relations Frontier and X used in the program are
auxiliary variables.
5
As usual, constants are viewed as zero-ary functions.
28 A. Blass, Y. Gurevich, and J. Van den Bussche

Path := tup-new{(x, y) | x = y};

Ref := {(p, x) | ∃y Path(x, y, p)};
Frontier := {(x, y, p, z) | Path(x, y, p) ∧ E(x, z) ∧ z 6= x};
while Frontier 6= ∅ do
X := tup-new{(x, y, p, z) | Frontier (x, y, p, z)};
Path := {(x, y, q) | Path(x, y, q) ∨ ∃p∃z X(x, y, p, z, q)};
Ref := {(q, z) | Ref (q, z) ∨ ∃x∃y∃p X(x, y, p, z, q)};
Child := {(p, q) | Child (p, q) ∨ ∃x∃y∃z X(x, y, p, z, q)};
Frontier := {(x, y, q, z 0 ) | ∃p∃z(X(x, y, p, z, q) ∧ z 6= y ∧ E(z, z 0 ))}
od;
Path := {(x, y, p) | ∃p0 (Path(x, y, p0 ) ∧ Ref (p0 , y))}.

Fig. 1. A while new program computing all-pairs shortest paths.

if Mode = 0 then
forall x ∈ Atoms do
Reached (x) := {x},
Paths(x, x) := {{x}},
Frontier (x) := {x}
enddo,
Mode := 1
endif,
if Mode = 1 then
forall x ∈ Atoms do
Old Frontier (x) := Frontier (x),
Frontier (x) := {y : y ∈ Atoms : y 6∈ Reached (x)
and {z : z ∈ Frontier (x) : E(z, y)} =
6 ∅}
enddo,
Mode := 2
endif,
if Mode = 2 then
forall x ∈ Atoms do
forall y ∈ Frontier (x) do
Paths(x, y) := {(p, y) : S
p ∈ {Paths(x, z) : z ∈ Old Frontier (x) : E(z, y)} : true}
enddo,
Reached (x) := Reached (x) ∪ Frontier (x)
enddo, S
Halt := {Frontier (x) : x ∈ Atoms : true} = ∅,
Mode := 1
endif.

Fig. 2. A BGS program computing all-pairs shortest paths.

 Abstract State Machines and Computationally Complete Query Languages 29

In the BGS program, a path x1 . . . xn is represented by a pair (x1 . . . xn−1 ,

xn ), where the x1 . . . xn−1 is again represented by a pair, recursively.6 The base
case n = 1 is represented by a singleton {x1 }. The program updates a dynamic
binary function Paths such that Paths(x, y) equals the set of shortest paths from
x to y. Other dynamic functions and constants used by the program to aid the
computation are Mode, Reached , Frontier , and Old Frontier . The comma bet-
ween rules denotes parallel composition, and is a shorthand for a trivial forall
do construct. The natural numbers 0, 1, and 2 assigned to Mode are in HF(D)
by their definition as von Neumann numerals: 0 is the empty set, and n > 0 is
{0, . . . , n − 1}, recursively [16]. The numbers 0 and 1 also play the role of the
booleans false and true.

3 BGS and while new under Polynomial Time

In this section, we define what it means for two classes of structures over the
same vocabulary to be separable in polynomial time by BGS programs, or by
while new programs. We then prove that there exists a pair that is separable in
polynomial time by a BGS program, but not by any while new program.
During the run of a BGS program on a structure with domain D, a certain
number of sets in HF(D) are activated, meaning that at some point they appear
in the extent of some dynamic function. Elements of active sets are also con-
sidered to be active, and this holds recursively. Similarly, during the run of a
while new program on a structure, a certain number of new elements are inven-
ted. Activated sets and invented elements yield measures of space usage by BGS
and while new programs, which are quite rough, but sufficient for our purposes.
Equally rough measures of time spent by BGS and while new programs can be
defined as follows: the time spent by a BGS program on a structure is the num-
ber of times the program is iterated until the halting condition is reached; the
time spent by a while new program on a structure is the number of times an FO
or tuple-new statement is executed during the run of the program.
In the following two paragraphs fix two disjoint classes K0 and K1 of struc-
tures over a common vocabulary.
Let Π be a BGS program using a boolean dynamic constant Output for
output. We say that Π separates K0 from K1 if for any structure A ∈ K0 ∪ K1 ,
the value of Output in Π(A) is false if A ∈ K0 , is true if A ∈ K1 . We say
that Π separates K0 from K1 in polynomial time if moreover, there exist two
polynomials p(n) and q(n) such that for any A ∈ K0 ∪ K1 , Π runs on A for at
most p(n) time, and activates at most q(n) sets, where n is the cardinality of
the domain of A.
Similarly, let Π be a while new program having some relation variable Output.
We say that Π separates K0 from K1 if Π(A) is defined for any structure A ∈
K0 ∪ K1 , and relation Output in Π(A) is empty if A ∈ K0 , and is not empty if
A ∈ K1 . We say that Π separates K0 from K1 in polynomial time if moreover,
6
Recall that ordered pairs (x, y) are by definition in HF(D), as {{x}, {x, y}} [16].
30 A. Blass, Y. Gurevich, and J. Van den Bussche

there exist two polynomials p(n) and q(n) such that for any A ∈ K0 ∪ K1 , Π
runs on A for at most p(n) time, and invents at most q(n) elements, where n is
the cardinality of the domain of A.
Since we do not care what the programs do on structures outside K0 and K1 ,
the above notion of separation is quite liberal. Still, we will be able to obtain a
negative result regarding the separating power of while new in polynomial time.
Also, in our definition, it is important to polynomially restrict the space used as
well as the time, because in BGS or while new it is possible to use an exponential
amount of space even in an only linear amount of time.
We can prove (proof omitted):

Theorem 1. There exist pairs of classes of structures that can be separated in

polynomial time by a BGS program, but not by a while new program.

Consider the vocabulary consisting of a single relation name P , which is

unary. For any natural number n, define a structure In over this vocabulary as
follows. The domain of In consists of 2n elements. Exactly n of these satisfy
the predicate P . The pair now for which we are going to prove the theorem
was already considered by Blass, Gurevich and Shelah [8] and is the following:
K0 = {In | n even}, and K1 = {In | n odd}. We can easily separate K0 from
K1 by a BGS program in polynomial time: the program generates all subsets of
P with even cardinality (which is in polynomial time because the cardinality of
the input domain is 2n ), and then checks whether P itself was generated.
We can actually show that K0 cannot be separated from K1 by any while new
program that can invent only a polynomial number of elements; the time spent
by the program will be irrelevant.
Because of the equivalence between while new and the generic machine model
of Abiteboul and Vianu [6], Theorem 1 implies that generic machines are strictly
weaker than BGS in the context of polynomial time computation. This result
corrects a tentative claim (‘the simulation in the reverse direction can, it seems,
be carried out using the “form and matter” considerations in Section 9’) near
the end of Section 1 of the BGS paper [8]. The form and matter considerations
mentioned there involve tuples rather than sets as “matter” and therefore run
into the same duplication problem as while new .

4 Polynomial Time Equivalence of BGS and while sets

new

In this section, we formally define notions of polynomial-time simulation of BGS

programs by while sets
new programs, and vice versa, and show that such simulations
exist.

4.1 Simulating while sets

new in BGS

To simulate while sets

new in BGS, we need some way to represent elements that are
invented by a while sets
new program by hereditarily finite sets that can be construc-
ted by a BGS program. For elements invented by a tup-new statement, we
 Abstract State Machines and Computationally Complete Query Languages 31

already did this in the previous section, where we described a list-construction

semantics for tup-new.7 So it remains to describe a set-construction semantics
for set-new.
To this end, recall the semantics of a set-new statement Y := set-new S
on a structure A (where S is a binary relation on A defined by some first-
order formula), which assigns to relation name Y the relation {(a, ι(ϕ(a))) | ∃b :
S(a, b)} for some bijection ι from the range of S (viewed as a set-valued function)
to new elements. We fix this bijection ι uniformly as follows. Assume it is the
mth time we are performing a tuple-new or set-new statement in the execution
of the program. Then ι(S(a)) is defined to be the pair

(S(a), λm ),

where λm is as defined in the previous section.

We now say that a BGS program Π 0 simulates a while sets new program Π if for
every input structure I, if Π(I) is defined then so is Π 0 (I), and for every relation
variable X of Π, say of arity r, there is an r-ary boolean dynamic function X b of
Π 0 , such that the tuples in X in Π(I) are exactly the tuples at which X b is true
in Π 0 (I). Moreover, we say that the simulation is linear-step, polynomial-space if
there exists a constant c and a polynomial p such that for every input structure
I where Π(I) is defined, the following holds. Let the time for which Π runs on
I be t, and let the number of invented elements during the run be s. Then Π 0
runs on I for at most ct time, activating at most p(n + s) sets, where n is the
cardinality of the domain of I.
Here, in close analogy to what we defined for while new programs at the begin-
ning of Section 3, we define the time spent by a while sets
new program on a structure
as the number of times an FO, tuple-new, or set-new statement is executed
during the run of the program.
Note that, while we allow a polynomial overhead in space usage, we allow
only a linear overhead in the running time of the simulation. A weaker notion
of polynomial time simulation could be defined, allowing a polynomial overhead
also for running time, but we will not need to consider this weaker notion as we
will be able to obtain positive results for our stronger notion.
We can show (proof omitted):

Theorem 2. Every while sets

new program can be linear-step, polynomial-space si-
mulated by a BGS program.

4.2 Simulating BGS in while sets

new

To simulate BGS in while sets

new , we need some way to represent hereditarily finite
sets by invented elements. To this end, we observe that for any finite domain
D, the structure (D ∪ HF(D), ∈) is an (infinite) directed acyclic graph. At any
stage in the run of a BGS program on a structure with domain D, the active
7
Lists are special kinds of sets: a list of length n is a mapping from {1, . . . , n} to the
set of members of the list, and a mapping is a set of ordered pairs.
32 A. Blass, Y. Gurevich, and J. Van den Bussche

sets, together with the elements of D, generate a finite subgraph of this graph.
The simulating while sets
new program will maintain a copy of that subgraph, where
the active sets are represented by invented elements, and the elements of D
are represented by themselves. The membership relation ∈ will be stored in a
relation variable Epsilon.
We now say that a while sets 0
new program Π simulates a BGS program Π if
for every input structure I, if Π(I) is defined then so is Π 0 (I), and for every
dynamic function name f of Π, say of arity r, there is an (r + 1)-ary relation
variable fb of Π 0 , such that fb in Π 0 (I) equals exactly the extent of f in Π(I),
under a representation of the active hereditarily finite sets by invented elements
as given in relation Epsilon in Π 0 (I). Moreover, we say that the simulation is
linear-step, polynomial-space if there exist a constant c and a polynomial p such
that for every input structure I where Π(I) is defined, the following holds. Let
the time for which Π runs on I be t, and let the number of sets activated during
the run be s. Then Π 0 runs on I for at most ct time, inventing at most p(s)
elements.8
We can show (proof omitted):

Theorem 3. Every BGS program can be linear-step, polynomial-space simulated

by a while sets
new program.

Acknowledgment. Thanks go to Marc Spielmann for proofreading an earlier

draft of this paper.

References

1. S. Abiteboul, R. Hull, and V. Vianu. Foundations of Databases. Addison-Wesley,

1995.
2. S. Abiteboul and P.C. Kanellakis. Object identity as a query language primitive.
Journal of the ACM, 45(5):798–842, 1998.
3. S. Abiteboul and V. Vianu. Procedural and declarative database update languages.
In Proceedings 7th ACM Symposium on Principles of Database Systems, pages 240–
250, 1988.
4. S. Abiteboul and V. Vianu. Fixpoint extensions of first-order logic and Datalog-
like languages. In Proceedings Fourth Annual Symposium on Logic in Computer
Science, pages 71–79. IEEE Computer Society Press, 1989.
5. S. Abiteboul and V. Vianu. Procedural languages for database queries and updates.
Journal of Computer and System Sciences, 41(2):181–229, 1990.
6. S. Abiteboul and V. Vianu. Generic computation and its complexity. In Proceedings
23rd ACM Symposium on the Theory of Computing, pages 209–219, 1991.
8
The reader will have noticed that, while here we require that Π 0 invents at most
p(s) elements, in the notion of polynomial-space simulation of while setsnew programs
by BGS programs as defined in the previous subsection, we allowed the simulating
BGS program to activate p(n + s) sets. The reason for this is that, even if a while sets
new
program Π does not invent any new elements (i.e., s = 0), a simulating BGS program
still needs to activate some sets just to evaluate the first-order formulas used in Π.
 Abstract State Machines and Computationally Complete Query Languages 33

7. Abstract state machines Web pages. hwww.eecs.umich.edu/gasmi.

8. A. Blass, Y. Gurevich, and S. Shelah. Choiceless polynomial time. Annals of Pure
and Applied Logic. To appear; also available from [7].
9. A. Chandra and D. Harel. Computable queries for relational data bases. Journal
of Computer and System Sciences, 21(2):156–178, 1980.
10. H.-D. Ebbinghaus and J. Flum. Finite Model Theory. Springer, 1995.
11. H.-D. Ebbinghaus, J. Flum, and W. Thomas. Mathematical Logic. Undergraduate
Texts in Mathematics. Springer-Verlag, 1984.
12. M. Fernández, D. Florescu, J. Kang, A. Levy, and D. Suciu. Catching the boat
with Strudel: Experiences with a Web-site management system. SIGMOD Record,
27(2):414–425, 1998. Proceedings ACM SIGMOD International Conference on
Management of Data.
13. Y. Gurevich. Evolving algebras: An attempt to discover semantics. Bulletin of the
European Association for Theoretical Computer Science, 43:264–284, 1991.
14. Y. Gurevich. Evolving algebra 1993: Lipari guide. In E. Börger, editor, Specification
and Validation Methods, pages 9–36. Oxford University Press, 1995.
15. Y. Gurevich. May 1997 draft of the ASM guide. Technical Report CSE-TR-336-97,
University of Michigan, EECS Department, 1997.
16. P. Halmos. Naive Set Theory. Van Nostrand Reinhold, 1960.
17. P.G. Kolaitis and M.Y. Vardi. Infinitary logics and 0-1 laws. Information and
Computation, 98(2):258–294, 1992.
18. J. Van den Bussche and J. Paredaens. The expressive power of complex values in
object-based data models. Information and Computation, 120:220–236, 1995.
19. J. Van den Bussche, D. Van Gucht, M. Andries, and M. Gyssens. On the comple-
teness of object-creating database transformation languages. Journal of the ACM,
44(2):272–319, 1997.
 On Verification of Refinements of
Timed Distributed Algorithms

J. Cohen1 and A. Slissenko1,2

1
UniversityParis-12
Dept. of Informatics
61 Av. du Gén. de Gaulle
94010, Créteil, France
{cohen,slissenko}@univ-paris12.fr
2
St.Petersburg Institute for
Informatics and Automation of
Russian Academy of Sciences

Abstract. This work is an attempt to apply Gurevich Abstract State

Machines methodology to the verification of refinements of real-time dis-
tributed asynchronous algorithms. We define refinements following the
semantical framework of observability, however, with respect to chosen
pieces of the program. The time we consider is continuous as our moti-
vation is related to systems of control that are usually specified within
continuous time framework; the same framework is valid for discrete time.
We remark that refinement of timed programs is not a simple replace-
ment of a part of a program by making it more detailed. As an example
to illustrate this we take Lamport’s Bakery Algorithm with real-time
constraints. However, one of the key questions related to the verification
of refinements is the preservation of verification proofs for the non refined
initial algorithm as much as possible when verifying the refinement. This
is the case for the notion of refinement we define. We introduce a notion
of asynchronous timed distributed algorithm, define its semantics and di-
scuss in what logic can be expressed its functioning. Then we introduce
notions of refinement, and consider a refinement of interprocess commu-
nication of real-time Lamport’s Bakery algorithm using parallel message
exchange. Such a refinement, contrary to our intuition, demands some
non evident transformation of the initial algorithm to make it correct.

1 Introduction

The goal of this paper is to define asynchronous timed algorithms and their re-
finements. To illustrate our system of notions we study a concrete algorithm,
namely a real-time version of Lamport’s Bakery, within this framework. (In fact,
the notions we introduce, are intended to model more complex algorithms, com-
posed from many modules with non trivial asynchronous interaction.) Though
the classical notion of asynchronous algorithm does not have any metric time
constraints, practical implementation of such an algorithm usually cannot have,

Y. Gurevich et al. (Eds.): ASM 2000, LNCS 1912, pp. 34–49, 2000.
c Springer-Verlag Berlin Heidelberg 2000

 On Verification of Refinements of Timed Distributed Algorithms 35

say, unlimited delays, and to be practical, some time constraints may be neces-
sary, at least at some level of refinement of such an algorithm. Moreover, the
time is very intuitive, and this is one of the reasons why time arguments are
largely used in reasoning about distributed algorithms.
What is more important concerning the time, is that it often appears in
initial requirements specification. So if we rewrite directly the initial specification
using time one can hardly require formal justification of this passage, but if
we eliminate time then, clearly, we modify the initial specification, and this
modification is to be justified. But formal justification will demand an adequate
notion of time to express what we are modifying. For example, Lamport in [7]
writes, giving a requirements specification of the critical section problem : ”At
any time, at most one computer may be in its critical section”. Clearly, we
can directly rewrite it as ∀ t (CS(t, p) → ∀ q 6= p ¬ CS(t, q)), where t is a time
variable, p and q are variables for processes and CS(t, r) is a predicate which
states that ”process r is in critical section at the moment t”. And that is how
we proceed here.
The paper [2] by E. Börger, Yu. Gurevich and D. Rozenzweig was a star-
ting point of our work. We use the methodology of Gurevich Abstract State
Machines (GASM) and not the exact notion from, say, [6]. More precisely, we
assemble in the style of GASM a minimal notion sufficient for our goal from
simple programming constructions starting from assignments as used in GASM.
The GASM methodology gives basic principles how to define semantics of such
algorithms. To be adequate to Lamport’s text [7] we introduce asynchronous
executions of our distributed algorithms with delays. To give to the algorithm
a real-time flavor we consider here restricted delays, and thus, slightly deviate
from the original Lamport’s Bakery.
Within our semantics the Lamport’s Bakery algorithm rewritten in our sy-
stem of notations can be proved to be correct following Lamport’s original proof.
However, our main goal is to define a notion of refinement of practical sig-
nificance and to study whether the verification proof for the non-refined initial
algorithm can be preserved for the verification of the refined one. In our defini-
tion of refinement we follow the idea of observational equivalence related to some
structural properties of the program. And for this notion the proof preservation
mentioned above takes place. Constructing a correct refinement remains a non-
trivial problem, and we illustrate it by giving a refinement of communications
in real-time Lamport’s Bakery. We also remark that a straightforward ‘local’
replacement of operators that we wish to refine by their refinements does not
give a correct refinement.
Though the notion of refinement was studied in many papers (e. g. see
[9,8,3,1,5]), our notion treats real-time algorithms, where the time can be used
without any restrictions and the refinement itself is being done with respect to
chosen pieces of the program. These properties are motivated by the analysis of
practical refinements and the verification problem, and all this differs our notion
from the notions of refinement considered in other papers. For example, in the
paper [8] the time is represented only as ticks. The properties of such systems are
36 J. Cohen and A. Slissenko

described in RTTL (Real-time Temporal Logic) where concrete bounds on the

number of ticks are admissible as real-time constraints. Treating refinement as
an observational equivalence the author proves a theorem about preservation of
properties under this equivalence. The limited formalisms of [8] may have some
algorithmic advantages, but we look for formalisms that permit to express much
richer properties.
The analysis of refinements and refinement mappings in [1] defined in terms
of runs of state machines touches many subtle questions in a semantical frame-
work. A state machines framework with many compositionality properties is also
developed in [5] aimed at compiler construction.
Paper [3], which gives many references on the study of the refinement, is
concentrated on interactive systems and compositionality. There is no time ex-
plicitly involved. With respect to compositionality our case is different because
of real-time constraints, and we have not yet studied this topic.

2 Asynchronous Distributed Algorithms and Their

Timed Semantics

Introducing explicit time we pursue two goals: first, to give a unified and intuitive
view on semantics and verification proofs and, second, to give possibility to
introduce time constraints in specifications of algorithm when refining it or when
making precisions of initial formulation.

2.1 Asynchronous Distributed Algorithm Definition

The vocabulary of a language to describe asynchronous distributed timed algo-
rithms (further, simply ”algorithms”) consists of sorts and functions (predicates
are treated as a particular case of functions). A part of the elements of the voca-
bulary will have a predefined interpretation, and the remaining elements will be
called abstract. Among abstract functions there are distinguished input or output
ones, the two sets being disjoint. A relation between input and output functions
constitutes the main part of the specification of functioning of the system under
consideration.
Sorts. The sort time T will be interpreted as R≥0 , and will be a subsort of the
sort of reals R. The sort ∞ is interpreted as an element which is greater than all
the elements of T . The sort T ∞ is T ∪ {∞}, and sort T is considered as subsort
of T ∞ . The sort Bool denotes Boolean values. There can be other sorts that are
presumed to be abstract and finite. Each sort is supplied with a set of variables,
these sets are disjoint for different sorts.
Functions. Pre-interpreted constants for reals are all Q. Nullary function CT
represents the current time. Relations {<, ≤} for T and T ∞ are usual, and there
is equality for each sort. Abstract functions (other than predicate of equality)
can be of type X → Z, where X is a product of finite abstract sorts, and Z is
an arbitrary sort.
The abstract functions are partitioned into external and internal. External
ones are those that cannot be changed by algorithm, internal ones are those that
 On Verification of Refinements of Timed Distributed Algorithms 37

can be changed by algorithm. External functions that depend on time are input
functions. Internal functions of zero arity, also called nullary functions, play the
role of identifiers, as well as terms f (X) for a fixed value of its arguments X.
The mentioned arguments X can be viewed as parameters. Identifiers f (X) and
f (Y ) are different if X 6= Y .
Syntax.
T erm ::= V ariable | F unction(T erm1 , ..., T ermn ),
here and in the next line F unction is an n-ary function symbol.
InternalT erm ::= InternalF unction(T erm1 , ..., T ermn ).
F ormula ::= standard first order formula.
Guard ::= Formula without time or real variables. It must be either closed or
may have free variables ω bounded by ForAll ω ∈ Ω, see below the definition
of the operator.
Assignment ::= InternalT erm := T erm.
Asynchronous timed distributed algorithm over a vocabulary V is a pair of
the form (IniState, {P1 , . . . , PN }), where IniState is a formula defining initial
state, i. e. the values of all the abstract functions of V (at time 0 if a function has
a time argument), and each Pi is a program or process that is defined below. (We
use the word ”operator” where in programming languages one uses ”statement”
and where Yu. Gurevich [6] usually uses the word ”rule”.)
P rogram ::= Prog Operator EndProg
Operator ::= Assignment – defined above
| Seq Operator; . . . ; Operator EndSeq – sequential block
| Par Operatork . . . kOperator EndPar – parallel block
| If Guard Then Operator EndIf – conditional operator
| Repeat Operator Until Guard EndRepeat – repeat operator
Extensions of the syntax like
(1) Par ForAll ω ∈ Ω Do Op1 (ω)k . . . kOpk (ω) EndPar, where Ω is a finite
set, (2) Repeat Operator EndRepeat, have obvious meaning.

2.2 Timed Semantics of Asynchronous Distributed Algorithms

The semantics we define presumes that the actions are instantaneous, but bet-
ween consecutive actions there is a delay. Imposing different constraints on de-
lays we can vary details of the semantics. We assume that delays are bounded
and can be different for different operators. Not to complicate the situation
too much, we suppose that there are two delays: one for ”internal” operations
of a process, and the other for ”external” information exchange. That means
that assignments or guards that use information from other processes have ”ex-
ternal” delays. We assume that there are positive constants δint and δext that
give bounds on the mentioned delays: t ≤int t0 ⇔ t < t0 ≤ (t + δint ) and
t ≤ext t0 ⇔ t < t0 ≤ (t + δext ). The notation t ≤delay t0 will mean that the delay
between t and t0 can be determined by syntactical criterion: the delay is internal
if the concerned operator deals only with the identifiers of the same process. For
example, an assignment of a process P is classified as external iff it is of the form
f (η) := θ0 , where f is an internal function of P , and either the terms of list η
or term θ0 contain functions of other processes. In this case we tacitly presume
38 J. Cohen and A. Slissenko

that the process has to send a query to another processes to get the necessary
information, that implies an external delay, and then to receive this information,
that contributes again to the external delay. Similar for guards.
Remark that we can eliminate Zero runs if we impose a positive lower bound
on delays.
The semantics is defined by a propagation of time moments of execution
of operators and by defining simultaneously the values of internal functions.
To simplify the presentation we will give first the rules of propagation of time
though, strictly speaking, some of them involve results of function evaluation.
Below P rog stands for P rogram, Op stands for Operator.
Time Propagation Rules.
Intuitive meaning of vertical arrows below is as follows: “Op ↓ (t)” means “t is
the termination time of the operator”, “↑ (t)Op” means “t is the start time of
the operator”. Double arrow is used when the algorithm takes values or change
values. So ↑ (t) goes down to an assignment, is changed to ↓ (t), and the latter goes
up to the end of the operator. Symbol ` means the expression at the left can be
rewritten as an expression at the right of `. Below t0 is arbitrary moment such
that t0 ≥int t, and t0 , t1 are arbitrary moments such that t1 ≥delay t0 ≥delay t.
• Given an algorithm (IniState, {P rog 1 , . . . , P rog N }) its initial time distribu-
tion is ` {↑ (0)P rog 1 , . . . , ↑ (0)P rog N }.
• The following transition says that the execution is finite and halts:
{P rog 1 ↓ (∞), . . . , P rog N ↓ (∞)} ` {P rog 1 , . . . , P rog N } ↓ (∞).
• ↑ (t)Prog ` Prog ↑ (t0 ). • ↓ (t)EndProg ` EndProg ↓ (∞).
• ↑ (t)Seq ` Seq ↑ (t0 ). • ↓ (t); ` ; ↑ (t0 ). • ↓ (t)EndSeq ` EndSeq ↓ (t0 ).
• ↑ (t)Par Op1 k . . . kOpn EndPar ` Par ↑ (t1 )Op1 k . . . k ↑ (tn )Opn EndPar
for any t1 , . . . , tn ≥int t.
• Par Op1 ↓ (t1 )k . . . kOpn ↓ (tn ) EndPar
` Par Op1 k . . . kOpn EndPar ↓ (t0 ) for any t0 ≥int max{t1 , . . . , tn }.
• ↑ (t)If ` If ↑ (t0 ). • ↓ (t)EndIf ` EndIf ↓ (t0 ).
• If ↑ (t)Guard Then Op EndIf ` If Guard  (t0 ) Then Op EndIf .
• If Guard
  (t) Then Op EndIf
If Guard Then ↑ (t0 )Op EndIf if Guard(t) is true,
`
If Guard Then Op EndIf ↓ (t0 ) otherwise.
• ↑ (t)Repeat ` Repeat ↑ (t0 ). • ↓ (t)Until ` Until ↑ (t0 ).
• Repeat Op Until ↑ (t)Guard EndRepeat
` Repeat Op Until Guard  (t0 ) EndRepeat.
• Repeat
 Op Until Guard  (t) EndRepeat
Repeat ↑ (t0 ) Op Until Guard EndRepeat, if Guard(t) is false,
`
Repeat Op Until Guard EndRepeat ↓ (t0 ), otherwise.
• ↑ (t)θ := θ0 ` θ  (t0 ) := θ0 . • θ  (t0 ) := θ0 ` θ := (t1 )θ0 .
0 0 0
• θ := (t1 )θ ` θ := θ ↓ (t ).
Let θ = f (η1 , . . . , ηn ), where f is an internal function. To change the value of
θ we are to find the values of arguments η1 , . . . , ηn of f and the value of θ0 to
assign. We take all these values at t0 and change the value of θ at t1 ≥delay t0 .
 On Verification of Refinements of Timed Distributed Algorithms 39

Definition of Run (Semantics of Algorithms).

We consider arbitrary inputs. The internal functions in our definition below will
be piecewise constant with pieces being left-open right-closed. Such a definition
was elaborated for another type of GASM in [4].
Given an algorithm A, an interpretation of sorts and values of inputs satis-
fying IniState for t = 0, an execution of A for these inputs defines the values of
its internal functions ρ(t) at least for some prefix T0 of T that together with the
inputs constitute a complete run {e ρ(t)}t∈T0 of the algorithm. Together with ρ
we will define other notions, either auxiliary or giving more detailed information
on the execution of the algorithm to use them in the analysis of refinements.
Clearly, the run ρe(t) which is defined at t determines an interpretation of
the vocabulary. We will denote by f [I], where I is an interpretation of a part
of the vocabulary containing f , the interpretation of f given by I. For example,
f [ρ(t)](X) or f (X)[ρ(t)] is the value of f (X) as defined by ρ(t). We extend this
notation on terms in a usual way. If we have an interpretation I(t) of V for every
time moment we can define for each f ∈ V : X → Z an interpretation I ◦ of its
timed image f ◦ : T × X → Z (symbol f ◦ is in some other vocabulary V ◦ ) as
f ◦ (t, X)[I ◦ ] = f (X)[I(t)]. Consider an algorithm A = (IniState, {P1 , . . . , PN })
over V . We assume that the initial state of A uniquely determines the values of
internal functions at time 0 being given an input. Let the external functions be
given and suppose that the algorithm starts its execution. Denote by ρ(t) the
values of internal functions at time t that are to be defined for t > 0. An internal
function can be changed only by assignments. So we look for the moments of
time at which the assignments are executed. We assume that all the guards
and operators occurring in A are enumerated so that we can speak about their
occurrences. By default, when mentioning a guard or an operator we mean its
occurrence in A.
Applying Time Propagation Rules to the algorithm A, starting from initial
time distribution, we get an expression that is a list of programs with occurrences
of arrow expressions ↓ (t), ↑ (t),  (t) and  (t) for concrete moments t. Such an
occurrence of time will be called arrowed occurrence. An algorithm with arrow
expressions will be called timed algorithm expression (TAE). Remark that to
make an assignment we take the values of all involved terms at the same point.
But the changed value will be valid only after this moment and on a left-open
interval. To memorize the value to assign we introduce for every moment T an
auxiliary value ρ+ (T ). So to define a run we define simultaneously ρ(t), ρ+ (t), a
sequence of TAE P(k) and some other notions.
By P(k) will be denoted the TAE obtained after k applications of Time
Propagation Rules. We define by induction on k time moments Tk , TAEs P(k),
functions ρ(t) for t ≤ Tk , ρ+ (t) for t < Tk and also:
– for every operator Op we define the set ExInt[Op, Tk ] of its intervals of
execution of the form [a, b) by describing consecutively their left and right ends
a and b situated to the left of Tk ,
– for every internal function f of type X → Z and X ∈ X we define a set
V al[f, X, Tk ] of values to take as f (X) at some moment greater or equal to Tk ,
40 J. Cohen and A. Slissenko

and the predicate T m[f, X, Tk ] which says that Tk is a moment to change the
value of f (X). By default, every element of the set V al is deleted from the set
after having been used.
– for each occurrence Assign of an assignment of the form f (η) := θ0 we
define a function Arg[Assign, Tk ] that gives X for which this assignment is to
change f (X) at Tk or later.
Time Propagation Rules are being applied to each occurrence of Tk in P(k).
To simplify the recursive definition, extend T to the left of 0 by, say, some
ε < 0, and set T0 = 0, take ρ(t) for t ∈ [ε, 0] as defined by the initial state, set
ρ+ (t) = ρ(0) for t ∈ [ε, 0) and P(0) = {↑ (0)P1 , . . . , ↑ (0)PN }. All the mentioned
sets are empty for T0 and predicate T m is false.
Suppose that Tk−1 < ∞, and all the mentioned functions and sets, have
been defined for Tk−1 for some k > 0. Let Tk = M inT m(P(k − 1)) be the
minimum of arrowed occurrences of time in P(k − 1). If Tk−1 < Tk then first
apply (UpDate) and then (Propagation) as described below, otherwise apply
directly (Propagation). The procedure (UpDate) is applied when all minimal
time occurrences with the same value have been eliminated – in this case the
time advances, and we make updates and extend all the values to the next
minimal time.
(UpDate): For all internal function f : X → Z and for all X ∈ X such that
T m[f, X, Tk−1 ] do: if #V al[f, X, Tk−1 ] > 1 then run is undefined for τ >
Tk−1 ; if V al[f, X, Tk−1 ] = {v} then set f ◦+ (τ ) = v, otherwise set f ◦+ (τ ) =
Lef tLimt→Tk−1 f ◦+ (t) for τ ∈ [Tk−1 , Tk ). Thus ρ+ (τ ) is extended for τ <
Tk . Now set ρ(τ ) = ρ+ (Tk−1 ) for τ ∈ (Tk−1 , Tk ]. After that, each used set
V al[f, X, Tk−1 ] becomes empty at Tk , and the others are extended to Tk .
(Propagation): If there is an occurrence of the form θ := (Tk )θ0 take it,
otherwise take any arrowed occurrence of Tk , and act according to the case:
(A.1) ↑ (Tk ) θ := θ0 . Replace this occurrence by θ  (t0 ) := θ0 choosing arbitrary
t0 ≥delay Tk . The point Tk is the left end of interval of execution of the assignment
under consideration.
(A.2) θ  (Tk ) := θ0 , where θ = f (η1 , . . . , ηn ). Let X0 = (η1 , . . . , ηn )[ρ(Tk )]
and v0 = θ0 [ρ(Tk )]. Add v0 to V al[f, X0 , Tk ] and set Arg[Assign, Tk ] = X0 ,
where Assign is the occurrence of θ := θ0 under consideration. Then choose any
t0 ≥delay Tk and replace (A.2) by θ := (t0 ) θ0 .
(A.3) θ := (Tk ) θ0 . Set T m[f, X0 , Tk ] and V al[f, X0 , Tk ] = V al[f, X0 , Tk−1 ]
for X0 = Arg[Assign, Tk−1 ]. Replace expression (A.3) by θ := θ0 ↓ (t0 ) for any
t0 ≥int Tk . The point t0 is the right end of interval of execution of the assignment
under consideration.
(B) Occurrences of ↑ (Tk ) and ↓ (Tk ) different from that of (A). Apply Time
Propagation Rules in a evident way. Set V al[f, X, Tk ] = V al[f, X, Tk−1 ] for all
X and Arg[Assign, Tk ] = Arg[Assign, Tk−1 ] for all Assign. The intervals of
execution of operators are defined also in a straightforward way, say, ↑ (t)Seq
gives the left end for the operator starting with this Seq, and EndSeq ↓ (t0 )
corresponding to this Seq gives the right end of the interval of execution.
 On Verification of Refinements of Timed Distributed Algorithms 41

Each application of the propagation rules as mentioned above give a new

TAE which is, by definition, P(k), and we set P(k − 1) ` P(k).
The induction step of the definition is over.
A run is total if it is defined over T or T ∪ {∞}. There are 2 reasons to get
non total run, namely, incompatibility of assignments or Zero run, i. e. a run
where Tk ≤ const < ∞ for k → ∞. By induction that follows the definition of
run one can prove

Lemma 1 In a total run the values of internal functions are piecewise constant
on intervals of the form (t, t0 ], t < t0 .

2.3 On Logic Representation of Runs

Runs of an asynchronous distributed algorithm A can be described as the set
of models of some logical formula. We sketch a first order timed logic over some
vocabulary W to give such a description. Our approach here is straightforward
and seems to be less efficient from the point of view of proof search than the
logic FOTL in [4] for synchronous block GASM.
The syntax of the logic we consider consists, in a way, of two sublanguages L0
and L1 . The language L0 corresponds to the language of A, and L1 describes how
time propagation rules determine runs. Thus, L1 gives an interaction of syntactic
part of the time propagation with the values of arguments and functions. So we
are to distinguish, say, term as a word and the value of the term. In particular
for this reason, we assume that A does not use time variables.
Notice that, given an algorithm, we can write down all occurrences of its
processes and operators. By Op below we will denote a meta-variable for such
occurrences. We could take a logic where Op would be a variable, but such a
logic would be more complicated from the point of view of formal verification.
The vocabulary W0 of L0 has the same sorts as the vocabulary V of A, the
same constants (static nullary functions) and in place of a function f of type
X → Z it has a symbol f 0 of type T × X → Z.
The vocabulary W1 of L1 is a vocabulary to write down TAEs with time
variables in arrow expressions. It contains vocabulary WT AE , words over which,
i. e. the set WT∗ AE , contain all TAEs. But we need also parameterized TAE. The
latter will be represented by terms built with the help of concatenation from
constants representing atoms to construct TAEs (key words, delimiters, arrows
etc.) and from variables for words and time. Thus, time variables are treated
from the point of view of their values when speaking about relations between
TAEs (equality, occurrence etc.). Functions and predicates of W1 permit to de-
scribe occurrences of operators, their form, occurrences of arrowed expressions,
arguments etc. Any occurrence of a word U in W can be described by the num-
ber of its first position in W or as (V1 , U V2 ) or (V1 U, V2 ) etc., where brackets
and comma used here are supposed to be not in the alphabet of W .
We assume that each occurrence of operator in A has a name, and Op will be
used as a variable for such occurrences. Letters Γ , ∆, U , V , W will be variables
for words over WT AE , f a meta-variable for internal function which type will be
42 J. Cohen and A. Slissenko

denoted as X → Z, X a list of variables of sorts X , v a variable for values of f

and T , t, T 0 ... time variables.
The description of run given above can be written as a formula that defines
a unary predicate T AE over WT∗ AE , binary predicate `, predicates
ExInt(Op, t, [a, b)), V al(f, X, t, v), T m(f, X, t) and function Arg(Assign, t) cor-
responding to sets and function with the same names used in the definition of run
(here Assign and Op are meta-variables, thus we have a function ArgAssign (t)
for each occurrence Assign of assignment and similar for ExInt).
Now run ρ◦ is a function of time and is represented as a finite set of symbols
f , each f ◦ being of the type T × X → Z and corresponding to f : X → Z of
◦

the vocabulary of A. To represent ρ◦+ we use symbols f ◦+ of type T × X → Z.

We describe how to write a formula representing runs with comments con-
cerning this description. First, we write a conjunction of formulas related to the
initial state.
(RunIni) ∀ t ∈ (ε, 0] IniState(ρ◦ (t)), initial values of functions constituting
◦
ρ (t) for t ∈ [ε, 0] are determined by the initial state.
(RunPlusIni) ∀ t ∈ [ε, 0) ρ◦+ (t) = ρ◦ (0), here we define the initial values of
functions constituting ρ◦+ (t) for t ∈ [ε, 0].
(TAEIni) T AE({↑ (0)P1 , . . . , ↑ (0)PN }) defines the initial TAE.
(ExecLftIni) 0 is the left end of interval of execution of P rog j , 1 ≤ j ≤ N .
(ValueToAssign) For all internal f : X → Z and X ∈ X predicate
V al(f, X, t, v) is f alse for all t ∈ [ε, 0] and for all v of sort of values of f .
This just described formula is connected by conjunction with the formula
below describing a recursive step. This formula is a conjunction of 2 formulas –
one corresponds to (Update) and the other to (Propagation).
Notations: • Φ=df T m(f, X, T ). • Predicate t = M inT m(W ) means “t is the
minimum of times thatVoccur in arrowed expressions of W ”.
(Update): The prefix f ∀ T ∀ T 0 ∀ Γ ∀ ∆ ∀ X ∀ v is followed by implication from
T AE(Γ ) ∧ T AE(∆) ∧ T = M inT m(Γ ) ∧ T 0 = M inT m(∆) ∧ T < T 0 ∧ Γ `


∆ ∧ V al(f, X, T, v) to:


∀ τ ∈ [T, T 0 ) Φ → f ◦+ (τ, X) = v ∧



¬ Φ → f ◦+ (τ, X) = Lef tLimt→T f ◦+ (t, X) ∧ V al(f, X, T 0 , v) ∧
 
∀ τ ∈ (T, T 0 ] f ◦ (τ, X) = f ◦+ (T, X)
(Propagation): This formula starts with ∀ T ∀ t ∀ Γ ∀ ∆ ∀ U ∀ V and some other
universal quantifiers of more technical nature (we do not mention them) and has
as its scope an implication. This implication has as its premise
(T AE(Γ ) ∧ T = M inT m(Γ )) and as its conclusion a conjunction of formulas
describing possible situations where one can meet an arrowed occurrence of T .
These situations correspond to time propagation rules as in the definition of run.
We consider only 3 situations (S1)–(S3).
(S1) Γ = U If Guard  (T ) Then Op EndIf V , where U and V give the
global context of the occurrence of  (T ) under consideration, Guard and Op are
a concrete guard and a concrete operator of A. We take the formula Guard◦ (T )
that is constructed from Guard by replacing all terms of the form f (η) by their
timed version f ◦ (T, η ◦ ); the operation ◦ goes down to variables and constants
 On Verification of Refinements of Timed Distributed Algorithms 43

that it does not change. So we continue writing the formula: Guard◦ (T ) and
T ≤delay t and ∆ = U If Guard Then ↑ (t)Op EndIf V implies T AE(∆) and
Γ ` ∆ and the extension of V al, T m, Arg to t.
(S2) Γ = U θ  (T ) := θ0 V , where θ = f (η1 , . . . , ηn ). Let v0 = θ0◦ (T ) and
X0 = (η1 , . . . , ηn )◦ (T ). Here, in addition to the discourse concerning ∆, T AE
and `, we set V al(f, X0 , T, v0 ) and Arg(Assign, T ) = X0 .
(S3) Γ = U θ := (T )θ0 V . Here in addition to defining T AE(∆) and Γ ` ∆
we say that for all v such that V al(f, X0 , T, v), where X0 = Arg(Assign, T ) and
Assign is the assignment under consideration, there take place T m(f, X0 , T )
and V al(f, X0 , T, v).

3 Refinement of Communication
3.1 Refinements
A refinement of a distributed algorithm is not a simple replacement of an opera-
tor by another one which is, in a way, more detailed. Though such a replacement
can take place, there may be other changes in the algorithm being refined. We
will see it later for Lamport’s Bakery. We define refinement semantically, in terms
of runs.
Assume that the runs under consideration are total, and the functions of the
vocabulary do not have udef as their value.
Let V ⊆ V1 be two vocabularies of the type described above. A run over
a vocabulary W is an interpretation of sorts and a mapping that give for any
t ∈ T a value of f (X ∗ ) for each f ∈ V : X → Z for every X ∗ ∈ X ∗ , where
X ∗ denotes the interpretation of sorts X . Now suppose that an interpretation of
sorts is fixed, that means in particular, that every sort in V is interpreted in V1
as in V .
Projection of a run ϕ1 over V1 onto V (notation: projV (ϕ1 )) is the run that
is the result of deleting from ϕ1 all identifiers of V1 \ V .
Let A and A1 be algorithms respectively over V and V1 .
We wish to compare runs of A and its refinement A1 modulo refined opera-
tors. As the latter are only in A but not in A1 we are to use some abstraction of
runs of A and that of A1 modulo some sequence of intervals (corresponding to
the intervals of execution of operators to refine) supplied with sets of identifiers
(corresponding to the identifiers being changed by these operators).
Operation W eed takes as its argument an interpretation ϕ of vocabulary
W (in particular, a run of A or that of A1 ) and a finite set of pairs Ω =
{(αi , Fi )}1≤i≤m , where each αi is a sequence of disjoint time intervals (in in-
creasing order) and each Fi is a finite set of identifiers of W different from input
or output ones. The value W eed(Ω, ϕ) is a mapping ϕ0 from time to values of ϕ
extended with udef obtained by the following procedure: for each interval αi (j)
set the values of each f (X) ∈ Fi equal to udef inside αi .
Consider a set S = {Op1 , . . . , Opm } of disjoint operators of A. Denote by
αi the sequence of intervals of execution of Opi and by Fi the set of identifiers
changed by Opi except the output ones. This gives the set Ω.
44 J. Cohen and A. Slissenko

An algorithm A1 is a refinement of A with respect S (operators to refine) if

for every run ϕ of A there exists a run ϕ1 of A1 such that W eed(Ω, projV (ϕ1 )) =
W eed(Ω, ϕ).
Algorithm A1 is a strong refinement of A with respect of a set of disjoint
operators of A (operators to refine) if it is a refinement of A and for every run
of A1 its projection into V is a run of A.
The disjointness of operators in the definition above is not essential as any
two (occurrences of) operators are either disjoint or one is a part of the other. We
suppose, without loss of generality, that the vocabulary of the refined algorithm
intersects the vocabulary of the initial algorithm exactly at inputs/outputs. As
a consequence of the fact that W eed do not touch inputs/outputs we have

Theorem 1 Strongly refined algorithm verifies the requirements iff the non-
refined one verifies them.

3.2 Lamport’s Bakery Algorithm

To illustrate how the introduced notions of run and refinement work we consider
Lamport’s Bakery Algorithm for the Critical Section Problem from [7] deleting
all questions of fault-tolerance vaguely discussed by L. Lamport. We rewrite it as
asynchronous distributed algorithm, constituted by N programs shown on Fig. 1,
where xp [1 : N ] is a local array of process p, and the strict order (a, b) < (c, d)
means that a < c, or a = c and b < d. The function choosing(i) (which role
is slightly obscure) in original Lamport’s Bakery is ensured by the semantics,
so it is omitted (as well as in [2]). The idea of the algorithm is as follows. A
process wishing to enter critical section (CS) gets a ticket (number) calculated
in line 3 and then enters Bakery with this ticket. Then he is waiting until all
previous tickets were satisfied, and enters CS. Tickets are natural numbers, the
number of processes involved is N , CSp is a predicate indicating that a process
p in CS. Instruction CSp means CSp := true, ¬ CSp means that CSp := f alse.
The array xp serves to get tickets of other processes. How p achieves access to
numberq is not discussed and will be the subject of refinement.
L. Lamport does not give rigorous semantics. Our algorithm on Fig. 1 and
semantics preserve the main idea of the Lamport’s Bakery. To be more precise
with respect to Lamport’s Bakery we are to permit an arbitrary delay after
line 8, however it does not change our reasoning on the whole, so we do not do
it. So in our case each process executes numberp := 0 infinitely often. Following
[7] one can prove, that BakeryAsyn satisfies:

(Safety): ∀ pq ∀ t p 6= q → ¬ (CSp (t) ∧ CSq (t)
(No two processes can be in CS at the same time.)
(Liveness):

∀ p ∀ t numberp (t) > 0 → ∃ t0 (t < t0 < t + C1 · δint + C2 · δext ∧ CSp (t0 )
for some natural constants C1 and C2 .
(Each process that wishes to enter CS eventually enters there.)
Strictly speaking, (Safety) and (Liveness) constitute only part of require-
ments specification, namely, specification of functioning. There must be present
 On Verification of Refinements of Timed Distributed Algorithms 45

BakeryAsynp :
Initial values (IniState): numberp = 0, ¬ CS, xp [q] = 0.
Repeat
Seq
-- Doorway:
1: numberp := 1; xp [p] := 1;
2: Par ForAll q 6= p Do xp [q] := numberq ; EndPar
3: xp [p] := 1 + maxq {xp [q]};
-- Bakery:
4: numberp := xp [p];
5: Par ForAll q 6= p Do
Repeat xp [q] := numberq

Until xp [q] = 0 ∨ (xp [p], p) < (xp [q], q)
EndRepeat;
EndPar
-- Critical Section:
6: CSp ;
7: ¬CS p ;
8: numberp := 0
EndSeq
EndRepeat

Fig. 1. Lamport’s Bakery as Distributed Asynchronous Algorithm.

specification of environment. But in our case, as the algorithm has no outputs,

the environment is represented by the semantics of algorithm.
Program for each p, shown on Figure 1, uses its local array xp [1 : N ],
and the interaction of the process p with q is represented as the assignment
xp [q] := numberq . Our goal is to refine this interaction and describe it in terms of
lower level interactions preserving, however, Lamport’s ”high-level” proof of cor-
rectness. To see the problem better we will refer to operator 2 as N umbers(p, xp )
and operator 5 as P riority(p, xp ).

3.3 Refinement of Lamport’s Bakery with Parallel Communications

So we wish to refine N umbersp and P riorityp for all p. In the algorithms descri-
bed above we may have simultaneous reads or writes. But simultaneous reads or
writes implicitly involves solving the problem of critical section, so we are in a
vicious circle which we wish to avoid. In Lamport’s algorithm there are two ty-
pes of parallelism, one concerns the interaction between processes, and the other
is the internal parallelism of one process. To manage the interaction between
the processes we will use a standard communication medium of asynchronous
distributed algorithms (based on send/receive messages).
Firstly, we introduce a new sort, namely Message. Each message µ is a
quadruple (message#, addressee, sender, contents), each component being of
46 J. Cohen and A. Slissenko

the following sort: message# ∈ N, addressee, sender ∈ P rocess, contents ∈

N ∪ {?}. Symbol ? will mean sending a query on the number.
The numbering of messages is necessary only to specify the communication
medium as a set – if the messages are not numbered we cannot distinguish two
equal messages sent at different moments of time that must be received twice.
We assume that a number is attributed automatically and will not mention it.
M edium is a finite set of messages of a priori non bounded size. We assume
that every sent message eventually reaches its destination within the external
delay. To send messages a process p make assignment to one of its internal
functions send that implies immediate sending of this message to M edium. The
messages arriving at p are placed in one or several queues. Extracting a message
is done by an assignment of the form z := F irstQueue which implies that the
first message becomes the value of z and disappears from the queue. If the
queue is empty then the value given by F irstQueuep is ⊥. Each queue has its
own F irstQueue.
A straightforward way of refinement of BakeryAbstr is to replace each
xp [q] := numberq by a sending/receiving of messages preserving the enveloping
parallelism. But it does not work, for example in the case of parallel communi-
cation, as defined below, we may even have deadlocks or other misfunctioning
(see Remark 1 below).
To model parallel communication facility we introduce for each process p
functions sendp,q (to send its queries), sendp,q (to send responses to queries of


others), F irstQueuep,q (to receive responses to its queries), F irstQueue


p,q (to
receive queries of others). In the description of environment one is to include
two queues for each p: Queue p,q and Queuep,q .


The refinement with parallel communication is on the Fig. 2 and is self-

explanatory.
Theorem 2 Algorithm BakeryRfnP is a strong refinement of BakeryAsyn if its
delays are appropriately diminished (that can be easily estimated).
Proof. We prove two assertions marked below by (A) and (B).
(A) ∀ ρ ∈ Runs(BakeryRf nP ) : proj(ρ) ∈ Runs(BakeryAsyn).
Property (A) is implied by the two following claims which are consequences
of the fact that every sent message arrives at its destination within the delay
δext , and that every query from another process will be responded. In the latter
case there may be some internal delay before sending a response, but having
been sent, the response will reach its destination within the external delay. We
do not estimate concrete delays as this is evident.
Claim 1. If [t, t1 ) is an interval of execution of operator 2 by p in the run ρ
then x◦p [q](τq ) = numberq◦ (τq0 ) for some t < τq0 < τq < t1 .
Proof. Suppose that p has sent a query to q executing Operator 2. There is
a moment of time such that Queue q,p contains ?. As each process q repeats
Operator 9.1, we can consider that the value of wq,p from this moment of time
is ?. Thus q sends its number to p within delay O(δint ).
The Medium will provide Queue p,q with numberq and thus, the assignment
zp,q := F irstQueue p,q will give number q as the value of zp,q and further as the
 On Verification of Refinements of Timed Distributed Algorithms 47

BakeryRf nPp :
Repeat
Seq
-- Doorway:
1: numberp := 1; xp [p] := 1;
2: Par ForAll q 6= p Do
2.1: send p,q := (q, p, ?);
2.2: Repeat zp,q := F irstQueue p,q Until Contents(zp,q ) ∈ N EndRepeat
2.3: xp [q] := Contents(zp,q );
EndPar;
3: xp [p] := 1 + maxq {xp [q]};
-- Bakery:
4: numberp := xp [p];
5: Par ForAll q 6= p Do
Repeat send p,q := (q, p, ?);
Repeat zp,q := F irstQueue p,q Until Contents(zp,q ) ∈ N EndRepeat
xp [q] := Contents(zp,q );

Until xp [q] = 0 ∨ (xp [p], p) < (xp [q], q)
EndRepeat;
EndPar
6–7: CSp ; ¬CSp ;
8: numberp := 0
EndSeq
EndRepeat
k
9: Par ForAll q 6= p Do
Repeat
9.1: wp,q := F irstQueue p,q ;
9.2: If Contents(wp,q ) =? Then send p,q := (q, p, numberp ) EndIf
EndRepeat
EndPar

Fig. 2. Lamport’s Bakery with Parallel Communication.

value of xp [q]. It is clear that to send/receive a message the algorithm spends at

most 2δext time plus some more delays δint . Remark that each Queueq,p has at
most one element at a given moment. 

Claim 2. If [t, t1 ) is an interval of execution of operator 5 by p in the run ρ

then for all q 6= p we have xp [q](τq ) = numberq (τq0 ) for some t < τq0 < τq < t1


and xp [q](τq ) = 0 ∨ (xp [p](t0 ), p) < (xp [q](τq ), q) for some t0 < t.
Proof. As in the previous proof, using the properties of the Medium, one can
easily deduce Claim 2. 

(B) ∀ρ0 ∈ Runs(BakeryAsyn)∃ρ ∈ Runs(BakeryRf nP ) :

W eed(proj(ρ)) = W eed(ρ0 ).
48 J. Cohen and A. Slissenko

Let ρ0 ∈ Runs(BakeryAsyn). We construct ρ by induction on the changes

of the values of ρ0 . For the moment 0 the initial values of the both considered
algorithms are compatible. Clearly the only cases to consider are the executions
of operator 2 and operator 5 of the BakeryAsyn algorithm which are refined
by BakeryRf nP . Remark that identifiers of operator 2 and operator 5 of diffe-
rent processes are different, and intervals of execution of operator 2 and that of
operator 5 for the same p are disjoint.
We take an interval ζ = [t, t1 ) of execution of operator 2. And within this
interval of time, we construct an appropriate execution of BakeryRf nP . Take all
the moments τq and τq0 in ζ such that x◦p [q](τq ) = numberq◦ (τq0 ) for BakeryAsyn.
Firstly, we choose delays such that for each q, Operator 2.1 ends before τq0 ,
say at tq in BakeryRf nPq that is Queue q,p contains the query of p at time
tq . Then we can choose delays such that q executes Operator 9.2 and gives
(p, q, numberq◦ (τq0 )) as the value of send
q,p . Then delays are chosen so that p
executes Operator 2.2 and 2.3 and finally gives numberq◦ (τq0 ) as the value of
xp [q] at time τq . The value of the array xp at the end of ζ is the same in both
executions of the two algorithms.
We can deal in the same way with operator 5. 
Theorems 2 and 1 imply
Corollary 1 The refined algorithm BakeryRf nP satisfies (Safety) and (Liven-
ess).

Remark 1. The well functioning of Lamport’s algorithm leans particularly

upon ‘good’ communications between processes. If we wish to make precise such
exchange of information, the very first idea is to focus only on the concerned
operators that is operator 2 and operator 5 in which communications occur.
However, this idea does not lead to a correct refinement. Suppose we autho-
rize only communications during the intervals of time when these two operators
are running. As a consequence, a process can only receive a value 1 or greater
as a response to its query since during the execution of these two operators
numberr 6= 0 for any process r.
Hence, a run of Lamport’s Bakery, where each process stays idle from time
0 with exception of one who reaches critical section, cannot be a run of our
‘refined’ algorithm.
Let p, q be two processes and let us consider the following run of Lamport’s
Bakery:
– numberq (t) = 0 for t ∈ [0, ∞)
– ¬CSq (t) for t ∈ [0, ∞)
– numberp (t) = 0 for t ∈ [0, 1)
– numberp (t) = 1 for t ∈ [1, 2)
– numberp (t) = 2 for t ∈ [2, 4)
– ¬CSp (t) for t ∈ [0, 3)
– CSp (t) for t ∈ [3, 4)
– numberp (t) = 0 for t ∈ [4, ∞)
– ¬CSp (t) for t ∈ [4, ∞)
 On Verification of Refinements of Timed Distributed Algorithms 49

where the intervals of execution of operators 2 and 5 for process p are respectively
[1, 2) and, say, [2.5, 3).
An execution of the ‘refined’ algorithm we consider is supposed to have the
same values for numberp , numberq , CSp and CSq except during these two
intervals of time. Therefore we must have numberq (t) = 0 and ¬CSq (t) for
t ∈ [0, 1) ∪ [2, 2.5) ∪ [3, ∞). To answer to the query of p during [1, 2), q must be
running operator 2 or operator 5. So q will execute operator 2 during [1, 2). In
order to have its number back to value 0, q must execute operators 3 − −8 before
time 2 : it is not possible because during that interval of time, numberp = 1 and
the loop in operator 5 cannot end.

References
1. M. Abadi and L. Lamport. The existence of refinement mappings. Technical
Report 29, DEC Systems Research Center, Palo Alto, California, August, 14
1988.
2. E. Börger, Y. Gurevich, and D. Rozenzweig. The bakery algorithm: Yet
another specification and verification. In E. Börger, editor, Specification and
Validation Methods, pages 231–243. Oxford University Press, 1995.
3. M. Broy. Compositional refinement of interactive systems. J. of the Assoc.
Comput. Mach, 44(6):850–891, 1997.
4. D. Beauquier and A. Slissenko. A first order logic for specification of timed
algorithms: Basic properties and a decidable class. 37 pages, 1999. To appear
in J. of Pure and Applied Logic.
5. G. Goos, A. Heberle, W. Loewe, and W. Zimmermann. On modular de-
finitions and implementations of programming languages. In Proc. of the
Intern. Workshop on Abstract State Machines (ASM’2000), March 20–24,
2000, Switzerland, Monte Veritá, Ticino, pages 174–208. ETH, Zürich, 2000.
6. Y. Gurevich. Evolving algebra 1993: Lipari guide. In E. Börger, editor,
Specification and Validation Methods, pages 9–93. Oxford University Press,
1995.
7. L. Lamport. A new solution of Dijkstra’s concurrent programming problem.
Communications of ACM, 17(8):453–455, 1974.
8. J. Ostroff. Composition and refinement of discrete real-time systems. ACM
Trans. on Software Engineering and Methodology, 8(1):1–48, 1999.
9. M. Wirsing. Algebraic specification. In J. van Leeuwen, editor, Handbook of
Theoretical Computer Science. Vol. B: Formal Models and Sematics, pages
677–788. Elsevier Science Publishers B.V., 1990.
 Objects + Views = Components?

Martin Odersky

École Polytechnique Fédérale de Lausanne

Abstract. In support of flexible composition, components need to be

adaptable and they need to have first-class plugs. Abstract state ma-
chines and object-oriented programming each satisfy one of these requi-
rements very well but satisfy the other only incompletely. This paper
describes views, a language construct which provides for both require-
ments in a balanced way.

1 Introduction

Increasingly, software is developed from pre-existing components rather than

being built from scratch. Component middleware such as Corba [25], COM [2]
or Java-Beans [6] allow applications to be constructed with the help of binary-
components and foster the development of a component market [26]. Compo-
nents can also take the form of services which are implemented as separate
software systems accessed over the internet [30].
Looking beyond the technical details, components are parts which exist to be
composed. The most important difference with respect to modules in traditional
top-down development is that future compositions of a component are unknown
to the component’s authors. To be useful, a component library therefore needs to
provide high flexibility in component reuse and composition. It has been argued
that this is a difficult task, in particular because the object-oriented languages
used to implement components do not provide the right abstractions [10].
This paper explores language design issues in the area of composition. We ar-
gue that a component is composable only if two requirements are fulfilled: The
component needs to be adaptable and its plugs should be first-class values.
A component is adaptable if clients can enrich its interface after the component
has been constructed and delivered. Such changes consist typically in adding
some data fields or methods which the original component did not yet support.
The methods are to be defined in terms of the component’s existing public
interface. Such additions should not affect a component’s source code. After all,
source code changes by clients are impossible for binary and service components
and pose severe version control problems for components existing in source form.
The need for adaptations arises because at the time when a component is desi-
gned and implemented one often does not know precisely in which contexts the

Y. Gurevich et al. (Eds.): ASM 2000, LNCS 1912, pp. 50–68, 2000.
c Springer-Verlag Berlin Heidelberg 2000

 Objects + Views = Components? 51

component will be used. As a simple example, consider the task of writing a class
for lists. Which methods should this class offer? A fairly small implementation
could be:

class List[a] = {
def head: a = ...
def tail: List[a] = ...
def isEmpty: Boolean = ...
}

Of course there are many more useful functions to be included in the List class.
Functions to select elements inside the lists, to form sublists, to concatenate
lists, to filter elements according to some criterion, or to map functions over all
elements come to mind. Furthermore, one might want to treat lists as represen-
tations of sets, in which case membership, union, intersection and set difference
also should be supported. And one can think of many other useful operations on
lists. The hard question is where to stop. Meyer [20] recommends a “shopping
list approach” to class design where essentially all useful methods one can think
of should be added to an abstraction. In the case of lists, this would probably
lead to many hundreds of methods, and even then clients of the list abstraction
will likely miss a method they require. It seems preferable to provide a more
compact list definition and let clients customize this definition as they require.
The criterion for a complete class design would then be that every useful method
for the class can be defined in terms of the public interface, not that the method
already is defined. The “can”-completeness criterion is much easier to fulfill; for
instance our simple implementation of lists above is already complete.

val SymTab {
type Scope
type Symbol = {
def name: String
def tp: Type
def location: Scope
...
}
def lookup (scope: Scope, name: String): Symbol = ...
def enter (scope: Scope, sym: Symbol): Boolean = ...
def newScope (outer: Scope): Scope = ...
}

Fig. 1. Symbol Table Module

As another example, consider the task of constructing a component for symbol

tables in a compiler. Specifications and algorithms for symbol tables are poten-
tially portable over many compiler implementations. Figure 1 shows a sketch of
52 M. Odersky

a generic implementation. However, it’s not clear what attributes should go into
a symbol, as is illustrated by the ... in the Symbol type. Clearly, symbols will
have a “name” attribute. One might also settle on fields that contain a symbol’s
type and location. But there are also other potential attributes which depend on
the situation in which a symbol table is used. For instance, if our compiler has a
code generator, symbols will need a field in which to store their address. On the
other hand, if the compiler is coupled with a source code browser, we might need
an additional attribute in a symbol which contains the list of all usage points
of that symbol. Again, the precise definition of some aspect of the symbol table
component depends on the contexts the component is used.
In classical structured programming, where data structures and code are sepa-
rated, symbol tables would be adapted by changing their code, filling in the ...
in the Symbol type as required by the application. But source code adaptation
is impossible or at least undesirable in a component setting. Furthermore, the
necessary adaptations to the symbol table module also violate the encapsulation
principle, since fields needed only by the code generator module are placed in the
common Symbol data structure which is accessible to all. Hence, better solutions
for adaptation are called for.
The other requirement for components has to do with how flexibly they can
be composed. One usually regards the run-time embodiments of a component’s
interfaces as its “plugs”. In object-oriented languages plugs are often realized as
objects which have methods which are accessible to clients. A plug is first-class
if it can be treated like a normal value. In particular, one should be able to
pass plugs as parameters to functions, and it should be possible to construct
data structures with plugs as elements. As a very simple example for this kind
of flexibility, consider the task of displaying the information associated with a
symbol. Since there are so many different devices on which this information
could be displayed, it makes sense to split the task in two. Symbols only provide
a method toString, which converts the information associated with the symbol
to a string. In other words, symbols support the interface

type Printable = {
def toString: String
}

Then, a display device could provide a general display service for objects that
“know” how to turn their information into strings:

def print(obj: Printable) = ...

If dev was such a device and sym was a symbol it would then be possible to
write dev.print(sym). Of course, this assumes that symbols are values that can
be passed to the print function. In particular, the type Symbol must be compatible
with the type Printable.
 Objects + Views = Components? 53

not first-class not adaptable

Structured Programming

adaptable
first-class

OOP ASM

OOP + Generics ASM + Type Classes

Objects + Views

Fig. 2. Support for adaptability and first-class plugs.

Since adaptability and first-class plugs are natural requirements for software
components, one might expect that mainstream languages would provide sup-
port for both. Maybe surprisingly, this has not been the case so far. Figure 2
summarizes the state of the art, which will be further discussed in the next sec-
tions. Even though structured programming languages such as Ada or Modula-2
provide support for modules, which can be seen as an embodiment of compo-
nents, the components thus defined are neither adaptable, nor do they have
first-class plugs. Object-oriented programming leads to components with first-
class plugs, but these components are not adaptable. Conversely, the approach
taken by abstract state machines lends itself to the construction of components
which are adaptable, but these components do not have first-class plugs. More
advanced type systems and design patterns can in each case provide some of
the missing functionality. Refining object-oriented programming with bounded
genericity can provide adaptability to some degree, whereas refining functional
programming and ASM’s with a construct such as Haskell’s type classes can
provide some of the benefits of first-class plugs.
This paper presents views as a simple technique to provide both adaptability and
first-class plugs. The technique uses standard object-oriented modeling techni-
ques to model first-class plugs. A view is a syntactic construct which provides
new methods for objects of an existing class, thus providing implementations for
additional object interfaces. Views that extend a component can be distributed
over arbitrary client components.
54 M. Odersky

The discussion in this paper is in the context of statically typed languages. As

notation for our examples, we use Funnel [24], together with some hypothetical
extensions.
The rest of this paper is structured as follows. Section 2 discusses strengths and
shortcomings of the abstract state machine approach to software composition.
Section 3 discusses the same for the object-oriented approach. Section 4 presents
views. Section 5 discusses related work and Section 6 concludes.

2 Abstract State Machines

The abstract state machine approach reverses the usual relationship between
functions and data. Rather than having mutable data structures which are ope-
rated on by immutable functions, we have immutable data, which are operated
on by both mutable and immutable functions. This makes use of the following
analogy between function application and selection.

f (x) ≡ x.f

In other words, a data selector can be seen as a single argument function over
the selected data domain and vice versa. This equivalence makes sense on the
left-hand sides of assignments as well:

f (x) := E ≡ x.f := E

We can thus re-interpret assignment of an expression to a variable field f of a

record x as an update of a mutable function f at the position x. If we apply this
re-interpretation consistently, we arrive at only immutable data, which serve as
index structures for possibly mutable functions. The immutable data structures
can be described as algebras, which leads to a simple formal treatment of state
change. This treatment lends itself to the verification of properties of imperative
programs with pointers, which are very hard to prove otherwise [23].
Abstract state machines add to this the idea of specifying change by a global
set of rules, which collectively specify an execution step. The question whether
execution is described by a global set of rules or by more conventional statements
with control flow constructs is irrelevant to the discussion in this paper and will
henceforth be ignored. In any case, the abstract state machine approach leads
very naturally to components which are adaptable.
For instance, consider the question which functions to provide together with
a List type. Abstract state machine programmers would do the same functional
programmers do – they would define List as an algebraic data type which can then
be accessed by pattern matching. Using Funnel as a notation, the algebraic data
type is expressed as a class with two cases and the pattern match is expressed
by a call to the match method of a list. For instance, here is the definition of a
list together with two representative functions over lists:
 Objects + Views = Components? 55

class List[a] = {
case Nil
case Cons (head: a, tail: List[a])
}

def append (xs: List[a], ys: List[a]) = xs.match {

case Nil: ys
case Cons (x, xs1): List.Cons (x, append (xs1, ys))
}
def map [a,b] (f: a → b, xs: List [a]) = xs.match {
case Nil: xs
case Cons (x, xs1): List.Cons (f (x), map (f, xs1))
}

All operations over lists can be expressed in terms of pattern matching. Hence, a
list operation can be placed in an arbitrary module, and new list operations can
be added in clients of the list abstraction as needed. In a word, lists designed as
algebraic data types are adaptable. This might be too trivial to mention except
for the fact that lists represented as objects which export all their methods are
not adaptable in the same way, as clients cannot add new methods to an existing
list. This point will be further discussed in the next section.
The abstract state machine approach also ensures adaptability for mutable func-
tions. For instance, consider the situation where a symbol table module as given
in Figure 1 is adapted to work with a code generator client. The code generator
needs to maintain with each symbol its address. This could be achieved by de-
fining a mutable function from symbols to their addresses in the code generator
module:

var adr (sym: Symbol): Int := 0

The := 0 part in the declaration above defines the initial value of the mutable
function to be 0 over all of its domain. The code generator module could then
enter a symbol’s address by an indexed assignment such as

adr (sym) := nextLocalAdr

and it could access a symbol’s address by calling adr(sym). No change to the

symbol table module is necessary. Furthermore, since addresses are dealt with
only in the code generator, the definition of adr can be kept local to the code
generator module, so that we have gained encapsulation as well as adaptability.
This is all good news. However, the abstract state machine approach does not
satisfy our second criterion, that components have first-class plugs. For instance,
we still cannot define a general print service with a method print(p: Printable)
which can be applied to symbols. Indeed, neither classical structured program-
ming nor abstract state machines provide a subtype relation which would let us
56 M. Odersky

treat a Symbol as a special instance of the more general Printable abstraction.

This is not an oversight either. To be useful, the Printable abstraction must de-
fine a method such as toString, which turns a Printable into a String. Hence, data
must be an abstraction which includes functions operating on the data. This
view is at odds with the data structuring approaches of functional programming
and abstract state machines, which treat data as domain structures for functions
defined elsewhere. In a sense our two criteria of adaptability and first-class plugs
seem to be at odds with each other. First-class plugs require grouping functions
with data whereas adaptability requires grouping functions separately.
The functional programming language Haskell [11] pioneered the concept of type
classes [29], which can to some degree make up for the lack of first-class plugs
in functional systems. It would be quite possible to extend this construct to
abstract state machine languages as well.
A type class represents a shared property of a set of types – alternatively, it can
be identified with the set of types which satisfy the property. A property states
that a type supports a given set of functions with given signatures. For instance,
here is a declaration of a type class1 .

type class Printable where {

toString :: a → String
}

This states that a type T belongs to the type class Printable if there is a function
toString, which takes a value of type T and yields a value of type String. Types
become members of type classes by explicit instance declarations. Example:

instance Printable Symbol where {

toString (sym) = ...
}

The instance declaration above declares Symbol to be an instance of type class

Printable and provides the implementation of the toString method of Printable.
One can now define functions which are polymorphic over all types which are
instances of a given class. Example:

print :: Printable a ⇒ a → ()
print x = ... toString (x) ...

This defines a function print which takes as parameters values of type a, where a
is an arbitrary instance of type class Printable. The call to toString in print will
invoke the method which is appropriate for the concrete type to which the type
variable a is instantiated. For example, if sym is a Symbol then the type variable
a is instantiated to Symbol and print (sym) invokes the toString method given in
1
We use Haskell syntax except for writing type class where Haskell uses class.
 Objects + Views = Components? 57

Symbol’s instance declaration. The type Printable a ⇒ a → () is called a qualified

type [12] and the (Printable ⇒) prefix is called a context.
Representing component interfaces as instances of type classes, one can hence
write generic functions which work with any component plugs which are in-
stances of a given type class. This yields many of the benefits of first-class plugs,
but some shortcomings remain. First, generic pluggability is achieved in a so-
mewhat roundabout fashion. Second, we still cannot define data structures over
component plugs. Consider for instance a data structure of lists whose elements
are printable components, not necessarily all of the same type. One might try
to express this as a List[Printable], but this would be ill-formed, since Printable
is a type class, not a type. The qualified type Printable a ⇒ List[a] would not
do either, since that type describes a list of objects which all have the same
instance type of Printable. One could go further, expressing heterogeneous data
structures by means of existential types [18], but this comes at the price of even
greater conceptual complexity.

3 Object-Orientation

Instead of starting with an adaptable architecture and trying to emulate first-

class plugs, mainstream programming has instead favored the object-oriented
approach, which has first-class plugs built in, yet needs additional provisions to
achieve some degree of adaptability.
In object-oriented design and programming, one groups data (called instance
variables) and functions operating on the data (called methods) together in an
object. Data and functions collectively are called the fields of an object. There
is a subtyping relation (written <:) between object types, which lets one use an
object with a certain set of fields in a context where only a subset of the fields
is required. This provides all we need for first-class plugs. For instance, to make
symbols printable, it suffices to make the type Symbol a subtype of Printable,
which requires the definition of a toString method in Symbol:

class SymTab = {
class Symbol <: Printable = {
... (fields as before) ...
def toString: String = name ++ ”:” ++ tp.toString
}
}

Then we can write:

val sym = new SymTab.Symbol

...
dev.print (sym)
58 M. Odersky

Are objects also adaptable? One might think so, because with inheritance it
is possible to define a subclass which adds new fields to an existing class. For
instance, it is possible to define in the code generator class CodeGen a new
class Symbol which inherits from the original symbol class as defined in module
SymTab:

class CodeGen = {
class Symbol <: SymTab.Symbol = {
var adr: Int
}
...
}

Instance objects of class CodeGen.Symbol have all the fields of original symbols,
plus the mutable adr field. So it seems we have managed to adapt symbol ta-
ble modules to a code generator client. But closer inspection shows us that this
is not really the case. There are two problems with adaptation through inhe-
ritance. One concerns types, the other object creation. Looking at types first,
we notice that function lookup in our symbol table module still returns objects
of type SymTab.Symbol, not objects of type CodeGen.Symbol. So the added adr
information does not show up in the type.

class Symtab {
...
class Symbol ...
def lookup (scope: Scope, name: String): Symbol
...
}

To access the adr field of a code generator symbol stored in the symbol table,
we need a dynamic type cast. In other words, we can gain adaptability only
by subverting the type system, and running the risk of dynamic type errors.
A safer alternative makes use of type parameterization. For instance, here is a
re-formulation of the symbol table class using bounded genericity:

class Symbol <: Printable = {

... (fields as before) ...
}
class SymTab [s <: Symbol] = {
def lookup (scope: Scope, name: String): s = ...
def enter (scope: Scope, sym: s): Boolean = ...
def newScope (outer: Scope): Scope = ...
}

The class for symbol tables is now parameterized by a type variable s which
represents the actual type of symbols stored in the table. The only requirement
 Objects + Views = Components? 59

on instantiations of s is that they are subtypes of Symbol. Some gluing is needed

at the top-level of an application to instantiate the class to the right symbol
type:

val symTab = new SymTab[CodeGen.Symbol]

The other problem with achieving adaptation with inheritance has to do with
object creation. Taking our symbol table example, we note that we have not
looked yet at code which creates symbols to be stored in the table. Let’s say
symbols are created during tree attribution, in class Attr:

class Attr = {
...
new Symbol(name, tp)
...
}

This is not yet precise, because we have not yet specified which symbol should
be created during tree attribution. If it is the original SymTab.Symbol that is
created, we are missing an adr if we want to couple tree attribution with code
generation. On the other hand, if it is a CodeGen.Symbol which is created, we
have hardwired the relationship between tree attribution and code generation.
Now it would be the tree attribution part which lacks adaptability, because it
would not be able to deal with different kind of symbols, for instance those
required by a source code browser. To overcome this problem, we can make use
of the Factory design pattern [7]. Factory objects would have an interface of the
form:
type Factory [t] = {
def make (...): t
}

There would have to be an appropriate factory for every subtype of symbol which
is being defined. E.g. in class CodeGen:

class CodeGen = {
...
val SymFactory: Factory[Symbol] = {
def make(name: String, tp: Type): Symbol =
new CodeGen.Symbol(name, tp)
}
...
}

The Attr class will be parameterized with the factory which is to be used for
symbol creation. We also have to parameterize Attr with the actual type of
symbols to be created. This yields:
60 M. Odersky

class Attr [s <: Symbol] (symFactory: Factory [s]) = {

...
symFactory.make (name, tp)
...
}

Even more gluing is needed at top-level, to obtain the correct instantiation of

the tree attribution module:

val symTab = new SymTab [CodeGen.Symbol]

val attr = new Attr [CodeGen.Symbol] (CodeGen.SymFactory)

The preceding discussion has shown that it is possible to obtain a certain de-
gree of adaptability in object-oriented designs, but this requires planning. If we
had not foreseen the possible later extension of the symbol type, we would not
have provided the necessary parameters for subtypes and factories. The hooks
necessary for enabling future extensions can clutter programs considerably, in
particular if we envisage multiple coexisting extensions. For example, the current
compiler framework would enable a single extension of Symbol, say for a code
generator or a browser, but it would not yet enable multiple coexisting exten-
sions, say where a symbol has both an adr field for a code generator and a uses
field for a browser. Multiple extensions can be supported by stacking extensions
on top of each other but this requires even more complex protocols.

4 Views

In the previous two sections, we have discussed two program structuring me-
thodologies which have complementary strengths. Each methodology supports
one of our two criteria of adaptability and first-class plugs, but supports the
other only incompletely and at the price of complex language features and de-
sign patterns. In this section, we propose a more symmetric combination which
can address both criteria equally well. This combination arises by translating
concepts first developed for type classes into an object-oriented setting.
When comparing type classes in functional programming and object-oriented
type systems, we find some strong analogies:

type class ≈ type

instance relation ≈ subtype relation
instance declaration ≈ subtyping clause

A type class corresponds to a type in the object-oriented setting. For instance,

instead of having the type class Printable, we would have the type Printable,
which has all “printable” types as subtypes. Consequently the instance relation
between types and type classes becomes the subtype relation between types.
 Objects + Views = Components? 61

Finally, an instance declaration in Haskell corresponds to a subtyping clause of

a class declaration. The analogy highlights an important difference between the
approach of Haskell and the one of object-oriented languages: in the latter, a
subtyping clause is textually part of the extending class whereas in Haskell, an
instance declaration can be given anywhere. Haskell’s approach is adaptable,
in that instance declarations for a type can be given after the type is defined.
The object-oriented approach is not, since adding new subtype edges requires
changes in the source code of the subtype.
Languages with structural subtyping do not need explicit subtyping clauses in
class declarations. Instead, the subtype relation is determined only by the struc-
ture of the classes: Classes with more fields become subclasses of classes with
fewer fields, as long as the types of common fields are the same or stand in turn
in a subtyping relation. This gives some more flexibility in the subclass relation.
But it still does not solve the adaptation problem since it does not provide a
way to augment a given class with new fields.
Seeing these disparities it seems promising to provide a construct on the object-
oriented side which matches more closely the distributed nature of instance decla-
rations. This construct is called a view. A view adds new fields to values of an
existing type. For instance, the following view declaration would add the method
toString to all values of type Symbol.

view (this: Symbol): Printable {

def toString: String = this.name.toString ++ ”:” ++ this.tp.toString
}

A view takes the form of an unnamed function, which takes a single parameter
of some type A and yields an object of type B. The view establishes a subtype
relationship between A and B by giving declarations of all fields of B which are
not yet present in A. All fields of A are inherited; i.e. they form implicitly part of
the resulting object. It is only possible to add new fields to a class with a view,
not to change existing fields.

Subtype Clauses

Subtype clauses can be regarded as syntactic sugar for view declarations with
empty bodies. For instance, our previous class declaration

class Symbol <: Printable = {

...
def toString: String = name ++ ”:” ++ tp.toString
}

would be equivalent to
62 M. Odersky

class Symbol = {
...
def toString: String = name ++ ”:” ++ tp.toString
}
view (this: Symbol): Printable = {}

Unlike a subtyping clause, a view declaration can be given anywhere in the

program; it need not appear textually adjacent to the declaration of the view’s
parameter class.

Variables in Views

Besides methods, views can also define mutable variables. Here is an example:

type Adr = { var adr: Int }

view (sym: SymTab.Symbol): Adr = {
var adr: Int
}

These declarations, when placed in the code generator module of a compiler,

would be equivalent to the mutable function which we have used in the abstract
state machine structuring:

var adr(sym: Symbol): Int

The only difference concerns selection syntax – we use the object-oriented version
sym.adr instead of the previous adr(sym).

Encapsulation

More interestingly, we can also have information encapsulated in views. Consider

for instance an encapsulated abstraction for address fields, which lets only even
addresses be stored. This is accomplished as follows.

type Adr = {
def setAdr (x: Int): ()
def getAdr: Int
}
view (sym: SymTab.Symbol): Adr = {
var adr: Int
def setAdr (x: Int) =
if (x % 2 == 0) adr := x
else error (”bad address”)
def getAdr = adr
}
 Objects + Views = Components? 63

This view defines a variable adr along with setter and getter methods. Since the
variable is not part of the view’s target type Adr, it is not accessible to clients
of the view.

Conditional Views

Sometimes, a parameterized type can implement a view only if some constraints

on the type parameters are satisfied. Consider for instance structural equality
as implemented by the interface

type Eq [a] = {
def equals (other: a): Boolean
}

A type T supports structural equality testing if it implements the unary equals

method which takes other values of type T as parameters. In other words, T sup-
ports structural equality testing if T <: Eq[T]. Now, consider the parameterized
type of lists, List[a]. Should lists support structural equality testing? Clearly, we
can compare lists structurally iff we can compare the list elements. This can be
expressed by the following, conditional view declaration.

view [a <: Eq [a]] (this: List[a]): Eq [List [a]] = {

def equals (that: List [a]) = {
this.isEmpty && that.isEmpty ||
!this.isEmpty && !that.nonEmpty &&
this.head.equals (that.head) && this.tail.equals (that.tail)
}
}

The [a <: Eq [a]] construct introduces a type parameter a for the view which is
bounded by Eq[a]. The type variable can hence be instantiated to any type T
which is a subtype of the bound Eq[T]. Note that this gives us a parameterized
type which has or lacks method equals, depending on whether or not the type of
the listed elements has the same method.

Semantics

The rest of this section discusses detailed design and implementation choices for
views. The most fundamental design choice concerns name resolution of fields
defined by views. Such resolution can be static, based on compile-time types or
it can be dynamic, based on run-time types. To see the difference, consider a
variant of our symbol table code. Assume that we want to further refine the type
of symbols into variable symbols and function symbols:
64 M. Odersky

type Symbol = {
... (fields as before) ...
}
type VarSymbol <: Symbol = { ... }
type FunSymbol <: Symbol = { ... }

In the code generator, one might need a function loadAdr which returns the
instruction for loading a given symbol’s address. The result of this function
will depend on the kind of symbol encountered. It is therefore reasonable to
implement loadAdr as a method of type Symbol, to be overridden in subtypes.
But as before, we want to keep the source code in Symtab unchanged. Again as
before, we resort to views for implementing the new functionality in the code
generator client. Instead of a single view we now need three views, one for each
kind of symbol:

type Loadable = {
def load: Code
}
view (sym: Symbol): Loadable = {
def loadAdr: Code = error (”can’t load”)
}
view (sym: VarSymbol): Loadable = {
def loadAdr: Code = LLA (sym.adr)
}
view (sym: FunSymbol): Loadable = {
def loadAdr: Code = LCA (sym.adr)
}

Can this work? Assume that the code generator calls the loadAdr method of
a symbol it retrieved (via lookup) from the symbol table. The static return
type of lookup is Symbol. So if we perform name resolution based on the static
type, the loadAdr will always result in error(”can’t load”). In an object-oriented
language such behavior would be counter-intuitive. We therefore pick dynamic
name resolution, where the loadAdr method is chosen according to the run-time
type of the symbol returned from lookup. If the returned symbol represents a
variable, we invoke the loadAdr method of the view for VarSymbol. If the symbol
represents a function, we invoke the loadAdr method of the view for FunSymbol
instead. Only if the returned symbol is neither a variable nor a function, the
loadAdr method of the view for the base class Symbol is invoked.
In the last example the call to loadAdr was resolved as if loadAdr was a regular
method in Symbol, which was overridden in VarSymbol and FunSymbol. We can
indeed often view a program with views as equivalent to a program with multiple
inheritance where all views are inlined in their base classes. View methods then
 Objects + Views = Components? 65

become regular methods and target types of views become elements of subtyping
clauses. The correspondence with multiple inheritance programs serves as a guide
in the definition of the semantics of views and the context-dependent syntax rules
governing them.
The context-dependent syntax rules are based on the concept of the view graph.
The view graph V G of a project has as nodes the set of types defined in all
source files of the project and as edges the set of views between types. Subtyping
clauses are also interpreted as view edges. Furthermore, we regard the methods
defined in a class as defined in an additional view which goes from a new node
representing a class implementation to the proper class type itself. A field which
occurs textually in a record type T is said to be defined by T . A field which
occurs textually in a view V is said to be implemented by V . The subtyping
ordering ≤ on classes induces an ordering on views as follows: A view from S to
T precedes a view from S 0 to T 0 if S ≤ S 0 . We then require the following:

– The view graph must be acyclic.

– A view from S to T may only implement fields which are not already defined
in S.
– When selecting a field f from an expression e of type T , there must be
exactly one type reachable from T in V G which defines f .
– Let I be the node representing a class implementation and let T be a type
reachable from I in V G. Then every field f in T must have a best implemen-
tation for I. This means: First, there must be a view on a path between I
and T which implements f . Second, among all the views on paths between I
and T which implement f , there must be one view which precedes all others
according to the induced subtyping ordering on views. This requirement cor-
responds to the usual requirement for multiple inheritance that inheritance
of fields must be unambiguous.

It is worth noting that some of these rules are global in that they require know-
ledge of the complete view graph. To check such global requirements in a system
with separate compilation, one could delay some type checking until link time.
As an alternative one could also maintain a project-wide repository in which
all current view declarations are maintained. Consistency of the view graph can
then be checked incrementally, as views are added and deleted.

5 Related Work

The shortcomings of the object-oriented approach in the area of adaptability

have been known for some time. They have given rise to a varied body of research.
The work on subject-oriented programming [9] identified the need for multiple,
independent roles played by a single object. Aspect-oriented programming [17],
hyperspaces [27], and adaptive programming [21] all address this requirement
with systems which act as source to source transformations. The idea is in each
66 M. Odersky

case that different aspects of a program are defined in separate source files which
are then merged using a “weaver” tool. Binary component adaptation [16] provi-
des mechanisms to perform similar changes on Java class files. Both weavers and
binary component adaptation are possible approaches for implementing views.
Compared to these approaches, views are more restrictive in that they support
only strict extensions of existing classes. One advantage of this restriction is that
strict extensions commute, so that the semantics of the weaver tool becomes a
non-issue.
Multi-methods [1,4] are another approach to extend the limitations of pure
object-oriented programming. Here, a method is no longer attached to a single
class, and method selection takes the dynamic type of all method arguments
into account. Like views, multi-methods can be defined independently of their
argument classes. But multi-methods are more general than views in that they
abandon the concept of a distinguished receiver argument and in that they re-
quire multiple dispatch. Open objects [19] are an idiom of multi-methods which
is comparable to views.
Like views, Haskell’s type classes allow a distributed definition of functionality
over data types. Qualification of type variables in a qualified type corresponds
to F-bounded qualification of views [3]. A number of different designs for type
classes have been developed [8,22,12,13,14]. Our view proposal is most closely
related to parametric type classes [5], in that the bound of a type variable a may
have additional parameters other than a itself. Type classes are usually studied
in systems without subtyping. Where subtyping is added [15], name resolution
is based on static types, whereas views are based on dynamic types.
A language construct called “view” has also been proposed for functional pro-
gramming [28]. These views serve as alternative decompositions of sum types
such as Haskell’s algebraic data types. By contrast, the views presented here
provide alternative accesses to product types such as records or objects. Both
designs have in common that a view can be defined in program parts other than
the one which defined its base type.

6 Conclusion
Do objects plus views equal components? This paper has argued that truly reus-
able component libraries require the ability to adapt components by extending
them with new functionality, and to compose components via first-class plugs.
Object-oriented programming and abstract state machines each provide one of
these two keys to successful composition. The view construct aims at providing
both of these keys.
Work is currently underway on a complete design and implementation of views
in the context of Funnel, an implementation language for functional nets. Only
future experience will tell whether the two keys are sufficient for the construction
of truly reusable component libraries. That’s why the title of this paper still ends
in a question mark.
 Objects + Views = Components? 67

References
1. D.G. Bobrow, K. Kahn, and G. et al. Kiczales. CommonLoops — merging Lisp
and object-oriented programming. In Proc. OOPSLA’86, pages 17–29, 1986.
2. Don Box. Essential COM. Addison Wesley, 1998.
3. Peter Canning, William Cook, Walter Hill, Walter Olthoff, and John C. Mitchell.
F-bounded polymorphism for object-oriented programming. In Functional Pro-
gramming Languages and Computer Architecture, pages 273–280, September 1989.
4. Craig Chambers. Object-oriented multi-methods in Cecil. In Proc. ECOOP, 1992.
5. Kung Chen, Paul Hudak, and Martin Odersky. Parametric type classes. In Proc.
ACM Conf. on Lisp and Functional Programming, pages 170–181, June 1992.
6. Robert Englander. Developing Java Beans. O’Reilly & Associates, 1997.
7. Erich Gamma, Richard Helm, Ralph Johnson, and John Vlissides. Design Patterns
: Elements of Reusable Object-Oriented Software. Addison-Wesley, 1994.
8. Cordelia Hall, Kevin Hammond, Simon Peyton Jones, and Philip Wadler. Type
classes in Haskell. In Proc. 5th European Symposium on Programming, pages 241–
256, 1994. Springer LNCS 788.
9. William Harrison and Harold Ossher. Subject-oriented programming (a critique
of pure objects). In Proc. OOPSLA, pages 411–428, September 1993.
10. Urs Hölzle. Integrating independently-developed components in object-oriented
languages. In Proc. ECOOP, volume 707 of Springer Verlag, Lecture Notes in
Computer Science, pages 36–56, 1993.
11. Paul Hudak, Simon Peyton Jones, and Philip Wadler. Report on the programming
language Haskell: a non-strict, purely functional language, version 1.2. SIGPLAN
Notices, 27(5), May 1992. more recent versions at www.haskell.org/definition.
12. Mark P. Jones. A theory of qualified types. In Proc. 4th European Symposium on
Programming, pages 287–306, February 1992. Springer LNCS 582.
13. Mark P. Jones. A system of constructor classes: Overloading and implicit higher-
order polymorphism. In Proc. ACM Conf. on Functional Programming Languages
and Computer Architecture, pages 52–61, June 1993.
14. Mark P. Jones. Type classes with functional dependencies. In Proc. European Sym-
posium on Programming, number 1782 in LNCS, pages 230–244. Springer Verlag,
March 2000.
15. Stefan Kaes. Type inference in the presence of overloading, subtyping, and re-
cursive types. In Proc. ACM Conf. on Lisp and Functional Programming, pages
193–204, June 1992.
16. Ralph Keller and Urs Hölzle. Binary component adaptation. In Proc. European
Conference on Object Oriented Programming, Lecture Notes in Computer Science.
Springer Verlag, 1998.
17. Gregor Kiczales, John Lamping, Anurag Mendhekar, Chris Maeda, Cristina Videira
Lopes, Jean-Marc Loingtier, and John Irwin. Aspect-oriented programming. In
Proc. of the European Conference on Object-Oriented Programming, Finland, June
1997.
18. Konstantin Läufer. Type classes with existential types. Journal of Functional
Programming, May 1996.
19. Todd Millstein and Craig Chambers. Modular statically typed multimethods. In
Proc. ECOOP, number 1628 in LNCS, pages 279–303. Springer Verlag, June 1999.
20. Betrand Meyer. Object-Oriented Software Construction. Prentice-Hall, 1988.
21. Mira Mezini, Linda Seiter, and Karl Lieberherr. Component integration with plug-
gable composite adapters. In Software Architectures and Component Technology.
Kluwer, 2000.
68 M. Odersky

22. Tobias Nipkow and Christian Prehofer. Type checking type classes. In Proc. 20th
ACM Symposium on Principles of Programming Languages, pages 409–418, 1993.
23. Martion Odersky. Programming with variable functions. In Proc. International
Conference on Functional Programming, September 1998.
24. Martin Odersky. Functional nets. In Proc. European Symposium on Programming,
number 1782 in LNCS, pages 1–25. Springer Verlag, March 2000.
25. Alan Pope. The Corba Reference Guide: Understanding the Common Object Re-
quest Broker Architecture. Addison Wesley, 1998.
26. Clemes Szyperski. Component Software : Beyond Object-Oriented Programming.
Addison-Wesley, 1998.
27. P. Tarr, H. Ossher, W. Harrison, and S. ”Sutton Jr.”. N degrees of separation:
Multi-dimensional separation of concerns. In Proc. ICSE, pages 107–119, 1999.
28. Philip Wadler. Views: a way for pattern matching to cohabit with data abstrac-
tion. In 14’th ACM Symposium on Principles of Programming Languages, Munich,
Germany, January 1987.
29. Philip Wadler and Stephen Blott. How to make ad-hoc polymorphism less ad-hoc.
In Proc. 16th ACM Symposium on Principles of Programming Languages, pages
60–76, January 1989.
30. Gio Wiederholt, Peter Wegner, and Stefano Ceri. Towards megaprogramming.
Communications of the ACM, 1992. November.
 Xasm– An Extensible, Component-Based
Abstract State Machines Language

Matthias Anlauff

GMD FIRST, Kekuléstr. 7

D-12489 Berlin, Germany
[email protected]

Abstract. The Abstract State Machine (ASM) [16] approach has al-
ready proven to be suitable for large-scale specifications of realistic sy-
stems [18,9,22,34]. Due to the fact that the ASM approach defines a
notion of executing specifications, it provides a perfect basis for a langu-
age, which can be used as a specification language as well as a high-level
programming language. However, in order to upgrade to a realistic pro-
gramming language, such a language must – besides other features – add
a modularization concept to the core ASM constructs in order to pro-
vide the possibility to structure large-scale ASM-formalizations and to
flexibly define reusable specification units. In this paper, the language
Xasm, which stands for Extensible ASM, is presented. Xasm realizes a
component-based modularization concept based on the notion of exter-
nal functions as defined in ASMs. This paper also briefly describes the
support environment of Xasm consisting of the Xasm-compiler transla-
ting Xasm programs to C source code, and the graphical debugging and
animation tool.

1 Introduction

The Abstract State Machine approach has been and is successfully used to model
a large number of case studies including industry-relevant ones. The simplicity of
the basic data and execution model of ASMs makes them perfectly suitable as the
basis for a language that on the one hand can be used as specification language
and on the other hand as a high-level programming language. In this paper,
the Xasm (Extensible ASM )1 language is presented which aims at providing
support for using ASMs as a programming language for producing efficient and
reusable programs. There exists a number of other ASM implementations which
all implement most of the ASM constructs as defined in the Lipari-Guide [16].
While the realization of the ASM constructs can be seen as the core functionality
which must be present in each ASM support system, the difference of an ASM
system compared to all others can be characterized by

1
formerly known as “Aslan”; the name has been changed because of a name conflict
with another tool.

