Programming in Standard ML

(D RAFT: V ERSION 1.2 OF 11.02.11.)

Robert Harper
Carnegie Mellon University
Spring Semester, 2011

c
Copyright 2011
by Robert Harper.
All Rights Reserved.

This work is licensed under the Creative Commons

Attribution-Noncommercial-No Derivative Works 3.0 United States
License. To view a copy of this license, visit
http://creativecommons.org/licenses/by-nc-nd/3.0/us/, or send a
letter to Creative Commons, 171 Second Street, Suite 300, San Francisco,
California, 94105, USA.

Preface
This book is an introduction to programming with the Standard ML programming language. It began life as a set of lecture notes for Computer
Science 15212: Principles of Programming, the second semester of the introductory sequence in the undergraduate computer science curriculum at
Carnegie Mellon University. It has subsequently been used in many other
courses at Carnegie Mellon, and at a number of universities around the
world. It is intended to supersede my Introduction to Standard ML, which
has been widely circulated over the last ten years.
Standard ML is a formally defined programming language. The Definition of Standard ML (Revised) is the formal definition of the language. It
is supplemented by the Standard ML Basis Library, which defines a common basis of types that are shared by all implementations of the language.
Commentary on Standard ML discusses some of the decisions that went into
the design of the first version of the language.
There are several implementations of Standard ML available for a wide
variety of hardware and software platforms. The best-known compilers
are Standard ML of New Jersey, MLton, Moscow ML, MLKit, and PolyML.
These are all freely available on the worldwide web. Please refer to The
Standard ML Home Page for up-to-date information on Standard ML and
its implementations.
Numerous people have contributed directly and indirectly to this text.
I am especially grateful to the following people for their helpful comments and suggestions: Brian Adkins, Nels Beckman, Marc Bezem, James
Bostock, Terrence Brannon, Franck van Breugel, Chris Capel, Matthew
William Cox, Karl Crary, Yaakov Eisenberg, Matt Elder, Mike Erdmann,
Matthias Felleisen, Andrei Formiga, Stephen Harris, Nils Jahnig, Joel Jones,
David Koppstein, John Lafferty, Johannes Laire, Flavio Lerda, Daniel R.
Licata, Adrian Moos, Bryce Nichols, Michael Norrish, Arthur J. ODwyer,

Frank Pfenning, Chris Stone, Dave Swasey, Michael Velten, Johan Wallen,
Scott Williams, and Jeannette Wing. Richard C. Cobbe helped with font selection. I am also grateful to the many students of 15-212 who used these
notes and sent in their suggestions over the years.
These notes are a work in progress. Corrections, comments and suggestions are most welcome.

Contents
Preface

ii

Overview

Programming in Standard ML
1.1 A Regular Expression Package . . . . . . . . . . . . . . . . .
1.2 Sample Code . . . . . . . . . . . . . . . . . . . . . . . . . . . .

II
2

The Core Language

13

Types, Values, and Effects

2.1 Evaluation and Execution . .
2.2 The ML Computation Model
2.2.1 Type Checking . . . .
2.2.2 Evaluation . . . . . . .
2.3 Types, Types, Types . . . . . .
2.4 Type Errors . . . . . . . . . . .
2.5 Sample Code . . . . . . . . . .
Declarations
3.1 Variables . . . . . . . . .
3.2 Basic Bindings . . . . . .
3.2.1 Type Bindings . .
3.2.2 Value Bindings .
3.3 Compound Declarations
3.4 Limiting Scope . . . . . .

3
3
12

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

15
15
16
17
19
21
23
23

.
.
.
.
.
.

24
24
25
25
26
27
28

CONTENTS
3.5
3.6
4

vi

Typing and Evaluation . . . . . . . . . . . . . . . . . . . . . . 29

Sample Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

Functions
4.1 Functions as Templates . . . .
4.2 Functions and Application . .
4.3 Binding and Scope, Revisited
4.4 Sample Code . . . . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

Products and Records

5.1 Product Types . . . . . . . . . . . . . . . .
5.1.1 Tuples . . . . . . . . . . . . . . . .
5.1.2 Tuple Patterns . . . . . . . . . . . .
5.2 Record Types . . . . . . . . . . . . . . . .
5.3 Multiple Arguments and Multiple Results
5.4 Sample Code . . . . . . . . . . . . . . . . .
Case Analysis
6.1 Homogeneous and Heterogeneous Types
6.2 Clausal Function Expressions . . . . . . .
6.3 Booleans and Conditionals, Revisited . .
6.4 Exhaustiveness and Redundancy . . . . .
6.5 Sample Code . . . . . . . . . . . . . . . . .
Recursive Functions
7.1 Self-Reference and Recursion
7.2 Iteration . . . . . . . . . . . .
7.3 Inductive Reasoning . . . . .
7.4 Mutual Recursion . . . . . . .
7.5 Sample Code . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

33
33
34
37
39

.
.
.
.
.
.

40
40
40
42
45
48
50

.
.
.
.
.

51
51
52
53
54
56
57
58
61
62
65
66

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

Type Inference and Polymorphism

8.1 Type Inference . . . . . . . . . .
8.2 Polymorphic Definitions . . . .
8.3 Overloading . . . . . . . . . . .
8.4 Sample Code . . . . . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

67
. 67
. 70
. 73
. 76

R EVISED 11.02.11

D RAFT

V ERSION 1.2

CONTENTS
9

vii

Programming with Lists

77
9.1 List Primitives . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
9.2 Computing With Lists . . . . . . . . . . . . . . . . . . . . . . 79
9.3 Sample Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

10 Concrete Data Types

10.1 Datatype Declarations . . . . .
10.2 Non-Recursive Datatypes . . .
10.3 Recursive Datatypes . . . . . .
10.4 Heterogeneous Data Structures
10.5 Abstract Syntax . . . . . . . . .
10.6 Sample Code . . . . . . . . . . .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

82
82
83
85
88
89
91

11 Higher-Order Functions
11.1 Functions as Values .
11.2 Binding and Scope .
11.3 Returning Functions
11.4 Patterns of Control .
11.5 Staging . . . . . . . .
11.6 Sample Code . . . . .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

92
92
93
95
97
99
102

.
.
.
.
.
.

103
104
104
105
107
110
112

.
.
.
.
.
.
.
.

113
113
115
116
118
119
120
122
124

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

12 Exceptions
12.1 Exceptions as Errors . . . . . . .
12.1.1 Primitive Exceptions . . .
12.1.2 User-Defined Exceptions .
12.2 Exception Handlers . . . . . . . .
12.3 Value-Carrying Exceptions . . . .
12.4 Sample Code . . . . . . . . . . . .

.
.
.
.
.
.

.
.
.
.
.
.

13 Mutable Storage
13.1 Reference Cells . . . . . . . . . . . .
13.2 Reference Patterns . . . . . . . . . .
13.3 Identity . . . . . . . . . . . . . . . . .
13.4 Aliasing . . . . . . . . . . . . . . . .
13.5 Programming Well With References
13.5.1 Private Storage . . . . . . . .
13.5.2 Mutable Data Structures . . .
13.6 Mutable Arrays . . . . . . . . . . . .
R EVISED 11.02.11

D RAFT

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.

V ERSION 1.2

CONTENTS

viii

13.7 Sample Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

14 Input/Output
127
14.1 Textual Input/Output . . . . . . . . . . . . . . . . . . . . . . 127
14.2 Sample Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
15 Lazy Data Structures
15.1 Lazy Data Types . . . . . . .
15.2 Lazy Function Definitions .
15.3 Programming with Streams
15.4 Sample Code . . . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

130
132
133
135
137

16 Equality and Equality Types

138
16.1 Sample Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
17 Concurrency
139
17.1 Sample Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

III

The Module Language

140

18 Signatures and Structures

18.1 Signatures . . . . . . . . . . . . .
18.1.1 Basic Signatures . . . . . .
18.1.2 Signature Inheritance . . .
18.2 Structures . . . . . . . . . . . . .
18.2.1 Basic Structures . . . . . .
18.2.2 Long and Short Identifiers
18.3 Sample Code . . . . . . . . . . . .
19 Signature Matching
19.1 Principal Signatures .
19.2 Matching . . . . . . .
19.3 Satisfaction . . . . . .
19.4 Sample Code . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

142
142
143
144
147
147
148
150

.
.
.
.

151
152
153
157
157

20 Signature Ascription
158
20.1 Ascribed Structure Bindings . . . . . . . . . . . . . . . . . . . 158
20.2 Opaque Ascription . . . . . . . . . . . . . . . . . . . . . . . . 159
R EVISED 11.02.11

D RAFT

V ERSION 1.2

CONTENTS

ix

20.3 Transparent Ascription . . . . . . . . . . . . . . . . . . . . . . 162

20.4 Sample Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
21 Module Hierarchies
165
21.1 Substructures . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
21.2 Sample Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
22 Sharing Specifications
174
22.1 Combining Abstractions . . . . . . . . . . . . . . . . . . . . . 174
22.2 Sample Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
23 Parameterization
23.1 Functor Bindings and Applications .
23.2 Functors and Sharing Specifications
23.3 Avoiding Sharing Specifications . . .
23.4 Sample Code . . . . . . . . . . . . . .

IV

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

Programming Techniques

182
182
185
187
191

192

24 Specifications and Correctness

24.1 Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . .
24.2 Correctness Proofs . . . . . . . . . . . . . . . . . . . . . . . .
24.3 Enforcement and Compliance . . . . . . . . . . . . . . . . . .

194
194
196
199

25 Induction and Recursion

202
25.1 Exponentiation . . . . . . . . . . . . . . . . . . . . . . . . . . 202
25.2 The GCD Algorithm . . . . . . . . . . . . . . . . . . . . . . . 207
25.3 Sample Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
26 Structural Induction
26.1 Natural Numbers . . . . . . . . .
26.2 Lists . . . . . . . . . . . . . . . . .
26.3 Trees . . . . . . . . . . . . . . . .
26.4 Generalizations and Limitations
26.5 Abstracting Induction . . . . . .
26.6 Sample Code . . . . . . . . . . . .

R EVISED 11.02.11

D RAFT

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

212
212
214
215
216
217
219

V ERSION 1.2

CONTENTS

27 Proof-Directed Debugging
220
27.1 Regular Expressions and Languages . . . . . . . . . . . . . . 220
27.2 Specifying the Matcher . . . . . . . . . . . . . . . . . . . . . . 222
27.3 Sample Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
28 Persistent and Ephemeral Data Structures
229
28.1 Persistent Queues . . . . . . . . . . . . . . . . . . . . . . . . . 232
28.2 Amortized Analysis . . . . . . . . . . . . . . . . . . . . . . . . 235
28.3 Sample Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
29 Options, Exceptions, and Continuations
29.1 The n-Queens Problem . . . . . . . .
29.2 Solution Using Options . . . . . . . .
29.3 Solution Using Exceptions . . . . . .
29.4 Solution Using Continuations . . . .
29.5 Sample Code . . . . . . . . . . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

239
239
241
242
244
246

30 Higher-Order Functions
30.1 Infinite Sequences . . . . . . . . . . . . .
30.2 Circuit Simulation . . . . . . . . . . . . .
30.3 Regular Expression Matching, Revisited
30.4 Sample Code . . . . . . . . . . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

247
248
251
254
256

31 Memoization
31.1 Cacheing Results . . . . . .
31.2 Laziness . . . . . . . . . . .
31.3 Lazy Data Types in SML/NJ
31.4 Recursive Suspensions . . .
31.5 Sample Code . . . . . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

257
257
259
261
263
264

.
.
.
.
.

265
266
266
268
272
273

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

32 Data Abstraction
32.1 Dictionaries . . . . . . . . . . . . . .
32.2 Binary Search Trees . . . . . . . . . .
32.3 Balanced Binary Search Trees . . . .
32.4 Abstraction vs. Run-Time Checking
32.5 Sample Code . . . . . . . . . . . . . .

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

33 Representation Independence and ADT Correctness

274
33.1 Sample Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
R EVISED 11.02.11

D RAFT

V ERSION 1.2

CONTENTS

xi

34 Modularity and Reuse

275
34.1 Sample Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
35 Dynamic Typing and Dynamic Dispatch
276
35.1 Sample Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276
36 Concurrency
277
36.1 Sample Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277

Appendices

278

A The Standard ML Basis Library

279

B Compilation Management
280
B.1 Overview of CM . . . . . . . . . . . . . . . . . . . . . . . . . . 281
B.2 Building Systems with CM . . . . . . . . . . . . . . . . . . . . 281
B.3 Sample Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
C Sample Programs

R EVISED 11.02.11

282

D RAFT

V ERSION 1.2

Part I
Overview

2
Standard ML is a type-safe programming language that embodies many
innovative ideas in programming language design. It is a statically typed
language, with an extensible type system. It supports polymorphic type
inference, which all but eliminates the burden of specifying types of variables and greatly facilitates code re-use. It provides efficient automatic
storage management for data structures and functions. It encourages functional (effect-free) programming where appropriate, but allows imperative (effect-ful) programming where necessary. It facilitates programming
with recursive and symbolic data structures by supporting the definition
of functions by pattern matching. It features an extensible exception mechanism for handling error conditions and effecting non-local transfers of
control. It provides a richly expressive and flexible module system for
structuring large programs, including mechanisms for enforcing abstraction, imposing hierarchical structure, and building generic modules. It is
portable across platforms and implementations because it has a precise
definition. It provides a portable standard basis library that defines a rich
collection of commonly-used types and routines.
Many implementations go beyond the standard to provide experimental language features, extensive libraries of commonly-used routines, and
useful program development tools. Details can be found with the documentation for your compiler, but heres some of what you may expect.
Most implementations provide an interactive system supporting on-line
program development, including tools for compiling, linking, and analyzing the behavior of programs. A few implementations are batch compilers that rely on the ambient operating system to manage the construction
of large programs from compiled parts. Nearly every compiler generates
native machine code, even when used interactively, but some also generate code for a portable abstract machine. Most implementations support separate compilation and provide tools for managing large systems
and shared libraries. Some implementations provide tools for tracing and
stepping programs; many provide tools for time and space profiling. Most
implementations supplement the standard basis library with a rich collection of handy components such as dictionaries, hash tables, or interfaces to
the ambient operating system. Some implementations support language
extensions such as support for concurrent programming (using messagepassing or locking), richer forms of modularity constructs, and support for
lazy data structures.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 1
Programming in Standard ML
1.1

A Regular Expression Package

To develop a feel for the language and how it is used, let us consider the
implementation of a package for matching strings against regular expressions. Well structure the implementation into two modules, an implementation of regular expressions themselves and an implementation of a
matching algorithm for them.
These two modules are concisely described by the following signatures.
signature REGEXP = sig
datatype regexp =
Zero | One | Char of char |
Plus of regexp * regexp |
Times of regexp * regexp |
Star of regexp
exception SyntaxError of string
val parse : string -> regexp
val format : regexp -> string
end
signature MATCHER = sig
structure RegExp : REGEXP
val accepts : RegExp.regexp -> string -> bool
end
The signature REGEXP describes a module that implements regular expressions. It consists of a description of the abstract syntax of regular expres-

1.1 A Regular Expression Package

sions, together with operations for parsing and unparsing them. The signature MATCHER describes a module that implements a matcher for a given
notion of regular expression. It contains a function accepts that, when
given a regular expression, returns a function that determines whether or
not that expression accepts a given string. Obviously the matcher is dependent on the implementation of regular expressions. This is expressed
by a structure specification that specifies a hierarchical dependence of an implementation of a matcher on an implementation of regular expressions
any implementation of the MATCHER signature must include an implementation of regular expressions as a constituent module. This ensures that
the matcher is self-contained, and does not rely on implicit conventions
for determining which implementation of regular expressions it employs.
The definition of the abstract syntax of regular expressions in the signature REGEXP takes the form of a datatype declaration that is reminiscent
of a context-free grammar, but which abstracts from matters of lexical presentation (such as precedences of operators, parenthesization, conventions
for naming the operators, etc..) The abstract syntax consists of six clauses,
corresponding to the regular expressions 0, 1, a, r1 + r2 , r1 r2 , and r .1 The
functions parse and format specify the parser and unparser for regular
expressions. The parser takes a string as argument and yields a regular
expression; if the string is ill-formed, the parser raises the exception SyntaxError with an associated string describing the source of the error. The
unparser takes a regular expression and yields a string that parses to that
regular expression. In general there are many strings that parse to the
same regular expressions; the unparser generally tries to choose one that
is easiest to read.
The implementation of the matcher consists of two modules: an implementation of regular expressions and an implementation of the matcher
itself. An implementation of a signature is called a structure. The implementation of the matching package consists of two structures, one implementing the signature REGEXP, the other implementing MATCHER. Thus the
overall package is implemented by the following two structure declarations:
structure RegExp :> REGEXP = ...
structure Matcher :> MATCHER = ...
The structure identifier RegExp is bound to an implementation of the REGEXP
1 Some

authors use for 0 and for 1.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

1.1 A Regular Expression Package

signature. Conformance with the signature is ensured by the ascription of

the signature REGEXP to the binding of RegExp using the :> notation.
Not only does this check that the implementation (which has been elided
here) conforms with the requirements of the signature REGEXP, but it also
ensures that subsequent code cannot rely on any properties of the implementation other than those explicitly specified in the signature. This helps
to ensure that modules are kept separate, facilitating subsequent changes
to the code.
Similarly, the structure identifier Matcher is bound to a structure that
implements the matching algorithm in terms of the preceding implementation RegExp of REGEXP. The ascribed signature specifies that the structure
Matcher must conform to the requirements of the signature MATCHER. Notice that the structure Matcher refers to the structure RegExp in its implementation.
Once these structure declarations have been processed, we may use the
package by referring to its components using paths, or long identifiers. The
function Matcher.match has type
Matcher.RegExp.regexp -> string -> bool,
reflecting the fact that it takes a regular expression as implemented within
the package itself and yields a matching function on strings. We may build
a regular expression by applying the parser, Matcher.RegExp.parse, to a
string representing a regular expression, then passing the resulting value
of type Matcher.RegExp.regexp to Matcher.accepts.2
Heres an example of the matcher in action:
val regexp =
Matcher.RegExp.parse "(a+b)*"
val matches =
Matcher.accepts regexp
val ex1 = matches "aabba"
(* yields true *)
val ex2 = matches "abac"
(* yields false *)
2 It

might seem that one can apply Matcher.accepts to the output of RegExp.parse,
since Matcher.RegExp.parse is just RegExp.parse. However, this relationship is not
stated in the interface, so there is a pro forma distinction between the two. See Chapter 22
for more information on the subtle issue of sharing.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

1.1 A Regular Expression Package

The use of long identifiers can get tedious at times. There are two typical methods for alleviating the burden. One is to introduce a synonym for
a long package name. Heres an example:
structure M = Matcher
structure R = M.RegExp
val regexp = R.parse "((a + %).(b + %))*"
val matches = M.accepts regexp
val ex1 = matches "aabba"
val ex2 = matches "abac"
Another is to open the structure, incorporating its bindings into the current environment:
open Matcher Matcher.RegExp
val regexp = parse "(a+b)*"
val matches = accepts regexp
val ex1 = matches "aabba"
val ex2 = matches "abac"
It is advisable to be sparing in the use of open because it is often hard to
anticipate exactly which bindings are incorporated into the environment
by its use.
Now lets look at the internals of the structures RegExp and Matcher.
Heres a birds eye view of RegExp:
structure RegExp :> REGEXP = struct
datatype regexp =
Zero | One | Char of char |
Plus of regexp * regexp |
Times of regexp * regexp |
Star of regexp
.
.
.
fun tokenize s = ...
.
.
.
fun parse s =
let
val (r, s) =
R EVISED 11.02.11

D RAFT

V ERSION 1.2

1.1 A Regular Expression Package

parse rexp (tokenize (String.explode s))

in
case s of
nil => r
| => raise SyntaxError "Bad input."
end
handle LexicalError =>
raise SyntaxError "Bad input."
.
.
.
fun format r =
String.implode (format exp r)
end
The elision indicates that portions of the code have been omitted so that
we can get a high-level view of the structure of the implementation.
The structure RegExp is bracketed by the keywords struct and end.
The type regexp is implemented precisely as specified by the datatype
declaration in the signature REGEXP. The parser is implemented by a function that, when given a string, explodes it into a list of characters, transforms the character list into a list of tokens (abstract symbols representing lexical atoms), and finally parses the resulting list of tokens to obtain
its abstract syntax. If there is remaining input after the parse, or if the
tokenizer encountered an illegal token, an appropriate syntax error is signalled. The formatter is implemented by a function that, when given a
piece of abstract syntax, formats it into a list of characters that are then
imploded to form a string. The parser and formatter work with character lists, rather than strings, because it is easier to process lists incrementally than it is to process strings.
It is interesting to consider in more detail the structure of the parser
since it exemplifies well the use of pattern matching to define functions.
Lets start with the tokenizer, which we present here in toto:
datatype token =
AtSign | Percent | Literal of char |
PlusSign | TimesSign |
Asterisk | LParen | RParen
exception LexicalError
fun tokenize nil = nil
R EVISED 11.02.11

D RAFT

V ERSION 1.2

1.1 A Regular Expression Package

|
|
|
|
|
|
|
|

tokenize (#"+" :: cs) = PlusSign :: tokenize cs

tokenize (#"." :: cs) = TimesSign :: tokenize cs
tokenize (#"*" :: cs) = Asterisk :: tokenize cs
tokenize (#"(" :: cs) = LParen :: tokenize cs
tokenize (#")" :: cs) = RParen :: tokenize cs
tokenize (#"@" :: cs) = AtSign :: tokenize cs
tokenize (#"%" :: cs) = Percent :: tokenize cs
tokenize (#"\\" :: c :: cs) =
Literal c :: tokenize cs
| tokenize (#"\\" :: nil) = raise LexicalError
| tokenize (#" " :: cs) = tokenize cs
| tokenize (c :: cs) = Literal c :: tokenize cs

The symbol @ stands for the empty regular expression and the symbol
% stands for the regular expression accepting only the null string. Concatentation is indicated by ., alternation by +, and iteration by *.
We use a datatype declaration to introduce the type of tokens corresponding to the symbols of the input language. The function tokenize
has type char list -> token list; it transforms a list of characters into
a list of tokens. It is defined by a series of clauses that dispatch on the first
character of the list of characters given as input, yielding a list of tokens.
The correspondence between characters and tokens is relatively straightforward, the only non-trivial case being to admit the use of a backslash
to quote a reserved symbol as a character of input. (More sophisticated
languages have more sophisticated token structures; for example, words
(consecutive sequences of letters) are often regarded as a single token of
input.) Notice that it is quite natural to look ahead in the input stream
in the case of the backslash character, using a pattern that dispatches on
the first two characters (if there are such) of the input, and proceeding accordingly. (It is a lexical error to have a backslash at the end of the input.)
Lets turn to the parser. It is a simple recursive-descent parser implementing the precedence conventions for regular expressions given earlier.
These conventions may be formally specified by the following grammar,
which not only enforces precedence conventions, but also allows for the

R EVISED 11.02.11

D RAFT

V ERSION 1.2

1.1 A Regular Expression Package

use of parenthesization to override them.

rexp
rtrm
rfac
ratm

::=
::=
::=
::=

rtrm | rtrm+rexp
rfac | rfac.rtrm
ratm | ratm*
@ | % | a | (rexp)

The implementation of the parser follows this grammar quite closely.

It consists of four mutually recursive functions, parse rexp, parse rtrm,
parse rfac, and parse ratm. These implement what is known as a recursive descent parser that dispatches on the head of the token list to determine
how to proceed.
fun parse rexp ts =
let
val (r, ts) = parse rtrm ts
in
case ts
of (PlusSign :: ts) =>
let
val (r, ts) = parse rexp ts
in
(Plus (r, r), ts)
end
| => (r, ts)
end
and parse rtrm ts = ...
and parse rfac ts =
let
val (r, ts) = parse ratm ts
in
case ts
of (Asterisk :: ts) => (Star r, ts)
| => (r, ts)
end
and parse ratm nil =
raise SyntaxError ("No atom")
| parse ratm (AtSign :: ts) = (Zero, ts)
| parse ratm (Percent :: ts) = (One, ts)
R EVISED 11.02.11

D RAFT

V ERSION 1.2

1.1 A Regular Expression Package

10

| parse ratm ((Literal c) :: ts) = (Char c, ts)

| parse ratm (LParen :: ts) =
let
val (r, ts) = parse rexp ts
in
case ts
of (RParen :: ts) => (r, ts)
| =>
raise SyntaxError "No close paren"
end
Notice that it is quite simple to implement lookahead using patterns that
inspect the token list for specified tokens. This parser makes no attempt
to recover from syntax errors, but one could imagine doing so, using standard techniques.
This completes the implementation of regular expressions. Now for the
matcher. The matcher proceeds by a recursive analysis of the regular expression. The main difficulty is to account for concatenation to match a
string against the regular expression r1 r2 we must match some initial segment against r1 , then match the corresponding final segment against r2 .
This suggests that we generalize the matcher to one that checks whether
some initial segment of a string matches a given regular expression, then
passes the remaining final segment to a continuation, a function that determines what to do after the initial segment has been successfully matched.
This facilitates implementation of concatentation, but how do we ensure
that at the outermost call the entire string has been matched? We achieve
this by using an initial continuation that checks whether the final segment
is empty.
Heres the code, written as a structure implementing the signature MATCHER:
structure Matcher :> MATCHER =
struct
structure RegExp = RegExp
open RegExp
fun match Zero k = false
| match One cs k = k cs
| match (Char c) cs k =
(case cs of nil => false | c::cs => (c=c) andalso (k cs ))
R EVISED 11.02.11

D RAFT

V ERSION 1.2

1.1 A Regular Expression Package

11

| match (Plus (r1, r2)) cs k =

match r1 cs k orelse match r2 cs k
| match (Times (r1, r2)) cs k =
match r1 cs (fn cs => match r2 cs k)
| match (Star r) cs k =
let
val mstar cs = k cs orelse match r cs mstar
in
mstar cs
end
fun accepts regexp string =
match regexp (String.explode string) (fn nil => true |
end

=> false)

Note that we incorporate the structure RegExp into the structure Matcher,
in accordance with the requirements of the signature. The function accepts
explodes the string into a list of characters (to facilitiate sequential processing of the input), then calls match with an initial continuation that ensures
that the remaining input is empty to determine the result. The type of
match is
RegExp.regexp -> char list ->

(char list -> bool) -> bool.

That is, match takes in succession a regular expression, a list of characters,

and a continuation of type char list -> bool; it yields as result a value
of type bool. This is a fairly complicated type, but notice that nowhere did
we have to write this type in the code! The type inference mechanism of
ML took care of determining what that type must be based on an analysis
of the code itself.
Since match takes a function as argument, it is said to be a higher-order
function. The simplicity of the matcher is due in large measure to the ease
with which we can manipulate functions in ML. Notice that we create a
new, unnamed function to pass as a continuation in the case of concatenation it is the function that matches the second part of the regular expression to the characters remaining after matching an initial segment against
the first part. We use a similar technique to implement matching against
an iterated regular expression we attempt to match the null string, but if
this fails, we match against the regular expression being iterated followed
R EVISED 11.02.11

D RAFT

V ERSION 1.2

1.2 Sample Code

12

by the iteration once again. This neatly captures the zero or more times
interpretation of iteration of a regular expression.
Important: the code given above contains a subtle error. Can
you find it? If not, see chapter 27 for further discussion!
This completes our brief overview of Standard ML. The remainder of
these notes are structured into three parts. The first part is a detailed introduction to the core language, the language in which we write programs in
ML. The second part is concerned with the module language, the means
by which we structure large programs in ML. The third is about programming techniques, methods for building reliable and robust programs. I
hope you enjoy it!

