This book – the first of two volumes – explores the syntactical constructs of the most common programming languages, and
264 96 9MB
English Pages 336 [329] Year 2021
Table of contents :
Cover
HalfTitle Page
Title Page
Copyright Page
Contents
Foreword
Preface
1 From Hardware to Software
1.1. Computers: a lowlevel view
1.1.1. Information processing
1.1.2. Memories
1.1.3. CPUs
1.1.4. Peripheral devices
1.2. Computers: a highlevel view
1.2.1. Modeling computations
1.2.2. Highlevel languages
1.2.3. From source code to executable programs
2 Introduction to Semantics of Programming Languages
2.1. Environment, memory and state
2.1.1. Evaluation environment
2.1.2. Memory
2.1.3. State
2.2. Evaluation of expressions
2.2.1. Syntax
2.2.2. Values
2.2.3. Evaluation semantics
2.3. Definition and assignment
2.3.1. Defining an identifier
2.3.2. Assignment
2.4. Exercises
Exercise 2.1
Exercise 2.2
3 Semantics of Functional Features
3.1. Syntactic aspects
3.1.1. Syntax of a functional kernel
3.1.2. Abstract syntax tree
3.1.3. Reasoning by induction over expressions
3.1.4. Declaration of variables, bound and free variables
3.2. Execution semantics: evaluation functions
3.2.1. Evaluation errors
3.2.2. Values
3.2.3. Interpretation of operators
3.2.4. Closures
3.2.5. Evaluation of expressions
3.3. Execution semantics: operational semantics
3.3.1. Simple expressions
3.3.2. Callbyvalue
3.3.3. Recursive and mutually recursive functions
3.3.4. Callbyname
3.3.5. Callbyvalue versus callbyname
3.4. Evaluation functions versus evaluation relations
3.4.1. Status of the evaluation function
3.4.2. Induction over evaluation trees
3.5. Semantic properties
3.5.1. Equivalent expressions
3.5.2. Equivalent environments
3.6. Exercises
Exercise 3.1
Exercise 3.2
Exercise 3.3
Exercise 3.4
Exercise 3.5
Exercise 3.6
Exercise 3.7
4 Semantics of Imperative Features
4.1. Syntax of a kernel of an imperative language
4.2. Evaluation of expressions
4.3. Evaluation of definitions
4.4. Operational semantics
4.4.1. Bigstep semantics
4.4.2. Smallstep semantics
4.4.3. Expressiveness of operational semantics
4.5. Semantic properties
4.5.1. Equivalent programs
4.5.2. Program termination
4.5.3. Determinism of program execution
4.5.4. Big steps versus small steps
4.6. Procedures
4.6.1. Blocks
4.6.2. Procedures
4.7. Other approaches
4.7.1. Denotational semantics
4.7.2. Axiomatic semantics, Hoare logic
4.8. Exercises
Exercise 4.1
Exercise 4.2
Exercise 4.3
5 Types
5.1. Type checking: when and how?
5.1.1. When to verify types?
5.1.2. How to verify types?
5.2. Informal typing of a program Exp2
5.2.1. A first example
5.2.2. Typing a conditional expression
5.2.3. Typing without type constraints
5.2.4. Polymorphism
5.3. Typing rules in Exp2
5.3.1. Types, type schemes and typing environments
5.3.2. Generalization, substitution and instantiation
5.3.3. Typing rules and typing trees
5.4. Type inference algorithm in Exp2
5.4.1. Principal type
5.4.2. Sets of constraints and unification
5.4.3. Type inference algorithm
5.5. Properties
5.5.1. Properties of typechecking
5.5.2. Properties of the inference algorithm
5.6. Typechecking of imperative constructs
5.6.1. Type algebra
5.6.2. Typing rules
5.6.3. Typing polymorphic definitions
5.7. Subtyping and overloading
5.7.1. Subtyping
5.7.2. Overloading
6 Data Types
6.1. Basic types
6.1.1. Booleans
6.1.2. Integers
6.1.3. Characters
6.1.4. Floating point numbers
6.2. Arrays
6.3. Strings
6.4. Type definitions
6.4.1. Type abbreviations
6.4.2. Records
6.4.3. Enumerated types
6.4.4. Sum types
6.5. Generalized conditional
6.5.1. C style switch/case
6.5.2. Pattern matching
6.6. Equality
6.6.1. Physical equality
6.6.2. Structural equality
6.6.3. Equality between functions
7 Pointers and Memory Management
7.1. Addresses and pointers
7.2. Endianness
7.3. Pointers and arrays
7.4. Passing parameters by address
7.5. References
7.5.1. References in C++
7.5.2. References in Java
7.6. Memory management
7.6.1. Memory allocation
7.6.2. Freeing memory
7.6.3. Automatic memory management
8 Exceptions
8.1. Errors: notification and propagation
8.1.1. Global variable
8.1.2. Record definition
8.1.3. Passing by address
8.1.4. Introducing exceptions
8.2. A simple formalization: MLstyle exceptions
8.2.1. Abstract syntax
8.2.2. Values
8.2.3. Type algebra
8.2.4. Operational semantics
8.2.5. Typing
8.3. Exceptions in other languages
8.3.1. Exceptions in OCaml
8.3.2. Exceptions in Python
8.3.3. Exceptions in Java
8.3.4. Exceptions in C++
Conclusion
Solutions to the Exercises
A.1. Introduction to language semantics Solution to exercise 2.1
Solution to exercise 2.2
A.2. Semantics of functional features Solution to exercise 3.1
Solution to exercise 3.2
Solution to exercise 3.3
Solution to exercise 3.4
Solution to exercise 3.5
Solution to exercise 3.6
Solution to exercise 3.7
A.3. Semantics of imperative features Solution to exercise 4.1
Solution to exercise 4.2
Solution to exercise 4.3
List of Notations
Index of Programs
References
Index
Other titles from iSTE in Computer Engineering
EULA
Concepts and Semantics of Programming Languages 1
Series Editor JeanCharles Pomerol
Concepts and Semantics of Programming Languages 1 A Semantical Approach with OCaml and Python Thérèse Hardin Mathieu Jaume François Pessaux Véronique Viguié DonzeauGouge
First published 2021 in Great Britain and the United States by ISTE Ltd and John Wiley & Sons, Inc.
Apart from any fair dealing for the purposes of research or private study, or criticism or review, as permitted under the Copyright, Designs and Patents Act 1988, this publication may only be reproduced, stored or transmitted, in any form or by any means, with the prior permission in writing of the publishers, or in the case of reprographic reproduction in accordance with the terms and licenses issued by the CLA. Enquiries concerning reproduction outside these terms should be sent to the publishers at the undermentioned address: ISTE Ltd 2737 St George’s Road London SW19 4EU UK
John Wiley & Sons, Inc. 111 River Street Hoboken, NJ 07030 USA
www.iste.co.uk
www.wiley.com
© ISTE Ltd 2021 The rights of Thérèse Hardin, Mathieu Jaume, François Pessaux and Véronique Viguié DonzeauGouge to be identified as the authors of this work have been asserted by them in accordance with the Copyright, Designs and Patents Act 1988. Library of Congress Control Number: 2021930488 British Library CataloguinginPublication Data A CIP record for this book is available from the British Library ISBN 9781786305305
Contents
Foreword . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
xi
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
xiii
Chapter 1. From Hardware to Software . . . . . . . . . . . . . . . . . . .
1
1.1. Computers: a lowlevel view . . . . . . . . . . 1.1.1. Information processing . . . . . . . . . . . 1.1.2. Memories . . . . . . . . . . . . . . . . . . . 1.1.3. CPUs . . . . . . . . . . . . . . . . . . . . . 1.1.4. Peripheral devices . . . . . . . . . . . . . . 1.2. Computers: a highlevel view . . . . . . . . . 1.2.1. Modeling computations . . . . . . . . . . . 1.2.2. Highlevel languages . . . . . . . . . . . . 1.2.3. From source code to executable programs
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
Chapter 2. Introduction to Semantics of Programming Languages 2.1. Environment, memory and state 2.1.1. Evaluation environment . . . 2.1.2. Memory . . . . . . . . . . . 2.1.3. State . . . . . . . . . . . . . . 2.2. Evaluation of expressions . . . . 2.2.1. Syntax . . . . . . . . . . . . 2.2.2. Values . . . . . . . . . . . . . 2.2.3. Evaluation semantics . . . . 2.3. Deﬁnition and assignment . . . 2.3.1. Deﬁning an identiﬁer . . . . 2.3.2. Assignment . . . . . . . . . . 2.4. Exercises . . . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
1 1 2 3 7 8 9 9 10 15
. . . . . . . . . . . .
16 16 18 20 21 21 22 24 26 26 29 31
vi
Concepts and Semantics of Programming Languages 1
Chapter 3. Semantics of Functional Features . . . . . . . . . . . . . . . 3.1. Syntactic aspects . . . . . . . . . . . . . . . . . . . . . 3.1.1. Syntax of a functional kernel . . . . . . . . . . . . 3.1.2. Abstract syntax tree . . . . . . . . . . . . . . . . . 3.1.3. Reasoning by induction over expressions . . . . . 3.1.4. Declaration of variables, bound and free variables 3.2. Execution semantics: evaluation functions . . . . . . 3.2.1. Evaluation errors . . . . . . . . . . . . . . . . . . . 3.2.2. Values . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.3. Interpretation of operators . . . . . . . . . . . . . 3.2.4. Closures . . . . . . . . . . . . . . . . . . . . . . . 3.2.5. Evaluation of expressions . . . . . . . . . . . . . . 3.3. Execution semantics: operational semantics . . . . . 3.3.1. Simple expressions . . . . . . . . . . . . . . . . . 3.3.2. Callbyvalue . . . . . . . . . . . . . . . . . . . . . 3.3.3. Recursive and mutually recursive functions . . . . 3.3.4. Callbyname . . . . . . . . . . . . . . . . . . . . . 3.3.5. Callbyvalue versus callbyname . . . . . . . . . 3.4. Evaluation functions versus evaluation relations . . . 3.4.1. Status of the evaluation function . . . . . . . . . . 3.4.2. Induction over evaluation trees . . . . . . . . . . . 3.5. Semantic properties . . . . . . . . . . . . . . . . . . . 3.5.1. Equivalent expressions . . . . . . . . . . . . . . . 3.5.2. Equivalent environments . . . . . . . . . . . . . . 3.6. Exercises . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
35 35 36 39 39 42 42 43 45 46 47 54 55 56 60 61 62 64 64 65 69 69 71 71
Chapter 4. Semantics of Imperative Features . . . . . . . . . . . . . . .
77
4.1. Syntax of a kernel of an imperative language 4.2. Evaluation of expressions . . . . . . . . . . . 4.3. Evaluation of deﬁnitions . . . . . . . . . . . 4.4. Operational semantics . . . . . . . . . . . . . 4.4.1. Bigstep semantics . . . . . . . . . . . . . 4.4.2. Smallstep semantics . . . . . . . . . . . 4.4.3. Expressiveness of operational semantics 4.5. Semantic properties . . . . . . . . . . . . . . 4.5.1. Equivalent programs . . . . . . . . . . . 4.5.2. Program termination . . . . . . . . . . . 4.5.3. Determinism of program execution . . . 4.5.4. Big steps versus small steps . . . . . . . 4.6. Procedures . . . . . . . . . . . . . . . . . . . 4.6.1. Blocks . . . . . . . . . . . . . . . . . . . 4.6.2. Procedures . . . . . . . . . . . . . . . . . 4.7. Other approaches . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
35
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
77 81 86 89 89 93 95 96 96 98 100 103 109 109 112 118
Contents
vii
4.7.1. Denotational semantics . . . . . . . . . . . . . . . . . . . . . . . . . 118 4.7.2. Axiomatic semantics, Hoare logic . . . . . . . . . . . . . . . . . . . 129 4.8. Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 Chapter 5. Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 5.1. Type checking: when and how? . . . . . . . . . . 5.1.1. When to verify types? . . . . . . . . . . . . . . 5.1.2. How to verify types? . . . . . . . . . . . . . . 5.2. Informal typing of a program Exp2 . . . . . . . . 5.2.1. A ﬁrst example . . . . . . . . . . . . . . . . . . 5.2.2. Typing a conditional expression . . . . . . . . 5.2.3. Typing without type constraints . . . . . . . . 5.2.4. Polymorphism . . . . . . . . . . . . . . . . . . 5.3. Typing rules in Exp2 . . . . . . . . . . . . . . . . 5.3.1. Types, type schemes and typing environments 5.3.2. Generalization, substitution and instantiation . 5.3.3. Typing rules and typing trees . . . . . . . . . . 5.4. Type inference algorithm in Exp2 . . . . . . . . . 5.4.1. Principal type . . . . . . . . . . . . . . . . . . 5.4.2. Sets of constraints and uniﬁcation . . . . . . . 5.4.3. Type inference algorithm . . . . . . . . . . . . 5.5. Properties . . . . . . . . . . . . . . . . . . . . . . . 5.5.1. Properties of typechecking . . . . . . . . . . . 5.5.2. Properties of the inference algorithm . . . . . 5.6. Typechecking of imperative constructs . . . . . . 5.6.1. Type algebra . . . . . . . . . . . . . . . . . . . 5.6.2. Typing rules . . . . . . . . . . . . . . . . . . . 5.6.3. Typing polymorphic deﬁnitions . . . . . . . . 5.7. Subtyping and overloading . . . . . . . . . . . . . 5.7.1. Subtyping . . . . . . . . . . . . . . . . . . . . . 5.7.2. Overloading . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . .
139 139 140 141 141 142 142 143 143 143 146 151 154 154 155 159 167 167 167 168 168 169 171 172 173 175
Chapter 6. Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179 6.1. Basic types . . . . . . . . . 6.1.1. Booleans . . . . . . . . 6.1.2. Integers . . . . . . . . . 6.1.3. Characters . . . . . . . 6.1.4. Floating point numbers 6.2. Arrays . . . . . . . . . . . . 6.3. Strings . . . . . . . . . . . 6.4. Type deﬁnitions . . . . . . 6.4.1. Type abbreviations . . . 6.4.2. Records . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
179 179 181 186 187 191 194 194 195 196
viii
Concepts and Semantics of Programming Languages 1
6.4.3. Enumerated types . . . . . 6.4.4. Sum types . . . . . . . . . 6.5. Generalized conditional . . . . 6.5.1. C style switch/case . . . . 6.5.2. Pattern matching . . . . . . 6.6. Equality . . . . . . . . . . . . . 6.6.1. Physical equality . . . . . . 6.6.2. Structural equality . . . . . 6.6.3. Equality between functions
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
200 202 205 205 208 216 217 218 220
Chapter 7. Pointers and Memory Management . . . . . . . . . . . . . . 223 7.1. Addresses and pointers . . . . . . . . 7.2. Endianness . . . . . . . . . . . . . . . 7.3. Pointers and arrays . . . . . . . . . . 7.4. Passing parameters by address . . . . 7.5. References . . . . . . . . . . . . . . . 7.5.1. References in C++ . . . . . . . . . 7.5.2. References in Java . . . . . . . . . 7.6. Memory management . . . . . . . . . 7.6.1. Memory allocation . . . . . . . . 7.6.2. Freeing memory . . . . . . . . . . 7.6.3. Automatic memory management
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
223 225 225 226 229 229 233 234 234 237 239
Chapter 8. Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243 8.1. Errors: notiﬁcation and propagation . . . . . . 8.1.1. Global variable . . . . . . . . . . . . . . . . 8.1.2. Record deﬁnition . . . . . . . . . . . . . . 8.1.3. Passing by address . . . . . . . . . . . . . . 8.1.4. Introducing exceptions . . . . . . . . . . . 8.2. A simple formalization: MLstyle exceptions 8.2.1. Abstract syntax . . . . . . . . . . . . . . . 8.2.2. Values . . . . . . . . . . . . . . . . . . . . . 8.2.3. Type algebra . . . . . . . . . . . . . . . . . 8.2.4. Operational semantics . . . . . . . . . . . . 8.2.5. Typing . . . . . . . . . . . . . . . . . . . . 8.3. Exceptions in other languages . . . . . . . . . 8.3.1. Exceptions in OCaml . . . . . . . . . . . . 8.3.2. Exceptions in Python . . . . . . . . . . . . 8.3.3. Exceptions in Java . . . . . . . . . . . . . . 8.3.4. Exceptions in C++ . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
243 245 245 245 246 247 247 248 248 248 250 250 251 251 253 254
Contents
Conclusion
ix
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
Appendix: Solutions to the Exercises . . . . . . . . . . . . . . . . . . . . 259 List of Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287 Index of Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289 References Index
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
Foreword
Computer programs have played an increasingly central role in our lives since the 1940s, and the quality of these programs has thus become a crucial question. Writing a highquality program – a program that performs the required task and is efﬁcient, robust, easy to modify, easy to extend, etc. – is an intellectually challenging task, requiring the use of rigorous development methods. First and foremost, however, the creation of such a program is dependent on an indepth knowledge of the programming language used, its syntax and, crucially, its semantics, i.e. what happens when a program is executed. The description of this semantics puts the most fundamental concepts into light, including those of value, reference, exception or object. These concepts are the foundations of programming language theory. Mastering these concepts is what sets experienced programmers apart from beginners. Certain concepts – like that of value – are common to all programming languages; others – such as the notion of functions – operate differently in different languages; ﬁnally, other concepts – such as that of objects – only exist in certain languages. Computer scientists often refer to “programming paradigms” to consider sets of concepts shared by a family of languages, which imply a certain programming style: imperative, functional, objectoriented, logical, concurrent, etc. Nevertheless, an understanding of the concepts themselves is essential, as several paradigms may be interwoven within the same language. Introductory texts on programming in any given language are not difﬁcult to ﬁnd, and a number of published books address the fundamental concepts of language semantics. Much rarer are those, like the present volume, which establish and examine the links between concepts and their implementation in languages used by programmers on a daily basis, such as C, C++, Ada, Java, OCaml and Python. The authors provide a wealth of examples in these languages, illustrating and giving life to the notions that they present. They propose general models, such as the kit
xii
Concepts and Semantics of Programming Languages 1
presented in Volume 2, permitting a uniﬁed view of different notions; this makes it easier for readers to understand the constructs used in popular programming languages and facilitates comparison. This thorough and detailed work provides readers with an understanding of these notions and, above all, an understanding of the ways of using the latter to create highquality programs, building a safer and more reliable future in computing. Gilles D OWEK Research Director, Inria Professor at the École normale supérieure, ParisSaclay Catherine D UBOIS Professor at the École nationale supérieure d’informatique pour l’industrie et l’entreprise January 2021
Preface
This twovolume work relates to the ﬁeld of programming. First and foremost, it is intended to give readers a solid grounding in the bases of functional or imperative programming, along with a thorough knowledge of the module and class mechanisms involved. In our view, the semantics approach is most appropriate when studying programming, as the impact of interlanguage syntax differences is limited. Practical considerations, determined by the material characteristics of computers and/or “smart” devices, will also be addressed. The same approach will be taken in both volumes, using both mathematical formulas and memory state diagrams. With this book, we hope to help readers understand the meaning of the constructs described in the reference manuals of programming languages and to establish solid foundations for reasoning and assessing the correctness of their own programs through critical review. In short, our aim is to facilitate the development of safe and reliable programs. Volume 1 begins with a presentation of the computer, in Chapter 1, ﬁrst at the material level – as an assemblage of components – then as a tool for executing programs. Chapter 2 is an intuitive, stepbystep introduction to language semantics, intended to familiarize readers with this approach to programming. In Chapter 3, we provide a detailed discussion on the subject, with a formal presentation of the execution semantics of functional features. Chapter 4 continues with the same topic, looking at the execution semantics of imperative features. In these two chapters, a clear mathematical framework is used to support our presentation. Also, all of the notions which we introduce in these chapters are implemented in both Python and OCaml to assist readers learning about the semantic concepts in question for the ﬁrst time. Multiple exercises, with detailed solutions, are provided in both cases. Chapter 5, on the subject of typing, begins by addressing typing rules, which are used to check programs; we then present the algorithm used to infer polymorphic types, along with the associated mathematical notions, all implemented in both languages. Finally, the extension of typing to imperative features is addressed. In Chapter 6, we
xiv
Concepts and Semantics of Programming Languages 1
present the main data types and methods of pattern matching, using a range of examples expressed in different programming languages. Chapter 7 focuses on lowlevel programming features: endianness, pointers and memory management; these notions are mostly presented using C and C++. Volume 1 ends with a discussion of error processing using exceptions, their semantics is presented in OCaml, and the exception management mechanisms used in Python, Java and C++ are also described (see Chapter 8). Thus, Volume 1 is intended to give a broad overview of the functional and imperative features of programming, from notions that can be modeled mathematically to notions that are linked to the hardware conﬁguration of computers themselves. Volume 2 focuses on modular and object programming, building on the foundations laid down in Volume 1 since modules, classes and objects are, in essence, the means of organizing functional or imperative constructs. Volume 2 ﬁrst analyzes the needs of developers in terms of tools for software architecture. Based on this study, an original semantic model, called a kit, is drawn up, jointly presenting all the features of the modules and objects that can meet these needs. The semantics of these kits are deﬁned in a rather informal way, as research in this ﬁeld has not yet led to a mathematical model of this set of features, while remaining relatively simple. From this model, we consider a set of emerging questions, the objective of which is to guide the acquisition of a language. This approach is then exempliﬁed by the study of the module systems of Ada, OCaml and C. Finally, the same approach will be used to deduce a semantic model of class and object features, which will serve to present classes in Java, C++, OCaml and Python from a uniﬁed perspective. This work is aimed at a relatively wide audience, from experienced developers – who will ﬁnd valuable additional information on language semantics – to beginners who have only written short programs. For beginners, we recommend working on the semantic concepts described in Volume 1 using the implementations in OCaml or Python to ease assimilation. All readers may beneﬁt from studying the reference manual of a programming language, while comparing the presentations of constructs given in the manual with those given here, guided by the questions mentioned in Volume 2. Note that we do not discuss the algorithmic aspect of data processing here. However, choosing the algorithm and the data representation that ﬁt the requirements of the speciﬁcation is an essential step in program development. Many excellent works have been published on this subject, and we encourage readers to explore the subject further. We also recommend using the standard libraries provided by the chosen programming language. These libraries include tried and tested implementations for many different algorithms, which may generally be assumed to be correct.
1 From Hardware to Software
This ﬁrst chapter provides a brief overview of the components found in all computers, from mainframes to the processing chips in tablets, smartphones and smart objects via desktop or laptop computers. Building on this hardwarecentric presentation, we shall then give a more abstract description of the actions carried out by computers, leading to a uniform deﬁnition of the terms “program” and “execution”, above and beyond the various characteristics of socalled electronic devices. 1.1. Computers: a lowlevel view Computer science is the science of rational processing of information by computers. Computers have the capacity to carry out a variety of processes, depending on the instructions given to them. Each item of information is an element of knowledge that may be transmitted using a signal and encoded using a sequence of symbols in conjunction with a set of rules used to decode them, i.e. to reconstruct the signal from the sequence of symbols. Computers use binary encoding, involving two symbols; these may be referred to as “true”/“false”, “0”/“1” or “high”/“low”; these terms are interchangeable, and all represent the two stable states of the electrical potential of digital electronic circuits. 1.1.1. Information processing Schematically, a computer is made up of three families of components as follows: – memories: store data (information) and executable code (the socalled von Neumann architecture); – one or more microprocessors, known as CPUs (central processing units), which process information by applying elementary operations;
Concepts and Semantics of Programming Languages 1: A Semantical Approach with OCaml and Python, First Edition. Thérèse Hardin, Mathieu Jaume, François Pessaux and Véronique Viguié DonzeauGouge. © ISTE Ltd 2021. Published by ISTE Ltd and John Wiley & Sons, Inc.
2
Concepts and Semantics of Programming Languages 1
– peripherals: these enable information to be exchanged between the CPU/memory couple and the outside. Information processing by a computer – in other terms, the execution of a program – can be summarized as a sequence of three steps: fetching data, computing the results and returning them. Each elementary processing operation corresponds to a conﬁguration of the logical circuits of the CPU, known as a logic function. If the result of this function is solely dependent on input, and if no notion of “time” is involved in the computations, then the function is said to be combinatorial; otherwise, it is said to be sequential. For example, a binary halfadder, as shown in Figure 1.1, is a circuit that computes the sum of two binary digits (input), along with the possible carry value. It thus implements a combinatorial logic function. Bit 0
or
Sum
and
Carry
Bit 1
Figure 1.1. Binary halfadder
The essential character of a combinatorial function is that, for the same input, the function always produces the same output, no matter what the circumstances. This is not true of sequential logic functions. For example, a logic function that counts the number of times its input changes relies on a notion of “time” (changes take place in time), and a persistent state between two inputs is required in order to record the previous value of the counter. This state is saved in a memory. For sequential functions, a same input value can result in different output values, as every output depends not only on the input, but also on the state of the memory at the moment of reading the new input. 1.1.2. Memories Computers use memory to save programs and data. There are several different technologies used in memory components, and a simpliﬁed presentation is as follows: – RAM (Random Access Memory): RAM memory is both readable and writeable. RAM components are generally fast, but also volatile: if electric power falls down, their content is lost;
From Hardware to Software
3
– ROM (Read Only Memory): information stored in a ROM is written at the time of manufacturing, and it is readonly. ROM is slower than RAM, but is nonvolatile, like, for example, a burned DVD; – EPROM (Erasable Programmable Read Only Memory): this memory is nonvolatile, but can be written using a speciﬁc device, through exposure to ultraviolet light, or by modifying the power voltage, etc. It is slower than RAM, for both reading and writing. EPROM may be considered equivalent to a rewritable DVD. Computers use the memory components of several technologies. Storage size diminishes as access speed increases, as fastaccess memory is more costly. A distinction is generally made between four different types of memory: – mass storage is measured in terabytes and is made either of mechanical disks (with an access time of ∼ 10 ms) or – increasingly – of solidstate drive (SSD) blocks. These blocks use an EEPROM variant (electrically erasable) with an access time of ∼ 0.1−0.3 ms, known as ﬂash memory. Mass storage is nonvolatile and is principally used for the ﬁle system; – RAM, which is external to the microprocessor. Recent home computers and smartphones generally possess large RAM capacities (measured in gigabytes). Embedded systems or consumer development electronic boards may have a much lower RAM capacity. The access time is around 40–50 ηs; – the cache is generally included in the CPU of modern machines. This is a small RAM memory of a few kilobytes (or megabytes), with an access time of around 5−10 ηs. There are often multiple levels of cache, and access time decreases with size. The cache is used to save frequently used and/or consecutive data and/or instructions, reducing the need to access slower RAM by retaining information locally. Cache management is complex: it is important to ensure consistency between the data in the main memory and the cache, between different CPUs or different cores (full, independent processing units within the same CPU) and to decide which data to discard to free up space, etc.; – registers are the fastest memory units and are located in the center of the microprocessor itself. The microprocessor contains a limited number (a few dozen) of these storage zones, used directly by CPU instructions. Access time is around one processor cycle, i.e. around 1 ns. 1.1.3. CPUs The CPU, as its name suggests, is the unit responsible for processing information, via the execution of elementary instructions, which can be roughly grouped into ﬁve categories: – data transfer instructions (copy between registers or between memory and registers);
4
Concepts and Semantics of Programming Languages 1
– arithmetic instructions (addition of two integer values contained in two registers, multiplication by a constant, etc.); – logical instructions (bitwise and/or/not, shift, rotate, etc.); – branching operations (conditional, nonconditional, to subroutines, etc.); – other instructions (halt the processor, reset, interrupt requests, testandset, compareandswap, etc.). Instructions are coded by binary words in a format speciﬁc to each microprocessor. A program of a few lines in a highlevel programming language is translated into tens or even hundreds of elementary instructions, which would be difﬁcult, error prone and time consuming to write out manually. This is illustrated in Figure 1.2, where a “Hello World!” program written in C is shown alongside its counterpart in x8664 instructions, generated by the gcc compiler.
# include i n t main printf return }
. s e c t i o n __TEXT . g l o b l _main . a l i g n 4 , 0 x90 _main : . cfi_startproc ## BB# 0 : pushq %r b p Ltmp0 : . c f i _ d e f _ c f a _ o f f s e t 16 Ltmp1 : . c f i _ o f f s e t %rbp , −16 movq %r s p , %r b p < s t d i o . h> Ltmp2 : () { . c f i _ d e f _ c f a _ r e g i s t e r %r b p ( " Hello world ! \ n " ) ; subq $16 , %r s p (0) ; leaq L_ . s t r (% r i p ) , %r d i movl $0 , −4(%r b p ) movb $0 , %a l callq _printf xorl %ecx , %e c x movl %eax , −8(%r b p ) movl %ecx , %e a x addq $16 , %r s p popq %r b p retq . cfi_endproc . s e c t i o n __TEXT L_ . s t r : . a s c i z " Hello world ! \ n "
Figure 1.2. “Hello world!” in C and in x8664 instructions
Put simply, a microprocessor is split into two parts: a control unit, which decodes and sequences the instructions to execute, and one or more arithmetic and logic units (ALUs) , which carry out the operations stipulated by the instructions. The CPU runs permanently through a threestage cycle:
From Hardware to Software
5
1) fetching the next instruction to be executed from the memory: every microprocessor contains a special register, the Program Counter (PC), which records the location (address) of this instruction. The PC is then incremented, i.e. the size of the fetched instruction is added to it; 2) decoding of the fetched instruction; 3) execution of this instruction. However, the next instruction is not always the one located next to the current instruction. Consider the function min in example 1.1, written in C, which returns the smallest of its two arguments. E XAMPLE 1.1.– C
int min (int a, int b) { if (a < b) return (a) ; else return (b) ; }
This function may be translated, intuitively and naively, into elementary instructions, by ﬁrst placing a and b into registers, then comparing them: min: load load compare
a, reg0 b, reg1 reg0, reg1
Depending on the result of the test – true or false – different continuations are considered. Execution continues using instructions for one or the other of these continuations: we therefore have two possible control paths. In this case, a conditional jump instruction must be used to modify the PC value, when required, to select the ﬁrst instruction of one of the two possible paths. branchgt load jump a_gt_b: load end: return
a_gt_b reg0, reg2 end reg1, reg2 reg2
The branchgt instruction loads the location of the instruction at label a_gt_b into the PC. If the result of the compare instruction is that reg0 > reg1, the next instruction is the one found at this address: load reg1, reg2. Otherwise, the next instruction is the one following branchgt: load reg0, reg2. This is followed by the unconditional
6
Concepts and Semantics of Programming Languages 1
jump instruction, jump, enabling unconditional modiﬁcation of the PC, loading it with the address of the end label. Thus, whatever the result of the comparison, execution ﬁnishes with the instruction return reg2. Conditional branching requires the use of a speciﬁc memory to determine whether certain conditions have been satisﬁed by the execution of the previous instruction (overﬂow, positive result, null result, superiority, etc.). Every CPU contains a dedicated register, the State Register (SR), in which every bit is assigned to signaling one of these conditions. Executing most instructions may modify all or some of the bits in the register. Conditional instructions (both jumps and more “exotic” variants) use the appropriate bit values for execution. Certain ARM® architectures [ARM 10] even permit all instructions to be intrinsically conditional. Every program is made up of functions that can be called at different points in the program and these calls can be nested. When a function is called, the point where execution should resume once the execution of the function is completed – the return address – must be recorded. Consider a program made up of the functions g() = k() + h() and f () = g() + h(), featuring several function calls, some of which are nested. g () = t11 = k () t12 = h () return t11 + t12 f () = v11 = g () v12 = h () return v11 + v12
A single register is not sufﬁcient to record the return addresses of the different calls. Calling k from g must be followed by calling h to evaluate t12. But this call of g was done by f, thus its return address in f should also be memorized to further evaluation of v12. The number of return addresses to record increases with the number of nested calls, and decreases as we leave these calls, suggesting very naturally to save these addresses in a stack. Figure 1.3 shows the evolution of a stack structure during successive function calls, demonstrating the need to record multiple return addresses. The state of the stack is shown at every step of the execution, at the moment where the line in the program is being executed. A dedicated register, the Stack Pointer (SP), always contains the address of the next free slot in the stack (or, alternatively, the address of the last slot used). Thus, in the case of nested calls, the return address is saved at the address indicated by the SP, and the SP is incremented by the size of this address. When the function returns, the PC is loaded with the saved address from the stack, and the SP is decremented accordingly.
From Hardware to Software
6 lin e
5
6 app
lin e
lin e
4
app
3
5 app
lin e
(Caller)
4
5 app
2
app
3
lin e
5 6
12
1
4
5 app
lin e
3
f () : v 11 = g () t 11 = k () t 12 = h () return t 11 + t v 12 = h () return v 11 + v 12
0
1 2
lin e
0
7
Figure 1.3. Function calls and return addresses
In summary, the internal state of a microprocessor is made up of its general registers, the program counter, the state register and the stack pointer. Note, however, that this is a highly simpliﬁed vision. There are many different varieties of microprocessors with different internal architectures and/or instruction sets (for example, some do not possess an integer division instruction). Thus, a program written directly using the instruction set of a microprocessor will not be executable using another model of microprocessor, and it will need to be rewritten. The portability of programs written in the assembly language of a given microprocessor is practically null. Highlevel languages respond to this problem by providing syntactic constructs, which are independent of the target microprocessors. The compiler or the interpreter have to translate these constructs into the language used by the microprocessor. 1.1.4. Peripheral devices As we saw in section 1.1.3, processors execute a constant cycle of fetching, decoding and executing instructions. Computations are carried out using data stored in the memory, either by the program itself or by an input/output mechanism. The results of computations are also stored in the memory, and may be returned to users using this input/output mechanism. The interest of any programmable system is inherently dependent on input/output capacities through which the system reacts to the external environment and may act on this environment. Even an assembly robot in a car factory, which repeats the same actions again and again, must react to data input from the environment. For example, the pressure of the grip mechanism must stop increasing once it has caught a bolt, and the time it takes to do this will differ depending on the exact position of the bolt. Input/output systems operate using peripherals, ancillary devices that may be electronic, mechanical or a combination of the two. These allow the microprocessor to acquire external information, and to transmit information to the exterior. Computer
8
Concepts and Semantics of Programming Languages 1
mice, screens and keyboards are peripherals used with desktop computers, but other elements such as motors, analog/digital acquisition cards, etc. are also peripherals. If peripherals are present, the microprocessor needs to devote part of its processing time to data acquisition and to the transmission of computed results. This interaction with peripherals may be directly integrated into programs. But in this case, the programs have to integrate regular checking of input peripherals to see if new information is available. It is technically difﬁcult (if not impossible) to include such a monitoring in every program. Furthermore, regular peripheral checks are a waste of time and energy if no new data is available. Finally, there is no guarantee that information would arrive exactly at the moment of checking, as data may be asynchronously emitted. This problem can be avoided by relying on the hardware to indicate the occurrence of new external events, instead of using software to check for these events. The interrupt mechanism is used to interrupt the execution of the current code and to launch the interrupt handler associated with the external event. This handler is a section of code, which is not explicitly called by the program being executed; it is located at an address known by the microprocessor. As any program may be interrupted at any point, the processor state, and notably the registers, must be saved before processing the interrupt. The code that is executed to process the interrupt will indeed use the registers and modify the SR, SP and PC. Therefore, previous values of registers must be restored in order to resume execution of the interrupted code. This context saving is carried out partially by the hardware and partially by the software. 1.2. Computers: a highlevel view The lowlevel vision of a von Neumann machine presented in section 1.1 provides a good overview of the components of a computer and of program execution, without going into detail concerning the operations of electronic components. However, this view is not particularly helpful in the context of everyday programming activity. Programs in binary code, or even assembly code, are difﬁcult to write as they need to take account of every detail of execution; they are, by nature, long and hard to review, understand and debug. The ﬁrst “highlevel” programming languages emerged very shortly after the ﬁrst computers. These languages assign names to certain values and addresses in the memory, providing a set of instructions that can be split into lowlevel machine instructions. In other terms, programming languages offer an abstract vision of the computer, enabling users to ignore lowlevel details while writing a program. The “hello world” program in Figure 1.2 clearly demonstrates the power of abstraction of C compared to the X86 assembly language.
From Hardware to Software
9
1.2.1. Modeling computations Any program is simply a description, in its own programming language, of a series of computations (including reading and writing), which are the only operations that a computer can carry out. An abstract view of a computer requires an abstract view – we call it a model – of the notion of computation. This subject was ﬁrst addressed well before the emergence of computers, in the late 19th century, by logicians, mathematicians and philosophers, who introduced a range of different approaches to the theory of calculability. The Turing machine [TUR 95] is a mathematical model of computation introduced in 1936. This machine operates on an inﬁnite memory tape divided into cells and has three instructions: move one cell of the tape right or left, write or read a symbol in the cell or compare the contents of two cells. It has been formally proven that any “imperative” programming language, featuring assignment, a conditional instruction and a while loop, has the same power of expression as this Turing machine. Several other models of the notion of algorithmic computation were introduced over the course of the 20th century, and have been formally proven to be equivalent to the Turing machine. One notable example is Kleene’s recursion theory [KLE 52], the basis for the “pure functional” languages, based on the notion of (potentially) recursive functions; hence, these languages also have the same power of expression as the Turing machine. Pure functional and imperative languages have developed in parallel throughout the history of highlevel programming, leading to different programming styles. 1.2.2. Highlevel languages Broadly speaking, the execution of a functional program carries out a series of function calls that lead to the result, with intermediate values stored exclusively in the registers. The execution of an imperative program carries out a sequence of modiﬁcations of memory cells named by identiﬁers, the values in the cells being computed during execution. The most widespread highlevel languages include both functional and imperative features, along with various possibilities (modules, object features, etc.) to divide source code into pieces that can be reused. Whatever the style of programming used, any program written in a highlevel language needs to be translated into binary language to be executed. These translations are executed either every time the program is executed – in which case the translation program is known as an interpreter or just once, storing the produced binary code – in which case the translator is known as a compiler. As we have seen, highlevel languages facilitate the coding of algorithms. They ease reviewing of the source code of a program, as the text is more concise than it
10
Concepts and Semantics of Programming Languages 1
would be for the same algorithm in assembly code. This does not, however, imply that users gain a better understanding of the way the program works. To write a program, a precise knowledge of the constructs used – in other terms, their semantics, what they do and what they mean – is crucial to understand the source code. Bugs are not always the result of algorithm coding errors, and are often caused by an erroneous interpretation of elements of the language. For example, the incrementation operator ++ in C exists in two forms (i++ or ++i), and its understanding is not as simple as it may seem. For example, the program: C
#include int main () { int i = 0 ; printf ("%d\n", i++) ; return (0) ; }
will print 0, but if i++ is replaced with ++i, the same program will print 1. There are a number of concepts that are common to all highlevel languages: value naming, organization of namespaces, explicit memory management, etc. However, these concepts may be expressed using different syntactic constructs. The ﬁeld of language semantics covers a set of logicomathematical theories, which describe these concepts and their properties. Constructing the semantics of a program allows to the formal veriﬁcation of whether the program possesses all of the required properties. 1.2.3. From source code to executable programs The transition from the program source to its execution is a multistep process. Some of these steps may differ in different languages. In this section, we shall give an overview of the main steps involved in analyzing and transforming source code, applicable to most programming languages. The source code of a program is made up of one or more text ﬁles. Indeed, to ease software architecture, most languages allow source code to be split across several ﬁles, known as compilation units. Each ﬁle is processed separately prior to the ﬁnal phase, in which the results of processing are combined into one single executable ﬁle. 1.2.3.1. Lexical analysis Lexical analysis is the ﬁrst phase of translation: it converts the sequence of characters that is indeed the source ﬁle into a sequence of words, assigning each to a category. Comments are generally deleted at this stage. Thus, in the following text presumed to be written in C
From Hardware to Software
11
/* This is a comment. */ if [x == 3 int +) cos ($v)
lexical analysis will recognize the keyword if, the opening bracket, the identiﬁer x, the operator ==, the integer constant 3, the type identiﬁer int, etc. No word in C can contain the character $, so a lexical error will be highlighted when $v is encountered. Lexical analysis may be seen as a form of “spell check”, in which each recognized word is assigned to a category (keyword, constant, identiﬁer). These words are referred to as tokens. 1.2.3.2. Syntactic analysis Every language follows grammar. For example, in English, a sentence is generally considered to be correctly formed if it contains a subject, verb and complement in an understandable order. Programming languages are no exception: syntactic analysis veriﬁes that the phrases of a source ﬁle conform with the grammar of their language. For example, in C, the keyword if must be followed by a bracketed expression, an instruction must end with a semicolon, etc. Clearly, the source text given in the example above in the context of lexical analysis does not respect the syntax of C. Technically, the syntactic analyzer is in charge of the complete grammatical analysis of the source ﬁle. It calls the lexical analyzer every time it requires a token to progress through the analyzed source. Syntactic analysis is thus a form of grammar veriﬁcation, and it also builds a representation of the source ﬁle by a data structure, which is often a tree, called the abstract syntax tree (AST). This data structure will be used by all the following phases of compilation, up to the point of execution by an interpreter or the creation of an executable ﬁle. 1.2.3.3. Semantic analyses The ﬁrst two analysis phases of compilation only concern the textual structure of the source. They do not concern the meaning of the program, i.e. its semantics. Source texts that pass the syntactic analysis phase do not always have meaning. The phrase “the sea eats a derivable rabbit” is grammatically correct, but is evidently nonsense. The bestknown semantic analysis is the typing analysis, which prohibits the combination of elements that are incompatible in nature. Thus, in the previous phase, “derivable” could be applicable to a function, but certainly not to a “rabbit”. Semantic analyses do not reduce to a form of typing analysis but they all interpret the constructs of a program according to the semantics of the chosen language. Semantic analyses may be used to eliminate programs, which leads to execution errors. They may also apply some transformations to program code in order to get an
12
Concepts and Semantics of Programming Languages 1
executable ﬁle (dependency analysis, closure elimination, etc.). These semantic analyses may be carried out during subsequent passes of source code processing, even after the code generation phase described in the following section. 1.2.3.4. Code interpretation/generation Once the abstract syntax tree (or a derived tree) has been created, there are two options. Either the tree may be executed directly via an interpreter, which is a program supplied by the programming language, or the AST is used to generate object code ﬁles, with the aim of creating an executable ﬁle that can be run independently. Let us ﬁrst focus on the second approach. The interpretation mechanism will be discussed later. Compilation uses the AST generated from the source ﬁle to produce a sequence of instructions to be executed either by the CPU or by a virtual machine (VM). The compilation is correct if the execution of this sequence of instructions gives a result, which conforms to the program’s semantics. Optimization phases may take place during or after object code generation, with the aim of improving its compactness or its execution speed. Modern compilers implement a range of optimizations, which study lies outside the scope of this book. Certain optimizations are “universal”, while others may be speciﬁc to the CPU for which the code is generated. The object code produced by the compiler may be either binary code encoding instructions directly or source text in assembly code. In the latter case, a program – known as the assembler – must be called to transform this lowlevel source code into binary code. Generally speaking, assemblers simply produce a mechanical translation of instructions written mnemonically (mov, add, jmp, etc.) into binary representations. However, certain more sophisticated assemblers may also carry out optimization operations at this level. Assembling mnemonic code into binary code is a very simple operation, which does not alter the structure of the program. The reference manual of the target CPU provides, for each instruction, the meaning of the bits of the corresponding binary word. For example, the reference manual for the MIPS32® architecture [MIP 13] describes the 32bit binary format of the instruction ADD rd, rs, rt (with the effect rd ← rs + rt on the registers) as: Bit weight
31
26
Bit value
000000
25
21 20
num. rs
16
num. rt
15
11
num. rd
10
000000100000
Figure 1.4. Coding the ADD instruction in MIPS32®
0
From Hardware to Software
13
Three packets of 6 bits are reserved for encoding the register numbers; the other bits in this word are ﬁxed and encode the instruction. The task of the assembler is to generate such bit patterns according to the instructions encountered in the source code. 1.2.3.5. Linking A single program may be made up of several source ﬁles, compiled separately. Once the object code from each source ﬁle has been produced, all these codes must be collected into a single executable ﬁle. Each object ﬁle includes “holes”, indicating unknown information at the moment of production of this object code. It is important to know where to ﬁnd this missing code, when calling functions deﬁned in a different compilation unit, or where to ﬁnd variables deﬁned in a location outside of the current unit. The linker has to gather all the object ﬁles and ﬁll all the holes. Evidently, for a set of object ﬁles to lead to an executable ﬁle, all holes must be ﬁlled; so the code of every function called in the source must be available. The linking process also has to integrate the needed code, if it comes from some libraries, whether from the standard language library or a thirdparty library. There is one ﬁnal question to answer, concerning the point at which execution should begin. In certain languages (such as C, C++ and Java), the source code must contain one, and only one, special function, often named main, which is called to start the execution. In other languages (such as Python and OCaml), deﬁnitions are executed in the order in which they appear, deﬁned by the ﬁle ordering during the linking process. Thus, “executing” the deﬁnition of a function does not call the function: instead, the “value” of this function is created and stored to be used later when the function is called. This means that programmers have to insert into the source ﬁle a call to the function which they consider to be the “starting point” of the execution. This call is usually the ﬁnal instruction of the last source ﬁle processed by the linker. A simpliﬁed illustration of the different transformation passes involved in source code compilation is shown in Figure 1.5.
source
lexemes lexical analysis
syntactic analysis
syntax tree
object code generation
executable link
Figure 1.5. Compilation process
1.2.3.6. Interpretation and virtual machines As we have seen, informally speaking, an interpreter “executes” a program directly from the AST. Furthermore, it was said that the code generation process may generate
14
Concepts and Semantics of Programming Languages 1
code for a virtual machine. In reality, interpreters rarely work directly on the tree; compilation to a virtual machine is often carried out as an intermediate stage. A virtual machine (VM) may be seen as a pseudomicroprocessor, with one or more stacks, registers and fairly highlevel instructions. The code for a VM is often referred to as bytecode. In this case, compilation does not generate a ﬁle directly executable by the CPU. Execution is carried out by the virtual machine interpreter, a program supplied by the programming language environment. So, the difference between interpretation and compilation is not clearcut. There are several advantages of using a VM: the compiler no longer needs to take the speciﬁcities of the CPU into account, the code is often more compact and portability is higher. As long as the executable ﬁle for the virtual machine interpreter is available on a computer, it will be possible to generate a binary ﬁle for the computer in question. The drawback to this approach is that the programs obtained in this way are often slower than programs compiled as “native” machine code.
2 Introduction to Semantics of Programming Languages
This chapter introduces intuitively the notions of name, environment, memory, etc., along with a ﬁrst formal description of these notions. It allows readers to familiarize themselves with the semantic approach of programming that we share with a number of other authors [ACC 92, DOW 09, DOW 11, FRI 01, WIN 93]. Any highlevel programming language uses names to denote the entities handled by programs. These names are generally known as identiﬁers, drawing attention to the fact that they are constructed in accordance with the syntactic rules of the chosen language. They may be used to denote programspeciﬁc values or values computed during execution. They may also denote locations (i.e. addresses in the memory), they are then called mutable variables. And identiﬁers can also denote operators, functions, procedures, modules, objects, etc., according to the constructs present in the language. For example, pi is often used to denote an approximate value of π; + is also an identiﬁer, denoting an addition operator and often placed between the two operands, i.e. in inﬁx position, as in 2 + 3. The expression 2 * x + 1 uses the identiﬁer x and to compute its value, we need to know the value denoted by x. Retrieving the value associated with a given identiﬁer is a mechanism at the center of any highlevel language. The semantics of a language provides a model of this mechanism, presented – in a simpliﬁed form – in section 2.1. All the formal deﬁnitions of languages, instructions, algorithms, etc., given in the following are coded in the programming languages OCaml and Python, trying to paraphrase these deﬁnitions and produce very similar versions of code in these two languages, even if developers in these languages may ﬁnd the programming style used here rather unusual. For readers not introduced to these languages, some very brief explanations are given in the codes’ presentation. But almost all features of OCaml and Python will be considered either in this ﬁrst volume or in the second,
Concepts and Semantics of Programming Languages 1: A Semantical Approach with OCaml and Python, First Edition. Thérèse Hardin, Mathieu Jaume, François Pessaux and Véronique Viguié DonzeauGouge. © ISTE Ltd 2021. Published by ISTE Ltd and John Wiley & Sons, Inc.
16
Concepts and Semantics of Programming Languages 1
where objectoriented programming is considered. We hope that these two encodings of formal notions can help readers who are not truly familiar with mathematical formalism. 2.1. Environment, memory and state 2.1.1. Evaluation environment Let X be a set of identiﬁers and V a set of values. The association of an identiﬁer x ∈ X with a value v ∈ V is called a binding (of the identiﬁer to its value), and a set Env of bindings is called an execution environment or evaluation environment. Env(x) denotes the value associated with the identiﬁer x in Env. The set of environments is denoted as E. In practice, the set of identiﬁers X that are actually used is ﬁnite: usually, we only consider those identiﬁers that appear in a program. An environment may thus be represented by a list of bindings, also called Env: [(x1 , Env(x1 )), (x2 , Env(x2 )), · · · , (xn , Env(xn ))] where {x1 , x2 , · · · , xn } denotes a ﬁnite subset of X, known as the domain of the environment and denoted as dom(Env). By convention, Env(x) denotes the value v, which appears in the ﬁrst (x, v) binding encountered when reading the list Env from the head (here, from left to right). In this model, a binding can be added to an environment using the operator ⊕. By convention, bindings are added at (the left of) the head of the list representing the environment: (xnew , vnew ) ⊕ [(x1 , v1 ), (x2 , v2 ), · · · , (xn , vn )] = [(xnew , vnew ), (x1 , v1 ), (x2 , v2 ), · · · , (xn , vn )] Suppose that a certain operation introduces a new binding of an identiﬁer, which is already present in the environment, for example (x2 , vnew ): (x2 , vnew ) ⊕ [(x1 , v1 ), (x2 , v2 ), · · · , (xn , vn )] = [(x2 , vnew ), (x1 , v1 ), (x2 , v2 ), · · · , (xn , vn )] The soobtained environment (x2 , vnew ) ⊕ Env contains two bindings for x2 . Searching for a binding starts at the head of the environment, and, with our convention, new bindings are added at the head. So the most recent addition, (x2 , vnew ), will be the ﬁrst found. The binding (x2 , v2 ) is not deleted, but it is said to be masked by the new binding (x2 , vnew ). Several bindings for a single identiﬁer x may therefore exist within the same environment, and the last binding added for x
Introduction to Semantics of Programming Languages
17
will be used to determine the associated value of x in the environment. Formally, the environment (x, v) ⊕ Env veriﬁes the following property: v if x = x ((x, v) ⊕ Env) (x ) = E nv(x ) if x = x By convention, the notation (x2 , v2 ) ⊕ (x1 , v1 ) ⊕ Env is used to denote the environment (x2 , v2 ) ⊕ ((x1 , v1 ) ⊕ Env). For example, ((x, v2 ) ⊕ (x, v1 ) ⊕Env)(x) = v2 When an environment is represented by a list of bindings, the value Env(x) is found as follows: Python
def valeur_de(env,x): for (x1,v1) in env: if x==x1: return v1 return None OCaml
let   val
rec valeur_de env x = match env with [] > None (x1, v1) :: t > if x = x1 then Some v1 else (valeur_de t x) valeur_de: (’a * ’b) list > ’a > ’b option
If no binding can be found in the environment for a given identiﬁer, this function returns a special value indicating the absence of a binding. In Python, the constant None is used to express this absence of value, while in OCaml, the predeﬁned sum type ’a option is used: OCaml
type ’a option = Some of ’a  None
The values of the type ’a option are either those of the type ’a or the constant None. The transformation of a value v of type ’a into a value of type ’a option is done by applying the constructor Some to v (see Chapter 5). The value None serves to denote the absence of value of type ’a; more precisely, None is a constant that is not a value of type ’a. This type ’a option will be used further to denote some kind of meaningless or absent values but that are needed to fully complete some deﬁnitions. The domain of an environment can be computed simply by traversing the list that represents it. A ﬁnite set is deﬁned here as the list of all elements with no repetitions.
18
Concepts and Semantics of Programming Languages 1
Python
def union_singleton(e,l): if e in l: return l else: return [e]+l def dom_env(env): r=[] for (x,v) in env: r = union_singleton(x,r) return r OCaml
let val let  val
rec union_singleton e l = if (List.mem e l) then l else e::l union_singleton : ’a > ’a list > ’a list rec dom_env env = match env with [] > []  (x, v) :: t > (union_singleton x (dom_env t)) dom_env : (’a * ’b) list > ’a list
Since the value returned by the function valeur_de is obtained by traversing the list from its head, adding a new binding (x, v) to an environment is done at the head of the list and the previous bindings of x (if any) are masked, but not deleted. Python
def ajout_liaison_env(env,x,v): return [(x,v)]+env OCaml
let ajout_liaison_env env x v = (x, v) :: env val ajout_liaison_env : (’a * ’b) list > ’a > ’b > (’a * ’b) list
2.1.2. Memory The formal model of the memory presented below makes no distinction between the different varieties of physical memory described in Chapter 1. The state of the memory is described by associating a value with the location of the cell in which it is stored. The locations themselves are considered as values, called references. As we have seen, highlevel languages allow us to name a location c, containing a value v, by an identiﬁer x bound to the reference r of c. Let R be a set of references and V a set of values. The association of a reference r ∈ R with a value v ∈ V is represented by a pair (r, v), and a set Mem of such pairs is called here a memory. Mem(r) denotes the value stored at the reference r in Mem. Let M be the set of memories. In practice, the set of references, which is actually used, is ﬁnite: once again, only those locations used by a program are generally considered. This means that the memory can be represented by a list, also called Mem: [(r1 , Mem(r1 )), (r2 , Mem(r2 )), · · · , (rn , Mem(rn ))]
Introduction to Semantics of Programming Languages
19
The existence of a pair (r, v) in a memory records that an initialization or a writing operation has been carried out at this location. Every referenced memory cell may be consulted through reading and can be assigned a new value by writing. In this case, the value previously held in the cell is deleted, it has “crashed”. Writing a value v at an address r transforms a memory Mem into a memory denoted as Mem[r := v]; if a value was stored at this location r in Mem, then this “old” value is replaced by v; otherwise, a new pair is added to Mem to take account of the writing operation. There is no masking, contrary to the case of the environments. Writing a new value v at a location r that already contains a value deletes the old value Mem(r): [(r1 , v1 ), · · · , (ri , vi ), · · · , (rn , vn )][ri := vi ] = [(r1 , v1 ), · · · , (ri , vi ), · · · , (rn , vn )] The domain of a memory dom(Mem) depends on the current environment and represents the space of references, which are accessible (directly or indirectly) from bound and nonmasked identiﬁers in the current execution environment. The addition of a binding (x, r) to an environment Env has a twofold effect, creating (x, r) ⊕ Env and extending Mem to Mem[r := v]. N OTE.– Depending on the language or program in question, the value v may be supplied just when x is introduced, or later, or never. If no value is provided prior to its ﬁrst use, the result of the program is unpredictable, leading to errors called initialization errors. Indeed, a location always contains a value, which does not need to be suited to the current computation if it has not been explicitly determined by the program. Note that the addition of a binding (x, r) in an environment Env, of which the domain contains x, may mask a previous binding of x in Env, but will not add a new pair (r, v) to Mem if r was already present in the domain of Mem. Thus, any list of pairs representing a memory cannot contain two different pairs for the same reference. The memory Mem[r := v] veriﬁes the following property: v if r = r M em[r := v](r ) = M em(r ) if r = r The memory (Mem[r1 := v1 ])[r2 := v2 ] is denoted as Mem[r1 := v1 ][r2 := v2 ]. For example, (Mem[r := v1 ][r := v2 ])(r) = v2 . The function valeur_ref computes the value stored at a given location. If nothing has previously been written at this location, the function returns a special value (None), indicating the absence of a known value (i.e. a value resulting from initialization or computation).
20
Concepts and Semantics of Programming Languages 1
Python
def valeur_ref(mem,a): if len(mem) == 0: return None else: a1,v1 = mem[0] if a == a1: return v1 else: return valeur_ref(mem[1:],a) OCaml
let   val
rec valeur_ref mem a = match mem with [] > None (a1, v1) :: t > if a = a1 then Some v1 else (valeur_ref t a) valeur_ref : (’a * ’b) list > ’a > ’b option
The following function writes a value into the memory: Python
def write_mem(mem,a,v): if len(mem) == 0: return [(a,v)] else: a1,v1 = mem[0] if a == a1: return [(a1,v)] + mem[1:] else: return [(a1,v1)] + write_mem(mem[1:],a,v) OCaml
let rec write_mem mem a v = match mem with  [] > [(a, v)]  (a1, v1) :: t > if a = a1 then (a1, v) :: t else (a1, v1) :: (write_mem t a v) val write_mem : (’a * ’b) list > ’a > ’b > (’a * ’b) list
2.1.3. State A state is deﬁned as a pair (Env, Mem) ∈ E × M such that any reference in the domain of Mem is accessible from a binding in Env. A reference is said to be accessible if its value can be read or written from an identiﬁer contained in Env by a series of operations of reading, writing, or reference manipulation. Given an environment Env, the set of identiﬁers X is partitioned into two subsets: cst Xref E nv , which contains the identiﬁers bound to a reference, and X E nv , which contains the others: Xref E nv = {x ∈ X  E nv(x) ∈ R}
Xcst E nv = {x ∈ X  E nv(x) ∈ V \ R}
The value associated with an identiﬁer x in Xref E nv is a reference E nv(x) = r where a value Mem(r) is stored, which can be modiﬁed by writing. Identiﬁers of Xref E nv are generally called mutable variables.