Y. Gurevich et al. (Eds.): ASM 2000, LNCS 1912, pp. 69–90, 2000.
c Springer-Verlag Berlin Heidelberg 2000

70 M. Anlauff

– its efficiency,
– the functionality of its support environment,
– its rule abstraction concept, and
– its interoperability with other languages and systems.

For example, all ASM implementations – including Xasm– define some macro
structures on top of the core ASM language in order to provide some kind of rule
abstraction concept. These additional features are indispensable for managing
large formalizations. In the ASM-Workbench [13], for instance, the a special
“Rule” construct is introduced being used to assemble ASM specifications from
smaller pieces.
Concerning these features, Xasm combines the advantages of using a formally
defined method with the features of a full-scale, component-based programming
language and its support environment.
The paper is organized as follows: In Section 2 an overview of Xasm is given.
Section 3 introduces the component-based module concept of Xasm, in Section 4
the external language interface of Xasm is described. In Section 6 the possibility
to specify the syntax of input languages using context-free grammar definitions
is presented, which is followed by the description of non-standard language con-
structs defined in Xasm in Section 5. Section 7 sketches the support environment
of Xasm; Section 8 contains concluding remarks and points out future work.

2 Overview of Xasm
Xasm is an implementation of sequential ASMs focusing on the generation of
efficient executable programs simulating the run of the specified ASM. In general,
the main design goals of Xasm can be given as follows:
– full support of the ASM language as defined in the Lipari-Guide;
– efficient execution of generated executables;
– comfortable animation and debugging of ASM specifications;
– component-based library concept for managing large-scale specifications;
– external language interface for integrating ASM specifications in other sy-
stems.
The scenario of building ASM-based programs using Xasm is depicted in Fi-
gure 1. Xasm source files are translated into C source by the Xasm-compiler.
Additionally, the user can integrate C-sources and -libraries using the external
language interface. As described below, Xasm introduces a notion of compo-
nents being stored in a special repository. During the translation process, the
Xasm-compiler retrieves registry information from the component in order to in-
tegrate pre-compiled Xasm-components in the current build process. The result
of such a build process is a binary being either an executable or a new element of
the component library. In either case, the binary contains the ASM algorithms
specified in the Xasm source files.
Basically, Xasm-programs are structured using “asm . . . endasm” con-
structs each of which containing a list of local function and universe declarations
 Xasm– Extensible, Component-Based ASMs 71

Xa
sm
-S
ou Xasm-
rc Compiler
e

Registry
Information

ge -S
C
ne ou
ra rc
te e
d
C-
So

Xasm Component
ur

Library
ce

C-Compiler&
Linker
C-libs&
obj.files

Executable

Fig. 1. Building Xasm applications

and a list of ASM rules representing a certain part of the overall specification.
In general, the structure of an Xasm-asm is shown in Figure 2. The meta infor-
mation part contains information concerning the role of the asm as a reusable
component; this part is described in more detail below.
As defined in the Lipari-Guide, types are not part of the core ASM language.
However, because typing has been proven to be very useful to avoid many kinds
of errors, in Xasm types can be supplied to the declaration of a function and
are used to detect static semantics inconsistencies of the formalization.

3 The Basic Structure of Xasm Programs: The Xasm

Component Model
In order to provide the full comfort of a modern programming language, pure
ASMs lack a concept of modularization which is indispensable for structuring
large-scale formalizations. Macros, which are normally used in the ASM litera-
ture to structure large ASM formalizations, only provide limited functionality
with respect to the advantages one expects from a module concept. However,
macros are a good means for “ASM-programming-in-the-small”, but they fail
72 M. Anlauff

asm A(a1 : T1 , . . . , an : Tn ) → a0 : T0
hasm meta informationi
is
huniverse, function, and subasm declarationsi
hinitialization rulesi
hasm rulesi
endasm

Fig. 2. The general structure of an Xasm-asm

to provide a basis for writing ASM formalizations that can be re-used in other
formalizations.
Therefore, Xasm uses a more powerful modularization concept which is based
on the notion of a component as it is used in component-based systems (e.g. [28,
31]).

3.1 The Use Modes of “asm”-Constructs

As mentioned above, an Xasm formalization is structured using “asm. . . end-

asm” units. In order to explain the relationships that can exist between these
units, we will first introduce the possible “use modes”: An asm can be accessed
by other asms in either of the following two ways:
If an asm A uses B as sub-asm, it means that B – possibly together
with arguments, if the arity of B > 0 – is used as a rule in the body of A.
If this rule fires, the rules of asm
B fire, which may result in upda-
ting locations of functions decla-
red in A. The sub-asm-use relation A A . . . A

between asms may contain cycles;

lazy evaluation techniques are used
to avoid an infinite number of ru- B B . . . B

les. The call as subasm is illustra-

asm step call as subasm
ted in the figure. The sub-asm B
and its parent asm A step simultaneously; formally they can be seen as one
single ASM.
Asm A uses B as a function, if B is defined as external function in A. In
this case, B – possibly together with arguments, if the arity of B > 0 – is used as a
 Xasm– Extensible, Component-Based ASMs 73

term in the body of A. Recursion

is allowed, so that the function-
use relation between asms may
A A . . .
contain cycles. The call as fun-
ction is illustrated in the figure.
During the run of the function-
asm B, its parent A doesn’t make B B . . . B

any step; from A’s point of view asm step call as function
B’s run happens in zero time. As
depicted in the figure, B behaves like a “normal” asm, the iterations shown here
are caused by the steps of the B-asm itself.
In each of the above cases, we call A the parent-asm of B, if A uses B as
sub-asm or as function. In any case, the asm must be declared in the parent
asm. As part of its meta information, an asm can be marked as a function or
as a sub-asm, so that it can only be used by other asms in the specified way.
For example, if B and C are asms defined as follows

asm B(x : Int) → Int asm C(x : Int)

used as function used as subasm
is is
... ...
endasm endasm

then B can only be used as function and C as sub-asm in other asms. This
is reflected by corresponding declarations of B and C:

asm A
is
subasm C(x : Int)
external function B(x : Int) → Int
...
endasm

Example

A typical situation for using sub-asms is given, when the specification can be
split up naturally into several sub-specifications each of which modeling a certain
aspect of the overall specification.
74 M. Anlauff

asm Robot is asm Robot is standing

universe M odeV alue = {standing, moving} used as subasm
subasms Robot is standing, is
Robot is moving ...
function mode → M odeV alue mode := moving
... ...
if mode = standing then endasm
Robot is standing asm Robot is moving
elseif mode = moving then used as subasm
Robot is moving is
endif ...
... mode := standing
endasm ...
endasm

In this case, the specification introduces the notion of a mode which can
be used to structure the formalization. In the example, it is assumed, that the
sub-asms update the value of the mode function to some value.

3.2 Xasm Components

The declaration of sub-asms and external functions that refer to other asms
in the specification requires that the existence and functionality of these asms
is known at specification time. This is comparable to a static module concept
and is useful for defining sub-parts of one specific formalization. In order to be
“component-based” like announced above, the module concept must be enriched
with some other features allowing a more flexible and comfortable definition of
reusable units.
For example, consider the following asm that may be used in the context
of a programming language semantics specification. It checks, whether a given
variable is defined in the current block, or in one of the parent blocks. The
information whether a variable is defined in a certain block is stored in the ASM
function DeclT able; the block structure is stored in the function P arentBlock
mapping blocks to its corresponding parent blocks:2
2
the meaning of the return rule is explained later
 Xasm– Extensible, Component-Based ASMs 75

asm check blockvar(block : Str, var : Str) → Bool

used as function
accesses functions DeclT able(Block : Str, var : Str) → ,
P arentBlock(Block : Str) → Str
is
function current block ← block
if DeclT able(current block, var) 6= undef then
return true
else
current block := P arentBlock(current block);
if current block = undef then
return false
endif
endif
endasm

Note, that these function works correctly without recursively calling itself; it
iterates until no update changes the internal state of the asm.
The “accesses” construct is used to specify the functions the asm expects
from its parent asm. Now, with this additional meta information, the asm can
be regarded as a component, because its provides information necessary to be
processed as stand-alone unit. The asm can be separately compiled and put into
the Xasm component library; other formalization can reuse it provided that they
declare the required functions.
Besides the “accesses” construct, which allows to read the locations of the
corresponding functions provided by the parent asm, the Xasm “updates”
construct marks the corresponding function as readable and writable for the
sub-asm or ASM function. In the previous example, the mode function must be
marked as “updated” the two sub-asms, because it is updated in the body of
each of them:

asm Robot is standing asm Robot is moving

used as subasm in Robot used as subasm in Robot
updates function updates function
mode → M odeV alue mode → M odeV alue
is is
... ...
mode := moving mode := standing
... ...
endasm endasm

Like the accessed functions, the updated functions must be declared in the
parent asm. In order to avoid repetitions in the source code, the notation
“used as subasm in A” can be used as an abbreviation of accessing all fun-
ctions and universes declared in A except those that are explicitly marked as
“updated” by the sub-asm (analogously for asms that are used as functions).
76 M. Anlauff

Besides functions, sub-asms can also be contained in the “accesses”-list of an

asm-component. The accessed sub-asms are used in the rule section of the asm
as if they have been declared locally.

3.3 Gluing of Xasm Components

In order to provide a high degree of flexibility in interconnecting Xasm-com-
ponents, it is possible to define local derived functions using so-called “with”
definitions. For example, a asm A wants to use the check blockvar as introduced
above, but A doesn’t declare a function named “DeclT able” as it is required by
the check blockvar asm. However, the DeclT able must be somehow expressible
using existing functions in A. The “with”-statement can be used to provide the
called asm with the necessary function declaration, as illustrated in the following
example:

asm A is
...
function currentmodule → Str
function SymT able(mod : Str, block : Str, v : Str) → Int
external function check blockvar(b : String, v : Str) → Bool
with DeclT able(b : Str, v : Str) ==
SymT able(currentmodule, b, v)
...
endasm

In a similar way, accessed sub-asms can be specified in the context of a “with”

statement.
Using Xasm components together with this kind of gluing mechanism provi-
des a powerful means to structure large specifications using smaller and reusable
units.

3.4 Informal Semantics of “Accesses” and “Updates” Declarations

The semantics of the “accesses” and “updates” declaration in asms depends on
whether the asm is used as a sub-asm or as a function. In the following, the
semantics of these constructs in each of these cases is explained briefly.

Accessed and updated functions in Sub-Asms. If an asm B is used in A

as a sub-asm, the accessed and updated functions in B are directly linked to the
corresponding functions in A. That means that if B updates a function declared
in A, the update is visible for both A and B in the subsequent step. This can
be done in this way, because A and B step simultaneously; the rules of B are
regarded as part of the rules of A. Similarly, if B accesses a sub-asm C, then
firing C in B has the same effect than firing C in A.
 Xasm– Extensible, Component-Based ASMs 77

Accessed and updated functions in External Functions. The more com-

plicated case is given when an asm B being used as an external function in A
accesses and updates functions declared in A. Due to the fact that A doesn’t
make any step during the run of B, rules in B updating dynamic functions decla-
red in A are actually not performed from B’s point of view. Therefore, in Xasm
the semantics of updated functions in external asm functions is defined in a way,
that these kind of unintuitive behavior is avoided:

– For each function f being marked as “updated” a local function with the
same name is (internally) declared in B;
– this local function is initialized with the values of the original function in A;
– during the run of B, the functions marked as “updated” can be accessed like
any other local function in B;
– on termination of B the updated locations of these function are propagated
to the original function declared in the parent asm A.
This ensures, that all updates of an “updates” function are accessible in B, and
that only the last updates are forwarded to the parent A. In general, the updates
of functions declared in A and updated in B are treated as being part of the
update set of A’s current step. As a consequence, multiple invocations of B in
the same step of A do not influence each other.
Consider the following – somewhat artificial – example:

asm B → Int
used as function
updates function f (x : Int) → Int
asm A is is
function f (x : Int) → Int function i ← 0
function v → Int function r
external function B → Int if i < 3 then
... r := f (0)
v := B f (0) := i
... i := i + 1
endasm else
return r
endif
endasm

In each step of the run of asm B the value of the updated function f (0) is
updated with a new value. The semantics of the “updates” declaration in Xasm
ensures that all updates of f are accessible in B and that only the last update of
f in B is propagated to the parent asm A. In this example, the update f (0) := 2
78 M. Anlauff

is propagated to A, while all other updates occurring in the “internal” steps in

B 3 have only local effects in B.4

3.5 Access Modes of External Functions

In order to allow different kinds of accesses to external entities, external functions
can be declared either as “monitored” or as “output” functions. In the first case,
the external function can be read, but not written (e.g user input), in the second
case, the external function can be written, but not read (e.g. output channels
like stdout and stderr).
As a restriction, an external function can be either a monitored or an output
function, not both. If a function would have both modes, reasoning about the
values of that function would require special case distinctions: If a location of
such a function is updated in one step of the ASM by means of an update rule, it
cannot be guaranteed that the location has the updated value in the next step,
because the environment might have changed it in the meantime.
In the following example, the 0-ary function error is defined as an external
function with “output” access mode; it is used in the parent asm for displaying
an error message on “stderr” and for setting an ok f lag to false.

asm A is
relation checkok
external [output] function error → Str
with ok f lag == checkok
...
error := ”...”
...
endasm

asm error → msg : Str

used as function
updates relation ok f lag
is
use stdio
stderr := msg
ok f lag := f alse
endasm

The value that is used for updating the external function can be accessed
using the named result parameter msg. The use construct includes pre-defined
header files containing function declaration that are in this case used to declare
the external functions stdout, stdin, and stderr.
3
in step 1: f (0) := 0; in step 2: f (0) := 1
4
that means that in the second step the update r := 0 is performed in B, in the third
step r := 1
 Xasm– Extensible, Component-Based ASMs 79

If no access mode is specified in the declaration of an external function, the

mode “monitored” is assumed.

3.6 Scoping
In the context of use-relations between asms, Xasm distinguishes between the
parent-asm and the caller-asm:

The parent-asm of an asm B is the asm where B is declared (either as

sub-asm or as external function), while the caller-asm is the asm where
the call actually takes place.

In the easiest case, parent and caller are the same, as in the above example: asm
error is declared in and called by asm A. In the following example, this is not
the case:
asm A is asm error → msg : Str
relation checkok used as function
external [output] function error → Str updates
with ok f lag == checkok relation ok f lag
subasm B is
... use stdio
B stderr := msg
... ok f lag := f alse
endasm endasm
asm B
updates function error → Str
is
error := ”...”
endasm

Here, A is the parent-asm and B the caller-asm of asm error. As a conse-

quence, the exported relation ok f lag is taken from the parent-asm, rather than
from the caller-asm. That means, that the update of error in B has the con-
sequence that the checkok relation is updated in A. This distinction has been
made, in order to completely abstract from the actual realization of exported
and accessed functions and sub-asms. In this case, B doesn’t need to “know”
that error is an external function.
The scoping rule for Xasm-asms is similar to static scoping in programming
languages and can be summarized as follows:
Exported and accessed functions and sub-asms of an asm B are always taken
from the asm where B has been declared either as external function or as
sub-asm.
80 M. Anlauff

3.7 Returning Values From External Functions

As already frequently used in the examples in this document, the return con-
struct is used to specify the return value of a monitored external function. In
terms of ASMs, “return ” is realized as follows: In each asm B that is used as
monitored function, a 0-ary dynamic function B result is declared and initiali-
zed with no result, a specific element of the superuniverse. Let R be the rule of
B as defined by the user, then the internally used rule representing the body of
B is given by the following conditional:

if B result = no result then

R
endif
In other words, updating B result with value different from the special
no result element directly forces the asm to terminate. The notation “return
t” in an asm B is then simply an abbreviation for the update “B result := t”.

4 The Xasm External Language Interface

In order to integrate ASM algorithms into other applications, Xasm defines an
external language interface. In the current version, this interface is implemented
for the connection of Xasm programs with programs written in the C language.
Interfaces to other language, like Java, are in preparation.
In principle, there are two alternatives how the interconnection to the exter-
nal application can be realized:
– C-functions are used to implement external ASM functions, or
– Xasm-asms are called from the C main program.
In the first case, the main control of the application is handled by the Xasm-part
of the system, while in the second case the C-application has the main control.
This is also reflected by the definition of the “main” C function: in the first case
it is contained in the Xasm-part, in the second case, the C-part must provide it.
The corresponding interfaces of Xasm for these two alternatives are explained
in the following.

4.1 External C-Functions

In the previous section we have shown, that external function can be specified
in Xasm using the asm construct. Alternatively, external functions can be im-
plemented in C. The corresponding Xasm-declaration is given as follows:
external "C:c name" [access mode] function
xasm name(a1 : T1 , . . . , an : Tn ) → T
The c name specifies the name of the C function; it can be omitted, if it is
equal to the Xasm-name of the function. Depending on the access mode of the
Xasm-function, the corresponding C-function prototypes differ slightly:
 Xasm– Extensible, Component-Based ASMs 81

– If the access mode is “monitored”, the C-functions are defined as follows:

ASMOBJ c name(ASM a, int argc, ASMOBJ* argv);
where “ASMOBJ” is the C-type representing elements of the superuniverse;
“ASM” is a C-struct containing information related to parent asm. These
types are specified in the header file “xasm.h” which must be included in
those files containing external function implementations. The arguments of
the function call can be accessed via the “argv” field using “argc” as argu-
ment count. The first argument, argv[0], always contains the name of the
corresponding Xasm-function as string element. The result of the external
function that is accessible in the calling asm is returned by the C-function
as “ASMOBJ”.
– If the access mode is “output”, the C-functions are defined as follows:
void cname(ASM a, int argc, ASMOBJ* argv, ASMOBJ val);
In this case, additionally the value that is used in the update representing
the call of the external function can be accessed using the “val” parameter.

As an example, the following C-code contains the implementation of the

“stderr” function previously used in one of the examples:
void xasm_stderr (ASM a, int argc, ASMOBJ* argv, ASMOBJ val) {
if (argc != 1) {
error("wrong # args for external function ’%s’.\n",
c_stringvalue(argv[0]));
return;
}
fprintf(stderr,"%s",str_obj(val));
}
The Xasm-library function “str obj” returns the string representation of an
ASMOBJ. The corresponding declaration of the external C-function in an asm
has the following format:

external "C:xasm_stderr" [output] function stderr → String

4.2 Embedding Xasm-Programs in C-Applications

If the Xasm-part of a system should provide services for a C-based application,

the main asm and all sub-asms of it can be called from the C-code. In this
case, the Xasm-compiler must be invoked with a special option that prevents
the generation of the “main”-function.
Before any of the asms can be invoked, the Xasm-part must be initiali-
zed. For this purpose, the generated C-code defines the function “asm main” as
follows:
int asm main(int argc, char **argv);
82 M. Anlauff

The arguments to this function are given as strings which are parsed and
transformed to corresponding “ASMOBJs”. The actual invocation of the main
asm can be made using the C-function “run mainasm”:
ASMOBJ run mainasm();
This kind of embedding is actually used in the implementation of the Xasm-
compiler itself: the static semantics check is carried out by an algorithm specified
in Xasm.
A number of external C-functions are already integrated into the runtime
system. For example, external functions to communicate using UNIX-Sockets,
string manipulation functions, file access functions etc. .

5 Non-standard Language Constructs of Xasm

Besides the implementation of the ASM core constructs, Xasm provides a num-
ber of useful extensions that can all be directly mapped to the original ASM
constructs. In the following, some of these extensions are described briefly; a
full version of the language specification is in preparation and will be available
shortly.

5.1 Constructor Terms

Xasm provides the possibility to define and use constructor terms. The concept
of constructor terms can be mapped to the ASM core language as follows: Ac-
cording to [17] each of the function names contained in the vocabulary V of an
ASM may be marked as relational or static, or both. In addition, we allow static
functions to be marked as constructive. Let Fc be the set containing all functions
in V marked as constructive, Fc ⊆ V . Let fc ∈ Fc , arity of fc = n, then the
following conditions hold for all states A of the ASM:
(i) ∀t1 , . . . , tn , ValA (ti ) 6= undef , 1 ≤ i ≤ n •
fc (t1 , . . . , tn ) 6= undef
(ii) ∀gc ∈ Fc , arity of gc is m; t1 , . . . , tn , ValA (ti ) 6= undef •
fc (t1 , . . . , tn ) = gc (s1 , . . . , sm ) ⇔
fc = gc ∧ n = m
∧ValA (ti ) = ValA (si ), 1 ≤ i ≤ n
where ValA (t) stands for the evaluation of term t in state A of the ASM. Infor-
mally speaking that means that each constructive function is (i) defined at all
locations and that (ii) the content of each location is a unique element of the
superuniverse w.r.t. the set of locations of all constructive functions. If fc ∈ Fc ,
then fc is called a constructor, and the terms fc (t1 , . . . , tn ) are called construc-
tor terms. In Xasm, the declaration of a constructor is part of the function
declarations, for example

constructor nil, cons( , )

universe BinT ree = {empty, children(l : BinT ree, r : BinT ree)}
 Xasm– Extensible, Component-Based ASMs 83

introduces the constructors nil, cons, empty, and children, where terms con-
structed using the latter two constructors are elements of the universe BinT ree.
Xasm also provides pattern matching functionality like it is used in many
other languages. Syntactically, pattern matching terms are used as condition
terms in conditionals, e. g.: 5

if b =∼ children(&l, &r) then

R(&l, &r)
else
...
endif

There are three kinds of pre-defined, commonly-used constructors in Xasm:

sets, sequences, and tuples. These constructors are specified using their usual
representation: {x1 , . . . , xn } for sets, [x1 , . . . , xn ] for sequence, and (x1 , . . . , xn )
for n-tuples. For sequences, the notation [H|T ] can be used in pattern matching
terms for accessing head and tail of a sequence.

5.2 Regular Expressions

In practice, strings are widely used as a common data format for exchanging
information between different systems. Xasm therefore provides a special kind
of pattern matching based on regular expressions as they are used in UNIX (e.g.
in the “sed” program) as well as in many scripting languages like Perl [33] and Tcl
[26]. The regular expression pattern matching is invoked using the =∼ operator
like for pattern matching with constructor terms. If both operands of the =∼
operator are strings, then the right operand is interpreted as regular expression
and the left operand as string being match against the regular expression. For
example, the regular expression pattern matching expression

s =∼ ’ˆ[A-Z]’

evaluates to true, if s is a string starting with a capital letter.

In regular expressions, parenthesis “\(..\)” can be used to mark certain parts
of the expression that correspond to sub-strings of the left-hand-side string, if the
pattern matching has been successful. For that, Xasm provide a special form of
regular expression pattern matching: If the left-hand-term of a pattern matching
expression evaluates to a string object s and the right-hand-term evaluates to a
tuple the first argument of which represents a string object r and the remaining
arguments are pattern matching variables &v1 , . . . , &vn , the string s is matched
against the regular expression r and the sub-matches are put into the pattern
matching variables &v1 , . . . , &vn , if the match has been successful.

5
Xasm uses a special syntax for pattern matching variables and equality symbol
84 M. Anlauff

if "AnyString" =∼ (’ˆ\(.\)\(.*\)$’, &hd, &tl) then

...
endif

In this example, the regular expression contains two sub-matches, the first one
matches the string "A", the second one the string "nyString".6 The submatches
can be accessed in the then-part of the conditional rules as values of the pattern
matching variables &hd and &tl.

5.3 The “Once”-Rule

A common situation occurring in ASM formalizations is that certain rules should
fire only once during the run of an ASM. Normally, one has to introduce extra
functions in order to ensure this behavior. Xasm introduces a “once” rule for
these situations:
if once R1
else R2
endif
which is – in terms of ASMs – equivalent to the conditional rule
if once(n) then
R1
once(n) := f alse
else
R2
endif
where n is a unique number representing the n’th occurrence of a “once”-rule
in the ASM, and once is a one-ary dynamic relation initialized with true for all
these numbers n. The notation “once R” is an abbreviation for
if once R else skip endif

6 Grammar Definitions in Xasm

Historically, Xasm has been developed as underlying ASM implementation for

Montages, a semi-visual method for specifying the syntax and semantics of pro-
gramming languages, see [21,2,3]. As a consequence, the support for program-
ming language related features has been integrated into the Xasm language as a
means to extend the original syntax with domain-specific constructs. The syntax
and semantics of these extensions can be specified using the Montages method
together with tool support environment Gem-Mex which translates user-defined
language definitions into Xasm-code.
6
Single quotes are used for the regular expressions in order to prevent interpretation
of special symbols (like “\”) in the string. To understand the example: ˆ stands for
the beginning, $ for the end of a string, the dot represents any character, and the
asterisk stand for zero or more occurrences of the preceeding expression.
 Xasm– Extensible, Component-Based ASMs 85

The grammar definitions used for the translation of Montages-specification

can also be used directly in Xasm. For that purpose, nonterm and token-
declarations can be given in Xasm, resulting in the generation of a parser for
the specified language. As an example for using grammar definitions in Xasm,
Figure 3 contains the specification of a parser that accepts empty XML tags.
The generated C-function can be accessed using an external function returning
the root node of the parse tree being constructed during parsing.

asm xmlparser(inpf ile) is

use syntax
token Ident = "[A-Za-z][A-Za-z0-9_]*"
startnonterm Elements
nonterm Elements[Element]
base /* empty */
cont Elements Element
endnonterm
nonterm Element = ElementEmpty endnonterm
nonterm ElementEmpty ::= ”<”Ident Attributes”/ >”;
endnonterm
nonterm Attributes[Attribute];
base /* empty */
cont Attributes Attribute
endnonterm
nonterm Attribute ::= Ident”=”String token;
lhs ← (Ident.N ame, String token.N ame)
endnonterm
external "C" function parse Elements(f ilename : String) → Elements
function RootN ode → Elements
RootN ode := parse Elements(inpf ile);
endasm

Fig. 3. A grammar specification in Xasm for parsing empty XML Elements

7 The Xasm Support Environment

The support environment of Xasm consists of the Xasm-compiler, the runtime
system and the graphical debugging and animation interface. In this paper, these
tools are only sketched briefly, a more detailed description will be contained in
the Xasm user manual which is currently under development. A description of
the graphical animation tool is also contained in [2,3].
86 M. Anlauff

7.1 The Xasm-Compiler and Runtime System

As already mentioned in Section 2, the Xasm-compiler xasmc translates Xasm
source code into C-code implementing the executable version of the ASMs speci-
fied in the Xasm source files. The syntax analysis part is implemented in C using
“lex” and “yacc” for generating scanner and parser code. The type checking part
is implemented in Xasm itself using the C-interface as introduced above.
The Xasm runtime system implements the core functionality of the Xasm
language. In here, all algorithms and data structures are realized being used
to transform an ASM into an executable program. At the heart of the run-
time system is the implementation of update and access functionality for ASM
functions. For that, a hashing mechanism is used to provide optimized access
to values of ASM functions. The runtime system also contains garbage collec-
tion facilities, which is indispensable, if ASM algorithms are used for continuous
control systems, as described in [5].

7.2 The Xasm Graphical Animation and Debugging Interface

In order to be able to animate and/or debug the Xasm program, a graphical
animation and debugging tool has been realized that enables to stepwisely exe-
cute the ASM, to trace updates that has been performed in each step, and to
view function values in each step. In case a grammar has been specified as input
format for the Xasm program, a special kind of graphical animation window can
be used to display function values that refer to node in the parse tree. Figure
4 shows a screen-dump of a debugging session. Additionally, an integrated de-
sign environment, incorporating the graphical user interface is currently under
development.

7.3 The Xasm-LATEX Package

As an additional support feature, an LATEX-package “asm.sty” is defined for
typesetting Xasm specifications. The LATEX-files can directly be used as input
file to the Xasm-compiler, so that no additional work is necessary to produce a
high-quality documentation from a running Xasm specification. The Xasm code
parts in this document are produced using the asm style being realized based on
the LATEX “program” style as defined in [14].

8 Conclusion
In this paper, the ASM based language Xasm has been presented focusing on
the additional features provided by the language with respect to the ASM core
concepts as defined in the Lipari Guide. A novel concept for structuring ASM
specifications based on the notion of components has been presented. This con-
cept perfectly fits into the basic model of the ASM approach, because it allows
to choose the level of abstraction for describing that fits best to a given pro-
blem without regarding technical constraints. Furthermore, this concept allows
 Xasm– Extensible, Component-Based ASMs 87

Fig. 4. Snapshot of a Xasm-debugger session

efficient development cycles, because asm’s can be designed as reusable compo-

nents by exactly specifying what is expected from the environment. An algebraic
view of a similar structuring concept has been given by Wolfgang May in [25].
He tries to formalize the notion of “hierarchical ASMs” which is used in early
papers of Gurevich and Börger. Later in [10] special cases of the May’s concepts
are proposed for ASM composition. These cases as well as May’s general notion
of “hierarchical ASMs” together with means for data encapsulation are directly
supported by Xasm.
The language presented in this paper is fully implemented. The system is
used as the basis for the Montages/Gem-Mex, where generated Xasm code is
translated into an interpreter for the language specified using Montages. As a
case study the full specification of ANSI C has been entered by Wuwei Shen [19].
Other case studies are currently under development, see for example an applica-
tion to microprocessor simulation in [32] and the application of Xasm as gluing
code in legacy systems [4].
As next steps, the Xasm compiler, runtime system and graphical support
environment will be further optimized. Also, a concept how tools for the (auto-
matic) verification, model checking and analysis of the ASM formalizations can
be integrated into the system is currently under development and will be part of
the support system in future versions of the tool. Furthermore, the connection
to repository systems will be subject to future considerations concerning the
Xasm-language. For example, certain ASM functions may be marked as “per-
sistent”, meaning that the values of the locations of these functions are stored
88 M. Anlauff

in the repository system, so that they can be access the next time the Xasm
is executed. This kind of extension is currently part of work carried out in an
industry-based research project running at GMD In this project, it is currently
considered to use Xasm for formulating certain consistency check algorithms
occurring in the context of this project.
Additional applications outside the ASM area are possible, since ASMs can
be considered as an instance of so called transitions system models, which form
as well the basis for other popular formalisms such as UNITY [11], TLA [23],
SPL [24] and the SAL intermediate language [29]. As mentioned in Section 6 the
Xasm system is integrated with the Montages method so that new or alternative
constructs can be easily supported by Xasm. Using Montages, both syntax and
semantics of new or alternative Xasm constructs can be developed in an inte-
grated development environment called Gem-Mex. Such an extensible system
architecture allows the Xasm tool to be tailored to one of the above-mentioned
formalisms based on transition systems.
An adapted version of Xasm would be especially useful in a context like
the SAL architecture [7], where various other tools are integrated for tasks like
theorem proving and model checking. The development cycle supported by the
tools integrated with SAL could be extended with a tool like Xasm for debugging
and animation of the transition system under consideration. After the developer
gains confidence in his specification by debugging and animating them with
Xasm, other tools can be used to further analyze the specification. As a result of
the further analysis, the original specification typically undergoes major changes.
To avoid simple errors introduced by these changes, the specification can again
be debugged and animated using Xasm before it is passed to the other tools.
We plan to adapt Xasm to architectures like SAL in order to make it ac-
cessible to a large base of users. We can thus adapt the Montages method to
formalisms like SAL by providing a language semantics with SAL. A specific
language may then be developed so that the SAL semantics of programs written
in that language can be easily analyzed using the available tools. The concrete
syntax of such a language can be adapted to the terminology of domain experts
using the language to describe their system. Examples of using Montages to
develop languages for special domains are given in [20,1,4].

Acknowledgments. I very much thank Philipp W. Kutter and Alfonso Pieran-

tonio for their collaboration and fruitful discussions which had a great influence
in the design and implementation of Xasm. I also like to thank Yuri Gurevich
and Egon Börger for their interest in my work presented in this paper and for
always helping me solving the right problems. My dearest special thanks go to
Asuman Sünbül. Her work in the field of component-based software engineering
has very much influenced the component model presented in this paper.
 Xasm– Extensible, Component-Based ASMs 89

References
1. M. Anlauff, A. Bemporad, S. Chakraborty, P.W. Kutter, D. Mignone, M. Morari,
A. Pierantonio, and L. Thiele. From ease in programming to easy maintenance:
Extending DSL usability with Montages. Technical Report 84, Institute TIK, ETH
Zürich, December 19999.
2. M. Anlauff, P. Kutter, and A. Pierantonio. Formal Aspects of and Development
Environments for Montages. In M. Sellink, editor, 2nd International Workshop
on the Theory and Practice of Algebraic Specifications, Workshops in Computing,
Amsterdam, 1997. Springer.
3. M. Anlauff, P. Kutter, and A. Pierantonio. Enhanced control flow graphs in Monta-
ges. In A.Zamulin D.Bjoerner, M.Broy, editor, Perspective of System Informatics,
volume 1755 of LNCS, pages 40 – 53, 1999.
4. M. Anlauff, P.W. Kutter, A. Pierantonio, and Asuman Sünbül. Using domain-
specific languages for the realization of component composition. In Proceedings
Formal Approaches in Software Engineering FASE00, LNCS, 2000.
5. M. Anlauff and A. Sünbül. An ASM specification of an elevator control system.
1999.
6. M. Anlauff and A. Sünbül. Software architecture based composition of components.
In GI-Workshop Sicherheit und Zuverlässigkeit software-basierter Systeme, 1999.
7. S. Bensalem, V. Ganesh, Y. Lakhnech, C. Muñoz, S. Owre, H. Rueß, J. Rushby,
V. Rusu, H. Saı̈di, N. Shankar, E. Singerman, and A. Tiwari. An overview of SAL.
In C. Michael Holloway, editor, LFM 2000: Fifth NASA Langley Formal Methods
Workshop, June 2000. to appear.
8. J. A. Bergstra and P. Klint. The ToolBus coordination architecture. In Ciancarini
and Hankin [12], pages 75–88.
9. E. Börger and D. Rosenzweig. A Mathematical Definition of Full Prolog. In Science
of Computer Programming, volume 24, pages 249–286. North-Holland, 1994.
10. E. Börger and J. Schmid. Composition and Submachine Concepts for Sequential
ASMs. In P. Clote and H. Schwichtenberg, editor, Gurevich Festschrift CSL 2000,
LNCS. Springer-Verlag, 2000. to Appear.
11. M. Chandy and J. Misra. Parallel Program Design: A Foundation. Addison-Wesley,
Reading, MA, 1988.
12. Paolo Ciancarini and Chris Hankin, editors. Coordination and models, Proceedings
of the first international conference, Cesena, Italy, number 1061 in LNCS. Springer
Verlag, 1996.
13. G. Del Castillo. The ASM Workbench: an Open and Extensible Tool Environment
for Abstract State Machines. In Proceedings of the 28th Annual Conference of the
German Society of Computer Science. Technical Report, Magdeburg University,
1998.
14. Michel Goossens, Frank Mittelbach, and Alexander Samarin. The LATEX Compa-
nion. Tools and Techniques for Computer Typesetting. Addison-Wesley, Reading,
MA, USA, second edition, 1994.
15. F. Griffel. Componentware. dpunkt.verlag, 1998.
16. Y. Gurevich. Evolving Algebras 1993: Lipari Guide. In E. Börger, editor, Specifi-
cation and Validation Methods, pages 9–36. Oxford University Press, 1995.
17. Y. Gurevich. May 1997 Draft of the ASM Guide. Department Technical Report
CSE-TR-336-97, University of Michigan, 1997.
18. Y. Gurevich and J. Huggins. The Semantics of the C Programming Language. In
E. Börger, H. Kleine Büning, G. Jäger, S. Martini, and M. M. Richter, editors,
Computer Science Logic, volume 702 of LNCS, pages 274–309. Springer, 1993.
90 M. Anlauff

19. J.K. Huggins and W. Shen. The static and dynamic semantics of C. In Local
Proceedings of ASM2000, TIK Report Nr. 87, 2000.
20. P. W. Kutter, D. Schweizer, and L. Thiele. Integrating formal domain-specific
language design in the software life cycle. In Current Trends in Applied Formal
Methods, LNCS. Springer, October 1998.
21. P.W. Kutter and A. Pierantonio. Montages: Specifications of Realistic Program-
ming Languages. Journal of Universal Computer Science, 3(5):416–442, 1997.
22. P.W. Kutter and A. Pierantonio. The Formal Specification of Oberon. Journal of
Universal Computer Science, 3(5):443–503, 1997.
23. L. Lamport. The temporal logic of actions. ACM TOPLAS, 16(3):872–923, 1994.
24. Z. Manna and A. Pnueli. The Temporal Logic of Reactive and Concurrent Systems,
Volume 1: Specification. Springer-Verlag, New York, NY, 1992.
25. W. May. Specifying complex and structured systems with evolving algebras. In
M. Bidoit and M. Dauchet, editors, Proceedings of TAPSOFT’97: Theory and Prac-
tice of Software Development, number 1214 in LNCS, pages 535 –549, 1997.
26. John K. Ousterhout. Scripting: Higher level programming for the 21st century.
IEEE Computer, 31(3):23–30, March 1998.
27. J.-G. Schneider and O. Nierstrasz. Scripting: Higher-level programming for
component-based systems. In OOPSLA 1998, 1998. Tutorial.
28. Jean-Guy Schneider and Oscar Nierstrasz. Components, scripts and glue. In Leonor
Barroca, Jon Hall, and Patrick Hall, editors, Software Architectures – Advances and
Applications, pages 13–25. Springer, 1999.
29. N. Shankar. Symbolic Analysis of Transition Systems. In This volume.
30. Asuman Sünbül. Architectural Design of Evolutionary Software Systems. PhD
thesis, Technical University Berlin, 1999. in preparation.
31. Clemens Szyperski. Component Software: Beyond Object-Oriented Programming.
ACM Press and Addison-Wesley, New York, N.Y., 1998.
32. J. Teich, P.W. Kutter, and R. Weper. Description and simulation of microprocessor
instruction sets using asms. In This volume.
33. Larry Wall and Randal L. Schwartz. Programming Perl. O’Reilly Associates, Inc.,
Sebastopol, CA, 1990.
34. C. Wallace. The Semantics of the Java Programming Language: Preliminary Ver-
sion. Technical Report CSE-TR-355-97, EECS Dept., University of Michigan, De-
cember 1997.
 ?
Generic Facilities in Object-Oriented ASMs

A. V. Zamulin

Institute of Informatics Systems

Siberian Branch of the Russian Academy of Sciences
630090, Novosibirsk, Russia
e-mail: [email protected]
fax: +7 3832 323494
phone: +7 3832 396258

Abstract. Facilities for defining generic object types, generic type ca-
tegories, generic functions and generic procedures in an object-oriented
ASM are described in the paper. These facilities permit one to specify
algorithms over complex data structures abstracting both from the type
of the structure components and the structure itself. The use of the fa-
cilities is demonstrated by the specifications of some important parts of
Standard Template Library for C++.

Keywords: ASM, object types, object categories, generic specifications.

1 Introduction

Object-oriented ASMs as a variant of traditional ASMs [1,2] are formally in-

troduced in [3]. They permit the specification of a dynamic system in terms of
mutable and constant objects belonging to different object types. Object types
are partially ordered according to a type–subtype relationship which serves for
modeling inheritance. Unfortunately, the technique described in the paper does
not provide facilities for the specification of generic object types and generic
algorithms. At the same time, such facilities are highly needed if one wishes to
write reusable specifications.
As a typical example, let us consider the problem of the specification of
Standard Template Library (STL) for C++ [4] which will be used in the paper
as a running example. Currently the semantics of the library components is
given informally as a set of requirements presented partly in English and partly
as C++ program fragments. As a result, the semantics remains incomplete and
imprecise, depending heavily on reader’s (and library implementor’s) intuition
and knowledge of C++. Therefore, a formal description of STL independent of
a particular programming language is highly needed.
STL is based on the notion of container which is a data structure consisting
of a number of elements of the same type. Several container classes are defined
?
This research is supported in part by Russian Foundation for Basic Research under
Grant 98-01-00682.

Y. Gurevich et al. (Eds.): ASM 2000, LNCS 1912, pp. 91–111, 2000.
c Springer-Verlag Berlin Heidelberg 2000

92 A.V. Zamulin

in STL: vectors, lists, deques, sets, multisets, maps, and multimaps. Other con-
tainer classes can be defined if needed. Each container class is parameterized by
the component type. Thus, for each data structure one can write an algorithm
abstracting from the component type. This provides the first level of genericity
typical of C++.
To abstract from the container’s structure, STL introduces a notion of ite-
rator which is a generalization of the pointer notion. Iterators are grouped into
different iterator categories providing abstract data-access methods. There are
categories of input iterators, output iterators, forward iterators, bidirectional ite-
rators, and random-access iterators. Iterator categories build a hierarchy. This
means that a forward iterator is also an input iterator and an output iterator, a
bidirectional iterator is also a forward iterator, and a random-access iterator is
also a bidirectional iterator. Algorithms can now be written in terms of iterators
abstracting from a concrete data structure. Most important here is the fact that
an algorithm requiring, say, an input iterator can also use a forward or bidirec-
tional or random-access iterator. This provides the second level of genericity.
One of the ways of the formal STL definition is the use of classical alge-
braic specifications whose advantages are sound mathematical foundation and
existence of specification languages and tools. Taking into account the generic
nature of the data structures and iterator categories, a specification language
like Spectrum [5] providing parameterized sorts and sort classes can be used for
this purpose. However, the notions of iterator and container subsume a notion of
state which can change when containers are searched or updated. The modeling
of the state by classical algebraic specifications involves extra data structures
representing the state which must be explicitly passed to a function as argument
and yielded as result. Algebraic specifications are also a poor tool for describing a
data structure with an arbitrary order of component insertion and deletion. This
leads to very complex specifications with problems of describing the differences
between different types of containers.
At the same time it seems very natural to consider containers and iterators
as objects possessing state and to define container classes and iterator classes as
generic object types parameterized by the component type. To define formally
iterator categories and thus represent the hierarchy of iterator classes, we need
a notion of type category similar to that of sort classes of Spectrum and a notion
of generic type category absent in conventional algebraic specifications. Thus,
the adaptation and extension of generic facilities of algebraic specifications for
enhancing object-oriented ASMs are the main tasks of this paper.
The paper is organized in the following way. Concrete object types defining
sets of states of potentially mutable objects and object-oriented dynamic systems
generalizing the communities of object states and their transitions are introdu-
ced in Section 2. Unconstrained generic object types are defined in Section 3.
Generic vector type and generic list type as typical representatives of Standard
Template Library are specified in Section 4 and Section 5, respectively. Object
type categories permitting the classification of object types on the base of their
operations are introduced in Section 6. Generic object types, generic type cate-
 Generic Facilities in Object-Oriented ASMs 93

gories and generic functions constrained by object type categories are defined in
Section 7. Some related work is discussed in Section 8, and some conclusions are
given in Section 9.
It is assumed that the reader is familiar with the basic ASMs notions which
can be found in [1,2]. The familiarity with the formal aspects of object-oriented
ASMs [3] is desirable, but not obligatory.

2 Basic Notions

2.1 Object Types

We distinguish between data and objects and, respectively, between data types
and object types. A data identifies itself and is immutable. An object possesses
a unique identifier and state which can be mutable. A data type defines a set
of immutable values and operations over them. An object type defines a set
of states of a potentially mutable object and a number of methods capable to
observe or update the object’s state. This distinction between data types and
object types requires different methods of their specification.
We consider that data types are specified by means of one of the algebraic
specification languages like Spectrum [5], Ruslan [6] or CASL [7]. Let (Σ0 , Ax),
where Σ0 is a signature and Ax is a set of axioms, be the specification of a number
of data types. Σ0 is called the data signature in the sequel. An algebra A0 of this
signature is called a data algebra. An object-structured signature, Σobj , extends
Σ0 with a set of object type signatures [3]. Respectively, an object-structured spe-
cification, (Σobj , Axobj ), extends (Σ0 , Ax) with a set of object type specifications.
Constructing the object-structured specification, we accompany each object type
signature with the corresponding axioms. Therefore, the notions of object type
signature and object type specification correspond, respectively, to the notions
of class declaration and class definition of C++.
The object-structured specification is defined in the following way. Let OT Y P E
be a set of (object type) names and OΦ a set of object type specifications.
The object-structured specification is then a triple hOT Y P E, OΦ, into i, where
into is a function mapping OT Y P E into OΦ. If T ∈ OT Y P E, ots ∈ OΦ, and
into (T ) = ots, then the maplet hT 7→ otsi is the specification of the object type
T (in this case we sometimes write that ots is marked with T ).
The specification of an object type T in this paper has the following form:
class T = spec
[object-type-signature]
{axioms},
where object-type-signature is a triple set-of-attribute-declarations; set-of-mutator-
declarations; set-of-observer-declarations.
An attribute or observer declaration has the following form: operation-name:
operation-profile. The operation-profile is either T or T1 , . . . , Tn −→ T where
T, Ti are data/object type names indicating the types of attribute/observer pa-
rameters (if any) and the result type. A mutator declaration is either just a
94 A.V. Zamulin

mutator name or mutator-name: mutator-profile, where mutator-profile is a se-

quence of data/object type names indicating the types of mutator parameters.
Intuitively, a tuple of attributes defines an object’s state, an observer is a
function computing something at a given object’s state, and a mutator is a
procedure changing an object’s state. Attributes are often called instance va-
riables, and observers and mutators are often called methods in object-oriented
programming languages. All of them are called methods in the sequel.
Notation: in the following examples sets of attribute, observer, and mutator
declarations in the object type signature are preceded by keywords attribute,
observer, and mutator, respectively.
As an example, let us consider a simple model of memory consisting of loca-
tions storing integer numbers. A consecutive number of locations forms a vector
which is a structure allowing both sequential and random access to its elements.
A location representing a vector element is called a vector iterator in the sequel.
Such an iterator can be understood as an object possessing a unique identifier
(address) and the attribute value stored updated by the mutator put value and
observed by the observer get value. Since we wish to provide both sequential and
random access to the vector elements, we define several extra observers as well.
As a result, we create the following object type signature:

class VecIt = spec

[attribute value stored: Integer;
mutator put value: Integer;
observer get value: Integer;
advance, retreat: VecIt;
plus, minus: Nat −→ VecIt;
difference: VecIt −→ Nat;
eq, neq, less, greater, leq, geq: VecIt −→ Boolean]

In the above example, the community of vector elements is represented by

the class VecIt (vector iterator) with several methods giving possibilities to move
either to the next (advance) or previous (retreat) element, to jump over several
elements forward (plus) or backward (minus), to calculate the distance between
two elements (difference) and to compare the element (location) identifiers (eq,
neq, less, greater, leq, geq). Note that the methods advance, retreat, plus and
minus are observers since they produce objects and do not update the states of
the corresponding objects.
Let Σ 0 = hΣ0 , Σobj i. A Σ 0 -algebra, A, is constructed by extending a Σ0 -
algebra, A0 , with a set of elements, called object identifiers, for each object
type name T . For example, a set of location addresses can be associated with
the object type V ecIt. We denote by AT the set of elements associated with a
data/object type T in the algebra A. In addition to the set of object identifiers,
for each object type T , a (partial) function atAT : AT −→ (AT01 , . . . , AT0n −→ AT0 )
is associated in the algebra A with the attribute name at : T10 , . . . , Tn0 −→ T 0 (a
function atAT : AT −→ AT0 is associated with the attribute name at : T 0 ) declared
in the signature of T ; such a function is called an attribute function. If id ∈ AT ,
 Generic Facilities in Object-Oriented ASMs 95

then atAT (id) is an attribute of id. Note that, like in object-oriented languages, the
object type is always the first parameter type of its attribute function although
it is never indicated in the attribute declaration. An object identifier is used as
argument when the corresponding object type is a function parameter type, and
an object identifier is produced when the corresponding object type is the result
type of a function.
An object is a pair (id, obs) where id is an object identifier and obs is a tuple
of its attributes called object’s state. For example, if id is an object identifier of
type VecIt in algebra A, then the pair hid, value storedA (id)i is an object with the
state value storedA (id).
To define the interpretation of observer and mutator names, a notion of
dynamic system is introduced.

2.2 Dynamic System

We discussed above only the functions defined inside the frames of object types.
In a more general case, an algebra can possess a number of ”stand-alone” fun-
ctions and constants defined outside the frames of object types. In addition to
this, higher-order functions for observing and transforming algebras can be de-
fined. Therefore, we introduce a notion of dynamic system possessing a number
of states and a number of operations for observing and updating the states.
Generally, the specification of an object-oriented dynamic system is repre-
sented by the following tuple:
h(Σ0 , Ax), (Σobj , Axobj ), Σdin , (Σdep , Axdep ), (Σproc , Axproc )i.
Its first component is the specification (the signature and axioms) of a number of
data types and related functions, and the second component is the specification
of a number of object types discussed above. The third component is the signa-
ture of a set of dynamic functions/constants which can be different in different
states of a dynamic system.
A dynamic function is declared as follows:
dynamic function function-name: T1 , ..., Tn −→ T ;
where T1 , ..., Tn are parameter types and T is a result type. A function without
parameters is called a constant and declared as follows:
dynamic const constant-name: T ;
where T is the type of the constant.
Examples.
dynamic const an iterator: VecIt;
dynamic function matrix: Nat, Nat −→ VecIt;

Let Σ 0 = hΣ0 , Σobj i, Σ = hΣ 0 , Σdin i, and A0 be a Σ 0 -algebra. A Σ-algebra

A is constructed by extending A0 with an element cA ∈ A0T for a constant decla-
ration c : T from Σdin and a (partial) function f A : A0T1 × . . . × A0Tn −→ A0T for a
function declaration f : T1 , . . . , Tn −→ T from Σdin . Such an algebra is called
an instance algebra or state. It represents a number of objects and dynamic fun-
ctions. An update of an object’s state or dynamic function as well as creation
96 A.V. Zamulin

or deletion of an object leads to the transformation of one instance algebra into

another. Note that the carrier of A is the same as that of A0 .
The forth component of the specification of a dynamic system is the signature
and axioms of dependent functions. They are declared in the same way as the
functions in Σdin with the use of the keyword depend. The values produced by
these functions depend on the state.
The fifth component of the specification of a dynamic system is the signa-
ture and axioms of procedures serving for state transformations. A procedure is
declared as follows:
proc procedure-name: T1 , ..., Tn ;
where T1 , ..., Tn are parameter types (n can be zero, i.e., a procedure can have
no parameter). Example:
proc allocate: Nat – allocation of a number of vector iterators;
Let now OID be the type of object identifiers isomorphic to natural num-
bers. This means that there is the sort name OID and two function names, say
nat to id and id to nat in Σ0 and the axiom id to nat(nat to id(x)) = x, where
x is a universally quantified variable of type N at, in Ax. Given a Σ0 -algebra
A0 , we assume that all object identifiers are selected from the set of values of
this type sequentially so that no object identifier is selected twice. We denote
by DS(A0 ) a dynamic system based on a data algebra A0 and by |DS(A0 )| the
set of instance algebras satisfying the following condition: the restriction of each
algebra of |DS(A0 )| to the data signature equals A0 , i.e., each algebra in |DS(A0 )|
is based on the same data algebra.
Given a dynamic system signature, a data algebra A0 with a set of object
identifiers, and a set of instance algebras |DS(A0 )|, we construct a dynamic sy-
stem DS(A0 ) by associating:

– with each observer name b : T1 , . . . , Tn −→ T 0 in an object type signature

DS(A )
marked with T , a partial function, bT 0 , mapping a pair hA, hid, a1 , . . . , an ii
(a pair hA, idi for an observer name b : T 0 ) to an element a0 ∈ AT0 , where
A ∈ |DS(A0 )|, id ∈ AT , and ai ∈ ATi , i = 1, . . . , n;
– with each mutator name m : T1 , . . . , Tn in an object type signature marked
DS(A )
with T , a partial function, mT 0 , mapping a pair hA, hid, a1 , . . . , an ii (a
pair hA, idi for a mutator name m without a profile) to an update set1 ,
where A ∈ |DS(A0 )|, id ∈ AT , and ai ∈ ATi , i = 1, . . . , n.
– with each dependent function name f : T1 , . . . , Tn −→ T , a partial function,
f DS(A0 ) , mapping a pair hA, ha1 , . . . , an ii (an algebra A for a function name
f : T ) to an element a ∈ AT , where A ∈ |DS(A0 )| and ai ∈ ATi , i = 1, . . . , n;
– with each procedure name p : T1 , . . . , Tn , a partial function, pDS(A0 ) , mapping
a pair hA, ha1 , . . . , an ii (an algebra A for a procedure name p without a profile)
to an update set, where A ∈ |DS(A0 )| and ai ∈ ATi , i = 1, . . . , n.

Thus, an object-oriented dynamic system DS(A0 ) of signature DΣ consists

of a set of instance algebras |DS(A0 )| called the carrier of DS(A0 ) and a set
1
See one of [1,2,3] for the definition of an update set and its use for producing a new
state.
 Generic Facilities in Object-Oriented ASMs 97

of observers, mutators, dependent functions and procedures. For example, the

function associated with the mutator name put value declared in the class V ecIt
will produce an update set needed for updating the content of a location, and the
function associated with the observer name advance will produce the identifier
considered to be next to a given identifier in a given state. A call of the procedure
allocate will eventually produce a new state by attaching a number of locations
to the current state.

2.3 Terms and Their Interpretations

There are several rules for creating terms of object types following the conven-
tions used in object-oriented languages. Moreover, a special kind of term called
transition term is introduced to denote transitions from one algebra to another
(transition rules are special cases of transition terms). The main rules for creating
the terms are the following (we omit interpretation where it is self-evident).
1. If at : T1 , . . . , Tn −→ T 0 is an attribute/observer declaration from the object
type signature marked with T , t1 , . . . , tn are terms of types T1 , . . . , Tn , res-
pectively, and t is a term of type T , then t.at(t1 , . . . , tn ) is a term of type T 0 .
Examples: vit.value stored, vit.advance, vit.plus(2), vit.less(vit1), where
vit and vit1 are the names of dynamic constants of type V ecIt. The term
is interpreted by invoking the corresponding attribute/observer function in
the current state.
2. If t is a term of type T , then D(t) is a term of type Boolean. Interpretation:
D(t)A = trueA if t is defined in A, and D(t)A = falseA otherwise. The interpre-
tation of this term allows us to check whether the argument term is defined
in the current state. For example, the interpretation of D(vit.advance) will
let us know whether there is an object identifier next to vit in the current
state.
3. If m : T1 , ..., Tn is a mutator declaration from the object type signature
marked with T , t1 , ..., tn are terms of types T1 , ..., Tn , respectively, and t is a
term of type T , then t.m(t1 , ..., tn ) is a transition term called a mutator call.
Example: vit.put value(3). This kind of term serves for indicating an update
of a mutable object. The term is interpreted by invoking the corresponding
mutator.
4. If f : T1 , . . . , Tn −→ T is a dynamic/dependent function declaration and
t1 , . . . , tn are terms of types T1 , . . . , Tn , respectively, then f (t1 , . . . , tn ) is
a term of type T . The term is interpreted by invoking the corresponding
dynamic/dependent function.
5. If p : T1 , ..., Tn is a procedure declaration and t1 , ..., tn are terms of types
T1 , ..., Tn , respectively, then p(t1 , ..., tn ) is a transition term called a procedure
call. The term is interpreted by invoking the corresponding procedure.

2.4 Axioms
Axioms accompanying the declarations of dependent functions are data equations
of the form t1 == t2 where t1 and t2 are two terms of the same type. The
98 A.V. Zamulin

terms are composed of universally quantified variables, operation names from

Σ0 , method names from Σobj , and function names from Σdin and Σdep . The
interpretation of such a term produces an algebra element. A data equation
t1 == t2 is satisfied in a dynamic system iff the interpretation of both t1 and
t2 produces the same algebra element in any its state. Since a term containing
the name of a dynamic function can evaluate differently in different states, the
function whose specification contains such a term generally produces different
results in different states (i.e., the result depends on the state).
Axioms accompanying the declarations of procedures are dynamic equations
of the form t1 == t2 where t1 and t2 are two transition terms. A transition term
is either a procedure call or a mutator call or a transition rule. The interpretation
of either of them produces an update set. In a dynamic equation t1 == t2 , the
first term is normally a procedure call or mutator call and the second one is
a transition rule. A dynamic equation is satisfied in a dynamic system iff the
interpretation of both terms in any its state produces the same update set.
Transition rules are conventionally created like in traditional ASMs [2] with
an additional possibility of using procedure calls and mutator calls in rule con-
structors. There is, however, an important difference in the treatment of the
assignment of an undefined value to a location. There can be no single undef
value for all data types. Therefore, all dynamic functions are partial, and in an
update rule
f (t1 , ..., tn ) := undef
undef is just a keyword indicating that f (t1 , ..., tn ) becomes undefined.
We can now write the following axioms for the above VecIt signature (typical
axioms for equality and non equality are omitted):

{forall i, i1: VecIt, x, n: Nat. – declaration of universally quantified variables

i.put value(x) == i.value stored := x; – dynamic axiom
i.get value == i.value stored; – this one and all the others are data axioms
i.advance == nat to id(id to nat(i) + 1);
i.retreat == nat to id(id to nat(i) – 1);
i.plus(n) == nat to id(id to nat(i) + n);
i.minus(n) == nat to id(id to nat(i) – n);
i.difference(i1) == id to nat(i) – id to nat(i1)
i.less(i1) == id to nat(i) < id to nat(i1);
i.greater(i1) == id to nat(i) > id to nat(i1);
i.leq(i1) == id to nat(i) ≤ id to nat(i1);
i.geq(i1) == id to nat(i) ≥ id to nat(i1)}

3 Generic Object Types

We cannot be satisfied with only concrete object types because many object
types can have a similar structure and it would be tiresome to specify them
again and again. Therefore a notion of generic object type is introduced. We
start with the simplest case, unconstrained generic object types.
 Generic Facilities in Object-Oriented ASMs 99

An object-structured specification Σ is defined above as a triple hOT Y P E,

OΦ, into i, where into is a function mapping OT Y P E into OΦ. We propose the
following way of constructing the elements of OT Y P E, using two non intersec-
ting sets of names, T and R, called names of concrete object types and names
of generic object types, respectively: if T ∈ T , then T ∈ OT Y P E; if T1 , ..., Tn
are elements of T and R ∈ R, then R(T1 , ..., Tn ) ∈ OT Y P E. The elements of
OT Y P E are called type terms in the sequel.
Let now into (R(T11 , ..., T1n )) = Spec1 and into (R(T21 , ..., T2n )) = Spec2,
where R(T11 , ..., T1n ) and R(T21 , ..., t2n ) are type terms and Spec1 and Spec2
are object type specifications. We say that the object type R(T11 , ..., T1n ) is a
sibling of the object type R(T21 , ..., T2n ) if the replacement of each T1i with T2i ,
i = 1, ..., n, in Spec1 converts it in Spec2. Thus, we can say that the object types
V ector(Integer) and V ector(Char) are siblings if the replacement of Integer
with Char in the specification of V ector(Integer) converts it in the specification
of V ector(Char) and vice virsa. We can propose a special way of constructing a
part of the function into for a family of object type siblings.
Let q1 , ..., qn be names (of type parameters). A pair hR(q1 , ..., qn ), Speci,
where R ∈ R and Spec is an object type specification additionally using q1 , ..., qn
as type terms in method declarations, is part of the function into , such that for
any type name Ti ∈ T , i = 1, ..., n, into (R(T1 , ..., Tn )) = Spec[T1 /q1 , ..., Tn /qn ],
where Spec[T1 /q1 , ..., Tn /qn ] is an object type specification produced by repla-
cing each qi in Spec with Ti .
A pair hR(q1 , ..., qn ), Speci is called a generic type specification. The repla-
cement of type parameters with type names in both parts of a generic type
specification is called a generic type instantiation. Note that due to the use of
the function into , we do not need to introduce a special semantics for generic ob-
ject types. A generic type specification in this approach is just a way of defining
a part of this function. This corresponds one to one to the practice of modern
programming languages (like C++) regarding generic object types as templates.
Two type terms R(T11 , ..., T1n ) and R(T21 , ..., T2n ) are equivalent if T1i and
T2i , i = 1, ..., n, are the same type name.
Example. Taking into account that in the real memory we would like to have
locations storing values of different types, we specify a generic vector iterator
type:

class VecIt(T) = spec

[attribute value stored: T;
mutator put value: T;
observer get value: T;
advance, retreat: VecIt(T);
plus, minus: Nat −→ VecIt(T);
difference: VecIt(T) −→ Nat;
eq, neq, less, greater, leq, geq: VecIt(T) −→ Boolean]
{forall x: T, i, i1: VecIt(T), n: Nat. – declaration of universally quantified variables
i.put value(x) == i.value stored := x; – dynamic axiom
i.get value == i.value stored; – this one and all the others are data axioms
100 A.V. Zamulin

i.advance == nat to id(id to nat(i) + 1);

i.retreat == nat to id(id to nat(i) – 1);
i.plus(n) == nat to id(id to nat(i) + n);
i.minus(n) == nat to id(id to nat(i) – n);
i.difference(i1) == id to nat(i) – id to nat(i1)
i.less(i1) == id to nat(i) < id to nat(i1);
i.greater(i1) == id to nat(i) > id to nat(i1);
i.leq(i1) == id to nat(i) ≤ id to nat(i1);
i.geq(i1) == id to nat(i) ≥ id to nat(i1)}.

4 Generic Vector Type

We can now give a formal specification of the generic vector type informally defi-
ned in [4]. It is assumed that vector components (which are vector iterators) are
numbered starting from zero. The allocated number of the vector components is
defined by the attribute max size, the current number of initialized components
is defined by the attribute size. Vector components in the range 0..size − 1 are
called vector elements in the sequel. The clause dom in the specification in-
dicates the domain of a partial function: in a domain specification dom t : b,
the term t is defined if and only if b evaluates to true. A transition rule of the
form set R1 , ..., Rn end indicates the parallel execution of the transition terms
R1 , ..., Rn .
class Vector(T) = spec
[attribute comp: Nat −→ VecIt(T); – vector components, allocated by an allocator
size: Nat; – the current size of a vector, initialized by an allocator
max size: Nat; – the maximum size of a vector, initialized by an allocator
mutator empty vec; – default constructor
initialized vec: Nat, T; – initialization of the content of the first n components
copy: Vector(T); – copy constructor
push back: T; – append an element at the end of a vector
pop back; – delete the last element of a vector
insert: VecIt(T), T; – insert an element at the position indicated
erase: VecIt(T); – remove the element indicated
swap: Vector(T); - swap the contents of two vectors
observer empty: Boolean; – is a vector empty?
[ ]: Nat −→ T; – fetch an element of a vector
front, back: T; – the first and last elements of a vector
begin, end: VecIt(T)] – the starting and terminating iterators of a vector
{forall x: T, iv: VecIt(T), n: Nat, v, v1: Vector(T).
dom v[n]: n ≥ 0 & n ≤ v.size;
dom v.initialized vec (n, x): n ≤ v.max size;
dom v.copy(v1): v.max size ≥ v1.size;
dom v.push back(x): v.size < v.max size;
dom v.insert(iv, x): v.size < v.max size & iv.geq(v.begin) & iv.less(v.end);
dom v.erase(iv): iv.geq(v.begin) & iv.less(v.end) & v.size > 0;
 Generic Facilities in Object-Oriented ASMs 101

v.empty vec == v.size := 0;

v.initialized vec (n, x) == set v.size := n,
forall i: 0 .. n–1. v.comp(i).value stored := x end;
v.copy(v1) == set v.size := v1.size,
forall i: 0 .. v1.size–1. v.comp(i).value stored := v1.comp(i).value stored end;
v.push back(x) == set v.comp(v.size).value stored := x, v.size := v.size + 1 end;
v.pop back == v.size := v.size – 1;
v.insert(iv, x) == set v.size := v.size + 1,
forall i: 0 .. v.size-1. if v.comp(i).geq(iv)
then v.comp(i+1).value stored := v.comp(i).value stored,
iv.value stored := x end;
v.erase(iv) == set v.size := v.size – 1,
forall i: 0 .. v.size-1. if v.comp(i).geq(iv)
then v.comp(i).value stored := v.comp(i+1).value stored end;
v.empty == v.size =0;
v[n] == v.comp(n).value stored;
v.begin := v.comp(0);
v.end == v.begin.plus(v.size);
v.front == v.comp(0).value stored;
v.back == v.comp(v.size-1).value stored }.

5 Generic List Type

A list is a double-linked sequence of elements ordered according to the use of
the constructor operations ”push back” (inserts an element at the end of a list),
”push front” (inserts an element at the beginning of a list), and ”insert” (inserts
an element in the middle of a list). All list elements are numbered starting with
one for the first element. The number of the last element is equal to the number
of list elements. A list element is an object of the corresponding iterator type
ListIt. We define it first.
class ListIt(T) = spec
[attribute value stored: T;
pred, next: ListIt(T);
mutator put value: T;
observer get value: T;
advance, retreat: ListIt(T);
eq, neq: ListIt(T) −→ Boolean]
{forall i: ListIt(T), x: T.
i.put value(x) == i.value stored := x;
i.get value == i.value stored;
i.advance == i.next;
i.retreat == i.pred}.
Note that, in contrast to the object type VecIt, this object type does not
possess the methods plus and minus since list elements are accessed only se-
quentially.
102 A.V. Zamulin

Now we can define the generic list type. Due to space limitations, the spe-
cification of some mutators is left to the reader. Note that a transition rule of
the form seq R1 , ..., Rn end indicates the sequential execution of the transition
terms R1 , ..., Rn , and the transition rule of the form while b do R end indicates
the repetition of the execution of the transition term R. The formal semantics
of these rules can be found in [3].

class List(T) = spec

[attribute begin, end: ListIt(T); – the starting and terminating iterators of a list
size: Nat; – current size of a list
mutator empty list; – empty list constructor
initialized list: Nat, T; – initialized list constructor
copy: List(T); – copy constructor
push front: T; – insert an element at the beginning of a list
push back: T; – insert an element at the end of a list
pop front; – delete the first element of a list
pop back; – delete the last element of a list
insert1: ListIt(T), T; – insert one element at the position indicated
insertN: ListIt(T), Nat, T; – insert N elements at the position indicated
erase1: ListIt(T); – remove the element indicated
eraseN: ListIt(T), ListIt(T); – remove the elements between the iterators
observer empty: Boolean; – is a list empty?
front, back: T; – first and last elements of a list
has: ListIt(T) −→ Boolean; – an auxiliary operation checking the presence
– of an element in the list
precedes: ListIt(T), ListIt(T) −→ Boolean] – an auxiliary operation verifying
– that the first iterator precedes the second one in a list
{forall x: T, n: Nat, l, l1: List(T), i, i1: ListIt(T).
dom l.pop front: l.size > 0; dom l.pop back: l.size > 0;
dom l.insert1(i, x): l.size > 0 & ¬l.has(i);
dom l.insertN(i, n, x): l.size > 0 & ¬l.has(i);
dom l.eraseN(i, i1): l.size > 0 & l.has(i) & l.has(i1) & l.precedes(i, i1);
dom l.front: l.size > 0; dom l.back: l.size > 0;
l.empty list == import new elem: ListIt(T) in
set l.begin := new elem, l.end := new elem, l.size := 0 end ;
– an iterator beyond the list component iterators is created; according to
- STL, it is the terminating list iterator
l.push back(x) == import new elem: ListIt(T) in
set new elem.pred := l.end, l.end.next := new elem, l.end.value stored := x,
l.end := new elem, l.size := l.size + 1 end ;
l.initialized list(n, x) == seq l.empty list, while l.size < n do l.push back(x) end;
l.copy(l1) == seq sequence of transition rules end;
l.push front(x) == import new elem: ListIt(T) in set of transition rules;
l.pop front == set l.begin := l.begin.next, l.begin.next.pred := undef,
size := size – 1 end;
l.pop back == set of transition rules;
 Generic Facilities in Object-Oriented ASMs 103

l.insert1(i, x) == import new: ListIt(T) in set i.next := new, i.next.pred := new,

new.pred := i, new.next := i.next, new.value stored := x, l.size := l.size + 1 end;
l.insertN(i, n, x) == seq l.insert1(i, x), if n > 1 then l.insertN(i.next, n–1, x) end ;
l.erase1(i) == if l.size = 1 then l.pop front – first element is deleted
else set i.pred.next := i.next, i.next.pred := i.pred, l.size := l.size – 1 end ;
l.eraseN(i, i1) == set l.erase1(i), if i.next.neq(i1) then l.eraseN(i.next, i1) end ;
l.front == l.begin.value stored; l.back == l.end.pred.value stored;
l.has(i) == Boolean term evaluating to ”true” if l has i;
l.precede(i, i1) == Boolean term evaluating to ”true” if i precedes i1 in l}.

6 Object Type Categories

The generic constructs defined above permit us to specify algorithms over data
structures, such as vectors or lists, abstracted from the type of the component
of the structure. One of the STL requirements is the ability to specify also
algorithms abstracted from the data structure itself, i.e. to specify an algorithm
capable, for example, to manipulate both vectors and lists. Object type categories
serve this purpose.
An object type category resembles a sort class of Spectrum [5] or type class
of Ruslan [6] both based on type classes introduced in [8] and sypes introduced
in [9]. It defines a set of object type specifications with some common properties.
Let CAT be a set of names, CS a set of specifications constructed like object
type specifications with the use of an extra object type name ”@”, and intc :
CAT −→ CS a function mapping names in CAT to specifications in CS. If
C ∈ CAT , cs ∈ CS, and cs = intc (C), then the maplet hC 7→ csi is the
specification of the object type category C.
Let hC 7→ csi be the specification of an object type category and hT 7→
otsi the specification of an object type. It is said that the object type T is a
type of the category C (or T belongs to C) if cs[T /@] ∈ ots, where cs[T /@] is
the specification cs with the symbol ”@” replaced with T . That is, an object
type belonging to a certain type category must include its specification as a
subspecification. Example:

classcat Equal = spec – category of object types with an equality operation

[observer eq, neq: @ −→ Boolean]
{forall x, y: @, exist z: @.
x.eq(x) == true; x.eq(y) == y.eq(x);
x.eq(z) & z.eq(y) == x.eq(y);
x.neq(y) == ¬x.eq(y)}.

Any object type possessing the methods eq and neq specified as above is
the type of the category Equal. We also consider that T Y P E is the name of
the category with the empty specification and, therefore, any data/object type
belongs to this category.
Constructing a type category, we can inherit the specification of an existing
type category producing the union of the specifications as result.
104 A.V. Zamulin

Like an object type, an object type category can be generic, i.e., it can
use type parameters in its specification. The definition of a generic object type
category is the same as the definition of a generic object type. A generic object
type belongs to a generic type category if their parameters match and it includes
the specification of the type category as its subspecification. Iterator categories
serve as examples.
1. Each iterator type of the following category has methods advance and
get value in addition to the methods of the category ”Equal”.2

classcat InputIterator(T) = spec Equal

[observer advance: @; – produces the succeeding iterator
get value: T] – reads the value stored in an iterator

2. Each iterator type of the following category has methods advance and
put value in addition to the methods of the category ”Equal”.3

classcat OutputIterator(T) spec Equal

[observer advance: @; – produces the succeeding iterator
mutator put value: T] – stores a new value in an iterator

3. Each iterator type of the following category has a mutator put value in
addition to the methods of the category ”InputIterator”.

classcat ForwardIterator(T) = spec InputIterator(T)

[mutator put value: T]
{forall i: @, x: T.
i.put value(x) == i.value stored := x;
i.get value == i.value stored}

4. Each iterator type of the following category has the method retreat in
addition to the methods of the category ”ForwardIterator”.

classcat BidirectionalIterator(T) = spec ForwardIterator(T)

[observer retreat: @] - produces the preceding iterator

5. Each iterator type of the following category has several methods in addition
to the methods of the category ”BidirectionalIterator”.

classcat RandomAccessIterator(T) = spec BidirectionalIterator(T)

[observer plus, minus: Nat −→ @;
difference: @ −→ Nat;
less, greater, leq, geq: @ −→ Boolean]
2
The category of input iterators is introduced in STL to allow iteration over input
streams in the same way as, say, over vectors.
3
The category of output iterators is introduced in STL to allow iteration over output
streams in the same way as, say, over vectors.
 Generic Facilities in Object-Oriented ASMs 105

According to the definitions, an object type VecIt(T) belongs to the type cate-
gories RandomAccessIterator(T), BidirectionalIterator(T), ForwardIterator(T),
OutputIterator(T), InputIterator(T), and Equal. An object type ListIt(T) be-
longs to the type categories BidirectionalIterator(T), ForwardIterator(T), Ou-
tputIterator(T), InputIterator(T), and Equal (it does not belong to the type
category RandomAccessIterator(T), however). Thus, a vector iterator can be
used in any algorithm requiring either a random access iterator or bidirectional
iterator or forward iterator or input iterator or output iterator. In the same way
a list iterator can be used in any algorithm except that one which requires a
random access iterator.

7 Constrained Genericity

According to the definitions of generic components in Section 3, any type can be

used as instantiation argument. At the same time, it is often needed that only
a type belonging to a certain type category could be substituted. Therefore, the
definitions of generic components should be changed according to this constraint.
The constrained generic object type is defined as follows.
Let q1 : C1 , ..., qk : Ck be names (of type parameters) indexed with type
category names. A pair hR(q1 : C1 , ..., qk : Ck ), Speci, where R ∈ R and Spec
is an object type specification additionally using q1 , ..., qk as type terms in me-
thod declarations and operators from C1 , ..., Ck in axioms, is part of the fun-
ction into , such that for any type term Ti of type category Ci , i = 1, ..., k,
into (R(T1 , ..., Tk )) = Spec[T1 /q1 , ..., Tk /qk ], where Spec[T1 /q1 , ..., Tk /qk ] is an
object type specification produced by replacing each qi in Spec with Ti .
The constrained generic type category is defined in a similar way.
A generic function declaration is a triple hgf, (q1 : C1 , ..., qk : Ck ), F P q i,
where gf is a (generic) function name, q1 : C1 , ..., qk : Ck are names (of type
parameters) indexed with type category names and F P q is a function profile
additionally using q1 , ..., qk as type terms.
If gf : h(q1 : C1 , ..., qk : Ck ), F P q i is a generic function declaration and
T1 , ...., Tk are type terms of type categories C1 , ..., Ck , respectively, then
gf (T1 , ..., Tk ) : F P is an instantiated function declaration, where F P is a fun-
ction profile produced by replacing each qi in F P q with Ti ; gf (T1 , ..., Tk ) is
called an instantiated function name. Instantiated function names are used for
producing data terms in the same way as ordinary function names are used.
If gf : h(q1 : C1 , ..., qk : Ck ), F P q i is the declaration of a dynamic function,
then an algebra A of the given signature is provided with a function mapAgf binding
an instantiated function name gf (T1 , ..., Tk ) to a function as it is described in
Section 2.2 for dynamic functions.
If gf : h(q1 : C1 , ..., qk : Ck ), F P q i is the declaration of a dependant function,
DS(A )
then a dynamic system DS(A0 ) is provided with a function mapgf 0 binding
an instantiated function name gf (T1 , ..., Tk ) to a function as it is described in
Section 2.2 for dependant functions.
106 A.V. Zamulin

A generic function specification consists of a generic function declaration and

a set of axioms using instantiated function names in their terms.
Notation: we use the brackets gen ... profile to embrace type parameters
in a generic function declaration.
Examples:
1. Specify a function which looks for a value in a data structure between
f irst and last iterators and returns the iterator storing the value if it is found
and the last iterator if the value is not found.

function find: gen I: InputIterator, T: TYPE profile: I(T), I(T), T −→ I(T);

{forall first, last: I(T), value: T.
find(I, T)(first, last, value) ==
if first.get value = value ∨ first.eq(last) then first
else find(I, T)(first.advance, last, value)}.

Now, if we have the following declarations:

vec: Vector(Integer);
list: List(Char);
we can invoke the function in the folowing ways:
find(VecIt, Nat)(vec.begin, vec.end, 7);
find(ListIt, Char)(list.begin, list.end, ’a’);
In this case both vector iterators and list iterators can be used because both
belong to the category of input iterators required in the function specification.
Note the substitution of the generic types V ecIt and ListIt for the generic type
category I. In the next example list iterators cannot be used.
2. Specify a function which performs binary search in a structure containing
ordered components and returns the iterator containing the element to be found
(for simplicity we assume that the structure really contains the element).

function binary find: gen I: RandomAccessIterator, T: Ordered

profile I(T), I(T), T −→ I(T);
{forall first, last: I(T), value: T.
binary find(I, T)(first, last, value) == let d = last.difference(first), h = d/2,
current = first.plus(h), cv = current.get value in
if cv = value then current
elseif value < cv then binary find(I, T)(first, current, value)
else binary find(I, T)(current, last, value)}.

Now, we can call the function with vector iterators like the following:
binary find(VectIt, Nat)(vec.begin, vec.end, 7)
because vector iterators belong to the class RandomAccessIterator, and we can-
not call it with list iterators. Note the use of the type category Ordered (not
defined here) which is needed to make sure that the type of the components
contains the operation ”<”.
The above examples show us in which way the functions abstracting both
from the structure and the type of its components can be specified.
 Generic Facilities in Object-Oriented ASMs 107

Generic procedures are declared similarly to generic functions and are speci-
fied similarly to ordinary procedures. The vector allocator, a special procedure
which allocates memory for a particular vector (according to STL, such a pro-
cedure is associated with each container class), can serve as an example. The
import rule of Gurevich ASMs [2] is used in the specification for creating a new
object identifier.

proc vec allocator: gen T: TYPE profile Vector(T), Nat;

{forall v: Vector(T), n: Nat.
vec allocator(T)(v, n) ==
set v.max size := n, v.size := 0,
forall i = 0 .. n-1. import new elem: Vector(T) in v.comp(i) := new elem
end – vector attributes are initialized.

If intvec is a dynamic constant of type V ector(Integer), then vec allocator(

Integer)(intvec, 100) will allocate 100 noninitialized locations to intvec and set
initial values to its attributes max size and size.

8 Related Work
We are not going to discuss here the approaches representing object states as
elements of the same algebra. The work along this approach is heavily based on
traditional algebraic specification methods. We can only repeat after F. Parisi-
Presicce and A. Pierantonio that ”the algebraic framework so far has been ina-
dequate in describing the dynamic properties of objects and their state transfor-
mation as well as more complex notions typical of the object oriented paradigm
such as object identity and persistency of objects” [10]. The interested reader
can refer to [11,12,13,14].
We review here several works considering object states as algebras. Object-
oriented extensions of the prominent specification methods VDM [15] and Z [16]
are the first of them.
VDM++ [17] introduces a notion of class definition as a template for a collec-
tion of objects possessing a number of instance variables (internal state) and
methods (external protocol). The definitions of the methods of existing classes
can be inherited when a new class is defined (multiple inheritance). Object’s
initial state and invariant can be specified. A set of statements typical of impe-
rative programming languages is provided. Unfortunately, the description of the
semantics of the language is done rather informally, in the way the semantics
of programming languages is usually described. As a result, the user gets an
impression that VDM++ is a programming language provided with some speci-
fication facilities (pre- and post-conditions) rather than a specification language.
No genericity is provided in the language.
Object-Z [18] practically has the same features as VDM++ with the dif-
ference that it is based on Z. A class is here a set of attributes and a set of
operations acting upon these attributes. In contrast to VDM++, the semantics
108 A.V. Zamulin

of Object-Z is formally given. The state is considered as a function from a set

of identifiers (attributes) to the set of all possible values. One can say that it
corresponds to a homogeneous algebra whose signature contains only constant
symbols (compare it with a more general case of ASM where functions as alge-
bra components play a significant role). A class definition can be supplied with
a number of type parameters, a kind of unconstrained genericity is provided in
this way. No notion of class category and, respectively, constrained genericity
exists in the language. Object creation is also not provided by the language.
Therefore a specification similar to the specification of list types given above is
not possible.
Z++ [19,20] is another development based on Z providing facilities compara-
ble to those of Object-Z. The main difference is that its syntax is quite different
from that of Object-Z (which actually follows the syntax of Z) and is chosen to
stress the commonalty of the language with object-oriented programming langu-
ages like Eiffel. Formal semantics of the language is defined by means of category
theory. However, the main attention is paid to the formal definition of refine-
ments between classes while such important notions as state, class, object, etc.
are considered well-known and not defined in the semantics. Like in Object-Z, a
class definition in Z++ can be supplied with a number of type parameters provi-
ding unconstrained genericity. Again, no notion of class category (in our sense)
and, respectively, constrained genericity exists in the language. The semantics
of a generic class specification is also not reported.
In [21] objects are elements of instance structures which are quadruples of
algebras of different signatures. Specifications of the algebras resemble traditio-
nal algebraic specifications. One of the algebras is extended with extra ”state
function symbols” mapping object identifiers to their values. Dynamic operati-
ons serving for object evolution are modeled by algebra morphisms. The author
believes that the specification of these operation should have an imperative na-
ture, but he does not suggest a method of specification. The approach is further
formalized with a heavy use of category theory in [10]. In contrast to all of this,
we represent the state by a single algebra, we believe that there is no neces-
sity for state function symbols since user-defined observers perfectly serve for
this purpose, and we suggest a concrete method of specification by means of
transition rules.
The “Hidden Sorted Algebra” approach [11], where some sorts are distinguis-
hed as hidden and some other as visible, treats states as values of hidden sorts.
Visible sorts are used to represent values which can be observed in a given state.
States are explicitly described in the specification in contrast to our approach.
This work combined with Meseguer’s rewriting logic [22] has served as the basis
of the dynamic aspects of the language CafeOBJ [23]. In this language states
and transitions are modeled, respectively, as objects and arrows belonging to the
same rewrite model which is a categorical extension of the algebraic structure.
Meseguer’s rewriting logic is also the basis of the specification language Maude
[24].
 Generic Facilities in Object-Oriented ASMs 109

The specification language Troll [25] should be mentioned as one of the main
practical achievements in the field. Troll is oriented to the specification of ob-
jects where a method (event) is specified by means of evaluation rules similar
to equations on attributes. Although the semantics of Troll is given rather in-
formally, there is a strong mathematical foundation of its dialect Troll-light [26]
with the use of data algebras, attribute algebras and event algebras. A relation
constructed on two sets of attribute algebras and a set of event algebra, called
object community, formalizes transitions from one attribute algebra to another.
Although Troll possesses generic facilities, non of them is formalized in [26].
The problem of constrained genericity in object-oriented class libraries is di-
scussed in [27]. Abstract classes resembling our object type categories are used
there for constraining the generic parameters, and a notion of syntactical con-
formance is introduced for checking whether a concrete class is an instance of a
given abstract class. There is an important difference between an abstract class
and a concrete class since one cannot create an object of an abstract class. In
fact, an abstract class specifies some methods common for a community of con-
crete classes. For this reason me make a clear difference between object types
and object type categories independently specified. An additional feature of our
approach is generic object type categories standing for communities of generic
object types.
Finally, a fundamental work [28] formalizing bounded parametric polymor-
phism similar to our constrained genericity should be paid attention. Here gene-
ricity is constrained by allowing only those type arguments which are subtypes
of the parameter type. In this respect, this approach is very similar to the pre-
vious one. Another peculiarity of the work is that an object does not possess a
unique identifier, it is just a tuple of methods, and object updates are simulated
by method overrides generally producing new objects.

9 Conclusion

The mechanisms for the specification of generic object types and type categories
are introduced in the paper. With the use of these mechanisms, many generic
algorithms abstracting from the type of the data structure being manipulated can
be easily specified. Although the general technique of generic type specification
is well-known and can be found in the literature, the novelty of this work is the
extension of the technique to the case of object-oriented ASMs. Another novelty
is the definition and use of generic type categories for the specification of generic
algorithms.
The described technique has permitted us to specify some components of the
Standard Template Library for C++. The library thus specified can be easily
adapted to another object-oriented language. The experience obtained in the
process of the specification has proved the power of the technique. Its main
features can be summarized as follows:
1. We represent immutable values by data types and specify them algebrai-
cally.
110 A.V. Zamulin

2. We represent mutable objects possessing states by object types and specify

them by means of transition rules.
3. We define generic (data, object) types to abstract from the type of the
component.
4. We define (data, object) type categories to abstract from the structure.
5. We define a generic algorithm by means of transition rules manipulating
the objects thus specified.
Tools supporting this style of specification remain the subject of further work.

Acknowledgements. I thank anonymous referees for their helpful comments.

References
1. Y. Gurevich. Evolving Algebras 1993: Lipary Guide. In: Börger, E., ed., Specifica-
tion and Validation Methods, Oxford University Press, 1995, pp. 9-36.
2. Y. Gurevich. May 1997 Draft of the ASM Guide. University of Michigan EECS
Departmental Technical Report CSE-TR-336-97 (available electronically from
http://www.eecs.umich.edu/gasm/).
3. A.V. Zamulin. Object-Oriented Abstract State Machines. Proc. ASM workshop,
Magderburg, Germany, 21-22 September, 1998, pp. 1-21 (available electronically
from http://www.eecs.umich.edu/gasm/).
4. D.R. Musser and A. Saini. STL Tutorial and Reference Guide. Addison-Wesley,
1996.
5. Broy, M., Facchi, C., Grosu, R., ea. The Requirement and Design Specification Lan-
guage Spectrum, An Informal Introduction, Version 1.0. Technische Universitaet
Muenchen, Institut fuer Informatik, April 1993.
6. A.V. Zamulin, The Database Specification Language RUSLAN: Main Features.
East-West Database Workshop (proc. Second International East-West Database
Workshop, Klagenfurt, Austria, September 25-28, 1994), Springer (Workshops in
Computing), 1994, 315-327.
7. P.D. Mosses. CoFI: The Common framework Initiative for Algebraic Specification
and Development. In: M. Bidoit and M. Dauchet, eds., TAPSOFT’97: Theory and
Practice of Software Development, LNCS, vol. 1214, pp. 115-137.
8. P. Wadler and S. Blott. How to make ad-hoc polymorphism less ad-hoc. Conf.
Record of the 16th ACM Annual Symp. on Principles of Progr. Lang., Austin,
Texas, January 1989.
9. Nakajima, R., Honda, M., and Nakahara, H. Hierarchical Program Specification: a
Many-sorted Logical Approach. Acta Informatica 14, pp. 135-155 (1980).
10. F. Parisi-Presicce and A. Pierantonio. Dynamic behavior of Object Systems. In:
Recent trends in Data Type Specification. LNCS, vol. 906, 1995, pp. 406-419.
11. J.A. Goguen and R. Diaconescu. Towards an Algebraic Semantics for the Object
Paradigm. Recent Trends in Data Type Specification, LNCS, 1994, vol. 785, pp.
1-29.
12. H.-D. Ehrig and A. Sernadas. Local Specification of Distributed Families of Sequen-
tial Objects. In: Recent Trends in Data Type Specifications. LNCS, vol. 906, 1994,
pp. 219-235.
13. F. Parisi-Presicce and A. Pierantonio. An Algebraic Theory of Class Specification.
ACM Transactions on Software Engineering and Methodology, April 1994, vol. 3,
No. 2, pp. 166-169.
 Generic Facilities in Object-Oriented ASMs 111

14. C. Cristea. Coalgebra Semantics for Hidden Algebra: Parameterised objects and
Inheritance. in: Recent Trends in Algebraic development Techniques. LNCS, vol.
1374, 1997, pp. 174-189.
15. C. B. Jones. Systematic Software Development using VDM. Prentice Hall, 1990.
16. J. M. Spivey. understanding Z. A specification language and its formal semantics.
Cambridge University Press, 1988.
17. E.H. D,rr and J. van Katwijk. A Formal Specification Language for Object Oriented
Designs. In: Computer Systems and Engineering (Proceedings of CompEuro’92).
IEEE Compute Society Press, 1992, pp. 214-219.
18. R. Duke, P. King, G.A. Rose, and G. Smith. The Object-Z specification language.
In: T. Korson, V. Vaishnavi, and B. Meyer, eds., Technology of Object-Oriented
Languages and Systems: TOOLS 5, Prentice hall, 1991, pp. 465-483.
19. K. Lano and H. Houghton, eds., Object-Oriented Specification case Studies. Pren-
tice Hall (object-oriented series), 1994.
20. K. Lano and H. Houghton. The Z++ manual. October 1994. Available from
ftp://theory.doc.ic.ac.uk/theory/papers/Lano/z++.ps.Z
21. A. Pierantonio. Making Statics Dynamic. In: G. Hommel, editor, Proc. Interna-
tional Workshop on Communication based Systems, Kluwer Academic Publishers,
1995, pp. 19-34.
22. J. Meseguer. Conditional rewriting logic as a unified model of concurrency. Theo-
retical Computer Science, vol. 96, No 1 (April 1992), pp. 73–155.
23. R. Diaconescu and K. Futatsugi. CafeOBJ Report.World Scientific Publishing Co.
Pte. Ltd, AMAST series in Computing”, vol. 6,1998.
24. J. Meseguer. A logical theory of concurrent objects and its realization in the Mode
language. In: Research Directions in Concurrent Object-Oriented Programming.
The MIT Press, Cambridge, Mass., 1993, pp. 314-390.
25. T. Hartmann, G. Saake, R. Jungclaus, P. Hartel, and J. Kush. Revised Ver-
sion of the Modelling Language TROLL. Technishe Universitaet Braunschweig,
Informatik-Berichte 94-03, 1994.
26. M. Gogolla and R. Herzig. An Algebraic Semantics for the Object Specification
Language TROLL-light. In: Recent Trends in Data Type Specifications, LNCS,
vol. 906, 1995, pp. 290–306.
27. A. Frick, G. Goos, R. Newmann, W. Zimmermann. Construction of Robust Class
Hierarchies. Software—Practice and Experience, 2000 (to be published).
28. M. Abadi and L. Cardelli. A Theory of Objects. Springer-Verlag, 1996.
 Towards an ASM Thesis for Unconventional
Algorithms

Wolfgang Reisig

Humboldt-Universität zu Berlin
[email protected]

Abstract. All descriptions of algorithms, be they formal or informal,

employ data structures, operations on them, and some policy to cause
operations be applied to data.
Gurevich calls a formal description technique for algorithms algorithm
universal if it allows for each informally described algorithm a formal
representation that would essentially make precise the notions used in
the informal description, not employing additional data, operations or
steps.
Gurevich’s ASM thesis claims Abstract State Machines be algorithm
universal for conventional, sequential algorithms.
Here we are behind properties of formal presentations that are algorithm
universal for unconventional, distributed algorithms.

1 Conventional and Unconventional Algorithms

1.1 Algorithmic Ideas
The design of a computer embedded solution to a real world problem always in-
cludes phases of algorithmic ideas. An algorithmic idea describes intuitive, tenta-
tive, not yet fully specified aspects of an intended algorithm. Though inevitably
vague, an algorithmic idea is intended to be transformed into an unambiguous,
formally presented algorithm.
Experience reveals that all algorithmic ideas cover three kinds of components:
First, a choice of passive components. This includes data items in various roles
(e.g. data stored in data bases, data in a transient buffer, or data amenable for
processing), variables with their respective actual values, and potential location
of control items.
Second, a choice of active components. This includes operations to observe
or to change (update) passive components, e.g. reading a variable, writing a
variable, establishing new control items, transporting data items to other places
(as e.g. in case of input and output), providing new roles or contexts to data
items.
Third, a policy to activate active components and to cause their action. This
includes sequential policies that cause sequences of action occurences, as well as
parallel policies that cause sequences of sets of action occurrences, and distributed
policies, that cause actions to occur in partially unspecified order.

Y. Gurevich et al. (Eds.): ASM 2000, LNCS 1912, pp. 112–130, 2000.
c Springer-Verlag Berlin Heidelberg 2000

 Towards an ASM Thesis for Unconventional Algorithms 113

1.2 Conventional Algorithms

We denote an algorithm conventional if it sticks to the conventional paradigm
of computing: Upon all input available, the algorithm starts. Elementary actions
are executed in sequence, conditional alternative, and conditional iteration. Runs
should terminate and provide output. Non-terminating runs are considered er-
rornous, providing no output. Such an algorithm computes a partial function
from the set of potential inputs to the set of potential outputs.
The intermediate choice of actions may not be unique, in which case one of
them is selected nondeterministically. The algorithm computes a relation in this
case.

1.3 Reactive Algorithms

As a first liberalization of the conventional paradigm, not all input is required to
be initially available, nor is all output postponed until termination: Input and
output may occur during a run, as outlined in Fig. 1.

Fig. 1. Outline of a run of a reactive algorithm

This yields a number of consequences: Firstly, a run may wait for fresh input.
‘Waiting’ is a really new aspect, not present in any conventional theories of com-
putation. Secondly, a run of such an algorithm is sensible also if it diverges.
Different diverging runs may execute different interactions with their environ-
ment.
Each run of this kind of algorithm transforms a finite or infinite sequence of
input data into a finite or infinite sequence of output data: The algorithm reacts
to each new input item with fresh output items; hence the term of reactive
algorithms.
Typical examples of such algorithms are technical control algorithms, such
as e.g. lift control.

1.4 Cooperating Algorithms

Two reactive algorithms A and B may cooperate, with output of A used as
input for B and vice versa, output of B used as input for A. In addition, the two
algorithms may cooperate with their environment. Fig.2 outlines this situation.
Typical examples of such algorithms are communication protocols.
114 W. Reisig

Fig. 2. Outline of a run of cooperating algorithms

1.5 Network Algorithms

The idea of cooperating algorithms can be generalized to networks of local algo-

rithms. Each local algorithm in such a network is directly linked to its neighbor
algorithms, as outlined in Fig.3.

Fig. 3. Network of local algorithms A, . . . , E

They may perform a task in cooperation, sending messages to neighbors, as

sketched in Fig.4.

1.6 Shared Memory and Threads on-the-Flay

A shared memory algorithm consists of two or more threads of actions, performed

on storage private to the threads, as well as on common storage. Threads of
control may be generated as well as destroyed during an algorithm’s run.
 Towards an ASM Thesis for Unconventional Algorithms 115

Fig. 4. Outline of a run of the network in Fig.3

Fig. 5. Regularly structured threads of control

Fig. 5 shows a regularly structured split of thread of control.In contrast, not

all parts of thread of control can uniquely be assigned to a pair of fork and join
in Fig.6.