1.2

Sample Code

Here is the complete code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Part II
The Core Language

14
All Standard ML is divided into two parts. The first part, the core
language, comprises the fundamental programming constructs of the language the primitive types and operations, the means of defining and
using functions, mechanisms for definining new types, and so on. The
second part, the module language, comprises the mechanisms for structuring programs into separate units and is described in Part III. Here we
introduce the core language.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 2
Types, Values, and Effects
2.1

Evaluation and Execution

Most familiar programming languages, such as C or Java, are based on an

imperative model of computation. Programs are thought of as specifying
a sequence of commands that modify the memory of the computer. Each
step of execution examines the current contents of memory, performs a
simple computation, modifies the memory, and continues with the next
instruction. The individual commands are executed for their effect on the
memory (which we may take to include both the internal memory and
registers and the external input/output devices). The progress of the computation is controlled by evaluation of expressions, such as boolean tests
or arithmetic operations, that are executed for their value. Conditional
commands branch according to the value of some expression. Many languages maintain a distinction between expressions and commands, but
often (in C, for example) expressions may also modify the memory, so that
even expression evaluation has an effect.
Computation in ML is of a somewhat different nature. The emphasis
in ML is on computation by evaluation of expressions, rather than execution
of commands. The idea of computation is as a generalization of your experience from high school algebra in which you are given a polynomial in a
variable x and are asked to calculate its value at a given value of x. We proceed by plugging in the given value for x, and then, using the rules of
arithmetic, determine the value of the polynomial. The evaluation model
of computation used in ML is based on the same idea, but rather than re-

2.2 The ML Computation Model

16

strict ourselves to arithmetic operations on the reals, we admit a richer

variety of values and a richer variety of primitive operations on them.
The evaluation model of computation enjoys several advantages over
the more familiar imperative model. Because of its close relationship to
mathematics, it is much easier to develop mathematical techniques for
reasoning about the behavior of programs. These techniques are important tools for helping us to ensure that programs work properly without
having to resort to tedious testing and debugging that can only show the
presence of errors, never their absence. Moreover, they provide important
tools for documenting the reasoning that went into the formulation of a
program, making the code easier to understand and maintain.
What is more, the evaluation model subsumes the imperative model
as a special case. Execution of commands for the effect on memory can be
seen as a special case of evaluation of expressions by introducing primitive operations for allocating, accessing, and modifying memory. Rather
than forcing all aspects of computation into the framework of memory
modification, we instead take expression evaluation as the primary notion. Doing so allows us to support imperative programming without destroying the mathematical elegance of the evaluation model for programs
that dont use memory. As we will see, it is quite remarkable how seldom
memory modification is required. Nevertheless, the language provides for
storage-based computation for those few times that it is actually necessary.

2.2

The ML Computation Model

Computation in ML consists of evaluation of expressions. Each expression

has three important characteristics:
It may or may not have a type.
It may or may not have a value.
It may or may not engender an effect.
These characteristics are all that you need to know to compute with an
expression.
The type of an expression is a description of the value it yields, should
it yield a value at all. For example, for an expression to have type int is to
R EVISED 11.02.11

D RAFT

V ERSION 1.2

2.2 The ML Computation Model

17

say that its value (should it have one) is a number, and for an expression
to have type real is to say that its value (if any) is a floating point number.
In general we can think of the type of an expression as a prediction of
the form of the value that it has, should it have one. Every expression is
required to have at least one type; those that do are said to be well-typed.
Those without a type are said to be ill-typed; they are considered ineligible
for evaluation. The type checker determines whether or not an expression
is well-typed, rejecting with an error those that are not.
A well-typed expression is evaluated to determine its value, if indeed
it has one. An expression can fail to have a value because its evaluation
never terminates or because it raises an exception, either because of a runtime fault such as division by zero or because some programmer-defined
condition is signalled during its evaluation. If an expression has a value,
the form of that value is predicted by its type. For example, if an expression evaluates to a value v and its type is bool, then v must be either true
or false; it cannot be, say, 17 or 3.14. The soundness of the type system
ensures the accuracy of the predictions made by the type checker.
Evaluation of an expression might also engender an effect. Effects include such phenomena as raising an exception, modifying memory, performing input or output, or sending a message on the network. It is important to note that the type of an expression says nothing about its possible
effects! An expression of type int might well display a message on the
screen before returning an integer value. This possibility is not accounted
for in the type of the expression, which classifies only its value. For this
reason effects are sometimes called side effects, to stress that they happen
off to the side during evaluation, and are not part of the value of the
expression. We will ignore effects until chapter 13. For the time being we
will assume that all expressions are effect-free, or pure.

2.2.1

Type Checking

What is a type? What types are there? Generally speaking, a type is defined by specifying three things:
a name for the type,
the values of the type, and
the operations that may be performed on values of the type.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

2.2 The ML Computation Model

18

Often the division of labor into values and operations is not completely
clear-cut, but it nevertheless serves as a very useful guideline for describing types.
Lets consider first the type of integers. Its name is int. The values
of type int are the numerals 0, 1, 1, 2, 2, and so on. (Note that negative numbers are written with a prefix tilde, rather than a minus sign!)
Operations on integers include addition, +, subtraction, -, multiplication,
*, quotient, div, and remainder, mod. Arithmetic expressions are formed
in the familiar manner, for example, 3*2+6, governed by the usual rules
of precedence. Parentheses may be used to override the precedence conventions, just as in ordinary mathematical practice. Thus the preceding
expression may be equivalently written as (3*2)+6, but we may also write
3*(2+6) to override the default precedences.
The formation of expressions is governed by a set of typing rules that
define the types of expressions in terms of the types of their constituent expressions (if any). The typing rules are generally quite intuitive since they
are consistent with our experience in mathematics and in other programming languages. In their full generality the rules are somewhat involved,
but we will sneak up on them by first considering only a small fragment
of the language, building up additional machinery as we go along.
Here are some simple arithmetic expressions, written using infix notation for the operations (meaning that the operator comes between the
arguments, as is customary in mathematics):
3
3 + 4
4 div 3
4 mod 3
Each of these expressions is well-formed; in fact, they each have type
int. This is indicated by a typing assertion of the form exp : typ, which
states that the expression exp has the type typ. A typing assertion is said to
be valid iff the expression exp does indeed have the type typ. The following
are all valid typing assertions:
3
3
4
4

: int
+ 4 : int
div 3 : int
mod 3 : int

R EVISED 11.02.11

D RAFT

V ERSION 1.2

2.2 The ML Computation Model

19

Why are these typing assertions valid? In the case of the value 3, it
is an axiom that integer numerals have integer type. What about the expression 3+4? The addition operation takes two arguments, each of which
must have type int. Since both arguments in fact have type int, it follows that the entire expression is of type int. For more complex cases we
reason analogously, for example, deducing that (3+4) div (2+3): int by
observing that (3+4): int and (2+3): int.
The reasoning involved in demonstrating the validity of a typing assertion may be summarized by a typing derivation consisting of a nested
sequence of typing assertions, each justified either by an axiom, or a typing rule for an operation. For example, the validity of the typing assertion
(3+7) div 5 : int is justified by the following derivation:
1. (3+7): int, because
(a) 3 : int because it is an axiom
(b) 7 : int because it is an axiom
(c) the arguments of + must be integers, and the result of + is an
integer
2. 5 : int because it is an axiom
3. the arguments of div must be integers, and the result is an integer
The outermost steps justify the assertion (3+4) div 5 : int by demonstrating that the arguments each have type int. Recursively, the inner
steps justify that (3+4): int.

2.2.2

Evaluation

Evaluation of expressions is defined by a set of evaluation rules that determine how the value of a compound expression is determined as a function
of the values of its constituent expressions (if any). Since the value of an
operator is determined by the values of its arguments, ML is sometimes
said to be a call-by-value language. While this may seem like the only sensible way to define evaluation, we will see in chapter 15 that this need not
be the case some operations may yield a value without evaluating their
arguments. Such operations are sometimes said to be lazy, to distinguish
R EVISED 11.02.11

D RAFT

V ERSION 1.2

2.2 The ML Computation Model

20

them from eager operations that require their arguments to be evaluated

before the operation is performed.
An evaluation assertion has the form expval. This assertion states that
the expression exp has value val. It should be intuitively clear that the
following evaluation assertions are valid.
5 5
2+3 5
(2+3) div (1+4) 1
An evaluation assertion may be justified by an evaluation derivation,
which is similar to a typing derivation. For example, we may justify the
assertion (3+7) div 5 2 by the derivation
1. (3+7) 10 because
(a) 3 3 because it is an axiom
(b) 7 7 because it is an axiom
(c) Adding 3 to 7 yields 10.
2. 5 5 because it is an axiom
3. Dividing 10 by 5 yields 2.
Note that is an axiom that a numeral evaluates to itself; numerals are fullyevaluated expressions, or values. Second, the rules of arithmetic are used
to determine that adding 3 and 7 yields 10.
Not every expression has a value. A simple example is the expression
5 div 0, which is undefined. If you attempt to evaluate this expression
it will incur a run-time error, reflecting the erroneous attempt to find the
number n that, when multiplied by 0, yields 5. The error is expressed
in ML by raising an exception; we will have more to say about exceptions
in chapter 12. Another reason that a well-typed expression might not have
a value is that the attempt to evaluate it leads to an infinite loop. We dont
yet have the machinery in place to define such expressions, but we will
soon see that it is possible for an expression to diverge, or run forever, when
evaluated.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

2.3 Types, Types, Types

2.3

21

Types, Types, Types

What types are there besides the integers? Here are a few useful base types
of ML:
Type name: real
Values: 3.14, 2.17, 0.1E6, . . .
Operations: +, -, *, /, =, <, . . .
Type name: char
Values: #"a", #"b", . . .
Operations: ord,chr,=, <, . . .
Type name: string
Values: "abc", "1234", . . .
Operations: , size, =, <, . . .
Type name: bool
Values: true, false
Operations: if exp then exp1 else exp2
There are many, many (in fact, infinitely many!) others, but these are
enough to get us started. (See Appendix A for a complete description of
the primitive types of ML, including the ones given above.)
Notice that some of the arithmetic operations for real numbers are written the same way as for the corresponding operation on integers. For example, we may write 3.1+2.7 to perform a floating point addition of two
floating point numbers. This is called overloading; the addition operation
is said to be overloaded at the types int and real. In an expression involving addition the type checker tries to resolve which form of addition
(fixed point or floating point) you mean. If the arguments are ints, then
fixed point addition is used; if the arguments are reals, then floating addition is used; otherwise an error is reported.1 Note that ML does not perform any implicit conversions between types! For example, the expression
1 If

the type of the arguments cannot be determined, the type defaults to int.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

2.3 Types, Types, Types

22

3+3.14 is rejected as ill-formed! If you intend floating point addition, you

must write instead real(3)+3.14, which converts the integer 3 to its floating point representation before performing the addition. If, on the other
hand, you intend integer addition, you must write 3+round(3.14), which
converts 3.14 to an integer by rounding before performing the addition.
Finally, note that floating point division is a different operation from
integer quotient! Thus we write 3.1/2.7 for the result of dividing 3.1 by
2.7, which results in a floating point number. We reserve the operator div
for integers, and use / for floating point division.
The conditional expression
if exp then exp1 else exp2
is used to discriminate on a Boolean value. It has type typ if exp has type
bool and both exp1 and exp2 have type typ. Notice that both arms of the
conditional must have the same type! It is evaluated by first evaluating
exp, then proceeding to evaluate either exp1 or exp2 , according to whether
the value of exp is true or false. For example,
if 1<2 then "less" else "greater"
evaluates to "less" since the value of the expression 1<2 is true.
Note that the expression
if 1<2 then 0 else (1 div 0)
evaluates to 0, even though 1 div 0 incurs a run-time error. This is because evaluation of the conditional proceeds either to the then clause or to
the else clause, depending on the outcome of the boolean test. Whichever
clause is evaluated, the other is simply discarded without further consideration.
Although we may, in fact, test equality of two boolean expressions, it
is rarely useful to do so. Beginners often writen conditionals of the form
if exp = true then exp1 else exp2 .
But this is equivalent to the simpler expression
if exp then exp1 else exp2 .
Similarly, rather than write
R EVISED 11.02.11

D RAFT

V ERSION 1.2

2.4 Type Errors

23

if exp = false then exp1 else exp2 ,

it is better to write
if not exp then exp1 else exp2
or, better yet, just
if exp then exp2 else exp1 .

2.4

Type Errors

Now that we have more than one type, we have enough rope to hang
ourselves by forming ill-typed expressions. For example, the following expressions are not well-typed:
size
#"1"
#"2"
3.14

45
+ 1
"1"
+ 2

In each case we are misusing an operator with arguments of the wrong

type.
This raises a natural question: is the following expression well-typed
or not?
if 1<2 then 0 else ("abc"+4)
Since the boolean test will come out true, the else clause will never be
executed, and hence need not be constrained to be well-typed. While this
reasoning is sensible for such a simple example, in general it is impossible
for the type checker to determine the outcome of the boolean test during
type checking. To be safe the type checker assumes the worst and insists
that both clauses of the conditional be well-typed, and in fact have the same
type, to ensure that the conditional expression can be given a type, namely
that of both of its clauses.

2.5

Sample Code

Here is the complete code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 3
Declarations
3.1

Variables

Just as in any other programming language, values may be assigned to

variables, which may then be used in expressions to stand for that value.
However, in sharp contrast to most familiar languages, variables in ML do
not vary! A value may be bound to a variable using a construct called a
value binding. Once a variable is bound to a value, it is bound to it for
life; there is no possibility of changing the binding of a variable once it has
been bound. In this respect variables in ML are more akin to variables in
mathematics than to variables in languages such as C.
A type may also be bound to a type constructor using a type binding.
A bound type constructor stands for the type bound to it, and can never
stand for any other type. For this reason a type binding is sometimes called
a type abbreviation the type constructor stands for the type to which it is
bound.1
A value or type binding introduces a new variable or type constructor, distinct from all others of that class, for use within its range of significance, or scope. Scoping in ML is static, or lexical, meaning that the range
of significance of a variable or type constructor is determined by the program text, not by the order of evaluation of its constituent expressions.
(Languages with dynamic scope adopt the opposite convention.) For the
time being variables and type constructors have global scope, meaning that
1 By

the same token a value binding might also be called a value abbreviation, but for
some reason it never is.

3.2 Basic Bindings

25

the range of significance of the variable or type constructor is the rest

of the program the part that lexically follows the binding. However,
we will soon introduce mechanisms for limiting the scopes of variables or
type constructors to a given expression.

3.2
3.2.1

Basic Bindings
Type Bindings

Any type may be given a name using a type binding. At this stage we have
so few types that it is hard to justify binding type names to identifiers, but
well do it anyway because well need it later. Here are some examples of
type bindings:
type float = real
type count = int and average = real
The first type binding introduces the type constructor float, which subsequently is synonymous with real. The second introduces two type constructors, count and average, which stand for int and real, respectively.
In general a type binding introduces one or more new type constructors simultaneously in the sense that the definitions of the type constructors
may not involve any of the type constructors being defined. Thus a binding such as
type float = real and average = float
is nonsensical (in isolation) since the type constructors float and average
are introduced simultaneously, and hence cannot refer to one another.
The syntax for type bindings is
type tycon1 = typ1
and ...
and tyconn = typn
where each tyconi is a type constructor and each typi is a type expression.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

3.2 Basic Bindings

3.2.2

26

Value Bindings

A value may be given a name using a value binding. Here are some examples:
val m : int = 3+2
val pi : real = 3.14 and e : real = 2.17
The first binding introduces the variable m, specifying its type to be int
and its value to be 5. The second introduces two variables, pi and e, simultaneously, both having type real, and with pi having value 3.14 and
e having value 2.17. Notice that a value binding specifies both the type
and the value of a variable.
The syntax of value bindings is
val var1 : typ1 = exp1
and ...
and varn : typn = expn ,
where each vari is a variable, each typi is a type expression, and each expi
is an expression.
A value binding of the form
val var : typ = exp
is type-checked by ensuring that the expression exp has type typ. If not,
the binding is rejected as ill-formed. If so, the binding is evaluated using
the bind-by-value rule: first exp is evaluated to obtain its value val, then val
is bound to var. If exp does not have a value, then the declaration does not
bind anything to the variable var.
The purpose of a binding is to make a variable available for use within
its scope. In the case of a type binding we may use the type variable introduced by that binding in type expressions occurring within its scope. For
example, in the presence of the type bindings above, we may write
val pi : float = 3.14
since the type constructor float is bound to the type real, the type of the
expression 3.14. Similarly, we may make use of the variable introduced
by a value binding in value expressions occurring within its scope.
Continuing from the preceding binding, we may use the expression
R EVISED 11.02.11

D RAFT

V ERSION 1.2

3.3 Compound Declarations

27

Math.sin pi
to stand for 0.0 (approximately), and we may bind this value to a variable
by writing
val x : float = Math.sin pi
As these examples illustrate, type checking and evaluation are context
dependent in the presence of type and value bindings since we must refer
to these bindings to determine the types and values of expressions. For
example, to determine that the above binding for x is well-formed, we
must consult the binding for pi to determine that it has type float, consult
the binding for float to determine that it is synonymous with real, which
is necessary for the binding of x to have type float.
The rough-and-ready rule for both type-checking and evaluation is that
a bound variable or type constructor is implicitly replaced by its binding
prior to type checking and evaluation. This is sometimes called the substitution principle for bindings. For example, to evaluate the expression
Math.cos x in the scope of the above declarations, we first replace the occurrence of x by its value (approximately 0.0), then compute as before,
yielding (approximately) 1.0. Later on we will have to refine this simple
principle to take account of more sophisticated language features, but it is
useful nonetheless to keep this simple idea in mind.

3.3

Compound Declarations

Bindings may be combined to form declarations. A binding is an atomic

declaration, even though it may introduce many variables simultaneously.
Two declarations may be combined by sequential composition by simply
writing them one after the other, optionally separated by a semicolon.
Thus we may write the declaration
val m : int = 3+2
val n : int = m*m
which binds m to 5 and n to 25. Subsequently, we may evaluate m+n to obtain the value 30. In general a sequential composition of declarations has
the form dec1 . . . decn , where n is at least 2. The scopes of these declarations
R EVISED 11.02.11

D RAFT

V ERSION 1.2

3.4 Limiting Scope

28

are nested within one another: the scope of dec1 includes dec2 , . . . , decn , the
scope of dec2 includes dec3 , . . . , decn , and so on.
One thing to keep in mind is that binding is not assignment. The binding
of a variable never changes; once bound to a value, it is always bound to
that value (within the scope of the binding). However, we may shadow a
binding by introducing a second binding for a variable within the scope
of the first binding. Continuing the above example, we may write
val n : real = 2.17
to introduce a new variable n with both a different type and a different
value than the earlier binding. The new binding eclipses the old one,
which may then be discarded since it is no longer accessible. (Later on, we
will see that in the presence of higher-order functions shadowed bindings
are not always discarded, but are preserved as private data in a closure.
One might say that old bindings never die, they just fade away.)

3.4

Limiting Scope

The scope of a variable or type constructor may be delimited by using let

expressions and local declarations. A let expression has the form
let dec in exp end
where dec is any declaration and exp is any expression. The scope of the
declaration dec is limited to the expression exp. The bindings introduced
by dec are discarded upon completion of evaluation of exp.
Similarly, we may limit the scope of one declaration to another declaration by writing
local dec in dec0 end
The scope of the bindings in dec is limited to the declaration dec0 . After
processing dec0 , the bindings in dec may be discarded.
The value of a let expression is determined by evaluating the declaration part, then evaluating the expression relative to the bindings introduced by the declaration, yielding this value as the overall value of the
let expression. An example will help clarify the idea:

R EVISED 11.02.11

D RAFT

V ERSION 1.2

3.5 Typing and Evaluation

29

let
val m : int = 3
val n : int = m*m
in
m*n
end
This expression has type int and value 27, as you can readily verify by
first calculating the bindings for m and n, then computing the value of m*n
relative to these bindings. The bindings for m and n are local to the expression m*n, and are not accessible from outside the expression.
If the declaration part of a let expression eclipses earlier bindings, the
ambient bindings are restored upon completion of evaluation of the let
expression. Thus the following expression evaluates to 54:
val m : int = 2
val r : int =
let
val m : int = 3
val n : int = m*m
in
m*n
end * m
The binding of m is temporarily overridden during the evaluation of the
let expression, then restored upon completion of this evaluation.

3.5

Typing and Evaluation

To complete this chapter, lets consider in more detail the context-sensitivity

of type checking and evaluation in the presence of bindings. The key ideas
are:
Type checking must take account of the declared type of a variable.
Evaluation must take account of the declared value of a variable.
This is achieved by maintaining environments for type checking and
evaluation. The type environment records the types of variables; the value
R EVISED 11.02.11

D RAFT

V ERSION 1.2

3.5 Typing and Evaluation

30

environment records their values. For example, after processing the compound declaration
val m : int = 0
val x : real = Math.sqrt(2.0)
val c : char = #"a"
the type environment contains the information
val m : int
val x : real
val c : char
and the value environment contains the information
val m = 0
val x = 1.414
val c = #"a"
In a sense the value declarations have been divided in half, separating
the type from the value information.
Thus we see that value bindings have significance for both type checking and evaluation. In contrast type bindings have significance only for
type checking, and hence contribute only to the type environment. A type
binding such as
type float = real
is recorded in its entirety in the type environment, and no change is made
to the value environment. Subsequently, whenever we encounter the type
constructor float in a type expression, it is replaced by real in accordance
with the type binding above.
In chapter 2 we said that a typing assertion has the form exp : typ, and
that an evaluation assertion has the form exp val. While two-place typing
and evaluation assertions are sufficient for closed expressions (those without variables), we must extend these relations to account for open expressions (those with variables). Each must be equipped with an environment
recording information about type constructors and variables introduced
by declarations.
Typing assertions are generalized to have the form
R EVISED 11.02.11

D RAFT

V ERSION 1.2

3.5 Typing and Evaluation

31

typenv ` exp : typ

where typenv is a type environment that records the bindings of type constructors and the types of variables that may occur in exp.2 We may think
of typenv as a sequence of specifications of one of the following two forms:
1. type typvar = typ
2. val var : typ
Note that the second form does not include the binding for var, only its
type!
Evaluation assertions are generalized to have the form
valenv ` exp val
where valenv is a value environment that records the bindings of the variables that may occur in exp. We may think of valenv as a sequence of specifications of the form
val var = val
that bind the value val to the variable var.
Finally, we also need a new assertion, called type equivalence, that determines when two types are equivalent, relative to a type environment.
This is written
typenv ` typ1 typ2
Two types are equivalent iff they are the same when the type constructors
defined in typenv are replaced by their bindings.
The primary use of a type environment is to record the types of the
value variables that are available for use in a given expression. This is
expressed by the following axiom:
. . . val var : typ . . . ` var : typ
2 The

turnstile symbol, `, is simply a punctuation mark separating the type environment from the expression and its type.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

3.6 Sample Code

32

In words, if the specification val var : typ occurs in the type environment,
then we may conclude that the variable var has type typ. This rule glosses
over an important point. In order to account for shadowing we require
that the rightmost specification govern the type of a variable. That way
re-binding of variables with the same name but different types behaves as
expected.
Similarly, the evaluation relation must take account of the value environment. Evaluation of variables is governed by the following axiom:
. . . val var = val . . . ` var val
Here again we assume that the val specification is the rightmost one governing the variable var to ensure that the scoping rules are respected.
The role of the type equivalence assertion is to ensure that type constructors always stand for their bindings. This is expressed by the following axiom:
. . . type typvar = typ . . . ` typvar typ
Once again, the rightmost specification for typvar governs the assertion.

3.6

Sample Code

Here is the complete code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 4
Functions
4.1

Functions as Templates

So far we just have the means to calculate the values of expressions, and to
bind these values to variables for future reference. In this chapter we will
introduce the ability to abstract the data from a calculation, leaving behind
the bare pattern of the calculation. This pattern may then be instantiated
as often as you like so that the calculation may be repeated with specified
data values plugged in.
For example, consider the expression 2*(3+4). The data might be taken
to be the values 2, 3, and 4, leaving behind the pattern  * ( + ),
with holes where the data used to be. We might equally well take the
data to just be 2 and 3, and leave behind the pattern  * ( + 4). Or
we might even regard * and + as the data, leaving 2  (3  4) as the
pattern! What is important is that a complete expression can be recovered
by filling in the holes with chosen data.
Since a pattern can contain many different holes that can be independently instantiated, it is necessary to give names to the holes so that instantiation consists of plugging in a given value for all occurrences of a name
in an expression. These names are, of course, just variables, and instantiation is just the process of substituting a value for all occurrences of a
variable in a given expression. A pattern may therefore be viewed as a
function of the variables that occur within it; the pattern is instantiated by
applying the function to argument values.
This view of functions is similar to our experience from high school

4.2 Functions and Application

34

algebra. In algebra we manipulate polynomials such as x2 + 2x + 1 as

a form of expression denoting a real number, with the variable x representing a fixed, but unknown, quantity. (Indeed, variables in algebra are
sometimes called unknowns, or indeterminates.) It is also possible to think
of a polynomial as a function on the real line: given a real number x, a
polynomial determines a real number y computed by the given combination of arithmetic operations. Indeed, we sometimes write equations
such as f ( x ) = x2 + 2x + 1, to stand for the function f determined by
the polynomial. In the univariate case we can get away with just writing
the polynomial for the function, but in the multivariate case we must be
more careful: we may regard the polynomial x2 + 2xy + y2 to be a function of x, a function of y, or a function of both x and y. In these cases
we write f ( x ) = x2 + 2xy + y2 when x varies and y is held fixed, and
g(y) = x2 + 2xy + y2 when y varies for fixed x, and h( x, y) = x2 + 2xy + y2 ,
when both vary jointly.
In algebra it is usually left implicit that the variables x and y range
over the real numbers, and that f , g, and h are functions on the real line.
However, to be fully explicit, we sometimes write something like
f : R R : x R 7 x2 + 2x + 1
to indicate that f is a function on the reals sending x R to x2 + 2x + 1
R. This notation has the virtue of separating the name of the function, f ,
from the function itself, the mapping that sends x R to x2 + 2x + 1.
It also emphasizes that functions are a kind of value in mathematics
(namely, a certain set of ordered pairs), and that the variable f is bound to
that value (i.e., that set) by the declaration. This viewpoint is especially important once we consider operators, such as the differential operator, that
map functions to functions. For example, if f is a differentiable function
on the real line, the function D f is its first derivative, another function on
the real line. In the case of the function f defined above the function D f
sends x R to 2x + 2.