Introduction to Semantics of Programming Languages
21
2.2. Evaluation of expressions The value of an expression is computed according to an evaluation environment and a memory, i.e. in a given state. This computation is deﬁned by the evaluation semantics of the expression. 2.2.1. Syntax The language of expressions Exp1 used here will be extended in Chapters 3 and 4. Its syntax is deﬁned in Table 2.1. e ::=   e1 
k Integer constant (k ∈ Z) x Identiﬁer (x ∈ X) + e2 Addition (e1 , e2 ∈ Exp1 ) !x Dereferencing (x ∈ X)
Table 2.1. Language of expressions Exp1
Thus, an expression e ∈ Exp1 is either an integer constant k ∈ Z, an identiﬁer x ∈ X, an expression obtained by applying an addition operator to two expressions in Exp1 or an expression of the form !x denoting the value stored in the memory at the location bound to the mutable variable x. Thus, this is an inductive deﬁnition of the set Exp1 . Note that Exp1 does not include an assignment construct. This is a deliberate choice. This point will be discussed in greater detail in section 2.3 by means of an extension of Exp1 . N OTE.– The symbol + used in deﬁning the syntax of expressions does not denote the integer addition operator. It could be replaced by any other symbol (for example ). Its meaning will be assigned by the evaluation semantics. The same is true of the constant symbols: for example, the symbol 4 may be interpreted as a natural integer, a relative integer or a character. E XAMPLE 2.1.– !x + y is an expression of Exp1 in the same way as (x + 2) + 3. Parentheses are used here to structure the expression, they are part of the socalled concrete syntax and will disappear in the AST. The set Exp1 of wellformed expressions of the language is deﬁned by induction and expressed directly by a recursive sum type. Types of this kind can be constructed in OCaml, but not in Python; in the latter case, they can be partially simulated by deﬁning a class for each sumtype constructor. Each class must contain a method with arguments corresponding exactly to the arguments of the sum type constructors it implements. An implementation of this type in Python is naïve, and users must ensure that these classes are used correctly. We know that there are possibilities of programming dynamic type veriﬁcation mechanisms in Python, which simulate strong
22
Concepts and Semantics of Programming Languages 1
typing (similar to that used in OCaml) and ensure that the code is used correctly; however, these techniques lie outside of the scope of this book. The objective of all implementations shown in this book is simply to illustrate and intuitively justify the correct handling of concepts. As we have already done, we choose this approach to implement sum types. Using Python, we deﬁne the following classes to represent the constructors of the set Exp1 : Python
class Cste1: def __init__(self,cste): self.cste = cste class Var1: def __init__(self,symb): self.symb = symb
class Plus1: def __init__(self,exp1,exp2): self.exp1 = exp1 self.exp2 = exp2 class Bang1: def __init__(self,symb): self.symb = symb
For example, the expression e1 = !x + y deﬁned in example 2.1 is written as: Python
ex_exp1 = Plus1(Bang1("x"),Var1("y"))
Using OCaml, the type of arithmetic expressions is deﬁned directly as: OCaml
type ’a exp1 = Cste1 of int  Var1 of ’a  Plus1 of ’a exp1 * ’a exp1  Bang1 of ’a
Values of this type are thus obtained using either the Cste1 constructor applied to an integer value, in which case they correspond to a constant expression, or using the Var1 constructor applied to a value of type ’a, corresponding to the type used to represent identiﬁers (the type ’a exp1 is thus polymorphic, as it depends on another type), or by applying the Plus1 constructor to two values of the type ’a exp1, or by applying the Bang1 constructor to a value of type ’a. For example, the expression e1 = !x + y is written as: OCaml
let ex_exp1 = Plus1 (Bang1 ("x"), Var1 ("y")) val ex_exp1 : string exp1
2.2.2. Values Given a state (Env, Mem), we determine the evaluation semantics of an expression e ∈ Exp1 by computing the value of e in this state, i.e. by evaluating e in
Introduction to Semantics of Programming Languages
23
this state. Values may be relative integers or references, hence V = Z ∪ R. An additional, speciﬁc value Err is added to the set V; this result is returned as the value of “meaningless” expressions. The result of the evaluation of an expression in Exp1 will therefore be a value belonging to the set V = V ∪ {Err}. Values in V are either relative integers or references. By deﬁning a sum type, these two collections of values can be grouped into a single type. Python
class CInt1: def __init__(self,cst_int): self.cst_int = cst_int
class CRef1: def __init__(self,cst_adr): self.cst_adr = cst_adr
Each class possesses a (object) constructor with the same name as the class: the constant k obtained from integer n (or, respectively, from reference r) is thus written as CInt1(n) (respectively, CRef1(r)), and this integer (respectively, reference) can be accessed from (the object) k by writing k.cst_int (respectively k.cst_adr). With OCaml, the type of elements in V is deﬁned directly, as follows: OCaml
type ’a const1 = CInt1 of int  CRef1 of ’a
A value of this type is obtained either using the constructor CInt1 applied to an integer value or using the constructor CRef1 applied to a value of type ’a corresponding to the type used to represent references. A type grouping the elements of V = V ∪ {Err} is deﬁned by applying the same method: Python
class VCste1: def __init__(self,cste): self.cste = cste
class Erreur1: pass
An element v in V is either a value in V obtained from a constant k and written as VCste1(k), or an object in the class Erreur1 (pass is used here to express the fact that the (object) constructor has no argument). With OCaml, the type of the elements in V is deﬁned directly as follows: OCaml
type ’a valeurs1 = VCste1 of ’a const1  Erreur1
24
Concepts and Semantics of Programming Languages 1
2.2.3. Evaluation semantics There are several formalisms that may be used to describe the evaluation of an expression. These will be introduced later. Let us construct an evaluation function: ___ : E × M × Exp1 → V The evaluation of the expression e in the environment Env and memory state Mem em is denoted as eM E nv = v with v ∈ V. Table 2.2 contains the recursive deﬁnition of _ the function __ . em kM E nv = k
(k ∈ Z)
em xM E nv = E nv(x)
if x ∈ X and x ∈ dom(Env)
em xM E nv = Err
if x ∈ X and x ∈ / dom(Env)
em M em M em M em M em e1 + e2 M E nv = e1 E nv + e2 E nv if e1 E nv ∈ Z and e2 E nv ∈ Z em e1 + e2 M E nv = Err em !xM E nv =
em em if e1 M / Z or e2 M /Z E nv ∈ E nv ∈
M em( E nv(x))
em !xM E nv = Err
if x ∈ Xref E nv if x ∈ / Xref E nv
Table 2.2. Evaluation of the expressions of Exp1
The value of an integer constant is the integer that it represents. The value of an identiﬁer is that which is bound to it in the environment, or Err. The value of an expression constructed with an addition symbol and two expressions e1 and e2 is obtained by adding the relative integers resulting from the evaluations of e1 and e2 ; the result will be Err if e1 or e2 is not an integer. The value of !x is the value stored at the reference Env(x) when x is a mutable variable, and Err otherwise. Thus, if e is evaluated as a reference, then e can only be an identiﬁer. Furthermore, certain expressions in Exp1 are syntactically correct, but meaningless: for example, the expression !x when x is not a mutable variable, i.e. when x does not bind a reference in the environment, or x1 + x2 when x1 (or x2 ) is a mutable variable. On the other hand, !x + y is a meaningful expression that denotes a value when y binds an integer and x binds a reference to an integer. E XAMPLE 2.2.– Let us evaluate the expression !x + y in the state Env = [(x, rx ), (y, 2)] and Mem = [(rx , 3)]: em M em M em !x + yM E nv = !x E nv + y E nv = Mem(Env(x)) + Env(y) = Mem(rx ) + 2 = 3 + 2 = 5
Introduction to Semantics of Programming Languages
25
The evaluation function ___ : E × M × Exp1 → V is obtained directly as follows: Python
def eval_exp1(env,mem,e): if isinstance(e,Cste1): return VCste1(CInt1(e.cste)) if isinstance(e,Var1): x = valeur_de(env,e.symb) if isinstance(x,CInt1) or isinstance(x,CRef1): return VCste1(x) return Erreur1() if isinstance(e,Plus1): ev1 = eval_exp1(env,mem,e.exp1) if isinstance(ev1,Erreur1): return Erreur1() v1 = ev1.cste ev2 = eval_exp1(env,mem,e.exp2) if isinstance(ev2,Erreur1): return Erreur1() v2 = ev2.cste if isinstance(v1,CInt1) and isinstance(v2,CInt1): return VCste1(CInt1(v1.cst_int + v2.cst_int)) return Erreur1() if isinstance(e,Bang1): x = valeur_de(env,e.symb) if isinstance(x,CRef1): y = valeur_ref(mem,x.cst_adr) if y is None: return Erreur1() return VCste1(y) return Erreur1() raise ValueError OCaml
let rec eval_exp1 env mem e = match e with  Cste1 n > VCste1 (CInt1 n)  Var1 x > (match valeur_de env x with Some v > VCste1 v  _ > Erreur1)  Plus1 (e1, e2) > ( match ((eval_exp1 env mem e1), (eval_exp1 env mem e2)) with  (VCste1 (CInt1 n1), VCste1 (CInt1 n2)) > VCste1 (CInt1 (n1 + n2))  _ > Erreur1)  Bang1 x > (match valeur_de env x with  Some (CRef1 a) > (match valeur_ref mem a with Some v > VCste1 v  _ > Erreur1)  _ > Erreur1) val eval_exp1 : (’a * ’b const1) list > (’b * ’b const1) list > ’a exp1 > ’b valeurs1
Considering example 2.2, we obtain: Python
ex_env1 = [("x",CRef1("rx")),("y",CInt1(2))] ex_mem1 = [("rx",CInt1(3))] >>> (eval_exp1(ex_env1,ex_mem1,ex_exp1)).cste.cst_int 5
26
Concepts and Semantics of Programming Languages 1
OCaml
let ex_env1 = [ ("x", CRef1 ("rx")); ("y", CInt1 (2)) ] val ex_env1 : (string * string const1) list let ex_mem1 = [ ("rx", CInt1 (3)) ] val ex_mem1 : (string * ’a const1) list # (eval_exp1 ex_env1 ex_mem1 ex_exp1) ;;  : string valeurs1 = VCste1 (CInt1 5)
2.3. Deﬁnition and assignment 2.3.1. Deﬁning an identiﬁer The language Def 1 extends Exp1 by adding deﬁnitions of identiﬁers. There are two constructs that make it possible to introduce an identiﬁer naming a mutable or nonmutable variable (as deﬁned in section 2.1.3). Note that, in both cases, the initial value must be provided. This value corresponds to a constant or to the result of a computation speciﬁed by an expression e ∈ Exp1 . These constructs modify the em current state of the system; after computing eM E nv , the next step in evaluating M em let x = e; is to add the binding (x, eEnv ) to the environment, while the evaluation em of var x = e; adds a binding (x, rx ) to the environment and writes the value eM E nv to the reference rx . In this case, we assume that the location denoted by the reference rx is computed by an external mechanism responsible for memory allocation. d ::= let x = e; Deﬁnition of a nonmutable variable (x ∈ X, e ∈ Exp1 )  var x = e; Deﬁnition of a mutable variable (x ∈ X, e ∈ Exp1 ) Table 2.3. Language Def 1 of deﬁnitions
The evaluation of a deﬁnition is expressed as follows: let x = e; em (Env, Mem) −−−−−−−→Def 1 (x, eM E nv ) ⊕ E nv, M em var x = e; em (Env, Mem) −−−−−−−−→Def 1 ((x, rx ) ⊕ Env, Mem[rx := eM E nv ])
[2.1]
This evaluation →Def 1 deﬁnes a relation between a state, a deﬁnition and a resulting state, or, in formal terms: →Def 1 ⊆ (E × M) × Def 1 × (E × M) Starting with a ﬁnite sequence of deﬁnitions d = [d1 ; · · · ; dn ] and an initial state (Env0 , Mem0 ), this relation produces the state (Envn , Memn ): d
d
d
1 2 n (Env0 , Mem0 ) −→ Def 1 ( E nv1 , M em1 ) −→Def 1 · · · −→Def 1 ( E nvn , M emn )
Introduction to Semantics of Programming Languages
27
d
This sequence of transitions may, more simply, be noted (Env0 , Mem0 ) →Def 1 (Envn , Memn ). E XAMPLE 2.3.– Starting with a memory with no accessible references and an “empty” environment, the sequence [var y = 2; let x =!y + 3;] builds the following state: var y = 2; ([ ], [ ]) −−−−−−−−→Def 1 ([(y, ry )], [(ry , 2)]) let x =!y + 3; ([(y, ry )], [(ry , 2)]) −−−−−−−−−−→Def 1 ([(x, 5), (y, ry )], [(ry , 2)]) cst In the environment Env = [(x, 5), (y, ry )], we obtain Xref E nv = {y} and X E nv = {x}.
N OTE.– In the deﬁnition of the two transitions in [2.1], we presume that the result of em the evaluation of the expression e, denoted as eM E nv , is not an error result. In the case of an error, no state will be produced and the evaluation stops. The abstract syntax of language Def 1 may be deﬁned as follows: Python
class Let_def1: def __init__(self,var,exp): self.var = var self.exp = exp
class Var_def1: def __init__(self,var,exp): self.var = var self.exp = exp
OCaml
type ’a def1 = Let_def1 of ’a * ’a exp1  Var_def1 of ’a * ’a exp1
We choose to construct a value corresponding to a reference using a constructor applied to an identiﬁer. Python
class Ref_Var1: def __init__(self,idvar): self.idvar = idvar OCaml
type ’a refer = Ref_Var1 of ’a
Hence, rx will be represented by Ref_Var1(”x”). As the relation →Def 1 deﬁnes a function, it can be implemented directly as follows:
28
Concepts and Semantics of Programming Languages 1
Python
def trans_def1(st,d): (env,mem) = st if isinstance(d,Let_def1): v = eval_exp1(env,mem,d.exp) if isinstance(v,VCste1): return (ajout_liaison_env(env,d.var,v.cste),mem) raise ValueError if isinstance(d,Var_def1): v = eval_exp1(env,mem,d.exp) if isinstance(v,VCste1): r = Ref_Var1(d.var) return (ajout_liaison_env(env,d.var,CRef1(r)), write_mem(mem,r,v.cste)) raise ValueError raise ValueError OCaml
let trans_def1 (env, mem) d = match d with  Let_def1 (x, e) > (match eval_exp1 env mem e with  VCste1 v > ((ajout_liaison_env env x v), mem)  Erreur1 > failwith "Erreur")  Var_def1 (x, e) > (match eval_exp1 env mem e with  VCste1 v > ((ajout_liaison_env env x (CRef1 (Ref_Var1 x))), (write_mem mem (Ref_Var1 x) v))  Erreur1 > failwith "Erreur") val trans_def1 : (’a * ’a refer const1) list * (’a refer * ’a refer const1) list > ’a def1 > (’a * ’a refer const1) list * (’a refer * ’a refer const1) list
By iterating this function, we obtain an implementation of →Def 1 . Python
def trans_def1_exec(st,ld): (env,mem) = st if len(ld) == 0: return (env,mem) else: return trans_def1_exec(trans_def1((env,mem),ld[0]),ld[1:]) OCaml
let trans_def1_exec (env, mem) ld = (List.fold_left trans_def1 (env, mem) ld) val trans_def1_exec : (’a * ’a refer const1) list * (’a refer * ’a refer const1) list > ’a def1 list > (’a * ’a refer const1) list * (’a refer * ’a refer const1) list
Introduction to Semantics of Programming Languages
29
Now, considering example 2.3, we obtain: Python
ex_ld0 = [Var_def1("y",Cste1(2)), Let_def1("x",Plus1(Bang1("y"),Cste1(3)))] (ex_e0,ex_m0) = trans_def1_exec(([],[]),ex_ld0) >>> eval_exp1(ex_e0,ex_m0,Var1("x")).cste.cst_int 5 >>> eval_exp1(ex_e0,ex_m0,Bang1("y")).cste.cst_int 2 OCaml
let ex_ld0 = [ Var_def1 ("y", Cste1 2); Let_def1 ("x", Plus1 (Bang1 "y", Cste1 3)) ] val ex_ld0 : string def1 list # (trans_def1_exec ([], []) ex_ld0) ;;  : (string * string refer const1) list * (string refer * string refer const1) list = ([("x", CInt1 5); ("y", CRef1 (Ref_Var1 "y"))], [(Ref_Var1 "y", CInt1 2)])
2.3.2. Assignment The language Lang1 extends Def 1 by adding assignment. The syntax of an assignment instruction is: x := e where x ∈ X and e ∈ Exp1 . When the mutable variable x is already bound in the current environment, this instruction enables us to modify the value of !x. Formally, execution of the instruction x := e modiﬁes the memory of the current state, and it is described by the following transition: x:=e
em (Env, Mem) −−−→Lang1 (Env, Mem[Env(x) := eM E nv ])
N OTE.– Once again, if the identiﬁer x is not bound in the environment or if the evaluation of e results in an error, no state is generated and evaluation stops. E XAMPLE 2.4.– Based on the state obtained in example 2.3, the following two assignments can be executed: ([(x, 5), (y, ry )], [(ry , 2)]) y:=!y+x
y:=8
−−−−−→Lang1 ([(x, 5), (y, ry )], [(ry , 7)]) −−−→Lang1 ([(x, 5), (y, ry )], [(ry , 8)]) Representing the abstract syntax of the assignment x := e by the pair (x, e), the relation − →Lang1 and the iteration of this relation from a sequence of assignments are implemented as follows:
30
Concepts and Semantics of Programming Languages 1
Python
def trans_lang1(st,a): (env,mem) = st (x,e) = a v = valeur_de(env,x) if isinstance(v,CRef1): ve = eval_exp1(env,mem,e) if isinstance(ve,VCste1): return (env,write_mem(mem,v.cst_adr,ve.cste)) raise ValueError def trans_lang1_exec(st,la): (env,mem) = st if len(la) == 0:return (env,mem) else: return trans_lang1_exec(trans_lang1((env,mem),la[0]),la[1:])
OCaml
let trans_lang1 (env,mem) (x, e) = match valeur_de env x with  Some (CRef1 (Ref_Var1 y)) > (match eval_exp1 env mem e with  VCste1 v > (env, (write_mem mem (Ref_Var1 y) v))  Erreur1 > failwith "Eval error")  _ > failwith "Undefined var" val trans_lang1 : (’a * ’b refer const1) list * (’b refer * ’b refer const1) list > ’a * ’a exp1 > (’a * ’b refer const1) list * (’b refer * ’b refer const1) list let trans_lang1_exec (env, mem) la = (List.fold_left trans_lang1 (env, mem) la) val trans_lang1_exec : (’a * ’b refer const1) list * (’b refer * ’b refer const1) list > (’a * ’a exp1) list > (’a * ’b refer const1) list * (’b refer * ’b refer const1) list
Considering example 2.4, we obtain: Python
ex_la0 = [("y",Plus1(Bang1("y"),Var1("x"))), ("y",Cste1(8))] (ex_e1,ex_m1) = trans_lang1_exec((ex_e0,ex_m0),ex_la0) >>> eval_exp1(ex_e1,ex_m1,Bang1("y")).cste.cst_int 8 OCaml
let ex_la0 = [ ("y", Plus1 (Bang1 "y",Var1 "x")); ("y", Cste1 8) ] val ex_la0 : (string * string exp1) list # (trans_lang1_exec (trans_def1_exec ([],[]) ex_ld0) ex_la0);;  : (string * string refer const1) list * (string refer * string refer const1) list = ([("x", CInt1 5); ("y", CRef1 (Ref_Var1 "y"))], [(Ref_Var1 "y", CInt1 8)])
Introduction to Semantics of Programming Languages
31
N OTE.– In the presentation above, assignment concerns only mutable variables. Similarly, the dereferencing operator ! can only be applied to a reference to obtain the stored value at this location. Thus, the assignment of a value to a mutable variable from a given reference requires here the use of the dedicated syntactic structures: x := ! x + 1. However, in many languages, assignment does not explicitly mention the dereferencing operator, and the syntactic use of variables is identical on both sides of the assignment: x = x + 1. Hence, an identiﬁer x denotes two different notions: the value Env(x) on the left of the assignment symbol, and the value M em( E nv(x)) on the right of the assignment symbol. These languages mask the different roles of a variable according to its position in the assignment. When a variable x is positioned to the left of the assignment, it is known as an lvalue and it denotes a location where the value to be assigned should be stored. In this way, it acts as a pointer. The variable x on the right of the assignment is known as the rvalue, and implicitly acts as a dereferenced pointer: the value to fetch is found at the location denoted by the variable. Thus, in the expression x = x + 1, even if x is used in the same way from a syntactic perspective, the instance on the right implicitly denotes ! x. In some languages, variable declaration implicitly involves the creation of a reference, unless otherwise stated; the name of the variable represents the location where the compiler will store the values assigned to it. 2.4. Exercises Exercise 2.1 Consider the state etat0 deﬁned by: E nv0
= [(x, 3), (y, ry ), (x, 8), (z, rz )] and Mem0 = [(rz , 2), (ry , 5)]
cst 1) Compute Xref E nv0 and X E nv0 .
2) Give a sequence of deﬁnitions to obtain the state etat0 from the empty state ([ ], [ ]). Are there other sequences that can be used to obtain this state? em0 3) Compute (!y + !z) + xM E nv0 .
4) Give three examples of expressions that generate three different error types when evaluated in the state etat0 .
32
Concepts and Semantics of Programming Languages 1
5) Consider the sequence let y = x; y:=x etat0 −−−→Lang1 etat1 −−−−−−−→Def 1 etat2 . Determine the states etat1 = (Env1 , Mem2 ) and etat2 = (Env2 , Mem2 ) and the cst ref cst sets Xref E nv1 , X E nv1 , X E nv1 and X E nv1 . 6) Consider the sequence let x = y; x:=4 etat0 −−−−−−−→Def 1 etat3 −−−→Lang1 etat4 . Determine the states etat3 = (Env3 , Mem3 ) and etat4 = (Env4 , Mem4 ). Does em0 M em4 !yM E nv0 = !y E nv4 ? Exercise 2.2 The two constructs x + + and + + x: e = k  x  e1 + e2 !x  x + +  + + x are added to the language Exp1 . In these two new constructs, the identiﬁer x must be a mutable variable. Intuitively, we see that the evaluation of the expression x + + produces the value stored in the location denoted by x and increments the value in the memory by 1. The expression + + x is evaluated differently: the value stored at the location denoted by x is incremented by 1, and this new value is the result of the evaluation. 1) Deﬁne an evaluation function: ___ : E × M × Exp1 → V × (E × M) em for expressions in the language Exp1 , extended so that eM E nv = (v, ( E nv , M em )) expresses the fact that evaluation of the expression e in the state (Env, Mem) transforms this state into a state (Env , Mem ) and produces the value v.
2) Let Env = [(x, rx )] and Mem = [(rx , 6)]. Compute: em !x + (x + +)M E nv
et
em (x + +) + !xM E nv
em M em Does !x + (x + +)M E nv = (x + +) + !xE nv ?
3) Show that, for any state (Env, Mem): em M em + + xM E nv = (x + +) + 1E nv
Introduction to Semantics of Programming Languages
33
Now, we extend the language Lang1 by considering the expressions of the extended Def 1 language and adding the construction x+ := e. In informal terms, the execution of this instruction in a state (Env, Mem) consists of ﬁrst ﬁnding the value vx stored at the reference v in Mem, then evaluating the expression e in this state to obtain its value ve and a new state (Env , Mem ), and ﬁnally, assigning to v the result of the addition of ve and vx . If at least one of the two values ve and vx is not an integer, the execution fails. 4) Redeﬁne the relation − →Lang1 for the extended Lang1 language. 5) Let Env0 = [(x, rx )] and Mem0 = [(rx , 2)]. Determine the state etat1 such that x+:=++x (Env0 , Mem0 ) −−−−−−→Lang1 etat1 . 6) Do the assignments x := x + +, x := + + x and x+ := 1 produce the same states when executed in the same state?
3 Semantics of Functional Features
In this chapter, we examine the semantics of the constructs that form the kernel of functional languages. We shall present several methods for expressing the meaning of these constructs, along with different ways of describing their semantics. The functional language studied here corresponds to the functional kernel of MLstyle languages. Many other published presentations of the semantics of functional features of languages exist, for example [DOW 11, FRI 01]. 3.1. Syntactic aspects 3.1.1. Syntax of a functional kernel The language studied here, denoted as Exp2 , includes only one syntactic category: expressions. Its abstract syntax is deﬁned from a set X of (symbols of) variables, the set K = Z ∪ IB of constant values (where Z is the set of relative integers and IB the set of booleans) and a set of primitive operators. Table 3.1 presents the nine construction rules that enable the inductive deﬁnition of the set Exp2 . An expression in this language can therefore be: 1) a constant denoting a value of K, which will be either a relative integer or true or f alse; 2) an identiﬁer, here called a variable to maintain consistency with the usual terminology of expressions; 3) an application of a unary operator to an expression; 4) an application of a binary operator to two expressions with inﬁx notation; 5) a conditional expression;
Concepts and Semantics of Programming Languages 1: A Semantical Approach with OCaml and Python, First Edition. Thérèse Hardin, Mathieu Jaume, François Pessaux and Véronique Viguié DonzeauGouge. © ISTE Ltd 2021. Published by ISTE Ltd and John Wiley & Sons, Inc.
36
Concepts and Semantics of Programming Languages 1
6) an expression let x = e1 in e2 which declares the variable x, called the local variable in the expression e2 , and deﬁnes it by the expression e1 (in other words, the value of e1 is named x in the expression e2 ); 7) a function of a variable x, the body of which is an expression e, obtained by the abstraction of x in e. The variable x is generally called the parameter (or the formal parameter) of the function; 8) a recursive function, designated by an identiﬁer f , of which the parameter is x and the body is an expression e, which may contain recursive calls to f ; 9) an application of an expression e1 to an expression e2 (the expression e2 is called the argument or the actual parameter of e1 ). Constant (v ∈ K) Variable (x ∈ X) Application of a primitive unary operator  op e1 (op ∈ {not}) Application of a primitive binary operator  e1 op e2 (op ∈ {+, −, ∗, /, and, or, =, ≤})  if e1 then e2 else e3 Conditional expression  let x = e1 in e2 Local binding (x ∈ X)  fun x → e Function deﬁnition (x ∈ X)  rec f x = e Recursive function deﬁnition (x, f ∈ X)  (e1 e2 ) Application of a function e ::= v x
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Table 3.1. Syntax of the language of expressions Exp2
3.1.2. Abstract syntax tree Each expression in this language can be represented by a tree, the abstract syntax tree (AST), every node of which is annotated with the label of the applied construction rule. For example, the expression: fun x1 → x1 + x2
[3.1]
is obtained by abstraction of x1 in the expression x1 + x2 , which becomes the body of the function constructed in this way. Similarly, the expression ef act : rec f act n = if n ≤ 1 then 1 else (f act (n − 1)) ∗ n
[3.2]
recursively deﬁnes the function computing n! by abstraction of n in the function body.
Semantics of Functional Features
Its AST is:
37
(2)
(1) n 1 (2) (4) f act n−1 (2) (1) (9) (2) n 1 (f act (n − 1)) n (4) (1) (4) n ≤ 1 1 (f act (n − 1)) ∗ n (5) if n ≤ 1 then 1 else (f act (n − 1)) ∗ n (8) rec f act n = if n ≤ 1 then 1 else (f act (n − 1)) ∗ n
All operators introduced in the programs, written in Python and OCaml, end with 2 in reference to the name Exp2 . The set of constants K contains the relative integers and booleans, and can be represented as follows: Python
class CInt2: def __init__(self,cst_int): self.cst_int = cst_int
class CBool2: def __init__(self,cst_bool): self.cst_bool = cst_bool
OCaml
type const2 = CInt2 of int  CBool2 of bool
Let us deﬁne the operators used in expressions of Exp2 : Python
class Not2: pass
class Plus2: pass
class Minus2: pass
class Mult2: pass
class And2: pass
class Or2: pass
class Eq2: pass
class Leq2: pass
class Div2: pass
OCaml
type unary_op2 = Not2 type bin_op2 = Plus2  Minus2  Mult2  Div2  And2  Or2  Eq2  Leq2
Expressions of Exp2 can now be deﬁned as follows: Python
class Cste2: def __init__(self,cste): self.cste = cste class Var2: def __init__(self,symb): self.symb = symb class U_Op2: def __init__(self,op,exp1): self.op = op self.exp1 = exp1
class Letin2: def __init__(self,symb,exp1,exp2): self.symb = symb self.exp1 = exp1 self.exp2 = exp2 class Fun2: def __init__(self,param,corps): self.param = param self.corps = corps
38
Concepts and Semantics of Programming Languages 1
class B_Op2: def __init__(self,op,exp1,exp2): self.op = op self.exp1 = exp1 self.exp2 = exp2 class If2: def __init__(self,exp1,exp2,exp3): self.exp1 = exp1 self.exp2 = exp2 self.exp3 = exp3
class Rec2: def __init__(self,nom,param,corps): self.nom = nom self.param = param self.corps = corps class App2: def __init__(self,exp1,exp2): self.exp1 = exp1 self.exp2 = exp2
OCaml
type ’a exp2 =  Cste2 of const2   U_Op2 of unary_op2 * ’a exp2   If2 of ’a exp2 * ’a exp2 * ’a exp2   Fun2 of ’a * ’a exp2   App2 of ’a exp2 * ’a exp2
Var2 of ’a B_Op2 of bin_op2 * ’a exp2 * ’a exp2 Letin2 of ’a * ’a exp2 * ’a exp2 Rec2 of ’a * ’a * ’a exp2
For example, the expressions fun x1 → x1 + x2 , let x1 = x1 + x2 in x1 + x3 and the expression ef act deﬁned in [3.2] are written as: Python
ex_exp1 = Fun2("x1",B_Op2(Plus2(),Var2("x1"),Var2("x2"))) ex_exp2 = Letin2("x1",B_Op2(Plus2(),Var2("x1"),Var2("x2")), B_Op2(Plus2(),Var2("x1"),Var2("x3"))) a_fact = \ Rec2("fact","n", If2(B_Op2(Leq2(),Var2("n"),Cste2(CInt2(1))), Cste2(CInt2(1)), B_Op2(Mult2(),App2(Var2("fact"), B_Op2(Minus2(),Var2("n"),Cste2(CInt2(1)))), Var2("n")))) OCaml
let ex_exp1 = Fun2 ("x1", B_Op2 (Plus2, Var2 "x1", Var2 "x2")) val ex_exp1 : string exp2 let ex_exp2 = Letin2 ("x1", B_Op2 (Plus2, Var2 "x1", Var2 "x2"), B_Op2 (Plus2, Var2 "x1", Var2 "x3")) val ex_exp2 : string exp2 let a_fact = Rec2 ("fact", "n", If2 (B_Op2 (Leq2, Var2 "n", Cste2 (CInt2 1)), Cste2 (CInt2 1), B_Op2 (Mult2, App2 (Var2 "fact", B_Op2 (Minus2, Var2 "n", Cste2 (CInt2 1))), Var2 "n"))) val a_fact : string exp2
Semantics of Functional Features
39
3.1.3. Reasoning by induction over expressions The set Exp2 is deﬁned inductively: each expression e ∈ Exp2 is obtained by applying the construction rules of Exp2 a ﬁnite number of times. This process means that a property P can be proved over the elements of Exp2 by applying the structural induction scheme shown in Box 3.1: we simply prove the property P for each construct assuming that P is satisﬁed by each component of the construct. This reasoning by induction will be used several times in the remainder of the book. if ∀v ∈ K P (v) and ∀x ∈ X P (x) and ∀e1 ∈ Exp2 P (e1 ) ⇒ P (op e1 ) and ∀e1 , e2 ∈ Exp2 (P (e1 ) and P (e2 )) ⇒ P (e1 op e2 ) and ∀e1 , e2 , e3 ∈ Exp2 (P (e1 ) and P (e2 ) and P (e3 )) ⇒ P (if e1 then e2 else e3 ) and ∀x ∈ X ∀e1 , e2 ∈ Exp2 (P (e1 ) and P (e2 )) ⇒ P (let x = e1 in e2 ) and ∀x ∈ X ∀e ∈ Exp2 P (e) ⇒ P (fun x → e) and ∀x, f ∈ X ∀e ∈ Exp2 P (e) ⇒ P (rec f x = e) and ∀e1 , e2 ∈ Exp2 (P (e1 ) and P (e2 )) ⇒ P ((e1 e2 )) then ∀e ∈ Exp2 P (e) Box 3.1. Structural induction scheme over e ∈ Exp2
3.1.4. Declaration of variables, bound and free variables An identiﬁer can be used several times within the same expression, and each of its uses is called an occurrence of the identiﬁer. The occurrences of the same identiﬁer do not all play the same role. In construct (6) let x = e1 in e2 , the identiﬁer x is declared and bound to the value of e1 in the environment, which will be used to evaluate e2 . Construct (7) abstracts the variable x in the expression e. This parameter x is said to be bound (or abstracted) in the expression fun x → e. It can be renamed without modifying the deﬁnition of the function; thus, it is sometimes called a dummy variable. For example, replacing x with z in fun x → x + y gives fun z → z + y, which is equivalent. Evidently, we cannot replace x with y since y already occurs in the function body x + y. In construct (8), the identiﬁer f is declared and the variable x is bound in the expression e. For example, the expression: let x(1) = 25 in let f = fun x(2) → x(3) + 3 in (f 7) + x(4)
[3.3]
40
Concepts and Semantics of Programming Languages 1
contains four occurrences of x, indexed from 1 to 4. Occurrence 2 abstracts occurrence 3 in the body of f . These occurrences of the dummy variable x may be renamed. Occurrence 1 declares the identiﬁer x, which is used to give a value to occurrence 4. When an occurrence of a variable in an expression is neither declared nor bound, it is said to be free in this expression. For e ∈ Exp2 , the set F(e) of variables, which has at least one free occurrence in e, is deﬁned inductively by: F : Exp2 → ℘(X) F(v) = ∅ (v ∈ K) F(x) = {x} (x ∈ X) F(op e1 ) = F(e1 ) F(e1 op e2 ) = F(e1 ) ∪ F(e2 ) F(if e1 then e2 else e3 ) = F(e1 ) ∪ F(e2 ) ∪ F(e3 ) F(let x = e1 in e2 ) = F(e1 ) ∪ (F(e2 ) \ {x}) F(fun x → e) = F(e) \ {x} F(rec f x = e) = F(e) \ {x, f } F((e1 e2 )) = F(e1 ) ∪ F(e2 ) Expressions that do not contain a free variable are said to be closed. This is the case for a function fun x → e, which is said to be closed if it is a constant function (F(e) = ∅ and x does not occur in e) or if F(e) = {x} (i.e. the body e has no free variable except x). For example, the function fun x → x + 1 is closed. This is also the case for fun x1 → fun x2 → x1 + x2 , but not for fun x2 → x1 + x2 since x1 is free in this instance: the notion is not stable when switching to a subexpression (the fact that an expression is closed does not necessarily imply that its subexpressions are also closed). E XAMPLE 3.1.– For the expressions ef act deﬁned in [3.2] and e25 deﬁned in [3.3], we have F(ef act ) = F(e25 ) = ∅. In expression e25 , the occurrence 3 of x is free in the subexpression x + 3 but becomes bound in the expression fun x → x + 3. The following two examples are not closed expressions: F(fun x1 → x1 + x2 ) = {x2 } F(let x1 = x1 + x2 in x1 + x3 ) = {x1 , x2 , x3 } Intuitively, to evaluate an expression, we need to know the values of its variables. The value of a free variable in an expression is searched for in the evaluation environment. For example, to evaluate the expression x + 1, we need to know the value of x. The expression let x = 3 in 2 ∗ x, on the other hand, is closed and can be evaluated without “external” information concerning the (bound) variables it contains.
Semantics of Functional Features
41
The set of free variables of an expression in Exp2 can be deﬁned recursively as follows: Python
def free_var2(e): if isinstance(e,Cste2): return emptyset if isinstance(e,Var2): return [e.symb] if isinstance(e,U_Op2): return free_var2(e.exp1) if isinstance(e,B_Op2): return union(free_var2(e.exp1),free_var2(e.exp2)) if isinstance(e,If2): return union(union(free_var2(e.exp1),free_var2(e.exp2)), free_var2(e.exp3)) if isinstance(e,Letin2): return union(free_var2(e.exp1), diff_set(free_var2(e.exp2),[e.symb])) if isinstance(e,Fun2): return diff_set(free_var2(e.corps),[e.param]) if isinstance(e,Rec2): return diff_set(free_var2(e.corps),[e.param,e.nom]) if isinstance(e,App2): return union(free_var2(e.exp1),free_var2(e.exp2)) raise ValueError OCaml
let          val
rec free_var2 e = match e with Cste2 _ > emptyset Var2 x > [x] U_Op2 (op, e1) > free_var2 e1 B_Op2 (op, e1, e2) > union (free_var2 e1) (free_var2 e2) If2 (e1, e2, e3) > union (union (free_var2 e1) (free_var2 e2)) (free_var2 e3) Letin2 (x, e1, e2) > union (free_var2 e1) (diff_set (free_var2 e2) [x]) Fun2 (x, e1) > diff_set (free_var2 e1) [x] Rec2 (f, x, e1) > diff_set (free_var2 e1) [x;f] App2 (e1, e2) > union (free_var2 e1) (free_var2 e2) free_var2 : ’a exp2 > ’a list
Considering the expressions in example 3.1, we obtain: Python
>>> free_var2(ex_exp1) [’x2’] >>> free_var2(ex_exp2) [’x2’, ’x1’, ’x3’] >>> free_var2(a_fact) [] OCaml
# # # 
free_var2 ex_exp1 ;; : string list = ["x2"] free_var2 ex_exp2 ;; : string list = ["x1"; "x2"; "x3"] free_var2 a_fact ;; : string list = []
42
Concepts and Semantics of Programming Languages 1
3.2. Execution semantics: evaluation functions The execution semantics (or dynamic semantics) of a programming language describes the execution of all syntactically correct programs in this language. In this section, we study the “pure” functional language Exp2 . As operations acting on memory are not provided by this language, only the notion of an execution environment is required. Note that, unlike memory (with modiﬁable contents), an execution environment (also called evaluation environment) can only be extended by a new binding of an identiﬁer to its value; any existing bindings for this identiﬁer are simply masked by the new binding. A program in Exp2 is an expression. Its execution consists of evaluating the expression, i.e. of computing its value. To do this, we begin by deﬁning an evaluation function, as in Chapter 2: __ : E × Exp2 → V The value eEnv is the result of the evaluation of expression e in environment Here, an environment Env ∈ E is a list of bindings (x, v) ∈ X × V. The set V (deﬁned in section 3.2.2) is the set of values associated with the variables, and V is the union of V and of a set of special values used to indicate evaluation failures. E nv.
3.2.1. Evaluation errors We have chosen to distinguish between three different kinds of errors produced during the evaluation of an expression e ∈ Exp2 . 1) The value of a free occurrence of a variable in e must be found in the current execution environment. If this variable does not belong to the environment domain, evaluation of e returns an error denoted as ErrU (with U signifying unbound). 2) Arithmetic errors, such as overﬂow or mathematically undeﬁned values, are denoted as ErrE (E for exception, the mechanism used by the processor to indicate an “exceptional” situation which makes it impossible to continue normal execution). For our purposes, we shall limit our considerations to division by zero. 3) The expressions of Exp2 denote either integers, booleans or functions. Thus, attempts may be made to add a boolean and an integer (3 + true), or to apply an integer to an expression (8 (x + 1)). Such expressions are syntactically correct, but cannot denote a value in V. In our model, the evaluation of such expressions leads to an error called typing error ErrT . Detection of typing errors prior to execution will be presented in Chapter 5. Note that certain languages assign semantics to expressions, which we consider to be invalid in Exp2 . In Python (v3.6), for example, the expression 3 + True returns the value 4.
Semantics of Functional Features
43
The set of errors considered here is thus IE = {ErrU , ErrE , ErrT }. This choice is somewhat arbitrary: it would be possible to use only one sort of error, or, on the contrary, to make a ﬁner distinction between different causes of errors. This can be done by modifying the error handling rules. In Exp2 , the evaluation of some expressions e may not terminate, and in this case, the value eEnv is undeﬁned: no value – not even an error value – will be returned. The evaluation function __ is thus a partial function. The formal processing of nontermination lies outside the scope of this book. 3.2.2. Values The set V contains integers, booleans and the values of functions called closures, presented in section 3.2.4. For a nonrecursive function (fun x → e), the closure is a triple x, e, Env made up of the name of the function parameter (x), the function body (e) and the environment in which the deﬁnition of the function is evaluated (Env). For a recursive function (rec f x = e), the closure is a quadruple f, x, e, Env made up of the name of the function (f ), the name of the function parameter (x), the function body (e) and the environment in which the deﬁnition of the function is evaluated (Env). Hence, the set of closures IF is deﬁned by: IF = (X × Exp2 × E) ∪ (X × X × Exp2 × E) Every value bound to an identiﬁer in an environment is an element in the set V = Z ∪ IB ∪ IF. The set V = V ∪ IE comprises all possible results of the evaluation of an expression. Sets IE, V and V can be deﬁned as follows: Python
class E_Typ2: pass class E_Exec2: pass class E_Undef2: pass class VInt2: def __init__(self,val_int): self.val_int = val_int class VBool2: def __init__(self,val_bool): self.val_bool = val_bool class Clos2: def __init__(self,param,corps,envdef):
44
Concepts and Semantics of Programming Languages 1
self.param = param self.corps = corps self.envdef = envdef class ClosRec2: def __init__(self,nomf,param,corps,envdef): self.nomf = nomf self.param = param self.corps = corps self.envdef = envdef class Val2: def __init__(self,val): self.val = val class Err2: def __init__(self,err): self.err = err OCaml
type erreur2 = E_Typ2  E_Exec2  E_Undef2 type ’a valeur2 =  VInt2 of int  VBool2 of bool  Clos2 of ’a * ’a exp2 * (’a * ’a valeur2) list  ClosRec2 of ’a * ’a * ’a exp2 * (’a * ’a valeur2) list type ’a val_err2 = Val2 of ’a valeur v2 v1 ∈ /Z ErrT ErrT
Table 3.2. Interpretation of operators
ErrT ErrT
Semantics of Functional Features
45
3.2.3. Interpretation of operators The primitive operators in Exp2 are symbols that are used to construct new expressions from existing expressions. To evaluate an expression, we must decide which operation opV on the values of V is represented by each symbol op; in other terms, we must interpret these symbols. The interpretation of the operators used in Exp2 is deﬁned in Table 3.2. The interpretation of primitive operators is deﬁned as follows: Python
def apply2_unary_op(op,v): if isinstance(op,Not2): if isinstance(v,VBool2): return Val2(VBool2(not v.val_bool)) return Err2(E_Typ2) raise ValueError def apply2_binary_op(op,v1,v2): if isinstance(op,Plus2): if isinstance(v1,VInt2) and isinstance(v2,VInt2): return Val2(VInt2(v1.val_int + v2.val_int)) return Err2(E_Typ2) if isinstance(op,Minus2): if isinstance(v1,VInt2) and isinstance(v2,VInt2): return Val2(VInt2(v1.val_int  v2.val_int)) return Err2(E_Typ2) if isinstance(op,Mult2): if isinstance(v1,VInt2) and isinstance(v2,VInt2): return Val2(VInt2(v1.val_int * v2.val_int)) return Err2(E_Typ2) if isinstance(op,Div2): if isinstance(v1,VInt2) and isinstance(v2,VInt2): if v2.val_int == 0: return Err2(E_Exec2) return Val2(VInt2(v1.val_int / v2.val_int)) return Err2(E_Typ2) if isinstance(op,And2): if isinstance(v1,VBool2) and isinstance(v2,VBool2): return Val2(VBool2(v1.val_bool and v2.val_bool)) return Err2(E_Typ2) if isinstance(op,Or2): if isinstance(v1,VBool2) and isinstance(v2,VBool2): return Val2(VBool2(v1.val_bool or v2.val_bool)) return Err2(E_Typ2) if isinstance(op,Eq2): if isinstance(v1,VBool2) and isinstance(v2,VBool2): return Val2(VBool2(v1.val_bool == v2.val_bool)) if isinstance(v1,VInt2) and isinstance(v2,VInt2): return Val2(VBool2(v1.val_int == v2.val_int)) return Err2(E_Typ2)
46
Concepts and Semantics of Programming Languages 1
if isinstance(op,Leq2): if isinstance(v1,VInt2) and isinstance(v2,VInt2): return Val2(VBool2(v1.val_int (match v with  VBool2 b > Val2 (VBool2 (not b))  _ > Err2 E_Typ2) val apply2_unary_op : unary_op2 > ’a valeur2 > ’b val_err2 let apply2_binary_op op v1 v2 = match op with  Plus2 > (match (v1, v2) with  (VInt2 n1, VInt2 n2) > Val2 (VInt2 (n1 + n2))  _ > Err2 E_Typ2)  Minus2 > (match (v1, v2) with  (VInt2 n1, VInt2 n2) > Val2 (VInt2 (n1  n2))  _ > Err2 E_Typ2)  Mult2 > (match (v1, v2) with  (VInt2 n1, VInt2 n2) > Val2 (VInt2 (n1 * n2))  _ > Err2 E_Typ2)  Div2 > (match (v1, v2) with  (VInt2 n1, VInt2 n2) > if n2 = 0 then Err2 E_Exec2 else Val2 (VInt2 (n1 / n2))  _ > Err2 E_Typ2)  And2 > (match (v1, v2) with  (VBool2 b1, VBool2 b2) > Val2 (VBool2 (b1 && b2))  _ > Err2 E_Typ2)  Or2 > (match (v1, v2) with  (VBool2 b1, VBool2 b2) > Val2 (VBool2(b1  b2))  _ > Err2 E_Typ2)  Eq2 > (match (v1, v2) with  (VInt2 n1, VInt2 n2) > Val2 (VBool2 (n1 = n2))  (VBool2 b1, VBool2 b2) > Val2 (VBool2 (b1 = b2))  _ > Err2 E_Typ2)  Leq2 > (match (v1, v2) with  (VInt2 n1, VInt2 n2) > Val2 (VBool2 (n1 Err2 E_Typ2) val apply2_binary_op : bin_op2 > ’a valeur2 > ’b valeur2 > ’c val_err2
3.2.4. Closures The language Exp2 allows us to deﬁne functions and apply them to actual parameters, also called arguments. Let f be a function with formal parameter x and body ef . To evaluate (f e) in a current environment Enva , we must evaluate e in this environment to obtain a value v. The body ef must then be evaluated in an environment Env in which x is bound to v: this binding allows us to replace the formal parameter with the actual parameter. So what is this environment Env? Let f be the expression fun x1 → x1 + x2 : evaluation of (f 2) creates the binding (x1 , 2),
Semantics of Functional Features
47
and it remains to ﬁnd the value of x2 – but in which environment? Two choices are possible: this value will be supplied either by the environment Enva in which application (f 2) is evaluated, or by the environment in which f was deﬁned. In the ﬁrst case, we speak of dynamic scope, whereas in the second case, we speak of static or lexical scope. For Exp2 , we adopt lexical scope: closures therefore record the environment in which functions are deﬁned. In this way, the values of free variables of the function body (except the formal parameter) are those at the function deﬁnition stage, i.e. those bound to these variables in the deﬁnition environment of the function. Compilers generally work with lexical scope, whereas interpreters often choose dynamic scope. Compilers supply optimized implementations of the closure notion, avoiding code duplication. In a language using lexical scope, the function body is only evaluated when the function in question is applied to an actual parameter. This does not, however, imply that the source code of this body is retained; the function body is translated into object code, which contains the instructions needed to record the deﬁnition environment (at least the bindings of the free variables of the function in this environment), along with instructions that bind the formal parameter to the value of the actual parameter, provided by a call. Execution of application of a function to an actual parameter leads to execution of the object code of its body, in the right environment. 3.2.5. Evaluation of expressions Table 3.3 deﬁnes the evaluation function for the expressions in Exp2 by induction, following the rules of abstract syntax. Let us examine the rows in this table. The value of a constant is the constant itself. Binding a variable in an environment provides the value of that variable. To evaluate an expression of the form op e or e1 op e2 , we must evaluate the subexpressions e or e1 and e2 . The evaluation of a conditional expression if e1 then e2 else e3 depends on the result returned by the evaluation of e1 . To evaluate let x = e1 in e2 , we evaluate e1 getting v, then extend the current environment with the binding (x, v) to enable evaluation of e2 . The value of a function is a closure. If this function is recursive then the closure will, in addition, contain the name of the function f . In order to evaluate an application (e1 e2 ), we must ﬁrst evaluate e1 and e2 in the current environment Env. The function body is then evaluated in the environment Envf contained in the closure, extended by the binding (x, v). In the case of a recursive function f , adding the name f to the closure means that we can evaluate its body ef in an environment where f is bound to the closure. Another way of constructing the closure of a recursive function f (used in section 3.3.3) consists of extending the deﬁnition environment of f to include the binding of f with its own closure, making it possible to evaluate recursive calls: rec f x = eEnv = F
with F = x, e, (f, F ) ⊕ Env
48
Concepts and Semantics of Programming Languages 1
These two approaches are equivalent. eEnv
e v
v E nv(x)
x op e1
if e1 then e2 else e3
if x ∈ dom(Env)
(x ∈ X)
ErrU opV e1 Env
if x ∈ / dom(Env) if e1 Env ∈ V (op ∈ {not})
e1 Env
if e1 Env ∈ IE
e1 Env opV e2 Env e1 op e2
(v ∈ K)
(op ∈ {+, −, ∗, /, and, or, =, ≤}) if e1 Env ∈ V and e2 Env ∈ V
e1 Env
if e1 Env ∈ IE
e2 Env
if e1 Env ∈ V and e2 Env ∈ IE
e2 Env
if e1 Env = true
e3 Env
if e1 Env = f alse
ErrT
if e1 Env ∈ V \ IB
e1 Env
if e1 Env ∈ IE
let x = e1 in e2
e2 (x,v)⊕Env
if e1 Env = v ∈ V
e1 Env
if e1 Env ∈ IE
fun x → e
x, e, Env
rec f x = e
f, x, e, Env ef (x,v)⊕Envf
(e1 e2 )
ef (f,f,x,ef ,Envf )⊕(x,v)⊕Envf e2 Env ErrT
if e1 Env = x, ef , Envf and e2 Env = v ∈ V if e1 Env = f, x, ef , Envf and e2 Env = v ∈ V if e1 Env ∈ IF and e2 Env ∈ IE if e1 Env ∈ / IF
Table 3.3. Evaluation of expressions in Exp2
R EMARK.– Unlike the language Def 1 presented in Chapter 2, Exp2 does not supply the construct let x = e; required to add a binding to the current environment. However, the construct let x = e1 in e2 makes it possible to mimic it. To evaluate an expression e in an environment [(x1 , v1 ), (x2 , v2 ), · · · , (xn , vn )], we simply consider the following expression: let x1 = v1 in let x2 = v2 in ··· let xn = vn in e
Semantics of Functional Features
49
E XAMPLE 3.2.– The expression deﬁned in [3.3] is evaluated as follows: let x = 25 in let f = fun x → x + 3 in (f 7) + xEnv0 = let f = fun x → x + 3 in (f 7) + xEnv1 = (f 7) + xEnv2 = (f 7)Env2 + xEnv2 = x + 3(x,7)⊕Env1 + 25 = 7 + 3 + 25 = 35 where: E nv1 E nv2
= (x, 25) ⊕ Env0 since 25Env0 = 25 = (f, x, x + 3, Env1 ) ⊕ Env1 since fun x → x + 3Env1 = x, x + 3, Env1
Note that each occurrence of the variable x is processed in a way adapted to its meaning. E XAMPLE 3.3.– The expression (ef act 3) where ef act is the expression deﬁned in [3.2] is evaluated as follows: ((rec f act n = if n ≤ 1 then 1 else (f act (n − 1)) ∗ n) 3)Env0 ef act
= if n ≤ 1 then 1 else (f act (n − 1)) ∗ nEnv1 (since ef act Env0 = Ff act and 3Env0 = 3) = (f act (n − 1)) ∗ nEnv1 (since n ≤ 1Env1 = f alse since nEnv1 = 3) = (f act (n − 1))Env1 ∗ nEnv1 = if n ≤ 1 then 1 else (f act (n − 1)) ∗ nEnv2 ∗ nEnv1 (since f actEnv1 = Ff act and n − 1Env1 = nEnv1 − 1 = 3 − 1 = 2) = (f act (n − 1)) ∗ nEnv2 ∗ nEnv1 (since n ≤ 1Env2 = f alse since nEnv2 = 2) = (f act (n − 1))Env2 ∗ nEnv2 ∗ nEnv1 = if n ≤ 1 then 1 else (f act (n − 1)) ∗ nEnv3 ∗ nEnv2 ∗ nEnv1 (since f actEnv2 = Ff act and n − 1Env2 = nEnv2 − 1 = 2 − 1 = 1) = 1Env3 ∗ nEnv2 ∗ nEnv1 (since n ≤ 1Env3 = true since nEnv3 = 1) =1∗2∗3=6 where: Ff act = f act, n, if n ≤ 1 then 1 else (f act (n − 1)) ∗ n, Env0 E nv1 = (f act, Ff act ) ⊕ (n, 3) ⊕ E nv0 E nv2 = (f act, Ff act ) ⊕ (n, 2) ⊕ E nv0 E nv3 = (f act, Ff act ) ⊕ (n, 1) ⊕ E nv0 Note that each recursive call is evaluated in the environment Env0 contained in the closure Ff act , extended with the binding of n to its current value.
50
Concepts and Semantics of Programming Languages 1
The evaluation function for the expressions of Exp2 shown in Table 3.3 can be directly deﬁned as follows: Python
def eval2(env,e): if isinstance(e,Cste2): if isinstance(e.cste,CInt2): return Val2(VInt2(e.cste.cst_int)) if isinstance(e.cste,VBool2): return Val2(VBool2(e.cste.cst_bool)) if isinstance(e,Var2): x = valeur_de(env,e.symb) if x is None: return Err2(E_Undef2) return Val2(x) if isinstance(e,U_Op2): a = eval2(env,e.exp1) if isinstance(a,Val2): return apply2_unary_op(e.op,a.val) return a if isinstance(e,B_Op2): a1 = eval2(env,e.exp1) if isinstance(a1,Val2): a2 = eval2(env,e.exp2) if isinstance(a2,Val2): return apply2_binary_op(e.op,a1.val,a2.val) return a2 return a1 if isinstance(e,If2): a1 = eval2(env,e.exp1) if isinstance(a1,Val2): if isinstance(a1.val,VBool2): if a1.val.val_bool == True: return eval2(env,e.exp2) return eval2(env,e.exp3) return Err2(E_Typ2) return a1 if isinstance(e,Letin2): a1 = eval2(env,e.exp1) if isinstance(a1,Val2): ne = ajout_liaison_env(env,e.symb,a1.val) return eval2(ne,e.exp2) return a1 if isinstance(e,Fun2): return Val2(Clos2(e.param,e.corps,env)) if isinstance(e,Rec2): return Val2(ClosRec2(e.nom,e.param,e.corps,env)) if isinstance(e,App2): a1 = eval2(env,e.exp1) if isinstance(a1,Val2): if isinstance(a1.val,Clos2): a2 = eval2(env,e.exp2) if isinstance(a2,Val2): ne = ajout_liaison_env(a1.val.envdef,a1.val.param,a2.val) return eval2(ne,a1.val.corps) return a2 if isinstance(e,Fun2): return Val2(Clos2(e.param,e.corps,env)) if isinstance(e,Rec2): return Val2(ClosRec2(e.nom,e.param,e.corps,env))
Semantics of Functional Features
if isinstance(e,App2): a1 = eval2(env,e.exp1) if isinstance(a1,Val2): if isinstance(a1.val,Clos2): a2 = eval2(env,e.exp2) if isinstance(a2,Val2): ne = ajout_liaison_env(a1.val.envdef,a1.val.param,a2.val) return eval2(ne,a1.val.corps) return a2 if isinstance(a1.val,ClosRec2): a2 = eval2(env,e.exp2) if isinstance(a2,Val2): ne = ajout_liaison_env(ajout_liaison_env(a1.val.envdef, a1.val.nomf, a1.val), a1.val.param,a2.val) return eval2(ne,a1.val.corps) return a2 return Err2(E_Typ2) return a1 raise ValueError OCaml
let     

   
val
rec eval2 env e = match e with Cste2 (CInt2 n) > Val2 (VInt2 n) Cste2 (CBool2 n) > Val2 (VBool2 n) Var2 x > (match valeur_de env x with  Some v > Val2 v  None > Err2 E_Undef2) U_Op2 (op, e1)> (match eval2 env e1 with  Val2 v > apply2_unary_op op v  z > z) B_Op2 (op, e1, e2) > (match eval2 env e1 with  Val2 v1 > (match eval2 env e2 with  Val2 v2 > apply2_binary_op op v1 v2  z > z)  z > z) If2 (e1, e2, e3) > (match eval2 env e1 with  Val2 (VBool2 true) > eval2 env e2  Val2 (VBool2 false) > eval2 env e3  Val2 x > Err2 E_Typ2  z > z) Letin2 (x, e1, e2) > (match eval2 env e1 with  Val2 v > eval2 (ajout_liaison_env env x v) e2  z > z) Fun2 (x, e1) > Val2 (Clos2 (x, e1, env)) Rec2 (f, x, e1) > Val2 (ClosRec2 (f, x, e1, env)) App2 (e1, e2) > (match eval2 env e1 with  Val2 (Clos2 (pf, cf, ef)) > (match eval2 env e2 with  Val2 v > eval2 (ajout_liaison_env ef pf v) cf  z > z)  Val2 (ClosRec2 (nf, pf, cf, ef)) > (match eval2 env e2 with  Val2 v > eval2 (ajout_liaison_env (ajout_liaison_env ef nf (ClosRec2(nf,pf,cf,ef))) pf v) cf  z > z)  _ > Err2 E_Typ2) eval2 : (’a * ’a valeur2) list > ’a exp2 > ’a val_err2
51
52
Concepts and Semantics of Programming Languages 1
The evaluation of the expressions in examples 3.2 and 3.3 is obtained in the following way: Python
e25 = Letin2("x",Cste2(CInt2(25)), Letin2("f",Fun2("x",B_Op2(Plus2(),Var2("x"),Cste2(CInt2(3)))), B_Op2(Plus2(),App2(Var2("f"),Cste2(CInt2(7))),Var2("x")))) >>> (eval2([],e25)).val.val_int 35 >>> (eval2([],App2(a_fact,Cste2(CInt2(3))))).val.val_int 6 OCaml
let e25 = Letin2 ("x", Cste2 (CInt2 25), Letin2("f", Fun2 ("x", B_Op2 (Plus2, Var2 "x", Cste2 (CInt2 3))), B_Op2 (Plus2, App2 (Var2 "f", Cste2 (CInt2 7)), Var2 "x"))) val e25 : string exp2 # (eval2 [] e25);;  : string val_err2 = Val2 (VInt2 35) # (eval2 [] (App2(a_fact,Cste2(CInt2(3)))));;  : string val_err2 = Val2 (VInt2 6)
3.2.5.1. Order of evaluation of subexpressions The order in which the subexpressions of an expression are evaluated has an effect on the result of the ﬁnal evaluation. The evaluation function deﬁned in Table 3.3 arbitrarily establishes the order of evaluation of the subexpressions of e1 op e2 : we have chosen to evaluate e1 and then e2 . If evaluation of e1 triggers an error, then e2 will not be evaluated. The implementation of a mathematically commutative binary operator thus loses this property. For example, if a variable x E nv, then does not belong to the domain of environment (3 / 0) + xEnv = x + (3 / 0)Env . Indeed, we have: (3 / 0) + xEnv = 3 / 0Env = ErrE
x + (3 / 0)Env = xEnv = ErrU
The same goes for some socalled “lazy” boolean operators. For example, when evaluating e1 and e2 , if e1 is evaluated as f alse, the result f alse may be returned without e2 being evaluated (as the conjunction value will be f alse in any case); hence, execution errors in e2 will not be detected. On the other hand, there is no speciﬁed evaluation order for e1 and e2 in the evaluation of application (e1 e2 ). Nevertheless, when the semantics is implemented in a compiler or an interpreter, a choice must be made. For example, in the interpreter eval2 presented earlier, e1 is evaluated before e2 . In some “lazy” languages, argument e2 of a function is evaluated only when necessary (if e1 is a constant function, there is no need to evaluate the argument; this means that execution errors in the argument will not be detected).
Semantics of Functional Features
53
For other expressions, evaluation order is dictated by the construct semantics. For example, evaluation of if e1 then e2 else e3 begins with the evaluation of e1 and continues with evaluation of either e2 or e3 depending on the returned boolean value. Evidently, if the evaluation of expression e1 results in an error, then no subexpression will be evaluated. 3.2.5.2. Lexical and dynamic scope The same program may produce different results depending on the chosen scope mechanism. Consider the following expression: let x2 = 3 in let f = fun x1 → x1 + x2 in let x2 = 8 in (f 2)
[3.4]
Using lexical scope, this expression is evaluated as follows: let x2 = 3 in let f = fun x1 → x1 + x2 in let x2 = 8 in (f 2)Env = let f = fun x1 → x1 + x2 in let x2 = 8 in (f 2)(x2 ,3)⊕Env = let x2 = 8 in (f 2)(f,x1 ,x1 +x2 ,(x2 ,3)⊕Env)⊕(x2 ,3)⊕Env = (f 2)(x2 ,8)⊕(f,x1 ,x1 +x2 ,(x2 ,3)⊕Env)⊕(x2 ,3)⊕Env = x1 + x2 (x1 ,2)⊕(x2 ,3)⊕Env = 2 + 3 = 5 Using dynamic scope, the evaluation rules shown in Table 3.3 are used, except those concerning deﬁnition and application of functions: in this case, the deﬁnition environment of the function is not recorded in the closure. If an application is evaluated in an environment Enva , the function body is evaluated in the extension of E nva with the binding of the formal parameter to the value of the argument. In the previous example, this evaluation is done as follows: let x2 = 3 in let f = fun x1 → x1 + x2 in let x2 = 8 in (f 2)Env = let f = fun x1 → x1 + x2 in let x2 = 8 in (f 2)(x2 ,3)⊕Env = let x2 = 8 in (f 2)(f,x1 ,x1 +x2 )⊕(x2 ,3)⊕Env = (f 2)(x2 ,8)⊕(f,x1 ,x1 +x2 )⊕(x2 ,3)⊕Env = x1 + x2 (x1 ,2)⊕(x2 ,8)⊕(f,x1 ,x1 +x2 )⊕(x2 ,3)⊕Env = 2 + 8 = 10 The dynamic scope mechanism presents two drawbacks. First, if the body of a function contains free variables, its meaning depends on the current value (in Enva ) of the free variables that may change from one call to another. This is particularly problematic when attempting to review source code: the meaning of the function being modiﬁed along the execution. Furthermore, a free variable may be bound to a value of a different type in the course of execution, and this may lead to a typing error when evaluating the function application. For example, if x2 in expression [3.4] is associated with a boolean value instead of the integer value 8, then evaluating (f 2) will involve adding 2 to a boolean, resulting in an error. However, during the deﬁnition
54
Concepts and Semantics of Programming Languages 1
of function f , the value associated with x2 was an integer, 3, and the application of the operator + in the body of f does not pose any problem in this case. Function f was thus “correctly typed” during deﬁnition. Using the lexical scope mechanism, the values associated with the free occurrences of variables in a function body are determined by the environment in which the function is deﬁned. Thus, in the body of f in the previous example, the value of x2 will always be 3 and expression [3.4] will always be evaluated as 5. The lexical scope mechanism will be used throughout the remainder of the chapter. 3.3. Execution semantics: operational semantics Operational semantics describes the speciﬁcation of a language’s compiler through a set of inference rules as: (R)
h1 · · · hn (I) c
This rule, named R, means that, if hypotheses h1 , · · · , hn are true, we can deduce c, or, in an equivalent form, that c can be deduced by establishing h1 , · · · , hn . I corresponds to additional information (condition for application of the rule, or intermediate computation). For example, consider a rule specifying the evaluation of the addition of two expressions: (R)
E nv
e 1 v1 E nv e2 v2 (v = v1 + v2 ) E nv e1 + e2 v
In this case, Env e v expresses the fact that, in environment Env, the evaluation of expression e returns a value v. The order of evaluation of e1 and e2 is not speciﬁed, and the two rules below leave the choice of the error propagation to the (developer of the) compiler: (R1 )
e 1 v1 (v1 ∈ IE) e 1 + e2 v1
E nv E nv
(R2 )
e 2 v2 (v2 ∈ IE) e1 + e2 v2
E nv E nv
This is an evaluation relation, not a function. Function __ , deﬁned in section 3.2, implements rule R1 . Function eval2 presented in section 3.2 deﬁnes a very precise evaluation strategy, establishing an order for evaluating subexpressions. This function therefore provides a fully deterministic deﬁnition of the evaluation. Function __ , on the other hand, leaves certain choices open (for example the order in which the subexpressions of an application are to be evaluated) while ﬁxing others (such as the order of evaluation of the subexpressions of an operator).