Fig. 6. Irregularly structured threads of control

Fork and join may dynamically be generated during a run of an algorithm.

The overall number of parallel threads thus is not fixed or predictable before-
hand. The notion of determined thread of control fades away.
116 W. Reisig

1.7 Liberal Policies of Action

Some versions of unconventional algorithms do with quite liberal policies of
action. As an - admittedly extreme - example, consider the set of symbols
{a, b, c, d, e} with the standard alphabetic order, <, and the set {1, . . . , 5} of
indices with their natural order. Furthermore, assume an initial pairing of let-
ters and indices, e.g. (a, 5), (b, 1), (c, 3), (d, 2), (e, 4). Our task is to order the
symbols properly, resulting in the pairs (a, 1), . . . (e, 5). The only available opera-
tion, swap ((x, i), (y, j)), is defined for letters x, y and indices i, j. This operation
is activated if x > y and i < j, in which case it returns (x, j) and (y, i).
The crucial point is the policy of action. We describe it intuitively by con-
ceiving each pair (x, i) as a fish, x, with a sticker, i.

Fig. 7. Outline of distributed sorting

The fishes are swimming around in an aquarium, as sketched in Fig.7. Whe-

never two fish meet, they exchange their stickers in case the order of fishes and
stickers do not coincide.
Is this an algorithm at all? What minimal requirements guarantee termina-
tion? How detect termination?

1.8 Progress and Quiescence

Conventional algorithms are always to proceed. For example, execution of a PAS-
CAL program does not stall with an enabled statement: all statements of PAS-
CAL are progressing. Unconventional algorithms may have actions that are not
progressing, but quiescent: they may remain enabled forever.
As an example, any reasonable algorithm that organizes mutual exclusion
(mutex) must do with at least three states for each involved site. Calling those
states quiet, pending, and critical, a quiet site must spontaneously be able to
 Towards an ASM Thesis for Unconventional Algorithms 117

become pending, but must also be able to remain quiet forever; the step from
quiet to pending is “quiescent”. The algorithm furthermore must guarantee that
pending will eventually lead to critical and back to quiet. Hence all other actions
are “progressing”.
Algorithms are frequently composed with other algorithms along quiescent
actions. Quiescence then hints at additional preconditions that are not explicitely
stated in the algorithm itself.

1.9 Fairness Assumptions

Nondeterministic choice among two or more actions can be required to be per-

formed fairly: a run is discarded, if it infinitely often is to decide the same
alternative and decides always in favor of the same action (up to finitely many
exceptions).
Fairness assumptions are irrevelant for conventional algorithms. Fairness as-
sumptions leave the scope of computable functions; they can be verified or fal-
sified only with infinite behavior. Fairness can not be implemented literally, one
usually implements a stronger assumption, ruling out more runs.
As an example, all mutex algorithms need a fairness assumption [11]. This
assumption is frequently shifted into fair reading of variables. Then the debate of
forthcoming Chapter 2 applies: If fair reading of shared variables is guaranteed,
the problem is solved.

1.10 Schemata for Algorithms

Algorithms frequently don’t determine all aspects of computation. The most

prominent parameter for computation is input. Unconventional algorithms do
with further parameters.
For example, the Echo algorithm is a schema to broadcast acknowledged
messages within connected, undirected networks of agents and message lines.
Algorithmic behavior, concurrent runs etc. manifests in single networks only.
The Echo algorithm is just a schema for many algorithms: each concrete graph
defines a specific algorithm.

1.11 Specification of Algorithms

A specification of an algorithm does not formulate operational behavior, but

just states its crucial properties. For example, each mutex algorithm assumes
two partners, l and r, with states quiet, pending and critical, and the steps as
discussed in 1.8. In addition, never both partners are coincidently critical. Any
algorithm that meets those requirements, is a perfect mutex algorithm.
More involved specifications do with requirements that guarantee distinguis-
hed behavior only if they can rely on proper input.
118 W. Reisig

1.12 Algorithm Universal Formalisms

A formal presentation may stick more or less tightly to an algorithmic idea.

The most liberal fashion would suggest that any algorithmic idea essentially
describes a computable function. A formal presentation of the algorithmic idea
then formalizes this function by any mathematical means. The tightmost fashion
was a formal presentation that would give the passive and active components
of the algorithmic idea a precise meaning, avoiding any new components, and
likewise would reflect the algorithmic ideas’ activation policy in such a way that
the steps of the formal presentation bijectively correspond to the steps of the
algorithmic idea. Gurevich in [8] calls such a formal description technique for
algorithms algorithm universal.
The sequential ASM thesis of [8] claims that sequential ASMs are algorithm
universal for conventional algorithms, as described in 1.1, as well as for reactive
algorithms as described in 1.2.
Here we are behind the problem of universal description techniques for all
unconventional algorithms, some of which we outlined above.
We will not commit ourselves as to whether or not parallel and distributed
ASMs would meet this property. Instead, in addition to the above described
issues, in the rest of the paper we discuss some aspects of unconventional algo-
rithms that must be respected by any universal description technique.
The above classification takes up an ongoing discussion on algorithmic con-
cepts that would not just compute partial functions. [16] made reactive algo-
rithms popular. For cooperating and Network algorithms, cf. [15], [1] and [20].
Shared memory and dynamically generated threads have been studied for deca-
des cf. [4]. Liberal Policies of actions and schemata are covered in [20], among
others. Specification issues are mainly discussed in the framework of Temporal
Logic, such as e.g. in Unity [CM88] or TLA [14]. Fundamental aspects of liberal
policies of algorithms are discussed in [17] and [21].

2 The Issue of Reading

It may come as a surprise that we conceive reading the actual value of variables
the first issue to be considered here. Reading is a quite obvious, simple operation
in conventional algorithms.
Unconventional algorithms, however, frequently consist of local agents that
share variables: A shared variable can be addressed by two or more agents. If they
do so coincidently, the outcome is debatable. There are two main approaches to
tackle this problem.

2.1 Reading Doesn’t Matter

The idea of this approach is one distinguished agent who controls the variable
and may both read (test) and write (change) it. All other agents are only allowed
to read. Reading is assumed not to really interfere with writing. Lamport in [13]
 Towards an ASM Thesis for Unconventional Algorithms 119

nicely exemplifies this approach by some kind of a flag that can be risen and
lowered by one person only. Others may observe the scenery.
The observer (reader) of the flag does not affect the flag in any respect. In
particular, the observer of the flag can not prevent the flag’s owner from rising
and lowering.

2.2 Reading Matters

Lamport’s above described approach properly reflects our everyday physical en-
vironment. It includes myriads of light waves that are reflected by items’ surface,
e.g. flags, and later on catched by humans eyes. This is why observation does
not affect the observed.
Not all physical universes are structured like this. An example is the universe
of computer hardware.
Reading a variable, i.e. copying the actual value of a register, is physically
realized as a sequence of actions that indeed affect the register’s contents. This
contents may be unreliable, inpredictable or in fact different from the expected
value at intermediate states. While being read, the register can not be written
at the same time.
A register like this behaves like a notebook, used by a group of agents. Both
actions of reading and writing require exclusive control. It is the agents’ access
to the notebook that counts for the purpose of writing and reading alike.

2.3 Conclusion

To what extent do the above considerations affect the construction of algorithms?

Conventional algorithms are not affected at all, but unconventional algorithms
are affected to a great extent. For example, a lot of algorithms that organize
mutual exclusion are sensible only in the framework of the first approach, where
reading is assumed not to affect a variable.
There is a traditional bias to conceive the assumptions of the first approach
more convenient and natural than the second. This is why operating systems
frequently offer a surface to processes that pretend the assumptions of the first
approach, in particular interference free reading. The second approach will howe-
ver turn out more lucid for a systematic approach to unconventional algorithms.
120 W. Reisig

3 Sequential and Concurrent Runs

3.1 The Notion of Runs

The notion of a single run is fundamental for algorithms. In fact, an algorithm’s

primarly purpose is to characterize infinitely many (and, frequently, infinitely
long) runs.
A run comprizes occurences of actions, called events. Events can be ordered
according to different aspects. One aspect assumes a (totally ordered) time scale,
defining a time stamp for each event. A run then is lock step if two or more events
may be given the same time stamp. Otherwise it is sequential.
Another aspect describes causal dependency. Any event causes preconditions
for its direct successor events. Transitive closure of the direct successor relation
then is a partial order which describes the causal dependencies among events.
Runs with this order are concurrent. Long standing disputes emphasize advan-
tages and disadvantages of sequential and concurrent runs.
Sequential runs are frequently considered more natural, convenient and wi-
despread. In the context of algorithm universal formalisms however, concurrent
runs are inevitable. Forthcoming sections have the details.

3.2 An Example

As an example consider Dijkstra’s well known system of five philosophers a, . . . , e,

sitting around a table. Each philosopher alternates the actions of thinking and
eating. As an additional requirement, neighbored philosophers are never to eat
together. Each pair of neighbors shares a fork to this end. A philosopher can
start eating only if he finds his adjacent two forks available. In this case he
picks them up (in which moment they are no longer available) and makes them
available again only upon terminating eating.
ASMs suggest to freely choose a level of abstraction at our convenience. So
we choose action A to comprize philosopher a picking up his adjacent forks,
eating, and releasing his forks. Actions B, C, D and E are defined accordingly.

3.3 Sequential Runs

What are the sequential runs of the above algorithm? Apparently, each infi-
nite sequence w ∈ {A, . . . , E}∞ ! On the given level of abstraction, neighbored
philosophers, never eating together, can not be distinguished from detached phi-
losophers who very well might eat together.
x
Writing y to denote concurrent events x and y, one may construct lockstep runs

x1 x2
··· ···
y1 y2
 Towards an ASM Thesis for Unconventional Algorithms 121

where xi and yi denote events of detached philosophers (i = 1, 2, . . . ). This

works nicely for four philosophers. However, lockstep runs can not adequately
represent, for example, an occurence of A while an occurrence of C is followed
by an occurence of D.

3.4 Refinement

To overcome the shortcomings of sequential and lockstep runs, one may refine
the actions A, . . . , E. For example, A may be replaced by three actions A1 , A2
and A3 , denoting the philosopher a to pick up his adjacent forks, to eat and to
release his forks, respectively. A sequence beginning A1 A2 B1 was then ruled
out, for obvious reasons.
Interleaved and lockstep runs consisting of refined actions could more truly
express causal relations than the above runs of unrefined actions can. But ASMs
insist in the specifiers right to choose an abstraction level at his convenience. A
proper way to solve the problem was the use of concurrent runs, to be considered
next.

3.5 Concurrent Runs

We introduce concurrent runs of the philosophers algorithm on the above intro-

duced level of actions A, . . . , E. Those runs very well allow to distinguish the
requirement of neighbors not eating concurrently. Graphically, let

& %
A
% &

represent the above described action A, with ingoing arcs denoting the forks of
A being available to A, and the outgoing arcs denoting the forks of A being
released by A.
Fig.8 then shows one (out of infinitely many) concurrent runs of the philoso-
phers’ algorithms: Philosophers a and c start concurrently, followed by b twice
in a row. d acts after c, and e after a and d.

3.6 Concurrent Determinism

Concurrent runs nicely depict a whole bunch of important aspects of behavior.

Concurrent determinism is one of them: An algorithm is concurrently determini-
stic if each initial state yields exactly one concurrent run. As an example assume
the above described system of five philosophers, with the additional requirement
of decent behavior : Neighbored philosophers alternate use of their shared fork.
Any fixed initial appointment of each fork to one of its users then yields a
unique concurrent run. Fig.9 and Fig.10 exemplify two such decent runs. Fig.11
outlines different patterns in those runs.
122 W. Reisig

A B B

C D E

Fig. 8. A concurrent run of the philosophers algorithm

glue glue
1 2

B B B B

A A A A A

E E E E

D D D D D

C C C C

B B B B

glue glue
1 2

Fig. 9. A decent run of the philosophers system

3.7 Local States in Concurrent Runs

The above examples of concurrent runs consist of occurences of five actions,

A, . . . , E. Direct succession of two such occurences, graphically depicted by e.g.
C → D in Fig.9 implicitely includes a local state that describes the common fork
of C and D as being released by C, apt to be picked up by D.
 Towards an ASM Thesis for Unconventional Algorithms 123

glue glue
1 2

B B B B

C C C C C

D D D D

E E E E E

A A A A

B B B B

glue glue
1 2

Fig. 10. A further decent run of the philosophers system

A A

D C

C D

Fig. 11. Different patterns in the runs of Fig.9 and Fig.10

Local states in concurrent runs are frequently much more prominent: There
are concurrent runs that are essentially characterized by their local states. The
runs of the sorting algorithm of Chapter 1.7 illustrate this.
With xi denoting the pair (x, i) consisting of ‘fish x with sticker i’, Fig. 12
and Fig. 13 show two runs of the sorting algorithm.
An event is represented as

& %

% &
124 W. Reisig

Fig. 12. A concurrent run of distributed sorting

Fig. 13. A further concurrent run of distributed sorting

Each event is an instance of the action of two fishes swapping their stickers.
The involved fishes and stickers are indicated by the local states presented at
the arcs starting and ending nodes, respectively.

3.8 Is There Really More in Concurrent Runs?

The sequential and concurrent runs of an algorithm are tightly interwined: A

concurrent run, i.e. a partially ordered set of events, uniquely defines a set of total
extension of its order. This yields sequences of actions. Each such sequence of
course is a sequential run of the algorithm. The total extensions of all concurrent
runs are just the sequential runs.
 Towards an ASM Thesis for Unconventional Algorithms 125

Vice versa, can the concurrent runs be derived from the sequential runs? This
depends on knowledge about the involved actions’ scope, the topic of the next
chapter.

3.9 A Short History of Concurrent Runs

Research into concurrent runs started with [10] as well as [9] , suggesting acyclic
graphs with nodes denoting events and arcs denoting causal precedence. Petri
in [18] describes how to construct the concurrent runs of condition - event sy-
stems, the most elementary class of Petri nets. Lamport advocates concurrent
runs in [12] , but renounces them in [14]. Concurrent runs have been suggested
for several versions of Petri Nets [5] , [3] , and other system models [7]. They
are extensively employed in [20]. The issue has been debated in an electronic
discussion on the concurrency mailing list, reprinted in [19].

4 The Scope of Actions

The scope of an action includes all variables that the action addresses. This
notion is not too important for conventional algorithms. It is however decisive
when unconventional algorithms are to be specified, constructed and analyzed.

4.1 An Example

As an example consider two integer variables x and y, and a Boolean variable,

z. Let

init ≡ x := 0; z := true; y := 0

Furthermore, let

X≡ while true do x := x + 1;
Y ≡ while true do y := y + 1;

program ≡ init ; X kY

The meaning of the program is obvious: initialization init is followed by parallel

execution of X and Y . With X and Y atomic actions, the sequential runs of pro-
gram are all infinite words w ∈ {X, Y }∞ with both X and Y occuring infinitely
often in w.
program has exactly one concurrent run, to be depicted as in Fig.14
126 W. Reisig

x=0 X → X → X → ···
z = true
y=0 Y → Y → Y → ···

Fig. 14. A concurrent run of program

4.2 A Variant
As a variant of the above program, assume the variables x, y, z and the initia-
lization init as above.
Furthermore, let

X≡ while z do x := x + 1;
Y ≡ while z do y := y + 1;

and

program ≡ init; X kY

The sequential runs of program and program coincide.

The concurrent runs of program depend on the approach to reading of variable
z, as discussed in Chapter 2. Concurrent runs are based on the assumption that
reading affects the variable, as described in 2.2. program then exihibits infinitely
many concurrent runs. One of them can be depicted as in Fig.15.

Fig. 15. A concurrent run of program

4.3 Void Arguments

As a further example for the crucial role of the scope of actions, let x and y be
variables over some domain, A, and let f : A → A be a function. The semantics
of x := f (x), in a concurrent run can be depicted as in Fig. 16.
In this figure, f denotes an event that transforms the local state of x, actually
x = n, into the local state x = f (n), thereby not touching the local state of y,
given by y = m.
 Towards an ASM Thesis for Unconventional Algorithms 127

x=n f x = f (n)

y=m

Fig. 16. Occurrence of x := f (x)

Now, let g : A × A → A be defined by g(a, a0 ) = f (a), i.e. the second argument

of g does not affect the outcome of g. Hence, f (x) and g(x, y) yield the same
result for all values of x and y. Replacing f (x) by g(x, y) does not change any
result of a sequential run.
The scope of f (x) is {x}, whereas the scope of g(x, y) is {x, y}. The semantics
of x := g{x, y} in a concurrent run is depicted in Fig. 17.

Fig. 17. Occurrence of x := g (x, y )

In this figure, g denotes an event that transforms the local states of both x
and y, actually x = n and g = m, respectively, into the local states x = f (n)
and y = m.
Summing up, f (x) and g(x, y) would cause equal effects in sequential runs,
but different effects in concurrent runs.

4.4 Deriving Concurrent Runs from Sequential Runs

Summing up the above observation, concurrent runs explicitely respect, repre-

sent and exploit the scope of actions, whereas sequential runs don’t. This yields
different notions of equivalence: The two programs of 4.1 and 4.2 as well as the
above assignment statements x := f {x} and x := g{x, y} may be mutually sub-
stituted in sequential runs. They may however not be substituted in concurrent
runs, because they would change the actions’ scope.
We are now prepared to return to the question stated in 3.7 and state the
proposition:

The concurrent runs of an algorithm can be derived from its sequential

runs if and only if the involved actions’ scope is known.
128 W. Reisig

5 Proper Data Structures

5.1 Inconsistent Update
A typical example of inconsistent update was an ASM - sequence (or ‘block’)
formed

f (a) := b
f (a) := c

with b 6= c. [8] requires computation to stall in this case. [7] suggests nondeter-
ministic choice.
Both alternatives assume a runtime system or a global controller or whate-
ver means to detect and to manage this case instantly. Distributed algorithms
prevent this kind of regime, however: If f (a) := b and f (a) := c are concur-
rently executed by two different agents, there is no means to detect inconsistency
instantly.

5.2 How Manage Distributed Inconsistency?

Static analysis of an algorithm may reveal that distributed inconsistency never
occurs. But in case it does occur, it can not be detected instantly upon occur-
rence. Hence it can neither be prevented nor treated in a particular way. So we
must provide a framework that would allow distributed inconsistency to occur.
An obvious choice for such a framework were relations, i.e. sets of pairs
R ⊆ A × B, for sets A and B. A function f : A → B is then conceived as the
relation fˆ := {(a, f (a)) | a ∈ A}. Distributed updates f (a) := b and f (a) := c
would then extend fˆ by (a, b) and (a, c).
The idea of relations is not yet satisfactory. As an example, consider three
agents A, B, C, concurrently executing f (a) := b, f (a) := c and f (a) := undef,
respectively. This would not yield a determined effect: It was now necessary to
explicitely state the element to be removed, i.e. to replace f (a) := undef by
remove (a, b) from f or remove (a, c) from f .
Still, a difficulty remains. It can be exemplified by the example of three
agents, A, B, C, concurrently executing f (a) := b, f (a) := b and remove (a, b)
from f , respectively. This, again, yields no determined effect. There is a solution,
however: With f (a) := b, A inserts one item (a, b) to f . With the same assig-
nment statement, B inserts another item (a, b) to f . With remove (a, b) from f ,
C removes one of those items. Removing both of them would require another
occurence of remove (a, b) from f .
Summing up, a data structure can concurrently be updated if it allows for
multiple occurrences of items; the structure is to be a multiset over a set M . That
multisets are the adequate basis for datastructures of distributed system, is well
known for decades: Multiple occurences of identical tokens is a basis notion of
Petri Nets. Multisets are likewise used in the Chemical Abstract Machine of [2].
 Towards an ASM Thesis for Unconventional Algorithms 129

The set M is frequently a relation or a graph of a function; it may be ordered,

be equipped with other operations or distinguished elements. It may be a matter
of fact for a given algorithm, that no element occurs at different locations at the
same time. Nevertheless, a framework is required to capture the case of multiple
occurrences. To this end we suggest two universal operations for all multisets
over some set A, and a ∈ A:
insert (f, a)
is always enabled and inserts another item a to f ;
remove (f, a)
is enabled if there exists at least one item a in f ; occurrence of remove (f, a)
then removes a from f .
Both these operations may occur concurrently. A nice illustration was a large
buffet table with several waiters inserting plates of food and many clients, re-
moving the plates from the buffet table.
The buffet paradigm does not imply, however, that the most general case
of multisets always applies. The structure of an algorithm may guarantee that
a dynamically changing data structure always meets distinguished properties.
Typically, items may actually occur at most once; they may consist in pairs of
items that form a tree, or are right unique (i.e. form a partial function), etc.
Even total functions can be gained by help of an operator update (f, a, b, c) that
consists of remove (f, (a, b)) and insert (f, (a, c)).

References
1. Hagit Attiya and Jenifer Welch. Distributed Computing. McGraw Hill, 1998.
2. Jean-Pierre Banâtre and Daniel Le Métayer. Programming by multiset transfor-
mation. CACM, 36(1):98–111, 1993.
3. Eike Best and Cesar Fernandez. Nonsequential Processes. Springer-Verlag, 1988.
4. W. P. de Roever et al. Concurrency verification: Introduction to compositional
and noncompositional proof methods. to appear 2000.
5. U. Goltz and W. Reisig. The non-sequential behaviour of Petri nets. Information
and Control, 57(2-3):125–147, 1983.
6. Yuri Gurevich. Evolving algebras–a tutorial introduction. Bulletin of the EATCS,
(43):264–284, 1991.
7. Yuri Gurevich. Evolving algebras 1993 : Lipari guide. In E. Börger, editor, Speci-
fication and Validation Methods, pages 9–36. Oxford University Press, 1995.
8. Yuri Gurevich. The sequential asm thesis. Bulletin of the EATCS 67, pages 93–124,
1999.
9. Anatol Holt. Introduction to occurrence systems. In Associative Information
Techniques, pages 175–203. American Elsevir, New York, 1971.
10. Anatol Holt, H. Saint, R. Shapiro, and S. Warshall. Final report on the information
systems theory project. Technical Report RADC-TR-68-305, Rome Air Develope-
ment Center, Griffis Air Force Base, New York, 1968. Distributed by Clearinghouse
for Scientific and Technical Information, US Department of Commerce, 352 pages.
11. Ekkart Kindler and Rolf Walter. Mutex needs fairness. Information Processing
Letters 62, pages 31–39, 1997.
12. Leslie Lamport. Time, clocks, and the ordering of events in a distributed system.
CACM, 21(7):558–565, 1978.
130 W. Reisig

13. Leslie Lamport. Solved problems, unsolved problems and non-problems in concur-
rency. In Proceedings of the 3rd Symposium on Principles of Distributed Computing,
1983, pages 34–44, 1984.
14. Leslie Lamport. The temporal logic of actions. ACM Transactions on Programming
Languages and Systems, 16(3):872–923, 1994.
15. Nancy Lynch. Distributed Algorithms. Kaufman, Morgan, Los Altos, CA, 1996.
16. Zohar Manna and Amir Pnueli. The Temporal Logic of Reactive and Concurrent
Systems. Springer-Verlag, Berlin, 1992.
17. Robin Milner. Elements of interaction. CACM, 36:78–89, 1993.
18. C.A. Petri. Non-sequential processes. Technical Report Internal Report, GMD-
ISF-77-5, Gesellschaft für Mathematik und Datenverarbeitung, Bonn(Germany),
1977.
19. Vaugham Pratt (ed.). Debate ’90: An electronic discussion on true concurrency.
In Peled, Pratt, and Holzmann, editors, Partial Order Methods in Verification,
volume 29 of DIMACS Series, pages 359–403. 1997.
20. Wolfgang Reisig. Elements of Distributed Algorithms. Springer-Verlag, 1998.
21. Peter Wegner. Interactive foundations of computing. Theoretical Computer
Science, 192:315–351, 1998.
 Partially Ordered Runs: A Case Study

Yuri Gurevich1 and Dean Rosenzweig2

Microsoft Research, USA

[email protected] and University of Zagreb, Croatia
[email protected]

Abstract. We look at some sources of insecurity and difficulty in rea-

soning about partially ordered runs of distributed ASMs, and propose
some techniques to facilitate such reasoning. As a case study, we prove
in detail correctness and deadlock–freedom for general partially ordered
runs of distributed ASM models of Lamport’s Bakery Algorithm.

Introduction
Distributed ASMs [4] is a general concurrent model of multi–agent computation.
It was intended, in generalization of its more limited precursors [5,3], to allow
as much concurrency as logically possible. Although the definition has been in
print for several years, its notion of partially ordered runs has remained largely
unexploited in its generality—most of its uses in the literature have recoursed
to some kind of specialization to linear time, discrete or continuous. The general
partially ordered runs seem to be somehow difficult to handle and to reason
about 1 .
Apart from deeply engrained intuitions of linear time, we feel that some
rather technical sources of this difficulty can be detected.
Sequential runs [4] have two properties which greatly facilitate reasoning.
1. Every move is executed in a well–defined state.
2. ‘External’ (or ‘monitored’, cf. below) changes can be located in time, and
thought of as actions by the environment. The environment thus becomes
just another (typically implicit) agent, whose behaviour can be specified in
declarative terms. The judicious splitting of dynamics to a part given by the
program and a part that can be specified declaratively is a natural way to
separate concerns and to abstract in ASMs.
In a partially ordered run neither of the above properties hold in general—a
(global) state in which a move is executed is in general not uniquely defined,
and it is not at all clear how to locate external changes in a partial order. These
seem to be important sources of insecurity and difficulty in reasoning about
partially ordered runs.
1
This remark is not limited to the ASM context—most formal methods modelling
concurrency tend to fall back, one way or another, to some kind of interleaving,
sequential semantics

Y. Gurevich et al. (Eds.): ASM 2000, LNCS 1912, pp. 131–150, 2000.
c Springer-Verlag Berlin Heidelberg 2000

132 Y. Gurevich and D. Rosenzweig

We address both issues here, by developing some techniques to salvage as

much of properties 1. and 2. as needed for partially ordered runs. In order to
convince the reader that, with such means, nontrivial reasoning about partially
ordered runs is feasible, we do it: as a case study, we apply the techniques to a
nontrivial correctness proof. The case study also demonstrates that the proper
setting for analysis of concurrent algorithms involves truly concurrent runs—
mapping to linear time may well miss some important points.
In section 2 we deduce some simple consequences of the coherence condition
of [4], providing sufficient conditions for moves in a distributed run to have at
least a significant portion of state well–defined when they execute. This largely
reconstructs property 1. for partially ordered runs of many distributed programs.
In the same section we introduce ‘external change’ in form of ‘monitored’ mo-
ves, by unknown agents with unknown programs, located exactly in the partial
order. This is an extension of the standard practice (of having ‘the environment’
as a single, typically implicit, unknown agent) reconstructing largely property
2. for partially ordered runs.
In the rest of the paper we explain a nontrivial correctness proof for partially
ordered runs: we prove in detail correctness and deadlock–freedom of (distributed
ASM models of) Lamport’s Bakery Algorithm [6]. We proceed on three different
abstraction levels there. The abstraction level of our primary model, B1 , corre-
sponds precisely to that of Lamport’s algorithm—the part specified declaratively
is exactly what Lamport’s algorithm doesn’t define. A higher level description,
of the model B2 , allows us to deduce correctness and deadlock–freedom from
abstract properties, shown to hold of B1 . A lower level description, of the mo-
del B0 , is an ASM shown to implement programatically exactly all behaviours
allowed by B1 .
Egon Börger and we wrote about the Bakery Algoirithm earlier; see [2] where
a correctness proof was given for ASM models of the Bakery Algorithm with runs
embedded in continuous linear time. Our new models and proofs are similar to
those of [2], but they are also different. The proofs of [2] rely essentially on
continuous linear time. As we will see later (see section 7), certain information
about partially ordered runs is obfuscated in linear runs. Here we remove the
linear time crutches and work directly with partially ordered runs. As in [2],
we borrow ideas from [6] and [1]. In order for the paper to be reasonably self–
contained we spell the entire construction out in full.

1 Preliminaries

We presume that the reader is familiar with [4]. Consider a one-agent program
π and let f be a basic function of π which is dynamic so that the values of f
can change in runs of π. Egon Börger suggested to use for ASMs the following
terminology borrowed from Parnas. f is controlled if only π can change it. f is
monitored if only the environment can change it. f is shared if both π and the
environment can change it. (In [4], controlled functions were called internal, and
monitored functions were called external.)
 Partially Ordered Runs: A Case Study 133

The terminology extends naturally to a multi-agent program Π. Let X be

a set of agents of Π. A dynamic basic function f is controlled by X if only
agents in X can change it. f is monitored by X if none of the agents in X can
change it. The terminology also extends to particular locations rather than whole
functions.

2 Partially Ordered Runs

We rely on the notion of partially ordered run of [4].
This means that we shall consider partially ordered sets of moves with the
‘finite history’ property: {y : y < x} is finite for all x (we shall refer to the
ordering relation by <, using also ≤, >, ≥). Each move is performed by an agent
and, since agents are sequential, moves by one agent form a sequence; since there
are finitely many agents, all antichains are finite.
A (global) state σ(I) is associated with every finite initial segment I (a
downwards closed finite subset) of a run, resulting from performing all moves in
I so that if s < t then s is executed earlier than t. In particular, if I is the empty
segment then σ(I) is the initial state of the run.
Using a partial order implies that moves s, t may be concurrent, i.e. incom-
parable: neither s ≤ t nor t ≤ s holds. The global state ‘resulting’ from a move
is then in general not uniquely determined. It depends on what global state we
see as the one in which the move is performed. Thus states are subject to a
coherence condition [4], which allows us to give the following definitions.
Let Post(t) be the set of all finite initial segments in which a move t is
maximal, and let a be the agent performing t. For each I ∈ Post(t), the state
σ(I) is the state obtained when a executes its program in the state σ(I \ {t})
‘The global state’ in which a move is performed is also not in general uniquely
defined. Let Pre(t) = {I \ {t} : I ∈ Post(t)}. Several different states are thus in
general associated with Pre(t), which makes reasoning about partially ordered
runs somewhat difficult. The coherence condition may however, as we shall see
below, impose that some term has a unique value at all states I ∈ Pre(t).
In order to express such requirements succinctly, we extend term valuation
from states to some statesets, saying that
n
ValPre(t) (u) = c if ∀I ∈ Pre(t) (c = Valσ(I) (u))
undef if no such value exists
where t is a move in a distributed run, ValS (u) is the value of term u in state
S, [4]. We shall often shorten ValPre(t) (u) to uPre(t) . When the value of uPre(t)
is given by the first clause, i.e. when there is a c such that ∀I ∈ Pre(t) (c =
Valσ(I) (u)), we shall say that uPre(t) is indisputable (or that its value is indis-
putable). Notice that an indisputable value may also be undef, but whenever
uPre(t) 6= undef then its value is indisputable.
For future reference, let us note some immediate properties of Pre(t).
Fact 1. Let t be a move in a partially ordered run. The set Pre(t) has the fol-
lowing properties.
134 Y. Gurevich and D. Rosenzweig

1. Pre(t) has a minimal element min Pre(t) = {s : s < t} and a maximal

element max Pre(t) = {s : s 6≥ t}.
2. Pre(t) is the set of all initial segments I such that min Pre(t) ⊆ I ⊆ max Pre(t),
and is hence closed under unions and intersections.
3. Let s be another move. The following are equivalent:
(a) s is concurrent with t.
(b) min Pre(s) ∪ min Pre(t) is an initial segment that belongs to Pre(s) ∩
Pre(t).
(c) Pre(s) ∩ Pre(t) 6= ∅.
(d) There exist I, J ∈ Pre(t) with s ∈ I \ J.

Statement 3 may need an argument. We prove that (a) implies (b) implies
(c) implies (d) implies (a).
(a) implies (b). Assume that s is concurrent with t and let I = min Pre(s) ∪
min Pre(t). Clearly I is an initial segment. Check that min Pre(t) ⊆ I ⊆ max Pre(t),
so that I ∈ Pre(t). By symmetry, I ∈ Pre(s).
(b) implies (c). Trivial.
(c) implies (d). Assume that Pre(s) ∩ Pre(t) 6= ∅ and let J be any member
of Pre(s) ∩ Pre(t). Set I = J ∪ {s}. Clearly, I ∈ Pre(t).
(d) implies (a). Suppose that I, J ∈ Pre(t) and s ∈ I \ J. If s < t then
s ∈ min Pre(t) ⊆ J so that s ∈ J which is impossible. The dual argument shows
that s > t is impossible as well.
We shall say that a move t (in a given partially ordered run) may change
the value of term u if, for some I ∈ Pre(t), we have Valσ(I) (u) 6= Valσ(I∪{t}) (u)
(equivalently, if t changes the value of u in some linearization of the run). If the
above holds for all I ∈ Pre(t) (equivalently, if t changes the value of u in all
linearizations), we shall say that t must change the value of u.
Recall that a linearization of a partially ordered run is a run with the same
moves, and a linear order extending the given partial order. It was noted in [4]
that, in view of the coherence condition, all linearizations of a finite run have
the same final state.

Example 1. Take for instance two agents a, b, such that a executes the program

x := 1

and b executes the program

if mode = first then

mode := second
y := 1
endif
if mode = second then
y := max(x,y)+1
mode := final
endif
 Partially Ordered Runs: A Case Study 135

Now assume that x = y = 0, mode = first initially and consider a run with a
move s of a, concurrent with two consecutive moves t1 , t2 of b. Then both s and
t1 may but not must change the value of max(x, y), while t2 must change it.
In these terms, we have
Lemma 1. If uPre(t) is not indisputable, then there is a move s concurrent with
t which may change u.
Proof. Assume the conclusion is false, that no move concurent with t may
change u. To go from σ(min Pre(t)) to σ(I), by Fact 1, we have to execute only
some moves concurrent to t, none of which may change u. Thus Valσ(I) (u) =
Valσ(min Pre(t)) (u) for all I ∈ Pre(t), and uPre(t) is indisputable, in contradiction
to the premise. t
u

Lemma 2. If there is move s concurrent with t which must change u, then

uPre(t) is not indisputable.
Proof. Since s, t are concurrent, by Fact 1 there is an initial segment I ∈ Pre(s)∩
Pre(t). Then both I and I ∪ {s} are in Pre(t), and, since s must change u, they
have different values of u. t
u
We shall say that a term u is focused (in a given run) if any move which may
change its value, also must change it. For a focused term u it is unambiguous
to say that a move changes the value of u. In many cases the property of being
focused will be obvious from the programs. Here we will also use the following
lemma.
Lemma 3. If the value of a term may only be changed by a single agent, then
the term is focused.
Proof. Suppose that u may be changed only by agent a and that t is a move
by a. It suffices to show that uPre(t) is indisputable, since then uPost(t) is also
indisputable, and is different from uPre(t) iff t changes u. To see that uPre(t)
is indisputable, note that the agents involved in all moves concurrent to t are
different from a, and by the premise none of them may change u. Then by lemma
1 uPre(t) is indisputable. t
u
Putting the above lemmata together, we have
Fact 2. If term u is focused, then uPre(t) is indisputable iff t compares, in the
given partial order, with all moves changing the value of u.
The above example shows that the assumption of focus cannot be dropped for
one direction of fact 2: max(x, y)Pre(t2 ) is indisputably 1, and t2 is concurrent to
s, which may change the value. The nice property that every location is changed
by one agent only does not help with the term max(x, y). Note also that fact 2
is useful for ‘counter’ updates like c := c + 1; if c is changed only in such a way,
even concurrently, then it is focused.
136 Y. Gurevich and D. Rosenzweig

Fact 3. If the values of uPre(t) and uPre(s) are both indisputable but different,
then s < t or t < s.

Proof. Assume uPre(t) and uPre(s) are both indisputable, and s and t are con-
current. Then there is an initial segment I ∈ Pre(s) ∩ Pre(t), and uPre(s) =
Valσ(I) (u) = uPre(t) . t
u

A novelty of this paper is in allowing a run of a multi-agent program Π

to have ‘monitored moves’, that is moves of unknown agents. We have some
known agents with known programs. The moves of these agents are by definition
controlled.
In addition, there may be some number of unknown agents whose programs
are unknown as well. Their moves are by definition monitored.
A dynamic function of Π, or a location, will be called controlled if it is
controlled by the known agents.
It will be called monitored if it is monitored by the known agents.
The presence of unknown agents is consistent with [4], even though the stan-
dard practice has been so far to assume that all explicit moves belong to known
agents with known programs, though the active environment could make some
implicit moves.
The moves by the environment now become explicit, and the unique moni-
tored agent of standard practice, ‘the environment’, is now allowed to split to a
number of possibly different, unknown agents.
The total number of agents involved in any given state is still finite. The
coherence condition applies to all moves, controlled or monitored (even though
we may have no direct way to verify instances of the coherence condition that
involve monitored moves). Therefore facts 2 and 3 remain valid.
The presence of monitored moves allows us to separate concerns and to ab-
stract. Parts of the algorithm can be formulated in form of more or less decla-
rative statements about monitored moves in runs (which blurs to an extent the
distinction between the algorithm and its environment).

3 Lamport’s Algorithm

For arbitrary but fixed N let P1 , . . . , PN be processes (we shall also talk about
‘customers’) that may want from time to time to access a ‘critical section’ CS
of code. Any mutual exclusion protocol—which each Pi is supposed to execute
in order to enter the critical section—has to prevent two processes from being
in the critical section simultaneously. The Bakery Algorithm provides each Pi
with a (shared) register Ri and a (private) array n[1], . . . , n[N ] holding natural
numbers. Only Pi is allowed to write to Ri but every process can read the
register. We assume each register to be initialized with value 0.
The algorithm was presented by Lamport with the following piece of pseu-
docode.
 Partially Ordered Runs: A Case Study 137

Start
n[i] := 1
write(Ri ,n[i])
Doorway
for all j6=i, read(Rj ,n[j])
Ticket
n[i] := 1 + maxj n[j]
write(Ri ,n[i])
Wait
for all j6=i, repeat
read(Rj ,n[j]) until
n[j]=0 or n[j]>n[i] or (n[j]=n[i] and j>i)
Critical Section
Finale
Ri := 0
The Bakery Algorithm is divided into six consecutive phases: start, doorway,
ticket assignment, wait section, critical section and finale.
To declare its interest in accessing the critical session, a process Pi writes 1
into array variable ni and then posts the written value in its register.
In the doorway section, Pi copies all the other registers into its array. It
then computes a ticket, which is the least integer greater than all integers in its
private array, writes the ticket into ni and posts the written value in its register.
During the subsequent wait section, process Pi keeps reading, into its array,
the registers of each other process Pj , until the resulting array value n[j] = 0 or
n[j] > n[i] or n[j] = n[i] ∧ j > i.
The meaning of the condition is the following: if n[j] = 0, then Pj is not
interested in entering the critical section, and it has no right to block Pi . If
n[j] > n[i] > 0, then Pi has a smaller ‘ticket’ and has the right to go before Pj .
The last clause resolves the case of two customers obtaining the same ‘ticket’:
then one with smaller identifier goes first. Note that by ordering pairs of positive
integers lexicographically:
(i, j) < (k, l) ←→ [i < k or (i = k and j < l)]
one can write the until condition as follows: n[j]=0 or (n[j],j)>(n[i],i).
Once permitted to go, Pi enters the critical section. Upon leaving CS, as
finale, Pi sets its register to 0.
Note also that the for-all commands in the doorway and the wait section may
be executed in many ways, in various sequences, all at once, concurrently etc.
It may be worth mentioning the following. The process first writes into n[i]
and then posts the written value at Ri . Obviously it could do the two actions in
the reverse order. Intuitively, the order between the two actions is immaterial,
but the sequential character of the pseudo-code imposes one.
138 Y. Gurevich and D. Rosenzweig

4 The Primary Model: ASM B1

The doorway section in Lamport’s program does not give us any indication how
customer i is supposed to perform reading. Should it read the registers Rj in
the order given by the indices, in the reversed order? Should it get help and use
vassal agents, one per each Rj ? There are many other possibilities. To reflect the
situation in proper generality, our primary ASM model B1 includes no reading
instructions whatsoever. Instead, we will require that runs of B1 satisfy certain
provisos that guarantee that reading is performed.

4.1 The Program of B1

The ASM has only one program, used by all customers, which has five rules. The
array A(X, Y ) represents the array n[Y ] of the program, private to customer X.
We assume that initially all registers have value 0, all customers are in mode
satisfied, and all elements of the array A(X, Y ) are undef. We assume that the
identifiers of the N customers are distinct natural numbers < N . Variables X, Y
will range over customers.

Start
if mode(me) = satisfied then
A(me,me) := 1, R(me) := 1, mode(me) := doorway

Ticket

if mode(me) = doorway and ∀Y 6= me (A(me,Y) 6= undef) then

A(me,me) := 1 + maxY A(me,Y), R(me) := 1 + maxY A(me,Y)
mode(me) := wait

Entry

if mode(me) = wait and

∀ Y 6= me (A(me,Y)=0 or (A(me,Y),id(Y)) > (A(me,me),id(me))) then
mode(me) := CS

Exit

if mode(me) = CS then
mode(me) := done

Finale

if mode(me) = done then

R(me) := 0, mode(me) := satisfied
∀ Y 6= me A(me,Y) := undef
 Partially Ordered Runs: A Case Study 139

4.2 Semantics of B1

We would like to assume that, in any mode different from Satisfied, no custo-
mer stalls forever; eventually it makes a move (provided a move is continuously
enabled from some time on).
Since in ASMs we have no explicit notion of a move (or program) being
enabled, and in partially ordered runs we have no explicit notion of time, both
‘enabled’ and ‘continuously from some time on’ need definitions.
There are two obvious candidates for the notion of a program being enabled
in a state. One is based on the intuition that a program is enabled if it ‘gets
down to updates’, i.e. if in the given state it generates a nonempty set of updates.
The other possibility is that it really changes the state, i.e. that the updateset is
nonempty and also nontrivial. We are happy to sidestep the issue here, since for
all programs of this paper the two notions will coincide—whenever a nonemtpy
set of updates is generated, it will also be nontrivial. Thus we can say that a
program is enabled in state σ if it produces a nonempty set of updates in σ.
We say that an agent X stalls forever in a run if (a) X has a last move, say
t, and (b) after t a move by X (the program of X) is eventually always enabled
(in all σ(J) for J ⊇ I, for some initial segment I 3 t).
We thus assume

Progress Proviso. No customer, in mode other than Satisfied, stalls forever.

We consider the runs of B1 containing enabled moves by customers executing

their programs, subject to the Progress Proviso, and also some monitored moves.
Our entire knowledge about monitored moves will be encapsulated in explicit
requirements D, W1, W2 below.
We now define intervals characterized by the successive executions, by a
process X, of its rules Start, Ticket, Entry, Exit, Finale (also in a partial order we
refer to open intervals (a, b) = {x : a < x < b}).

Definition 1. Suppose a is the move of X executing Start rule, and b is the

next move by X (which has to execute the Ticket rule).
Then the interval x = (a, b) is a doorway of X, and a = Start(x), b =
Ticket(x). If b is the last execution of X then the wait interval W (x) = {t : t > b}
is incomplete and the CS interval CS(x) is undefined. Suppose that c is the next
move of X after b (necessarily executing Entry rule), d is the next move of X after
c (necessarily executing Exit rule), and e is the next move of X after d (necessa-
rily executing Finale rule). Then W (x) = (b, c) and CS(x) = (c, d), c = Entry(x),
d = Exit(x), e = Finale(x).

By Progress Proviso and requirement D below, every doorway is complete,

i.e. each execution of Start is followed by execution of Ticket. So is every cri-
tical section, i.e. each execution of Entry is followed by executions of Exit (and
subsequently Finale).
140 Y. Gurevich and D. Rosenzweig

The program of customer X writes to locations mode(X), R(X), A(X, Y )2 ,

where locations A(X, Y ) with Y 6= X are only cleaned up (that is set to undef)
by X (in Finale) and somebody else writes more meaningful information into
these locations.
Our program covers all but the reading actions. Since our definitions do not
allow ‘partially known programs’, i.e. a controlled agent can do no more than
what his program instructs him to do, more meaningful values have to be written
there by the environment, i.e. by somebody else.
We assume that locations mode(X), R(X), A(X, X) are also controlled by X,
i.e. that no other (known or unknown) agent may write there. This is justified
by Lamport’s algorithm: other customers, as well as the environment, have no
business writing to mode(X), R(X), A(X, X).
This assumption implies that R(Y ) is focused, for all customers Y , and, by
the program and fact 2, for all moves t in a run of B1 ,

Corollary 1. R(Y )Pre(t) is indisputable iff t compares to all Start, Ticket and
Finale moves of Y.

To avoid repetitive case distinctions for customers which (being satisfied)

have register 0, and of customers which happen to receive the same ticket, we
introduce the following notation. If f is a function from customers to natural
numbers, let

N · f (X) + id(X), if f (X) > 0;
f 0 (X) =
∞, otherwise.

Let X, Y range over customers, and x, y over doorways of customers X, Y

respectively.
We abbreviate 1 + maxY A(X, Y )Pre(Ticket(x)) as T (x).
The declarative requirements, saying what reads need be done, are then (with
x being an arbitrary doorway of customer X)

D Each execution of Start by X completes to a doorway x. For each x, for each

Y 6= X there is a move b ∈ x, such that A(X, Y )Pre(Ticket(x)) = R(Y )Pre(b)
(thus T (x) > R(Y )Pre(b) 6= undef).
W1 If W (x) is complete, then for each Y 6= X there is a move b ∈ W (x), such
that R(Y )Pre(b) = A(X, Y )Pre(Entry(x)) (thus T 0 (x) < R0 (Y )Pre(b) 6= undef).
W2 If W (x) is incomplete, then for some Y 6= X there is an infinite chain
b1 < b2 < · · · of moves in W (x), such that, for each n, R0 (Y )Pre(bn ) < T 0 (x)
(thus also R(Y )Pre(bn ) 6= undef).

D tells us that the value of R(Y ), appearing in the array at Ticket(x), is

read in x. W1 says that a permission to go is obtained by executing, for each
Y , a successful read in W (x), while W2 tells us that X may be prevented from
2
by [4] the official notation for these locations is (mode, X), (R, X), (A, (X, Y )); since
in the simple cases occurring in this paper, no ambiguity may arise, we shall use the
applicative term notation as above also for locations.
 Partially Ordered Runs: A Case Study 141

going only by executing, for some Y 6= X, an infinite sequence of unsuccessful

reads in W (x), where a read b ∈ W (x) from R(Y ) on behalf of X is successful
if R0 (Y )Pre(b) > T 0 (x). It turns out that D, W1 and W2 is all that we need to
know about reading actions in order to prove correctness and deadlock–freedom
of B1 .
Requirements D and W1 say that, for each Y , there is a move b(Y ) in x,
respectively W (x), having some property. Remark that, without loss of genera-
lity, we can assume that these moves b(Y ) are all distinct. Suppose namely that,
in D or W1, we have b = b(Y1 ) = · · · = b(Yk ). Then we can replace b with k
distinct monitored moves which, in the partial order, follow and precede exactly
the same moves as b does. It is easy to see that this replacement leaves us with
a legitimate run, with exactly the same partial order of customers’ moves. A
remark of the same kind applies also to sequences of moves claimed by W3, but
we shall not need that case.
The reader familiar with [2] might notice that, what in similar requirements
there were temporal conditions on some monitored locations, takes here (and in
the next section) the shape of conditions on behaviour of unknown agents. The
role of some time moments in proofs of [2] thus turns out to be that of place
holders for monitored moves.

5 Correctness and Deadlock–Freedom: The ASM B2

We define an ASM expressing a ‘higher level’ view of the Bakery Algorithm,

similar to B1 but with the array abstracted away. The relevant datum to be
described abstractly is the ticket assigned to a customer X (and written into
its register R(X)) when X leaves the doorway and enters the wait section. We
introduce for this purpose two monitored functions, boolean valued Ready and
integer valued T , expressing, respectively, readiness of the ticket and its value.
The relevant moment to be analyzed is the moment at which a process with
a ticket is allowed to enter the critical section. This ‘permission to go’ will also
be represented by a monitored function, Go.
We will impose requirements on the environment and monitored moves, re-
sponsible for the values of Ready, T and Go, which will be shown to guarantee
the correctness and deadlock–freedom of the higher level ASM B2 . We will then
show that these requirements are correctly implemented in B1 .

5.1 The Program of B2

Start

if mode(me) = satisfied then

R(me) := 1, mode(me) := doorway
142 Y. Gurevich and D. Rosenzweig

Ticket
if mode(me) = doorway and Ready(me) then
R(me) := T(me), mode(me) := wait
Entry
if mode(me) = wait and Go(me) then
mode(me) := CS
Exit
if mode(me) = CS then
mode(me) := done
Finale
if mode(me) = done then
mode(me) := satisfied, R(me) := 0

5.2 Semantics of B2
The ASM B2 is similar to that of B1 except for the fact that the array is gone.
In particular we assume the Progress Proviso (for known agents, i.e. customers).
The role of the array is taken over by three monitored functions, Ready, T
and Go. Looking at B1 , Ready(X) and T (X) can be seen as standing for the
abbreviations used there, while Go(X) can be interpreted as the guard of the
Entry rule, ∀Y 6= X(A(X, Y ) = 0 or (A(X, Y ), id(Y )) > (A(X, X), id(X))).
The ASM B2 provides however no means to compute Ready, T and Go.
Our first requirement says that every interested customer eventually obtains
his ticket:
C0 Each execution of Start, by a customer X, completes to a doorway x. For
each x the value T (X)Pre(Ticket(x)) is indisputable.
The indisputable value of T (X)Pre(Ticket(x)) will be, like before, denoted by
T (x). In order to express the rest of our conditions on the array in terms of T
and Go, we need some additional notation and terminology.
For open intervals in a partial order we also use (a, b) < (c, d) if b ≤ c,
and say that the two intervals are concurrent if neither b ≤ c nor d ≤ a. Note
that concurrency does not necessarily imply overlap, i.e. existence of common
elements; it in general just allows it.3
Sometimes we shall also compare elements with intervals: c < (a, b) if c ≤ a,
likewise for >.
This ordering will help us to formalize the idea that tickets increase together
with doorways (see C2 below). This should also apply in a way to concurrent
3
Note however that, if intervals are interpreted as intervals on the partial order of
initial segments, with (a, b) containing all segments containing a but not b, then
concurrent intervals indeed overlap.
 Partially Ordered Runs: A Case Study 143

doorways; these are ordered by the following relation ≺, borrowed from its linear
order analog of [1].
Let X 6= Y , and let x, y range over doorways of X, Y respectively.

Definition 2. x
y if x and y are concurrent and T 0 (x) < T 0 (y). Further, x ≺ y
if x
y or x < y.

Lemma 4. x ≺ y or y ≺ x.

Proof. Note that T 0 (y) 6= T 0 (x) for X 6= Y , while two doorways of the same
customer can never be concurrent. t
u

Our other conditions are then

C1 T (x) is a positive integer > 1.
C2 If y < x then either Finale(y) < Ticket(x) or T 0 (y) < T 0 (x).
C3 If W (x) is complete, then, for every Y 6= X, there exists a move b ∈ W (x),
such that T 0 (x) < R0 (Y )Pre(b) (thus R(Y )Pre(b) 6= undef).
C4 If W (x) is incomplete, then there is a y ≺ x with W (y) incomplete.
Intuitively, C2 says that tickets respect the temporal precedence of doorways
with concurrent wait periods, C4 is an induction principle, and C3 expresses
that permission to go is obtained by checking the ticket against competitors’
registers. C2 (together with C0) is easily seen to be an abstract version of D, C3
is an abstract version of W1, while the fact, to be proved below, that C4 follows
from W2 together with D,W1, is the essence of deadlock–freedom for the Bakery
algorithm.
An immediate consequence of C3 is finite concurrency of doorways:

Corollary 2. The set of doorways concurrent to any given doorway is finite.

Proof. Let x < x0 be two doorways of X both concurrent to y. By C3 applied

to x, there is a move b, with x < b < x0 . Since R(Y )Pre(b) is indisputable, by
corollary 1 b compares to both ends of y; b ≤ Start(y) would imply x < y,
while b ≥ Ticket(y) would imply y < x0 , both contradicting the assumption of
concurrency. Thus b ∈ y. But, by finite history, there can be only finitely many
such b’s. t
u

5.3 Correctness and Fairness of B2

Lemma 5. (First Comes, First Served) If y ≺ x and W (x) is complete,
then W (y) is complete and CS(y) < CS(x).

Proof. Assume the premise is satisfied and the conclusion is false, i.e. that there
is no move Finale(y) < Entry(x). Take b as given by C3.
Claim 1 : T 0 (y) < T 0 (x).
Claim 2 : Ticket(y) < b.
144 Y. Gurevich and D. Rosenzweig

Given the claims, we have T 0 (y) < T 0 (x) < R0 (Y )Pre(b) 6= undef, and thus Y
must be writing to R(Y ) by a move in (Ticket(y), b). But the first such write
after Ticket(y) must be a Finale move, which contradicts the assumption that
the conclusion of the lemma is false.
Claim 1 follows immediately from definition of ≺ in case of concurrency, and
from C2 otherwise.
To prove Claim 2, we first note that b is comparable to both ends of y, and,
in view of y ≺ x, b ≤ Start(y) is impossible. It also impossible that Start(y) <
b ≤ Ticket(y), since then R(Y )Pre(b) = 1, which contradicts the choice of b. ut

Lemma 6. ≺ is transitive.

Proof. by contradiction. Suppose x ≺ y ≺ z ≺ x. Count the number n of <’s

in the above sequence of ≺ signs. In case n = 0 the statement follows from the
fact that the order of integers (tickets) is transitive, and in cases n = 2, 3 the
statement follows the fact that the partial order < of open intervals in a partial
order is transitive. In case n = 1, by symmetry, we may assume x
y
z < x
and therefore T 0 (x) < T 0 (y) < T 0 (z). Given the assumption x ≺ y ≺ z ≺ x,
by Lemma 5, if one of the waiting sections is complete then so are the other
two, and we have CS(x) < CS(y) < CS(z) < CS(x) which is impossible. So all
three waiting sections must be incomplete. Thus we can apply C2 to obtain also
T 0 (z) < T 0 (x), which is impossible. t
u

Lemma 7. (Deadlock freedom) Every W (x) is complete.

Proof. By corollary 2 (and finite history) ≺ is well–founded. Then C4 is precisely

the induction principle required to establish the claim. t
u

This section is summarized in the following

Theorem 1. Doorways are linearly ordered by ≺. All waiting sections are com-
plete, and x ≺ y implies CS(x) < CS(y).

5.4 B1 Implements B2 Correctly

We check that the requirements are satisfied in B1 (i.e. follow from D, W1, W2),
where Ready(X) = (∀Y 6= X(A(X, Y ) 6= undef)), T (X) = 1 + maxY A(X, Y ),
and Go(X) means that the condition of the rule Entry is satisfied.
C0 is enforced by requirement D and the Progress Proviso for B1 .
C1 is satisfied since the maximum in the rule Ticket is taken over all Y ,
including X which at that moment has register value R(X) = 1.
C2. Take b as given by D. Since R(Y )Pre(b) is indisputable, the move b com-
pares to all Start, Ticket and Finale moves of Y . With Ticket(y) ≤ Start(x) < b,
it is meaningful to ask whether Y executes the Finale move in (Ticket(y), b). If
it does, we are done; if it doesn’t, R(Y )Pre(b) = T (y), and, by D, T 0 (y) < T 0 (x).
C3 follows immediately from W1.
 Partially Ordered Runs: A Case Study 145

C4. By contradiction, suppose that the premise is satisfied but the conclusion
is false, i.e. W (x) is incomplete but W (y) is complete for all y ≺ x. Let Y and
b1 < b2 < · · · be the customer and the sequence of moves as given by W2.
Claim: There is a move b ∈ W (x), with R(Y )Pre(b) 6= undef, such that the
following two properties hold for each y:

(i) b > Ticket(y) (ii) if y ≺ x then b > Finale(y).

First we derive the desired contradiction from the claim, and second we prove
the claim.
So suppose that the claim is true and let b be as in the claim. Then R(Y )Pre(b)
has an indisputable value, and b thus compares to all moves of Y that change
R(Y ). What is the value of R(Y ) in Pre(b)? We have two possible scenarios.
Scenario A: all y ≺ x; then b succeeds every Finale(y) and thus R(Y )Pre(b) = 0.
Scenario B: there is some y with Ticket(y) < b ≤ Finale(y); then R(Y )Pre(b) =
T (y). To summarize, if b is as in the claim, then R(Y )Pre(b) is either 0 or T (y),
so that R0 (Y )Pre(b) ≥ T 0 (y).
The values of R(Y )Pre(b) and of R(Y )Pre(bn ) for every n are indisputable.
Moves b and bn thus all compare with every move of Y which changes R(Y ).
It is easy to see that any bn 6< b satisfies (i) and (ii) in the claim. But then, as
shown above, R0 (Y )Pre(bn ) ≥ T 0 (y), which contradicts the property of bn in W2.
Thus every bn < b, which contradicts finite history.
It remains to prove the claim.
To prove the claim, note that, for y ≺ x, CS(y) is defined and complete, by
the assumption that W (y) is complete and the Progress Proviso. It suffices to
prove that there is at most one y > x. The sequence of doorways of Y is then
finite: by finite history, it has finitely many elements < x, by corollary 2 finitely
many elements concurrent to x. Thus Y has a last move, say eY . By Progress
Proviso, eY can only be a Ticket or a Finale move. Since all bn , by corollary 1,
compare to eY , by finite history, for sufficiently large n we have bn > eY . We
can then, for claimed b, take any bn > eY .
It remains to prove that Y has at most one doorway > x. Suppose x < y.
Then, by C2 (with x, y playing y, x respectively), T 0 (x) < T 0 (y) (since W (x)
is incomplete). If W (y) were complete, by C3 there would be a c ∈ W (y) such
that R0 (X)Pre(c) > T 0 (y). But since x < y < c, we also have c ∈ W (x) and
R(X)Pre(c) = T (x), so T 0 (x) > T 0 (y), which is impossible. Thus W (y) is incom-
plete, and y is the last doorway of Y .
We have thus verified that C0–C4 hold of arbitrary runs of B1 . It follows that
the results of the previous subsection, summarized in Theorem 1, hold of B1 as
well.
146 Y. Gurevich and D. Rosenzweig

6 Realizing the Model with Reading Agents: the ASM B0

6.1 The Program of B0

Until now the universe of known agents was populated by customers only. Now we
also have another kind of agents. They present one way of making the unknown
agents of B1 known.
Formally the universe of agents splits into two disjoint universes: Customer
and Reader. Customers and readers are related by several functions. If X and
Y are distinct customers, then in each state there is at most one reader–agent
r(X, Y ), and there are no other readers.
If r is the reader r(X, Y ), then Lord(r) = X and Subject(r) = Y . The
readers will be created on the fly, when needed (at Start and Ticket moves), and
will self–destruct when their task is completed.

Customer Start

if mode(me) = satisfied then

A(me,me) := 1, R(me) := 1, mode(me) := doorway
∀ Y 6= me create-reader(me, Y, doorway)

Ticket

if mode(me) = doorway and (∀Y 6=me) A(me,Y) 6= undef then

A(me,me) := 1 + maxY A(me,Y), R(me) := 1 + maxY A(me,Y)
mode(me) := wait
∀ Y 6= me create-reader(me, Y, wait)

Entry

if mode(me) = wait and

∀ Y 6= me (A(me,Y)=0 or (A(me,Y),id(Y)) > (A(me,me),id(me))) then
mode(me) := CS

Exit

if mode(me) = CS then
mode(me) := done

Finale

if mode(me) = done then

R(me) := 0, mode(me) := satisfied
(∀ Y 6= me) A(me,Y) := undef

where create-reader(X, Y, m) abbreviates the rule

 Partially Ordered Runs: A Case Study 147

create r
agent(r) := true, Reader(r) := true
program(r) := reader-program
Lord(r) := X, Subject(r) := Y
mode(r) := m
endcreate

Reader
A(Lord(me),Subject(me)) := R(Subject(me))
if mode(me) = doorway then
destroy-reader(me)
if mode(me) = wait then
if R(Subject(me)) = 0
or (R(Subject(me)),id(Subject(me)))
> (A(Lord(me),Lord(me)),id(Lord(me))) then
destroy-reader(me)
where destroy-reader(a) abbreviates the rule

agent(a) := false, Reader(a) := false

program(a) := undef, Lord(a) := undef, Subject(a) := undef

6.2 Semantics of B0
Semantics of B0 is like that of B1 , but considerably simpler, since all locations
are controlled by the known agents, and there are no moves monitored by the
known agents, to put constraints on—it is all in the programs for B0 . The reader
agents are one way to realize the requirement that those ‘for-all commands in
the doorway and the wait section of Lamport’s pseudocode may be executed in
many ways, in various sequences, all at once, concurrently etc.’ In fact the reader
agents capture all ways to realize that requirement, see below.
The only assumption we have to make, outside of the program, is the Progress
Proviso, applying here to all agents, both customers and readers:
Progress Proviso. No reader, and no customer in mode other than Satisfied,
stalls forever.
The reader–agents are created on the fly, and destroyed upon completion of
their task: the effect of destroy-reader(a), if a is a reader–agent, is returning a
to the reserve.

6.3 B0 Realizes B1
The constraints D, W1, W2 can be read as a rather direct description of what the
reader–agents do for their customers in B0 . The fact that every run of B0 satisfies
148 Y. Gurevich and D. Rosenzweig

D, W1, W2 follows from the programs and the Progress Proviso (together with
the semantics described above or in [4]).
D is satisfied in B0 since, for every move t of X executing Start, for every Y 6=
X, there is a reader r(X, Y ) at Post(Start(x)). By programs and the Progress
Proviso each of these readers makes a single self destructive move, which is the
b required by D; by programs and the Progress Proviso X eventually executes
Ticket.
By programs and Progress Proviso, for every Y 6= X there is a reader r(X, Y )
at Post(Entry(x)). That reader makes a move in W (x). For W1, W2 it then
suffices to note

Fact 4. A move t by r(X, Y ) in W (x) is the last such move iff it is successful,
i.e. T 0 (x) < R0 (Y )Pre(t) .

W1 is namely satisfied in B0 since, for each Y 6= X, we can take the last

waiting section move of r(X, Y ) for the claimed b.
W2 is satisfied in B0 since, if all r(X, Y ) for Y 6= X have a last move in W (x),
by Progress Proviso X must eventually execute Entry. Thus for some Y 6= X
the reader r(X, Y ) keeps reading forever—take the sequence of his moves for
b1 < b2 < · · · as claimed.
We have thus established that every run ρ0 of B0 can be viewed as a run
of B1 . Since B1 , as far as reading is concerned, can be viewed as a declarative
description of algorithmic behaviour, rather then an algorithm proper, ρ0 can
also be seen as a realization of behaviour prescribed by B1 .
To be more precise, let us introduce an appropriate implementation relation
between moves and runs of the two ASMs.
A move t0 by customer X in B0 implements a move t1 by the same customer
in B1 if the indisputable portions of states (values of mode(X), R(X), A(X, X))
at Pre(t0 ), Post(t0 ) coincide with those at Pre(t1 ), Post(t1 ) respectively.
A run ρ0 of B0 implements a run ρ1 of B1 if the partial order of customers’
moves in ρ0 is order–isomorphic to the partial order of customers’ moves in ρ1 ,
implementing it pointwise: whenever the isomorphism maps a move t0 in ρ0 to
a move t1 in ρ1 , then t0 implements t1 .
In these terms, we have established that B1 (more specifically, D, W1, W2)
provides a sound description of algorithmic behaviour: B0 is an algorithm beha-
ving like that. For the record,

Lemma 8. (Soundness of B1 ) Each run of B0 implements a run of B1 .

We can actually claim more. The requirements D, W1, W2 allow many dif-
ferent behaviours. Is there a behaviour, allowed by B1 , which is not captured by
the reader–agents of B0 ? Not really. This is the content of the following lemma,
expressing a kind of completeness property.

Lemma 9. (Completeness of B1 ) Each run ρ1 of B1 is implemented by a run

ρ0 of B0 .
 Partially Ordered Runs: A Case Study 149

Proof. The idea is to transform ρ1 to ρ0 by implementing reading moves of ρ1

with appropriate moves of reader–agents, possibly ignoring some inconsequential
monitored moves of ρ1 . The replacement process is done in the order of ρ1 , that
is earlier read moves are replaced (or discarded) earlier. The following conditions
will be guaranteed by induction for every move b introduced by replacement, for
every customer X and every doorway x of X. At Pre(b) there is a reader agent
r = r(X, Y ) for some Y 6= X. If b is a move of r in x, then mode(r)Pre(b) =
doorway (so that r self–destructs at b), and if b is a move of r in W (x), then
mode(r)Pre(b) = wait (so that r self–destructs at its last move in W (x)).
Now let x = (a, b) be a doorway of X. By D, for each Y 6= X there is a
move b(Y ) ∈ x such that A(X, Y )Pre(Ticket(x)) = R(Y )Pre(b) . Recall that we can
assume that b(Y ) are all distinct. Then implement each b(Y ) with a move of
r = r(X, Y ). By induction condition, r is in mode doorway before b(Y ), and
therefore self–destructs there.
The case of read moves in W (x) is similar. Since W (x) is complete, W1
guarantees that, for each Y 6= X there is a move b(Y ) ∈ W (x) such that
R(Y )Pre(b) = A(X, Y )Pre(Entry(x)) . Without loss of generality, all moves b(Y ) are
distinct. Replace every b(Y ) with a move of r = r(X, Y ). By the induction con-
dition, r is in mode wait before b(Y ), and therefore self–destructs at the move. If
b ∈ W (x) is a monitored move different from all b(Y ), we can discard it—also, if
there is a b(Y ) > b, we can implement b with an unsuccessful, nonselfdestructive
read of r(X, Y ).
Finally, remove all remaining monitored moves in ρ1 . The result is the desired
run ρ0 of B0 , implementing ρ1 . t
u

7 Concluding Remarks

Corrollary 2 implies the following cofiniteness property for the partial runs of the
Bakery algorithm: for each move t the set {s|s 6> t} is finite. This is a property
of the Bakery Algorithm, and not of the modelling framework: in spite of finite
history, it is easy to concoct legitimate partially ordered runs of some algorithm
which violate the cofiniteness property. In the case of the Bakery Algorithm, the
cofiniteness property implies that any two infinitely active customers have to
synchronize infinitely many times.
The cofiniteness property is an example of a property of partial runs that
is obfuscated in linear runs (since for linears runs it amounts to finite history).
This indicates that concurrent computations may have significant properties
which cannot be discovered by studying only their linearizations. Concurrent
computations should be analyzed directly.

Acknowledgements. We thank Ante Djerek, Robert Eschbach and Igor Ur-

biha, who have provided very useful remarks on a draft version of this paper.
150 Y. Gurevich and D. Rosenzweig

References
1. Uri Abraham. Bakery algorithms. Unpublished manuscript, pp. 35, 1993.
2. Egon Börger, Yuri Gurevich, and Dean Rosenzweig. The bakery algorithm: Yet ano-
ther specification and verification. In E. Börger, editor, Specification and Validation
Methods, pages 231–243. Oxford University Press, 1995.
3. Paola Glavan and Dean Rosenzweig. Communicating evolving algebras. In
E. Börger, H. Kleine Büning, G. Jäger, S. Martini, and M. M. Richter, editors,
Computer Science Logic, number 702 in Lecture Notes in Computer Science, pages
182–215. Springer, 1993.
4. Yuri Gurevich. Evolving algebra 1993: Lipari guide. In E. Börger, editor, Specifica-
tion and Validation Methods, pages 9–36. Oxford University Press, 1995.
5. Yuri Gurevich and Larry Moss. Algebraic operational semantics and Occam. In
E. Börger, H. Kleine Büning, and M. M. Richter, editors, CSL’89, 3rd Workshop on
Computer Science Logic, number 440 in Lecture Notes in Computer Science, pages
176–192. Springer, 1990.
6. Leslie Lamport. A new solution of Dijkstra concurrent programming problem. Com-
munications of the ACM, 17(8):453–455, 1974.
 Investigating Java Concurrency
Using Abstract State Machines

Yuri Gurevich1 , Wolfram Schulte1 , and Charles Wallace2

1
Microsoft Research, Redmond, WA 98052-6399, USA
{gurevich,schulte}@microsoft.com
2
University of Delaware, Newark, DE 19716, USA
[email protected]

Abstract. We present a mathematically precise, platform-independent

model of Java concurrency using the Abstract State Machine method.
We cover all aspects of Java threads and synchronization, gradually ad-
ding details to the model in a series of steps. We motivate and explain
each concurrency feature, and point out subtleties, inconsistencies and
ambiguities in the official, informal Java specification.

1 Introduction

The Java programming language [7,13] provides sophisticated support for con-
currency. The fundamental operations of concurrent programs are implemented
as built-in features of the language. Many of the notoriously complicated details
of concurrent programming are hidden from the programmer, simplifying the de-
sign of concurrent applications. Furthermore, a platform-neutral memory model
is included as part of the language. The incorporation of such intricate, subtle
operations into Java calls for a precise specification. As interest in the language’s
concurrency model increases, Java developers are examining and exploiting its
details [15,17,16,14]. The popularity of Java and, more importantly, its empha-
sis on cross-platform compatibility make the need for such a specification even
stronger.
We present a model of the concurrent features of Java, using the formal
operational specification method of Abstract State Machines (ASMs) 1 [9,22,23].
We use the Java Language Specification manual (JLS) [7], as our reference for the
language. The JLS is an informal specification, and due to the ambiguity which
pervades natural language, it can be interpreted in different ways. Our model
gives an unambiguous specification which reflects our interpretation of the JLS.
Throughout the paper, we indicate where ambiguities and omissions in the JLS
give rise to other interpretations. The formal specification process also uncovers
many subtle but important issues which the JLS does not bring to light. Our
goal is a specification that is not only precise but accessible to its readers, even
those not familiar with Java, concurrent programming, or ASMs. As part of this
1
ASMs were formerly known as Evolving Algebras.

Y. Gurevich et al. (Eds.): ASM 2000, LNCS 1912, pp. 151–176, 2000.
c Springer-Verlag Berlin Heidelberg 2000

152 Y. Gurevich, W. Schulte, and C. Wallace

project, we implement the ASM specifications using the ASMGofer interpreter

[20]. This implementation serves as a convenient tool for prototyping and testing.
It is important to distinguish unintentional ambiguity from intentional un-
derspecification. The authors of the JLS envision support for Java concurrency
taking many different forms: “by having many hardware processors, by time-
slicing a single hardware processor, or by time-slicing many hardware proces-
sors [7]”. The JLS leaves some details unspecified in order to give implementers
of Java a certain amount of freedom: “the intent is to permit certain standard
hardware and software techniques that can greatly improve the speed and effi-
ciency of concurrent code [7]”. As is usual with the ASM methodology [8,2], we
design our ASMs to model Java concurrency at its natural level of abstraction,
capturing the essence of the JLS without committing to any implementation
decisions.
While most of the JLS specification is imperative in nature, some of the
concurrency model is described in a declarative style: rather than explain how to
execute a concurrent program, the JLS gives conditions that a correct execution
must meet. This shift in presentation style is most likely due to the notion that
only a declarative specification can be truly implementation-independent.
There are two ways for us to deal with the declarative portions of the JLS.
The obvious and simple way is to reformulate them declaratively as conditions
on the execution of our ASM. The other way is to implement them in the ASM
itself. We choose to do the latter, whenever it is natural to do so, but in a way
that does not sacrifice generality. Our model obeys the conditions established
in the JLS, yet it is fully general in the sense that for any concurrent program
execution that follows the rules of the JLS, there is a corresponding run of the
ASM. We find our imperative approach helpful in creating a clear mental picture
of the concurrency model, and in uncovering hidden assumptions, inconsistencies
and ambiguities in the JLS.
There are several formalizations of Java concurrency in the literature. Börger
and Schulte [3] also uses ASMs, to give a semantic analysis of Java that exposes
a hierarchy of natural sublanguages. It covers the language features for con-
currency, but its focus is on the high-level language rather than the lower-level
concurrency model, so it does not provide a full specification of Java concurrency.
There are several works [1,4,5] using the Structural Operational Semantics (SOS)
methodology [18]. Attali et al. [1] do not concentrate on the lower-level details of
Java concurrency, while Cenciarelli et al. [4] give them in a declarative style, in
keeping with the presentation in the JLS. As mentioned earlier, we believe that
there are advantages to our imperative approach. Coscia and Reggio [5] expli-
citly deal with the details of the concurrency model and propose some changes to
its design. Gontmakher and Schuster [6] present a non-imperative specification,
with the purpose of comparing Java’s memory behavior with other well-known
notions of consistency. Our focus is different, however: while their work assumes
a particular interpretation of the JLS and proceeds from there, our goal is first
to come to a clear understanding of the concurrency model as presented in the
 Investigating Java Concurrency 153

JLS, and then to discuss its consequences. As we shall see, there are several
issues within the JLS which make this initial understanding difficult.
We introduce in §2 the basic rules by which agents (threads) interact with
regard to shared variables, and in §3 the special considerations for variables
marked as volatile. In §4 we introduce locks, which are used to limit concurrency
within a program. In §5 we discuss prescient store actions. In §6 we describe
thread objects. In §7, we cover waiting and notification.
Our technical report [11] provides a complete presentation of our ASM speci-
fication. In it, we also prove that our specification is as general as possible while
obeying all the correctness criteria of the JLS.

Acknowledgment. This work was partially supported by NSF grant CCR-95-

04375.

2 Threads and Variables

In a single-threaded computation, a single agent executes the instructions of a
program, one at a time. In contrast, a run of a Java program may be multi-
threaded, (intuitively) involving multiple agents that execute instructions con-
currently. Each agent executes sequentially, the sequence of its actions forming a
thread of execution. In the parlance of distributed computing, the term “thread”
is used to refer not only to an agent’s computation but also to the agent itself.
Threads may execute on a single processor or multiple processors. The atomic
actions of different threads may be interleaved; they may even be concurrent in
the case of multiple processors.
Since different threads may access and update common data, the order of
their actions affects the results of an execution. The JLS establishes some condi-
tions on the interaction of threads with respect to shared data, but intentionally
leaves some freedom to implementers of the language. Consequently, the results
of executing a multithreaded program may vary between different Java plat-
forms, and even between different executions on the same platform.
A data value is either an instance of a primitive type, such as int or boolean,
or an object, a dynamically created value. A value resides at a location either
in the working memory of a thread or in the main memory. Certain memory
locations serve as manifestations of a variable. The master copy of a variable is
the unique location in the main memory that manifests the variable. A thread’s
working copy of a variable is the unique location in the thread’s working memory
that manifests the variable. A thread accesses or updates only its working copies.
The JLS speaks of the main memory as an agent which communicates with
threads and updates variable master values. While this view is adequate for
describing a system with a centralized main memory, it does not capture the
possibility of a distributed main memory [16], where updates of different variable
master copies may occur concurrently. We find it convenient to imagine that each
variable has a master agent, which accesses and updates the variable’s master
copy. Since there is a one-to-one correspondence between variables and their
masters, we may identify a variable master with the variable it controls.
154 Y. Gurevich, W. Schulte, and C. Wallace

We create an ASM to model threads’ actions on variables. The agents of

are threads (members of Thread) and variable masters (members of Var). At
the beginning of each run, there is a single member of the universe Thread.
Since we identify variable masters with the variables they control, a member of
the universe Var is both a variable and a variable master. The universe Value
represents the set of values that master or working copies of variables may bear.
The function masterValue maps each variable to the current value of its master
copy. The function workingValue, given a thread and variable, returns the current
value of the thread’s working copy of that variable.
A thread’s execution engine is the component of the thread that executes the
Java program. It may perform actions on variables: creating a variable, assigning
a new value to a variable, or using a previously assigned value of a variable.
There may be more than one value associated with a variable, but a thread has
immediate access only to the value of its working copy. A use or assign action is
internal to a thread, involving only its working copy. Since we are not interested
here in what a thread computes during its execution, our view of use and assign
actions is simple: an assign action just changes the value of a variable’s working
copy, and a use action does nothing. (Although what happens in a use action is
not important to us, we shall see in §3 that the occurrence of a use action may
be significant, even at our level of abstraction.) A thread’s execution engine may
do one other sort of action of interest to us: creating another thread, which may
execute concurrently with its creator.
Threads pass values of shared variables among themselves via the main me-
mory. A thread may update the master copy of a variable with a freshly assig-
ned value, or it may request the current value of the master copy. This is done
through asynchronous communication with a variable’s master agent. To trans-
fer the value of its master copy to a thread’s working memory, (the master of)
a variable issues a read action. This is followed by a load action by the thread,
which installs the value into its working memory. To transfer its working value
of a variable to the main memory, a thread issues a store action. This is followed
by a write action by (the master of) the variable, which installs the value into
the main memory.
Somehow information is passed from variable to thread in a read-load se-
quence, and from thread to variable in a store-write sequence. To give a more
imperative character to our description, without any essential loss of generality,
we introduce an explicit but quite abstract notion of messages passing between
threads and variables. When a variable issues a read action, it sends an access
message to a thread, and the target thread then receives the message by issuing
a load action. When a thread performs a store action, it sends an update mes-
sage to a variable, and the variable then receives the message by issuing a write
action. So for example, we may speak of a thread storing out to a variable by
sending it an access message, or a variable writing in from a thread by installing
the value of an update message.
We define a universe Msg comprising the universe AccMsg (messages from
variables to threads) and the universe UpdMsg (messages from threads to varia-
 Investigating Java Concurrency 155

bles). The function var maps each access message to the variable that sent it,
and maps each update message to the variable that is its intended recipient. The
function thread maps each update message to the thread that sent it, and maps
each access message to the thread that is its intended recipient. The function
value returns the value contained in a given message.
In any state, the members of the universe AccMsg that have been sent from
a variable v to a thread t form a “subuniverse” of AccMsg, which we call
AccMsg(v, t). Likewise, the members of the universe UpdMsg that have been
sent from t to v form a subuniverse of UpdMsg, which we call UpdMsg(t, v).
While the JLS does not dictate a specific policy on the access of shared va-
riables, it does impose some rules on the concurrent behavior of threads and
variables. We present these rules as they appear in the JLS and give our inter-
pretations of them.

JLS rules 1

1. [p. 403] “The actions performed by any one thread are totally ordered; that
is, for any two actions performed by a thread, one action precedes the other.”
2. [p. 403] “The actions performed by the main memory for any one variable are
totally ordered; that is, for any two actions performed by the main memory
on the same variable, one action precedes the other.”
3. [p. 403] “It is not permitted for an action to follow itself.”

In other words, actions on variables form a partial order (which we denote by

<) in which the actions of any single thread are linearly ordered, and the actions
of any single variable master are linearly ordered. Note that this supports our
view of independent variable master agents as opposed to a monolithic main
memory. The read and write actions on a single variable are ordered linearly,
but read and write actions on different variables may occur concurrently. This
follows naturally if we think of the master copy of each variable as controlled by
an independent agent.

JLS rules 2

1. [p. 403] “Each load action by a thread is uniquely paired with a read action
by the main memory such that the load action follows the read action.”
2. [p. 403] “Each store action by a thread is uniquely paired with a write action
by the main memory such that the write action follows the store action.”
3. [p. 405] “For every load action performed by any thread t on its working
copy of a variable v, there must be a corresponding preceding read action by
the main memory on the master copy of v, and the load action must put into
the working copy the data transmitted by the corresponding read action.”
4. [p. 405] “For every store action performed by any thread t on its working
copy of a variable v, there must be a corresponding following write action
by the main memory on the master copy of v, and the write action must
put into the master copy the data transmitted by the corresponding store
action.”
156 Y. Gurevich, W. Schulte, and C. Wallace

5. [p. 405] “Let action A be a load or store by thread t on variable v, and let
action P be the corresponding read or write by the main memory on variable
v. Similarly, let action B be some other load or store by thread t on that
same variable v, and let action Q be the corresponding read or write by the
main memory on variable v. If A precedes B, then P must precede Q.”

These rules restrict the ways in which a thread’s load and store actions may
be interleaved with a variable’s read and write actions. First, we note that the
term “uniquely paired” in Rules 2.1 and 2.2 is ambiguous. A statement of the
form “Each element x of X is uniquely paired with an element y of Y such that
φ(x, y)” has a lenient interpretation: “For all x in X there is a unique y in Y
such that φ(x, y)”. It also has a strict interpretation: “For all x in X there is
a unique y in Y such that φ(x, y), and for all y in Y there is a unique x in X
such that φ(x, y)”. The lenient interpretation allows spurious y’s; that is, y’s
that are not paired with any x. Which interpretation shall we use? The strict
interpretation is what the authors of the JLS intended [21], so we adopt it here.
However, we find that a lenient interpretation of the term in Rule 2.1 has some
advantages that make it worth discussing.
Read/load order. For every load action L by a thread t on a variable v, there
is a read action R < L by v such that L loads in the value read out at R
(Rules 2.1 and 2.3). This supports our conception of access messages: every load
action must load in the value of an access message issued by a previous read
action. Furthermore, a thread may not load in the same access message twice
(Rule 2.5).
The choice of interpretation of “uniquely paired” determines whether every
access message must be loaded in. The strict interpretation, under which every
access message is loaded in, is simpler and more appealing from a logical perspec-
tive, which is why it is the intended interpretation [21]. But note that spurious
read actions are innocuous: failing to load in a given access message has only
the effect of making a thread’s working memory less up-to-date than possible.
Furthermore, the lenient interpretation allows a higher degree of independence
of threads and variables: a variable is free to issue access messages without con-
cern for whether they are eventually loaded in. For instance, updated master
values of an important variable could be broadcast to threads, with each thread
deciding whether to load the value in; or blocks (pages) of master values could
be sent to a thread, with the thread selecting which values to load in.
Store/write order. For every store action S by a thread t on a variable v,
there is a write action W > S by v such that W writes in the value stored out
at S (Rules 2.2 and 2.4). Using our message-passing parlance, every store action
sends an update message which must be written in by a following write action.
As with the case of read and load actions, how we interpret “uniquely paired”
determines whether every write action is an action of writing in an update mes-
sage. Here, adopting the strict interpretation seems less controversial. It is not
clear what value a spurious write action writes. At best, it simply rewrites the
value that is already in main memory, in which case it is useless. But if it writes
 Investigating Java Concurrency 157

a different value, it may obliterate a current value in main memory, replacing it

with an outdated or invalid value.
Read/write order follows load/store order. Each thread performs load and
store actions on a variable in a certain order, and the variable must perform the
corresponding read and write actions in the same order. Let us examine this
more closely.
Load/load. If a thread t performs a load action L1 (with corresponding read
action R1 ) on a variable v, and another load action L2 (with corresponding read
action R2 ) on v, Rule 2.5 postulates that (L1 < L2 ) ⇒ (R1 < R2 ). If v sends
access messages m and m0 to t, then t may load in m before m0 only if m was
sent before m0 . In other words, t loads in ever newer access messages from v.
Without loss of generality, we may think of t as discarding m when it loads it
in.
The conjunction of Rule 2.3 and Rule 2.5 implies Rule 2.1. Rule 2.1 asserts
that each load action has a preceding unique read action. Rule 2.3 asserts that
every load action has a preceding read action and Rule 2.5 asserts that different
load actions must have different corresponding read actions, ensuring uniqueness.
Store/store. If a thread t performs a store action S1 on v (with corresponding
write action W1 ) and another store action S2 on v (with corresponding write
action W2 ), Rule 2.5 postulates that (S1 < S2 ) ⇒ (W1 < W2 ). If t sends an
update message m to v before sending another update message m0 to v, then m
must be written before m0 . In other words, v writes in ever newer values stored
out by t. Thus writing m0 is allowed only if m has been delivered through a write
action. We can think of v as discarding m when it writes it.
Store/load. If a thread t performs a store action S (with corresponding write
action W ) on v, and t performs a load action L (with corresponding read action
R) on v, Rule 2.5 postulates that (S < L) ⇒ (W < R). In other words, if t
sends an update message m to v before loading an access message m0 in from v,
then v must write m in before sending m0 . If t stores its working value out to v
and then immediately loads in from v, and other threads do not store out to v
in the meantime, then the value it receives from the load is the value it stored.
Thus sending m0 is allowed only after the point in time when m is written in.
If m0 is sent but not loaded in before this point in time, it can never be loaded
in, and thus Rule 2.1 is violated (under the strict interpretation). To avoid this,
v must not issue an access message to t as long as there is a pending update
message from t to v, and t must not issue an update message to v as long as
there is a pending access message from v to t. Furthermore, v and t must be
prevented from sending messages to each other concurrently.
Under the assumption that messages take no time to travel, we may require
that v checks for incoming update messages from t before sending an access
message to t, and t checks for incoming access messages from v before sending
an update message to v. But this does not eliminate the possibility of concurrent
actions by v and t. Thus some means of avoiding concurrent actions by v and t is
necessary. Note that this complication disappears if we were to adopt the lenient
158 Y. Gurevich, W. Schulte, and C. Wallace

interpretation of “uniquely paired” for the pairing of read and load actions. In
our example, t could simply ignore the access message m0 and still have its
update message m written.
To represent the relative age of messages, we introduce the relation <, and
the term oldest? to determine whether a given message has no predecessors.
term oldest?(m): (not ∃m0 : Msg) m0 < m

The JLS does not impose any particular means of ordering messages, so we
avoid doing so here by making the function external. We restrict attention to
runs in which the function < behaves in the expected way.

JLS rules 3 Let t be a thread and v be a variable.

1. [p. 404] “A use or assign action by t of v is permitted only when dictated by
execution by t of the Java program according to the standard Java execution
model.”
2. [p. 404] “A store action by t on v must intervene between an assign by t of
v and a subsequent load by t of v.”
3. [p. 404] “An assign action by t on v must intervene between a load or store
by t of v and a subsequent store by t of v.”
4. [p. 404] “After a thread is created, it must perform an assign or load action
on a variable before performing a use or store action on that variable.”
5. [p. 405] “After a variable is created, every thread must perform an assign or
load action on that variable before performing a use or store action on that
variable.”

These rules impose constraints on the exchange between a thread’s execution

engine and working memory.
Use and assign actions. Threads may only issue use and assign actions when
needed for progress by the execution engine (Rule 3.1). At the level of abstrac-
tion we have chosen, we are not concerned with what threads actually compute,
so this rule does not affect our model. Nevertheless, we should note that the
term “dictated” is somewhat ambiguous here. For example, let var be an ex-
pression that requires a use action on a variable v. In evaluating the expression
(var + var), how many use actions on v are dictated by the program? The
strictest interpretation would require a use action for each reference to a varia-
ble (our example would require two use actions), and this is what the authors of
the JLS intended [21].
Store actions. If a thread t performs a load or store action LS on v and then
performs a store action S on v, Rule 3.3 postulates that t performs an assign
action A on v such that LS < A < S. In other words, spurious store actions
are forbidden: t may only store out to v if it has a new value to transmit. This
prevents the current contents of the main memory from being overwritten by
older (i.e., previously stored) values. Without this rule, a thread could store out
outdated values it loaded in long ago, overwriting more current values that other
threads stored out in the interim.
 Investigating Java Concurrency 159

If a thread t performs an assign action A on a variable v and then performs

a load action L on v, there is a store action S by t on v such that A < S < L
(Rule 3.2). In other words, a thread must store out a value of a variable v after
issuing a sequence of assigns to v, transmitting at least the last assigned value
to the main memory, before attempting a load; the assigns are not completely
forgotten. A thread may load an access message in from a given variable only if
the last working value it assigned has been stored out. Note that a thread may
assign to a variable and then use the variable with no intervening load. In this
case, no intervening store is required, and the assigned value may simply remain
in the working memory.
This makes the coordination between threads and variables more complex.
Once t performs an assign on v, its next action on v cannot be a load. So if t
assigns a value to v concurrently with v sending an access message to t, t may
not subsequently load the message in until it stores its working value out to v.
But it cannot perform this store action, as the write action corresponding to
the store would follow v’s read action; thus Rules 2.1 and 2.3 are violated. To
avoid this, v must not issue an access message to t if there is an assign action
by t on v that has not been followed by a store action on v, and t must not
assign to v if there is a pending access message. So the actions which threads
and variables must avoid doing concurrently are not reads and stores, as Rule 2.5
might suggest, but reads and assigns.
Use/store actions preceded by an assign/load action. If a thread t per-
forms a use or store action U S on a variable v, Rules 3.4 and 3.5 postulate that
there is an assign or load action AL < U S by t on v. For use and store actions
to behave properly, there must be a valid value in working memory. Initially, all
the contents of a thread’s working memory are invalid, but as a result of these
rules, no invalid value of the working memory is ever used or stored out.
We define the following functions to help enforce the conditions described
above. Rule 3.3 allows a thread t to store out its working value of a variable
v only if it has assigned a fresh value to v but has not stored it out; Rule 3.2
allows t to load an access message in from v only if it has no freshly assigned
working value of v to store out. The function freshAssign? determines whether
a given thread has assigned a fresh value to a given variable without storing it
out. Rules 3.4–5 allow t to use its working value of v only if there is a valid
value of v in its working memory, installed there by a load or assign action. The
function usableValue? determines whether a given thread has a valid value for a
given variable in its working memory.
The actions of a variable master consist of issuing a read message to a thread,
and writing in an update message.
module Var: choose among
choose t: Thread: readOK?(t)
Read out to t
choose t: Thread
choose m: UpdMsg(t, Self): writeOK?(m, t)
Write m in from t
160 Y. Gurevich, W. Schulte, and C. Wallace

rule Read out to t: extend AccMsg(Self, t) with m

value(m) := masterValue(Self)
rule Write m in from t: masterValue(Self) := value(m)
UpdMsg(m) := false

The terms readOK? and writeOK? determine whether the given action is
allowed by the JLS rules. Rules 2.5 and 3.2 allow a variable v to read out to a
thread t only if every assign action by t on v has been followed by corresponding
store and write actions. Rule 2.5 allows t to write a message in from v only if
the message is the oldest pending message from t to v.
term readOK?(t): not (freshAssign?(t, Self) or (∃m: UpdMsg(t, Self)))
term writeOK?(m, t): oldest?(m)

The actions of a thread consist of storing a value out through an update

message, loading in an access message, or taking a step in executing the Java
program. Program execution may involve using the working value of a variable,
assigning a value to the working copy of a variable, creating a variable, or creating
a thread.
module Thread: choose among
Execute program WM-MM transfer
rule Execute program: choose among
WM-EE transfer Create var Create thread
rule WM-EE transfer : choose among
choose v: Var: useOK?(v)
Use v
choose v: Var: assignOK?(v)
Assign to v
rule WM-MM transfer : choose among
choose v: Var
choose m: AccMsg(v, Self): loadOK?(m, v)
Load m in from v
choose v: Var: storeOK?(v)
Store out to v

We define rules for load, store, use and assign actions, as well as the actions
of creating a variable and creating a thread.
rule Load m in from v: workingValue(Self, v) := value(m)
usableValue?(Self, v) := true
AccMsg(m) := false
rule Store out to v: freshAssign?(Self, v) := false
extend UpdMsg(Self, v) with m
value(m) := workingValue(Self, v)
 Investigating Java Concurrency 161

rule Use v: skip

rule Assign to v: freshAssign?(Self, v) := true
usableValue?(Self, v) := true
choose val: Value
workingValue(Self, v) := val
rule Create var : create Var agent v
rule Create thread : create Thread agent t

The terms loadOK?, storeOK?, useOK? and assignOK? determine whether the
given action is allowed by the JLS rules. Rule 2.5 allows a thread t to load a
message in from a variable v only if the message is the oldest pending message
from v to t. Rules 3.3–5 allow t to store out to v only if there has been an assign
action by t on v without a following store action. Rules 3.4–5 allow t to use v
only if there has been an assign or load action that put a value in t’s working
copy of v. Rules 2.5 and 3.2 allow t to assign to v only if every access message
from v to t has been loaded in.
term loadOK?(m, v): oldest?(m) and not freshAssign?(Self, v)
term storeOK?(v): freshAssign?(Self, v)
term useOK?(v): usableValue?(Self, v)
term assignOK?(v): (not ∃m: AccMsg(v, Self))

3 Volatile Variables

The basic model for threads and shared variables, as presented in §2, permits
optimizations that reduce the amount of communication between agents and
thus enhance the performance of multithreaded programs.
Communication between threads and the main memory can be lessened
through caching: keeping values in working memory without transmitting them
to main memory or getting new values from main memory. Instead of consulting
the main memory every time it uses a variable, a thread may service several use
actions on the same variable with a single load action. It may also service several
assign actions with a single store action, sending only the last of these assigned
values to main memory.
Caching may have undesirable results, particularly for variables that are as-
signed to and used frequently by different threads. Cached values may become
outdated as other threads store out to the main memory. Also, caching assigned
values prevents other threads from viewing the newly assigned values as they
arise.
Communication between variables can be avoided altogether, allowing them
to operate independently of one another. While Rule 2.5 dictates that the order
of a thread’s actions on a single variable is also followed by the variable, the order
162 Y. Gurevich, W. Schulte, and C. Wallace

of its actions over different variables need not be followed in the main memory.
Consequently, variables need not coordinate their actions among themselves.
Hence for certain variables, a stricter discipline is necessary. Java allows the
programmer to declare variables volatile at the time of their creation. Volatile
variables follow a policy that disallows the optimizations described above.
We modify the ASM to model operations on volatile variables. We add a uni-
verse VolVar, representing the set of volatile variables. Every member of VolVar
is also a member of Var. The rule Create var requires a slight change: a new
variable may be marked as volatile.
The JLS imposes the following additional conditions on volatile variables.

JLS rules 4 [p. 407] “Let t be a thread and let v and w be volatile variables.”
1. “A use action by t on v is permitted only if the previous action by t on v
was load, and a load action by t on v is permitted only if the next action by
t on v is use. The use action is said to be “associated” with the read action
that corresponds to the load.”
2. “A store action by t on v is permitted only if the previous action by t on
v was assign, and an assign action by t on v is permitted only if the next
action by t on v is store. The assign action is said to be “associated” with
the write action that corresponds to the store.”
3. “Let action A be a use or assign by thread t on variable v, let action F be
the load or store associated with A, and let action P be the read or write of
v that corresponds to F . Similarly, let action B be a use or assign by thread
t on variable w, let action G be the load or store associated with B, and let
action Q be the read or write of v that corresponds to G. If A precedes B,
then P must precede Q.”