4.2

Functions and Application

The treatment of functions in ML is very similar, except that we stress

the algorithmic aspects of functions (how they determine values from arguments), as well as the extensional aspects (what they compute). As in
R EVISED 11.02.11

D RAFT

V ERSION 1.2

4.2 Functions and Application

35

mathematics, a function in ML is a kind of value, namely a value of function type of the form typ -> typ0 . The type typ is the domain type (the type of
arguments) of the function, and typ0 is its range type (the type of its results).
We compute with a function by applying it to an argument value of its domain type and calculating the result, a value of its range type. Function
application is indicated by juxtaposition: we simply write the argument
next to the function.
The values of function type consist of primitive functions, such as addition and square root, and function expressions, which are also called lambda
expressions,1 of the form
fn var : typ => exp
The variable var is called the parameter, and the expression exp is called
the body. It has type typ->typ0 provided that exp has type typ0 under the
assumption that the parameter var has the type typ.
To apply such a function expression to an argument value val, we add
the binding
val var = val
to the value environment, and evaluate exp, obtaining a value val0 . Then
the value binding for the parameter is removed, and the result value, val0 ,
is returned as the value of the application.
For example, Math.sqrt is a primitive function of type real->real that
may be applied to a real number to obtain its square root. For example, the
expression Math.sqrt 2.0 evaluates to 1.414 (approximately). We can,
if we wish, parenthesize the argument, writing Math.sqrt (2.0) for the
sake of clarity; this is especially useful for expressions such as Math.sqrt
(Math.sqrt 2.0). The square root function is built in. We may write the
fourth root function as the following function expression:
fn x : real => Math.sqrt (Math.sqrt x)
It may be applied to an argument by writing an expression such as
(fn x : real => Math.sqrt (Math.sqrt x)) (16.0),
1 For

purely historical reasons.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

4.2 Functions and Application

36

which calculates the fourth root of 16.0. The calculation proceeds by binding the variable x to the argument 16.0, then evaluating the expression
Math.sqrt (Math.sqrt x) in the presence of this binding. When evaluation completes, we drop the binding of x from the environment, since it is
no longer needed.
Notice that we did not give the fourth root function a name; it is an
anonymous function. We may give it a name using the declaration
forms introduced in chapter 3. For example, we may bind the fourth root
function to the variable fourthroot using the following declaration:
val fourthroot : real -> real =
fn x : real => Math.sqrt (Math.sqrt x)
We may then write fourthroot 16.0 to compute the fourth root of 16.0.
This notation for defining functions quickly becomes tiresome, so ML
provides a special syntax for function bindings that is more concise and
natural. Instead of using the val binding above to define fourthroot, we
may instead write
fun fourthroot (x:real):real = Math.sqrt (Math.sqrt x)
This declaration has the same meaning as the earlier val binding, namely
it binds fn x:real => Math.sqrt(Math.sqrt x) to the variable fourthroot.
It is important to note that function applications in ML are evaluated
according to the call-by-value rule: the arguments to a function are evaluated before the function is called. Put in other terms, functions are defined
to act on values, rather than on unevaluated expressions. Thus, to evaluate
an expression such as fourthroot (2.0+2.0), we proceed as follows:
1. Evaluate fourthroot to the function value fn x : real => Math.sqrt
(Math.sqrt x).
2. Evaluate the argument 2.0+2.0 to its value 4.0
3. Bind x to the value 4.0.
4. Evaluate Math.sqrt (Math.sqrt x) to 1.414 (approximately).
(a) Evaluate Math.sqrt to a function value (the primitive square
root function).
R EVISED 11.02.11

D RAFT

V ERSION 1.2

4.3 Binding and Scope, Revisited

37

(b) Evaluate the argument expression Math.sqrt x to its value, approximately 2.0.
i. Evaluate Math.sqrt to a function value (the primitive square
root function).
ii. Evaluate x to its value, 4.0.
iii. Compute the square root of 4.0, yielding 2.0.
(c) Compute the square root of 2.0, yielding 1.414.
5. Drop the binding for the variable x.
Notice that we evaluate both the function and argument positions of an
application expression both the function and argument are expressions
yielding values of the appropriate type. The value of the function position
must be a value of function type, either a primitive function or a lambda
expression, and the value of the argument position must be a value of the
domain type of the function. In this case the result value (if any) will be of
the range type of the function. Functions in ML are first-class, meaning that
they may be computed as the value of an expression. We are not limited to
applying only named functions, but rather may compute new functions
on the fly and apply these to arguments. This is a source of considerable
expressive power, as we shall see in the sequel.
Using similar techniques we may define functions with arbitrary domain and range. For example, the following are all valid function declarations:
fun
fun
fun
fun
fun
fun

srev (s:string):string = implode (rev (explode s))

pal (s:string):string = s (srev s)
double (n:int):int = n + n
square (n:int):int = n * n
halve (n:int):int = n div 2
is even (n:int):bool = (n mod 2 = 0)

Thus pal "ot" evaluates to the string "otto", and is even 4 evaluates to
true.

4.3

Binding and Scope, Revisited

A function expression of the form

R EVISED 11.02.11

D RAFT

V ERSION 1.2

4.3 Binding and Scope, Revisited

38

fn var:typ => exp

binds the variable var within the body exp of the function. Unlike val
bindings, function expressions bind a variable without giving it a specific
value. The value of the parameter is only determined when the function
is applied, and then only temporarily, for the duration of the evaluation of
its body.
It is worth reviewing the rules for binding and scope of variables that
we introduced in chapter 3 in the presence of function expressions. As before we adhere to the principle of static scope, according to which variables
are taken to refer to the nearest enclosing binding of that variable, whether
by a val binding or by a fn expression.
Thus, in the following example, the occurrences of x in the body of the
function f refer to the parameter of f, whereas the occurrences of x in the
body of g refer to the preceding val binding.
val x:real = 2.0
fun f(x:real):real = x+x
fun g(y:real):real = x+y
Local val bindings may shadow parameters, as well as other val bindings.
For example, consider the following function declaration:
fun h(x:real):real =
let val x:real = 2.0 in x+x end * x
The inner binding of x by the val declaration shadows the parameter x of
h, but only within the body of the let expression. Thus the last occurrence
of x refers to the parameter of h, whereas the preceding two occurrences
refer to the inner binding of x to 2.0.
The phrases inner and outer binding refer to the logical structure,
or abstract syntax of an expression. In the preceding example, the body
of h lies within the scope of the parameter x, and the expression x+x
lies within the scope of the val binding for x. Since the occurrences of x
within the body of the let lie within the scope of the inner val binding,
they are taken to refer to that binding, rather than to the parameter. On
the other hand the last occurrence of x does not lie within the scope of the
val binding, and hence refers to the parameter of h.
In general the names of parameters do not matter; we can rename them
at will without affecting the meaning of the program, provided that we
R EVISED 11.02.11

D RAFT

V ERSION 1.2

4.4 Sample Code

39

simultaneously (and consistently) rename the binding occurrence and all

uses of that variable. Thus the functions f and g below are completely
equivalent to each other:
fun f(x:int):int = x*x
fun g(y:int):int = y*y
A parameter is just a placeholder; its name is not important.
Our ability to rename parameters is constrained by the static scoping
rule. We may rename a parameter to whatever wed like, provided that
we dont change the way in which uses of a variable are resolved. For
example, consider the following situation:
val x:real = 2.0
fun h(y:real):real = x+y
The parameter y to h may be renamed to z without affecting its meaning.
However, we may not rename it to x, for doing so changes its meaning!
That is, the function
fun h(x:real):real = x+x
does not have the same meaning as h, because now both occurrences of x
in the body of h refer to the parameter, whereas in h the variable x refers
to the outer val binding, whereas the variable y refers to the parameter.
While this may seem like a minor technical issue, it is essential that you
master these concepts now, for they play a central, and rather subtle, role
later on.

4.4

Sample Code

Here is the complete code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 5
Products and Records
5.1

Product Types

A distinguishing feature of ML is that aggregate data structures, such as

tuples, lists, arrays, or trees, may be created and manipulated with ease.
In contrast to most familiar languages it is not necessary in ML to be concerned with allocation and deallocation of data structures, nor with any
particular representation strategy involving, say, pointers or address arithmetic. Instead we may think of data structures as first-class values, on a
par with every other value in the language. Just as it is unnecessary to
think about allocating integers to evaluate an arithmetic expression, it is
unnecessary to think about allocating more complex data structures such
as tuples or lists.

5.1.1

Tuples

This chapter is concerned with the simplest form of aggregate data structure, the n-tuple. An n-tuple is a finite ordered sequence of values of the
form
(val1 ,...,valn ),
where each vali is a value. A 2-tuple is usually called a pair, a 3-tuple a
triple, and so on.
An n-tuple is a value of a product type of the form
typ1 *... *typn .

5.1 Product Types

41

Values of this type are n-tuples of the form

(val1 ,...,valn ),
where vali is a value of type typi (for each 1 i n).
Thus the following are well-formed bindings:
val pair : int * int = (2, 3)
val triple : int * real * string = (2, 2.0, "2")
val quadruple
: int * int * real * real
= (2,3,2.0,3.0)
val pair of pairs
: (int * int) * (real * real)
= ((2,3),(2.0,3.0))
The nesting of parentheses matters! A pair of pairs is not the same as
a quadruple, so the last two bindings are of distinct values with distinct
types.
There are two limiting cases, n = 0 and n = 1, that deserve special
attention. A 0-tuple, which is also known as a null tuple, is the empty
sequence of values, (). It is a value of type unit, which may be thought of
as the 0-tuple type.1 The null tuple type is surprisingly useful, especially
when programming with effects. On the other hand there seems to be no
particular use for 1-tuples, and so they are absent from the language.
As a convenience, ML also provides a general tuple expression of the
form
(exp1 ,...,expn ),
where each expi is an arbitrary expression, not necessarily a value. Tuple
expressions are evaluated from left to right, so that the above tuple expression evaluates to the tuple value
(val1 ,...,valn ),
provided that exp1 evaluates to val1 , exp2 evaluates to val2 , and so on. For
example, the binding
1 In Java (and other languages) the type unit is misleadingly written void, which suggests that the type has no members, but in fact it has exactly one!

R EVISED 11.02.11

D RAFT

V ERSION 1.2

5.1 Product Types

42

val pair : int * int = (1+1, 5-2)

binds the value (2, 3) to the variable pair.
Strictly speaking, it is not essential to have tuple expressions as a primitive notion in the language. Rather than write
(exp1 ,...,expn ),
with the (implicit) understanding that the expi s are evaluated from left to
right, we may instead write
let val x1 = exp1
val x2 = exp2
.
.
.
in

val xn = expn
(x1,...,xn) end

which makes the evaluation order explicit.

5.1.2

Tuple Patterns

One of the most powerful, and distinctive, features of ML is the use of

pattern matching to access components of aggregate data structures. For
example, suppose that val is a value of type
(int * string) * (real * char)
and we wish to retrieve the first component of the second component of
val, a value of type real. Rather than explicitly navigate to this position
to retrieve it, we may simply use a generalized form of value binding in
which we select that component using a pattern:
val (( , ), (r:real, )) = val
The left-hand side of the val binding is a tuple pattern that describes a
pair of pairs, binding the first component of the second component to the
variable r. The underscores indicate dont care positions in the pattern
their values are not bound to any variable. If we wish to give names to
all of the components, we may use the following value binding:
val ((i:int, s:string), (r:real, c:char)) = val
R EVISED 11.02.11

D RAFT

V ERSION 1.2

5.1 Product Types

43

If wed like we can even give names to the first and second components of
the pair, without decomposing them into constituent parts:
val (is:int*string,rc:real*char) = val
The general form of a value binding is
val pat = exp,
where pat is a pattern and exp is an expression. A pattern is one of three
forms:
1. A variable pattern of the form var:typ.
2. A tuple pattern of the form (pat1 ,...,patn ), where each pati is a pattern. This includes as a special case the null-tuple pattern, ().
3. A wildcard pattern of the form .
The type of a pattern is determined by an inductive analysis of the form
of the pattern:
1. A variable pattern var:typ is of type typ.
2. A tuple pattern (pat1 ,...,patn ) has type typ1 * *typn , where each
pati is a pattern of type typi . The null-tuple pattern () has type unit.
3. The wildcard pattern has any type whatsoever.
A value binding of the form
val pat = exp
is well-typed iff pat and exp have the same type; otherwise the binding is
ill-typed and is rejected.
For example, the following bindings are well-typed:
val
val
val
val

(m:int, n:int) = (7+1,4 div 2)

(m:int, r:real, s:string) = (7, 7.0, "7")
((m:int,n:int), (r:real, s:real)) = ((4,5),(3.1,2.7))
(m:int, n:int, r:real, s:real) = (4,5,3.1,2.7)

In contrast, the following are ill-typed:

R EVISED 11.02.11

D RAFT

V ERSION 1.2

5.1 Product Types

44

val (m:int,n:int,r:real,s:real) = ((4,5),(3.1,2.7))

val (m:int, r:real) = (7+1,4 div 2)
val (m:int, r:real) = (7, 7.0, "7")
Value bindings are evaluated using the bind-by-value principle discussed
earlier, except that the binding process is now more complex than before.
First, we evaluate the right-hand side of the binding to a value (if indeed
it has one). This happens regardless of the form of the pattern the righthand side is always evaluated. Second, we perform pattern matching to
determine the bindings for the variables in the pattern.
The process of matching a value against a pattern is defined by a set
of rules for reducing bindings with complex patterns to a set of bindings
with simpler patterns, stopping once we reach a binding with a variable
pattern. The rules are as follows:
1. The variable binding val var = val is irreducible.
2. The wildcard binding val

= val is discarded.

3. The tuple binding

val (pat1 ,...,patn ) =
(val1 ,...,valn )
is reduced to the set of n bindings
.
.
.

val pat1 = val1

val patn = valn
In the case that n = 0 the tuple binding is simply discarded.

These simplifications are repeated until all bindings are irreducible, which
leaves us with a set of variable bindings that constitute the result of pattern
matching.
For example, evaluation of the binding
val ((m:int,n:int), (r:real, s:real)) = ((2,3),(2.0,3.0))
proceeds as follows. First, we compose this binding into the following two
bindings:
R EVISED 11.02.11

D RAFT

V ERSION 1.2

5.2 Record Types

45

val (m:int, n:int) = (2,3)

and (r:real, s:real) = (2.0,3.0).
Then we decompose each of these bindings in turn, resulting in the following set of four atomic bindings:
val
and
and
and

m:int = 2
n:int = 3
r:real = 2.0
s:real = 3.0

At this point the pattern-matching process is complete.

5.2

Record Types

Tuples are most useful when the number of positions is small. When the
number of components grows beyond a small number, it becomes difficult
to remember which position plays which role. In that case it is more natural to attach a label to each component of the tuple that mediates access to
it. This is the notion of a record type.
A record type has the form
{lab1 :typ1 ,...,labn :typn },
where n 0, and all of the labels labi are distinct. A record value has the
form
{lab1 =val1 ,...,labn =valn },
where vali has type typi . A record pattern has the form
{lab1 =pat1 ,...,labn =patn }
which has type
{lab1 :typ1 ,...,labn :typn }
provided that each pati has type typi .
A record value binding of the form

R EVISED 11.02.11

D RAFT

V ERSION 1.2

5.2 Record Types

46

val
{lab1 =pat1 ,...,labn =patn } =
{lab1 =val1 ,...,labn =valn }
is decomposed into the following set of bindings
val pat1 = val1
and ...
and patn = valn .
Since the components of a record are identified by name, not position,
the order in which they occur in a record value or record pattern is not
important. However, in a record expression (in which the components may
not be fully evaluated), the fields are evaluated from left to right in the
order written, just as for tuple expressions.
Here are some examples to help clarify the use of record types. First,
let us define the record type hyperlink as follows:
type hyperlink =
{ protocol : string,
address : string,
display : string }
The record binding
val mailto rwh : hyperlink =
{ protocol="mailto",
address="[email protected]",
display="Robert Harper" }
defines a variable of type hyperlink. The record binding
val { protocol=prot, display=disp, address=addr } = mailto rwh
decomposes into the three variable bindings
val prot = "mailto"
val addr = "[email protected]"
val disp = "Robert Harper"
which extract the values of the fields of mailto rwh.
Using wild cards we can extract selected fields from a record. For example, we may write
R EVISED 11.02.11

D RAFT

V ERSION 1.2

5.2 Record Types

47

val {protocol=prot, address= , display= } = mailto rwh

to bind the variable prot to the protocol field of the record value mailto rwh.
It is quite common to encounter record types with tens of fields. In
such cases even the wild card notation doesnt help much when it comes
to selecting one or two fields from such a record. For this we often use
ellipsis patterns in records, as illustrated by the following example.
val {protocol=prot,...} = intro home
The pattern {protocol=prot,...} stands for the expanded pattern
{protocol=prot, address= , display= }
in which the elided fields are implicitly bound to wildcard patterns.
In general the ellipsis is replaced by as many wildcard bindings as are
necessary to fill out the pattern to be consistent with its type. In order for
this to occur the compiler must be able to determine unambiguously the type of
the record pattern. Here the right-hand side of the value binding determines
the type of the pattern, which then determines which additional fields to
fill in. In some situations the context does not disambiguate, in which case
you must supply additional type information, or avoid the use of ellipsis
notation.
Finally, ML provides a convenient abbreviated form of record pattern
{lab1 ,...,labn }
which stands for the pattern
{lab1 =var1 ,...,labn =varn }
where the variables vari are variables with the same name as the corresponding label labi . For example, the binding
val { protocol, address, display } = mailto rwh
decomposes into the sequence of atomic bindings
val protocol = "mailto"
val address = "[email protected]"
val display = "Robert Harper"
This avoids the need to think up a variable name for each field; we can just
make the label do double duty as a variable.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

5.3 Multiple Arguments and Multiple Results

5.3

48

Multiple Arguments and Multiple Results

A function may bind more than one argument by using a pattern, rather
than a variable, in the argument position. Function expressions are generalized to have the form
fn pat => exp
where pat is a pattern and exp is an expression. Application of such a function proceeds much as before, except that the argument value is matched
against the parameter pattern to determine the bindings of zero or more
variables, which are then used during the evaluation of the body of the
function.
For example, we may make the following definition of the Euclidean
distance function:
val dist
: real * real -> real
= fn (x:real, y:real) => Math.sqrt (x*x + y*y)
This function may then be applied to a pair (a two-tuple!) of arguments to
yield the distance between them. For example, dist (2.0,3.0) evaluates
to (approximately) 4.0.
Using fun notation, the distance function may be defined more concisely as follows:
fun dist (x:real, y:real):real = Math.sqrt (x*x + y*y)
The meaning is the same as the more verbose val binding given earlier.
Keyword parameter passing is supported through the use of record patterns. For example, we may define the distance function using keyword
parameters as follows:
fun dist {x=x:real, y=y:real} = Math.sqrt (x*x + y*y)
The expression dist {x=2.0,y=3.0} invokes this function with the indicated x and y values.
Functions with multiple results may be thought of as functions yielding tuples (or records). For example, we might compute two different notions of distance between two points at once as follows:
R EVISED 11.02.11

D RAFT

V ERSION 1.2

5.3 Multiple Arguments and Multiple Results

49

fun dist2 (x:real, y:real):real*real

= (Math.sqrt (x*x+y*y), Math.abs(x-y))
Notice that the result type is a pair, which may be thought of as two results.
These examples illustrate a pleasing regularity in the design of ML.
Rather than introduce ad hoc notions such as multiple arguments, multiple
results, or keyword parameters, we make use of the general mechanisms
of tuples, records, and pattern matching.
It is sometimes useful to have a function to select a particular component from a tuple or record (e.g., the third component or the component
with a given label). Such functions may be easily defined using pattern
matching. But since they arise so frequently, they are pre-defined in ML
using sharp notation. For any tuple type
typ1 * *typn ,
and each 1 i n, there is a function #i of type
typ1 * *typn ->typi
defined as follows:
fun #i ( , ..., , x, , ..., ) = x
where x occurs in the ith position of the tuple (and there are underscores
in the other n 1 positions).
Thus we may refer to the second field of a three-tuple val by writing
#2(val). It is bad style, however, to over-use the sharp notation; code is
generally clearer and easier to maintain if you use patterns wherever possible. Compare, for example, the following definition of the Euclidean
distance function written using sharp notation with the original.
fun dist (p:real*real):real
= Math.sqrt((#1 p)*(#1 p)+(#2 p)*(#2 p))
You can easily see that this gets out of hand very quickly, leading to unreadable code. Use of the sharp notation is strongly discouraged!
A similar notation is provided for record field selection. The following
function #lab selects the component of a record with label lab.
fun #lab {lab=x,...} = x
Notice the use of ellipsis! Bear in mind the disambiguation requirement:
any use of #lab must be in a context sufficient to determine the full record
type of its argument.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

5.4 Sample Code

5.4

50

Sample Code

Here is the complete code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 6
Case Analysis
6.1

Homogeneous and Heterogeneous Types

Tuple types have the property that all values of that type have the same
form (n-tuples, for some n determined by the type); they are said to be
homogeneous. For example, all values of type int*real are pairs whose
first component is an integer and whose second component is a real. Any
type-correct pattern will match any value of that type; there is no possibility of failure of pattern matching. The pattern (x:int,y:real) is of
type int*real and hence will match any value of that type. On the other
hand the pattern (x:int,y:real,z:string) is of type int*real*string
and cannot be used to match against values of type int*real; attempting
to do so fails at compile time.
Other types have values of more than one form; they are said to be heterogeneous types. For example, a value of type int might be 0, 1, 1, . . . or
a value of type char might be #"a" or #"z". (Other examples of heterogeneous types will arise later on.) Corresponding to each of the values of
these types is a pattern that matches only that value. Attempting to match
any other value against that pattern fails at execution time with an error
condition called a bind failure.
Here are some examples of pattern-matching against values of a heterogeneous type:
val 0 = 1-1
val (0,x) = (1-1, 34)
val (0, #"0") = (2-1, #"0")

6.2 Clausal Function Expressions

52

The first two bindings succeed, the third fails. In the case of the second,
the variable x is bound to 34 after the match. No variables are bound in
the first or third examples.

6.2

Clausal Function Expressions

The importance of constant patterns becomes clearer once we consider

how to define functions over heterogeneous types. This is achieved in
ML using a clausal function expression whose general form is
fn pat1 => exp1
.
| .
.
| patn => expn
Each pati is a pattern and each expi is an expression involving the variables
of the pattern pati . Each component pat=>exp is called a clause, or a rule.
The entire assembly of rules is called a match.
The typing rules for matches ensure consistency of the clauses. Specifically, there must exist types typ1 and typ2 such that
1. Each pattern pati has type typ1 .
2. Each expression expi has type typ2 , given the types of the variables in
pattern pati .
If these requirements are satisfied, the function has the type typ1 ->typ2 .
Application of a clausal function to a value val proceeds by considering
the clauses in the order written. At stage i, where 1 i n, the argument
value val is matched against the pattern pati ; if the pattern match succeeds,
evaluation continues with the evaluation of expression expi , with the variables of pati replaced by their values as determined by pattern matching.
Otherwise we proceed to stage i + 1. If no pattern matches (i.e., we reach
stage n + 1), then the application fails with an execution error called a
match failure.
Heres an example. Consider the following clausal function:
val recip : int -> int =
fn 0 => 0 | n:int => 1 div n
R EVISED 11.02.11

D RAFT

V ERSION 1.2

6.3 Booleans and Conditionals, Revisited

53

This defines an integer-valued reciprocal function on the integers, where

the reciprocal of 0 is arbitrarily defined to be 0. The function has two
clauses, one for the argument 0, the other for non-zero arguments n. (Note
that n is guaranteed to be non-zero because the patterns are considered in
order: we reach the pattern n:int only if the argument fails to match the
pattern 0.)
The fun notation is also generalized so that we may define recip using
the following more concise syntax:
fun recip 0 = 0
| recip (n:int) = 1 div n
One annoying thing to watch out for is that the fun form uses an equal
sign to separate the pattern from the expression in a clause, whereas the
fn form uses a double arrow.
Case analysis on the values of a heterogeneous type is performed by
application of a clausally-defined function. The notation
case
of
|
|

exp
pat1 => exp1
...
patn => expn

is short for the application

(fn pat1 => exp1
| ...
| patn => expn )
exp.
Evaluation proceeds by first evaluating exp, then matching its value successively against the patterns in the match until one succeeds, and continuing with evaluation of the corresponding expression. The case expression fails if no pattern succeeds to match the value.

6.3

Booleans and Conditionals, Revisited

The type bool of booleans is perhaps the most basic example of a heterogeneous type. Its values are true and false. Functions may be defined
R EVISED 11.02.11

D RAFT

V ERSION 1.2

6.4 Exhaustiveness and Redundancy

54

on booleans using clausal definitions that match against the patterns true
and false.
For example, the negation function may be defined clausally as follows:
fun not true = false
| not false = true
The conditional expression
if exp then exp1 else exp2
is short-hand for the case analysis
case exp
of true => exp1
| false => exp2
which is itself short-hand for the application
(fn true => exp1 | false => exp2 ) exp.
The short-circuit conjunction and disjunction operations are defined
as follows. The expression exp1 andalso exp2 is short for
if exp1 then exp2 else false
and the expression exp1 orelse exp2 is short for
if exp1 then true else exp2 .
You should expand these into case expressions and check that they behave
as expected. Pay particular attention to the evaluation order, and observe
that the call-by-value principle is not violated by these expressions.

6.4

Exhaustiveness and Redundancy

Matches are subject to two forms of sanity check as an aid to the ML

programmer. The first, called exhaustiveness checking, ensures that a wellformed match covers its domain type in the sense that every value of the

R EVISED 11.02.11

D RAFT

V ERSION 1.2

6.4 Exhaustiveness and Redundancy

55

domain must match one of its clauses. The second, called redundancy checking, ensures that no clause of a match is subsumed by the clauses that precede it. This means that the set of values covered by a clause in a match
must not be contained entirely within the set of values covered by the preceding clauses of that match.
Redundant clauses are always a mistake such a clause can never be
executed. Redundant rules often arise accidentally. For example, the second rule of the following clausal function definition is redundant:
fun not True = false
| not False = true
By capitalizing True we have turned it into a variable, rather than a constant pattern. Consequently, every value matches the first clause, rendering
the second redundant.
Since the clauses of a match are considered in the order they are written, redundancy checking is correspondingly order-sensitive. In particular, changing the order of clauses in a well-formed, irredundant match can
make it redundant, as in the following example:
fun recip (n:int) = 1 div n
| recip 0 = 0
The second clause is redundant because the first matches any integer value,
including 0.
Inexhaustive matches may or may not be in error, depending on whether
the match might ever be applied to a value that is not covered by any
clause. Here is an example of a function with an inexhaustive match that
is plausibly in error:
fun is numeric #"0" =
| is numeric #"1"
| is numeric #"2"
| is numeric #"3"
| is numeric #"4"
| is numeric #"5"
| is numeric #"6"
| is numeric #"7"
| is numeric #"8"
| is numeric #"9"
R EVISED 11.02.11

true
= true
= true
= true
= true
= true
= true
= true
= true
= true
D RAFT

V ERSION 1.2

6.5 Sample Code

56

When applied to, say, #"a", this function fails. Indeed, the function never
returns false for any argument!
Perhaps what was intended here is to include a catch-all clause at the
end:
fun is numeric #"0" = true
| is numeric #"1" = true
| is numeric #"2" = true
| is numeric #"3" = true
| is numeric #"4" = true
| is numeric #"5" = true
| is numeric #"6" = true
| is numeric #"7" = true
| is numeric #"8" = true
| is numeric #"9" = true
| is numeric = false
The addition of a final catch-all clause renders the match exhaustive, because any value not matched by the first ten clauses will surely be matched
by the eleventh.
Having said that, it is a very bad idea to simply add a catch-all clause
to the end of every match to suppress inexhaustiveness warnings from the
compiler. The exhaustiveness checker is your friend! Each such warning is
a suggestion to double-check that match to be sure that youve not made
a silly error of omission, but rather have intentionally left out cases that
are ruled out by the invariants of the program. In chapter 10 we will see
that the exhaustiveness checker is an extremely valuable tool for managing
code evolution.

6.5

Sample Code

Here is the complete code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 7
Recursive Functions
So far weve only considered very simple functions (such as the reciprocal
function) whose value is computed by a simple composition of primitive
functions. In this chapter we introduce recursive functions, the principal
means of iterative computation in ML. Informally, a recursive function is
one that computes the result of a call by possibly making further calls to
itself. Obviously, to avoid infinite regress, some calls must return their
results without making any recursive calls. Those that do must ensure
that the arguments are, in some sense, smaller so that the process will
eventually terminate.
This informal description obscures a central point, namely the means
by which we may convince ourselves that a function computes the result
that we intend. In general we must prove that for all inputs of the domain type, the body of the function computes the correct value of result
type. Usually the argument imposes some additional assumptions on the
inputs, called the pre-conditions. The correctness requirement for the result is called a post-condition. Our burden is to prove that for every input
satisfying the pre-conditions, the body evaluates to a result satisfying the
post-condition. In fact we may carry out such an analysis for many different pre- and post-condition pairs, according to our interest. For example,
the ML type checker proves that the body of a function yields a value of
the range type (if it terminates) whenever it is given an argument of the
domain type. Here the domain type is the pre-condition, and the range
type is the post-condition. In most cases we are interested in deeper properties, examples of which we shall consider below.
To prove the correctness of a recursive function (with respect to given

7.1 Self-Reference and Recursion

58

pre- and post-conditions) it is typically necessary to use some form of inductive reasoning. The base cases of the induction correspond to those
cases that make no recursive calls; the inductive step corresponds to those
that do. The beauty of inductive reasoning is that we may assume that the
recursive calls work correctly when showing that a case involving recursive calls is correct. We must separately show that the base cases satisfy
the given pre- and post-conditions. Taken together, these two steps are
sufficient to establish the correctness of the function itself, by appeal to an
induction principle that justifies the particular pattern of recursion.
No doubt this all sounds fairly theoretical. The point of this chapter is
to show that it is also profoundly practical.

7.1

Self-Reference and Recursion

In order for a function to call itself, it must have a name by which it can
refer to itself. This is achieved by using a recursive value binding, which are
ordinary value bindings qualified by the keyword rec. The simplest form
of a recursive value binding is as follows:
val rec var:typ = val.
As in the non-recursive case, the left-hand is a pattern, but here the righthand side must be a value. In fact the right-hand side must be a function
expression, since only functions may be defined recursively in ML. The
function may refer to itself by using the variable var.
Heres an example of a recursive value binding:
val rec factorial : int->int =
fn 0 => 1 | n:int => n * factorial (n-1)
Using fun notation we may write the definition of factorial much more
clearly and concisely as follows:
fun factorial 0 = 1
| factorial (n:int) = n * factorial (n-1)
There is obviously a close correspondence between this formulation of
factorial and the usual textbook definition of the factorial function in

R EVISED 11.02.11

D RAFT

V ERSION 1.2

7.1 Self-Reference and Recursion

59

terms of recursion equations:

0! = 1
n! = n (n 1)!