Semantics of Functional Features
55
N OTE.– When the result of the evaluation of an expression depends on the choice of implementation used by the compiler to evaluate subexpressions, the expression itself is considered to be semantically incorrect. The value associated with the expression depends on the compiler. The reference handbooks of certain programming languages detail these situations. Operational semantics allows us to safely ignore certain constraints, such as the order of evaluation of arguments in an operator. For example, a compiler designer may prefer not to follow an imposed order for argument computation, often for reasons relating to optimization. The obligations of a compiler may be speciﬁed by using a semantics that establishes the meaning of constructs in a language, while maintaining a degree of freedom regarding certain implementation choices. The evaluation order for arguments of a commutative binary operator may, for example, be left unspeciﬁed in the semantics. This means that several different values may be associated with the same expression, for example if the semantics contains the two rules R1 and R2 deﬁned above. 3.3.1. Simple expressions The evaluation relation for simple expressions in language Exp2 , denoted as , is deﬁned using the inference system presented in Box 3.2, We have chosen to deﬁne an operational semantics, which respects the choices used in the function __ . This results in two equivalent semantics.
(F1 ) (F3 )
(F5 ) (F7 )
E nv
vv
(v ∈ K)
(F2 )
e1 v1 (v ∈ V) op e1 opV v1 1
E nv E nv
e1 v1 E nv e1 v1 E nv e2 v2 v1 ∈ V (v1 ∈ IE) (F4 ) E nv op e1 v1 E nv e1 op e2 v1 opV v2 v2 ∈ V E nv
e1 v1 E nv e1 v1 E nv e2 v2 v1 ∈ V (v1 ∈ IE) (F6 ) E nv e1 op e2 v1 E nv e1 op e2 v2 v2 ∈ IE E nv
E nv e1 true E nv e2 v E nv if e1 then e2 else e3 v
(F9 )
E nv
(F10 )
(F8 )
E nv e1 f alse E nv e3 v E nv if e1 then e2 else e3 v
E nv e1 v (v ∈ V \ IB) if e1 then e2 else e3 ErrT
E nv
E nv e1 v (v ∈ IE) if e1 then e2 else e3 v
Box 3.2. Operational semantics of simple expressions in Exp2
56
Concepts and Semantics of Programming Languages 1
Rule F1 expresses the fact that a constant in the language is evaluated as itself. Rules F2 , F3 , F4 , F5 and F6 indicate that the evaluation of the application of a unary or binary operator to one or two expressions is obtained by applying this operator to the result of the evaluation of the subexpressions involved. Of course, if an error is encountered when evaluating one of these expressions, any application of the operator will return the same error. Rules F7 , F8 , F9 and F10 deal with the evaluation of conditional expressions of the form if e1 then e2 else e3 . Depending on the result of the evaluation of e1 , either expression e2 or e3 will be evaluated. Once again, if an error is encountered when evaluating e1 , evaluation of the conditional expression will result in the same error. Furthermore, if the evaluation of e1 returns a nonboolean value, then a type error (ErrT ) will occur when evaluating the conditional expression. 3.3.2. Callbyvalue D EFINITION 3.1.– Callbyvalue: the callbyvalue mechanism deﬁnes a strategy for evaluating function applications, indicating that the argument (i.e. the actual parameter) must always be evaluated. This strategy is extended to the expressions let x = e1 in e2 , where e1 is always evaluated. The rules set out in Box 3.3 deﬁne the operational semantics of Exp2 . These rules apply a callbyvalue mechanism.
(x ∈ / dom(Env)) (x ∈ dom(Env)) (F12 ) x Env(x) E nv x ErrU E nv e1 v1 (x, v1 ) ⊕ E nv e2 v2 (v1 ∈ V) (F13 ) E nv let x = e1 in e2 v2 E nv e1 v1 (F14 ) (v1 ∈ IE) E nv let x = e1 in e2 v1 (F16 ) (F15 ) E nv fun x → e x, e, E nv E nv rec f x = e f, x, e, E nv E nv e1 x, ef , E nvf E nv e2 v2 (x, v2 ) ⊕ Envf ef v (F17 ) / IE) (v2 ∈ E nv (e1 e2 ) v E nv e1 f, x, ef , E nvf E nv e2 v2 (f, f, x, ef , Envf ) ⊕ (x, v2 ) ⊕ Envf ef v (F18 ) / IE) (v2 ∈ E nv (e1 e2 ) v E nv e1 v1 E nv e2 v2 (F19 ) (v1 ∈ IF, v2 ∈ IE) E nv (e1 e2 ) v2 E nv e1 v (v ∈ / IF) (F20 ) E nv (e1 e2 ) ErrT
(F11 )
E nv
Box 3.3. Operational semantics of Exp2 : callbyvalue
Semantics of Functional Features
57
Rules F11 and F12 are used to obtain the value associated with a variable x in an environment Env. If this variable does not belong to the domain of Env, the value ErrU is returned. Rules F13 and F14 describe the evaluation of expressions of the form let x = e1 in e2 in an environment Env. First, expression e1 in Env is evaluated, leading to a value noted v1 . Next, e2 is evaluated in Env, extended with the binding (x, v1 ), to obtain the ﬁnal result. If an error occurs when evaluating expression e1 , then evaluating expression let x = e1 in e2 will result in the same error. Rules F15 and F16 deﬁne the value of a function, i.e. a closure containing the current environment (the one in which the deﬁnition of the function takes place). Rules F17 and F18 describe the evaluation of an application (e1 e2 ). This evaluation will not produce an error if e1 is evaluated as a closure and if the evaluation of e2 does not result in an error. Rule F17 applies to the case of an expression e1 , which is evaluated as a nonrecursive function, i.e. as a closure of the form x, ef , Envf . Argument e2 is evaluated in the current environment (in which the call takes place), Env, giving a value v2 . The body of function e1 , i.e. ef , is evaluated in Envf , the deﬁnition environment of e1 extended by the binding of the formal parameter to the value of the actual parameter, (x, v2 ). Rule F18 applies to cases where e1 is evaluated as a recursive function, i.e. a closure of the form f, x, ef , Envf . The evaluation of this application is similar to that deﬁned by rule F17 , but Envf contains an additional binding (f, f, x, ef , Envf ), used to evaluate applications corresponding to recursive calls to the function in ef . Note the use of the lexical binding mechanism: the free variables in a function body keep the values they had when the function was deﬁned, no matter what happens later. E XAMPLE 3.4.– Again, let us consider the expression deﬁned in [3.3]: let x = 25 in let f = fun x → x + 3 in (f 7) + x e2 e1
The evaluation tree for this expression is given as: (F15 ) (F1 ) (F13 )
E nv0
25 25
(F13 )
E nv1
e2 x, x + 3, Env1 Ff
(x, 25) ⊕ Env0 e1 35 E nv1
E nv0
let x = 25 in let f = fun x → x + 3 in (f 7) + x 35
1
58
Concepts and Semantics of Programming Languages 1
where
1
is the evaluation tree: (F11 )
(F17 )
E nv2
f Ff
(F1 )
E nv2
77
2 (F11 ) (f 7) 10 E nv2 x 25 (f, Ff ) ⊕ Env1 (f 7) + x 35
E nv2
(F4 )
E nv2
where
2
is the evaluation tree: (F11 )
(F4 )
(F11 ) (x, 7) ⊕ Env1 x 7 (x, 7) ⊕ Env1 3 3 (x, 7) ⊕ Env1 x + 3 10
E XAMPLE 3.5.– The evaluation of (ef act 3) where ef act is the expression deﬁned in [3.2]: (rec f act n = if n ≤ 1 then 1 else (f act (n − 1)) ∗ n) e2 e1 ef act
is done as follows: (F16 )
E nv0
ef act f act, n, e1 , Env0
where
1
E nv0
33
1
Ff act
(F18 )
(F1 )
E nv0
(ef act 3) 6
is the evaluation tree:
(F11 ) (F4 ) (F1 ) E nv1 n 3 E nv1 n 3 E nv1 1 1 2 (F4 ) (F4 ) E nv1 n ≤ 1 f alse E nv1 e2 6 (F8 ) (f act, Ff act ) ⊕ (n, 3) ⊕ Env0 e1 6 E nv1
where
2
is the evaluation tree:
(F11 ) (F18 )
E nv1
f act Ff act
(F11 ) (F1 ) E nv1 n 3 E nv1 1 1 (F4 ) E nv1 n − 1 2 3
E nv1
(f act (n − 1)) 2
Semantics of Functional Features
where
3
59
is the evaluation tree:
(F11 ) (F4 ) (F1 ) E nv2 n 2 E nv2 n 2 E nv2 1 1 4 (F4 ) (F4 ) E nv2 n ≤ 1 f alse E nv2 e2 2 (F8 ) (f act, Ff act ) ⊕ (n, 2) ⊕ Env0 e1 2 where
4
E nv2
is the evaluation tree:
(F11 ) (F18 ) where
5
E nv2
f act Ff act
(F11 ) (F1 ) E nv2 n 2 E nv2 1 1 (F4 ) E nv2 n − 1 1 5
E nv2
(f act (n − 1)) 2
is the evaluation tree: (F4 ) (F4 )
(F1 ) n1 E nv3 1 1 (F1 ) E nv3 n ≤ 1 true E nv3 1 1 (f act, Ff act ) ⊕ (n, 1) ⊕ Env0 e1 1
E nv3
(F8 )
E nv3
R EMARK.– Termination of the evaluation mechanism: using the syntactic construct rec f x = e, expressions may be deﬁned for which evaluation does not terminate. For example, from expression ef oo deﬁned by rec f x = (f x)
[3.5]
we can construct the expression: let f oo = (rec f x = (f x)) in (f oo 2)
[3.6]
The evaluation of this expression does not terminate. The construction of the evaluation tree for this expression is neverending: (F16 ) E nv rec f x = (f x) f, x, (f x), E nv 1 (F13 )
Ff oo
let f oo = rec f x = (f x) in (f oo 2) where the construction of 1 is given as: (F11 ) (F1 ) E nv1 f oo Ff oo E nv1 2 2 2 (F18 ) (f oo, Ff oo ) ⊕ Env (f oo 2) E nv
E nv1
60
Concepts and Semantics of Programming Languages 1
where the construction of
2
is given as:
.. . (F18 ) (F11 ) (F11 ) E nv2 f Ff oo E nv2 x 2 E nv2 (f x) (F18 ) (f, Ff oo ) ⊕ (x, 2) ⊕ Env (f x) E nv2
This semantics can only be used to describe terminating executions. In other terms, Env e v expresses the fact that v is the value of e obtained by a ﬁnite evaluation tree, constructed using the inference rules presented in Box 3.2 and Box 3.3. There is no such evaluation tree for expression [3.6]. This semantics is called bigstep operational semantics: Env e v does not describe the computation, only the result. To describe and reason using computations, including nonterminating computations, the semantics may be formalized as smallstep semantics (or reduction semantics), a subject discussed at length in Chapter 4. 3.3.3. Recursive and mutually recursive functions Two occurrences of ef occur on either side of the symbol in rule F18 , which concerns the application of a recursive function. At ﬁrst glance, this may seem paradoxical: evaluating ef on the right leads us to evaluate ef on the left. As we saw in section 3.2.5, this question may be expressed directly by using rule F16 and binding f to the closure: F = x, e, (f, F ) ⊕ Env Once again, F appears on both sides of the equals sign. The existence of value F can be proven using the mathematical theory of the least ﬁxpoint (LFP), brieﬂy presented in section 4.7.1; a more detailed description can be found in [JAU 16]. Taking this approach, rules F16 , F17 and F18 can be condensed into two rules: (N16 )
E nv
rec f x = e x, e, (f, F ) ⊕ Env
E nvf
=F
E nv e1 x, e, E nvf E nv e2 v2 (x, v2 ) ⊕ Envf e v (N17_18 ) E nv (e1 e2 ) v
Applications of recursive and nonrecursive functions are based on the same mechanism. The LFP theory allows us to be sure that the evaluation of any expression with a ﬁnite number of recursive calls will terminate. Furthermore, the same theory can be used to prove that the two approaches deﬁne the same semantics for Exp2 .
Semantics of Functional Features
61
Now, let us consider the case of mutually recursive functions. For example, the following functions determine the parity of an integer: rec even n = if n = 0 then true else odd(n − 1) rec odd n = if n = 0 then f alse else even(n − 1) The body of function even contains a call to function odd, of which the body also contains a call to function even. These two functions are said to be mutually recursive, and some programming languages use a syntactic construct to express this form of dependency between two functions. When two functions are mutually recursive, it is impossible to give them two independent values: for example, the deﬁnition environment contained in the closure associated with function even must contain a binding for function odd and vice versa. One simple way of solving this problem is to handle it syntactically. For example, the construct let_rec f x = ef and g y = eg ; can be used to deﬁne two mutually recursive functions f and g. Evaluation of this construct simultaneously adds the bindings of f and g to the current environment. Considering the approach presented at the start of this section for closures, the evaluation of these deﬁnitions extends the current environment Env as follows: (g, G) ⊕ (f, F ) ⊕ Env with
F = x, ef , (g, G) ⊕ (f, F ) ⊕ Env G = y, eg , (g, G) ⊕ (f, F ) ⊕ Env
3.3.4. Callbyname D EFINITION 3.2.– Callbyname: the callbyname mechanism deﬁnes a strategy for evaluating function applications and other expressions. The subexpressions of an expression e will only be evaluated if their values are needed for the evaluation of e. To evaluate (f e), we evaluate f in the current environment Env. The expression e is only evaluated if the evaluation of the body of f in the deﬁnition environment Envf of f requires a value for e in Env. The pair e, Env, called a frozen expression, must therefore be retained. An environment thus becomes a list of associations, where the elements are pairs (x, e, Enve ), where x is a “frozen” variable, e is the expression deﬁning x and Enve is the environment in which e is to be evaluated: E nv
= [(x1 , e1 , Enve1 ), · · · , (xn , en , Enven )]
The notion of closure is thus generalized to all expressions, not only functions. Bigstep operational semantics with the callbyname strategy is deﬁned in Box 3.2 and Box 3.4. The evaluation rules for simple expressions, deﬁned in Box 3.2,
62
Concepts and Semantics of Programming Languages 1
are valid for both call strategies. For example, the evaluation of a conditional expression is always dependent on the evaluation of the condition, and the evaluation of an arithmetic expression requires us to evaluate at least one operand. The rules shown in Box 3.4 use the new notion of environment. To evaluate a variable x in an environment Env, which associates ex , Envx with x, we simply evaluate ex in the and F12 . To evaluate an expression of the environment Envx , i.e. apply the rules F11 form let x = e1 in e2 in an environment Env, we evaluate e2 in the environment (x, e1 , Env) ⊕ Env; this is an application of rule F13 . The evaluation of a function results in an additional closure. The formulation presented in section 3.3.3 is again and F15 . The application of a function f to an argument e in used here for rules F14 an environment Env ﬁrst evaluates f in Env as a closure x, ef , Envf , and then . evaluates ef in environment (x, e, Env) ⊕ Envf , following rule F16 (F11 )
E nvx E nv
ex v ) (x ∈ / dom(Env)) ((x, ex , Envx ) ∈ Env) (F12 xv E nv x ErrU ) (F13
) (F14
) (F15
) (F16
(x, e1 , Env) ⊕ Env e2 v E nv let x = e1 in e2 v E nv
fun x → e x, e, Env
E nv
rec f x = e x, e, (f, rec f x = e, Env) ⊕ Env
E nv
e1 x, ef , Envf (x, e2 , Env) ⊕ Envf ef v E nv (e1 e2 ) v ) (F17
e1 v (v ∈ / IF) (e1 e2 ) ErrT
E nv E nv
Box 3.4. Operational semantics of Exp2 : callbyname
3.3.5. Callbyvalue versus callbyname The two call strategies do not necessarily produce the same results. For example, the evaluation of expression ((fun x → 3) (8 / 0)) using a callbyvalue strategy operates as follows: (F1 ) (F1 ) E nv 8 8 E nv 0 0 (F15 ) (F4 ) E nv 8 / 0 ErrE E nv fun x → 3 x, 3, E nv (F19 ) E nv ((fun x → 3) (8 / 0)) ErrE
Semantics of Functional Features
63
If a callbyvalue strategy produces a ﬁnite evaluation tree, then a callbyname strategy will also produce a ﬁnite tree. On the other hand, obtaining a ﬁnite tree using a callbyname strategy does not imply that a ﬁnite tree will necessarily be produced using callbyvalue. For example, the evaluation of the expression ((fun x → 3) (ef oo 2)) does not terminate using a callbyvalue strategy (see section 3.3.2), whereas the value 3 is obtained using callbyname: (F1 ) ) (F14 E nv fun x → 3 x, 3, E nv (x, (ef oo 2), Env) ⊕ Env 3 3 (F16 ) E nv ((fun x → 3) (ef oo 2)) 3
The two strategies also differ in their capacity to share evaluations of subexpressions. For example, using callbyvalue, ((fun x → x + x) e) is evaluated as follows: .. . (R) (F15 ) E nv ef x, x + x, E nv E nv e ve (F17 ) E nv ((fun x → x + x) e) v ef
where
is the tree: (F11 )
(F4 )
(F11 ) x ve E nv1 x ve (x, ve ) ⊕ Env x + x v
E nv1
E nv1
Here, the value v is obtained by computing the value ve of e only once. Using a callbyname mechanism, on the other hand, we obtain: E nv e ve E nv e ve (F11 ) (F11 ) E nv1 x ve E nv1 x ve (F14 (F4 ) ) E nv ef x, x + x, E nv (x, e, Env) ⊕ Env x + x v E nv1 (F16 ) E nv ((fun x → x + x) e) v ef
The value ve of e is computed twice to obtain the value v of the expression x + x. The callbyneed strategy is an implementation of the callbyname strategy which attempts to maximize sharing of subexpressions, in order to evaluate the arguments of functions once only. R EMARK.– In languages that allow side effects (reading/writing/assignment), callbyvalue is often used because this strategy makes it possible to predict when
64
Concepts and Semantics of Programming Languages 1
side effects will occur; this is not the case with a callbyname strategy. Compilers for languages using a callbyname strategy research and evaluate subexpressions producing side effects using a callbyvalue strategy. 3.4. Evaluation functions versus evaluation relations In this section, we examine the determinism of the evaluation function with the callbyvalue strategy. This leads us to study a new form of induction. 3.4.1. Status of the evaluation function The question examined here is whether, for every expression e ∈ Exp2 , there exists a single value v ∈ V such that eEnv = v. The answer is no. If e is an expression that requires evaluation of the application of a nonterminating recursive function (such as expression [3.6], called “foo”, in section 3.3.2), then we cannot obtain a value v ∈ V such that eEnv = v. So __ is not a total function. However, if such a value v exists, it is provably unique (thus __ is a partial function): ∀e ∈ Exp2 ∀Env ∈ E ∀v, v ∈ V eEnv = v and eEnv = v ⇒ v = v P (e)
Let us call this property P . We may try to prove that P holds for all expressions e ∈ Exp2 by induction over e, using the induction scheme deﬁned in Box 3.1. This is possible for most of the syntactic constructs, for example for expressions of the form let x = e1 in e2 . Let Env be an environment and v and v two values. If let x = e1 in e2 Env = v and let x = e1 in e2 Env = v , then we can prove that v = v . First, note that by induction hypothesis, P (e1 ) is veriﬁed, and thus if e1 Env = v1 and e1 Env = v1 , then v1 = v1 . Two cases are now possible. If v1 ∈ V, then, by deﬁnition, v = e2 (x,v1 )⊕Env and v = e2 (x,v )⊕Env and, since 1 v1 = v1 , by induction hypothesis 1 over e2 , we obtain v = v . Otherwise, if v1 ∈ IE, then by deﬁnition, v = v1 and v = v1 , and since v1 = v1 , we directly obtain v = v . Similar reasoning can be used to prove P for all constructs with the exception of function applications, i.e. expressions of the form (e1 e2 ). In this case, the only available induction hypotheses are related to the subexpressions e1 and e2 in (e1 e2 ). However, when e1 is evaluated as a closure containing the expression ef , the evaluation of the application (e1 e2 ) leads to an evaluation of the body ef , which is 1 Note that the order of quantiﬁcations in the statement P impacts the form of the induction hypotheses. For example, quantifying ﬁrst over the environment (∀Env ∈ E ∀e ∈ Exp2 ∀v, v ∈ V) allows us to use the induction hypothesis over e1 since e1 is evaluated in E nv but forbids us to use the induction hypothesis over e2 as e2 is evaluated in an environment that is different to Env.
Semantics of Functional Features
65
not a subexpression of (e1 e2 ) (for example, e1 may be only the name of the function). We cannot suppose that P holds for ef and we cannot conclude. Thus, it is not possible to prove P (e) through reasoning by structural induction over the expression e. This property cannot be proved by induction over the structure of expressions and requires an induction over the “size” of the computations involved in evaluation. Considering the evaluation function __ , this can be done using the least ﬁxpoint theory for continuous functions, as described in section 4.7.1. The property will be proved here using bigstep operational semantics, which allows us to reason by induction over the evaluation trees; this is a way to consider the “size” of the computations involved in evaluation. 3.4.2. Induction over evaluation trees Our aim here is to prove a property P which states that if an expression is evaluated as a value, then this value is unique. Formally, within the framework of bigstep operational semantics, we wish to prove the following proposition: P ROPOSITION 3.1.– Determinism: ∀e ∈ Exp2 ∀Env ∈ E ∀v, v ∈ V (Env e v and Env e v ) ⇒ v = v The set of evaluation (or inference) trees for the expressions in Exp2 can be deﬁned inductively, which means that we can reason by induction over these trees, i.e. over the “size” of the computations carried out during evaluations. More generally, given an inference system S, i.e. a set of inference rules of the form (R)
e1 · · · e n e
the set AS of elements e that can be inferred (sometimes called the set of theorems of S) is deﬁned inductively as follows: – for every axiom in S, i.e. for every rule of the form (R) , the conclusion e e belongs to AS ; – if e1 , · · · , en are elements in AS , and if (R)
e1 · · · en e
66
Concepts and Semantics of Programming Languages 1
is a rule inS, then e ∈ AS . Indeed if e1 , · · · , en are elements in AS , then there exist trees 1 , · · · , n for each of these elements, and we can therefore construct the following tree: 1
(R)
···
n
en
e1 e
Proving a property P over the elements of AS comes down to proving that this property holds for all elements concluding an axiom in S and proving, for each rule (R)
e1 · · · en e
in S that if all e1 , · · · , en satisfy the property P , then e also satisﬁes the property P . This results in an inductive reasoning scheme built on the inference tree of e. In the case of evaluation trees for the expressions in Exp2 , this enables us to reason by induction over the computations carried out during the evaluation. Using this technique, and taking SExp2 to be the inference system deﬁning the bigstep operational semantics of the expressions of Exp2 , proposition 3.1 can be proved by considering the following equivalent formulation: ∀ Env e v ∈ ASExp2 ∀v ∈ V Env e v ∈ ASExp2 ⇒ v = v [3.7] P ( E nvev)
P ROOF.– Let P be the property to prove and let us show that this property is satisﬁed by all evaluation trees Env e v ∈ ASExp2 . If Env e v has been obtained using rule F1 , then e = v is a constant in K, and since F1 is the only rule that can be used to evaluate a constant, v = v . If Env e v has been obtained using rule F2 , then e = op e1 , Env e1 v1 , v1 ∈ V and v = opV v1 . By induction hypothesis, if Env e1 v1 , then v1 = v1 ∈ V and since F2 is the only applicable rule for evaluating e, we obtain v = opV v1 = opV v1 = v. If Env e v has been obtained using rule F3 , then e = op e1 , Env e1 v1 , v1 ∈ IE and v = v1 . By induction hypothesis, if Env e1 v1 , then v1 = v1 ∈ IE. F3 is the only applicable rule for evaluating e, and we obtain v = v1 = v1 = v. If Env e v has been obtained using rule F4 , then e = e1 op e2 , Env e1 v1 , Env e2 v2 , v1 , v2 ∈ V and v = v1 opV v2 . By induction hypothesis, if Env e1 v1 , then v1 = v1 ∈ V and if Env e2 v2 , then v2 = v2 ∈ V. Rule F4 is the only applicable rule for evaluating e and we obtain v = v1 opV v2 = v1 opV v2 = v.
Semantics of Functional Features
67
If Env e v has been obtained using rule F5 , then e = e1 op e2 , Env e1 v1 , v1 ∈ IE and v = v1 . By induction hypothesis, if Env e1 v1 , then v1 = v1 ∈ IE. Rule F5 is the only applicable rule for evaluating e and we obtain v = v1 = v1 = v. If Env e v has been obtained using rule F6 , then e = e1 op e2 , Env e1 v1 , Env e2 v2 , v1 ∈ V, v2 ∈ IE and v = v2 . By induction hypothesis, if E nv e1 v1 , then v1 = v1 ∈ V and if E nv e2 v2 , then v2 = v2 ∈ IE. Rule F6 is the only applicable rule for evaluating e and we obtain v = v2 = v2 = v. If Env e v has been obtained using rule F7 (respectively, F8 ), then e is the conditional expression if e1 then e2 else e3 , Env e1 true (respectively, E nv e1 f alse) and E nv e2 v (respectively, E nv e3 v). By induction hypothesis, if: E nv
e1 v1 then v1 = true (respectively v1 = f alse)
E nv
e2 v (respectively, Env e3 v ) then v = v
and if:
Rule F7 (respectively, F8 ) is the only applicable rule for evaluating e and we obtain v = v. If Env e v has been obtained using rule F9 (respectively, F10 ), then e = if e1 then e2 else e3 , Env e1 v1 , v1 ∈ V \ IB (respectively, v1 ∈ IE) and v = ErrT (respectively, v = v1 ). By induction hypothesis, if Env e1 v1 , then v1 = v1 . Rule F9 (respectively, F10 ) is the only applicable rule for evaluating e and we obtain v = v. If Env e v has been obtained using rule F11 (respectively, F12 ), then e = x is a variable in X such that x ∈ dom(Env) (respectively, x ∈ / dom(Env)) and v = E nv(x) (respectively, v = ErrU ). Rule F11 (respectively, F12 ) is the only applicable rule for evaluating e and we obtain v = v. If Env e v has been obtained using rule F13 , then e = let x = e1 in e2 , e1 v1 , (x, v1 ) ⊕ Env e2 v2 , v1 ∈ V and v = v2 . By induction hypothesis, if Env e1 v1 , then v1 = v1 ∈ V and if (x, v1 ) ⊕ Env e2 v2 , then v2 = v2 . Rule F13 is the only applicable rule for evaluating e, and since v1 = v1 , then (x, v1 ) ⊕ Env = (x, v1 ) ⊕ Env and we obtain v = v2 = v2 = v. E nv
If Env e1 then v1 = v1 v = v1 = v1 E nv
e v has been obtained using rule F14 , then e = let x = e1 in e2 , v1 , v1 ∈ IE and v = v1 . By induction hypothesis, if Env e1 v1 , ∈ IE and F14 is the only applicable rule for evaluating e and we obtain = v.
68
Concepts and Semantics of Programming Languages 1
If Env e v has been obtained using rule F15 (respectively, F16 ), then e = fun x → ef and v = x, ef , Env (respectively, e = rec f x = ef and v = f, x, ef , Env) and since F15 (respectively, F16 ) is the only applicable rule for evaluating a function (respectively, a recursive function), then v = v . If Env e v has been obtained using rule F17 (respectively, F18 ), then e = (e1 e2 ) and: e1 x, ef , Envf (respectively, Env e1 f, x, ef , Envf ) e2 v2 (x, v2 ) ⊕ Envf ef v (respectively, (f, f, x, ef , Envf ) ⊕ (x, v2 ) ⊕ Envf ef v) E nv E nv
By induction hypothesis, if: E nv
e1 v1 then v1 = x, ef , Envf (respectivelyv1 = f, x, ef , Envf )
E nv
e2 v2 then v2 = v2
and if:
Hence: (x, v2 ) ⊕ Envf = (x, v2 ) ⊕ Envf (respectively (f, v1 ) ⊕ (x, v2 ) ⊕ Envf = (f, f, x, ef , Envf ) ⊕ (x, v2 ) ⊕ Envf ) The only applicable rule is F17 (respectively, F18 ) and, by induction hypothesis, since: (x, v2 ) ⊕ Envf ef v (respectively, (f, f, x, ef , Envf ) ⊕ (x, v2 ) ⊕ Envf ef v) we obtain v = v . If Env e v has been obtained using rule F19 , then e = (e1 e2 ), Env e1 v1 , Env e2 v2 , v1 ∈ IF, v2 ∈ IE and v = v2 . By induction hypothesis, if E nv e1 v1 , then v1 = v1 ∈ IF and if E nv e2 v2 , then v2 = v2 ∈ IE. Rule F19 is the only applicable rule for evaluating e and we obtain v = v2 = v2 = v. If Env e v has been obtained using rule F20 , then e = (e1 e2 ), Env / IF and v = ErrT . By induction hypothesis, if Env e1 v1 , then e1 v1 , v1 ∈ / IF. Rule F20 is the only applicable rule for evaluating e and we obtain v1 = v1 ∈ v = v. The proposition below establishes an equivalence between the semantics deﬁned by the evaluation function __ and the operational semantics with callbyvalue deﬁned in Box 3.2 and Box 3.3.
Semantics of Functional Features
69
P ROPOSITION 3.2.– Equivalence: ∀e ∈ Exp2 ∀Env ∈ E ∀v ∈ V
eEnv = v ⇔ Env e v
The implication ⇒ can be proved using the LFP theory but it is out of the scope of this book. The proof of the implication ⇐ is obtained using the same technique applied for proposition 3.1, by induction over the evaluation tree for Env e v. For each rule: (Fi )
E nv1
e1 v1 · · · Envn en vn E nv e v
of the inference system SExp2 , which deﬁnes the bigstep operational semantics of expressions in Exp2 , we must prove that eEnv = v, under the induction hypothesis that e1 Env1 = v1 , etc., en Envn = vn . This proof is immediate, as the rules for evaluating the operational semantics were deﬁned following exactly the same choices used to deﬁne the evaluation function. 3.5. Semantic properties The usual techniques used to prove semantic properties over expressions in a programming language are presented as follows. 3.5.1. Equivalent expressions D EFINITION 3.3.– Equivalent expressions: two expressions are equivalent if and only if, in each environment, their evaluations return the same value: [3.8] ∀e1 , e2 ∈ Exp2 e1 ≡ e2 ⇔ ∀Env ∈ E e1 Env = e2 Env In this deﬁnition, we used the evaluation function __ for the evaluation of the expressions. The relation for evaluating bigstep operational semantics can also be used to deﬁne the notion of equivalent expressions: ∀e1 , e2 ∈ Exp2 e1 ≡ e2 ⇔ (∀Env ∈ E ∀v ∈ V Env e1 v ⇔ Env e2 v)
[3.9]
Using this second deﬁnition, proving that e1 ≡ e2 amounts to proving how the evaluation tree for Env e2 v can be obtained from the tree for Env e1 v and vice versa. Following proposition 3.2, these two approaches are equivalent. We can easily prove that ≡ deﬁnes a congruence, i.e. an equivalence relation compatible with the internal composition laws of Exp2 . For example, if e1 ≡ e2 and e1 ≡ e2 , then (e1 + e1 ) ≡ (e2 + e2 ).
70
Concepts and Semantics of Programming Languages 1
E XAMPLE 3.6.– We can prove that let x = e1 in e2 ≡ ((fun x → e2 ) e1 ). Using deﬁnition [3.8], this equivalence is proved via proof by cases, using the evaluation of e1 . If e1 Env = v ∈ V, then let x = e1 in e2 Env = e2 (x,v)⊕Env and since fun x → e2 Env = x, e2 , Env, we obtain ((fun x → e2 ) e1 )Env = e2 (x,v)⊕Env , which concludes the proof. If e1 Env = v ∈ IE, it is also possible to complete the proof. In this case, we have let x = e1 in e2 Env = v = ((fun x → e2 ) e1 )Env . The equivalence can be established by considering deﬁnition [3.9]. First, take a tree Env let x = e1 in e2 v. There are two possibilities. If this tree has been obtained by applying rule F13 , it will take the form: 1
(F13 )
E nv
2
e1 v1 (x, v1 ) ⊕ Env e2 v (v1 ∈ V) E nv let x = e1 in e2 v
[3.10]
and a tree for Env ((fun x → e2 ) e1 ) v can be constructed easily from trees and 2 : (F15 ) (F17 )
E nv
1
2 f x, e2 , Env Env e1 v (x, v ) ⊕ Env e2 v 1 1 1 (v1 ∈ / IE) E nv ((fun x → e2 ) e1 ) v [3.11] f
Second, tree Env let x = e1 in e2 v has been obtained using rule F14 : (F14 )
E nv
E nv e1 v (v ∈ IE) let x = e1 in e2 v
[3.12]
Once again, a tree for Env ((fun x → e2 ) e1 ) v can be constructed easily from : (F15 ) (F19 )
f x, e2 , Env Env e1 v (v ∈ IE) E nv ((fun x → e2 ) e1 ) v
E nv
[3.13]
f
Reciprocally, a tree for Env ((fun x → e2 ) e1 ) v can be used to construct a tree for Env let x = e1 in e2 v since trees [3.10] and [3.12] can be obtained from trees [3.11] and [3.13], respectively.
Semantics of Functional Features
71
3.5.2. Equivalent environments As we said informally in section 3.1.4, the result of evaluating an expression e in an environment Env depends only on the values associated with the free variables of e by Env. In other terms, the evaluation result of e is the same in all environments, which associate the same values with the free variables of e. Giving a formal deﬁnition of this property requires the introduction of the notion of equivalent environments over a set of variables as follows: D EFINITION 3.4.– Equivalent environments: let F ⊆ X be a set of variables and and Env2 two environments:
E nv1
E nv1
≡F Env2 ⇔ ∀x ∈ F Env1 (x) = Env2 (x)
Obviously, given two subsets of X, F1 and F2 , if F1 ⊆ F2 and Env1 ≡F2 Env2 , then Env1 ≡F1 Env2 . It is now possible to express, formally, that for any expression e, the value of eEnv depends only on the values bound in Env to the free variables of e (F(e)). P ROPOSITION 3.3.– For any expression e ∈ Exp2 : ∀Env1 , Env2 ∈ E E nv1 ≡F (e) E nv2 ⎞ ⎛ eEnv1 = eEnv2 ∈ V \ IF ⎟ ⎜ or e E nv1 = x, e , E nv1 and e E nv2 = x, e , E nv2 ⎟ ⎜ ⎟ ⎜ ⇒ ⎜ and Env1 ≡F (e )\{x} Env2 ⎟ ⎠ ⎝ or e E nv1 = f, x, e , E nv1 and e E nv2 = f, x, e , E nv2 and Env1 ≡F (e )\{x} Env2 The proof of this proposition is long and complex. Exercise 3.5 establishes this property for a language containing no functional expressions. In this case, proof is obtained by induction over the syntactic structure of the expression. In cases where functional values are present, this induction is no longer sufﬁcient, and the proof requires the use of theorems of the ﬁxpoint theory. 3.6. Exercises Exercise 3.1 Given an environment Env, compute: fun x → let y = 3 in x + yEnv and let y = 3 in fun x → x + yEnv
72
Concepts and Semantics of Programming Languages 1
Exercise 3.2 Let Env be the following environment deﬁned from an expression e ∈ Exp2 and an initial environment Env0 : E nv
= (x, 4) ⊕ (f, y, e, (x, 3) ⊕ Env0 ) ⊕ (x, 3) ⊕ Env0
1) Give a condition for the expression e such that: let f = fun y → e in (f x)Env = (f x)Env 2) Compute let f = fun x → x + (f x) in (f x)Env when e = y + x. Exercise 3.3 Consider the following expression e: let x = 3 in (let f = e1 in (let g = e2 in (let h = e3 in e4 ))) where: e1 : fun x → (let y = x in y) e2 : let y = x in fun x → y e3 : fun k → fun x → (k x) In this exercise, the callbyvalue strategy is applied to evaluate expressions. 1) Build a “partial” evaluation tree for the expression e in an environment Env0 to determine the environment Env4 , which will be used to evaluate expression e4 . 2) Complete the tree for e4 = ((h f ) 2). 3) Complete the tree for e4 = ((h g) 2). Exercise 3.4 Consider the following expression e0 : let x = 2 in (let f = e1 in e) where e1 is the expression fun y → y + x. In this exercise, the callbyname strategy is applied to evaluate expressions. 1) Build a “partial” evaluation tree for the expression e0 in an environment Env0 to determine the environment Env2 , which will be used to evaluate expression e.
Semantics of Functional Features
73
2) Complete the tree for the following expression e: let x = 3 in (f x). 3) Complete the tree for the following expression e: let x = 3 in ((fun y → y + x) x) 4) Complete the tree for the following let g = e2 in ((g (x − 1)) f ), where e2 is the expression:
expression
e:
rec g z = fun k → if z = 0 then (k x) else ((g (z − 1)) k) Exercise 3.5 In this exercise, we consider the subset of Exp2 containing only simple arithmetic expressions, retaining only constructs (1), (2), (4) and (6) from Box 3.1 with K = Z and op ∈ {+, −, ∗}. The syntax of this language, denoted as EA , is thus: e ::= k  x  e1 + e2  e1 − e2  e1 ∗ e2  let x = e1 in e2 For the sake of simplicity, we suppose here that the environments used to evaluate any expression e always contain a binding for each free variable occurring in e; thus, we do not need to account for cases of error. Hence, the operational semantics for evaluating the expressions in EA can be deﬁned from the set of values V = V = Z (since IE = ∅), and by only considering rules F1 , F4 , F11 and F13 from Box 3.2 and Box 3.3. Prove that the result of the evaluation of an expression a ∈ EA in an environment Env only depends on the value associated with the free variables of a in environment Env. ∀a ∈ EA ∀Env1 , Env2 ∈ E Env1 =F (a) Env2 ⇒ (Env1 a n ⇔ Env2 a n) Exercise 3.6 In this exercise, we consider the language EA deﬁned in exercise 3.5 without the letin construct. Replacing a variable x ∈ X by an expression e ∈ EA in an expression e ∈ EA leads to an expression inductively deﬁned by: ⎧ k if e = k ∈ Z ⎪ ⎪ ⎨ e if e = x ∈ X e[x ← e ] = y if e = y ∈ X et y = x ⎪ ⎪ ⎩ e1 [x ← e ] op e2 [x ← e ] if e = e1 op e2
74
Concepts and Semantics of Programming Languages 1
Prove that: ∀e, e ∈ EA ∀x ∈ X ∀Env ∈ E e[x ← e ]Env = e
(x,e
Env )⊕Env
Exercise 3.7 Smallstep operational semantics. In this exercise, we consider the language EA deﬁned in exercise 3.5 without the letin construct (the operational semantics is deﬁned using only rules F1 , F4 and F11 from Box 3.2 and Box 3.3). A smallstep operational semantics can be deﬁned to evaluate the expressions in EA . The smallstep semantics allows us to take into account the steps leading to the result of a computation at a level ﬁner than that obtained by using the bigstep semantics. A conﬁguration is deﬁned as a pair e, Env, where e ∈ EA and Env ∈ E. We deﬁne a conﬁguration transition relation using the inference system presented in Box 3.5. Intuitively, e, Env → e , Env means that the evaluation of expression e in environment Env leads us to evaluate the “simpler” expression e in the same environment. For example:
3 + (2 ∗ 8), Env → 3 + 16, Env
(AS1 )
x, Env → Env(x), Env (AS2 )
n1 + n2 , Env → n, Env
(AS3 )
n1 , n 2 ∈ Z
(x ∈ X)
n1 ∗ n2 , Env → n, Env
(n = n1 + n2 ) (n = n1 ∗ n2 )
(n = n1 − n2 )
n1 − n2 , Env → n, Env
e1 , Env → e1 , Env
e2 , Env → e2 , Env (AS5 ) (AS6 )
e1 + e2 , Env → e1 + e2 , Env
n1 + e2 , Env → n1 + e2 , Env
e1 , Env → e1 , Env
e2 , Env → e2 , Env (AS7 ) (AS8 )
e1 ∗ e2 , Env → e1 ∗ e2 , Env
n1 ∗ e2 , Env → n1 ∗ e2 , Env
e1 , Env → e1 , Env
e2 , Env → e2 , Env (AS9 ) (AS10 )
e1 − e2 , Env → e1 − e2 , Env
n1 − e2 , Env → n1 − e2 , Env (AS4 )
Box 3.5. Smallstep operational semantics
A transition sequence is a potentially inﬁnite sequence of the form:
e0 , Env → e1 , Env → e2 , Env → · · · such that for all i ≥ 0, there exists an inference tree for ei , Env → ei+1 , Env. A conﬁguration is said to be terminal if it takes the form n, Env where n ∈ Z
Semantics of Functional Features
75
(as, in this case, the result of the evaluation is available). Relation → is then deﬁned by e, Env → n if, and only if, there is a transition sequence of ﬁnite positive or zero length:
e0 , Env → e1 , Env → e2 , Env → · · · → em , Env such that conﬁguration em , Env is terminal with em = n. 1) Give the transition sequence used to describe the evaluation of the expression (z + (x − y)) ∗ (x + 7)
[3.14]
in an environment Env such that Env(x) = 1, Env(y) = 3 and Env(z) = 2. 2) Since an expression is “simpliﬁed” at each step of a transition sequence, we prove that all transition sequences are of ﬁnite length. a) From the wellfounded order relation 2 ≤ over the set of natural integers IN, deﬁne a wellfounded order relation over IN × IN. b) Given an expression e ∈ EA , let NV (e) be the number of occurrences of variables in e and NO (e) the number of occurrences of operators appearing in e. Prove that, if e, Env → e , Env, then: (NV (e ), NO (e )) ≺ (NV (e), NO (e)) where ≺ is the strict order associated with . c) From this result, deduce that all transition sequences are ﬁnite. d) Is this result still true when the rule below is added? Explain your answer. (AS11 )
e1 + e2 , Env → e2 + e1 , Env
3) Prove that if e Env n, then e, Env → n. 4) Prove that if e1 , Env → e2 , Env and Env e2 v, then Env e1 v.
5) Using your result from the previous question, prove that if e, Env → n, then e Env n. 2 An order relation ≤ over a set E is wellfounded if there is no strictly decreasing inﬁnite sequence e1 > e2 > · · · of elements of E.
76
Concepts and Semantics of Programming Languages 1
6) Now we want to describe the evaluation of expression [3.14] in the same environment Env using the following computation sequence:
(z + (x − y)) ∗ (x + 7), Env → (z + (x − y)) ∗ (1 + 7), Env → (z + (x − y)) ∗ 8, Env → (z + (x − 3)) ∗ 8, Env → (z + (1 − 3)) ∗ 8, Env → (z + −2) ∗ 8, Env → (2 + −2) ∗ 8, Env → 0 ∗ 8, Env → 0, Env Propose a new inference system to deﬁne the relation → such that for each step in this transition sequence there is an inference tree using this system.
4 Semantics of Imperative Features
This chapter presents the three classical semantic approaches to the most common imperative features of programming languages, as they are usually treated (see [NIE 07, WIN 93]). 4.1. Syntax of a kernel of an imperative language We consider here a language, Lang3 , corresponding to the kernel of most imperative languages. It allows explicit handling of references of locations, i.e. memory addresses. This language is built from two languages: Def 3 for deﬁnitions and Exp3 for expressions. Both are variations of languages Def 1 and Exp1 , introduced in Chapter 2. The language Def 3 , deﬁned in Table 4.1, contains a single construct that introduces a binding into an execution environment. d ::= let x = e; deﬁnition of a nonmutable variable (x ∈ X, e ∈ Exp3 ) Table 4.1. Language of deﬁntions Def 3
The language Exp3 , used for expressions, aims at manipulating values and modifying the memory by allocating locations for new values. The syntax of Exp3 is presented in Table 4.2. Thus, an expression e ∈ Exp3 is either: – a constant k ∈ K = Z ∪ IB; – an identiﬁer x ∈ X; – an expression obtained by applying a unary (or binary) operator to an expression (or two expressions);
Concepts and Semantics of Programming Languages 1: A Semantical Approach with OCaml and Python, First Edition. Thérèse Hardin, Mathieu Jaume, François Pessaux and Véronique Viguié DonzeauGouge. © ISTE Ltd 2021. Published by ISTE Ltd and John Wiley & Sons, Inc.
78
Concepts and Semantics of Programming Languages 1
– an allocation ↑ e which reserves a location, stores the result of the evaluation of e at this location, and returns the reference of this location r ∈ R; – dereferencing ↓ e, which evaluates the expression e and, if the value of e is a reference, returns the value stored at this location. e ::= k Constant (k ∈ K)  x Variable (x ∈ X)  op e1 Application of a primitive unary operator  e1 op e2 Application of a primitive binary operator  ↑e Indirection, memory allocation  ↓e Dereferencing
(1) (2) (3) (4) (5) (6)
Table 4.2. Language of expressions Exp3
We shall not describe the possible primitive operators here; depending on the chosen language, these operators may handle integers, booleans or references (pointer arithmetics). The set of constants is deﬁned by the union of integers, booleans and references. Using integers to represent references, the following deﬁnitions are used to represent K ∪ R. Python
class CInt4: def __init__(self,cst_int): self.cst_int = cst_int class CBool4: def __init__(self,cst_bool): self.cst_bool = cst_bool
class CRef4: def __init__(self,cst_adr): self.cst_adr = cst_adr
OCaml
type const4 = CInt4 of int  CBool4 of bool  CRef4 of int
For our purposes, we shall use the same classical operators introduced in Chapter 3. Python
class Not4: pass
class Minus4: pass
class Div4: pass
class Or4: pass
class Plus4: pass
class Mult4: pass
class And4: pass
class Eq4: pass
OCaml
class Leq4: pass
type unary_op4 = Not4 type bin_op4 = Plus4  Minus4  Mult4  Div4  And4  Or4  Eq4  Leq4
Semantics of Imperative Features
79
Expressions of Exp3 are deﬁned as follows: Python
class Cste4: def __init__(self,cste): self.cste = cste class Var4: def __init__(self,symb): self.symb = symb class U_Op4: def __init__(self,op,exp1): self.op = op self.exp1 = exp1
class B_Op4: def __init__(self,op,exp1,exp2): self.op = op self.exp1 = exp1 self.exp2 = exp2 class Ref4: def __init__(self,exp): self.exp = exp class Deref4: def __init__(self,exp): self.exp = exp
OCaml
type ’a exp4 =  Cste4 of const4  Var4 of ’a  U_Op4 of unary_op4 * ’a exp4  B_Op4 of bin_op4 * ’a exp4 * ’a exp4  Ref4 of ’a exp4  Deref4 of ’a exp4
Finally, language Lang3 includes assignment (already presented for language Lang1 in Chapter 2) along with the classical control structures (sequence, conditional, loop), occurring in most imperative languages. Syntax of Lang3 is deﬁned in Table 4.3. c ::= x := e Assignment (x ∈ X, e ∈ Exp3 )  skip  c1 ; c 2 Sequence  if e then c1 else c2 Conditional (e ∈ Exp3 )  while e do c1 Loop (e ∈ Exp3 )
(1) (2) (3) (4) (5)
Table 4.3. Language of commands Lang3
At this step, the language of deﬁnitions Def 3 and the language of programs Lang3 can be deﬁned separately. However, in section 4.6, Def 3 will be extended in order to deﬁne procedures. The body of a procedure is a program in Lang3 , and Lang3 will be extended with local deﬁnitions of variables introduced by expressions in Def 3 . Languages Def 3 and Lang3 are thus deﬁned in a mutually recursive way as follows (these deﬁnitions will be extended in section 4.6). Python
class Let_def4: def __init__(self,symb,exp): self.symb = symb self.exp = exp
class ICond4: def __init__(self,exp,pgm1,pgm2): self.exp = exp self.pgm1 = pgm1
80
Concepts and Semantics of Programming Languages 1
class IAffect4: def __init__(self,symb,exp): self.symb = symb self.exp = exp class ISkip4: pass class ISeq4: def __init__(self,pgm1,pgm2): self.pgm1 = pgm1 self.pgm2 = pgm2
self.pgm2 = pgm2 class IWhile4: def __init__(self,exp,pgm): self.exp = exp self.pgm = pgm
OCaml
type ’a def4 =  Let_def4 of ’a * ’a exp4 and ’a pgm4 =  IAffect4 of ’a * ’a exp4  ISeq4 of ’a pgm4 * ’a pgm4  IWhile4 of ’a exp4 * ’a pgm4  ... procedure calls
 ... procedure definitions  ISkip4  ICond4 of ’a exp4 * ’a pgm4 * ’a pgm4  ... blocks with local variables
E XAMPLE 4.1.– Program for computing GCD: the fragment of program represented in the abstract syntax tree below computes the GCD of two natural integers already stored in the memory at locations x and y: (1)
(1) y :=↓ y− ↓ x x :=↓ x− ↓ y if ↓ x ≤↓ y then y :=↓ y− ↓ x else x :=↓ x− ↓ y (5) while not(↓ x =↓ y) do if ↓ x ≤↓ y then y :=↓ y− ↓ x else x :=↓ x− ↓ y (4)
[4.1] This program is written as follows: Python
ex_pgm4_1 = IWhile4(U_Op4(Not4(), B_Op4(Eq4(),Deref4(Var4("x")),Deref4(Var4("y")))), ICond4(B_Op4(Leq4(),Deref4(Var4("x")),Deref4(Var4("y"))), IAffect4("y",B_Op4(Minus4(),Deref4(Var4("y")), Deref4(Var4("x")))), IAffect4("x",B_Op4(Minus4(),Deref4(Var4("x")), Deref4(Var4("y")))))) OCaml
let ex_pgm4_1 = IWhile4 (U_Op4 (Not4, B_Op4 (Eq4, Deref4 (Var4 "x"), Deref4 (Var4 "y"))), ICond4 (B_Op4 (Leq4, Deref4 (Var4 "x"), Deref4 (Var4 "y")), IAffect4 ("y", B_Op4 (Minus4, Deref4 (Var4 "y"), Deref4 (Var4 "x"))), IAffect4 ("x", B_Op4 (Minus4, Deref4 (Var4 "x"), Deref4 (Var4 "y")))))
Semantics of Imperative Features
81
Constructs used to deﬁne functions and procedures will be introduced later. However, the following theorem characterizes the expressiveness of the language Lang3 deﬁned at this stage. T HEOREM 4.1.– Böhm and Jacopini [BOE 66]: Any algorithm can be expressed as a composition of three control structures: sequence, conditional instruction and iteration. Deﬁning a semantics for the language Lang3 allows us to have a precise description of the memory modiﬁcations resulting from assignments and control instructions. Throughout this chapter, we consider that all expressions and programs over which a semantics is deﬁned are syntactically correct and welltyped (see Chapter 5): evaluation of an expression or execution of a program never generates a typing error. 4.2. Evaluation of expressions As in Chapter 2, an expression in Exp3 is evaluated in a state made up of an execution environment Env ∈ E and a memory Mem ∈ M. However, the language of expressions Exp3 contains an operator ↑ , which allocates memory, stores the value of its argument in this new location and returns its reference. Thus, while evaluating an expression never modiﬁes the execution environment, it may trigger a modiﬁcation of the memory state (known as a side effect of the evaluation). Evaluating an expression returns a value that may be an integer, a boolean or a reference, hence V = Z ∪ IB ∪ R. Since we are only considering welltyped expressions (i.e. expressions for which evaluation does not result in a typing error), the only possible errors are deﬁnition errors ErrU (access to an identiﬁer which is not deﬁned in the environment) and execution errors ErrE (for example in the case of a division by 0). Hence, IE = {ErrU , ErrE } and V = V ∪ IE is the set of all possible results of expression evaluation. We thus deﬁne an evaluation function: ___ : (E × M × Exp3 ) → (V × M) em such that eM E nv = (v, M em ) means that v is the value obtained by evaluating the expression e in the state (Env, Mem) and Mem is the memory issued from Mem when computing v. This function is deﬁned in Table 4.4.
We suppose that each operator op has a semantics deﬁning how to interpret it by an operator opV over the domain of values. We also suppose that during allocation (operator ↑) an external mechanism provides locations in the memory and returns their references. For expressions of the form e1 op e2 , we have chosen to evaluate
82
Concepts and Semantics of Programming Languages 1
expression e1 ﬁrst, then – depending on the result – to evaluate e2 . Furthermore, when an error is generated during expression evaluation, the returned value corresponds to this error, and the memory is that produced by the last errorfree evaluation. em eM E nv = (v, M em )
e k
x
op e1
e1 op e2
↑ e1
↓ e1
v=k (k ∈ K) = Mem v = Env(x) if x ∈ dom(Env) (x ∈ X) M em = M em v = ErrU if x ∈ / dom(Env) (x ∈ X) M em = M em em v = opV v1 if e1 M E nv = (v1 , M em1 ) M em = M em1 and v1 ∈ V em v = v1 if e1 M E nv = (v1 , M em1 ) M em = M em and v1 ∈ IE em v = v1 opV v2 if e1 M = (v1 , Mem1 ) and v1 ∈ V E nv em1 M em = M em2 and e2 M = (v2 , Mem2 ) and v2 ∈ V E nv v = v1 M em if e1 Env = (v1 , Mem1 ) and v1 ∈ IE M em = M em em v = v2 if e1 M E nv = (v1 , M em1 ) and v1 ∈ V M em1 M em = M em1 and e2 Env = (v2 , Mem2 ) and v2 ∈ IE v=r em if e1 M E nv = (v1 , M em1 ) and v1 ∈ V M em = M em1 [r := v1 ] v = v1 em if e1 M E nv = (v1 , M em1 ) and v1 ∈ IE M em = M em v = Mem1 (v1 ) em if e1 M E nv = (v1 , M em1 ) and v1 ∈ R M em = M em1 v = v1 em if e1 M E nv = (v1 , M em1 ) and v1 ∈ IE M em = M em M em
Table 4.4. Evaluation of expressions in Exp3
In this chapter, we have chosen to use the exception mechanism (see Chapter 8) instead of introducing special values for errors. If an error is triggered by evaluating an expression, the exception Eval_Error is raised and the evaluation stops. The set of values V thus includes the constants from K and the references from R. Other values will be introduced later, in addition to the deﬁnitions given below. Python
class VCste4: def __init__(self,cst): self.cst = cst OCaml
type ’a valeurs4 =  VCste4 of const4  ...