Caching is disallowed. If a thread t performs a load action L on v, Rule 4.1

postulates that there is a use action U > L by t on v, and there is no action
Act by t on v such that L < Act < U . In other words, each value loaded in from
the main memory is used exactly once. In contrast, multiple use actions on a
non-volatile variable may use a value loaded in by a single load action.
Rule 4.2 postulates that a thread t performs a store action S on a volatile
variable v if and only if there is an assign action A < S by t on v and there is
no action Act by t on v such that A < Act < S. In other words, every assigned
value must be stored out and written to the main memory exactly once. For
non-volatile variables, assign actions may overwrite one another without having
their values stored out.
Read/write order follows use/assign order. Each thread performs use and
assign actions on volatile variables in a certain order; Rule 4.3 ensures that the
variables must perform the corresponding read and write actions in the same
order. This condition is similar to Rule 2.5, but different in some important
ways. First, the thread actions mentioned in Rule 4.3 are use and assign actions,
as opposed to store and load actions as in Rule 2.5. Also, the ordering holds over
actions on all volatile variables, not just those on a single variable.
 Investigating Java Concurrency 163

To account for the behavior of volatile variables, we modify our view of

message passing between threads and variables. We think of a volatile variable
access message as ending in a use action, as opposed to a load action. Similarly,
we think of a volatile variable update message as originating with an assign
action, as opposed to a store action. By Rules 2.4 and 4.2, all volatile update
messages must be stored out and written, and by Rules 2.3 and 4.1, all volatile
access messages must be loaded in and used. We examine this more closely.
Use/Use. If a thread t performs a use action U1 (with corresponding read action
R1 ) on a volatile variable v, and another use action U2 (with corresponding read
action R2 ) on a volatile variable w, Rule 4.3 postulates that (U1 < U2 ) ⇒ (R1 <
R2 ). Note that R1 and R2 may be actions by different variables. Given an access
message m from v and an access message m0 from w, t may use m before m0
only if m was sent before m0 . In other words, t uses ever newer volatile access
messages.
The conjunction of Rule 4.3 with Rules 4.1 and 3.1 implies that volatile va-
riables must work closely with threads’ execution engines. Once a thread loads
in a read message from a volatile variable, by Rule 4.1 it must use that variable,
and by Rule 4.3 it must do so before performing any other assign or use ac-
tion. Rule 3.1 forbids use actions not “dictated” by the execution engine, so the
message must be sent with knowledge of what is needed by the Java program.
Consider a scenario where a thread has just loaded an access message from a
volatile variable v, but the next operation by the execution engine is an assign
action on v, or an action on another volatile variable. The only way for t to pro-
gress is to issue a gratuitous use action on v, but this is forbidden by Rule 3.1.
Therefore, volatile variables must be careful when issuing access messages to
a thread, doing so only when it is clear that the thread’s execution needs one
immediately.
Rule 4.3 also implies that volatile variables must not read out to the same
thread concurrently. As the actions of a single thread are linearly ordered, by
Rule 4.3 the corresponding read actions must be linearly ordered as well. This
may require some coordination between volatile variables; if several volatile va-
riables wish to read out to a single thread, only a single variable at a time may
do so.
Assign/Assign. If a thread t performs an assign action A1 (with corresponding
write action W1 ) on a volatile variable v, and another assign action A2 (with
corresponding write action W2 ) on a volatile variable w, Rule 4.3 postulates that
(A1 < A2 ) ⇒ (W1 < W2 ). Note that W1 and W2 may be actions by different
variables. In other words, given an update message m from t to v and an update
message m0 from t to w, v may write m before m0 only if m was sent before m0 .
So volatile update messages sent by t are written in the order in which they are
sent.
Assign/Use. If a thread t performs a use action U (with corresponding read
action R) on a volatile variable v, and an assign action A (with corresponding
write action W ) on a volatile variable w, Rule 4.3 postulates that (A < U ) ⇒
164 Y. Gurevich, W. Schulte, and C. Wallace

(W < R). In other words, if t sends a volatile update message m before using a
volatile access message m0 , then m must be written in before m0 is sent.
Thus sending m0 is allowed only after the point in time when m and all other
volatile update messages from t have been written. If m0 is sent but not used
before this point in time, it can never be used, hence violating Rule 4.1. To avoid
this, v must not issue an access message to t as long as there is a pending volatile
update message from t, and t must not issue a volatile update message as long
as there is a pending volatile access message to t. Furthermore, volatile access
and update messages must not be sent to and from t concurrently.
We define some additional functions to help enforce these conditions. Rule 4.1
requires a thread t to use a volatile variable v if and only if it has loaded in a
message from v without using its value. If a thread has loaded in a message from
a volatile variable but not used its value, the function msgToUse returns this
message. If a thread has assigned a value to a volatile variable but not stored
the value out, the function msgToStore returns the volatile update message.
The terms readOK? and writeOK? include extra conditions on volatile va-
riables. Rule 4.3 allows a volatile variable v to read out to a thread t only if
all assign actions by t on volatile variables have been followed by correspon-
ding store and write actions. To readOK? we add the conjunct VolVar(Self) ⇒
(∀v: VolVar) (not ∃m: UpdMsg(t, v)). Rule 4.3 allows v to write an update mes-
sage from t in only if the message has been stored out and is the oldest pen-
ding volatile message from t. To writeOK? we add the conjunct VolVar(Self) ⇒
m 6= msgToStore(t, Self).
We modify the rules for load and store actions. A volatile access message is
removed when it is used, rather than when it is loaded in. A volatile update
message is created through an assign action, rather than through a store action.

rule Load m in from v: workingValue(Self, v) := value(m)

usableValue?(Self, v) := true
if VolVar(v) then msgToUse(Self, v) := m
else AccMsg(m) := false

rule Store out to v: freshAssign?(Self, v) := false

if VolVar(v) then msgToStore(Self, v) := undef
else extend UpdMsg(Self, v) with m
value(m) := workingValue(Self, v)

The term loadOK? includes an extra condition on volatile variables. Rule 4.1
allows a thread t to load a message in from a volatile variable v only if it has used
all previous loaded access messages from v. Rule 4.2 allows t to load a message
in from v only if it has stored out all previously assigned values to v. To loadOK?
we add the conjunct VolVar(v) ⇒ not usableValue?(Self, v).
We modify the rules for use and assign actions. A use action on a volatile
variable removes the volatile access message from the AccMsg universe. An assign
action on a volatile variable generates a UpdMsg.
 Investigating Java Concurrency 165

rule Use v: if VolVar(v) then

usableValue?(Self, v) := false
msgToUse(Self, v) := undef
AccMsg(msgToUse(Self, v)) := false
rule Assign to v: freshAssign?(Self, v) := true
if not VolVar(v) then usableValue?(Self, v) := true
choose val: Value
workingValue(Self, v) := val
if VolVar(v) then
extend UpdMsg(Self, v) with m
val(m) := val
msgToStore(Self, v) := m

The terms useOK? and assignOK? include extra conditions on volatile varia-
bles. Rule 4.1 allows a thread t to use its working value of a variable v only
if it has loaded in a message from v without using its value. Rule 4.3 requires
this message to be the oldest pending volatile access message to t. To useOK?
we add the conjunct VolVar(v) ⇒ oldest?(msgToUse(Self, v)). Rule 4.2 allows t
to assign a value to v only if all of t’s previous assigns to t have been stored
out, and Rule 4.3 requires that there be no pending volatile access messages
to t. To assignOK? we add the conjunct VolVar(v) ⇒ not (freshAssign?(Self, v)
or usableValue?(Self, v) or (∃m: AccMsg(v 0 , Self))).
If two volatile variable variables read out to t concurrently, the corresponding
load actions will be ordered, since t can only perform one load action at a time.
But the read actions will not be ordered, thereby violating Rule 4.3. We restrict
attention to runs in which this does not occur.
Rules 4.1 and 4.2 ensure that if a load or assign action on a VolVar occurs,
a corresponding use or store action will occur sometime in the future. Similarly
to Rules 2.1 and 2.2, these rules can be flouted by delaying a use or store action
indefinitely. We restrict attention to runs that avoid this situation.

4 Locks
Certain situations require that a thread be able to perform a series of operations
without interference from other threads. In general, the programs of threads
t1 . . . tn may have critical regions that they should avoid executing concurrently.
It seems reasonable to try to prevent any ti and tj from entering their critical
regions simultaneously. A natural way to solve this problem in an object-oriented
framework is to have one critical object related to all these critical regions. All
the threads may refer to the critical object, but only one thread may own the
object at any given time, and only the owner can execute its critical region. Java
provides support for this sort of mutual exclusion. The critical regions are called
synchronized regions.
In Java, each object has a unique lock associated with it. A thread gains
entry into a synchronized code region of its program by performing a lock action
166 Y. Gurevich, W. Schulte, and C. Wallace

on the lock of the critical object. The thread then holds the lock until it exits
the synchronized code region and performs an unlock action on the lock.
Various threads may issue lock or unlock actions on the same lock. The lock
and unlock actions on a single lock are performed in conjunction with an arbiter,
which restricts the number of threads holding a single lock to one at a time. The
JLS speaks of main memory as this arbiter. As with variable master copies, we
prefer to think of each lock as controlled by a master agent, rather than by the
(possibly distributed) main memory.
We modify the ASM to model operations on locks. The agents are threads,
variable masters, and lock masters. We represent locks as members of the uni-
verse Lock. As with variables, we identify lock master agents with the locks they
control, so a member of the universe Lock is both a lock and a lock master.
Objects are members of the universe Object. A LockActionType is either lock or
unlock, and a LockAction is a pair consisting of a LockActionType and a Lock on
which to perform the given action. The function lock maps each object to its
lock.
To see how the introduction of locks ensures that only one thread enters its
synchronized region, we must consider the JLS rules for the concurrent behavior
of locks.

JLS rules 5

1. [p. 403] “The actions performed by the main memory for any one lock are
totally ordered; that is, for any two actions performed by the main memory
on the same lock, one action precedes the other.”
2. [p. 403] “Each lock or unlock action is performed jointly by some thread
and the main memory.”

Rule 5.1 supports our view of independent lock master agents. The actions
on a single lock are ordered linearly, but actions on different locks may occur
concurrently.

JLS rules 6 Let T be a thread and L be a lock.

1. [p. 406] “A lock action by T on L may occur only if, for every thread S
other than T , the number of preceding unlock actions by S on L equals the
number of preceding lock actions by S on L.”
2. [p. 406] “An unlock action by thread T on lock L may occur only if the
number of preceding unlock actions by T on L is strictly less than the number
of preceding lock actions by T on L.”

Rule 6.1 ensures that a thread’s hold on a lock is exclusive, and Rule 6.2
ensures that for every unlock action there is a unique preceding lock action.
But note that the rules allow several lock actions on the same lock to occur in
succession, with no intervening unlock action. We find it convenient to think of a
thread as building up a number of claims on a lock. A lock action adds a claim,
and an unlock action takes one away. Rule 6.1 states that only one thread may
 Investigating Java Concurrency 167

have a positive number of claims on the lock; furthermore, a thread may acquire
multiple claims on a lock and surrenders the lock when and only when it has
given up all claims. Rule 6.2 states that a thread may not release a claim on a
lock it does not hold; the number of claims it has on a lock cannot be negative.
These rules ensure that the synchronization mechanism restricts synchronized
code regions to one thread at a time. At most one thread at a time may have a
positive number of claims on a given lock, and so if several threads need to lock
the same lock, they must compete for it. Only the one with a positive number of
claims on the lock is allowed into its synchronized region; the others must wait.
The function claims returns the number of claims a given thread has on
a given lock. The function synchAction returns the action that the thread is
requesting (if any).
A thread may only continue the execution of its program if it is active;
i.e., not synchronizing. We change the thread module accordingly, by guarding
the rule Execute program with the term active?(Self).
term active?(t): undef?(synchAction(t))

To the rule Execute program, we add the options Synch and Create object. We
add rules for threads’ synchronization actions with locks. A thread synchronizes
with a lock for either a lock action or an unlock action.
rule Synch on ` to perform act: synchAction(Self) := (act, `)
rule Synch: choose obj: Object
choose act: LockActionType
Synch on lock(obj) to perform act

The rule Create object creates an object and associates it with a new lock.
We also modify Create thread to associate each new thread with a new lock.
rule Create object: extend Object with obj
extend Lock with `
lock(obj) := `

The following JLS rules place some guarantees on the contents of a thread’s
working memory before and after synchronization actions.

JLS rules 7 [p. 407] “Let T be any thread, let V be any variable, and let L be
any lock.”
1. “Between an assign action by T on V and a subsequent unlock action by T
on L, a store action by T on V must intervene; moreover, the write action
corresponding to that store must precede the unlock action, as seen by main
memory.”
2. “Between a lock action by T on L and a subsequent use or store action by
T on a variable V , an assign or load action on V must intervene; moreover,
if it is a load action, then the read action corresponding to that load action
must follow the lock action, as seen by main memory.”
168 Y. Gurevich, W. Schulte, and C. Wallace

If a thread t issues an assign action A on a variable v and then issues an

unlock action U l on a lock `, Rule 7.1 postulates that t issues a store action S
on v (with corresponding write action W ) such that A < S < W < U l. Note that
` and v are independent; for an unlock action on a particular lock `, Rule 7.1
applies to all variables v. Thus when a thread releases a claim on a lock, it is
ensured that all its assigned values have been stored out and written to the main
memory.
If a thread t issues a lock action Lk on a lock ` and then issues a use or
store action U S on a variable v, Rule 7.2 postulates that either t issues an
assign action A on v such that Lk < A < U S, or t issues a load action L (with
corresponding read action R) on v such that Lk < R < L < U S. For a lock
action on a particular lock `, Rule 7.2 applies to all variables v. So this rule
ensures that by the time a thread acquires a claim on a lock, all the values
cached in its working memory have been flushed out to main memory. When
a thread acquires a claim on a lock, it acts as if its working memory is empty.
Only values assigned or loaded in after the lock action may be used, and only
values assigned after the lock action may be stored out.
The conditions imposed by these rules have some ramifications for earlier
rules. A read message issued before a lock action cannot be loaded in after the
lock action, but Rule 2.1 dictates that any such message must be loaded in.
Therefore, all pending read messages must be loaded in before a lock action.
Also, a value assigned before a lock action cannot be stored out after the lock
action, but Rule 3.2 dictates that it must be stored out. Therefore, all assigned
values must be stored out before a lock action.
The actions of a lock master consist of granting a claim to a lock and taking
one away.

module Lock: choose among

choose t: Thread: lockOK?(t)
Lock for t
choose t: Thread: unlockOK?(t)
Unlock for t

rule Lock for t: claims(t, Self) := claims(t, Self) + 1

synchAction(t) := undef
do-forall v: Var: usableValue?(t, v)
usableValue?(t, v) := false

rule Unlock for t: claims(t, Self) := claims(t, Self) - 1

synchAction(t) := undef

The terms lockOK? and unlockOK? determine whether a given action is allo-
wed. A lock action for a thread t on a lock ` is allowed only if t is synchronizing
for a lock action on `. Rule 6.1 allows such a lock action only if all threads other
than t have no claims on `. Rules 3.2 and 7.2 require that all previously assigned
values have been stored out. Rules 2.1, 2.3 and 7.2 require that all access messa-
 Investigating Java Concurrency 169

ges have been loaded in, and Rules 4.1 and 7.2 require that all access messages
from volatile variables have been used.
term lockOK?(t):
synchAction(t) = (lock, Self) and ((∀t0 : Thread: t0 6= t) claims(t0 , Self) = 0)
and (not ∃v: Var) freshAssign?(t, v) or (∃m: AccMsg(v, t))
and (not ∃v: VolVar) usableValue?(t, v)
term unlockOK?(t):
synchAction(t) = (unlock, Self) and claims(t, Self) > 0
and (not ∃v: Var) freshAssign?(t, v) or (∃m: UpdMsg(t, v))

5 Prescient Store Actions

A standard store action sends a value to the main memory from the working
memory, where it was put by a preceding assign action. Under certain circum-
stances, Java allows an optimization in which a store action may precede its
corresponding assign action, allowing the value to be transmitted to the main
memory sooner than otherwise possible. This type of store action is called pre-
scient, as it must be known ahead of time what the value of the assign action
will be.
A prescient store action differs from a normal store action in that the value
it sends to the main memory is not the current contents of working memory,
but rather some fresh value. We call the following assign action retroactive, since
the value it puts into working memory is not a fresh value, but rather the value
of the preceding prescient store action. We find it natural to speak of prescient
store actions and retroactive assign actions as distinct from regular store and
assign actions.
We modify the ASM to model prescient store actions. If a given thread has
issued a prescient store action on a given variable, without a corresponding retro-
active assign action, the function presStoreVal returns the value of the prescient
store.
We define rules for the new actions of prescient store and retroactive assign.
Rule 8.4 disallows a store action on v. We also modify the rules WM-EE transfer
and WM-MM transfer , adding prescient store and retroactive assign as options.
rule Store presciently to v: choose v: Var: presStoreOK?(Self, v)
freshAssign?(Self, v) := false
choose val: Value
presStoreVal(Self, v) := val
extend UpdMsg(Self, v) with m
val(m) := val
rule Assign retroactively to v: usableValue?(Self, v) := true
presStoreVal(Self, v) := undef
workingValue(Self, v) := presStoreVal(Self, v)
170 Y. Gurevich, W. Schulte, and C. Wallace

Prescient store actions must obey the following rules set out in the JLS:

JLS rules 8 [p. 408] “Suppose that a store by [a thread] T of [a non-volatile

variable] V would follow a particular assign by T of V according to the rules
of the previous sections, with no intervening load or assign by T of V . The
special rule allows the store action to instead occur before the assign action, if
the following restrictions are obeyed:

1. If the store action occurs, the assign is bound to occur.

2. No lock action intervenes between the relocated store and the assign.
3. No load of V intervenes between the relocated store and the assign.
4. No other store of V intervenes between the relocated store and the assign.
5. The store action sends to the main memory the value that the assign action
will put into the working memory of thread T .”

If a thread t performs a prescient store action P S on a variable v, there is

a retroactive assign action RA > P S by t on v such that RA assigns the value
stored out at P S (Rules 8.1 and 8.5). Rules 8.1 and 8.5 ensure that the value t
sends to main memory via a prescient store action ends up in t’s working memory
via a following retroactive assign action.
Furthermore, there is no action Act by t, where Act is a lock action or a
load or store action on v, such that P S < Act < RA (Rules 8.2-4). These rules
ensure that relocating t’s store action on v (i.e., making it prescient) does not
affect program behavior, in the sense that the effects of a prescient store action
on t’s working memory and the main memory are no different from those of the
corresponding non-prescient store action.
The terms presStoreOK? and retroAssignOK? determine whether a prescient
store action or retroactive assign action is allowed by the JLS rules, respectively.
Rules 2.5 and 3.2 allow t to store to v presciently only if every AccMsg from v to
t has been loaded in and there is no retroactive assign pending. Rule 8 restricts
prescient store actions to volatile variables, and allows a retroactive assign ac-
tion only if there is a preceding prescient store action without a corresponding
retroactive assign action.
term presStoreOK?(v): not VolVar(v) and undef?(presStoreVal(Self, v))
and (not ∃m: AccMsg(v, Self))
term retroAssignOK?(v): def?(presStoreVal(Self, v))

Notice that t may issue a use action U on v between P S and RA. If the
presciently stored value were to appear in t’s working memory before t put it
there at RA, t could end up using the presciently stored value prematurely. To
prevent this, Rule 8.3 prohibits t from loading in from v between P S and RA.
For the same reason, albeit less obviously, Rule 8.4 prohibits any lock action for
t between P S and RA. This is needed because if a lock action for t were to occur
and then t were to use v between P S and RA, Rule 7.2 would require a load
action between P S and U , and such a load action is forbidden by Rule 8.3.
 Investigating Java Concurrency 171

Relocating a store action makes sense only if there is no attempt to store

between the relocated (prescient) store) action and its (retroactive) assign ac-
tion. If t were to issue a store action S on v (with corresponding assign action
A) between P S and RA, we would get the following undesirable result. S would
follow P S, but the order of the corresponding assign actions would be different:
RA would follow A. Thus the value stored through P S would be newer (i.e., as-
signed more recently) than the value stored through the following action S. But
Rule 2.5 would dictate that the newer value be overwritten in main memory by
the older value. To prevent this, Rule 8.4 prohibits t from storing out a value of
v between P S and RA.
Some of the rules introduced in previous sections (namely, 3.2–5 and 7.1–
2) refer to store and assign actions and so must be modified to accommodate
prescient store and retroactive assign actions. The rules as they appear in the
JLS are not modified when prescient store actions are introduced.
We modify the terms lockOK? and loadOK? so that they enforce Rules 8.2–
8.4. If a thread t has stored out presciently to v but has not yet performed the
corresponding retroactive assign action, Rule 8.2 disallows any lock action for
t, Rule 8.3 disallows a load action by t on v, and Rule 8.4 disallows any store
action by t on v. To lockOK?(t), loadOK?(t) and storeOK?(t) we add the conjunct
(∀v: Var) undef?(presStoreVal(t, v).
Rule 8.1 ensures that if a prescient store action on a variable occurs, a cor-
responding retroactive assign action will occur sometime in the future. As with
previous rules, this rule can be flouted by delaying a retroactive assign action
indefinitely. We restrict attention to runs that avoid this situation.

6 Thread Objects

We now consider how Java represents threads as objects. An object is an instance

of a programmer-defined type called a class. Objects are created dynamically
during a thread’s computation. A class defines the state variables of its instances
and the methods (operations) that may be invoked upon them. Each thread is
represented by an object which is an instance of the class Thread. As there is a
one-to-one correspondence between threads and their representative objects, we
may identify a thread with its object. Information on a thread’s state is held in
its object. The methods defined in Thread allow a thread to access or modify its
own state information and that of other threads.
We modify the ASM to model operations on threads. Since we identify a
thread with its Thread instance, a member of the universe Thread is both a
thread and a Thread instance. Note that Thread instances are members of both
Thread and Object.
After a thread is created, it may be started by invoking the method start
upon it. Once started, a thread is alive until it stops, which occurs when the
thread’s execution engine terminates or the method stop is invoked upon it.
If stop is invoked on a thread, the thread throws an exception, signaling the
occurrence of an unexpected event. Once a thread stops, it is no longer alive
172 Y. Gurevich, W. Schulte, and C. Wallace

and cannot be restarted. It is possible to stop a thread that has not started; if
this happens, the thread will never become alive. It is not stated explicitly in
the JLS what happens if start is invoked upon a thread that is already alive.
However, the intent seems to be that the invoker of start throws an exception
and nothing happens to the invokee [12].
We define functions started? and stopped? to represent the status of each
thread. We also define rules Start and Stop, which update them appropriately.
The JLS rules impose some obligations on threads which they may not be
able to fulfill after they have stopped. By Rules 2.1 and 2.3 (following the strict
interpretation of “uniquely paired”), every access message must be loaded in.
By Rule 4.1, every load action on a volatile variable must be followed by a
corresponding use action, and by Rule 4.2, every assign action to a volatile
variable must be followed by a corresponding store action. The JLS does not
make it clear whether these obligations extend to threads that have stopped.
Furthermore, it is not clear which choice is the more reasonable one. For instance,
in some cases it may be desirable to have the last assign actions of a thread stored
out, while in other cases it may be clearly undesirable; for example, consider a
thread which is stopped because it is computing erroneous results. Officially, this
is still an open issue [21]. For simplicity, we take the view that stopped threads
are not obligated to do any further actions.
The JLS rules impose some obligations on variables which may not be fulfilled
by the time the execution of all threads terminates; in particular, by Rule 2.4,
every store message must be written. Thus it may still be necessary for variables
to write some update messages in even after all threads have terminated.
A thread that is alive can be suspended, in which case it remains alive but its
execution engine does nothing. A thread is suspended when some thread (pos-
sibly itself) invokes the method suspend upon it. A suspended thread resumes
(becomes unsuspended) when some other thread invokes the method resume
upon its Thread instance. It may be useful to suspend a thread when it can-
not make progress by itself, freeing resources for other threads that can make
progress. Can a suspended thread issue load or store actions? Officially, this is
another unresolved issue [21]. Here we choose the more permissive interpretation,
allowing suspended threads to issue read and load actions.
We add the function suspended? to represent the suspended status of each
thread. We also define rules Suspend and Resume, which update this function.
A thread may be marked as a daemon. Such threads are intended to exe-
cute only in conjunction with non-daemon threads, performing “housekeeping”
actions which assist non-daemon threads, but no actions useful in themselves.
For this reason, once all non-daemon threads have stopped, execution of all
threads halts. A program starts with a single non-daemon thread. A new thread
starts with the daemon status of the thread that creates it, but its status can
be changed via the setDaemon method.
To our ASM we add the function daemon?, which determines a thread’s
current daemon status, and the rule Set daemon status, which updates this fun-
ction. We also modify the rule Create thread . A thread is added to the Object
 Investigating Java Concurrency 173

universe and is initialized as unstarted, unstopped and unsuspended. It inherits

the daemon status of the thread that created it.
We define a rule Invoke thread method , which chooses among the options
Start, Stop, Suspend and Set daemon status. We modify the Execute program
rule, adding the option Invoke thread method . To active? we add the conjuncts
not suspended?(t) and undef?(synchAction(t)). We also modify the Thread mo-
dule, guarding Execute program with alive?(Self) and not HaltAllThreads?
We define the term alive? to represent the running status of a thread. Also,
the term HaltAllThreads? determines whether all non-daemon threads have ter-
minated.
term alive?(t): started?(t) and not stopped?(t)
term HaltAllThreads?: (∀t: Thread: alive?(t)) daemon?(t)

We modify the modules for variables and locks. If all non-daemon threads
have terminated, the only action a variable may take is to write in update mes-
sages, so in the Var module we guard the read option with not HaltAllThreads?.
There are no actions for locks to take once all non-daemon threads have termi-
nated, so in the Lock module we guard the lock and unlock options with the
term not HaltAllThreads?.

7 Waiting and Notification

While a thread holds a particular lock, other threads in need of that lock are
not allowed to proceed. In certain cases, a thread holding a lock may reach a
state from which it cannot progress by itself; in such a case, suspending the
thread may not be a solution, as the thread would still hold the lock even when
suspended. It would be desirable for the thread to give up control and release
its locks temporarily, so that other threads may acquire them.
Java has built-in support for this, through the mechanisms of waiting and
notification. If a thread has a claim on the lock of an object, it may release
all its claims on the lock and disable itself by invoking the method wait of
class Object on the object. Another thread may signal to the waiting thread
that further progress is possible through the notify or notifyAll methods. In
this way, a waiting thread may resume execution when it is possible, without
repeatedly checking its state. Each object has a wait set, empty when the object
is created, which contains all threads waiting for the lock on the object.
The method wait is invoked by a thread t on an object obj. The thread t must
hold the lock of obj; if it does not, an exception is thrown. Let k be the number
of claims t has on the lock of obj; then k unlock actions are performed, and t is
added to the wait set of obj and disabled. When t is re-enabled, it attempts to
regain k claims on the lock; it may need to compete with other threads to do
this. Once the lock claims are restored, the wait method terminates.
We modify the ASM to model waiting and notification actions. A thread goes
through several phases while it waits on a lock. First it synchronizes with the
174 Y. Gurevich, W. Schulte, and C. Wallace

lock to release all claims on the lock. Then it waits to be notified by another
thread. Once it is notified, it synchronizes with the lock to regain all the claims it
released. For a thread that is waiting on a lock, the function waitMode gives the
current phase of the thread, the function waitLock gives the lock on which it is
waiting, and claimsToRegain gives the number of claims that it has (temporarily)
released.
We introduce a rule for a thread’s actions while waiting on a lock. Waiting
involves synchronization with the lock during the unlock and relock phases. Note
that the waiting thread does not change its own mode from wait to relock; this
is done by another thread, through a notification.
rule Start to wait: choose obj: Object: claims(Self, lock(obj)) > 0
waitMode(Self) := unlock
waitLock(Self) := lock(obj)
claimsToRegain(Self) := claims(Self, lock(obj))
rule Continue to wait:
if waitMode(Self) = unlock then
if claims(Self, waitLock(Self)) > 0 then
Synch on waitLock(Self) to perform unlock
else waitMode(Self) := wait
elseif waitMode(Self) = relock then
if claims(Self, waitLock(Self)) < claimsToRegain(Self) then
Synch on waitLock(Self) to perform lock
else waitMode(Self) := undef

The method notify is also invoked by a thread t on an object obj. The

thread t must hold the lock of obj; if it does not, an exception is thrown. If the
wait set of obj is not empty, one thread is removed from it and enabled. The
method notifyAll operates similarly but removes and enables all threads from
the wait set. Neither method releases any of t’s claims on the lock of obj, so there
is no guarantee that the lock will be made available to the notified threads; this
is left up to the programmer.
We introduce rules for notification. The method notify operates on a par-
ticular thread waiting on a particular lock; the method notifyAll operates on
all threads waiting on a particular lock.
rule Notify t: waitMode(t) := relock
rule Notify:
choose obj: Object
choose t: Thread: waitMode(t) = wait and waitLock(t) = lock(obj)
Notify t
rule NotifyAll :
choose obj: Object
do-forall t: Thread: waitMode(t) = wait and waitLock(t) = lock(obj)
Notify t
 Investigating Java Concurrency 175

When a thread stops, the method notifyAll is invoked upon its object.
This allows a straightforward implementation of a technique called thread joins.
Consider a scenario where a thread t0 is expected to start executing only after
another thread t has stopped. One way of implementing this is to have t0 join
with t by waiting until t has stopped before proceeding. Since notifyAll is
invoked when a thread stops, t0 may join with t by simply waiting on t’s object.
Since stopping a thread involves invoking notifyAll on its representative object,
we change our ASM rule for stopping threads accordingly.
We point out a subtle issue. It is not clear from the JLS whether notifica-
tion precedes stopping or vice versa when a thread stops. However, it seems
reasonable that if stopping and notification do not occur as a single action, then
notification must precede stopping (at least in the case where a thread stops
itself). This leaves open the possibility that other threads may be notified of the
thread’s stop action before the thread has actually stopped, particularly if the
notification process takes a relatively long time.
Threads may execute waiting or notification actions. Execution of a Java pro-
gram does not proceed while a thread is waiting, so to the term active? we add the
conjunct waitMode(t) 6= wait. To the rule Execute program, we add the options
Start to wait, Notify and NotifyAll . We also add the guard def?(waitMode(Self));
if this evaluates to true, then the rule Continue to wait fires; otherwise, one of
the other options is chosen.

8 Conclusion
The Java concurrency model is currently in a state of flux. Researchers are pro-
posing modifications to the model, to allow common compiler optimizations and
programming practices that are currently prohibited by the model [19]. It is our
belief that a satisfactory successor to the current model must start with a firm
foundation. The specification of the model must not be subject to multiple in-
terpretations, as is the description in the JLS. It is equally important that the
specification be accessible to programmers who wish to use concurrency. While
multithreaded programming is inherently complicated, the method of documen-
tation should not exacerbate the problem.
We feel that this work has several things to offer the Java community. Our
alternative view of Java concurrency brings to light some issues that might other-
wise remain hidden in the JLS description. As the ASM model is mathematically
precise, it can stand as an unambiguous cross-platform standard for the langu-
age. Our account of Java concurrency has an imperative flavor that is familiar to
programmers, yet we are not restricted to a purely imperative approach; we are
free to use a declarative style wherever it is more natural to do so. By implemen-
ting the concurrency model (albeit in a completely abstract way) we can readily
see how the various constraints of the JLS definition interact with one another.
This is much less obvious if constraints are presented in isolation, as they are in
the JLS. Finally, programmers interested in multithreaded Java can explore the
model by using the ASMGofer prototype. We believe that ASMs are a useful
specification tool for both the current and future models of Java concurrency.
176 Y. Gurevich, W. Schulte, and C. Wallace

References
1. I. Attali, D. Caromel, and M. Russo. A formal executable semantics for Java. In
Proceedings of the OOPSLA ’98 Workshop on the Formal Underpinnings of Java,
1998.
2. E. Börger. Why use Evolving Algebras for hardware and software engineering? In
Proceedings of SOFSEM, 1995.
3. E. Börger and W. Schulte. A programmer friendly modular definition of the seman-
tics of Java. In J. Alves-Foss, editor, Formal Syntax and Semantics of Java. Springer,
1998.
4. P. Cenciarelli, A. Knapp, B. Reus, and M. Wirsing. ¿From sequential to multi-
threaded Java: An event-based operational semantics. In M. Johnson, editor, Alge-
braic Methodology and Software Technology, pages 75–90. Springer, 1997.
5. E. Coscia and G. Reggio. A proposal for a semantics of a subset of multi-threaded
“good” Java programs. In Proceedings of the OOPSLA ’98 Workshop on the Formal
Underpinnings of Java, 1998.
6. A. Gontmakher and A. Schuster. Java consistency: Non-operational characterizati-
ons for Java memory behavior. Technion/CS Technical Report CS0922, 1997.
7. J. Gosling, B. Joy, and G. Steele. The Java Language Specification. Addison-Wesley,
1996.
8. Y. Gurevich. Evolving Algebras: an attempt to discover semantics. In G. Rozenberg
and A. Salomå, editors, Current Trends in Theoretical Computer Science, pages 266–
292. World Scientific, 1993.
9. Y. Gurevich. Evolving Algebras 1993: Lipari guide. In E. Börger (editor), Specifica-
tion and Validation Methods, Oxford University Press, 1995, 9–36.
10. Y. Gurevich. May 1997 draft of the ASM guide. Available at [22], 1997.
11. Y. Gurevich, W. Schulte and C. Wallace. Investigating Java concurrency using
Abstract State Machines. Technical Report 2000-04, Department of Computer &
Information Sciences, University of Delaware, 1999.
12. C. Horstmann and G. Cornell. Core Java 1.1, volume II: Advanced Features. Sun
Microsystems Press, 1998.
13. Sun Microsystems. Java technology home page. http://java.sun.com/.
14. A. Jolin. Java’s atomic assignment: The key to simpler data access across threads.
Java Report 3(8), 27–36, 1998.
15. D. Lea. Concurrent Programming in Java. Addison-Wesley, 1997.
16. M. MacBeth, K. McGuigan and P. Hatcher. Executing Java threads in parallel in
a distributed memory environment. In Proceedings of IBM CASCON, 1998.
17. S. Oaks and H. Wong. Java Threads. O’Reilly and Associates, 1997.
18. G. Plotkin. Structural Operational Semantics (Lecture notes). Technical Report
DAIMI FN-19, Aarhus University, 1981.
19. W. Pugh. Fixing the Java memory model. In Proceedings of ACM Java Grande,
1999.
20. Joachim Schmid. ASMGofer home page. http://www.tydo.de/AsmGofer/.
21. G. Steele. Personal communication.
22. Univ. of Michigan. ASM home page. http://www.eecs.umich.edu/gasm/.
23. Univ. of Paderborn. ASM home page. http://www.uni-paderborn.de/cs/asm/.
 Verifying Compilers and ASMs
or
ASMs for Uniform Description of
Multistep Transformations

Gerhard Goos and Wolf Zimmermann

Universität Karlsruhe
Fakultät für Informatik
{ggoos,zimmer}@ipd.info.uni-karlsruhe.de

Abstract. A verifying compiler ensures that the compiled code is al-

ways correct but the compiler may also terminate with an error mesage
and then fails to generate code. We argue that with respect to compi-
ler correctness this is the best possible result which can be achieved in
practice. Such a compiler may even include unverified code provided the
results of such code can be proven correct independently from how they
are generated. We then show how abstract state machines (ASMs) can be
used to uniformly describe the dynamic semantics of the programs being
compiled across the various intermediate transformation steps occurring
within a compiler. Besides being a convenient tool for describing dyna-
mic semantics the fact that we do not have to switch between different
descriptional methods is found to be extremely useful.

1 Introduction

Many statistics attribute more than 50% of all software failures to problems in
requirement engineering whereas programming errors account for a much smal-
ler percentage, cf. [23,49]. But what if everything in software development was
correct except that errors were introduced during compilation? This case is not
as infrequent as some people think, cf. [52,37,12,13,14]. Especially programs of-
ten show different behavior with and without optimization. The generated code
should exactly behave like the source program; this is of utmost importance
especially in safety-critical applications. The requirement can only be ensured
by a compiler which verifiably produces correct code. But despite many efforts,
cf. e. g. the efforts for validating Ada-compilers, no compiler on the market for
a realistic programming language such as Ada, C, C++ or Java is fulfilling
this requirement. It is therefore no surprise that safety surveillance institutions
such as the ‘‘Technische Überwachungsvereine’’ (TÜV) in Germany often do not
accept software written in high-level languages but instead check the resulting
assembly code and require the use of non-optimizing compilers.
McCarthy and Painter, [36], were first considering correctness of compi-
lation but for arithmetic expressions only. Many people dealt with the problem

Y. Gurevich et al. (Eds.): ASM 2000, LNCS 1912, pp. 177–202, 2000.
c Springer-Verlag Berlin Heidelberg 2000

178 G. Goos and W. Zimmermann

thereafter as discussed in Sect. 9 but nobody succeeded in producing a correct

compiler for a realistic programming language with a typical machine language
as target. In our opinion two main causes have lead to this:
– The problem was considered in an idealistic, i. e. mathematical setting in
which common data types such as integer and float were considered like
integers and reals in mathematics, i. e. the range and precision limitations
on computers were ignored. In the same vain storage limitations at compile-
or run-time and other types of resource limitations were ignored.
– The formal methods chosen for describing the source and target language
and of the intermediate languages in the compiler made the treatment of
realistic languages too difficult. As a consequence attention was restricted to
relatively small programming languages disregarding the complexities and
dark sides of realistic languages. For the same reason code generation was
mostly restricted to stack machines using the data types of the source lan-
guage. Also, the software architecture for correct compilers in past efforts
was specially adapted to the verification needs.
In the present contribution we first discuss the notion of correctness. We ar-
rive at the conclusion that the main goal should be the correctness of the target
program produced, not primarily the correctness of the compiler itself. This in-
sight allows for reusing the results of compiler research of the past decades and
for using conventional compiler architectures. This discussion is closely related
to the problem how to deal with resource limitations of all kinds. We then intro-
duce Gurevich’s abstract state machines (ASMs) and show that this method
is a suitable means for formalizing the (dynamic) semantic of programming lan-
guages on all levels of representation. The advantages include the facts that we
do not have to specially adapt the compiler architecture to verification needs,
and that we do not have to switch between different formalisms for describing
semantics during the compilation process.

2 Correctness of Compilation
We consider a pair (PL,ML) consisting of an imperative or object-oriented pro-
gramming language PL such as Ada, C, C++, Sather-K, [28], or Java and
a microprocessor such as the DEC Alpha represented by its machine language
ML. For simplicity we restrict ourselves to sequential programming languages.
The translation of a source program π in PL to a target program π 0 = C(π)
in ML is certainly correct if π 0 shows the same behavior as π, i. e. on all inputs
π 0 is producing the same outputs as π.
A state of a computation is called observable if it is the initial or final state or,
if the operation leading into this state is an input or output operation. Observa-
ble states are the only states representing an effect noticeable in the environment
of a computation.1 If we consider states q, q 0 of the execution of π and π 0 res-
pectively as sets of state variables then the requirement same behavior asks for
1
The programmer or compiler writer may, at his discretion, designate further states,
e. g. procedure entries or exits, as observable.
 Verifying Compilers and ASMs 179

identifying corresponding state variables in q and q 0 holding the same values in

corresponding observable states. We represent this requirement by a relation ρ
between states of π 0 and π.
Correctness of translation then requires that a sequence q0 , q1 , . . . , qi , . . . of
observable states in the execution of π translates into a sequence q00 , q10 , . . . , qi0 , . . .
of observable states of π 0 and ρ(qi ) = qi0 holds for corresponding states.
The behavior of π may be indeterministic: Fig. 1 shows the acyclic graph of
possible execution paths for the program
do true -> x := 1
[] true -> x := 0
od
in Dijkstra’s language of guarded commands. The implementation may choose

x := ?

x := 1 x := 0

x := ?

x := 1 x := 0

x := ?

Fig. 1. Execution paths for an indeterministic program

an arbitrary path:
do true -> x := 1
od
is a correct translation of this program. For this reason we can only require that
each sequence of observable states of the translated program π 0 = C(π) possesses
a corresponding state sequence of the source program π; the converse is not true.
Ideally we want to preserve the termination behavior of the source program.
For reasons discussed in the next section this is not possible in a realistic envi-
ronment: we must admit that the target program π 0 is prematurely terminating
due to excessive resource demands. Thus, during compilation we cannot pre-
serve total but only partial correctness. Altogether this leads to the following
definition of correct translation:
Definition 1. A target program π 0 is a correct translation of a source program
π, π 0 = C(π), if for all inputs one of the following conditions applies:
180 G. Goos and W. Zimmermann

– To each sequence S 0 = q00 , q10 , . . . , qk0 , . . . of observable states of π 0 there is a

sequence S = q0 , q1 , . . . , qk , . . . of observable states of π with qi0 = ρ(qi ) for
i = 0, . . . , k, . . . If S 0 regularly terminates after k steps then also S.
– To each sequence S 0 = q00 , q10 , . . . , qk0 , qk+1
0
of observable states of π 0 termina-
0
ting with an error message due to a resource violation in state qk+1 there is
0
a sequence S = q0 , q1 , . . . , qk , . . . of observable states of π with qi = ρ(qi ) for
i = 0, . . . , k.
In both cases the state sequences must be complete with respect to observability.

3 Dealing with Resource Limitations

Mathematically speaking the length of nearly all programs exceeds k lines where
k is an arbitrarily large number, e. g. the number of atoms in the universe. A
compiler C running on a computer in the real world cannot cope with such
programs. It must instead be prepared for refusing source programs because the
compilation would demand more resources, e. g. storage, than is available. We
could even argue that
begin
print("program not compilable due to excessive resource requirements")
end
is a correct compiler for arbitrary pairs (PL,ML) without regard how those
languages are defined.
Thus, correct compilation is primarily an engineering issue: we want to pro-
duce this error message for as few programs as possible and deliver a correct
target program for the remaining source programs. But requiring that a compi-
ler is delivering correct target programs for all source programs is unrealistic.
For becoming realistic we must restrict our goal: We do not want to construct
correct compilers but are instead interested in verifying compilers: they ensure
the correctness of the target program in case they deliver one; and, of course,
we want to make the set of correctly translated programs as large as possible.
For similar reasons a compiler cannot be kept responsible if at runtime of
the generated program an integer multiplication overflows, or if the number and
size of objects to be allocated exceeds certain bounds, or if the operating system
on the target machine cannot cope with the number of simultaneously acces-
sed external devices and communication lines, etc. In a realistic environment
the programmer of the source program must take precautions against these pro-
blems, not the compiler writer. For all such reasons we must allow not only the
compiler but also the generated target program to prematurely terminate with
an error message due to violation of resource limitations.
If the processor running the compiler or the target program is too slow to
achieve the desired result in acceptable time we also have a violation of resource
limitations. In practice, when real-time constraints are exceeded, it is infeasible
to decide whether the programmer was demanding too much or the compiler
generated bad code or the target processor was too slow. We thus do not count
execution speed of the target program amongst the correctness criteria.
 Verifying Compilers and ASMs 181

4 Compiler Architecture

Let us assume that we have subdivided a compilation C : π → π 0 into a sequence

of two or more translation steps C 0 : π → π 00 , C 00 : π 00 → π 0 , and that each step
yields correct results according to definition 1. Then, as long as there are no
resource violations, each observable state q 0 of π 0 has a corresponding observable
state q 00 with q 0 = ρ00 (q 00 ) and in turn there is an observable state q with q 00 =
ρ0 (q) and hence q 0 = ρ00 (ρ0 (q)) with appropriately defined relations ρ0 and ρ00 .
If we define ρ(q) = ρ00 (ρ0 (q)) then definition 1 is therefore also fulfilled for the
sequence C = C 0 ; C 00 of these steps.
We term this property vertical compositionality of translation steps. It allows
the introduction of an arbitrary number of intermediate program representati-
ons π 00 , . . . into the compilation process. If each individual compilation step is
correct then also the sequential composition of these steps. Traditional compi-

Analysis Mapping Coding

ass. + phases
symbols syntax semantics code
transf. optim.
gen. linking

symbol structure attributed code target

source structure intermediate
sequence tree graph program representations
tree

symbol- definition global

table table data

Fig. 2. Compiler architecture

ler architecture is using vertical compositionality for decomposing compilations

into a sequence of simpler steps each of which can be dealt with by appropriate
methods and tools, cf. Fig. 2 and e. g. [54,39].
Actually, the first row of Fig. 2 shows a decomposition of a compiler into modules
only, not necessarily into sequential steps; however, locally for each element of an
intermediate representation the phases follow each other sequentially from left to right.
Conceptually it is thus justified to consider the phases as sequential steps although the
actual merging in time may be much more complex.
The semantics of programming languages and thus the behavior of program
executions is usually defined by attaching meaning to certain phrases of a pro-
gram as given by the (context-free) grammar of the language. The meaning of
182 G. Goos and W. Zimmermann

larger phrases and the program as a whole is then derived by substituting the
semantics of subphrases into the semantics of larger ones. This property is called
horizontal compositionality (of language elements).
There may be syntactic and semantic limitations for the admissibility of sub-
phrases. A program is termed statically correct if these conditions are fulfilled.
Compilers check these conditions during semantic analysis and refuse to gene-
rate code for incorrect programs. When discussing verifying compilers we may
restrict attention to statically correct programs.
Horizontal compositionality permits to deal with one language element at
a time during the vertically decomposed translation steps. It thus adds to the
structuring of the compiler. The method is, of course, not applicable before we
have determined the phrase structure and during optimization steps which on
purpose consider several language elements together.
This raises the question how we can certify the correctness of the abstract
syntax tree for a given program text. The requirement same behavior does not
help since the dynamic behavior is only defined for the phrases visible in the
syntax tree whereas the program text is only serving as a vehicle for deriving
this tree. Similarly, as soon as we have decided which instruction sequences cor-
rectly represent the intended behavior of phrases in the source program we have
established the correctness of the target program. The remaining assembly and
linking phase only decides about the proper encoding of the resulting instruc-
tions but does not change the semantic meaning. So to say, the assembly phase
acts like the inverse of the transformation source text → syntax tree, but on a
different level of abstraction.
Thus, basically only the mapping phase in Fig. 2 is concerned with the se-
mantic correctness of compilation. The remaining phases deal with decoding and
encoding the structures which carry semantic meaning. Especially the transla-
tion process seen as a mapping between source and target semantics is confined
to the mapping phase only. This insight has been ignored for a long time in the
discussion about correct compilation.
For lexical analysis, parsers, code selectors and other phases routinely gene-
rators are used for deriving the corresponding compiler phase. Verifying such
phases would require to verify the correctness of the corresponding generators,
a very heavy and burdensome task. Previous work on correct compilation thus
relied on hand-writing these phases and verifying the hand-written code. In the
next section we show that there is an easier way of dealing with this problem
when we relax the requirements and only request that (for certain source pro-
grams) the output of a phase is correct instead of requesting that the code of
the compiler phase is producing correct results under all circumstances. Under
this condition we can use unverified generators and other tools as in traditional
compilers and nevertheless trust the resulting code.
 Verifying Compilers and ASMs 183

5 Verifying Compilers: The Case for Program Checking

Program checking was originally introduced for algorithms by Blum and Kannan,
[1]. We present it here as in [26] for verifying the outputs of systems such as com-
pilers or parts thereof.
Let C be a program implementing a function f : I → O with precondition
P (x) and postcondition Q(x, C(x)) on input x. Let checker (x, y) : Bool be a
function that returns the value of Q(x, y). Consider the program

function C 0 (x : I) : O
y := C(x);
if checker (x, y) then return y
else abort;
end;

and assume that C 0 (x) does not abort. Then the result Q(x, C 0 (x)) holds if
checker is correctly implemented. Thus, if C 0 does not abort then the result
y = C(x) fulfills its postcondition no matter how it was computed. In particular,
the partial correctness of C 0 does not depend on the partial correctness of C.
Therefore we only need to verify checker but not C for getting a verified result.
In practice the verification of checker is often much simpler than the verifi-
cation of C. This is particularly true if C contains a lot of optimizations which,
from the correctness point of view, are irrelevant as long as they maintain some
simple correctness conditions. E. g. the register allocator within a compiler is
correct as long as the values assigned to the same register have non-overlapping
lifetimes; an optimizing register allocator must maintain this condition while it
is devoting most of its effort to avoiding spill code as far as possible; its quality
in this respect does not influence its correctness.

Table 1. Lines of program code to verify for a program-checked IS front-end

C/Sather-K Binary Prog.

Lines Byte Byte

Generators COCKTAIL 110.000 2.4 MB 1.2 MB

Generated C Code
22.000 600 KB 300 KB
Impl. IS-Frontend

500 (Parser) 14 KB
Checker (Sather-K) +100 (Compare) 3 KB 200 KB
+700 (AST) 30 KB
184 G. Goos and W. Zimmermann

Table 2. Lines of program code to verify for example back-end

Modula/Sather-K Binary Prog.

Lines Byte Byte

Generator BEG (Modula) 35.000 2 MB 1.1 MB

Generated C Code
18.000 400 KB 500 KB
Impl. MIS Code-Selection

500 (Parser)
Checker (Sather-K) +300 (Checker) 200 KB
+400 (MIS)

Industrial: ANDF ⇒ ST9 140.000 6.4 MB 3.5 MB

We applied this approach to a compiler front-end for a C subset IS, [19,

32]. Details can be found in [31]. Tab. 1 shows the results. Tab. 2 shows similar
results for the code selection phase using the back-end generator BEG, [22,21].
The checker has been written in the object-oriented language Sather-K, [28]; a
subset of Sather-K comparable to Java will be the first realistic language for
which we expect to present a verifying compiler. As can be seen in both cases
the verification effort for the checker is by orders of magnitude lower than for
the generator or the generated compiler phase.

6 What Has to Be Verified

Program checking only verifies that a result y obeys its postcondition Q(x, y).
Nevertheless the result could be wrong because the postcondition Q does not
ensure the required properties of y. For a verifying compiler we must distinguish
several proof obligations:
1. Verification of the compiling specification: The compiling specification de-
scribes the correspondence of source and target language in formal terms.
Source and target languages are usually given by informal descriptions, e. g.
an ISO-norm for the source language and processor descriptions by hardware
manufacturers.
2. Verification of the compiler specification: This specification introduces the
various compiler phases, intermediate representations and data structures,
e. g. as in Fig. 2. Formally it must be verified that vertical composition of the
compiler phases leads to the mapping given by the compiling specification.
3. Verification of the compiler implementation: A proof is required that the va-
rious parts of the compiler architecture are correctly implemented according
to the compiler specification.
 Verifying Compilers and ASMs 185

4. Verification of the compiler encoding: No matter whether the compiler is

implemented as a program in a high-level language or directly in assembly
code hardware only accepts bit sequences as inputs. Thus we have to ensure
that the symbolic code is properly encoded by bit sequences as required by
the hardware.
The compiling specification can only be verified by hand; verification tools
cannot be applied since the original descriptions cannot be explored by formal
means.
In practice compiling and compiler specification are usually combined. Then
the specifications underlying the compiler phases such as the regular expression
describing the lexemes, the context-free grammar, etc. must be verified by hand.
This combination reveals that the specification of the mapping phase relating
notions and concepts of the source and the target language is the crucial part of
all these specifications.
As far as theorem provers can be applied the Verifix project relies on the use
of PVS, cf. [20]. Only the verification of the implementation can be (partially)
replaced by program checking.
Verification of the encoding can be achieved by program checking based on
the specification of the mapping of symbolic assembly code to binary code. In
principle also the results of the linking phase and of the system loader must be
ckecked; this is presently not part of our project.
Compiler phases and program checkers are programs themselves and their
correctness depends on the correctness of the compilers used to translate them.
To break the cycle that a verifying compiler assumes the existence of another
verifying compiler, a suitable bootstrap process is discussed in [27].

7 Abstract State Machines

So far we have discussed what has to be done for achieving a verifying compiler
and argued that such a compiler could be built along the guidelines of traditio-
nal compiler architecture. Underlying was the assumption that we have formal
description methods at our disposal by which we can formalize the requirement
same behavior and verify that it is maintained during a compilation.

q q
i i+1

ρ ρ

q'
q'
i i+1

Fig. 3. Corresponding observable states

The requirement asks for establishing the relation ρ between corresponding

observable states as in Def. 1; the diagram Fig. 3 must commute. To this end
186 G. Goos and W. Zimmermann

we must look into the details of the representation of states and how they are
dynamically transformed by executing statements and instructions respectively.
According to our experience abstract state machines, [29, ], provide adequate
means to formally handle this problem.
An Abstract State Machine (ASM) is a tuple (Σ, ΦInit , ΦInv , Trans) where Σ
is a signature, ΦInit and ΦInv are sets of Σ-formulas (the initial and invariant
conditions), and Trans is a finite set of transition rules. A transition rule is a
pair (ϕ, Updates), denoted by
if ϕ then Updates
where ϕ is a Σ-formula and Updates is a set of pairs t1 := t2 of Σ-terms, called
updates. The set of states is the set Alg(Σ) of Σ-algebras that satisfy the Σ-
formulas in ΦInv , i. e. q is a model for ΦInv in the sense of logic, q |= ΦInv ; [[·]]q
denotes the interpretation function of symbols of Σ in Σ-algebra q. A state is
initial iff q |= ΦInit .
Remark 1. This definition of ASMs slightly differs from the usual definition.
Instead of defining the states to be Σ-algebras, we define them as sets of Σ-
algebras that satisfy a set of formulas. By this modification we can prove desired
properties of ASMs by logical calculi more easily.
The state transition relation → is based on transition rules. Intuitively, a
transition rule (ϕ, Updates) fires if its condition ϕ is satisfied. Formally speaking
the condition is true in state q iff q |= ϕ. The updates t1 := t2 of the rule are
then executed and change the interpretation of some symbols in the state. A
state q is final iff no rule fires in q.
Let Exec(q) be the set of updates executed in state q. Exec(q) is consistent iff
. . .
for all2 f (t1 , . . . , tk ) := t, f (t01 , . . . , t0k ) := t0 , q |= t1 = t01 ∧ · · · ∧ tk = t0k ⇒ t = t0
holds. Then the state  transition relation q → q̄ is defined by

[[t]]q ∃ f (t1 , . . . , tn ) := t ∈ Exec(q)
[[f ]]q̄ (a1 , . . . , an ) = with [[ti ]]q = ai , i = 1, . . . , n


[[f ]]q (a1 , . . . , an ) otherwise
for each k-ary function symbol f of Σ iff Exec(q) is consistent, q̄ |= ΦInv and
[[T ]]q = [[T ]]q̄ for all sorts of Σ (i.e. the interpretation of sorts is not changed by
the state transition).

8 Compilers as Multistep Transformation Systems

In this section, we focus on proving the correctness of the compiling specification,
i. e. that the observable behavior of a program is preserved during compilation.
We assume that the dynamic semantics of a program is given by an ASM (not
only its observable behavior).
For sequential languages and processors a state transition q → q̄ can be split
into two parts:
2 .
t1 = t2 denotes a Σ-equation.
 Verifying Compilers and ASMs 187

1. determine the flow of control, i. e. the next state q̄;

2. execute the remaining updates occurring during this transition.

The next statement or instruction (after a conditional jump) q̄ possibly depends

on a Boolean expression with a value already known in state q but unknown
during compilation. q̄ is represented in language and processor descriptions by
distinguishing a state variable next task indicating the next state. As long as
we solely consider the flow of control, its refinements and its translation we can
abstract from the details of ASMs and concentrate on this state variable only.
This leads to the notion of state transition systems (STS) as defined in Sect. 8.1.
State transition systems deal with dynamic sequencing of states. When a
compiler transforms the state sequence then it also transforms the sequence of
observable states. Thus, the observable states to be preserved can be defined
within state transition systems although the proof that the behavior is actually
preserved requires details only available in the full ASM descriptions.
The principle of constructing the operational semantics of a program by ho-
rizontal composition as explained in Sect. 4 leads to compositions of state tran-
sition systems from ‘‘smaller’’ ones defined in Sect. 8.2. Refinements of composed
state transitions systems can be obtained from refinements of their components.

8.1 State Transition Systems

Formally, a state transition system (STS) is a triple S = (Q, I, →) where Q is

a set of states, I ⊆ Q is a set of initial states and →⊆ Q × Q is a transition
relation. A state q ∈ Q is final iff there is no q 0 ∈ Q0 with q → q 0 . A run qq
of S is a finite or infinite sequence of states hq0 , q1 , q2 , . . .i such that q0 ∈ I and
qi−1 → qi for all 0 < i. A run is complete iff it is infinite or its last state is final.
A sub-run of a run qq is a subsequence of qq. A sequence of states is a sub-run
of S iff it is a sub-run of some run of S. A STS is trivial iff →= ∅ and basic iff
→⊆ I × (Q \ I). We only consider non-trivial STSs. Every run of a non-trivial
basic STS consists exactly of one state transition.
An ASM A = (Σ, ΦInit , ΦInv , Trans) defines a STS SA = (Q, I, →) with

– Q = {q ∈ Alg(Σ) | q |= ΦInv }
– I = {q ∈ Alg(Σ) | q |= ΦInv ∪ ΦInit }, and
– → is the transition relation of A.

Let S = (Q, I, →), S 0 = (Q0 , I 0 , →0 ) be STSs and φ : Q → Q0 be a partial

function with domain DOM(φ). Compilers transform programs frequently by
refining them: a single state transition of S 0 is replaced by a sub-run of a more
detailed state transition system S. S φ-refines S 0 iff

(i) I ⊆ DOM(φ),
(ii) φ(I) ⊆ I 0 ,
(iii) φ(q0 ) →0 φ(qn ) for every sub-run hq0 , . . . , qn i of S with q0 , qn ∈ DOM(φ),
φ(q0 ) 6= φ(qn ), and q1 , . . . , qn−1 6∈ DOM(φ), cf. Fig. 4.
188 G. Goos and W. Zimmermann

qi0 qj0

φ φ

q0 ··· qi ··· qj ··· qn

Fig. 4. Refinement and abstraction of STSs

q0 q1 q2 q3 ···

φ φ φ φ

q00 ··· q10 ··· 0

q2 ··· q30 ···

Fig. 5. Horizontal decomposition

Conversely S 0 is called a φ-abstraction of S. The refinement (abstraction)

is 1:1 iff φ is injective and φ(q 0 ) → φ(q̄ 0 ) for every q 0 →0 q̄ 0 . The refinement
(abstraction) is complete iff φ is surjective. The refinement (abstraction) is total
iff φ is a total function with the additional properties that φ(q) is final in S 0 if q
is final in S and there is no infinite sub-run hqi : i ∈ Ni with φ(qi ) final for any
i ∈ N. S and S 0 are isomorphic iff S 0 φ-refines S for a total bijective function
φ : Q → Q0 .

Remark 2. For final states q of S it is not ensured that q ∈ DOM(φ) and φ(q) is
final in S 0 . It seems desirable that properties analogous to (i) and (ii) hold for
final states. However, for constructing correct compilers the definition above is
exactly what is needed: A final state q 6∈ DOM(φ) represents an abortion of an
execution due to violation of resource limitations.

Remark 3. The semantics of a program will usually define many intermediate

states between two observable states. We may abstract the ASM describing
such a program to a STS or, conversely, we may consider the ASM description
as a refinement of a STS which only contains observable states. Then it is this
abstraction, not the details of the original ASM, which must be preserved during
compilation. The abstraction function φ is thus a special case of the more general
relation ρ between states of two STSs.

If S 0 refines a STS S then every run of S 0 induces a run of S, cf. Fig. 5:

Theorem 1. Let S = (Q, I, →) and S 0 be two STSs, and φ : Q → Q0 a partial

function. S φ-refines S 0 iff each run of S 0 is a sequence of sub-runs hq00 , . . . , qi0 i,
hqi0 , . . . , qj0 i, · · · such that for each sub-run hqi0 , . . . , qj0 i one of the following con-
ditions holds:

– There is a sub-run hqi , . . . , qk , . . . , qj i of S and φ(qk ) = qk0 for k = i, . . . , j.

 Verifying Compilers and ASMs 189

– j = i + 1 and there is a sub-run hq (0) , . . . , q (m) i of S with φ(q (0) ) = qi0 ,

φ(q (m) ) = qj0 = qi+1
0
and q (k) ∈
/ DOM(φ) for 0 < k < m.

The refinement relation on STS is transitive:

Theorem 2 (Stepwise Refinement). Let Si = (Qi , Ii , →i ), i = 1, 2, 3, be

three state transition systems. If S3 φ2 -refines S2 and S2 φ1 -refines S1 , then S3
φ1 ◦ φ2 -refines S1 . S3 is a total refinement of S1 if both refinements are total.

We are interested in these stepwise refinements since they are used in horizon-
tal composition and by translation steps which act like macro-substitutions. But
compilers based on pure macro-expansion produce inefficient code in particular
for RISC-processors. Stepwise refinement for example would forbid optimizati-
ons that change the order of state transitions such as instruction scheduling and
code motion.
To cope with this problem we replace in Theorem 2 the abstraction function
φ2 by a more general relation ρ ⊆ Q3 ×Q2 but still request that S3 is a refinement
of S1 . Then the relation φ0 = φ1 ◦ ρ must be a partial function. Furthermore
conditions (i)–(iii) for refinements must be ensured by φ0 .
For condition (i) we require I3 ⊆ DOM(ρ) = {q (3) ∈ Q3 | ∃q (2) ∈ Q2 .q (3) ρ
(2)
q } and ρ(I3 ) ⊆ I2 . Since I2 ⊆ DOM(φ1 ) we then have ρ(I3 ) ⊆ DOM(φ1 ) and
thus I3 ⊆ DOM(φ0 ).
ρ(I3 ) ⊆ I2 also implies condition (ii): since S2 φ1 -refines S1 we have φ1 (I2 ) ⊆
I1 . Hence φ0 (I3 ) ⊆ φ1 (I2 ) ⊆ I1 .
(3) (3) (3) (3)
For condition (iii) let hq0 , . . . , qn i be a sub-run of S3 with q0 , qn ∈
(3) (3) (3) (3) (3)
DOM(φ0 ), q1 , . . . , qn−1 6∈ DOM(φ0 ), and φ0 (q0 ) 6= φ0 (qn ). Then φ0 (q0 ) →1
(3) (2) (3)
φ0 (qn ) must hold. Obviously, there must be states q0 ∈ ρ(q0 ) ∩ DOM(φ) and
(2) (3) (2) (2)
qm ∈ ρ(qn ) ∩ DOM(φ) such that φ(q0 ) 6= φ(qm ). Hence, we must require
(2) (2)
q 0 →+ 3
2 qm for at least one pair of these states. Furthermore, there must be at
(2) (2) (2) (2)
least one sub-run hq0 , . . . , qm i with q1 , . . . , qm−1 6∈ DOM(φ); otherwise there
(3) (3)
would be at least two state transitions from φ0 (q0 ) to φ0 (qn ), cf. Fig. 6(c).
This property can be ensured by requiring DOM(φ) ⊆ RAN(ρ) = {q (2) ∈ Q2 |
∃q (3) ∈ Q3 .q (2) ρ q (3) }. Together we have

Theorem 3 (Vertical Decomposition). Let Si = (Qi , Ii , →i ), i = 1, 2, 3, be

three state transition systems such that S2 φ-refines S1 for a partial function
φ : Q2 → Q1 . Let ρ ⊆ Q3 × Q2 be a relation with the following properties:

(i) φ0 = φ ◦ ρ is a partial function.

(ii) I3 ⊆ DOM(ρ) and ρ(I3 ) ⊆ I2
(iii) DOM(φ) ⊆ RAN(ρ).
(3) (3) (3) (3)
(iv) For every sub-run hq0 , . . . , qn i with q0 , qn ∈ DOM(φ0 ) there are states
(2) (3) (2) (3) (2) (2)
q0 ∈ ρ(q0 ) ∩ DOM(φ) and qm ∈ ρ(qn ) ∩ DOM(φ) such that q0 →+ 2 qm .

Then S3 φ0 -refines S for a function φ0 : Q3 → Q1 .

190 G. Goos and W. Zimmermann

q0
(1) (1)
q1 q (1)

φ φ φ φ
(2) (2) (2) (2)
q0 ··· qm q0 ··· qm

ρ ρ ρ ρ

(2) (3) (2) (3)

q0 ··· qn q0 ··· qn
(3) (3)
(a) Condition (iv), φ0 (q0 ) 6= φ0 (qn
(3)
) (b) Condition (iv), φ0 (q0 ) = φ0 (qn
(3)
)

(1) (1) (1)

q0 q1 q2

φ φ φ
(2) (2) (2)
q0 ··· qj ··· qm

ρ ρ

(2) (3)
q0 ··· qn

(c) Forbidden by Condition (iv)

Fig. 6. Vertical decomposition

We say that S3 ρ-simulates S2 if ρ ⊆ Q3 × Q2 under the conditions of theorem 3.

ρ-simulation allows for optimizations that reorder execution.

Remark 4. Condition (iv) considers all sub-runs beginning and ending with sta-
tes and is more general than argued above. However, every such sub-run can
be decomposed into sub-runs that do not contain intermediate states in the do-
(3) (3)
main of φ0 . If φ0 (q0 ) 6= φ0 (qn ) then condition (i) implies that there are states
(2) (3) (2) (3) (2) (2)
q0 ∈ ρ(q0 ) ∩ DOM(φ) and qm ∈ ρ(qn ) ∩ DOM(φ) such that q0 6= qm .
(2) (2) (2) (2)
Fig. 6(a) and (b) illustrate condition (iv) for the case q0 6= qm and q0 = qm
respectively.

Mapping the data of a high-level language program to bits and bytes is

a data-refinement. When using ASMs we change signatures and implement the
algebras by other algebras. Given a partial injective mapping φ : Alg(Σ 0 , Φ0Inv ) →
Alg(Σ, ΦInv ) then an ASM A0 = (Σ 0 , Φ0Init , Φ0Inv , Trans 0 ) is a φ-data-refinement
of A = (Σ, ΦInit , ΦInv , Trans) iff A0 1:1-refines A and for each transition rule
(ϕ0 , Updates 0 ) there is exactly one transition rule (ϕ, Updates) such that q 0 |= ϕ0
implies φ(q 0 ) |= ϕ.

Example 1. Assume the operational semantics a programming language is defi-

ned by an ASM. Assume the state of a program in a given programming language
PL contains data objects allocated in an environment (the procedure stack) and
on a heap. A compiler for PL must map these objects to the concepts of the
3
As usual, →+ denotes the transitive closure of a binary relation →.
 Verifying Compilers and ASMs 191

target machine, i. e. to byte adressable memory. Besides the given operational

semantics A of PL, given by an ASM, we need a second operational semantics
A0 for PL using the concepts of the target machine. A0 is a data-refinement of A.
But given the limited memory of the target machine A0 contains bounds beyond
which it will refuse to allocate additional objects.
Refinement functions φ are not required to be total. The example exhibits
the reason: the stack and the heap of a high-level programming language are
unlimited, whereas the memory of the target machine is limited. Hence, a run
of A0 may end in an exceptional state qe where no more data can be stored.
Obviously, qe 6∈ DOM(φ).

8.2 Programs as State Transition Systems: Composition of STSs

By using horizontal compositionality state transition systems for statements and
expressions are combined for obtaining the STS for sequential or conditional
execution of statements, for loops, etc. We view the composed STSs and its
components as graphs and describe the composition as a graph building process.
The building blocks are state transition systems as in Fig. 7 with initial and final
state (sets) I and T . Basic STSs are drawn by ovals instead of rectangles.

I S T

Fig. 7. Basic Components for Composition

For two STSs S1 , S2 an edge S1 → S2 as in Fig. 8(a) means sequential

execution of S1 and S2 . It requires that the postcondition T1 of S1 implies the
precondition I2 of S2 , i.e. each final state of S1 is an initial state of S2 . S1 could
also have several final states with different postconditions to which we may attach
different successors; then S1 acts as a selector; by arriving at a certain final state
it selects amongst its successors. In this case, we annotate the corresponding
edge with the set of final states. In order to avoid new non-determinism, the sets
associated with the out-edges of an STS must be pairwise-disjoint. Furthermore,
an STS can only be entered at an initial state and left at a final state. Finally,
the initial states of a composed STS must be defined. This is modeled by the
special vertex I: an initial state of a component S is initial in the composed STS
iff there is an edge (I, S).
Then a directed labeled graph G = (V, E) with V = {S1 , . . . , Sk ∪ {I} with
initial node I (precondition) is a composition of the k STSs S1 = (Q1 , I1 , →1
), . . . , Sk = (Qk , Ik , →k ) if each edge (Si , Sj ) is associated with a subset Fi,j of
final states Fi of Si , respectively, and the following conditions are satisfied:
(i) There exists at least one index i and an edge (I, Si ). (There is a component
Si to begin with.)
192 G. Goos and W. Zimmermann

(ii) Fi,j ⊆ Fi ∩ Ij . (For every edge (Si , Sj ), each state of Fi,j is final in Si and
initial in Sj .)
(iii) Fi,j ∩ Fi,k = ∅ for every j 6= k. (Each final state of Si unambigously
determines its successor.)
(iv) A common state q ∈ Qi ∩ Qj of two components Si , Sj either belongs to
Fi,j ∩ Fj,i , or there is a Sk such that q ∈ Fi,k ∩ Fj,k . (If two components have
common states then either there is an edge between the two components
and the state is in the set associated with the edge, or they have a common
successor Sk and the state is in the sets associated with both edges (Si , Sk )
and (Sj , Sk )).

Let Sj1 , . . . , Sjl be the successors of I in G. The STS SG = (Q1 ∪ · · · ∪

Qk , Ij1 ∪ · · · ∪ Ijl , →1 ∪ · · · ∪ →k ) is the STS described by G. For convenience,
we assume Fi,j = ∅ if there is no edge (Si , Sj ).
Condition (iii) ensures that no new non-determinism is introduced by com-
position, condition (iv) forbids undesired ‘‘jumps’’ out of a STS.
If the STSs are abstractions from ASMs then the conditions (ii)–(iv) about
the state sets Fi,j translate into formulae about pre- and postconditions of ASMs:
(i,j) (i) (j) (i) (j) (i) (j) (i,j)
(ii’) Φfinal |= ΦInv ∪ ΦInv ∪ Φfinal ∪ ΦInit and ΦInv ∪ ΦInv |= Φfinal for each edge
(Ai , Aj ).
(i,h) (i,j)
(iii’) For every component Ai with successors Ah , Aj , the set Φfinal ∪ Φfinal is
inconsistent.
(i) (j)
(iv’) For every pair Ai , Aj and every formula ϕ such that ΦInv |= ϕ and ΦInv |=
(i,j) (j,i)
ϕ either Φfinal |= ϕ, Φfinal |= ϕ, or Ai and Aj have a common successor Ah
(i,h) (j,h)
such that Φfinal |= ϕ and Φfinal |= ϕ.

Remark 5. A state q of the STS SG described by the composition G is final iff

it is a final state of some component Si and there is no out-edge (Si , Sj ) with
q ∈ Fi,j . We may make this property explicit by adding a new special vertex T ,
an edge (Si , T ) and associate with q the set Fi,T of those final states of Si which
are final in SG .

Example 2. The STS of the sequential composition Stats 1 ; Stats 2 is defined by

the sequential composition of the STS of Stats 1 and Stats 2 , cf. Fig. 8(a). A
conditional statement if cond then Stats 1 else Stats 2 is defined as the sequential
composition of the STS for cond , a single state transition if , and the disjoint
union of the STSs for Stats 1 and Stats 2 , cf. Fig. 8(b). A loop while cond do
Stats is defined as in Fig. 8(c) by a cyclic graph consisting of the STS for cond ,
a single state transition while, and the STS for Stat.

Obviously, this example only defines the possible control-flow of a program

viewed as sequences of possible state transitions. But how e. g. the while node
decides amongst its possible successors is yet unspecified. With ASMs this detail
can be added:
 Verifying Compilers and ASMs 193

I Stats 1 Stats 2 T I cond while T

(a) Stats 1 ; Stats 2 Stats

(c) while cond do Stats

I cond if Stats 1 T

Stats 2 T

(b) if cond then Stats 1 else Stats 2

Fig. 8. STS representing control-flow

Example 3. Fig. 9 shows the signature of ASMs for a while-language. The signa-
ture contains sorts TASK representing the sort of tasks (or instructions) belon-
ging to various instruction classes. Tasks are represented by constant symbols,
i. e. nullary functions of the algebraic signature Σ, for each instruction occurring
in a program. Similarly, variables are represented by a constant for each occur-
ring variable. Fig. 10 shows state transition rules for instruction classes. These
are used for the basic STSs defined in Fig. 8. For example, the oval vertex if
.
represents an ASM with the transition rule for IF and ΦInit = {CT = if j } for a
j ∈ {1, . . . , k1 }. The function NT is a partial function. A last task is final. The
conditions (i), (ii), (iv) require that each ASM in the composition has different
tasks. Let A = (Σ, ΦInit , ΦInv , Trans) be a non-basic ASM used in the compo-
. .
sition. Then, ΦInv = {CT = task 1 ∨ · · · CT = task k } where task 1 , . . . , task k
are a set of tasks, some of which are initial in successor ASMs. ΦInit and Φfinal
.
have the same form, e. g. ΦInit specify a disjunction of equations CT = task i
that are initial. An edge (A, A0 ) also associated with a disjunction of equations
.
CT = task where each of these equations occur in Φfinal as well as in Φ0Init .

8.3 Montages
During compilation a program is represented by an abstract syntax tree. The
elements of such trees carry semantic meaning and the meaning of a program is
composed from the meaning of these elements. The abstract syntax ‘‘normalizes’’
the concrete syntax in several respects: it no longer cares about representational
details such as use of keywords or rules for operator precedence etc. These details
are encoded either in node names of subtrees or in the structure of an expression
tree. Also selection rules such as
statement ::= assignment | procedure call | loop | ...
which only enumerate the possible alternatives, are avoided: whenever a state-
ment is allowed then one of the alternatives is directly inserted in the abstract
194 G. Goos and W. Zimmermann

Σ≡
sorts INT , VAR, TASK
subsorts DECISION , ASSIGN , EXPR < TASK , DES , PLUS , CONST < EXPR
WHILE , IF < DECISION

functions . . . , −2, −1, 0, 1, 2, . . . :→ INT

x1 , x2 , . . . , xk0 :→ VAR where x1 , x2 , . . . , xk0 are all identifiers of a program
if 1 , . . . , if k1 :→ IF
while 1 , . . . , while k2 :→ WHILE
assign 1 , . . . , assign k3 :→ ASSIGN
des 1 , . . . , des k4 :→ DES
plus 1 , . . . , plus k5 :→ PLUS
const 1 , . . . , const k2 :→ CONST

content :VAR → INT the memory

CT : → TASK the task pointer
value :EXPR → INT the value of an expression
+ :INT × INT → INT
id :DES → VAR variable of a designator
const :CONST → INT value of a constant expression
NT :TASK → TASK the next task
TT :DECISION → TASK alternative decision
dest :ASSIGN → DES destination of an assignment
src :ASSIGN → EXPR source of an assignment
cond :DECISION → EXPR condition for a decision
lop, rop :PLUS → EXPR left and right operand

Fig. 9. Signature of ASMs for a While-Language

syntax tree.4 Thus, the tree only contains leaves and so-called composition pro-
ductions composing a language element from smaller units (subtrees). Every tree
element has a semantic meaning which influences program execution.
Semantic analysis is concerned with evaluating context conditions i. e. name
analysis, type checking/inference, and operator identification, and checking their
consistency; it works on symbolic information and does not care about the dyna-
mic interpretation of a program: whether an operation named ‘‘integer addition’’
really means an addition (not a subtraction) is of no concern. Even the question
which algebraic laws are applicable is uninteresting during this phase. Traditio-
nally semantic analysis is described by an attribute grammar, based on abstract
syntax.

4
Of course, given the non-orthogonal design of many languages the selection may
depend on context conditions which must be checked during semantic analysis.
 Verifying Compilers and ASMs 195

.
if CT is IF ∧ ¬value(cond (CT ) = 0 then
CT := TT (CT )
.
if CT is WHILE ∧ ¬value(cond (CT ) = 0 then
CT := TT (CT )
if CT is ASSIGN then
content(id (dest(CT ))) := value(src(CT ))
CT := NT (CT )
if CT is DES then
value(CT ) := content(id (CT ))
CT := NT (CT )
.
if CT is IF ∧ value(cond (CT )) = 0 then
CT := NT (CT )
.
if CT is WHILE ∧ value(cond (CT )) = 0 then
CT := NT (CT )
if CT is PLUS then
value(CT ) := value(lop(CT )) + value(rop(CT ))
CT := NT (CT )
if CT is CONST then
value(CT ) := const(CT ))
CT := NT (CT )

Fig. 10. Transition Rules for a While-Language

Only when we enter the mapping phase of a compiler we get concerned with
the actual (dynamic) semantics of a program: We need to know the control and
data flow and certain information about the dynamic interpretation of opera-
tions. The required static information can again be represented by attribute
grammars. The dynamic semantics must be correctly transformed by the map-
ping phase; its rôle is that of invariants which must be preserved.
This description can be represented for each production by a variant of Kut-
ter’s montages, [35]. Montages were developed for graphically describing the
dynamic semantics of source languages by help of ASMs. We use it here for
specifying semantic analysis and the mapping phase of a compiler. The main
differences to Kutter’s monateges are that our context-free grammar defines
abstract, not concrete syntax, and the static semantics is described by attribute
grammars (not by ASMs).
A montage in this sense is a tuple consisting of a composition production
p : X0 ::= X1 · · · Xk , a control- and data-flow graph Gp , attribution rules R,
conditions Cp , and a set of transition rules Trans p . The graph Gp consists of
four classes of vertices and edges. A task vertex is graphically represented by
an ellipse. There is at most one task vertex representing the start of execution
of the montage; this vertex must be labeled with X0 . Nonterminal vertices are
labeled with one of the nonterminals of the RHS of p. I and T represent initial
and terminal vertices, respectively.
A control-flow edge is drawn dashed, a data-flow edge is drawn solid. The
origin of named edges must be task vertices. The origin of a data-flow edge must
196 G. Goos and W. Zimmermann

be a task vertex. Each out-edge of a data-flow edge must be named. The out-
edges of a square vertex Xi are labeled with integers. If Xi defines n terminal
sets, then Xi has n out-edges. The destination of named control-flow edges must
be either task vertices or the set of initial tasks has exactly one element. Analo-
gously, the destination of named data-flow edges must be either task vertices or
the set of terminal task has exactly one element.

Remark 6. This definition does not exclude non-determinism. It requires that

each desired non-determinism is explicitly specified.

Fig. 11 shows the montage for the while-loop. The decision whether the loop
ends or is continued rests on the value of the loop condition. This dependence
is indicated by a data-flow edge. There may be additional ways for exiting the
loop if (as in C) the body contains continue or break-statements.

While ::= Expr Stats

cond

NT
I Expr While T

1 3
TT

Stats 2 T

Stats.loopEnd := choose(While.next)
Stats.loopCont := choose(Expr .initial
Expr .defs := While.defs
Stats.defs := While.defs

if CT is While then
.
if value(cond(CT )) = 0 then
CT := NT (CT )
else
CT := TT (CT )

Fig. 11. A Montage for While-Loops

There are two views on the graph Gp : First it defines a set of attributes and
attribution rules representing the task structure of programs. Together with the
nodes and edges of the graph we have an ASM ASL for the source language.
Additionally we have a parametrized ASM ATL for the target language when
names and data-flow edges are removed from Gp . The mapping is implicitly spe-
cified as a conditional graph rewrite system. A montage thus definea a mapping
from abstract syntax to control- and data-flow graphs.
The operational semantics of a program can be composed from the monta-
ges representing the involved language elements according to the structure of
the syntax tree. Each square vertex represents a different subtree of the AST;
conditions (ii)–(iv) of Section 8.2 are satisfied, since each value of CT specifies
the currently executed ASM.
 Verifying Compilers and ASMs 197

8.4 Horizontal Decomposition

The program transformations used by compilers transform the control- and data-
flow graphs. We already discussed that the observable behaviour O of the pro-
gram must be preserved, i.e. the ASM assigned to the transformed program
must be a refinement of O. The program transformations can be described by a
finite set of local graph rewrite rules. Horizontal compositionality implies that
the correctness of such rules can be proven locally. Since O is given by a (hierar-
chical) composition G of STSs, it must be ensured that composing refinements
by the corresponding composition (i.e. using the same graph and replacing each
component by a refinement) is a refinement of O.
A composed STS (ASM) can be refined by refining its components. Let
G = (V, E) be a composition of S1 , . . . , Sk , S10 , . . . , Sk0 be φ1 , . . . , φk -refinements
of S1 , . . . , Sk , respectively, and φ = φ1 ∪ · · · ∪ φk . Furthermore, let G0 = (V 0 , E 0 )
be the graph obtained from G by replacing S1 , . . . , Sk by S10 , . . . , Sk0 , respectively,
0
and associating the sets Fi,j = φ−1 (Fi,j ) with edges (Si0 , Sj0 ). We now investigate
the requirements for G being a composition of S10 , . . . , Sk0 and the STS obtained
0

by G0 being a refinement of G.
For G0 being a composition of S10 , . . . , Sk0 the condition Section 8.2(i) is satis-
fied whereas conditions (ii)–(iv) could be violated. However, this can be cured
by ‘‘renaming’’ states. We leave the formal proof of this to the reader. When
using montages, such a renaming is not necessary since the task pointer CT
unambigously defines the current ASM.
Let S, S 0 be the STS described G, G0 respectively. φ must be a partial function
for S 0 being a φ-refinement of S. This is only the case iff φi (q) = φj (q) for
q ∈ DOM(φi ) ∩ DOM(φj ). It is not hard to see that conditions (i) and (ii)
for S 0 being a φ-refinement of S are satisfied. It remains to prove condition
(iii). Obviously, each sub-run of a component Si0 is also a sub-run of S 0 . Hence,
the only sub-runs that may violate condition (iii) are sub-runs hq00 , . . . , qn0 i with
q00 , qn0 ∈ DOM(φ), and q10 , . . . , qn−1
0
6∈ DOM(φ) where q00 and qn0 strictly belong
to the states of two different STSs. However, there are no such sub-runs because
0
Fi,j ⊆ Ij0 ⊆ DOM(φj ) ⊆ DOM(φ) for all i, j.
Hence, we obtain the

Theorem 4. Let G, S1 , . . . , Sk , S10 , . . . , Sk0 , G0 = (V 0 , E 0 ), φ, and Fi,j

0
be defined
0 0 0
as above. If G is a composition of S1 , . . . , Sk and if φi (q) = φj (q) for all i, j,
i 6= j, q ∈ Q0i ∩ Q0j then the STS S 0 obtained from G0 is a φ-refinement of the
STS S obtained from G.

Again, it is not hard to see that these conditions can be ensured by simply
renaming the states. In the case of montages, only φi (q) = φj (q) for q ∈ Qi ∩ Qj
has to be proven. If no data are refined, the condition is always true. Hence, the
compositional refinement of montages is easy to prove whereas data-refinements
(e.g. memory mapping) require additional work.
198 G. Goos and W. Zimmermann

8.5 Correct Compilations

A compilation from one language to another consists of two tasks: data mapping
and operation mapping. In this subsection, we demonstrate the use of the above
theory for showing the correctness of compilations. We use the ASMs ASL and
ATL as given by montages for implicitly describing data and operation map-
ping. Our goal is to independently prove the correctness of data mappings and
conditional graph rewrite rules.
Consider first the data mapping. With STS, we decomposed ASMs in a
‘‘behavioral’’ part and in a data part. A data mapping only maps the data part
and keeps the behavioural part. In particular, it assigns a new semantics (by me-
ans of an ASM A0SL ) to the source language using the concepts of the data part
of the target language (e.g. implementing a runtime stack using an address-based
memory). Since the behavioural part is kept, the correctness of this mapping can
be shown by proving that A0SL 1-1-refines ASL . In order to complete the cor-
rectness proof, we have to show according to Theorem 3 that ATL ρ-simulates
A0SL .
Theorem4 can be used to prove that ATL ρ-simulates A0SL . Observe that
ASMs are STSs and we now use the behavioural part of an ASM, i.e. its view as
a STS. Theorem 4 allows to independently prove the correctness of any graph
rewrite rule. The relation ρ relates the initial and final states of the left-hand
side and the right-hand side of the graph rewrite rule, respectively.
If a graph rewrite rule has an application condition, the correctness proof can
assume that this condition is satisfied. The compiler verifies the satisfaction of
this condition using program checking (in general it checks a stronger condition
because the application condition may be state dependent).

9 Related Work

Correctness of compilers was first considered in [36] but focused on the compi-
lation of arithmetic expressions. Thereafter most people explored the potential
of denotational semantics, e. g. [15,40,41,44,45,48,56], or of refinement calculi,
e. g. [6,4,8,16,17,33,38,43], structural operational semantics, e. g. [18] and alge-
braic models, e. g. [50]. Other approaches use abstract state machines, e. g. [6,
4,8]. Most of these projects did not compile into machine language. Instead,
they designed abstract machines, and compiled for interpreters of these abstract
machines.
These semantics-based approaches lead to monolithic compilers, cf. [24,53].
They do neither allow for reuse of traditional compiler technology nor do they
prepare for program reorganizations, as necessary for global optimization on
the machine code level. E. g., expressions are usually refined into postfix form
and then interpreted on a stack machine. The efficiency of the generated code
is by magnitudes worse than that of other compilers and thus does not meet
practical requirements, cf. [18,44]. Except [38], even projects which really gene-
rated machine language, e. g. [6,4,42,43], and ProCos, [33], chose the transputer,
 Verifying Compilers and ASMs 199

i. e., a stack machine, as their target. [38] discusses compilation of a stack-based

intermediate language Piton into an assembly language of the register-based pro-
cessor FM9001. The correctness proofs of the transformations as well as their
implementation (in ACL2 logic) are checked mechanically using the ACL2 inter-
preter. In contrast to our work, the compilation is a macro-expansion and the
source programs must be terminating regularly.
The idea of program checking was originally applied to algorithms in [1] and
continued in [2,3,57]. [26] discusses its application to constructing correct sy-
stems. [46,47] apply the idea to translating synchronous languages (SIGNAL,
Lustre, Statecharts) to C-programs; however, their assumptions allow only for
(reactive) source programs consisting of a single loop; the loop body must im-
plement a function from inputs to outputs; only the loop body is checked.
Many languages have been described so far using abstract state machines,
e. g. C [30], C++ [55], Prolog/WAM [8], Occam/Transputer [6,4], Java [11], Java
Virtual Machine [9,10,58], APE100 [5], ARM2 [34], DEC-Alpha [25], DLX [7].

10 Conclusions

We have introduced a notion of compiler verification which remains feasible also

in the presence of the unavoidable limitations of realistic harware and software.
This notion together with the idea of program checking has laid the ground for a
compiler architecture very similar to traditional architectures. We thus can also
generate efficient target code comparable to the code quality of other compilers.
Nevertheless constructing verifying compilers will remain a tedious task for the
foreseeable future.
In this work state transition systems and abstract state machines have shown
their usefulness as a unifying descriptional tool for specifying the dynamic se-
mantics of the programs being compiled on all levels of internal representation
within the compiler. ASMs have not only lead to easily understood and natural
specifications. The main advantage is that we do not have to change horses in
the middle of the river and can apply the same specification seamlessly to all
transformation steps. The expressive power of ASMs is only partially used in this
work: The number of cases in the proofs of many theorems can be considerably
reduced by adhering to a common discipline of the ASM features being used.

Acknowledgments. We are grateful to Hans Langmaack, Friedrich W. von

Henke, Axel Dold, Thilo Gaul, Wolfgang Goerigk, Andreas Heberle, Ulrich Hoff-
mann, Markus Müller-Olm, Holger Pfeifer, Harald Rueß and many students in
Karlsruhe, Kiel and Ulm for their contributions to the Verifix project which
made this paper possible. The Verifix project is supported by the Deutsche
Forschungsgemeinschaft under contract numbers Go 323/3-2, He 2411/2-2, La
426/15-2.
200 G. Goos and W. Zimmermann

References
1. M. Blum and S. Kannan. Program correctness checking . . . and the design of
programs that check their work. In Proceedings 21st Symposium on Theory of
Computing, 1989.
2. M. Blum, M. Luby, and R. Rubinfeld. Self–testing/correcting with applications
to numerical problems. In Proceedings 22nd Symposium on Theory of Computing,
1990.
3. Manuel Blum and Sampath Kannan. Designing programs that check their work.
Journal of the Association for Computing Machinery, 42(1):269–291, January 1995.
4. Egon Boerger, Igor Durdanovic, and Dean Rosenzweig. Occam: Specification and
Compiler Correctness.Part I: The Primary Model. In U. Montanari and E.-R. Olde-
rog, editors, Proc. Procomet’94 (IFIP TC2 Working Conference on Programming
Concepts, Methods and Calculi). North-Holland, 1994.
5. E. Börger, G. Del Castillo, P. Glavan, and D. Rosenzweig. Towards a Mathemati-
cal Specification of the APE100 Architecture: the APESE Model. In B. Pehrson
and I. Simon, editors, IFIP 13th World Computer Congress, volume I: Techno-
logy/Foundations, pages 396–401, Elsevier, Amsterdam, the Netherlands, 1994.
6. E. Börger and I. Durdanovic. Correctness of compiling occam to transputer. The
Computer Journal, 39(1):52–92, 1996.
7. E. Börger and S. Mazzanti. A Practical Method for Rigorously Controllable Hard-
ware Design. In J.P. Bowen, M.B. Hinchey, and D. Till, editors, ZUM’97: The
Z Formal Specification Notation, volume 1212 of LNCS, pages 151–187. Springer,
1997.
8. E. Börger and D. Rosenzweig. The WAM-definition and Compiler Correctness.
Technical Report TR-14/92, Dip. di informatica, Univ. Pisa, Italy, 1992.
9. E. Börger and W. Schulte. A Modular Design for the Java VM architecture. In
E. Börger, editor, Architecture Design and Validation Methods. Springer, 1998.
10. E. Börger and W. Schulte. Defining the Java Virtual Machine as Platform for
Provably Correct Java Compilation. In 23rd International Symposium on Mathe-
matical Foundations of Computer Science, LNCS. Springer, 1998. To appear.
11. E. Börger and W. Schulte. Programmer Friendly Modular Definition of the Se-
mantics of Java. In J. Alves-Foss, editor, Formal Syntax and Semantics of Java,
LNCS. Springer, 1998.
12. Borland/Inprise. Official borland/inprise delphi-3 compiler bug list.
http://www.borland.com/devsupport/delphi/fixes/3update/compiler.html,
jul 1999. Delphi3 Compiler Bug List.
13. Borland/Inprise. Official borland/inprise delphi-4 compiler bug list.
http://www.borland.com/devsupport/delphi/fixes/delphi5/compiler.html, jul
1999. Delphi4 Compiler Bug List.
14. Borland/Inprise. Official borland/inprise delphi-5 compiler bug list.
http://www.borland.com/devsupport/delphi/fixes/delphi5/compiler.html, jan
2000. Delphi5 Compiler Bug List.
15. D. F. Brown, H. Moura, and D. A. Watt. Actress: an action semantics directed
compiler generator. In Compiler Compilers 92, volume 641 of Lecture Notes in
Computer Science, 1992.
16. B. Buth, K.-H. Buth, M. Fränzle, B. v. Karger, Y. Lakhneche, H. Langmaack, and
M. Müller-Olm. Provably correct compiler development and implementation. In
U. Kastens and P. Pfahler, editors, Compiler Construction, volume 641 of LNCS.
Springer-Verlag, 1992.
 Verifying Compilers and ASMs 201

17. Bettina Buth and Markus Müller-Olm. Provably Correct Compiler Implementa-
tion. In Tutorial Material – Formal Methods Europe ’93, pages 451–465, Denmark,
April 1993. IFAD Odense Teknikum.
18. Stephan Diehl. Semantics-Directed Generation of Compilers and Abstract Machi-
nes. PhD thesis, University of the Saarland, Germany, 1996.
19. A. Dold, T. Gaul, W. Goerigk, G. Goos, A. Heberle, F. von Henke, U. Hoffmann,
H. Langmaack, H. Pfeifer, H. Ruess, and W. Zimmermann. Definition of the Langu-
age IS. Verifix Working Paper [Verifix/UKA/1], University of Karlsruhe/Kiel/Ulm,
1995.
20. A. Dold, T. Gaul, and W. Zimmermann. Mechanized verification of compiler
back-ends. In B. Steffen and T. Margaria, editors, Proceedings of the Internatio-
nal Workshop on Software Tools for Technology Transfer STTT ’98, pages 13–24,
Aalborg, Denmark, 1998.
21. H. Emmelmann. Code selection by regularly controlled term rewriting. In R. Gie-
gerich and S.L. Graham, editors, Code Generation - Concepts, Tools, Techniques,
Workshops in Computing. Springer-Verlag, 1992.
22. H. Emmelmann, F.-W. Schröer, and R. Landwehr. Beg - a generator for efficient
back ends. In ACM Proceedings of the Sigplan Conference on Programming Lan-
guage Design and Implementation, June 1989.
23. Albert Endres. An analysis of errors and their causes in system programs. SIG-
PLAN Notices, 10(6):327–336, 1975.
24. David A. Espinosa. Semantic Lego. PhD thesis, Columbia University, 1995.
25. T.S. Gaul. An Abstract State Machine Specification of the DEC-Alpha Processor
Family. Verifix Working Paper [Verifix/UKA/4], Universität Karlsruhe, 1995.
26. W. Goerigk, T.S. Gaul, and W. Zimmermann. Correct Programs without Proof?
On Checker-Based Program Verification. In Proceedings ATOOLS’98 Workshop on
“Tool Support for System Specification, Development, and Verification”, Advances
in Computing Science, Malente, 1998. Springer Verlag.
27. Wolfgang Goerigk. Towards Rigorous Compiler Implementation Verification. In
Proc. of the 1997 Workshop on Programming Languages and Fundamentals of Pro-
gramming, Avendorf, Germany, November 1997.
28. Gerhard Goos. Sather-K — The Language. Software — Concepts and Tools,
18:91–109, 1997.
29. Y. Gurevich. Evolving Algebras: Lipari Guide. In E. Börger, editor, Specification
and Validation Methods. Oxford University Press, 1995.
30. Y. Gurevich and J. Huggins. The Semantics of the C Programming Language. In
CSL ’92, volume 702 of LNCS, pages 274–308. Springer-Verlag, 1993.
31. A. Heberle, T. Gaul, W. Goerigk, G. Goos, and W. Zimmermann. Construction of
Verified Compiler Front-Ends with Program-Checking. In Proceedings of PSI ’99:
Andrei Ershov Third International Conference on Perspectives Of System Infor-
matics, volume 1755 of Lecture Notes in Computer Science, Novosibirsk, Russia,
1999. Springer Verlag.
32. Andreas Heberle and Dirk Heuzeroth. The formal specification of IS. Technical
Report [Verifix/UKA/2 revised], IPD, Universität Karlsruhe, January 1998.
33. C.A.R. Hoare, He Jifeng, and A. Sampaio. Normal Form Approach to Compiler
Design. Acta Informatica, 30:701–739, 1993.
34. J. Huggins and D. Van Campenhout. Specification and Verification of Pipelining
in the ARM2 RISC Microprocessor. ACM Transactions on Design Automation of
Electronic Systems, 3(4):563–580, October 1998.
35. P. W. Kutter and A. Pierantonio. Montages specifications of realistic programming
languages. Journal of Universal Computer Science, 3(5):416–442, 1997.
202 G. Goos and W. Zimmermann

36. John McCarthy and J. Painter. Correctness of a compiler for arithmetic expressi-
ons. In Schwartz [51], pages 33–41.
37. Sun Microsystems. Sun official java compiler bug database.
http://java.sun.com/products/jdk/1.2/bugs.html, mar 2000.
38. J S. Moore. Piton, A Mechanically Verified Assembly-Level Language. Kluwer
Academic Publishers, 1996.
39. C. Robert Morgan. Building an Optimizing Compiler. Digital Press, Februar 1998.
ISBN 155558179X.
40. P. D. Mosses. Abstract semantic algebras. In D. Bjørner, editor, Formal description
of programming concepts II, pages 63–88. IFIP IC-2 Working Conference, North
Holland, 1982.
41. P. D. Mosses. Action Semantics. Cambridge University Press, 1992.
42. Markus Müller-Olm. An Exercise in Compiler Verification. Internal report, CS
Department, University of Kiel, 1995.
43. Markus Müller-Olm. Modular Compiler Verification. PhD thesis, Techn. Fakultät
der Christian-Albrechts-Universität, Kiel, June 1996.
44. Jens Palsberg. Provably Correct Compiler Generation. PhD thesis, Department of
Computer Science, University of Aarhus, 1992. xii+224 pages.
45. L. Paulson. A compiler generator for semantic grammars. PhD thesis, Stanford
University, 1981.
46. A. Pnueli, M. Siegel, and E. Singermann. Translation validation. In Tools and
Algorithms for the Construction and Analysis of Systems, volume 1384 of Lecture
Notes in Computer Science, pages 151–166. Springer-Verlag, 1998.
47. Amir Pnueli, O. Shtrichman, and M. Siegel. The code validation tool (cvt). Int.
J. on Software Tools for Technology Transfer, 2(2):192–201, 1998.
48. W. Polak. Compiler specification and verification. In Lecture Notes in Computer
Science, number 124 in LNCS. Springer-Verlag, 1981.
49. Robert M. Poston. Preventing software requirements specification errors with ieee
830. IEEE Software, 2(1):83–86, January 1985.
50. T. Rus. Algebraic processing of programming languages. Theoretical Computer
Science, 199:105–143, 1998.
51. J. T. Schwartz, editor. Mathematical Aspects of Computer Science, Proc. Symp.
in Appl. Math., RI, 1967. Am. Math. Soc.
52. Reinier Sterkenburg. Borland pascal compiler bug list.
http://www.dataweb.nl/ r.p.sterkenburg/bugsall.htm, feb 2000.
53. M. Tofte. Compiler Generators. Springer-Verlag, 1990.
54. William M. Waite and Gerhard Goos. Compiler Construction. Springer-Verlag,
1984.
55. C. Wallace. The Semantics of the C++–Programming Language. In E. Börger,
editor, Specification and Validation Methods. Oxford University Press, 1995.
56. M. Wand. A semantic prototyping system. SIGPLAN Notices, 19(6):213–221, June
1984. SIGPLAN 84 Symp. On Compiler Construction.
57. Hal Wasserman and Manuel Blum. Software reliability via run-time result-
checking. Journal of the ACM, 44(6):826–849, November 1997.
58. W. Zimmermann and T. Gaul. An Abstract State Machine for Java Byte Code.
Verifix Working Paper [Verifix/UKA/12], University of Karlsruhe, 1997.
 An ASM Dynamic Semantics for Standard ML

Steven C. Cater and James K. Huggins

Kettering University
Computer Science Program
1700 W. Third Avenue
Flint, MI 48504-4898 USA
{scater,jhuggins}@kettering.edu

Abstract. The Abstract State Machines (ASM) methodology is a meth-

odology for formally specifying computing systems. We use the ASM me-
thodology to give the dynamic semantics of the functional programming
language Standard ML. We give an operational semantics for Standard
ML by means of an interpreter for (appropriately pre-processed) Stan-
dard ML programs; the effect of a Standard ML instruction can be seen
in terms of the corresponding actions performed by the ASM.