( n > 0)

Recursive value bindings are type-checked in a manner that may, at

first glance, seem paradoxical. To check that the binding
val rec var : typ = val
is well-formed, we ensure that the value val has type typ, assuming that
var has type typ. Since var refers to the value val itself, we are in effect
assuming what we intend to prove while proving it!
(Incidentally, since val is required to be a function expression, the type
typ will always be a function type.)
Lets look at an example. To check that the binding for factorial
given above is well-formed, we assume that the variable factorial has
type int->int, then check that its definition, the function
fn 0 => 1 | n:int => n * factorial (n-1),
has type int->int. To do so we must check that each clause has type
int->int by checking for each clause that its pattern has type int and
that its expression has type int. This is clearly true for the first clause of
the definition. For the second, we assume that n has type int, then check
that n * factorial (n-1) has type int. This is so because of the rules for
the primitive arithmetic operations and because of our assumption that
factorial has type int->int.
How are applications of recursive functions evaluated? The rules are
almost the same as before, with one modification. We must arrange that
all occurrences of the variable standing for the function are replaced by
the function itself before we evaluate the body. That way all references
to the variable standing for the function itself are indeed references to the
function itself!
Suppose that we have the following recursive function binding
val rec var : typ =
fn pat1 => exp1
| ...
| patn => expn
R EVISED 11.02.11

D RAFT

V ERSION 1.2

7.1 Self-Reference and Recursion

60

and we wish to apply var to the value val of type typ. As before, we consider each clause in turn, until we find the first pattern pati matching val.
We proceed, as before, by evaluating expi , replacing the variables in pati by
the bindings determined by pattern matching, but, in addition, we replace
all occurrences of the var by its binding in expi before continuing evaluation.
For example, to evaluate factorial 3, we proceed by retrieving the
binding of factorial and evaluating
(fn 0=>1 | n:int => n*factorial(n-1))(3).
Considering each clause in turn, we find that the first doesnt match, but
the second does. We therefore continue by evaluating its right-hand side,
the expression n * factorial(n-1), after replacing n by 3 and factorial
by its definition. We are left with the sub-problem of evaluating the expression
3 * (fn 0 => 1 | n:int => n*factorial(n-1))(2)
Proceeding as before, we reduce this to the sub-problem of evaluating
3 * (2 * (fn 0=>1 | n:int => n*factorial(n-1))(1)),
which reduces to the sub-problem of evaluating
3 * (2 * (1 * (fn 0=>1 | n:int => n*factorial(n-1))(0))),
which reduces to
3 * (2 * (1 * 1)),
which then evaluates to 6, as desired.
Observe that the repeated substitution of factorial by its definition
ensures that the recursive calls really do refer to the factorial function itself.
Also observe that the size of the sub-problems grows until there are no
more recursive calls, at which point the computation can complete. In
broad outline, the computation proceeds as follows:
1. factorial 3
2. 3 * factorial 2
R EVISED 11.02.11

D RAFT

V ERSION 1.2

7.2 Iteration

61

3. 3 * 2 * factorial 1
4. 3 * 2 * 1 * factorial 0
5. 3 * 2 * 1 * 1
6. 3 * 2 * 1
7. 3 * 2
8. 6
Notice that the size of the expression first grows (in direct proportion to
the argument), then shrinks as the pending multiplications are completed.
This growth in expression size corresponds directly to a growth in runtime storage required to record the state of the pending computation.

7.2

Iteration

The definition of factorial given above should be contrasted with the

following two-part definition:
fun helper (0,r:int) = r
| helper (n:int,r:int) = helper (n-1,n*r)
fun factorial (n:int) = helper (n, 1)
First we define a helper function that takes two parameters, an integer
argument and an accumulator that records the running partial result of the
computation. The idea is that the accumulator re-associates the pending
multiplications in the evaluation trace given above so that they can be performed prior to the recursive call, rather than after it completes. This reduces the space required to keep track of those pending steps. Second, we
define factorial by calling helper with argument n and initial accumulator value 1, corresponding to the product of zero terms (empty prefix).
As a matter of programming style, it is usual to conceal the definitions
of helper functions using a local declaration. In practice we would make
the following definition of the iterative version of factorial:

R EVISED 11.02.11

D RAFT

V ERSION 1.2

7.3 Inductive Reasoning

62

local
fun helper (0,r:int) = r
| helper (n:int,r:int) = helper (n-1,n*r)
in
fun factorial (n:int) = helper (n,1)
end
This way the helper function is not visible, only the function of interest is
exported by the declaration.
The important thing to observe about helper is that it is iterative, or tail
recursive, meaning that the recursive call is the last step of evaluation of an
application of it to an argument. This means that the evaluation trace of a
call to helper with arguments (3,1) has the following general form:
1. helper (3, 1)
2. helper (2, 3)
3. helper (1, 6)
4. helper (0, 6)
5. 6
Notice that there is no growth in the size of the expression because there
are no pending computations to be resumed upon completion of the recursive call. Consequently, there is no growth in the space required for an
application, in contrast to the first definition given above. Tail recursive
definitions are analogous to loops in imperative languages: they merely
iterate a computation, without requiring auxiliary storage.

7.3

Inductive Reasoning

Time and space usage are important, but what is more important is that
the function compute the intended result. The key to the correctness of
a recursive function is an inductive argument establishing its correctness.
The critical ingredients are these:
1. An input-output specification of the intended behavior stating pre-conditions
on the arguments and a post-condition on the result.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

7.3 Inductive Reasoning

63

2. A proof that the specification holds for each clause of the function,
assuming that it holds for any recursive calls.
3. An induction principle that justifies the correctness of the function as
a whole, given the correctness of its clauses.
Well illustrate the use of inductive reasoning by a graduated series
of examples. First consider the simple, non-tail recursive definition of
factorial given in section 7.1. One reasonable specification for factorial
is as follows:
1. Pre-condition: n 0.
2. Post-condition: factorial n evaluates to n!.
We are to establish the following statement of correctness of factorial:
if n 0, then factorial n evaluates to n!.
That is, we show that the pre-conditions imply the post-condition holds
for the result of any application. This is called a total correctness assertion
because it states not only that the post-condition holds of any result of
application, but, moreover, that every application in fact yields a result
(subject to the pre-condition on the argument).
In contrast, a partial correctness assertion does not insist on termination,
only that the post-condition holds whenever the application terminates.
This may be stated as the assertion
if n 0 and factorial n evaluates to p, then p = n!.
Notice that this statement is true of a function that diverges whenever it is
applied! In this sense a partial correctness assertion is weaker than a total
correctness assertion.
Let us establish the total correctness of factorial using the pre- and
post-conditions stated above. To do so, we apply the principle of mathematical induction on the argument n. Recall that this means we are to
establish the specification for the case n = 0, and, assuming it to hold for
n >= 0, show that it holds for n + 1. The base case, n = 0, is trivial:
by definition factorial n evaluates to 1, which is 0!. Now suppose that
n = m + 1 for some m >= 0. By the inductive hypothesis we have that
R EVISED 11.02.11

D RAFT

V ERSION 1.2

7.3 Inductive Reasoning

64

factorial m evaluates to m! (since m 0), and so by definition factorial

n evaluates to
n m! = (m + 1) m!
= ( m + 1) !
= n!,
as required. This completes the proof.
That was easy. What about the iterative definition of factorial? We
focus on the behavior of helper. A suitable specification is given as follows:
1. Pre-condition: n 0.
2. Post-condition: helper (n, r) evaluates to n! r.
To show the total correctness of helper with respect to this specification,
we once again proceed by mathematical induction on n. We leave it as an
exercise to give the details of the proof.
With this in hand it is easy to prove the correctness of factorial if
n 0 then factorial n evaluates to the result of helper (n, 1), which
evaluates to n! 1 = n!. This completes the proof.
Helper functions correspond to lemmas, main functions correspond to
theorems. Just as we use lemmas to help us prove theorems, we use helper
functions to help us define main functions. The foregoing argument shows
that this is more than an analogy, but lies at the heart of good programming style.
Heres an example of a function defined by complete induction (or strong
induction), the Fibonacci function, defined on integers n >= 0:
(* for n>=0, fib n yields the nth Fibonacci number *)
fun fib 0 = 1
| fib 1 = 1
| fib (n:int) = fib (n-1) + fib (n-2)
The recursive calls are made not only on n-1, but also n-2, which is why
we must appeal to complete induction to justify the definition. This definition of fib is very inefficient because it performs many redundant computations: to compute fib n requires that we compute fib (n-1) and fib
(n-2). To compute fib (n-1) requires that we compute fib (n-2) a second time, and fib (n-3). Computing fib (n-2) requires computing fib
R EVISED 11.02.11

D RAFT

V ERSION 1.2

7.4 Mutual Recursion

65

(n-3) again, and fib (n-4). As you can see, there is considerable redundancy here. It can be shown that the running time of fib is exponential in
its argument, which is quite awful.
Heres a better solution: for each n >= 0 compute not only the nth
Fibonacci number, but also the (n 1)st as well. (For n = 0 we define the
1st Fibonacci number to be zero). That way we can avoid redundant
recomputation, resulting in a linear-time algorithm. Heres the code:
(* for n>=0, fib n evaluates to (a, b), where
a is the nth Fibonacci number, and
b is the (n-1)st *)
fun fib 0 = (1, 0)
| fib 1 = (1, 1)
| fib (n:int) =
let
val (a:int, b:int) = fib (n-1)
in
(a+b, a)
end
You might feel satisfied with this solution since it runs in time linear in
n. It turns out (see Graham, Knuth, and Patashnik, Concrete Mathematics
(Addison-Wesley 1989) for a derivation) that the recurrence
F0 = 1
F1 = 1
Fn = Fn1 + Fn2
has a closed-form solution over the real numbers. This means that the
nth Fibonacci number can be calculated directly, without recursion, by using floating point arithmetic. However, this is an unusual case. In most
instances recursively-defined functions have no known closed-form solution, so that some form of iteration is inevitable.

7.4

Mutual Recursion

It is often useful to define two functions simultaneously, each of which

calls the other (and possibly itself) to compute its result. Such functions
R EVISED 11.02.11

D RAFT

V ERSION 1.2

7.5 Sample Code

66

are said to be mutually recursive. Heres a simple example to illustrate the

point, namely testing whether a natural number is odd or even. The most
obvious approach is to test whether the number is congruent to 0 mod
2, and indeed this is what one would do in practice. But to illustrate the
idea of mutual recursion we instead use the following inductive characterization: 0 is even, and not odd; n > 0 is even iff n 1 is odd; n > 0 is
odd iff n 1 is even. This may be coded up using two mutually-recursive
procedures as follows:
fun
|
and
|

even 0 = true
even n = odd (n-1)
odd 0 = false
odd n = even (n-1)

Notice that even calls odd and odd calls even, so they are not definable
separately from one another. We join their definitions using the keyword
and to indicate that they are defined simultaneously by mutual recursion.

7.5

Sample Code

Here is the complete code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 8
Type Inference and
Polymorphism
8.1

Type Inference

So far weve mostly written our programs in what is known as the explicitly typed style. This means that whenever weve introduced a variable,
weve assigned it a type at its point of introduction. In particular every
variable in a pattern has a type associated with it. As you may have noticed, this gets a little tedious after a while, especially when youre using
clausal function definitions. A particularly pleasant feature of ML is that
it allows you to omit this type information whenever it can be determined
from context. This process is known as type inference since the compiler is
inferring the missing type information based on context.
For example, there is no need to give a type to the variable s in the
function
fn s:string => s "\n".
The reason is that no other type for s makes sense, since s is used as an
argument to string concatenation. Consequently, you may write simply
fn s => s "\n",
leaving ML to insert :string for you.
When is it allowable to omit this information? Almost always, with
very few exceptions. It is a deep, and important, result about ML that

8.1 Type Inference

68

missing type information can (almost) always be reconstructed completely

and unambiguously where it is omitted. This is called the principal typing
property of ML: whenever type information is omitted, there is always a
most general (i.e., least restrictive) way to recover the omitted type information. If there is no way to recover the omitted type information, then the
expression is ill-typed. Otherwise there is a best way to fill in the blanks,
which will (almost) always be found by the compiler. This is an amazingly
useful, and widely under-appreciated, property of ML. It means, for example, that the programmer can enjoy the full benefits of a static type system
without paying any notational penalty whatsoever!
The prototypical example is the identity function, fn x=>x. The body
of the function places no constraints on the type of x, since it merely returns x as the result without performing any computation on it. Since the
behavior of the identity function is the same for all possible choices of type
for its argument, it is said to be polymorphic. Therefore the identity function
has infinitely many types, one for each choice of the type of the parameter
x. Choosing the type of x to be typ, the type of the identity function is
typ->typ. In other words every type for the identity function has the form
typ->typ, where typ is the type of the argument.
Clearly there is a pattern here, which is captured by the notion of a type
scheme. A type scheme is a type expression involving one or more type
variables standing for an unknown, but arbitrary type expression. Type
variables are written a (pronounced ), b (pronounced ), c (pronounced ), etc.. An instance of a type scheme is obtained by replacing
each of the type variables occurring in it with a type scheme, with the same
type scheme replacing each occurrence of a given type variable. For example, the type scheme a->a has instances int->int, string->string,
(int*int)->(int*int), and (b->b)->(b->b), among infinitely many
others. However, it does not have the type int->string as instance, since
we are constrained to replace all occurrences of a type variable by the same
type scheme. However, the type scheme a->b has both int->int and
int->string as instances since there are different type variables occurring
in the domain and range positions.
Type schemes are used to express the polymorphic behavior of functions. For example, we may write fn x:a=>x for the polymorphic identity function of type a->a, meaning that the behavior of the identity
function is independent of the type of x. Similarly, the behavior of the
function fn (x,y)=>x+1 is independent of the type of y, but constrains the
R EVISED 11.02.11

D RAFT

V ERSION 1.2

8.1 Type Inference

69

type of x to be int. This may be expressed using type schemes by writing this function in the explicitly-typed form fn (x:int,y:a)=>x+1 with
type int*a->int.
In these examples we needed only one type variable to express the
polymorphic behavior of a function, but usually we need more than one.
For example, the function fn (x,y) = x constrains neither the type of x
nor the type of y. Consequently we may choose their types freely and independently of one another. This may be expressed by writing this function in the form fn (x:a,y:b)=>x with type scheme a*b->a. Notice
that while it is correct to assign the type a*a->a to this function, doing
so would be overly restrictive since the types of the two parameters need
not be the same. However, we could not assign the type a*b->c to this
function because the type of the result must be the same as the type of
the first parameter: it returns its first parameter when invoked! The type
scheme a*b->a precisely captures the constraints that must be satisfied for the function to be type correct. It is said to be the most general or
principal type scheme for the function.
It is a remarkable fact about ML that every expression (with the exception of a few pesky examples that well discuss below) has a principal type
scheme. That is, there is (almost) always a best or most general way to infer
types for expressions that maximizes generality, and hence maximizes flexibility in the use of the expression. Every expression seeks its own depth
in the sense that an occurrence of that expression is assigned a type that is
an instance of its principal type scheme determined by the context of use.
For example, if we write
(fn x=>x)(0),
the context forces the type of the identity function to be int->int, and if
we write
(fn x=>x)(fn x=>x)(0)
the context forces the instance (int->int)->(int->int) of the principal
type scheme for the identity at the first occurrence, and the instance int->int
for the second.
How is this achieved? Type inference is a process of constraint satisfaction. First, the expression determines a set of equations governing the relationships among the types of its subexpressions. For example, if a function
R EVISED 11.02.11

D RAFT

V ERSION 1.2

8.2 Polymorphic Definitions

70

is applied to an argument, then a constraint equating the domain type of

the function with the type of the argument is generated. Second, the constraints are solved using a process similar to Gaussian elimination, called
unification. The equations can be classified by their solution sets as follows:
1. Overconstrained: there is no solution. This corresponds to a type error.
2. Underconstrained: there are many solutions. There are two sub-cases:
ambiguous (due to overloading, which we will discuss further in section 8.3), or polymorphic (there is a best solution).
3. Uniquely determined: there is precisely one solution. This corresponds
to a completely unambiguous type inference problem.
The free type variables in the solution to the system of equations may be
thought of as determining the degrees of freedom or range of polymorphism of the type of an expression the constraints are solvable for any
choice of types to substitute for these free type variables.
This description of type inference as a constraint satisfaction procedure
accounts for the notorious obscurity of type checking errors in ML. If a
program is not type correct, then the system of constraints associated with
it will not have a solution. The type inference procedure attempts to find
a solution to these constraints, and at some point discovers that it cannot
succeed. It is fundamentally impossible to attribute this inconsistency to
any particular constraint; all that can be said is that the constraint set as
a whole has no solution. The checker usually reports the first unsatisfiable equation it encounters, but this may or may not correspond to the
reason (in the mind of the programmer) for the type error. The usual
method for finding the error is to insert sufficient type information to narrow down the source of the inconsistency until the source of the difficulty
is uncovered.

8.2

Polymorphic Definitions

There is an important interaction between polymorphic expressions and

value bindings that may be illustrated by the following example. Suppose
that we wish to bind the identity function to a variable I so that we may
refer to it by name. Weve previously observed that the identity function
R EVISED 11.02.11

D RAFT

V ERSION 1.2

8.2 Polymorphic Definitions

71

is polymorphic, with principal type scheme a->a. This may be captured

by ascribing this type scheme to the variable I at the val binding. That is,
we may write
val I : a->a = fn x=>x
to ascribe the type scheme a->a to the variable I. (We may also write
fun I(x:a):a = x
for an equivalent binding of I.) Having done this, each use of I determines a
distinct instance of the ascribed type scheme a->a. That is, both I 0 and I I
0 are well-formed expressions, the first assigning the type int->int to I,
the second assigning the types
(int->int)->(int->int)
and
int->int
to the two occurrences of I. Thus the variable I behaves precisely the same
as its definition, fn x=>x, in any expression where it is used.
As a convenience ML also provides a form of type inference on value
bindings that eliminates the need to ascribe a type scheme to the variable
when it is bound. If no type is ascribed to a variable introduced by a
val binding, then it is implicitly ascribed the principal type scheme of the
right-hand side. For example, we may write
val I = fn x=>x
or
fun I(x) = x
as a binding for the variable . The type checker implicitly assigns the principal type scheme, a->a, of the binding to the variable I. In practice we
often allow the type checker to infer the principal type of a variable, but it
is often good form to explicitly indicate the intended type as a consistency
check and for documentation purposes.
The treatment of val bindings during type checking ensures that a
bound variable has precisely the same type as its binding. In other words
R EVISED 11.02.11

D RAFT

V ERSION 1.2

8.2 Polymorphic Definitions

72

the type checker behaves as though all uses of the bound variable are implicitly replaced by its binding before type checking. Since this may involve replication of the binding, the meaning of a program is not necessarily preserved by this transformation. (Think, for example, of any expression that opens a window on your screen: if you replicate the expression
and evaluate it twice, it will open two windows. This is not the same as
evaluating it only once, which results in one window.)
To ensure semantic consistency, variables introduced by a val binding
are allowed to be polymorphic only if the right-hand side is a value. This
is called the value restriction on polymorphic declarations. For fun bindings this restriction is always met since the right-hand side is implicitly a
lambda expression, which is a value. However, it might be thought that
the following declaration introduces a polymorphic variable of type a ->
a, but in fact it is rejected by the compiler:
val J = I I
The reason is that the right-hand side is not a value; it requires computation to determine its value. It is therefore ruled out as inadmissible for
polymorphism; the variable J may not be used polymorphically in the remainder of the program. In this case the difficulty may be avoided by
writing instead
fun J x = I I x
because now the binding of J is a lambda, which is a value.
In some rare circumstances this is not possible, and some polymorphism is lost. For example, the following declaration of a value of list
type1
val l = nil @ nil
does not introduce an identifier with a polymorphic type, even though the
almost equivalent declaration
val l = nil
does do so. Since the right-hand side is a list, we cannot apply the trick
of defining l to be a function; we are stuck with a loss of polymorphism in
1 To

be introduced in chapter 9.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

8.3 Overloading

73

this case. This particular example is not very impressive, but occasionally
similar examples do arise in practice.
Why is the value restriction necessary? Later on, when we study mutable storage, well see that some restriction on polymorphism is essential if the language is to be type safe. The value restriction is an easilyremembered sufficient condition for soundness, but as the examples above
illustrate, it is by no means necessary. The designers of ML were faced
with a choice of simplicity vs flexibility; in this case they opted for simplicity at the expense of some expressiveness in the language.

8.3

Overloading

Type information cannot always be omitted. There are a few corner cases
that create problems for type inference, most of which arise because of
concessions that are motivated by long-standing, if dubious, notational
practices.
The main source of difficulty stems from overloading of arithmetic operators. As a concession to long-standing practice in informal mathematics
and in many programming languages, the same notation is used for both
integer and floating point arithmetic operations. As long as we are programming in an explicitly-typed style, this convention creates no particular problems. For example, in the function
fn x:int => x+x
it is clear that integer addition is called for, whereas in the function
fn x:real => x+x
it is equally obvious that floating point addition is intended.
However, if we omit type information, then a problem arises. What are
we to make of the function
fn x => x+x ?
Does + stand for integer or floating point addition? There are two distinct reconstructions of the missing type information in this example, corresponding to the preceding two explictly-typed programs. Which is the
compiler to choose?
When presented with such a program, the compiler has two choices:
R EVISED 11.02.11

D RAFT

V ERSION 1.2

8.3 Overloading

74

1. Declare the expression ambiguous, and force the programmer to provide enough explicit type information to resolve the ambiguity.
2. Arbitrarily choose a default interpretation, say the integer arithmetic, that forces one interpretation or another.
Each approach has its advantages and disadvantages. Many compilers
choose the second approach, but issue a warning indicating that it has
done so. To avoid ambiguity, explicit type information is required from
the programmer.
The situation is actually a bit more subtle than the preceding discussion implies. The reason is that the type inference process makes use of
the surrounding context of an expression to help resolve ambiguities. For
example, if the expression fn x=>x+x occurs in the following, larger expression, there is in fact no ambiguity:
(fn x => x+x)(3).
Since the function is applied to an integer argument, there is no question
that the only possible resolution of the missing type information is to treat
x as having type int, and hence to treat + as integer addition.
The important question is how much context is considered before the
situation is considered ambiguous? The rule of thumb is that context is
considered up to the nearest enclosing function declaration. For example,
consider the following example:
let
val double = fn x => x+x
in
(double 3, double 4)
end
The function expression fn x=>x+x will be flagged as ambiguous, even
though its only uses are with integer arguments. The reason is that value
bindings are considered to be units of type inference for which all ambiguity must be resolved before type checking continues. If your compiler
adopts the integer interpretation as default, the above program will be accepted (with a warning), but the following one will be rejected:

R EVISED 11.02.11

D RAFT

V ERSION 1.2

8.3 Overloading

75

let
val double = fn x => x+x
in
(double 3.0, double 4.0)
end
Finally, note that the following program must be rejected because no
resolution of the overloading of addition can render it meaningful:
let
val double = fn x => x+x
in
(double 3, double 3.0)
end
The ambiguity must be resolved at the val binding, which means that the
compiler must commit at that point to treating the addition operation as
either integer or floating point. No single choice can be correct, since we
subsequently use double at both types.
A closely related source of ambiguity arises from the record elision
notation described in chapter 5. Consider the function #name, defined by
fun #name {name=n:string, ...} = n
which selects the name field of a record. This definition is ambiguous because the compiler cannot uniquely determine the domain type of the
function! Any of the following types are legitimate domain types for
#name, none of which is best:
{name:string}
{name:string,salary:real}
{name:string,salary:int}
{name:string,address:string}
Of course there are infinitely many such examples, none of which is clearly
preferable to the other. This function definition is therefore rejected as
ambiguous by the compiler there is no one interpretation of the function
that suffices for all possible uses.
In chapter 5 we mentioned that functions such as #name are pre-defined
by the ML compiler, yet we just now claimed that such a function definition is rejected as ambiguous. Isnt this a contradiction? Not really,
R EVISED 11.02.11

D RAFT

V ERSION 1.2

8.4 Sample Code

76

because what happens is that each occurrence of #name is replaced by the

function
fn {name=n,...} = n
and then context is used to resolve the local ambiguity. This works well,
provided that the complete record type of the arguments to #name can be
determined from context. If not, the uses are rejected as ambiguous. Thus,
the following expression is well-typed
fn r : {name:string,address:string,salary:int} =>
(#name r, #address r)
because the record type of r is explicitly given. If the type of r were omitted, the expression would be rejected as ambiguous (unless the context
resolves the ambiguity.)

8.4

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 9
Programming with Lists
9.1

List Primitives

In chapter 5 we noted that aggregate data structures are especially easy

to handle in ML. In this chapter we consider another important aggregate
type, the list type. In addition to being an important form of aggregate
type it also illustrates two other general features of the ML type system:
1. Type constructors, or parameterized types. The type of a list reveals the
type of its elements.
2. Recursive types. The set of values of a list type are given by an inductive definition.
Informally, the values of type typ list are the finite lists of values of
type typ. More precisely, the values of type typ list are given by an inductive definition, as follows:
1. nil is a value of type typ list.
2. if h is a value of type typ, and t is a value of type typ list, then h::t
is a value of type typ list.
3. Nothing else is a value of type typ list.
The type expression typ list is a postfix notation for the application
of the type constructor list to the type typ. Thus list is a kind of function mapping types to types: given a type typ, we may apply list to it

9.1 List Primitives

78

to get another type, written typ list. The forms nil and :: are the value
constructors of type typ list. The nullary (no argument) constructor nil
may be thought of as the empty list. The binary (two argument) constructor :: constructs a non-empty list from a value h of type typ and another
value t of type typ list; the resulting value, h::t, of type typ list, is pronounced h cons t (for historical reasons). We say that h is consd onto t,
that h is the head of the list, and that t is its tail.
The definition of the values of type typ list given above is an example
of an inductive definition. The type is said to be recursive because this definition is self-referential in the sense that the values of type typ list are
defined in terms of (other) values of the same type. This is especially clear
if we examine the types of the value constructors for the type typ list:
val nil : typ list
val (op ::) : typ * typ list -> typ list
The notation op :: is used to refer to the :: operator as a function, rather
than to use it to form a list, which requires infix notation.
Two things are notable here:
1. The :: operation takes as its second argument a value of type typ
list, and yields a result of type typ list. This self-referential aspect
is characteristic of an inductive definition.
2. Both nil and op :: are polymorphic in the type of the underlying elements of the list. Thus nil is the empty list of type typ list for
any element type typ, and op :: constructs a non-empty list independently of the type of the elements of that list.
It is easy to see that a value val of type typ list has the form
val1 ::(val2 :: ( ::(valn ::nil) ))
for some n 0, where vali is a value of type typ for each 1 i n.
For according to the inductive definition of the values of type typ list,
the value val must either be nil, which is of the above form, or val1 ::val0 ,
where val0 is a value of type typ list. By induction val0 has the form
(val2 :: ( ::(valn ::nil) ))

R EVISED 11.02.11

D RAFT

V ERSION 1.2

9.2 Computing With Lists

79

and hence val again has the specified form.

By convention the operator :: is right-associative, so we may omit the
parentheses and just write
val1 ::val2 :: ::valn ::nil
as the general form of val of type typ list. This may be further abbreviated using list notation, writing
[ val1 , val2 , ..., valn ]
for the same list. This notation emphasizes the interpretation of lists as
finite sequences of values, but it obscures the fundamental inductive character of lists as being built up from nil using the :: operation.