Semantics of Imperative Features
83
Since references are represented by integers, a simple way of carrying out memory allocation is to increment the last created reference by 1. This gives us the new allocated reference r, and the binding (r, v) can be added at the head of the list representing the memory. The function below carries out this operation and returns a pair made up of the new reference and the list representing the new memory. Python
def add_mem(mem,v): if len(mem) == 0: return (0,[(0,v)]) else: n1,v1 = mem[0] return (n1+1,[(n1+1,v)]+mem) OCaml
let   val
add_mem mem v = match mem with [] > (0, [(0, v)]) (n1, v1) :: mem1 > (n1 + 1, ((n1 + 1), v) :: mem) add_mem : (int * ’a) list > ’a > int * (int * ’a) list
Interpretation of operators is deﬁned as in Chapter 3. Error handling is replaced by the raising of the exception Eval_Error. Python
def apply4_unary_op(op,v): if isinstance(op,Not4) and isinstance(v,VCste4) \ and isinstance(v.cst,CBool4): return VCste4(CBool4(not v.cst.cst_bool)) raise ValueError def apply4_binary_op(op,v1,v2): if isinstance(v1,VCste4) and isinstance(v2,VCste4): vc1 = v1.cst vc2 = v2.cst if isinstance(op,Plus4): if isinstance(vc1,CInt4) and isinstance(vc2,CInt4): return VCste4(CInt4(vc1.cst_int + vc2.cst_int)) raise ValueError if isinstance(op,Minus4): if isinstance(vc1,CInt4) and isinstance(vc2,CInt4): return VCste4(CInt4(vc1.cst_int  vc2.cst_int)) raise ValueError if isinstance(op,Mult4): if isinstance(vc1,CInt4) and isinstance(vc2,CInt4): return VCste4(CInt4(vc1.cst_int * vc2.cst_int)) raise ValueError if isinstance(op,Div4): if isinstance(vc1,CInt4) and isinstance(vc2,CInt4) and vc2.cst_int != 0: return VCste4(CInt4(vc1.cst_int / vc2.cst_int)) raise ValueError
84
Concepts and Semantics of Programming Languages 1
if isinstance(op,And4): if isinstance(vc1,CBool4) and isinstance(vc2,CBool4): return VCste4(CBool4(vc1.cst_bool and vc2.cst_bool)) raise ValueError if isinstance(op,Or4): if isinstance(vc1,CBool4) and isinstance(vc2,CBool4): return VCste4(CBool4(vc1.cst_bool or vc2.cst_bool)) raise ValueError if isinstance(op,Eq4): if isinstance(vc1,CInt4) and isinstance(vc2,CInt4): return VCste4(CBool4(vc1.cst_int == vc2.cst_int)) raise ValueError if isinstance(op,Leq4): if isinstance(vc1,CInt4) and isinstance(vc2,CInt4): return VCste4(CBool4(vc1.cst_int (match v with  VCste4 (CBool4 b) > VCste4 (CBool4 (not b))  _ > raise Eval_Error) val apply4_unary_op : unary_op4 > ’a valeurs4 > ’b valeurs4 let apply4_binary_op op v1 v2 = match op with  Plus4 > (match (v1, v2) with  (VCste4 (CInt4 n1), VCste4 (CInt4 n2)) > VCste4 (CInt4 (n1 + n2))  _ > raise Eval_Error)  Minus4 > (match (v1, v2) with  (VCste4 (CInt4 n1), VCste4 (CInt4 n2)) > VCste4 (CInt4 (n1  n2))  _ > raise Eval_Error)  Mult4 > (match (v1, v2) with  (VCste4 (CInt4 n1), VCste4 (CInt4 n2)) > VCste4 (CInt4 (n1 * n2))  _ > raise Eval_Error)  Div4 > (match (v1, v2) with  (VCste4 (CInt4 n1), VCste4 (CInt4 n2)) > if n2 = 0 then raise Eval_Error else VCste4 (CInt4 (n1 / n2))  _ > raise Eval_Error)  And4 > (match (v1, v2) with  (VCste4 (CBool4 b1), VCste4 (CBool4 b2)) > VCste4 (CBool4 (b1 && b2))  _ > raise Eval_Error)  Or4 > (match (v1, v2) with  (VCste4 (CBool4 b1), VCste4 (CBool4 b2)) > VCste4 (CBool4 (b1  b2))  _ > raise Eval_Error)
Semantics of Imperative Features
85
 Eq4 > (match (v1, v2) with  (VCste4 (CInt4 n1), VCste4 (CInt4 n2)) > VCste4 (CBool4 (n1 = n2))  _ > raise Eval_Error)  Leq4 > (match (v1, v2) with  (VCste4 (CInt4 n1), VCste4 (CInt4 n2)) > VCste4 (CBool4 (n1 raise Eval_Error) val apply4_binary_op : bin_op4 > ’a valeurs4 > ’b valeurs4 > ’c valeurs4
It is now possible to deﬁne the function that evaluates an expression. Throughout this chapter, a state will be represented by an environment binding some values of V to variable symbols, and by a memory binding constants in K ∪ R to references (represented by integers). Python
def eval_exp4(env,mem,e): if isinstance(e,Cste4): return (VCste4(e.cste),mem) if isinstance(e,Var4): x = valeur_de(env,e.symb) if x is None: raise ValueError return (x,mem) if isinstance(e,U_Op4): v,memv = eval_exp4(env,mem,e.exp1) return (apply4_unary_op(e.op,v),memv) if isinstance(e,B_Op4): v1,mem1 = eval_exp4(env,mem,e.exp1) v2,mem2 = eval_exp4(env,mem1,e.exp2) return (apply4_binary_op(e.op,v1,v2),mem2) if isinstance(e,Ref4): v,mem1 = eval_exp4(env,mem,e.exp) r,mem2 = add_mem(mem1,v.cst) return (VCste4(CRef4(r)),mem2) if isinstance(e,Deref4): v1,mem1 = eval_exp4(env,mem,e.exp) if isinstance(v1,VCste4) and isinstance(v1.cst,CRef4): v2 = valeur_ref(mem1,v1.cst.cst_adr) if v2 is None: raise ValueError else: return (VCste4(v2),mem1) raise ValueError raise ValueError raise ValueError OCaml
let rec eval_exp4 env mem e = match e with  Cste4 k > (VCste4 k, mem)  Var4 x > ( match valeur_de env x with Some(v) > (v, mem)  _ > raise Eval_Error)  U_Op4 (op, e1) > (match eval_exp4 env mem e1 with  (VCste4 v, mem1) > ((apply4_unary_op op (VCste4 v)), mem1)  _ > raise Eval_Error)  B_Op4(op,e1,e2) > (match eval_exp4 env mem e1 with  (VCste4 v1, mem1) > (match eval_exp4 env mem1 e2 with  (VCste4 v2, mem2) > ((apply4_binary_op op (VCste4 v1) (VCste4 v2)), mem2)
86
Concepts and Semantics of Programming Languages 1
 _ > raise Eval_Error)  _ > raise Eval_Error)  Ref4 e1 > (match eval_exp4 env mem e1 with  (VCste4 v, mem1) > let (a, mem2) = (add_mem mem1 v) in (VCste4 (CRef4 a), mem2)  _ > raise Eval_Error)  Deref4 e1 > (match eval_exp4 env mem e1 with  (VCste4 (CRef4 a), mem1) > ( match valeur_ref mem1 a with  Some v > (VCste4 v, mem1)  _ > raise Eval_Error)  (VCste4 _, mem1) > raise Eval_Error  _ > raise Eval_Error) val eval_exp4 : (’a * ’b valeurs4) list > (int * const4) list > ’a exp4 > ’b valeurs4 * (int * const4) list
4.3. Evaluation of deﬁnitions Deﬁnitions introduced with language Def 3 only modify execution environments. Their semantics is deﬁned by: let x = e; em (Env, Mem) −−−−−−−→Def 3 ((x, v) ⊕ Env, Mem ) if eM E nv = (v, M em ) This deﬁnition adds the binding (x, v) to the environment Env and sets the em memory to Mem when eM E nv = (v, M em ). Using a ﬁnite sequence of deﬁnitions d = [d1 , · · · , dn ] and an initial state (Env0 , Mem0 ), this relation computes the state (Envn , Memn ): d
d
d
1 2 n (Env0 , Mem0 ) −→ Def 3 ( E nv1 , M em1 ) −→Def 3 · · · −→Def 3 ( E nvn , M emn )
d
This sequence of transitions is denoted by: (Env0 , Mem0 ) →Def 3 (Envn , Memn ). State transformations generated by a deﬁnition, respectively, a list of deﬁnitions, are deﬁned as follows. These deﬁnitions will be completed further as the language Def 3 is extended. Python
def trans_def4(e,d): env,mem = e if isinstance(d,Let_def4): v,mem1 = eval_exp4(env,mem,d.exp) return (ajout_liaison_env(env,d.symb,v),mem1) if ... raise ValueError def trans_def4_exec(e,ld): if len(ld) == 0: return e else: return trans_def4_exec(trans_def4(e,ld[0]),ld[1:])
Semantics of Imperative Features
87
OCaml
let trans_def4 (env, mem) d = match d with  Let_def4 (x, e) > (match eval_exp4 env mem e with (VCste4(v),mem1) > ((ajout_liaison_env env x v),mem1))  ... val trans_def4 : (’a * ’a valeurs4) list * (int * const4) list > ’a def4 > (’a * ’a valeurs4) list * (int * const4) list let trans_def4_exec (env, mem) ld = (List.fold_left trans_def4 (env, mem) ld) val trans_def4_exec : (’a * ’a valeurs4) list * (int * const4) list > ’a def4 list > (’a * ’a valeurs4) list * (int * const4) list
Unlike the language of deﬁnitions Def 1 presented in Chapter 2, Def 3 includes no commands of the form var x = e; to deﬁne mutable variables. From a semantic perspective, however, the same effect can be obtained using the ↑ operator. Considering an expression e, of which the evaluation produces a value v and does not em = (v, Mem), then we have modify the memory, i.e. such that eM E nv M em ↑ eEnv = (r, Mem[r := v]), and hence: var x = e; (Env, Mem) −−−−−−−−→Def 1 ((x, r) ⊕ Env, Mem[r := v]) let x =↑ e; (Env, Mem) −−−−−−−−→Def 3 ((x, r) ⊕ Env, Mem[r := v]) E XAMPLE 4.2.– The transition sequence in Box 4.1 shows how, from the following sequence of deﬁnitions: let x1 =↑ 4; let x2 = x1 ; let x3 = 2+ ↓ x1 ; let x4 =↑ x1 ; let x5 =↑↑ x1 ; [4.2] and a state (Env, Mem), we obtain the state: (x5 , rx5 ) ⊕ (x4 , rx4 ) ⊕ (x3 , 6) ⊕ (x2 , rx1 ) ⊕ (x1 , rx1 ) ⊕ Env M em[rx1 := 4][rx4 := rx1 ][ra := rx1 ][rx5 := ra ]
[4.3]
The sequence deﬁned in [4.2] is coded as: Python
ex_ld4_1 = [Let_def4("x1",Ref4(Cste4(CInt4(4)))), Let_def4("x2",Var4("x1")), Let_def4("x3",B_Op4(Plus4(),Cste4(CInt4(2)),Deref4(Var4("x1")))), Let_def4("x4",Ref4(Var4("x1"))), Let_def4("x5",Ref4(Ref4(Var4("x1"))))] OCaml
let ex_ld4_1 = [ Let_def4 ("x1", Let_def4 ("x2", Let_def4 ("x3", Let_def4 ("x4", Let_def4 ("x5",
Ref4 (Cste4 (CInt4 4))) ; Var4 "x1") ; B_Op4 (Plus4, Cste4 (CInt4 2), Deref4 (Var4 "x1"))) ; Ref4 (Var4"x1")) ; Ref4 (Ref4 (Var4 "x1")))]
88
Concepts and Semantics of Programming Languages 1
And its execution started in a state with an empty environment is obtained as follows: Python
>>> [ ( ( ( [ ( (
print_state(trans_def4_exec(([],[]),ex_ld4_1)) x5 , VCste4 (’CRef4’, 3) ), ( x4 , VCste4 (’CRef4’, 1) ), x3 , VCste4 (’CInt4’, 6) ), ( x2 , VCste4 (’CRef4’, 0) ), x1 , VCste4 (’CRef4’, 0) ), ] 3 , (’CRef4’, 2) ), ( 2 , (’CRef4’, 0) ), ( 1 , (’CRef4’, 0) ), 0 , (’CInt4’, 4) ), ]
OCaml
# trans_def4_exec ([], []) ex_ld4_1 ;;  : (string * string valeurs4) list * (int * const4) list = ([("x5", VCste4 (CRef4 3)); ("x4", VCste4 (CRef4 1)); ("x3", VCste4 (CInt4 6)); ("x2", VCste4 (CRef4 0)); ("x1", VCste4 (CRef4 0))], [(3, CRef4 2); (2, CRef4 0); (1, CRef4 0); (0, CInt4 4)])
Taking rx1 = 0, rx4 = 1, rx5 = 3 and ra = 0, we thus obtain the state deﬁned in [4.3]. (⎛Env, Mem)
⎞
let x1 =↑ 4; −−−−−−−−−→Def 3
⎟ ⎜ ⎝(x1 , rx1 ) ⊕ Env, Mem[rx1 := 4]⎠
E nv1 M em ⎛ ⎞1
let x2 = x1 ; −−−−− −−−−→Def 3
⎜ ⎟ ⎝(x2 , rx1 ) ⊕ Env1 , Mem1 ⎠
E nv2 ⎛ ⎞
(ii)
let x3 = 2+ ↓ x1 ; ⎜ ⎟ −−−−−−−−−−−−−→Def 3 ⎝(x3 , 6) ⊕ Env2 , Mem1 ⎠
E nv3 ⎛ let x4 =↑ x1 ; −−−−−−−−−−→Def 3 let x5 =↑↑ x1 ; −−−−−−−−−−−→Def 3
(iii) ⎞
⎜ ⎟ ⎝(x4 , rx4 ) ⊕ Env3 , Mem1 [rx4 := rx1 ]⎠
E nv4 M em2 ⎛
(iv) ⎞
⎜ ⎟ ⎝(x5 , rx5 ) ⊕ Env4 Mem2 [ra := rx1 ][rx5 := ra ]⎠ (v)
E nv5
(i) (ii) (iii) (iv) (v)
(i)
M em3
em 4M E nv = (4, M em) M em1 x1 Env1 = (Env1 (x1 ), Mem1 ) = (rx1 , Mem1 ) em1 2+ ↓ x1 M E nv2 = (2 + M em1 ( E nv2 (x1 )), M em1 ) = (6, M em1 ) M em1 ↑ x1 Env3 = (rx4 , Mem1 [rx4 := Env3 (x1 )]) = (rx4 , Mem1 [rx4 := rx1 ]) em2 ↑↑ x1 M E nv4 = (rx5 , M em2 [ra := rx1 ][rx5 := ra ]) em2 since ↑ x1 M E nv4 = (ra , M em2 [ra := E nv4 (x1 )]) = (ra , M em2 [ra := rx1 ])
Box 4.1. Example: execution of a sequence of deﬁnitions
Semantics of Imperative Features
89
4.4. Operational semantics The instruction at the heart of all imperative languages is assignment, which changes the state of the memory, i.e. modiﬁes its contents. Execution of an assignment x := e in a state (Env, Mem) consists of evaluating the expression e in this state to obtain a value v and a new memory state Mem , searching the value r = Env(x) ∈ R of x in environment Env, and then, ﬁnally, writing the value v at the location r to obtain memory Mem [r := v]. R EMARK.– Few highlevel languages include a syntactic construct to distinguish between the reference of the location of a variable and the value stored at this location. These two distinct values are often denoted by the variable name alone; it is left to users to determine which value is to be used for each occurrence of the variable name in an expression. As we saw at the end of section 2.3, in these languages, the meaning of variable names occurring in assignments depends on their position to the left or to the right of the assignment symbol, on the left variable names denote a location, and on the right, they denote a value. The execution of an imperative program consists of executing a sequence of assignments and tests. The control structures included in a program specify the assignments to execute, the order of execution and the number of times they should be executed. Given an execution environment, the semantics of Lang3 describes the memory transformations associated with each instruction in Lang3 . These transformations are described by a relation, following the approach ﬁrst introduced in [PLO 04]. 4.4.1. Bigstep semantics Given an execution environment Env, the bigstep operational semantics of language Lang3 deﬁnes a transition relation: Env ⊆ M × Lang3 × M such that Mem, c Env Mem means that the execution of instruction c in state (Env, Mem) leads to the memory state Mem . This relation is deﬁned by induction, using the inference system presented in Box 4.2. Rule C1 deﬁnes the semantics of the assignment x := e: the evaluation of e in state (Env, Mem) must not raise an error, and evaluation of the variable x in the environment Env must produce a reference, otherwise the execution is aborted. As the premises for this rule do not take the form Mem, c Env Mem , they must be considered as conditions for applying the rule.
90
Concepts and Semantics of Programming Languages 1
(C1 )
E nv(x)
em1 =r∈R eM = (v, Mem2 ) ∈ V × M E nv
Mem1 , x := e Env Mem2 [r := v]
(C2 )
Mem, skip Env
M em
Mem1 , c2 Env
Mem, c1 Env Mem1 (C3 )
Mem, c1 ; c2 Env Mem2 (C4 ) (C5 )
M em2
em eM
Mem1 , c1 Env E nv = (true, M em1 )
Mem, if e then c1 else c2 Env Mem2
em eM
Mem1 , c2 Env E nv = (f alse, M em1 )
Mem, if e then c1 else c2 Env Mem2
M em2 M em2
Mem1 , c Env Mem2
Mem2 , while e do c Env
Mem, while e do c Env Mem3
em eM E nv = (true, M em1 )
(C6 ) (C7 )
M em3
em eM E nv = (f alse, M em1 )
Mem, while e do c Env Mem1
Box 4.2. Bigstep operational semantics of Lang3
Rule C2 simply indicates that executing the instruction skip does not affect the state. Rule C3 describes the evolution of the memory during the execution of a sequence of two instructions. Rules C4 and C5 specify the memory transformation associated with a conditional; they contain a condition to apply them which depends on the result of the evaluation of expression e. Finally, rules C6 and C7 relate to the while e do c loop, and specify how many times the execution of the body c of the loop is to be repeated, according to the value of the conditional expression e. Unlike the other instructions, while e do c allows us to write programs where execution does not terminate, and which, consequently, cannot be described by an inference tree since such a tree is a ﬁnite structure. For such nonterminating programs c, there is no memory state Mem such that
Mem, c Env Mem . For example, taking the program f oo = while true do c, there is no memory state Mem such that Mem, f oo Env Mem . In fact, this program contains an inﬁnite loop, hence it is impossible to build a ﬁnite execution tree for it: .. em . trueM E nv (C ) (C6 ) = (true, Mem) 2 Mem, skip Env Mem
Mem, f oo Env ? (C6 )
Mem, f oo Env ? since Mem, f oo Env ? occurs both in the root and in the premise. Bigstep operational semantics uses inference trees – which are ﬁnite – and cannot, therefore, take account of the execution of nonterminating programs. A ﬁner approach,
Semantics of Imperative Features
91
suitable for describing the execution of nonterminating programs, will be presented in section 4.4.2. Let us implement a function that, given a state (Env, Mem) and a program c, computes the new memory Mem such that Mem, c Env Mem . Note however that, just like the evaluation function deﬁned in Chapter 3, this does not correspond to a mathematical function. Indeed, this function may not terminate if it is applied to a program whose execution in the given state does not terminate. This is the case when execution leads to an inﬁnite loop (as we saw in Chapter 3 for expressions corresponding to the application of a recursive function, where evaluation resulted in an inﬁnite number of recursive calls). This function reproduces the execution mechanism described by the semantic rules in Box 4.2. As such, it is not compositional: the execution of a program does not simply correspond to the execution of each of its components. Rule C6 describes the execution of the instruction while e do c using the execution of the c component, but also the execution of the loop itself. We shall return to this point in section 4.7.1. As we assume that executions of programs do not generate typing errors, we just have to introduce here a single exception Exec_Error, which will be raised in case of error of execution. Once again, the deﬁnition of this function will be extended as language Lang3 is enriched. Python
def exec_pgm4(env,mem,c): if isinstance(c,IAffect4): r = valeur_de(env,c.symb) if r is None: raise ValueError if isinstance(r,VCste4) and isinstance(r.cst,CRef4): v,mem1 = eval_exp4(env,mem,c.exp) if isinstance(v,VCste4): return write_mem(mem1,r.cst.cst_adr,v.cst) raise ValueError raise ValueError if isinstance(c,ISkip4): return mem if isinstance(c,ISeq4): return exec_pgm4(env,exec_pgm4(env,mem,c.pgm1),c.pgm2) if isinstance(c,ICond4): b,mem1 = eval_exp4(env,mem,c.exp) if isinstance(b,VCste4) and isinstance(b.cst,CBool4): if b.cst.cst_bool: return exec_pgm4(env,mem1,c.pgm1) else: return exec_pgm4(env,mem1,c.pgm2) raise ValueError if isinstance(c,IWhile4): b,mem1 = eval_exp4(env,mem,c.exp) if isinstance(b,VCste4) and isinstance(b.cst,CBool4): if b.cst.cst_bool: return exec_pgm4(env,exec_pgm4(env,mem1,c.pgm),c) else: return mem1 raise ValueError if ... raise ValueError
92
Concepts and Semantics of Programming Languages 1
OCaml
exception Exec_Error let rec exec_pgm4 env mem c = match c with  IAffect4 (x, e) > ( match ((valeur_de env x), (eval_exp4 env mem e)) with  (Some (VCste4 (CRef4 a)), (VCste4 v, mem1)) > write_mem mem1 a v  _ > raise Exec_Error)  ISkip4 > mem  ISeq4 (c1, c2) > exec_pgm4 env (exec_pgm4 env mem c1) c2  ICond4(e,c1,c2) > (match eval_exp4 env mem e with  (VCste4 (CBool4 true), mem1) > exec_pgm4 env mem1 c1  (VCste4 (CBool4 false), mem1) > exec_pgm4 env mem1 c2  _ > raise Exec_Error)  IWhile4 (e, c1) > (match eval_exp4 env mem e with  (VCste4 (CBool4 true), mem1) > exec_pgm4 env (exec_pgm4 env mem1 c1) c  (VCste4 (CBool4 false), mem1) > mem1  _ > raise Exec_Error)  ... val exec_pgm4 : (’a * ’a valeurs4) list > (int * const4) list > ’a pgm4 > (int * const4) list
E XAMPLE 4.3.– The execution of the program from example 4.1: while not(↓ x =↓ y) do if ↓ x ≤↓ y then y :=↓ y− ↓ x else x :=↓ x− ↓ y e c1 c2 ew cw
in a state (Env, Mem) = ((y, ry ) ⊕ (x, rx ) ⊕ Env0 , Mem0 [rx := 4][ry := 2]) is described by the following inference tree: em ew M = (true, M em) E nv (C6 )
where
1
1
is the tree: em = rx ↓ x− ↓ yM E nv = (2, M em)
Mem, c2 Env Mem[rx := 2]
Mem, cw Env Mem[rx := 2]
em eM E nv = (f alse, M em)
(C5 ) and
2
Mem, while ew do cw Env Mem[rx := 2]
2
(C1 )
E nv(x)
is the tree: M em[r :=2]
(C5 )
= (f alse, Mem[rx := 2]) ew Env x
Mem[rx := 2], while ew do cw Env Mem[rx := 2]
Semantics of Imperative Features
93
Returning to program ex_pgm4_1, introduced in example 4.1, this execution is obtained as follows: Python
ex_ld4_2 = [Let_def4("x",Ref4(Cste4(CInt4(4)))), Let_def4("y",Ref4(Cste4(CInt4(2))))] (env_ex_ld4_2,mem_ex_ld4_2) = trans_def4_exec(([],[]),ex_ld4_2) >>> print_state((env_ex_ld4_2,exec_pgm4(env_ex_ld4_2,mem_ex_ld4_2,ex_pgm4_1))) [ ( y , VCste4 (’CRef4’, 1) ), ( x , VCste4 (’CRef4’, 0) ), ] [ ( 1 , (’CInt4’, 2) ), ( 0 , (’CInt4’, 2) ), ] OCaml
let ex_ld4_2 = [ Let_def4 ("x", Ref4 (Cste4 (CInt4 4))); Let_def4 ("y", Ref4 (Cste4 (CInt4 2)))] # let (env, mem) = trans_def4_exec ([],[]) ex_ld4_2 in exec_pgm4 env mem ex_pgm4_1 ;;  : (int * const4) list = [(1, CInt4 2); (0, CInt4 2)]
4.4.2. Smallstep semantics The bigstep operational semantics presented in section 4.4.1 handles elements of the form Mem1 , c Env Mem2 , which are only able to specify the ﬁnal state Mem2 of the memory obtained after the execution of a program c, without providing any precise information concerning the steps, which led up to it (this information can be retrieved in the tree for Mem1 , c Env Mem2 ). Furthermore, as inference trees are, by convention, ﬁnite objects, this semantics is unable to account for the execution of nonterminating programs, as we saw earlier. Smallstep operational semantics provides ﬁner, more detailed information concerning program execution. To present it, we deﬁne a conﬁguration as a pair c, Mem ∈ Lang3 × M of a program c and a memory state Mem. We then deﬁne a transition relation between conﬁgurations using the inference system deﬁned in Box 4.3. Intuitively, c1 , Mem1 →Env c2 , Mem2 means that executing one step of the program c1 in a state (Env, Mem1 ) – which corresponds to the conﬁguration
c1 , Mem1 – leads to a state (Env, Mem2 ) in which c2 is the rest of the program to be executed: this corresponds to the conﬁguration c2 , Mem2 . A conﬁguration is said to be terminal if it takes the form skip, Mem, and in this case, the memory M em will no longer be modiﬁed. Formally, the only two rules leading to a terminal conﬁguration correspond to an assignment (S1 ) and to an exit of loop (S7 ).
94
Concepts and Semantics of Programming Languages 1
(S1 ) (S2 )
em1 = r ∈ R eM E nv = (v, M em2 ) ∈ V × M
x := e, Mem1 →Env skip, Mem2 [r := v]
E nv(x)
c1 , Mem1 →Env c1 , Mem2
c1 ; c2 , Mem1 →Env c1 ; c2 , Mem2
(S3 )
skip ; c, Mem →Env c, Mem
(S4 )
em1 eM E nv = (true, M em2 )
if e then c1 else c2 , Mem1 →Env c1 , Mem2
(S5 )
em1 eM E nv = (f alse, M em2 )
if e then c1 else c2 , Mem1 →Env c2 , Mem2
(S6 )
em1 eM E nv = (true, M em2 )
while e do c, Mem1 →Env c ; while e do c, Mem2
(S7 )
em1 eM E nv = (f alse, M em2 )
while e do c, Mem1 →Env skip, Mem2
Box 4.3. Smallstep operational semantics of Lang3
A sequence of computations is a sequence, possibly inﬁnite, of the form:
c0 , Mem0 →Env c1 , Mem1 →Env c2 , Mem2 →Env · · · such that there exists an inference tree for ci , Memi →Env ci+1 , Memi+1 for all i ≥ 0. The relation →Env can then be deﬁned by c, Mem →Env Mem if, and only if, there exists a sequence of computations, containing only a ﬁnite number of transitions (or zero transitions)
c0 , Mem0 →Env c1 , Mem1 →Env c2 , Mem2 →Env · · · →Env ck , Memk such that c = c0 , Mem = Mem0 and Mem = Memk and such that the conﬁguration
ck , Memk is terminal (i.e. ck = skip). If we know the length of the sequence, as in this case, we write c, Mem →kEnv Mem . E XAMPLE 4.4.– Consider the execution of the program fragment shown in example 4.1: while not(↓ x =↓ y) do if ↓ x ≤↓ y then y :=↓ y− ↓ x else x :=↓ x− ↓ y e c1 c2 ew cw c
Semantics of Imperative Features
95
in a state (Env, Mem) = ((y, ry ) ⊕ (x, rx ) ⊕ Env0 , Mem0 [rx := 4][ry := 2]). The smallstep semantics of this program builds the following sequence: →Env →Env →Env →Env →Env
em (S6 ) ew M E nv = (true, M em) em (S2 , S5 ) eM E nv = (f alse, M em) (S2 , S1 ) Env(x) = rx and em ↓ x− ↓ yM E nv = (2, M em)
skip ; c, Mem[rx := 2] (S3 ) M em[r :=2]
c, Mem[rx := 2] (S7 ) ew Env x = (f alse, Mem[rx := 2])
skip, Mem[rx := 2]
c, Mem
cw ; c, Mem
c2 ; c, Mem
So we get: c, Mem →Env Mem[rx := 2]. 4.4.3. Expressiveness of operational semantics As shown by examples 4.3 and 4.4, when used to describe the execution of terminating programs in Lang3 , bigstep and smallstep operational semantics produce the same result. This semantic equivalence is proved in section 4.5.4. Nevertheless, as it provides a ﬁner description of executions of programs, smallstep semantics can handle certain features of some programming languages, which cannot be described using bigstep semantics. Parallelism is one such feature. Consider the construction c1 c2 which denotes the parallel execution of two programs c1 and c2 . Bigstep semantics only allows us to describe memory states after the execution of programs c1 and c2 ; in this approach, the only possible ways of executing c1 c2 are to execute c1 then execute c2 , or to execute c2 then execute c1 : (CP1 )
Mem, c1 Env Mem Mem , c2 Env Mem
Mem, c1 c2 Env Mem
(CP2 )
Mem, c2 Env Mem Mem , c1 Env Mem
Mem, c1 c2 Env Mem
For example, using these rules, there are only two possible executions of program c: x := 2 (x := 3 ; x :=↓ x∗ ↓ x) giving us:
Mem, c Env Mem[Env(x) := 2] and Mem, c Env Mem[Env(x) := 9]
96
Concepts and Semantics of Programming Languages 1
This semantics is unable to take account of possible interleavings between the executions of c1 and c2 . Smallstep semantics, on the other hand, allows us to describe the execution of just one part of a program: (SP1 )
c1 , Mem →Env c1 , Mem
c1 c2 , Mem →Env c1 c2 , Mem
(SP2 )
c2 , Mem →Env c2 , Mem
c1 c2 , Mem →Env c1 c2 , Mem
(SP3 )
c1 , Mem →Env skip, Mem
c2 , Mem →Env skip, Mem 4 (SP )
c1 c2 , Mem →Env c2 , Mem
c1 c2 , Mem →Env c1 , Mem
These rules can be used to describe the following three executions of the program c:
c, Mem →Env →Env →Env →Env
x := 3 ; x :=↓ x∗ ↓ x, Mem[Env(x) := 2]
skip ; x :=↓ x∗ ↓ x, Mem[Env(x) := 3]
x :=↓ x∗ ↓ x, Mem[Env(x) := 3]
skip, Mem[Env(x) := 9]
c, Mem →Env →Env →Env →Env
x := 2 (skip ; x :=↓ x∗ ↓ x), Mem[Env(x) := 3]
x := 2 x :=↓ x∗ ↓ x, Mem[Env(x) := 3]
x := 2, Mem[Env(x) := 9]
skip, Mem[Env(x) := 2]
c, Mem →Env →Env →Env →Env
x := 2 (skip ; x :=↓ x∗ ↓ x), Mem[Env(x) := 3]
x := 2 x :=↓ x∗ ↓ x, Mem[Env(x) := 3]
x :=↓ x∗ ↓ x, Mem[Env(x) := 2]
skip, Mem[Env(x) := 4]
4.5. Semantic properties In this section, we present applications of the semantic deﬁnitions given earlier, along with a few classical proof techniques using these formalisms. 4.5.1. Equivalent programs As an operational semantics describes program behavior, it is possible to identify “equivalent” programs, i.e. programs whose executions terminate and lead to the same memory state. Using bigstep operational semantics, this equivalence relation can be deﬁned formally as follows:
Semantics of Imperative Features
97
D EFINITION 4.1.– Equivalent programs: two programs c1 and c2 are equivalent, noted c1 ≡ c2 , if and only if: ∀Env ∈ E ∀Mem, Mem ∈ M
Mem, c1 Env Mem ⇔ Mem, c2 Env Mem
E XAMPLE 4.5.– Consider the following two programs: c1 : (if e then c else c ) ; c
c2 : if e then c ; c else c ; c
To prove that c1 ≡ c2 , we build an inference tree for Mem, c1 Env Mem from an inference tree for Mem, c2 Env Mem and vice versa. To this end, we examine all the possible rules for constructing such a tree, distinguishing between two possible cases for the evaluation of e. We must then simply “combine” the inference trees. If e is evaluated as true, then, from the tree em eM 1 E nv (Ci ) = (true, Mem1 )
Mem1 , c Env Mem2 2 (C4 ) (Cj )
Mem, if e then c else c Env Mem2
Mem2 , c Env Mem (C3 )
Mem, (if e then c else c ) ; c Env Mem we can obtain the tree
1 2 (Ci ) (Cj ) em
M em , c M em
M em , c Env Mem eM 1 2 2 E nv E nv (C3 ) = (true, Mem1 )
Mem1 , c ; c Env Mem (C4 )
Mem, if e then c ; c else c ; c Env Mem
and vice versa. A similar reasoning applies when e is evaluated as f alse. E XAMPLE 4.6.– Consider the following two programs: c1 : c ; (if e then c else c )
c2 : if e then c ; c else c ; c
To show that c1 ≡ c2 , it is enough to build a counterexample where c, c and c are instructions x := 0, x :=↓ x + 1 and x :=↓ x + 5, respectively, and where e is the expression 1 ≤↓ x. Starting from a state (Env, Mem) such that Env(x) = rx ∈ R and M em(rx ) = 2, we obtain:
Mem, c ; (if e then c else c ) Env Mem1 with Mem1 (x) = 5
Mem, if e then c ; c else c ; c Env Mem2 with Mem2 (x) = 1
98
Concepts and Semantics of Programming Languages 1
4.5.2. Program termination The existence of an inference tree describing the execution of a program guarantees that this execution is ﬁnite. As all inference trees are ﬁnite, showing that a program c terminates from a starting state (Env, Mem) using bigstep operational semantics comes down to proving the existence of a memory state Mem such that
Mem, c Env Mem . E XAMPLE 4.7.– Returning to the program fragment used in example 4.1: while not(↓ x =↓ y) do if ↓ x ≤↓ y then y :=↓ y− ↓ x else x :=↓ x− ↓ y e c1 c2 ew cw c
let us show that the execution of this program terminates when starting from any memory state storing strictly positive integer values at locations x and y. Formally, we must prove that: ∀Env ∈ E ∀Mem ∈ M (Mem(Env(x)) ≥ 1 and Mem(Env(y)) ≥ 1) ⇒ ∃Mem ∈ M Mem, c Env Mem To prove this property, given an environment Env ∈ E, we deﬁne the set: MEnv = {Mem  Mem(Env(x)) ≥ 1 and Mem(Env(y)) ≥ 1} ⊆ M and we must prove that: ∀Mem ∈ MEnv ∃Mem ∈ M Mem, c Env Mem P E nv ( M em)
The proof is obtained by wellfounded induction. First, we deﬁne a relation Env over the elements of MEnv . Then we show that this relation is a wellfounded order relation, that is, there is no strictly decreasing inﬁnite sequence of values according to Env . Finally, we prove that for any element Mem ∈ MEnv , under the hypothesis PEnv (Mem ) for all Mem ≺Env Mem, we can get PEnv (Mem). Since at least one value associated with the two variables x and y decreases for each execution of the loop body, the relation Env over MEnv can be deﬁned as follows: ⎛ ⎞ M em1 ( E nv(x)) ≤ M em2 ( E nv(x)) ⎠ M em1 E nv M em2 ⇔ ⎝ and M em1 ( E nv(y)) ≤ M em2 ( E nv(y)) and ∀z ∈ X \ {x, y} Mem1 (Env(z)) = Mem2 (Env(z))
Semantics of Imperative Features
99
The strict order ≺Env associated with Env is thus deﬁned by: ⎛ ⎞ M em1 ( E nv(x)) = M em2 ( E nv(x)) ⎜ ⎟ or Mem1 (Env(y)) = Mem2 (Env(y)) ⎜ ⎟ ⎟ M em ( E nv(x)) ≤ M em ( E nv(x)) and M em1 ≺ E nv M em2 ⇔ ⎜ 1 2 ⎜ ⎟ ⎝ and Mem1 (Env(y)) ≤ Mem2 (Env(y)) ⎠ and ∀z ∈ X \ {x, y} Mem1 (Env(z)) = Mem2 (Env(z)) In the below proof, it is helpful to note that if Mem1 ≺Env Mem2 , then < Mem2 (Env(x)) or Mem1 (Env(y)) < Mem2 (Env(y)). Relation Env clearly deﬁnes an order relation (i.e. it is reﬂexive, antisymmetric and transitive). M em1 ( E nv(x))
Let us prove that Env is well founded. Suppose that there exists a strictly decreasing inﬁnite sequence: M em1
Env Mem2 Env · · ·
The memories in MEnv only associate strictly positive values to variables x and y and these values themselves decrease or remain the same. As there is no inﬁnite decreasing sequence of strictly positive integers, there exists some rank k such that all elements of this inﬁnite sequence of rank greater than k are equal. Thus, we would have: ∀j ≥ k
M emk ( E nv(x))
= Memj (Env(x)) and Memk (Env(y)) = Memj (Env(y))
But, by deﬁnition, if Memk Env Memk+1 , then: M emk ( E nv(x))
= Memk+1 (Env(x)) or Memk (Env(y)) = Memk+1 (Env(y))
This is a contradiction: our hypothesis is false. Now, let us prove PEnv (Mem) by wellfounded induction. Take Mem ∈ MEnv . By induction hypothesis, we have ∀Mem ≺Env Mem, PEnv (Mem ) and we must prove the property PEnv (Mem). Let Mem(Env(x)) = m and Mem(Env(y)) = n. Two cases are possible: 1) if m = n, then we get PEnv (Mem) since we can construct the following tree: (C7 )
em ew M E nv = (f alse, M em)
Mem, c Env Mem
2) if m = n, then there are two possible subcases: a) if m ≤ n, then we can construct the tree 1 : em ↓ y− ↓ xM E nv = (n − m, M em)
Mem, c1 Env Mem[Env(y) := n − m]
Mem, cw Env Mem[Env(y) := n − m]
em eM E nv = (true, M em)
(C4 )
(C1 )
100
Concepts and Semantics of Programming Languages 1
But, by deﬁnition, Mem[Env(y) := n − m] ≺Env Mem, and by induction hypothesis, we obtain the tree 2 : .. . (Ci )
Mem[Env(y) := n − m], cw Env Mem b) if m > n, then we can build the tree 1 : em ↓ x− ↓ yM E nv = (m − n, M em)
Mem, c2 Env Mem[Env(x) := m − n]
Mem, cw Env Mem[Env(x) := m − n]
em eM E nv = (f alse, M em)
(C5 )
(C1 )
Once again, we obtain Mem[Env(x) := m − n] ≺Env Mem, and using the induction hypothesis, we obtain the tree 2 : .. . (Cj )
Mem[Env(x) := m − n], cw Env Mem In all cases, we get the result since we can construct the following tree: em ew M E nv = (true, M em) (C6 )
1
Mem, c Env Mem
2
This proof is typically used to show that a program terminates: ﬁrst, a wellfounded order must be built (a point which is not always immediately obvious), then the proof can be done by wellfounded induction. 4.5.3. Determinism of program execution To ensure that the same program executed from the same state will always produce the same result, we prove that the execution of instructions is deterministic. P ROPOSITION 4.1.– Determinism of instruction execution: ∀c ∈ Lang3 ∀Env ∈ E ∀Mem, Mem1 , Mem2 ∈ M ( Mem, c Env Mem1 and Mem, c Env Mem2 ) ⇒ Mem1 = Mem2 We ﬁrst attempt to prove this proposition by induction over the program structure. For the program while e do c, by induction hypothesis, we may suppose that the property is veriﬁed for c. Using the following two trees .. .. . . (Cj ) (Cj )
Mem , c Env Mem1
Mem , c Env Mem1 1 2 we therefore obtain Mem1 = Mem1 .
Semantics of Imperative Features
101
In the case where e is evaluated as true, we must prove the equality between two memory states Mem2 and Mem2 , obtained in the following manner (note that the evaluation of expressions is deterministic): .. em . eM E nv (C ) k = (true, Mem )
M em1 , while e do c E nv M em2 1 (C6 )
Mem, while e do c Env Mem2 .. em . eM E nv (Ck ) = (true, Mem )
Mem1 , while e do c Env Mem2 2 (C6 )
Mem, while e do c Env Mem2 The induction hypothesis over c establishes the equality Mem1 = Mem1 . To deduce Mem2 = Mem2 , an induction hypothesis over the instruction while e do c would also be needed but there is no way to get it. Hence, reasoning by induction over the structure of programs is not sufﬁcient to prove the determinism of program execution. In this case, instead of using the induction scheme generated by the inductive deﬁnition of Lang3 , we need to use the induction scheme generated by the inference system deﬁning the relation
Mem, c Env Mem . This leads to reason by induction over executions instead of programs. A formal deﬁnition of this induction scheme is shown in Box 4.4. em ∀Mem ∈ M Env(x) = r ∈ R and eM E nv = (v, M em ) ∈ V × M ⇒ P ( Mem, x := e Env Mem [r := v]) and ∀Mem ∈ M P ( Mem, skip Env Mem) and ∀c1 , c2 ∈ Lang3 ∀Mem, Mem1 , Mem2 ∈ M (P ( Mem, c1 Env Mem1 ) and P ( Mem1 , c2 Env Mem2 )) ⇒ P ( Mem, c1 ; c2 Env Mem2 ) and ∀c1 , c2 ∈ Lang3 ∀Mem, Mem1 , Mem2 ∈ M ∀e ∈ Exp3 em (eM E nv = (true, M em1 ) and P ( M em1 , c1 E nv M em2 )) ⇒ P ( Mem, if e then c1 else c2 Env Mem2 ) and ∀c1 , c2 ∈ Lang3 ∀Mem, Mem1 , Mem2 ∈ M ∀e ∈ Exp3 em (eM E nv = (f alse, M em1 ) and P ( M em1 , c2 E nv M em2 )) ⇒ P ( Mem, if e then c1 else c2 Env Mem2 ) and ∀c ∈ Lang3 ∀Mem, Mem1 , Mem2 , Mem3 ∈ M ∀e ∈ Exp3 em (eM E nv = (true, M em1 ) and P ( M em1 , c E nv M em2 ) and P ( Mem2 , while e do c Env Mem3 )) ⇒ P ( Mem, while e do c Env Mem3 ) and ∀c ∈ Lang3 ∀Mem, Mem1 ∈ M ∀e ∈ Exp3 em eM E nv = (true, M em1 ) ⇒ P ( M em, while e do c E nv M em1 ) then ∀c ∈ Lang3 ∀Mem, Mem ∈ M P ( Mem, c Env Mem )
if
Box 4.4. Reasoning scheme by structural induction over Mem, c Env Mem
102
Concepts and Semantics of Programming Languages 1
P ROOF.– (This is the proof of proposition 4.1) The property to prove P ( Mem, c Env Mem1 ) is deﬁned by: ∀Mem2 ∈ M Mem, c Env Mem2 ⇒ Mem1 = Mem2 We prove the following points. – The form of the tree for Mem, x := e Env Mem1 can only be: (C1 )
E nv(x)
em =r∈R eM E nv = (v, M em ) ∈ V × M
Mem, x := e Env Mem [r := v]
hence Mem1 = Mem [r := v]. As the evaluation of expressions is deterministic, this is the only possible tree for Mem, x := e Env Mem2 , so Mem1 = Mem2 . – The form of the tree for Mem, skip Env Mem1 can only be: (C2 )
Mem, skip Env Mem
This is also the only possible tree for Mem, skip Env Mem2 , so Mem1 = = Mem2 .
M em
– The tree for Mem, c1 ; c2 Env Mem1 must take the form: .. .. . . (Ci ) (Cj )
Mem, c1 Env Mem
Mem , c2 Env Mem1 (C3 )
Mem, c1 ; c2 Env Mem1 Similarly, the tree for Mem, c1 ; c2 Env Mem2 must take the form: .. .. . . (Ci ) (Cj )
Mem, c1 Env Mem
Mem , c2 Env Mem2 (C3 )
Mem, c1 ; c2 Env Mem2 By induction hypothesis, we have P ( Mem, c1 Env Mem ) and thus Mem = M em and since we have, still from the induction hypothesis, P ( M em , c2 E nv M em1 ), we ﬁnally obtain M em1 = M em2 . em – When eM E nv = (true, M em ), the tree for M em, if e then c1 else c2 E nv must take the form: .. . em eM = (true, M em ) (C ) i E nv
Mem , c1 Env Mem1 (C4 )
Mem, if e then c1 else c2 Env Mem1
M em1
Semantics of Imperative Features
103
Similarly, the tree for Mem, if e then c1 else c2 Env Mem2 must take the form: .. . M em eEnv = (true, Mem ) (Cj )
Mem , c1 Env Mem2 (C4 )
Mem, if e then c1 else c2 Env Mem2 By induction hypothesis, we have P ( Mem , c1 Env Mem1 ) and thus Mem1 = em M em2 . The reasoning used when e M E nv = (f alse, M em ) is similar. em – When eM E nv = (true, M em ), the tree for M em, while e do c E nv M em1 must take the form: em eM E nv = (Ci )
Mem , c (true, Mem )
(C6 )
.. .. . . (Cj ) Env Mem1
Mem1 , while e do c Env Mem1
Mem, while e do c Env Mem1
Similarly, the tree for Mem, while e do c Env Mem2 must take the form: em eM E nv = (Ci )
Mem , c (true, Mem )
(C6 )
.. .. . . (C ) j Env Mem2
Mem2 , while e do c Env Mem2
Mem, while e do c Env Mem2
By induction hypothesis, we have P ( Mem , c Env Mem1 ) and thus M em1 = M em2 , and since we have, here again by induction hypothesis, P ( Mem1 , while e do c Env Mem1 ), we ﬁnally obtain Mem1 = Mem2 . em – When eM E nv = (f alse, M em ), the tree for M em, while e do c E nv M em1 must take the form:
(C7 )
em eM E nv = (f alse, M em )
Mem, while e do c Env Mem
This is also the only possible tree for Mem, while e do c Env Mem2 and we therefore obtain Mem1 = Mem = Mem2 . 4.5.4. Big steps versus small steps Bigstep and smallstep operational semantics provide two ways to describe the execution of programs in Lang3 . These descriptions are based on different formalisms, but are identical from a semantic perspective. As we can see from examples 4.3 and 4.4, the same ﬁnite execution of a program may be described by a
104
Concepts and Semantics of Programming Languages 1
tree, in the case of bigstep semantics, or by a transition sequence in the case of smallstep semantics; the results obtained using the two approaches are identical. We now formally prove this semantic equivalence. We begin by proving the following lemma. L EMMA 4.1.– If c1 ; c2 , Mem →kEnv Mem , then there exists Mem ∈ M and two integers k1 and k2 such that c1 , Mem →kE1nv Mem and c2 , Mem →kE2nv Mem with k = k1 + k2 + 1. P ROOF.– We proceed by induction over k. If k = 0, the property is trivial (since in this case the hypothesis c1 ; c2 , Mem →0Env Mem cannot be veriﬁed, since
c1 ; c2 , Mem is not a terminal conﬁguration). If k = k0 + 1, two cases are possible: 1) If the transition sequence in the hypothesis is:
c1 ; c2 , Mem →Env c1 ; c2 , Mem1 →Env · · · →Env skip, Mem sequence of length k0 with c1 , Mem →Env c1 , Mem1 , then, by induction hypothesis, there exists two k k integers k1 and k2 such that c1 , Mem1 →E1nv Mem1 and c2 , Mem1 →E2nv Mem with k0 = k1 + k2 + 1. We construct the following sequence of length k1 + 1: k
c1 , Mem →Env c1 , Mem1 →E1nv Mem1 which allows us to conclude with k1 = k1 + 1 and k2 = k2 . 2) If the transition sequence in the hypothesis is:
skip ; c2 , Mem →Env c2 , Mem →Env · · · →Env skip, Mem sequence of length k0 then we can conclude directly, with k1 = 0 and k2 = k0 .
The equivalence between the bigstep and smallstep operational semantics can now be proved. P ROPOSITION 4.2.– ∀c ∈ Lang3 ∀Env ∈ E ∀Mem1 , Mem2 ∈ M:
Mem1 , c Env Mem2 ⇔ c, Mem1 →Env Mem2 P ROOF.– (⇒) We proceed by induction over Mem1 , c Env Mem2 (applying the induction scheme presented in Box 4.4).
Semantics of Imperative Features
105
If Mem1 , c Env Mem2 has been obtained using rule C1 : (C1 )
E nv(x)
em1 =r∈R eM E nv = (v, M em) ∈ V × M
Mem1 , x := e Env Mem[r := v]
then rule S1 gives us x := e, Mem1 →Env skip, Mem[r := v] and since
skip, Mem[r := v] is a terminal conﬁguration, we get:
x := e, Mem1 →Env Mem[r := v] If Mem1 , c Env Mem2 has been obtained using rule C2 , then c is the instruction skip and Mem1 = Mem2 , and since skip, Mem2 is a terminal conﬁguration, by considering a sequence of length zero we get:
skip, Mem1 →Env Mem2 If Mem1 , c Env Mem2 has been obtained using rule C3 : .. .. . . (Ci ) (Cj )
Mem1 , c1 Env Mem
Mem, c2 Env Mem2 (C3 )
Mem1 , c1 ; c2 Env Mem2 by induction hypothesis, we have c1 , Mem1 →Env Mem and c2 , Mem →Env M em2 . Hence, we have two transition sequences:
c1 , Mem1 →Env · · · →Env skip, Mem
[4.4]
c2 , Mem →Env · · · →Env skip, Mem2
[4.5]
Applying rule S2 to each of the transitions in sequence [4.4], we can build the following sequence:
c1 ; c2 , Mem1 →Env · · · →Env skip ; c2 , Mem
[4.6]
and since rule S3 allows us to construct the transition:
skip ; c2 , Mem →Env c2 , Mem
[4.7]
by concatenating [4.6], [4.7] and [4.5], we get the following sequence:
c1 ; c2 , Mem1 →Env · · · →Env skip, Mem2 and thus c1 ; c2 , Mem1 →Env Mem2 . – If Mem1 , c Env Mem2 has been obtained using rule C4 : .. .
Mem, c1 Env Mem2
Mem1 , if e then c1 else c2 Env Mem2
em1 eM E nv = (true, M em)
(C4 )
(Ci )
106
Concepts and Semantics of Programming Languages 1
then, by induction hypothesis, we have c1 , Mem1 →Env Mem and there exists a transition sequence:
c1 , Mem1 →Env · · · →Env skip, Mem from which, by applying rule S4 , we can obtain:
if e then c1 else c2 , Mem1 →Env c1 , Mem1 →Env · · · →Env skip, Mem which gives us if e then c1 else c2 , Mem1 →Env Mem2 . – If Mem1 , c Env Mem2 has been obtained using rule C5 , the reasoning is similar to that used in the previous case. – If Mem1 , c Env Mem2 has been obtained using rule C6 : em1 eM E nv (Ci ) =
Mem, c0 (true, Mem)
(C6 )
.. .. . . (C ) j Env Mem
Mem , while e do c0 Env Mem2
Mem1 , while e do c0 Env Mem2
then, by induction hypothesis, we get c0 , Mem →Env M em and
while e do c0 , Mem →Env Mem2 , and we thus have two transition sequences:
c0 , Mem →Env · · · →Env skip, Mem
while e do c0 , Mem →Env · · · →Env skip, Mem2
[4.8] [4.9]
Applying rule S2 to each of the transitions in sequence [4.8], the following sequence can be constructed:
c0 ; while e do c0 , Mem →Env · · · →Env skip ; while e do c0 , Mem [4.10] and since rules S3 and S6 allow us to construct transitions:
skip ; while e do c0 , Mem →Env while e do c0 , Mem
[4.11]
while e do c0 , Mem1 →Env c0 ; while e do c0 , Mem [4.12] by concatenating [4.12], [4.10], [4.11] and [4.9], we obtain the following sequence:
while e do c0 , Mem1 →Env · · · →Env skip, Mem2 hence while e do c0 , Mem1 →Env Mem2 . – If Mem1 , c Env Mem2 has been obtained using rule C7 , then c is the em1 = (f alse, Mem2 ). Rule S7 gives us the instruction while e do c0 and eM E nv
Semantics of Imperative Features
107
transition while e do c, Mem1 →Env skip, Mem2 and since skip, Mem2 is a terminal conﬁguration, we obtain while e do c0 , Mem1 →Env Mem2 . (⇐) Since, by hypothesis c, Mem1 →Env Mem2 , there exists a sequence of transitions of length k:
c, Mem1 →Env c1 , Mem1 →Env · · · →Env ck , Memk = skip, Mem2 Proof is obtained by wellfounded induction over k then by case over the ﬁrst transition in this sequence. If k = 0, then the starting conﬁguration is terminal, M em1 = M em2 and rule C2 allows us to conclude. Otherwise, there are several possible cases. – If the ﬁrst transition is x := e, Mem1 →Env skip, Mem[r := v] (obtained using rule S1 ), then k = 1 since skip, Mem[r := v] is a terminal conﬁguration, and using rule C1 we get:
Mem1 , x := e Env Mem[r := v] – If the ﬁrst transition is obtained using rule S2 :
c1 ; c2 , Mem1 →Env c1 ; c2 , Mem1 →Env · · · →Env ck , Memk c1 =skip, M em2 with c1 , Mem1 →Env c1 , Mem1 , then, by lemma 4.1, there exists two integers k1 and k2 such that c1 , Mem1 →kE1nv Mem and c2 , Mem →kE2nv Mem2 with k1 + k2 + 1 = k. Hence k1 < k and k2 < k and, by induction hypothesis, we get Mem1 , c1 Env Mem and Mem , c2 Env Mem2 and we can conclude by building the tree: .. .. . . (Ci ) (Cj )
Mem1 , c1 Env Mem
Mem , c2 Env Mem2 (C3 )
Mem1 , c1 ; c2 Env Mem2 – If the ﬁrst transition is obtained using rule S3 :
skip ; c1 , Mem1 →Env c1 , Mem1 →Env · · · →Env ck , Memk =skip, M em2 with Mem1 = Mem1 , then we get Mem1 , c1 Env Mem2 by induction hypothesis and we can conclude by constructing the tree: .. .. . . (C2 ) (Ci )
Mem1 , skip Env Mem1
Mem1 , c1 Env Mem2 (C3 )
Mem1 , skip ; c1 Env Mem2
108
Concepts and Semantics of Programming Languages 1
– If the ﬁrst transition is obtained using rule S4 :
if e then c1 else c2 , Mem1 →Env c1 , Mem1 →Env · · · →Env ck , Memk =skip, M em2 em1 with eM = (true, Mem1 ), then Mem1 , c1 Env Mem2 by induction E nv hypothesis, and we conclude by constructing the tree: .. . M em1 eEnv = (true, Mem1 ) (Ci )
Mem1 , c1 Env Mem2 (C4 )
Mem1 , if e then c1 else c2 Env Mem2
– Similar reasoning is used if the ﬁrst transition is obtained using rule S5 . – If the ﬁrst transition is obtained using rule S6 :
while e do c0 , Mem1 →Env c0 ; while e do c0 , Mem1 →Env · · · →Env ck , Memk c1 =skip, M em2 em1 with eM = (true, Mem1 ), then Mem1 , c1 Env Mem2 by induction E nv hypothesis, and we can construct the tree: .. . M em1 eEnv = (true, Mem1 ) (Ci )
Mem1 , c1 Env Mem2 (C4 )
Mem1 , if e then (c0 ; while e do c0 ) else skip Env Mem2
and since we can show (see exercise 4.1) that: while e do c0 ≡ if e then (c0 ; while e do c0 ) else skip we obtain Mem1 , while e do c0 Env Mem2 . – If the ﬁrst transition is obtained using rule S7 :
while e do c0 , Mem1 →Env skip, Mem2 em1 with eM E nv = (f alse, M em2 ) (and thus k = 1), we can construct the tree:
(C7 )
em1 eM E nv = (f alse, M em2 )
Mem1 , while e do c0 Env Mem2
Semantics of Imperative Features
109
4.6. Procedures Procedures are used to structure the code of a program, isolating certain parts and making them accessible from any point in the program (by means of a procedure call instruction). We shall begin by introducing the notion of a block, which conﬁnes deﬁnitions of variables and procedures in a portion of code. We will then build on this notion to introduce procedures, which trigger the execution of a block, and outline different possible modes of parameter transmission. The concepts of functions and procedures give the basis of program architecture and will be extended with modular and object features in Volume 2. The bigstep operational semantics formalism will be used throughout this section as a basis for presenting new constructs. 4.6.1. Blocks The syntax of Lang3 is extended by adding the construct: begin d c end where d is a sequence (which may be empty) of deﬁnitions in language Def 3 (which may be extended in order to allow local deﬁnition of procedures) and c is a program in Lang3 . This new construct ensures that the variables and procedures deﬁned by d will be accessible during an execution of c. The execution of begin d c end in a state (Env, Mem) is a twostep process: 1) construction of the state in which program c is to be executed by evaluating the deﬁnitions in d: d
(Env, Mem) →Def 3 (Env1 , Mem1 ) where →Def 3 is the relation deﬁned in section 4.2 (adapted for language Def 3 ); 2) execution of the program c in state (Env1 , Mem1 ):
Mem1 , c Env1 Mem2 To take account of this new construct, the rule C8 given in Box 4.5 is added to the rules in Box 4.2. d
(Env, Mem) →Def 3 (Env1 , Mem1 )
Mem1 , c Env1 (C8 )
Mem, begin d c end Env Mem2
Box 4.5. Bigstep operational semantics: blocks
M em2
110
Concepts and Semantics of Programming Languages 1
Once again, only the memory can be modiﬁed by executing a block: the evaluation of the declarations d transforms Mem into Mem1 , leading to the state where the execution of the instruction c starts; this execution leads to Mem2 . The starting environment is only temporarily modiﬁed during execution of c. We introduce the syntactic construct that adds blocks to language Lang3 . First, with Python, we deﬁne the following class: Python
class IBloc4: def __init__(self,list_def,pgm): self.list_def = list_def self.pgm = pgm
With OCaml, the following constructor is added to the deﬁnition of the sum type ’a pgm4 (see section 4.1): OCaml
 IBloc4 of (’a def4) list * ’a pgm4
The function exec_pgm4, which executes programs in Lang3 (see section 4.1), is extended to handle this new construct. With Python, we simply insert the following instruction into the function code: Python
if isinstance(c,IBloc4): env1,mem1 = trans_def4_exec((env,mem),c.list_def) return exec_pgm4(env1,mem1,c.pgm)
With OCaml, the following pattern is added to the function: OCaml
 IBloc4 (d, c1) > let (env1, mem1) = trans_def4_exec (env, mem) d in exec_pgm4 env1 mem1 c1
E XAMPLE 4.8.– In a state (Env, Mem) deﬁned by Env = (y, ry ) ⊕ (x, rx ) ⊕ Env0 and Mem = Mem0 [rx := 3][ry := 0] where (Env0 , Mem0 ) is any given state, the execution of a block c deﬁned by: begin (let z = x; let w =↑ ↓ x; let x = 5;) z := x ; y := w end d
cB
Semantics of Imperative Features
111
is obtained by, ﬁrst, constructing the state in which the block instructions are to be executed: (Env, Mem) let z = x; ((z, rx ) ⊕ Env, Mem) −−−−−−−→Def 3 let w =↑ ↓ x; −−−−−−−−−−→Def 3 ((w, ⎛ rw ) ⊕ (z, rx ) ⊕ Env, Mem[rw := 3])
⎞
⎜ ⎟ ⎝(x, 5) ⊕ (w, rw ) ⊕ (z, rx ) ⊕ Env, Mem[rw := 3]⎠
let x = 5; −−−−−−−→Def 3
E nv1
M em1
Using rule C8 , we obtain the following execution tree:
d
1 2 (Env, Mem) →Def 3 (Env1 , Mem1 ) (C3 )
Mem1 , cB Env1 Mem (C8 )
Mem, c Env Mem1 [rx := 5][ry := rw ]
M em
where
1
and
1
2
2
: (C1 )
are the trees: em1 = rx xM E nv1 = (5, M em1 )
Mem1 , z := x Env1 Mem1 [rx := 5]
E nv1 (z)
M em [r :=5]
= ry wEnv1 1 x = (rw , Mem1 [rx := 5]) : (C1 )
Mem1 [rx := 5], y := w Env1 Mem1 [rx := 5][ry := rw ] E nv1 (y)
During execution of block c, the environment Env1 binds the variable x to the value 5, and the variable z to the reference rx . Hence, assignment z := x stores the value 5 in the location denoted by rx . Allocation also takes place during execution of the block: the environment Env1 binds variable w to a reference rw . At the end of execution of this block, environment Env1 is no longer accessible and this binding ceases to be available. Nevertheless, the assignment y := w stored the reference rw into the location denoted by ry , and the contents of the location denoted by rw thus remain accessible by evaluating the expression ↓ ↓ y. The deﬁnitions of state (Env, Mem) and block c (both introduced at the beginning of this example), and the construction of state (Env, Mem ) such that
Mem, c Env Mem , are obtained as follows:
112
Concepts and Semantics of Programming Languages 1
Python
(env_ex_ld4_3,mem_ex_ld4_3) \ = trans_def4_exec(([],[]), [Let_def4("x",Ref4(Cste4(CInt4(3)))), Let_def4("y",Ref4(Cste4(CInt4(0))))]) ex_pgm4_2 = IBloc4([Let_def4("z",Var4("x")), Let_def4("w",Ref4(Deref4(Var4("x")))), Let_def4("x",Cste4(CInt4(5)))], ISeq4(IAffect4("z",Var4("x")),IAffect4("y",Var4("w")))) >>> print_state((env_ex_ld4_3,exec_pgm4(env_ex_ld4_3,mem_ex_ld4_3,ex_pgm4_2))) [ ( y , VCste4 (’CRef4’, 1) ), ( x , VCste4 (’CRef4’, 0) ), ] [ ( 2 , (’CInt4’, 3) ), ( 1 , (’CRef4’, 2) ), ( 0 , (’CInt4’, 5) ), ] OCaml
# let (env,mem) = (trans_def4_exec ([],[]) [Let_def4("x",Ref4(Cste4(CInt4(3)))); Let_def4("y",Ref4(Cste4(CInt4(0))))]) in let ex_pgm4_2 = IBloc4([Let_def4("z",Var4("x")); Let_def4("w",Ref4(Deref4(Var4("x")))); Let_def4("x",Cste4(CInt4(5)))], ISeq4(IAffect4("z",Var4("x")),IAffect4("y",Var4("w")))) in (env,(exec_pgm4 env mem ex_pgm4_2));;  : (string * string valeurs4) list * (int * const4) list = ([("y", VCste4 (CRef4 1)); ("x", VCste4 (CRef4 0))], [(2, CInt4 3); (1, CRef4 2); (0, CInt4 5)])
4.6.2. Procedures 4.6.2.1. Language of deﬁnitions A new syntactic construct is added to Def 3 : proc p(x){c}. This construct deﬁnes a procedure named p (which is an identiﬁer), having a parameter denoted by a variable x and a body which is a program c in Lang3 . For the sake of simplicity, we shall limit our discussion to procedures with a single parameter. The descriptions below can easily be extended to procedures with multiple parameters, for example by replacing the single parameter with an nuple. For procedures without arguments, the argument is a value of the “empty” type. We implement the syntactic construct used to deﬁne procedures in Def 3 . With Python, we deﬁne the following class: Python
class Let_proc4: def __init__(self,name,param,pgm): self.name = name self.param = param self.pgm = pgm
Semantics of Imperative Features
113
With OCaml, the deﬁnition of the sum type ’a def4 (see section 4.1) is extended with the following constructor: OCaml
 Let_proc4 of ’a * ’a * ’a pgm4
The execution of a procedure is triggered by a call (also called an invocation) to it. As in Chapter 3, where we describe the application of functions, there are several possible choices for determining the state in which a procedure should be executed. Using a static (or lexical) scope mechanism, a procedure is executed in the state in which it was deﬁned (augmented with the binding for its parameter), and the value associated with the procedure is therefore a closure containing its deﬁnition environment. Using a dynamic scope mechanism, a procedure is executed in the state in which the call is done, and the value associated with the procedure is therefore a closure containing only the name of the parameter and the procedure body. In both cases, the binding of the procedure name to its value is added to the execution environment, allowing deﬁnitions of recursive procedures. Table 4.5 describes how relation →Def 3 is extended in the two cases. proc p(x){c} (Env, Mem) −−−−−−−−−−→Def 3 ((p, x, c, Env) ⊕ Env, Mem) proc p(x){c} Dynamic scope (Env, Mem) −−−−−−−−−−→Def 3 ((p, x, c) ⊕ Env, Mem) Static scope
Table 4.5. Extension of relation →Def 3 : procedures
Allowing the deﬁnition of procedures requires deﬁning a new value corresponding to a closure. Let us ﬁrst consider a static scope mechanism. With Python, we simply add the following class: Python
class Ferm4: def __init__(self,param,pgm,env): self.param = param self.pgm = pgm self.env = env
With OCaml, the sum type ’a valeurs4 (see section 4.2) must be extended with the following constructor: OCaml
 Ferm4 of ’a * ’a pgm4 * (’a * ’a valeurs4) list
114
Concepts and Semantics of Programming Languages 1
Using a dynamic scope mechanism, with Python, we add the following class: Python
class Ferm4: def __init__(self,param,pgm): self.param = param self.pgm = pgm
while with OCaml, the deﬁnition of the sum type ’a valeurs4 is extended with the constructor: OCaml
 Ferm4 of ’a * ’a pgm4
To extend the relation →Def 3 whilst taking into account procedure deﬁnitions, we need to modify the function trans_def4 deﬁned in section 4.3. With Python, in case of static scope, this is done by inserting the following instruction into the function code: Python
if isinstance(d,Let_proc4): return (ajout_liaison_env(env,d.name,Ferm4(d.symb,d.pgm,d.env)),mem)
and, in case of dynamic scope, the following instruction is inserted: Python
if isinstance(d,Let_proc4): return (ajout_liaison_env(env,d.name,Ferm4(d.symb,d.pgm)),mem)
With OCaml, with static scope, we extend the function trans_def4 with the following pattern: OCaml
 Let_proc4(p,x,c) > ((ajout_liaison_env env p (Ferm4(x,c,env))),mem)
with dynamic scope, we add the following pattern: OCaml
 Let_proc4(p,x,c) > ((ajout_liaison_env env p (Ferm4(x,c))),mem)
Semantics of Imperative Features
115
4.6.2.2. Procedure call mechanisms 4.6.2.2.1. Call by value When calling a procedure, the call by value mechanism transmits a value, which is the result of evaluating an expression e in the current environment of the call. Using this mechanism, the syntactic construct of a procedure call in Lang3 is given as: call p(e) To add the syntactic construct for a procedure call with Python, we deﬁne the following class: Python
class ICall4: def __init__(self,name,exp): self.name = name self.exp = exp
With OCaml, the deﬁnition of the sum type ’a pgm4 (see section 4.1) is extended with the following constructor: OCaml
 ICall4 of ’a * ’a exp4
Box 4.6 presents different semantics of this call mechanism. In the ﬁrst version, the transmitted value is seen as a constant: the binding of the parameter to the value is directly present in the environment (and thus this value cannot be modiﬁed). In the second version, the parameter of the procedure is considered as a mutable variable local to the procedure, which is initialized using the transmitted value. In this case, an allocation is done, the parameter references this location and the contents of this location may be modiﬁed by the instructions in the procedure body. As in the case of blocks, this reference may remain accessible once the procedure has been executed, if it was assigned to a variable “external” to the procedure. Both versions are presented for static and dynamic scope mechanisms.