1 Introduction
The Abstract State Machines (ASM) methodology [11] is a methodology for
formally specifying computing systems (software, hardware, or mixed). First
introduced by Gurevich [9] (under their former name, “evolving algebras”) the
ASM methodology is mathematically precise, yet general enough to be applicable
to a wide variety of problem domains [3,7,15]. The ASM Thesis [12] asserts that
any computing system can be described at its natural level of abstraction by an
appropriate ASM.
Standard ML (SML) is a functional programming language. It has been de-
scribed as a “safe, modular, strict, functional, polymorphic programming lan-
guage with compile-time type checking and type inference, garbage collection,
exception handling, immutable data types and updatable references, abstract
data types, and parametric modules.” [18]
In this paper, we describe the dynamic semantics of SML using the ASM
methodology, using Milner [19] as our definition of SML. We describe the dyna-
mic semantics of SML by describing an ASM which acts as an SML interpreter,
executing an (appropriately pre-processed) SML program. This provides an ope-
rational semantics for SML; the effect of a given SML instruction can be seen in
terms of the corresponding actions performed by the ASM.
We separate concerns in this work and restrict our attention to the dynamic
semantics of SML, assuming that all static semantic analysis and checking has
been performed in a pre-processing stage. We also follow Milner [19] in restric-
ting our attention to the so-called “bare” language of SML: a subset of SML
from which all other language constructs in SML can be derived. (For example,
arbitrary tuples of the form (t1 , t2 , . . . , tn ) are translated into records of the form

Y. Gurevich et al. (Eds.): ASM 2000, LNCS 1912, pp. 203–222, 2000.
c Springer-Verlag Berlin Heidelberg 2000

204 S.C. Cater and J.K. Huggins

{1 = t1 , 2 = t2 , . . . , n = tn }.) Notice that this does not minimize the scope of

the work; using appropriate substitutions, any SML program can be translated
into the bare language and thus given semantics by our ASM descriptions.
For brevity, we omit discussion of SML constructs dealing with structure
restrictions, functor applications, layered patterns, declarations, and value bin-
dings; a full discussion of these constructs may be found in [6].
We assume familiarity with sequential ASMs in the following sections; see
[10] for an introduction to sequential ASMs.

2 ASM for SML: Simple Expressions

We now describe an ASM which acts as an intepreter for SML programs. Since
we are solely concerned with the dynamic semantics of SML rather than the
static semantics, we assume that all static analysis has already been performed
on the program to be interpreted. The input program will thus be represented
in an abstract form (to be described). The ASM operates by walking through
this form of the program, performing the required calculations and changes to
the system state.
In this section, we focus on the evaluation of arbitrary SML expressions
written in the “bare” SML language. In succeeding sections we expand our focus
to include other features of SML.

2.1 Some Common Universes, Functions, and Abbreviations

An SML program is a collection of terms; consequently, we define a universe

of Terms which is used to represent a given program. The inter-relationship
between various terms in a program are described by various unary functions,
most of which are described as they are needed. For example, we use the unary
function Next: Terms → Terms to indicate the next term in a sequence of terms
being evaluated linearly. Similarly, we use the unary function Parent: Terms →
Terms to indicate the smallest enclosing term of a given term. For example, (the
term representing) the expression “1+2” is the parent of (the term representing)
the expression “1”.
The nullary function CurTerm: Term is used to indicate the current term
being evaluated (similar to the role of an instruction counter). CurTerm acts
as the focus of attention for the ASM; the ASM at each step examines various
information regarding CurTerm and makes the appropriate changes to the state.
To identify particular terms, we define a universe of Kinds and a unary
function Kind: Terms → Kinds used to label individual terms. The members
of Kinds will be specified as we proceed. For example, a term t representing
a function application satisfies Kind(t) = functionApp, where functionApp is a
nullary function indicating a unique element of Kinds. (A more intuitive name
for Kinds might be “Types”; we choose not to use that name to avoid confusion
with SML’s extensive type system.)
 An ASM Dynamic Semantics for Standard ML 205

if OK and CurTerm.Kind = constant then

CurTerm.Value := CurTerm.ConstValue
KEEPGOING
endif

Fig. 1. Evaluating constants.

The result of evaluating an SML expression is a value; consequently, we

define a universe of Values comprising the set of possible results of expression
evaluation. We use a unary function Value: Terms → Values to record the results
of evaluating a particular term during the course of a computation; the ASM
updates Value throughout the computation as values propagate from child terms
to parent terms.
SML uses special values called exceptions to indicate the presence of errors
during a computation; the existence of an exception alters the normal flow of
control. We use a universe Exceptions (a subset of Values) to represent the set
of these special values; a nullary function CurException: Exceptions indicates
the current exception (value) which is propagating through the system. If no
exception is propagating, CurException has the special value undef.
We use two simple abbreviations in our rules. OK abbreviates the term
“CurException = undef ”; it indicates that the computation in process is procee-
ding normally (i.e., no exception is present). KEEPGOING abbreviates the rule
“CurTerm := CurTerm.Next”; it is used when the current term has been pro-
cessed and attention should be passed to the next term in the sequence of terms.
These abbreviations are introduced to improve the readability of the forthco-
ming rules; additionally, we will later re-define the KEEPGOING abbreviation
as the need for a more complicated ”keep-going” command becomes necessary.

2.2 Constant Expressions

Evaluating a constant requires almost no work; the result of evaluating a con-
stant is the corresponding constant value. We use a static function ConstValue:
Terms → Values to map a constant expression to its corresponding value. The
corresponding ASM rule, shown in Fig. 1, is straightforward.

2.3 Identifiers
In SML, all expressions are evaluated with respect to an environment: a finite
collection of name bindings. Evaluating an identifier in SML results in the cor-
responding value stored in the current environment. Thus, to give the rules for
identifier evaluation, we need to define how environments are represented in our
ASM.
We use universes of Envs (for environments) and EnvEntries (for entries in
an environment) to represent a given environment. An environment contains a
206 S.C. Cater and J.K. Huggins

if OK and CurTerm.Kind = identifier then

CurTerm.Value := Lookup(CurTerm.Identifier, CurEnv).ValueEntry
KEEPGOING
endif

Fig. 2. Identifiers.

finite number of entries, indexed by identifier names; we use a universe of Identi-

fiers to represent these names. The binary function Lookup: Envs × Identifiers →
EnvEntries returns the environment entry corresponding to the specified identi-
fier in the specified environment. The nullary function CurEnv: Envs indicates
the environment currently being used for evaluating expressions.
Environment entries are not simply values; each entry also has an additional
datum indicating whether the value corresponds to a value variable, a value con-
structor, or an exception constructor. We use a universe of ValueTags to indicate
these kinds of data in an environment, along with unary (projection) functions
ValueEntry: EnvEntries → Values and TagEntry: EnvEntries → ValueTags to
extract the needed information from a given environment entry.
Having described how environments work in SML, the rule for evaluating an
identifier simply accesses the identifier (given by a simple function Identifier:
Terms → Identifiers) and extracts the corresponding value from the current
environment. The rule is shown in Fig. 2.

2.4 Records
In SML, a record is represented as a sequence of label-expression pairs; the result
of evaluating a record is a finite mapping of the labels in the sequence to the
corresponding values (as evaluated in the current environment).
Thus, we need a universe of Maps corresponding to these finite pairs of labels
and expressions. (Note that Maps is not the same universe as Envs; environments
carry additional information which record maps do not.) We use functions Crea-
teMap: Identifiers × Values → Maps and AddToMap: Maps × Identifiers ×
Values → Maps to construct these maps.
Record expressions are represented in our ASM as a labeled expression optio-
nally followed by a record expression (following the natural recursive definition
of a sequence). Our ASM rules for evaluating records (shown in Fig. 3) move
through this sequence of record expressions, constructing the corresponding map.
(The rule below uses functions Expr: Terms → Terms and NextRecord: Terms
→ Terms to indicate the two components of a record expression; we implicitly
introduce component functions such as these in future rules without comment.)

2.5 Reference Terms

In SML, references are pointers to storage locations maintained in system me-
mory. References are used to store and retrieve values outside of the usual en-
 An ASM Dynamic Semantics for Standard ML 207

if OK and CurTerm.Kind = recordExpr then

if CurTerm.Expr.Value = undef then CurTerm := CurTerm.Expr
elseif CurTerm.NextRecord = undef then
CurTerm.Value := CreateMap(CurTerm.Label, CurTerm.Expr.Value)
CurTerm.Expr.Value := undef
KEEPGOING
elseif CurTerm.NextRecord.Value = undef then
CurTerm := CurTerm.NextRecord
else
CurTerm.Value := AddToMap(CurTerm.NextRecord.Value,
CurTerm.Label, CurTerm.Expr.Value)
CurTerm.Expr.Value := undef, CurTerm.NextRecord.Value := undef
KEEPGOING
endif
endif

Fig. 3. Records.

if OK and CurTerm.Kind = reference then

if CurTerm.Argument.Value = undef then CurTerm := CurTerm.Argument
else
extend Addresses with a
CurTerm.Value := a, Store(a) := CurTerm.Argument.Value
endextend
CurTerm.Argument.Value := undef
KEEPGOING
endif
endif

Fig. 4. Reference terms.

vironment and expression evaluation mechanism; they thus break the “pure”
functional programming paradigm.
Evaluating reference term of the form “ref t” causes a new memory location
to be allocated and returned as the value of the expression. Additionally, the
new memory location is initialized to the value of the argument expression t, as
evaluated in the current environment.
Thus, we need a new universe of Addresses to represent references, and a
unary function Store: Addresses → values to represent the current memory store.
Evaluating a reference term thus requires allocating a new memory address and
initializing the store appropriately. The resulting rule is shown in Fig. 4.
208 S.C. Cater and J.K. Huggins

if OK and CurTerm.Kind = assign then

if CurTerm.Argument.Value = undef then CurTerm := CurTerm.Argument
else
Store(MapLookup(CurTerm.Argument.Value,“1”)) :=
MapLookup(CurTerm.Argument.Value,“2”))
CurTerm.Value := Unit, CurTerm.Argument.Value := undef
KEEPGOING
endif
endif

Fig. 5. Assignment expressions.

2.6 Assignment
Assignment terms change the values stored in system memory. Syntactically,
an assignment appears in the “bare” language as the application of the special
function “:=” to a single value: a record in which the label “1” is bound to the
memory address to be changed and the label “2” is bound to the new value.
Recall that the value of a record is a map. In order to extract the neces-
sary values from the argument record, we use a binary function MapLookup:
Maps × Identifiers → Values. The rule (shown in Fig. 5) extracts the necessary
information and performs the assignment.
All expressions in SML evaluate to a value; however, assignment expressions
are evaluated for their side-effects rather than any value they might return. SML
uses the special value “unit” to represent the value of an assignment expression;
we use a corresponding distinguished element Unit: Values.

2.7 Raising and Propagating Exceptions

A “raise” expression takes an argument which evaluates to an exception value.
Evaluating such an expression causes the argument to be evaluated; the resulting
value is then set to propagate through the system (in our ASMs, this is performed
by assigning the value to CurException).
Exceptions propagate by moving from child term to parent term repeatedly
until a handle expression is found. The rules shown in Fig. 6 show how exceptions
are raised and propagated. Note that the presence of a propagating exception in
CurException falsifies the OK term, thus making most other rules inapplicable
in this situation.

3 ASM for SML: Function Applications

In this section, we focus on the evaluation of SML expressions involving function
application, which involve evaluating expressions in environments other than the
current environment. We also discuss other SML expressions whose semantics
involve changing the current evaluation environment.
 An ASM Dynamic Semantics for Standard ML 209

if OK and CurTerm.Kind = raise then

if CurTerm.Argument.Value = undef then CurTerm := CurTerm.Argument
else
CurException := CurTerm.Argument.Value
CurTerm.Argument.Value := undef, CurTerm := CurTerm.Parent
endif
endif
if CurException 6= undef and CurTerm.Kind 6= handle then
CurTerm.Value := undef, CurTerm := CurTerm.Parent
endif

Fig. 6. Raising and propagating exceptions.

if OK and CurTerm.Kind = functionClosure then

CurTerm.Value := MakeFunction (CurTerm.MatchTerm, CurEnv)
KEEPGOING
endif

Fig. 7. Function closures.

3.1 Function Closures

A function closure is a match rule (in brief, a pattern-matching rule for evaluating
the function) along with the environment to be used in evaluating that match
rule. Function closures are created by evaluating statements of the form “fn
match”, where match is a match rule. The environment currently in use is bound
to the specified match rule for later use.
We consequently make use of a universe of FunctionClosures representing
this information, with construction function MakeFunction: Terms × Envs →
FunctionClosures and projection functions MatchBinding: FunctionClosures →
Terms and EnvBinding: FunctionClosures → Envs. The rule for evaluating fun-
ction closure expressions is shown in Fig. 7.

3.2 Function Applications: Preliminaries and Special Cases

Several forms of function application are present in SML. In all of these cases, a
function application consists of two terms; a term corresponding to the function
to be applied, and a term corresponding to the argument for that function.
(All functions in the “bare” language are unary; functions of higher arity are
simulated by records.) The rule shown in Fig. 8 performs this evaluation in all
cases.
One special case occurs when the function expression is not a function at
all, but a value constructor (i.e., an identifier). In this case, the value of the
expression is a pair consisting of the constructor and the argument value. We
210 S.C. Cater and J.K. Huggins

if OK and CurTerm.Kind = functionApp then

if CurTerm.LeftExpr.Value = undef then CurTerm := CurTerm.LeftExpr
elseif CurTerm.RightExpr.Value = undef then CurTerm := CurTerm.RightExpr
endif
endif

Fig. 8. Evaluating arguments for function application.

if OK and CurTerm.Kind = functionApp and CurTerm.RightExpr.Value 6= undef

then
if CurTerm.LeftExpr.Value.Identifier then
CurTerm.Value := Pair(CurTerm.LeftExpr.Value, CurTerm.RightExpr.Value)
CurTerm.LeftExpr.Value := undef, CurTerm.RightExpr.Value := undef
KEEPGOING
endif
endif

Fig. 9. Constructor application.

use the function Pair: Values × Values → Values to construct such pairs. The
rule for constructor application is shown in Fig. 9.
A second special case occurs when the function to be applied is a built-in
operator. We use a universe PrimitiveOps to represent such primitive operations,
as well as a function Apply: PrimitiveOps × Values → Values to represent the
definition of such primitive operators. Thus, our rule simply uses the Apply
function to generate the appropriate value.
Applying a primitive operator may also generate an exception; thus, if the
value returned by Apply is an exception, we need to initiate exception handling
(as in the case of a raise expression). Thus, the rule shown in Fig. 10 checks
the resulting value before passing it along.

3.3 Function Application to Function Closures

Here we handle the case of applying a user-defined function closure to an ar-

gument term. As seen above, function closures include a match expression and
a current environment. Evaluating such a function application involves evalua-
ting the specified match expression against the value of the function argument;
however, such evaluation occurs not in the current environment, but in the envi-
ronment specified in the function closure. (We defer our discussion of evaluating
match terms against values.)
Of course, evaluating this match term in the new environment could involve
evaluating other function applications, requiring evaluating other match terms
in other environments, and so on. It becomes clear that we need a stack-like
 An ASM Dynamic Semantics for Standard ML 211

if OK and CurTerm.Kind = functionApp and CurTerm.RightExpr.Value 6= undef

then
if CurTerm.LeftExpr.Value.PrimitiveOp then
let Result = Apply(CurTerm.LeftExpr.Value, CurTerm.RightExpr.Value) in
if Result.Exception then CurException := Result
else
CurTerm.Value := Result
KEEPGOING
endif
endlet
CurTerm.LeftExpr.Value := undef, CurTerm.RightExpr.Value := undef
endif
endif

Fig. 10. Primitive operator application.

structure to use in evaluating terms in environments different from the current

environment, while still maintaining the current environment for future use.
The elements which need to be stored in an activation record upon our stack
are the following: the intermediate expression values already generated and sto-
red in the function Value, the current focus of attention stored in CurTerm,
the current evaluation environment stored in CurEnv, and the current excep-
tion propagating stored in CurException. Rather than creating an explicit stack
structure (and following the precedent in [13]), we use the universe Naturals of
natural numbers and redefine the functions named above to have the following
signatures:

Value: Terms × Naturals → Values CurTerm: Naturals → Terms

CurException: Naturals → Exceptions CurEnv: Naturals → Envs

The extra numerical argument to each of these functions is used to indicate

the level of the stack at which the given value is stored. For example, CurTerm(3)
indicates the term currently being evaluated at level 3 of the stack. A new
distinguished function StackLevel: Naturals indicates the current position on
the stack; initially, StackLevel = 0. Thus, CurTerm(StackLevel) indicates the
term currently being evaluated.
This requires changes to all the rules previously presented. In order to sim-
plify our presentation (and following the precedent in [4]), we abbreviate Cur-
Term(StackLevel) by CurTerm, Value(X,StackLevel) by Value(X), and so on. In
any function hereafter with an argument belonging to Naturals, we suppress that
argument if its value is StackLevel ; we explicitly specify the argument if it is a
term other than StackLevel or if clarity would be better served by explicit specifi-
cation. This set of abbreviations allows us to ignore the stack in situations where
its presence is semantically uninteresting, and also allows us to use the previous
rules without modification (other than, of course, the above abbreviations).
212 S.C. Cater and J.K. Huggins

EVALUATE(term, env, matchval) ::=

Target(StackLevel + 1) := term
CurTerm(StackLevel + 1) := term
CurEnv(StackLevel + 1) := env
CurMatchVal(StackLevel + 1) := matchval
StackLevel := StackLevel + 1

EVALUATE(term, env) ::= EVALUATE(term, env, undef )

Fig. 11. EVALUATE abbreviation.

if OK and CurTerm.Kind = functionApp and CurTerm.RightExpr.Value 6= undef

and CurTerm.LeftExpr.Value.FunctionClosure then
if ReturnValue = undef then
EVALUATE(CurTerm.LeftExpr.Value.MatchBinding,
CurTerm.LeftExpr.Value.EnvBinding,
CurTerm.RightExpr.Value)
else
CurTerm.Value := ReturnValue, ReturnValue := undef
CurTerm.LeftExpr.Value := undef, CurTerm.RightExpr.Value := undef
KEEPGOING
endif
endif

Fig. 12. Function application to function closures.

We introduce some functions used in the process of making a context-switch

dealing with the call stack. Target: Naturals → Terms indicates the top-level
term which is to be evaluated at this level of the stack. CurMatchValue: Naturals
→ Values indicates the value which is to be matched against Target as part of
this evaluation process. ReturnValue: Values is used to store the return value
when a given term evaluation on the stack completes.
We often make use of the EVALUATE abbreviation, shown in Fig. 11. The
idea is that EVALUATE(t, e, v) attempts to match term t against value v using
environment e. The abbreviation works by implicitly creating the next entry on
the call stack and transferring control to term t. When that term completes its
evaluation, ReturnValue will have the desired result and StackLevel will be reset
to the proper value. The alternate form EVALUATE(t,e) is used in situations
where a match value is not required.
Finally, we can present the rule for performing a function application to a
function closure. The rule shown in Fig. 12 simply invokes the specified function
closure, attempting to match the specified argument against the function closure
term.
 An ASM Dynamic Semantics for Standard ML 213

KEEPGOING ::=
if CurTerm = Target then CurTerm := ReturnTerm
else CurTerm := CurTerm.Next
endif

Fig. 13. New definition of the KEEPGOING abbreviation.

if CurTerm = ReturnTerm then

if CurException = undef then ReturnValue := Value(Target)
else
ReturnValue := CurException, CurException(StackValue-1) := CurException
endif
Target.Value := undef, CurException := undef
StackLevel := StackLevel - 1
endif

Fig. 14. Returning from a function call.

3.4 Returning from Function Closures

The rules presented in the last section provide only half of what is needed for
handling calls to function closures. The rules show how to invoke the evaluation
of another expression in another environment; they fail to show what to do when
evaluation of that expression has been completed. Here we complete the picture.
We begin by re-defining the abbreviation KEEPGOING. Recall that KEEP-
GOING works by moving CurTerm to the next expression to be evaluated in
the program. This behavior should be suspended when CurTerm has returned to
Target; at this point, the value stored at Target should be returned to the caller.
The first half of this process is shown in the new definition of KEEPGOING
in Fig. 13; it transfers control in such situations to a new distinguised element
ReturnTerm: Terms, where we perform the actual return.
We now show what happens when control reaches ReturnTerm. Here we
simply have to keep the promises we made earlier: place the proper return value
into ReturnValue and re-establish the previous evaluation context. Note that we
have to deal with situations in which an exception is propagating as well; in such
situations, we return the exception in both ReturnValue and CurException. The
rule is shown in Fig. 14.

3.5 Handle Expressions

A handle expression is used to encapsulate exception processing over a given
term. The argument term is evaluated; if no exception occurs while evaluating
the argument, the handle expression does nothing and control continues on as
usual. The rule for this case is shown in Fig. 15; notice that the truth of the guard
214 S.C. Cater and J.K. Huggins

if OK and CurTerm.Kind = handle then

if CurTerm.Expr.Value = undef then CurTerm := CurTerm.Expr
else
CurTerm.Value := CurTerm.Expr.Value, CurTerm.Expr.Value := undef
KEEPGOING
endif
endif

Fig. 15. Normal processing of handle expressions.

if CurException 6= undef and CurTerm.Kind = handle then

if ReturnValue = undef then
EVALUATE(CurTerm.Match, CurEnv, CurException)
elseif ReturnValue = Fail or ReturnValue.Exception then
CurTerm.Value := undef, CurTerm := CurTerm.Parent
else
CurTerm.Value := ReturnValue, CurException := undef
KEEPGOING
endif
ReturnValue := undef
endif

Fig. 16. Exception processing of handle expressions.

OK indicates that no exceptions have occurred while processing the argument

term.
Associated with each handle expression is a match rule (similar to that used
in function definitions). If a propagating exception reaches a handle expression,
we attempt to match the propagating exception (which is, after all, a value)
against the associated match rule, as if we made a function call to the match
rule. If the match rule returns a value or an exception, we allow that value
or exception to continue propagating. The special value Fail: Values may be
returned if the match rule fails to generate a result; in such situations, we allow
the old exception to continue propagating. We use our context-switch mechanism
introduced in the last section to perform this evaluation. The corresponding rule
is shown in Fig. 16.

3.6 Let Expressions

Let expressions contain a set of bindings (which generate an environment) and
an expression which should be evaluated with respect to those bindings and the
current environment. Those bindings should only be added to the current envi-
ronment to evaluate the target expression; afterwards, the current environment
should revert its previous value.
 An ASM Dynamic Semantics for Standard ML 215

if OK and CurTerm.Kind = let then

if CurTerm.LeftExpr.Value = undef then CurTerm := CurTerm.LeftExpr
elseif ReturnValue = undef then
EVALUATE(CurTerm.RightExpr,
CombineEnv(CurEnv, CurTerm.LeftExpr.Value))
else
CurTerm.Value := ReturnValue, ReturnValue := undef
CurTerm.LeftExpr.Value := undef
KEEPGOING
endif
endif

Fig. 17. Let expressions.

We can use the same context-switch mechanism introduced previously to pro-

vide semantics for let expressions. We first evaluate the set of attached bindings,
which generates an environment. We then combine that environment with the
current environment, using a function CombineEnv: Envs × Envs → Envs to
perform the combination appropriately. The new combined environment is then
used as the basis for a context-switch to evaluate the target expression; thus, the
new bindings implicitly disappear (as desired) when the call returns. The rule
for evaluating let expressions is shown in Fig. 17.

4 ASM for SML: Matches

In this section, we focus on the evaluation of SML match rules. An SML match
rule consists of a pattern and an associated expression. Match rules are always
evaluated with respect to a target value. A successful attempt to match a value
against a pattern results in an environment, representing the bindings required
to match the value against that pattern; in such situations, the associated ex-
pression is then evaluated with respect to the current environment augmented
by the new bindings, and its value returned. An unsuccessful attempt to match
a value against a pattern results in the special value Fail: Values.

4.1 Match Lists

Matches usually occur in a list. Evaluating a list of matches is relatively straight-
forward; one evaluates each match in the list sequentially until a non-Fail value
is generated, which is returned as the value of the list expression. Should all
matches in the list return Fail, the list returns Fail as well.
The rule for evaluating match lists is shown in Fig. 18. Note that we asso-
ciate with every match list its component rule (found with MatchRule: Terms
→ Terms) and the remainder of the list which follows that rule (found with
MatchList: Terms → Terms).
216 S.C. Cater and J.K. Huggins

if OK and CurTerm.Kind = matchList then

if CurTerm.MatchRule.Value = undef then CurTerm := CurTerm.MatchRule
elseif CurTerm.MatchRule.Value 6= Fail then
CurTerm.Value := CurTerm.MatchRule.Value
CurTerm.MatchRule.Value := undef
KEEPGOING
elseif CurTerm.MatchList = undef then
CurTerm.Value := Fail, CurTerm.MatchRule.Value := undef
KEEPGOING
elseif CurTerm.MatchList.Value = undef then CurTerm := CurTerm.MatchList
else
CurTerm.Value := CurTerm.MatchList.Value
CurTerm.MatchList.Value := undef, CurTerm.MatchRule.Value := undef
KEEPGOING
endif
endif

Fig. 18. Match lists.

4.2 Match Rules

The semantics for evaluating match rules were explained in the introduction to
this section; the corresponding rule is shown in Fig. 19.

4.3 Default Patterns

Here we begin to present rules for handling various types of patterns. Recall that
CurMatchVal is used to hold the value against which the current pattern is to
be matched; the result of a pattern match is either an environment or the special
value Fail.
The simplest patterns to match are the underscore pattern “_” and the ellipsis
pattern “. . .”. Either pattern matches against any value and creates no bindings;
the empty environment is returned. We use the distinguished element EmptyEnv:
Envs to represent the environment containing no bindings. The corresponding
simple rule is shown in Fig. 20.

4.4 Constant Patterns

One can match values against special patterns representing constants (e.g., nu-
meric literals). The match succeeds if and only if the constant pattern corre-
sponds to the current match value; no bindings are generated. Using the function
ConstValue: Terms → Values to indicate the true value of this constant term,
the rule shown in Fig. 21 performs this operation.
 An ASM Dynamic Semantics for Standard ML 217

if OK and CurTerm.Kind = matchRule then

if CurTerm.Pattern.Value = undef then CurTerm := CurTerm.Pattern
elseif CurTerm.Pattern.Value = Fail then
CurTerm.Value := Fail, CurTerm.Pattern.Value := undef
KEEPGOING
elseif ReturnValue = undef then
EVALUATE(CurTerm.Expr, CombineEnv(CurEnv, CurTerm.Pattern.Value))
else
CurTerm.Value := ReturnValue, ReturnValue := undef
CurTerm.Pattern.Value := undef
KEEPGOING
endif
endif

Fig. 19. Match rules.

if OK and (CurTerm.Kind = “_” or CurTerm.Kind = “. . .”) then

CurTerm.Value := EmptyEnv
KEEPGOING
endif

Fig. 20. Default patterns.

4.5 Simple Identifiers

Matching a value against an identifier in the current environment succeeds in
three cases:
1. The identifier is currently unbound. The resulting value is a single-entry
environment, binding the identifier to the value.
2. The identifier is currently bound to a free variable. The resulting value is a
single-entry environment, binding the identifier to the variable.
3. The identifier is already bound to the desired value. No additional bindings
are generated.

if OK and CurTerm.Kind = specialConstant then

if CurMatchVal = CurTerm.ConstValue then CurTerm.Value := EmptyEnv
else CurTerm.Value := Fail
endif
KEEPGOING
endif

Fig. 21. Constant patterns.

218 S.C. Cater and J.K. Huggins

if OK and CurTerm.Kind = identifier then

let id = CurTerm.Identifier in
if Lookup(CurEnv, id) = undef
or Lookup(CurEnv, id).TagEntry = ValueVariable then
CurTerm.Value := CreateEnv(id, Pair(CurMatchVal, ValueVariable))
elseif Lookup(CurEnv, id).ValueEntry = CurMatchVal then
CurTerm.Value := EmptyEnv
else CurTerm.Value := Fail
endif
KEEPGOING
endlet
endif

Fig. 22. Identifier patterns.

We use a new function CreateEnv: Identifiers × Values → Envs to create a

single-entry environment with the specified binding. The corresponding rule is
shown in Fig. 22.

4.6 Records: Labeled Patterns

Recall that the value of a record in SML is a finite name/value mapping. To
match a record value against a sequence of labeled patterns, we simply need to
ensure that the labeled sequence agrees with the record. That is, if label ` is
associated with expression e, the current match value should associate ` with a
value v such that v successfully (recursively) matches against e.
Consequently, our context-switch rules come in handy again, as we need to
repeatedly check that each labeled expression successfully matches against the
corresponding values in the match record. Our rule, shown in Fig. 23, simply
proceeds along the list of labeled patterns, combining the resulting environments
as needed.

4.7 Constructed Patterns

Constructed patterns are patterns consisting of a constructor and an argument.
A constructed pattern successfully matches against a value which is itself a
constructed value with the same identifier and whose value (recursively) matches
against the pattern argument. The rule for constructed patterns is shown in Fig.
24.

4.8 Memory References

A memory reference term successfully matches against a value if the value is an
address, and the value stored at that address in memory (recursively) matches
against the argument of the reference term. The rule for memory reference terms
is given in Fig. 25.
 An ASM Dynamic Semantics for Standard ML 219

if OK and CurTerm.Kind = labeledPattern then

if ReturnValue = undef then
EVALUATE(CurTerm.Pattern, CurEnv,
MapLookup(CurMatchVal, CurTerm.Label))
else
if ReturnValue = Fail then
CurTerm.Value := Fail
KEEPGOING
elseif CurTerm.MoreBindings = undef then
CurTerm.Value := ReturnValue
KEEPGOING
elseif CurTerm.MoreBindings.Value = undef then
CurTerm.Pattern.Value := ReturnValue
CurTerm := CurTerm.MoreBindings
else
if CurTerm.MoreBindings.Value = Fail then CurTerm.Value := Fail
else CurTerm.Value := CombineEnv(CurTerm.Pattern.Value,
CurTerm.MoreBindings.Value)
endif
CurTerm.Pattern.Value := undef
CurTerm.MoreBindings.Value := undef
KEEPGOING
endif
endif
endif

Fig. 23. Record matching.

5 Discussion
ASMs were first proposed as a methodology for specifying the semantics of pro-
gramming languages [8]. ASMs have been applied to a wide variety of program-
ming languages: imperative languages such as C/C++ [13,20], logic program-
ming languages such as Prolog [4] and its variants, object-oriented languages
such as Java [5,21] and Oberon [17], and hardware languages such as VHDL [2].
To the best of our knowledge, this case study in Standard ML is the first appli-
cation of ASMs to provide the semantics of a functional programming language.
The official semantics of Standard ML is given by Milner [19], using an axio-
matic semantics called Natural Semantics. The rules given in Milner, while defi-
nitive, rely heavily on axiomatic notation and proof rules which can be difficult
to read. With the presence of an official semantics, there appear to be few other
treatments of SML using other semantic techniques. A recent paper by Watt [22]
announces the use of Action Semantics to give semantics to SML; another work
by Harper and Stone [14] translates SML into a typed variant of the lambda
calculus to which an operational semantics is given. All of these other works
give both static and dynamic semantics for SML.
220 S.C. Cater and J.K. Huggins

if OK and CurTerm.Kind = constructedPattern then

let id = CurTerm.Identifier in
if Lookup(id, CurEnv).ValueEntry.Second 6= ValueConstructor
and Lookup(id, CurEnv).ValueEntry.Second 6= ExceptionConstructor then
CurTerm.Value := Fail
KEEPGOING
elseif Lookup(id, CurEnv).ValueEntry.First 6= CurMatchVal.First then
CurTerm.Value := Fail
KEEPGOING
elseif ReturnValue = undef then
EVALUATE(CurTerm.Argument, CurEnv, CurMatchVal.Second)
else
CurTerm.Value := ReturnValue, ReturnValue := undef
KEEPGOING
endif
endlet
endif

Fig. 24. Constructed pattern matches.

if OK and CurTerm.Kind = memoryReference then

if ReturnValue = undef then
EVALUATE(CurTerm.Argument, CurEnv, Store(CurMatchVal))
else
CurTerm.Value := ReturnValue, ReturnValue := undef
KEEPGOING
endif
endif

Fig. 25. Reference matches.

One interesting feature of our ASM semantics for SML is our use of the
EVALUATE macro, used to evaluate a term in an environment differing from
the current environment. The EVALUATE macro appears in roughly one-third
of the rules given above (not counting rules omitted for brevity), suggesting that
term-environment evaluation plays an important role in the dynamic semantics
of SML. This is natural; SML is, after all, a functional programming language,
which draws its roots from the lambda calculus. Since function application (i.e.,
term evaluation in a different environment) is at the heart of the lambda calculus,
it seems natural that this feature should exhibit itself so prominently in our ASM
semantics.
We would argue that the ASM dynamic semantics given above are as precise
and accurate as conventional semantic treatments such as the official definition
[19]. The rules and explanations above are somewhat longer than in an axiomatic
 An ASM Dynamic Semantics for Standard ML 221

approach, as the ASM notation lends itself to natural language notations rather
than axiomatic proof rules. However, one may find the readability of these rules is
substantially better. Such explanations can provide a better basis for explaining
and understanding the language with a minimal amount of notational overhead.
Several possible extensions of this work are being contemplated. One clear
gap in the descriptions above is any treatment of the static semantics of SML,
including the extensive type system. The official definition of SML [19] spends
extensive time defining the static aspects of the language, which can also be
confusing at times. Recently Montages [16] have been used to describe both static
and dynamic semantics of programming languages using ASMs; we consider
extending this work using Montages to provide static semantics for SML as well.
Extending this work to the Montages framework has another important be-
nefit: executability. At this time, the ASM description given here has not been
tested by any executable tools. The Gem-Mex tool [1] allows one to execute Mon-
tage descriptions directly; we expect that doing so will give us further evidence
of the correctness of these descriptions.
Alternatively, ASMs have been used for various proofs of correctness, such as
compilation techniques; we consider specifying a compilation technique for SML
to some intermediate language and proving its correctness.

References
1. M. Anlauff, P. Kutter, and A. Pierantonio. Formal Aspects of and Development
Environments for Montages. In M. Sellink, editor, 2nd International Workshop
on the Theory and Practice of Algebraic Specifications, Workshops in Computing,
Amsterdam, 1997. Springer.
2. E. Börger, U. Glässer, and W. Müller. The Semantics of Behavioral VHDL’93
Descriptions. In EURO-DAC’94. European Design Automation Conference with
EURO-VHDL’94, pages 500–505, Los Alamitos, California, 1994. IEEE CS Press.
3. E. Börger and J. Huggins. Abstract State Machines 1988-1998: Commented ASM
Bibliography. Bulletin of EATCS, 64:105–127, February 1998. (An updated version
is available from [15].).
4. E. Börger and D. Rosenzweig. A Mathematical Definition of Full Prolog. In Science
of Computer Programming, volume 24, pages 249–286. North-Holland, 1994.
5. E. Börger and W. Schulte. Programmer Friendly Modular Definition of the Se-
mantics of Java. In J. Alves-Foss, editor, Formal Syntax and Semantics of Java,
LNCS. Springer, 1998.
6. S. Cater and J. Huggins. An ASM Dynamic Semantics for Standard ML. Technical
Report CPSC-1999-2, Kettering University, 1999.
7. U. Glässer. Abstract State Machines Europe Home Page. http://www.uni-
paderborn.de/cs/asm/.
8. Y. Gurevich. Reconsidering Turing’s Thesis: Toward More Realistic Semantics
of Programs. Technical Report CRL-TR-38-84, EECS Department, University of
Michigan, 1984.
9. Y. Gurevich. Logic and the Challenge of Computer Science. In E. Börger, editor,
Current Trends in Theoretical Computer Science, pages 1–57. Computer Science
Press, 1988.
222 S.C. Cater and J.K. Huggins

10. Y. Gurevich. Evolving Algebras. A Tutorial Introduction. Bulletin of EATCS,

43:264–284, 1991. (Reprinted in G. Rozenberg and A. Salomaa, eds., Current
Trends in Theoretical Computer Science, World Scientific, 1993, 266–292.).
11. Y. Gurevich. Evolving Algebras 1993: Lipari Guide. In E. Börger, editor, Specifi-
cation and Validation Methods, pages 9–36. Oxford University Press, 1995.
12. Y. Gurevich. Sequential Abstract State Machines Capture Sequential Algorithms.
ACM Transactions on Computational Logic, page to appear, 2000.
13. Y. Gurevich and J. Huggins. The Semantics of the C Programming Language. In
E. Börger, H. Kleine Büning, G. Jäger, S. Martini, and M. M. Richter, editors,
Computer Science Logic, volume 702 of LNCS, pages 274–309. Springer, 1993.
14. Robert Harper and Chris Stone. A type-theoretic interpretation of Standard ML.
In Gordon Plotkin, Colin Stirling, and Mads Tofte, editors, Robin Milner Fest-
schrifft. MIT Press, 1998.
15. J. Huggins. Abstract State Machines Home Page.
http://www.eecs.umich.edu/gasm/.
16. P.W. Kutter and A. Pierantonio. Montages: Specifications of Realistic Program-
ming Languages. Journal of Universal Computer Science, 3(5):416–442, 1997.
17. P.W. Kutter and A. Pierantonio. The Formal Specification of Oberon. Journal of
Universal Computer Science, 3(5):443–503, 1997.
18. Lucent Technology. Standard ML of New Jersey Home Page. http://cm.bell-
labs.com/cm/cs/what/smlnj/.
19. R. Milner, M. Tofte, R. Harper, and D. MacQueen. The Definition of Standard
ML (Revised). MIT Press, 1997.
20. C. Wallace. The Semantics of the C++ Programming Language. In E. Börger,
editor, Specification and Validation Methods, pages 131–164. Oxford University
Press, 1995.
21. C. Wallace. The Semantics of the Java Programming Language: Preliminary Ver-
sion. Technical Report CSE-TR-355-97, EECS Dept., University of Michigan, De-
cember 1997.
22. D. Watt. The Static and Dynamic Semantics of Standard ML. In Proceedings of the
Second International Workshop on Action Semantics (AS’99), number NS-99-3 in
BRICS Notes Series, pages 155–172. Department of Computer Science, University
of Aarhus, May 1999.
Modeling the Dynamics of UML State Machines

Egon Börger1 , Alessandra Cavarra2 , and Elvinia Riccobene2

1
Dipartimento di Informatica - Università di Pisa - C.so Italia, 40 - 50125 Pisa
[email protected]
(currently visiting Microsoft Research, Redmond)
2
Dipartimento di Matematica e Informatica - Università di Catania -
V.le A. Doria, 6 - 95125 Catania
{cavarra, riccobene}@dmi.unict.it

Abstract. We define the dynamic semantics of UML State Machines

which integrate statecharts with the UML object model. The use of
ASMs allows us (a) to rigorously model the event driven run to com-
pletion scheme, including the sequential execution of entry/exit actions
(along the structure of state nesting) and the concurrent execution of
internal activities; (b) to formalize the object interaction, by combining
control and data flow features in a seamless way; and (c) to provide a
precise but nevertheless provably most general computational meaning
to the UML terms of atomic and durative actions/activities. We borrow
some features from the rigorous description of UML Activity Diagrams
by ASMs in [7].

1 Introduction

The Unified Modeling Language [2,5,20] is a standardized notation based on a

set of diagrams to describe the structure and the behavior of a software system.
In [5] it is stated that “UML is more than just a graphical language. Rather,
behind every part of its graphical notation there is a specification that provides a
textual statement of the syntax and semantics of that building block” although
the official document [4] for the UML semantics only gives an unambiguous
textual definition of the syntax for UML notations and leaves the behavioral
content of various UML constructs largely open. The necessity to develop the
UML as a precise (i.e. well defined) modeling language is widely felt [10,9,19]
and the pUML (precise UML) group has been created to achieve this goal [18].
In this paper we analyze one of the principal diagram types which are used
in UML for the description of dynamical system behavior, namely statechart or
state diagrams, and provide a rigorous definition of their dynamics. Many papers
on the semantics of statecharts [16,21,11,17] exist in the literature, in particular
in relation to their implementation in STATEMATE [15] and RHAPSODY [14].
Nevertheless, the debate is still ongoing on what exactly should be considered as
the authoritative definition of UML State Machines which integrate statecharts
with the UML object model. One major difficulty here concerns the mechanisms
for object interaction [14,19].

Y. Gurevich et al. (Eds.): ASM 2000, LNCS 1912, pp. 223–241, 2000.
c Springer-Verlag Berlin Heidelberg 2000

224 E. Börger, A. Cavarra, and E. Riccobene

ASMs [1,12] provide a technique to solve such specification problems and to

clarify the relevant issues. In this paper, we propose an ASM model that (a)
rigorously defines the UML event handling scheme in a way which makes all
its “semantic variation points” explicit, including the event deferring and the
event completion mechanism; (b) encapsulates the run to completion step in
two simple rules (Transition Selection and Generate Completion Events)
where the peculiarities relative to entry/exit or transition actions and sequen-
tial, concurrent or history states are dealt with in a modular way; (c) integrates
smoothly the state machine control structure with the data flow; (d) clarifies
various difficulties concerning the scheduling scheme for internal ongoing (really
concurrent) activities; (e) describes all the UML state machine features that
break the thread-of-control; (f) provides a precise computational content to the
UML terms of atomic and durative actions/activities, without loosing the inten-
ded generality of these concepts (see footnote 10), and allows one to clarify some
dark but semantically relevant points in the UML documents on state machines.
We do not take any position on which UML concepts or understandings of
them are reasonable or desirable. Through our definitions we however build a
framework for rigorous description and analysis of logically consistent interpre-
tations of the intuitions which underly UML concepts. In fact, exploiting the
abstract nature of ASMs it is easy to adapt our definitions to changing requi-
rements. We hope that this will contribute to their rational reconstruction, for
the standardization, and to the comparison of different implementations. Our
model can also serve as reference model for implementing tools for code gene-
ration, simulation and verification of UML models. This work can be viewed as
a continuation of [7] where a rigorous semantics of UML activity diagrams has
been provided.
The paper is organized as follows. Section 2 introduces the basic concepts
underlying UML statechart diagrams and ASMs. The ASM model for the be-
havioral meaning of these diagrams is defined in section 3. In section 4, the
semantical equivalence between some state machine building blocks is discussed.
In section 5 we compare our model with related work and show that it satisfies
the UML meta-model requirements for state machines.

2 Basic Concepts

In this section we sketch the basic concepts underlying UML state machines and
ASMs and review the notation.

2.1 UML Statechart Diagrams

Statechart diagrams are one of the five diagrams in the UML for modeling the
dynamic aspects of systems. Statecharts were invented by David Harel [15,16],
the semantics and the notation of UML statecharts are substantially those of
Harel’s statecharts with adaptations to the object-oriented context [3].
 Modeling the Dynamics of UML State Machines 225

Statechart diagrams focus on the event-ordered behavior of an object, a fea-

ture which is specially useful in modeling reactive systems. A statechart diagram
shows the event triggered flow of control due to transitions which lead from state
to state, i.e. it describes the possible sequences of states and actions through
which a model element can go during its lifetime as a result of reacting to di-
screte events. A state reflects a situation in the life of an object during which this
object satisfies some condition, performs some action, or waits for some event.
According to the UML meta-model [4], states can belong to one of the following
categories: simple states, composite states (sequential, concurrent, submachine),
final, and pseudostates (initial, history, stub, junction, synch).
Transitions are viewed in UML as relationships between two states indicating
that an object in the first state will enter the second state and perform specific
actions when a specified event occurs provided that certain conditions are satis-
fied [3]. UML statecharts include internal, external and completion transitions.
The semantics of event processing in UML state machines is based on the
run to completion (rtc) assumption: events are processed one at a time and when
the machine is in a stable configuration, i.e. a new event is processed only when
all the consequences of the previous event have been exhausted. Therefore, an
event is never processed when the state machine is in some intermediate, unstable
situation.
Events may be specified by a state as being possibly deferred. They are
actually deferred if, when occurring, they do not trigger any transition. This
will last until a state is reached where they are no more deferred or where they
trigger a transition.

2.2 Abstract State Machines

ASMs are transition systems, their states are multi-sorted first-order structures,
i.e. sets with relations and functions, where for technical convenience relations
are considered as characteristic boolean-valued functions. The transition relation
is specified by rules describing the modification of the functions from one state
to the next, namely in the form of guarded updates (“rules”)
if Condition then Updates
where U pdates is a set of function updates f (t1 , . . . , tn ) := t, which are simul-
taneously executed when Condition is true.
We use multi-agent ASMs [12,1] to model the concurrent substates and the
internal activities which may appear in a UML statechart diagram. A multi-
agent ASM is given by a set of (sequential) agents, each executing a program
consisting of ASM rules. Their distributed runs are defined in [12].
Since ASMs offer the most general notion of state, namely structures of ar-
bitrary data and operations which can be tailored to any desired level of abstrac-
tion, this allows us on the one side to reflect in a simple and coherent way the
integration of control and data structures, resulting from mapping statecharts to
the UML object model. In fact, machine transitions are described by ASM rules
where the actions become updates of data (function values for given arguments).
226 E. Börger, A. Cavarra, and E. Riccobene

On the other side also the interaction between objects is naturally reflected by
the notion of state of multi-agent (distributed) ASMs.
For the constructs of sequentialization, iteration and submachine of sequen-
tial ASMs we use the definitions which have been given in [8]. They provide the
concept of “stable” state needed to guarantee that the event triggered sequen-
tial exit from and entry into nested diagrams is not interrupted by a too early
occurrence of a next event.

3 ASM Model for UML Statechart Diagrams

In this section we model the event governed run to completion step in statechart
diagrams. We first introduce the signature of UML statecharts and then define
the execution rules. Statechart diagrams are made up of (control) states1 and
transitions, belonging to the abstract sets STATE and TRANSITION.

3.1 Control State

The set STATE is partitioned into simple, composite (with substructure), and
pseudo states. Composite states are partitioned into sequential and concurrent
ones. Pseudostates are partitioned into initial and history states.
Simple states (1) are of the form state(entry,exit,do(A),defer)2 , where the pa-
rameters entry/exit denote actions that have to be performed as soon as the
state is entered/exited, do(A) denotes the internal ongoing activity A that must
be executed as long as the state is active, and defer is a set of events that are
candidate to be retained in that state.
Sequential composite states (2) are of the form state(entry,exit,do(A),defer,i-
nit,final,history), where entry/exit, do(A) and defer have the same meaning as
for simple states, init denotes the initial state of the submachine state associated
to this composite state, final denotes its final state, and history its associated
history state (see below for details on initial, final and history states). Sequential
composite states contain one or more substates, exactly one of which is required
to be active when the composite state is active.
Concurrent composite states (3) are of the form state(entry,exit,do(A),defer,
concurrentComp), where entry/exit, do(A) and defer are as above, and concur-
rentComp yields the set of the concurrent (sequential) substates3 composing the
state. When a concurrent state is active, all of its subcomponents are active.
1
This notion of control state, deriving from the finite state machine notion of “inter-
nal” state, is only a tiny fraction of the overall system state which is reflected by the
ASM notion of state as structure, i.e. domains (of to be instantiated objects) with
operations and relations.
2
Simple and composite states may have a name, i.e. a string denoting uniquely the
state. We omit the name parameter in the signature of such states as it is not relevant
for our model.
3
Each substate is sequential because it must enclose an initial and a final state.
 Modeling the Dynamics of UML State Machines 227

According to [4], an event that is deferred in a composite state is automati-

cally deferred in all its directly or transitively nested substates. For reasons of
simplicity, but without loss of generality, we assume that the defer set of each
state explicitly contains all the inherited events to be deferred.

(1) (2) (3)

Initial states • indicate where to start by default when the enclosing (com-
posite sequential) state is invoked. A History state, associated to a sequential
composite state say S, is a pseudostate that can be of two types: shallow history
H and deep history H∗ . The shallow history state records, upon exiting S,
only the most recent active state directly contained in S and restores the recor-
ded state when the history state is invoked. The deep history state records the
most recent active hierarchical configuration of S, and restores this configuration
when the history state is invoked. To keep track of the configuration, we use a
dynamic function
memory : STATE −→ STATE∗
that is initialized to the empty sequence for each state which has never been
accessed. To guarantee the correct entering order, we handle memory as a LIFO
list. In case ofJshallow history state memory contains at most one state.
Final states • are special states whose activation indicates that the enclosing
state is complete.
We denote by SimpleState, SequentialState, ConcurrentState, PseudoState,
FinalState the characteristic functions of the corresponding subsets of STATE.
Any state which is enclosed within a composite state is called a substate
of that state. In particular, it is called direct substate when it is not contained
in any other state; otherwise it is referred to as a transitively nested substate.
The nesting structure of statechart diagrams is encoded by the following static
functions:

– UpState: STATE −→ STATE ∪ {undef}, such that UpState(s) = t iff s is a

direct substate of a compound state t.
– DownState: STATE ×STATE −→ BOOL, such that DownState(t, s) = true
iff s is direct substate of a compound state t.
– UpChain : STATE × STATE −→ STATE∗ ,
UpChain(s, t) = [S1 , . . . , Sn ] where n > 1 &
S1 = s & Sn = t & ∀i = 2 . . . n, UpState(Si−1 ) = Si

– DownChain : STATE × STATE −→ STATE∗ ,

DownChain(s, t) = [S1 , . . . , Sn ] where n > 1 &
S1 = s & Sn = t & ∀i = 1 . . . n − 1, DownState(Si , Si+1 )
228 E. Börger, A. Cavarra, and E. Riccobene

Upchain and DownChain yield empty sequences on each pair of not nested
states. We write Up/DownChain(s1 , sb2 ) to indicate the right open sequence
U p/DownChain(s1 , sb2 ) = [T1 , . . . , Tn [, if it exists. Notice that Up/DownCh-
ain(s, sb) = [].

3.2 Transitions

The set TRANSITION is partitioned into internal and external transitions.

External transitions are of form trans(source,target,event,guard,action), where
source/target represent the source/target states of the transition, event denotes
the triggering event which may enable the transition to fire, guard is a boolean
expression that is evaluated as soon as the event occurs (if it evaluates to false
the transition does not fire), action is an action that is executed at the time the
transition fires.
Internal transitions are of the form trans(source,event,guard,action), where
all the parameters have the same meaning as for the external transitions. Inter-
nal transitions have a source state but no target state because the active state
does not change when they fire, and no exit or entry actions are executed. We
distinguish between external and internal transitions using a predicate internal
on TRANSITION.
Statechart diagrams include also completion transitions, namely transitions
with an implicit “completion event” indicating the completion of the state the
transition leaves. We can handle completion transitions as special trigger tran-
sitions, labeled by completionEvent(S), where S is the source state, and assume
that all transitions in a statechart diagram are labeled with an event. The only
transitions outgoing pseudostates are completion transitions [20].
For each type of state and transition parameter, we use a (static) function
param which applied to the related states or transitions yields the corresponding
parameter. For example entry(state) yields the entry action associated to state,
source(trans) the source state of the transition trans, etc. We often suppress
parameters notationally.

3.3 Agents

Let AGENT be the set of agents which move through the statechart diagram,
each executing what is required for its currently active state. A state becomes
active when it is entered as result of some transition, and becomes inactive if it
is exited as result of a transition. “When dealing with composite and concurrent
states, the simple term current state can be quite confusing. In a hierarchical
state machine more then one state can be active at once. If the control is on a
simple state that is contained in a composite state, then all the composite states
that either directly or transitively contain the simple state are also active” [4].
Therefore, to maintain what in UML is called the current configuration of active
states, we introduce a dynamic function
currState : AGENT → P(STATE)
 Modeling the Dynamics of UML State Machines 229

whose updates follow the control flow of the given statechart diagram. The fun-
ction deepest : AGENT −→ STATE yields the last (innermost) state reached by
an agent.
The agents execute UML statechart diagrams, i.e. they all use the same
program (or ASM Rule). As a consequence, in the formulation of these rules
below, we use the 0-ary function Self which is interpreted by each agent a as a.
When a new agent is created to perform a concurrent subcomputation (defined
by one of the substates in a concurrent composite state), it is linked to the parent
agent by the dynamic function
parent : AGENT → AGENT ∪ {undef}
We assume that this function yields undef for the main agent who is not part
of any concurrent flow. The active subagents of an agent a are collected in the
set SubAgent(a) = {a0 ∈AGENT | parent(a0 ) = a}
At the beginning of the computation, we require that there is a unique agent,
positioned on the initial state of the top state, and whose program consists of
the rules Transition Selection and Generate Completion Event described
below.

3.4 Event Handling

In UML it is assumed that a state machine processes one event at a time and
finishes all the consequences of that event before processing another event [5,
20]. “An event is received when it is placed on the event queue of its target. An
event is dispatched when it is dequeued from the event queue and delivered to
the state machine for processing. At this point, it is referred as the current event.
Finally, it is consumed when event processing is complete. A consumed event is
no longer available for processing” [4].
We therefore assume that one event is processed at a time. Since the par-
ticular event enqueuing and dispatching mechanisms are deliberately not furt-
hermore specified in UML, we model them here explicitly as semantic variation
points and therefore use a monitored predicate dispatched indicating which event
is dequeued to be processed. At any moment, the only transitions that are eli-
gible to fire when an event e occurs are the ones departing from an active state
(i.e. whose source state belongs to currState) whose associated guard evaluates
to true4 . This is expressed by the following condition
enabled(t, e) ≡ event(t) = e & guard(t) & source(t) ∈ currState
It is possible for more than one transition to be enabled by the same event,
but UML allows only those transitions to be fired simultaneously which occur
in concurrent substates [4]. In all the other cases, the enabled transitions are
said to be in conflict with each other. One can distinguish three types of conflict
situations: (1) an internal transition in an active state conflicts with a transition
outgoing from that state, (2) two or more transitions originating from the same
source in an active state are enabled by e, and (3) two or more transitions with
different source states but belonging to the same active state are enabled by the
4
If no guard is associated to a transition t, we assume guard(t) = true.
230 E. Börger, A. Cavarra, and E. Riccobene

occurrence of e. In UML the selection among conflicting transitions is constrained

only for case (3) by giving priority to the innermost enabled transition. We now
formalize this priority for (3), whereas in the cases (1) and (2) we reflect the
choice between different scheduling mechanisms as a semantic variation point,
namely by using abstract selection functions; see the Transition Selection rule
below.
Let enabled(e) = {t ∈ TRANSITION | enabled(t, e)} be the set of all transi-
tions enabled by e. We define an equivalence relation ∼ on enabled(e) as follows:
∀ t1 , t2 ∈ enabled(e), t1 ∼ t2 iff source(t1 ) = source(t2 ).
The nesting of states induces the total order relation5 ≤ on the quotient set
enabled(e)/ ∼, defined as [t1 ] ≤ [t2 ] iff source(t1 ) is a direct or a transitively
nested substate of source(t2 ).
Let FirableTrans(e) be the minimum equivalence class in enabled(e)/ ∼. It re-
flects the UML requirement that among transitions enabled by the same event
and with different source states, priority is given to an innermost one. The choice
among those innermost ones is left open as semantic variation point (see the
choose construct in the Transition Selection rule).
If a dispatched event does not trigger any transition in the current state, it
is lost unless it occurs in the deferred set of the deepest active state. This is
formalized by the following predicate deferrable on EVENT:
deferrable(e) = true ⇔ enabled(e) = ∅ & e ∈ defer(deepest)
As suggested in [20], to store deferred events we associate to each agent a
list6 of events deferQueue that is dynamically updated during the computation
(see rule Transition Selection). We can therefore define deferred(e) to mean e ∈
deferQueue.
We call a deferred event releasable when it becomes ready to be consumed,
i.e. when it can trigger a transition in the current state configuration
releasable(e) = true ⇔ deferred(e) & enabled(e) 6= ∅

3.5 Statechart Diagram Main Rules

In this subsection we define the ASM rules for the execution of statecharts, i.e.
we specify the sequences of states that an object goes through, and of the actions
it takes, in response to events which occur during its lifetime [20].
Apparently, UML leaves it unspecified how to choose between dispatched
and releasable events. We reflect this by using a selection function which, at any
moment, chooses either a dispatched event triggering a transition, or an event
that has been deferred. A dispatched event, if deferrable, has to be inserted
into the deferQueue. A releasable event, when chosen for execution, has to be
5
Observe that ≤ is total since all the source states of the transitions in enabled belong
to currState and therefore they are nested.
6
Apparently, this list is meant to be a set, leaving the exact ordering of elements open
as a semantic variation point. A similar remark applies also to other lists occurring
in the UML texts.
 Modeling the Dynamics of UML State Machines 231

deleted from deferQueue7 . This implies that when choosing an event which is
simultaneously dispatched and releasable, that event will be deleted from the
deferred events.8
We define in the next section the exact meaning of the state machine execu-
tion of a transition, namely by a parameterized macro stateMachineExecution.
This leads us to the following main rule for selecting the machine transition to
be executed next.
Rule Transition Selection
choose e : dispatched(e) ∨ releasable(e)
choose trans in FirableTrans(e)
stateMachineExecution(trans)
if deferrable(e) then insert(e,deferQueue)
if releasable(e) then delete(e,deferQueue)
The rule for selecting and executing a transition fires simultaneously, at each
“run to completion step”, with a rule to generate completion events.
Completion events are generated when an active state satisfies the completion
condition [4]. They trigger a transition outgoing such states. An active state is
considered completed if one of the following cases occurs: (1) it is an active
pseudostate, (2) it is a sequential composite state with active final state, (3)
the state internal activity terminates while the state is still active, or (4) it is a
concurrent composite state and all its direct substates have reached their final
state. We formalize this by the predicate
completed(S) = true ⇐⇒ PseudoState(S) or
(SequentialState(S) & final(S) ∈ currState) or
terminated(A(S)) or
(ConcurrentState(S) &
∀Si ∈ concurrentComp(S) ∀ai ∈ SubAgent(Self)
final(Si ) ∈ currState(ai ))
where terminated(A(S)) is a derived predicate that holds if and only if the run
of the ASM A(S), which formalizes the internal activity of S, reaches a final
state.
Each time the completion condition evaluates to true for an active state
S that is not a direct substate of a concurrent state9 a completion event is
generated. This is expressed by the rule Generate Completion Event that is
executed simultaneously for each state S ∈ currState.
7
If upon execution of transition trans, a deferred event e ∈ defer(source(trans)) does
not belong to defer(target(trans)), then it must be deleted from deferQueue, as spe-
cified as part of the enterState macro below.
8
Should another interpretation be intended, we would probably change the guard
“if releasable(e)” in the Transition selection rule to e.g. “if releasable(e) & not
dispatched(e)”.
9
This restriction reflects that in UML no direct substate of a concurrent state can
generate a transition event. Such substates are required to be sequential composite
states.
232 E. Börger, A. Cavarra, and E. Riccobene

Rule Generate Completion Event

do forall S ∈ currState
if completed(S) & ¬ ConcurrentState(UpState(S))
then generate(completionEvent(S))
Although the order of event dequeuing is not defined, it is explicitly required that
completion events must be dispatched before any other queued events [4]. We
reflect this requirement as a constraint on the monitored predicate dispatched.
The above two rules, which fire simultaneously at each run to completion step,
define the top level behavior of UML state machines. It remains to define in
more detail the meaning of the macros appearing in those rules.
The UML requirement that an object is not allowed to remain in a pseu-
dostate, but has to immediately move to a normal state [20], cannot be guaran-
teed by the rules themselves, but has to be imposed as an integrity constraint
on the permissible runs.

3.6 The Rule Macros

We define now the subrule stateMachineExecution where parameterization by
transitions allows us to modularize the definition for the different types of tran-
sitions and the involved states.

State Machine Execution. If an internal transition is triggered, then the cor-

responding action is executed (there is no change of state and no exit or entry
actions must be performed). Otherwise, if an external transition is triggered, we
must determine the correct sequence of exit and entry actions to be executed
according to the transition source and target state. Transitions outgoing from
composite states are inherited from their substates so that a state may be exi-
ted because a transition fires that departs from some of its enclosing states. If a
transition crosses several state boundaries, several exit and entry actions may be
executed in the given order. To this purpose, we seek the innermost composite
state that encloses both the source and the target state, i.e. their least common
ancestor. Then the following actions are executed sequentially: (a) the exit ac-
tions of the source state and of any enclosing state up to, but not including,
the least common ancestor, innermost first (see macro exitState); (b) the action
on the transition; (c) the entry actions of the target state and of any enclosing
state up to, but not including, the least common ancestor, outermost first (see
macro entryState); finally (d) the “nature” of the target state is checked and the
corresponding operations are performed.
The sequentialization and iteration constructs defined for ASMs in [8] provide
the combination of black box – atomic step – view and the white box – durative
– view which is needed here to guarantee that when the two ASM rules defined
above are executed, all the updates which occur in the macros defined below
are performed before the next event is dispatched or becomes releasable. This
behavior is reflected by the parameterized macro stateMachineExecution (which
constitutes the body of the Transition Selection Rule). The macros appearing in
this rule are described below.
 Modeling the Dynamics of UML State Machines 233

stateMachineExecution(trans) ≡
if internal(trans) then action(trans)
else seq
exitState(source(trans),ToS)
action(trans)
enterState(FromS,target(trans))
case target(trans)
SequentialState: enterInitialState(target(trans))
ConcurrentState: startConcurrComput(target(trans))
HistoryState: restoreConfig(target(trans))
endcase
where anc = lca(source(trans),target(trans))
ToS = directSubState(anc,UpChain(source(trans),anc))
FromS = directSubState(anc,DownChain(anc,target(trans)))

and directSubState: STATE × STATE∗ −→ STATE is defined by

directSubState(s,L) = s0 iff s0 ∈ L & UpState(s0 ) = s, i.e. s0 is the only direct
substate of s belonging to the list L.
It remains to define the macros for exiting and entering states, and for the
additional actions for sequential, concurrent and history states.

Exiting States. If a transition that crosses the boundary of a composite state

fires, we must distinguish two cases to perform the exits from nested states in
an order which respects the hierarchical structure (see macro exitState below):

1. The agent is not inside a concurrent flow (i.e. parent(Self) = undef). If the
agent is (1) not parent of concurrent subagents or (2) it is parent of concur-
rent subagents but each subagent already performed its exit actions, then
for each state from the source state up to, but excluding, the source/target
least common ancestor state (see the stateMachineExecution rule above),
innermost first, it sequentially (a) stops the internal ongoing activities, (b)
performs the exit actions , and (c) removes those states from the agent’s
current state. Moreover, it (d) updates the history (if defined and provided
that the final state has not been reached), memorizing in it all the states it
is exiting in case of deep history, or only its direct active substate in case
of shallow history, and (e) updates deferQueue by deleting all those events
which are no more deferred (see macro sequentialExit which uses a macro
abortInternalActivity defined below). In case (2) the agent must furthermore
update its deferQueue to retain all deferred events of its own but none of
those processed by its subagents. Finally it disconnects all its deactivated
subagents (see the corresponding macro defined below).
2. The agent is inside a concurrent flow (i.e. parent(Self) 6= undef). We have
to consider the two cases, whether the trigger event is relevant for all the
subagents running in parallel within the concurrent state or not. To this
purpose we check that the transition source state belongs to the active state
of both the agent and its parent (when a subagent is created, it inherits
234 E. Börger, A. Cavarra, and E. Riccobene

its parent’s current state, therefore at any time the currState of the parent
agent is a subset of its subagents’ currState). In this case, each subagent
performs the same sequentialExit macro as in the first case, i.e. starting
from its deepest state up to, but excluding, its parent’s deepest state, it
sequentially (a) stops the internal ongoing activities, (b) performs the exit
actions and (c) removes those states from the agent’s current state. Moreover,
it (d) updates the history (if defined and provided that the final state has
not been reached) memorizing in it all the states it is exiting in case of deep
history, or only its direct active substate in case of shallow history, and (e)
updates deferQueue by deleting all those events which are no more deferred
(see macro sequentialExit). Finally, the agent is deactivated, meaning that
its rule is set to undef and its current state to the empty set (see macro
deactivate).
Now consider the case that the transition source state belongs to the active
state of at least one but not to all subagents of an agent. Then the event is
relevant only for this subagent, and this agent performs the sequential exit
as in case 1.

exitState(s,t) ≡ if parent(Self) = undef

then if SubAgent(Self) = ∅
then sequentialExit(s,t)
elseif noActiveSubAgents
then seq deferQueue(Self) := defer(deepest(Self))
\ ∩
deferQueue(ai )
ai ∈SubAgent(Self)
sequentialExit(s,t)
disconnectSubAgents
if parent(Self) 6= undef
then if s ∈ currState(Self) &
s ∈ currState(parent(Self))
then
sequentialExit(S, S 0 )
deactivate(Self)
else sequentialExit(s,t)

where S = deepest(Self)
S 0 = deepest(parent(Self))
noActiveSubAgents = ∀ai ∈ SubAgent(Self) :
currState(ai ) = ∅

For the definition of the macro sequentialExit we use the macro abortInternalAc-
tivity which will be defined below. In defining sequentialExit we use a function
hist(s, S) whose value depends on whether S is a deep history state or not.
b for deep history, directSubState(S,UpChain(s,S))
hist(s, S) yields UpChain(s,S)
for shallow history.
 Modeling the Dynamics of UML State Machines 235

sequentialExit(s,t) ≡ loop through S ∈ UpChain(s,t)

seq
abortInternalActivity(S)
exit(S)
currState := remove(S,currState)
endseq
if history(S) 6= undef & final(S) 6∈ currState
then memory(history(S)) := hist(s, S)
endloop

deactivate(a) ≡ Rule(a) := undef

currState(a) := ∅

disconnectSubAgents ≡ do forall ai ∈ SubAgent(Self)

parent(ai ) = undef

Entering States. A transition may have a target state nested at any depth
in a composite state. Therefore, any state enclosing the target one up to, but
excluding, the least common ancestor will be entered in sequence, outermost first.
Entering a state means that (a) the state is activated, i.e. inserted in currState,
(b) its entry action is performed, and (c) the state internal activity (if any) is
started. This is realized by the macro enterState for which we use the macro
startActivity defined below. The agent’s deferQueue is updated by deleting all
those events which are no more deferred in the target state.

enterState(s,t) ≡ loop through S ∈ DownChain(s,t)

seq
currState := insert(S,currState)
entry(S)
startActivity(S)
endseq
deferQueue := deferQueue ∩ defer(S)
endloop

Internal Activities. When a state is active, its internal activity (if any) is re-
quired to be executed. Apparently, internal activities are intended as concurrent
and [4] imposes no particular scheduling conditions for them. We model this by
creating a new worker agent whose job is to execute the activity of its associated
state. The worker agent is created when the state is entered and after its entry
action has been executed. It receives as program the ASM A(S) formalizing the
state activity.

startActivity(S) ≡ extend AGENT with a

Rule(a) := A(S)
worker(S) := a
236 E. Börger, A. Cavarra, and E. Riccobene

Using an ASM as rigorous replacement for the intuitive notion of “internal UML
activity”, we obtain a mathematically rigorous definition without loosing gene-
rality10 . In addition we make the UML notion of “ongoing” activity precise by
defining it as steps of an ASM in a multi-agent distributed ASM run.
If an activity is aborted prior to its termination as result of the firing of an
outgoing transition, then before leaving the state its associated worker agent is
deactivated since its job is terminated. This is performed by the following macro
which is used for defining sequentialExit.

abortInternalActivity(S) ≡ Rule(worker(S)) := undef

worker(S) := undef

Sequential Composite States. A transition drawn to the boundary of a se-

quential composite state is equivalent to a transition to its initial pseudostate
[4]. Therefore, when a composite sequential state is the target state of a trig-
gered transition, the control passes to its initial state that is inserted in currState.

enterInitialState(S) ≡ currState := insert(init(S),currState)

History States. If a transition incoming to a history state within a composite

state fires, the configuration of active states stored in its memory is restored.
Therefore, each state in the history is activated, i.e. it is inserted in currState,
its entry action is performed, its activity is executed, and the state is removed
from the history. The right entering order is guaranteed by the LIFO structure
of memory.
Observe that when a state is entered for the first time or its most recently
active state prior to its exit was the final state, its history (if any) must be empty
[4]. This is guaranteed in our model since we initialize each history state memory
to the empty sequence, delete it after using it, and store nothing in it when its
enclosing state is exited by a final state.

restoreConfig(H) ≡ loop through S ∈ memory(H)

seq
currState := insert(S,currState)
entry(S)
startActivity(S)
memory(H) := delete(S,memory(H))
endseq
deferQueue := deferQueue ∩ defer(S)
endloop

Concurrent Composite States. If a transition incoming to a concurrent com-

posite state fires, the flow of control is split into two or more flows of control. The
currently active agent creates a new agent ai for each concurrent component Si .
All the subagents inherit their parent’s program to execute statechart diagrams,
10
That no generality is lost derives from Gurevich’s proof of the ASM thesis in [13].
 Modeling the Dynamics of UML State Machines 237

its currState configuration and the parent’s list of active deferred events. As a
transition drawn to the boundary of a concurrent composite state is equivalent
to a transition to any of its concurrent components and consequently to the com-
ponent initial state, each agent ai activates the component Si and its associated
initial state.

startConcurrComput(S) ≡
let S1 , . . . , Sn = concurrentComp(S)
extend AGENT with a1 , . . . , an
do forall 1 ≤ i ≤ n
parent(ai ) := Self
deferQueue(ai ) := deferQueue(Self)
Rule(ai ) := Rule(Self)
currState(ai ) := insert({Si ,init(Si )},currState)

The parent agent will stand idle waiting for the termination of its subagents’
computation. This is enforced by the definition of when a concurrent state is com-
pleted to trigger the completion event which may enable the transition exiting
the concurrent state. The running subagents can finish their job either because
of a completion event generated by their parent11 or by the firing of an explicit
event labeling a transition outgoing their enclosing state. In our model the sub-
states’ exit action and internal activity abortion are performed by the exitState
macro, in a synchronized fashion. Other choices are easily defined modeling our
rules appropriately. The UML documents seem not to mention this semantically
relevant issue at all.
Remark. In a concurrent compound state S 0 , a transition trans(e,g,a) out-
going from a state S in a concurrent component and incoming to a state S 00 sib-
ling of S 0 (see Fig. 1.a), can be viewed as split into two transitions (see Fig. 1.b):
a transition trans(e,g,Send exitEvent(S)) from S to S 0 , where Send exitEvent(S)
is an event generation action12 , and a transition trans(exitEvent(S),true,a) from
S 0 to S 00 . To guarantee the expected semantics of UML statecharts, we impose,

S’ S’

S S

e/[g]Send exitEvent

e/[g]a exitEvent/[true]a
S’’ S’’

(a) (b)

Fig. 1.

11
In this case they all must have reached their final state.
12
According to [4] an action labeling a transition may consist in sending a signal.
238 E. Börger, A. Cavarra, and E. Riccobene

as an integrity constraint on the permissible runs, that the event exitEvent(S)

must be dispatched before any other event (see the event handling mechanism
in section 3.4).

4 Semantical Equivalence among Building Blocks

UML statecharts encompass for notational convenience some constructs which
can be defined in terms of basic constructs. Not to overload our model, we
decided to include only the basic notations and to sketch here how to replace
the remaining constructs by combinations of basic constructs.

Fork-Join Pseudostates. Fork and join pseudostates split and merge transiti-
ons arriving at, emanating from, concurrent states. A transition to the boundary
of a concurrent compound state is equivalent to a transition to each of its direct
substates (and therefore to their initial states), and a transition from the bound-
ary of a concurrent compound state is equivalent to a transition from the final
states of each of its substates. Therefore, the fork (resp. join) semantics can be
obtained by allowing only incoming (resp. outgoing) transitions that terminate
on (resp. depart from) the boundary of concurrent states, and imposing that
each concurrent substate must enclose an initial and a final state.

Junction Pseudostate. Junction states are used only to chain together mul-
tiple transitions – this is known as merge –, or to split an incoming transition
into multiple outgoing transitions labeled with different guard conditions – this
is known as conditional branch [4].

Submachine States. UML statecharts provide also submachine states, a syntac-

tical convenience to facilitate reuse and modularity [4]. A submachine state is
only a shorthand that implies a macro-like expansion by another state machine
and is semantically equivalent to a composite state. According to the UML me-
tamodel, a submachine state is of the form state(entry,exit,do(A),include(S 0 )).
One can assume that each occurrence of a submachine state is substituted by
the sequential13 composite state defined by entry, exit, do(A), S’. Moreover, we
identify transitions directly incoming to, respectively outgoing from, the subma-
chine state with transitions directly incoming to, respectively outgoing from, the
resulting sequential composite state.

Stub States. A stub state is nothing else then an alias for an entry point to or an
exit point from a state s in S 0 of a submachine state state(entry,exit,do(A),inclu-
de(S 0 )).

Additional Constructs. Synch states are used to synchronize the execution

of concurrent substates. Their semantics can be given by slightly modifying the
above formalization of concurrent states.
13
Submachine states are never concurrent [4].
 Modeling the Dynamics of UML State Machines 239

5 Conclusion and Related Work

In this section we discuss some ambiguities in the official semantics of UML [4,5,
20] which are resolved in the ASM model. We also show how UML requirements
for state machines are satisfied by our model.
The state machine execution is formalized through the macro stateMachine-
Execution (invoked by the rule Transition Selection) that reflects the scheme of
a generic control machine. These ASM statecharts generalize the Mealy ASMs
defined in [6].
Our model reflects all the characteristics of the state machines metamodel in
[4] and adds to its structural, static definition the underlying control flow seman-
tics. A subtle question regards the execution of ongoing state activities. What
does happen when an internal transition occurs? Does the activity interrupt and
then restart from the same computation point, or does it never interrupt? The
way we model internal activities guarantees the second, to our understanding
reasonable, alternative. However, our model can be easily adapted to formalize
other behaviors.
By replacing the undefined UML terms of “action” and “activity” with (possi-
bly structured, in the sense of [8]) “ASM rule”, we provide a precise mathematical
content to these terms without loosing the generality intended by the designers
of UML (see in this connection Gurevich’s proof of the ASM thesis [13]). Our
model also provides a precise meaning of the vague UML term “ongoing internal
activity”, namely as execution of an ASM in a multi-agent distributed run as
defined in [12]. The sequentialization, iteration and submachine constructs de-
fined for ASMs in [8] clarify in what sense sequences of nested exit and entry
actions can be guaranteed to be executed in one “run to completion step”, as
postulated by the UML documents, namely before the next event may trigger
the next “step”. Our model also makes some semantically relevant features14
explicit which seem not to have been considered in the official UML documents.
Several semantics for statecharts have been proposed in the literature [21].
Most of these are concerned with modeling Harel’s statecharts, whose semantics
is rather different from UML state machines (e.g. in the event handling policy).
Although our model can be adapted to grasp such differences, our intent is
to define the UML state machine semantics up to the degree of precision one
can reach without compromising the desired freedom of the so called “semantic
variation points”.
Differently from the formalization of UML state machines in [11,19], our
model reflects the original structure of machines as described in the UML do-
cuments, without imposing any graphical transformation or flattening of dia-
grams. [11] uses graph rewriting techniques to transform UML state machines
into a “normal form” machine, without considering the execution of actions
and activities. The model in [19] leaves out some state machines features, and
some are covered by means of semantical equivalences which, however, do not
14
E.g. whether abortion of internal activities and exit actions of concurrent agents
should be synchronized or not.
240 E. Börger, A. Cavarra, and E. Riccobene

always respect the UML metamodel constraints (see [4], pp. 2-126). For instance,
entry/exit actions in a state are replaced by attaching such actions respectively
to the state incoming/outgoing transitions, whereas the metamodel asserts that
the multiplicity of Action in Transition is 0..1, that is no or exactly one action
may label a transition.
In [17] the operational behavior of UML state machine constructs is described
using pseudo-code in a way which in many places includes specific implementa-
tion decisions (mostly without stating them), whereas we tried to let the intended
semantic variation points of UML stand out explicitly as such.

References
1. Abstract State Machines. http://www.eecs.umich.edu/gasm/.
2. Rational Software Corporation, Unified Modeling Language UML, version 1.3,
1999.
3. UML 1.3 Notation, 1999. (Published as part of [2]).
4. UML 1.3 Semantics, 1999. (Published as part of [2]).
5. G. Booch, J. Rumbaugh, and I. Jacobson. The Unified Modeling Language User
Guide. Addison Wesley, 1999.
6. E. Börger. High Level System Design and Analysis using Abstract State Machines.
In D. Hutter and W. Stephan and P. Traverso and M. Ullmann, editor, Current
Trends in Applied Formal Methods (FM-Trends 98), number 1641 in LNCS, pages
1–43. Springer-Verlag, 1999.
7. E. Börger, A. Cavarra, and E. Riccobene. An ASM Semantics for UML Activity
Diagrams. In T. Rus, editor, AMAST2000, volume 1816 of LNCS, pages 293–308.
Springer Verlag, May 2000.
8. E. Börger and J. Schmid. Composition and Submachine Concepts for Sequential
ASMs. In Gurevich Festschrift CSL 2000, 2000. (To appear).
9. A. S. Evans, J-M. Bruel, K. Lano R. France, and B. Rumpe. Making UML Precise.
In In OOPSLA’98 Workshop on Formalizing UML. Why and How?, October 1998.
10. R. B. France, A. S. Evans, K. C. Lano, and B. Rumpe. Developing the UML as a
formal modeling notation. Computer Standards and Interfaces: Special Issues on
Formal Development Techniques, Accepted for publication, 1998.
11. Martin Gogolla and Francesco Parisi-Presicce. State diagrams in UML: A formal
semantics using graph transformations. In Manfred Broy, Derek Coleman, Tom
S. E. Maibaum, and Bernhard Rumpe, editors, Proceedings PSMT’98 Workshop
on Precise Semantics for Modeling Techniques. Technische Universität München,
TUM-I9803, 1998.
12. Y. Gurevich. Evolving Algebras 1993: Lipari Guide. In E. Börger, editor, Specifi-
cation and Validation Methods, pages 9–36. Oxford University Press, 1995.
13. Y. Gurevich. Sequential Abstract State Machines capture Sequential Algorithms.
ACM Transactions on Computational Logic, 1, 2000. (To appear).
14. D. Harel and E. Gery. Executable Object Modeling with Statecharts. Computer,
IEEE Computer Society, 30(7):31–42, 1997.
15. D. Harel and A. Naamad. The STATEMATE Semantics of Statecharts. ACM
Trans. Soft. Eng. method, 5(4):293–333, 1996.
16. D. Harel and M. Politi. Modeling Reactive Systems with Statecharts. McGraw-Hill,
1998.
 Modeling the Dynamics of UML State Machines 241

17. Ivan Paltor and Johan Lilius. Formalising UML state machines for model checking.
In Robert France and Bernhard Rumpe, editors, UML’99 - The Unified Modeling
Language. Beyond the Standard. Second International Conference, Fort Collins,
CO, USA, October 28-30. 1999, Proceedings, volume 1723 of LNCS. Springer, 1999.
18. The precise UML group. http://www.cs.york.ac.uk/puml/.
19. G. Reggio, E. Astesiano, C. Choppy, and H. Hussmann. Analysing UML Active
Classes and Associated State Machines – A Lightweight Formal Approach. In
FASE 2000 - Fundamental Approaches to Software Engineering. Lecture Notes in
Computer Science, 2000. (To appear).
20. J. Rumbaugh, I. Jacobson, and G. Booch. The Unified Modeling Language Refe-
rence Manual. Addison Wesley, 1999.
21. M. von der Beek. A Comparison of Statechart Variants. In Formal Techniques in
Real-Time and Fault-Tolerant Systems, volume 526. Lecture Notes in Computer
Science, 1994.
 On the Formal Semantics of SDL-2000:
A Compilation Approach Based on an Abstract
SDL Machine

Robert Eschbach1 , Uwe Glässer2 , Reinhard Gotzhein1 , and Andreas Prinz3

1
Department of Computing Science, University of Kaiserslautern,
D-67653 Kaiserslautern, Germany,
2
Heinz Nixdorf Institute, University of Paderborn,
D-33102 Paderborn, Germany,
3
Department of Computer Science, Humboldt-University Berlin,
D-12489 Berlin, Germany

Abstract. In November 1999, a new version of SDL (Specification and

Description Language) called SDL-2000 has passed ITU, an internatio-
nal standardization body for telecommunication. SDL is a fairly complex,
graphical formal description technique for the development of distributed
systems, and has been broadly used in industry for many years. Efforts to
define the semantics of SDL-2000 formally have started early in 1998. By
now, a draft formal semantics is available, which is determined to become
the official formal SDL semantics after its approval in 2000. It is based
on the formalism of Abstract State Machines (ASMs), which has been
selected for several reasons including intelligibility and executability.
The formal semantics of SDL addresses the static semantics, transforma-
tion rules, and the dynamic semantics. The approach taken to define the
dynamic semantics is particularly interesting. Although basically being
operational, it differs from existing approaches in several ways. In this
paper, we address and highlight some of these differences, using a simpli-
fied specification language called SSL instead of SDL. In defining a formal
dynamic semantics for SSL, we formally describe an abstract machine, a
compilation function mapping SSL specifications to code of this machine,
and an operational definition of the set of initial states, using ASM as
the underlying formalism. Furthermore, we present in some detail the
semantics of SSL procedure calls.

1 Introduction

Since 1976, SDL (Specification and Description Language) is an ITU standar-

dized language1 for the development of distributed real-time systems in gene-
ral, and telecommunication systems in particular. With its graphical syntax, its
object-oriented features, its direct support for reuse, and its integration with
1
SDL is defined by the ITU-T, the Telecommunication Standardization Sector of the
International Telecommunication Union (ITU), in their Recommendation Z.100 [20].

Y. Gurevich et al. (Eds.): ASM 2000, LNCS 1912, pp. 242–265, 2000.
c Springer-Verlag Berlin Heidelberg 2000

 On the Formal Semantics of SDL-2000 243

other development languages, SDL satisfies the primary needs of system deve-
lopers. Today, SDL is being broadly applied in industry, and is being supported
by powerful commercial tool environments.
In 1988, the semantics of SDL has been defined formally, upgrading the langu-
age to a formal description technique (FDT). This formal definition has been
updated and maintained for subsequent versions of the SDL language in 1992
and 1996. Building on a combination of the VDM meta language Meta-IV with
a CSP-like communication mechanism, the formal semantic model provides a
comprehensive formalization of all static and dynamic aspects of SDL, though,
it is hardly manageable because of its size2 . Essentially, the formal semantics
is given by a set of Meta-IV programs that take an SDL specification as input,
determine the correctness of its static semantics, perform a number of transfor-
mations to replace non-basic language constructs, and interpret the specification.
It has been argued that this style of defining the formal semantics is particularly
suitable for tool builders.
In November 1999, a new version of the language, called SDL-2000, has been
approved as international standard. In addition to the usual language mainte-
nance, SDL-2000 offers important new features, including object-oriented data
type definitions, a unified concept for blocks, processes, and composite states,
and exception handling. Based on the assessment that the existing Meta-IV pro-
grams would be too difficult to update and maintain, it was decided to conceive
a new formal semantics for SDL-2000 from scratch. For this purpose, a special
task force, the SDL semantics group, consisting of experts from Germany and
China including the authors of this paper, was formed. The draft formal seman-
tics defined by this group (see [24]) has been determined to become the official
formal SDL semantics after its approval, which is expected for November 2000.
Among the primary design objectives of this approach is intelligibility and con-
ciseness of the formal description as a prerequisite for acceptance, correctness
and maintainability. Additionally, executability has become a key issue. Exe-
cutability calls for an operational formalism with readily available tool support.
For this and other reasons (cf. [11]), Abstract State Machines (ASMs) have fi-
nally been selected as the underlying formalism. To support executability, the
formal semantics defines, for each SDL specification, corresponding ASM code.
This differs substantially from the interpreter view taken in previous work, and
enables SDL-to-ASM-compilers. Thus, given the availability of an SDL-to-ASM-
compiler, it will be straightforward to generate reference implementations.
The approach taken by the SDL semantics group to formalize the dynamic se-
mantics of SDL-2000, though basically operational, differs in several aspects from
the approach taken to define the semantics of SDL-88 through SDL-96, and also
from the style of other ASM approaches (cf. [12,10]):

– The style of the semantics has the characteristics of a compiler approach,

as compared to the interpreter approach chosen, for instance, to define the
2
The complete formal semantics for SDL-96, as defined in Annex F to Z.100, is do-
cumented on more than 500 pages of Meta-IV descriptions.
244 R. Eschbach et al.

semantics of SDL-88 through SDL-96. A compiler usually performs lexical

and syntactic analysis functions, just like an interpreter does. Unlike an
interpreter, compilers produce executables for some kind of machine, instead
of executing a version of the source directly. Thus, the source is viewed as a
collection of instructions rather than a data structure.
– In practice, processing specifications into an abstract syntax representation
is simpler and faster than compiling it into some kind of executable code. On
the other hand, compilation into code directly enables executability. There
is a technique to combine the advantages of both approaches, namely: the
generation of abstract code ”running” on an abstract machine. The SDL
semantics group has devised an SDL Abstract Machine, SAM for short, which
is defined in terms of a distributed real-time ASM. SDL specifications are
then compiled into the instruction set of this machine.
– Compared to the standard technique for defining the set of initial states of
an ASM model—usually this is done in an axiomatic or declarative style3 —
a realistic semantic model of SDL requires a different approach. Following
the view of Z.100 on how initial system states are derived from given SDL
specifications, the SAM model starts its execution from a set of so-called
pre-initial states that are canonically defined. The actual set of initial states
is then determined in an operational way as result of an initialization phase.
This way, the initialization phase and the execution phase may even overlap.

Due to the complexity of SDL-2000, the formal semantics is of substantial size.

Therefore, in order to present and illustrate the application of the basic techni-
ques above to the definition of the SDL semantics, we have chosen a very small
language called SSL (Simple Specification Language). In the following, we define
an abstract machine for SSL, a compilation function mapping SSL specifications
to code running on this machine, and the initialization and execution phases.
Moreover, we present in more detail the semantics of SSL procedure calls. Our
model of procedure semantics allows for call-by-value and call-by-reference se-
mantics and supports the concept of static binding.
Taking into account that the work reported here is not an academic exercise,
but takes place in a real-life industrial setting, pragmatic compromises are una-
voidable. For this reason, some decisions taken in the formalization of the SDL
(and hence SSL) semantics are rather pragmatic or strongly influenced by the
SDL experts. In our opinion, such work is crucial to demonstrate the usefulness
of formalisms such as ASMs. In this sense, the paper reports on work that is still
in progress, but which has reached a stage where most ”design” decisions have
been taken, and a sufficiently stable draft is available (cf. [24]). A preliminary
version of this paper can be found in [7].
This paper is organized as follows. Section 2 gives an overview on related work.
Pragmatic aspects of our formalization approach are briefly discussed in Section
3, while Section 4 provides an overview of the formal Semantics for SDL-2000.
3
Quite often, an initial state S0 is defined through a set of first-order formulas
ϕ1 , . . . , ϕn such that S0 |= ϕ1 ∧ . . . ∧ ϕn .
 On the Formal Semantics of SDL-2000 245

Section 5 introduces the language SSL. Section 6 defines the abstract compi-
lation of SSL specifications to SSLM programs. Section 7 defines the SSL ab-
stract machine (SSLM). Section 8 describes the initialization. Section 9 presents
conclusions. Throughout this paper we assume the reader to be familiar with
Gurevich’s notion of ASM (cf. [14,15,16]).

2 Related Work
Regarding the importance of SDL as a practical tool for industrial systems en-
gineering, it is certainly not astonishing that in addition to activities within
ITU, there is a considerable variety of attempts to formalize the semantics of
SDL using various formal methods. According to the principal objectives be-
hind these attempts, one can distinguish two basically different aspects, namely:
(1) approaches that are mainly concerned with analysis and verification of SDL
system specifications; (2) approaches that are mainly concerned with formal
documentation, maintenance and validation of the SDL language definition.

2.1 Analysis and Verification

In [1], Bergstra and Middleburg define a process algebra semantics of a restricted
version of SDL, which they call ϕSDL (where ϕ stands for “flat”), as it does not
cover the structural aspects of SDL. Additionally, they introduce a restricted
notion of time to simplify time related concepts of SDL. The authors claim to
have convincing pragmatic justification for their choices; for instance, they argue
that ”a dramatically simplified version of SDL” and an adequate semantics for it
”are prerequisites for advanced analysis and formal verification”. After all, they
deploy an impressive mathematical apparatus making it difficult to see whether
ϕSDL is relevant in practice.
Broy [4], Holz and Stølen [19], and Hinkel [18] model various subsets of (es-
sentially) Basic SDL using denotational semantics based on stream processing
functions of FOCUS [5]. While it may be natural to model SDL process com-
munication as discrete streams of signals, the functional view does neither sup-
port the concept of system states and state transitions nor allows the stream
formalism for an adequate treatment of time. Even the most comprehensive mo-
del [18] builds on a fundamentally restricted notion of global system time not
allowing to express time quantities explicitly. In particular, it is not possible to
specify any quantitative delays as required for dealing with the timer concept of
SDL. Since the expiration time of a timer does not refer to a global system time
(as represented by the SDL expression now), the meaning of an active timer is
reduced to the fact that a timer signal will eventually be generated (after some
finite delay). This results in a severe oversimplification.4
4
It is certainly right that the current definition of SDL needs further improvements
allowing a more detailed and more precise inspection of timing behavior (see for
instance [22]), e.g. to determine delays caused by time consuming operations; on the
other hand, any reasonable approach to overcome these deficiencies cannot mean to
replace the current time concept by an even more primitive one, as done in [18].
246 R. Eschbach et al.