9.2

Computing With Lists

How do we compute with values of list type? Since the values are defined
inductively, it is natural that functions on lists be defined recursively, using
a clausal definition that analyzes the structure of a list. Heres a definition
of the function length that computes the number of elements of a list:
fun length nil = 0
| length ( ::t) = 1 + length t
The definition is given by induction on the structure of the list argument.
The base case is the empty list, nil. The inductive step is the non-empty
list ::t (notice that we do not need to give a name to the head). Its definition is given in terms of the tail of the list t, which is smaller than the
list ::t. The type of length is a list -> int; it is defined for lists of
values of any type whatsoever.
We may define other functions following a similar pattern. Heres the
function to append two lists:
fun append (nil, l) = l
| append (h::t, l) = h :: append (t, l)
This function is built into ML; it is written using infix notation as exp1 @
exp2 . The running time of append is proportional to the length of the first
list, as should be obvious from its definition.
Heres a function to reverse a list.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

9.2 Computing With Lists

80

fun rev nil = nil

| rev (h::t) = rev t @ [h]
Its running time is O(n2 ), where n is the length of the argument list. This
can be demonstrated by writing down a recurrence that defines the running time T (n) on a list of length n.
T (0) = O (1)
T ( n + 1) = T ( n ) + O ( n )
Solving the recurrence we obtain the result T (n) = O(n2 ).
Can we do better? Oddly, we can take advantage of the non-associativity
of :: to give a tail-recursive definition of rev.
local
fun helper (nil, a) = a
| helper (h::t, a) = helper (t, h::a)
in
fun rev l = helper (l, nil)
end
The general idea of introducing an accumulator is the same as before, except that by re-ordering the applications of :: we reverse the list! The
helper function reverses its first argument and prepends it to its second
argument. That is, helper (l, a) evaluates to (rev l) @ a, where we
assume here an independent definition of rev for the sake of the specification. Notice that helper runs in time proportional to the length of its first
argument, and hence rev runs in time proportional to the length of its
argument.
The correctness of functions defined on lists may be established using
the principle of structural induction. We illustrate this by establishing that
the function helper satisfies the following specification:
for every l and a of type typ list, helper(l, a) evaluates to
the result of appending a to the reversal of l.
That is, there are no pre-conditions on l and a, and we establish the postcondition that helper (l, a) yields (rev l) @ a.
The proof is by structural induction on the list l. If l is nil, then helper
(l,a) evaluates to a, which fulfills the post-condition. If l is the list h::t,
R EVISED 11.02.11

D RAFT

V ERSION 1.2

9.3 Sample Code

81

then the application helper (l, a) reduces to the value of helper (t,
(h::a)). By the inductive hypothesis this is just (rev t) @ (h :: a),
which is equivalent to (rev t) @ [h] @ a. But this is just rev (h::t) @
a, which was to be shown.
The principle of structural induction may be summarized as follows.
To show that a function works correctly for every list l, it suffices to show
1. The correctness of the function for the empty list, nil, and
2. The correctness of the function for h::t, assuming its correctness for
t.
As with mathematical induction over the natural numbers, structural induction over lists allows us to focus on the basic and incremental behavior
of a function to establish its correctness for all lists.

9.3

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 10
Concrete Data Types
10.1

Datatype Declarations

Lists are one example of the general notion of a recursive type. ML provides
a general mechanism, the datatype declaration, for introducing programmerdefined recursive types. Earlier we introduced type declarations as an abbreviation mechanism. Types are given names as documentation and as
a convenience to the programmer, but doing so is semantically inconsequential one could replace all uses of the type name by its definition
and not affect the behavior of the program. In contrast the datatype declaration provides a means of introducing a new type that is distinct from
all other types and that does not merely stand for some other type. It is the
means by which the ML type system may be extended by the programmer.
The datatype declaration in ML has a number of facets. A datatype
declaration introduces
1. One or more new type constructors. The type constructors introduced may, or may not, be mutually recursive.
2. One or more new value constructors for each of the type constructors
introduced by the declaration.
The type constructors may take zero or more arguments; a zero-argument,
or nullary, type constructor is just a type. Each value constructor may
also take zero or more arguments; a nullary value constructor is just a
constant. The type and value constructors introduced by the declaration
are new in the sense that they are distinct from all other type and value

10.2 Non-Recursive Datatypes

83

constructors previously introduced; if a datatype re-defines an old type

or value constructor, then the old definition is shadowed by the new one,
rendering the old ones inaccessible in the scope of the new definition.

10.2

Non-Recursive Datatypes

Heres a simple example of a nullary type constructor with four nullary

value constructors.
datatype suit = Spades | Hearts | Diamonds | Clubs
This declaration introduces a new type suit with four nullary value constructors, Spades, Hearts, Diamonds, and Clubs. This declaration may be
read as introducing a type suit such that a value of type suit is either
Spades, or Hearts, or Diamonds, or Clubs. There is no significance to the
ordering of the constructors in the declaration; we could just as well have
written
datatype suit = Hearts | Diamonds | Spades | Clubs
(or any other ordering, for that matter). It is conventional to capitalize the
names of value constructors, but this is not required by the language.
Given the declaration of the type suit, we may define functions on it
by case analysis on the value constructors using a clausal function definition. For example, we may define the suit ordering in the card game of
bridge by the function
fun
|
|
|
|
|
|
|

outranks
outranks
outranks
outranks
outranks
outranks
outranks
outranks

(Spades, Spades) = false

(Spades, ) = true
(Hearts, Spades) = false
(Hearts, Hearts) = false
(Hearts, ) = true
(Diamonds, Clubs) = true
(Diamonds, ) = false
(Clubs, ) = false

This defines a function of type suit * suit -> bool that determines whether
or not the first suit outranks the second.
Data types may be parameterized by a type. For example, the declaration
R EVISED 11.02.11

D RAFT

V ERSION 1.2

10.2 Non-Recursive Datatypes

84

datatype a option = NONE | SOME of a

introduces the unary type constructor a option with two value constructors, NONE, with no arguments, and SOME, with one. The values of type typ
option are
1. The constant NONE, and
2. Values of the form SOME val, where val is a value of type typ.
For example, some values of type string option are NONE, SOME "abc",
and SOME "def".
The option type constructor is pre-defined in Standard ML. One common use of option types is to handle functions with an optional argument.
For example, here is a function to compute the base-b exponential function
for natural number exponents that defaults to base 2:
fun expt (NONE, n) = expt (SOME 2, n)
| expt (SOME b, 0) = 1
| expt (SOME b, n) =
if n mod 2 = 0 then
expt (SOME (b*b), n div 2)
else
b * expt (SOME b, n-1)
The advantage of the option type in this sort of situation is that it avoids
the need to make a special case of a particular argument, e.g., using 0 as
first argument to mean use the default exponent.
A related use of option types is in aggregate data structures. For example, an address book entry might have a record type with fields for various
bits of data about a person. But not all data is relevant to all people. For
example, someone may not have a spouse, but they all have a name. For
this we might use a type definition of the form
type entry = { name:string, spouse:string option }
so that one would create an entry for an unmarried person with a spouse
field of NONE.
Option types may also be used to represent an optional result. For
example, we may wish to define a function reciprocal that returns the
reciprocal of an integer, if it has one, and otherwise indicates that it has
R EVISED 11.02.11

D RAFT

V ERSION 1.2

10.3 Recursive Datatypes

85

no reciprocal. This is achieve by defining reciprocal to have type int ->

int option as follows:
fun reciprocal 0 = NONE
| reciprocal n = SOME (1 div n)
To use the result of a call to reciprocal we must perform a case analysis of
the form
case (reciprocal exp)
of NONE => exp1
| SOME r => exp2
where exp1 covers the case that exp has no reciprocal, and exp2 covers the
case that exp has reciprocal r.

10.3

Recursive Datatypes

The next level of generality is the recursive type definition. For example,
one may define a type typ tree of binary trees with values of type typ at
the nodes using the following declaration:
datatype a tree =
Empty |
Node of a tree * a * a tree
This declaration corresponds to the informal definition of binary trees with
values of type typ at the nodes:
1. The empty tree Empty is a binary tree.
2. If tree 1 and tree 2 are binary trees, and val is a value of type typ, then
Node (tree 1, val, tree 2) is a binary tree.
3. Nothing else is a binary tree.
The distinguishing feature of this definition is that it is recursive in the
sense that binary trees are constructed out of other binary trees, with the
empty tree serving as the base case.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

10.3 Recursive Datatypes

86

(Incidentally, a leaf in a binary tree is here represented as a node both of

whose children are the empty tree. This definition of binary trees is analogous to starting the natural numbers with zero, rather than one. One can
think of the children of a node in a binary tree as the predecessors of that
node, the only difference compared to the usual definition of predecessor
being that a node has two, rather than one, predecessors.)
To compute with a recursive type, use a recursive function. For example, here is the function to compute the height of a binary tree:
fun height Empty = 0
| height (Node (lft, , rht)) =
1 + max (height lft, height rht)
Notice that height is called recursively on the children of a node, and is
defined outright on the empty tree. This pattern of definition is another
instance of structural induction (on the tree type). The function height
is said to be defined by induction on the structure of a tree. The general
idea is to define the function directly for the base cases of the recursive
type (i.e., value constructors with no arguments or whose arguments do
not involve values of the type being defined), and to define it for non-base
cases in terms of its definitions for the constituent values of that type. We
will see numerous examples of this as we go along.
Heres another example. The size of a binary tree is the number of
nodes occurring in it. Heres a straightforward definition in ML:
fun size Empty = 0
| size (Node (lft, , rht)) =
1 + size lft + size rht
The function size is defined by structural induction on trees.
A word of warning. One reason to capitalize value constructors is to
avoid a pitfall in the ML syntax that we mentioned in chapter 2. Suppose
we gave the following definition of size:
fun size empty = 0
| size (Node (lft, , rht)) =
1 + size lft + size rht
The compiler will warn us that the second clause of the definition is redundant! Why? Because empty, spelled with a lower-case e, is a variable, not
R EVISED 11.02.11

D RAFT

V ERSION 1.2

10.3 Recursive Datatypes

87

a constructor, and hence matches any tree whatsoever. Consequently the

second clause never applies. By capitalizing constructors we can hope to
make mistakes such as these more evident, but in practice you are bound
to run into this sort of mistake.
The tree data type is appropriate for binary trees: those for which each
node has exactly two children. (Of course, either or both children might
be the empty tree, so we may consider this to define the type of trees with
at most two children; its a matter of terminology which interpretation you
prefer.) It should be obvious how to define the type of ternary trees, whose
nodes have at most three children, and so on for other fixed arities. But
what if we wished to define a type of trees with a variable number of children? In a so-called variadic tree some nodes might have three children,
some might have two, and so on. This can be achieved in at least two
ways. One way combines lists and trees, as follows:
datatype a tree =
Empty |
Node of a * a tree list
Each node has a list of children, so that distinct nodes may have different
numbers of children. Notice that the empty tree is distinct from the tree
with one node and no children because there is no data associated with
the empty tree, whereas there is a value of type a at each node.
Another approach is to simultaneously define trees and forests. A
variadic tree is either empty, or a node gathering a forest to form a tree;
a forest is either empty or a variadic tree together with another forest. This
leads to the following definition:
datatype a tree =
Empty |
Node of a * a forest
and a forest =
None |
Tree of a tree * a forest
This example illustrates the introduction of two mutually recursive datatypes.
Mutually recursive datatypes beget mutually recursive functions. Heres
a definition of the size (number of nodes) of a variadic tree:

R EVISED 11.02.11

D RAFT

V ERSION 1.2

10.4 Heterogeneous Data Structures

fun
|
and
|

size
size
size
size

88

tree Empty = 0
tree (Node ( , f)) = 1 + size forest f
forest None = 0
forest (Tree (t, f)) = size tree t + size forest f

Notice that we define the size of a tree in terms of the size of a forest, and
vice versa, just as the type of trees is defined in terms of the type of forests.
Many other variations are possible. Suppose we wish to define a notion
of binary tree in which data items are associated with branches, rather than
nodes. Heres a datatype declaration for such trees:
datatype a tree =
Empty |
Node of a branch * a branch
and a branch =
Branch of a * a tree
In contrast to our first definition of binary trees, in which the branches
from a node to its children were implicit, we now make the branches themselves explicit, since data is attached to them.
For example, we can collect into a list the data items labelling the branches
of such a tree using the following code:
fun collect Empty = nil
| collect (Node (Branch (ld, lt), Branch (rd, rt))) =
ld :: rd :: (collect lt) @ (collect rt)

10.4

Heterogeneous Data Structures

Returning to the original definition of binary trees (with data items at the
nodes), observe that the type of the data items at the nodes must be the
same for every node of the tree. For example, a value of type int tree has
an integer at every node, and a value of type string tree has a string at
every node. Therefore an expression such as
Node (Empty, 43, Node (Empty, "43", Empty))
is ill-typed. The type system insists that trees be homogeneous in the sense
that the type of the data items is the same at every node.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

10.5 Abstract Syntax

89

It is quite rare to encounter heterogeneous data structures in real programs. For example, a dictionary with strings as keys might be represented as a binary search tree with strings at the nodes; there is no need
for heterogeneity to represent such a data structure. But occasionally one
might wish to work with a heterogeneous tree, whose data values at each
node are of different types. How would one represent such a thing in ML?
To discover the answer, first think about how one might manipulate
such a data structure. When accessing a node, we would need to check
at run-time whether the data item is an integer or a string; otherwise we
would not know whether to, say, add 1 to it, or concatenate "1" to the
end of it. This suggests that the data item must be labelled with sufficient
information so that we may determine the type of the item at run-time. We
must also be able to recover the underlying data item itself so that familiar
operations (such as addition or string concatenation) may be applied to it.
The required labelling and discrimination is neatly achieved using a
datatype declaration. Suppose we wish to represent the type of integeror-string trees. First, we define the type of values to be integers or strings,
marked with a constructor indicating which:
datatype int or string =
Int of int |
String of string
Then we define the type of interest as follows:
type int or string tree =
int or string tree
Voila! Perfectly natural and easy heterogeneity is really a special case of
homogeneity!

10.5

Abstract Syntax

Datatype declarations and pattern matching are extremely useful for defining and manipulating the abstract syntax of a language. For example, we
may define a small language of arithmetic expressions using the following
declaration:

R EVISED 11.02.11

D RAFT

V ERSION 1.2

10.5 Abstract Syntax

90

datatype expr =
Numeral of int |
Plus of expr * expr |
Times of expr * expr
This definition has only three clauses, but one could readily imagine adding
others. Here is the definition of a function to evaluate expressions of the
language of arithmetic expressions written using pattern matching:
fun eval (Numeral n) = Numeral n
| eval (Plus (e1, e2)) =
let
val Numeral n1 = eval e1
val Numeral n2 = eval e2
in
Numeral (n1+n2)
end
| eval (Times (e1, e2)) =
let
val Numeral n1 = eval e1
val Numeral n2 = eval e2
in
Numeral (n1*n2)
end
The combination of datatype declarations and pattern matching contributes enormously to the readability of programs written in ML. A less
obvious, but more important, benefit is the error checking that the compiler can perform for you if you use these mechanisms in tandem. As an
example, suppose that we extend the type expr with a new component for
the reciprocal of a number, yielding the following revised definition:
datatype expr =
Numeral of int |
Plus of expr * expr |
Times of expr * expr |
Recip of expr
First, observe that the old definition of eval is no longer applicable to
values of type expr! For example, the expression
R EVISED 11.02.11

D RAFT

V ERSION 1.2

10.6 Sample Code

91

eval (Plus (Numeral 1, Numeral 2))

is ill-typed, even though it doesnt use the Recip constructor. The reason is
that the re-declaration of expr introduces a new type that just happens
to have the same name as the old type, but is in fact distinct from it. This
is a boon because it reminds us to recompile the old code relative to the
new definition of the expr type.
Second, upon recompiling the definition of eval we encounter an inexhaustive match warning: the old code no longer applies to every value of
type expr according to its new definition! We are of course lacking a case
for Recip, which we may provide as follows:
fun
|
|
|

eval (Numeral n) = Numeral n

eval (Plus (e1, e2)) = ... as before ...
eval (Times (e1, e2)) = ... as before ...
eval (Recip e) =
let
val Numeral n = eval e
in
Numeral (1 div n)
end

The value of the checks provided by the compiler in such cases cannot be
overestimated. When recompiling a large program after making a change
to a datatype declaration the compiler will automatically point out every
line of code that must be changed to conform to the new definition; it is
impossible to forget to attend to even a single case. This is a tremendous
help to the developer, especially if she is not the original author of the code
being modified and is another reason why the static type discipline of ML
is a positive benefit, rather than a hindrance, to programmers.

10.6

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 11
Higher-Order Functions
11.1

Functions as Values

Values of function type are first-class, which means that they have the same
rights and privileges as values of any other type. In particular, functions
may be passed as arguments and returned as results of other functions,
and functions may be stored in and retrieved from data structures such
as lists and trees. We will see that first-class functions are an important
source of expressive power in ML.
Functions which take functions as arguments or yield functions as results are known as higher-order functions (or, less often, as functionals or
operators). Higher-order functions arise frequently in mathematics. For
example, the differential operator is the higher-order function that, when
given a (differentiable) function on the real line, yields its first derivative
as a function on the real line. We also encounter functionals mapping functions to real numbers, and real numbers to functions. An example of the
former is provided by the definite integral viewed as a function of its integrand, and an example of the latter is the definite integral of a given
function on the interval [ a, x ], viewed as a function of a, that yields the
area under the curve from a to x as a function of x.
Higher-order functions are less familiar tools for many programmers
since the best-known programming languages have only rudimentary mechanisms to support their use. In contrast higher-order functions play a
prominent role in ML, with a variety of interesting applications. Their
use may be classified into two broad categories:

11.2 Binding and Scope

93

1. Abstracting patterns of control. Higher-order functions are design patterns that abstract out the details of a computation to lay bare the
skeleton of the solution. The skeleton may be fleshed out to form a
solution of a problem by applying the general pattern to arguments
that isolate the specific problem instance.
2. Staging computation. It arises frequently that computation may be
staged by expending additional effort early to simplify the computation of later results. Staging can be used both to improve efficiency and, as we will see later, to control sharing of computational
resources.

11.2

Binding and Scope

Before discussing these programming techniques, we will review the critically important concept of scope as it applies to function definitions. Recall
that Standard ML is a statically scoped language, meaning that identifiers
are resolved according to the static structure of the program. A use of the
variable var is considered to be a reference to the nearest lexically enclosing
declaration of var. We say nearest because of the possibility of shadowing; if we re-declare a variable var, then subsequent uses of var refer to the
most recent (lexically!) declaration of it; any previous declarations are
temporarily shadowed by the latest one.
This principle is easy to apply when considering sequences of declarations. For example, it should be clear by now that the variable y is bound
to 32 after processing the following sequence of declarations:
val
val
val
val

x
y
x
y

=
=
=
=

2
x*x
y*x
x*y

(*
(*
(*
(*

x=2 *)
y=4 *)
x=8 *)
y=32 *)

In the presence of function definitions the situation is the same, but it can
be a bit tricky to understand at first.
Heres an example to test your grasp of the lexical scoping principle:

R EVISED 11.02.11

D RAFT

V ERSION 1.2

11.2 Binding and Scope

val
fun
val
val

x
f
x
z

=
y
=
=

94

2
= x+y
3
f 4

After processing these declarations the variable z is bound to 6, not to 7!

The reason is that the occurrence of x in the body of f refers to the first
declaration of x since it is the nearest lexically enclosing declaration of the
occurence, even though it has been subsequently re-declared.
This example illustrates three important points:
1. Binding is not assignment! If we were to view the second binding of
x as an assignment statement, then the value of z would be 7, not 6.
2. Scope resolution is lexical, not temporal. We sometimes refer to the
most recent declaration of a variable, which has a temporal flavor,
but we always mean nearest lexically enclosing at the point of occurrence.
3. Shadowed bindings are not lost. The old binding for x is still
available (through calls to f), even though a more recent binding has
shadowed it.
One way to understand whats going on here is through the concept
of a closure, a technique for implementing higher-order functions. When
a function expression is evaluated, a copy of the environment is attached
to the function. Subsequently, all free variables of the function (i.e., those
variables not occurring as parameters) are resolved with respect to the environment attached to the function; the function is therefore said to be
closed with respect to the attached environment. This is achieved at
function application time by swapping the attached environment of the
function for the environment active at the point of the call. The swapped
environment is restored after the call is complete. Returning to the example above, the environment associated with the function f contains the
declaration val x = 2 to record the fact that at the time the function was
evaluated, the variable x was bound to the value 2. The variable x is subsequently re-bound to 3, but when f is applied, we temporarily reinstate
the binding of x to 2, add a binding of y to 4, then evaluate the body of the
function, yielding 6. We then restore the binding of x and drop the binding
of y before yielding the result.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

11.3 Returning Functions

11.3

95

Returning Functions

While seemingly very simple, the principle of lexical scope is the source of
considerable expressive power. Well demonstrate this through a series of
examples.
To warm up lets consider some simple examples of passing functions
as arguments and yielding functions as results. The standard example of
passing a function as argument is the map function, which applies a given
function to every element of a list. It is defined as follows:
fun map (f, nil) = nil
| map (f, h::t) = (f h) :: map (f, t)
For example, the application
map (fn x => x+1, [1,2,3,4])
evaluates to the list [2,3,4,5].
Functions may also yield functions as results. What is surprising is that
we can create new functions during execution, not just return functions
that have been previously defined. The most basic (and deceptively simple) example is the function constantly that creates constant functions:
given a value k, the application constantly k yields a function that yields
k whenever it is applied. Heres a definition of constantly:
val constantly = fn k => (fn a => k)
The function constantly has type a -> (b -> a). We used the fn notation for clarity, but the declaration of the function constantly may also
be written using fun notation as follows:
fun constantly k a = k
Note well that a white space separates the two successive arguments to
constantly! The meaning of this declaration is precisely the same as the
earlier definition using fn notation.
The value of the application constantly 3 is the function that is constantly 3; i.e., it always yields 3 when applied. Yet nowhere have we defined the function that always yields 3. The resulting function is created
by the application of constantly to the argument 3, rather than merely
R EVISED 11.02.11

D RAFT

V ERSION 1.2

11.3 Returning Functions

96

retrieved off the shelf of previously-defined functions. In implementation terms the result of the application constantly 3 is a closure consisting
of the function fn a => k with the environment val k = 3 attached to it.
The closure is a data structure (a pair) that is created by each application of
constantly to an argument; the closure is the representation of the new
function yielded by the application. Notice, however, that the only difference between any two results of applying the function constantly lies in
the attached environment; the underlying function is always fn a => k. If
we think of the lambda as the executable code of the function, then this
amounts to the observation that no new code is created at run-time, just
new instances of existing code.
This also points out why functions in ML are not the same as code
pointers in C. You may be familiar with the idea of passing a pointer to
a C function to another C function as a means of passing functions as arguments or yielding functions as results. This may be considered to be a
form of higher-order function in C, but it must be emphasized that code
pointers are significantly less powerful than closures because in C there
are only statically many possibilities for a code pointer (it must point to one
of the functions defined in your code), whereas in ML we may generate dynamically many different instances of a function, differing in the bindings
of the variables in its environment. The non-varying part of the closure,
the code, is directly analogous to a function pointer in C, but there is no
counterpart in C of the varying part of the closure, the dynamic environment.
The definition of the function map given above takes a function and list
as arguments, yielding a new list as result. Often it occurs that we wish to
map the same function across several different lists. It is inconvenient (and
a tad inefficient) to keep passing the same function to map, with the list
argument varying each time. Instead we would prefer to create a instance
of map specialized to the given function that can then be applied to many
different lists. This leads to the following definition of the function map:
fun map f nil = nil
| map f (h::t) = (f h) :: (map f t)
The function map so defined has type (a->b) -> a list -> b list.
It takes a function of type a -> b as argument, and yields another function of type a list -> b list as result.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

11.4 Patterns of Control

97

The passage from map to map is called currying. We have changed a

two-argument function (more properly, a function taking a pair as argument) into a function that takes two arguments in succession, yielding after the first a function that takes the second as its sole argument. This
passage can be codified as follows:
fun curry f x y = f (x, y)
The type of curry is
(a*b->c) -> (a -> (b -> c)).
Given a two-argument function, curry returns another function that, when
applied to the first argument, yields a function that, when applied to the
second, applies the original two-argument function to the first and second
arguments, given separately.
Observe that map may be alternately defined by the binding
fun map f l = curry map f l
Applications are implicitly left-associated, so that this definition is equivalent to the more verbose declaration
fun map f l = ((curry map) f) l

11.4

Patterns of Control

We turn now to the idea of abstracting patterns of control. There is an

obvious similarity between the following two functions, one to add up the
numbers in a list, the other to multiply them.
fun
|
fun
|

add
add
mul
mul

up
up
up
up

nil = 0
(h::t) = h + add up t
nil = 1
(h::t) = h * mul up t

What precisely is the similarity? We will look at it from two points of view.
One view is that in each case we have a binary operation and a unit
element for it. The result on the empty list is the unit element, and the
result on a non-empty list is the operation applied to the head of the list
and the result on the tail. This pattern can be abstracted as the function
reduce defined as follows:
R EVISED 11.02.11

D RAFT

V ERSION 1.2

11.4 Patterns of Control

98

fun reduce (unit, opn, nil) =

unit
| reduce (unit, opn, h::t) =
opn (h, reduce (unit, opn, t))
Here is the type of reduce:
val reduce : b * (a*b->b) * a list -> b
The first argument is the unit element, the second is the operation, and the
third is the list of values. Notice that the type of the operation admits the
possibility of the first argument having a different type from the second
argument and result.
Using reduce, we may re-define add up and mul up as follows:
fun add up l = reduce (0, op +, l)
fun mul up l = reduce (1, op *, l)
To further check your understanding, consider the following declaration:
fun mystery l = reduce (nil, op ::, l)
(Recall that op :: is the function of type a * a list -> a list that
adds a given value to the front of a list.) What function does mystery
compute?
Another view of the commonality between add up and mul up is that
they are both defined by induction on the structure of the list argument,
with a base case for nil, and an inductive case for h::t, defined in terms of
its behavior on t. But this is really just another way of saying that they are
defined in terms of a unit element and a binary operation! The difference
is one of perspective: whether we focus on the pattern part of the clauses
(the inductive decomposition) or the result part of the clauses (the unit
and operation). The recursive structure of add up and mul up is abstracted
by the reduce functional, which is then specialized to yield add up and
mul up. Said another way, the function reduce abstracts the pattern of defining
a function by induction on the structure of a list.
The definition of reduce leaves something to be desired. One thing to
notice is that the arguments unit and opn are carried unchanged through
the recursion; only the list parameter changes on recursive calls. While
this might seem like a minor overhead, its important to remember that
R EVISED 11.02.11

D RAFT

V ERSION 1.2

11.5 Staging

99

multi-argument functions are really single-argument functions that take

a tuple as argument. This means that each time around the loop we are
constructing a new tuple whose first and second components remain fixed,
but whose third component varies. Is there a better way? Heres another
definition that isolates the inner loop as an auxiliary function:
fun better reduce (unit, opn, l) =
let
fun red nil = unit
| red (h::t) = opn (h, red t)
in
red l
end
Notice that each call to better reduce creates a new function red that uses
the parameters unit and opn of the call to better reduce. This means that
red is bound to a closure consisting of the code for the function together
with the environment active at the point of definition, which will provide
bindings for unit and opn arising from the application of better reduce
to its arguments. Furthermore, the recursive calls to red no longer carry
bindings for unit and opn, saving the overhead of creating tuples on each
iteration of the loop.