116
Concepts and Semantics of Programming Languages 1
Call by value: constant parameter em = x, c, Envp eM E nv = (v, M em ) E nv = (p, x, c, E nvp ) ⊕ (x, v) ⊕ E nvp
Mem , c Env Mem (C9 )
Mem, call p(e) Env Mem
E nv(p)
Static scope
em = x, c eM E nv = (v, M em ) = (p, x, c) ⊕ (x, v) ⊕ Env
Mem , c Env Mem Dynamic scope (C9 )
Mem, call p(e) Env Mem
E nv(p) E nv
Call by value: mutable parameter em = x, c, Envp eM E nv = (v, M em ) = (p, x, c, Envp ) ⊕ (x, rx ) ⊕ Envp
Mem [rx := v], c Env Mem (C9 )
Mem, call p(e) Env Mem
E nv(p) E nv
Static scope
em = x, c eM E nv = (v, M em ) E nv = (p, x, c) ⊕ (x, rx ) ⊕ E nv
Mem [rx := v], c Env Mem Dynamic scope (C9 )
Mem, call p(e) Env Mem
E nv(p)
Call by reference = xp , c, Envp Env(x) = rx ∈ R = (p, xp , c, Envp ) ⊕ (xp , rx ) ⊕ Envp
Mem, c Env Mem (C9 )
Mem, call p(x) Env Mem E nv(p) E nv
Static scope
= xp , c Env(x) = rx ∈ R = (p, xp , c) ⊕ (xp , rx ) ⊕ Env
Mem, c Env Mem Dynamic scope (C9 )
Mem, call p(x) Env Mem E nv(p) E nv
Box 4.6. Bigstep operational semantics: calls to a procedure
Semantics of Imperative Features
117
Using a static scope mechanism and call by value with a constant parameter, we add the following instruction to the code of the Python function exec_pgm4 (see section 4.4.1): Python
if isinstance(c,ICall4): v,mem1 = eval_exp4(env,mem,c.exp) fp = valeur_de(env,c.name) if fp is None: raise ValueError if isinstance(fp,Ferm4) and isinstance(v,VCste4): enve = ajout_liaison(ajout_liaison_env(fp.env,fp.param,v), c.name,fp) return exec_pgm4(enve,mem1,fp.pgm) raise ValueError raise ValueError
In this case, with OCaml, the deﬁnition of function exec_pgm4 is extended to include the following pattern: OCaml
 ICall4(p,e) > (match ((valeur_de env p),(eval_exp4 env mem e)) with (Some(Ferm4(xp,cp,envp)),(VCste4(v),mem1)) > let enve = (ajout_liaison_env (ajout_liaison_env envp xp (VCste4(v))) p (Ferm4(xp,cp,envp))) in (exec_pgm4 enve mem1 cp)  _ > raise Exec_Error)
Using a dynamic scope mechanism and call by value with a constant parameter, we add the following instruction to the code of the Python function exec_pgm4: Python
if isinstance(c,ICall4): v,mem1 = eval_exp4(env,mem,c.exp) fp = valeur_de(env,c.name) if fp is None: raise ValueError if isinstance(fp,Ferm4) and isinstance(v,VCste4): enve = ajout_liaison(ajout_liaison_env(env,fp.param,v), c.name,fp) return exec_pgm4(enve,mem1,fp.pgm) raise ValueError
In this case, with OCaml, the deﬁnition of function exec_pgm4 is extended to include the following pattern: OCaml
 ICall4(p,e) > (match ((valeur_de env p),(eval_exp4 env mem e)) with (Some(Ferm4(xp,cp)),(VCste4(v),mem1))
118
Concepts and Semantics of Programming Languages 1
> let enve = (ajout_liaison_env (ajout_liaison_env env xp (VCste4(v))) p (Ferm4(xp,cp))) in (exec_pgm4 enve mem1 cp)  _ > raise Exec_Error)
We consider only the implementation of the two cases described above. The other cases (mutable parameter and call by reference) are obtained in similar ways. 4.6.2.2.2. Call by reference The call by reference mechanism transmits to the procedure a value, which is the reference of a mutable variable present in the current environment of the call. Using this mechanism, the syntactic construct of the procedure call in Lang3 is: call p(x) where x is a mutable variable in the calling environment. Its contents may be directly modiﬁed when executing the body of p. The formal deﬁnition of this mechanism is presented in Box 4.6. Unlike the call by value mechanism, where the parameter is considered as a variable that is local to the procedure, there is no allocation in this case. This mechanism will be discussed in more detail in section 7.4, where we explain its implementation using pointers. 4.7. Other approaches Only smallstep and bigstep operational semantics have been considered up to this point. Both cases involve the deﬁnition of a transition relation, which characterizes the transformation of states resulting from program execution (and from the evaluation of expressions occurring in the program, as the selected expression language permits sideeffects). In this section, we present two other classical approaches to the semantics of imperative languages. For the sake of simplicity, we suppose that the evaluation of expressions does not modify the states em of the memory, and we write eM E nv = v to express that the evaluation of e in a state (Env, Mem) returns the value v. We use the language Lang3 (without blocks and procedures), whose syntax is deﬁned in Table 4.3. 4.7.1. Denotational semantics 4.7.1.1. Partial functions associated with a program The function exec_pgm4, deﬁned in section 4.4.1, computes the state of the memory Mem obtained after executing a program c in a state (Env, Mem). However, as we said, exec_pgm4 does not correspond to a mathematical function as
Semantics of Imperative Features
119
computation of Mem does not necessarily terminate. We now wish to deﬁne this semantics by a true mathematical function: __ : (E × Lang3 ) → (M M) This is the purpose of denotational semantics. Some properties over programs and their executions are easier to prove within this framework, developed in detail in [STO 77]. Of course, this function __ and the operational semantics must be semantically equivalent. In formal terms: ∀Env ∈ E ∀c ∈ Lang3 cEnv (Mem) = Mem ⇔ Mem, c Env Mem ∀Mem, Mem ∈ M [4.13] Thus, cEnv is a partial function from M to M (noted M M) which, given a memory state Mem, computes – in a ﬁnite amount of time – the memory state Mem obtained after executing a program c in an environment Env. If the execution in state (Env, Mem) of the program c does not terminate, then the function cEnv is not deﬁned for Mem (we denote this case by cEnv (Mem) = undef). Hence, the denotational semantics deﬁnes a function _Env providing for every program c a function describing the memory state transformations, which occur when c is executed. This function _Env is partial and also compositional: function cEnv is obtained by composing the functions corresponding to the components of c. This is not the case for exec_pgm4, which deﬁnes the semantics of while e do c in a certain state by recursively using the semantics of the same instruction in a different state, the one obtained after one execution of the body of the loop. x := eEnv =
M em
→
em := v] if Env(x) = r ∈ R and v = eM E nv ∈ V undef otherwise
M em[r
skipEnv = id = c1 ; c2 Env = c2 Env ◦ c1 Env =
M em
M em
→
M em
→ c2 Env (c1 Env (Mem))
if e then c1 else c2 Env = cond(eEnv , c1 Env , c2 Env ) cond : (M IB) × (M ⎧ M) × (M M) → (M M) ⎨ f1 (Mem) if fb (Mem) = 1 cond(fb , f1 , f2 ) = Mem → f2 (Mem) if fb (Mem) = 0 ⎩ undef if fb (Mem) = undef n while e do cEnv = lfp(F ) = {F (⊥)  n ≥ 0} F : (M M) → (M M) ⊥ : Mem → undef F : f → cond(eEnv , f ◦ cEnv , id) Box 4.7. Denotational semantics
120
Concepts and Semantics of Programming Languages 1
The deﬁnition of cEnv is given in Box 4.7. The assignment instruction x := e is associated with the function that transforms a memory state by storing the result of evaluating the expression e (if this does not raise an error) at the location denoted by x (if x denotes a reference). The function associated with the instruction skip is the identity function. For the sequence c1 ; c2 , we compose the two functions associated with c1 and c2 ; moreover, if one of these is applied to a state for which it is undeﬁned, then the result of applying the composition is also undeﬁned. The function associated with if e then c1 else c2 uses the function cond(fb , f1 , f2 ) which, given a memory state, selects which function to apply to this state (f1 or f2 ) according to the result of the boolean function (fb ) in this state. Finally, consider the instruction while e do c. Using the bigstep operational semantics formalism, it is easy to show (see exercise 4.1) that: while e do c ≡ if e then (c ; while e do c) else skip and thus, using the deﬁnitions given above, the function associated with the loop while e do c could be: while e do cEnv = cond(eEnv , while e do cEnv ◦ cEnv , id) However, such a deﬁnition using while e do cEnv is not compositional. One solution to this problem is to abstract the function while e do cEnv to get the function: F : f → cond(eEnv , f ◦ cEnv , id) This function F is known as a functional as its argument is itself a function. Its body uses only components of while e do c, F is therefore compositional. Furthermore, it veriﬁes: F (while e do cEnv ) = while e do cEnv In other words, while e do cEnv is a ﬁxpoint of F . To adopt such an approach, we must establish the existence of at least one ﬁxpoint for F and, when several ﬁxpoints exist, determine which to use. For example, consider the program while ↓ x = 0 do skip, which may be associated with the functional F : f → cond(↓ x = 0Env , f, id) as skipEnv = id and f ◦ id = f . According to the deﬁnition of cond, we therefore have: ⎧ ⎛ ⎞ ⎨ f (Mem) if Mem(Env(x)) = 0 and Env(x) ∈ R if Mem(Env(x)) = 0 and Env(x) ∈ R ⎠ F : f → ⎝Mem → Mem ⎩ undef else Given a function h : M M, we deﬁne the following function gh : M M: ⎧ ⎨ h(Mem) if Mem(Env(x)) = 0 and Env(x) ∈ R if Mem(Env(x)) = 0 and Env(x) ∈ R gh : Mem → Mem ⎩ undef else
Semantics of Imperative Features
121
According to the deﬁnition of F , we therefore have: ⎧ ⎨ gh (Mem) if Mem(Env(x)) = 0 and Env(x) ∈ R if Mem(Env(x)) = 0 and Env(x) ∈ R F (gh ) : Mem → Mem ⎩ undef else and from the deﬁnition of gh we obtain: ⎧ ⎨ h(Mem) if Mem(Env(x)) = 0 and Env(x) ∈ R if Mem(Env(x)) = 0 and Env(x) ∈ R F (gh ) : Mem → Mem ⎩ undef else Hence F (gh ) = gh and so for every function h : M M, gh is a ﬁxpoint of F . This example shows that there exist several ﬁxpoints of the functional F . We now need to determine which of these ﬁxpoints characterize exactly the same states as the operational semantics (see property [4.13]). For a memory state Mem such that Mem(Env(x)) = 2, the body of the loop is not executed, and the memory is not modiﬁed. For a memory state Mem such that Mem(Env(x)) = 0, the body of the loop is executed an inﬁnite number of times and the program does not terminate, so we wish that while ↓ x = 0 do skipEnv (Mem) = undef. From the function k : M em → undef, we deﬁne gk : M em if M em( E nv(x)) = 0 and E nv(x) ∈ R gk : Mem → undef else We see that gk is a ﬁxpoint of F , and this is the ﬁxpoint that we wish to associate to while ↓ x = 0 do skipEnv . For any other ﬁxpoint gh in F , if gk (Mem) = undef, then gk (Mem) = gh (Mem). As we will see, by deﬁning a partial order relation over partial functions from M to M, this ﬁxpoint corresponds to the least ﬁxpoint of the functional F , denoted as lfp(F ). The proofs of the existence and uniqueness of this ﬁxpoint, along with a procedure used to compute it, are presented in the following. 4.7.1.2. Existence, uniqueness and computation of the least ﬁxpoint of a functional 4.7.1.2.1. Partial order over partial functions To deﬁne the notion of the least ﬁxpoint of a functional, we introduce the order relation ! over the set noted [A B] of partial functions from A to B: f1 ! f2 ⇔ (∀x ∈ A f1 (x) = undef ⇒ f1 (x) = f2 (x)) Intuitively, f1 ! f2 if and only if f2 extends f1 , i.e. f2 keeps all the results of f1 and possibly provides some new ones. We can easily prove that this relation deﬁnes
122
Concepts and Semantics of Programming Languages 1
a partial order noted (!, [A B]): the relation is reﬂexive, antisymmetric and transitive. We recall that, given a partial order (D, !), the upper bounds and lower bounds of a subset Y ⊆ D are deﬁned by: M aj(Y ) = {d ∈ D  ∀y ∈ Y y ! d}
M in(Y ) = {d ∈ D  ∀y ∈ Y d ! y}
The supremum "Y of Y ⊆ D is the least upper bound of Y , if there exists: ∀d ∈ D (∀y ∈ Y y ! d) ⇒ "Y ! d The least element (minimum) of Y ⊆ D, if it exists, is the only element in M in(Y ) ∩ Y . The proposition below can be proved easily. P ROPOSITION 4.3.– ⊥ : Mem → undef is the least element of [M M]. Given a partial order (D, !) and a function f : D → D, an element d ∈ D is a ﬁxpoint of f if and only if f (d) = d. If it exists, the least ﬁxpoint of f , denoted as lfp(f ), is a ﬁxpoint of f such that: ∀d ∈ D (f (d) = d ⇒ lfp(f ) ! d) 4.7.1.2.2. Complete partial orders The least ﬁxpoint of a functional is obtained in an incremental way by constructing a sequence of increasingly precise approximations. The least ﬁxpoint is the least upper bound of this sequence of approximations. Complete partial orders provide a framework for establishing the existence of a least upper bound in these situations. Given a partial order (D, !), a subset Y ⊆ D is a chain if its elements are totally ordered: ∀y1 , y2 ∈ Y y1 ! y2 or y2 ! y1 A ccpo – chaincomplete partial order – is a partial order (D, !) such that, for any chain Y ⊆ D, "Y exists. If "Y exists for every subset Y ⊆ D, the partial order (D, !) is a complete lattice (and consequently a ccpo). L EMMA 4.2.– Every ccpo (D, !) has a least element ⊥ = "∅. P ROOF.– ∅ is clearly a chain, and since (D, !) is a ccpo, "∅ exists. Since "∅ ∈ D, to show that "∅ is the least element in D, it sufﬁces to prove that it is a lower bound of D. Using the deﬁnition of "∅, we have: ∀d ∈ D (∀x ∈ ∅ x ! d) ⇒ "∅ ! d
Semantics of Imperative Features
123
and since ∀x ∈ ∅ x ! d is always true, we have ∀d ∈ D " ∅ ! d and thus "∅ is a lower bound of D. The following lemma, which is an immediate consequence of deﬁnitions, will be used as follows. L EMMA 4.3.– Given a ccpo (D, !) and a chain Y ⊆ D, we have: "(Y ∪ {⊥}) = "Y The following proposition characterizes the least upper bound of a chain of partial functions from M to M. P ROPOSITION 4.4.– ([M M], !) is a ccpo such that for any chain Y : f (Mem) if there exists f ∈ Y such that f (Mem) = undef "Y : Mem → undef else P ROOF.– Let Y be a chain in ([M M], !). First, let us prove that "Y deﬁnes a function. If a function f ∈ Y exists such that f (Mem) = undef, then for any function f ∈ Y such that f (Mem) = undef, we have f (Mem) = f (Mem). Since Y is a chain, by deﬁnition, either f ! f and thus f (Mem) = f (Mem) since f (Mem) = undef, or f ! f and since f (Mem) = undef we have f (Mem) = f (Mem). Now, let us show that "Y deﬁnes the least upper bound of Y . Given f ∈ Y and Mem ∈ M, if f (Mem) = undef, then by construction we have ("Y )(Mem) = f (Mem) and f ! "Y . So "Y deﬁnes an upper bound of Y . Now we show that this is the least upper bound of Y : let g be an upper bound of Y , we have to prove that "Y ! g. Take Mem ∈ M such that ("Y )(Mem) = undef; by deﬁnition, there exists f ∈ Y such that f (Mem) = ("Y )(Mem) and since f ! g (since g is an upper bound of Y ), we obtain ("Y )(Mem) = f (Mem) = g(Mem) and thus the conclusion. 4.7.1.2.3. Continuous functions The computation of the least ﬁxpoint of a functional F is based on the continuity of this function. Given two partial orders (D1 , !1 ) and (D2 , !2 ), a function f : D1 → D2 is monotonic when it preserves the order, i.e. if and only if: ∀x, y ∈ D1
x !1 y ⇒ f (x) !2 f (y)
The lemma below is easy to prove since ! is a transitive relation. L EMMA 4.4.– Let (D1 , !1 ), (D2 , !2 ) and (D3 , !3 ) be three partial orders. If the functions f : D1 → D2 and f : D2 → D3 are monotonic, then f ◦ f : D1 → D3 is monotonic.
124
Concepts and Semantics of Programming Languages 1
The following lemma establishes a property over the least upper bounds of a chain and of its image by a monotonic function. L EMMA 4.5.– Let (D1 , !1 ), (D2 , !2 ) be two ccpo and f : D1 → D2 a monotonic function. If Y is a chain of (D1 , !1 ), then f (Y ) is a chain of (D2 , !2 ) such that "2 f (Y ) !2 f ("1 Y ). P ROOF.– If Y = ∅, then f (Y ) = ∅ is a chain of (D2 , !2 ) and since (D2 , !2 ) is a ccpo, following lemma 4.2, "2 ∅ is the least element in D2 and thus "2 ∅ !2 f ("1 ∅). If Y = ∅, then if x2 , y2 ∈ f (Y ), by deﬁnition, there exists x1 , y1 ∈ Y such that x2 = f (x1 ) and y2 = f (y1 ). Since Y is a chain of (D1 , !1 ), either x1 !1 y1 and so x2 = f (x1 ) !2 f (y1 ) = y2 , or y1 !1 x1 and so y2 = f (y1 ) !2 f (x1 ) = x2 (as f is monotonic). f (Y ) is therefore a chain of (D2 , !2 ). Let us prove that "2 f (Y ) !2 f ("1 Y ). By deﬁnition, ∀y ∈ Y y !1 "1 Y , and since f is monotonic, ∀y ∈ Y f (y) !2 f ("1 Y ), hence f ("1 Y ) is an upper bound of f (Y ). We know that "2 f (Y ) is the least upper bound of Y and so "2 f (Y ) !2 f ("1 Y ). Given two partial orders (D1 , !1 ) and (D2 , !2 ), a function f : D1 → D2 is said to be continuous if and only if it is monotonic and keeps the least upper bounds of the chains, i.e. if and only if f is monotonic and "2 f (Y ) = f ("1 Y ) for any nonempty chain Y in (D1 , !1 ). The function f is said to be strict if the property is also veriﬁed for an empty chain: "2 f (∅) = "2 ∅ = ⊥2 = f (⊥1 ) = f ("1 ∅) Once again, the composition of two continuous functions is itself a continuous function. L EMMA 4.6.– Take three ccpo (D1 , !1 ), (D2 , !2 ) and (D3 , !3 ). If f : D1 → D2 and f : D2 → D3 are continuous, then f ◦ f : D1 → D3 is continuous. P ROOF.– Functions f and f are continuous, so we can show that f ◦ f is monotonic, as stated in lemma 4.4. Let Y be a nonempty chain in (D1 , !1 ). Since f is continuous, by deﬁnition, "2 f (Y ) = f ("1 Y ) and f is monotonic. According to lemma 4.5, f (Y ) is a nonempty chain in (D2 , !2 ) and since f is continuous, we obtain "3 f (f (Y )) = f ("2 f (Y )); since "2 f (Y ) = f ("1 Y ), we obtain "3 (f ◦ f )(Y ) = f (f ("1 Y )) and thus "3 (f ◦ f )(Y ) = (f ◦ f )("1 Y ), which proves that f ◦ f is continuous. It is now possible to establish that the following functional is continuous: F : f → cond(eEnv , f ◦ cEnv , id) To do this, we consider the functional as a composition of two functions, whose continuity is proved as follows:
Semantics of Imperative Features
125
P ROPOSITION 4.5.– Take two partial functions fb : M IB and f0 : M M. Function F : f → cond(fb , f, f0 ) is continuous. P ROOF.– First, let us prove that the function F is monotonic. Let f1 and f2 be two functions in [M M] such that f1 ! f2 , and let us show that F (f1 ) ! F (f2 ). Assume Mem ∈ M such that (F (f1 ))(Mem) = undef; we show that (F (f1 )) (Mem) = (F (f2 ))(Mem). If fb (Mem) = 1, then: (F (f1 ))(Mem) = f1 (Mem)
and
(F (f2 ))(Mem) = f2 (Mem)
Since f1 ! f2 and f1 (Mem) = (F (f1 ))(Mem) = undef, we have f1 (Mem) = f2 (Mem) and therefore (F (f1 ))(Mem) = (F (f2 ))(Mem), establishing that F (f1 ) ! F (f2 ). If fb (Mem) = 0, then (F (f1 ))(Mem) = f0 (Mem) and (F (f2 )) (Mem) = f0 (Mem), hence (F (f1 ))(Mem) = (F (f2 ))(Mem), which allows us to establish that F (f1 ) ! F (f2 ). If fb (Mem) = undef, then (F (f1 ))(Mem) = undef, which is contradictory. Let us show that for any nonempty chain Y in ([M "F (Y ) = F ("Y ).
M], !),
Since we know that F is monotonic, by lemma 4.5, we get "F (Y ) ! F ("Y ). It remains to prove that F ("Y ) ! "F (Y ), i.e. F ("Y ) ! "{F (f )  f ∈ Y }. Take M em1 ∈ M. Let us prove that if (F ("Y ))( M em1 ) = M em2 = undef, then ("F (Y )) = Mem2 . If fb (Mem1 ) = 1, then (F ("Y )) (Mem1 ) = ("Y ) (Mem1 ) = Mem2 and, by proposition 4.4, there exists f ∈ Y such that f (Mem1 ) = Mem2 . By deﬁnition, for all f ∈ Y , we have F (f ) ! "F (Y ) and we therefore obtain ("F (Y ))(Mem1 ) = Mem2 . Otherwise, if fb (Mem1 ) = 0, then (F ("Y ))(Mem1 ) = f0 (Mem1 ) = Mem2 and for any function f ∈ Y : (F (f ))(Mem1 ) = f0 (Mem1 ) = Mem2 = undef Once again, by deﬁnition, for all f ∈ Y , we have F (f ) ! "F (Y ) and we thus obtain ("F (Y ))(Mem1 ) = Mem2 . Finally, if fb (Mem1 ) = undef, then (F ("Y )) (Mem1 ) = undef, which is contradictory. P ROPOSITION 4.6.– Take f0 : M M. The function F : f → f ◦ f0 is continuous. P ROOF.– First, let us prove that F is monotonic. Let f1 and f2 be two functions of [M M] such that f1 ! f2 , and let us prove that F (f1 ) ! F (f2 ). Take Mem ∈ M such that (F (f1 ))(Mem) = undef: we can show that (F (f1 ))(Mem) = (F (f2 ))(Mem). By deﬁnition, (F (f1 ))(Mem) = f1 (f0 (Mem)) = undef, hence
126
Concepts and Semantics of Programming Languages 1
f0 (Mem) = undef and since f1 ! f2 , we obtain f1 (f0 (Mem)) = f2 (f0 (Mem)), allowing us to conclude. Now, we must prove that F is continuous: (F ("Y ))(Mem1 ) = ("Y )(f0 (Mem1 )) = Mem2 ⇔ ∃f ∈ Y f (f0 (Mem1 )) = Mem2 (see proposition 4.4) ⇔ ∃f ∈ Y (F (f ))(Mem1 ) = Mem2 ⇔ ("F (Y ))(Mem1 ) = Mem2 (see proposition 4.4)
P ROPOSITION 4.7.– The functional F : f → cond(eEnv , f ◦ cEnv , id) is a continuous function. P ROOF.– Taking F1 : f → cond(eEnv , f, id) and F2 : f → f ◦ cEnv , we have F (f ) = F1 (F2 (f )) = F1 (f ◦ cEnv ) = cond(eEnv , f ◦ cEnv , id) hence F = F1 ◦ F2 . According to proposition 4.5, function F1 is continuous, and according to proposition 4.6, function F2 is continuous. Lemma 4.6 thus gives the conclusion. 4.7.1.2.4. Least ﬁxpoint of a continuous function over a ccpo Now let us show how the least ﬁxpoint of the functional F can be computed incrementally. This construction is based on the theorem as follows. T HEOREM 4.2.– Let (D, !) be a ccpo and f : D → D a continuous function: lfp(f ) = "({f n (⊥)  n ≥ 0}) P ROOF.– First, let us prove the existence of "({f n (⊥)  n ≥ 0}). It is easily done by a proof by induction over n: ∀n ∈ IN ∀m ∈ IN m ≥ n ⇒ f n (⊥) ! f m (⊥) (since, according to lemma 4.2, ⊥ is the least element of D and f is continuous, and thus monotonic). {f n (⊥)  n ≥ 0} is therefore a chain, and since (D, !) is a ccpo, "({f n (⊥)  n ≥ 0}) exists. Now, we can show that "({f n (⊥)  n ≥ 0}) is a ﬁxpoint of f : f ("({f n (⊥)  n ≥ 0})) = "(f ({f n (⊥)  n ≥ 0})) as f is continuous = "({f n (⊥)  n ≥ 1}) = "({f n (⊥)  n ≥ 1} ∪ {⊥}) = "({f n (⊥)  n ≥ 0}) It remains to prove that "({f n (⊥)  n ≥ 0}) is the least ﬁxpoint of f . Let d ∈ D be a ﬁxpoint of f . According to lemma 4.2, ⊥ ! d and thus for any natural number n, we have f n (⊥) ! f n (d) (which can be proved by induction over n since f is continuous and thus monotonic). Furthermore, we get ∀n ∈ IN f n (d) = d, by induction over n since d is a ﬁxpoint of f . d is therefore an upper bound of
Semantics of Imperative Features
127
{f n (⊥)  n ≥ 0} and since "({f n (⊥)  n ≥ 0}) is the least upper bound of {f n (⊥)  n ≥ 0}, we obtain "({f n (⊥)  n ≥ 0}) ! d. 4.7.1.2.5. Example Consider the following program, which computes ↓ y =↓ x! where ↓ x is a strictly positive integer: y := 1; while not(↓ x = 1) do (y :=↓ y× ↓ x; x :=↓ x − 1) c1 c4 c5 e c3 c2 c
Let us compute the result of executing this program in a state (Env0 , Mem0 ) such that M em0 ( E nv0 (x)) = 5 and E nv0 (y) ∈ R. In what follows, we suppose that E nv(x) and E nv(y) are elements of R noted rx and ry respectively. By deﬁnition, the functions associated with each of the components of this program are as follows: c1 Env : Mem → Mem[ry := 1] eEnv : Mem →
1 if Mem(rx ) = 1 0 if Mem(rx ) = 1
c3 Env = c5 Env ◦ c4 Env c3 Env : Mem → c5 Env (Mem[ry := Mem(ry ) × Mem(rx )]) = Mem[ry := Mem(ry ) × Mem(rx )][rx := Mem(rx ) − 1] c2 Env = lfp(F ) where F : f → cond(e E nv , f ◦ c3 E nv , id) f c3 Env (Mem) if Mem(rx ) = 1 F (f ) : Mem → M em if Mem(rx ) = 1 The function c3 Env describes the state obtained after executing the body of the loop. When the body of the loop is executed twice, function c3 2Env = c3 ; c3 Env = c3 Env ◦ c3 Env is deﬁned by: ⎞ ⎛ M em [ry := M em(ry ) × M em(rx )] ⎟ ⎜ [rx := Mem(rx ) − 1] ⎟ M em → ⎜ ⎝ [ry := Mem(ry ) × Mem(rx ) × (Mem(rx ) − 1)] ⎠ [rx := Mem(rx ) − 1 − 1] M em [ry := M em(ry ) × M em(rx ) × ( M em(rx ) − 1)] = [rx := Mem(rx ) − 2]
128
Concepts and Semantics of Programming Languages 1
More generally, when the body of the loop is executed n times, the function c3 nEnv is deﬁned by: ⎞ ⎛ ry := Mem(ry ) × Mem(rx ) × (Mem(rx ) − 1) M em ⎠ × · · · × (Mem(rx ) − (n − 1)) M em → ⎝ [rx := Mem(rx ) − n] Thus, in a state such that Mem(rx ) = p, we have: M em [ry := M em(ry ) × p × (p − 1) × · · · × 2] p−1 c3 Env (Mem) = [rx := Mem(rx ) − 1] = Mem[ry := Mem(ry ) × p!][rx := Mem(rx − 1] Using theorem 4.2, we can compute lfp(F ) = "({F n (⊥)  n ≥ 0}) from F (⊥) = ⊥ : Mem → undef. By deﬁnition, we have: undef if Mem(rx ) = 1 1 F (⊥) = F (⊥) : Mem → M em if M em(rx ) = 1 0
Similarly:
F 2 (⊥) = F (F 1 (⊥)) : Mem →
F 1 (⊥)(c3 Env (Mem)) if Mem(rx ) = 1 M em if Mem(rx ) = 1
and since c3 Env (Mem)(rx ) = Mem(rx ) − 1, we have: ⎧ if Mem(rx ) = 1 and Mem(rx ) = 2 ⎨ undef F 2 (⊥) : Mem → c3 Env (Mem) if Mem(rx ) = 2 ⎩ M em if Mem(rx ) = 1 With a further iteration, we obtain: F (⊥) = F (F (⊥)) : Mem → 3
hence:
2
F 2 (⊥)(c3 Env (Mem)) if Mem(rx ) = 1 M em if Mem(rx ) = 1
⎧ undef ⎪ ⎪ ⎪ ⎪ ⎨ F 3 (⊥) : Mem → c3 2Env (Mem) ⎪ ⎪ ⎪ c3 Env (Mem) ⎪ ⎩ M em
By induction, we obtain: ⎧ undef ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎨ c3 n−1 E nv ( M em) n F (⊥) : Mem → c3 n−2 E nv ( M em) ⎪ ⎪ · · · ⎪ ⎪ ⎪ ⎪ c 1 (Mem) ⎪ ⎪ ⎩ 3 0Env c3 Env (Mem)
if Mem(rx ) = 1 and Mem(rx ) = 2 and Mem(rx ) = 3 if Mem(rx ) = 3 if Mem(rx ) = 2 if Mem(rx ) = 1 if Mem(rx ) = 1 and Mem(rx ) = 2 and · · · and Mem(rx ) = n if Mem(rx ) = n if Mem(rx ) = n − 1 if Mem(rx ) = 2 if Mem(rx ) = 1
Semantics of Imperative Features
129
which can be written in a simpler form: undef if Mem(rx ) < 1 or Mem(rx ) > n n F (⊥) : Mem → ( M em) if Mem(rx ) = p and 1 ≤ p ≤ n c3 p−1 E nv and the function c2 Env = lfp(F ) = "n≥0 {F n (⊥)} is thus deﬁned by: undef if Mem(rx ) < 1 c2 Env : Mem → ( M em) if Mem(rx ) = p and p ≥ 1 c3 p−1 E nv Using the expression of c3 p−1 E nv established earlier, we ﬁnally obtain: ⎧ ⎨ undef if Mem(rx ) < 1 c2 Env : Mem → Mem[ry := Mem(ry ) × p!][rx := Mem(rx ) − 1] ⎩ if Mem(rx ) = p and p ≥ 1 It is now possible to compute cEnv0 (Mem0 ) as follows: cEnv0 (Mem0 ) = c2 Env0 (c1 Env0 (Mem0 )) = c2 Env0 (Mem0 [ry := 1]) = Mem0 [ry := 5!][rx := 1] At the end of execution of the program c started in the state (Env0 , Mem0 ), the value contained in the cell designated by y is thus Mem0 (Env0 (x))! = 5! = 120, and the value contained in the cell designated by x is 1. 4.7.2. Axiomatic semantics, Hoare logic The operational and denotational semantics of the language Lang3 presented in this chapter characterize the transformations of the state along the execution of a program. The state descriptions provided by these semantics are “complete” in the sense that they determine the values contained in every memory cell. Axiomatic semantics describes these states in terms of the properties they verify. These properties, which are often called assertions, may concern the contents of any location, but it is also possible to consider properties deﬁned only for a subset of the locations. Moreover, this axiomatic semantics is generally used to prove properties concerning the expected results of programs, rather than to describe their execution. It also provides a formalism for establishing the correctness of a program with regard to desired properties, i.e. with regard to a program speciﬁcation. This formalism, introduced in [HOA 83], is based on a speciﬁcation language used to create formal descriptions of tasks assigned to programs, and to express the relation between program input and output. The correctness property is expressed using a Hoare triple: {P1 }c{P2 }
130
Concepts and Semantics of Programming Languages 1
where P1 and P2 are expressions given in some language of speciﬁcations and c is a program. Intuitively, this triple expresses the notion that if the memory state satisﬁes the assertion P1 (the precondition), then the execution of program c will lead to a memory state satisfying assertion P2 (the postcondition). Axiomatic semantics characterizes Hoare triples that are “true” (for example the triple {↓ x = 2}x :=↓ x + 1{↓ x = 3} is “true”, while the triple {↓ x = 2}x :=↓ x + 1{↓ x = 5} is “false”). The logical formalism used to express the assertions may correspond to a fragment of ﬁrstorder logic (with deduction methods), or to an extensional approach where an assertion P is considered to be a predicate over em states such that P (( Env, Mem)) is satisﬁed if and only if P M E nv = 1. In this case, assertions are assimilated to the boolean expressions used in the language. We shall take this approach here. 4.7.2.1. Partial correctness 4.7.2.1.1. Provable Hoare triples The inference system in Box 4.8 contains rules that characterize a set of provable Hoare triples expressing the partial correctness of c for precondition P1 and postcondition P2 . (Skp)
{P }skip{P } (Seq)
(Cond)
(Aﬀ)
{P [↓ x ← e]}x := e{P }
{P2 }c2 {P3 } {P1 }c1 {P2 } {P1 }c1 ; c2 {P3 }
{P1 and e}c1 {P2 } {P1 and not e}c2 {P2 } {P1 }if e then c1 else c2 {P2 }
(WP )
{PI and e}c{PI } {PI }while e do c{PI and not e}
(Cons)
{P1 }c{P2 } {P1 }c{P2 }
P1 ⇒ P1 P2 ⇒ P2
Box 4.8. Axiomatic semantics: partial correctness
Since the instruction skip does not modify the state of the memory, rule (Skp) simply indicates that the states before and after the execution of skip verify the same properties. Rule (Aﬀ) states that for a property to be veriﬁed for the contents of a memory cell designated by x after execution of x := e, it must be veriﬁed for e prior to the execution of the assignment. For example, we have: (Aﬀ)
{↓ x + 1 = 3}x :=↓ x + 1{↓ x = 3}
Semantics of Imperative Features
131
Note here that the precondition ↓ x + 1 = 3 (equivalent to ↓ x = 2) corresponds to the postcondition ↓ x = 3 where ↓ x (where x is the left member of the assignment) has been replaced by ↓ x + 1 (the right member of the assignment), noted (↓ x + 3) [↓ x ←↓ x+1]. In this rule, P [↓ x ← e] designates the assertion obtained by replacing ↓ x by e in P : the substitution mechanism depends on the speciﬁcation language used to express the assertions. Using the extensional approach described above, the assertion P [↓ x ← e] satisﬁes:
em M em M em P [↓ x ← e]M E nv = P E nv where M em = M em[ E nv(x) := e E nv ]
Rule (Seq) states that if the execution of c1 in a state that satisﬁes P1 leads to a state that satisﬁes P2 and if the execution of c2 in such a state leads to a state that satisﬁes P3 , then the execution of the sequence c1 ; c2 in a state that satisﬁes P1 leads to a state that satisﬁes P3 . For example, we have: (Aﬀ) (Aﬀ) {0 + x ≥ 0}a := 0{↓ a + x ≥ 0} {↓ a + x ≥ 0}b := x{↓ a+ ↓ b ≥ 0} (Seq) {0 + x ≥ 0}a := 0 ; b := x{↓ a+ ↓ b ≥ 0} In this example, we wish to write the precondition in a simpler form, replacing 0 + x ≥ 0 with x ≥ 0 and expressing the postcondition in the form ↓ a ≥ − ↓ b. This transformation is possible using rule (Cons): if every state which satisﬁes x ≥ 0 also satisﬁes 0 + x ≥ 0 and if the execution of the sequence a := 0 ; b := x from such a state leads to a state which satisﬁes ↓ a+ ↓ b ≥ 0, which must verify ↓ a ≥ − ↓ b, then {x ≥ 0}a := 0 ; b := x{↓ a ≥ − ↓ b} is a provable triple: .. . (Seq) {0 + x ≥ 0}a := 0 ; b := x{↓ a+ ↓ b ≥ 0} x ≥ 0 ⇒ 0 + x ≥ 0 (Cons) ↓ a+ ↓ b ≥ 0 ⇒↓ a ≥ − ↓ b {x ≥ 0}a := 0 ; b := x{↓ a ≥ − ↓ b} The transformation of the assertions of a Hoare triple done by this rule is based on logical implications, which must be established. With the extensional approach, em proving P1 ⇒ P2 requires us to show that for any state (Env, Mem), if P1 M E nv = 1, M em then P2 Env = 1. It is possible to only transform one of the two assertions in a Hoare triple, using one of the following two rules: (ConsG )
{P1 }c{P2 } (P1 ⇒ P1 ) {P1 }c{P2 }
(ConsD )
{P1 }c{P2 } (P ⇒ P2 ) {P1 }c{P2 } 2
These rules can be deduced directly from rule (Cons). Rule (Cond) expresses the postcondition of a conditional instruction as a function of the boolean expression
132
Concepts and Semantics of Programming Languages 1
found in this instruction and of the postconditions associated with the programs found in the two branches of the instruction. For example, we have: (Aﬀ) (Aﬀ) {y = 0}x := y{↓ x = 0} {not y = 0}x := 0{↓ x = 0} ⎫ ⎫ (ConsG ) ⎧ (ConsG ) ⎧ ⎨ true ⎬ ⎨ true ⎬ and x := y{↓ x = 0} and x := 0{↓ x = 0} ⎩ ⎭ ⎩ ⎭ y=0 not y = 0 (Cond) {true}if y = 0 then x := y else x := 0{↓ x = 0} Finally, considering a property PI , the rule (WP ) establishes that: – PI is veriﬁed at the beginning of the execution of a loop; – if both PI and the condition to enter the loop are veriﬁed, PI remains true after executing the body of the loop, then after the execution of the loop, PI holds and the condition to enter the loop is no longer satisﬁed. PI is known as a loop invariant and is used to express properties of the computation done by the loop. For example: (Aﬀ) (ConsG ) (WP )
{↓ x + 1 ≤ b}x :=↓ x + 1{↓ x ≤ b} P1
{↓ x ≤ b and ↓ x < b}x :=↓ x + 1{↓ x ≤ b}
(P1 ⇒ P1 )
P1
{↓ x ≤ b}while ↓ x < b do x :=↓ x + 1{↓ x ≤ b and not ↓ x < b} P2
(ConsD )
(P2 ⇒ P2 ) {↓ x ≤ b}while ↓ x < b do x :=↓ x + 1{↓ x = b} P2
4.7.2.1.2. Validity and completeness A program c is partially correct with regard to precondition P1 and postcondition P2 if and only if for every state that satisﬁes P1 and in which the execution of c terminates, the state obtained after execution satisﬁes P2 . Since the operational semantics presented in section 4.4.1 can be used to characterize terminating executions and the states resulting from these executions, we can establish that provable Hoare triples {P1 }c{P2 } (with the extensional approach used here) correspond exactly to triples expressing the partial correctness of c in relation to P1
Semantics of Imperative Features
133
and P2 . First, the inference system c can be shown to be valid (if a triple is provable, then the program is partially correct): {P1 }c{P2 } ⇒ ∀(Env, Mem) ∈ E × M
em P1 M E nv = 1 and Mem, c Env Mem
em ⇒ P2 M E nv = 1
Second, we can show that this system is complete (if a program is partially correct, then the corresponding Hoare triple is provable): em =1 P1 M M em E nv ⇒ P2 Env = 1 ∀(Env, Mem) ∈ E × M and Mem, c Env Mem ⇒ {P1 }c{P2 } Proofs of these two properties can be found in [NIE 07]. 4.7.2.2. Total correctness Partial correctness of a program relates to properties of the states obtained when the execution of the program terminates. For nonterminating executions, the properties that are obtained are not correct, even if the inference trees of the associated Hoare triples are ﬁnite. A typical example of this is the program while true do skip, for which it is possible to prove the two triples {true}while true do skip{true} and {true}while true do skip{f alse}: (Skp) {true}skip{true} (ConsG ) ((true and true) ⇒ true) {true and true}skip{true} (WP ) {true}while true do skip{true and not true} (ConsD ) (I) {true}while true do skip{true/f alse} since the expression true and not true is never veriﬁed, so I can correspond to each of the implications (true and not true) ⇒ true and (true and not true) ⇒ f alse. A program c is totally correct with respect to a precondition P1 and postcondition P2 if and only if for any state satisfying P1 , the execution of c in this state terminates in a state that satisﬁes P2 . To characterize provable triples expressing the total correctness of a program, we simply modify rule (WP ) (as nontermination results from the construction while e do c alone). The new rule (WT )
{PI and e and V = n}c{PI and V < n} (PI ⇒ V ≥ 0) {PI }while e do c{PI and not e}
introduces a loop variant V that must be positive when the invariant PI is veriﬁed, and must strictly decrease at each iteration of the loop. This guarantees that the loop
134
Concepts and Semantics of Programming Languages 1
will terminate, as no strictly decreasing inﬁnite sequence exists over the set of natural numbers IN. For example, we can prove the total correctness expressed by the following triple: {↓ x ≤ b}while ↓ x < b do x :=↓ x + 1{↓ x = b} by considering the variant V = b− ↓ x. The precondition guarantees that V ≥ 0 and V strictly decreases at each iteration of the loop (as ↓ x is strictly increasing). (Aﬀ) ⎧ ⎪ ⎪ ⎪ ⎨
⎫ ⎪ ⎪ ⎪ ⎬
↓ x + 1 ≤ b and x :=↓ x + 1{A5 } b− ↓ x − 1 < n ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎭ ⎩ A6 ⎫ (I3 ) (ConsG ) ⎧ ⎪ ⎪ ⎫ ⎪ ⎪ ⎧ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ↓x≤b ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ and ⎬ ⎨ ⎬ ⎨ ↓x≤b ⎪ ↓x 0}cd {x = (↓ a × y)+ ↓ b and ↓ b < y}
5 Types
The syntax of the language Exp2 , i.e. the rules deﬁning the correct form of the expressions that make up the language, was presented in Chapter 3. The execution semantics of this language, which explains how programs written in the language are executed, was described using a variety of approaches. However, when introducing expressions of the language (in particular, the applications, see section 3.1.1), then the evaluation function (see section 3.2) and then the interpretation of operators (see section 3.2.3), we noted that the syntax does not prevent us from writing expressions that cannot be evaluated, such as 3 + true or (1 2) or 1 = f alse. In other words, the execution semantics chosen for this language cannot assign a value to these expressions. The existing syntax also allows us to write the expression f = g where f and g are functions, raising substantive questions concerning the meaning of such an expression. In the following, the semantics of these syntactically correct expressions having no clear meaning is more deeply discussed. In the expression 3 + true, the semantics for the primitive + is the addition over Z. The semantics assigned to true is not an element in Z, and thus does not belong to the domain of deﬁnition of this addition. Hence, this expression cannot be evaluated. However, it would be possible to assign a semantics to true in Z by associating it with an integer, such as 1. This might be unwise, as true would then have two roles, denoting a truth value as well as the integer 1. Developers, compilers, interpreters and code reviewers would then need to determine the role of each occurrence of true in a program to assign it the right semantics (creating a signiﬁcant risk of confusion, and, thus, error). Coding true in source code by an integer is thus errorprone. For this reason, we introduced a set IB for assigning semantics to booleans. A function application must have a functional value as its left hand operand. Once the formal parameter has been replaced by the right operand, i.e. the argument, the
Concepts and Semantics of Programming Languages 1: A Semantical Approach with OCaml and Python, First Edition. Thérèse Hardin, Mathieu Jaume, François Pessaux and Véronique Viguié DonzeauGouge. © ISTE Ltd 2021. Published by ISTE Ltd and John Wiley & Sons, Inc.
138
Concepts and Semantics of Programming Languages 1
evaluation of the body of the functional value returns the value of this application. In the expression “(1 2)”’, the constant 1 is not a function, and therefore has no body; this application cannot be evaluated. In mathematics, an equality is a binary relation deﬁned over elements of a same set. As the semantics of Exp2 is deﬁned mathematically, Z and IB are two disjoint sets, and their elements cannot be compared. For example, the semantics of the expression 1 = f alse is undeﬁned. Some languages get rid of this link with mathematics. For example, in Python there is a deﬁned equality relation between integers in booleans. For example, 1 = f alse is evaluated as f alse, while 1 = true is evaluated as true but 42 = true is evaluated as f alse. This holds true for other relational operators, so, for example, true < 2 is evaluated as true in Python. There are several possible semantics for an equality between functions such as f = g. It is best to avoid testing equalities of this kind in programs, as the results are often unpredictable. If the domains or codomains of f and g are different, they clearly cannot be equal. Their types will be said to be “incompatible” (a notion which will be discussed later). However, equalities between two functions with the same domain and codomain are often used in mathematics, for example, in the following b statement: “If two functions f and g over reals verify f = g, then a f (t)dt = b g(t)dt”. To use this theorem, we ﬁrst need to prove that f = g. The evaluation of a this expression along program execution requires an algorithm, either written in the source code or imported from a library, which is able to prove that, for any value x of the domain of f and g, the equality f (x) = g(x) is true. At the beginning of the 20th century, logicians proved that there can be no general algorithm for determining whether any two functions are “identical” in a ﬁnite time. The problem of equality of functions is said to be undecidable. If the domain is ﬁnite, then it is still possible to ﬁnd an adapted algorithm, but the time required to execute it depends on the number of elements in the domain. In Exp2 , no semantics is assigned to equalities between functions as the meaning of these equalities heavily depends on the domain of the functions. Proving the equality of two functions over the set of booleans is easy. On the other hand, proving the equality of two functions over the set of integers (restricted to integers representable in the memory) may take some time. Nevertheless, for a variety of reasons, the designers of certain programming languages have chosen to allow function equality tests. The semantics assigned to an expression f = g are consequently rather ad hoc. One version involves the syntactic equality of deﬁnitions: x → x + 2 is equal to itself, but is not equal to x → 1 + x + 1, nor to x → 2 + x or to y → y + 2. Another option focuses on the equality of the locations of the codes for f and g in the memory after the compilation. We encourage readers who use a language that tests function equality to consult the relevant reference manual concerning the meaning assigned to f = g. However, we do not advise using expressions of this kind.
Types
139
5.1. Type checking: when and how? In the examples presented at the start of this chapter, we referred to the notion of a mathematical set while explaining why certain expressions should be rejected. For reasons issued from mathematical logical studies, this notion of “set” is replaced by the notion of type in programming languages studies. In a typed language, all expressions have a type, as do the functions themselves. For example, the addition of integers is a function, of which the parameters are two integers and the result is an integer. The addition of ﬂoating point numbers (called ﬂoats) is a function, of which the parameters and result are ﬂoats. However, the two additions cannot use the same algorithm as the internal representations of integers and ﬂoating points in memory are totally different: the structure of their encodings is different and the bits do not have the same meaning. To evaluate the expression 3 + 4.5, for example, we must ﬁrst determine which algorithm to use for the addition. Some strongly typed languages reject this expression as it is “illtyped”. Some others will accept it but 3 will be considered as a ﬂoating point value. Such conversions often remain implicit; at best, they can lead to programs, which are difﬁcult to understand, and at worst, they can lead to execution errors, which can be hard to detect. These implicit conversions are hard to take into account in typing rules. Typechecking a program consists of verifying that the construction of all its expressions respect the rules of wellformedness regarding types, which are called the typing rules. If the typing rules are well designed, a welltyped program can be shown to be consistent, that is, any execution of this program, if it does not loop, provides a result of the type given by the typechecking. In this case, the language in question is said to be strongly typed. If a type system does not guarantee this property, it is said to be weakly typed. This is the case, for example, when a system permits explicit conversions between types: programmers can force the compiler to consider that the type of an expression is arbitrarily different from the one declared or computed by the compiler itself. Typechecking signiﬁcantly reduces the risk of writing and executing incorrect programs. The type language and the typing rules of the expressions of a language deﬁne the typing semantics of this language, which is formally studied in this chapter. 5.1.1. When to verify types? There are two possible typechecking mechanisms. The ﬁrst consists of verifying types during the execution of the program and determining whether execution can safely go on or whether a runtime error will occur. This semantics, known as dynamic typing, interacts with the dynamic execution semantics of the language. When an expression is recognized as being correctly typed, its evaluation returns a
140
Concepts and Semantics of Programming Languages 1
value that cannot be a typing error; otherwise, an error will be raised (this mechanism was ﬁrst presented in Chapter 3 in relation to the ErrT error). In this case, the semantic rules must include typing rules, and must account for errors resulting from type incompatibilities. The second mechanism consists of carrying out typechecking prior to execution: this is known as static typing. A program can only be executed if its typechecking did not detect type incompatibilities. Typing error detection therefore occurs earlier than in the case of dynamic typing, and the execution is guaranteed to be free of typing errors. The execution semantics only needs to deal with welltyped programs, making it easier to handle. 5.1.2. How to verify types? Once a static or dynamic typing mechanism has been selected, there are two options for veriﬁcation. Some programming languages ask programmers to declare the types of their variables explicitly. In C, for example, the function to compute n!, presented in example 3.3, is written as: C
int fact (int n) { if (n type_t
The type algebra is implemented in OCaml using a sum type. Variables are differentiated using names (character strings). We also deﬁne a function new_ty_var used to create a new type variable whose name differs from all alreadycreated ones. This unicity in terms of names is obtained using a counter, which is local to the function; this counter is incremented each time a variable is created, and its value supplies the name of the variable. As in the case of previous implementations, classes are used in Python to encode the sum types available in OCaml. Python
class T_int: def __init__ (self): pass
class T_var: def __init__ (self, vname): self.vname = vname
class T_bool: def __init__ (self): pass
var_cpt = 0 def new_ty_var (): global var_cpt var_cpt = var_cpt + 1 return T_var ("t" + str (var_cpt))
class T_fun: def __init__ (self, t1, t2): self.t1 = t1 self.t2 = t2
Each class is deﬁned to represent a constructor of the type algebra. While OCaml allows the creation of a persistent reference local to a function, in Python the counter used to assign a unique name to each new variable must be deﬁned as a global variable. Obviously, only the function new_ty_var will use this counter. For example, the ﬁrst application of this function will return the string "t0", the second will return "t1", etc.
Types
145
Instead of giving typing rules using natural language, we introduce a mathematical deﬁnition of the typing relation Env e : τ , which expresses that, in a typing environment Env, the expression e is of type τ . To type an expression, we need a typing environment that associates a type scheme to each identiﬁer, as we saw, in section 5.2.4, that a single expression may have several types issued from the same type scheme. In this case, each type is derived from a “skeleton” type scheme: this is the deﬁnition of polymorphism. D EFINITION 5.1.– Type scheme: the syntax of a type scheme is deﬁned by the following grammar: σ ::= ∀α1 . . . αn . τ where τ is the body of the scheme and α1 , . . . , αn are the generalized variables. The order in which universally quantiﬁed variables αi appear in a scheme is not important. It is assumed that their names are all different. For example, ∀α1 α2 . α1 → α2 and ∀α2 α1 . α1 → α2 denote the same type scheme. Furthermore, two type schemes differing only by quantiﬁed variable names are considered to be identical (these schemes may also be said to be equivalent modulo αconversion). For example, ∀α1 α2 . α1 → α2 and ∀α3 α4 . α3 → α4 denote the same scheme. A typing environment is represented by a list of bindings: E nv
= [(x1 , σ1 ), · · · , (xn , σn )]
where each identiﬁer xi is bound to its type scheme σi . OCaml
type scheme_t = (var_name_t list * type_t)
A type scheme is encoded in OCaml as a pair, the ﬁrst component being the list of generalized variables and the second component being the type representing the body of the scheme. Python does not permit type deﬁnitions, so an encoding in this language cannot be given. However, we use the same data structure in Python and OCaml (i.e. a pair in which the ﬁrst component is a list). The representation of typing environments is similar to the one used for execution environments. The values bound to identiﬁers are simply replaced by type schemes. OCaml
type ty_env_t = (string * scheme_t) list exception Unbound_var of string let   val
rec find_id_ty idname = function [] > raise (Unbound_var idname) h :: q > if fst h = idname then snd h else find_id_ty idname q find_id_ty : string > (string * ’a) list > ’a
146
Concepts and Semantics of Programming Languages 1
Python
class Unbound_var (Exception): def __init__ (self, id): self.id = id
def find_id_ty (idname, env) : for (n, ty) in env : if n == idname : return ty raise Unbound_var (idname)
The typing environment is deﬁned as an association list between identiﬁers and their type schemes. The function find_id_ty searches the ﬁrst occurrence of its ﬁrst argument (an identiﬁer) in its second argument (the list encoding the current environment). It starts from the head of the list and raises a speciﬁc error if this argument is not found when reaching the end of this list. This representation is similar to the one used in execution environments; the values bound to the identiﬁers are simply replaced by type schemes. 5.3.2. Generalization, substitution and instantiation To construct an environment, identiﬁers must be bound to type schemes, and so types must be transformed into schemes. Two steps are needed for this transformation. However, there are a few cases where no generalization should be done, so the type is converted to a socalled trivial type scheme. D EFINITION 5.2.– Trivial type scheme: a trivial scheme is a scheme with the form ∀∅. τ . OCaml
let trivial_scheme ty = (([], ty) : scheme_t) val trivial_scheme : type_t > scheme_t
Without the explicit type annotation scheme_t put on trivial_scheme, OCaml would have associated the “more general” type, ’a > (’b list * ’a), with trivial_scheme. The annotation leads to a more precise type, a point that can help the developer. Python
def trivial_scheme (ty): return ([], ty)
Conversely, generalizing variables appearing in a type expression needs a type generalization operation, which returns a polymorphic type scheme. However, only certain variables should be generalized: – variables that are not found anywhere in the environment; – variables that are generalized in the type schemes present in the environment; these variables can be renamed without changing the denotation of the scheme and are said to be bound variables.
Types
147
In other words, variables that are present in the environment but not bound by a ∀ are not generalized: these variables are called the free variables of the environment in question. A free variable denotes a yet unknown type and all the free occurrences of this variable in the environment denote the same type. D EFINITION 5.3.– Free variables of a type scheme: the set F(σ) of free variables of the type scheme σ contains the type variables of σ, that is, those which are not generalized. This set is deﬁned by: F(∀α1 . . . αn . τ ) = F(τ ) \ {α1 , . . . , αn } where F(τ ) is deﬁned by: F(int) = F(bool) = ∅ F(α) = {α} F(τ1 → τ2 ) = F(τ1 ) ∪ F(τ2 ) This deﬁnition can be extended to typing environments: F(σ) F(Env) = (x,σ)∈ E nv
D EFINITION 5.4.– Generalization: the generalization of a type τ in an environment E nv is the type scheme deﬁned by: Gen (τ, Env) = ∀α1 . . . αn . τ with {α1 . . . αn } = F(τ ) \ F(Env) E XAMPLE 5.1.– With the environment: E nv
= [(x, ∀α1 . α1 → α1 ), (y, ∀α1 . α1 → α2 ), (z, ∀∅. α3 ), (w, ∀∅. int)]
we have F(Env) = {α2 , α3 }, and: Gen ((α1 → α2 ) → (α3 → α4 ), Env) = ∀α1 α4 . (α1 → α2 ) → (α3 → α4 ) OCaml
let find_vars_in_type in_ty = let rec rec_find accu = function  T_int  T_bool > accu  T_fun (t1, t2) > let new_accu = rec_find accu t1 in rec_find new_accu t2  T_var v_name > if not (List.mem v_name accu) then v_name :: accu else accu in rec_find [] in_ty val find_vars : type_t > var_name_t list
148
Concepts and Semantics of Programming Languages 1
Python
def find_vars_in_type (in_ty): def rec_find (accu, ty): if isinstance (in_ty, T_int) or isinstance (in_ty, T_bool): return [] elif isinstance (in_ty, T_fun): return find_vars_in_type (in_ty.t1) + find_vars_in_type (in_ty.t2) elif isinstance (in_ty, T_var): if in_ty.vname in accu: return accu else: return [in_ty.vname] + accu else: raise ValueError return rec_find ([], in_ty)
The function used to ﬁnd free variables within a type operates recursively over the type structure. In the case of a type variable, we use the List.mem function from the standard library to verify whether an element belongs to a list; this prevents multiple insertions of the variable into the accumulator. OCaml
let rec find_vars_in_env = function  [] > []  (_, (gen_vars, body)) :: rem > let vars_of_body = find_vars_in_type body in let sch_free_vars = List.filter (fun var > not (List.mem var gen_vars)) vars_of_body in sch_free_vars @ (find_vars_in_env rem) val find_vars_in_env : (’a * (var_name_t list * type_t)) list > var_name_t list Python
def find_vars_in_env (env): accu = [] for (_, (gen_vars, body)) in env: vars_of_body = find_vars_in_type (body) sch_free_vars = \ list (filter (lambda var: not (var in gen_vars), vars_of_body)) accu = accu + sch_free_vars return accu OCaml
let generalize env ty = let vars_of_ty = find_vars_in_type ty in let vars_of_env = find_vars_in_env env in let gen_vars = List.filter (fun var > not (List.mem var vars_of_env)) vars_of_ty in (gen_vars, ty) val generalize : (’a * (’b * type_t)) list > type_t > (var_name_t list * type_t) Python
def generalize (env, ty) : vars_of_ty = find_vars_in_type (ty) vars_of_env = find_vars_in_env (env) gen_vars = list (filter (lambda var: not (var in vars_of_env), vars_of_ty)) return (gen_vars, ty)
Types
149
The generalization function proceeds in three stages. First, all variables present in the type to be generalized are fetched. Then the set of free variables in the environment is computed. The ﬁnal step computes the list of variables present in the type, excluding those found in the environment. In OCaml, this ﬁltering operation is carried out using the List.filter function from the standard library, which only retains those elements of a list that satisfy a given predicate. In Python, ﬁltering is carried out by deﬁning this predicate using the construct lambda. Conversely, to type an expression which is an identiﬁer, we need to generate a type from the type scheme bound to this identiﬁer in the environment. This is done by the instantiation operation, which returns an instance of a scheme by “specializing” it. The formal deﬁnition of an instance uses the notion of substitution, presented as follows. D EFINITION 5.5.– Substitution: a substitution is a partial function from the set of type variables to the type algebra, which has a ﬁnite domain. [α1 ← τ1 ; . . . ; αn ← τn ] denotes the substitution θ. The domain of θ, denoted as dom(θ), is the set {α1 , . . . , αn }. The codomain of θ, denoted as codom(θ), is the set {τ1 , . . . , τn }. θ associates the type τi to the variable αi . For the sake of simplicity, in what follows, we suppose that the variables of dom(θ) are all different and that: dom(θ) ∩ F(τ ) = ∅ τ ∈ codom(θ)
Substitutions naturally extend to a partial function from types to types, still with a ﬁnite domain, in the following manner: θ(int) = int θ(bool) = bool θ(α) if α ∈ dom(θ) θ(α) = α else θ(τ1 → τ2 ) = θ(τ1 ) → θ(τ2 ) From now on, we shall consider the substitution θ and its extension θ to be one and the same, and both shall be denoted as θ. The hypotheses used in deﬁning a substitution guarantee that, for any type τ , θ(θ(τ )) = θ(τ ) (in this case, θ is said to be idempotent). OCaml
type subst_t = (var_name_t * type_t) list let empty_subst = ([] : subst_t) val empty_subst : subst_t let singleton var_name ty = ([(var_name, ty)] : subst_t) val singleton : var_name_t > type_t > subst_t
150
Concepts and Semantics of Programming Languages 1
Python
empty_subst = [] def singleton (var_name, ty) : return [(var_name, ty)]
A substitution is represented by an association list, i.e. a list of pairs in which the ﬁrst component represents the name of the type variable and the second component the type to replace the variable. If the association list contains only one element, then the substitution is said to be elementary. OCaml
let rec apply_one transf = function  T_fun (t1, t2) > T_fun (apply_one transf t1, apply_one transf t2)  T_var var_name > if (fst transf) = var_name then snd transf else T_var var_name  any > any val apply_one : (var_name_t * type_t) > type_t > type_t Python
def apply_one (transf, ty) : if isinstance (ty, T_int) or isinstance (ty, T_bool) : return ty elif isinstance (ty, T_fun) : return T_fun (apply_one (transf, ty.t1), apply_one (transf, ty.t2)) elif isinstance (ty, T_var) : (fst_transf, snd_transf) = transf if fst_transf == ty.vname : return snd_transf else : return ty else : raise ValueError
The ﬁrst argument of the function apply_one is a list encoding an elementary substitution, and the second argument is a type. This function returns an update of this type, where the variable involved in the elementary substitution is replaced by its associated type. OCaml
let rec apply ty = function  [] > ty  h :: q > apply (apply_one h ty) q val apply : type_t > (var_name_t * type_t) list > type_t Python
def apply (ty, subst) : ty_res = ty for transf in subst : ty_res = apply_one (transf, ty_res) return ty_res
The apply function applies a substitution to a type. In other terms, the update of this type is done by iteratively performing the elementary substitutions contained in the list representing the substitution. This way of doing is correct by the hypotheses established in deﬁnition 5.5.