Fischer, Dimitrov and Tauber propose an extended Petri Net model, so-called
SDL Time Nets, as a formal basis to verify SDL protocol specifications [8], [9].
The transformation of SDL specifications into corresponding net models as well
as the transformation of results back to the SDL level is done automatically
within the SDL Integrated Tools Environment SITE. Though these transfor-
mations, in principle, remain invisible for the user, the authors concede that
certain knowledge is necessary about the internal data structures of the analy-
sis tools. Moreover, the integration of further net analysis algorithms is difficult
and requires detailed knowledge about the internals (like interfaces, structure
and representation formats) of the tools.

2.2 Documentation and Validation

An ambitious attempt aiming at a comprehensive formal semantics for SDL-92

as a basis for further development of SDL was started by Lau and Prinz with
their definition of BSDL (Base SDL) [21]. With the ultimate goal to simplify
the language definition, they proposed a modeling approach using Object-Z to
define a universal core for SDL as a conceptual framework to deal with the main
building blocks of the language. This core should be as clear and concise as pos-
sible, yet it should include the basic model of processes and communication, all
the object-oriented concepts and partly also structuring concepts. The meaning
of additional language constructs (not handled within the core) should then be
fixed by defining their transformation to the core. In that respect, the BSDL
approach is similar to the original formal model of SDL, but suggests a more
systematic and robust construction of a formal semantic model. However, the
Object-Z definition of BSDL was rejected by the ITU even before the work was
completed.
Regarding further developments of SDL towards SDL-2000, an approach based
on transition systems generated from attributed abstract syntax trees (ASTs)
is outlined in [13]. An attribute grammar is formed by adding evaluation rules
to the abstract grammar of the language, where attributes for instance may
represent actions. This way, one obtains for each complete SDL specification an
attributed AST containing all the information required to generate a behavior
model. The underlying mathematical model and the notation are not determined,
which leaves room for subsequent choices.
In [12] and [10], an ASM behavior model for Basic SDL-92 is defined. Starting
from an abstract operational view, the dynamic properties of Basic SDL-92 are
formalized in terms of an abstract SDL interpreter based on a distributed real-
time ASM model. This work provides a conceptual framework that has further
been developed and extended by combining it with the compiler-based approach
from [13], as well as certain BSDL concepts from [21], resulting in a robust formal
basis for the definition of an ASM behavior model for SDL-2000, as discussed
here.
 On the Formal Semantics of SDL-2000 247

3 Pragmatic Foundation

This section addresses fundamental questions concerning the practicability, sui-

tability and robustness of our formalization approach. The primary focus here is
on pragmatic aspects rather than on technical details. That is, we try to explain
at a conceptual level distinctive features of our SDL behavior model. In particu-
lar, we try to answer the question, ”Why and how does the abstract operational
view of the ASM method in combination with the compiler approach practiced
here naturally meet the demands on formal documentation, language mainten-
ance and tool development as arising from the SDL standardization process?”.
Industrial engineering of complex distributed systems based on SDL, as already
outlined in Sect. 1, has quite a tradition in telecommunication technology. Mean-
while, one can as well observe an increasing proliferation of SDL applications in
various fields of information technology (e.g., engineering of distributed control
systems). Like the development of SDL over its life cycle, this progress has been
influenced by the apparent convergence of information technology and commu-
nication technology.
To meet the needs of system design experts, the language has been continuously
improved over a period of more than 20 years, evolving from a primitive gra-
phical notation to a sophisticated formal description technique, according to the
demands on industrial systems technology. Consequently, the underlying langu-
age concepts closely reflect the view and the intuitive understanding of system
behavior and system structure of those who have mainly contributed (through
their practical experience) to the development of SDL (cf. the standard litera-
ture, e.g. [23,6]). Hence, in order to gain acceptance, any realistic attempt to
formalize the semantics of SDL with the intention to facilitate language design
and validation of implementations therefore cannot succeed without supporting
the intuitive understanding of well established SDL concepts (including functio-
nal, structural and timing aspects) as defined in Z.100.

3.1 Abstract Operational View

The problem of ”turning English into mathematics” is considerably simplified by

a close correspondence between the specification method and the semantic model
to be formalized. In fact, this allows for more natural and adequate abstractions
resulting in more comprehensible and more reliable descriptions.5 Given the defi-
nition of SDL in Z.100, it makes perfect sense to formalize dynamic properties of
SDL specifications as abstract machine runs since this style of modeling directly
reflects the abstract operational view of the informal language description. Con-
ceptually, the effect of actions and events as associated with active SDL objects
(i.e., those objects that have a behavior) on the global system state is stated in
terms of the effect of SAM operations on abstract machine states.
5
For instance, was it possible to eliminate a number of cumbersome in-language trans-
formations (specified by so-called Model Sections) from Z.100 by defining the mea-
ning of certain non-basic SDL constructs directly in the formal behavior model.
248 R. Eschbach et al.

More specifically, the resulting abstraction depends on the granularity of ab-

stract machine operations specified by the SAM instruction set. Ideally, this
granularity is such that all relevant effects on the system state are highlighted,
whereas minor operational details remain hidden in the functions and predicates
on top of which these operations are formulated. Additional details may then
be introduced through stepwise refinements such that the style of representa-
tion gradually turns from declarative to operational while crossing abstraction
levels.6
Essential for a direct and concise encoding of SDL behavior primitives in ASM
notation is of course a close correspondence of the underlying computation mo-
dels. In that respect, there are two observations concerning our modeling ap-
proach:

- The SDL view of distributed systems with real-time constraints and the
semantic modeling concept of distributed real-time ASM clearly coincide.
That is, essential properties of the underlying computation models – namely,
the notions of concurrency, reactivity and time as well as the notion of state –
are so tightly related that the common understanding of SDL can directly be
converted into a formal semantic model avoiding any formalization overhead.
- Even without direct support of object-oriented features in ASMs, the resul-
ting ASM model of the dynamic semantics is particularly concise, readable
and understandable. Furthermore, this model can easily be extended and
modified as required for an evolving technical standard. Beyond the purpose
addressed here, the model may as well be utilized for defining a bridging se-
mantics in order to combine SDL with other modeling languages (e.g., UML
and VHDL).

3.2 Correctness and Reliability

Defining semantics by a mapping to a well defined semantic basis, as realized by
the compilation of SDL specifications into SAM code, is a well-known technique,
but is often considered to be as dangerous as helpful. Of course one may argue
that this leads to the old ”the compiler defines the language semantics” problem:
”Who validates and/or verifies the correctness of this mapping and with respect
to which origin?”.
Now, the question here is: ”How can one establish that a formal semantic model
like the ASM behavior model of SDL faithfully reflects the intuitive understan-
ding of SDL, i.e. yields a valid interpretation of the definitions in the language
reference manual?”. Since there is no way of proving correctness (in a strict
mathematical sense), the approach taken here is to construct an ASM behavior
6
An illustrative example is the abstract definition of the behavior model for SDL
channels (see [24]). By exploiting the decentralized organization of the underlying
signal flow model, the entire static information on the system structure and the
reachability constraints are completely encapsulated into a predicate compatible,
which can then be refined by introducing technical details on how this information
is to be derived from a given specification.
 On the Formal Semantics of SDL-2000 249

model that reflects the dynamic properties of the language so closely that cor-
rectness can be established by observation and experimentation. In ASM terms,
such a model is called a ground model, cf. [2]. The quality of ground models
considerably depends on the underlying abstractions for dealing with real-world
phenomena, i.e. on the way in which basic objects and operations of a ground
model are related to basic entities and actions as observed in the real world.
Investigating the construction of ground models (as part of the analysis and
design process), one can identify general principles for justifying their appro-
priateness and validity. In [2], Egon Börger convincingly argues that this justifi-
cation process has three basic dimensions, namely: a conceptual justification, an
experimental justification, and a mathematical justification. According to this
understanding, the best we can achieve is a solid pragmatic foundation.

4 Overview of the Formal SDL Semantics

The definition of the formal semantics of SDL-2000 is structured into the follo-
wing major parts:

– grammar,
– well-formedness conditions (static semantics),
– transformation rules, and
– dynamic semantics.

The grammar defines the set of syntactically correct SDL specifications. In the
SDL standard, a concrete textual, a concrete graphical, and an abstract gram-
mar are defined using BNF with some extensions to capture graphical language
elements. The abstract grammar is obtained from the concrete grammars by
removing irrelevant details such as separators and lexical rules.
The well-formedness conditions impose additional context constraints on syntac-
tically correct specifications, such as which names it is allowed to use at a given
place, which kind of values to assign to variables, etc.
For some language constructs, the formal semantics is not provided directly, but
through transformation rules. These rules define how a given specification is to
be transformed in order to replace the language constructs by basic language
elements, and are formally represented as rewrite rules.
Finally, the dynamic semantics is given to syntactically correct SDL specificati-
ons satisfying the well-formedness conditions, after application of the transfor-
mation rules. The dynamic semantics consists of the following parts as illustrated
by figure 1.

– The SDL Abstract Machine (SAM) defines basic signal flow concepts of SDL
such as signals, timers, exceptions, and gates in terms of the ASM model.
Furthermore, ASM agents are specialized to model agents in the context of
SDL. Finally, several signal processing and behavior primitives - in a sense
the abstract machine instructions of the SAM - are predefined.
250 R. Eschbach et al.

Abstract Syntax
Structure Behaviour Data
Interface
Compilation
Initialization

Signal Flow Agents Primitives

SDL Abstract Machine

ASM

Fig. 1. Overview of the Dynamic Semantics

– An SDL Abstract Machine Program defines a set of computations. Com-

putations consist of an initialization phase and an execution phase. SAM
programs have predefined parts that are the same for all SDL specifications,
and variable parts that are generated from the abstract syntax representa-
tion of a given SDL specification.
– The initialization handles static structural properties of the specification
by recursively unfolding all the static objects of the specification. In fact,
the same happens during the execution phase when new SDL agents are
created dynamically. From this point of view, the initialization is merely
the instantiation of the SDL system agent.
– The compilation function maps behavior representations into the SAM
primitives. This function amounts to an abstract compiler taking the
abstract syntax tree (AST) of an SDL specification as input and trans-
forming it to the abstract machine instructions.
– The data semantics is separated from the rest of the dynamic semantics by
an interface. The use of an interface is intentional at this place. It will allow
to exchange the data model, if for some domain another data model is more
appropriate than the SDL built-in model. Moreover, also the SDL built-in
model can be changed this way without affecting the rest of the semantics,
thus enhancing its maintainability.

As in the past, the new formal semantics is defined starting from the abstract
syntax of SDL, which is documented in the SDL standard. From this abstract
syntax, a behavior model that can be understood as abstract code generated
 On the Formal Semantics of SDL-2000 251

from an SDL specification is derived. The dynamic semantics associates, with

each SDL specification, a particular distributed, real-time ASM. Intuitively, an
ASM consists of a set of autonomous agents cooperatively performing concurrent
machine runs. This approach differs substantially from the interpreter view taken
in the previous SDL semantics definition, and will enable SDL-to-ASM compilers.
Conceptually, one can identify three basic layers at which our semantic model
of SDL is actually defined (either explicitly or implicitly), namely:

(2) the compilation function,

(1) the SAM architecture, and
(0) the semantic model of distributed real-time ASM.

Regarding layers (2),(1), it is not a priori fixed which aspects of behavior are
subject to modeling at which one of these two layers. Though, in principle,
specific aspects depending on a given SDL specification typically ought to be
encoded into the compilation function, whereas invariant dynamic properties
are more naturally expressed in terms of the SAM model. Finally, to keep the
formalization as concise and elegant as possible, all aspects that both computa-
tion models (SDL and ASM) have in common are preferably given implicitly by
mapping SDL behavior primitives directly onto corresponding primitives of the
underlying ASM model at layer (0).
By splitting the definition of the SDL behavior model into the above layers, one
obtains a flexible framework for semantic modeling that allows for natural and
adequate abstractions, transparent and well defined formalization, and flexible
rearrangements on the basis of an executable semantics.

5 SSL: A Simple Specification Language

In order to illustrate our approach to the formal SDL-2000 semantics and the
application of the basic techniques introduced in section 1 we have chosen a
very small language called SSL (Simple Specification Language). SSL is unty-
ped, and is based on a small number of statements including assignment and the
conditional statement. Furthermore, an SSL-Specification may contain proce-
dure definition and procedure calls. A procedure definition can contain reference
parameters and value parameters.

5.1 Basics
The syntax of SSL is given below. We use Name as a predefined lexical unit and
describe the syntax using a variant of BNF (Backus-Naur-Form). More precisely
we will just give an abstract syntax of the language. We assume the concrete
syntax to be handled by an ordinary parser that provides us with the abstract
representation in terms of the abstract syntax.
Please note, that we use prefixes in the syntax to make the role of the items
clear. Prefixes are separated from the nonterminal name by a ”-”. Prefixes do
not have a semantics.
252 R. Eschbach et al.

5.2 Specifications and Procedures

An SSL specification describes a concurrent system. It is given as a list of varia-

bles, a list of local specifications, a list of locally defined procedures, and a main
expression. To execute an SSL specification means to first start the execution of
the enclosed SSL specifications and then to evaluate the main-expression. The
enclosed specifications evaluate their main-expression concurrently with.

sslSpec ::= spec-Name variableDecl∗ sslSpec∗ procedureDef∗

main-expression
variableDecl ::= Name

specification Hanoi
procedure hanoi (ref x,y,z, val n)
if n = 1 then
x := x - 1; z := z + 1; return;
else
call hanoi(x,z,y,n-1); call hanoi(x,y,z,1);
call hanoi(y,x,z,n-1); return;
end hanoi

var x,y,z,n
begin Hanoi
x := 4; y := 0; z := 0; n := 4; call hanoi(x,y,z,n);
end Hanoi

Fig. 2. Specification Hanoi

A procedure is given as a list of its parameters and an expression for the body.

procedureDef ::= proc-Name refParam∗ valParam∗ variableDecl∗

procedureDef∗ main-expression
refParam ::= Name
valParam ::= Name

5.3 Expressions

Expressions in the scope of SSL include also the usual behavior of statements. In
this case, their return value is simply ignored. The nonterminal function denotes
constants (like 0 and 1) as well as functions (as + and -). The meaning of
these constructs is supposed to be part of the language. The semantics of ifExpr,
variable, assignment, procCall, procReturn and equality is as usual. For sequence,
 On the Formal Semantics of SDL-2000 253

the semantics is the sequential interpretation of the subexpressions. The resulting

value of the sequence is the last one.

expression ::= function | ifExpr | variable | equality | sequence |

assignment | procCall | procReturn
function ::= Name expression∗
ifExpr ::= test-expression then-expression [else-expression]
variable ::= Name
equality ::= lhs-expression rhs-expression
sequence ::= fst-expression nxt-expression
assignment ::= Name expression
procCall ::= Name ref-Name∗ expression∗
procReturn ::= expression

Please note, that the syntax domain names are also used as the constructors for
the syntax tree. A sample SSL specification can be found in figure 2.

6 Compilation

The compilation defines a mapping from the behavioral part of the abstract
syntax into ASM behavior. In fact, an abstract code is defined as the result of
this mapping. The behavior belonging to the behavior primitives in the abstract
code is defined within section 7.

6.1 Basics

In this section, we define a compilation function for SSL specifications and for
SSL procedures.
compileSpec : sslSpecification → AbstractCode
compileProc : procedureDef → AbstractCode
For the representation of code, we use a special labeling. Each instruction has
a label attached to it. The specification keeps the information of the current
label and selects the instruction to process according to this label. Therefore, we
introduce a static domain Label. The abstract code is then a set of instructions
and an instruction a tuple consisting of a label and a primitive.
AbstractCode = Instruction-set
Instruction = Label × Primitive.
The compiled code of all specifications and procedures is merged together in the
initial state. Specifications and procedures are characterized by the functions
entrySpec : sslSpecification → Label,
entryProc : Name → Label,
254 R. Eschbach et al.

which return the entry labels. For SSL we introduce the following primitives.
Again, the prefixes are without semantic relevance.

Primitive = Pfun ∪ Pif ∪ Pvar ∪ Passign ∪ Pcall ∪ Preturn

Pfun = Name × value- Label∗ × next- Label
Pif = condition- Label × then- Label × next- Label
Pvar = Name × next- Label
Passign = Name × value- Label × next- Label
Pcall = Name × value- Label∗ × ref- Name∗ × next- Label
Preturn = value- Label

In order to define the compilation function, we introduce two auxiliary functions.

unique : expression → Label
This function is provided by the compiler: a unique labeling of the expressions.
Entry labels of expressions are determined by a recursive function
entryLabel : expression → Label
In the remaining text, we use the following convention: we often write x.f instead
of f(x). Function entryLabel is recursively defined as follows.
entryLabel(e : expression) ≡
if e ∈ sequence then entryLabel(e.fst-expression)
elsif e ∈ equality then entryLabel(e.lhs-expression)
elsif e ∈ assignment then entryLabel(e.expression)
elsif e ∈ ifExpr then entryLabel(e.test-expression)
elsif e ∈ procReturn ∧ e.expression 6= undef
then entryLabel(e.expression)
elsif e ∈ function ∧ e.expression-seq 6= empty
then entryLabel(e.expression-seq[1])
elsif e ∈ procCall ∧ e.expression-seq 6= empty
then entryLabel(e.expression-seq[1])
else unique(e)
Given these domains and functions, we can define the compilation function and
the entry labels as below.

6.2 Compilation of Expressions

The compilation of expressions is the main part of the compilation function.
We present below a part of it in order to take the idea clear. There is nothing
complicated to the omitted parts. We have introduced appropriate comments
marked with ’//’.
compileExpr(e : expression, next : Label) ≡
if e ∈ variable // evaluate the variable
then {mk-Instruction(unique(e), mk-Pvar(e.Name, next))}
 On the Formal Semantics of SDL-2000 255

elsif e ∈ ifExpr // first evaluate the test, then switch to then or to else
then compileExpr(e.test-expression, unique(e)) ∪
{ mk-Instruction(unique(e),
mk-Pif(unique(e.test-expression),
entryLabel(e.then-expression), elseLabel)} ∪
compileExpr(e.then-expression, next) ∪
compileExpr(e.else-expression, next)
else . . .
where
elseLabel ≡
if e.else-expression = undef then next
else entryLabel(e.else-expression)

6.3 Compilation of Specifications and Procedures

The compilation of a specification is just the compilation of its main expression.
The entry label of a specification is the entry label of its main expression.
compileSpec(s : sslSpecification) ≡
compileExpr(s.main-expression, undef)
entrySpec(p : sslSpecification) ≡
entryLabel(s.main-expression)
The compilation of a procedure is just the compilation of its main expression.
The entry label of a procedure is the entry label of its main expression.
compileProc(p : procedureDef) ≡
compileExpr(p.main-expression, undef)∀
p ∈ procedureDef : entryProc(p.Name) =
entryLabel(p.main-expression)
The simplified result of the compilation of the sample SSL specification (see
figure 2) without procedure hanoi can be found below in figure 3. Apart from
the generated code, the following equation hold:
specEntry(Hanoi) = H0

7 SSLM: An Abstract Machine for SSL

In this section we introduce an abstract machine for SSL called SSLM, corre-
sponding to the SDL Abstract Machine. The intention of our approach is to
compile each SSL-instruction to a set of ”simple” SSLM instructions, and to
give these ”simple” instructions a rigorous meaning. Like in [17], and [3], we
present the definition of SSLM in several steps. In each step we add some details
of a particular aspect of the language SSL. After introducing some basics, we
present in the first step the semantics of “simple” primitives like the evaluation
of assignments. In a second step we add the meaning of procedure calls and
returns. In the last step we add the distribution to SSLM.
256 R. Eschbach et al.

Label Primitive Next Comment

H0 fun(4,hi) H1 value of 4
H1 assign(x,H0) H2 x:=4
H2 fun(0,hi) H3 value of 0
H3 assign(y,H2) H4 y:=0
H4 fun(0,hi) H5 value of 0
H5 assign(z,H4) H6 z:=0
H6 fun(4,hi) H7 value of 4
H7 assign(n,H6) H8 n:=4
H8 var(n) H9 value of n
H9 proc(hanoi,hH8i,hx,y,zi) undef hanoi(x,y,z,n)

Fig. 3. Compilation of Specification Hanoi

7.1 Basics

In this section we describe how states are realized in SSLM. Usually, a state is
a mapping of variable names to values. In SSLM locality and scope of variables
are given by a model in which variable names are mapped to (memory) locati-
ons containing values. This model allows for call-by-value and call-by-reference
semantics, i.e. procedures can have reference and value parameters. To this aim
we introduce a dynamic domain Location for locations and dynamic, controlled
functions
loc : VarName → Location
cont : Location → Value.
We use here and in the following VarName, FuncName, and ProcName as syno-
nyms for Name. We use in the following a static, controlled function
splmInstruction : Label → Primitive,
which treats the sets of pairs returned by the compilation functions as a function.
This is possible since the compilation functions always compute a graph of a
function. Furthermore, we introduce a kind of program counter, modeled by
currentLabel :→ Label,
which returns the label associated with the instruction to be executed next.
During the dynamic execution of a program intermediate values can be stored
at labels. More precisely, we introduce the dynamic, controlled function
labelValue : Label → Value,
for this purpose. In case of a sequence seq = (l1 , . . . ln ) of labels we simply write
labelValues(seq) for the sequence (labelValue(l1 ), . . . , labelValue(ln )).
Function names are evaluated by a static function that is provided by the data
semantics of the language. This would again be part of the data interface in the
case of SDL.
evalFunc : FuncName × Value∗ → Value
 On the Formal Semantics of SDL-2000 257

7.2 Simple Primitives

In this section we describe the semantics of simple primitives of SSLM. These
primitives are used for the evaluation of expressions, assignments, and conditions.
The evaluation of expressions is realized by macros EvalFunc and EvalVar. Con-
stant names and function names are evaluated with the macro EvalFunc. This
macro requires three formal parameters: a function name funcName, a sequence
of labels labels, and a continue label next. The intention is that values stored at
labels are extracted with function valueLabel. Function funcName is evaluated
with these values, and the resulting value is stored at currentLabel.
EvalFunc(f uncN ame, labels, next) ≡
labelValue(currentLabel) := evalFunc(f uncN ame, labelValues(labels)),
currentLabel := next
Variable names are evaluated with the macro EvalVar in a similar way as function
names.This macro requires two formal parameters: a variable name varName,
and a continue label next.
EvalVar(varN ame, next) ≡
labelValue(currentLabel) := cont(loc(varN ame)),
currentLabel := next
Assignments are evaluated with the macro EvalAssign. This macro requires three
formal parameters: a variable name varName, a value label valueLabel , and a
continue label next. Macro EvalAssign essentially sets the value of variable name
varName to the value stored in label and jumps to continue label next.
EvalAssign(varN ame, label, next) ≡
cont(loc(varN ame)) := labelValue(label),
currentLabel := next
Furthermore, there exists a macro EvalCondition which evaluates conditions. Due
to its simplicity we omit the corresponding definition.

7.3 Procedures
Before we can define macros for procedure call and return we must have some
means for storing multiple values of several incarnations of a function. To this aim
we introduce frames which are organized as stacks. The universe Frame comprises
at each point of time a finite set of frames which initially contains one element.
A dynamic, controlled function topFrame :→ Frame indicates the current top
of the stack. The function prev : Frame → Frame returns for each frame its
predecessor (if there is any). The following dynamic, controlled functions are
introduced for the procedure call and return mechanism. The former function
corresponds to a kind of return address, the latter gives the information at which
label the result of a procedure must be stored.
returnLabel : Frame → Label
resultLabel : Frame → Label
258 R. Eschbach et al.

We change the declaration of function loc to

loc : Name × Frame → Value.
This means we have to rewrite almost every rule defined so far. Instead of doing
so we simply state that every previous occurrence of f (x) should be replaced by
f (x, topFrame), where f is one of the functions {loc, labelValue} (cf. figure 4).
For example, rule EvalVar becomes now

EvalVar(varN ame, next) ≡

labelValue(currentLabel, topFrame) :=
cont(loc(varN ame, topFrame)),
currentLabel := next

Macro EvalCall requires four formal parameters: a procedure name procName, a

sequence of labels labels, a sequence of variable names varNames, and a return
label label . A procedure call leads to the creation of a new frame. The new
frame is pushed on the stack of frames, and return-label and result-label for this
frame are set. New Variable bindings are established by a parameter passing
mechanism or by the creation of local variables.

EvalCall(procN ame, labels, varN ames, label) ≡

extend Frame with f rame
PushNewFrame(f rame, label),
ParameterPassing(procN ame, labels, varN ames, f rame),
CreateLocalVars(procN ame, topF rame),
currentLabel := procEntry(procN ame)

Macro PushNewFrame pushes frame on the stack of frames, and updates returnLabel
and resultLabel.

PushNewFrame(f rame, label) ≡

topFrame := f rame, prev(f rame) := topFrame,
returnLabel(f rame) := label, resultLabel(f rame) := currentLabel

Macro ParameterPassing realizes the parameter passing mechanism, i.e. it handles

passing of value and reference parameters.

ParameterPassing(procN ame, labels, varN ames, f rame) ≡

PassRefParameters(procN ame, varN ames),
PassValParameters(procN ame, labels),
StaticBinding(procN ame, f rame)

Macro PassRefParameters has two formal arguments: the procedure name procName
and the sequence of actual reference parameters varNames. This macro realizes
the parameter passing mechanism for reference parameters. This means referen-
ces of variable names to locations in current state are changed appropriately.
We use a function refPar : ProcName × Nat → VarName to determine reference
parameters. Function |.| denotes the length function for sequences.
 On the Formal Semantics of SDL-2000 259

PassRefParameters(procN ame, varN ames) ≡

forall i ∈ {1, ..., |varN ames|} do
loc(refPar(procN ame, i), topFrame) :=
loc(varN ames(i), topFrame)

Macro PassValParameters creates for each value parameter a new location. Each
label in the sequence labels has a value in frame. We write simply labels(i)
to project on the i-th component of this sequence. The value of a new loca-
tion location is set to the value stored in labels and frame where i ranges over
{1, . . . , |labels|}. The location for the i-th value parameter given by a function
valPar : ProcName × Nat → VarName is set to a new location. We assume that
the size of labels equals the number of value parameters.

PassValParameters(procN ame, labels) ≡

forall i ∈ {1, ..., |labels|} do
extend Location with location
cont(location) := labelValue(labels(i), topFrame),
loc(valPar(procN ame, i), topFrame) := location

Let p be a procedure name. We introduce a static function

binding : VarName × ProcName → ProcName
which returns for each variable name and each procedure name the name of a
procedure to which variable name is bound according to the static scope rules.
Function findLocation : VarName × ProcName × Frame → Location evaluates for
a given variable name x the stack starting with frame in order to determine the
first frame F such that the equation binding(x, procName) = proc(F ) holds. The
dynamic, controlled function proc : Frame → ProcName returns for each frame
the name of the corresponding procedure.

findLocation(varN ame, procN ame, f rame) ≡

if binding(varN ame, procN ame) = proc(f rame)
then loc(varN ame, f rame)
else findLocation(varN ame, procN ame, prev(f rame))

Macro StaticBinding saves procName in frame creates a variable binding for those
variables which are not defined in the body of procName.

StaticBinding(procN ame, f rame) ≡

proc(f rame) := procN ame,
forall varN ame ∈ VarName do
if binding(varN ame, procN ame) 6= procN ame
then loc(varN ame, topFrame) :=
findLocation(varN ame, procN ame, topFrame)

Macro CreateLocalVars creates new location for local variables associated with a
procedure or a program. Local variables are determined by a means of a static
function locVars : Name → Name-set.
260 R. Eschbach et al.

CreateLocalVars(name, f rame) ≡
forall varN ame ∈ locVars(name) do
extend Location with location
loc(varN ame, f rame) := location
Macro EvalReturn essentially removes the top frame of the stack, saves the result
according to the result label, and jumps to the return label.
EvalReturn(label) ≡
topFrame := prev(topFrame),
labelValue(resultLabel(topFrame), prev(topFrame)) :=
labelValue(label, topFrame),
currentLabel := returnLabel(topFrame)

Label
VarName
L7
returnLabel
L6 x y
Frame
L5
resultLabel
L4 F2
L3 loc

L2 F1
L1
L0 labelValue F0 l0 l1 Location
cont

10 14 14 Value

Fig. 4. State and Control Information

7.4 Distribution
In this section we refine the SSLM defined so far to a distributed abstract ma-
chine. The underlying distribution model consists of several agents which execute
their programs and communicate by means of shared variables. The refinement
step is done in a similar way as in the preceding section, namely by changing the
declaration of some existing functions. We provide each agent with its own pro-
gram counter. To this aim we change the declaration of functions currentLabel,
and topFrame to
currentLabel : Agent → Label,
topFrame : Agent → Frame.
Every previous occurrence of f inside macros should be replaced by f (Self),
where f is one of the functions {currentLabel, topFrame}.
 On the Formal Semantics of SDL-2000 261

8 Initialization and Execution

With the abstract machine for SSL being defined, we can now assign a formal
semantics to SSL specifications by associating, with each specification, a set of
initial states and an ASM program. Please note that conceptually, it is better
to think of SSLM programs, because the instructions considered here are SSLM
instructions. Formally, the programs below are ASM programs.

8.1 Initialization
In the ASM model, the set of initial states is usually defined by constraints
imposed on domains, functions, and predicates. These constraints are required
to hold in the first state of each run of the abstract state machine. Initial con-
straints are stated by formulae that are preceded by the keyword initially, i.e. in
a property-oriented style.
As it turned out during work on the formal SDL semantics, a property-oriented
definition of the set of initial states, given by the set of SDL agents, SDL agent
sets, and link agents together with the initial values of all functions and predi-
cates defined on these agents is rather complex and hard to follow. Therefore, a
different approach has been chosen, which may be characterized as a combination
of constraints and initial computations: several (simple) constraints determine
the so-called set of pre-initial states. The actual set of initial states is determined
by the initialization phase that starts in a pre-initial state, and creates a set of
agents as well as the system structure.
To give a flavor of the approach taken in the formal SDL semantics, we adapt
it here to define the set of initial states associated with an SSL specification. A
nullary function
main :→ Agent
is introduced to refer to the pre-initial system agent. Furthermore, we define
topLevelSpec :→ sslSpec
mySpec : Agent → sslSpec
The set of pre-initial states is defined by the following set of constraints (strictly
speaking, the set of pre-initial states is the set of initial states of the ASM model):

initially Agent = {main}

initially sslmInstruction
S =
S {p ∈ sslSpec : p.compileSpec} ∪
{p ∈ procedureDef : p.proc-Name.compileProc}
initially main.mySpec = topLevelSpecification
initially main.currentLabel = topLevelSpecification.entrySpec
initially main.Mod = InitModule
initially Frame = {fstFrame}
initially main.topFrame = fstFrame
The value of the static function sslmInstruction is determined by the compila-
tion defined in Section 6. It is the complete code of all behavior parts within
262 R. Eschbach et al.

the specification. Furthermore, the pre-initial values of certain functions are

defined. The function mySpec associates with each agent the abstract syntax
description of its specification. For agent main, this is defined by the nullary
function topLevelSpecification, that is, the main specification. The start label of
a specification is obtained from the function entrySpec, as defined in Section 6.
Finally, the behavior of agent main is determined by assigning the ASM module
InitModule, which determines the steps of main during the initialization phase:

InitModule ≡
forall p ∈ Self.mySpec.sslSpec-seq do
extend Agent with a
extend Frame with f rame
a.mySpec := p, a.topFrame := f rame
a.currentLabel := p.entrySpec, a.Mod := InitModule,
CreateLocalVars(Self.mySpec.spec-Name, Self.topFrame)
Self.Mod := ExecModule

Firing of InitModule creates, for each SSL specification defined in the specifi-
cation of the current agent, one agent executing the InitModule. Furthermore
local variables of the corresponding SSL specification are created. This leads to
a recursive unfolding of the tree of SSL specifications defined by the top level
specification. Macro CreateLocalVars is defined in section 7.3.
In the SDL semantics, it is important that the initialization phase is comple-
ted before the execution phase starts. This is achieved by a mechanism that
blocks agents after their initialization until the initialization phase has ended.
For simplicity, we have not modeled this mechanism in the solution presented
here.

8.2 Execution of SSL Specifications

The execution of an SSL specification is given by the following ASM module:

ExecModule ≡ EvalInstruction(Self.currentLabel.sslmInstruction)

The evaluation of an instruction is given by macro EvalInstruction. Depending

on the type of an instruction a corresponding macro is executed. For example,
an instruction of type Passign leads to the execution of macro EvalAssign as
illustrated in the following.

EvalInstruction(p) ≡
if . . .
elsif p ∈ Passign
then EvalAssign(p.Name, pvalue-Label, p.next-Label)
elsif p ∈ Pcall
then EvalCall(p.Name, p.value-Label-seq, p.ref-Name-seq, p.next-Label)
...
 On the Formal Semantics of SDL-2000 263

9 Conclusions
In this paper, we have presented the transformational/operational approach
which has been successfully applied to define the semantics of SDL-2000 for-
mally. We have illustrated some basic techniques of the approach using SSL, a
simple specification language.
The following topics have been covered:
– Abstract Machine: The semantics is based on an abstract machine. This
machine is an extension of the ASM model and is specifically tailored towards
the needs of the specification technique. It introduces an abstract code for the
FDT and could therefore also be used for its implementation. The abstract
machine comprises a set a machine instructions and their semantics as well
as some domains and functions. An important part of the abstract machine
is the built-in procedure model that allows for call-by-value and call-by-
reference parameter passing semantics. Furthermore, the concept of static
binding is built into the abstract machine.
– Compilation: Any program/specification of the specification technique is gi-
ven a semantics by transforming it into the abstract machine. This is es-
sentially an abstract compilation function. The compilation transforms the
domain of the abstract syntax trees into the domain of abstract machine
instruction sets. The execution of these sets is then covered within the de-
scription of the abstract machine.
– Initialization: In order to run a specification on the abstract machine, a spe-
cial initialization has to be performed. To this end, the abstract syntax tree
is used to determine the internal parts of the specification to be initialized.
These parts are then initialized as well. This way, the initialization takes
care of structural information, while the behavioral information is handled
by the compilation function. The same initialization behavior could also be
used for dynamic creation of specifications.
The subdivision of the semantics into the parts presented above has proved very
useful in the context of SDL, as it is essential here to be able to adjust the
semantics to the development of the FDT. This is achieved by using two layers
of abstraction: (1) abstract the concrete syntax into an abstract syntax, (2)
transform the abstract syntax to an abstract machine.
It should be noted that the formal SDL semantics has been conceived in parallel
to the language definition itself. During this process, several substantial changes
to the language definition, which turned out to be a ”moving target”, were made.
These changes of course affected the formal semantics definition, but usually did
not touch the SAM.
Finally, it should again be stressed that the work reported here is not an academic
exercise, but takes place in a real-life, industrial setting. In our opinion, it is
this kind of work academic efforts should eventually lead to. The successful
application of mathematical formalism to real-world problems and their approval
by industry is a strong selling point for having formalisms at all. In this sense, the
work in progress reported in this paper is one important step in this direction.
264 R. Eschbach et al.

Acknowledgment. We thank the anonymous referees and the organizers of

the ASM2000 workshop for their constructive criticism on a preliminary version
of this paper.

References
1. J. A. Bergstra and C. A. Middleburg. Process Algebra Semantics of ϕSDL. Techni-
cal Report UNU/IIST Report No. 68, UNU/IIST, The United Nations University,
April 1996.
2. E. Börger. High Level System Design and Analysis using Abstract State Machines”.
In Traverso Ullman Hutter, Stephan, editor, Current Trends in Applied Formal
Methods (FM-Trends 98), LNCS. Springer, 1999.
3. E. Börger and W. Schulte. A Programmer Friendly Modular Definition of the
Semantics of Java. In J. Alves-Foss, editor, Formal Syntax and Semantics of Java,
LNCS. Springer, 1998.
4. M. Broy. Towards a Formal Foundation of the Specification and Description Lan-
guage SDL. Formal Aspects of Computing 3, (3):21–57, 1991.
5. M. Broy, F. Dederichs, C. Dendorfer, M. Fuchs, T. F. Gritzner, and R. Weber.
The design of distributed systems – an introduction to FOCUS. Technical Report
TUM-19202-2 (SFB-Bericht Nr. 342/2-2/92/A), Institut für Informatik, Techni-
sche Universität München, January 1993.
6. J. Ellsberger, D. Hogrefe, and A. Sarma. SDL: Formal Object-oriented Language
for Communicating Systems. Prentice Hall Europe, 1997.
7. R. Eschbach, R. Gotzhein, and A. Prinz. On the Formal Semantics of SDL-2000: A
Compilation Approach using Abstract State Machines. In Local Proc. of ASM2000,
Monte Verità, Switzerland, 2000.
8. J. Fischer and E. Dimitrov. Verification of SDL protocol specifications using ex-
tended Petri Nets. In Proc. of the Workshop on Petri Nets and Protocols of the
16th Intern. Conf. on Application and Theory of Petri Nets, pages 1–12. Torino,
Italy, 1995.
9. J. Fischer, E. Dimitrov, and U. Taubert. Analysis and formal verification of SDL‘92
specifications using extended Petri Nets, 1995. internal report.
10. U. Glässer. ASM semantics of SDL: Concepts, methods, tools. In Y. Lahav,
A. Wolisz, J. Fischer, and E. Holz, editors, Proc. of the 1st Workshop of the SDL
Forum Society on SDL and MSC (Berlin, June 29 - July 1, 1998), volume 2, pages
271–280, 1998.
11. U. Glässer, R. Gotzhein, and A. Prinz. Towards a New Formal SDL Semantics
Based on Abstract State Machines. In SDL’99 - The Next Millennium, Proc. of
the 9th SDL Forum. Elsevier Science B.V., 1999.
12. U. Glässer and R. Karges. Abstract State Machine semantics of SDL. Journal of
Universal Computer Science, 3(12):1382–1414, 1997.
13. R. Gotzhein, B. Geppert, F. Rößler, and P. Schaible. Towards a new formal SDL
semantics. In Y. Lahav, A. Wolisz, J. Fischer, and E. Holz, editors, Proc. of the
1st Workshop of the SDL Forum Society on SDL and MSC (Berlin, June 29 - July
1, 1998), volume 1, pages 55–64, 1998.
14. Y. Gurevich. Evolving Algebras 1993: Lipari Guide. In E. Börger, editor, Specifi-
cation and Validation Methods. Oxford Univ. Press, 1995.
15. Y. Gurevich. May 1997 draft of the ASM guide. Technical Report CSE-TR-336-97,
University of Michigan, 1997.
 On the Formal Semantics of SDL-2000 265

16. Y. Gurevich. The sequential ASM thesis. Bulletin of the European Association for
Theoretical Computer Science, 67:93–124, February 1999.
17. Y. Gurevich and J. K. Huggins. The Semantics of the C Programming Language.
In Selected papers from CSL’92 (Computer Science Logic), LNCS, pages 274–308.
Springer, 1993.
18. U. Hinkel. Formale, semantische Fundierung und eine darauf abgestützte Verifika-
tionsmethode für SDL. Ph.D., Techn. Univ. München, 1998.
19. E. Holz and K. Stølen. An Attempt to Embed a Restricted Version of SDL as a
Target Language in Focus. In St. Leue D. Hogrefe, editor, Proc. of Forte ’94, pages
324–339. Chapmann & Hall, 1994.
20. ITU-T. Recommendation Z.100: Specification and Description Language (SDL),
International Telecommunication Union (ITU). Geneva, 1999.
21. St. Lau and A. Prinz. BSDL: The Language – Version 0.2. Department of Computer
Science, Humboldt University Berlin, August 1995.
22. A. Mitschele-Thiel. Performance evaluation of SDL systems. In Y. Lahav, A. Wo-
lisz, J. Fischer, and E. Holz, editors, Proc. of the 1st Workshop of the SDL Forum
Society on SDL and MSC (Berlin, June 29 - July 1, 1998), volume 1, pages 35–44,
1998.
23. A. Olsen, O. Færgemand, B. Møller-Pedersen, R. Reed, and J. R. W. Smith. Sy-
stems Engineering Using SDL-92. Elsevier Science B. V., 1994.
24. A. Prinz, R. Gotzhein, U. Glässer, and R. Eschbach. SDL For-
mal Semantics, Z.100.F, ITU-T, International Telecommunication Union.
www.informatik.hu-berlin.de/˜prinz/Semantics/Z100F_Geneva.doc.
 Description and Simulation of Microprocessor
Instruction Sets Using ASMs

Jürgen Teich1 , Philipp W. Kutter2 , and Ralph Weper1

1
Electrical Engineering Department,
University of Paderborn, Germany
{teich, weper}@date.uni-paderborn.de
2
Department of Electrical Engineering,
Swiss Federal Inst. of Tech. Zurich, Switzerland
[email protected]

Abstract. In this paper, we describe how cycle-accurate processor beha-

vior may be efficiently described using Abstract State Machines (ASMs).
Given a register transfer description of the target processor, an extrac-
tion mechanism is described following the approach in [26] that extracts
so called guarded register transfer patterns from the processor descrip-
tion. It will be shown that these may be directly transformed into a set
of ASM rules which in turn provide an executable model of the proces-
sor for simulation purposes. Here, we use the ASM description language
XASM from which the Gem-Mex tool [2] automatically generates a gra-
phical simulator of a given architecture. The feasibility of this approach
is demonstrated for an ARM microprocessor.

1 Motivation
Areas such as digital signal processing and process control have brought up new
kinds of processor architectures which are dedicated to their specific needs and
frequently called ASIPs (application-specific instruction set processors). Exam-
ples are digital signal processors (DSPs) and micro-controllers. Especially in the
area of embedded systems, efficiency of both the architecture and the compiler
for code generation is of utmost importance.
Hence, a new era dealing with the problem of simultaneously developing
processor architecture and compiler has just emerged – architecture and compiler
co-design.
Here, we describe a method that starts with an extraction method following
the approach of Leupers [26] that enables an efficient description and extrac-
tion of the instruction set of a given register transfer description of the target
architecture. The instruction set of a processor may be described by a set of gu-
arded register transfer patterns. Here, these patterns are translated directly into
a set of Gurevich’s ASM [17] (abstract state machine) rules, a model for parallel
computation that has already been widely used for describing the semantics of
as well problem-oriented languages (like C [18], Prolog [11], Oberon [22], Java
[31]) as hardware description languages (e.g., behavioral VHDL [8]). Several case

Y. Gurevich et al. (Eds.): ASM 2000, LNCS 1912, pp. 266–286, 2000.
c Springer-Verlag Berlin Heidelberg 2000

 Description and Simulation of Microprocessor Instruction Sets Using ASMs 267

studies to describe processor architecture hardware (e.g., APE100 [7,6], ARM

microprocessor [19]) have successfully demonstrated the capabilities of ASMs for
modeling the behavior of hardware and show how those specifications may be
refined, e.g., to express pipelining [10,5].
In [12], Del Castillo and Hardt use the specification language ASM-SL for
modeling processor behaviors by ASMs. ASM-SL is the basis of a tool environ-
ment called “The ASM Workbench” which supports syntax- and type-checking
of ASM specifications as well as simulation and debugging of ASMs. Unfortu-
nately, Del Castillo and Hardt do not show how to derive ASM descriptions
automatically. So what existing studies of using ASMs for hardware description
miss (see [9] for a reference bibliography), is a systematics to automatically de-
rive ASM descriptions from known models, such as register transfer descriptions,
e.g., from specifications written in hardware description languages (VHDL [8],
MIMOLA [4], nML [14], etc.).
In the following, we list some significant approaches aiming at code generation
for retargetable processors.

– The RECORD system [26], developed at the University of Dortmund, aims at

automatic code generation for fixed-point DSPs with a fixed instruction word
length. The application program, written in a data flow language (DFL),
is internally represented by a control-/data flow graph. The target proces-
sor has to be specified by the user in the structural hardware description
language MIMOLA [4]. RECORD is a further development of the machine-
independent MSSQ compiler [28,27] which generates microcode based on a
MIMOLA description of the target architecture.
– The hardware description language nML [14] describes a processor in terms
of its instruction set. A machine description is organized as an attribute
grammar with the derivation rules structuring the description and the attri-
butes defining the instructions’ properties.
For nML, a rather simple model of execution is proposed: a running machine
executes a single thread of instructions which are held in a memory and
are addressed via a program counter. The nML language permits concise,
hierarchical processor description in a behavioral style.
– The CBC/SIGH/SIM environment [13] consists of the retargetable code ge-
nerator CBC, which uses a standard code-generator generator for instruction
selection, and the instruction set simulator SIGH/SIM. Target architectures
are described via nML, application programs are given in a high level lan-
guage. The basic idea of SIGH/SIM is to generate a single C++ function
for each instruction of the target processor simulating the impact of the in-
struction’s execution on the processor’s state. However, in case of pipelined
architectures, this leads to enormous costs because the C++ functions have
to be copied for each pipeline stage.
– Another tool using nML is CHESS [16,25], developed at IMEC in Leuven,
Belgium, a retargetable code generation environment for fixed-point DSP
processors. CHESS uses a mixed behavioral/structural processor model sup-
porting load/store architectures and both homogeneous and heterogeneous
268 J. Teich, P.W. Kutter, and R. Weper

register structures. The application is given as a C- or DFL-program and is

represented by a control-/data flow graph. The processor description is re-
presented by a so called instruction set graph containing the register set and
a compact description of the instruction set. CHESS supports automatic bit
alignment and generates machine code to be simulated by the instruction
set simulator CHECKERS [25].
– FLEXWARE (SGS Thompson, Bell Northern Research) [29] uses a mixed
model for behavioral/structural processor description. Target processors are
described by three separate structures: the set of available instruction pat-
terns, a graph model representing the data-path, and a resource classification
that accounts for special-purpose registers. The whole framework consists of
the retargetable code generator CODESYN and the instruction set simulator
INSULIN. CODESYN takes one or more algorithms expressed in a high-level
language and maps them onto a user defined instruction set to produce op-
timized machine code for a target ASIP or a commercial processor core.
INSULIN is based on a reconfigurable VHDL model of a generic instruction
set processor. Due to the use of standard VHDL tools, the advantage of this
approach is not the generation of highly efficient code but the time saving
in processor model development [29].
– CASTLE [24] is an automata-theoretic approach describing the behavior of
the data-path as extended finite state machines which are extracted from a
register transfer level description of the target processor based on a VHDL-
template.

In this paper, we show how a simulator of a processor architecture given

either a netlist or a graphical description of its data-path, may be automatically
generated as follows:

1. Instruction set extraction: Using ideas of the work of Leupers [26], the in-
struction set of a processor may be automatically extracted from a processor
description resulting in a set of so called guarded register transfer patterns
that describe the cycle-accurate behavior of the instruction set architecture.
2. ASM generation: From this set of patterns, we show that an ASM descrip-
tion that also reflects the cycle-accurate behavior of the processor may be
automatically generated. In particular, we present a code translation metho-
dology that generates an ASM description into the language XASM [1].
3. Simulator/Debugger Generation: In the last step, the existing ASM simula-
tion and prototyping environment Gem-Mex [23] is used in order to generate
a simulator/debugger for this ASM. Major advantage of using this metho-
dology is its efficiency in the number of simulation cycles per second that
may be obtained for simulating the execution of a microprocessor.
4. Generic Simulation Core: Another highlight of our approach is that the
functional behavior of instructions of many different number representations
and word-lengths may be considered using an arbitrary precision integer
engine.
 Description and Simulation of Microprocessor Instruction Sets Using ASMs 269

I.(12:5)

I.(4)
RAM
inp
PC +1 I.(20:13)

IM MUX I.(3:2)

I.(20:0) I.(1:0)
ALU

REG

control-path data-path

Fig. 1. A simple processor architecture

5. Library based computing engine: Based on the above simulation core, a li-
brary of generic instructions including arithmetic, transport operations, and
number conversions encountered in almost any microprocessor is supported.

2 Generic Description of Instruction Sets by Register

Transfer Patterns

During the execution of a machine instruction, several register transfers may take
place. In [26], the set of these register transfers is represented by so called regi-
ster transfer patterns (RTP) which are templates of register transfers. A register
transfer (RT) in turn reads input values from registers, memory cells, and/or
input ports, performs a computation, and assigns the result to a destination.
That way, a RTP reflects the capabilities of the data-path of a microprocessor
under question.

Example 1. Fig. 1 shows the architecture of a simple processor consisting of its

control- and data-path. I stands for instruction word, inp denotes an input port,
PC and REG denote (scalar) registers. RAM and IM denote memories where IM
is the instruction memory. The terms in parentheses (hi:lo) denote the respective
bit index subrange of the instruction word I.

Given a processor with a set of registers REG, a set of memories M EM , a

set of input ports PIN , and a set of output ports POU T , we give the following
definitions which are close to Leupers’ [26]:

Definition 1 (RT Expression). A register transfer (RT) expression is either

– a binary constant B ∈ {0, 1}+ ,

270 J. Teich, P.W. Kutter, and R. Weper

– a scalar read access readr(r) or readp(p) with r ∈ REG, p ∈ PIN ,

– an indexed read access readm(m, a) where a is a register transfer expression
and m ∈ M EM ,
– a complex expression op(e1 , · · · , ek ) where op is an operator of arity k ≥ 1
and e1 , . . . , ek are RT expressions, or
– a subrange expression e0 = e.(hi : lo) where e is an RT expression and
(hi : lo) denotes a bit index subrange.

Definition 2 (RT Destination). A register transfer (RT) destination is either

– a scalar write access writer(r) or writep(p) with r ∈ REG, p ∈ POU T , or
– an indexed write access writem(m, a), where a is a register transfer expres-
sion and m ∈ M EM .

Definition 3 (RT Pattern). A register transfer pattern (RTP) is a pair (d, e),
where d is a register transfer destination and e is a register transfer expression.

The set of RTPs of a processor under consideration may be extracted automa-

tically by traversing the signal paths from an arbitrary source to each destination
d of the target architecture which are executable in one machine cycle.
Example 2. In Fig. 2, the data-path of a simple processor is shown. The RTPs
that write R5 ∈ REG are given by the set:
{(writer(R5), +(readr(R1), readr(R3))), (writer(R5), +(readr(R2), readr(R3))),
(writer(R5), +(readr(R1), readr(R4))), (writer(R5), +(readr(R2), readr(R4)))}.
In general, if register files are the source and target of an instruction, the
number of RTPs may be represented by templates, hence, only one RTP is
needed for a class of targets, e.g., a register file.
From a hardware-oriented point of view, register transfers take place during
the execution of each machine cycle. The execution of a particular register trans-
fer, however, depends on the control-path and the machine state. Therefore, each
RTP is coupled with a condition for its execution, a so-called register transfer
condition (RTC).
Typical conditions for the execution on an RT are
– static conditions, e.g., conditions involving condition code, mode registers
M R ⊆ REG, and instruction bits:
The mode register variables are defined as the set of Boolean variables
[
M RV = {Mµ,k | Mµ,k ∈ {0, 1}; k ∈ {0, · · · , width(µ) − 1)}}
µ∈M R

with width(µ) denoting the width of the mode register µ, the index k deno-
ting the k-th bit position of the mode register µ.
Instruction bit variables are denoted by:
IBV = {Ik | Ik ∈ {0, 1}; k ∈ {0, · · · , Lw − 1}}
with Lw denoting the processor’s instruction word length and k indicating
a specified bit position.
 Description and Simulation of Microprocessor Instruction Sets Using ASMs 271

databus

I.(2) I.(3)
D1 D2

R1 R2

0 1
MR1 MUX1 R3 R4

0 1 0 1
I.(0)
MUX2 MUX3 MR2

ADD
I.(1)
R5

Fig. 2. Simple data-path from which RTPs may be extracted. I.(0) − I.(3) denote
corresponding bits of the instruction word I of the control-path

– dynamic (run-time) conditions, e.g., the equality of two register contents. Let
COM P denote the set of all comparisons permissible on a processor, i.e.,
the set of RT expressions of the form op(e, e0 ) where op ∈ {=, 6=, <, >, ≤, ≥}
is a comparison operator. The dynamic control variables are defined as:

DCV = {Dc | Dc ∈ {0, 1}; c ∈ COM P }

Definition 4 (RT Condition). A register transfer condition (RTC) is a Boole-

an function F : {0, 1}K → {0, 1} on the set of Boolean variables IBV ∪ M RV ∪
DCV with K = |IBV ∪ M RV ∪ DCV |.
A register transfer is executed if and only if its RTC evaluates to true for a
given machine state. With the concept of RTCs, we can consider a processor as
a machine which in every control step executes the same set of parallel guarded
operations, each of which has the form

IF <register_transfer_condition> THEN <register_transfer_pattern>

In general, RTCs may be arbitrary combinations of static and dynamic con-

ditions, and therefore may be considerably complex.
Definition 5 (Guarded Register Transfer Pattern (GRTP)). A GRTP
is a pair (F, T ) where F is an RTC and T is an RTP.
272 J. Teich, P.W. Kutter, and R. Weper

Now, instruction set extraction is the task of computing all GRTPs for a
given processor. For a computational procedure for instruction set extraction,
Leupers uses an internal graph data structure that is generated from a MIMOLA
netlist of the processor, see [26] for details. By recursively traversing this graph,
the complete set of RTPs may be extracted efficiently.
GRTPs capture all information required for code generation. RTPs represent
the RPs that a processor can execute, while RTCs represent the required machine
state in terms of partial instructions and mode register states, as well as possible
dynamic conditions.

Example 3. Consider again the simple data-path shown in Fig. 2. I stands for
instruction word. MR1 and MR2 denote mode registers. Suppose we want to
analyze the RTC for RTP ”R5:= R2 + R3”.1 Routing the value of R2 to the left
input of module ADD via multiplexers MUX1, MUX2 can be realized by control
codes MM R1 = 1 and I.(0) = 0. By setting MM R2 = 0, the content of R3 is
switched to the right input of ADD. Furthermore, loading R5 with the result of
the functional unit ADD can be enabled by setting I.(1) = 1. The corresponding
RTC is
F1 = I.(0) · I.(1) · MM R1 · M M R2
There exists also an alternative route for transportation of R2 to ADD: If bus
driver D1 is activated by setting I.(2) = 1, and MUX2 passes its right input
(I.(0)=1), then the value of R2 is routed via databus. A potential bus conflict
needs to be avoided by setting driver D2 to a tristate mode (I.(3) = 0). The
control codes I.(1) = 1 and MM R2 = 0 remain unchanged. The resulting RTC is

F2 = I.(0) · I.(1) · I.(2) · I.(3) · M M R2

Since F1 and F2 are alternative conditions for the same operation, we obtain the
GRTP
((F1 ∨ F2 ), R5 := R2 + R3)

2.1 Modeling of Guarded Register Transfer Patterns by ASM

Update Rules

In mathematics, domains,2 functions, and relations constitute what is called a

structure. Structures without relations are traditionally called algebras. A se-
quential evolving algebra, or now called ASM (abstract state machine) [17], can
be defined by a set of transition rules

IF <Cond> THEN <Updates>

Note that this formula already looks very much like a GRTP in case Cond is
an RTC and Updates is an RTP. In general, Cond may be an arbitrary boolean
1
Short expression for the RTP (writer(R5), +(readr(R2), readr(R3)).
2
in the following, also called universes [17]
 Description and Simulation of Microprocessor Instruction Sets Using ASMs 273

valued expression (first-order logic formula) and Updates is a finite set of updates
to be described later.
The operational semantics of an ASM may be described as follows [17]: The
effect of a transition rule R when applied to an algebra A is to produce another
algebra A0 which differs from A by the new values for those functions at those
arguments where the values are updated by the rule R. If Cond is true, the rule
can be executed by simultaneously executing each update in the set of updates.
In order to understand this, we have to look at the definition of (function)
updates because these are the mechanism by which the dynamics of arbitrary
systems can be explicitly described.

Definition 6 (Function Update). Let f denote an arbitrary n−ary function

and t1 , · · · , tn denote a sequence of parameters. Then, a function update

f (t1 , · · · , tn ) := t

sets the value of f at (t1 , · · · , tn ) to t.

ASMs allow function updates with arbitrary functions f and expressions ti ,t

of any complexity or level of abstraction. Functions whose values can change are
called dynamic in contrast to static functions which do not change.

Example 4. Again, we consider the guarded register transfer pattern (GRTP)

((F1 ∨ F2 ), R5 := R2 + R3)

from Example 3 with

F1 = I.(0) · I.(1) · MM R1 · M M R2

F2 = I.(0) · I.(1) · I.(2) · I.(3) · M M R2

This pattern may be possibly described by the following guarded function update
IF F1 ∨ F2 THEN contents(R5) := contents(R2) + contents(R3)
in which the set of registers may be modeled by a universe REG and the function
contents : REG → W ORD models the RTP for write (read) depending on
whether the expression contents(Ri ) stands on the left hand side or right hand
side of the update. Hence, R5 := R2 + R3 in Example 3 reads the contents of
registers R2 and R3, adds these values, and writes the sum into register R5.

In Example 4, we have seen that a guarded register transfer pattern (GRTP)

naturally corresponds to a guarded ASM update rule.
In terms of ASM rules, a block of update rules is itself a rule. Hence, in a
given state, the updates of all rules whose guards evaluate to true, are executed
simultaneously.
Hence, these semantics naturally describe a single-clock hardware implemen-
tation: In each clock cycle, all conditions (these are register transfer conditions)
are evaluated. Simultaneously, the combinational circuits evaluate the terms they
274 J. Teich, P.W. Kutter, and R. Weper

compute (these are RT expressions). At the end (or beginning) of each clock cy-
cle, the state of the system, as represented by the contents of the registers and
memories, is updated. This is represented by guarded updates in which the fun-
ctions correspond to RT destinations. An ASM description of an instruction set
simultaneously defines an interpreter for the instruction set.

3 Generic Processor Simulation - Library and Simulation

Core

In this section, it is our purpose to present a set of necessary ASM universe

and function definitions that are needed for describing number representations,
interconnections, registers and memories, etc. generically.
For functional simulation, we maintain a library of bit-true arithmetic fun-
ctions that may be used to simulate arbitrary architectures. We define a si-
mulation engine based on arbitrary word-length integer operations such that
for different number systems, word-lengths, and word representations (e.g. little
versus big endian), only two functions that convert a number into an integer,
and, after simulating an operation, back to its number representation, have to
be written once.

3.1 Typing and Subtyping

Hardware typically operates on bundles (vectors) of wires, also termed bit-vectors

or bit-strings in hardware description languages such as VHDL. Therefore, it is
necessary to define a simple universe called BIT ∈ {0, 1} where 0 and 1 denote
the numeric values that we usually associate with the Boolean values f alse and
true, respectively.
With WORD(n) we denote the universe of all {0,1}-bit-vectors of length n.

W ORD(n) = {w | w ∈ {0, 1}n ; n ∈ IN}

The same notation is used for registers from the universe REG and memories
from the universe M EM in case they have to be distinguished by different
word-lengths. For example, let REG(16) denote the subset of registers having a
word-length of 16 bits.3
For the reason we would like to present processor architectures independent
from the number representation used (e.g., signed, one’s or two’s complement,
etc.), we introduce a function4

val : W ORD(t) → IN T EGER

3
In case all word-lengths are equal for all memories and registers, we will omit the
index n for short.
4
However, we will restrict ourselves in this paper to fixed-point arithmetic.
 Description and Simulation of Microprocessor Instruction Sets Using ASMs 275

that returns the value of a word in the corresponding number system and its
inverse5
val−1 : IN T EGER → W ORD(t)

3.2 Modeling Combinational Modules

A combinational module with n Inputs i1 , · · · , in and m Outputs o1 , · · · , om may

be described by a (vector-) function

F : i1 × · · · × in → o1 × · · · × om

Here, i1 , · · · , in are each inputs of a subset of W ORD, and o1 , · · · , om are each

outputs of a subset of W ORD. In a given number system, the value val(oj ) of
each output oj , j = 1, · · · , m may be given in terms of the values of the inputs
of F by equations of the form:

val(oj ) = Fj (val(i1 ), val(i2 ), · · · , val(in )).

3.3 Modeling Registers and Memories

The model for registers (universe REG) and memories (universe M EM ) has
already been shown. The contents of registers are modeled by unary functions

contents : REG(n) → W ORD(n)

for each word-length n ∈ IN T EGER. For memories, the indexed access func-
tions
contents : M EM (n) × IN DEX → W ORD(n)
are defined for each memory m ∈ M EM (n) where IN DEX is defined over the
universe {0, · · · , size(m) − 1} and the function

size : M EM (n) → IN T EGER

denoting the size of memory m ∈ M EM (n) given in terms of the number of

words of type W ORD(n).

3.4 Modeling Interconnections

It remains to show how to model interconnections, how to extract subwords

out of words, how to merge words to larger words, and how to copy bundles of
signals.
5
For non-redundant number systems (e.g., two’s complement), the representation of
a value value ∈ IN T EGER by a bit-vector of length n is unique. In case the
conflict of number representation for the number 0 can be uniquely resolved, also
non-redundant number systems such as one’s complement and sign-magnitude re-
presentations may be dealt with.
276 J. Teich, P.W. Kutter, and R. Weper

Given an element w ∈ W ORD(n) the function bit of the form

bit : W ORD(n) × IN T EGER → BIT

returns the ith bit b of w — or b = bit(w, i) (or simply b = w.(i)) — such

that val(b) = val(bit(w, i)) in case 0 ≤ i < n, and is undefined else.
In order to compose arbitrary vectors of bits, we introduce a function to
model the extraction of subwords from a word, and thus realize the expression
e0 = e.(hi : lo)

extract : W ORD(n) × IN T EGER × IN T EGER → W ORD(m)

where
e0 = extract(e, hi, lo)
returns a word e0 of word-length m = hi − lo + 1 in case lo ≤ hi < n that satisfies
∀i ∈ {0, hi − lo} : val(e0 .(i)) = val(e.(i + lo)). Else it is undefined.
Moreover, the function

merge : W ORD(n) × W ORD(m) → W ORD(n + m)

or c = merge(a, b) allows to merge two words a ∈ W ORD(n) and b ∈ W ORD(m)

to form a composed word c ∈ W ORD(n + m) that satisfies ∀i ∈ {0, · · · , n − 1} :
val(c.(i)) = val(a.(i)) and ∀i ∈ {n, · · · , n + m − 1} : val(c.(i)) = val(b.(i − n)).
Finally, we define the function

split : W ORD(n) → W ORD(n)

that allows to copy signals: b = split(a) with a ∈ W ORD(n) and b ∈ W ORD(n)

∀i ∈ {0, · · · , n − 1} : val(b.(i)) = val(a.(i)).

Example 5. Lets consider Fig. 3 where a bundle of wires is grouped from a

signals b ∈ W ORD(6), a ∈ W ORD(8), and 5 leading zeros into a 16 bit signal
c ∈ W ORD(16). Using merge and extract, this grouping may be functionally
described as follows:

c = merge(merge(extract(b, 5, 3), extract(a, 0, 0)), merge(extract(a, 7, 1), zero))

3.5 Retargetability to Multiple Number Systems

In the following, we assume that BIT = W ORD(1) ∈ {0, 1} is modeling the
Boolean values of f alse and true, respectively.
A bit-vector x ∈ W ORD(n) represents a binary vector of length n. This is
an ordered sequence of the form (xn−1 xn−2 · · · x1 x0 ) where the xi = bit(x, i)
are the bits (also called digits). We have introduced the function val returning
the value of a bit-string. Intentionally, this has been done independently from
the number system used, e.g., non-redundant, weighted, and positional number
systems. Our intention is to simulate processor architectures bit-true at a high
 Description and Simulation of Microprocessor Instruction Sets Using ASMs 277

a b
5 8 6
zero

3
12 4

16

c
Fig. 3. Example of extract and merge of signals

level (i.e., RT-level) of abstraction, fast and independent from the number system
of the architectures’ arithmetic-logical units.
Therefore, we implement a simulation engine that computes functions based
on (arbitrary long) integer representations for implementation of highly complex
arithmetic functions, and use the two functions

val : W ORD(n) → IN T EGER and val−1 : IN T EGER → W ORD(n)

to convert numbers from different number systems. Only these two functions
have to be implemented to retarget the simulator to a different arithmetic num-
ber system, see Fig. 4.6 The functions val, respectively val−1 are the interfaces
of the integer engine to a processor with a different number representation and
word-length. Note that in some cases, it may be more convenient to perform ope-
rations on the values of extracted subwords (e.g. single bits) of operands using
extract and not on the values of the complete operands.
In [30], details are given for the definition of the functions val and val−1 for
typical number systems used in computer systems.
In the following example, it is shown how a typical operation of two operands
is simulated generically using the above retargetable simulation engine.

Example 6. Let c = op(a, b) with a ∈ W ORD(n), b ∈ W ORD(m), and c ∈

W ORD(q) denote an arbitrary operation with two operands a and b. The fun-
ction op is evaluated as follows:

1. compute val(a) based on its word-length n;

2. compute val(b) based on its word-length m;
3. calculate val(c0 ) in terms of val(a) and val(b) according to op;
6
Note that the view shown is completely opposite to the view one has when simulating
hardware. The user (environment) thinks in terms of values (e.g., integers) whereas
internally, the hardware realizes operations on bit-vectors in a binary number system.
278 J. Teich, P.W. Kutter, and R. Weper

x (WORD(n)) y (WORD(m))

x (WORD(n)) y (WORD(m)) val val

val(x) val(y)

integer
engine
F F
int

val(z)

z (WORD(q)) -1
val

z (WORD(q))

Fig. 4. Concept of a retargetable simulation engine

4. let val(c0 ) := val(c0 ) mod 2q (truncation to result word-length default);7

5. let c = val−1 (val(c0 )).
In case n 6= m, the above extension of the number representation might be
necessary in hardware, however not in our simulation engine as the function op
is generic. Such an operation is typical for arithmetic computation.
In [30], we present a catalogue of standard computations (e.g. logical and
standard arithmetic operations) that have been implemented in the library.

4 An Environment for Processor Simulation and

Debugging Using XASM
In this section, we show how the exemplary description of the ARM [15] micro-
processor is derived using the above methodology, how the instruction set may
be simulated, and how assembler programs of this processor may be debugged.
The motivation of the following ARM-processor case study is a hand written
ASM-description as given in [19,20]. In our methodology, a corresponding ASM-
description will be obtained by instruction set extraction and ASM generation
from a netlist or graphical description of the ARM’s data-path.
In Section 4.1 we sketch the architecture of the ARM processor and show
how its description is implemented using the ASM language XASM.
In addition to entering the ASM rules, we had to implement the library
of arithmetic and logical functions described in Section 3. These functions are
implemented in C and interfaced to XASM. Further, a parser both for assembler
7
Different other behaviors like rounding etc. may be handled similarly.
 Description and Simulation of Microprocessor Instruction Sets Using ASMs 279

A[31:0]
control
address register

PC
incrementer

PC
register
bank instruction

R14
R15
decode
R0
ALU Bus

&
multiply control
A Bus

register

barrel
shifter
B Bus

ALU

data out register data in register

D[31:0]

Fig. 5. Architecture of the ARM processor

code and binary code was written using the existing parser support of XASM.
Some details are given in Section 4.2. Given the XASM specification, the Gem-
Mex tool [2] generates a graphical debugger and simulation tool, see Section 4.3.

4.1 The ARM2 Processor

The ARM2 Processor of Advanced Risc Machines [15] is a 32 bit microprocessor

architecture with a set of 16 registers, see Figure 5. The dedicated register R15
contains the program counter (PC) and status flags, R14 is reserved for the re-
turn address of branch instructions. The registers R0 to R13 are general purpose
registers.
The ARM processor supports a three staged pipeline. Instructions are dispat-
ched in three phases: During the first phase instruction fetch, an instruction is
fetched from the instruction memory, in the second phase decode, the instruc-
280 J. Teich, P.W. Kutter, and R. Weper

tion is decoded and assigned to the respective execution unit. During the phase
execute, the instruction will be executed and the result will be stored.
One can think of these phases as being executed sequentially, but in fact the
architecture has parallel logic for all three phases, so in a stable pipeline state,
the ARM processor is able to execute three operations of a machine program in
parallel. In [19], a simple ASM-model of the ARM microprocessor for sequential
execution mode is given and stepwise refined to a model executing the stages in
parallel. The correctness of the refinements is proven.
We started by entering the specifications of the different refinements, and
debugged them. The disadvantage of considering several refinements simulta-
neously is that it is almost impossible to maintain all versions if changes must
be made. Thus, we decided to integrate the different models into one model that
can be transformed into any of the refinements by exchanging some macro defi-
nitions. This version of the description is given in [21]. Depending on the boolean
value of the macro PIPELINED, either the stages are executed sequentially or in
pipelined mode. In sequential mode, a nullary function
function Stage
is used. The values of Stage are fetch, decode, or execute. In XASM syntax, the
three constants can be introduced by enumeration:
universe STAGES = {fetch, decode, execute}
The control structure simulating both the sequential and the parallel execution of
stages is structured as follows. Assume, code[Contents(PC)] fetches the current
instruction from memory.
if PIPELINED or Stage = fetch then
Instr := code[Contents(PC)]
if PIPELINED then
DecodeInstr := Instr
else
Stage := decode
endif
R_FETCH
endif
if PIPELINED or Stage = decode then
if PIPELINED then
ExecuteInstr := DecodeInstr
else
Stage := execute
endif
R_DECODE
endif
if PIPELINED or Stage = execute then
if PIPELINED then
else
 Description and Simulation of Microprocessor Instruction Sets Using ASMs 281

Stage := fetch
endif
R_EXECUTE
endif

In pipelined mode, all three parts of the rule are executed in parallel, otherwise
the parts are executed sequentially in cyclic mode.
In the three remaining rules R FETCH, R DECODE, and R EXECUTE of the com-
plete ARM description not given here, the active instructions must be accessed
by three macros defined as follows:

#define FETCHINSTR Instr

#define DECODEINSTR (if PIPELINED then DecodeInstr else Instr)
#define EXECUTEINSTR (if PIPELINED then ExecuteInstr else Instr)

If these macros are used to refer to the instruction, almost the complete functio-
nality of the rules R FETCH, R DECODE, and R EXECUTE can be defined independent
of whether the stages are executed in parallel or sequentially.
In order to make the description retargetable to other architectures, we used
the ASM-library of Section 3. The adaption to our methodology involved mainly
the adaption of the signature. The structure of the rules remained unchanged.

4.2 Parser Generation for Assembler and Binary Code

XASM provides a number of built in constructs for generating a parser. The

main structure of the assembler parser starts with the declaration of a list of
instructions, separated by semicolons.

nonterm ARMcode[Instruction];
base Instruction
cont ARMcode ‘‘;’’ Instruction
endnonterm

Instructions are alternatively alu-, branch-, single transfer-, or multiple transfer

instructions. In XASM this alternative is written as

nonterm Instruction = AluInstr | BranchInstr |

SingleTransferInstr | MultipleTransferInstr
endnonterm

Finally, the syntax of single instructions is given, for instance for a branch:

nonterm BranchInstr ::= ‘‘B’’ (link=)b_link (cond=)Conds

(adr=)address;
R
endnonterm
282 J. Teich, P.W. Kutter, and R. Weper

A[31:0]
control
ARM.xasm address register
testprg.ass

PC
incrementer

PC
register
bank instruction
decode
&
multiply control

A Bus
register

barrel
shifter
Architecture Instruction

B Bus
Description Description

XASM Assembler
Retargetable code code
Simulation instr.c
Engine
C code
Gem-Mex
Function
Library
C code Tcl/Tk

Debugging
and
Simulation
Environment

Fig. 6. Designflow from architecture description and program input to simulation

where in the rule R one can access the components of the branch instruction
as link, cond, and adr. In the rule R certain informations about the instruction
are precomputed.
If a further analysis of the code is desired, e.g. an abstract verification of
certain properties, the Montages [3] support of Gem-Mex can be used. We re-
stricted the use of Gem-Mex to the generation of graphics for simulation and
debugging.

4.3 Simulation of the ARM Processor

In Figure 6, we display the designflow from architecture description and pro-

gram input to simulation. From a graphical description of the architecture, the
GRTP description is extracted and then transformed into ASM rules. The upper
left part shows the transformation from the architecture description into XASM
code resulting in a file ARM.xasm. The upper right part displays the input of
the simulator’s algorithmic load which consists of an application program writ-
ten in assembler notation, the file testprg.ass. The XASM-compiler translates
 Description and Simulation of Microprocessor Instruction Sets Using ASMs 283

Fig. 7. Debugging of a program using the Gem-Mex Environment

the architecture description into C-code and calls the gcc-compiler to compile
these C-files. Thereby, the C-library of arithmetic functions described in Section
3 is linked. As a result, an executable is generated and which interacts with
generic Tcl/Tk scripts provided by the Gem-Mex simulation and debugging en-
vironment. The visual aspects of the simulation and debugging environment are
shown at the bottom of the figure.
Figure 7 displays a snapshot of a simulation run both of the pipelined (top)
and the sequential (bottom) implementation. The algorithmic load respectively
consists of a small assembler program, see Figure 8.
Here, after nine steps, the sequential implementation has reached the end of
an execution phase and sets the Stage flag to fetch indicating that the next
operation has to be a fetch operation. The parallel implementation only needs
three cycles to reach the same machine state due to pipelining. The right hand
picture shows the contents of the respective register file.

5 Conclusions
In this paper, we present foundations for the simulation of a processor archi-
tecture and its instruction set given the syntax of the instruction and either a
netlist or graphical description of the architecture’s data-path.
A major advantage of our approach with respect to other approaches is that
the ASM-formalism itself is suited for architecture description. The generated
ASMs can thus serve as additional documentation. Changes may be made di-
rectly at the ASM level and the corresponding changes in the netlist/graphical
description can be regenerated.
284 J. Teich, P.W. Kutter, and R. Weper

Fig. 8. The application program in ARM assembler language

Furthermore, we present a methodology for generic processor simulation that

is independent of a processor’s specific parameters like instruction word length
and number representation. In order to accomplish this, we build a library of
bit-true arithmetic functions and construct an integer engine as a simulation
core which can be adapted to almost any microprocessor by specific interface
functions.
Here, we demonstrate the feasibility of our methodology for simulating a
realistic pipelined RISC architecture, the ARM microprocessor.

References
1. M. Anlauff. Xasm - an extensible, component-based abstract state machines lan-
guage. In This Volume.
2. M. Anlauff, P.W. Kutter, and A. Pierantonio. Formal aspects of and development
environments for Montages. In Second Int. Workshop on the Theory and Practice
of Algebraic Specifications, Workshops in Computing. Springer-Verlag, 1997.
3. M. Anlauff, P.W. Kutter, and A. Pierantonio. Enhanced control flow graphs in
Montages. In Perspective of System Informatics, number 1726 in LNCS, 1999.
4. S. Bashford, U. Bieker, B. Harking, R. Leupers, P. Marwedel, A. Neumann, and
D. Voggenauer. The MIMOLA language - version 4.1. Technical report, University
of Dortmund, 1994.
5. E. Börger. Why use evolving algebras for hardware and software engineering? In
M. Bartosek, J. Staudek, and J. Wiederman, editors, SOFSEM’95, 22nd Seminar
on Current Trends in Theory and Practice of Informatics, volume 1012 of Springer
Lecture Notes on Computer Science (LNCS), pages 236–271, 1995.
 Description and Simulation of Microprocessor Instruction Sets Using ASMs 285

6. E. Börger and G. Del Castillo. A formal method for provably correct composition
of a real-life processor out of basic components (the APE100 reverse engineering
study). In B. Werner, editor, Proc. of the First Int. Conf. on Engineering of
Complex Computer Systems (ICECCS’95), pages 145–148, November 1995.
7. E. Börger, G. Del Castillo, P. Glavan, and D. Rosenzweig. Towards a mathema-
tical specification of the APE100 architecture: the APESE model. In B. Pehrson
and I. Simon, editors, IFIP 13th World Computer Congress, volume I : Techno-
logy/Foundations, pages 396–401, Elsevier, Amsterdam, The Netherlands, 1994.
8. E. Börger, U. Glässer, and W. Müller. The semantics of behavioral VHDL’93
descriptions. In European Design Automation Conference (EURO-DAC’94), pages
500–505, IEEE CS Press, Los Alamitos, CA, U.S.A., 1994.
9. E. Börger and J. K. Huggins. Abstract state machines 1988-1998: Commented
ASM bibliography. In H. Ehrig, editor, Formal Specification Column, Bulletin of
the EATCS64. February 1998.
10. E. Börger and S. Mazzanti. A practical method for rigorously controllable hardware
design. In J. P. Bowen, M. B. Hinchey, and D. Till, editors, ZUM’97: The Z Formal
Specification Notation, volume 1212 of Springer Lecture Notes on Computer Science
(LNCS), pages 151–187, 1996.
11. E. Börger and D. Rosenzweig. A mathematical definition of full Prolog. Science
of Computer Programming, 24:249–286, 1995. North-Holland.
12. G. Del Castillo and W. Hardt. Fast dynamic analysis of complex hw/sw-systems
based on abstract state machine models. In Proc. of the 6th Int. Workshop on
Hardware/Software Co-Design (Code/Cashe98), Seattle, pages 77–81, March 1998.
13. A. Fauth. Beyond tool-specific machine descriptions. In P. Marwedel and G. Goos-
sens, editors, Code Generation for Embedded Processors, pages 138–152. Kluwer
Academic Press, 1995.
14. A. Fauth, J. Van Praet, and M. Freericks. Describing instruction set processors
using nML. In Proceedings on the European Design and Test Conference, Paris,
France, pages 503–507, March 1995.
15. S . Furber. ARM System Architecture. Addison-Wesley Longman, 1996.
16. G. Goossens, J. Van Praet, D. Lanneer, W. Geurts, and F. Thoen. Programma-
ble chips in consumer electronics and telecommunications. In G. de Micheli and
M. Sami, editors, Hardware/Software Co-Design, volume 310 of NATO ASI Series
E: Applied Sciences, pages 135–164. Kluwer Academic Press, 1996.
17. Y. Gurevich. Evolving algebras 1993: Lipari guide. In E. Börger, editor, Specifica-
tion and Validation Methods, pages 9–36, Oxford University Press, 1995.
18. Y. Gurevich and J. Huggins. The semantics of the C programming language.
In E. Börger, H. Kleine Büning, G. Jäger, S. Martini, and M. M. Richter, edi-
tors, Computer Science Logic, volume 702 of Springer Lecture Notes on Computer
Science (LNCS), pages 274–309, 1993.
19. J. Huggins and D. Van Campenhout. Specification and verification of pipelining in
the ARM2 RISC microprocessor. Technical report, CSE-TR-321-96, EECS Dept.,
Univ. of Michigan, 1996.
20. J. Huggins and D. Van Campenhout. Specification and verification of pipelining
in the ARM2 RISC microprocessor. ACM Transactions on Design Automation of
Electronic Systems, 3(4):563–580, 1998.
21. S. Küng. Simulation mit Abstrakten Zustandsmaschinen, Studienarbeit, 1998.
22. P.W. Kutter and A. Pierantonio. The formal specification of Oberon. Journal of
Universal Computer Science, 3(5):443–503, 1997.
23. P.W. Kutter and A. Pierantonio. Montages: Specification of realistic programming
languages. Journal of Universal Computer Science, 3(5):416–442, 1997.
286 J. Teich, P.W. Kutter, and R. Weper

24. M. Langevin, E. Cerny, J. Wilberg, and H.-T. Vierhaus. Local microcode genera-
tion in system design. In P. Marwedel and G. Goossens, editors, Code Generation
for Embedded Processors, pages 171–187. Kluwer Academic Press, 1995.
25. D. Lanneer, J. Van Praet, A. Kifli, K. Schoofs, W. Geurts, F. Thoen, and G. Goos-
sens. Chess: Retargetable code generation for embedded dsp processors. In P. Mar-
wedel and G. Goossens, editors, Code Generation for Embedded Processors, pages
85–102. Kluwer Academic Press, 1995.
26. R . Leupers. Retargetable Code Generation for Digital Signal Processors. Kluwer
Academic Publishers, Dordrecht, The Netherlands, 1997.
27. R. Leupers and P. Marwedel. Retargetable code compilation based on structu-
ral processor descriptions. Design Automation for Embedded Systems, 3(1):1–36,
January 1998.
28. L. Nowak. Graph based retargetable microcode compilation in the MIMOLA design
system. In 20th Annual Workshop on Microprogramming (MICRO-20), pages 126–
132, 1987.
29. P. Paulin, C. Liem, T. May, and S. Sutarwala. Flexware: A flexible firmware
development environment for embedded systems. In P. Marwedel and G. Goossens,
editors, Code Generation for Embedded Processors, pages 67–84. Kluwer Academic
Press, 1995.
30. J. Teich, P.W. Kutter, and R. Weper. A retargatable engine for the simulation of
microprocessors using asms; DATE-report 04/00. 2000.
31. C. Wallace. The Semantics of the Java Programming Language: Preliminary Ver-
sion. Technical Report CSE-TR-355-97, EECS Dept., University of Michigan, De-
cember 1997.
 Symbolic Analysis of Transition Systems?

Natarajan Shankar

Computer Science Laboratory

SRI International
Menlo Park CA 94025 USA
{shankar, owre}@csl.sri.com
URL: http://www.csl.sri.com/˜shankar/
Phone: +1 (650) 859-5272 Fax: +1 (650) 859-2844

Abstract. We give a brief overview of the Symbolic Analysis Laboratory

(SAL) project. SAL is a verification framework that is directed at ana-
lyzing properties of transition systems by combining tools for program
analysis, model checking, and theorem proving. SAL is built around a
small intermediate language that serves as a semantic representation for
transition systems that can be used to drive the various analysis tools.

The transition system model of a program consists of a state type, an in-

itialization predicate on this state type, and a binary next-state relation. The
execution of a program starts in a state satisfying the initialization predicate so
that each state and its successor state satisfy the next-state relation. Transition
systems are a simple low-level model that have none of the semantic complica-
tions of high-level programming languages. Constructs such as branches, loops,
and procedure calls can be modelled within a transition system through the use
of explicit control variables. The transition system model forms the basis of se-
veral formalisms for several popular formalisms including UNITY [15], TLA [28],
SPL [31], and ASMs [20]. It also underlies verification tools such as SMV [30],
Murphi [18], and STeP [32].
If we focus our attention on the verification of properties of transition sy-
stems, we find that even this simple model poses some serious challenges. The
verification of transition systems is performed by showing that the system sa-
tisfies an invariance or progress property, or that it refines another transition
system. It is easy to write out proof rules for the verification of such properties
but the actual application of these proof rules requires considerable human in-
genuity. For example, the verification of invariance properties requires that the
invariant be inductive, i.e., preserved by each transition. A valid invariant might
need to be strengthened before it can be shown to be inductive. Fairness con-
straints and progress measures have to be employed for demonstrating progress
?
This work was funded by the Defence Advanced Research Projects Agency under
Contract NO. F30603-96-C-0204, and NSF Grants No. CCR-9712383 and CCR-
9509931. The SAL project is a combined effort between SRI International, Stanford
University, and the University of California, Berkeley.

Y. Gurevich et al. (Eds.): ASM 2000, LNCS 1912, pp. 287–302, 2000.
c Springer-Verlag Berlin Heidelberg 2000

288 N. Shankar

properties. It takes a fair amount of effort and ingenuity to come up with suitable
invariant strengthenings and progress measures.
Methods like model checking [13] that are based on state-space exploration
have the advantage that they are largely automatic and seldom require the fine-
grain interaction seen with deductive methods. Since these methods typically
explore the reachable state space (i.e., the strongest invariant), there is no need
for invariant strengthening. Progress measures are also irrelevant since the size of
the whole state space is bounded. However, model checking methods apply only
to a limited class of systems that possess small, essentially finite state spaces.
Theorem proving or model checking are not by themselves adequate for effec-
tive verification. It is necessary to combine the expressiveness of the deductive
methods with the automation given by model checking. This way, small, finite-
state systems can be directly verified using model checking. For larger, possi-
bly infinite-state systems, theorem proving can be used to construct property-
preserving abstractions over a smaller state space. Such abstractions convert
data-specific characteristics of a computation into control-specific ones. The
finite-state model constructed by means of abstraction can be analyzed using
model checking. It is easy to actually compute the properties of a system from a
finite-state approximation and map these properties back to the original system.
We give an overview of an ongoing effort aimed at constructing a general
framework for the integration of theorem proving, model checking, and program
analysis. We use the term symbolic analysis to refer to the integration of these
analysis techniques since they all employ representations based on symbolic logic
to carry out a symbolic interpretation of program behavior. The framework also
emphasizes analysis, i.e., the extraction of a large number of useful properties,
over correctness which is the demonstration of a small number of important
properties. The framework is called the Symbolic Analysis Laboratory (SAL).
We motivate the need for symbolic analysis and describe the architecture and
intermediate language of SAL.

1 A Motivating Example

We use a very simple and artificial example to illustrate how symbolic analysis
can bring about a synergistic combination of theorem proving, model checking,
and program analysis. The example consists of a transition system with a state
contain a (control) variable PC ranging over the scalar type {inc, dec}, and two
integer variables B and C. Initially, control is in state inc and the variables B
and C are set to zero. There are three transition rules shown below as guarded
commands:

1. When PC = inc, then B is incremented by two, C is set to zero, and control

is transferred to state dec.

PC = inc −→ B 0 = B + 2; C 0 = 0; PC 0 = dec;
 Symbolic Analysis of Transition Systems 289

B:=B+2; C:= 0;
B:=0;
C:=0;

inc dec

B>0 -> B:= B-2; B>0 -> B:= B-2;

C:=C+1; C:=C+1;

Fig. 1. A Simple Transition System: Twos

2. When PC = dec, B is decremented by two, and C is incremented by one,

and control is transferred to state decl .

PC = dec ∧ B > 0 −→ B 0 = B − 2; C 0 = C + 1; PC 0 = inc;

3. Same as transition rule 2, but control stays in dec.

PC = dec ∧ B > 0 −→ B 0 = B − 2; C 0 = C + 1;

There is also an implicit stuttering transition from state dec to itself when none
of the guards of the other transitions holds, i.e., when B ≤ 0. Since the inc state
has a transition with a guard that is always true, there is no need for a stuttering
transition on inc. The transition system is shown diagrammatically in Figure 1.

The transition system Twos satisfies a number of interesting invariants.

1. B is always an even number.
2. B and C are always non-negative.
3. B is always either 0 or 2.
4. B is always 0 in state inc.
5. C is always either 0 or 1.
6. In state dec, B = 2 iff C = 0.
The purpose of symbolic analysis is to find and validate such properties with
a high degree of automation and minimal human guidance and intervention.
While efficient automation is essential for analyzing large transition systems,
the intended outcome of symbolic analysis is human insight. The analysis should
therefore not rule out human interaction.

2 Some Symbolic Analysis Techniques

We enumerate some symbolic analysis techniques and assess their utility on the
Twos example. For this purpose, we focus on the invariant (1) below.

B =0∨B =2 (1)
290 N. Shankar

Note that the transition system Twos is a potentially infinite state system since
variables B and C range over the integers.
Some mathematical preliminaries are in order. A transition system P is given
by a pair hIP , NP i consisting of an initialization predicate on states IP , and a
binary next-state relation on states NP . We constrain the next-state relation N
to be total so that ∀s : ∃s0 : N (s, s0 ). The metavariables s, s0 range over states.
We treat a set of states as equivalent to its characteristic predicate. The boolean
connectives ∧, ∨, ⊃, are lifted from the booleans to the level of predicates and
correspond to the set-theoretic operations ∩, ∪, and ⊆, respectively. An assertion
is a predicate on states. The metavariables φ, ψ range over assertions. A predi-
cate transformer is a map from predicates to predicates. A monotone predicate
transformer τ preserves the subset or implication ordering on predicates so that
if φ ⊃ ψ, then τ (φ) ⊃ τ (ψ). The fixed point of a monotone predicate transformer
τ is an assertion φ such that φ ≡ τ (φ). As a consequence of the Tarski–Knaster
theorem, every monotone predicate transformer has a least fixed point lf p(τ )
and a greatest fixed point gf p(τ ) such that

lf p(τ ) = τ (lf p(τ )) ⊃ gf p(τ ) = τ (gf p(τ )).

Let ⊥ represent the empty set of states, > the set of all states, and ω the set
of natural numbers. If the state space is finite, then the least fixed point lf p(τ )
can be calculated as

⊥ ∨ τ (⊥) ∨ τ 2 (⊥) ∨ . . . ∨ τ n (⊥)

for some n, and similarly, gf p(τ ) can be calculated as

> ∧ τ (>) ∧ τ 2 (>) ∧ . . . ∧ τ n (>),

for some n. W W
If τ is ∨-continuous (i.e., τ ( i∈ω φi ) = i∈ω τ (φi ) for φi such that whenever
i < j, φi ⊃ φj ), then _
lf p(τ ) = τ i (⊥) (2)
i∈ω
V V
Similarly, if τ is ∧-continuous (i.e., τ ( i∈ω φi ) = i∈ω τ (φi ) for φi such that
whenever i < j, φj ⊃ φi ), then
^
gf p(τ ) = τ i (>) (3)
i∈ω

Equations (2) and (3) provide an iterative way of computing the least and grea-
test fixed points but these on infinite-state spaces, the computations might not
converge in a bounded number of steps.
Typical examples of monotone predicate transformers include
1. Strongest postcondition of a transition relation N , sp(N ), which is defined
as
sp(N )(φ) ≡ (∃s0 : φ(s0 ) ∧ N (s0 , s)).
 Symbolic Analysis of Transition Systems 291

2. Strongest postcondition of a transition system P , sp(P ) is defined as

sp(P )(φ) ≡ IP ∨ sp(NP )(φ).

3. Weakest precondition of a transition relation N , wp(N ) is defined as

wp(N )(φ) ≡ (∀s0 : N (s, s0 ) ⊃ φ(s)).

2.1 Invariant Proving

The invariance rule is the most heavily used proof rule in any program logic [23,
33]. Given a transition system P as a pair hIP , NP i, consisting of an initialization
IP and a next-state relation NP , the invariance rule usually has the form:

` ψ(s) ⊃ φ(s)
` IP (s) ⊃ ψ(s)
` ψ(s0 ) ∧ NP (s0 , s1 ) ⊃ ψ(s1 )

P |= invariant φ

In this rule, the assertion ψ is a strengthening of the assertion φ. Such a

strengthening is needed since the assertion φ may not be inductive, i.e., satisfy
the premises ` IP (s) ⊃ φ(s) and ` φ(s0 ) ∧ NP (s0 , s1 ) ⊃ φ(s1 ).
In the Twos example, the invariant (1) is not inductive. It fails because it is
not preserved by transition 1 since we cannot establish

` (PC = inc ∧ (B = 0 ∨ B = 2))

∧ (PC = inc ∧ B 0 = B + 2 ∧ C 0 = 0 ∧ PC 0 = dec)
⊃ (B 0 = 0 ∨ B 0 = 2).

The invariant has to be strengthened with the observation that when PC = inc,
B is always 0 so that it now reads

B = 0 ∨ (PC 6= inc ∧ B = 2). (4)

The strengthened invariant (4) is inductive. The need for invariant strengt-
hening in program proofs is the key disadvantage of the deductive methods with
respect to model checking. Quite a lot of effort is needed to turn a putative
invariant into an inductive one. Once an invariant has been strengthened in this
manner, it can contain a large number of conjuncts that generate a case explosion
in the proof. Much of the focus of symbolic analysis is on supplementing deduc-
tive verification with the means of automatically obtaining useful invariants and
invariant strengthenings.

2.2 Enumerative Model Checking

The early approaches to model checking were based on the feasibility of com-
puting fixed point properties for finite-state systems. The reachable states of a
292 N. Shankar

finite-states can be computed by starting from the set of initial states and explo-
ring the states reachable in n consecutive transitions. Any property that holds
on all the reachable states is a valid invariant. There are many variations on
this basic theme. Many modern enumerative model checkers such as Murphi [18]
and SPIN [24] carry out a depth-first search exploration of the transition graph
while maintaing a hash-table to record states that have already been visited. In
SPIN, the LTL model checking problem is transformed into one of emptiness for
ω-automata, i.e., automata that recognize infinite strings [38,19].
In enumerative model checking, properties written in a branching-time tem-
poral logic CTL can be verified in time proportional to N × F where N is
the size of the transition graph and F the size of the temporal formula. Model
checking linear-time temporal logic formulas is more expensive and takes time
proportional to N × 2F where N is the size of the model and F is of the formula.
The Twos example succumbs rather fortuitously to enumerative model
checking. Even though the potential state space of Twos is unbounded, only
a bounded part of the state space is reachable since B is either 0 or 2, and C is
either 0 or 1. The success of enumerative model checking is somewhat anomalous
since this method is unlikely to terminate on typical infinite-state systems. Even
on finite-state systems, an enumerative check is unlikely to succeed because the
size of the searchable state space can be exponential in the size of the program
state. Still, enumerative model checking is an effective debugging technique that
can often detect and display simple counterexamples when a property fails.