11.5

Staging

An interesting variation on reduce may be obtained by staging the computation. The motivation is that unit and opn often remain fixed for many
different lists (e.g., we may wish to sum the elements of many different
lists). In this case unit and opn are said to be early arguments and the
list is said to be a late argument. The idea of staging is to perform as
much computation as possible on the basis of the early arguments, yielding a function of the late arguments alone.
In the case of the function reduce this amounts to building red on the
basis of unit and opn, yielding it as a function that may be later applied to
many different lists. Heres the code:

R EVISED 11.02.11

D RAFT

V ERSION 1.2

11.5 Staging

100

fun staged reduce (unit, opn) =

let
fun red nil = unit
| red (h::t) = opn (h, red t)
in
red
end
The definition of staged reduce bears a close resemblance to the definition
of better reduce; the only difference is that the creation of the closure
bound to red occurs as soon as unit and opn are known, rather than each
time the list argument is supplied. Thus the overhead of closure creation
is factored out of multiple applications of the resulting function to list
arguments.
We could just as well have replaced the body of the let expression with
the function
fn l => red l
but a moments thought reveals that the meaning is the same.
Note well that we would not obtain the effect of staging were we to use
the following definition:
fun curried reduce (unit, opn) nil = unit
| curried reduce (unit, opn) (h::t) =
opn (h, curried reduce (unit, opn) t)
If we unravel the fun notation, we see that while we are taking two arguments in succession, we are not doing any useful work in between the
arrival of the first argument (a pair) and the second (a list). A curried function does not take significant advantage of staging. Since staged reduce
and curried reduce have the same iterated function type, namely
(b * (a * b -> b)) -> a list -> b
the contrast between these two examples may be summarized by saying
not every function of iterated function type is curried. Some are, and some
arent. The interesting examples (such as staged reduce) are the ones
that arent curried. (This directly contradicts established terminology, but
it is necessary to deviate from standard practice to avoid a serious misapprehension.)
R EVISED 11.02.11

D RAFT

V ERSION 1.2

11.5 Staging

101

The time saved by staging the computation in the definition of staged reduce
is admittedly minor. But consider the following definition of an append
function for lists that takes both arguments at once:
fun append (nil, l) = l
| append (h::t, l) = h :: append(t,l)
Suppose that we will have occasion to append many lists to the end of a
given list. What wed like is to build a specialized appender for the first
list that, when applied to a second list, appends the second to the end of
the first. Heres a naive solution that merely curries append:
fun curried append nil l = l
| curried append (h::t) l = h :: curried append t l
Unfortunately this solution doesnt exploit the fact that the first argument
is fixed for many second arguments. In particular, each application of the
result of applying curried append to a list results in the first list being
traversed so that the second can be appended to it.
We can improve on this by staging the computation as follows:
fun staged append nil = (fn l => l)
| staged append (h::t) =
let
val tail appender = staged append t
in
fn l => h :: tail appender l
end
Notice that the first list is traversed once for all applications to a second argument. When applied to a list [v1 ,...,vn ], the function staged append
yields a function that is equivalent to, but not quite as efficient as, the
function
fn l => v1 :: v2 :: ... :: vn :: l.
This still takes time proportional to n, but a substantial savings accrues
from avoiding the pattern matching required to destructure the original
list argument on each call.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

11.6 Sample Code

11.6

102

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 12
Exceptions
In chapter 2 we mentioned that expressions in Standard ML always have a
type, may have a value, and may have an effect. So far weve concentrated
on typing and evaluation. In this chapter we will introduce the concept of
an effect. While its hard to give a precise general definition of what we
mean by an effect, the idea is that an effect is any action resulting from
evaluation of an expression other than returning a value. From this point
of view we might consider non-termination to be an effect, but we dont
usually think of failure to terminate as a positive action in its own right,
rather as a failure to take any action.
The main examples of effects in ML are these:
1. Exceptions. Evaluation may be aborted by signaling an exceptional
condition.
2. Mutation. Storage may be allocated and modified during evaluation.
3. Input/output. It is possible to read from an input source and write to
an output sink during evaluation.
4. Communication. Data may be sent to and received from communication channels.
This chapter is concerned with exceptions; the other forms of effects will
be considered later.

12.1 Exceptions as Errors

12.1

104

Exceptions as Errors

ML is a safe language in the sense that its execution behavior may be understood entirely in terms of the constructs of the language itself. Behavior such as dumping core or incurring a bus error are extra-linguistic
notions that may only be explained by appeal to the underlying implementation of the language. These cannot arise in ML. This is ensured by
a combination of a static type discipline, which rules out expressions that
are manifestly ill-defined (e.g., adding a string to an integer or casting an
integer as a function), and by dynamic checks that rule out violations that
cannot be detected statically (e.g., division by zero or arithmetic overflow).
Static violations are signalled by type checking errors; dynamic violations
are signalled by raising exceptions.

12.1.1

Primitive Exceptions

The expression 3 + "3" is ill-typed, and hence cannot be evaluated. In

contrast the expression 3 div 0 is well-typed (with type int), but incurs
a run-time fault that is signalled by raising the exception Div. We will
indicate this by writing
3 div 0 raise Div
An exception is a form of answer to the question what is the value
of this expression?. In most implementations an exception such as this
is reported by an error message of the form Uncaught exception Div,
together with the line number (or some other indication) of the point in
the program where the exception occurred.
Exceptions have names so that we may distinguish different sources
of error in a program. For example, evaluation of the expression maxint
* maxint (where maxint is the largest representable integer) causes the
exception Overflow to be raised, indicating that an arithmetic overflow
error arose in the attempt to carry out the multiplication. This is usefully
distinguished from the exception Div, corresponding to division by zero.
(You may be wondering about the overhead of checking for arithmetic
faults. The compiler must generate instructions that ensure that an overflow fault is caught before any subsequent operations are performed. This
can be quite expensive on pipelined processors, which sacrifice precise delivery of arithmetic faults in the interest of speeding up execution in the
R EVISED 11.02.11

D RAFT

V ERSION 1.2

12.1 Exceptions as Errors

105

non-faulting case. Unfortunately it is necessary to incur this overhead if

we are to avoid having the behavior of an ML program depend on the
underlying processor on which it is implemented.)
Another source of run-time exceptions is an inexhaustive match. Suppose we define the function hd as follows
fun hd (h:: ) = h
This definition is inexhaustive since it makes no provision for the possibility of the argument being nil. What happens if we apply hd to nil? The
exception Match is raised, indicating the failure of the pattern-matching
process:
hd nil raise Match
The occurrence of a Match exception at run-time is indicative of a violation of a pre-condition to the invocation of a function somewhere in the
program. Recall that it is often sensible for a function to be inexhaustive,
provided that we take care to ensure that it is never applied to a value
outside of its domain. Should this occur (because of a programming mistake, evidently), the result is nevertheless well-defined because ML checks
for, and signals, pattern match failure. That is, ML programs are implicitly bullet-proofed against failures of pattern matching. The flip side is
that if no inexhaustive match warnings arise during type checking, then
the exception Match can never be raised during evaluation (and hence no
run-time checking need be performed).
A related situation is the use of a pattern in a val binding to destructure
a value. If the pattern can fail to match a value of this type, then a Bind
exception is raised at run-time. For example, evaluation of the binding
val h:: = nil
raises the exception Bind since the pattern h:: does not match the value
nil. Here again observe that a Bind exception cannot arise unless the compiler has previously warned us of the possibility: no warning, no Bind
exception.

12.1.2

User-Defined Exceptions

So far we have considered examples of pre-defined exceptions that indicate fatal error conditions. Since the built-in exceptions have a builtR EVISED 11.02.11

D RAFT

V ERSION 1.2

12.1 Exceptions as Errors

106

in meaning, it is generally inadvisable to use these to signal programspecific error conditions. Instead we introduce a new exception using an
exception declaration, and signal it using a raise expression when a runtime violation occurs. That way we can associate specific exceptions with
specific pieces of code, easing the process of tracking down the source of
the error.
Suppose that we wish to define a checked factorial function that ensures that its argument is non-negative. Heres a first attempt at defining
such a function:
exception Factorial
fun checked factorial n =
if n < 0 then
raise Factorial
else if n=0 then
1
else n * checked factorial (n-1)
The declaration exception Factorial introduces an exception Factorial,
which we raise in the case that checked factorial is applied to a negative
number.
The definition of checked factorial is unsatisfactory in at least two
respects. One, relatively minor, issue is that it does not make effective use
of pattern matching, but instead relies on explicit comparison operations.
To some extent this is unavoidable since we wish to check explicitly for
negative arguments, which cannot be done using a pattern. A more significant problem is that checked factorial repeatedly checks the validity
of its argument on each recursive call, even though we can prove that if
the initial argument is non-negative, then so must be the argument on each
recursive call. This fact is not reflected in the code. We can improve the
definition by introducing an auxiliary function:
exception Factorial
local
fun fact 0 = 1
| fact n = n * fact (n-1)
in
fun checked factorial n =
if n >= 0 then
R EVISED 11.02.11

D RAFT

V ERSION 1.2

12.2 Exception Handlers

107

fact n
else
raise Factorial
end
Notice that we perform the range check exactly once, and that the auxiliary
function makes effective use of pattern-matching.

12.2

Exception Handlers

The use of exceptions to signal error conditions suggests that raising an

exception is fatal: execution of the program terminates with the raised
exception. But signaling an error is only one use of the exception mechanism. More generally, exceptions can be used to effect non-local transfers
of control. By using an exception handler we may catch a raised exception
and continue evaluation along some other path. A very simple example
is provided by the following driver for the factorial function that accepts
numbers from the keyboard, computes their factorial, and prints the result.
fun factorial driver () =
let
val input = read integer ()
val result =
toString (checked factorial input)
in
print result
end
handle Factorial => print "Out of range."
An expression of the form exp handle match is called an exception handler. It is evaluated by attempting to evaluate exp. If it returns a value, then
that is the value of the entire expression; the handler plays no role in this
case. If, however, exp raises an exception exc, then the exception value is
matched against the clauses of the match (exactly as in the application of a
clausal function to an argument) to determine how to proceed. If the pattern of a clause matches the exception exc, then evaluation resumes with
the expression part of that clause. If no pattern matches, the exception
R EVISED 11.02.11

D RAFT

V ERSION 1.2

12.2 Exception Handlers

108

exc is re-raised so that outer exception handlers may dispatch on it. If no

handler handles the exception, then the uncaught exception is signaled
as the final result of evaluation. That is, computation is aborted with the
uncaught exception exc.
In more operational terms, evaluation of exp handle match proceeds
by installing an exception handler determined by match, then evaluating
exp. The previous binding of the exception handler is preserved so that
it may be restored once the given handler is no longer needed. Raising
an exception consists of passing a value of type exn to the current exception handler. Passing an exception to a handler de-installs that handler,
and re-installs the previously active handler. This ensures that if the handler itself raises an exception, or fails to handle the given exception, then
the exception is propagated to the handler active prior to evaluation of the
handle expression. If the expression does not raise an exception, the previous handler is restored as part of completing the evaluation of the handle
expression.
Returning to the function factorial driver, we see that evaluation
proceeds by attempting to compute the factorial of a given number (read
from the keyboard by an unspecified function read integer), printing the
result if the given number is in range, and otherwise reporting that the
number is out of range. The example is trivialized to focus on the role of
exceptions, but one could easily imagine generalizing it in a number of
ways that also make use of exceptions. For example, we might repeatedly
read integers until the user terminates the input stream (by typing the end
of file character). Termination of input might be signaled by an EndOfFile
exception, which is handled by the driver. Similarly, we might expect that
the function read integer raises the exception SyntaxError in the case
that the input is not properly formatted. Again we would handle this
exception, print a suitable message, and resume.
Heres a sketch of a more complicated factorial driver:
fun factorial driver () =
let
val input = read integer ()
val result =
toString (checked factorial input)
val = print result
in
R EVISED 11.02.11

D RAFT

V ERSION 1.2

12.2 Exception Handlers

109

factorial driver ()
end
handle EndOfFile => print "Done."
| SyntaxError =>
let
val = print "Syntax error."
in
factorial driver ()
end
| Factorial =>
let
val = print "Out of range."
in
factorial driver ()
end
We will return to a more detailed discussion of input/output later in these
notes. The point to notice here is that the code is structured with a completely uncluttered normal path that reads an integer, computes its factorial, formats it, prints it, and repeats. The exception handler takes care
of the exceptional cases: end of file, syntax error, and domain error. In the
latter two cases we report an error, and resume reading. In the former we
simply report completion and we are done.
The reader is encouraged to imagine how one might structure this program without the use of exceptions. The primary benefits of the exception
mechanism are as follows:
1. They force you to consider the exceptional case (if you dont, youll
get an uncaught exception at run-time), and
2. They allow you to segregate the special case from the normal case in
the code (rather than clutter the code with explicit checks).
These aspects work hand-in-hand to facilitate writing robust programs.
A typical use of exceptions is to implement backtracking, a programming technique based on exhaustive search of a state space. A very simple, if somewhat artificial, example is provided by the following function
to compute change from an arbitrary list of coin values. What is at issue
is that the obvious greedy algorithm for making change that proceeds
R EVISED 11.02.11

D RAFT

V ERSION 1.2

12.3 Value-Carrying Exceptions

110

by doling out as many coins as possible in decreasing order of value does

not always work. Given only a 5 cent and a 2 cent coin, we cannot make
16 cents in change by first taking three 5s and then proceeding to dole out
2s. In fact we must use two 5s and three 2s to make 16 cents. Heres a
method that works for any set of coins:
exception Change
fun change 0 = nil
| change nil = raise Change
| change (coin::coins) amt =
if coin > amt then
change coins amt
else
(coin :: change (coin::coins) (amt-coin))
handle Change => change coins amt
The idea is to proceed greedily, but if we get stuck, we undo the most
recent greedy decision and proceed again from there. Simulate evaluation
of the example of change [5,2] 16 to see how the code works.

12.3

Value-Carrying Exceptions

So far exceptions are just signals that indicate that an exceptional condition has arisen. Often it is useful to attach additional information that
is passed to the exception handler. This is achieved by attaching values to
exceptions.
For example, we might associate with a SyntaxError exception a string
indicating the precise nature of the error. In a parser for a language we
might write something like
raise SyntaxError "Integer expected"
to indicate a malformed expression in a situation where an integer is expected, and write
raise SyntaxError "Identifier expected"
to indicate a badly-formed identifier.
To associate a string with the exception SyntaxError, we declare it as
R EVISED 11.02.11

D RAFT

V ERSION 1.2

12.3 Value-Carrying Exceptions

111

exception SyntaxError of string.

This declaration introduces the exception SyntaxError as an exception
carrying a string as value. This declaration introduces the exception constructor SyntaxError.
Exception constructors are in many ways similar to value constructors.
In particular they can be used in patterns, as in the following code fragment:
... handle SyntaxError msg => (print "Syntax error: " ^ msg)
Here we specify a pattern for SyntaxError exceptions that also binds the
string associated with the exception to the identifier msg and prints that
string along with an error indication.
Recall that we may use value constructors in two ways:
1. We may use them to create values of a datatype (perhaps by applying
them to other values).
2. We may use them to match values of a datatype (perhaps also matching a constituent value).
The situation with exception constructors is symmetric.
1. We may use them to create an exception (perhaps with an associated
value).
2. We may use them to match an exception (perhaps also matching the
associated value).
Value constructors have types, as we previously mentioned. For example, the list constructors nil and :: have types a list and a * a list
-> a list, respectively. What about exception constructors? A bare
exception constructor (such as Factorial above) has type exn and a valuecarrying exception constructor (such as SyntaxError) has type string ->
exn. Thus Factorial is a value of type exn, and
SyntaxError "Integer expected"
is a value of type exn.
The type exn is the type of exception packets, the data values associated
with exceptions. The primitive operation raise takes any value of type
R EVISED 11.02.11

D RAFT

V ERSION 1.2

12.4 Sample Code

112

exn as argument and raises an exception with that value. The clauses of
a handler may be applied to any value of type exn using the rules of pattern matching described earlier; if an exception constructor is no longer in
scope, then the handler cannot catch it (other than via a wild-card pattern).
The type exn may be thought of as a kind of built-in datatype, except
that the constructors of this type are not determined once and for all (as
they are with a datatype declaration), but rather are incrementally introduced as needed in a program. For this reason the type exn is sometimes
called an extensible datatype.

12.4

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 13
Mutable Storage
In this chapter we consider a second form of effect, called a storage effect,
the allocation or mutation of storage during evaluation. The introduction
of storage effects has profound consequences, not all of which are desirable. (Indeed, one connotation of the phrase side effect is an unintended
consequence of a medication!) While it is excessive to dismiss storage effects as completely undesirable, it is advantageous to minimize the use of
storage effects except in situations where the task clearly demands them.
We will explore some techniques for programming with storage effects
later in this chapter, but first we introduce the primitive mechanisms for
programming with mutable storage in ML.

13.1

Reference Cells

To support mutable storage the execution model that we described in chapter 2 is enriched with a memory consisting of a finite set of mutable cells. A
mutable cell may be thought of as a container in which a data value of a
specified type is stored. During execution of a program the contents of
a cell may be retrieved or replaced by any other value of the appropriate
type. Since cells are used by issuing commands to modify and retrieve
their contents, programming with cells is called imperative programming.
Changing the contents of a mutable cell introduces a temporal aspect
to evaluation. We speak of the current contents of a cell, meaning the value
most recently assigned to it. We also speak of previous and future values of a
reference cell when discussing the behavior of a program. This is in sharp

13.1 Reference Cells

114

contrast to the effect-free fragment of ML, for which no such concepts apply. For example, the binding of a variable does not change while evaluating within the scope of that variable, lending a permanent quality
to statements about variables the current binding is the only binding
that variable will ever have.
The type typ ref is the type of reference cells containing values of type
typ. Reference cells are, like all values, first class they may be bound
to variables, passed as arguments to functions, returned as results of functions, appear within data structures, and even be stored within other reference cells.
A reference cell is created, or allocated, by the function ref of type typ ->
typ ref. When applied to a value val of type typ, ref allocates a new cell,
initializes its content to val, and returns a reference to the cell. By new
we mean that the allocated cell is distinct from all other cells previously
allocated, and does not share storage with them.
The contents of a cell of type typ is retrieved using the function ! of type
typ ref -> typ. Applying ! to a reference cell yields the current contents
of that cell. The contents of a cell is changed by applying the assignment
operator op :=, which has type typ ref * typ -> unit. Assignment is
usually written using infix syntax. When applied to a cell and a value, it
replaces the content of that cell with that value, and yields the null-tuple
as result.
Here are some examples:
val
val
val
val
val
val
val
val

r = ref 0
s = ref 0
= r := 3
x = !s + !r
t = r
= t := 5
y = !s + !r
z = !t + !r

After execution of these bindings, the variable x is bound to 3, the variable

y is bound to 5, and z is bound to 10.
Notice the use of a val binding of the form val = exp when exp is
to be evaluated purely for its effect. The value of exp is discarded by the
binding, since the left-hand side is a wildcard pattern. In most cases the
R EVISED 11.02.11

D RAFT

V ERSION 1.2

13.2 Reference Patterns

115

expression exp has type unit, so that its value is guaranteed to be the nulltuple, (), if it has a value at all.
A wildcard binding is used to define sequential composition of expressions in ML. The expression
exp1 ; exp2
is shorthand for the expression
let
val

= exp1

in
exp2
end
that first evaluates exp1 for its effect, then evaluates exp2 .
Functions of type typ->unit are sometimes called procedures, because
they are executed purely for their effect. This is apparent from the type: it
is assured that the value of applying such a function is the null-tuple, (),
so the only point of applying it is for its effects on memory.

13.2

Reference Patterns

It is a common mistake to omit the exclamation point when referring to

the content of a reference, especially when that cell is bound to a variable.
In more familiar languages such as C all variables are implicitly bound
to reference cells, and they are implicitly de-referenced whenever they are
used so that a variable always stands for its current contents. This is both a
boon and a bane. It is obviously helpful in many common cases since it alleviates the burden of having to explicitly dereference variables whenever
their content is required. However, it shifts the burden to the programmer in the case that the address, and not the content, is intended. In C
one writes & x for the address of (the cell bound to) x. Whether explicit or
implicit de-referencing is preferable is to a large extent a matter of taste.
The burden of explicit de-referencing is not nearly so onerous in ML as
it might be in other languages simply because reference cells are used so
infrequently in ML programs, whereas they are the sole means of binding
variables in more familiar languages.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

13.3 Identity

116

An alternative to explicitly de-referencing cells is to use ref patterns. A

pattern of the form ref pat matches a reference cell whose content matches
the pattern pat. This means that the cells contents are implicitly retrieved
during pattern matching, and may be subsequently used without explicit
de-referencing. In fact, the function ! may be defined using a ref pattern
as follows:
fun !(ref a) = a
When called with a reference cell, it is de-referenced and its contents is
bound to a, which is returned as result. In practice it is common to use
both explicit de-referencing and ref patterns, depending on the situation.

13.3

Identity

Reference cells raise delicate issues of equality that considerably complicate reasoning about programs. In general we say that two expressions (of
the same type) are equal iff they cannot be distinguished by any operation
in the language. That is, two expressions are distinct iff there is some way
within the language to tell them apart. This is called Leibnizs Principle of
identity of indiscernables we equate everything that we cannot tell apart
and the indiscernability of identicals that which we deem equal cannot
be told apart.
What makes Leibnizs Principle tricky to grasp is that it hinges on what
we mean by a way to tell expressions apart. The crucial idea is that we
can tell two expressions apart iff there is a complete program containing one
of the expressions whose observable behavior changes when we replace that
expression by the other. That is, two expressions are considered equal iff
there is no such scenario that distinguishes them. But what do we mean by
complete program? And what do we mean by observable behavior?
For the present purposes we will consider a complete program to be
any expression of basic type (say, int or bool or string). The idea is that a
complete program is one that computes a concrete result such as a number.
The observable behavior of a complete program includes at least these
aspects:
1. Its value, or lack thereof, either by non-termination or by raising an
uncaught exception.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

13.3 Identity

117

2. Its visible side effects, include visible modifications to mutable storage or any input/output it may perform.
In contrast here are some behaviors that we will not count as observations:
1. Execution time or space usage.
2. Private uses of storage (e.g., internally-allocated reference cells).
3. The name of uncaught exceptions (i.e., we will not distinguish between terminating with the uncaught exception Bind and the uncaught exception Match.
With these ideas in mind, it should be plausible that if we evaluate
these bindings
val r = ref 0
val s = ref 0
then r and s are not equivalent. Consider the following usage of r to compute an integer result:
(s := 1 ; !r)
Clearly this expression evaluates to 0, and mutates the binding of s. Now
replace r by s to obtain
(s := 1 ; !s)
This expression evaluates to 1, and mutates s as before. These two complete programs distinguish r from s, and therefore must be considered
distinct.
Had we replaced the binding for s by the binding
val s = r
then the two expressions that formerly distinguished r from s no long
do so they are, after all, bound to the same reference cell! In fact, no
program can be concocted that would distinguish them. In this case r and
s are equivalent.
Now consider a third, very similar scenario. Let us declare r and s as
follows:
R EVISED 11.02.11

D RAFT

V ERSION 1.2

13.4 Aliasing

118

val r = ref ()
val s = ref ()
Are r and s equivalent or not? We might first try to distinguish them by
a variant of the experiment considered above. This breaks down because
there is only one possible value we can assign to a variable of type unit
ref! Indeed, one may suspect that r and s are equivalent in this case,
but in fact there is a way to distinguish them! Heres a complete program
involving r that we will use to distinguish r from s:
if r=r then "its r" else "its not"
Now replace the first occurrence of r by s to obtain
if s=r then "its r" else "its not"
and the result is different.
This example hinges on the fact that ML defines equality for values
of reference type to be reference equality (or, occasionally, pointer equality).
Two reference cells (of the same type) are equal in this sense iff they both
arise from the exact same use of the ref operation to allocate that cell;
otherwise they are distinct. Thus the two cells bound to r and s above are
observably distinct (by testing reference equality), even though they can
only ever hold the value (). Had equality not been included as a primitive,
any two reference cells of unit type would have been equal.
Why does ML provide such a fine-grained notion of equality? True
equality, as defined by Leibnizs Principle, is, unfortunately, undecidable
there is no computer program that determines whether two expressions
are equivalent in this sense. ML provides a useful, conservative approximation to true equality that in some cases is not defined (you cannot test
two functions for equality) and in other cases is too picky (it distinguishes
reference cells that are otherwise indistinguishable). Such is life.

13.4

Aliasing

To see how reference cells complicate programming, let us consider the

problem of aliasing. Any two variables of the same reference type might
be bound to the same reference cell, or to two different reference cells. For
example, after the declarations
R EVISED 11.02.11

D RAFT

V ERSION 1.2

13.5 Programming Well With References

119

val r = ref 0
val s = ref 0
the variables r and s are not aliases, but after the declaration
val r = ref 0
val s = r
the variables r and s are aliases for the same reference cell.
These examples show that we must be careful when programming
with variables of reference type. This is particularly problematic in the
case of functions, because we cannot assume that two different argument
variables are bound to different reference cells. They might, in fact, be
bound to the same reference cell, in which case we say that the two variables are aliases for one another. For example, in a function of the form
fn (x:typ ref, y:typ ref) => exp
we may not assume that x and y are bound to different reference cells. We
must always ask ourselves whether weve properly considered aliasing
when writing such a function. This is harder to do than it sounds. Aliasing
is a huge source of bugs in programs that work with reference cells.

13.5

Programming Well With References

Using references it is possible to mimic the style of programming used in

imperative languages such as C. For example, we might define the factorial function in imitation of such languages as follows:

R EVISED 11.02.11

D RAFT

V ERSION 1.2

13.5 Programming Well With References

120

fun imperative fact (n:int) =

let
val result = ref 1
val i = ref 0
fun loop () =
if !i = n then
()
else
(i := !i + 1;
result := !result * !i;
loop ())
in
loop (); !result
end
Notice that the function loop is essentially just a while loop; it repeatedly
executes its body until the contents of the cell bound to i reaches n. The
tail call to loop is essentially just a goto statement to the top of the loop.
It is (appallingly) bad style to program in this fashion. The purpose of the
function imperative fact is to compute a simple function on the natural numbers. There is nothing about its definition that suggests that state
must be maintained, and so it is senseless to allocate and modify storage
to compute it. The definition we gave earlier is shorter, simpler, more efficient, and hence more suitable to the task. This is not to suggest, however,
that there are no good uses of references. We will now discuss some important uses of state in ML.

13.5.1

Private Storage

The first example is the use of higher-order functions to manage shared

private state. This programming style is closely related to the use of objects to manage state in object-oriented programming languages. Heres
an example to frame the discussion:

R EVISED 11.02.11

D RAFT

V ERSION 1.2

13.5 Programming Well With References

121

local
val counter = ref 0
in
fun tick () = (counter := !counter + 1; !counter)
fun reset () = (counter := 0)
end
This declaration introduces two functions, tick of type unit -> int and
reset of type unit -> unit. Their definitions share a private variable
counter that is bound to a mutable cell containing the current value of
a shared counter. The tick operation increments the counter and returns
its new value, and the reset operation resets its value to zero. The types
of the operations suggest that implicit state is involved. In the absence
of exceptions and implicit state, there is only one useful function of type
unit->unit, namely the function that always returns its argument (and
its debatable whether this is really useful!).
The declaration above defines two functions, tick and reset, that share
a single private counter. Suppose now that we wish to have several different instances of a counter different pairs of functions tick and reset that
share different state. We can achieve this by defining a counter generator (or
constructor) as follows:
fun new counter () =
let
val counter = ref 0
fun tick () = (counter := !counter + 1; !counter)
fun reset () = (counter := 0)
in
{ tick = tick, reset = reset }
end
The type of new counter is
unit -> { tick : unit->int, reset : unit->unit }.
Weve packaged the two operations into a record containing two functions that share private state. There is an obvious analogy with class-based
object-oriented programming. The function new counter may be thought
of as a constructor for a class of counter objects. Each object has a private instance variable counter that is shared between the methods tick and reset
of the object represented as a record with two fields.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

13.5 Programming Well With References

122

Heres how we use counters.

val c1 = new counter ()
val c2 = new counter ()
#tick c1 ();
(* 1 *)
#tick c1 ();
(* 2 *)
#tick c2 ();
(* 1 *)
#reset c1 ();
#tick c1 ();
(* 1 *)
#tick c2 ();
(* 2 *)
Notice that c1 and c2 are distinct counters that increment and reset independently of one another.

13.5.2

Mutable Data Structures

A second important use of references is to build mutable data structures.