Types
151
The notion of substitution is used to deﬁne instances of a type scheme. D EFINITION 5.6.– Instance: τ is said to be an instance of the type scheme σ = ∀α1 . . . αk . τ and we note it σ ≤ τ if there exists a substitution θ such that dom(θ) = {α1 , . . . , αk } and τ = θ(τ ). In this case, the type τ is said to be compatible with τ . 5.3.3. Typing rules and typing trees We can now give the rules used to ensure that an expression is welltyped. These rules are shown in Box 5.1, with the notation already used for the description of the bigstep operational semantics of Exp2 in section 3.3. We give a formal deﬁnition of the relation Env e : τ , which means that in a typing environment Env, the expression e is of type τ . (Ttrue )
E nv
(Tvar )
true : bool
(Tfalse )
E nv(x) ≤ τ E nv x : τ
(Tbinop )
E nv
(Tite )
(Tunop )
E nv
f alse : bool
E nv
E nv
(Tint )
E nv
i : int
op : τ1 → τ2 Env e : τ1 E nv op e : τ2
op : τ1 → τ2 → τ3 Env e1 : τ1 E nv e1 op e2 : τ3
E nv
e 2 : τ2
e1 : bool Env e2 : τ Env e3 : τ E nv if e1 then e2 else e3 : τ
(Tfun ) (Tlet )
E nv
e 1 : τ1
(x, ∀∅. τ1 ) ⊕ Env e : τ2 fun x → e : τ1 → τ2
E nv
σ = Gen (τ1 , Env) (x, σ) ⊕ Env e2 : τ2 let x = e1 in e2 : τ1 → τ2
E nv
(Trec )
(f, ∀ ∅. τ1 → τ2 ) ⊕ (x, ∀ ∅. τ1 ) ⊕ Env e : τ2 E nv rec f x = e : τ1 → τ2
(Tapp )
E nv
e1 : τ2 → τ1 Env e2 : τ2 E nv (e1 e2 ) : τ1
Box 5.1. Typing rules in Exp2
Rules Ttrue , Tfalse and Tint indicate that a boolean constant has the type bool and an integer constant has the type int. According to rule Tvar , the type of an identiﬁer is an instance of the scheme to which it is bound in the typing environment and Env(x) denotes the type scheme bound to x in Env. Rule Tunop states that the type of the application of a unary operator to an argument is that returned by the operator, as
152
Concepts and Semantics of Programming Languages 1
long as the argument is of the type required by this operator. Similarly, rule Tbinop indicates that the type of the application of a binary operator to two arguments is that returned by the operator, as long as the two arguments are of the types required by this operator. Rule Tite indicates that the type of a conditional expression is that shared by the two subexpressions making up its then and else branches, as long as the condition subexpression is of type bool. According to rule Tfun , a function of a variable x of which the body is an expression e is of type τ1 → τ2 if the type of e is τ2 when the type of x is taken to be τ1 . Note that the scheme bound to the argument x is trivial: the type variables that may appear in τ1 are not generalized. If these variables were generalized, it would be possible to type incorrect programs. For example, consider the expression fun x → if x then x + 1 else x + 2. Typing this expression must fail, as x is used both as a boolean (rule Tite ) and as an integer (rule Tbinop ). If the type of x was generalized to a scheme used to typecheck the body of the function, each occurrence of x would be typed using a new instance of the scheme: the ﬁrst instance would be compatible with booleans, while the second one would be compatible with integers. For this reason, type τ1 is not generalized, meaning that the same instance must be used for all occurrences of x in the body of the function. Rule Tlet indicates that the type of an expression let x = e1 in e2 is that of the subexpression e2 , under the hypothesis that x is bound to the scheme that generalizes the type of e1 . According to rule Trec , a recursive function f of a variable x, of which the body is an expression e, is of type τ1 → τ2 where τ2 is the type of e under the hypothesis that the type of x is τ1 and the hypothesis that the type scheme of f is compatible with τ1 → τ2 . Finally, rule Tapp states that, in order to apply an expression e1 to an argument e2 , e1 must be of the functional type τ1 → τ2 and e2 must be of the type τ1 . The type of the overall expression is that of the result of the function, i.e. τ2 . Note that rule Trec introduces a binding between the recursively deﬁned function and a trivial type scheme into the environment. Thus, the expression rec f x = let y = (f 7) in (f true) is rejected by the typechecking process. Indeed, f is used in the expression with both the type int → α and the type bool → α. Note that generalizing the type of f , i.e. assigning type β → α to f , would still have been correct in terms of the execution semantics used in evaluating expressions. The scheme of f must be limited to a trivial scheme in order to guarantee the decidability of type inference (i.e. the possibility of designing a type computation algorithm which is guaranteed to terminate).
Types
153
N OTE.– This restriction is also applicable to languages that permit the deﬁnition of mutually recursive functions, although if the mutually deﬁned functions do not perform actual recursive calls. Thus, the following program – which is evidently correct – will be rejected by the typing process: OCaml
let rec map f l = match l with [] > []  h :: q > (f h) :: (map f q) and map_succ l = map (fun x > x + 1) l and map_not l = map (fun x > not x) l ;; Error: This expression has type int but an expression was expected of type bool
This program redeﬁnes the function map that iterates a function f, given as an argument over a list l. Within this mutually recursive deﬁnition (introduced by let rec and and), it also deﬁnes a function that adds 1 to each element of a list of integers, and a function that complements each element of a list of booleans. This joint deﬁnition of the three functions is rejected by the OCaml compiler, as map is applied to a list of integers and to a list of booleans in the same deﬁnition. Typechecking of the subexpression map does not generalize its type. Generalization would have been done if map had been recursively deﬁned independently of its two uses. OCaml
let rec map f l = match l with [] > []  h :: q > (f h) :: (map f q) ;; val map : (’a > ’b) > ’a list > ’b list = OCaml
let and val val
map_succ l = map (fun x > x + 1) l map_not l = map (fun x > not x) l ;; map_succ : int list > int list = map_not : bool list > bool list =
To avoid such difﬁculties, OCaml provides the keyword rec to annotate the functions recognized as recursive by their author. Using this annotation avoids implicitly considering all functions as recursive ones (as is done in many languages without type inference), which would forbid the generalization of their types, hence cancelling any possibility of polymorphism. Note that there are other advantages to distinguishing between recursive and nonrecursive functions during compilation, including the capacity to generate more efﬁcient code. The different stages involved in computing the type of an expression can be described in the form of a typing tree, similar to the trees used to represent the evaluation of expressions in Chapter 3.
154
Concepts and Semantics of Programming Languages 1
E XAMPLE 5.2.– The typing tree of the expression let f = fun x → x in (f 5) is: β → β ≤ int → int α≤α E nv 5 : int (Tvar ) (Tvar ) (x, ∀∅. α) ⊕ Env x : α E nv f : int → int (Tfun ) (Tapp ) E nv fun x → x : α → α (f, ∀β. β → β) ⊕ Env (f 5) : int
E nv (Tlet ) E nv let f = fun x → x in (f 5) : int
5.4. Type inference algorithm in Exp2 The rules presented in section 5.3 are used to verify the type of a program including type annotations. They also deﬁne the typing system of the language and, as we shall see in section 5.5, guarantee certain safety properties of welltyped programs. In this section, we describe an algorithm that generates types from expressions in a program. By extension, this inference algorithm can also be used to check whether a program is welltyped. Obviously, the type inference algorithm must be correct with regard to the type system and the associated rules. The advantage of type inference lies in the fact that programmers do not have to provide explicit type annotation within programs. 5.4.1. Principal type The type inference algorithm is described formally by a set of rules. These rules establish the fact that, in a typing environment Env, an expression e is of type τ when a set of constraints described by a substitution θ over the types can be satisﬁed (for example, this set of constraints must not require a type variable α to be equal to both int and bool). This is noted Env e (τ, θ). Furthermore, in case of polymorphism, the type τ computed by this algorithm must be as general as possible: it is called the principal type of e. For example, the type of fun x → fun y → (x y) is τ0 = (α → β) → α → β, but it could also be: τ1 : (int → int) → int → int τ2 : (int → bool) → int → bool τ3 : (int → bool → int) → int → bool → int However, type τ0 is the most general, as it allows this expression to be used in the highest number of contexts. The other possible types all result from a substitution applied to τ0 . For example, types τ1 , τ2 and τ3 may be obtained in the following manner: τ1 = τ0 [α ← int, β ← int] τ2 = τ0 [α ← int, β ← bool] τ3 = τ0 [α ← int, β ← (bool → int)]
Types
155
Considering the principal type of an expression prevents some type conﬂicts. If the type inference algorithm does not use the principal type of an expression, it may reject expressions which are indeed welltyped. For this reason, any expression must have a principal type belonging to the type system (this is the case in the Exp2 language), otherwise inference will be impossible or users will be forced to provide type annotation for certain parts of their programs. 5.4.2. Sets of constraints and uniﬁcation The type inference algorithm establishes that an expression e is of type τ when a set of constraints can be satisﬁed, i.e. when a solution exists. A constraint introduces an equality between two types. For example, from the expression if (f 4) then 5 else (g true), we deduce the following set of constraints to be satisﬁed: – the type τf of f is functional: τf = α → β; – the argument of f is an integer: α = int; – as the result of f being used as a condition, it is a boolean: β = bool; – the type τg of g is functional: τg = γ → δ; – the argument of g is a boolean: γ = bool; – the result of g is of the same type as the expression 5, meaning that it is an integer: δ = int. If all these equality constraints can be satisﬁed together, then the expression is welltyped. This is true here if we take α = int, β = bool, γ = bool and δ = int. The nonconstrained variables are then generalizable (with certain conditions), and contribute to establishing the polymorphism of the type. Proving that a constraint τ1 = τ2 is satisﬁed requires the ability to replace certain variables in the left and/or right of the equals sign with types, which may themselves contain their own variables. In other words, a substitution θ must exist such that θ(τ1 ) = θ(τ2 ). Determining whether such a substitution exists is known as a type uniﬁcation problem. If this substitution exists, then τ1 and τ2 are said to be uniﬁable, and substitution θ is said to be an uniﬁer of terms τ1 and τ2 . Thus, two terms are uniﬁable if all or part of their variables can be instantiated by the same substitution in a way which makes them structurally equal. As several types may be assigned to an expression, several substitutions may exist to identify two terms (this is the reason for type plurality). Consider the two terms τ1 = α → β and τ2 = α → bool. The two substitutions θ1 = [α ← int; β ← bool] and θ2 = [β ← bool] are uniﬁers of τ1 and τ2 since θ1 (τ1 ) = θ1 (τ2 ) = int → bool and θ2 (τ1 ) = θ2 (τ2 ) = α → bool.
156
Concepts and Semantics of Programming Languages 1
A principal uniﬁer, also often called a most general uniﬁer, of τ1 and τ2 is a substitution μ such that, for any other uniﬁer θ, θ(τ1 ) is an instance of μ(τ1 ), i.e. θ(τ1 ) can be obtained by applying a substitution θ to μ(τ1 ). This is essentially the same as saying that the principal uniﬁer μ is a minimal substitution, and thus all other uniﬁers θ include this substitution and can be written as θ = θ ◦ μ. In the previous example, θ2 is a principal uniﬁer, whereas θ1 is not. In this case, we have θ1 = [α ← int] ◦ θ2 . The principal uniﬁer of two types τ1 and τ2 is computed by solving the system of equations {τ1 = τ2 }. The uniﬁcation algorithm Mgu presented in section 5.4.2 can be used to solve systems of this kind. It takes a set Q of equalities between types, called constraints, as an argument. The solution of a system Q, when it exists, is a substitution θ such that, for every equation τ1 = τ2 in Q, we have θ(τ1 ) = θ(τ2 ). This algorithm is recursive and computes the most general uniﬁer of system Q. The system Q[α ← τ ] denotes the system Q in which α is replaced by τ in every equation. If none of the listed cases apply, the operation fails, triggering the failure of the type inference process. Notably, if one member of an equality is a variable, the other cannot be a type containing this variable (unless it is itself a variable). This constraint is needed to avoid inﬁnite looping over problems of the form Mgu ({α = int → α}) for which there is no type of ﬁnite size which solves this equation. If α = int → α, we have α = int → int → α, α = int → int → int → α, and so on. Mgu (∅) = id Mgu ({α = α} ∪ Q) = Mgu (Q) Mgu ({α = τ } ∪ Q) = Mgu (Q[α ← τ ]) ◦ [α ← τ ] if α ∈ / F (τ ) Mgu ({τ = α} ∪ Q) = Mgu (Q[α ← τ ]) ◦ [α ← τ ] if α ∈ / F (τ ) Mgu ({int = int} ∪ Q) = Mgu (Q) Mgu ({bool = bool} ∪ Q) = Mgu (Q) Mgu ({τ1 → τ1 = τ2 → τ2 } ∪ Q) = Mgu ({τ1 = τ2 ; τ1 = τ2 } ∪ Q) Box 5.2. Type uniﬁcation algorithm for Exp2
E XAMPLE 5.3.– Let us compute the principal uniﬁer of τ1 = α1 → (α2 → int) and τ2 = int → (α3 → α1 ). Mgu ({α1 → (α2 → int) = int → (α3 → α1 )}) = Mgu ({α1 = int; α2 → int = α3 → α1 }) = Mgu ({α2 → int = α3 → int}) ◦ [α1 ← int] = Mgu ({α2 = α3 ; int = int}) ◦ [α1 ← int] = Mgu ({int = int}) ◦ [α2 ← α3 ] ◦ [α1 ← int] = Mgu (∅) ◦ [α2 ← α3 ] ◦ [α1 ← int] = id ◦ [α2 ← α3 ] ◦ [α1 ← int] = [α2 ← α3 ; α1 ← int] = θ
Types
We can verify that θ(τ1 ) = θ(τ2 ) = int → (α2 → int). OCaml
exception Cycle exception Incompatible_tys of (type_t * type_t) let    val
rec occur_check v_name = function T_int  T_bool > false T_fun (t1, t2) > (occur_check v_name t1)  (occur_check v_name t2) T_var n > v_name = n occur_check : var_name_t > type_t > bool
let rec unify t1 t2 = match (t1, t2) with  (T_int, T_int)  (T_bool, T_bool) > empty_subst  (T_fun (a, b), T_fun (c, d)) > let subst_left = unify a c in let b_substed = apply b subst_left in let d_substed = apply d subst_left in compose (unify b_substed d_substed) subst_left  (T_var v, other)  (other, T_var v) > if occur_check v other then raise Cycle ; singleton v other  _ > raise (Incompatible_tys (t1, t2)) val unify : type_t > type_t > subst_t
Python
class Cycle (Exception) : pass class Incompatible_tys (Exception) : def __init__ (self, ty1, ty2) : self.ty1 = ty1 self.ty2 = ty2 def occur_check (v_name, ty) : if isinstance (ty, T_int) or isinstance (ty, T_bool) : return False elif isinstance (ty, T_fun) : return occur_check (v_name, ty.t1) or occur_check (v_name, ty.t2) elif isinstance (ty, T_var) : return ty.vname == v_name else : raise ValueError Python
def unify (ty1, ty2) : if (isinstance (ty1, T_int) and isinstance (ty2, T_int)) or\ (isinstance (ty1, T_bool) and isinstance (ty2, T_bool)) : return empty_subst elif isinstance (ty1, T_fun) and isinstance (ty2, T_fun) : a, b, c, d = ty1.t1, ty1.t2, ty2.t1, ty2.t2 subst_left = unify (a, c) b_substed = apply (b, subst_left) d_substed = apply (d, subst_left) return compose (unify (b_substed, d_substed), subst_left)
157
158
Concepts and Semantics of Programming Languages 1
elif isinstance (ty1, T_var) or isinstance (ty2, T_var) : (other, v) = (ty1, ty2.vname) if isinstance (ty2, T_var) \ else (ty2, ty1.vname) if occur_check (v, other) : raise Cycle return singleton (v, other) else : raise Incompatible_tys (ty1, ty2)
The veriﬁcation “if α ∈ / F(τ )” is carried out using the function occur_check, which checks whether a variable appears in a type. As we saw from the rules, this function is used in the two cases when a variable must be uniﬁed with a type. The function unify works through the structure of both types, verifying equalities in the case of basic types, carrying out recursive calls for functional types and extending the substitution result in cases where one of the two types is a variable. The exception Incompatible_tys has two parameters: the two types that raise an uniﬁcation conﬂict. Such a possibility is useful for creating error messages, providing users with more details concerning the nature of the problem. For reasons of efﬁciency, the implementation given here does not operate over a set of constraints, but only over the two types to unify. The uniﬁcation function unify can now be tested using example 5.3, deﬁning the two types of this example using Python syntax. Python
type1 = T_fun ((T_var ("t1")), (T_fun ((T_var ("t2")), (T_int ())))) type2 = T_fun ((T_int ()), (T_fun ((T_var ("t3")), (T_var ("t1"))))) mgu = unify (type1, type2) print_subst (mgu) [(’t1’, ), (’t3’, )] (t1 ty  T_fun (t1, t2) > let t1_copied = copy t1 in let t2_copied = copy t2 in T_fun (t1_copied, t2_copied)  T_var v_name > (try List.assoc v_name fresh_assoc with Not_found > ty) in copy body val specialize : (var_name_t list * type_t) > type_t Python
class Not_found (Exception) : pass def list_assoc (key, l) : for (k, v) in l : if k == key : return v raise Not_found def specialize (sch) : (vars, body) = sch fresh_assoc = list (map (lambda var : (var, new_ty_var ()), vars)) def copy (ty) : if isinstance (ty, T_int) or isinstance (ty, T_bool) : return ty elif isinstance (ty, T_fun) : t1_copied = copy (ty.t1) t2_copied = copy (ty.t2) return T_fun (t1_copied, t2_copied)
160
Concepts and Semantics of Programming Languages 1
elif isinstance (ty, T_var) : try : return list_assoc (ty.vname, fresh_assoc) except Not_found : return ty else : raise ValueError return copy (body)
The function specialize, used to obtain an instance of a type scheme, creates an association list that binds each generalized variable of the scheme to a fresh variable. It then follows a recursive path, copying the structure of the body of the scheme and replacing each generalized variable with its corresponding fresh version (otherwise, the nongeneralized variable is retained). The implementation in OCaml uses the function List.assoc from the standard library, which searches through an association list for the value associated with a key, and, if none is found, the exception Not_found is raised. This function does not exist in Python, so we deﬁne it under the name list_assoc. As the environment contains type schemes, we also need to be able to apply a substitution to a scheme and to an environment. D EFINITION 5.8.– Substitution applied to a type scheme: take a type scheme σ = ∀α1 . . . αn . τ and a substitution θ such that: (dom(θ) ∪ F(θ(τ ))) ∩ {α1 , . . . , αn } = ∅ we deﬁne θ(σ) = ∀α1 . . . αn . θ(τ ). This deﬁnition notably prohibits substitution of the generalized variables of a type scheme. In the following implementation, we do not verify that the variables from the substitution domain and codomain are not present in the scheme. This choice has no effect on the correctness of the inference algorithm, which only establishes equality constraints between types. The types here never include generalized variables, as they are produced by the instantiation operation that replaces generalized variables with fresh variables. Thus, the constructed substitutions will never contain generalized variables. N OTE.– A more general deﬁnition of the application of a substitution θ to a type scheme σ, with no hypotheses concerning σ and θ, is possible. This is obtained by renaming the generalized variables α1 . . . αn of σ with variables found neither in the domain of θ nor in θ(τ ). Applying a substitution to an environment comes down to creating an environment in which the bodies of the schemes involved in the bindings have been transformed by the substitution.
Types
161
OCaml
let subst_env subst env = List.map (fun (name, (vars, body)) > let body’ = apply body subst in let sch’ = (vars, body’) in (name, sch’)) env val subst_env : subst_t > (’a * (’b * type_t)) list > (’a * (’b * type_t)) list Python
def subst_env (subst, env) : def do_binding (binding) : (name, sch) = binding (vars, body) = sch body_new = apply (body, subst) sch_new = (vars, body_new) return (name, sch_new)
5.4.3.2. Composition of substitutions The inference algorithm works through expressions recursively, determining equality constraints. At each step, it constructs a substitution and a type. The type is that of the current expression, and the substitution memorizes the equality constraints to be satisﬁed to obtain this type. The substitutions that are constructed for each subexpression must be combined. The composition of two substitutions is deﬁned as follows. Composing two substitutions θ1 and θ2 , represented by the lists θ1 = [(α1 , τ1 ), · · · , (αn , τn )]
θ2 = [(β1 , τ1 ), · · · , (βk , τk )]
comes down to computing the substitution θ1 ◦ θ2 , such that θ1 ◦ θ2 (x) = θ1 (θ2 (x)). We therefore apply θ1 to all the types contained in θ2 ; the only elements of substitution θ1 that are retained are the bindings for variables, which do not occur in the domain of θ2 . The substitution θ1 ◦ θ2 is thus represented by the list: [(β1 , θ1 (τ1 )), · · · , (βk , θ1 (τk )), (α1 , τ1 ), · · · , (αn , τn )] from which we have eliminated all pairs (βi , θ1 (τi )) such that βi = θ1 (τi ) (since, in this case, βi no longer belongs to the domain of θ1 ◦ θ2 ) and all pairs (αi , τi ) such that αi ∈ {β1 , · · · , βk } (these pairs are no longer useful, as the substitution θ2 is applied before the substitution θ1 ). E XAMPLE 5.4.– Given two substitutions: θ1 = [(α1 , τ1 ), (α2 , α4 )]
and
θ2 = [(α3 , α1 → τ2 ), (α1 , τ3 ), (α4 , α2 )]
162
Concepts and Semantics of Programming Languages 1
the composition θ1 ◦ θ2 is the substitution: θ1 ◦ θ2 = [(α1 , τ3 ), (α3 , τ1 → τ2 ), (α2 , α4 )] In this case, the binding (α1 , τ1 ) from θ1 is not retained since α1 belongs to the domain of θ2 and the binding (α4 , θ1 (α2 )) is not retained since α4 = θ1 (α2 ). OCaml
let compose subst1 subst2 = let subt_1_2 = (List.map (fun (name, ty) > (name, apply ty subst1)) subst2) @ (List.filter (fun (name, _) > not (List.mem_assoc name subst2)) subst1) in List.filter (fun (name, ty) > match ty with Types.T_var name’ > name name’  _ > true) subt_1_2 val compose : subst_t > subst_t > subst_t Python
def list_mem_assoc (key, l) : for (k, v) in l : if k == key : return True return False def compose (subst1, subst2) : def subst1_on_binding (binding) : (name, ty) = binding return (name, apply (ty, subst1)) def filter_not_in_subst2 (binding) : (name, _) = binding return not (utils.list_mem_assoc (name, subst2)) def filter_not_identity (binding) : (name, ty) = binding if isinstance (ty, T_var) : return name != ty.vname return True part1 = list (map (subst1_on_binding, subst2)) part2 = list (filter (filter_not_in_subst2, subst1)) return list (filter (filter_not_identity, part1 + part2))
The implementation uses the function List.mem_assoc from the standard OCaml library. The arguments of this function are a key and an association list, and it returns the boolean value stating whether or not the key occurs in the domain of the association list. This function does not exist in Python, so we deﬁne it under the name list_mem_assoc. 5.4.3.3. Type inference In order to limit the number of inference rules, we have chosen to consider applications of unary and binary operators as speciﬁc cases of the general application of a function. Usually (and with no loss of generality), the type of an operator is
Types
163
found in the initial environment of any program. The name of an operator plays exactly the same role as that of an identiﬁer declared by the programmer. In the rules presented in Box 5.3, id denotes the identity substitution. E XAMPLE 5.5.– The tree used to infer the type of expression e deﬁned by: let f = fun x → fun y → fun z → if x then y else z in (((f true) 2) 3) e4 e5 e3 e6 e2 e7 e1
is: (Ilet )
E nv
1
2
e (int, θ)
where 1 is the typing tree Env e1 (bool → (α3 → (α3 → α3 )), θ1 ) shown below, and obtained considering: = Mgu ({α1 = bool}) = Mgu (∅) ◦ [α1 ← bool] = id ◦ [α1 ← bool] = [α1 ← bool] E nv3 = μ1 (id( E nv3 )) = [(z, ∀∅. α3 ), (y, ∀∅. α2 ), (x, ∀∅. bool)] μ2 = Mgu ({id(α2 ) = α3 }) = Mgu (∅) ◦ [α2 ← α3 ] = [α2 ← α3 ] θ1 = μ2 ◦ id ◦ id ◦ μ1 ◦ id = μ2 ◦ μ1 = [α2 ← α3 ; α1 ← bool] μ1
(Itrue )
E nv
(Iint )
(Iite )
true (bool, id)
(If alse )
f alse (bool, id) x ∈ dom(Env)
E nv
(Ivar ) v (int, id) E nv x (Inst ( E nv(x)), id) E nv e1 (τ1 , θ1 ) μ1 = Mgu ({τ1 = bool}) μ1 (θ1 ( E nv)) e2 (τ2 , θ2 ) θ2 (μ1 (θ1 (Env))) e3 (τ3 , θ3 ) μ2 = Mgu ({θ3 (τ2 ) = τ3 }) E nv
if e1 then e2 else e3 (μ2 (τ3 ), μ2 ◦ θ3 ◦ θ2 ◦ μ1 ◦ θ1 ) E nv e1 (τ1 , θ1 ) θ1 ( E nv) e2 (τ2 , θ2 ) α fresh μ = Mgu ({θ2 (τ1 ) = τ2 → α}) (Iapp ) E nv (e1 e2 ) (μ(α), μ ◦ θ2 ◦ θ1 ) α fresh (x, ∀∅. α) ⊕ Env e (τ1 , θ1 ) (If un ) E nv fun f → e (θ1 (α) → τ1 , θ1 ) E nv e1 (τ1 , θ1 ) (x, Gen (τ1 , θ1 ( E nv))) ⊕ θ1 ( E nv) e2 (τ2 , θ2 ) E nv
(Ilet )
(Irec )
E nv let x = e1 in e2 (τ2 , θ2 ◦ θ1 ) α, β fresh (f, ∀∅. α → β) ⊕ (x, ∀∅. α) ⊕ Env e (τ1 , θ1 ) μ = Mgu ({θ1 (β) = τ1 }) E nv
rec f x = e (μ(θ1 (α → β)), μ ◦ θ1 )
Box 5.3. Type inference algorithm for Exp2
164
Concepts and Semantics of Programming Languages 1
The typing tree
1
is thus:
(Ivar ) (Ivar ) (Ivar ) E nv3 x (α1 , id) E nv3 y (α2 , id) E nv3 z (α3 , id) (Iite ) (z, ∀∅. α3 ) ⊕ Env2 e4 (α3 , θ1 ) E nv3 (If un ) (y, ∀∅. α2 ) ⊕ Env1 e3 (α3 → α3 , θ1 ) E nv2 (If un ) (x, ∀∅. α1 ) ⊕ Env e2 (α3 → (α3 → α3 ), θ1 ) E nv1 (If un ) E nv e1 (bool → (α3 → (α3 → α3 )), θ1 ) and where 2 is the typing tree: (Ivar ) E nv4 f (bool → (α4 → (α4 → α4 )), id) 5 (Iapp ) E nv4 e5 (α4 → (α4 → α4 ), μ3 ◦ id) 4 (Iapp ) E nv4 e6 (int → int, μ4 ◦ id ◦ μ3 ) 3 (Iapp ) E nv4 e7 (int, μ5 ◦ id ◦ μ4 ◦ id ◦ μ3 ) where Env4 = (x, ∀α3 . (bool → (α3 → (α3 → α3 )))) ⊕ Env and the typing of e5 is obtained considering: μ3 = Mgu ({id(bool → (α4 → (α4 → α4 ))) = bool → α5 }) = Mgu ({bool = bool; α4 → (α4 → α4 ) = α5 }) = Mgu ({α4 → (α4 → α4 ) = α5 }) = Mgu ({∅}) ◦ [α5 ← α4 → (α4 → α4 )] = [α5 ← α4 → (α4 → α4 )] μ3 (α5 ) = α4 → (α4 → α4 ) to type e6 : μ3 (Env4 ) = Env4 μ4 = Mgu ({id(α4 → (α4 → α4 )) = int → α6 }) = Mgu ({α4 = int; α4 → α4 = α6 }) = Mgu ({int → int = α6 }) ◦ [α4 ← int] = Mgu (∅) ◦ [α6 ← int → int] ◦ [α4 ← int] = [α6 ← int → int; α4 ← int] μ4 (α6 ) = int → int to type e7 : μ4 (Env4 ) = Env4 μ5 = Mgu ({id(int → int) = int → α7 }) = Mgu ({int = int; int = α7 }) = Mgu ({int = α7 }) = Mgu (∅) ◦ [α7 ← int] = [α7 ← int] μ5 (α7 ) = int
Types
165
θ = μ5 ◦ id ◦ μ4 ◦ id ◦ μ3 = μ5 ◦ μ4 ◦ μ3 = [α7 ← int; α6 ← int → int; α4 ← int; α5 ← int → (int → int)] 3, 4 and 5 are, respectively, the following trees: (Iint )
E nv4
3 (int, id) (Itrue )
OCaml
let    
E nv4
(Iint )
E nv4
2 (int, id)
true (bool, id)
rec type_expr env = function Cste2 (CInt2 _) > (T_int, empty_subst) Cste2 (CBool2 _) > (T_bool, empty_subst) Var2 v > (specialize (find_id_ty v env), empty_subst) Fun2 (arg_name, body) > let arg_ty = new_ty_var () in let (body_ty, body_subst) = type_expr ((arg_name, (trivial_scheme arg_ty)) :: env) body in (T_fun ((apply arg_ty body_subst, body_ty)), body_subst)  Letin2 (v_name, e1, e2) > let (e1_ty, e1_subst) = type_expr env e1 in let new_env = (v_name, (generalize env e1_ty)) :: env in let new_env2 = subst_env e1_subst new_env in let (e2_ty, e2_subst) = type_expr new_env2 e2 in (e2_ty, compose e2_subst e1_subst)  App2 (e1, e2) > let (e1_ty, e1_subst) = type_expr env e1 in let (e2_ty, e2_subst) = type_expr (subst_env e1_subst env) e2 in let fun_res_ty = new_ty_var () in let e1_ty_substed = (apply e1_ty e2_subst) in let mu = unify (T_fun (e2_ty, fun_res_ty)) e1_ty_substed in ((apply fun_res_ty mu), (compose mu (compose e2_subst e1_subst)))  If2 (cond_e, then_e, else_e) > ( let (cond_ty, cond_subst) = type_expr env cond_e in let mu1 = unify cond_ty T_bool in let env_then = subst_env mu1 (subst_env cond_subst env) in let (then_ty, then_subst) = type_expr env_then then_e in let env_else = subst_env then_subst env_then in let (else_ty, else_subst) = type_expr env_else else_e in let mu2 = unify (apply then_ty else_subst) else_ty in ((apply else_ty mu2), (compose mu2 (compose else_subst (compose then_subst (compose mu1 cond_subst))))) )
166
Concepts and Semantics of Programming Languages 1
 Rec2 (f_name, arg_name, body) > let arg_ty = new_ty_var () in let fun_res_ty = new_ty_var () in let fun_ty = T_fun (arg_ty, fun_res_ty) in let new_env = (f_name, (trivial_scheme fun_ty)) :: (arg_name, (trivial_scheme arg_ty)) :: env in let (body_ty, body_subst) = type_expr new_env body in let mu = unify (apply fun_res_ty body_subst) body_ty in ((apply (apply fun_ty body_subst) mu), (compose mu body_subst)) val type_expr : (string * scheme_t) list > exp2 > (type_t * subst_t) Python
def type_expr (env, exp): if isinstance (exp, Cste2) : if isinstance (exp.cst, CInt2) : return (T_int (), empty) elif isinstance (exp.cst, CBool2) : return (T_bool (), empty) else : raise AstValError elif isinstance (exp, Var2) : return (specialize (find_id_ty (exp.name, env)), empty) elif isinstance (exp, Fun2) : arg_ty = new_ty_var () arg_name = exp.id_arg body = exp.body (body_ty, body_subst) = \ type_expr ([(arg_name, trivial_scheme (arg_ty))] + env, \ body) return (T_fun (apply (arg_ty, body_subst), body_ty), \ body_subst) elif isinstance (exp, Letin2) : (e1_ty, e1_subst) = type_expr (env, exp.exp1) new_env = [(exp.name, generalize (env, e1_ty))] + env new_env2 = subst_env (e1_subst, new_env) (e2_ty, e2_subst) = type_expr (new_env2, exp.exp2) return (e2_ty, compose (e2_subst, e1_subst)) elif isinstance (exp, App2) : (e1_ty, e1_subst) = type_expr (env, exp.exp1) (e2_ty, e2_subst) = \ type_expr (subst_env (e1_subst, env), exp.exp2) fun_res_ty = new_ty_var () e1_ty_substed = apply (e1_ty, e2_subst) mu = unify.unify (T_fun (e2_ty, fun_res_ty), e1_ty_substed) return \ (apply (fun_res_ty, mu),\ compose (mu, compose (e2_subst, e1_subst))) elif isinstance (exp, If2) : (cond_ty, cond_subst) = type_expr (env, exp.cond_e) mu1 = unify.unify (cond_ty, T_bool ()) env_then = subst_env (mu1, subst_env (cond_subst, env)) (then_ty, then_subst) = type_expr (env_then, exp.then_e) env_else = subst_env (then_subst, env_then) (else_ty, else_subst) = type_expr (env_else, exp.else_e)
Types
167
mu2 = unify.unify (apply (then_ty, else_subst), else_ty) return \ (apply (else_ty, mu2), \ compose \ (mu2,\ compose \ (else_subst,\ compose \ (then_subst, compose (mu1, cond_subst))))) elif isinstance (exp, Rec2) : arg_ty = new_ty_var () fun_res_ty = new_ty_var () fun_ty = T_fun (arg_ty, fun_res_ty) new_env = \ [(f_name, trivial_scheme (fun_ty))] + \ [(arg_name, trivial_scheme (arg_ty))] + env (body_ty, body_subst) = type_expr (new_env, body) mu = unify.unify (apply (fun_res_ty, body_subst), body_ty) return \ (apply (apply (fun_ty, body_subst), mu),\ compose (mu, body_subst)) else: raise AstValError
The function type_expr is deﬁned by case over the structure of the expressions. It returns a pair in which the ﬁrst component is the type of the typed expression and the second component is the substitution resulting from the typing. As we explained earlier, the cases of unary and binary operators will not be addressed separately, as these are simply speciﬁc cases of the general application App2.
5.5. Properties 5.5.1. Properties of typechecking We can establish several properties of the typing system set out in Box 5.1 which guarantee its correctness with regard to the semantics of the language. The most important of these properties is set out here, although we shall not provide a demonstration; the relevant proof can be found in [DAM 82]. P ROPOSITION 5.1.– Strong safety: let e be an expression, τ a type, Envt a typing environment and Env an evaluation environment such that ∀x ∈ X, Envt Env(x) : E nvt (x). If E nvt e : τ and E nv e v, then v is a value of type τ . 5.5.2. Properties of the inference algorithm Just as typing rules need to satisfy a certain number of properties, the algorithm itself must meet four requirements: termination, correctness, completeness and
168
Concepts and Semantics of Programming Languages 1
principality. Termination guarantees that the type of an expression will be inferred in a ﬁnite period of time (the compiler will not enter into an inﬁnite loop). The algorithm is correct if, for a given expression e and a typing environment Env, it returns a type which is correct with respect to the typing rules set out in Box 5.1. This correctness means that the inference algorithm will not produce wrong results. Completeness means that if a type exists for a given expression e and a typing environment Env, then it will be found by the algorithm: in other words, the algorithm will not “miss” a solution. Principality guarantees that the algorithm will always compute the most general type, in the sense presented in section 5.4. P ROPOSITION 5.2.– Correctness of the inference algorithm: if Env e (τ, θ), then θ(Env) e : τ . P ROPOSITION 5.3.– Completeness and principality of the inference: let Env be an environment and e an expression. If there exists a substitution θ and a type τ such that θ (Env) e : τ , then Env e (τ, θ) and there exists a substitution θ such that τ = θ (τ ) and θ = θ ◦ θ. 5.6. Typechecking of imperative constructs To introduce typechecking of imperative languages, we shall deﬁne rules for the language Lang3 introduced in Chapter 4, extended using the block structure (see section 4.6.1). The type system given here is not an extension of the rules shown in Box 5.1, as we are not considering functional features, only mutable variables. We thus need to type the language of deﬁnitions (given in Table 4.1), the language of expressions (given in Table 4.2) and the instructions (described in Table 4.3). For reasons of simplicity, the same symbol will be used to denote all three typing relations. 5.6.1. Type algebra As references belong to the set of values, they must be assigned a type, which is called ref. To handle, in a consistent manner, the value stored at the address denoted by a reference, the type ref must also indicate the type of the referenced value. The kernel of Lang3 contains no function nor polymorphism and only integer and boolean constants; the type algebra is thus very simple. This algebra, presented in Table 5.2, contains an integer type int, a boolean type bool and a type of references on a value of type τ , denoted as τ ref. As polymorphism is not considered, the typing environment does not bind identiﬁers to type schemes, but simply to types.
Types
169
τ := int Integer  bool Boolean  τ ref Reference Table 5.2. Type algebra of Lang3
5.6.2. Typing rules Let us begin by setting out the typing rules for expressions in the language. As in the case of the rules presented in Box 5.1, the type relation Env e : τ establishes that, in a typing environment Env, an expression e is of type τ .
(Tint )
k∈Z E nv
k : int (Tunop )
(Tbinop )
k ∈ IB
(Tbool )
E nv
op : τ1 → τ2 → τ3 E nv
(Tref )
E nv E nv
k : bool
op : τ1 → τ2 E nv
e:τ
↑ e : τ ref
(Tvar )
E nv
E nv(x) E nv
=τ
x:τ
e : τ1
op e : τ2 E nv
e 1 : τ1
E nv
e 2 : τ2
e1 op e2 : τ3 (Tderef )
E nv
e : τ ref
E nv
↓e:τ
Box 5.4. Typing rules for expressions in Lang3
Now, let us consider the typing of variable declarations and blocks. A variable declaration d may be used to extend the typing environment by binding the declared identiﬁer to the type of its deﬁning expression. Thus, the relation Env d : Env expresses the fact that, in an environment Env, typing the declaration d will create, from Env, a new environment noted Env . We recall that a block is written begin d c end, where d is a (possibly empty) sequence of declarations (see section 4.6.1) and c an instruction. We thus need rules for typing a sequence of declarations. Let d d denote a sequence beginning with a declaration d, followed by the sequence d. The empty sequence is noted ∅. We have chosen to consider that an instruction c has no type as it has no value. Nevertheless, given a typing environment Env, the typing algorithm must check that the expressions involved in an instruction are correctly typed. This typing relation is noted Env c : W T (WT denotes welltyped). Note that in certain languages, such as OCaml or certain extensions of C (such as those available with the GCC compiler),
170
Concepts and Semantics of Programming Languages 1
instructions are expressions, meaning that they have values and, consequently, types. The nature of these types depends on choices inherent to the language.
(Tlet ) (Tedseq )
E nv
E nv E nv
let x = e; : (x, τ ) ⊕ Env (Tdseq )
∅ : Env
e:τ
E nv
d : Env1 E nv
E nv1
d : Env2
d d : Env2
Box 5.5. Typing rules for declarations in Lang3
(Tassign ) (Tseq )
E nv
E nv(x)
= τ ref
E nv
c1 : W T E nv
E nv
e:τ
E nv
c2 : W T
(Tloop )
c1 ; c2 : W T
(Tcond )
E nv
(Tskip )
x := e : W T
e : bool E nv
(Tblock )
E nv
E nv
E nv
c1 : W T
E nv
skip : W T
e : bool
E nv
E nv
c : WT
while e do c : W T
E nv
c2 : W T
if e then c1 else c2 : W T
d : Env1 E nv
E nv1
c : WT
begin d c end : W T
Box 5.6. Typing rules for instructions in Lang3
E XAMPLE 5.6.– Consider the GCD computation program presented in section 4.1 and deﬁned by: begin let x = 45;let y = 5; while not(↓ x =↓ y) do if ↓ x ≤↓ y then y :=↓ y− ↓ x else x :=↓ x− ↓ y e2 i2 i3 e1 end
i1
Its typechecking tree is: (Tint ) (Tint ) E nv 45 : int E nv1 5 : int (Tlet ) (Tlet ) E nv let x = 45; : E nv1 E nv1 let y = 5; : E nv2 (Tdeseq ) E nv let x = 45;let y = 5; : E nv2 1 (Tblock ) E nv begin let x = 45;let y = 5; while e1 do i1 end : W T
Types
171
where Env1 = (x, int ref) ⊕ Env , Env2 = (y, int ref) ⊕ Env1 and 1 is the typing tree while e1 do i1 , written as: E nv2 = : int → int → bool E nv2
(Tloop ) 2
E nv2
and
4
3
(Tcond ) 5
E nv2 (y) = int ref (Tvar ) E nv2 y : int ref (Tderef ) E nv2 ↓ y : int
is the typing tree for if e2 then i2 else i3 : ≤ : int → int → bool
(Tbinop )
and
6
4
while e1 do i1 : W T
are the typing trees for ↓ x and ↓ y:
E nv2
and
E nv2
E nv2 (x) = int ref (Tvar ) E nv2 x : int ref (Tderef ) E nv2 ↓ x : int
and
3
↓ x =↓ y : bool E nv2 not(↓ x =↓ y) : bool
(Tunop )
and
2
not : bool → bool(Tbinop )
E nv2 E nv2
2
3
e2 : bool
5
6
if e2 then i2 else i3 : W T
are the typing trees for instructions i2 and i3 : E nv2
E nv2 (y)
(Tassign )
E nv2 (x)
(Tassign )
= int ref (Tbinop )
− : int → int → int
3
2
↓ y− ↓ x : int E nv2 i2 : W T E nv2 − : int → int → int
= int ref (Tbinop )
E nv2
2
3
↓ x− ↓ y : int E nv2 i3 : W T E nv2
5.6.3. Typing polymorphic deﬁnitions In a language that accepts polymorphism, the introduction of references creates an additional difﬁculty. Consider the deﬁnition let x =↑ [ ]; where [ ] is the value of an empty list. Suppose that the type of this list is polymorphic, and thus of the form α list, where α is the type of the elements. If we assign the type scheme ∀α. (α list) ref to x, then the rules of polymorphism can be used, for example, in: x := [1] ; x := [true] ; 1 + head (↓ x) ;
172
Concepts and Semantics of Programming Languages 1
where the function head (supposed to exist) returns the ﬁrst element of a (nonempty) list. In this program, as the type of x is polymorphic, it is possible to assign x with any given list. The ﬁnal instruction in the program is welltyped: x is of type (α list) ref, so ↓ x is of type α list, and so head returns a value of type α, the latter being instantiated in int in the addition expression. However, in practice, x contains a boolean value, so execution will necessarily raise an error, as addition is only deﬁned for integers. Conversely, consider the following program: let f = fun x > ↑ x ;
which creates a reference upon an input value. The type of the function f is α → α ref. The generalization of α is not problematic. Function f can be applied to a value of any type t, and will return a reference of type t ref, which is not polymorphic. One simple solution to this problem is to only generalize certain variables. The types of expressions that correspond to constants, variables or functions are generalized, and those of applications of functions or operators are not generalized. In the example above, in the declaration let x =↑ [ ]; the expression ↑ [ ] is an application of the operator ↑, hence the type α list ref should not be generalized. This solution relies on a syntactic criterion alone. Other methods have been developed (e.g. [LER 91, PIE 02]) to relax this restriction and increase the number of safe generalizations, but these lie outside the scope of this book. 5.7. Subtyping and overloading Many languages offer extensions to the notion of type, designed to make the type algorithm accept expressions which it would otherwise reject. This increases ﬂexibility when writing code, but makes it harder to review: in addition to typing rules, users must also master the rules of subtyping and overloading. Subtyping makes it possible to relax the constraint concerning the type t1 of the argument of a function, which may then be applied to an expression of type t2 , as long as the values of type t2 share at least the characteristics of the values of type t1 . Overloading enables the same name f to be used to denote different values (usually functions), leaving the typing algorithm to determine which binding to use for each occurrence of f in an expression. The aim, in this case, is to relax the constraint brought by the typing discipline: an identiﬁer must have one, and only one, type. In this section, we provide a brief presentation of subtyping and overloading. These features will be discussed in greater detail in Volume 2, in Chapters 1 and 3, with examples provided in Chapters 2 and 4.
Types
173
5.7.1. Subtyping If the typing algorithm used by a language satisﬁes the partial correctness property, any welltyped program will be evaluated as a value of the expected type or will loop. Certain programs rejected by the typing process also respect this property of partial correctness, but the typechecking is unable to determine this. For this reason, the type language and the typing algorithm may be enriched in order to reduce program rejection, while maintaining the partial correctness property. Subtyping is a way of reﬁning the typing process using a comparison relation on types noted ts tb 1 and a typing algorithm that permits any expression e of type ts to be accepted as an expression of type tb . The comparison relation must be reﬂexive and transitive: in other words, this relation is a partial preorder (there is no need to compare any two types using ). We have already encountered such a preorder in section 5.4. The relation “ τ is an instance of the type scheme σ = ∀α1 . . . αk . τ ”, denoted as σ ≤ τ , is indeed a partial preorder over polymorphic types. For example, type schemes ∀∅. int and ∀∅. bool are not comparable, and ∀α. α → α is smaller than int → int. A preorder on types can be deﬁned in the following manner. We ﬁrst choose the family F of types, which we already know how to compare. Next, we extend (when necessary: the preorder does not need to be total) the preorder chosen for F to other types. Since type languages are often constructed by induction, extension is also usually carried out by induction. This gives some guarantees on the consistency between inductive construction and comparison. For example, consider ts tb and ss sb . If the type language includes product types, it seems consistent to extend this order by taking ts × ss tb × sb . Suppose now that the order on types is the order of inclusion of their sets of values: ts tb if the set of values of type ts is included in the set of values of type tb . For example, this would be the case if values of type tb represent individual people, and values of type ts represent individual employees. Suppose that ts tb and ss sb . consider a function f : tb → ss . In this case, f allows us to deﬁne a function g : ts → sb by simply restricting f to the subset of values of type ts . Furthermore, as g takes values in ss , it also takes values in sb , as shown in Figure 5.1. Function f is deﬁned over a domain that contains the domain of g and the codomain of f is included in that of g. Thus, anything which is treated by g is also treated by f , and any value returned by f is included in the values returned by g. From this intuitive interpretation, it is possible to prove that the subtyping relation deﬁned over the type family F can be extended by taking the following rule: if ts tb and ss sb , then (tb → ss ) (ts → sb ). 1 Read tsmall for ts and tbig for tb .
174
Concepts and Semantics of Programming Languages 1
tb
ss
sb
ts
f:
g: f tb ts
ss sb g
Figure 5.1. Covariance and contravariance. For a color version of this ﬁgure, see www.iste.co.uk/hardin/programming1.zip
A preorder over types that verify this rule is said to be contravariant over the type of the function arguments and covariant over the type of the results. It can be proved that a typing algorithm using such a preorder can be constructed from the two rules shown in Box 5.7 (and, potentially, from similar rules for certain other type constructors).
(T )
E nv
e : ts ts tb e : tb
E nv
(Tapp )
E nv
e1 : tb → t Env e1 : ts E nv (e1 e2 ) : t
ts tb
Box 5.7. Subtyping rules
Rule T deﬁnes the semantics of the order relation over types: any expression e of a type ts that is smaller than a type tb is also of type tb . Rule Tapp is a consequence of this semantics: any function that requires an expression of type tb as its argument will accept an argument with a type ts , which is a subtype of tb . This order is frequently used in the case described below using our example. Consider a function f , taking individuals as its parameter and returning, for example, their birth year (f is therefore of the type tb → int). f may be restricted to data of type ts , i.e. for individual employees. Noting this restriction g, the type of g is thus ts → int. We have (tb → int) (ts → int), ensuring the consistency of this restriction in terms of typing. Conventionally, the name f is retained for the restriction (instead of introducing a new name g). Programming language manuals often specify that any function of type t1 → t2 can be applied to any value of type t, if t is a subtype of t1 , and the type of the result t2 remains unchanged. This approach does not exploit the full potential of functional type comparison. We shall return to this point in Volume 2, Chapter 1.
Types
175
These rules extending the typing relation cannot be generalized indiscriminately to any operation; however, this does not decrease the utility of subtyping, as only a partial preorder is required. For example, the subtyping relation cannot be extended to assignment. Consider ts tb : in our example, ts is the type of individual employees and gives their name and salary, and tb is the type of all individuals, giving only their name. Now, suppose that the values of type ts are stored over a greater number of bytes than the values of type tb . This is the case in our example. The assignment of v : tb to x : ts ref will leave a number of “free” bytes in the zone pointed by x; evidently, this is not acceptable, unless the compiler identiﬁes the issue and handles it in a suitable manner. Now, let us examine how this “common sense” constraint is obtained through a detailed examination of extension to assignment. Consider an assignment operator := of type tb ref × tb → unit. If we extend subtyping to this operator by induction, then := will also be of the type ts ref × tb → unit. This means that a reference to a value of type ts will become a reference to a value of type tb . This raises the question of the effective meaning of the word “reference” (see Chapter 6). We do not intend to go into detail concerning subtyping and its properties here; our presentation is limited to the most commonly encountered notion. Other approaches exist, for example, using different orders induced over functional types, which may be covariant on the argument types. In all cases, complex theoretical studies are required to prove the partial correctness properties of these notions of subtyping, and, ﬁnally, to create effective and efﬁcient typing algorithms. A number of examples of subtyping will be studied in Volume 2, notably in Chapters 2 and 4. 5.7.2. Overloading Identiﬁer overloading is a syntactic facility proposed by certain programming languages. In this case, the same identiﬁer can be used for functions or operators with different deﬁnitions, and, consequently, different types and values. Overloading is often used to give the same name to operations, which have very similar speciﬁcations, but are applied to different data types. For example, the expressions 3 + 4, 3.5 + 4.6 and “aa” + “bb” all use +, but the code and type of the operations denoted by this identiﬁer are very different. The compiler is responsible for choosing the correct version of + to use in evaluating each of these expressions. This task is known as overloading resolution. In order to take account of overloading, the environment model must be adapted. Without overloading, adding a new binding for an identiﬁer f , which is already bound in an environment E, will mask the preexisting bindings of f in E. In cases where a licit overloading is present, several bindings of f may be accessible in E. The typing algorithm must determine which binding of f to use when typing the application f (e), according to the type of e. Overloading resolution may be carried out during compilation or execution of the program. The overloading resolution
176
Concepts and Semantics of Programming Languages 1
algorithm used is highly dependent on choices made by the language. Different forms, such as operator overloading, function and procedure overloading, constant overloading and method overloading, may or may not be included, independently of one another. The overloading resolution method may be described, informally, as follows. Let exp be an expression containing an overloaded identiﬁer f . To type this expression and thus determine which value of f to use, we must: 1) associate a type variable with each occurrence of f , denoting one of the possible types for this identiﬁer; 2) establish the system of equations which these variables must satisfy, using the typing rules of the language. Additional information may also be used, such as the names of formal parameters if these are used in passing the actual parameters; 3) at the end of this step, each type variable is associated with at least a type; 4) verify that there is only one solution to the system of equations. To do this, we substitute each type variable, one by one, with one of its associated types. If there is no solution, then exp is illtyped. If several solutions exist, then we have an ambiguity and exp is rejected by the typing process. Consider the following example, written in a pseudolanguage that permits overloading for functions and for constants deﬁning enumerated types, i.e. types deﬁned by a set of explicitly named and nonmutable values: type color is (blue, green, red) type light is (green, red) function F(c:color):light = ... function F(x:light):color = ...
Let F_a be the ﬁrst declaration of F and let F_b be the second declaration. Let green_C and red_C be the overloaded values of the type color and green_F and red_F the overloaded values of the type light. The expression F(blue) is clearly typable, taking F_a for F since blue is clearly and unambiguously of type color. The expressions F(green) and F(F(green)), on the other hand, are evidently ambiguous and will therefore be rejected. This ambiguity is removed in F(F(green)) = F(blue): typing this expression leads us to resolve the overloading of F and green. To do this, we write the expression in the form F_1(F_2(green)) = F_3(blue). Let τ1 = τ11 → τ12 be the type of F_1, τ2 = τ21 → τ22 the type of F_2 and τ3 = τ31 → τ32 that of F_3, where tij denotes the type variables. The possible values of τ1 are light → color and color → light. Since F_1 is applied to the result of F_2, we deduce the equation τ11 = τ22 . The results of F_1 and F_3 must be of the same type, giving us the equation τ12 = τ32 . Since F_3 is applied to blue, we have τ31 = color, hence
Types
177
τ32 = light, hence τ12 = light and τ11 = τ22 = color, and thus the type of green is light. The expression F(F(green)) = F(blue) is then analyzed by the typing algorithm, following resolution, as F_a(F_b(green)) = F_a(blue). The rules used in overloading resolution may be more complex in cases where functions may have a variable number of arguments, in the presence of optional arguments, where the language includes object features that permit overloading for methods interacting with inheritance and subclassing. A brief presentation of overloading is given in Chapters 1 and 3 of Volume 2, and the subject is also discussed in relation to each of the languages presented in Chapters 2 and 4 of Volume 2. Note that overloading is not permitted in C. It is important to note that, although overloading makes it easier to write programs, it can make them difﬁcult to review; users must have an indepth knowledge of the associated semantics, which may differ from one language to the next. This point will be addressed further in Volume 2, Chapters 3 and 4.
6 Data Types
6.1. Basic types The languages Exp2 , . . . , Lang3 that we used to describe the notions of semantics and typing only consider two data types: integers and booleans. However, most programming languages include many more basic types. In this section, we shall present the most widespread data types, along with the ways in which they differ between languages. 6.1.1. Booleans For certain languages, such as OCaml, Java, Ada and Pascal, the boolean type bool is a constructor of the type algebra and has two possible values: true and f alse. Other languages, such as C and C++, encode boolean values as integers, with a value of 0 conventionally used to mean "false" and any other value corresponding to “true”. Some of these languages, such as C++, include a type bool and predeﬁned identiﬁers false and true but these values are fully compatible with integer values. Any integer expression may be used as the condition for a loop or an ifthenelse. Any boolean expression may be used as an index in an array, or as an operand in an arithmetic operation. This makes the source text harder to understand and may lead to errors, which are difﬁcult to detect. Python behaves somewhat differently. It supplies two values, True and False, but allows integers and booleans to be mixed. It is thus possible to test whether or not an integer value is equal or different to a boolean value, or to use an integer value as the test for a condition. These practices are not recommended, see the examples given in the following. Remember that an integer, such as the value 2 in example 6.1, may be the result of evaluating a more complex expression. If using this value 2 as a boolean leads to an execution error, it may be difﬁcult to retrieve its origin in the source code.
Concepts and Semantics of Programming Languages 1: A Semantical Approach with OCaml and Python, First Edition. Thérèse Hardin, Mathieu Jaume, François Pessaux and Véronique Viguié DonzeauGouge. © ISTE Ltd 2021. Published by ISTE Ltd and John Wiley & Sons, Inc.
180
Concepts and Semantics of Programming Languages 1
E XAMPLE 6.1.– Python
$ python3.7 >>> 0 == True False >>> 0 == False True >>> 1 == True True >>> 1 == False False >>> 2 == True False >>> 2 == False False
Example 6.1 shows that version 3.7 of Python interprets False as the value 0 and True as the value 1. It is interesting to note that the value 2 (or any other value which is not 0 or 1) is neither True nor False. Thus, the two expressions in example 6.2 – which have identical intuitive semantics – give different results. E XAMPLE 6.2.– Python
>>> if 2 == True : ... print ("2 == ... else : ... print ("2 != ... 2 != True >>> if 2 == False : ... print ("2 != ... else : ... print ("2 == ... 2 == True
True") True")
True") True")
As the tests 2 == True and 2 == False are both always false, these two expressions evaluate the else branch, displaying contradictory results. However, if we use the integer value 2 directly as a condition, the then branch will be evaluated, as if 2 was equivalent to True. E XAMPLE 6.3.– Python
>>> if 2 : ... print ("Then branch") ... else : ... print ("Else branch") ... Then branch
Data Types
181
6.1.2. Integers Given the ﬁnite size of a register, it is only possible to represent a ﬁnite number of binary conﬁgurations. As every integer is encoded in binary (base 2), a register can only encode a ﬁnite number of integers. For positive integers, the bit of rank n in a binary conﬁguration represents the value 2n . The value of the integer encoded by a binary conﬁguration is the sum of its powers. For example, the conﬁguration 01100101 over 8 bits corresponds to 26 +25 +22 +20 = 101. Positive and negative integers are differentiated by the use of a sign bit, which is the most signiﬁcant bit (MSB) in the binary representation of the integer. Furthermore, negative integers are encoded using two’s complement. To obtain the opposite of a positive number, we invert the bits of the binary representation of the absolute value, and then add 1 to the result. Thus, the value 4 is represented over 4 bits as 0100, and the value − 4 is represented as 1100. This encoding guarantees the uniqueness of the value 0 and provides an addition instruction, which is not dependent on the signs of the operands. 6.1.2.1. Size of integers Up to this point, we have only used one integer type, int, representing relative integers. Some languages, such as C, C++ and Java, include several integer types: char, short, long, long long and int. These differ in terms of the number of bits used to encode values. In Java, the number of bits for each integer type is clearly speciﬁed; in C, however, successive versions of the standard simply indicate a minimum size for each integer type. The values of int in C/C++ are generally coded over 32 bits and those of long int are coded over 64 bits, at least in current architectures. Java byte short int long
Bits 8 16 32 64
C/C++ char short (int) int long (int) long long (int)
Bits ≥8 ≥ 16 ≥ 16 ≥ 32 ≥ 64
Table 6.1. Integer size by language
In other languages, such as OCaml or Python 3 (Python 2 permitted the speciﬁcation of long), integer size cannot be speciﬁed, and is set by the standard of the language. In any case, users should keep integer size in mind in order to avoid certain errors.