2.3 Symbolic Model Checking

The use of symbolic representation for the state sets was proposed in order
to combat the state explosion problem in enumerative model checking [4,30]. A
symbolic representation for boolean functions based on binary decision diagrams
(BDDs) [10] has proved particularly successful. A finite state can be represented
as a bit-vector. Then sets of bit-vectors are just boolean functions and can be
represented as BDDs. In particular, the initial set, a given invariant claim, the
transition relation, and the reachable state set, can all be represented as BDDs.
The BDD operations can be used to compute images of state sets with respect to
the transition relation. This allows predicate transformers such as strongest post-
condition and weakest precondition to be applied to the BDD representation of a
state set. The reachable state set can be computed by means of a fixed point ite-
ration of the strongest postcondition computation starting from the initial state
set. Every intermediate iteration of the reachable state set is also represented
as a BDD. There are several advantages to the use of BDDs. Sometimes even
sets of large cardinality might have compact symbolic representations. BDDs
are a canonical representation for boolean functions so that equivalence tests
are cheap. BDDs are especially good at handling the boolean quantification that
is needed in the image computations. Automata-theoretic methods can also be
represented in symbolic form. Some symbolic model checkers include SMV [30]
 Symbolic Analysis of Transition Systems 293

Such symbolic representations do require the state to be explicitly finite. This

means that the Twos example cannot be coded directly in a form that can be
directly understood by a symbolic model checker. Some work has to be done in
order to reduce the problem to finite-state form so that it can be handled by a
symbolic model checker.

2.4 Invariant Generation

Automatic invariant generation techniques have been studied since the 1970s [14,
21,27,36], and more recently in the work of Bjørner, Browne, and Manna [3], and
Bensalem, Lakhnech, and Saı̈di [9,34,7].
As in model checking, the basic operation in invariant generation is that
of taking the strongest postcondition or weakest precondition of a state set X
with respect to the transition relation N . Some of the techniques for computing
invariants are described briefly below.

Least Fixed Point of the Strongest Postcondition. The invariant computed here
corresponds to the reachability state set. It is computed by starting with an
initial symbolic representation of the initial state set given by the program. This
set is successively enlarged by taking its image under the strongest postcondition
operation until a fixed point is reached, i.e., no new elements are added to the set.
We term this method LFP-SP. It yields a symbolic representation of the set of
reachable states which is the strongest invariant. However, LFP-SP computation
often does not terminate since the computation might not converge to a fixed
point in a finite number of steps. Take, for example, a program that successively
increments by one, a variable x that is initially zero. This program has a least
fixed point, i.e., x is in the set of natural numbers, but the iterative computation
does not converge.
For the Twos example, the LFP-SP computation does terminate with the
desired invariant as seen in the calculation below.

Inv 0 = (P C = inc ∧ B = 0 ∧ C = 0)
Inv 1 = Inv 0 ∨ (P C = dec ∧ B = 2 ∧ C = 0)
Inv 2 = Inv 1 ∨ (B = 0 ∧ C = 1)
Inv 3 = (B = 0 ∧ C = 1) ∨ (P C = dec ∧ B = 2 ∧ C = 0)
= Inv 2

The resulting invariant easily implies the strengthened inductive invari-

ant (4). The LFP-SP computation terminates precisely because the reachable
state set is bounded. In more typical examples, approximation techniques based
on widening will be needed to accelerate the convergence of the least fixed point
computation.

Greatest Fixed Point of the Strongest Postcondition. The greatest fixed point
iteration starts with the entire state space and strengthens it in each iteration
294 N. Shankar

by excluding states that are definitely unreachable. This approach, which we call
GFP-SP, yields a weaker invariant than the least fixed point computation. The
GFP-SP computation also need not terminate. Even when it does terminate, the
resulting invariant might not be strong enough. In the case of the program with
single integer variable x that is initially zero and incremented by one in each
transition, the GFP-SP computation returns the trivial invariant true. However
the GFP-SP method has the advantage that it can be made to converge more
easily than the LFP-SP method, and any intermediate step in the computation
already yields a valid invariant.
The greatest fixed point invariant computation for Twos (ignoring the varia-
ble C) can be carried out as follows. Here Inv i (pc) represents the i iteration of
the invariant for control state pc.

Inv 0 (inc) = (B = 0 ∨ B ≥ −1) = (B ≥ −1)

Inv 0 (sub) = true

Inv 1 (inc) = (B ≥ −1)

Inv 1 (sub) = (B ≥ 1 ∨ B ≥ −1) = (B ≥ −1)

Inv 2 (inc) = (B ≥ −1)

Inv 2 (sub) = (B ≥ −1)

The invariant B ≥ −1 is not all that useful since this information contributes
nothing to the invariants that we wish to establish. Still, the GFP-SP method
is not without value. It is especially useful for propagating known invariants.
For example, if we start the iteration with invariant (1), then we can use the
GFP-SP method to deduce that the strengthened invariant (4).

Greatest Fixed Point of the Weakest Precondition. Both LFP-SP and GFP-
SP compute inductive invariants that are valid, whereas the method GFP-WP
takes a putative invariant and strengthens it in order to make it inductive. The
computation starts with a putative invariant S, and successively applies the
weakest precondition operation wp(P)(S) to it. If this computation terminates,
then either the resulting assertion is a strengthening of the original invariant
that is also inductive, or the given invariant is shown to be invalid.
With the Twos example, the weakest precondition with respect to the puta-
tive invariant (1) yields the strengthened invariant (4).

2.5 Abstract Interpretation

Many of the invariant generation techniques are already examples of abstract
interpretation which is a general framework for lifting program execution from
the concrete domain of values to a more abstract domain of properties. Examples
of abstract interpretation include sign analysis (positive, negative, or zero) of
 Symbolic Analysis of Transition Systems 295

variables, interval analysis (computing bounds on the range of values a variable

can take), live variable analysis (the value of a variable at a control point might
be used in the computation to follow), among many others.
We can apply an interval analysis to the Twos example. Initially, the interval
for B is [0, 0] for PC = inc. This yields an interval of [2, 2] for B when PC = dec.
In the next step, we have an approximation of [0, 2] for B when PC = dec, and
[0, 0] when PC = inc. The next round, we get an approximation of [−1, 0] for
the range of B when PC = inc, and [0, 2] for the range of B when PC = dec.
At this point the computation converges, but the results of the analysis are still
too approximate and do not discharge the invariant (1).

2.6 Property Preserving Abstractions

Since model checking is unable to cope with systems with infinite or large
state spaces, abstraction has been studied as a technique for reducing the state
space [12,29,35]. In data abstraction, a variable over an infinite or large type is
reduced to one over a smaller type. The smaller type is essentially a quotient with
respect to some equivalence relation of the larger type. For example, a variable
ranging over the integers can be reduced to boolean form by considering only
the parity (odd or even) of the numbers. Predicate abstraction is an extension
of data abstraction that introduces boolean variables for predicates over a set of
variables. For example, if x and y are two integer variables in a program, it is
possible to abstract the program with respect to the predicates such as x < y,
x = y. These variables are then replaced by boolean variables p and q such that
p corresponds to the x < y and q corresponds to x = y. Even though predicate
abstraction introduces only boolean variables, it is possible to simulate a data
abstraction of a variable to one of finite type by using a binary encoding of the
finite type.
In general, an abstraction is given by means of a concretization map γ such
that γ(a) for an abstract variable a returns its concrete counterpart. In the case
of the abstraction where x < y is replaced by p and x = y by q, γ(a) = (x < y)
and γ(b) = (x = y). The more difficult direction is computing an abstraction
α(C) given a concrete predicate C. The construction of α requires the use of
theorem proving as described below.
There are also two ways of using abstractions in symbolic analysis. In one
approach, the abstract reachability set [35,17] is constructed by the following
iteration
ARG(P )(s) = lf p(α(IP ) ∨ α ◦ sp(P ) ◦ γ).
We can then check if p is an invariant of P by verifying γ(ARG(P )) ⊃ p.
A second way of using abstraction is by actually constructing the abstracted
version of the program and the property of interest [8,16,37]. This can be more
efficient since the program and property are usually smaller than the abstract
reachability graph.
In the Twos example, the predicate abstraction is suggested by the predicates
B = 0 and B = 2 in the putative invariant. The abstract transition system by
replacing the predicate B = 0 by c and B = 2 by d is shown in Figure 2.
296 N. Shankar

inc dec
c, -d -c, d

dec
c, -d

Fig. 2. Abstract Twos

The abstract transition system computed using predicate abstraction can

easily be model checked to confirm that invariant (1) holds. The stronger inva-
riant (4) can also be extracted from the reachable state space of the abstract
transition system.
Predicate abstraction affords an effective integration of theorem proving and
model checking where the former is used to construct a finite-state property-
preserving abstraction that can be analyzed using the latter. The abstraction
loses information so that a property can fail to hold in the abstract system even
when its concrete counterpart is valid for the concrete system. In this case, the
abstraction has to be refined by introducing further predicates for abstraction [8,
11].

3 SAL: A Symbolic Analysis Laboratory

We have already seen a catalog of symbolic analysis techniques. The idea of
a symbolic analysis laboratory is to allow these techniques to coexist so that
the analysis of a transition system can be carried out by successive applications
of a combination of these techniques [6]. With such a combination of analysis
techniques, one could envisage a verification methodology where
1. A cone-of-influence reduction is used to discard irrelevant variables.
2. Invariant generation is used to obtain small but useful invariants.
3. These invariants are used to obtain a reasonably accurate abstraction to a
finite-state transition system.
4. Model checking is used to compute useful invariants of the finite-state ab-
straction.
5. The invariants computed by model checking over the abstraction are used
propagated using invariant generation techniques.
6. This cycle can be repeated until no further useful information is forthcoming.
SAL provides a blackboard architecture for symbolic analysis where a collec-
tion of tools interact through a common intermediate language for transition
 Symbolic Analysis of Transition Systems 297

SVC
Esterel
SAL
Murphi
ASM
Mocha

JAVA
PVS

SMV

Programs
Verification
conditions
Abstractions
Properties

Fig. 3. The Architecture of SAL

systems. The individual analyzers (theorem provers, model checkers, static ana-
lyzers) are driven from this intermediate language and the analysis results are
fed back to this intermediate level. In order to analyze systems that are written
in a conventional source language, the transition system model of the source pro-
gram has to be extracted and cast in the SAL intermediate language.1 The model
extracted in the SAL intermediate language essentially captures the transition
system semantics of the original source program.
The SAL architecture is shown in Figure 3 The SAL architecture is constrai-
ned so that the different analysis tools do not communicate directly with each
other, but do so through the SAL intermediate language. The interaction bet-
ween the tools must therefore be at a coarse level of granularity, namely in terms
of transition systems, their properties, and property-preserving transformations
between transition systems. Allowing the tools to communicate directly to each
other would require a quadratic number of different maps (for a given number
of tools) between these analysis tools.

3.1 The SAL Intermediate Language

The intermediate language for SAL2 serves as

1. The target of translations from source languages.
2. The source for translations to the input formats of different analysis tools.
3. A medium for communication between different analysis tools.
1
We are currently working on a translator from a subset of Verilog to SAL, and
another from a subset of Java to SAL.
2
The SAL intermediate language was designed in collaboration with Prof. David Dill
of Stanford, Prof. Tom Henzinger at UC Berkeley, and several colleagues at SRI,
Stanford, and UC Berkeley.
298 N. Shankar

The SAL intermediate language is based on languages and models such

as SMV [30], Murphi [18], Reactive Modules [2], ASM [20], UNITY [15], and
TLA [28], among others. The unit of specification in SAL is a context which
contains declarations of types, constants, transition system modules, and as-
sertions. A SAL module is a transition system unit. A basic SAL module is a
state transition system where the state consists of input, output, local , and global
variables, where

– An input variable to a module can be read but not written by the module.
– An output variable to a module can be read and written by the module, and
only read by an external module.
– A local variable to a module can be read and written by the module, but is
not read or written by the module.
– A global variable to a module can be read and written by the module as well
as by an external module
A basic module also specifies the initialization and transition steps. These can
be given by a combination of definitions or guarded commands. A definition is of
the form x = expression or x0 = expression, where x0 refers to the new value of
variable x in a transition. A definition can also be given as a selection of the form
x0 ∈ set which means that the new value of x is nondeterministically selected
from the value of of set. A guarded command is of the form g −→ S, where g
is a boolean guard and S is a list of definitions of the form x0 = expression or
x0 ∈ set.
As in synchronous language such as Esterel [5] and Lustre [22], SAL allows
synchronous, i.e., Mealy machine, interaction so that the new value of a local or
output variable can be determined by the new value of a variable. Such inter-
action introduces the possibility of a causal cycle where each variable is defined
to react synchronously to the other. Such causal cycles are ruled out by using
static analysis to generate proof obligations demonstrating that such cycles are
not reachable. The UNITY and ASM models do not admit such synchronous
interaction since the new values of a variable in a transition are completely de-
termined by the old values of the variables. SMV allows such interaction but the
semantics is not clearly specified, particularly when causal cycles are possible.
The Reactive Modules [2] language uses a static partial ordering on the variables
that breaks causal loops by allowing synchronous interaction in one direction of
the ordering but not the other. In TLA [28], two modules are composed by conjoi-
ning their transition relations. TLA allows synchronous interaction where causal
loops can be resolved in any manner that is compatible with the conjunction of
the transition relations is satisfied.
SAL modules can be composed

– Synchronously, so that M1 kM2 is a module that takes M1 and M2 transitions

in lockstep, or
– Asynchronously, so that M1 [] M2 is a module that takes an interleaving of
M1 and M2 transitions.
 Symbolic Analysis of Transition Systems 299

There are rules that govern the usage of variables within a composition.
Two modules engaged in a composition must not share output variables and nor
should the output variables of one module overlap with the global variables of
another. The modules can can share input and global variables, and the input
variables of one module can be the output or global variables of the other. Two
modules that share a global variable cannot be composed synchronously, since
this might create a conflict when both modules attempt to write the variable
synchronously. The rules governing composition allow systems to be analyzed
modularly so that system properties can be composed from module properties [2].
The N-fold synchronous and asynchronous compositions of modules are also
expressible in SAL. Module operations include those for hiding and renaming
of variables. Any module defined by means of composition and other module
operations can always be written as a single basic module, but with a significant
loss of succinctness.
SAL does not contain features other than the rudimentary ones described
above. There are no constructs for synchronization, synchronous message pas-
sing, or dynamic process creation. These have to explicitly implemented by me-
ans of the transition system mechanisms available in SAL. While these features
are useful, their introduction into the language would place a greater burden on
the analysis tools.
The SAL language is thus similar in spirit to Abstract State Machines [20]
in that both serve as basic conceptual models for transition systems. However,
machines described in SAL are not abstract compared with those in ASM nota-
tion since SAL is intended as a front-end to various popular model checking and
program analysis tools.

4 Conclusions

Powerful automated verification technologies have become available in the form

of model checkers for finite, timed, and hybrid systems, decision procedures,
theorem provers, and static analyzers. Individually, these technologies are quite
limited in the range of systems or properties they can handle with a high degree
of automation. These technologies are complementary in the sense that one is
powerful where the other is weak. Static analysis can derive properties by me-
ans of a syntactic analysis. Model checking is best suited for control-intensive
systems. Theorem proving is most appropriate for verifying mathematical pro-
perties of the data domain. Symbolic analysis is aimed at achieving a synergistic
integration of these analysis techniques. The unifying ideas are

1. The use of transition systems as a unifying model, and

2. Fixed point computations over symbolic representations as the unifying ana-
lysis scheme.
3. Abstraction as the key technique for reducing infinite-state systems to finite-
state form.
300 N. Shankar

Implementation work on the SAL framework is currently ongoing. The preli-

minary version of SAL consists of a parser, typechecker, causality checker, an
invariant generator, translators from SAL to SMV and PVS, and some other
tools. SAL is intended as an experimental framework for studying the ways in
which different symbolic analysis techniques can be combined to achieve greater
automation in the verification of transition systems.

Acknowledgments. Many collaborators and colleagues have contributed ideas

and code to the SAL language and framework, including Saddek Bensalem, Da-
vid Dill, Tom Henzinger, Luca de Alfaro, Vijay Ganesh, Yassine Lakhnech, Cesar
Muñoz, Sam Owre, Harald Rueß, John Rushby, Vlad Rusu, Hassen Saı̈di, Eli
Singerman, Mandayam Srivas, Jens Skakkebæk, and Ashish Tiwari.

References
1. Rajeev Alur and Thomas A. Henzinger, editors. Computer-Aided Verifi-
cation, CAV ’96, volume 1102 of Lecture Notes in Computer Science, New
Brunswick, NJ, July/August 1996. Springer-Verlag.
2. R. Alur and T.A. Henzinger. Reactive modules. Formal Methods in System
Design, 15(1):7–48, 1999.
3. Nikolaj Bjørner, I. Anca Browne, and Zohar Manna. Automatic generation
of invariants and intermediate assertions. Theoretical Computer Science,
173(1):49–87, 1997.
4. J. R. Burch, E. M. Clarke, K. L. McMillan, D. L. Dill, and L. J. Hwang.
Symbolic model checking: 1020 states and beyond. Information and Com-
putation, 98(2):142–170, June 1992.
5. G. Berry and G. Gonthier. The Esterel synchronous programming lan-
guage: Design, semantics, and implementation. Science of Computer Pro-
gramming, 19(2):87–152, 1992.
6. Saddek Bensalem, Vijay Ganesh, Yassine Lakhnech, César Muñoz, Sam
Owre, Harald Rueß, John Rushby, Vlad Rusu, Hassen Saı̈di, N. Shankar,
Eli Singerman, and Ashish Tiwari. An overview of SAL. In C. Michael Hol-
loway, editor, LFM 2000: Fifth NASA Langley Formal Methods Workshop,
Hampton, VA, June 2000. NASA Langley Research Center. To appear.
7. Saddek Bensalem and Yassine Lakhnech. Automatic generation of invari-
ants. Formal Methods in Systems Design, 15(1):75–92, July 1999.
8. Saddek Bensalem, Yassine Lakhnech, and Sam Owre. Computing abstrac-
tions of infinite state systems compositionally and automatically. In Hu
and Vardi [26], pages 319–331.
9. Saddek Bensalem, Yassine Lakhnech, and Hassen Saı̈di. Powerful techni-
ques for the automatic generation of invariants. In Alur and Henzinger [1],
pages 323–335.
10. R. E. Bryant. Graph-based algorithms for Boolean function manipulation.
IEEE Transactions on Computers, C-35(8):677–691, August 1986.
11. Edmund Clarke, Orna Grumberg, Somesh Jha, Yuan Lu, and Helmut Veith.
Counterexample-guided abstraction refinement. In E. A. Emerson and A. P.
Sistla, editors, Computer-Aided Verification, Lecture Notes in Computer
Science. Springer-Verlag, 2000. To appear.
 Symbolic Analysis of Transition Systems 301

12. Edmund M. Clarke, Orna Grumberg, and David E. Long. Model checking
and abstraction. ACM Transactions on Programming Languages and Sy-
stems, 16(5):1512–1542, September 1994.
13. E. M. Clarke, Orna Grumberg, and Doron Peled. Model Checking. MIT
Press, 1999.
14. P. Cousot and N. Halbwachs. Automatic discovery of linear restraints
among variables. In 5th ACM Symposium on Principles of Programming
Languages. Association for Computing Machinery, January 1978.
15. K. Mani Chandy and Jayadev Misra. Parallel Program Design: A Founda-
tion. Addison-Wesley, Reading, MA, 1988.
16. M. A. Col on and T. E. Uribe. Generating finite-state abstractions of re-
active systems using decidion procedures. In Hu and Vardi [26], pages
293–304.
17. Satyaki Das, David L. Dill, and Seungjoon Park. Experience with predicate
abstraction. In Halbwachs and Peled [25], pages 160–171.
18. David L. Dill. The Murφ verification system. In Alur and Henzinger [1],
pages 390–393.
19. R. Gerth, D. Peled, M. Y. Vardi, and P. Wolper. Simple on-the-fly auto-
matic verification of linear temporal logic. In Proc. 15th Work. Protocol
Specification, Testing, and Verification, Warsaw, June 1995. North-Holland.
20. Yuri Gurevich. Evolving algebras 1993: Lipari guide. In Egon Börger, edi-
tor, Specification and Validation Methods, International Schools for Com-
puter Scientists, pages 9–36. Oxford University Press, Oxford, UK, 1995.
21. S. M. German and B. Wegbreit. A synthesizer for inductive assertions.
IEEE Transactions on Software Engineering, 1(1):68–75, March 1975.
22. N. Halbwachs, P. Caspi, P. Raymond, and D. Pilaud. The synchronous da-
taflow programming language Lustre. Proceedings of the IEEE, 79(9):1305–
1320, September 1991.
23. C. A. R. Hoare. An axiomatic basis of computer programming. Commu-
nications of the ACM, 12(10):576–580, October 1969.
24. G. J. Holzmann. Design and Validation of Computer Protocols. Prentice
Hall, 1991.
25. Nicolas Halbwachs and Doron Peled, editors. Computer-Aided Verification,
CAV ’99, volume 1633 of Lecture Notes in Computer Science, Trento, Italy,
July 1999. Springer-Verlag.
26. Alan J. Hu and Moshe Y. Vardi, editors. Computer-Aided Verification,
CAV ’98, volume 1427 of Lecture Notes in Computer Science, Vancouver,
Canada, June 1998. Springer-Verlag.
27. S. Katz and Z. Manna. Logical analysis of programs. Communications of
the ACM, 19(4):188–206, April 1976.
28. Leslie Lamport. The temporal logic of actions. ACM TOPLAS, 16(3):872–
923, May 1994.
29. C. Loiseaux, S. Graf, J. Sifakis, A. Bouajjani, and S. Bensalem. Property
preserving abstractions for the verification of concurrent systems. Formal
Methods in System Design, 6:11–44, 1995.
30. Kenneth L. McMillan. Symbolic Model Checking. Kluwer Academic Publis-
hers, Boston, MA, 1993.
31. Zohar Manna and Amir Pnueli. The Temporal Logic of Reactive and Con-
current Systems, Volume 1: Specification. Springer-Verlag, New York, NY,
1992.
302 N. Shankar

32. Zohar Manna and The STeP Group. STeP: Deductive-algorithmic verifi-
cation of reactive and real-time systems. In Alur and Henzinger [1], pages
415–418.
33. A. Pnueli. The temporal logic of programs. In Proc. 18th Symposium on
Foundations of Computer Science, pages 46–57, Providence, RI, November
1977. ACM.
34. Hassen Saı̈di. A tool for proving invariance properties of concurrent systems
automatically. In Tools and Algorithms for the Construction and Analysis
of Systems TACAS ’96, volume 1055 of Lecture Notes in Computer Science,
pages 412–416, Passau, Germany, March 1996. Springer-Verlag.
35. Hassen Saı̈di and Susanne Graf. Construction of abstract state graphs with
PVS. In Orna Grumberg, editor, Computer-Aided Verification, CAV ’97,
volume 1254 of Lecture Notes in Computer Science, pages 72–83, Haifa,
Israel, June 1997. Springer-Verlag.
36. N. Suzuki and K. Ishihata. Implementation of an array bound checker.
In 4th ACM Symposium on Principles of Programming Languages, pages
132–143, January 1977.
37. Hassen Saı̈di and N. Shankar. Abstract and model check while you prove.
In Halbwachs and Peled [25], pages 443–454.
38. Moshe Y. Vardi and Pierre Wolper. An automata-theoretic approach to
automatic program verification (preliminary report). In Proceedings 1st
Annual IEEE Symp. on Logic in Computer Science, pages 332–344. IEEE
Computer Society Press, 1986.
 Encoding Abstract State Machines in PVS

Angelo Gargantini1 and Elvinia Riccobene2

1
Dipartimento di Elettronica e Informazione - Politecnico di Milano -
[email protected]
2
Dipartimento di Matematica e Informatica - Università di Catania -
[email protected]

Abstract. In this paper we show how the specification and verification

system PVS (Prototype Verification System) can provide tool support for
Abstract State Machines (ASMs), especially oriented towards automatic
proof checking and mechanized proving of properties. Useful templates
are presented which allow encoding of ASM models into PVS without
any extra user’s skill. We prove the transformation preserves the ASM
semantics and provide a framework for an automatic tool, prototypically
implemented, which translates ASM specifications in PVS. The ASM
specification of the Production Cell given in [4] is taken as case study to
show how to formalize multi-agent ASMs in PVS and prove properties.

1 Introduction
The Gurevich’s Abstract State Machines (ASMs) [9] have been successfully used
for design and analysis of complex hardware/software systems [3,2]. Through
real case studies, ASMs have shown to be a practical method for rigorous sy-
stem development and to meet the requirements, addressed by Heitmeyer in [10],
that formal methods need to have “to be useful to practitioners”: a user-friendly
notation, useful, easy to understand feedback, integration into standard develop-
ment process. ASMs use only the standard language and standard methods of
programming and mathematics. They are models easy to read, understand and
inspectable by the customer, but they are flexible to change, and precise and
complete to match the designer’s requirements. However, we believe that the
success of a formal method in industrial field also depends on the availability of
related automated tools helping the user during the process of specification and
subsequent verification and validation. Therefore, machine support for the use
of ASMs is extremely useful.
During the last years various investigations have been started in order to
verify standard mathematical reasoning about ASMs by interactive or fully au-
tomated proof tools. Some encouraging results are already reported in literature
(see discussion in [3]). Among theorem provers, PVS (Prototype Verification
System) has been used in [7] to show the correctness of bottom-up rewriting
specification for back-end compilers from an intermediate language into binary
RISC processor code. Dold et al. state that “erroneous rules have been found
using PVS and, through failed proof attempts, errors were corrected by ins-
pection of the proof state”. Mechanical checking of formal specifications and

Y. Gurevich et al. (Eds.): ASM 2000, LNCS 1912, pp. 303–322, 2000.
c Springer-Verlag Berlin Heidelberg 2000

304 A. Gargantini and E. Riccobene

properties is useful both as a final confirmation of their correctness and for sim-
plifying the process of discovering inconsistencies and proving properties. In fact,
experience shows that even the most carefully crafted formal specifications and
proofs, when done by hand, can still contain inconsistencies and other errors,
and that such errors are discovered only using tools capable to perform some
sort of analysis (also simply name checking, type consistency checking or other
similar simple analysis). Proofs can also be so long and tedious that machine
support can reduce the human effort.
In this paper, we show how PVS can provide tool support for ASMs, es-
pecially oriented towards automatic proof checking and mechanized proving of
properties. PVS [13], developed by SRI, is a higher order logic specification and
verification environment with a Gentzen-like deduction system. PVS provides
both a highly expressive specification language and automation of proof steps.
Starting from the work of Dold et al. in [8,7] we develop suitable PVS theories
to encode ASM models in PVS and we prove that our transformation preserves
the semantics of ASMs. Our approach differs from that of Dold et al. in the way
of encoding the ASM transition system. Our goal is to find a transformation from
ASMs to PVS that preserves the layout of the original rule transition system. To
achieve this goal we do not take any strong assumption as “only one rule enabled
at a time”, indeed we allow more rules to be simultaneously executed. Therefore,
we do not demand that the user should transform the whole system of rules in
one meta rule, transformation which can require skill and ingenuity, especially
in case of multi-agent models. Instead, we propose an algorithmic approach of
transformation which keeps the set of transition rules as a set of different rules.
We provide useful templates for helping the user during the encoding phase,
and useful proof schemes for requirements verification. The suggested encoding
scheme is mechanizable and we provide a framework for a tool, that we pro-
totypically implemented, supporting automatic transformation in PVS of ASM
models given in the ASM-SL language [5].
Our PVS encoding of ASMs has been tested for multi agent models. We
report here the PVS specification of the Production Cell given in [4], and we
discuss the results of the mechanized proofs of safety and liveness properties.
Several proofs are obtained with assumptions, usually regarding the physical
environment behavior, that are implicit in the formal description but need to
be added in the machine supported approach. Thanks to the PVS features, we
formulate such assumptions in a declarative way instead of using the operational
way required by the model checking approach used by Winter in [14] to show
the correctness of the Production Cell specification.
The article is organized as follows. The problem of encoding ASMs in PVS
is discussed in Sections 2.1-2.3. Section 2.4 shows how to validate specifications
using PVS. Sections 2.5 and 2.6 present proof schemes to support verification of
invariant properties and trace properties. Section 3 contains the PVS specifica-
tion of the Production Cell case study and the results of the verification task.
In Section 4 we discuss our work in comparison with related results, and we
conclude outlining future research directions.
 Encoding Abstract State Machines in PVS 305

2 An Encoding of ASMs in PVS

In this section we present our encoding of ASMs in PVS. It consists of a set of
PVS theories, hierarchically organized, which define types and functions mode-
ling ASM universes and rules, and of a set of PVS strategies defined in order to
simplify the proof conduction. Universe and function representation is based on
the work of Dold et al. [8,7]. PVS encoding of rules and dynamics of ASMs is
significantly changed.

2.1 Abstract State

An ASM state models a(n abstract) machine state, i.e. the collection of elements
the machine “knows”, and the functions and predicates it uses to manipulate
such elements. Mathematically a state is defined as a structure which consists
of a collections of domains (sets, also called universes, each of which stands for
a particular type of elements) with arbitrary functions and relations defined on
them. Universes representation. Universes have different encoding depending
on their being static or dynamic:
– A static universe U , i.e. a universe which does not change during computa-
tion, is encoded as uninterpreted type U:TYPE.
– A dynamic universe U , i.e. a universe which may grow during computation,
is encoded as a PVS set U:setof[T] of elements of type T.
As elements can not be added to PVS types, only such encoding allows to
expand universes. To this purpose, the functions add and union, (pre-)defined
in the PVS theory sets, can be used.
Remark 1. The concept of type in PVS is very different from the usual mea-
ning of type in programming languages. PVS types abstractly represent entities
with common operations. There are only a few built-in PVS types such as num-
bers, integers and boolean. Other types may be defined either as uninterpreted
types, or through the construct DATATYPE, or in terms of already defined types
by usual type constructors (as lists, records, etc.).
It is possible to encode the superuniverse itself as an uninterpreted type S
and define every universe U as a subset of S – as already noted by Dold in [8]–,
or as uninterpreted subtype (with the same meaning as subset). However, in
the last case it is necessary to introduce axioms to guarantee properties on sets,
as, for example, disjointness between universes, so not exploiting the strong and
powerful type system of PVS. For this reason we suggest to encode universes
as uninterpreted types whenever possible, and use the construct setof only for
dynamic universes.

Functions and states. ASM basic functions are classified in static functions
which remain constant, and dynamic functions which may change interpretation
during computation. Dynamic functions are furthermore classified in controlled,
monitored and shared. Controlled functions occur inside function updates and
306 A. Gargantini and E. Riccobene

can change by application of transition rules. Monitored functions can change

only due to the environment. Shared functions can change by application of a
transition rule but also by some other agents; they formalize the interactions in
case of multi-agent computations and combined updates of locations. Functions
defined in terms of the basic ones are called derived.
We encode functions almost as proposed in [8].
– Static functions are encoded as usual PVS functions.
– Monitored functions are modeled as PVS functions adding the type ENV to
the arguments of their domain. ENV is an uninterpreted type which represents
the external environment, that is everything outside the system’s control. A
boolean dynamic monitored variable is therefore encoded as a function from
ENV to bool, whereas a generic monitored function f : S → T is translated
into the function f:[ENV,S->T].
Our way of interpreting ENV is different from that proposed in [8] where ENV
is the record of all monitored functions. Our view of ENV has the advantage
of allowing a designer to easily add and remove monitored functions.
– Controlled functions are regarded as fields of a record CTRLSTATE which
represents the controlled part of an ASM state. Therefore, if f : A → B and
g : Int → Bool are controlled functions, the controlled part of the state is
defined as CTRLSTATE: TYPE [# f:[A->B], g:[int->bool] #].
– Shared function are defined as controlled functions. In a multi-agent ASM,
the controlled part of a “global” state is viewed as the collection of all con-
trolled agent’s states. Therefore, shared functions, as controlled by at least
one agent, are defined as components of the CTRLSTATE.
– Derived function are encoded as functions defined in terms of their function
components in a straightforward way from their ASM specification.
An ASM state is compound of its controlled part and the environment. It is
defined as the record STATE: TYPE [# env: ENV, ctrl: CTRLSTATE #].

2.2 ASM Rules

ASM (transition) rules model the actions performed by the machine to manipu-
late elements of its domains and which will result in a new state. A transition
rule Ri has the general form:
Ri : if condi then update i
condi , also called the guard of the rule, is a statement expressing the condition
under which Ri must be applied; update i consists of finitely many functions up-
dates:
f (t1 , . . . , tn ) := t
which are executed simultaneously. f is an arbitrary n-ary function and t1 , ..., tn
is the tuple of arguments at which the value of the function is set to t.

How to serialize the parallel rule application. In this section we tackle the
problem of finding an algorithm to serialize firing of ASM rules which guarantees
the semantics of their parallel application as stated in [3]:
 Encoding Abstract State Machines in PVS 307

An ASM M is a finite set of rules [. . . ]. Applying one step of M to a

state A produces as next state another algebra A0 , of the same signature,
obtained as follows.
First evaluate in A, using the standard interpretation of classical logic, all
the guards of all the rules of M. Then compute in A, for each of the rules
of M whose guard evaluates to true, all the arguments and all the values
appearing in the updates of this rule. Finally replace, simultaneously for
each rule and for all the locations in question, the previous A function
value by the newly computed value.

Let {R1 , ..., Rn } be a set of rules. The naive approach of computing the next
state A0 of the current state A by applying rules Ri , i = 1 . . . n, sequentially,1
does not guarantee the ASM semantics for two reasons: (a) guards condi of Ri
are not evaluated in the same state A forall i; (b) terms occurring in the left and
right sides of function updates of Ri are not all evaluated in A. The following
example may help to convince that this approach is wrong.
Example. Let A = (x → 0, y → 1, z → 2) be the current state consisting of
three integer variables x, y, z, and {R1 , R2 , R3 } be the transition system with
rules defined as follows:
R1 : if x = 0 then y := 5 R2 : if y = 5 then x := 2 R3 : z := x + y
By sequential rules application, it results:
       
x→0 x→0 x→ 2 x→2
R1 R2 R3
A :  y → 1  →  y → 5  →  y → 5  →  y → 5  : A0
z→2 z→2 z→2 z→ 7

whereas by correct computation would be A0 = (x → 0, y → 5, z → 1).

An algorithm which overcomes the errors of the naive attempt is given below.
Let {R1 ,..., Rn } be a set of transition rules, Ri of the form if condi then
updatei with update i a sequence of function updates f (t1 , t2 , . . . , tn ) := t, and A
be the current state.
The algorithm can be stated as follows:
1. For all rules R1 ,..., Rn , mark Ri if enabled at the state A.
2. For all i such that marked Ri , evaluate in A all terms t1 , t2 , . . . , tn , t occurring
in function updates f (t1 , t2 , . . . , tn ) := t of update i .
3. For all i such that marked Ri , sequentially perform the function updates in
update i using the values computed at step 2.
The proposed algorithm initially evaluates all the guards condi in A and only
afterwards, sequentially, performs all the function updates. Applying the algo-
rithm to the above example we get the following correct transformation:

1
I.e., first evaluate the guard of R1 and, if it is true, perform the updates defined in
R1 to obtain A1 ; then apply R2 to A1 and obtain A2 ; and so on till computing the
final state An to be taken as A0 .
308 A. Gargantini and E. Riccobene

       
x→0 x→0 x→0 x→0
R R R
A :  y → 1  →1  y → 5  →2  y → 5  →3  y → 5  : A0
z→2 z→2 z→2 z→ 1

Rule R2 does not affect the state A because not enabled, and R3 correctly
computes the new value of z using the values of x and y in A.

Inconsistent updates. If a location is updated to different values by simul-

taneously firing rules, then we speak of “inconsistent updates”. Although the
location value should remain unchanged due to the non-execution of inconsi-
stent updates [9], according to our algorithm, the location would take the last
value assigned to it. For this reason, at the moment, we are able to consider only
consistent ASM models. It would be possible to discover inconsistent updates
through simulation of suitable test cases. However, simulation can never gua-
rantee absence of inconsistency. Applying a formal verification technique similar
to that used for proving invariant properties in Section 2.4, we might discover
inconsistent updates. Checking inconsistency of ASM specifications is a problem
still under investigation. Note that the problem of firing inconsistent updates is
not taken in consideration in [8,7].

How to deal with extend. The proposed algorithm does not give the expected
result when dealing with combination of rules which simultaneously add elements
to a same set by the construct extend. This problem arose trying to encode
in PVS the ASM model of crypto-protocol given in [1]. In this model, different
agents can send at the same time messages into the TRAFFIC (a set of messages)
representing the network channel common to all agents.
To be convinced why extend must be dealt with particular attention, con-
sider the following example. Let A be a state where a set ∆ has value ∅. The
following rules are enabled in A:
R1 : if ∆ = ∅ then extend ∆ with a
R2 : if ∆ = ∅ then extend ∆ with b
Encoding “extend ∆ with u” as a function update of the form ∆ := ∆ ∪{u},
by our algorithm, we get the following transformation:
R R
A : (∆ → {}) →1 (∆ → {a}) →2 (∆ → {b}) : A0

However, as both rules are applicable, according to the correct semantics of

extend, we expect the result ∆ = {a, b} in the final state A0 .
The algorithm needs to be adjusted to deal with extend. A correct solution
should (a) evaluate expressions in the current state; (b) fire updates not involving
extend against the current state; (c) if a rule Ri contains an update extend U
with u, add u to the set U interpreted not in the current state, but in the state
obtained after firing all rules preceding Ri (otherwise we could loose updates of
U by other possible extends of previous rules).
 Encoding Abstract State Machines in PVS 309

The revisited version of the algorithm follows:

1. For all rules R1 ,..., Rn , mark Ri if enabled at the state A.

2. For all i such that marked Ri , evaluate in A all terms occurring in func-
tion updates of update i with exception of those representing universes as
arguments of extend.
3. Let, unless renaming, {R1 , ..., Rm } be the set of all marked rules. Assume
s0 = A. Sequentially, compute si from si−1 , i = 1 . . . m, performing function
updates of update i of rule Ri in si−1 : use values computed at step 2 for
all terms except those representing universes within extend which must be
evaluated in si−1 . The final state sm be A0 .

Remark 2: Reserve and multisets. According to the ASM semantics,

elements added to extended universes are “new” and imported from Reserve.
To guarantee that, Dold et al. introduce a predicate new to control an element
does not already belong to the set to be extended, and a function sort update
which extends a set by new elements. However, PVS does not allow sets to
contain different occurrences of the same element, and even Dold’s formalization
does not avoid such a problem. When an ASM specification requires a universe
to contain multiple occurrences of a same element, we suggest to encode the
universe as a multiset instead of a set. Keeping in mind that in PVS a set of
elements of type T is a function from T to bool, as natural extension, a multiset
can be encoded as a function from T to natural numbers: the value of the function
represents the number of occurrences of its argument in the multiset. Functions
and predicates on sets, like add, member, emptyset, etc., must be redefined. The
PVS theory to model multisets follows:

multisets [T: TYPE]: THEORY

BEGIN
multiset: TYPE = [T-> nat]
x, y: VAR T
a: VAR multiset
% an element x is member of a only if the occurrences
% of x in a are greater than 0
member(x, a): bool = a(x) > 0
empty?(a): bool = (FORALL x: NOT member(x, a))
emptyset: multiset = lambda x: 0
nonempty?(a): bool = NOT empty?(a)
% the function ‘‘add’’ returns the same multiset modified
% increasing by one the number of occurrences of the added
% element
add(x, a): (nonempty?) = lambda y:
if x = y then a(y)+ 1 else a(y) endif
END multisets

Remark 3. The rule encoding proposed in [8,7] is based on the (strong)

assumption that rules “concern distinct (disjoint) cases”. As a consequence, one
310 A. Gargantini and E. Riccobene

may infer that (a) at every step only one rule is enabled, and therefore (b) all
rules “may be combined into a single state transition function, defined by a case
construct or nested conditionals”. Such a single transition function is called “one-
step interpreter” by Dold et al. This assumption of disjoint cases is restrictive,
since in many applications more then one rule might fire at a time. In case
of possible simultaneous firing of many rules, a unique transformation function
might still model the ASM dynamics as well, however we prefer keeping the set
of transition rules as set of distinct rules, without forcing the user to define an
equivalent transformation function from the current state to the next one. The
advantages are two: first stylistic, because we preserve the ASMs structure, and
second practice, because writing one equivalent global transformation function
requires skill and ingenuity. Furthermore, due to the assumption that only one
rule can be applied in one step, the one-step interpreter is absolutely not suitable
in case of multi-agent ASMs.
Remark 4. Another possible approach for rule encoding would be to define
the value of every variable and function (location) in the next step by means of
axioms. For example, the update f (x) := t would be translated into an axiom
of the form update : AXIOM f(next(s))(x) = t, where next(s) stands for
the next state. In this way it would not be necessary to introduce the state
as a record and we could consider the state as simple uninterpreted type. This
approach is similar to that taken by [14] to translate ASMs in SMV. However,
this translation does not preserve the original form of the rules and problems
rise when a rule consists of more than one function update, or when more rules
update the same function (location). In both cases, a complex transformation is
needed as described in detail in [14].

2.3 Implementing ASM Rules in PVS

In this section we show how the proposed algorithm has been implemented in
PVS. Keep in mind that in our encoding a state consists of a controlled part
(CTRLSTATE) modifiable by the rules, and the environment (ENV).

Rule encoding. We consider a rule as a function Ri(current,intCtrl). The

first argument current: STATE represents the current state s (controlled part
and environment), and it is used to evaluate guards (step 1 of the algorithm)
and compute terms (step 2 of the algorithm). The second argument intCtrl:
CTRLSTATE represents an intermediate controlled state which is the result of
applying the previous i−1 rules, and is used to compute those updates extending
sets (step 3 of the algorithm). Ri(current,intCtrl) yields a new controlled
(intermediate) state.
In PVS a rule is defined as Ri:[STATE,CTRLSTATE -> CTRLSTATE].

Computation of the next state by composition of rules. To compute the

next state of a current one by application of a set of rules, we define the function
next as explained below.
 Encoding Abstract State Machines in PVS 311

Let R1 , R2 , . . . , Rn be a set of rules, and s the current state (its controlled

part be cs). The controlled part cs0 of the next state s0 is inductively defined as
follows:

1. cs0 = cs
2. for i = 1, . . . , n : csi = Ri (s, csi−1 )
3. cs’ = csn

R1 R2 Rn

s cs 1 cs 2 cs n-1 cs’=cs n

This algorithm defines only how to compute the controlled part of s0 since
the rules do not change the monitored part (i.e. the environment).
The set of all ASM rules is encoded as a list rules of functions from state
and controlled state to a controlled state:

rules: list[[STATE,CTRLSTATE->CTRLSTATE]]

The application of the list rules using as current state s0 and as intermediate
controlled state cs i, is given by the following recursive definition2 :

apply(s0,cs i,rules) : recursive CTRLSTATE =

if null?(rules) then cs i
else apply(s0,car(rules)(s0,cs i),cdr(rules))
endif measure length(rules)

The controlled part of the next state of s is defined applying the list rules of
all ASM rules, and taking s as initial current state and its controlled part as
initial intermediate controlled state:

nextCtrlState(s:STATE): CTRLSTATE = apply(s,s‘ctrl,rules)

The monitored part (the environment) of the next state of s is defined by

the following function returning a random value for ENV:

nextEnv(s:STATE) : ENV

The next state of s is the composition of the two next (controlled and mo-
nitored) parts:

next(s:STATE): STATE =
(#env:= nextEnv(s), ctrl:= nextCtrlState(s) #)

2
car and cdr are built-in PVS functions yielding head and tail of a list, respectively.
312 A. Gargantini and E. Riccobene

Templates for rules encoding. We report here the two templates for rules
encoding. They distinguish between rules with and without extending updates.
Let Ri ≡ if condi then updatei be a transition rule.
1 If update i is a sequence of function updates of the form f (t1 , . . . , tn ) := t, then
Ri is translated in PVS as follows:
Ri(current, intCtrl) : CTRLSTATE =
IF cond i(current) THEN intCtrl
WITH [f:=f(intCtrl) WITH [(t1,...,tn):=t]]
ELSE intCtrl ENDIF
If the guard cond i is true in the current state then the updated interme-
diate controlled state intCtrl is returned; otherwise the state intCtrl is
returned unchanged (all terms ti and t are computed in the current state).
2 If update i has the form extend Delta with alpha, then Ri is translated in
PVS as follows:
Ri(current, intCtrl) : CTRLSTATE =
IF cond i(current) THEN intCtrl
WITH [Delta := add(alpha,Delta(intCtrl))]
ELSE intCtrl ENDIF
If the guard cond i is true, the element alpha is added to Delta evaluated
in intCtrl; otherwise intCtrl is returned unchanged.
The following examples may help to better understand rules encoding.
Example 1. Let the ASM rules be
R1 : if x = 0 then y := 5 R2 : if y = 5 then x := 2 R3 : z := x + y
The controlled state is defined as a record of three variables:
CTRLSTATE: TYPE = [# x: int, y: int, z: int #]
The three rules are defined as follows:
R1(current, intCtrl) : CTRLSTATE =
IF x(current) = 0 THEN intCtrl WITH [y := 5]
ELSE intCtrl ENDIF
R2(current, intCtrl) : CTRLSTATE =
IF y(current) = 5 THEN intCtrl WITH [x := 2]
ELSE intCtrl ENDIF
R3(current, intCtrl) : CTRLSTATE =
intCtrl WITH [z := x(current) + y(current)]
Example 2 (with sets).
R1 : if ∆ = ∅ then extend ∆ with a
R2 : if ∆ = ∅ then extend ∆ with b
Assuming that elements of the set ∆ belong to a certain type elementType,
and a and b are two constants of that type:
elementType : TYPE
a,b : elementType
the controlled state is defined as a record containing only the set ∆:
 Encoding Abstract State Machines in PVS 313

CTRLSTATE: TYPE = [# Delta: SETOF[elementType]#]

The two rules become:
R1(current, intCtrl) : CTRLSTATE =
IF empty?(Delta(current)) THEN intCtrl
WITH [Delta := add(a,Delta(intCtrl))]
ELSE intCtrl ENDIF
R2(current, intCtrl) : CTRLSTATE =
IF empty?(Delta(current)) THEN intCtrl
WITH [Delta := add(b,Delta(intCtrl))]
ELSE intCtrl ENDIF

Non determinism. In ASM non-determinism can be expressed using the con-

structor choose which allows to fire a rule R(x) choosing randomly an x satis-
fying given conditions:
choose x in U s.t. g(x)
R(x)
In PVS choose can be encoded by a monitored function from the environment
to the subset of U consisting of elements satisfying the condition g(x).
chooseX : [ENV-> {x:U|g(x)}]
The subcase choose x in U is encoded by chooseX : [ENV->U]. The non-
determinism is captured by leaving undefined (without no specified mathemati-
cal low) the function chooseX. Therefore, for a given environment e, the value
of chooseX(e) in {x:U|g(x)} is not determined.

2.4 Validating Specifications through Simple Proofs

After having specified system universes, functions and rules, the designer should
check whether the specification is correct or not, i.e. if it meets users’ needs and
if it satisfies all the desired properties (requirements).
The first step in this direction is to check some possible behaviors of the
system as specified and compare them with the desired behaviors. This approach
is followed by ASM simulators (like ASM Workbench and ASM-Gofer). A similar
approach might be followed in our encoding, probing the specification by means
of “formal challenges”. With this term we mean putative theorems, i.e. properties
that should be true if the specification is correct. The designer should start
from formally specifying and proving very simple statements and then gradually
prove more complex properties. Only at the end he/she should try to prove the
complete requirements.
In our Example 1 a very simple property is: “if in state s x is 0, y is 1 and z
is 2, then in the next state x should be equal to 0, y to 5 and z to 1”. It can be
encoded as the following lemma:
prop1: lemma
s‘ctrl= (#x:=0,y:=1,z:=2#) => next(s) = (#x:=0,y:=5,z:=1#)
314 A. Gargantini and E. Riccobene

We have proved the lemma prop1 simply using the PVS decision procedures
of rewriting and symbolic execution. More complex and complete properties
might contain quantification on system quantities and, therefore, to be proven
they might need the use of more advanced PVS strategies.
The main goal of these short proofs is to gain a deeper understanding of
the system, to become confident in the correctness of the proposed formal de-
scription, and discover possible errors or faults as early as possible, since it is
widely acknowledged that the cost of correcting specification errors is order of
magnitudes higher in the later stages of the life cycle of the system (like during
testing or even during its normal functioning). At the end the designer should
be able to formally state and prove the actual requirements. The verification of
system requirements is the subject of the following sections.

2.5 Using Induction to Prove Invariants

In mathematics induction is widely used to prove properties that hold for every
natural number. In formal models that describe system behavior as a sequence
of states (as the ASM approach does), the same scheme can be used to prove
system invariants, i.e. properties holding in every state. In this case induction is
based on the following theorem:
Theorem 1. Let S0 be the set of all initial states and P (s) a property of the
state s. If
(i.) P (s0 ) holds ∀s0 ∈ S0 ;
(ii.) P (s) → P (s0 ), ∀s, s0 states such that s0 = next(s)
then P is an “invariant”.
In our encoding we have defined and proved theorem 1 as follows:
induction: THEOREM
(forall(s:STATE): P(s)=> P(next(s)) and
(forall(s:(init)): P(s)) implies INV(P)
where (init) denotes the set of initial states and P the property to prove as
invariant (INV(P) means that P is true in the initial state and in every reachable
state).
This theorem, along with an “ad hoc” strategy that we have defined in the
file pvs-strategies, provides an induction scheme that can be used to prove
invariants in ASM models.

2.6 Trace Properties

A trace is an (infinite) sequence s0 , s1 , . . . , sn , . . . of states – any state si should
be considered as compound of the controlled part and the environment – satis-
fying the property that: (a) s0 is a valid initial state, and (b) for every pair of
subsequent states si , si+1 it holds that si+1 is the next state of si .
In our encoding a trace is formalized as a sequence of states satisfying the
property of being a trace:
 Encoding Abstract State Machines in PVS 315

trace: TYPE = {x : sequence[STATE] | member(first(x),init)

and forall n: nth(x,n+1) = next(nth(x,n))}
Trace properties are properties on traces, i.e. properties which are expressed
and proved in terms of traces. The two most common types of trace properties
are:

– properties holding in every state of every trace (always), or

– properties holding at least in one state in every trace (eventually).

We can express in PVS that a property stateProp holds in every state of the
trace t by the following predicate:
always(t,stateProp): bool = FORALL n: stateProp(nth(t,n))
We have proved the equivalence between this approach based on traces and
that based on invariants by proving the following lemma
Lemma 1. “stateProp” is an invariant iff it always holds in every trace of the
system.
In PVS: equivalence: LEMMA
INV(stateProp) <=> forall t: always(t,stateProp)

We express in PVS that a property reachProp holds in a state of the trace t by

the following predicate:

eventually(t,reachProp): bool = EXISTS n: reachProp(nth(t,n))

The always property is normally used to express “safety” requirements (“no-

thing bad will never occur”): properties that must be true in every state. The
eventually property normally expresses “liveness” properties (“something good
will eventually happen”) modeling requirements that must be eventually true.

3 A Case Study: Production Cell

In this section we explain this novel use of PVS as tool support for ASMs using
as case study the Production Cell model given in [4]. The main purpose of
this section is not to show that the Production Cell specification of Börger and
Mearelli satisfies safety and liveness properties. That has been already proved
in [14] by means of the model checker SMV. We simply like to show, through a
concrete example, how to apply our method of encoding ASM specifications in
PVS, and how to mechanize proofs.

3.1 A Brief Introduction of the Production Cell Case Study

The production cell control problem was posed in [11] as case study derived
from “an actual industrial installation in a metal-processing plant in Karlsruhe”
to obtain a “realistic, comparative survey” for testing “the usefulness of formal
316 A. Gargantini and E. Riccobene

methods for critical software systems and to prove their applicability to real-
world examples” [11]. Börger and Mearelli propose a solution of the production
cell control problem in [4], and show how to integrate the use of ASMs into a
complete software development life cycle.
. . . the production cell is composed of two conveyor belts, a positioning
table, a two-armed robot, a press, and a traveling crane. Metal plates
inserted in the cell via the feed belt are moved to the press. There, they
are forged and then brought out of the cell via the other belt and the
crane. [11]
The system is specified “as a distributed ASM with six modules, one for the
agents” – the Feed Belt, the Robot, the Press, the Deposit Belt, the Traveling
Crane, the Elevating Rotary Table – “composing the production cell, and working
together concurrently where each of the component ASMs follows its own clock.
Each of the agents represents a sequential process which can execute its rules as
soon as they become enabled. The sequential control of each agent is formalized
using a function currPhase: Agent → Phase which yields at each moment the
current phase of the agent”[4]. We refer the reader to [4] for further details.

3.2 The PVS Specification

For convenience, we report below the signature and the module of the Ground-
CELL Program for the Feed Belt (FB).
Monitored function: PieceInFeedBeltLightBarrier
Shared function: TableLoaded (between FB and the elavating rotary table ERT)
Derived functions:
TableInLoadPosition ≡ currPhase(ERT) = StoppedInLoadPosition
TableReadyForLoading ≡ TableInLoadPosition and not TableLoaded
Controlled functions: FeedBeltFree,
currPhase(FB) ∈ {NormalRun,Stopped,CriticalRun}.
Module:
FB NORMAL.
if currPhase = NormalRun and PieceInFeedBeltLightBarrier
then FeedBeltFree:= True
if TableReadyForLoading then currPhase:= CriticalRun
else currPhase:= Stopped
FB STOPPED.
if currPhase = Stopped and TableReadyForLoading
then currPhase:= CriticalRun

FB CRITICAL.
if currPhase = CriticalRun and PieceInFeedBeltLightBarrier
then currPhase:= NormalRun
TableLoaded:= True
Initialization: currPhase = NormalRun, FeedBeltFree = True,
PieceInFeedBeltLightBarrier = False
 Encoding Abstract State Machines in PVS 317

We now report the PVS encoding of the Feed Belt specification. To describe
the Feed Belt we define a type containing all the possible values of the feed belt
phase:
FBPhase : TYPE = {NormalRun,Stopped,CriticalRun}
The controlled function FeedBeltPhase is then included as component of the
record CTRLSTATE which represents the controlled part (controlled and shared
functions) of the (global) state:
CTRLSTATE : TYPE =
[# FeedBeltPhase : FBPhase,
FeedBeltFree : bool, % controlled by the FB
TableLoaded : bool, % controlled by the FB and ERT
...#]
The dots are replaced by the controlled part of the other five agents, that we
skip for the sake of conciseness. Monitored variable is defined as function from
the environment to its domain as:
PieceInFeedBeltLightBarrier : [ENV->bool]
Starting from definitions of monitored and controlled functions, derived functions
are defined as:
TableInLoadPosition(s:CTRLSTATE) : bool =
ERTPhase(s) = StoppedInLoadPosition
TableReadyForLoading(s:CTRLSTATE) : bool =
TableInLoadPosition(s) and not TableLoaded(s)
The initial state is modeled by a predicate over the states (we report only the
part concerning FB):
init(s:STATE) : bool =
FeedBeltPhase(s)= NormalRun and FeedBeltFree(s)
and not PieceInFeedBeltLightBarrier(s) ...
For the rules we report only the example of the FB NORMAL rule:
FB NORMAL(current,intCtrl): CTRLSTATE =
if FeedBeltPhase(current) = NormalRun and
PieceInFeedBeltLightBarrier(current)
then intCtrl with
[FeedBeltFree := true,
FeedBeltPhase := if TableReadyForLoading(current)
then CriticalRun
else Stopped
endif]
else intCtrl
endif
318 A. Gargantini and E. Riccobene

3.3 The Safety Properties

Using PVS we have proved all the safety properties (for the Feed Belt, the Robot,
the Press, the Deposit Belt, and the Traveling Crane) of the Production Cell
as given in [4]. For some of these properties, the encoding in PVS and the
proof are straightforward. Others require some user effort and skill. In order to
discuss the degree of interaction necessary for proving in PVS the requirements
of the Production Cell case study, we present some selected examples of proved
properties having different degree of complexity.
The Feed Belt Safety Property: the feed belt does not put metal blanks
on the table if the latter is already loaded or not stopped in loading position. It
has been quickly encoded considering that the feed belt puts metal blanks only
when it is in the CriticalRun phase3 :

FeedBeltSafety: theorem
FeedBeltPhase(s) = CriticalRun
=> ElevatingRotaryTablePhase(s) = StoppedInLoadPosition
and not TableLoaded(s)

The proof of this property, reported below, is immediate (as also its hand proof
in [4]):

(”” (ASM-INDUCT)
((”1” (COMPUTE-NEXT) (GRIND))
(”2” (TYPEPRED ”is!1”) (GRIND))))

(ASM-INDUCT) applies the induction theorem presented in Section 2.5. By

induction the proof is split in two parts: the induction step (proved by the
branch ”1”) and the initial state (branch ”2”). (COMPUTE-NEXT) is a strategy
defined in our encoding and expands the definitions of the next state and the
rules. (GRIND) is a PVS command that rewrites remaining definitions, splits the
cases and applies the decision procedures of PVS. (TYPEPRED ”is!1”) recalls
the type definition of the initial state ”is!1”. The last (GRIND) expands the
definition of initial state and applies the PVS decision procedures.
This proof case shows that the user effort to write properties and obtain
relative proofs might be very low. According to our experience, all simplest pro-
perties whose proof do not involve assumptions about the environment, require
a minimal user interaction and their proofs can be performed using induction,
case splitting, decision procedures, and rewriting rules. However, many other pro-
perties have to be proved taking in consideration the interaction of the system
with the environment. In these cases, proving properties might require greater
user effort and skill. As example, consider the first part of the Press Safety
Property 1: the press is not moved downward if it is in its bottom position. Its
PVS encoding is straightforward:
3
In this lemma and in the following ones, s must be considered as universally quan-
tified over the set of all reachable states from an initial state.
 Encoding Abstract State Machines in PVS 319

PressSafety1a: theorem
PressBottomPosition(s) =>
not PressMotorDown(Press(next(s)))
To prove this property we have to introduce an assumption about the moni-
tored function BottomPosition asserting that if the press is closed for forging,
then it is not already in the bottom position:
notBottom: axiom
PressPhase(s)= ClosedForForging => not BottomPosition(s)
This is an obvious implication considering how the system works, but we have
to explicitly state that by means of an axiom. Upon introducing the notBottom
axiom, the proof of the Press Safety Property 1 is obtained applying induction,
expanding the definitions and applying the PVS decision procedures. For other
properties we introduce similar assumptions by means of axioms, and recall these
axioms during proofs. These assumptions often concern the correct behavior of
the sensors, and sometimes are missing in the original description because im-
plicitly assumed. This again shows that automatic support may help to uncover
errors forcing the designer to precisely introduce every assumption. We also note
that these assumptions are introduced by means of logical statements, similar
to those given in [4], while a model checker would require us to express them in
an operational way.
Another example we like to report here is the Press Safety Property 2:
The press does only close when no robot arm is positioned inside it. In order
to encode it in a concise form, we introduce two (derived) boolean functions:
PressIsClosing and ArmInPress, defined as
PressIsClosing(s): bool = PressMot(PressPhase(s)) = up
i.e. the press is closing only when its motor is going up, and
ArmInPress(s): bool =
Arm1Ext(s) > 0 and Angle(s) = Arm1ToPress or
Arm2Ext(s) > 0 and Angle(s) = Arm2ToPress
Using these definitions the Press Safety property becomes:
PressSafety2: theorem
PressIsClosing(s) => not ArmInPress(s)
To prove this property we model the angle of the robot arm by the monitored
variable Angle encoded as function on the environment:
Angle :[ENV-> real]
Then we formalize all the assumptions about the movement of the robot: how
the robot rotates, how the angle changes, and how the sensors communicate to
the robot when to stop. We prove the Press Safety Property 2 using induction,
case analysis (considering all the possible robot phases in the current and in the
next state), recalling the assumptions about the robot movement, and applying
the automatic decision procedures of PVS.
320 A. Gargantini and E. Riccobene

3.4 The Liveness Property

We have proved the liveness property as well, i.e. that the system never goes in
deadlock. We have stated it as “every component in the system will eventually
change its state”. This statement is weaker than the property proved (by hand) in
[4]. Börger and Mearelli establish also a performance valuation about the number
of pieces that the cell is able to process cyclically. The proof of this performance
property is under investigation. Our property is similar to the “progress agent
property” stated in [4].
In order to prove liveness, we model and specify some assumptions about
the environment (those called “Cell Assumption” in [4]). For example we as-
sume that every object on the feed belt keeps moving until eventually it arrives
at its destination and the monitored function PieceInFeedBeltLightBarrier
becomes true:
FBAssumption: axiom
FeedBeltPhase(s) = NormalRun =>
exists(ns:(followingOrNow(s))):PieceInFeedBeltLightBarrier(ns)
followingOrNow(s) is the set of all states obtained repeatedly applying the
function next to the state s, and of the state s itself.
We make similar assumptions for every state of every component of the cell,
thus we assume that every component keeps moving till the monitored function of
interest (which selected on the basis of the state and the component) eventually
changes its value. Starting from these assumptions we prove, for example, the
liveness of the FeedBelt:
FeedBeltProgress: lemma
exists (ns: (following(s))):
not FeedBeltPhase(ns) = FeedBeltPhase(s)
Using the same approach we are able to prove similar properties for every agent.

4 Related Work and Conclusions

Several attempts of applying both theorem provers and model checkers to ASM
models have been performed. In [12] the KIV (Karlsruhe Interactive Verifier)
system has been used to mechanically verify the proof of correctness of the Pro-
log to WAM transformation. PVS has been used in [7,8] to perform mechanical
verification of the correctness of back-end rewrite system (BURS) specifications.
A model checker approach is reported in [14] where correctness of the Produc-
tion Cell specification of Börger and Mearelli has been proved through SMV.
Recently, an interface from the ASM Workbench to the SMV model checking
tool, based on an ASM-to-SMV transformation, has been presented in [6]. A di-
rect comparison of our approach can be done with the work of Dold et al. using
PVS and with the Winter’s work using SMV.
Along our presentation, especially in remarks 1,2 and 3, we have discussed
differences between Dold et al.’s approach and ours. We provide a more natural
 Encoding Abstract State Machines in PVS 321

translation of ASMs in PVS keeping the set of transition rules as a set of different
rules, instead of forcing the user to define an equivalent transformation function
in terms of one meta rule. We also provide the user with useful templates to guide
his/her formalization of ASMs in PVS. These templates have also allowed us to
provide a framework for a tool to automatically translate ASM specifications
in PVS. This tool is under development and we plan to integrate it into the
ASM-Workbench system. In addition we present proof schemes to encode and
prove invariants and properties on traces.
Although the approach based on the model checker SMV allows properties
verification in a completely automatic manner (unless the well known state ex-
plosion problem), the advantages of using our approach regard both the speci-
fication and the verification phase. PVS can easily manage specifications with
infinite sets and an unbounded number of agents, has a powerful language (to
represent functions, sets, lists and so on), has a strong type system, and can
use the usual logical constructs (like universal and existential quantifications).
Proof can be performed almost automatically in the simplest cases (as shown in
the Production Cell case study). For more complex properties, in any case, our
encoding can be used to check proofs done by hand or to support the user during
the proof in an interactive way. In connection with the results presented in [14],
we like to remark that the model checking approach can deal only with a finite
set of agents and each agent having a finite number of possible states. This is the
case of the Production Cell, under the assumption that continuous intervals (for
example, the robot angle values) can be treated as finite sets of discrete values.
This assumption was indeed used by Winter in [14], while it is not necessary in
our approach, since we are able to deal with infinite sets (for example, we treat
the robot angle as a real number). The correctness proof (with its results) of
the Production Cell specification as it is shown in [14] has to be related to the
added formalization of the environmental behavior. It is a mayor benefit of our
approach that the assumptions regarding the interaction of the system and the
environment can be formalized in a logical than in an operational way (i.e. in
terms of transition rules) as required in [14].
Concluding, we like to stress our confidence that the proposed PVS encoding
also works well for multi-agent ASMs with an unlimited number of agents, which
are very complex to treat. It is not so hard to imagine how difficult can be
performing mechanized proof verification of properties regarding interleaving
computations of agents. The case study we present here is an example of a
multi-agent system, but with a limited number of agents. However, the method
has been successfully applied to analyze properties of crypto-protocols, where
an unlimited number of agents run simultaneously. Security and authentication
properties of the ASM specification presented in [1] have been proved in PVS
using the technique of invariants. We have voluntarily left the presentation of
these results out because they would require a specific treatment.

Acknowledgments. We kindly like to thank Egon Börger for his useful advice.
We also thank anonymous referees for their helpful suggestions.
322 A. Gargantini and E. Riccobene

References
1. G. Bella and E. Riccobene. A Realistic Environment for Crypto-Protocol Analyses
by ASMs. In Proceedings of the 28th Annual Conference of the German Society of
Computer Science. Technical Report, Magdeburg University, 1998.
2. E. Börger. Why Use Evolving Algebras for Hardware and Software Engineering? In
M. Bartosek, J. Staudek, and J. Wiederman, editors, Proceedings of SOFSEM’95,
22nd Seminar on Current Trends in Theory and Practice of Informatics, volume
1012 of LNCS, pages 236–271. Springer, 1995.
3. E. Börger. High level system design and analysis using abstract state machines.
In D. Hutter, W. Stephan, P. Traverso, and M. Ullmann, editors, Current Trends
in Applied Formal Methods (FM-Trends 98), number 1641 in LNCS, pages 1–43.
Springer-Verlag, 1999.
4. E. Börger and L. Mearelli. Integrating ASMs into the Software Development Life
Cycle. Journal of Universal Computer Science, 3(5):603–665, 1997.
5. G. Del Castillo. The ASM Workbench: an Open and Extensible Tool Environment
for Abstract State Machines. In Proceedings of the 28th Annual Conference of the
German Society of Computer Science. Technical Report, Magdeburg University,
1998.
6. G. Del Castillo and K. Winter. Model Checking Support for the ASM High-Level
Language. Technical Report TR-RI-99-209, Universität-GH Paderborn, June 1999.
7. A. Dold, T. Gaul, V. Vialard, and W. Zimmerman. ASM-Based Mechanized Veri-
fication of Compiler Back-Ends. In Proceedings of the 28th Annual Conference of
the German Society of Computer Science. Technical Report, Magdeburg Univer-
sity, 1998.
8. Axel Dold. A formal representation of abstract state machines using pvs. Technical
Report Verifix Report Ulm/6.2, Universitat Ulm, July 1998.
9. Y. Gurevich. Evolving Algebras 1993: Lipari Guide. In E. Börger, editor, Specifi-
cation and Validation Methods, pages 9–36. Oxford University Press, 1995.
10. C. Heitmeyer. On the Need for Parctical Formal Methods. In Proceedings of
FTRTFT’98, 5th Intern. Symposium Real-Time Fault-Tolerant Systems, volume
1486 of LNCS, pages 18–26. Springer, 1998.
11. C. Lewerentz and T. Linder, editors. Formal Development of Reactive Systems. A
Case Study “Production Cell”. Number 891 in LNCS. Springer, 1995.
12. G. Schellhorn and W. Ahrendt. Reasoning about Abstract State Machines: The
WAM Case Study. Journal of Universal Computer Science, 3(4):377–413, 1997.
13. N. Shankar, S. Owre, and J. Rushby. The PVS proof checker: A reference manual.
Technical report, Computer Science Lab., SRI Intl., Menlo Park, CA, 1993.
14. K. Winter. Model Checking for Abstract State Machines. Journal of Universal
Computer Science, 3(5):689–701, 1997.
 Model Checking Abstract State Machines
and Beyond

Marc Spielmann

Mathematische Grundlagen der Informatik,

RWTH Aachen, D-52056 Aachen, Germany
[email protected]

Abstract. We propose a systematic investigation of the (semi-) auto-

matic verifiability of ASMs. As a first step, we put forward two verifi-
cation problems concerning the correctness of ASMs and investigate the
decidability and complexity of both problems.

1 Introduction
Abstract state machines (ASMs) [14,15,16] have become the formal foundation of
a successful methodology for specification and verification of complex hardware
and software systems. This is particularly witnessed by numerous publications
using the ASM formalism for rigorous mathematical correctness proofs of large-
scale applications. (See the recent bibliography [5] and the web site [24]. For
an introduction to the ASM verification method the reader is referred to [7].)
Interestingly, most of these contributions focus on manual verification, while the
number of publications where all or part of the verification process is automated
is rather small. (For exceptions see [25,26,12,6] and consult [24].) In a nutshell,
computer-aided verification of ASMs, i.e., (semi-) automatic verification of dy-
namic systems expressed in terms of ASMs, has not yet been well developed.
In view of the success of the ASM verification method in manual verification
we think there is need for a systematic investigation of the (semi-) automatic
verifiability of ASMs. The present paper can be viewed as an attempt to initiate
such an investigation.
As a first step toward a systematic investigation of the verifiability of ASMs,
we have to make precise what we actually mean by “verifying ASMs”. In its full
generality, the problem of verifying ASMs can be seen as a decision problem of
the following kind:
Given an ASM M (i.e., a formal description of some dynamic system)
and a specification ϕ (i.e., a formal description of a desirable property of
the system), decide whether M ‘satisfies’ ϕ.
One of our main goals in this paper is to identify decision problems of the above
kind such that solving these problems coincides with proving properties of ASMs.
We put forward two such problems, which we call the model-checking problem
and the verification problem for ASMs.

Y. Gurevich et al. (Eds.): ASM 2000, LNCS 1912, pp. 323–340, 2000.
c Springer-Verlag Berlin Heidelberg 2000

324 M. Spielmann

The model-checking problem for ASMs, denoted mc, can be stated informally
as follows:
mc: Given an ASM M , a specification ϕ, and an input I appropriate for
M , decide whether ϕ holds during all possible computations of M on I.
Note that, in general, mc cannot be solved by means of testing, i.e., by simply
running a given ASM M on a given input I and observing whether a given
specification ϕ is satisfied. There are two reasons for this. Firstly, M may be
non-deterministic or may access external functions or relations. For instance, if
M chooses a natural number n in the first step and then proceeds depending on
the choice of n, one would have to run M for each of the infinitely many possible
choices for n. Secondly, even if M is deterministic and does not interact with its
environment, it may not halt on I.
The above formulation of mc immediately raises the question of a specifica-
tion language for ASMs, i.e., a formal language suitable to express properties
of ASMs. Since in the literature there is no consensus on the choice of such a
language, we advocate here first-order branching temporal logic (FBTL) as spe-
cification language for ASMs. (FBTL is a straightforward extension of the well-
known propositional branching-time logic CTL∗ by first-order reasoning [11].
For details see the next section.) Another open issue is the notion of input. The
question is which type of input is suitable for ASMs and how does the initial
state of an ASM on some input look like? Concerning this question we follow
[2,4,13,3] and consider ASMs whose inputs are finite structures. To not impose
unnecessary restrictions on the initial states of ASMs, we associate with every
ASM M a function that maps every input I appropriate for M to an initial state
of M on I. In principle, this function can be any mapping from inputs to states.
An investigation of mc is motivated mainly by applications where it suffices
to ensure correctness of an ASM for only a small number of inputs. As an ex-
ample, consider an ASM M running on the notebook computer of a salesperson.
M receives as input a database D, say, the catalog of a supplier stored on a
CD-ROM, and interacts with the salesperson via some external relations. The
salesperson may inquire about the availability and prices of certain products
or may store customer orders in a dynamic relation of M . Since in this scena-
rio M runs on the given database D only, we can check whether M satisfies a
specification ϕ by deciding whether (M, ϕ, D) is a positive instance of mc.
Although a solution of mc is interesting for certain applications, there are
many applications where correctness of an ASM must be guaranteed for all ad-
missible inputs, of which there are often infinitely many. For instance, a compiler
is usually supposed to correctly translate an input program of arbitrary length.
In such cases one has to to check whether for a given ASM M and a given spe-
cification ϕ, (M, ϕ, I) is a positive instance of mc for every admissible input I.
This problem, which we call the verification problem for ASMs and denote by
verify, can be stated informally as follows:
verify: Given an ASM M and a specification ϕ, decide whether for every
input I appropriate for M , ϕ holds during all possible computations of
M on I.
 Model Checking Abstract State Machines and Beyond 325

Not surprisingly, both mc and verify are—in their full generality—unde-

cidable, as ASMs are computationally complete. (Indeed, every Turing machine
can be regarded as a particularly simple ASM. It is an easy exercise to define a
reduction of the halting problem for Turing machines to both problems.) The-
refore, we investigate the decidability of mc and verify with respect to classes
of restricted ASMs. The idea is to formulate conditions on ASMs such that mc
(resp. verify) becomes decidable for ASMs that satisfy these conditions. Once
decidability with respect to a class of ASMs is established, all ASMs in this class
are automatically verifiable in the sense that there exists a procedure which
checks whether a given ASM satisfies a given FBTL specification on a particular
input (resp. on all inputs).
It is worth noticing that for unrestricted ASMs, mc and verify collapse to
the same problem. For example, in order to decide whether (M, ϕ) is a positive
instance of verify, it suffices to decide whether (M 0 , ϕ, I) is a positive instance
of mc, where M 0 ignores I, non-deterministically chooses some other input I 0 (of
arbitrary size!), and then simulates M on I 0 . However, mc and verify do not
necessarily coincide for restricted ASMs. Consider, for example, ASMs defined
by means of a single rule of the form (if γ then accept), where the guard γ is a
first-order sentence. Checking whether such an ASM accepts a particular input
structure can be done in Pspace, whereas verifying that it accepts all input
structures is undecidable by Trakhtenbrot’s theorem [10].
Although deciding mc appears to be easier than deciding verify, it is a
priori not clear whether a solution of verify (for a class of ASMs) implies a
solution of mc (for the same class of ASMs). The question is whether mc is a
subproblem of verify in the sense that there exists a reduction of mc to verify
with respect to fixed classes of ASMs. We give some formal evidence that this
is indeed the case, thereby also providing a justification for the title of the paper.

Outline. In the next section, we recall first-order branching temporal logic which
will serve here as specification language for ASMs. In Section 3, we formally
define the model-checking problem for ASMs and present some results concerning
its decidability and complexity (with respect to classes of restricted ASMs). In
Section 4, we define the verification problem for ASMs, show that it subsumes
the model checking problem, and study its decidability and complexity.

2 A Specification Language for Abstract State Machines

In order to investigate the automatic verifiability of ASMs we need to fix a

formal language suitable to express properties of ASMs. We propose to specify
properties of ASMs in terms of first-order branching temporal logic (FBTL).
Formally, FBTL formulas express properties of computation graphs of ASMs.
These graphs can be viewed as collections of runs of ASMs and are introduced
next.
326 M. Spielmann

2.1 Computation Graphs

We assume the reader to be familiar with (sequential) ASMs [15,14]. To simplify
the definition of computation graphs, we first introduce some additional termi-
nology and notation.
ASM vocabularies. An ASM vocabulary Υ is a quadruple (Υin , Υstat , Υdyn , Υext )
of finite, pairwise disjoint vocabularies. The symbols in Υin , Υstat , Υdyn , and Υext
are called input, static, dynamic, and external symbols, respectively. Sometimes
we also denote by Υ the (ordinary) vocabulary Υin ∪ Υstat ∪ Υdyn ∪ Υext . The
intended meaning will be clear from the context. For technical reasons we shall
always assume that Υin contains the constant symbol 0.
ASM programs. Let Υ = (Υin , Υstat , Υdyn , Υext ) be an ASM vocabulary. An
ASM program Π over Υ is a rule in the sense of [15], each of whose static
(resp. dynamic, external) symbols occurs in Υin ∪ Υstat (resp. Υdyn , Υext ). Intui-
tively, the symbols in Υstat denote those relations and functions which are static
and not included in the input of an ASM (e.g., arithmetical operations).
States. A state S over an ASM vocabulary Υ is a structure over the (ordinary)
vocabulary Υ . We consider both finite and infinite states of ASMs.
Transitions. Let Π be an ASM program over Υ . The one-step semantics of Π
defines a binary transition relation T(Υ,Π) between states over Υ as usual. That
is, (S, S 0 ) ∈ T(Υ,Π) iff
1. S and S 0 are states over Υ , and
2. S 0 is the successor state (or sequel) of S with respect to Π in the sense of
[15].

Remark 1. (a) If Π contains choose, import, or external symbols, then T(Υ,Π)

is in general non-deterministic, i.e., there may exist states S, S 0 , S 00 with (S, S 0 ),
(S, S 00 ) ∈ T(Υ,Π) and S 0 6= S 00 .
(b) If S is finite and Π contains import, then it may happen that Π attempts
to import a new element, although the reserve of S is empty. One possible
solution to this problem is to modify the semantics of import so that in states
with empty reserve it always ‘imports’ the element denoted by 0. t
u
Inputs. An input I over an ASM vocabulary Υ is a finite structure over Υin .
Initialization mappings. An initialization mapping initial over an ASM vo-
cabulary Υ is a function which maps every input I over Υ to a state SI over Υ
satisfying the following three conditions:
1. The universe of I is a subset of the universe of SI .
2. For every relation symbol R ∈ Υin , every k-ary function symbol f ∈ Υin , and
every k-tuple ā of elements in SI ,
 I
f (ā) if ā consists of elements in I
RSI = RI and f SI (ā) =
0I otherwise.
3. For every input I 0 over Υ , if I 0 and I are isomorphic, so are SI 0 and SI .
 Model Checking Abstract State Machines and Beyond 327

Abstract state machines. An abstract state machine M is a triple (Υ, initial ,

Π) consisting of
– an ASM vocabulary Υ ,
– an initialization mapping initial over Υ , and
– an ASM program Π over Υ without free variables.
Υin , Υstat , Υdyn , and Υext are called the input, static, dynamic, and external vo-
cabulary of M , respectively. An input I appropriate for M is an input over Υ .
M is deterministic if Π does not contain choose.
We can now define the notion of a computation graph of an ASM.
Definition 2. Let M = (Υ, initial , Π) be an ASM and I an input appro-
priate for M . The computation graph of M on I, denoted CM (I), is a triple
(States, Trans, S0 ) where
– States is the set of those states over Υ whose universe is identical with the
universe of initial (I),
– Trans ⊆ States × States is the restriction of T(Υ,Π) to States, and
– S0 := initial (I).
A run of M on I is an infinite path in CM (I) starting at S0 . t
u
Notice that for every ASM M = (Υ, initial , Π) and for every input I appro-
priate for M , CM (I) is finite iff initial (I) is finite.
Example 3. In this and the next example (Example 5) we show that classical
model checking, i.e., checking whether a given finite-state system is a model of a
given temporal formula (see, e.g., [8]), is a special case of model checking ASMs.
To this end, we display below a simple ASM M whose computation graphs are
finite-state systems.
In classical model checking, finite-state systems are often represented as
Kripke structures. A Kripke structure K is a tuple

(S, T, s0 , P1 , . . . , Pk ) (1)

consisting of a finite set S of states, a binary transition relation T ⊆ S × S,

an initial state s0 ∈ S, and a sequence P1 , . . . , Pk of subsets of S. (Sometimes,
the sequence P1 , . . . , Pk is also given in the form of a labeling function L : {p1 ,
. . . , pk } → 2S with L(pi ) = Pi for each i ∈ {1, . . . , k}.) Since it is customary
in the context of model checking, we assume that every s ∈ S has at least one
T -successor, which may be s itself. K− := (S, T, s0 ) is called the transition graph
of K.
We define M = (Υ, initial , Π) so that for every Kripke structure K, the
computation graph of M on input K− is isomorphic to K− . The vocabulary
of M is defined as follows: Υin := {T, 0}, where T is a binary relation symbol
(which will be interpreted as the edge relation of the input transition graph),
Υdyn := {pebble}, where pebble is a nullary function symbol, Υstat := ∅, and
Υext := ∅. The program of M is:
328 M. Spielmann

program Π:
choose x : T (pebble, x)
pebble := x
Observe that the transition graph K− of any Kripke structure K is a finite
structure over Υin , and thus an input over Υ . The initialization mapping of M
− −
is defined by initial (K− ) := (K− , 0K ), where (K− , 0K ) denotes the state over
Υ in which T and 0 are interpreted as in K− , and the nullary dynamic function
−
symbol pebble is interpreted as 0K . One can now verify that S 7→ pebble S is an
isomorphism between CM (K− ) and K− . t
u

2.2 First-Order Branching Temporal Logic

Recall that all runs of an ASM M on a particular input I are embedded in the
computation graph of M on I. Hence, it is reasonable to express a property of
M (on any input) as a property of all computation graphs of M . To express pro-
perties of computation graphs we propose first-order branching temporal logic
(FBTL), which is a combination of first-order logic and the well-known proposi-
tional branching temporal logic CTL∗ [8,11]. On the one hand, first-order logic
allows us to reason about the states of an ASM. (Recall that the states of an
ASM are first-order structures.) On the other hand, CTL∗ enables us to reason
about the temporal and non-deterministic behavior of an ASM. For instance, one
can express that ‘good things eventually happen’ or that ‘there exists a good
run’ of an ASM. For the reader’s convenience we recall the definition of FBTL
here.

Definition 4. State formulas and path formulas of first-order branching tempo-

ral logic are defined by simultaneous induction:
(S1) Every atomic (first-order) formula is a state formula.
(S2) If ϕ is a path formula, then Eϕ is a state formula.
(P1) Every state formula is a path formula.
(P2) If ϕ and ψ are path formulas, then Xϕ, ϕUψ, and ϕBψ are path for-
mulas.
(SP1) If ϕ and ψ are state (resp. path) formulas, then ϕ ∨ ψ and ¬ϕ are state
(resp. path) formulas.
(SP2) If x is a variable and ϕ a state (resp. path) formula, then ∃xϕ is a state
(resp. path) formula.
The free and bound variables of state and path formulas are defined in the obvious
way. FBTL denotes the set of state formulas. t
u

Semantics of FBTL formulas. Intuitively, a state formula of the form Eϕ

expresses that there (E)xists an infinite path (presumably a run of an ASM)
such that the path formula ϕ holds along this path. The intuitive meaning of
path formulas of the form Xϕ, ϕUψ, and ϕBψ is as follows:
– Xϕ: ϕ holds in the ne(X)t state.
 Model Checking Abstract State Machines and Beyond 329

– ϕUψ: ψ holds eventually and ϕ holds (U)ntil then.

– ϕBψ: either ψ holds always or ϕ holds (B)efore ψ fails.
We formally define the semantics of FBTL formulas only with respect to compu-
tation graphs of ASMs. Let Υ be an ASM vocabulary, ϕ a state formula over Υ ,
ψ a path formula over Υ , and C = (States, Trans, S0 ) a computation graph of an
ASM of vocabulary Υ . W.l.o.g., we may assume that free(ϕ) = free(ψ) = {x̄}.
Let ρ = (Si0 )i∈ω be an infinite path in C (not necessarily starting at the initial
0
state S0 of C). For any j ∈ ω, let ρ|j denote the infinite path (Si+j )i∈ω , i.e., the
0 0
suffix Sj , Sj+1 , . . . of ρ.
Simultaneously for every state S in C, every infinite paths ρ in C, and all
interpretations ā of the variables x̄ (chosen from the universe of S0 ) define the
two satisfactory relations (C, S, ā) |= ϕ and (C, ρ, ā) |= ψ by induction on the
construction of ϕ and ψ:

(S1) (C, S, ā) |= ϕ :⇔ S |= ϕ[ā], where ϕ is an atomic formula

(S2) (C, S, ā) |= Eϕ :⇔ there is an infinite path ρ0 in C starting at S
such that (C, ρ0 , ā) |= ϕ
(P1) (C, ρ, ā) |= ϕ :⇔ (C, S00 , ā) |= ϕ, where S00 is the first state of ρ
(P2) (C, ρ, ā) |= Xϕ :⇔ (C, ρ|1, ā) |= ϕ
(C, ρ, ā) |= ϕUψ :⇔ there exists i ∈ ω such that (C, ρ|i, ā) |= ψ and
for all j < i, (C, ρ|j, ā) |= ϕ
(C, ρ, ā) |= ϕBψ :⇔ for every i ∈ ω, if (C, ρ|i, ā) |= ¬ψ, then there
exists j < i with (C, ρ|j, ā) |= ϕ

The semantics of formulas derived by means of rule (SP1) is standard. It remains

to declare the semantics of formulas derived by means of rule (SP2). Below, σ
stands for either a state S or an infinite path ρ in C, depending on whether ϕ
is a state or a path formula.

(SP2) (C, σ, ā) |= ∃yϕ(x̄, y) :⇔ there is an element b in the universe of S0

such that (C, σ, ā, b) |= ϕ(x̄, y)

For every FBTL sentence ϕ over Υ , let

C |= ϕ :⇔ (C, S0 ) |= ϕ.

The following abbreviations are customary and will be used frequently:

– Aϕ := ¬E¬ϕ (ϕ holds along every path).
– Fϕ := true Uϕ (ϕ holds eventually).
– Gϕ := false Bϕ (ϕ holds always).
Note that ϕBψ ≡ ¬(¬ϕU¬ψ).

Example 5. Recall the ASM M in Example 3. We observe that classical model

checking is a special case of model checking M . Let Υ 0 = {P1 , . . . , Pk } be a
vocabulary containing only set symbols, and pebble a nullary function symbol
(as in the vocabulary of M ). For every CTL∗ formula ϕ over Υ 0 , obtain the
330 M. Spielmann

FBTL sentence ϕ0 over Υ 0 ∪ {pebble} by replacing in ϕ every occurrence of Pi

with the atomic formula Pi (pebble). Let MΥ 0 be defined as M , except that the
input vocabulary of MΥ 0 is {T, 0} ∪ Υ 0 and the initialization mapping of MΥ 0
maps every Kripke structure K of form (1) to the state (K, 0K ). Now verify that
for every CTL∗ formula ϕ over Υ 0 and every Kripke structure K of form (1),
K |= ϕ ⇔ CMΥ 0 (K) |= ϕ0 .
In other words, we can check whether K is a model of ϕ by deciding whether
(MΥ 0 , ϕ0 , K) is a positive instance of the model-checking problem for ASMs. In
fact, in the next section we are going to define the model-checking problem for
ASMs such that K |= ϕ iff (MΥ 0 , ϕ0 , K) ∈ mc. t
u
Examples of FBTL specifications motivated by concrete applications can be
found in [22,23].

3 Model Checking Abstract State Machines

Recall the informal formulation of the model-checking problem for ASMs in the
introduction. If the reader agrees that FBTL (or some suitable extension of this
logic) is an appropriate specification language for ASMs, then the model-checking
problem for ASMs can be rephrased as follows:
mc0 : Given an ASM M , a FBTL sentence ϕ over the vocabulary of M ,
and an input I appropriate for M , decide whether CM (I) |= ϕ.
Notice, however, that this problem is still not a computational problem. The
difficulty here is that the initialization mapping of an ASM M may not be finitely
representable, in which case M cannot serve as input to a computational device.
Next, we further restrict the above problem so that it becomes a computa-
tional problem. Then, in the second part of this section, we present some results
concerning the decidability and complexity of the obtained computational pro-
blem.

3.1 The Model-Checking Problem

The problem of representing the initialization mappings of ASMs in finite terms
disappears if we concentrate on ASMs uniformly initialized in the following sense.
Definition 6. A class C of ASMs is uniformly initialized if, whenever M and
M 0 are two ASMs in C of the same vocabulary, then M and M 0 have that same
initialization mapping. t
u
With every uniformly initialized class C of ASMs and every ASM vocabulary
Υ , we can associate an initialization mapping initial C C
Υ , such that initial Υ is the
initialization mapping of every M ∈ C of vocabulary Υ . Observe that, in the
context of a uniformly initialized class C of ASMs, every M ∈ C is uniquely
determined by its vocabulary and program. More precisely, if Υ is the vocabulary
and Π the program of M , then necessarily M = (Υ, initial C Υ , Π). This motivates
the following definition.
 Model Checking Abstract State Machines and Beyond 331

Definition 7. The standard representation of an ASM (Υ, initial , Π) is given

by the pair (Υ, Π). t
u

We are now in the position to define the model-checking problem for ASMs
as a computational problem. Let C be a uniformly initialized class of ASMs
and F a fragment of FBTL. The model checking problem for C-ASMs and F -
specifications is the following decision problem:

mc(C, F ): Given the standard representation of an ASM M ∈ C, a sen-

tence ϕ ∈ F over the vocabulary of M , and an input I appropriate for
M , decide whether CM (I) |= ϕ.

Remark 8. Note the subtlety in the above definition of mc, namely that by
viewing an ASM M ∈ C as a finitely representable pair (Υ, Π) the problem of
actually representing the initialization mapping of M has not been solved, but
rather has been made part of the model-checking problem itself. t
u

Example 9. (a) For every vocabulary Υ 0 = {P1 , . . . , Pk } containing only set

symbols, and every CTL∗ formula ϕ over Υ 0 , let MΥ 0 and ϕ0 be defined as in
Example 5. Let C denote the class of all MΥ 0 , and set F = {ϕ0 : ϕ ∈ CTL∗ }. C
is uniformly initialized and we have (MΥ 0 , ϕ0 , K) ∈ mc(C, F ) iff K |= ϕ. That is,
mc(C, F ) coincides with the model checking problem for CTL∗ .
(b) One can define a class C of ASMs and a fragment F of FBTL such that
mc(C, F ) coincides with the symbolic model checking problem for CTL∗ . We
sketch the idea. Consider a Kripke structure K of form (1) and a CTL∗ formula
ϕ over {P1 , . . . , Pk }. Obtain the FBTL formula ϕ0 by replacing every occurrence
of Pi with the boolean symbol pi . (A boolean symbol is a nullary relation symbol.)
K can be represented by means of boolean formulas γT (x̄, ȳ), γP1 (x̄), . . . , γPk (x̄)
with x̄ and ȳ two n-tuples of boolean variables, free(γT ) = {x̄, ȳ}, and for each
Pi , free(γPi ) = {x̄}. For simplicity, let us assume that for each Pi , γPi (false)
evaluates to false. There is an ASM MK with input vocabulary {0}, dynamic
vocabulary {x̄, p̄}, and

program Π:
choose y1 , . . . , yn ∈ {true, false} : γT (x̄, ȳ)
x̄ := ȳ
if γP1 (ȳ) then p1 := true else p1 := false
...
if γPk (ȳ) then pk := true else pk := false

such that for any input I appropriate for MK , K |= ϕ ⇔ CMK (I) |= ϕ0 . t

u

In the remainder of this section we investigate the decidability and complexity

of mc(C, F ).
332 M. Spielmann

3.2 Decidability and Complexity

We first observe that the model-checking problem is decidable for uniformly
initialized ASMs whose initial states are finite and computable from the inputs.
Lemma 10. Let C be a uniformly initialized class of ASMs such that there
exists an algorithm AC as follows. For every ASM vocabulary Υ and every input
I over Υ , AC on input (Υ, I) outputs initial C C
Υ (I) (where initial Υ is defined as
below Definition 6). Then mc(C, FBTL) is decidable.
Proof. Given an instance (M, ϕ, I) of mc(C, FBTL), where the vocabulary of M
is, say, Υ , we can run AC on (Υ, I) to obtain the initial state of M on I. Since
this state is finite, so is CM (I). Hence, we can construct CM (I) explicitly and
check whether CM (I) |= ϕ holds. t
u
Applications of this lemma can be found in [25,9,26]. Note however that the
decision procedure outlined in the proof of the lemma is often too expensive to
be useful in practice, even if the complexity of the algorithm AC is low. For
example, the number of reachable states of an ASM M employing a non-nullary
dynamic relation in general grows exponentially in the size of the initial state.
Consequently, the space required for constructing a computation graph of M
explicitly can be exponential in the size of the input, independent of AC .
Next, we present a class C of ASMs and a fragment F of FBTL for which a
useful restriction of mc(C, F ) is in Pspace.
Definition 11. An ASM M = (Υ, initial , Π) is finitely initialized if for every
input I appropriate for M ,
1. the universe of initial (I) is that of I, and
2. for every relation symbol R ∈ Υ − Υin , every k-ary function symbol f ∈
Υ − Υin , and every k-tuple ā of elements in initial (I),
Rinitial(I) = ∅ and f initial(I) (ā) = 0.
t
u
Observe that the initialization mapping of every finitely initialized ASM M
is uniquely determined by the vocabulary of M . In particular, every class of
finitely initialized ASMs is uniformly initialized and for every such class C there
exists an algorithm AC as in Lemma 10.
Remark 12. (a) On the first glance, the second condition in Definition 11 may
seem to render finitely initialized ASMs useless for practical purposes, because in
applications the initial interpretations of the symbols in Υ −Υin must often satisfy
certain conditions. In particular, it may not be desirable to have all symbols in
Υ − Υin be initialized as in the second condition of the definition. But note that
the initial interpretations of these symbols can be seen as part of the input
of an ASM. For example, consider a finitely initialized ASM M whose nullary
dynamic function pebble should assume, instead of 0, the value 42 in the initial
state. This can be achieved by modifying M as follows. Let Π be the program
of M , pebbleInitial a new constant symbol, and running a new boolean symbol.
Replace Π with the following program:
 Model Checking Abstract State Machines and Beyond 333

program Π 0 :
if ¬running then
pebble := pebbleInitial
running := true
if running then
Π
Now, while model checking M , consider only input structures in which the input
symbol pebbleInitial is interpreted as 42.
(b) Suppose that the second condition in Definition 11 is relaxed as follows:
there is a distinguished symbol S ∈ Υ − Υin whose initial interpretation depends
on Υ but is not restricted otherwise. We display a class C of simple ASMs
finitely initialized in the relaxed sense and a simple fragment F of FBTL such
that mc(C, F ) is undecidable.
W.l.o.g., we may assume that S is a boolean symbol and that every ASM
vocabulary contains S. Choose some undecidable problem
P ⊆ {Υ : Υ is an ASM vocabulary}.
(For instance, let TM Υ be the Turing machine whose encoding—in some fixed
standard binary encoding—equals the binary representation of the number of
symbols in Υ . Then P := {Υ : TM Υ halts on the empty word} is undecidable,
as a reduction of the halting problem for Turing machines shows.) For every ASM
vocabulary Υ , let MΥ := (skip, initial Υ , Υ ) be a finitely initialized ASM, except
that now for every input I appropriate for M , initial Υ (I) |= S iff Υ ∈ P . Let
C denote the class of all MΥ , and set F = {S}. For any input I appropriate for
M , Υ 7→ (MΥ , S, I) is a reduction of P to mc(C, F ). This implies that mc(C, F )
is undecidable. t
u
In favor of a succinct formulation of the next theorem we introduce some
additional notation.
Definition 13. Let T denote the closure of the set of first-order formulas under
the rules for negation and disjunction, and the following rule:
(T) If ϕ and ψ are formulas, then Xϕ, ϕUψ, and ϕBψ are formulas.
The universal closure of T, denoted UT, is the set of formulas of the form A∀x̄ϕ
with ϕ ∈ T and free(ϕ) ⊆ {x̄}, where free(ϕ) is defined in the obvious way. u t
We define a restriction of the model-checking problem for ASMs. An inve-
stigation of this restriction is motivated by the observation that the arities of
relations and functions used in practice tend to be rather small. Indeed, for
practical purposes it often suffices to solve mc for ASMs whose vocabulary con-
tains only symbols of arity ≤ m, for some a priori fixed natural number m. Let
mcm (C, F ) denote the restriction of mc(C, F ) to instances where only symbols
of arity ≤ m occur.
The following theorem generalizes a result in [22]. Note that UT can be
viewed as a fragment of FBTL.
334 M. Spielmann

Theorem 14. Let C be the class of finitely initialized ASMs whose program
does not contain import or choose. For any natural number m, mcm (C, UT) is
Pspace-complete.
We already encounter Pspace-hardness of mcm (C, UT) if m = 0 and C is the
class of finitely initialized ASMs of the form (skip, initial Υ , Υ ). This follows by
a reduction of the Pspace-complete model-checking problem for propositional
temporal logic [20,11].
As already pointed out in the introduction, a decision procedure for mc(C, F )
can be useful for the verification of ASMs that are supposed to run correctly on
a small number of inputs only. However, for applications where correctness of an
ASM has to be ensured for a large number of inputs—or even infinitely many
inputs—a solution of mc(C, F ) does not suffice. For such applications we have
to solve the model-checking problem for all admissible inputs. This will be our
main concern in the next section.