The data structures (such as lists and trees) weve considered so far are
immutable in the sense that it is impossible to change the structure of the
list or tree without building a modified copy of that structure. This is
both a benefit and a drawback. The principal benefit is that immutable
data structures are persistent in that operations performed on them do not
destroy the original structure in ML we can eat our cake and have it too.
For example, we can simultaneously maintain a dictionary both before
and after insertion of a given word. The principal drawback is that if we
arent really relying on persistence, then it is wasteful to make a copy of a
structure if the original is going to be discarded anyway. What wed like
in this case is to have an update in place operation to build an ephemeral
(opposite of persistent) data structure. To do this in ML we make use of
references.
A simple example is the type of possibly circular lists, or pcls. Informally,
a pcl is a finite graph in which every node has at most one neighbor, called
its predecessor, in the graph. In contrast to ordinary lists the predecessor
R EVISED 11.02.11

D RAFT

V ERSION 1.2

13.5 Programming Well With References

123

relation is not necessarily well-founded: there may be an infinite sequence

of nodes arranged in descending order of predecession. Since the graph
is finite, this can only happen if there is a cycle in the graph: some node
has an ancestor as predecessor. How can such a structure ever come into
existence? If the predecessors of a cell are needed to construct a cell, then
the ancestor that is to serve as predecessor in the cyclic case can never be
created! The trick is to employ backpatching: the predecessor is initialized to Nil, so that the node and its ancestors can be constructed, then it is
reset to the appropriate ancestor to create the cycle.
This can be achieved in ML using the following datatype declaration:
datatype a pcl = Pcl of a pcell ref
and a pcell = Nil | Cons of a * a pcl;
A value of type typ pcl is essentially a reference to a value of type typ
pcell. A value of type typ pcell is either Nil, the cell at the end of a noncircular possibly-circular list, or Cons (h, t), where h is a value of type
typ and t is another such possibly-circular list.
Here are some convenient functions for creating and taking apart possiblycircular lists:
fun
fun
fun
fun

cons (h, t) =
nill () = Pcl
phd (Pcl (ref
ptl (Pcl (ref

Pcl (ref (Cons (h, t)));

(ref Nil);
(Cons (h, )))) = h;
(Cons ( , t)))) = t;

To implement backpatching, we need a way to zap the tail of a possiblycircular list.

fun stl (Pcl (r as ref (Cons (h, ))), u) =
(r := Cons (h, u));
If youd like, it would make sense to require that the tail of the Cons cell
be the empty pcl, so that youre only allowed to backpatch at the end of a
finite pcl.
Here is a finite and an infinite pcl.
val finite = cons (4, cons (3, cons (2, cons (1, nill ()))))
val tail = cons (1, nill());
val infinite = cons (4, cons (3, cons (2, tail)));
val = stl (tail, infinite)
R EVISED 11.02.11

D RAFT

V ERSION 1.2

13.6 Mutable Arrays

124

The last step backpatches the tail of the last cell of infinite to be infinite
itself, creating a circular list.
Now let us define the size of a pcl to be the number of distinct nodes
occurring in it. It is an interesting problem is to define a size function
for pcls that makes no use of auxiliary storage (e.g., no set of previouslyencountered nodes) and runs in time proportional to the number of cells in
the pcl. The idea is to think of running a long race between a tortoise and a
hare. If the course is circular, then the hare, which quickly runs out ahead
of the tortoise, will eventually come from behind and pass it! Conversely,
if this happens, the course must be circular.
local
fun race (Nil, Nil) = 0
| race (Cons ( , Pcl (ref c)), Nil) =
1 + race (c, Nil)
| race (Cons ( , Pcl (ref c)), Cons ( , Pcl (ref Nil))) =
1 + race (c, Nil)
| race (Cons ( , l), Cons ( , Pcl (ref (Cons ( , m))))) =
1 + race (l, m)
and race (Pcl (r as ref c), Pcl (s as ref d)) =
if r=s then 0 else race (c, d)
in
fun size (Pcl (ref c)) = race (c, c)
end
The hare runs twice as fast as the tortoise. We let the tortoise do the counting; the hares job is simply to detect cycles. If the hare reaches the finish
line, it simply waits for the tortoise to finish counting. This covers the
first three clauses of race. If the hare has not yet finished, we must continue with the hare running at twice the pace, checking whether the hare
catches the tortoise from behind. Notice that it can never arise that the
tortoise reaches the end before the hare does! Consequently, the definition
of race is inexhaustive.

13.6

Mutable Arrays

In addition to reference cells, ML also provides mutable arrays as a primitive data structure. The type typ array is the type of arrays carrying valR EVISED 11.02.11

D RAFT

V ERSION 1.2

13.6 Mutable Arrays

125

ues of type typ. The basic operations on arrays are these:

val
val
val
val

array : int * a -> a array

length : a array -> int
sub : a array * int -> a
update : a array * int * a -> unit

The function array creates a new array of a given length, with the given
value as the initial value of every element of the array. The function length
returns the length of an array. The function sub performs a subscript operation, returning the ith element of an array A, where 0 i < length(A).
(These are just the basic operations on arrays; please see Appendix A for
complete information.)
One simple use of arrays is for memoization. Heres a function to compute the nth Catalan number, which may be thought of as the number
of distinct ways to parenthesize an arithmetic expression consisting of a
sequence of n consecutive multiplications. It makes use of an auxiliary
summation function that you can easily define for yourself. (Applying
sum to f and n computes the sum of f 1 + + f n.)
fun C 1 = 1
| C n = sum (fn k => (C k) * (C (n-k))) (n-1)
This definition of C is hugely inefficient because a given computation may
be repeated exponentially many times. For example, to compute C 10 we
must compute C 1, C 2, . . . , C 9, and the computation of C i engenders the
computation of C 1, . . . , C i 1 for each 1 i 9. We can do better by
caching previously-computed results in an array, leading to an enormous
improvement in execution speed. Heres the code:
local
val limit : int = 100
val memopad : int option array =
Array.array (limit, NONE)
in
fun C 1 = 1
| C n = sum (fn k => (C k)*(C (n-k))) (n-1)
and C n =
if n < limit then
case Array.sub (memopad, n)
R EVISED 11.02.11

D RAFT

V ERSION 1.2

13.7 Sample Code

126
of SOME r => r
| NONE =>
let
val r = C n
in
Array.update (memopad, n, SOME r);
r
end

else
C n
end
Note carefully the structure of the solution. The function C is a memoized
version of the Catalan number function. When called it consults the memopad to determine whether or not the required result has already been
computed. If so, the answer is simply retrieved from the memopad, otherwise the result is computed, stored in the cache, and returned. The function C looks superficially similar to the earlier definition of C, with the
important difference that the recursive calls are to C, rather than C itself.
This ensures that sub-computations are properly cached and that the cache
is consulted whenever possible.
The main weakness of this solution is that we must fix an upper bound
on the size of the cache. This can be alleviated by implementing a more
sophisticated cache management scheme that dynamically adjusts the size
of the cache based on the calls made to it.

13.7

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 14
Input/Output
The Standard ML Basis Library (described in Appendix A) defines a threelayer input and output facility for Standard ML. These modules provide
a rudimentary, platform-independent text I/O facility that we summarize
briefly here. The reader is referred to Appendix A for more details. Unfortunately, there is at present no standard library for graphical user interfaces; each implementation provides its own package. See your compilers
documentation for details.

14.1

Textual Input/Output

The text I/O primitives are based on the notions of an input stream and an
output stream, which are values of type instream and outstream, respectively. An input stream is an unbounded sequence of characters arising
from some source. The source could be a disk file, an interactive user, or
another program (to name a few choices). Any source of characters can
be attached to an input stream. An input stream may be thought of as
a buffer containing zero or more characters that have already been read
from the source, together with a means of requesting more input from the
source should the program require it. Similarly, an output stream is an unbounded sequence of characters leading to some sink. The sink could be a
disk file, an interactive user, or another program (to name a few choices).
Any sink for characters can be attached to an output stream. An output
stream may be thought of as a buffer containing zero or more characters
that have been produced by the program but have yet to be flushed to the

14.1 Textual Input/Output

128

sink.
Each program comes with one input stream and one output stream,
called stdIn and stdOut, respectively. These are ordinarily connected to
the users keyboard and screen, and are used for performing simple text
I/O in a program. The output stream stdErr is also pre-defined, and is
used for error reporting. It is ordinarily connected to the users screen.
Textual input and output are performed on streams using a variety of
primitives. The simplest are inputLine and print. To read a line of input
from a stream, use the function inputLine of type instream -> string. It
reads a line of input from the given stream and yields that line as a string
whose last character is the line terminator. If the source is exhausted, return the empty string. To write a line to stdOut, use the function print
of type string -> unit. To write to a specific stream, use the function
output of type outstream * string -> unit, which writes the given string
to the specified output stream. For interactive applications it is often important to ensure that the output stream is flushed to the sink (e.g., so
that it is displayed on the screen). This is achieved by calling flushOut of
type outstream -> unit. The print function is a composition of output
(to stdOut) and flushOut.
A new input stream may be created by calling the function openIn of
type string -> instream. When applied to a string, the system attempts
to open a file with that name (according to operating system-specific naming conventions) and attaches it as a source to a new input stream. Similarly, a new output stream may be created by calling the function openOut
of type string -> outstream. When applied to a string, the system attempts to create a file with that name (according to operating systemspecific naming conventions) and attaches it as a sink for a new output
stream. An input stream may be closed using the function closeIn of type
instream -> unit. A closed input stream behaves as if there is no further input available; request for input from a closed input stream yield
the empty string. An output stream may be closed using closeOut of type
outstream -> unit. A closed output stream is unavailable for further output; an attempt to write to a closed output stream raises the exception
TextIO.IO.
The function input of type instream -> string is a blocking read operation that returns a string consisting of the characters currently available
from the source. If none are currently available, but the end of source has
not been reached, then the operation blocks until at least one character is
R EVISED 11.02.11

D RAFT

V ERSION 1.2

14.2 Sample Code

129

available from the source. If the source is exhausted or the input stream
is closed, input returns the null string. To test whether an input operation would block, use the function canInput of type instream * int ->
int option. Given a stream s and a bound n, the function canInput determines whether or not a call to input on s would immediately yield up
to n characters. If the input operation would block, canInput yields NONE;
otherwise it yields SOME k, with 0 k n being the number of characters
immediately available on the input stream. If canInput yields SOME 0, the
stream is either closed or exhausted. The function endOfStream of type
instream -> bool tests whether the input stream is currently at the end
(no further input is available from the source). This condition is transitive
since, for example, another process might append data to an open file in
between calls to endOfStream.
The function output of type outstream * string -> unit writes a string
to an output stream. It may block until the sink is able to accept the entire string. The function flushOut of type outstream -> unit forces any
pending output to the sink, blocking until the sink accepts the remaining
buffered output.
This collection of primitive I/O operations is sufficient for performing
rudimentary textual I/O. For further information on textual I/O, and support for binary I/O and Posix I/O primitives, see the Standard ML Basis
Library.

14.2

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 15
Lazy Data Structures
In ML all variables are bound by value, which means that the bindings of
variables are fully evaluated expressions, or values. This general principle
has several consequences:
1. The right-hand side of a val binding is evaluated before the binding
is effected. If the right-hand side has no value, the val binding does
not take effect.
2. In a function application the argument is evaluated before being passed
to the function by binding that value to the parameter of the function. If the argument does not have a value, then neither does the
application.
3. The arguments to value constructors are evaluated before the constructed value is created.
According to the by-value discipline, the bindings of variables are evaluated, regardless of whether that variable is ever needed to complete execution. For example, to compute the result of applying the function fn
x => 1 to an argument, we never actually need to evaluate the argument,
but we do anyway. For this reason ML is sometimes said to be an eager
language.
An alternative is to bind variables by name,1 which means that the binding of a variable is an unevaluated expression, known as a computation or a
1 The

lished.

terminology is historical, and not well-motivated. It is, however, firmly estab-

131
suspension or a thunk.2 This principle has several consequences:
1. The right-hand side of a val binding is not evaluated before the binding is effected. The variable is bound to a computation (unevaluated
expression), not a value.
2. In a function application the argument is passed to the function in
unevaluated form by binding it directly to the parameter of the function. This holds regardless of whether the argument has a value or
not.
3. The arguments to value constructor are left unevaluated when the
constructed value is created.
According to the by-name discipline, the bindings of variables are only
evaluated (if ever) when their values are required by a primitive operation.
For example, to evaluate the expression x+x, it is necessary to evaluate the
binding of x in order to perform the addition. Languages that adopt the
by-name discipline are, for this reason, said to be lazy.
This discussion glosses over another important aspect of lazy evaluation, called memoization. In actual fact laziness is based on a refinement
of the by-name principle, called the by-need principle. According to the byname principle, variables are bound to unevaluated computations, and are
evaluated only as often as the value of that variables binding is required
to complete the computation. In particular, to evaluate the expression x+x
the value of the binding of x is needed twice, and hence it is evaluated
twice. According to the by-need principle, the binding of a variable is
evaluated at most once not at all, if it is never needed, and exactly once if
it ever needed at all. Re-evaluation of the same computation is avoided by
memoization. Once a computation is evaluated, its value is saved for future
reference should that computation ever be needed again.
The advantages and disadvantages of lazy versus eager languages have
been hotly debated. We will not enter into this debate here, but rather content ourselves with the observation that laziness is a special case of eagerness.
(Recent versions of) ML have lazy data types that allow us to treat unevaluated computations as values of such types, allowing us to incorporate
laziness into the language without disrupting its fundamental character
2 For

reasons that are lost in the mists of time.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

15.1 Lazy Data Types

132

on which so much else depends. This affords the benefits of laziness, but
on a controlled basis we can use it when it is appropriate, and ignore it
when it is not.
The main benefit of laziness is that it supports demand-driven computation. This is useful for representing on-line data structures that are created
only insofar as we examine them. Infinite data structures, such as the sequence of all prime numbers in order of magnitude, are one example of
an on-line data structure. Clearly we cannot ever finish creating the sequence of all prime numbers, but we can create as much of this sequence
as we need for a given run of a program. Interactive data structures, such
as the sequence of inputs provided by the user of an interactive system,
are another example of on-line data structures. In such a system the users
inputs are not pre-determined at the start of execution, but rather are created on demand in response to the progress of computation up to that
point. The demand-driven nature of on-line data structures is precisely
what is needed to model this behavior.
Note: Lazy evaluation is a non-standard feature of ML that is supported
only by the SML/NJ compiler. The lazy evaluation features must be enabled by executing the following at top level:
Compiler.Control.lazysml := true;
open Lazy;

15.1

Lazy Data Types

SML/NJ provides a general mechanism for introducing lazy data types by

simply attaching the keyword lazy to an ordinary datatype declaration.
The ideas are best illustrated by example. We will focus attention on the
type of infinite streams, which may be declared as follows:
datatype lazy a stream = Cons of a * a stream
Notice that this type definition has no base case! Had we omitted the
keyword lazy, such a datatype would not be very useful, since there
would be no way to create a value of that type!
Adding the keyword lazy makes all the difference. Doing so specifies
that the values of type typ stream are computations of values of the form
Cons (val, val0 ),
R EVISED 11.02.11

D RAFT

V ERSION 1.2

15.2 Lazy Function Definitions

133

where val is of type typ, and val0 is another such computation. Notice how
this description captures the incremental nature of lazy data structures.
The computation is not evaluated until we examine it. When we do, its
structure is revealed as consisting of an element val together with another
suspended computation of the same type. Should we inspect that computation, it will again have this form, and so on ad infinitum.
Values of type typ stream are created using a val rec lazy declaration that provides a means for building a circular data structure. Here
is a declaration of the infinite stream of 1s as a value of type int stream:
val rec lazy ones = Cons (1, ones)
The keyword lazy indicates that we are binding ones to a computation,
rather than a value. The keyword rec indicates that the computation is
recursive (or self-referential or circular). It is the computation whose underlying value is constructed using Cons (the only possibility) from the integer
1 and the very same computation itself.
We can inspect the underlying value of a computation by pattern matching. For example, the binding
val Cons (h, t) = ones
extracts the head and tail of the stream ones. This is performed by
evaluating the computation bound to ones, yielding Cons (1, ones), then
performing ordinary pattern matching to bind h to 1 and t to ones.
Had the pattern been deeper, further evaluation would be required,
as in the following binding:
val Cons (h, (Cons (h, t)) = ones
To evaluate this binding, we evaluate ones to Cons (1, ones), binding h
to 1 in the process, then evaluate ones again to Cons (1, ones), binding
h to 1 and t to ones. The general rule is pattern matching forces evaluation
of a computation to the extent required by the pattern. This is the means by
which lazy data structures are evaluated only insofar as required.

15.2

Lazy Function Definitions

The combination of (recursive) lazy function definitions and decomposition by pattern matching are the core mechanisms required to support lazy
R EVISED 11.02.11

D RAFT

V ERSION 1.2

15.2 Lazy Function Definitions

134

evaluation. However, there is a subtlety about function definitions that requires careful consideration, and a third new mechanism, the lazy function
declaration.
Using pattern matching we may easily define functions over lazy data
structures in a familiar manner. For example, we may define two functions
to extract the head and tail of a stream as follows:
fun shd (Cons (h, )) = h
fun stl (Cons ( , s)) = s
These are functions that, when applied to a stream, evaluate it, and match
it against the given patterns to extract the head and tail, respectively.
While these functions are surely very natural, there is a subtle issue that
deserves careful discussion. The issue is whether these functions are lazy
enough. From one point of view, what we are doing is decomposing a
computation by evaluating it and retrieving its components. In the case
of the shd function there is no other interpretation we are extracting a
value of type typ from a value of type typ stream, which is a computation
of a value of the form Cons (exph , expt ). We can adopt a similar viewpoint about stl, namely that it is simply extracting a component value
from a computation of a value of the form Cons (exph , expt ).
However, in the case of stl, another point of view is also possible.
Rather than think of stl as extracting a value from a stream, we may instead think of it as creating a stream out of another stream. Since streams
are computations, the stream created by stl (according to this view) should
also be suspended until its value is required. Under this interpretation the
argument to stl should not be evaluated until its result is required, rather
than at the time stl is applied. This leads to a variant notion of tail that
may be defined as follows:
fun lazy lstl (Cons ( , s)) = s
The keyword lazy indicates that an application of lstl to a stream does
not immediately perform pattern matching on its argument, but rather sets
up a stream computation that, when forced, forces the argument and extracts the tail of the stream.
The behavior of the two forms of tail function can be distinguished
using print statements as follows:

R EVISED 11.02.11

D RAFT

V ERSION 1.2

15.3 Programming with Streams

135

val rec lazy s = (print "."; Cons (1, s))

val = stl s
(* prints "." *)
(* silent *)
val = stl s
val rec lazy s = (print "."; Cons (1, s));
val = lstl s
(* silent *)
(* prints "." *)
val = stl s
Since stl evaluates its argument when applied, the . is printed when it
is first called, but not if it is called again. However, since lstl only sets
up a computation, its argument is not evaluated when it is called, but only
when its result is evaluated.

15.3

Programming with Streams

Lets define a function smap that applies a function to every element of a

stream, yielding another stream. The type of smap should be (a -> b)
-> a stream -> b stream. The thing to keep in mind is that the application of smap to a function and a stream should set up (but not compute)
another stream that, when forced, forces the argument stream to obtain
the head element, applies the given function to it, and yields this as the
head of the result.
Heres the code:
fun smap f =
let
fun lazy loop (Cons (x, s)) =
Cons (f x, loop s)
in
loop
end
We have staged the computation so that the partial application of smap
to a function yields a function that loops over a given stream, applying
the given function to each element. This loop is a lazy function to ensure that it merely sets up a stream computation, rather than evaluating
its argument when it is called. Had we dropped the keyword lazy from
the definition of the loop, then an application of smap to a function and a
stream would immediately force the computation of the head element of
R EVISED 11.02.11

D RAFT

V ERSION 1.2

15.3 Programming with Streams

136

the stream, rather than merely set up a future computation of the same
result.
To illustrate the use of smap, heres a definition of the infinite stream
of natural numbers:
val one plus = smap (fn n => n+1)
val rec lazy nats = Cons (0, one plus nats)
Now lets define a function sfilter of type
(a -> bool) -> a stream -> a stream
that filters out all elements of a stream that do not satisfy a given predicate.
fun sfilter pred =
let
fun lazy loop (Cons (x, s)) =
if pred x then
Cons (x, loop s)
else
loop s
in
loop
end
We can use sfilter to define a function sieve that, when applied to a
stream of numbers, retains only those numbers that are not divisible by a
preceding number in the stream:
fun m mod n = m - n * (m div n)
fun divides m n = n mod m = 0
fun lazy sieve (Cons (x, s)) =
Cons (x, sieve (sfilter (not o (divides x)) s))
(This example uses o for function composition.)
We may now define the infinite stream of primes by applying sieve to
the natural numbers greater than or equal to 2:
val nats2 = stl (stl nats)
val primes = sieve nats2

R EVISED 11.02.11

D RAFT

V ERSION 1.2

15.4 Sample Code

137

To inspect the values of a stream it is often useful to use the following

function that takes n 0 elements from a stream and builds a list of those
n values:
fun take 0 = nil
| take n (Cons (x, s)) = x :: take (n-1) s
Heres an example to illustrate the effects of memoization:
val rec lazy s = Cons ((print "."; 1), s)
val Cons (h, ) = s;
(* prints ".", binds h to 1 *)
val Cons (h, ) = s;
(* silent, binds h to 1 *)
Replace print ".";1 by a time-consuming operation yielding 1 as result,
and you will see that the second time we force s the result is returned
instantly, taking advantage of the effort expended on the time-consuming
operation induced by the first force of s.

15.4

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 16
Equality and Equality Types
16.1

Sample Code

Here is the code for this chapter.

Chapter 17
Concurrency
Concurrent ML (CML) is an extension of Standard ML with mechanisms
for concurrent programming. It is available as part of the Standard ML of
New Jersey compiler. The eXene Library for programming the X windows
system is based on CML.

17.1

Sample Code

Here is the code for this chapter.

Part III
The Module Language

141
The Standard ML module language comprises the mechanisms for structuring programs into separate units. Program units are called structures. A
structure consists of a collection of components, including types and values, that constitute the unit. Composition of units to form a larger unit
is mediated by a signature, which describes the components of that unit.
A signature may be thought of as the type of a unit. Large units may be
structured into hierarchies using substructures. Generic, or parameterized,
units may be defined as functors.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 18
Signatures and Structures
The fundamental constructs of the ML module system are signatures and
structures. A signature may be thought of as a description of a structure,
and a structure may correspondingly be thought of as an implementation
of a signature. Many languages have similar constructs, often with different names. Signatures are often called interfaces or package specifications,
and structures are often called implementations or packages. Whatever
the terminology, the main idea is to assign a type to a body of code as a
whole.
Unlike many languages, however, the relationship between signatures
and structures is many-to-many, rather than many-to-one or one-to-one. A
signature may describe many different structures, and a structure may satisfy many different signatures. Thus, strictly speaking, it does not make
sense to speak of the signature of a structure or the structure matching a
signature, because there can be more than one in each case. By contrast
many languages impose much more stringent conditions such as requiring that each structure have a unique signature, or that each signature arise
from a unique structure. This is not the case for ML.

18.1

Signatures

A signature is a specification, or a description, of a program unit, or structure. Structures consist of declarations of type constructors, exception constructors, and value bindings. A signature specifies some requirements
on a structure, such as what type components it must have, and what

18.1 Signatures

143

value components it must have and what must be their types. A structure matches, or implements, a signature iff it meets these requirements in a
sense that will be made precise below. The requirements are to be thought
of as descriptive in that the structure may meet more stringent requirements
than are specified by a signature by, for example, having more components than are specified, but the structure nevertheless matches any less
stringent specification. (As a limiting case, any structure matches the null
signature that imposes no requirements on it!)

18.1.1

Basic Signatures

A basic signature expression has the form sig specs end, where specs is a
sequence of specifications. There are four basic forms of specification that
may occur in specs:1
1. A type specification of the form
type (tyvar1 ,...,tyvarn ) tycon [ = typ ],
where the definition typ of tycon may or may not be present.
2. A datatype specification, which has precisely the same form as a datatype
declaration.
3. An exception specification of the form
exception excon [ of typ ],
where the type typ of excon may or may not be present.
4. A value specification of the form
val id : typ.
Each specification may refer to the type constructors introduced earlier in
the sequence. No component may be specified more than once.
Signatures may be given names using a signature binding
signature sigid = sigexp,
1 There are two other forms of specification beyond these four, substructure specifications

and sharing specifications. These will be discussed in chapter 21.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

18.1 Signatures

144

where sigid is a signature identifier and sigexp is a signature expression.

Signature identifiers are abbreviations for the signatures to which they are
bound. In practice we nearly always bind signature expressions to identifiers and refer to them by name.
Here is an illustrative example of a signature definition. We will refer
back to this definition often in the rest of this chapter.
signature QUEUE =
sig
type a queue
exception Empty
val empty : a queue
val insert : a * a queue -> a queue
val remove : a queue -> a * a queue
end
The signature QUEUE specifies a structure that must provide
1. a unary type constructor a queue,
2. a nullary exception Empty,
3. a polymorphic value empty of type a queue,
4. two polymorphic functions, insert and remove, with the specified
type schemes.
Queues are polymorphic in the type of its entriesthe same operations
are used regardless of the entry type.

18.1.2

Signature Inheritance

Signatures may be built up from one another using two principal tools,
signature inclusion and signature specialization. Each is a form of inheritance
in which a new signature is created by enriching another signature with
additional information.
Signature inclusion is used to add more components to an existing signature. For example, if we wish to add an emptiness test to the signature
QUEUE we might define the augmented signature, QUEUE WITH EMPTY, using
the following signature binding:
R EVISED 11.02.11

D RAFT

V ERSION 1.2

18.1 Signatures

145

signature QUEUE WITH EMPTY =

sig
include QUEUE
val is empty : a queue -> bool
end
As the notation suggests, the signature QUEUE is included into the body of
the signature QUEUE WITH EMPTY, and an additional component is added.
It is not strictly necessary to use include to define this signature. Indeed, we may define it directly using the following signature binding:
signature QUEUE WITH EMPTY =
sig
type a queue
exception Empty
val empty : a queue
val insert : a * a queue -> a queue
val remove : a queue -> a * a queue
val is empty : a queue -> bool
end
There is no semantic difference between the two definitions of QUEUE WITH EMPTY.
Signature inclusion is a convenience that documents the history of how
the more refined signature was created.
Signature specialization is used to augment an existing signature with
additional type definitions. For example, if we wish to refine the signature
QUEUE to specify that the type constructor a queue must be defined as a
pair of lists, we may proceed as follows:
signature QUEUE AS LISTS =
QUEUE where type a queue = a list * a list
The where type clause patches the signature QUEUE by adding a definition for the type constructor a queue.
The signature QUEUE AS LISTS may also be defined directly as follows:

R EVISED 11.02.11

D RAFT

V ERSION 1.2

18.1 Signatures

146

signature QUEUE AS LISTS =

sig
type a queue = a list * a list
exception Empty
val empty : a queue
val insert : a * a queue -> a queue
val remove : a queue -> a * a queue
end
A where type clause may not be used to re-define a type that is already
defined in a signature. For example, the following is illegal:
signature QUEUE AS LISTS AS LIST =
QUEUE AS LISTS where type a queue = a list
If you wish to replace the definition of a type constructor in a signature
with another definition using where type, you must go back to a common
ancestor in which that type is not yet defined.
signature QUEUE AS LIST =
QUEUE where type a queue = a list
Two signatures are said to be equivalent iff they differ only up to the
type equivalences induced by type abbreviations. For example, the signature QUEUE where type a queue = a list is equivalent to the signature
signature QUEUE AS LIST =
sig
type a queue = a list
exception Empty
val empty : a list
val insert : a * a list -> a list
val remove : a list -> a * a list
end
Within the scope of the definition of the type a queue as a list, the two
are equivalent, and hence the specifications of the value components are
equivalent.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

18.2 Structures

18.2

147

Structures

A structure is a unit of program consisting of a sequence of declarations

of types, exceptions, and values. Structures are implementations of signatures; signatures are the types of structures.

18.2.1

Basic Structures

The basic form of structure is an encapsulated sequence of declarations

of the form struct decs end. The declarations in decs are of one of the
following four forms:
1. A type declaration defining a type constructor.
2. A datatype declaration defining a new datatype.
3. An exception declaration defining an exception with a specified argument type (if any).
4. A value declaration defining a new value variable with a specified
type.
These are, of course, just the declarations that were introduced in Part II
of this book.
A structure expression is well-formed iff it consists of a well-formed sequence of well-formed declarations (according to the rules given in Part II).
A structure expression is evaluated by evaluating each of the declarations
within it, in the order given. This amounts to evaluating the right-hand
sides of each value declaration in turn to determine its value, which is
then bound to the corresponding value identifier. Any effects that arise
from evaluating a value binding occur when the structure expression is
evaluated. A structure value is a structure expression in which all bindings are fully evaluated.
A structure may be bound to a structure identifier using a structure
binding of the form
structure strid = strexp
This declaration defines strid to stand for the value of strexp. Such a declaration is well-formed exactly when strexp is well-formed. It is evaluated by
R EVISED 11.02.11

D RAFT

V ERSION 1.2

18.2 Structures

148

evaluating the right-hand side, and binding the resulting structure value
to strid.
Here is an example of a structure binding:
structure Queue =
struct
type a queue = a list * a list
exception Empty
val empty = (nil, nil)
fun insert (x, (b,f)) = (x::b, f)
fun remove (nil, nil) = raise Empty
| remove (bs, nil) = remove (nil, rev bs)
| remove (bs, f::fs) = (f, (bs, fs))
end
Recall that a fun binding is really an abbreviation for a val rec binding,
and hence constitutes a value binding (of function type).