182
Concepts and Semantics of Programming Languages 1
6.1.2.2. Signed and unsigned integers Working in C and C++, programmers may choose to work with signed integers (used to represent relative integers) or unsigned integers (used to represent natural numbers). This choice deﬁnes the way the MSB is used. For signed integers, this bit is used for the sign, using two’s complement encoding. For unsigned integers, this bit is used in the same way as the others to encode values. This choice is not available in OCaml, Java and Python, where integers are always signed. The range of data that may be represented by an integer type depends on the presence of a sign and on the number of bits assigned to the type. This is presented in Table 6.2. Bits 8 8 16 16 32 32 64 64
Sign Min Max No 0 255 (28 − 1) Yes 128 (−27 ) 127 (27 − 1) No 0 65535 (216 − 1) 15 Yes 32768 (−2 ) 32767 (215 − 1) No 0 4294967294 (232 − 1) Yes 2147483648 (−231 ) 2147483647 (231 − 1) No 0 18446744073709551615 (264 − 1) 63 Yes 9223372036854775808 (−2 ) 9223372036854775807 (263 − 1) Table 6.2. Data range by type
6.1.2.3. Overﬂow As integer size is predetermined, overﬂow may occur in cases where an integer is “too big” or “too small” to be represented over the number n of bits assigned to its type. Integer operations and their properties do not, therefore, belong to Z but to arithmetic modulo Z/nZ. Consider the following example. The decimal value 9 is represented in Z in binary by 1001; the value 7 is represented by 0111 and the value of their sum is represented by 10000. Take a representation using unsigned integers, coded over 4 bits. The representations of 9 and 7 are those found in Z, as they only require 4 bits. The representation of 16, however, becomes 0000: the MSB has been lost. The actual computation is (9 + 7) mod 24 and not 9 + 7. Symmetrically, consider − 8 − 1: assuming that these are signed values, do we obtain − 9? The decimal value − 8 is represented in binary by 1000, and the value 1 is represented by 0001. Their difference is thus 11111, which now requires 5 bits. Again, the MSB has been lost, and we obtain a result of 1111, equal to − 1 in decimal. Similarly, if the location of a given integer variable is ﬁxed to n bits and if this variable is assigned with a value requiring m bits with m > n, then overﬂow will occur: (m − n) bits are lost.
Data Types
183
6.1.2.4. Mixing integers Assigning an unsigned variable with the signed result of a computation, and viceversa, may also result in overﬂow. The value of the computation must therefore be converted to a value with the same type as the variable. This “signed ↔ unsigned” conversion is somewhat dangerous. Let us consider the conversion of the signed number − 4 into an unsigned number using integers encoded over 4 bits. Using two’s complement, the encoding for − 4 is 1100. If an unsigned variable is assigned this bit sequence (without two’s complement), then its value is decoded as an unsigned value and represents the (necessarily) positive value 12. Now, consider the conversion of the unsigned number 15 into a signed number. The encoding for 15 is the sequence 1111, which is interpreted as a signed number using two’s complement as the value − 1. Similarly, mixing signed and unsigned integers in computations can lead to overﬂow. In such cases, the value of one of the operands must be converted in order to acquire the type of the other operand. Example 6.4 shows a case of mixing and its effects. E XAMPLE 6.4.– C
#include int main () { int i = 10 ; unsigned int u = 1 ; if (i + u < 0) printf ("%d + %u < 0\n", i, u) ; else printf ("%d + %u >= 0\n", i, u) ; return (0) ; } $ ./a.out 10 + 1 >= 0
If i + u is negative, the computation will have been carried out using int; otherwise, unsigned int will have been used. The C standard stipulates that if both signed and unsigned operands are present, the signed operand must be converted into an unsigned operand. The value of i is thus interpreted as unsigned, leading to overﬂow, as the sign bit is considered as a “normal” value bit and i no longer represents − 10, but 4294967286. Therefore, the result of the operation is positive. If the computation had been carried out using int, however, the value of i would have remained negative, and the result of the computation would have been negative. The (needed) use of modulo n arithmetic, where n is the size of the integers, creates particularly tricky overﬂow problems, often occurring at the bounds of the representable integer domains. Thus, a program may operate correctly for a time, for
184
Concepts and Semantics of Programming Languages 1
many different inputs, before suddenly and silently producing an inconsistent result (without raising any error) for a case that falls near to the limits. In the absence of an error message, this inconsistent result may propagate throughout the execution, far from the point at which it appeared; this makes the source of the error particularly difﬁcult to identify. 6.1.2.5. Integer arithmetic We shall conclude our study of integers in computer science according to their characteristics in most programming languages by considering some properties of integer arithmetic in a mathematical context. Addition is a total operation over N, extended to a total operation over Z. Subtraction is not total over N (5 − 25 cannot be computed in N) but becomes total in Z. The division encountered in N is the Euclidean division (with an integer quotient and remainder). The exact division operation is deﬁned in Q. In practice, the passage from N to Z is generally implicit: “the temperature is 10◦ , it falls by 25◦ , it is now − 15◦ ”. This implicit passage from N to Z or from Z to R corresponds to a type change in computing. In N, the expression 1 / 4 denotes a Euclidean division with a result of 0, while in R, 1 / 4 is equal to 0.25. Ignoring these shifts between number sets can produce some surprising results: for example, (1 / 4) ∗ 4 gives a result of 0, whereas (1 ∗ 4) / 4 gives a result of 1. These results are perfectly correct, since in integer arithmetic, / is used to denote Euclidean division. However, / is often interpreted implicitly as representing exact division, hence the surprise. When writing arithmetic expressions using integers, the order of evaluation of certain operations (such as multiplications and divisions) is essential, and should be speciﬁed using parentheses. Otherwise, in the expression 1 / 4 * 4, the compiler will choose to compute either the division or the multiplication ﬁrst, depending on the implementation. Switching to another compiler may therefore lead to a change in the result and, possibly, an error. This error does not necessarily become apparent immediately: it may remain latent, giving the impression that the program is correct. Note that Python 3, which is dynamically typed, may change the type of an arithmetic expression imperceptibly, moving from an integer expression to a ﬂoating expression, as we see in example 6.5. The division operator automatically returns a value of the ﬂoating (real) type. E XAMPLE 6.5.– Python
>>> x = 5 >>> type (x)
>>> type (x + x)
Data Types
185
>>> type (x / x)
>>> x / x 1.0
The variable x, initialized with the value 5, is taken to be of type integer. The result of x + x is also of type integer. The result of x / x, on the other hand, is automatically converted to a ﬂoat, as we see from the type of x / x and the displayed value, which includes a decimal element. 6.1.2.6. Implicit conversion of integers in C The C standard stipulates that intermediate computations in an arithmetic expression are to be applied to int by default. This semantic feature of C explains the result obtained in example 6.6, which otherwise appears surprising. E XAMPLE 6.6.– C
#include int main signed i = (i printf return }
(int argc, char *argv[]) { char i = 128 ;  1) / 2 ; ("%d\n", i) ; (0) ;
$ ./a.out 64
The variable i, which is signed over 8 bits (signed char), is initialized with the smallest possible value. Therefore, the evaluation of (i – 1) using modulo arithmetic results in an overﬂow, generating a value of 0; dividing this result by 2, we obtain 0. However, as the arithmetic expression is implicitly executed using int, i.e. 32bits integers, the computation of i – 1 does not generate an overﬂow and gives a result of − 129. This result is then divided by 2 using integer division, producing a ﬁnal result of − 64. This semantic feature prevented the expected overﬂow from occurring, but it can also be a cause of unexpected results, as we see from example 6.7. E XAMPLE 6.7.– C
#include int main (int argc, char *argv[]) { int a = 100000, b = 100000 ; long int c = a * b ; printf ("%ld\n", c) ; c = (long int) (a * b) ; printf ("%ld\n", c) ; c = (long int) a * b ;
186
Concepts and Semantics of Programming Languages 1
printf ("%ld\n", c) ; return (0) ; } $ ./a.out 1410065408 1410065408 10000000000
This program attempts to compute 105 × 105 , storing the result in a long int, which is sufﬁciently large to hold the result of 1010 . The result of the ﬁrst computation is not 1010 . The multiplication was carried out using two ints, resulting in overﬂow. The result was then stored in a long int c. Imposing a type for the result of the multiplication does not solve the problem, as we see from the second computation. The conversion occurs too late in the process: the overﬂow has already occurred. To obtain a correct result, the type of one of the operands, for example a, has to be converted to long, as shown in the third computation. 6.1.3. Characters Prior to the introduction of the Unicode standard and UTF8, UTF16 and UTF32 encodings, characters were represented by integers over 8 bits (one byte). Each of the ﬁrst 128 values was associated with a "letter", following the ASCII standard. The integer value used to represent each character was referred to as its ASCII code. The ASCII standard was extended, with different variants in which the 128 remaining values were used to represent other “letters” (including accented characters). Unicode extended character encoding by allowing “letters” to take up several bytes, vastly broadening the ﬁeld of possibilities. Certain languages have dedicated types for handling Unicode encoding; for example, Java uses the char type (an integer over 16 bits) in addition to the byte type to handle ASCII encoding. Others, such as OCaml and C/C++ only include a char type, used for integers over 8 bits. In this case, the programmer is left to encode characters manually using a different type if extended ASCII is not sufﬁcient. In Python 3, no difference is made between character strings and individual characters. At syntax level, anything that looks like a character (shown between apostrophes in many languages) produces a value of the same type used for character strings (usually shown between speech marks). Example 6.8 provides an illustration of this point.
Data Types
187
E XAMPLE 6.8.– Python
>>> u = ’c’ >>> u ’c’ >>> v = "c" >>> v ’c’ >>> type (u)
>>> type (v)
Similarly, the extraction of a character located at a given position in a string (e.g. the second character in example 6.9) returns a string made up of a single character. E XAMPLE 6.9.– Python
>>> str = "Aloha" >>> str[1] ’l’
Python 3 includes a further type, byte, which is – as the name suggests – used to represent bytes. This type is mainly used to represent data byte by byte, instead of encoding single characters. Byte values less than 128 correspond to characters in ASCII standard. Any higher values are not interpreted as characters, as we see in example 6.10. E XAMPLE 6.10.– Python
>>> b’\x41’ b’A’ >>> b’\xC3’ b’\xc3’
6.1.4. Floating point numbers Floating point numbers are encoded in an entirely different way to integers. The encoding of ﬂoating point numbers is governed by the IEEE 754 standard, which speciﬁes two different lengths: floats, over 32 bits, and doubles over 64 bits. The standard IEC60559 (which is compatible with IEEE 754 for 32 and 64bits ﬂoats) introduced an additional size, long double, which is not numerically ﬁxed. These long double integers are intended to enable greater precision than double, with at least a range as large as that of doubles. According to these standards, a ﬂoating point number is represented by three groups of bits with different meanings. The MSB is used to represents the sign. The
188
Concepts and Semantics of Programming Languages 1
next group of bits represents the exponent. Finally, the signiﬁcand (also known as a mantissa) is coded over the remaining bits. As a ﬁrst approximation – which is technically erroneous, but sufﬁcient to establish an intuitive understanding – we may consider that the value of a ﬂoating point number is sign × mantissa × 2exponent . As in the case of integers, certain languages, such as C/C++, allow users to specify the size of ﬂoating point numbers; others, such as Java, OCaml and Python, have a predeﬁned default size. SEEEEEEEEMMMMMMMMMMMMMMMMMMMMMMM Exponent
Mantissa
Figure 6.1. Encoding a ﬂoating point number over 32 bits. For a color version of this ﬁgure, see www.iste.co.uk/hardin/programming1.zip
Floating point numbers are encoded over a ﬁnite number of bits, meaning that the set of these numbers is ﬁnite. As ﬂoating point numbers are indeed rational, any rational (or real) number falling between two consecutive ﬂoating point numbers cannot be represented exactly. In other terms, there are “holes” between two consecutive rational numbers exactly represented using ﬂoating point numbers, as shown in example 6.11. E XAMPLE 6.11.– C
#include int main () { float fl1 = 48431.1231 ; float fl2 = 48431.1239 ; float fl3 = 48431.1250 ; printf ("fl1: %f fl2: %f fl3: %f\n", fl1, fl2, fl3) ; printf ("fl1 == fl2 ? %d\n", (fl1 == fl2)) ; printf ("fl2 == fl3 ? %d\n", (fl2 == fl3)) ; printf ("fl1 == fl3 ? %d\n", (fl1 == fl3)) ; fl1 = fl1 + 0.0001 ; printf ("fl1 + 0.0001: %f\n", fl1) ; return (0) ; } $ ./a.out fl1: 48431.125000 fl2: 48431.125000 fl1 == fl2 ? 1 fl2 == fl3 ? 1 fl1 == fl3 ? 1 fl1 + 0.0001: 48431.125000
fl3: 48431.125000
Data Types
189
Three variables of type float are declared and initialized with syntactically different constants. The display seems to indicate that they are equal, but the accuracy and precision of this display may be questionable. For this reason, we compare these values with each other and consider the results of the equality tests (booleans). This veriﬁcation indicates that the three variables are, in fact, equal: two of the initialization values were lost in a “hole”. The same holds true for the result obtained by adding 0.0001 to the ﬁrst variable, which does not change: this, too, is lost in a “hole”. Floating point numbers are not the ideal rational numbers of mathematics. As there is only a ﬁnite number of float values, approximation and rounding cannot be avoided in evaluating declarations and arithmetic operations. Example 6.12 shows the veriﬁcation of an expected result for a simple addition of doubles. E XAMPLE 6.12.– C
#include int main double double double double printf printf return }
() { fl1 = 0.1 ; fl2 = 0.2 ; fl3 = 0.3 ; fl4 = fl1 + fl2 ; ("fl1: %f fl2: %f fl3: %f fl4: %f\n", fl1, fl2, fl3, fl4) ; ("fl3 = fl4 ? %d\n", (fl3 == fl4)) ; (0) ;
$ ./a.out fl1: 0.100000 fl3 = fl4 ? 0
fl2: 0.200000
fl3: 0.300000
fl4: 0.300000
Displaying the variables and the result of the computation seems to indicate that 0.1 + 0.2 = 0.3. However, the equality test between the addition and the variable fl3 assigned to 0.3 tells a different story: the computation has been affected by rounding. Finally, the displayed values have also been subject to rounding, and do not correspond to the actual values stored in the memory. To avoid certain problems related to value rounding and computations using ﬂoating point numbers, the equality operators found in programming languages should never be used to test the equality of two ﬂoating point numbers v1 and v2 . It is better to use a socalled comparison (also said a test ± ), i.e. − ≤ v1 − v2 ≤ , or  v1 − v2 ≤ , as we show in example 6.13.
190
Concepts and Semantics of Programming Languages 1
E XAMPLE 6.13.– C
#include #include #define EPSILON (1e6) int main double double double double printf printf return }
/* Very small epsilon... */
() { fl1 = 0.1 ; fl2 = 0.2 ; fl3 = 0.3 ; fl4 = fl1 + fl2 ; ("fl1: %f fl2: %f fl3: %f fl4: %f\n", fl1, fl2, fl3, fl4) ; ("fl3 ~ fl4 ? %d\n", (fabs (fl3  fl4) < EPSILON)) ; (0) ;
$ ./a.out fl1: 0.100000 fl3 ~ fl4 ? 1
fl2: 0.200000
fl3: 0.300000
fl4: 0.300000
The fabs function found in the mathematical library of C is used to compute the absolute value of the difference between fl3 and fl4. The chosen value of allows us to ignore rounding that took place during computation, meaning that we can consider the two values to be equal. Note that the value of depends on the order of magnitude of the values involved in the computation. In astronomy, for example, if the result of a computation states that two stars are within 100 m of one another, this very likely means that the stars have, in fact, collided. In the context of particle physics, on the other hand, if a computation states that two particles are 0.0001 m apart, they have clearly not collided. The examples shown above all use relatively “simple” numbers, without long sequences of digits behind the decimal point. The problems that we have described may occur in any program using ﬂoating point numbers. Furthermore, as the encoding of ﬂoating point numbers is standardized, these problems apply equally to any programming language. 6.1.4.1. Mixing integers and ﬂoating point numbers Certain languages, such as OCaml, prohibit arithmetic operations, which include both integers and ﬂoating point numbers. As we see from example 6.14, an addition operation using integers is denoted as + and is of type int → int → int; an addition operation using ﬂoating point numbers is denoted as +. and is of type float → float → float.
Data Types
191
E XAMPLE 6.14.– OCaml
# (+) ;;  : int > int > int = # (+.) ;;  : float > float > float = # 3 + 4.5 ;; Error: This expression has type float but an expression was expected of type int
Any attempt to use a value of the wrong type will result in a typing error. Developers must explicitly specify the conversion of one of the operands using functions from the language’s standard library, as shown in example 6.15. E XAMPLE 6.15.– OCaml
# # 
(float_of_int 3) +. 4.5 ;; : float = 7.5 3 + (int_of_float 4.5) ;; : int = 7
This choice prevents any “silent” value conversions: the program does exactly what the developer speciﬁes. The loss of ﬂexibility resulting from this choice is counterbalanced by increased ease of understanding and more reliable maintainability. Other languages carry out implicit conversions between integers and ﬂoating point numbers according to their own speciﬁc rules. For example, the C standard (ISO/IEC 9899:201x) stipulates that if one of the operands of an arithmetic operator is a ﬂoating point number, then the other operand must be converted to a ﬂoating point number, and the result will also be of this type. In cases where the result is assigned to a variable, the type of the variable has no impact on the type of the arithmetic computation. Once obtained, the result will be converted to conform to the type of the variable. 6.2. Arrays An array is a data structure that groups data elements of the same type into a contiguous sequence of ﬁxed size. Any element in the array can be accessed in a constant time via an indexing mechanism. Given the location of the ﬁrst element of an array in memory (the base address) and the size of a single element, an element with index i can be accessed at the address base + i × size of an element. The indices of the elements in the array are therefore integer values (or, more generally in certain languages, values of a discrete and totally ordered type). As the size of an array is ﬁxed, attempting to access an element outside the array is a programming error, and may not immediately, or indeed necessarily, trigger an execution error: the error is
192
Concepts and Semantics of Programming Languages 1
thus latent. Indeed, attempts to access a location outside of the bounds of an array may very well concern a memory location that belongs to the program being executed; in this case, there will be no violation of access rights, and the operating system will not react. The program will read – or, in cases of assignment, silently destroy – a zone corresponding to other variables, making the error very hard to detect. Once a change in a variable behavior has been detected – which may take some time – the user needs to work backwards to ﬁnd the illegal access in the array which is responsible for corrupting the data. In the most favorable case, the zone which the program attempts to access does not belong to the program itself: this causes the program to be immediately aborted by the operating system. Certain languages include a mechanism to detect attempts to access locations outside the bounds of an array. For example, during the execution, Python, Ada and Java check that indices remain within the bounds of the array. These veriﬁcations may be deactivated in OCaml, resulting in a mechanism similar to that found in C/C++, where no veriﬁcation is carried out. However, the improvement in performance gained in this way (faster access due to the lack of veriﬁcation) comes at the price of reduced program safety. In some languages, such as OCaml, C/C++ and Java, indices run from 0 to the size of the array − 1. In other languages, such as Ada and Pascal, the index range can be deﬁned using upper and lower bounds. For example, an array might be deﬁned with indices varying from 3 to 15 inclusive. The compiler then translates these indices so that the minimum index corresponds to the base of the array, i.e. to a shift of 0. Python takes a very different approach to arrays. This combines some features of arrays in other languages, as described above, with list mechanisms. In other terms, it is possible to access elements directly via indexing, but it is also possible to add new elements or suppress some of them at any “place” in the array, say at index i, resulting in a modiﬁcation of indices of elements whose indices are greater than i. This ﬂexibility means that the complexity of insertion operations varies, as one or more other elements may need to be moved. The internal structure of an array in Python cannot simply be described by the address of the base. Similarly, the size of arrays is not ﬁxed, and it is adjusted at runtime as elements are added and deleted. Each time an array is full, Python reallocates a greater array by adding several cells at once to avoid too frequent extensions. When an array includes a certain proportion of unused cells, a smaller array is allocated. The precise enlargement/reduction policy used in Python lies outside of the scope of this book. Note that arrays behave in a more complex manner in Python than in other programming languages, as their internal structure is not so easy to grasp. Python also permits negative indices, which correspond to the nth element counted back from the end of the array, thus traversing it from “right” to “left”. It also allows users to manipulate slices of an array, taken as subarrays containing the elements of
Data Types
193
the array present between the two bounds deﬁning the slice, with possible application of an incrementation step, as shown in example 6.16. E XAMPLE 6.16.– Python
>>> t = [0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24] >>> t[2:10] [4, 6, 8, 10, 12, 14, 16, 18]
The constructed slice begins with the element of index 2 and ends at index 10 excluded. Let us continue this example. E XAMPLE 6.17.– Python
>>> t[2:10:3] [4, 10, 16]
With an incrementation step of 3, the slice still begins with the element of index 2 and runs to index 10 excluded, but it jumps two elements from the initial array each time (i.e. the index is incremented by 3 each time). The general syntax of a slice is array[start:end:increment]: any of the three components may be absent, positive, or negative, and in the latter case, the array is processed in reverse order. The ﬁrst slice from the array t in example 6.18 fetches all elements of the array in reverse order, starting with the ﬁfth from the end and working back to the ﬁrst element of the array. In other words, it fetches all elements in reverse order, with the exception of the ﬁnal four values in the initial array. The second slice only takes one element out of two. E XAMPLE 6.18.– Python
>>> t[5::1] [16, 14, 12, 10, 8, 6, 4, 2, 0] >>> t[5::2] [16, 12, 8, 4, 0]
This ﬂexibility in expressing arrays comes at a cost, as the meaning of these syntactic variants of array operations are more involved. Purely for entertainment purposes, readers may wish to determine the result of t[6:2:3], t[6:2:1] or t[6:2:1].
194
Concepts and Semantics of Programming Languages 1
6.3. Strings A character string (or string) is a contiguous sequence of characters. A string is, intrinsically, an array of characters, independent of the way an individual character is represented (see section 6.1.3). Many languages have a predeﬁned type for strings, along with a range of associated operations (concatenation, access to characters, length computation, etc.). Java and C++ offer a dedicated class for strings; OCaml and Python use a predeﬁned type and a module providing operations. In C, strings are encoded by the developer as a character array. A special character (known as the sentinel), with value 0 and noted ’\0’, is used to mark the end of a string. This representation is used internally by C++, ensuring compatibility with Cstyle strings. Depending on the chosen language, it may or may not be possible to modify some characters in a string. Python, Java and OCaml do not permit modiﬁcations, strings in these languages are not mutable. These languages provide alternative types, which implement sequences of characters using mutable bytes. However, a suitable function from the language’s standard library may be used to convert these sequences into strings. In C (and C++ if the dedicated class is not used), although strings take the form of arrays, mutability is not entirely straightforward. A string may be deﬁned in two ways, as shown in example 6.19. E XAMPLE 6.19.– C
char arr[] = "A string" ; char *str = "A string" ;
In the ﬁrst case, the compiler creates an array of characters that is stored in the memory zone of the program variables; characters in the string can be modiﬁed. In the second case, the compiler creates a nonmutable string of characters, which is stored in a readonly area of the memory. Any attempt to modify characters in the string will result in an execution error, as the operating system will detect an illegal attempt to write to a readonly zone. All of the languages, which include a speciﬁc type for strings, also supply comparison primitives, which are used to test equality or order relations between strings. And C, which does not include a dedicated type, does not provide comparison operators for strings. In this case, a comparison operator applied to two strings will simply compare the addresses of the respective arrays, rather than inspecting the contents of the strings (see section 6.6). The language’s standard library provides a string comparison function (strcmp). 6.4. Type deﬁnitions We presented typechecking and inference mechanisms in section 5.4. In this case, the type algebra was ﬁxed, and comprised basic types and the “function” type
Data Types
195
constructor. In section 6.1, we examined a number of basic types used in modern programming languages. In addition to these basic types added to the type algebra, all languages include constructs for deﬁning new types. Type deﬁnition provides a response to three orthogonal needs. First, users may wish to create new values different to those of the basic types (such as int or bool). Second, it is sometimes necessary to combine several types in order to create a more complex data structure. Finally, it may be helpful to rename an existing type to make code easier to read and maintain. In all cases, the deﬁnition of a type involves adding a new name called type constructor to the type algebra. The introduction of type deﬁnitions modiﬁes the existing equality relation between types. The equality of terms of the type algebra, i.e. structural equality, was used implicitly in Chapter 5; two types were considered to be equal if their terms were identical. When we introduce new type names – either to rename existing types or to create structured types – there are two possible meanings of equality of types. In the case of nominal typing, we consider that two types are equal if and only if they have the same name. Otherwise, the equality is deﬁned by verifying that the structures of the types bound to the names are equal: this is known as structural typing. The choice of one of these options is a semantical decision, and it does not depend on the forms of deﬁnition available in a language. 6.4.1. Type abbreviations Type abbreviation consists of naming a type expression. In C and C++, this naming operation is carried out using the typedef construct. In OCaml, a new name is introduced by means of a type deﬁnition, which must relate to a simple type expression (i.e. one comprising only type constructors). Leaving aside the conﬁnement mechanisms presented in Volume 2, Chapter 1. the creation of abbreviations is a means of shortening type expressions. Consider example 6.20, written in C. Instead of const unsigned long long int, we may wish to use a more concise alias, introducing the constructor cullint_t and using it in declarations. E XAMPLE 6.20.– C
typedef const unsigned long long int cullint_t ; cullint_t t[10] ; cullint_t add (cullint_t x, cullint_t x) { return (x + y) ; }
196
Concepts and Semantics of Programming Languages 1
Using type abbreviations also makes code easier to read and maintain. Consider a function for which the inputs are a pressure and a speed, and which returns a power order for a machine. The type of this function is float → float → float. With this type, it is not possible to differentiate between the float numbers representing these different kinds of values, and the type of the function provides no semantical information on its arguments. The introduction of three type abbreviations, as shown in example 6.21, allows a more informative function type. E XAMPLE 6.21.– OCaml
# type pressure_t = float ;; type pressure_t = float # type speed_t = float ;; type speed_t = float # type command_t = float ;; type command_t = float # let regulate (p : pressure_t) (s : speed_t) : command_t = ... ;; val regulate : pressure_t > speed_t > command_t =
In many languages, such as C, C++ and OCaml, type abbreviations do not imply nominal typing. An alias remains equal to the body of its deﬁnition. This makes it possible to deﬁne a variable p of type pressure_t and a variable s of type speed_t, and to compare them or use the operator +. of type float → float → float, as shown in example 6.22. E XAMPLE 6.22.– OCaml
# let p : pressure_t = 2.0 ;; val p : pressure_t = 2. # let s : speed_t = 25.0 ;; val s : speed_t = 25. # p +. 2.5 = s ;;  : bool = false
6.4.2. Records An individual person may be characterized by data including their surname, ﬁrst name and date of birth. Surnames and ﬁrst names are strings, whereas a date of birth is composed of three integers. As these data elements are of different types, they cannot be stored in an array. The notion of a record (or structure) is used to aggregate data of different types within a single type. Each data element is named and represents a ﬁeld of the record. Each ﬁeld is thus assigned the type corresponding to the data element, which it is intended to contain. The deﬁnition of a record type is given by the name of the type and the enumeration of the ﬁeld names, along with their respective types.
Data Types
197
E XAMPLE 6.23.– OCaml
type person = { surname : string ; firstname : string ; month : int ; day : int ; year : int ; }
E XAMPLE 6.24.– C
struct person { char *surname ; char *firstname ; int month ; int day ; int year ; };
As we saw in section 6.3, strings are not a base type in C, which explains the speciﬁc way in which they are declared. If e is an expression of a record type that contains a ﬁeld f ld, the “dot notation” e.f ld denotes the ﬁeld f ld of the value of e. Record type values are constructed by specifying the value associated with each ﬁeld. In OCaml, as we see from example 6.25, a record value is deﬁned by specifying the names of ﬁelds in order to enable type synthesis. One advantage of this seemingly verbose approach is that ﬁeld values can be deﬁned in any order. This also holds true for Ada despite the absence of a type inference mechanism. E XAMPLE 6.25.– OCaml
let p = { firstname = "John" ; month = 7 ; day = 5 ; year = 1900 ; surname = "Doe" }
In C, ﬁeld names are not shown, and so ﬁeld values must be speciﬁed in the same order used when declaring the record type. E XAMPLE 6.26.– C
struct person p = { "Doe", "John", 7, 5, 1900 } ;
In many programming languages, the ﬁelds of a record cannot be deﬁned in relation to one another, or in relation to the ﬁelds which precede them. Examples 6.27 and 6.28 illustrate this point.
198
Concepts and Semantics of Programming Languages 1
E XAMPLE 6.27.– OCaml
# type t = { a : int ; b : int } ;; type t = { a : int; b : int; } # let v = { a = 5 ; b = v.a } ;; Error: Unbound value v # let rec v = { a = 5 ; b = v.a } ;; Error: This kind of expression is not allowed as righthand side of ‘let rec’
E XAMPLE 6.28.– C
#include struct t { int a ; int b ; } ; void f () { struct t v = { 15, v.a } ; printf ("%d\n", v.b) ; } int main () { f () ; return (0) ; } $ gcc foo.c foo.c:7:22: warning: initialization struct t v = { 15, ~ 1 warning generated. $ ./a.out 0
variable ’v’ is uninitialized when used within its own [Wuninitialized] v.a } ; ^
In C, the compiler does not reject deﬁnitions of this type, but warns the user that the value of the record ﬁeld used in deﬁning the record itself is not initialized. Executing the program obtained in this way clearly shows that ﬁeld b was not initialized properly. 6.4.2.1. Record typing The deﬁnition of a record type introduces a new name (type constructor) into the typing environment. It is therefore important to consider how the type of a record value should be determined, and, by extension, how record types with different names but the same structure interact with one another. There are two main approaches to record typing. The ﬁrst method, which permits type inference, consists of guiding typing on the base of ﬁeld names found in the expression. The value of a record expression must therefore include the names of ﬁelds in addition to their values. This is the case in OCaml. Thus, a type deﬁnition type t = { a : int ; b : float } introduces
Data Types
199
two new names, a and b, into the namespace for record ﬁeld names in the environment, while recording the fact that these belong to the type t. When a record expression is analyzed, its ﬁelds must all belong to the same type, guiding its veriﬁcation. Hence, two different record types cannot share the same ﬁeld names (without signiﬁcant effort on the part of the programmer). However, if type deﬁnitions are carried out in different compilation units or modules, ambiguity can be avoided by specifying the compilation unit containing these deﬁnitions. This point is illustrated in example 6.29. E XAMPLE 6.29.– OCaml
# type foo = { a : int ; b : float } ;; type foo = { a : int; b : float; } # type bar = { a : int ; b : float } ;; type bar = { a : int; b : float; } # let v = { a = 42 ; b = 3.14 } ;; val v : bar = {a = 42; b = 3.14}
The expression bound to the value v is inferred as being of type bar, as it was responsible for the most recent binding of the ﬁeld name a to a record type. Values of the foo type can no longer be created as the corresponding ﬁeld names are masked. Note, however, that it is still possible to manipulate values of type foo by acting on their ﬁelds. Using a form of type annotation with the capacity to override inference, a function get_foo_a can be written to fetch the value of ﬁeld a from a value of type foo. E XAMPLE 6.30.– OCaml
# let get_foo_a (x : foo) = x.a ;; val get_foo_a : foo > int =
The second method, used in C, among others, makes use of the type assigned to variables which take record type values. In this case, there is no need to specify ﬁeld names in record values. Thus, the value { 42, 3.14 } does not have a type per se, except that of a twoﬁeld structure in which the ﬁrst ﬁeld is an integer and the second is a ﬂoating point number. However, at the point of assignment or initialization, the type given to the receiving variable at the time of its deﬁnition will be checked for compatibility with the ﬁelds of the expression. This means that two different record types may share the same ﬁeld names, as shown in example 6.31.
200
Concepts and Semantics of Programming Languages 1
E XAMPLE 6.31.– C
struct foo { int a ; float b ; } ; struct bar { int a ; float b ; } ; void f () { struct foo v1 = { 42, 3.14 } ; struct bar v2 = { 43, 3.15 } ; }
Note that, in C, record expressions can only be used to initialize variables. Thus, an expression of the form f ({ 42, 3.14 }) is illegal. In C, Java or OCaml, if two record types have the same structure but different names, they will not be compatible: the typing of records is nominal and not structural. Let us add an assignment of v2 to v1 in our previous program, written in C. In the example 6.32, the assignment of v2 to v1 is rejected, despite the fact that both of these variables are initialized by an expression of the same form. E XAMPLE 6.32.– C
... void f () { struct foo v1 = { 42, 3.14 } ; struct bar v2 = { 43, 3.15 } ; v1 = v2 ; } $ gcc Wall foo.c foo.c:14:6: error: assigning to ’struct foo’ from incompatible type ’struct bar’ v1 = v2 ; ^ ~~ 1 error generated.
6.4.3. Enumerated types The “values” of a trafﬁc light are red, yellow and green. We wish to write a function next, which takes the current state of a trafﬁc light and returns the following state. The ﬁrst solution is to choose a value from a basic type, such as int, for each value of the trafﬁc light: for example, take red = 54, yellow = 12 and green = 36. However, there is nothing to prevent combining the “trafﬁc light values” with any other integer value. Expressions such as red/2 or (next 23) are perfectly welltyped, although they are semantically incorrect. The function next is intended to operate on only three speciﬁc values – so what should it do if it receives an unexpected input value?
Data Types
201
Imagine that we inadvertently deﬁne both red = 36 and green = 36. No error will be raised, despite the fact that these two trafﬁc light values, which should be disjunct, are now equal. An expression used to test a trafﬁc light value will no longer differentiate between a red light and a green light. We explicitly chose the integer values to assign to the light values. There is no real reason for this choice, given that the values used to encode the colors have no inﬂuence on the algorithms acting upon them. Certain languages include a construct that allow users to perform tests enumerating all possible cases of a type’s values and verifying exhaustiveness (no values are left out). In a case where light values are encoded by integers, we would need to write, at best, a general case that should never occur, or, at worst, all other possible cases that should never occur. The capacity to deﬁne a new type with twobytwo disjunct values, only compatible with other values of this type, eliminates these drawbacks. An enumerated type deﬁnition of this kind enumerates the names given to values, known as value constructors. Languages such as Java, C and C++ allow these deﬁnitions to be created by means of the enum type construct. Once value constructors have been introduced, they can be used in the same ways as any other value. This is illustrated in example 6.33. E XAMPLE 6.33.– C
enum signal_t { Green, Yellow, Red } ; enum signal_t next (enum signal_t state) { if (state == Green) return Yellow ; if (state == Orange) return Red ; return Green ; }
The language compiler is in charge of encoding these values by integers internally. Value constructors may be more or less strongly typed depending on the language. In C, they remain integers in their own right: it is always possible to call the next function with any given integer. In C++ and Java, constructors are only values of the type which declaration introduced them and are not compatible with integers. For example, calling next using 0 as an argument in C++ results in a type error: foo.cpp:14:3: error: no matching function for call to ’next’ next (0) ; ^~~~ bar.cpp:4:15: note: candidate function not viable: no known conversion from ’int’ to ’enum signal_t’ for 1st argument enum signal_t next (enum signal_t state) ^ 1 error generated.
202
Concepts and Semantics of Programming Languages 1
6.4.4. Sum types OCaml uses sum types that are more elaborate than simple enumerations of constructors. A sum type is used to group the values of several types into a single type. Take two types t1 and t2 . Let us deﬁne a type t, in which all values are either the values of type t1 , marked by C1 , or the values of type t2 , marked by C2 . The construction of a sum type is similar to the disjoint sum operation on sets (hence the name sum type). C1 deﬁnes the injection of v1 of type t1 that results in the value C1 (v1 ) of type t; similarly, v2 of type t2 is injected as C2 (v2 ) in t. C1 and C2 are called the value constructors of type t. The OCaml code presented in this book has already used a lot of sum types to represent the abstract syntaxes of languages and the different algebras encountered. In example 6.34, playing cards are encoded by a sum type. E XAMPLE 6.34.– OCaml
type card_t = Jack  Queen  King  Number of int
A card may be a picture card, identiﬁed by its name alone. Alternatively, it may be a number card, and the number must be known in order to differentiate it from the other cards in the suit. The value constructor Number is applied to values of the int type. The value constructor Jack deﬁnes a unique value of the type card_t. A sum type may be polymorphic. Some of its constructors may be applied to types written using type variables, introduced by the deﬁnition of the sum type. The predeﬁned lists in the standard library of OCaml are given by a recursive and polymorphic type deﬁnition, similar to that shown in example 6.35. E XAMPLE 6.35.– OCaml
type ’a list = Nil  Cons of (’a * ’a list)
In C/C++, there is no syntactic construct for value constructors. Programmers must encode constructors themselves, using union and struct types. A union type is literally a disjointed sum of the different types mentioned in its deﬁnition. A value of the union is represented on a memory range whose size is the maximum of the sizes of the values of these different types. Thus, it differs from a record, where data is aggregated. Each ﬁeld of the union allows access to the data of the union, which will be interpreted as having the type associated with the ﬁeld. This point is illustrated in example 6.36.
Data Types
203
E XAMPLE 6.36.– C
#include union int_float_t { int as_int ; float as_float ; } ; int main () { union int_float_t v ; v.as_float = 3.14 ; printf ("As the float %f as the integer %d\n", v.as_float, v.as_int) ; return (0) ; } $ ./a.out As the float 3.140000 as the integer 1078523331
Any value of the union type int_float_t can receive two interpretations. The variable v is initialized by considering the union as a ﬂoat. Two views of this value are displayed. The ﬁrst one, as a ﬂoat, allows the interpretation of the data contained in memory in accordance with the interpretation made during writing. As such, we ﬁnd the value 3.14. The second display is performed by interpreting the value as an integer. Since the encoding of integers and ﬂoats are fundamentally different, we see a value that seems irrelevant. Union types in C do not have value constructors like sum types in OCaml. Thus, it is necessary to add a marker to know, at runtime, which ﬁeld of the union to use and how to interpret the value of this union. Example 6.36 shows that an incorrect interpretation gives semantically incorrect results. A suitable marker may be an enumerated type with at least as many values as the union has ﬁelds. This marker and the union should then be grouped together by means of a record. Let us see this with example 6.37. E XAMPLE 6.37.– C
#include enum tag_t { Val_int, Val_float } ; union int_float_t { int as_int ; float as_float ; } ; struct sum_t { enum tag_t tag ; union int_float_t val ; } ; struct sum_t read_int () { struct sum_t v ; scanf ("%d", &v.val.as_int) ; v.tag = Val_int ; return (v) ; } struct sum_t read_float () { struct sum_t v ;
204
Concepts and Semantics of Programming Languages 1
scanf ("%f", &v.val.as_float) ; v.tag = Val_float ; return (v) ; } void print (struct sum_t v) { if (v.tag == Val_int) printf ("%d\n", v.val.as_int) ; else if (v.tag == Val_float) printf ("%f\n", v.val.as_float) ; } int main () { print (read_int ()) ; print (read_float ()) ; return (0) ; } $ ./a.out 42 42 45.7 45.700001
This program deﬁnes a sum type, which is equivalent to the OCaml deﬁnition type sum_t = Val_int of int  Val_float of float. Two functions, read_int and read_float, allow users to input an integer and a ﬂoating point number, and to
generate a sum type value by initializing the union value with the appropriate marker and ﬁeld. The print function selects an option according to the marker, and calls printf with the print format (%f or %d) and the ﬁeld of the union, which corresponds to this marker. Note the presence of a rounding error: 45.7 is displayed as 45.700001. Java does not feature union or sum type constructs. The union notion may be replaced by a record, a ﬁeld being assigned to each type of the union and keeping the marker ﬁeld in order to discriminate between the other ﬁelds. However, in this case, the size of the data element will be the sum of the sizes of all ﬁelds, rather than the size of the largest ﬁeld, resulting in a memory waste. The usual solution to this problem is to create a class hierarchy, extending a base class, representing the sum type without its constructors (readers who are not familiar with the notions of objectoriented programming may wish to consult the Volume 2, Chapter 4, before returning to this explanation). Each value constructor is encoded by an inheriting class of the base class, its arguments being deﬁned by the attributes of this class. Moreover, the class must possess methods returning the values of these attributes. To discriminate between the different kinds of values in the union, there are some solutions: either the base class provides a marker, or a dynamic test on the types of the union is done, or the algorithms must be adapted to the objectoriented style. Python, too, does not propose union or sum type mechanisms. This is the reason why, in this book, the implementation of semantic notions with Python uses classes and the isinstance construct. This allows the mimicking of the OCaml pattern
Data Types
205
matching, in order to obtain rather similar codes. Note that this programming style in Python is rather “wordy” and does not always conform to bestpractice rules for objectoriented programming. An additional construct, known as pattern matching, used to discriminate between the values of sum or enumerated types on a casebycase basis will be presented in section 6.5. 6.4.4.1. Typing sum types In languages where enumerated types are reduced to integers (via the enum construct), such as C or Java, sum types expressions are typed in exactly the same way as integer type expressions. In OCaml, the typing of sum types is directed by the names of the value constructors present in the expression, just as the typing of records is directed by ﬁeld names. A deﬁnition type t = A  B of int introduces the names t, A and B into the typing environment; the fact that A and B are the constructors of t, the fact that A has no argument and the fact that B has an argument of type int are also recorded. Two different sum types cannot share the same constructor names; if this happens, the most recent value constructor will mask the same named constructor of the previous sum type. If deﬁnitions are done in different compilation units or modules, it is possible to distinguish between constructors by precising their compilation unit. 6.5. Generalized conditional Most programming languages include a construct for discriminating between the different values of a type in order to avoid cascading “if. . . then. . . if” situations. This is a form of generalized conditional that only tests equalities. Such a control structure allows to describe a treatment according to the different possible values of a given expression. 6.5.1. C style switch/case In its simplest form, the generalized conditional is an enumeration of some or all values of the type of the evaluated expression, where each case is bound to the code to execute. The switch construct in C allows the discrimination between values of scalar types (more precisely, types represented by integers: this includes enum but excludes ﬂoats). In Java, since version 7, strings can also be processed in this way. A construct of this kind in C is shown in example 6.38.
206
Concepts and Semantics of Programming Languages 1
E XAMPLE 6.38.– C
void f (int x) { switch (x) { case 0: printf ("x is 0\n") ; break ; case 1: printf ("x is 1\n") ; break ; default: printf ("x is something else\n") ; break ; } }
The function f evaluates x and executes the code of the case corresponding to the value of x. The execution of f (1) will therefore display “x is 1”. Each case is ended, hence separated from the next one, by a break instruction. Since the integer type has a very large (while not inﬁnite) number of values, a special case labeled default is used to consider all the values that do not appear in the explicit cases. It is important to note that the default case is not mandatory. If it is not supplied and if the value does not correspond to any of the listed cases, then no action will be applied (silently). For instance, if we were to remove the default case from our example before computing f (2), the function would not display anything. Using a switch without including a default case is not recommended. It is impossible to know, when reading code, if the case was forgotten or if it was simply not written, because in “all other cases” the “empty” behavior is suitable. In order to produce maintainable code, it is good practice to explicitly include an empty default case, immediately followed by a break. This inclusion costs nothing – the compiler will not generate any code for this case – and it will be evident, from the source code, that the author intended to do nothing in such cases. Since enum in C are simply integers, it is possible to discriminate over an enum type by enumerating its values (which constitute a ﬁnite set), as shown in example 6.39. E XAMPLE 6.39.– C
enum dir { North, East, South, West } ; enum dir invert (enum dir enum dir res ; switch (d) { case North: res = South case East: res = West ; case South: res = North case West: res = East ; } return (res) ; }
d) {
; break ; break ; ; break ; break ;
Data Types
207
In this example, there is no default case. Indeed, its utility is not evident since all possible values of the type have been considered. In practice, since the values of the enum are encoded by integers, the program does not cover all possible values in the computing domain, only those declared in the type. In C, the enum type is fully compatible with integer values: a function with a parameter of the enum type may be called using any integer. Hence, if no default case is present, erroneous calls to the invert function will go undetected. The addition of a default case is highly recommended as a means to increase robustness, and is essential when developing critical software. This seemingly regular case structure hides a more complex reality. The break instruction is used to cut off the instruction sequence for a case, just as it can be used to interrupt a loop. In example 6.40, if the break between two cases is omitted, the program will go on to execute the code for the second case straight after the ﬁrst. E XAMPLE 6.40.– C
void f (int x) { switch (x) { case 0: printf ("x is 0\n") ; case 1: printf ("x is 1\n") ; break ; default: printf ("x is something else\n") ; break ; } }
In example 6.40, the break at the end of case 0 was removed. Hence, a call f(0) will display “x is 0”, but also “x is 1”. Furthermore, cases may be seen as simple jump labels, as their order does not follow any particular structure. Cases can be arranged in any order (although default is, by pure convention, the last considered case, but this is not mandatory), and can be put anywhere within the scope of the curly brackets delimiting the switch. These points are illustrated in example 6.41. E XAMPLE 6.41.– C
void f (void) { int j = 24 ; switch (j) { case 23 : { printf ("case 23\n") ; while (j) { default : printf ("case default\n") ; } } break ; } }
208
Concepts and Semantics of Programming Languages 1
In this example, the default case appears in the code between the label for case 23 and the break terminating this case. Furthermore, it is located within the body of a loop. The program above is entirely legal in C, and compilation will not trigger any warning. On execution, as j is different from 23, the program will jump to the default case, i.e. to the middle of the while loop, and will continue to loop inﬁnitely as the condition is always true. 6.5.2. Pattern matching Certain languages, such as OCaml, Haskell, Scala, F# and C#, include a construct known as pattern matching. This is similar to the switch/case construct seen in section 6.5.1, but it is a lot more expressive as it permits the deconstruction of structured values (sum type values, records or objects). The switch/case construct tests the equality of the examined value with atomic constants. Pattern matching is a way of testing using the structure of a value. Consider the sum type type t = I of int  B of bool and v a value of t. We can state cases such as “v is of the form I of something or B of something”, and use the value “something” in the treatment devoted to the case. A pattern is a kind of skeleton built from value constructors, labels of records and variables, deﬁning a shape. Such a pattern matches all values having the same shape and this matching binds the variables of the pattern to the corresponding subterms of the value. The names of these variables may be used by the instructions associated with the pattern and their bindings will be used when the instructions are executed. For the type t above, the pattern I x matches the value I 46: they are both built with the same value constructor I. This matching binds x to 46 when the code associated with the pattern I x is executed. A more formal deﬁnition of pattern/value compatibility is given below. First, however, let us consider some examples, shown in OCaml and using the match/ with construct. E XAMPLE 6.42.– OCaml
type dir = North  East  South  West let invert d = match d with North > South  East > West  South > North  West > East
This function, deﬁned by cases, is identical to that of the example 6.39 written to illustrate the switch/case construct in C. As the type dir is made up of exclusively argumentfree constructors, pattern matching operates in the same way as a simple switch/case. If the argument of invert is West, then the fourth branch of the matching will be selected during the execution and the value East will be returned.
Data Types
209
E XAMPLE 6.43.– OCaml
type intlist = End  Cell of (int * intlist) let rec sum l = match l with Cell (i, remainder) > i + sum remainder  End > 0
The intlist type represents lists of integers, and the function sum returns the sum of the integers in a list. If l = Cell (a, r), then the variable i is bound to a and the variable remainder is bound to r. The bindings of these two pattern variables are used in the body of the case, enabling the addition of the current integer value and the sum of the remainder list, computed recursively. If l = End, the recursion stops and the value 0 is returned. Note that a “catchall” pattern, denoted as _, may replace a pattern variable if the value matched by this variable is unused in the body of the case. This pattern is equivalent to a variable, but avoids introducing a new name. It provides a clear indication to readers that this part of the matched value is not used. E XAMPLE 6.44.– OCaml
let is_empty l = match l with Cell (_, _) > false  _ > true
A list is empty if the Cell constructor has no occurrence in its value, which is therefore End. We ﬁrst consider the case of a value built with the constructor Cell, which will return false whatever is the integer and the remainder list arguments of this constructor. Thus, the corresponding pattern is Cell(_,_) with two occurrences of _. For all other cases (in our example, only End), the list is empty. Note that the ﬁnal _ “catches” the only remaining case End: as such, we could write  End > true in a completely equivalent manner. 6.5.2.1. ML style pattern matching: formalization Before establishing a formal semantics of pattern matching, we need to provide a more detailed explanation of the notion of term, which we have discussed brieﬂy in the context of type inference in section 5.4 where the word term referred to type expressions. More generally, terms are constructed from a set Σ of symbols, which each have an arity (the number of “arguments” of the symbol) and a set of variables V , according to the following inductive deﬁnition. D EFINITION 6.1.– Term: the set T (Σ, V ) of terms over Σ and V is deﬁned by: – every symbol of arity 0 is a term in T (Σ, V ); – every variable in V is a term in T (Σ, V ); – if t1 , . . . , tn are terms in T (Σ, V ), and if C is a symbol of arity n, then C(t1 , . . . , tn ) is a term in T (Σ, V ).
210
Concepts and Semantics of Programming Languages 1
As we deﬁned a substitution on types, replacing type variables by types, we deﬁne a substitution on terms as a partial ﬁnite domain function of the set of variables into the set of terms and extend it naturally into a function of terms into terms. Substitutions can be used to establish a preorder of generality on terms. A term t1 is less general than a term t2 if there exists a substitution φ such that t1 = φ(t2 ) and it is noted t1 ≤ t2 . Formally, matching a value v against a pattern m consists of considering the pattern m as a term and determining if v ≤ m. So we have to compute the substitution φ associating the variables of the pattern with subterms (subvalues) of v. If such a substitution φ exists, then m is said to match v and φ is used to extend the execution environment of the pattern matching case. 6.5.2.1.1. Order of patterns Pattern matching is done according to the patterns involved in the match/with construct and, before giving a formal semantics to this mechanism, we precise how the choice of the right pattern is done. Patterns are examined in the order in which they appear. If the ﬁrst pattern does not match the examined value, the second case is tested, and so on for as long as necessary. If no pattern is found that matches the value in question, an execution error is raised. Note that languages in the ML family analyze pattern exhaustiveness at compiletime and a warning is emitted by their compiler in case of nonexhaustive pattern matching. If a value matches more than one pattern, the order in which patterns are listed is decisive: the ﬁrst matching pattern is used, and the other matching patterns are ignored. If several patterns always match the same values, then the principle of choosing the ﬁrst one still holds; however, for example in OCaml, the compiler raises a warning to indicate the existence of this redundancy in the match/with construct. These different situations – partially overlapping cases, totally overlapping cases and missing cases – are illustrated in example 6.45. E XAMPLE 6.45.– OCaml
type logic_op = Or of (bool * bool)  And of (bool * bool) let eval e = match e with  Or (true, _)  Or (_, true)  Or (false, true)  And (true, true) > true
Data Types
211
 And (_, _) > false Warning 8: this patternmatching is not exhaustive. Here is an example of a case that is not matched: Or (false, false) Warning 11: this match case is unused. val eval : logic_op > bool =
The type logic_op represents boolean conjunction and disjunction expressions. Note that true and false are exceptions to the rule in OCaml syntax stating that sum type constructors must always start with an uppercase letter. The function eval returns the boolean value of an expression. If we compute eval (Or (false, true)), the ﬁrst pattern does not match, as the ﬁrst argument of Or in the pattern is true. We then examine the second case: this pattern matches, as any value is accepted as the ﬁrst argument, and true is the second argument in both the pattern and the matched value. The third pattern is redundant as all of the values that it captures have already been matched by the previous case. The compiler ﬂags this issue in the second warning. The case where both arguments in the pair are false, However, the case of a value of the form And (false, false) is not considered: the compiler issues a warning, see the ﬁrst message shown in the example. 6.5.2.1.2. Operational semantics We shall extend the abstract syntax of Exp2 , presented in section 3.1.1, as shown in Table 6.3. p ::= v Pattern variable  C(p1 , . . . , pn ) Value constructor e ::= ... Previous expressions  match e with p → e  _ → e Pattern matching Table 6.3. Extension of the syntax of Exp2 for pattern matching
First, we need to deﬁne the syntax of patterns, which are essentially composed of constructors and variables (the catchall pattern can be assimilated to a variable of which the name is irrelevant). We then add a construct for pattern matching with only two cases. The ﬁrst case is a given pattern, and the second is the catchall, which always succeeds. This structural limitation is used to simplify presentation, and it has no effect on the generality of the language. It is always possible to rewrite match e with P1 > e1  P2 > e2  P3 > e3
as
match e with  P1 > e1  _ > ( match e with  P2 > e2  _ > e3 )
212
Concepts and Semantics of Programming Languages 1
The function filter, used to check if a pattern p matches a value v, is deﬁned by recursively comparing the structures of p and v. It calls a function filter’ using an empty initial substitution. The substitution that is used as input for filter’ accumulates the bindings created by the comparison between the pattern and the value. Box 6.1 shows the deﬁnition of these two functions, providing only nominal cases. The matching process implicitly fails on all other cases. x ∈ dom(φ)
φ(x) = v
filter’(φ, x, v) = φ ⊕ (x, v)
filter’(φ, x, v) = φ
φ1 = filter’(φ, p1 , v1 )
...
φn = filter’(φn−1 , pn , vn )
filter’(φ, C(p1 , . . . , pn ), C(v1 , . . . , vn )) = φn filter(p, v) = filter’(∅, p, v) Box 6.1. Matching function
The rules of the operational semantics are shown in Box 6.2 and use the matching function of Box 6.1. The rule Ematch1 corresponds to the case where the ﬁrst pattern p matches the value v of the expression e, returning a substitution φ. It is interesting to note that the structures of a substitution and an environment are identical, that is, lists of associations of a variable and a value. Thus, we may extend the environment with the substitution φ returned by filter to evaluate e1 , the expression associated with the ﬁrst pattern. The rule Ematch2 corresponds to the case where the ﬁrst pattern does not match the value v of e, indicating that a substitution φ cannot be found. The result of the computation is the evaluation of the expression e2 associated with the second pattern, which captures any value. (Ematch1 ) (Ematch2 )
E nv
ev
φ = filter(p, v)
φ ⊕ Env e1 v1
match e with p → e1  _ → e2 v1 E nv e v ∃φ = filter(p, v) Env e2 v2 E nv E nv
match e with p → e1  _ → e2 v2
Box 6.2. Operational semantics of pattern matching
6.5.2.1.3. Typing rules There are similarities between the typing of the match/with construct and that of the conditional, in that all branches must return values of the same type. The type of a value constructor is a functional type, of which the argument type is the product of the types of its arguments and the result type is t, (the type of the value built with this constructor). We assume given a function TypeCstr that provides the type scheme of a constructor. The value constructors of sum types are bound to their type scheme in the typing environment when the sum type is deﬁned.
Data Types
213
Rules Tpvar and Tpcons typecheck patterns and return an environment extended by the bindings of pattern variables to their respective types. The rule Tmatch ensures that the pattern and the examined value are of the same type. It typechecks the body of the ﬁrst case in the environment extended by the bindings of pattern variables, and typechecks the body of the second case in the initial environment, verifying that the types of these bodies are the same.
(Tpcons ) (Tmatch )
(Tpvar ) Env x : τ ⇒ (x, τ ) ⊕ Env (τ1 ∗ . . . ∗ τn ) → τ ≤ TypeCstr(C) E nv p1 : τ1 ⇒ E nv1 . . . Envn−1 pn : τn ⇒ Envn
E nv
e:τ
C(p1 , . . . , pn ) : τ ⇒ Envn p : τ ⇒ Env Env e1 : τ
E nv E nv E nv
E nv
match e with p → e1  _ → e2 : τ
e2 : τ
Box 6.3. Typing rules for pattern matching
6.5.2.2. Extensions and other forms of matching 6.5.2.2.1. Guarded patterns Many of the languages that use pattern matching associate a condition, known as a guard, with a pattern (the when construct in OCaml, C#, and F#, or if in Scala). The associated branch will only be taken at runtime if the pattern matches the examined value and if the condition is satisﬁed, as shown in example 6.46. E XAMPLE 6.46.– OCaml
card type = Jack  Queen  King  Number of int let card_name c = match c with  Jack > "jack"  Queen > "queen"  King > "king"  Number 1 > "ace"  Number n when n > 1 && n < 11 > string_of_int n  Number _ > "illegal card!"
The card_name function handles the case Number 1 separately, enabling it to return a more usual string, i.e. “ace” instead of “1”. The other legal numerical values, from 2 to 10 inclusive, are handled separately. This is done using the match case Number n, which is limited to numerical values in this interval. Adding a guard to this pattern means that it cannot match all remaining values of the card type, even if they are structurally compatible. An additional case is therefore needed to handle values rejected by the guard.
214
Concepts and Semantics of Programming Languages 1
6.5.2.2.2. Disjunction of patterns Some situations require the same treatment to be applied to several cases. In this situation, instead of enumerating patterns one by one and repeating the same body each time, we may use a single pattern deﬁned by a disjunction of (sub)patterns, and a single processing may then be applied, as shown in example 6.47. E XAMPLE 6.47.– OCaml
let crowned_head c = match c with Queen > true  King > true  _ > false let crowned_head2 c = match c with Queen  King > true  _ > false
The function crowned_head, which speciﬁes whether a crown is shown on a playing card, produces the same code twice: once for Queen, and once for King. In the second version, the ﬁrst two patterns are replaced by a single pattern that accepts both Queen and King. This pattern may be read as “match Queen or King”. Using “or”patterns is particularly helpful in situations where the (identical) processing to apply to each case is a long piece of code. Sharing the code in these situations reduces the sizes of the source code and the executable. Given that a disjunction pattern uses the same processing for each subpattern, if one of these subpatterns contains a variable, this one must also occur – and be of the same type – in the other subpatterns. This condition is necessary to ensure that this pattern variable will be well bound, independently of the speciﬁc subpattern used to catch the examined value, and will only have one type. This point is illustrated in example 6.48. E XAMPLE 6.48.– OCaml
type exparith =  Plus of (exparith * exparith)  Minus of (exparith * exparith)  Mult of (exparith * exparith)  Div of (exparith * exparith)  Num of int let rec count_operators e = match e with  Plus (e1, e2)  Minus (e1, e2)  Mult (e1, e2)  Div (e1, e2) > 1 + (count_operators e1) + (count_operators e2)  Num _ > 0
The purpose of the count_operators function is to count the number of operators present in an arithmetic expression. The four cases concerning +, −, ∗, / all require two recursive calls, one for each subexpression. The treatment of the associated constructors can be factorized, as they allow the same variables e1 and e2 to be bound to the same type on every occasion. 6.5.2.2.3. Matching records Patterns syntax also includes records with named ﬁelds, as presented in section 6.4.2. In a pattern, only those ﬁelds that are relevant to a treatment need to be named. This avoids the need to include all ﬁelds with a _ as a subpattern, as we see in example 6.49.
Data Types
215
E XAMPLE 6.49.– OCaml
type transport = Bus  Metro  Tram type line = { number : int ; kind : transport ; company : string } let rec filter_metro l = match l with  [] > []  li :: rest > ( match li with  { number = n ; kind = Metro } > n :: (filter_metro rest)  _ > filter_metro rest )
The function filter_metro takes a list of public transport lines as input, and returns a list containing only the numbers of metro lines. Since the company ﬁeld is irrelevant to this function, it is left out of the pattern. In practice, the ﬁnal pattern _ has the record type line, implicitly signifying “any value for all ﬁelds of this type”. 6.5.2.2.4. Matching on types Some languages have the ability to test the type of the values encountered during execution. This ability is often linked to object mechanisms, which we shall present in greater detail in Volume 2, Chapter 4. For example, classes of C# or Scala, can be used to deﬁne a hierarchy of structured types, enriched incrementally, along with functions used to test the type of the objects which they handle and to adapt their treatment accordingly. E XAMPLE 6.50.– Scala
import scala.math.sqrt abstract class Shape case class Point (x : Double, y : Double) extends Shape case class Line (x1 : Double, y1 : Double, x2 : Double, y2 : Double) extends Shape case class Circle (x : Double, y : Double, radius : Double) extends Shape def perimeter (shape : Shape) : Double = { shape match { case Point (_, _) => 0 case Line (x1, y1, x2, y2) => sqrt ((x2  x1) * (x2  x1) + (y2  y1) * (y2  y1)) case Circle (_, _, r) => 2.0 * 3.14159 * r } }
216
Concepts and Semantics of Programming Languages 1
This program, written in Scala, deﬁnes a general type Shape then two types Point and Line which “extend” the ﬁrst type using their own ﬁelds. The perimeter function uses a value of the general type as input, and examines the precise type of this value. The body of each match case may then use the speciﬁc ﬁelds coming from the actual type of the examined value. At ﬁrst glance, case class could be perceived as a different syntactic approach to deﬁning sum types; however, the construct is not simply limited to matching and provides several additional mechanisms used in objectoriented programming. 6.6. Equality In section 6.5, we introduced constructs generalizing the conditional, ﬁrst to deal with more than two cases, then to provide a notion of “equality” broadened by the pattern matching. While the concept of equality used in programming languages appears straightforward, it merits more detailed examination. In a mathematical context, equality is deﬁned as a reﬂexive, symmetric and transitive relation between two elements of the same set. The intuitive semantics of the equality of two elements consists of stating that if, in a mathematical expression, the ﬁrst is replaced by the second (or vice versa), the value of the expression remains unchanged. The notion of equality was formalized in the 17th century by Gottfried Wilhelm Leibniz, who stated that two values x and y are equal if, and only if, for any predicate P , P (x) is true if and only if P (y) is true: (x = y) ⇔ ∀P, P (x) ⇔ P (y) The difﬁculty with this deﬁnition lies in the need for universal quantiﬁcation on predicates, i.e. on functions returning a truth value. Hence, this deﬁnition cannot be directly applied in computer science. It is often said that two “computer” values are considered to be equal if their representations in the memory are identical. In other terms, the sequences of bits encoding these values must be of the same length and contain twobytwo identical bits. Using the usual semantics for adding integers, it is easy to understand that 4 = 3 + 1; in other terms, the values resulting from the evaluation of the expressions 4 and 3 + 1 are equal, and both will be represented internally, in binary, as 100b . N OTE.– The syntax used for equality varies between languages. For example, in C, Java, Python and others, the notation == is used. In OCaml, on the other hand, equality is noted = (== has a different equality semantics) like in Basic, Pascal and Ada. Given that assignment is also noted = in many languages, users are strongly advised to study the syntax of any new language in order to avoid misinterpretations.