4 Verifying Abstract State Machines

We study the problem of model checking ASMs against FBTL specifications
for all inputs. After formally defining this problem, which we call the verifica-
tion problem for ASMs, we provide some evidence that it subsumes the model-
checking problem for ASMs and investigate its decidability and complexity.

4.1 The Verification Problem

Let C be a uniformly initialized class of ASMs and F a fragment of FBTL.
Verifying C-ASMs against F -specifications means solving the following decision
problem:
verify(C, F ): Given the standard representation of an ASM M ∈ C and
a sentence ϕ ∈ F over the vocabulary of M , decide whether for every
input I appropriate for M , CM (I) |= ϕ.

Remark 15. One may object that the above formulation of verify does not
adequately reflect real-life verification of ASMs, as in applications one is often
interested in verifying an ASM only for inputs that satisfy certain conditions.
Note however that these conditions can be viewed as part of a specification. For
example, suppose that the FBTL sentence ψ describes all admissible inputs of
an ASM M . That is, an input I is considered to be admissible for M iff I |= ψ.
Then (M, ψ → ϕ) is a positive instance of verify iff for every admissible input
I appropriate for M , CM (I) |= ϕ. t
u
Not surprising, verify(C, F ) is in general undecidable, even for classes C of
simple ASMs and simple fragments F of FBTL. For instance, recall the reduction
of the undecidable problem P to mc(C, F ) in Remark 12 (b) and observe that
the same reduction (with I now removed from the image of every Υ ) also reduces
P to verify(C, F ).
 Model Checking Abstract State Machines and Beyond 335

Proviso 16. In the remainder of this section we restrict ourselves to finitely

initialized ASMs (see Definition 11). t
u

4.2 Model Checking vs Verification

We show that mc(C, F ) is polynomial-time reducible to verify(C, F ) if C and F
satisfy certain closure conditions. In favor of a concise presentation of the result
we introduce a restricted type of first-order quantification, which we call witness-
bounded quantification. In the following, FO denotes first-order logic (with equa-
lity).
Witness-bounded quantification. A witness set W is a finite set of variables
and constant symbols. For aW witness set W and a variable x not in W , we
write (x ∈ W ) instead of ( ν∈W x = ν). Intuitively, (x ∈ W ) holds iff the
interpretation of x matches the interpretation of some symbol in W .
Definition 17. The witness-bound fragment of FO, denoted FOW , is obtained
from FO by replacing the rule for (unrestricted) quantification with the following
rule for witness-bounded quantification:
(WBQ) If W is a witness set, x a variable not in W , and ϕ a formula, then
(∃x ∈ W )ϕ and (∀x ∈ W )ϕ are formulas.
The free and bound variables of FOW formulas are defined as usual. In particular,
x occurs bound in (∃x ∈ W )ϕ and (∀x ∈ W )ϕ, while all variables in the witness
set W occur free. t
u
We view FOW as a fragment of FO where formulas of the form (∃x ∈ W )ϕ
and (∀x ∈ W )ϕ are mere abbreviations for ∃x(x ∈ W ∧ ϕ) and ∀x(x ∈ W → ϕ),
respectively.
Reduction of mc to verify. Proceeding toward a reduction of mc to verify,
we next define witness-bounded versions of ASMs and FBTL formulas. Let W be
a witness set containing only constant symbols. For every FBTL formula ϕ, ob-
tain ϕW from ϕ by replacing every first-order quantifier with the corresponding
W -bounded quantifier (e.g., replace ∃x with (∃x ∈ W ), and ∀x with (∀x ∈ W )).
For every ASM M = (Υ, initial , Π) with W ⊆ Υin , let M W := (Υ, initial , Π W )
where Π W is obtain by induction on the construction of Π:
– If Π is skip or an update, then Π W := Π.
– If Π = (if ϕ then Π0 ), then Π W := (if ϕW then Π0W ).
– If Π = (Π0 ||Π1 ), then Π W := (Π0W ||Π1W ).
– If Π = (do-forall x : ϕ, Π0 ), then Π W := (do-forall x ∈ W : ϕW , Π0W ).
If Π is a choose or import rule, proceed similar to the last case. For a class C
of ASMs and a fragment F of FBTL, let C W and F W be defined by:
CW := {M W : M = (Υ, initial , Π) ∈ C,
W ⊆ Υin is a set of constant symbols}
FW := W
{ϕ : ϕ ∈ F , W is a set of constant symbols}.
336 M. Spielmann

Lemma 18. Let C0 be the class of finitely initialized ASMs whose program does
not contain import and whose input (resp. external) vocabulary does not con-
tain function symbols of arity ≥ 1 (resp. ≥ 0). Furthermore, let F be a frag-
ment of FBTL containing for every boolean symbol b the formula EXb. For any
class C 0 ⊇ C0W and any fragment F 0 ⊇ F W closed under the boolean operators,
mc(C0 , F ) is polynomial-time reducible to verify(C 0 , F 0 ).

Proof. (Sketch.) Consider an instance (M, ϕ, I) of mc(C, F ) with M = (Υ,

initial Υ , Π). W.l.o.g., we may assume that every element in I is denoted by
some constant symbol in Υin . (If this is not the case, enrich Υin with new constant
symbols and add interpretations of the new symbols to I. This modification of
I is clearly polynomial-time computable.) We construct an instance (M 0 , ϕ0 ) of
verify(C W , FV 0
) which is positive iff (M, ϕ, I) is a positive instance of mc(C, F ).
Let χI := I|=γ γ where γ ranges in the set of atomic and negated atomic
sentences over Υin . χI is a quantifier-free FO sentence over Υin and can be con-
structed from I in polynomial time. For every finite structure I 0 over Υin we
have I 0 |= χI iff there exists a substructure of I 0 isomorphic to I.
To the definition of M 0 and ϕ0 . Let W be the set of constant symbols in Υin ,
and error a boolean symbol not in Υ . Set ϕ0 = (¬EXerror → ϕW ) and M 0 =
(Υ, initial Υ , Π 0 ) with Π 0 := (Π W || if ¬χI then error ). (M 0 , ϕ0 ) can be obtained
from (M, ϕ, I) in polynomial time and is an instance of verify(C W , F 0 ). t
u

Let C0 be as in the above lemma. The proof of the lemma shows that, if
C ⊆ C0 and F ⊆ FBTL have reasonable closure properties and C W ⊆ C and
F W ⊆ F , then mc(C, F ) is polynomial-time reducible to verify(C, F ).

4.3 Decidability and Complexity

The next proposition provides a sufficient condition for the decidability of the
verification problem. Let us first recall two decision problems from logic.
Finite satisfiability and finite validity. Let L be a logic, Υ a vocabulary,
and ϕ an L-sentence over Υ . ϕ is called finitely satisfiable if there exists a finite
structure A over Υ with A |= ϕ. ϕ is called finitely valid if for every finite struc-
ture A over Υ , A |= ϕ. By fin-sat(L) (resp. fin-val(L)) we denote the problem
of deciding finite satisfiability (resp. finite validity) of a given L-sentence.

Proposition 19. (a) Let C be a class of ASMs and F a fragment of FBTL. If

there exists a logic L satisfying the following two conditions, then verify(C, F )
is decidable.

1. fin-val(L) is decidable.
2. There exists a computable function which maps every instance (M, ϕ) of
verify(C, F ) to an L-sentence χM,ϕ over the input vocabulary of M , such
that for every input I appropriate for M ,

CM (I) |= ϕ ⇔ I |= χM,ϕ .
 Model Checking Abstract State Machines and Beyond 337

(b) Let ¬F denote the set of negated F -sentences. verify(C, ¬F ) is decidable if

in the first condition instead of fin-val(L), fin-sat(L) is decidable.
Proof. To assertion (a). The second condition immediately implies that (M, ϕ)
7 χM,ϕ is a reduction of verify(C, F ) to fin-val(L). The latter problem is
→
decidable by assumption. To assertion (b). Verify the following chain of equiva-
lences: (M, ¬ϕ) ∈ verify(C, ¬F ) iff CM (I) |= ¬ϕ for every I (appropriate for
M ) iff CM (I) 6|= ϕ for every I iff I 6|= χM,ϕ for every I iff χM,ϕ 6∈ fin-sat(L).
Hence, (M, ¬ϕ) 7→ χM,ϕ is a reduction of verify(C, ¬F ) to the complement of
fin-sat(L). Decidability of fin-sat(L) then implies assertion (b). t
u
Of course, the main challenge in applying Proposition 19 is to find a logic L
that satisfies the two conditions in the proposition. As possible candidates for L
we propose the following logics:
– existential transitive closure logic (E+TC),
– existential least fixed-point logic (E+LFP) [1,10], and
– SO(∃), i.e., the set of second-order formulas in prenex normal form, whose
quantifier prefix is an arbitrary string of SO quantifiers followed be a sting
of existential FO quantifiers [19].
Both fin-val and fin-sat are decidable for each of these logics if one restricts
attention to formulas over relational vocabularies [18,19].
In the remainder of this section we recall some results concerning the deci-
dability of verify from [21]. The main positive result there was obtained by
means of an application of Proposition 19.
Definition 20. Let Υ be an ASM vocabulary where Υdyn and Υext contain only
nullary symbols. Sequential nullary programs over Υ are defined inductively:
– Updates: For every relation symbol b ∈ Υdyn , every function symbol v ∈
Υdyn , and every term t over Υ , each of the following is a sequential nullary
program: b := true, b := false, v := t.
– Conditionals: If Π is a sequential nullary program and ϕ a witness-bounded
formula over Υ , then (if ϕ then Π) is a sequential nullary program.
– Parallel composition: If Π0 and Π1 are sequential nullary programs, then
(Π0 ||Π1 ) is a sequential nullary program.
– Non-deterministic choice: If Π is a sequential nullary program, x̄ a tuple
of pairwise distinct variables, and ϕ a witness-bounded formula over Υ such
that ∃x̄ϕ is finitely valid, then (choose x̄ : ϕ, Π) is a sequential nullary
program.
The free and bound variables of a sequential nullary program are defined in the
obvious way.
A sequential nullary ASM is a finitely initialized ASM whose program is a
sequential nullary program without free variables. SN-ASMrel (resp. SN-ASMfct )
denotes the class of sequential nullary ASMs whose input vocabulary contains
only relation and constant symbols (resp. contains at least two non-nullary sym-
bols, one of which is a function symbol). t
u
338 M. Spielmann

The next definition introduces two fragments of FBTL, denoted ETE and
UTU. Informally speaking, the formulas in ETE (resp. UTU) are built from ato-
mic formulas by means of disjunction, conjunction, existential (resp. universal)
quantification (applicable only to state formulas), the temporal operators X, U,
and B, and the path quantifier E (resp. A).

Definition 21. ETE denotes the set of FBTL formulas derivable by means of
rules (S1), (S2), (P1), (P2) in Definition 4, and the following two formula-for-
mation rules:
(SP1)’ If ϕ and ψ are state (resp. path) formulas, then ϕ ∨ ψ and ϕ ∧ ψ are
state (resp. path) formulas.
(SP2)’ If x is a variable and ϕ a state formula, then ∃xϕ is a state formula.
UTU denotes the set of negated ETE formulas. t
u

Let verifym (C, F ) denote the restriction of verify(C, F ) to instances where

only symbols of arity ≤ m occur. Again, solving verifym (C, F ) instead of
verify(C, F ) should be no serious obstacle for practical purposes.

Theorem 22 ([21]). For any natural number m, the problem

verifym (SN-ASMrel , ETE ∪ UTU)

is Pspace-complete. In other words, verifying sequential nullary ASMs with rela-

tional input against (ETE ∪ UTU)-specifications is a Pspace-complete problem,
given that the maximal arity of the employed input relations is a priori bounded.

The proof of the theorem closely follows a construction due to Immerman and
Vardi that was first presented in [17] as a translation of CTL∗ into (FO+TC).
Using this construction one can show that for C = SN-ASMrel , F = ETE, and
L = (E+TC) the second condition in Proposition 19 is satisfied. The containment
assertion of the theorem is then implied by the following two observations:
1. Finite validity and finite satisfiability of (E+TC) sentences over relational
vocabularies is decidable in Pspace if one imposes an upper bound m on
the arities of the occurring relation symbols [21,23].
2. There exists a polynomial-time computable function as in the second condi-
tion of Proposition 19.
verifym (C, F ) is already Pspace-hard if m = 0, F contains EXaccept, and
C includes all ASMs in SN-ASMrel whose external vocabulary is empty and
whose dynamic vocabulary contains only boolean symbols. This follows by a
reduction of the Pspace-complete satisfiability problem for quantified boolean
formulas.
Unfortunately, Theorem 22 does not hold for sequential nullary ASMs whose
inputs contain functions, as the following theorem shows. For the sake of bre-
vity we write liveness(C) instead of verify(C, {EFaccept}) and safety(C)
instead of verify(C, {AG¬error }).
 Model Checking Abstract State Machines and Beyond 339

Theorem 23 ([21]). The two problems liveness(SN-ASMfct ) and

safety (SN-ASMfct ) are both undecidable. Moreover, safety(C) is undecidable
for every class C that includes all deterministic ASMs in SN-ASMfct whose ex-
ternal vocabulary is empty and whose dynamic vocabulary consists of two nullary
function symbols and arbitrarily many nullary relation symbols.

The proof of this theorem is by reduction of the halting problem for Turing
machines. It is based on the following two observations:
1. Inputs that contain non-nullary functions suffice to encode bit-strings.
2. Two nullary dynamic functions suffice to check whether a bit-string encodes
an accepting computation of a Turing machine.
An analysis of the proof indicates that automatic verification of even very simple
ASMs whose inputs can be used for encoding bit-strings is not feasible. Nevert-
heless, Theorem 22 and results in [22] show that restricted variants of ASMs
with relational inputs can be verified automatically.

Acknowledgements. Thanks to the anonymous referees.

References
1. A. Blass and Y. Gurevich. Existential fixed-point logic. In E. Börger, editor,
Computation Theory and Logic, volume 270 of Lecture Notes in Computer
Science, pages 20–36. Springer-Verlag, 1987.
2. A. Blass and Y. Gurevich. The Linear Time Hierarchy Theorems for Abstract
State Machines. Journal of Universal Computer Science, 3(4):247–278, 1997.
3. A. Blass, Y. Gurevich, and J. Van den Bussche. Abstract State Machines
and Computationally Complete Query Languages. This volume.
4. A. Blass, Y. Gurevich, and S. Shelah. Choiceless Polynomial Time. Technical
Report MSR-TR-99-08, Mircosoft Research, 1999.
5. E. Börger and J. Huggins. Abstract State Machines 1988–1998: Commented
ASM Bibliography. Bulletin of the EATCS, 64:105–127, February 1998.
6. D. Beauquier and A. Slissenko. Verification of Timed Algorithms: Gurevich
Abstract State Machines versus First Order Timed Logic. Technical Report
TIK-Report 87, ETH Zürich, March 2000.
7. E. Börger. Why Use Evolving Algebras for Hardware and Software Enginee-
ring? In M. Bartosek, J. Staudek, and J. Wiederman, editors, Proceedings
of 22nd Seminar on Current Trends in Theory and Practice of Informatics
(SOFSEM ‘95), volume 1012 of Lecture Notes in Computer Science, pages
236–271. Springer Verlag, 1995.
8. E.M. Clarke, E.A. Emerson, and A.P. Sistla. Automatic Verification of
Finite State Concurrent Systems Using Temporal Logic. ACM Trans. on
Prog. Lang. and Sys., 8(2):244–263, April 1986.
9. G. Del Castillo and K. Winter. Model Checking Support for the ASM High-
Level Language. Technical Report TR-RI-99-209, Universität-GH Pader-
born, June 1999.
10. H. D. Ebbinghaus and J. Flum. Finite Model Theory. Springer-Verlag, 1995.
340 M. Spielmann

11. E.A. Emerson. Temporal and Modal Logic. In J. van Leeuwen, editor, Hand-
book of Theoretical Computer Science, volume B, pages 995–11072. Elsevier
Science Publishers B.V., 1990.
12. A. Gargantini and E. Riccobene. Encoding Abstract State Machines in PVS.
This volume.
13. E. Grädel and M. Spielmann. Logspace Reducibility via Abstract State
Machines. In J. Wing, J. Woodcock, and J. Davies, editors, World Congress
on Formal Methods (FM ‘99), volume 1709 of Lecture Notes in Computer
Science, pages 1738–1757. Springer-Verlag, 1999.
14. Y. Gurevich. Evolving Algebras 1993: Lipari Guide. In E. Börger, editor,
Specification and Validation Methods, pages 9–36. Oxford University Press,
1995.
15. Y. Gurevich. May 1997 Draft of the ASM Guide. Technical Report CSE-
TR-336-97, University of Michigan, May 1997.
16. Y. Gurevich. The Sequential ASM Thesis. Bulletin of the EATCS, 67:93–124,
1999.
17. N. Immerman and M.Y. Vardi. Model Checking and Transitive Closure
Logic. In Proceedings of 9th International Conference on Computer-Aided
Verification (CAV ‘97), volume 1254 of Lecture Notes in Computer Science,
pages 291–302. Springer-Verlag, 1997.
18. A. Levy, I. Mumick, Y. Sagiv, and O. Shmueli. Equivalence, Query-
Reachability, and Satisfiability in Datalog Extensions. In Proceedings of
12th ACM Symposium on Principles of Database Systems (PODS ‘93), pa-
ges 109–122, 1993.
19. E. Rosen. An existential fragment of second order logic. Archive for Mathe-
matical Logic, 38:217–234, 1999.
20. A.P. Sistla and E.M. Clarke. The Complexity of Propositional Linear Tempo-
ral Logics. Journal of the Association for Computing Machinery, 32(3):733–
749, July 1985.
21. M. Spielmann. Automatic Verification of Abstract State Machines. In Pro-
ceedings of 11th International Conference on Computer-Aided Verification
(CAV ‘99), volume 1633 of Lecture Notes in Computer Science, pages 431–
442. Springer-Verlag, 1999.
22. M. Spielmann. Verification of Relational Transducers for Electronic Com-
merce. In Proceedings of 19th ACM Symposium on Principles of Database
Systems (PODS 2000). ACM Press, 2000. To appear.
23. M. Spielmann. Abstract State Machines: Verification Problems and Comple-
xity. PhD thesis, RWTH Aachen. In preparation.
24. ASM Web Site. http://www.eecs.umich.edu/gasm. Maintained by J. Hug-
gins.
25. K. Winter. Model Checking for Abstract State Machines. Journal of Uni-
versal Computer Science, 3(5):689–701, 1997.
26. K. Winter. Methodology for Model Checking ASM: Lessons learned from
the FLASH Case Study. This volume.
 Towards a Methodology for Model Checking
ASM:
Lessons Learned from the FLASH Case Study

Kirsten Winter1

GMD FIRST, Germany

[email protected]

Abstract Gurevich’s Abstract State Machines (ASM) constitute a high-

level specification language for a wide range of applications. The exist-
ing tool support for ASM was extended, in a previous work, to provide
computer-aided verification, in particular by model checking. In this pa-
per, we discuss the applicability of the model checking approach in gen-
eral and describe the steps that are necessary to fit different kinds of
ASM models for the model checking process. Along the example of the
FLASH cache coherence protocol, we show how model checking can sup-
port development and debugging of ASM models. We show the necessary
refinement for the message passing behaviour in the protocol and give
examples for errors found by model checking the resulting model. We
conclude with some general remarks on the existing transformation al-
gorithm.

1 Introduction

An ASM model comprises the specification of the state space of the system and
its behaviour specified by state transitions. The state space is given by means
of universes and functions over these universes. If the domains and ranges of all
contributing functions are finite (and not too large) and fix an ASM model can
be transformed into a model checker language, e.g. the language of the SMV
model checker ([7]). (We call a universe fix if it may not be extended during a
run by firing some transition rules.)
A first schematic approach is published in [13]. In [2] the schema is extended
for coping with ASM with n-ary functions (n > 0). All n-ary functions will be
unfolded to get 0-ary functions that can be mapped to simple state variables
in the SMV model. Of course, model checking as a fully automatic approach is
limited with respect to the computational effort and thus not feasible for every
ASM model. However, our extended transformation approach can tackle a much
broader set of applications and yields a real improvement of the former simple
transformation.
Once a model is transformed into the model checker language we benefit
from checking properties concerning safety and liveness. If counterexamples can
be detected they may yield a good insight in the model under development.

Y. Gurevich et al. (Eds.): ASM 2000, LNCS 1912, pp. 341–360, 2000.
c Springer-Verlag Berlin Heidelberg 2000

342 K. Winter

On the other hand, our transformation tool supplies SMV with a high level
modelling language. The modelling task is facilitated by allowing the use of more
complex data types and n-ary functions for parametrisation. In general, we found
that ASM models, in comparison to similar SMV models (cf. section 6), are more
concise but yet more general by means of using parameters. Also, an ASM model
can be scaled up more easily than an SMV model, which have to come along
with simple state variables rather than n-ary functions.
This process of transforming the ASM model into SMV language, and check-
ing the resulting SMV model, is running automatically, i.e. without user interac-
tion, once safety and liveness properties are formalised in a temporal logic (for
SMV this is CTL). This is why the model checking approach is so attractive
for industry. No expertise seems to be necessary for using a model checker, it is
simply a “press button functionality”.
This works not only in principle, it works in practice too, as examples show.
In most cases, however, we have to adapt the model at hand in order to fit it to
the model checking process.

1.1 Fitting ASM Models for Model Checking Process

Most ASM — on a certain level of abstraction — do not comprise state and

state transition specification only, but some additional assumptions (cf. fig 1).

assumptions ASM 0
assump. ASM 1
ass. ASM 2
.. ..
.. ..
interface to
environment Implementation

Figure 1. Different Layers of Abstraction

Often these assumptions are necessary for the mathematical proof in order
to cope with parts that are abstract or external for the ASM model. They are
given informally or formalised by some logic expression in the text that surrounds
the specification of the transition rules. Assumptions related to system inherent
behaviour will be specified in terms of ASM transition rules at some lower level
of abstraction, whereas assumptions over the environment of the system are
dedicated to external behaviour that should not be specified further by rules
 Methodology for Model Checking ASM 343

in a refinement step, but rather stated in a declarative way on a high level of

abstraction (e.g. by means of temporal logic formulae). This is necessary for the
development of embedded systems that are not regarded to be robust against
arbitrary — possibly faulty — environment behaviour. We give some simple
examples:

– In the ASM model of the bakery algorithm ([1]) the ticket function T is
external and its behaviour is abstract in the most abstract model. But we
find that some logical properties are assumed that have to be satisfied by T .
This logical properties are then specified in the appropriately refined ASM.
– Also, for the specification of the bakery algorithm, we find that a fairness
assumption is necessary for proving correctness of the algorithm ([1]). At the
moment there is no feature in ASM language to express fairness assumptions.
It should be discussed to extend the language in this direction.
– For embedded systems like the production cell, the specification of the envi-
ronment is abstract ([8]) and not part of the ordinary transition system. The
behaviour of sensors is formalised by means of oracle functions. However, it
is necessary to assume that the behaviour of the environment is “reasonable”
in order to guarantee correctness of the ASM model. In [8] most assumptions
are given in terms of logical formulae that remain external for the refined
ASM, too.
– For the specification of protocols we might abstract from the underlying
communication model governing the transfer of messages, like in the ASM
model of the FLASH cache coherence protocol in [3]. But for the proofs we
have to assume that the messages are transfered according to a particular
strategy (e.g. the FIFO-strategy, such that the ordering of messages is pre-
served). At some lower level in the ASM hierarchy a proper message passing
behaviour has to be specified that implements the assumption made on the
order of transfered messages (cf. section 3).

Obviously, only states, state transitions, and assumptions on the model together
give the complete specification of the problem at hand. This view is sketched
in fig. 1: the dashed box comprises all parts of a model at a particular level of
abstraction.
In ASM, external or system inherent but (on a certain level) abstract be-
haviour is specified by oracle functions. Oracle functions – in case of finite
domains – can be simply transformed into corresponding non restricted state
variables within the SMV model (neither their initialisation nor their updating
is specified), i.e. we give a loose specification and a behaviour may be chosen
non-deterministically. Since model checking is complete testing over the whole
state space, every loose specification leads to a complete case distinction over all
possibilities. The problems arising with this are not only a matter of the known
state explosion problem, but if we fail to meet the additional assumptions on
oracle functions (those which are not expressed by means of transition rules),
the SMV model comprises behaviour that is excluded in the complete ordinary
ASM model.
344 K. Winter

For our approach of transformation, we have to add the assumptions, given

outside of the transition rules, to the SMV model:

– Fairness constraints can simply be added in the SMV code, since the SMV-
language offers the corresponding feature.
– For embedded systems, since the current ASM tools do not support the spec-
ification of assumptions in terms of some temporal logic, we have to specify
assumptions on the environment by means of ASM transition rules to get a
suitable input for the transformation. It is a difficult task, however, to model
behaviour accurately in a constructive way, because abstract properties have
to be encoded operationally. This process easily leads to under-specification
or over-specification, i.e. the specification is too open or too restrictive. It
must be carefully inspected if the specification meets the assumptions. Oth-
erwise errors may remain undetected or a “wrong counterexample” may be
generated (see below). We have to be aware that the results of model check-
ing will hold only with respect to the specified environment.
– Assumptions made on abstract parts of the internal model can simply be
added by means of refining the model appropriately. As an example the
reader is referred to the case study presented in the following sections. (In
particular, the matters of refinement are discussed in section 3).

Let MASM the set of ASM models, MSMV the set of SMV models, and RUN
the set of runs of an ordinary temporal structure over branching time, which
provides a semantics that both kinds of model have in common. To give a more
precise notion of missing assumptions, we introduce mod and mod be two func-
tions that yields the semantics of a model such that mod : MASM → RUN
and mod : MSMV → RUN . Problems may arise in two cases:

mod(MASM ) ⊂ mod(MSMV )

mod(MASM ) ⊃ mod(MSMV )
In the first case — the transformed SMV-model is strict greater than
the ASM-model because of a loose specification of assumptions —
mod(MSMV ) \ mod(MASM ) may contain runs, that violate the property to be
checked. One of these may give a counterexample that is, however, not a proper
run of the ASM model. No proposition on the model under investigation is
possible then since only one counterexample will be given. We call it wrong
counterexample, it obstructs the overall debugging process.
In the second case — when specifying assumptions too restrictive and the
SMV-model is strict smaller than the ordinary ASM-model — we may fail to
detect errors that occurs only in those runs of the ASM-model that are excluded
for the SMV-model. The runs in mod(MASM ) \ mod(MSMV ) will not be checked
by the model checker.
In the following we give an example for fitting an ASM-model to our model
checking approach. We show how the ASM model of the FLASH cache coherence
protocol that is given in [3] can be refined in order to add the assumption that
 Methodology for Model Checking ASM 345

are stated in the proof part. Also, the necessary fairness assumption are added in
the SMV code. Assumptions on environmental behaviour are not needed in this
particular case, we can keep the “environment” by means of oracles functions
without any restricting specification.
In section 2 we present the original abstract model of the FLASH protocol
from [3]. Section 3 introduces our refinements that are necessary to complete
the model to be transformed. The results that show the benefits from model
checking are presented in section 4. In section 5 the transformation and its
possible optimisations is discussed. We give remarks on related work in section 6
and conclude with section 7.

2 Case Study: The FLASH Cache Coherence Protocol

The Stanford FLASH multiprocessor (cf. [6]) integrates support for cache co-
herent shared memory for a large number of interconnected processing nodes.
That is, each node (or processor) holds some part of memory, which is accessible
for other nodes too, in order to read or write data from this part. One of the
problems to be investigated for this architecture is the coherence of the data,
since access to data is realized by working with a copy of it. It may happen that
one node is going to process a copy of the data when the data have changed in
the meanwhile (because of some writing access of another node).
The parts of the distributed memory are given as sets of lines, i.e. small pieces
of memory content. Each line is associated with a home node hosting the part
of the physical memory where the line resides. Whenever a process (or a node)
needs to have access to a particular line, we say that a read or write miss occurs.
Every read or write miss concerning a remote memory line triggers a line request
to its home node.
Being interconnected the nodes are able to communicate with each other.
They can send (and deliver) messages in order to request for a remote line or
to respond on a request by means of sending a copy of the needed data. To
provide coherence of the data additional book-keeping is necessary to prevent
from simultaneous reading and writing on the same line. That is, writing needs
exclusive access to a line whereas, reading is allowed in shared access. Message
passing and book-keeping of shared and exclusive access is the matter of the
protocol specification we consider in our ASM model.
The ASM model of the protocol in [3] is based on agents. Each agent models
one (processor) node that holds a certain block of the distributed memory. A
set of transition rules describes the behaviour of a single agent. Its behaviour is
determined by the incoming message that is to be processed, in fact the type
of a message is suitable for determination. This notion yields the clear model
structure shown in figures 2 and 3.
Incoming messages are requests from remote nodes (or from the receiving
node itself in order to simplify the model, this special case is called intra-node
communication further on). In figure 2 (lower part) the reader may find transi-
tion rules to be triggered when a request for shared access is received. In figure 3
346 K. Winter

(upper part) transition rules for handling requests for exclusive access are de-
picted. For readability the message types to be distinguished are surrounded by
boxes.
A read or write miss that causes a request to be sent is arbitrarily generated
by means of oracles. These oracles come along as messages as well, their types
have the form cc. (cf. figure 2, upper part). Whenever the agent receives such
kind of message it generates the corresponding request and enters a waiting
mode. Releasing an access is handled in the same way. If the agent sends a
request for releasing a line the state of the line is invalidated, that is, the node
has no reliable copy of the line any more.
In our adaptation of the model the parts related to data (to be sent within
the messages) are discarded. Since the data do neither influence the control flow
of the protocol behaviour (data do not control any of the guards in the transition
rules) nor determine the properties to be checked, we do not lose expressiveness
of the model and our checking results.
State Functions. Beside the message type the agent’s behaviour depends on
several state variables: curPhase(line) (phase of the current request), State(line)
(state of the local line copy in use), and pending(line) (flag for currently processed
request). Owner(line) and the set of Sharers of a line are also taken into account.
These functions are local for each agent, the additional parameter self is omitted
in figures 2 and 3.
Message Structure. A message is modelled as a quintuple consisting of the
type of the message, the addressed agent, the sender agent, the agent initiating
the request and the requested line.
The message types related to shared access are:
get: requesting a line from its home
put: granting a line to the requester (source of the request)
fwdget: forwarding the request to an exclusive owner of the line
swb: requesting a write-back of an owned line that is to be shared
nack, nackc: negatively acknowledging the request or forwarded request
(nackc), if it cannot be performed now.
The transition rules that are related to requests for shared access are given
in figure 2: the circulation of a “get”-request. A get-request may either be (a)
negatively acknowledged if another request on the line is already processing
(pending(line) is true), or (b) it is forwarded to the current owner if there is
already an exclusive access (owner(line) not undefined), or (c) it is granted to
the requester if no owner is noted (else case). If there is an owner (case (b)
above), the grant for shared access is given by the agent which “believes” to
be an owner, and, moreover, the owner has to release its exclusive copy. If an
agent gets a forward get-request and does not agree to be the owner of the line
(State(l) is not exclusive) then the request has to be negatively acknowledged
as well.
In analogy, message types related to exclusive access are:
getx: requesting a line for exclusive access from its home
putx: granting a line for exclusive access to the requester
 Methodology for Model Checking ASM 347

if M essT ype = cc.get ∧ CurP hase(l) = ready

then SendMsg(get, home(l), self , self , l)
CurP hase(l) : = wait

if M essT ype = cc.getx ∧ CurP hase(l) = ready

then SendMsg(getx, home(l), self , self , l)
CurP hase(l) : = wait

if M essT ype = cc.rpl ∧ CurP hase(l) = ready ∧ State(l) = shared

then SendMsg(rpl, home(l), self , self , l)
State(l) : = invalid

if M essT ype = cc.wb ∧ CurP hase(l) = ready ∧ State(l) = exclusive

then SendMsg(wb, home(l), self , self , l)
State(l) : = invalid

if M essT ype = get

then if pending(l) then SendMsg(nack, source, self , source, l)
else if Owner(l) 6= undef
then SendMsg(fwdget, Owner(l), self , source, l)
pending(l) : = true
else SendMsg(put, source, self , source, l)
Sharer(l, source) : = true

if M essT ype = fwdget

then if State(l) = exclusive then SendMsg(put, source, self , source, l)
SendMsg(swb, home(l), self , source, l)
State(l) : = shared
else SendMsg(nack, source, self , source, l)
SendMsg(nackc, home(l), self , source, l)

if M essT ype = put if M essT ype = swb

then curP hase(l) : = ready then Sharer(l, source) : = true
if curP hase(l) 6= invalid Sharer(Owner(l), l) : = true
then State(l) : = shared Owner(l) : = undef
pending(l) : = false

if M essT ype = nack if M essT ype = nackc

then curP hase(l) : = ready then pending(l) : = false

Figure 2. Responding on a request for shared access

348 K. Winter

if M essT ype = getx

then if pending(l)
then SendMsg(nack, source, self , source, l)
else if Owner(l) 6= undef
then SendMsg(fwdgetx, Owner(l), self , source, l)
pending(l) : = true
else if ∃u : Sharer(l, u)
then ∀u : Sharer(l, u)SendMsg(inv, u, self , source, l)
pending(l) : = true
else SendMsg(putx, source, self , source, l)
Owner(l) : = source

if M essT ype = fwdgetx

then if State(l) = exclusive
then SendMsg(putx, source, self , source, l)
SendMsg(fwdAck, home(l), self , source, l)
State(l) : = invalid
else SendMsg(nack, source, self , source, l)
SendMsg(nackc, home(l), self , source, l)

if M essT ype = putx if M essT ype = fwdAck

then State(l) : = exclusive then if Owner(l) 6= undef
CurP hase(l) : = ready then Owner(l) : = source
pending(l) : = true

if M essT ype = inv

then SendMsg(invAck, home(l), self , source, l)
if State(l) = shared
then State(l) : = invalid
else if curP hase(l) = wait
then curP hase(l) : = invalidP hase

if M essT ype = invAck

then Sharer(l, M essSender) : = false
if ∀a : Agents | a 6= M essSender ∧ Sharer(l, a) = false
then SendMsg(putx, source, self , source, l)
pending(l) : = false

if M essT ype = rpl if M essT ype = wb

then if ∃u : Sharer(l, u) ∧ ¬pending(l) then if Owner(l) 6= undef
then Sharer(l, u) : = undef then Owner(l) : = undef

Figure 3. Responding on a request for exclusive access

 Methodology for Model Checking ASM 349

fwdgetx: forwarding the request for exclusive access the owner of the line
inv: requesting a current sharer of the line to invalidate its local copy
invAck: acknowledging the invalidation of the line
fwdAck: owner’s granting according to a forwarded shared request.
The corresponding behaviour is shown in figure 3: the circulation of a “getx”-
request. Again, the exclusive request may be negatively acknowledged if another
request is in process already, or it is forwarded to the owner if there is one
noted. In addition to the get-request we have to take care for possible sharers
if there is no owner. Each of the sharers of the line has to be informed about
the request for exclusive access (by sending an inv-message). When getting an
inv-message, sharer has to invalidate its copy of the line and response with an
acknowledgement of invalidation (sending invAck-message). When each of the
sharers has sent its invalidation acknowledgement to home a grant for exclusive
access is send to the requester (i.e. sending putx-message is delayed until all
invAck-messages are received).
For releasing a shared or exclusive copy from its cache an agent sends a
write back (wb) or a replace message (rpl) to home. The agent has to be deleted
from the list of shares or not being owner any more.

3 Refinements of the ASM Model

Sending a message is given as a macro definition. In the abstract model of [3]

SendMsg adds a message to a (possibly infinite) set of messages in transit using
the extend rule constructor. The strategy for receiving a message from this set
is not specified. For the proof it is just assumed in [3] that the messages are
received in the right order. In order to formalise this assumption and to keep
the model finite we had to refine the model.1 By means of the refinement we
restrict the model excluding those runs, in which a message is sent but never
received, or sent but overwritten by a message that is sent later. That is, we
want to exclude that messages may received in a different order than being sent.
Without our refinements the model is either infinite (in [3] the set of messages
in transit is not restricted at all) and yield wrong counterexamples as well (if
we rely on restricting the set to be finite only). We cannot guarantee liveness if
messages may be received in a wrong order.
In order to avoid loss of messages and to guarantee that the order of sending
messages and receiving a message is proper, we add a strategy of synchronisa-
tion: (a) We introduce a queue for messages in transit for each agent. (b) A
message can be send only if the queue for messages in transit is not full. (c) We
interleave the protocol behaviour with synchronisation steps. Each step of pro-
tocol communication (specified by the rules given in figures 2 and 3) is followed
by one step of message passing through. The refinement and necessary changes
are described in the following subsections.
1
At http://www.first.gmd.de/˜kirsten the listing of the refined ASM-SL model
can be found.
350 K. Winter

3.1 Queue for Messages in Transit

Instead of handling messages as tuples we decided to keep the components of a

message as single functions. (That is to avoid unnecessary complexity caused
by the unfolding step in the transformation.) The formerly infinite universe
MessInTransit is replaced by finite queues for each agent that holds one of
the message components:
MessInTr: QLength * Agent -> Type,
SenderInTr: QLength * Agent -> Agent,
SourceInTr: QLength * Agent -> Agent, and
LineInTr: QLength * Agent -> Line.

The universes QLength, Agent, and Line are sets that are finitely restricted by
a maximal index: maxQ, maxAgent, or maxLine. These constants can easily be
adjusted in order to scale up the model under investigation.
Sending a message (SendMsg) is now specified as appending the message
components to the corresponding functions: If ni is the smallest index for which
the queue for messages in transit is empty, we update all message-component-
functions at this index. We indicate emptiness of queue(ni ) by means of message
type noMess for the message type component, i.e. if (MessInTr(ni , ai )=noMess)
is satisfied and for all nj < ni it is not, then ni marks the index for writing the
message to be append. This is specified by means of the following macro:
transition AppendToTransit(agent_,(sender_,mess_,source_,line_)) ==
if MessInTr(1,agent_)=noMess
then SenderInTr(1,agent_) := sender_
MessInTr(1,agent_) := mess_
SourceInTr(1,agent_):= source_
LineInTr(1,agent_) := line_
else do forall i in { 2..maxQ }
if MessInTr((i-1),agent_)!=noMess and MessInTr(i,agent_)=noMess
then SenderInTr(i,agent_) := sender_
MessInTr(i,agent_) := mess_
SourceInTr(i,agent_):= source_
LineInTr(i,agent_) := line_
endif
enddo
endif

For requests (i.e. message type is cc.get, cc.getx, cc.wb, cc.rpl) we in-
troduce an extra queue for request in transit MessInTrR. Analogously, we define
dynamic functions that hold one of the components of a request. Sending of a
request message is specified by means of the macro AppendRequestToTransit
that looks similar to the appending macro shown above.

3.2 Passing through of Messages in Transit

Sending of a message by means of appending it to the queue of messages in

transit, is only one half of the message exchange. As a second step we have to
 Methodology for Model Checking ASM 351

specify how the messages in transit are read by the agents to be addressed, that
is how messages are passing through (cf. Fig. 4).
Each agent holds a current message (which is seen to be empty if its type
equals to noMess). Again we model the current message by means of its com-
ponents: InMess, InSender, InSource, and InLine are dynamic functions over
domain Agent. The current message is delivered by passing through the first
element of the transit-queues MessInTr and MessInTrR (see Fig. 4). Note, that
requests have lower priority than messages, i.e. a request is passed through only
if there is no message in transit left.

MessInTr
message_passing
... SendMsg

InMess 1 2 3 maxQ

MessInTrR
SendRequest
message_passing ...
1 2 maxR

Figure 4. Message passing through from MessInTransit to Incoming Message

Processing the current message and delivering a new message operates on the
same dynamic functions of the model. In order to avoid racing we have to inter-
leave both steps: message processing of the agents and message passing through.
We extend the overall behaviour by means of a sub-step for synchronisation. In
the synchronisation step the messages are passed through to the addressed agent
in the proper order.
In an ASM model, introducing a sub-step (to be processed after each state
transition) is structure preserving: in addition to the ASM for message processing
we specify an ASM for the message passing through. That is, we do not have to
interfere the ordinary transition rules (cf. figures 2 and 3) with transition rules
for message passing through. An overall ASM main invokes both “sub-ASM”
agent behaviour and message passing in turn:
transition main ==
if toggle = behave
then agent_behaviour
toggle := sync
else (* if toggle = sync *)
message_passing
toggle := behave
endif

Taking this, we benefit from the clear and understandable structure of the
abstract model that is given by [3].
352 K. Winter

3.3 Racing Situation for Intra-node-Communication

Racing situations (i.e. simultaneous updating of the same location through dif-
ferent rules) is excluded since we apply an interleaving semantics to the agents
running in parallel, and introduce sub-steps for message passing through (as de-
scribed above). However, racing may be caused by intra-node communication:
To keep the ASM model small we do not distinguish whether the requesting
node is also home of the line under request or not. That is, if the home of a line
wants to have shared or exclusive access of its own line, it has to proceed in the
same way as all other nodes has to, i.e. sending a get/getx-message to itself.
If requester is also home of the line it may happen that two message are to
be send simultaneously to the same address. A racing situation occurs, one of
the messages (the first of two updates) will be lost. (In figures 2 and 3 the reader
may find such situations in the fwdget-rule and the fwdgetx-rule.)
In order to avoid racing we introduce three new message types that combine
two messages: put swb, putx fwdAck, and nack nackc. Whenever the source of
the message (i.e. the requester) is equal to the home of the line under request, we
send one of combined messages instead of sending two single messages. According
to the new message types we introduce also rules that combine both of the rules
that have to be fired when receiving such messages.

4 Results of Model Checking

We take model checking of the transformed ASM model as an evolutionary pro-

cess of debugging: we edit the ASM model, transform it automatically into an
SMV model, run SMV to check the properties under investigation, simulate the
resulting counterexample (if any) with the ASM Workbench, and edit the ASM
model again. As the debugging process is more efficient if the model checking
terminates in a reasonable span of time we keep our model in the beginning as
small as possible. We find that, even if the model is restricted to two commu-
nicating agents and one line of memory, we detect errors in the abstract model
as well as in our refinement. In the following we describe some of the detected
errors as examples.2 The errors raise when checking the model for safety and
liveness, i.e.:
– No two agents have exclusive access on the same line simultaneously.
– Each request will eventually be acknowledged.
– Whenever an agent gets shared access home will note it as a sharer.

2
The corresponding listings of the counterexamples can be found in the appendix of
the full version of this paper to be find on http://www.first.gmd.de/˜kirsten.
 Methodology for Model Checking ASM 353

We formalise these requirements in CTL, e.g.3 :

V
[AG (!(State(a(i),l)=exclusive & State(a(j),l)=exclusive))]
Vi6=j
Vi [AG (curPhase(a(i),l) = wait -> AF (curPhase(a(i),l) = ready))]
i [AG ( State(a(i),l)=shared -> AX (Sharer(l,a(i)) = true)
| Sharer(l,a(i)) = true -> AX (State(a(i),l)=shared)) ]

Our first counterexample shows simultaneous exclusive access. We simulate the

ASM model and run into an error that can also be found in the abstract ASM
model of [3]:
Whenever a putx-message is sent to grant exclusive access the addressed
requester has to be noted as owner of the line. This is specified for the
getx-rule but it is missing for the invAck-rule that might also cause a
putx-message to be sent (see also Fig. 3). The protocol is unsafe since
simultaneous exclusive access may occur, and written data may be lost.
Other counterexamples are dedicated to the problem of racing (i.e., conflicts
caused by simultaneous updates) on the finite message queue. Although our data
space is restricted to a very short queue, we can derive more general remarks
for the message passing via queues that concern the finite limitation of a queue
itself (which, in fact, is characteristic for each possible realization), e.g.:
Each sharer of a requested line has to process the rule for invalidation
(inv-rule). It sends an invAck-message to home for acknowledging the in-
validation. When receiving an invAck-message, home deletes the sender
from the list of sharers. If home is sharer too (in case of intra-node com-
munication (see Sec. 3.3)), a deadlock may occur if the number of sharers
is greater or equal than the length of the message queue: home may fail
to complete with the inv-rule when the queue is full and sending a mes-
sage is not possible (since every other sharer may have sent before); home
stays busy and can not process the incoming invAck-rule to clear the
queue. In general, we found out that the message queue must be larger
or equal than the number of agents since in the worst case each agent is
a sharer and will send simultaneously an invAck-message to the home
node.
We correct our ASM model in that we reply the guards of the request-rules
(i.e. rules for cc.get, cc.getx, cc.wb, and cc.rpl) when passing the requests
in transit through to the home of the requested line.
The examples show that helpful borderline cases can be detected more easily
by a model checker than by pure simulation. The computational effort for the
automated transformation of our models ranges from three to five seconds. The
size of the resulting SMV models is given below where the different columns yield
3
Though the third specification is rather weak, it yields helpful counterexamples. The
disjunction is necessary due to the fact that the sequence in that the corresponding
updates occur (noting a state to be shared and adding a sharer) changes for different
scenarios.
354 K. Winter

the comparing results when scaling up the parameters for the number of agents
and lines.4 The variable ordering is determined by the automatic reordering
facility that is given by the SMV.
resources used: 2 agents, 1 line 3 agents, 1 line 2 agents, 2 lines
user time/system time: 4.69 s/0.13 s 5687.52 s/0.6 s 17263.2 s/0.86 s
BDD nodes allocated: 70587 1612740 2975127
Bytes allocated: 4849664 37748736 54657024
BDD nodes repr. transition relation: 19261 + 78 288986 + 82 78365 + 96

Although checking our model of the FLASH protocol is only feasible for a
small number of agents and lines, the results show that the counterexamples
yield extremely helpful scenarios for locating errors.

5 Transformation Issues
The general approach for the transformation of ASM transition rules into SMV
code is already described in [13]. The extended transformation algorithm that
includes the transformation of n-ary dynamic functions (with n > 0) can be
found in [2]. Based on this foundation we want to add some more general remarks
here concerning extensibility and optimisation. In order to make this paper self
containing we recall our main ideas of transformation.

5.1 Transformation of ASM into SMV Language

The basic translation scheme introduced in [13] can be applied to transform into
SMV a subset ASM0 of ASM, where: (i) only nullary dynamic and external
functions are allowed; (ii) the only available data types are integers, booleans
and enumerated types; (iii) the only defined static functions are those corre-
sponding to predefined operations in SMV (boolean operations, +, -, etc.).
As the semantic models for ASM0 are essentially basic transition systems,
the translation of ASM into SMV is very close:

– non-static functions (i.e., dynamic and external functions) are identified with
locations and thus mapped one-to-one to SMV state variables;
– values of the ASM data types are mapped one-to-one to SMV constants;
– applications of static functions are translated to applications of the corre-
sponding built-in operators of SMV.

What remains to be done is to restructure the ASM program into a form where
updates of the same location, together with their guards, are collected together.
This is done in two steps. First, we transform an ASM program P into an
equivalent ASM program P 0 consisting only of a block of guarded updates (i.e.,
rules of the form if G then f (t) := t) by means of a “flattening” transformation:
4
The experiments were carried out on an UltraSPARC-II station with 296MHz and
2048 Mb memory, the operating system is Solaris 2.6.
 Methodology for Model Checking ASM 355
( 
if G1T then RT1  if G ∧ G1T then RT1


[[RT ]] = ... 
 ...
if Gn n 
T then RT
n n
if G ∧ GT then RT
( ⇒ [[if G then RT else RF ]] =
if G1Fthen 1
RF 
 if ¬G ∧ G1F then RF
1


[[RF ]] = ... ...

if Gm m
F then RF
if ¬G ∧ Gm m
F then RF

Second, we collect all guarded updates of the same location, thus obtaining,
for each location loc occurring on the left-hand side of an update in P 0 , a pair
(loc, {(G1 , t1 ), . . . , (Gn , tn )}) which maps loc to a set of pairs (guard, right-hand
side). Such a pair is translated into the following SMV assignment:
ASSIGN next(C [[loc ]]) :=
case C [[G1 ]] : C [[t1 ]] ; ... C [[Gn ]] : C [[tn ]] ; 1 : C [[loc ]] esac;

where C [[.]] denotes here the ASM → SMV compiling function for terms, which
is straightforward for ASM0 .
The ordinary transformation shown above is extended in order to tackle the
ASM language as complete as possible. We have to exclude, however, infinite
domains as well as the notion of import or export transition rules because
the domains of a model have to be fix and may not grow during a run. But
any other rule constructor as well as arbitrary data types and operations (in
particular, lists, finite sets, finite maps and user-definable freely generated types,
as provided by ASM-SL) can be used without restriction. Finite quantifications
are also supported.
In our approach we reduce an arbitrary (finite and fix) ASM to ASM0 . The
main problem here is that in general we do not know which location is updated
by f (t1 , . . . , tn ) := t (if n > 0) because the parameters ti may be dynamic
functions and thus can change their current value. Evaluation of a location is
possible if and only if all ti are evaluated, i.e. either they are static or have to be
unfold to cover all possible values. Thus, the basic idea of the transformation is
to iteratively unfold and simplify rules until all terms can be reduced to values
or locations.
Terms can be simplified by means of the transformation [[.]]ρ defined in Ta-
ble 1, which is then extended to rules in a canonical way.
The rule-unfolding transformation E, which operates on closed rules such as
the program P , is formally defined in Table 2. It works as follows:

– if the rule R consists of a block of update rules of the form location := value,
it terminates and yields R as result (there is nothing left to unfold);
– otherwise, it looks for the first location l occurring in R (but not as left-hand
side of some update rule) and unfolds R according to the possible values of l.
In turn, the unfolding has to be applied to the sub-rules [[R[ l/xi ]]] obtained
by substituting the values xi for l in R and simplifying.

Applying E to the (simplified) ASM program [[P ]]∅ yields a program P 0 =

E([[P ]]∅ ) which is essentially an ASM0 program.
356 K. Winter

Table 1. Term and Rule Simplification

Term Simplification
[[x]]ρ = x [[ l ]]ρ = l

x = ρ(v) if v ∈ dom(ρ)
[[ v ]]ρ =
v otherwise

[[ti ]]ρ = xi for each i ∈{1, . . . , n} ⇒

x = f (x1 , . . . , xn ) if f static function name
[[f (t1 , . . . , tn )]]ρ =
l = (f, (x1 , . . . , xn )) if f dynamic/external function name
0
[[ti ]]ρ = l or [[ti ]]ρ = f 0 (t ) for some i ∈ {1, . . . , n} ⇒
[[f (t1 , . . . , tn )]]ρ = f ([[t1 ]]ρ , . . . , [[tn ]]ρ )

 [[G]]ρ[v7→x1 ] op . . . op [[G]]ρ[v7→xn ]

if [[A]]ρ = x = {x1 , . . . , xn } (i.e., if [[A]]ρ is a value)
[[(Q v in A : G)]]ρ =
 (Q v in [[A]]ρ : [[G]](ρ\v) )

otherwise.
(where op ≡ ∧ if Q ≡ forall, op ≡ ∨ if Q ≡ exists).

Rule Simplification
[[skip]]ρ = skip
[[tL := tR ]]ρ = [[tL ]]ρ := [[tR ]]ρ
[[R1 . . . Rn ]]ρ = [[R1 ]]ρ . . . [[Rn ]]ρ
(
[[RT ]]ρ if [[G]]ρ = true
[[if G then RT else RF ]]ρ = [[RF ]]ρ if [[G]]ρ = false
if [[G]]ρ then [[RT ]]ρ else [[RF ]]ρ otherwise.
forall v in A with G R0 ]]ρ =
[[do 
0 0
 [[if G then R ]]ρ[v7→x1 ] . . . [[if G then R ]]ρ[v7→xn ]

if [[A]]ρ = x = {x1 , . . . , xn } (i.e., if [[A]]ρ is a value)
= 0

 do forall v in [[ A ]] ρ with [[ G ]] (ρ\v) [[ R ]] (ρ\v)
otherwise.

Table 2. Rule Unfolding

Rule Unfolding
If R has the form l1 := x1 . . . ln := xn , then E(R) = R.
Otherwise:
E(R) = if l = x1 then E([[R[l/x1 ]]]∅ )
else if l = x2 then E([[R[l/x1 ]]]∅ )
...
else if l = xn then E([[R[l/xn ]]]∅ )
where l is the first location occurring in R (but not as lhs of an update rule)
and {x1 , . . . , xn } is the range of location l.
 Methodology for Model Checking ASM 357

5.2 Interfacing Other Model Checkers

Using the SMV as a tool for checking ASM was the choice for our first approach
in this field. This decision was guided by the fact that the semantics of the SMV
modelling language (i.e. the target language for the transformation) is given by
simple transition systems. That is, it is very close to the ASM semantics. In
principle, the SMV can easily be substituted by any other model checker that is
based on transition systems, too, e.g., the SVE ([4]) or the VIS ([5]). Since our
transformation consists of several steps:
1. Unfolding the n-ary dynamic functions in order to get a simple ASM with
only 0-ary functions (named ASM0 ).
2. Flattening the nested structure of the transition rules of the ASM0 in order
to get sets of simple guarded updates for each location (i.e. sets of triples
(location, guard, update value)).
3. Transforming the individual sets of guarded updates into our notion of ab-
stract syntax of SMV language.
4. Pretty printing the model given in abstract syntax of the SMV language into
SMV code.
In order to develop an interface to any other model checker that treats a kind of
transition system it is obviously sufficient to substitute only the last two steps
of the transformation.

5.3 Optimisation Issues

Since efficiency is a crucial point for the model checking approach we also have
to focus on optimisation issues. The results of the investigations so far are sum-
marised in the following two points:
Location Unfolding. In [2] the rule unfolding is formally given by the
schema shown in section 5.1 (see table 2).
For the implementation of this schema we identified a crucial point for op-
timisation: The set of locations to be unfolded (i.e. substituted by their values)
should be more restricted. We regard all boolean operators, e.g. not, and, or,
equal as primitive functions. Only those locations that occur as function param-
eters of non-primitive functions or appear at the right hand side of updates or
equations are to be substituted. Locations that appear as 0-ary functions on
the left hand side of equations (i.e. in guards) or as parameters for primitive
functions should not be unfolded. Although this optimisation seems to be obvi-
ous it is crucial for our approach in order to keep the transformation efficient
(and feasible) even for large models. Since the transformation schema given in
[2] suggests the exhaustive use of induction (even more since we implement our
transformation by means of ML) it appeared that additional case distinctions
are crucial. With the resulting optimised transformation algorithm we are able
to generate SMV code in few seconds. Thus, the transformation is not any longer
the bottleneck of our approach.
Module Structure in SMV Code. In the current approach for the trans-
formation we choose the most simple way for treating agents in ASM: We con-
sider self (denoting the data space of a particular agent) as a simple oracle
358 K. Winter

function that is updated non-deterministically. The resulting SMV code has a

plain structure since we do not exploit the notion of modules or processes given
in the SMV language. Instead, the transition behaviour of all agents is specified
to be parallel and synchronous but guarded by the current value of self. Using a
more elaborated module structure in the target language we might benefit from
the notion of interleaving semantics of parallel process. In [7] it is stated that
the internal BDD structure (that represents the model) may be smaller in size
for interleaving asynchronous than for synchronous transition relations.
To investigate a possible optimisation of the transformation, we edited the
SMV code of our running example in that we introduced the model structure
appropriately for specifying interleaving agents. We compared the resulting SMV
models in their size of BDD, which is one of the measurements for the model
checking algorithm. (The other one is the CTL formula to be checked.) The
corresponding order of SMV state variables is produced by -reorder option of
SMV.

– without use of processes, normal transformation:

BDD nodes allocated: 55446
Bytes allocated: 6029312
BDD nodes representing transition relation: 12444 + 54

– with use of process but only global variables:

BDD nodes allocated: 76367
Bytes allocated: 3473408
BDD nodes representing transition relation: 11649 + 50

– with use of process and with locally declared process variables:

BDD nodes allocated: 89365
Bytes allocated: 3670016
BDD nodes representing transition relation: 13626 + 50

At least for our example of the FLASH protocol we found that the size of
BDDs that represent the corresponding models are comparable. Therefore we
keep the transformation without exploiting the notion of SMV modules in order
to keep the implementation as simple as possible.

6 Related Work
In [11] a different approach for automated verification for ASM is elaborated.
Spielmann represents an ASM model independently of its possible input by
means of a logic for computation graphs (called CGL*). The resulting formula
is combined with a CTL*-like formula which specifies properties and both are
checked by means of deciding their finite validity. This addresses the problem of
checking systems with arbitrary inputs. However, Spielmann concludes that this
approach is restricted to ASM with only 0-ary dynamic functions and relational
input. For many problems, however, we found that the input, or the environment
respectively, proper to the ASM model at hand is restricted by certain assump-
tions (as described in Sec. 1.1). The restriction to 0-ary functions outweighs this
 Methodology for Model Checking ASM 359

potential benefit. In [10] some helpful theoretical results of the model checking
and the verification problem for ASM are published. This investigation tends to
define the limit of both problems in general in terms of their decidability and
complexity.
In some cases the formalisation of an environment behaviour that is restricted
by assumptions is a complex task. Operational semantics (like ASM) may yield
more complex solutions than a logical approach (like the temporal logic LTL). It
might be useful to facilitate the user with LTL for the specification of additional
assumptions on the external behaviour. This approach is supported by the model
checker SVE ([4]). Thus, adapting our transformation to the SVE tool might be
promising.
The case study of FLASH protocol that is used here in order to give an
example for applying our approach is investigated within other approaches as
well. The most elaborated work is published in [9]. Park and Dill introduced
a method relying on so called aggregation function that yields the refinement
relation between specification and implementation. Their approach for proving
correctness is supported by means of the interactive theorem prover PVS. Giving
up on fully automated support they can tackle a broader range of requirements,
of course.
Wing and Vaziri-Farahani investigated in [12] a slightly similar kind protocol
for cache coherence (but not exactly the FLASH protocol) by means of model
checking, namely using SMV. As we do they claim that “Without loss of gen-
erality, we can analyse this cache coherence protocol by considering one client,
C, one server, S, and one file, f”. But the introduced SMV model, in contrast
to the ASM model we used, lacks the possibility to scale up easily by using pa-
rameters. We found the ASM model, being more general but yet concise at the
same time, leads to a more clear and understandable structure than modelling
in plain SMV language only.
Also McMillan in [7] applied the SMV to a distributed cache protocol. But
his SMV model of the protocol being spread over several modules is not fully
introduced and thus difficult to understand. Since the application is not exactly
the same as the FLASH protocol we cannot relate the effort for checking (BDD
size and checking time) properly.

7 Conclusions

In previous work ([13] and [2]) an interface from the ASM Workbench to SMV
was presented. The transformation from ASM to the SMV language comprises
the treatment of dynamic functions of arity n > 0, which is an crucial improve-
ment, as most ASM specifications benefit from the abundant use of parametric
dynamic functions.
In this paper we gave an overview of our experience with model checking
ASM so far. Although, our transformation can treat a great subset of the ASM
language we found that in practice most ASM models have to be refined or ex-
tended to fit for the model checker: Since assumptions on the ASM model that
360 K. Winter

are stated besides the ordinary transition rules are not covered by the trans-
formation we have to add a proper specification of these assumptions into the
resulting SMV-model. This can be done by simply exploiting language features
of SMV (e.g. the fairness construct), by further refinement of the ASM model
at hand, or by specifying environment behaviour by means of ASM transition
rules such that the given assumptions are covered by the transformation.
The practicability of the approach is demonstrated by a non-trivial case
study: the ASM model of the FLASH protocol. In order to fit the model for
the model checking process we have to refine the model in that we specify the
message passing behaviour. We gave some examples for errors found by the model
checker that can hardly be detected by pure mathematical proofs, and deduce
more general constraints for the model at hand from the counterexamples.
We conclude with some remarks on the implementation of the transformation
algorithm concerning extensibility and optimisation issues.

Acknowledgements. My thank belongs to the anonymous referees for their

helpful comments as well as Stephan Herrmann and Giuseppe Del Castillo for
many fruitful discussions.

References
1. E. Börger, Y. Gurevich, and D. Rosenzweig. The Bakery Algorithm: Yet Another
Specification and Verification. In E. Börger, editor, Specification and Validation
Methods. Oxford University Press, 1994.
2. G. Del Castillo and K. Winter. Model checking support for the ASM high-level
language. In S. Graf and M. Schwartzbach, editors, Proc. on 6th Int. Conf. TACAS
2000, volume 1785 of LNCS, pages 331–346, 2000.
3. A. Durand. Modeling cache coherence protocol - a case study with FLASH. In
U. Glässer and P. Schmitt, editors, Procs. of the 28th Conf. of German Society of
Computer Science, TR, Magdeburg University, 1998.
4. T. Filkorn et al. SVE Users’ Guide. Siemens AG, München, 1996.
5. The VIS Group. Vis: A system for verification and synthesis. In T. Henzinger
R. Alur, editor, 8th Int. Conf. CAV’96, volume 1102 of LNCS, 1996.
6. J. Kuskin, D. Ofelt, and M. Heinrich et. al. The stanford FLASH multiprocessor.
In 21th Int. Symp. on Computer Architecture, 1994.
7. K. McMillan. Symbolic Model Checking. Kluwer Academic Publishers, 1993.
8. L. Mearelli. An Evolving Algebra Model Of The Production Cell. Master’s thesis,
Universita di Pisa, 1996.
9. S. Park and D. Dill. Verification of cache coherence protocols by aggregatiuon of
distributed transactions. Theory of Computing Systems, 31:355–376, 1998.
10. M. Spielmann. Model Checking Abstract State Machines and Beyond. In This
Volume.
11. M. Spielmann. Automatic verification of abstract state machines. In N. Halbwachs
and D. Peled, editors, Computer Aided Verification, CAV ’99, number 1633 in
LNCS, pages 431–442, Trento, Italy, 1999.
12. J. M. Wing and M. Vaziri-Farahani. A case study in model checking software
systems. Science of Computer Programming, 28:273–299, 1997.
13. K. Winter. Model checking for abstract state machines. J.UCS Journal for Uni-
versal Computer Science (special issue), 3(5):689–702, 1997.
 Report on a Practical Application of ASMs in
Software Design

Egon Börger1 , Peter Päppinghaus2 , and Joachim Schmid2

1
Università di Pisa, Dipartimento di Informatica, I-56125 Pisa, Italy
[email protected]
2
Siemens AG, Corporate Technology, D-81730 Munich, Germany
{peter.paeppinghaus, joachim.schmid}@mchp.siemens.de

Abstract. ASMs have been used at Siemens Corporate Technology to

design a component in a software package called FALKO. Main purpose
of FALKO is the construction and validation of timetables for railway
systems. For simulation the whole closed-loop traffic control system is
modelled within FALKO. The railway process model part of FALKO
was formally specified using the ASM approach. C++ code is generated
from the formal specification and compiled together with the handwrit-
ten C++ code of the other components to obtain the FALKO executable.
The project started in May 1998 and was finished in March 1999. Since
then FALKO is used by the Vienna Subway Operator for the validation
of the whole subway operational service.

1 FALKO

Abstract State Machines (ASMs) [3,1] have been used at Siemens Corporate
Technology to design a component in a software package called FALKO. This
name is the German acronym for “Timetable Validation and Timetable Con-
struction” describing concisely the main functionality of this tool. Detailed ope-
rational timetables including e.g. vehicle roster plan can be automatically calcu-
lated from raw data like train frequency and infrastructure of a railway system.
For these calculations estimations of trip times are used. Timetables – whether
constructed with FALKO or other tools – have to be validated for operability
and robustness. Conventionally this is done by (physically) driving trial runs.
With FALKO this costly form of validation can be replaced by dynamic simula-
tion of operational timetables modelling quantitative aspects like velocities and
trip times as accurately as needed.
To perform dynamic simulation of timetables, the whole closed-loop traffic
control system is modelled within FALKO. The model is designed in a mo-
dular way with three main components: train supervision/train tracking, in-
terlocking system, and railway process model. Software components for train
supervision/train tracking and for the interlocking system are nowadays part
of real train operation. The railway process model on the other hand serves to
replace the real physical system (trains, signals, switches etc.) in the simulation.

Y. Gurevich et al. (Eds.): ASM 2000, LNCS 1912, pp. 361–366, 2000.
c Springer-Verlag Berlin Heidelberg 2000

362 E. Börger, P. Päppinghaus, and J. Schmid

Fig. 1. Screenshot of FALKO GUI

The modelling is based on discrete event simulation. The three components

communicate via events, which are tagged with certain data and stamped with
the (fictitious) time of their occurrence. Let us give an example. Train supervi-
sion, e.g., sends an event at time t0 to the interlocking system requesting route
r3 . The interlocking system, possibly after having received an event which frees
route r3 , sends an event at time t1 to the railway process model requesting switch
sw5 to change its position. The railway process model then calculates the dura-
tion of rotating this switch and sends an event to the interlocking system to the
effect that switch sw3 has arrived at the requested position at time t2 , etc.
Besides the three main components modelling the traffic control system there
is as a fourth hidden component the event handler, which controls the simula-
tion by receiving events from the visible components and sending them to the
respective addressee in the appropriate order.
The railway process model is based on a physical model of driving according
to which the crucial numerical calculations for determining speeds, trip times etc.
are done. This physical model is another hidden component of FALKO realized
as a C++ library.
Further parts of FALKO are a comfortable GUI and extensive facilities to
analyze and graphically display data of a simulation run. The GUI can also be
used for visualization of simulation runs (see figure 1).
 Report on a Practical Application of ASMs 363

Fig. 2. Screenshot of HTML documentation

2 Design of Railway Process Model with ASMs

2.1 Design Process and Development Tools
The railway process model has been designed with ASMs [3,1]. All the other
components of FALKO have been designed and implemented conventionally with
handwritten C++ code.1 Along with this design a prototypical development en-
vironment has been built up, which can also be used to maintain this component
of FALKO.
This development environment supports a seamless flow from the specifi-
cation down to the executable code. The single source for FALKO’s railway
process model is a specification consisting of formal parts together with infor-
mal explanations. The formal parts are ASM rules and static, dynamic, derived
and external functions written in ASM-SL, the language of the ASM workbench
[2]. The specification comes as a collection of HTML documents (see figure 2).
Hyperlinks to navigate between uses and definitions of rules and functions are
generated automatically. The formal parts can be extracted automatically from
the HTML documents to be fed into the front end of the ASM workbench, which
is used for syntax and type analysis.
1
The physical model of driving is not included in the railway process model. It has
been designed conventionally as a component of its own.
364 E. Börger, P. Päppinghaus, and J. Schmid

In the design phase of the project the ASM workbench was also used for
early tests of the ASM model(s). At the end of the design phase it was decided
to attempt code generation rather than hand coding the already debugged ASM
model.
In the implementation phase a code generator has been developed, automa-
tically generating C++ code from type correct ASM-SL code. In addition some
wrapper code for interfacing the generated code to the remaining components,
and some low-level “library code” was hand coded.
Formal verification of the ASM model has not been attempted, this was not
a goal of the project.

2.2 Effort for Design and Implementation

Design Phase

– Requirement specification based on predecessor system, developed in mee-

tings of the design team, documented by minutes of the meetings (4 persons
2 weeks)
– Design of 1st draft of executable ASM model (1 person 8 weeks)
– Several cycles of testing and debugging using the ASM workbench [2] (1
person 8 weeks + 1 person 11 weeks)
– Review of 2nd draft of ASM model by design team plus external reviewers
(6 persons 1 week)
– Several cycles of improving, testing and debugging (2 persons 5 weeks)

Implementation Phase

– Development of ASM-SL to C++ code generator (1 person 4 weeks)

– Specification and implementation of additional handwritten C++ code (1
person 2 weeks)
– Integration of FALKO system including testing and debugging (3 persons 3
weeks)
– Documentation of railway process model component and final polish (1 per-
son 6 weeks)

Summing up that part of the above listed effort, which was spent on behalf of
the railway process model component of FALKO, this yields a total effort of 66
person weeks.
It is not possible, of course, to compare this effort reliably to the correspon-
ding effort in a conventional software design process. But a rough estimation
done by an experienced programmer intimately familiar with FALKO says that
the ASM design (including development of the C++ code generator) has excee-
ded a conventional effort by about 10%.
 Report on a Practical Application of ASMs 365

2.3 Size of ASM Model and C++ Code

ASM Model (source of C++ code generation)
– ca. 3 000 lines of ASM workbench code
– 120 rules
– 315 functions and relations (240 functions, 75 relations)
– 71 dynamic
– 69 external
– 59 static
– 116 derived

C++ Code
– ca. 9 000 lines of generated C++ code
– ca. 2 900 additional lines of handwritten C++ code, consisting of
– ca. 400 lines wrapper code for interfacing to other components of FALKO
– ca. 2 500 lines low-level library code
In the prototypical predecessor system of FALKO the railway process model
consisted of ca. 20 000 lines of (handwritten) C++ code. To be fair one has to
take into account, however, that this component – having been experimented
with – had grown over time (and become difficult to maintain, which was the
reason for redesigning it completely).

3 Experiences
It turned out that one of the main advantages of ASMs for this design was the
parallel update view. This made it possible to model the railway process in such
a way that each state of the ASM model corresponds to a “snapshot” of the
virtual physical process taken at the occurrence of a “relevant discrete event”.
Due to this, one can always have a clear picture of the physical state snapshot,
on which auxiliary computations (specified by static and derived functions) are
based.
FALKO developers and reviewers not concerned with formal methods had no
problems to understand the ASM model. The possibility of early tests by execu-
ting the ASM model with the ASM workbench was very helpful and uncovered
bugs also in other components of FALKO at an early stage.
No serious attempt has been made to measure the potential performance loss
due to the use of generated C++ code. Comparison to the predecessor system of
FALKO has been possible only on one example, for which the data happened to
be available in both systems. In this small example performance of the previous
system was about 30% better than that of the current FALKO system. For the
time being the performance problem has been left aside, since it turned out that
FALKO’s performance is good enough for the purpose the product is used for.
FALKO is used in four installations at the Vienna Subway Operator since
March 1999, one of these installations being in daily use. Up to now (i.e. March
366 E. Börger, P. Päppinghaus, and J. Schmid

2000) the customer reported no bugs. After finishing the first version of FALKO,
two bugs in the railway process model have, however, been discovered in tests
during development of the second version. The FALKO developers have not yet
familiarized themselves with the ASM specific tools. The generated C++ code
being readable enough they chose to implement temporary fixes of the two bugs
by handhacking the generated C++ code, and to postpone correction of the
ASM model until the faults are analyzed more thoroughly.

Acknowledgments. We thank our colleagues of the FALKO team for their

open attitude towards this pioneering use of ASMs and for a year of enjoyable
and fruitful cooperation. Furthermore we thank Giuseppe Del Castillo for the
permission to use early versions of his ASM workbench and for the support we
obtained from him.

References
1. E. Börger and J. Huggins. Abstract State Machines 1988-1998: Commented ASM
Bibliography. Bulletin of EATCS, 64:105–127, February 1998. Updated bibliography
available at http://www.eecs.umich.edu/gasm.
2. Giuseppe Del Castillo. The ASM Workbench – A tool environment for computer
aided analysis and validation of ASM models. PhD thesis, University of Paderborn,
to appear.
3. Yuri Gurevich. Evolving Algebras 1993: Lipari Guide. In E. Börger, editor, Speci-
fication and Validation Methods, pages 9–36. Oxford University Press, 1995.
 Using Abstract State Machines at Microsoft:
A Case Study

Mike Barnett, Egon Börger? , Yuri Gurevich,

Wolfram Schulte, and Margus Veanes

Microsoft Research, One Microsoft Way, Redmond, WA 98052, USA

{mbarnett, boerger, gurevich, schulte, margus}@microsoft.com

Abstract. Our goal is to provide a rigorous method, clear notation and

convenient tool support for high-level system design and analysis. For
this purpose we use abstract state machines (ASMs). Here we describe a
particular case study: modeling a debugger of a stack based runtime en-
vironment. The study provides evidence for ASMs being a suitable tool
for building executable models of software systems on various abstraction
levels, with precise refinement relationships connecting the models. High
level ASM models of proposed or existing programs can be used throug-
hout the software development cycle. In particular, ASMs can be used
to model inter component behavior on any desired level of detail. This
allows one to specify application programming interfaces more precisely
than it is done currently.

1 Introduction
This paper describes a case study on the use of ASMs as support for design
and analysis of software at Microsoft. In order to use ASMs we needed a tool
for executing ASMs integrated with the Microsoft programming environment,
in particular with the Component Object Model or COM [5]. We developed a
prototype called AsmHugs [8] by extending the Hugs system [2] which is an
implementation of the lazy functional programming language Haskell. AsmHugs
is in many ways similar to AsmGofer [1] but Hugs enabled us to use H/Direct [6,
7] for integration with COM. A detailed technical report of this case study is in
preparation [3].

1.1 What Is the Case Study about?

We present a model for a command-line debugger of a stack-based runtime envi-
ronment. The model was reverse-engineered from the debugger which is written
in C++ and which uses a particular application programming interface (API).
We have in fact three models of the debugger, each at a different level of ab-
straction. They form a refinement hierarchy where each refinement relationship
is a procedural abstraction.
?
Visiting researcher from Università di Pisa, Dipartimento di Informatica, I-56125
Pisa, Italy, [email protected]

Y. Gurevich et al. (Eds.): ASM 2000, LNCS 1912, pp. 367–379, 2000.
c Springer-Verlag Berlin Heidelberg 2000

368 M. Barnett et al.

1.2 Why This Particular Case Study?

Our motivation was to study the applicability (and integrability) of ASMs to

Microsoft software. Software at Microsoft is generally written in modules, or
components. These often form client/server relationships where a component
may function both as a server and as a client. A client and a server interact
solely through an established API. An API is a set of procedures to which all
communication is restricted. The debugger in question is an application program
for end-users. It is a client of the debug services.
Suppose you want to understand how the API works without reading the
code. You have a specification in the Interface Definition Language (IDL [5])
which gives you only the relevant signatures. The inner workings of the API’s
methods are hidden; the only additional information you may find is an informal
natural-language description. Such descriptions may be incomplete and become
often inconsistent with the code, as the code evolves over time. Obviously, there
is no way to formally enforce correspondence between code and its natural-
language description. The two main problems with using program code itself
as documentation are these. First, the code is usually huge and includes too
many irrelevant details. Second, the code might not be available for proprietary
reasons.
In this study, we model a particular program but our interests are broader:
how to use ASMs to better specify the APIs by which different components
interact. An ASM model provides the missing information at the appropriate
level of abstraction so that the user of a given component can understand both
the behavior of the component and the protocol that the user is supposed to
follow in order to exploit the behavior. Each method is specified as a rule, and
the valid patterns of calls are reflected in the state.
As part of a broader project, we also built an ASM model of the other side
of the API: the debug services (to be described in a forthcoming report); the
debugger model was a valuable source in that context.

1.3 COM

Microsoft software is usually composed of COM components. These are really

just static containers of methods. In your PC, you will find dynamic-link li-
braries (DLLs); a library contains one or more components (in compiled form).
COM is a language-independent as well as machine-independent binary standard
for component communication. An API for a COM component is composed of
interfaces; an interface is an access point through which one accesses a set of
methods. A client of a COM component never accesses directly the component’s
inner state, or even cares about its identity; it only makes use of the functionality
provided by different methods behind the interface (or by requesting a different
interface).
 Using Abstract State Machines at Microsoft 369

1.4 Integration with COM

Writing an executable model for a client of a COM component requires that
either you model also the server side of the interfaces (and then maybe the COM
components for which that server is a client, and so on) or you integrate your
model into the COM environment and thus make it possible for your model to
actually call the COM component. We do the latter. H/Direct provides a means
for a Haskell program to communicate with COM components both as a client
and as a server. For the debugger model, both modes (the client and the server)
are needed. The server mode is needed because the debug services API specifies
the callback interface that the client must implement. The debug services use the
callback interface to make asynchronous procedure calls to the model. However,
Hugs is a sequential system. In order to use it in an asynchronous multi-threaded
COM environment, a modification of the Hugs runtime was required. Other
modifications of the Hugs runtime were needed to allow the outside world to
invoke Hugs functions.
All three of the debugger models in the case study, even the most refined one
(the ground model), are sequential ASMs (enriched with some Haskell functio-
nalities). The ground model communicates with the outside world. The question
arises how to view that communication in ASM terms. Calls from the ground
model to the debug services are invocations of external functions. Calls in the
other direction (the callbacks) happen to be more specific: callbacks are updates
of nullary monitored functions.

2 The Case Study

The case study is to model a sample debugger of a particular runtime environ-
ment. The main goal of this debugger is to illustrate a correct way to use the
debugging services provided by that runtime. See Figure 1. Nonetheless, the de-
bugger has more than 30k lines of C++ code and exhibits complex behavior,
partly due to the involvement of several asynchronous agents (the user, the run-
time and the operating system), but mostly because of the complexity of the
debug services that expose about 50 interfaces and 250 methods.
A careful analysis of the debugger, mainly by analyzing the source code, led
us to a refinement hierarchy consisting of three models. By analyzing the causal
dependencies of different actions we realized that the complications due to the
seemingly asynchronous behaviour could be completely avoided, which enabled
us to model the debugger by a sequential ASM. The design decisions which are
reflected by the resulting debugger model were later validated by running the
model as a replacement of the actual debugger.

1. Control model. The abstraction level of this model is at the control level.
The user can enter a command if the debugger is in a mode where the user
has a prompt. The runtime environment can issue a callback to the debugger
only if the latter is expecting a callback. At this level, the only effect of a
callback or user command is to change the control state.
370 M. Barnett et al.

Debugger IDL
Debug
Services

Runtime

Fig. 1. Debugger and the runtime debug services.

2. Object model. This model reflects the compile time or static structure of
the underlying architecture, namely that modules contain classes, classes
contain functions, functions contain code, etc. Furthermore, it provides a
restricted view of the run time structure, namely that processes contain
threads, threads contain frames, etc. At this level exactly those commands
are modeled that are intimately connected to the compile time structure and
to the restricted view of the run time structure, such as execution control,
breakpoint control, and stepping commands.
3. Ground Model. This model has the same core functionality as the debugger.
It provides a more detailed view of the run time structure than the object
model. User commands dealing with inspection of the run time stack, and
contents of individual frames, like inspection of specific object fields, are
modeled here.

AsmHugs
Control Object Ground
Model Model Model H/Direct

Runtime
Debug Services

Fig. 2. Connections between the models and the runtime.

 Using Abstract State Machines at Microsoft 371

In principal, all models are executable, either without the runtime or through
the refinements using the actual runtime (see Figure 2). However, in the former
case, a proper simulation harness has to be provided. An example of a test
harness by using a “wizard” is illustrated in Section 3.2. In a related project, we
constructed an ASM model that approximates the runtime. This model could
be used here to embody the runtime in the harness.
When executing the control model using the runtime, each action is interpre-
ted by AsmHugs via the indicated refinements; COM connectivity is provided
by H/Direct.

3 Control Model of the Debugger

In this model we consider the main control modes of the debugger. These modes
reflect who has the control: the user, the runtime, or the debugger itself. When
the user has control, the debugger presents a prompt and waits for input. The
debugger waits for a response from the runtime when the runtime has the control.
Once the runtime has responded with an event, the debugger has control and
has to decide what mode to transit to. As we will see, this decision is based on
particular properties of the event and further input from the runtime.
Although the communication between the real debugger and the runtime is
asynchronous, the documentation of the debug API specifies that the runtime
issues at most one asynchronous event at a time; before issuing the next event
the runtime must receive an acknowledgement from the debugger. Furthermore,
the communication protocol between the debugger and the runtime ensures that
at most one of them may have control. Therefore, sequential ASMs suffice for
our modeling purposes.
The remainder of the section is structured as follows. First, we will lay out
the control model ASM in full detail, introducing the necessary state components
and the rules as needed. Second, we will run a particular user scenario that will
indicate a possible discrepancy between the runtime and the model. Finally, we
will remedy the model.
This example shows a typical use of a high level model during the design
phase in a software development cycle.

3.1 Control Model ASM

The debugger can be in one of four modes. In Init mode the debugger either
hasn’t been started yet, or it has been terminated. In Break mode the user
has the control. In Run mode the runtime has the control. In Break? mode the
debugger has the control. The dynamic ASM function, dbgMode, records the
mode of the debugger in the current state; it has the initial value Init. The top
level rule of the debugger is dbg. Below, do is short for do in-parallel.

dbgMode = initVal Init

dbg = do
372 M. Barnett et al.

if dbgMode == Break or dbgMode == Init then handleCommands

if dbgMode == Run then handleResponses
if dbgMode == Break? then handlePendingEvents

There are two monitored functions, command and response, that are updated
by the user and the runtime, respectively. Initially, both monitored functions
have a value that indicates that no input has been entered by either.

command = initVal "nothing"

response = initVal "nothing"

User commands. We can partition user commands into three groups: com-
mands for starting and quitting the debugger, commands that hand the control
over to the runtime (e.g. execution control and stepping commands), and com-
mands that do not affect the control mode (e.g. state inspection and breakpoint
setting).
The user can issue commands only if the debugger is either in Init mode
or in Break mode. In the first case, the only meaningful action is to start the
debugger.

handleCommands = do onStart
onExit
onBreakingCommand
onRunningCommand
command := "nothing"

onStart = if dbgMode == Init and command == "start"

then do doCommand("start")
dbgMode := Break

In Break mode the debugger handles normal debugging commands and it

may switch to Run mode or back to Init mode, depending on the command.

onExit = if dbgMode == Break and command == "exit"

then do doCommand("exit")
dbgMode:= Init

onBreakingCommand = if dbgMode == Break and isBreakingCommand(command)

then do doCommand(command)

onRunningCommand = if dbgMode == Break and isRunningCommand(command)

then do doCommand(command)
dbgMode:= Run

Firing of execution control commands and stepping commands implies that

the control is handed over to the runtime.

isRunningCommand(x) = x in? {"run <pgm>", "continue", "kill",

"step into", "step over", "step out"}
 Using Abstract State Machines at Microsoft 373

Other commands have no effect on the control mode.

isBreakingCommand(x) =
not(isRunningCommand(x)) and x != "exit" and x != "start"

As mentioned above, in the control model, handling of commands is a skip

operation. This rule is refined in the object model.

doCommand(x) = skip

Callbacks. The debugger can handle callbacks or responses from the runtime
only in the Run mode. The value of the monitored function response is a callback
message from the runtime notifying the debugger about a runtime event. In the
debugger, each event is classified either as a stopping event or as a non-stopping
event.

handleResponses = do onStoppingEvent
onNonStoppingEvent
response := "nothing"

onStoppingEvent =
if isStoppingEvent(response) then do dbgMode:= Break?
doCallback(response)

onNonStoppingEvent =
if not(isStoppingEvent(response)) then doCallback(response)

Breakpoint hit events, step complete events, and process exit events are al-
ways stopping events. Initially isStoppingEvent is the following unary relation
(unary Boolean function).

isStoppingEvent(x) = x == "step completed" or x == "breakpoint hit" or

x == "process exited"

However, the relation is dynamic and may evolve during a run as a result of
a specific user command or a specific runtime event.
In the control model the actual handling of callbacks is a skip operation. This
rule is refined in the object model.

doCallback(x) = skip

Pending Events. In Break? mode a stopping event has happened and the
debugger should hand the control over to the user. This happens only if there
are no pending events in the runtime. Otherwise the control goes back to the
runtime and the debugger continues the current process
374 M. Barnett et al.

handlePendingEvent = do onPendingEvent
onNoPendingEvent

onPendingEvent = if isEventPending then do dbgMode := Run

isEventPending := False
doCommand("continue")

onNoPendingEvent = if not(isEventPending) then do dbgMode := Break

The boolean function isEventPending is monitored by the runtime.

The control model is summarized by a state diagram in Figure 3.

onBreakingCommand onNonStoppingEvent
onRunningCommand

Break Run
onPendingEvent
onNoPendingEvent
onStoppingEvent
onStart onExit Break?

Init

Fig. 3. Control model of the debugger.

3.2 A Wizard-of-Oz Experiment

We want to explore the behavior of the model on a given set of user scenarios,
without having access to the real runtime, and possibly expose contradictions in
the model. Since we reverse-engineered the debugger, we cannot claim that our
model is truly faithful to it. In fact, any error that might show up may very well
have sneaked into the model without being in the debugger. The point we want
to make is that this form of testing would be useful if used during the design
phase.
Since the actual runtime is missing, we will ask a “wizard” to play its role.
Table 1 shows a run, with row i indicating the part of the state that has changed
after the i’th ASM step of the model.
 Using Abstract State Machines at Microsoft 375

Table 1. A run of the control model.

dbgMode command response isEventPending

0: Init start
1: Break bp hello.cpp:7
2: Break run hello.exe
3: Run created process
4: Run loaded module
5: Run created thread
6: Run hit breakpoint
7: Break? True
8: Run loaded class
9: Run ...

The run shows that after we hit a breakpoint there is a pending event in
the runtime. According to the model, the current process is continued and the
control is passed to the runtime. It turns out that the pending event was a
non-stopping event (class was loaded). Obviously, this behaviour contradicts the
expected consequence of reaching a breakpoint, namely that the user should
get the control. At this point we have two options to solve the problem: if we
can constrain the runtime to meet the restriction that only stopping events can
cause the isEventPending flag to become true, then the current model is correct;
otherwise we have to modify the model. We choose the latter, see Figure 4.

onBreakingCommand onNonStoppingEvent
onRunningCommand

Break Run

onNoPendingEvent
onStoppingEvent
onStart onExit Break?
onPendingEvent’

onAnyEvent
Init Run’

Fig. 4. New control model of the debugger.

376 M. Barnett et al.

4 Object Model

The static and the dynamic structures of the runtime architecture are reflected
in the object model with just enough detail so that one can model features
that deal with execution control, breakpoints, and stepping, and explore their
interaction.
Let us consider some examples. The first example is the refinement of the
user command that starts the debugger. The next two examples specify what
happens in the object model when a module is loaded or unloaded in the runtime.
Recall that all those rules are just skip rules in the control model.

4.1 User Commands

Starting the debugger

doCommand("start") = do in-sequence
coInitialize
shell := newShell(services = newServicesInterface(...),
callback = newCallbackInterface(...),
modules = {},
breakpoints = {},
debugee = undef)
shell.services.setCallback(shell.callback)

The first rule initializes the COM environment. The second rule creates a new
debugger shell with a pointer to the services and a new callback. The third rule
invokes the setCallback method of the debug services with the new callback
interface pointer as argument, thus providing the services access to the client’s
callback methods.

4.2 Callbacks

Loading of modules

doCallback(<"load module",mod>) = do
shell.modules := shell.modules ++ {mod}
do in-sequence
forall bp in shell.breakpoints do bp.bind(mod)
shell.debugee.continue()

The new module is recorded in the shell. The shell attempts to bind each of
its set of breakpoints to that module. Once all new bindings have been made (if
any), the debugee is continued through an external call.
We have omitted the bind rule that checks if the location that the breakpoint
refers to indeed exists in the module. If the location exists, a real breakpoint
is created at that location through an external call to the services; otherwise,
nothing is done.
 Using Abstract State Machines at Microsoft 377

Unloading of modules

doCallback (<"unload module",mod>) = do

shell.modules := shell.modules \ {mod}
do in-sequence
forall bp in shell.breakpoints do bp.unbind(mod)
shell.debugee.continue()

The effect of unloading a module is to update the shell and to remove all the
breakpoints from that module.

4.3 Some Comments

There are a couple of things worth noticing about the above rules. First, the
design decision to bind a breakpoint that may already be bound implies that if
there is a breakpoint referring to a location that exists in two or more distinct
modules, then the breakpoint is associated to all of them. Second, all breakpoints
are handled simultaneously; there are no ordering constraints between them.
This is a typical situation: in the actual (C++) code there is a linearly ordered
structure maintaining the elements that are then processed sequentially in the
order determined by the structure.
When constructing the object model we detected a mismatch between the
way loading and unloading of module callbacks was implemented. Namely, alt-
hough loading followed the specification above, during unloading of a given mo-
dule, if a breakpoint was bound to this module, then that breakpoint was not
only removed from this module but from all other modules as well. We did not
discover this when studying the code, but only when constructing the object
model. In fact, it is hard to see it from the code, but readily apparent from the
model.

5 Ground Model

The ground model has the same core functionality as the debugger and can be
executed as an AsmHugs program that communicates with the runtime by using
the H/Direct generated glue code. It gives a more detailed view of the run time
structure than the object model. All the user commands, such as inspection of
specific object fields, are modeled here. The ground model is fully described in
the technical report [3].

6 Conclusion

The case study illustrates some particular ways that ASMs can help the system
designer:
378 M. Barnett et al.

– Concisely describe complex systems. The debugger involves several asynchro-

nous agents and involves about 50 interfaces and about 250 methods. The
size of the debugger is about 30K lines of C++ code. The size of the ASM
specification (which can be run, though not as fast, and which provides es-
sentially the same functionality) is only 4K lines. Due to built-in parallelism,
the runs of our ASM are shallow. In fact it takes only bounded many steps
(less than 10) to process one user command in contrast to the C++ code
which may require unbounded many steps for the purpose (the number of
steps may depend on the program being debugged). The parallelization was
obtained, e.g., by abstraction from irrelevant sequentialization in the C++
code (see Section 4.3).
– Abstract away the environment as needed. We could easily separate the mo-
dels, but still formally tie them via refinements, without losing executability.
– Interactively explore the design on all abstraction levels. It is very difficult
to detect at the source code level such high level bugs as the ones that we
detected with relative ease with the Wizard-of-Oz experiment. The object
model enabled us to detect inconsistencies in the callback management.

To repeat, our goal is to provide a rigorous method, clear notation and con-
venient tool support for high-level system design and analysis. Our main tool
will execute ASMs (and will help you to write ASMs). Several issues that arose
during the case study have influenced the design of that tool.

– We found the parallel synchronous construct forall very useful. Similar

conclusions were drawn from the Falko project [4].
– Set and list comprehensions turned out to be very convenient.
– We found object oriented notation useful to structure specs, to improve their
readability, and to link their execution to the object-oriented programming
paradigm.
– We realized that in order for ASMs to be useful in Microsoft (or indeed
anywhere COM is used), ASM models must achieve full COM interoperabi-
lity.

The first and the third point are illustrated by the examples in Section 4.2.

Acknowledgments. We thank Sigbjorn Finne for helping us with H/Direct

during the course of this work.

References
1. AsmGofer. http://www.tydo.de/AsmGofer/.
2. Hugs98. http://www.haskell.org/hugs/.
3. Mike Barnett, Egon Börger, Yuri Gurevich, Wolfram Schulte, and Margus Veanes.
Using ASMs at Microsoft: A case study. Technical report, Microsoft Research,
Redmond, USA, 2000.
4. Egon Börger, Peter Päppinghaus, and Joachim Schmid. Report on a practical ap-
plication of ASMs in software design. In This Volume.
 Using Abstract State Machines at Microsoft 379

5. Don Box. Essential COM. Addison-Wesley, Reading, MA, 1998.

6. Sigbjorn Finne, Daan Leijen, Erik Meijer, and Simon Peyton Jones. H/Direct: A
binary foreign language interface for Haskell. In Proc ACM Sigplan International
Conference on Functional Programming (ICFP’98), Baltimore, pages 153–162, 1998.
7. Sigbjorn Finne, Daan Leijen, Erik Meijer, and Simon Peyton Jones. Calling hell
from heaven and heaven from hell. In Proc ACM Sigplan International Conference
on Functional Programming (ICFP’99), Paris, France, 1999.
8. Foundations of Software Engineering, Microsoft Research. AsmHugs.
http://www.research.microsoft.com/foundations/.
 Author Index

Anlauff, Matthias 69 Prinz, Andreas 242

Päppinghaus, Peter 361
Barnett, Margus 367
Blass, Andreas 9, 22 Reisig, Wolfgang 112
Börger, Egon 1, 223, 361, 367 Riccobene, Elvinia 223, 303
Rosenzweig, Dean 131
Cater, Steven C. 203
Cavarra, Alessandra 223 Schmid, Joachim 361
Cohen, J. 34 Schulte, Wolfram 151, 367
Shankar, Natarajan 287
Eschbach, Robert 242 Slissenko, A. 34
Spielmann, Marc 323
Gargantini, Angelo 303
Teich, Jürgen 266
Glässer, Uwe 242
Goos, Gerhard 177 Van den Bussche, Jan 22
Gotzhein, Reinhard 242 Veanes, Margus 367
Gurevich, Yuri 22, 131, 151, 367
Wallace, Charles 151
Huggins, James K. 203 Weper, Ralph 266
Winter, Kirsten 341
Kutter, Philipp W. 266
Zamulin, A.V. 91
Odersky, Martin 50 Zimmermann, Wolf 177