Data Types
217
There are two possible interpretations when testing the equality of two identiﬁers x1 and x2 : we may wish to check whether Env(x1 ) = Env(x2 ), or whether M em( E nv(x1 )) = M em( E nv(x2 )). 6.6.1. Physical equality In languages such as C, the equality operator == tests the identity of values in the compared expressions. These values must be scalar, otherwise the test will be rejected during the compilation. This point has signiﬁcant implications for comparing data structures. Generally, the encoding used to construct evolved data structures involves pointers, which are scalar values. Equalities between data structures can relate to the addresses of these structures, in the case of physical equality, or to the values of the structures, known as structural equality, a notion presented in section 6.6.2. In C, when the equality == is used in relation to data structures, it denotes physical equality. The most striking example of this concerns the comparison of strings. As we saw in section 6.3, strings in C are encoded as arrays of characters (terminating with the character ’\0’) and the arrays are denoted by pointers (see section 7.3). In testing the equality of two variables representing strings, we are, in fact, testing the equality of the pointers representing these strings, and not equality of the data stored at these locations. This is illustrated in example 6.51. E XAMPLE 6.51.– C
#include int main () { char *s1 = "Foo" ; char s2[4] = { ’F’, ’o’, ’o’, ’\0’ } ; if (s1 == s2) printf ("Equal\n") ; else printf ("Different\n") ; return (0) ; }
This program declares the variable s1, containing the string "Foo" and the variable s2 as a character array containing the same letters as the string "Foo" and ending with a ’\0’. Thus, s1 and s2 “contain” identical strings. However, at runtime, this program will return Different, as the locations denoted by s1 and s2 differ. The semantics of equality in C is even more complex when we consider the case of constant string declarations. The compiler operates by creating only one occurrence of each constant string in a program. Thus, different variables initialized with the same constant string will denote the same reference. A modiﬁed version of example 6.51 is shown below, in which s1 and s2 are initialized with the same string "Foo".
218
Concepts and Semantics of Programming Languages 1
E XAMPLE 6.52.– C
#include int main () { char *s1 = "Foo" ; char *s2 = "Foo" ; if (s1 == s2) printf ("Equal\n") ; else printf ("Different\n") ; return (0) ; }
At runtime, the program will now return Equal since the compiler initialized s1 and s2 using the location of a single occurrence of the string "Foo", assigned once and for all in the program code. As we saw in section 6.3, the difference between declaring a string as a pointer or an array adds another subtle level of difﬁculty. A declaration of an array will necessarily result in a distinct memory allocation, and will thus produce a negative result in an equality test, even if two variables of this type are initialized using the same string. In the example below, a subtle change has been made to the program in 6.52. Instead of declaring s2 as a char*, we have chosen to declare it as a char[], i.e. as an array. E XAMPLE 6.53.– C
#include int main () { char *s1 = "Foo" ; char s2[] = "Foo" ; if (s1 == s2) printf ("Equal\n") ; else printf ("Different\n") ; return (0) ; }
At runtime, this program will return Different since the array designated by s2 is necessarily found at a separate location to the string "Foo" used to initialize both cases. 6.6.2. Structural equality Structural equality relates to the contents of addresses. Two values that are physically equal will be structurally equal, but the reverse is not necessarily true.
Data Types
219
Structural equality is only possible in languages with native data structures, and the size and form of the information composing the structure must be known at compiletime. This is the case in both OCaml and Python, which allow structural comparison of lists, arrays, etc. In these languages, the compiler chooses the representation of these structures, unlike in C, where encoding is left to the programmer. The internal implementation of equality draws on knowledge of the compiler’s representation choices. Note that these languages also have the capacity to test for physical equality (== in OCaml, is in Python). In languages such as C or C++, the compiler cannot automatically determine which semantics to use in comparing two arbitrary data structures. There is no means of knowing whether any given pointer represents either the address of a single entity, or that of an array (and thus of multiple values). This task can only be carried out by the developer, who knows the semantics of the structures which she or he has created. Thus, in the case of a linked list structure, the programmer must implement a comparison which follows the links in the list, knowing that each pointer only provides the address of the next element. In a situation involving arrays of linked lists, the programmer must develop an operation to compare consecutive objects, knowing that the pointer for an array designates the beginning of a sequence of elements. This point is illustrated in example 6.54. E XAMPLE 6.54.– C
#include #include struct list { int val ; struct list *next ; };
/* Cell value. */ /* Address of the next cell. */
/* Equality check of two lists denoted by the address of their head cells. */ bool list_equal (struct list *l1, struct list *l2) { if (l1 == NULL) { if (l2 == NULL) return (true) ; else return (false) ; } else { if (l2 == NULL) return (false) ; else return ((l1>val == l2>val) && list_equal (l1>next, l2>next)) ; } } /* Equality check of two arrays of lists. The sizes of the arrays are used as arguments, as they are not included in the array structures. */ bool tab_list_equal (int long1, struct list *t1, int long2, struct list *t2) { int i ;
220
Concepts and Semantics of Programming Languages 1
if (long1 != long2) return (false) ; else { for (i = 0; i < long1; i++) { if (! list_equal (&t1[i], &t2[i])) return (false) ; } return (true) ; } }
Although both functions list_equal and tab_list_equal use pointers over struct list, their behavior is radically different due to the fact that, semantically, these pointers do not represent the same data structure. Note that in C, the size of an array is not associated with its structure; the size of each array must therefore be supplied as an argument for function tab_list_equal. Since arrays t1 and t2 contain lists and not pointers over lists, function list_equal is called specifying the addresses of the elements in the array to compare. The decision to deﬁne t1 and t2 as arrays of lists instead of list addresses is arbitrary; our aim is essentially to show that, even if the two comparison functions are of similar types, their algorithms are fundamentally different. Note that in practice, where data structures of this kind are created dynamically, t1 and t2 would most likely contain pointers to lists, and would therefore be of the type struct list**. For character strings, the standard library of C includes a function strcmp for structural comparison purposes. This function is easy to write, as the strings are arrays terminating with a sentinel marker (’\0’). In objectoriented languages, such as Java or C++, strings are encapsulated within classes from their standard library, which include a structural comparison function. Nevertheless, this function is not one of the primitives of these languages and was written by the library’s developers. In languages that include an operator overloading mechanism, such as C++ (see section 5.7.2), it is possible to change the behavior of == when used with data structures in order to assign a dedicated semantics, for example structural equality. This redeﬁnition is done during the implementation of the data structure. Once again, this approach relies on explicit development work carried out by the designer of the data structure based on a thorough knowledge of the semantics of this structure. 6.6.3. Equality between functions We discussed the semantics of equality tests for functions at the start of Chapter 5, indicating that there cannot exist a satisfactory response to this problem. The choices used by a number of languages are described as follows. 6.6.3.1. C, C++, Python These languages compare the addresses of the code of the compared functions. The equality is thus implicitly physical in nature. The only way to obtain a positive test result is to compare a function identiﬁer with itself.
Data Types
221
6.6.3.2. OCaml In this case, any attempt to test the structural equality of two functions will result in an error message, even if the two operands are the same function identiﬁer. Nevertheless, physical equality can still be tested, using the same semantics found in C or C++, as shown in example 6.55. E XAMPLE 6.55.– OCaml
# let f x = x ;; val f : ’a > ’a = # f = f ;; Exception: Invalid_argument "compare: functional value". Raised by primitive operation at unknown location Called from file "toplevel/toploop.ml", line 180, characters 1756 # let g x = x ;; val g : ’a > ’a = # f == f ;;  : bool = true # g == f ;;  : bool = false
6.6.3.3. Java Function comparison is illegal in Java, and this is indicated during the compilation. There is no implementation of the predeﬁned == operator for functions, so the compiler simply rejects these expressions. Furthermore, functions (methods) are not objects belonging to classes, so the equal method, used to test equality between objects in a class, is not applicable. 6.6.3.4. Coq This tool is both an environment for formal proof and a programming language. It is widely used to construct proofs for properties veriﬁed by computer, such as those concerning communication protocols [DUB 18]. Users may write expressions, including functions, and prove properties over them. Under certain conditions – which lie beyond the scope of this book – OCaml code can be extracted automatically from a development in Coq. Coq can manipulate terms representing a program, and thus can rewrite certain terms and deduce an equality. This even holds true for certain functions, meaning that it is possible to prove an equality between functions which are a priori syntactically different, as shown in example 6.56.
222
Concepts and Semantics of Programming Languages 1
E XAMPLE 6.56.– Coq
Welcome to Coq 8.8.2 (November 2018) Coq < Definition f (x : nat) := x. f is defined Coq < Definition g (y : nat) := if true then y else 0. g is defined Coq < Goal f = g. 1 subgoal ============================ f = g Unnamed_thm < compute. 1 subgoal ============================ (fun x : nat => x) = (fun y : nat => y) Unnamed_thm < auto. No more subgoals. Unnamed_thm < Qed. Unnamed_thm is defined
We begin by deﬁning two functions f and g and precise our goal which is to prove their equality (using the Goal command). We use a tactic that is native to Coq to compute and rewrite the terms found in the goal to prove (using the compute command). Coq is then able to eliminate the test (false) found in g, producing two function bodies that are identical, apart from the names of the bound variables. Finally, a conclusion of equality is reached automatically (auto command) despite this difference in the variable names. Of course, this power comes both from the fact that the developer guides Coq through the proof process and from the fact that the functions being compared are relatively simple. The automatic process only works for relatively simple functions. In more complex cases, the responsibility for proof falls to the user, and may become arbitrarily complicated (if not impossible) depending on the way in which functions are deﬁned. For example, the equality of the functions f x = x + x and g y = 2 * y cannot be veriﬁed automatically, and the proof can be done, but requires signiﬁcant work.
7 Pointers and Memory Management
This chapter is dedicated to the notions of pointers and references, already mentioned in Chapter 4, with illustrations mostly written in C and C++. These languages include explicit pointers, which sometimes reveal technical details closely linked to computer architecture. Pointers and references are essential to the notion of objects, which will be discussed in greater detail in the context of objectoriented languages (Volume 2, Chapters 3 and 4) in relation to Java (see Volume 2, section 4.1), Python (see Volume 2, section 4.4) and C++ (see Volume 2, section 4.2). 7.1. Addresses and pointers In terms of lowlevel implementation, a computer memory may be seen as a large “array” of bytes. This “array” is made up of zones corresponding to program variables and, more generally, of zones explicitly or silently allocated by the program. Some of these zones are created automatically by the program’s execution mechanisms, some by compilation mechanisms and some by demands formulated explicitly in the source code. The information manipulated by a program is stored in the memory at a certain location, indicated by an address: this notion has already been discussed in Chapters 2 and 4. An address may thus be seen as the index in the memory “array” where an item of information is stored. An address is an element of information like any other: it may be recorded, manipulated or used to read or write at the location it denotes. A value of the address type is called a pointer. A variable containing an address is therefore of address type. The term “pointer” is often used indifferently to designate an address type or an address value according to the context. As we saw in our study of typing imperative features, a pointer type indicates the type of the pointed values. This rule holds true for languages such as C and C++,
Concepts and Semantics of Programming Languages 1: A Semantical Approach with OCaml and Python, First Edition. Thérèse Hardin, Mathieu Jaume, François Pessaux and Véronique Viguié DonzeauGouge. © ISTE Ltd 2021. Published by ISTE Ltd and John Wiley & Sons, Inc.
224
Concepts and Semantics of Programming Languages 1
Memory
and guarantees that values stored at an address will be manipulated in a way that is consistent with their type. In C, for example, suppose that we are reading the value at a location denoted by a pointer p. If p denotes the address of an integer (int), there will be 4 bytes to read. In the case of a ﬂoat (double), there will be 8 bytes to read. Figure 7.1 shows a schematic representation of the integer 23 in the memory, beginning at address 10. The way bytes are organized in the memory will be discussed in detail in section 7.2. 0
1
2
3
8
9
10
11
00 14
15
16
4
12
00 17
6
13
00 18
7
14
23 19
20
...
Figure 7.1. Schematic representation of memory and address. For a color version of this ﬁgure, see www.iste.co.uk/hardin/programming1.zip
In C, the pointer type is denoted by the constructor *. Thus, int * denotes a pointer over int, i.e. an address where integer type information will be found. The notation int *x; declares the variable x as a pointer to an integer: literally, the value pointed by x has type int. Variables may be created to record addresses, and these variables are also stored at an address in the memory. Thus, int** represents a pointer to a pointer to an integer. The information found at the address designated by such a pointer is itself an address.
Memory
Figure 7.2 shows the integer introduced in Figure 7.1 alongside a pointer, the value of which represents the address of this previous integer. For the purposes of this representation, the pointer was arbitrarily considered to be of the same size as an integer, but this is not always the case in current architectures. 0
1
2
3
8
9
10
00 14
15
16
4
11
00 17
00
6
12
00 18
00
7
13
14
23 19
00
20
10
...
Figure 7.2. Pointer to a pointer. For a color version of this ﬁgure, see www.iste.co.uk/hardin/programming1.zip
Pointers and Memory Management
225
7.2. Endianness The way in which integers are stored in the memory varies from processor to processor. Storage of integers needs several bytes, which may be stored in the memory from left to right or from right to left. This order, known as endianness, is imposed by the processor’s architecture. Processors that store the least important bytes of an integer at the lowest addresses are said to be “little endian”. In an architecture of this kind, for the hexadecimal value 01020304, the byte 01 would be placed at the address at the start of the integer, 02 would be placed at the following address, and so on. Intel x86 and ARM processors are little endian. Processors that store bytes in the reverse order, with the most important byte at the lowest address, are said to be “big endian”. In this case, for the value 01020304, the byte 04 would be stored at the address of the start of the integer, followed by 03, and so on. The MC68000 family of Motorola processors are big endian. Certain processors, such as MIPS or PowerPC models, can be conﬁgured (using hardware, software, or both) to operate in either mode. Of course, endianness can lead to compatibility problems if data structures are directly extracted from memory and transmitted (over a network, or as a ﬁle on a disk) to a processor with a different endianness. In these cases, a program must be used to reorder the bytes in question before attempting to access the received integer values. 7.3. Pointers and arrays As we have seen, an address represents a “location” in the memory. As such, it may be the location of a variable, or the start of a contiguous data zone. In certain languages, arrays – which are contiguous sequences of data of the same type – are simply represented by the address of the start of the memory zone in which they are located. Thus, as we saw in section 6.2, an element in an array is accessed by computing its address from the start of the array, the size of each element and the index. Computation using addresses is often referred to as pointer arithmetic. In C and C++, the addition operation p + n, using a pointer p over a type t and an integer n, is interpreted by the compiler as a “shift of n elements of type t from p” and this is the only permitted arithmetic operation on pointers. For example, if p is declared by int *p, the memory access operation for expression p[5] will occur at address p + 5 * 4, since an integer is made up of 4 bytes. In practice, the bracketed notation p[5] is identical to the dereferencing of the address p + 5. In C, this is written *(p + 5) (as addition is commutative, the notation i[tab], which is semantically equivalent to tab[i] in C, also denotes cell i in the array tab but should be avoided for readability reasons).
226
Concepts and Semantics of Programming Languages 1
7.4. Passing parameters by address In section 4.6, we examined different procedure call mechanisms, deﬁning different modes of communication between the caller and the called procedure. In the case of passing by value, the caller transmits the value to the called procedure: this value is either a constant or a mutable variable, which is local to the procedure and initialized by the transmitted value. In the case of passing by reference, communication between the caller and the called procedure operates in both directions: the address of a mutable variable from the calling environment is transmitted to the procedure, and the procedure can then modify the value of this variable. Box 4.6 describes the possible semantics of these different call mechanisms. In this section, we shall present the use of pointers in passing arguments: this is known as passing an argument by address. This is a technique for implementing passing by reference but also, for certain compilers, a means of optimizing the transmission of large arguments. Several situations in which passing by address is helpful are described below. By deﬁnition, a function returns a single value. However, we may require several values to be returned. One simple example is a function designed to take an array of integers and return the numbers of negative, null and positive values found in the array. In languages with native support for nuples, a single triple value will be returned. Example 7.1 shows the deﬁnition of this function in Python. E XAMPLE 7.1.– Python
def count (t) : vneg = 0 vnul = 0 vpos = 0 i = 0 while i < len (t) : if t[i] < 0 : vneg = vneg + 1 elif t[i] == 0 : vnul = vnul + 1 else : vpos = vpos + 1 i = i + 1 return (vneg, vnul, vpos)
If the chosen language does not support nuples – as in C – another mechanism must be used, such as creating a record type containing three integers and returning a value of this type, as shown in example 7.2.
Pointers and Memory Management
227
E XAMPLE 7.2.– C
struct three_ints { int vneg ; int vnul ; int vpos ; } ; struct three_ints count (int *t, int size_t) { int i ; struct three_ints res = { 0, 0, 0 } ; for (i = 0; i < size_t; i++) { if (i < 0) res.vneg++ ; else if (i == 0) res.vnul++ ; else res.vpos++ ; } return res ; }
The drawback to this solution is that we need to create a “utility” structure for every type of multiple values, which we wish to return. Thus, if another function needs to return two integers and a ﬂoat, a new structure must be created for this purpose and so on, increasing the size of the code. Furthermore, the deﬁnitions of these structures “pollute” the source code and limit maintainability, to the detriment of those deﬁnitions that help to reveal the algorithmic architecture of the program. Another solution is to create global variables dedicated to the function, in which the function writes the values to return. Evidently, a global variable will be needed for each value to return, just as multiple structures were required in the previous approach. This makes the program more complicated to maintain, as we need to ensure that each function writes to its own global variables, and that each calling function reads the results from the right global variables. The solution also limits reuse: the code for a function cannot be reused in another program without also importing the declarations of the associated global variables. These issues can be avoided by passing arguments by address, an approach involving pointers. The function takes arguments representing the addresses where the results to transmit to the caller should be stored. The formal and actual parameters of this function are therefore of pointer type. The actual parameters transmitted by the caller are the addresses of the variables where the results of the function call will be stored. This point is illustrated in example 7.3. E XAMPLE 7.3.– C
void count (int *t, int size_t, int *vneg, int *vnul, int *vpos) { int i ; for (i = 0; i < size_t; i++) { if (t[i] < 0) (*vneg)++ ;
228
Concepts and Semantics of Programming Languages 1 else if (t[i] == 0) (*vnul)++ ; else (*vpos)++ ;
} } void caller () { int t[6] = { 1, 1, 0, 0, 0, 2 } ; int vneg, vnul, vpos ; count (t, 6, &vneg, &vnul, &vpos) ; printf ("Negative: %d Null: %d Positive: %d\n", vneg, vnul, vpos) ; }
The caller function transmits the addresses of the three local variables vneg, vnul and vpos, which will act as “mailboxes” to count. The count function will store the values returned by its computation in these locations. The computation increments the values found at the speciﬁed addresses (use of * in (*vneg)++, (*vnul)++ and (*vpos)++). We have chosen not to use return in this function: all results will be transmitted using the passing by address mechanism. In example 7.3, the count function uses variables passed by address in writeonly mode. This mode is often known as “out”, indicating that the parameters are the output of the function. These arguments may also be used in both reading and writing: this is known as “inout” mode. In this case, the function begins by reading the values located at the addresses in order to carry out a computation; it then overwrites them in order to store the results of the computation. These modes, with a dedicated syntax, for passing arguments by address are also found in Ada, a language that is brieﬂy presented in Volume 2, section 2.2. Consider the function in example 7.4, which multiplies a vector of dimension 2 by a scalar. The resulting vector may be rewritten “over” the argument vector. E XAMPLE 7.4.– C
void mult_vect_scal (float *x, float *y, float n) { *x = *x * n ; *y = *y * n ; }
The function mult_vect_scal must ﬁrst retrieve the value of each component of the vector before multiplying it by the scalar, then writing the result of this product in the original vector. Passing arguments by address is also justiﬁed when the argument to pass to a function is large. Instead of copying the entire contents of the structure to transmit to the called function – which can be costly in terms of both time and memory – we may simply transmit the address of the structure. However, this strategy is not without its risks if the contents of the structure must be left as given by the called function: if this function has access to the memory zone of the caller, there is nothing to prevent it
Pointers and Memory Management
229
from making modiﬁcations. Certain languages include type modiﬁers that can be used to declare a variable or formal parameter as constant. The code in the scope of the declaration prohibits itself any modiﬁcation of the variable following initialization. This modiﬁer is named const in C and C++, in in Ada, and final in Java. In example 7.5, the function f attempts to write to the address designated by i, but this address is deﬁned as a constant. The program is thus rejected during the compilation. E XAMPLE 7.5.– C
void f (const int *i) { *i = *i + 1 ; } $ gcc foo.c c faux.c:5:6: error: readonly variable is not assignable *i = *i + 1 ; ~~ ^ 1 error generated.
7.5. References N OTE.– The notion of reference, as described here, arises from the term used in C++ and Java. It is not the same as the references used in the context of mutable variables in Chapters 2 and 4. Unfortunately, both uses of the term are ﬁrmly established in the context of programming languages, so it is important to be aware of the speciﬁc meaning used in any given language in order to prevent ambiguity. As we saw in section 7.4, arguments can be passed to a function by address, allowing the function to modify the corresponding values in the caller. In C, this is done using pointers, and the called function must read or write to the addresses received as arguments. The notion of a reference produces the same effect without recourse to pointers, simplifying the way the source code is written. 7.5.1. References in C++ In the context of C++, a reference is a variable that is an alias of a preallocated memory zone. References are initialized using an expression that represents a memory zone, accessed by a name (a more precise deﬁnition of the lvalue introduced in section 2.3.2). Such a name can be a variable name, a cell in an array, an object’s attribute, etc., hence it cannot be an arbitrary expression, such as a constant or the result of an intermediate computation. Since the introduction of C++11, however, users can deﬁne references that are initialized by other expressions, known as rvalue references; a description of this mechanism lies outside the scope of this book. Even in older standards, references may be initialized by an rvalue if the reference is declared as
230
Concepts and Semantics of Programming Languages 1
a constant. Readers interested in further details on this topic may read specialized documents about C++. Just as every pointer is a pointer “to a type”, every reference is a reference “to a type”. The syntax used in C++ to declare a reference to a type t is t&. Example 7.6 shows the declaration of a reference, and the effect of a modiﬁcation on the variable to which it refers. E XAMPLE 7.6.– C++
#include int main () { int i = 46 ; int& j = i ; i = i + 1 ; j = j + 1 ; printf ("i = %d j = %d\n", i, j) ; return 0 ; }
The variable j is declared as a reference to int and initialized with the address of i. When i is modiﬁed, j is also modiﬁed and vice versa. During execution, this program will thus display i = 48 j = 48 since two incrementations have been made, one on the identiﬁer i and the other on j. Example 7.7 is a modiﬁed version of example 7.6 in which the reference is replaced by a (constant) pointer. Note the explicit manipulation of the pointer, both in extracting the address of i and in dereferencing j. E XAMPLE 7.7.– C
#include int main () { int i = 46 ; int *j = &i ; i = i + 1 ; *j = *j + 1 ; printf ("i = %d j = %d\n", i, *j) ; return (0) ; }
The program in example 7.8, in which we attempt to initialize the reference j with a constant and the reference k with the value returned by a function call, does not compile: the constant 10 and the integer returned by times2 do not represent allocated memory zones.
Pointers and Memory Management
231
E XAMPLE 7.8.– C++
#include int times2 (int v) { return (2 * v) ; } int main () { int i = 46 ; int& j = 10 ; int& k = times2 (i) ; return 0 ; }
Attempts to compile this program using the g++ compiler will result in an error regarding the initializations of j and k: $ g++ refs2.cpp refs2.cpp:11:8: error: nonconst lvalue reference to type ’int’ cannot bind to a temporary of type ’int’ int& j = 10 ; ^ ~~ refs2.cpp:12:8: error: nonconst lvalue reference to type ’int’ cannot bind to a temporary of type ’int’ int& k = times2 (i) ; ^ ~~~~~~~~~ 2 errors generated.
Using an array to represent a memory zone, it is possible to create a reference to a cell, as shown in example 7.9. E XAMPLE 7.9.– C++
#include int main () { int t[2] = { 12, 21 } ; int& i = t[1] ; i = i + 1 ; printf ("i = %d t[1] = %d\n", i, t[1]) ; return 0 ; }
The variable i is a reference to an integer representing the second cell in the array t. As t denotes a zone in the memory, the second cell will also denote a zone in the memory. Running this program will return the expected result i = 22 t[1] = 22. References can also be used as function parameters in order to “silently” pass a parameter by address, as shown in example 7.10.
232
Concepts and Semantics of Programming Languages 1
E XAMPLE 7.10.– C++
#include void f (int& v) { v = v + 1 ; } int main () { int i = 46 ; f (i) ; printf ("i = %d\n", i) ; return 0 ; }
The formal parameter v of f is a reference to an integer. Thus, during the call f(i), v becomes an alias of i; in other terms, i and v denote the same location in the memory, although the two identiﬁers are not in the same scope. Thus, the incrementation of v in f is visible in i in the caller function, and the program returns i = 47. Given that a reference is an alias, it is entirely possible to create a reference to a reference, or to pass a reference to a function accepting a reference as an argument; the reference always refers to the same initial variable. This is illustrated in example 7.11. E XAMPLE 7.11.– C++
#include void f (int& v) { v = v + 1 ; } int main () { int i = 46 ; int &j = i ; int &k = j ; f (i) ; f (j) ; f (k) ; printf ("i = %d j = %d k = %d\n", i, j, k) ; return (0) ; }
Creating j as a reference to i, and k as a reference to j, identiﬁers j and k are both aliases of i. In the context of calls to f with each of these identiﬁers as arguments, the formal parameter v also becomes an alias of i. Three incrementations are therefore applied, and the program returns i = 49 j = 49 k = 49. The same applies in the context of a function call: the argument passed during the call for a parameter of reference type must designate a memory area. Once again, the purpose of this mechanism is to create aliases for identiﬁers. Thus, example 7.12 will be rejected by the compiler.
Pointers and Memory Management
233
E XAMPLE 7.12.– C++
#include void f (int& v) { v = v + 1 ; } int main () { f (67) ; return 0 ; }
The call to f uses an argument that does not represent a zone in the memory, resulting in an error message during the compilation. $ g++ refs.cpp refs.cpp:8:3: error: no matching function for call to ’f’ f (67) ; ^ refs.cpp:3:6: note: candidate function not viable: expects an lvalue for 1st argument void f (int& v) ^ 1 error generated.
As in the case of passing by address, references may be used to transmit large arguments of a call, avoiding the need to copy them. A constant reference may be declared (in the same way as a constant pointer) to prevent the called function from modifying the memory zone transmitted by the caller function. 7.5.2. References in Java While Java does not include an explicit reference mechanism, some authors consider that the language passes an argument by value if the argument is of a primitive type (integer, boolean or ﬂoat), and by reference if this argument has an object type (including character strings, seen here as objects). This is not strictly true: in reality, Java always passes by value. The difference between a primitive value and an object value is that objects are always dynamically allocated, and are thus represented by pointers. When the argument of a call by value is an object value, the value which is actually passed as the argument is a copy of the pointer. As Java does not use explicit pointers, each time an object is manipulated, it is done silently through its address.
234
Concepts and Semantics of Programming Languages 1
7.6. Memory management In this section, we shall discuss memory allocation and disallocation using manual or automatic approaches. In highlevel languages, data are stored in the memory by the execution process, without direct intervention from the developer. In certain languages, developers also have the option to use lower level features to obtain full or partial control over memory storage. A relatively brief presentation of memory management by the execution process, sometimes with the participation of the programmer, is now discussed. 7.6.1. Memory allocation When a compiler analyzes the deﬁnition of a variable, it assigns an area of memory in which to store the value bound to this variable. This operation is known as memory allocation. For variables, and arrays whose size can be determined at compiletime, the memory requirements are known and the allocation can be planned during the compilation. This is known as static allocation. In example 7.13, the size of variables i and s cannot vary from one execution to the next: their allocation is static. E XAMPLE 7.13.– C
int i = 42 ; char s[20] ;
7.6.1.1. Dynamic allocation In other cases, the quantity of memory required for all possible executions of a program cannot be determined in advance. For example, the designers of the text editor used to write this book had no means of knowing how many characters each text ﬁle would contain, meaning that they had no knowledge of the amount of memory required during the writing process. In other words, memory requirements are unknown at compiletime. The programmer can certainly reserve a “fairly large” amount of memory (for example using a very large array), and say to themselves that “that will be enough”. This solution is unsatisfactory for several reasons. First of all, when an execution uses less memory than expected at the time of allocation, the unused space cannot be used for the execution of another program. Recall that, as a resource, memory is shared by all programs, and its size is necessarily limited. Then, if the default memory allocation is insufﬁcient, all that remains is to modify the source code to request a larger size. This is unacceptable because, on the one hand, a software user
Pointers and Memory Management
235
does not necessarily know (or need to know) how to modify this source code; on the other hand, this source code is not always available, especially when it comes from proprietary and paid software. The execution system (in collaboration with the operating system) must therefore include a mechanism, which allocates memory during program execution in order to satisfy the actual needs of each execution. This mechanism is called dynamic allocation. The result of a dynamic allocation is the start address of the area just allocated, which must be stored in a pointer. A memory allocation may be rejected, for example if there is insufﬁcient space. Usually in this case, the primitive used for memory allocation returns an “illegal” address, often referred to as the NULL pointer (from the name of the constant used to deﬁne this address in many languages). This raises two important points. You have to check whether memory allocation has been successful before attempting to access the returned address. In addition, any pointer in a program is likely to contain (for whatever reason) this illegal NULL value. This value must be detected, because accessing this illegal address interrupts the execution of the program. 7.6.1.2. Allocation for large structures Memory allocation for very large data, e.g. of multiple megabytes, can raise certain issues which differ according to the scope of the binding of the data. Consider the case of a local variable bound to such data. Memory allocation is often done on the execution stack, for example, this is the case in C. As the size of the stack is limited, it can become totally full and overﬂow, resulting in an execution error. This is illustrated in example 7.14, despite the 16 gigabytes of RAM available in the machine. E XAMPLE 7.14.– C
#include #include #define SIZE (10000000L) int main () { int t[SIZE] ; t[0] = 5 ; printf ("%d\n", t[0]) ; return 0 ; }
It is not possible to allocate 10 million integers to the stack, and so executing this program generates a segmentation error.
236
Concepts and Semantics of Programming Languages 1
$ gcc foo.c $ ./a.out Segmentation fault: 11
Now, consider the case of a global variable bound to a very large data element. If this variable is not initialized, most compilers will generate a memory allocation to be performed when loading the executable. If the data are initialized, it will appear in extenso in the code generated by the compiler; this signiﬁcantly increases the size of the executable (and the compilation time), as we see from example 7.15. E XAMPLE 7.15.– C
#include #include #define SIZE (10000000L) int t[SIZE] = { 1 } ; int main () { t[0] = 5 ; printf ("%d\n", t[0]) ; return (0) ; }
The executable generated by example 7.15 contains the sequence of statically allocated bytes, initialized for t, which is added to the code produced for the functions. The displayed size of the binary produced in this case is clearly different to that of the same program using dynamic allocation that is smaller, around only 10 kB. $ gcc foo.c $ ls l a.out rwxrxrx 1 user
user
40005896
8 mar 09:34 a.out
Finally, consider the case of a function, transmitting a local memory zone containing results to its caller, when this result is large (see section 7.4). In example 7.16, f creates a local variable that is an array, applies a treatment to this array and returns the array to the program, which is called f. E XAMPLE 7.16.– C
int* f () { int t[10] ; ... /* Processing. */ return t ; } int* g () { ... f () /* Call to f */ }
Pointers and Memory Management
237
Example 7.16 is perfectly acceptable for languages with automatic memory management systems, or where arrays are native data structures. In languages such as C and C++, however, such a program is incorrect. Since the local variables of a function are allocated on the stack, the local variable t would be “stacked” at the beginning of the execution of f. In practice, this means that the 10 integers making up the array are placed on the stack, and t contains the address (in the stack) of the ﬁrst element of the array. At the end of the execution of f, the local variables of f are removed from the stack. However, the return t instruction does not return the contents of the array t but simply the address of its ﬁrst element in the stack. If the function g that calls f uses this address, there is no guarantee that the current data at this address still has any link with t as the stack may have been modiﬁed in the meantime. These problems can be avoided through the use of dynamic allocation, even if the size of the memory allocation is known at compiletime. In example 7.16, the array t may be allocated dynamically. At the end of the execution of f, the local variables will disappear but the allocated zone remains accessible, as long as it has not been explicitly freed. 7.6.1.3. Dynamic allocation operators Dynamic memory allocation is performed by calling a dedicated function in the language. In C, this function is called malloc, and takes as the argument the number of bytes to allocate. It returns an untyped pointer (void*), which is the start address of the area allocated by the memory allocator. The malloc function also exists in C++, but the more elaborate new operator is often used instead. Unlike malloc, new calls the constructor of the created object (studied in Volume 2, section 4.2). The number of bytes to allocate is not required as the argument by new, since the compiler handles this computation using the type of the object being allocated. Note that, as an operator, new can be overloaded (see section 5.7.2). 7.6.2. Freeing memory When a dynamically allocated memory zone is no longer needed, it must be given back to the system for reuse. In certain languages, the programmer is responsible for this restitution mechanism and must specify points in the execution where some memory zones will no longer be used and can be freed. This form of memory management can be extremely complex in the case of large programs, and different kinds of errors may occur. First of all, the programmer may misjudge when an area is no longer needed and release it too soon. Therefore, the program continues to run instructions using values stored in that zone. But, since this zone has been deallocated, it no longer “belongs” to the running program. If it attempts to access it, the program will be interrupted
238
Concepts and Semantics of Programming Languages 1
for violating memory space. Indeed, the operating system prohibits programs from accessing any areas of the memory which have not been assigned to them. Premature freeing of memory can result in a signiﬁcantly worse version of this error, when the memory area has been reallocated to the program for another use. In this case, the program continues to access this area that belongs to it, so the operating system does not interrupt it. However, the processing using the area released too early and the one allocated this area after this release, read and write their data concurrently in this area, leading to unpredictable data modiﬁcations. This silent data corruption leads to errors that can be difﬁcult to ﬁnd, and correcting the problem is particularly difﬁcult. Memory may also be freed too late, or not at all. In this case, the program “hogs” memory resources for no good reason. The worstcase scenario is that there will not be enough memory available to satisfy other requirements, resulting in an execution error. To a lesser extent, modern operating systems have the ability to virtualize memory: external storage peripherals (hard disks) act as buffers to temporarily free up memory from programs or processes that are not currently using it. The result is an overall slowdown of the machine and a performance collapse, which can lead to more or less total paralysis of the entire system. For memory to be freed, the address of each allocated zone must be saved up until the point of release. If this address is lost, it will no longer be possible to free the zone, resulting in the effects described above. This is known as a memory leak. In example 7.17, the zone allocated by the primitive malloc is not freed at the end of the function f. The address of this zone, memorized in the local variable p, disappears at the end of the execution of f. The address thus becomes unknown and, consequently, it is impossible to free the space. E XAMPLE 7.17.– C
int f (int size) { int *p = malloc (size) ; for (int i = 0; i < size; i++) { do_something (p[i]) ; } return (...) ; }
Other common memory freeing errors include releasing a zone that has not been dynamically allocated, and making multiple attempts to free the same zone. Just like allocation, memory release is achieved by calling a construct of the language. In C, this construct is the free function, whose argument is the start
Pointers and Memory Management
239
address of a zone allocated by malloc. The size to be freed is not to be speciﬁed by the developer, as this information is memorized by the memory manager during allocation. This is all the more important as malloc does not necessarily allocate exactly the requested amount of memory: the actual allocation may be somewhat larger, for reasons we will not go into here. The programmer cannot know exactly how much memory to free up, hence the importance of ofﬂoading this issue. In C++, all memory allocated using new must be freed up using the delete operator. During a call to delete, the destructor of the allocated object (described in Volume 2, section 4.2) is called automatically; this would not be the case using free. As an operator, delete may also be overloaded. 7.6.3. Automatic memory management To overcome the difﬁculty of memory release and the errors it induces, languages such as Java, OCaml, Python and Lisp include a mechanism in their runtime; this automatically retrieves and frees areas of memory that are no longer accessible and are therefore useless. This mechanism, called the garbage collector (GC), increases language safety as users are relieved of the responsibility of memory management. To manage memory automatically, the execution manager needs to know the size of each allocated memory zone. The set of memory blocks (zones) used by the program (known as live blocks) is organized into a memory graph. Its starting points, called roots, are the registers, the execution stack and the global variables. The edges of this graph are the pointers located in the blocks. Any live block can be accessed by transitively following pointers from the roots: this block is still useful at executiontime. The socalled dead blocks are inaccessible from the roots so they can be retrieved. Thus, the essential task of a GC is to browse the memory graph during execution, looking for dead blocks. It must therefore be able to differentiate pointers from other types of data. The efﬁciency of a GC is measured by the extent to which it increases the executiontime, by the amount of memory space it requires and by its ability to free up all the memory that can be freed. It depends on the frequency of allocations and the frequency of assignments within programs. Different GCs use different techniques, none of which is notably better than the others. A basic outline is provided in the following sections. 7.6.3.1. Reference counting GC A reference counting GC associates a counter with each allocated memory block. This counter is incremented each time a pointer to the block is created, and decremented each time a pointer is destroyed. If the counter reaches 0, this means that the block is no longer used and can be freed; this may result in the deletion of pointers to other blocks. If this is the case, the counters of these latter blocks will be
240
Concepts and Semantics of Programming Languages 1
a dat
a
dat
size cnt = 1
size cnt = 0
a dat
size cnt = 2
size cnt
decremented; if any of these counters reach 0, a new pass of the GC is triggered. Pointer management in blocks comes at a cost, which may be signiﬁcant. This GC mechanism is not able to free cyclic data structures, that is which reference themselves, as their counters never reach zero: this is shown in Figure 7.3. This phenomenon results in memory leaks.
a
dat
Figure 7.3. Cycle and reference counter. For a color version of this ﬁgure, see www.iste.co.uk/hardin/programming1.zip
7.6.3.2. Markandsweep GC A markandsweep GC goes through the memory graph, marking live blocks. Unmarked blocks are therefore dead and the GC adds them to the free memory, represented by a linked list of free blocks. Memory
Used /
Unused
NULL Figure 7.4. List of free blocks. For a color version of this ﬁgure, see www.iste.co.uk/hardin/programming1.zip
Markandsweep GCs have the ability to free cyclic structures. However, they are not able to merge consecutive free blocks, as each block is reinserted into the list of free blocks independently. This results in fragmentation of the memory, as shown in Figure 7.5, which can prevent the allocation of a structure that is bigger than any individual free block, even if the overall quantity of free memory is sufﬁcient. 7.6.3.3. Copying GC A copying GC uses two memory zones which are equal in size. Only one of these zones is used to allocate memory at any time. When the quantity of available memory becomes insufﬁcient, the GC goes through the memory graph and copies all
Pointers and Memory Management
241
live blocks, consecutively, into the unused zone. At the end of this process, the previously unused zone contains all live blocks and the GC reverses the roles of the two zones. Block copying prevents memory fragmentation, but the pointers in the copied blocks must be updated. There are two main drawbacks to copying GC: ﬁrst, program execution stops during the GC phase, and second, it doubles the amount of memory needed. Free blocks 1 and 2
2
1 NULL
NULL Not merged
Figure 7.5. Freeing and fragmentation. For a color version of this ﬁgure, see www.iste.co.uk/hardin/programming1.zip
Allocations
Allocations
GC
Heap 1
Heap 2
Heap 1
Heap 2
Figure 7.6. Copying and compacting in a copying GC. For a color version of this ﬁgure, see www.iste.co.uk/hardin/programming1.zip
7.6.3.4. Generational GC Generational GC is based on the empirical observation that recently allocated (socalled young) blocks tend to stay alive for a short period of time. These young blocks mostly store values of temporary variables, of local variables of blocks or functions that disappear once the block or function has been executed. A global variable, which must remain accessible throughout the whole execution, will have its value stored in a block that will “age”. Memory is thus split into several areas containing blocks of similar ages, each referred to as a generation. Allocation is done in the zone containing the youngest blocks. If this zone is not large enough to satisfy an allocation request, the GC scans the graph for the living blocks of this younger generation and copies them into the slightly older generation. Once this older generation becomes saturated by the copying, the GC looks for dead blocks to
242
Concepts and Semantics of Programming Languages 1
eliminate them. In a twogeneration GC, the processings of the young and old generations are often referred to as minor and major GC, respectively. As in the case of copybased GC, memory is automatically compacted. Furthermore, if the language causes a large number of temporary allocations, the minor GC will be able to free large amounts of memory simply by examining the young generation. The execution of generational GC implies recording and managing pointers from old blocks to young blocks, resulting in additional costs during assignment operations. Again, like copying GCs, this approach interrupts the execution of the program; however, the duration of the interruption may be reduced by adjusting the size parameters of the different generations. 7.6.3.5. Conservative GC To add an automatic memory recovery mechanism to a language relying on “manual” allocation/deallocation, a conservative GC can be used. It scans the memory assuming that any sequence of bytes that can represent an address value in the memory is a pointer. It frees all blocks that are deﬁnitely dead. But some blocks are wrongly identiﬁed as alive because they contain a value assimilated by the GC to a pointer. Such a block can indeed be dead but will be considered alive for the duration of the execution. This GC does not allow memory compaction and can lead to memory leaks. To conclude this section, the memory management method is a demarcation point between high and low level languages. Manual memory management is a recognized source of errors due to its complexity. There are very few tools to check this manual management. In this section, we have provided a brief overview of GCs, which partly remedy these difﬁculties. Other more complex solutions, such as concurrent or parallel GCs operating over multiprocessor architectures [DOL 93, DOL 94] will not be addressed here, as they lie outside the scope of this book.
8 Exceptions
8.1. Errors: notiﬁcation and propagation Errors may occur when executing a program. They may result from choices made by the programmer who may specify certain conditions, which must stop the execution. For example, a program to compute the average of a list of ages will not be able to handle a negative age value in the same way as a positive value. Other kinds of errors may occur that are not directly related to the programmer’s intent, for example an attempt to divide by 0 or an attempt to illegally access a ﬁle due to missing permissions. To produce a robust program, any programmer must be aware of these possible errors and should prevent their consequences when they occur. The most basic solution is to include multiple checks throughout a program and to implement a mechanism to highlight incorrect results, from error detection to management. The result obtained from each function call should be checked and if this value indicates that an error occurred during the execution of this call, then an error must be returned. A number of possible approaches for error notiﬁcation and handling are outlined below using a simple pseudolanguage. f () ... if ... then return error ... g () r = f () if r is error then return error ... h () r = g ()
Concepts and Semantics of Programming Languages 1: A Semantical Approach with OCaml and Python, First Edition. Thérèse Hardin, Mathieu Jaume, François Pessaux and Véronique Viguié DonzeauGouge. © ISTE Ltd 2021. Published by ISTE Ltd and John Wiley & Sons, Inc.
244
Concepts and Semantics of Programming Languages 1
if r is error then return error ...
This mechanism, which is entirely reliant on the programmer, presents several drawbacks. Error propagation is explicit in the algorithm. Furthermore, forgetting to check the validity of the result of a function call could go unnoticed. In addition, it must be possible to notify an error in the returned value. This is easy when the codomain of a function is partial: just encode the error in a value outside the codomain. For example, a function computing the absolute value of the difference between its arguments, which are assumed to be strictly positive, can only return positive or null integer values. Hence, errors can be represented using − 1, − 2, etc., as follows: abs_diff_positive_ints (x, y) if x = 0 or y = 0 then return (1) if x < 0 or y < 0 then return (2) if x > y then return (x  y) else return (y  x)
This mechanism for encoding error values becomes impossible if the codomain of the function is total: all the values that can be returned are legal values. In other words, there is no available value to represent an error. In the following (simplistic) example computing the acceleration of a mass subjected to a force, if the mass is null, we do not know what value to return. accel (force, mass) if mass = 0 then return ??? else return (force / mass)
Some languages natively include nuples. In this case, a pair may be returned, the ﬁrst component being a ﬂag stating whether or not this result is meaningful, the other one being the value of the result. If the ﬂag indicates an error, any value may be chosen for the result, as it is known to be meaningless. accel (force, mass) if mass = 0 then return (false, 42) else return (true, (force / mass))
In languages with sum types, the type of returned values can also be modiﬁed to include a notion of validity or invalidity. type result = Error  Valid (float) accel (force, mass) if mass = 0 then return (Error) else return (Valid (force / mass))
For languages without native nuples or sum types, there are three possible solutions.
Exceptions
245
8.1.1. Global variable The ﬁrst solution, which is extremely difﬁcult to maintain, consists of using a global variable to act as a validity ﬂag. var validity accel (force, mass) if mass = 0 then validity = false return 42 else validity = true return (force / mass) f () ... v = accel (...) if validity then ... else ...
As we saw in section 7.4, this solution reduces maintainability and readability. Deﬁning as many global variables as there are functions needing to return errors, or sharing a same global variable between several functions does not ease code review and maintenance. Nevertheless, this mechanism is used in the C standard library to detail errors in certain system calls and functions. The global variable errno is modiﬁed by these functions as a side effect, adding error information. 8.1.2. Record deﬁnition As we saw in section 7.4, even if a language does not natively support nuples, it remains possible to create a twoﬁeld record. However, a record type must be created for every possible type of returned value, so this solution to the error notiﬁcation problem is also unsatisfactory. 8.1.3. Passing by address Some languages, such as C, allow arguments to be passed by address (see section 7.4). Instead of passing a “conventional” value, we may pass the address of a memory zone, for example the address of a variable of the calling function. Thus, the called function will write at this address. In this way, information can be transferred from the called function to the caller, both for errorfree results and for error notiﬁcation, as shown in example 8.1.
246
Concepts and Semantics of Programming Languages 1
E XAMPLE 8.1.– C
float accel (float force, float mass, bool *valid) { if (mass == 0) { *valid = false ; return 42 ; } *valid = true ; return (force / mass) ; }
8.1.4. Introducing exceptions All of the techniques for reporting and propagating errors presented above have the common disadvantage of being explicitly the responsibility of the programmer. Many languages (Ada, Modula, ML, Java, C++, Python, etc.) use a different mechanism, known as exceptions. Exceptions are used to interrupt a computation in order to take account of an error or an abnormal conﬁguration. When a running program encounters a situation of this kind, it raises an exception to warn that a valid result will not be supplied. If the program fails to handle the exception, its execution is immediately stopped. Taking into account an exception consists of explicitly foreseeing in the program the case where such an exception would occur and writing the algorithm to handle this case. This is what is called catching an exception by an exception handler. Raising an exception at a point in the program transfers control to the closest handler for that exception in the program’s call stack, as will be explained in example 8.2. E XAMPLE 8.2.– exception NullMass accel (force, mass) if mass = 0 then raise NullMass return (force / mass) f () ... try { v = accel (...) ... } with x { if x = NullMass then error ("Error, mass is null.") else raise x }
In example 8.2, raise is the construct for raising an exception. If it is raised, the execution of accel is interrupted. The exception will propagate through the chain of
Exceptions
247
function calls that led up to accel, until a dedicated handler is found. There is a handler (the try ... with construct) in the immediate calling function, f, which catches the exception and executes the prescribed treatment. If there had been no handler for NullMass in f, the exception would have propagated to the function that called f, and so on as long as no handler is found, possibly as far as the root function of the program, interrupting it as well, thus killing the execution of the global program. The handler in function f catches a possible exception, which is then bound to the variable x. If the caught exception is not the expected one, the program explicitly raises it again to ensure its detection. As we shall see, many languages automatically handle the raising of an exception if it is not one of the cases expected by the handler. We have chosen here a simpliﬁed construction of the handlers, without any loss of generality, in order to ease the presentation. Exceptions provide a simple and ﬂexible way of handling errors in software. If an exception is not explicitly handled by the call which raised it, it is automatically propagated upstream in the functions calls graph until a function which “knows” how to handle the exception is found. If there is no handler for this exception, then the execution of the program is interrupted, highlighting the existence of a conﬁguration not foreseen in the algorithm. We can see that this mechanism contrasts radically with the one previously used to report errors in the traditional style, where the “impossible values” returned by functions must be encoded and propagated manually. Exceptions can be “entities” of different kinds, depending on the language. Nevertheless, the same two functions – raising and catching exceptions – are always present. 8.2. A simple formalization: MLstyle exceptions In this section, we shall give a formal presentation of the syntax, the execution semantics and the exceptions typechecking for our functional language Exp2 . This presentation is based on the model used by the ML family of languages. We deliberately consider only constant exceptions in order to simplify the presentation. We will brieﬂy discuss the approaches used in certain other languages in section 8.3, highlighting the main differences between these approaches and the model presented here. These differences do not affect the fundamental principle of exceptions. 8.2.1. Abstract syntax Let us extend the abstract syntax of Exp2 given in section 3.1.1. Table 8.1 presents three additional rules, added to those in Table 3.1, to deﬁne constructs relating to exceptions: 1) declaration of a constant exception by its name Exc;
248
Concepts and Semantics of Programming Languages 1
2) raising an exception expression e; 3) catching an exception, binding the identiﬁed exception to the variable x. e ::= . . . previous expressions  exception Exc in e Declaration of an exception  raise e Raising an exception  try e1 with x → e2 Catching an exception Table 8.1. Extension of the syntax of Exp2 to include exceptions
When executing a try/with construct, if the evaluation of e1 raises an exception Exc, then the result of the execution is the evaluation of e2 in which x is bound to Exc. 8.2.2. Values In Exp2 , an exception is a constant value just like an integer in Z, a boolean in IB or a closure in IF. The fact that exceptions are values means that they can be results of evaluating arbitrary expressions, before – possibly – being raised. We must therefore differentiate between a result being an exception value and one being an explicitly raised exception. In the ﬁrst case, no error is reported; the result is simply a value allowing a later error notiﬁcation. In the second case, the result is an actual and characterized error. In section 3.2.2, “error results” (denoted by IE) were “highlighted” errors and not values. In addition to the existing values, we must introduce values for nonraised the set of raised exceptions. exceptions. Let IE be the set of exception values and IE Thus, the set of values in our language is now deﬁned by V = Z ∪ IB ∪ IF ∪ IE. The set containing all possible results of the evaluation of an expression, that is either a value or a raised exception, is now deﬁned as V = V ∪ IE. 8.2.3. Type algebra Since exceptions are values, they must have a type. We have chosen to name this type exn. It extends the type algebra presented in Table 5.1. 8.2.4. Operational semantics Now, let us extend the rules of the bigstep operational semantics using call by value presented in Box 3.3. Intuitively, evaluation results can now be represented by a sum type as follows: type ’a result = Raised of exn  Value of ’a
Exceptions
249
To verify whether a result belongs to the “error” or “value” sets, it sufﬁces to look at the result constructor. When an exception is declared, its existence is recorded in the environment by binding its name to a value. For simplicity’s sake, we shall consider that this value is simply the name of the exception. In the languages of the ML family, the exceptions are only the constructors of a particular sum type that can be extended throughout the program. For this reason, exceptions can be parameterized using arguments, just like any other sum type constructor. For example, an exception declared as exception SyntaxError of (string * int * int)
may indicate the presence of a syntax error by specifying the name of the ﬁle, the line number and the number of the character in the line where the propagated error was detected.
(Edexc )
(Exc, E XC) ⊕ Env e v E nv
E nv e v (Eraise ) (v ∈ IE) exception Exc in e v E nv raise e v
E nv e1 v1 (Etry ) E nv try e1 with x → e2 v1 E nv e1 v 1 (x, v1 ) ⊕ E nv e2 v2 (Ecatch ) E nv try e1 with x → e2 v2
Box 8.1. Operational semantics of exceptions
The rule Edexc is used to declare an exception by binding its name to a new value of the same name in the environment. The rule Eraise describes the evaluation of an error raising expression. The result of this evaluation is not a value, but a raised expression Rules Etry and Ecatch are used for error catching. Rule Etry makes belonging to IE. the hypothesis that the evaluation of the expression e1 in the environment Env does not raise an exception, so the exception handler is not used and the returned value is the one of e1 . Rule Ecatch makes the hypotheses that the evaluation of the expression e1 in the environment Env raises an exception (v1 ) and that this exception v1 is caught by the handler. Thus, the handler expression e2 is executed in Env extended by a binding of x to the intercepted exception value. The environment in which the exception v1 is raised for the ﬁrst time, following the evaluation of an expression e required in order to evaluate e1 , is not recorded in this rule. As mentioned in section 8.1.4, if an exception is not explicitly handled by the call which raised it, then the exception mechanism is in charge to propagate it as a result to calls, ﬁguring into the graph of function calls, until ﬁnding a function which
250
Concepts and Semantics of Programming Languages 1
handles its treatment. Note that this transmission of v1 is described by the subtree of which the conclusion is the left premise of this rule Ecatch ; this subtree is built with the evaluation rules of function applications or operations seen in Chapter 3. 8.2.5. Typing The typing rules presented in Box 5.1 can now be extended by three new rules, as shown in Box 8.2. (Tcatch )
(Exc, ∀ ∅. exn) ⊕ Env e : τ
(Traise )
α fresh
exception Exc in e : τ E nv raise e : α E nv e1 : τ (x, ∀ ∅. exn) ⊕ E nv e2 : τ (Ttry ) E nv try e1 with x → e2 : τ
E nv
Box 8.2. Typing rules for exceptions in Exp2
The rule Tcatch simply extends the environment with the name of the exception (which is also its value) bound to the type exn of all the exceptions. The rule Traise is used to type the raising of an exception. Since this construct raises an error instead of returning a value, it does not have, strictly speaking, a true type. The rule assigns it a polymorphic type denoted by α, which is a fresh type variable. This choice keeps the properties of the typing algorithm and allows an exception to be raised in any type context without type constraints. Thus, if ... then 1 else raise Exc is welltyped, as is if ... then "Ok" else raise Exc. Finally, the rule Ttry requires the type of the result of the expression to be the same, whether or not the exception handler is used. In other words, whether or not the evaluation of e1 ends in an error, the result of the expression must be of the same type. Researchers have extended the type system of an ML language (OCaml) to include static detection of uncaught exceptions [PES 99]. The type algebra in this case is more complex, as it is annotated using exception expressions, and requires the implementation of cyclic types. To obtain more precise analytical results, the type inference requires the identiﬁers of recursive functions to be generalized in order to type their bodies (see section 5.3.3). A heuristic on the size of type terms allows the decidability of the inference process to be maintained. 8.3. Exceptions in other languages In this section, we shall provide a brief overview of the way in which exceptions are handled in different languages. In C++, Python and Java, for example, exceptions are closely linked to the object mechanisms of the language. An exception is deﬁned by a class inheriting from a base “exception” class. Objects in this class may be raised
Exceptions
251
and caught by constructs similar to those presented in section 8.1.4. The option to create exceptions inheriting from other exceptions introduces subtle differences in handlers’ ability to catch raised exceptions. A handler able to catch a given exception will also be able to catch all exceptions inheriting from this exception. Thus, the order of exception matching cases in a handler is signiﬁcant in deﬁning the case which effectively catches an exception. 8.3.1. Exceptions in OCaml The match/with construct presented earlier is a simpliﬁed version of the construct used in OCaml. Instead of just one catching case, any number of cases may be used, as in a pattern matching. Thus, for a handler to catch different exceptions, instead of writing: E XAMPLE 8.3.– OCaml
try e with  x > ( match x with  Ex1 (v) > ...  Ex2 > ... )
we may write, more directly: E XAMPLE 8.4.– OCaml
try e with  Ex1 (v) > ...  Ex2 > ...
8.3.2. Exceptions in Python In example 8.5, the class MyExc inherits directly from the Exception base class in Python and introduces an exception. E XAMPLE 8.5.– Python
class MyExc (Exception) : def what (self) : return ("MyExc") class MyDerExc (MyExc) : def what (self) : return ("MyDerExc")
252
Concepts and Semantics of Programming Languages 1
These two class deﬁnitions introduce two exceptions deﬁned by the developer. MyDerExc inherits from MyExc. These two classes use a what method, returning a
message associated with the expression that they represent. Let us continue to examine example 8.5, deﬁning a handler to catch the exceptions of Exception (after raising them). Python
try : raise (MyExc) except Exception as e : print (e.what ()) try : raise (MyDerExc) except Exception as e : print (e.what ())
Since MyExc and MyDerExc are Exceptions, the handler catches them and displays the message bound to each exception: MyExc MyDerExc
Now, let us deﬁne a twocase handler with the same example; the ﬁrst case catches MyDerExc and the second catches MyExc. We raise the exception MyDerExc. Python
try : raise (MyDerExc) except MyDerExc as e : print ("MyDerExc", e.what ()) except MyExc as e : print ("MyExc", e.what ())
The ﬁrst case in the handler is able to catch MyDerExc; thus, this case is used, and the program displays: MyDerExc MyDerExc
Now, let us invert the order of cases in the handler. Case MyExc is encountered ﬁrst. Python
try : raise (MyDerExc) except MyExc as e : print ("MyExc", e.what ()) except MyDerExc as e : print ("MyDerExc", e.what ())
Exceptions
253
The program now displays: MyExc MyDerExc
showing that the ﬁrst case in the handler was used: this is the case used to catch MyExc but the printed value of e.what() is the name of the catched exception MyDerExc. The inheritance relation means that the exception MyDerExc can be caught by a handler of MyExc. Thus, the second case in the handler is redundant and will never be accessible. 8.3.3. Exceptions in Java Writing the program from example 8.5 in Java results in a compilation error, and the program is rejected, as shown in example 8.6. E XAMPLE 8.6.– Java
class MyExc extends Exception { public String what () { return ("MyExc") ; } } ; class MyDerExc extends MyExc { public String what () { return ("MyDerExc") ; } } ; public class exn_class { public static void main (String[] args) { try { throw new MyExc () ; } catch (Exception e) { e.printStackTrace () ; } try { throw new MyDerExc () ; } catch (Exception e) { e.printStackTrace () ; } try { throw new MyDerExc () ; } catch (MyDerExc e) { System.out.println (e.what ()) ; } catch (MyExc e) { System.out.println (e.what ()) ; } try { throw new MyDerExc () ; } catch (MyExc e) { System.out.println (e.what ()) ; } catch (MyDerExc e) { System.out.println (e.what ()) ; } } }
$ javac exn_class.java exn_class.java:20: exception MyDerExc has already been caught catch (MyDerExc e) { System.out.println (e.what ()) ; } ^ 1 error
The catch occurring in the last line of the source code has been recognized as redundant by the compiler, any exception inheriting from Myexc being catched by a handler of Myexc.
254
Concepts and Semantics of Programming Languages 1
8.3.4. Exceptions in C++ In C++, using example 8.5, a warning is generated during compilation, but the resulting program can be executed and behaves in the same way as it did in Python, as we see from example 8.7. E XAMPLE 8.7.– C
#include #include using namespace std ; class MyExc : public exception { public : virtual const char* what () const throw () { return ("MyExc") ; } } ; class MyDerExc : public MyExc { public : virtual const char* what () const throw () { return ("MyDerExc") ; } } ; class MyExc exc ; class MyDerExc dexc ; int main () { try { throw exc ; } catch (exception& e) { cout