18.2.2

Long and Short Identifiers

Once a structure has been bound to a structure identifier, we may access

its components using paths, or long identifiers, or qualified names. A path has
the form strid.id.2 It stands for the id component of the structure bound
to strid. For example, Queue.empty refers to the empty component of the
structure Queue. It has type a Queue.queue (note well the syntax!), stating
that it is a polymorphic value whose type is built up from the unary type
constructor Queue.queue. Similarly, the function Queue.insert has type
a * a Queue.queue -> a Queue.queue and Queue.remove has type a
Queue.queue -> a * a Queue.queue.
Type definitions permeate structure boundaries. For example, the type
a Queue.queue is equivalent to the type a list because it is defined to
be so in the structure Q. We will shortly introduce the means for limiting
the visibility of type definitions. Unless special steps are taken, the definitions of types within a structure determine the definitions of the long
identifiers that refer to those types within a structure. Consequently, it is
correct to write an expression such as
2 In

chapter 21, we will generalize this to admit an arbitrary sequence of strids separated by a dot.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

18.2 Structures

149

val q = Queue.insert (1, ([6,5,4],[1,2,3]))

even though the list [6,5,4] was not obtained by using the operations
from the structure Q. This is because the type int Queue.queue is equivalent to the type int list, and hence the call to insert is well-typed.
The use of long identifiers can get out of hand, cluttering the program, rather than clarifying it. Suppose that we are frequently using the
Queue operations in a program, so that the code is cluttered with calls to
Queue.empty, Queue.insert, and Queue.remove. One way to reduce clutter is to introduce a structure abbreviation of the form structure strid =
strid0 that introduces one structure identifier as an abbreviation for another. For example, after declaring
structure Q = Queue
we may write Q.empty, Q.insert, and Q.remove, rather than the more verbose forms mentioned above.
Another way to reduce clutter is to open the structure Queue to incorporate its bindings directly into the current environment. An open declaration
has the form
open strid1 ... stridn
which incorporates the bindings from the given structures in left-to-right
order (later structures override earlier ones when there is overlap). For
example, the declaration
open Queue
incorporates the body of the structure Queue into the current environment
so that we may write just empty, insert, and remove, without qualification,
to refer to the corresponding components of the structure Queue.
Although this is surely convenient, using open has its disadvantages.
One is that we cannot simultaneously open two structures that have a
component with the same name. For example, if we write
open Queue Stack
where the structure Stack also has a component empty, then uses of empty
(without qualification) will stand for Stack.empty, not for Queue.empty.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

18.3 Sample Code

150

Another problem with open is that it is hard to control its behavior,

since it incorporates the entire body of a structure, and hence may inadvertently shadow identifiers that happen to be also used in the structure. For
example, if the structure Queue happened to define an auxiliary function
helper, that function would also be incorporated into the current environment by the declaration open Queue, which may not have been intended.
This turns out to be a source of many bugs; it is best to use open sparingly,
and only then in a let or local declaration (to limit the damage).

18.3

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 19
Signature Matching
When does a structure implement a signature? The structure must provide all of the components and satisfy all of the type definitions required by
the signature. Type components must be provided with the same number
of arguments and with an equivalent definition (if any) to that given in
the signature. Value components must be present with the type specified
in the signature, up to the definitions of any preceding types. Exception
components must be present with the type of argument (if any) specified
in the signature.
These simple principles have a number of important consequences.
To minimize bureaucracy, a structure may provide more components
than are strictly required by the signature. If a signature requires
components x, y, and z, it is sufficient for the structure to provide x,
y, z, and w.
To enhance reuse, a structure may provide values with more general
types than are required by the signature. If a signature demands a
function of type int->int, it is enough to provide a function of type
a->a.
To avoid over-specification, a datatype may be provided where a type
is required, and a value constructor may be provided where a value
is required.
To increase flexibility, a structure may consist of declarations presented
in any sensible order, not just the order specified in the signature,
provided that the requirements of the specification are met.

19.1 Principal Signatures

19.1

152

Principal Signatures

There is a most stringent, or most precise, signature for a structure, called its
principal signature. A structure may be considered to match a signature exactly when the specified signature is no more restrictive than the principal
signature of the structure. To determine whether a structure matches a signature, it is enough to check whether the principal signature of that structure is no weaker than the specified signature. For the purposes of type
checking, the principal signature is the official proxy for the structure. We
need never examine the code of the structure durng type checking, once
its principal signature has been determined.
A structure expression is assigned a principal signature by a componentby-component analysis of its constituent declarations. The principal signature of a structure is obtained as follows:1
1. Corresponding to a declaration of the form
type (tyvar1 ,...,tyvarn ) tycon = typ,
the principal signature contains the specification
type (tyvar1 ,...,tyvarn ) tycon = typ
The principal signature includes the definition of tycon.
2. Corresponding to a declaration of the form
datatype (tyvar1 ,...,tyvarn ) tycon =
con1 of typ1 | ... | conk of typk
the principal signature contains the specification
datatype (tyvar1 ,...,tyvarn ) tycon =
con1 of typ1 | ... | conk of typk
The specification is identical to the declaration.
3. Corresponding to a declaration of the form
1 These

rules gloss over some technical complications that arise only in unusual circumstances. See The Definition of Standard ML [3] for complete details.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

19.2 Matching

153

exception id of typ
the principal signature contains the specification
exception id of typ
4. Corresponding to a declaration of the form
val id = exp
the principal signature contains the specification
val id : typ
where typ is the principal type of the expression exp (relative to the
preceding declarations).
In brief, the principal signature contains all of the type definitions, datatype
definitions, and exception bindings of the structure, plus the principal
types of its value bindings.

19.2

Matching

A candidate signature sigexpc is said to match a target signature sigexpt iff

sigexpc has all of the components and all of the type equations specified by
sigexpt . More precisely,
1. Every type constructor in the target must also be present in the candidate, with the same arity (number of arguments) and an equivalent
definition (if any).
2. Every datatype in the target must be present in the candidate, with
equivalent types for the value constructors.
3. Every exception in the target must be present in the candidate, with
an equivalent argument type.
4. Every value in the target must be present in the candidate, with at
least as general a type.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

19.2 Matching

154

The candidate may have additional components not mentioned in the target, or satisfy additional type equations not required in the target, but it
cannot have fewer of either. The target signature may therefore be seen
as a weakening of the candidate signature, since all of the properties of the
latter are true of the former.
The matching relation is reflexiveevery signature matches itself
and transitive if sigexp1 matches sigexp2 and sigexp2 matches sigexp3 , then
sigexp1 matches sigexp3 . Two signatures are equivalent (in the sense of
Chapter chapter 18) iff each matches the other, which is to say that they
are equivalently restrictive.
It will be helpful to consider some examples. Recall the following signatures from chapter 18.
signature QUEUE =
sig
type a queue
exception Empty
val empty : a queue
val insert : a * a queue -> a queue
val remove : a queue -> a * a queue
end
signature QUEUE WITH EMPTY =
sig
include QUEUE
val is empty : a queue -> bool
end
signature QUEUE AS LISTS =
QUEUE where type a queue = a list * a list
The signature QUEUE WITH EMPTY matches the signature QUEUE, because
all of requirements of QUEUE are met by QUEUE WITH EMPTY. The converse
does not hold, because QUEUE lacks the component is empty, which is required by QUEUE WITH EMPTY.
The signature QUEUE AS LISTS matches the signature QUEUE. It is identical to QUEUE, apart from the additional specification of the type a queue.
The converse fails, because the signature QUEUE does not satisfy the requirement that a queue be equivalent to a list * a list.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

19.2 Matching

155

Matching does not distinguish between equivalent signatures. For example, consider the following signature:
signature QUEUE AS LIST = sig
type a queue = a list
exception Empty
val empty : a list
val insert : a * a list -> a list
val remove : a list -> a * a list
val is empty : a list -> bool
end
At first glance you might think that this signature does not match the signature QUEUE, since the components of QUEUE AS LIST have superficially
dissimilar types from those in QUEUE. However, the signature QUEUE AS LIST
is equivalent to the signature QUEUE with type a queue = a list, which
matches QUEUE for reasons noted earlier. Therefore, QUEUE AS LIST matches
QUEUE as well.
Signature matching may also involve instantiation of polymorphic types.
The types of values in the candidate may be more general than required
by the target. For example, the signature
signature MERGEABLE QUEUE =
sig
include QUEUE
val merge : a queue * a queue -> a queue
end
matches the signature
signature MERGEABLE INT QUEUE =
sig
include QUEUE
val merge : int queue * int queue -> int queue
end
because the polymorphic type of merge in MERGEABLE QUEUE instantiates to
its type in MERGEABLE INT QUEUE.
Finally, a datatype specification matches a signature that specifies a
type with the same name and arity (but no definition), and zero or more
R EVISED 11.02.11

D RAFT

V ERSION 1.2

19.2 Matching

156

value components corresponding to some (or all) of the value constructors

of the datatype. The types of the value components must match exactly the
types of the corresponding value constructors; no specialization is allowed
in this case. For example, the signature
signature RBT DT =
sig
datatype a rbt =
Empty |
Red of a rbt * a * a rbt |
Black of a rbt * a * a rbt
end
matches the signature
signature RBT =
sig
type a rbt
val Empty : a rbt
val Red : a rbt * a * a rbt -> a rbt
end
The signature RBT specifies the type a rbt as abstract, and includes two
value specifications that are met by value constructors in the signature
RBT DT.
One way to understand this is to mentally rewrite the signature RBT DT
in the (fanciful) form2
signature RBT DTS =
sig
type a rbt
con Empty : a rbt
con Red : a rbt * a * a rbt -> a rbt
con Black : a rbt * a * a rbt -> a rbt
The rule is simply that a val specification may be matched by a con specification.
2 Unfortunately,

the signature RBT DTS is not a legal ML signature!

R EVISED 11.02.11

D RAFT

V ERSION 1.2

19.3 Satisfaction

19.3

157

Satisfaction

Returning to the motivating question of this chapter, a candidate structure

implements a target signature iff the principal signature of the candidate
structure matches the target signature. By the reflexivity of the matching relation it is immediate that a structure satisfies its principal signature. Therefore any signature implemented by a structure is weaker than
the principal signature of that structure. That is, the principal signature is
the strongest signature implemented by a structure.

19.4

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 20
Signature Ascription
Signature ascription imposes the requirement that a structure implement a
signature and, in so doing, weakens the signature of that structure for all
subsequent uses of it. There are two forms of ascription in ML. Both require that a structure implement a signature; they differ in the extent to
which the assigned signature of the structure is weakened by the ascription.
1. Transparent, or descriptive ascription. The structure is assigned the
signature obtained by propagating type definitions from the principal
signature to the candidate signature.
2. Opaque, or restrictive ascription. The structure is assigned the target
signature as is, without propagating any type definitions.
In either case the components of the structure are cut down to those specified in the signature; the only difference is whether type definitions are
propagated from the principal signature or not.

20.1

Ascribed Structure Bindings

The most common form of signature ascription is in a structure binding.

There are two forms, the transparent
structure strid : sigexp = strexp
and the opaque

20.2 Opaque Ascription

159

structure strid :> sigexp = strexp

Transparent ascription is written using a single colon, :, whereas opaque
ascription is written using :>.
Ascribed structure bindings are type checked as follows. First, we
check that strexp implements sigexp according to the rules given in chapter 19. The principal signature sigexp0 of strexp is determined, and checked
to match sigexp. The matching process determines an augmentation sigexp0
of sigexp in which we propagate type equations from the principal signature, sigexp0 , to sigexp0 . Second, the structure identifier is assigned a signature based on the form of ascription. For an opaque ascription, it is
assigned the signature sigexp; for a transparent ascription, it is assigned
the signature sigexp0 .
Ascribed signature bindings are evaluated by first evaluating strexp.
Then a view of the resulting value is formed by dropping all components
that are not present in the target signature, sigexp. The structure variable
strid is bound to the view. The formation of the view ensures that the
components of a structure may always be accessed in constant time, and
that there are no space leaks because of components that are present in
the structure, but not in the signature.

20.2

Opaque Ascription

The primary use of opaque ascription is to enforce data abstraction. A

good example is provided by the implementation of queues as, say, pairs
of lists.
structure Queue :> QUEUE =
struct
type a queue = a list * a list
val empty = (nil, nil)
fun insert (x, (bs, fs)) = (x::bs, fs)
exception Empty
fun remove (nil, nil) = raise Empty
| remove (bs, f::fs) = (f, (bs, fs))
| remove (bs, nil) = remove (nil, rev bs)
end

R EVISED 11.02.11

D RAFT

V ERSION 1.2

20.2 Opaque Ascription

160

The use of opaque ascription ensures that the type a Queue.queue is abstract. No definition is provided for it in the signature QUEUE, and therefore it has no definition in terms of other types of the language; the type
a Queue.queue is abstract.
For the type a Queue.queue to be abstract means that the only operations that may be performed on values of that type are empty, insert, and
remove. Importantly, we may not make use of the fact that a queue is really
a pair of lists on the grounds that it is implemented this way. We have
obscured this fact by opaquely ascribing a signature that does not provide
a definition for the type a queue. All clients of the structure Queue are
insulated from the details of how queues are implemented. Consequently,
the implementation of queues can be changed without breaking any client
code, as long as the new implementation satisfies the same signature.
Hiding the representation of a type allows us to isolate the enforcement
of representation invariants on a data structure. We may think of the type
a Queue.queue as the type of states of an abstract machine whose sole
instructions are empty (the initial state), insert, and remove. Internally
to the structure Queue we may wish to impose invariants on the internal
state of the machine. The beauty of data abstraction is that it provides
an elegant means of enforcing such invariants, called the assume-ensure,
or rely-guarantee, method. It reduces the enforcement of representation
invariants to these two requirements:
1. All initialization instructions must ensure that the invariant holds
true of the machine state after execution.
2. All state transition instructions may assume that the invariant holds
of the inputs states, and must ensure that it holds of the output state.
By induction on the number of instructions executed, the invariant must
hold for all statesit must really be invariant!
Suppose that we wish to implement an abstract type of priority queues
for an arbitrary element type. The queue operations are no longer polymorphic in the element type because they actually touch the elements to
determine their relative priorities. Here is a possible signature for priority
queues that expresses this dependency:1
1 In chapter 21 well introduce better means for structuring this module, but the central

points discussed here will not be affected.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

20.2 Opaque Ascription

161

signature PQ =
sig
type elt
val lt : elt * elt -> bool
type queue
exception Empty
val empty : queue
val insert : elt * queue -> queue
val remove : queue -> elt * queue
end
Now let us consider an implementation of priority queues in which
the elements are taken to be strings. Since priority queues form an abstract type, we would expect to use opaque ascription to ensure that its
representation is hidden. This suggests an implementation along these
lines:
structure PrioQueue :> PQ =
struct
type elt = string
val lt : string * string -> bool = (op <)
type queue = ...
.
.
.
end
But not only is the type PrioQueue.queue abstract, so is PrioQueue.elt!
This leaves us no means of creating a value of type PrioQueue.elt, and
hence we can never call PrioQueue.insert. The problem is that the interface is too abstract it should only obscure the identity of the type
queue, and not that of the type elt.
The solution is to augment the signature PQ with a definition for the
type elt, then opaquely ascribe this to PrioQueue:
signature STRING PQ = PQ where type elt = string
structure PrioQueue :> STRING PQ = ...
Now the type PrioQueue.elt is equivalent to string, and we may call
PrioQueue.insert with a string, as expected.
The moral is that there is always an element of judgement involved in
deciding which types to hold abstract, and which to make opaque. In the
R EVISED 11.02.11

D RAFT

V ERSION 1.2

20.3 Transparent Ascription

162

case of priority queues, the determining factor is that we specified only the
operations on elt that were required for the implementation of priority
queues, and no others. This means that elt could not usefully be held
abstract, but must instead be specified in the signature. On the other hand
the operations on queues are intended to be complete, and so we hold the
type abstract.

20.3

Transparent Ascription

Transparent ascription cuts down on the need for explicit specification of

type definitions in signatures. As we remarked earlier, we can always replace uses of transparent ascription by a use of opaque ascription with a
hand-crafted augmented signature. This can become burdensome. On the
other hand, excessive use of transparent ascription impedes modular programming by exposing type information that would better be left abstract.
The prototypical use of transparent ascription is to form a view of a
structure that eliminates the components that are not necessary in a given
context without obscuring the identities of its type components. Consider
the signature ORDERED defined as follows:
signature ORDERED =
sig
type t
val lt : t * t -> bool
end
This signature specifies a type t equipped with a comparison operation
lt.
It should be clear that it would not make sense to opaquely ascribe
this signature to a structure. Doing so would preclude ever calling the lt
operation, for there would be no means of creating values of type t. Such
a signature is only useful once it has been augmented with a definition for
the type t. This is precisely what transparent ascription does for you.
For example, consider the following structure binding:

R EVISED 11.02.11

D RAFT

V ERSION 1.2

20.3 Transparent Ascription

163

structure String : ORDERED =

struct
type t = string
val clt = Char.<
fun lt (s, t) = ... clt ...
end
This structure implements string comparison in terms of character comparison (say, to implement the lexicographic ordering of strings). Ascription of the signature ORDERED ensures two things:
1. The auxiliary function clt is pruned out of the structure. It was intended for internal use, and was not meant to be externally visible.
2. The type String.t is equivalent to string, even though this fact is
not present in the signature ORDERED. Transparent ascription computes an augmentation of ORDERED with this definition exposed. The
true signature of String is the signature
ORDERED where type t = string
which makes clear the underlying definition of t.
A related use of transparent ascription is to document an interpretation of a type without rendering it abstract. For example, we may wish
to consider the integers ordered in two different ways, one by the standard arithmetic comparison, the other by divisibility. We might make the
following declarations to express this:
structure IntLt : ORDERED =
struct
type t = int
val lt = (op <)
end
structure IntDiv : ORDERED =
struct
type t = int
fun lt (m, n) = (n mod m = 0)
end

R EVISED 11.02.11

D RAFT

V ERSION 1.2

20.4 Sample Code

164

The ascription specifies the interpretation of int as partially ordered, in

two senses, but does not hide the type of elements. In particular, IntLt.t
and IntDiv.t are both equivalent to int.

20.4

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 21
Module Hierarchies
So far we have confined ourselves to considering flat modules consisting of a linear sequence of declarations of types, exceptions, and values.
As programs grow in size and complexity, it becomes important to introduce further structuring mechanisms to support their growth. The ML
module language also supports module hierarchies, tree-structured configurations of modules that reflect the architecture of a large system.

21.1

Substructures

A substructure is a structure within a structure. Structures bindings (either opaque or transparent) are admitted as components of other structures. Structure specifications of the form
structure strid : sigexp
may appear in signatures. There is no distinction between transparent
and opaque specifications in a signature, because there is no structure to
ascribe!
The type checking and evaluation rules for structures are extended
to substructures recursively. The principal signature of a sub-structure
binding is determined according to the rules given in chapter 19. A substructure binding in one signature matches the corresponding one in another iff their signatures match according to the rules in chapter 19. Evaluation of a sub-structure binding consists of evaluating the structure expression, then binding the resulting structure value to that identifier.

21.1 Substructures

166

To see how substructures arise in practice, consider the following programming scenario. The first version of a system makes use of a polymorphic dictionary data structure whose search keys are strings. The signature
for such a data structure might be as follows:
signature MY STRING DICT =
sig
type a dict
val empty : a dict
val insert : a dict * string * a -> a dict
val lookup : a dict * string -> a option
end
The return type of lookup is a option, since there may be no entry in the
dictionary with the specified key.
The implementation of this abstraction looks approximately like this:
structure MyStringDict :> MY STRING DICT =
struct
datatype a dict =
Empty |
Node of a dict * string * a * a dict
val empty = Empty
fun insert (d, k, v) = ...
fun lookup (d, k) = ...
end
The omitted implementations of insert and lookup make use of the builtin lexicographic ordering of strings.
The second version of the system requires another dictionary whose
keys are integers, leading to another signature and implementation for
dictionaries.
signature MY INT DICT =
sig
type a dict
val empty : a dict
val insert : a dict * int * a -> a dict
val lookup : a dict * int -> a option
end
R EVISED 11.02.11

D RAFT

V ERSION 1.2

21.1 Substructures

167

structure MyIntDict :> MY INT DICT =

sig
datatype a dict =
Empty |
Node of a dict * int * a * a dict
val empty = Empty
fun insert (d, k, v) = ...
fun lookup (d, k) = ...
end
The ellided implementations of insert and lookup make use of the primitive comparison operations on integers.
At this point we may observe an obvious pattern, that of a dictionary
with keys of a specific type. To avoid further repetition we decide to abstract out the key type from the signature so that it can be filled in later.
signature MY GEN DICT =
sig
type key
type a dict
val empty : a dict
val insert : a dict * key * a -> a dict
end
Notice that the dictionary abstraction carries with it the type of its keys.
Specific instances of this generic dictionary signature are obtained using where type.
signature MY STRING DICT =
MY GEN DICT where type key = string
signature MY INT DICT =
MY GEN DICT where type key = int
A string dictionary might then be implemented as follows:
structure MyStringDict :> MY STRING DICT =
struct
type key = string
datatype a dict =
Empty |
R EVISED 11.02.11

D RAFT

V ERSION 1.2

21.1 Substructures

168

Node of a dict * key * a * a dict

val empty = Empty
fun insert (Empty, k, v) = Node (Empty, k, v, Empty)
fun lookup (Empty, ) = NONE
| lookup (Node (dl, l, v, dr), k) =
if k < l then
(* string comparison *)
lookup (dl, k)
else if k > l then
(* string comparison *)
lookup (dr, k)
else
v
end
By a similar process we may build an implementation MyIntDict of the
signature MY INT DICT, with integer keys ordered by the standard integer
comparison operations.
Now suppose that we require a third dictionary, with integers as keys,
but ordered according to the divisibility ordering.1 This implementation,
say MyIntDivDict, makes use of modular arithmetic to compare operations, but has the same signature MY INT DICT as MyIntDict.
structure MyIntDivDict :> MY INT DICT =
struct
type key = int
datatype a dict =
Empty |
Node of a dict * key * a * a dict
fun divides (k, l) = (l mod k = 0)
val empty = Empty
fun insert (None, k, v) = Node (Empty, k, v, Empty)
fun lookup (Empty, ) = NONE
| lookup (Node (dl, l, v, dr), k) =
if divides (k, l) then
(* divisibility test *)
lookup (dl, k)
else if divides (l, k) then (* divisibility test *)
lookup (dr, k)
else
1 For

which, m < n iff m divides n evenly.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

21.1 Substructures

169

v
end
Notice that we required an auxiliary function, divides, to implement the
comparison in the required sense.
With this in mind, let us re-consider our initial attempt to consolidate
the signatures of the various versions of dictionaries in play. In one sense
there is nothing to do the signature MY GEN DICT suffices. However,
as weve just seen, the instances of this signature, which are ascribed to
particular implementations, do not determine the interpretation. What
wed like to do is to package the type with its interpretation so that the
dictionary module is self-contained. Not only does the dictionary module
carry with it the type of its keys, but it also carries the interpretation used
on that type.
This is achieved by introducing a substructure binding in the dictionary structure. To begin with we first isolate the notion of an ordered
type.
signature ORDERED =
sig
type t
val lt : t * t -> bool
val eq : t * t -> bool
end
This signature describes modules that contain a type t equipped with an
equality and comparison operation on it.
An implementation of this signature specifies the type and the interpretation, as in the following examples.
(* Lexicographically ordered strings. *)
structure LexString : ORDERED =
struct
type t = string
val eq = (op =)
val lt = (op <)
end
(* Integers ordered conventionally. *)
structure LessInt : ORDERED =
R EVISED 11.02.11

D RAFT

V ERSION 1.2

21.1 Substructures

170

struct
type t = int
val eq = (op =)
val lt = (op <)
end
(* Integers ordered by divisibility.*)
structure DivInt : ORDERED =
struct
type t = int
fun lt (m, n) = (n mod m = 0)
fun eq (m, n) = lt (m, n) andalso lt (n, m)
end
Notice that the use of transparent ascription is very natural here, since
ORDERED is not intended as a self-contained abstraction.
The signature of dictionaries is re-structured as follows:
signature DICT =
sig
structure Key : ORDERED
type a dict
val empty : a dict
val insert : a dict * Key.t * a -> a dict
val lookup : a dict * Key.t -> a option
end
The signature DICT includes as a substructure the key type together with
its interpretation as an ordered type.
To enforce abstraction we introduce specialized versions of this signature that specify the key type using a where type clause.
signature STRING DICT =
DICT where type Key.t=string
signature INT DICT =
DICT where type Key.t=int
These are, respectively, signatures for the abstract type of dictionaries whose
keys are strings and integers.
How are these signatures to be implemented? Corresponding to the
layering of the signatures, we have a layering of the implementation.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

21.1 Substructures

171

structure StringDict :> STRING DICT =

struct
structure Key : ORDERED = LexString
datatype a dict =
Empty |
Node of a dict * Key.t * a * a dict
val empty = Empty
fun insert (None, k, v) = Node (Empty, k, v, Empty)
fun lookup (Empty, ) = NONE
| lookup (Node (dl, l, v, dr), k) =
if Key.lt(k, l) then
lookup (dl, k)
else if Key.lt (l, k) then
lookup (dr, k)
else
v
end
Observe that the implementation of insert and lookup make use of the
comparison operations Key.lt and Key.eq.
Similarly, we may implement IntDict, with the standard ordering, as
follows:
structure LessIntDict :> INT DICT =
struct
structure Key : ORDERED = LessInt
datatype a dict =
Empty |
Node of a dict * Key.t * a * a dict
val empty = Empty
fun insert (None, k, v) = Node (Empty, k, v, Empty)
fun lookup (Empty, ) = NONE
| lookup (Node (dl, l, v, dr), k) =
if Key.lt(k, l) then
lookup (dl, k)
else if Key.lt (l, k) then
lookup (dr, k)
else
v
R EVISED 11.02.11

D RAFT

V ERSION 1.2

21.1 Substructures

172

end
Similarly, dictionaries with integer keys ordered by divisibility may be
implemented as follows:
structure IntDivDict :> INT DICT =
struct
structure Key : ORDERED = IntDiv
datatype a dict =
Empty |
Node of a dict * Key.t * a * a dict
val empty = Empty
fun insert (None, k, v) = Node (Empty, k, v, Empty)
fun lookup (Empty, ) = NONE
| lookup (Node (dl, l, v, dr), k) =
if Key.lt(k, l) then
lookup (dl, k)
else if Key.lt (l, k) then
lookup (dr, k)
else
v
end
Taking stock of the development, what we have done is to structure
the signature of dictionaries to allow the type of keys, together with its
interpretation, to vary from one implementation to another. The Key substructure may be viewed as a parameter of the signature DICT that is instantiated by specialization to specific types of interest. In this sense substructures subsume the notion of a parameterized signature found in some
languages. There are several advantages to this:
1. A signature with one or more substructures is still a complete signature. Parameterized signatures, in contrast, are incomplete signatures that must be completed to be used.
2. Any substructure of a signature may play the role of a parameter.
There is no need to designate in advance which are arguments and
which are results.
In chapter 23 we will introduce the mechanisms needed to build a
generic implementation of dictionaries that may be instantiated by the key
type and its ordering.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

21.2 Sample Code

21.2

173

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 22
Sharing Specifications
In chapter 21 we illustrated the use of substructures to express the dependence of one abstraction on another. In this chapter we will consider the
problem of symmetric combination of modules to form larger modules.

22.1

Combining Abstractions

The discussion will be based on a representation of geometry in ML based

on the following (drastically simplified) signature.
signature GEOMETRY =
sig
structure Point : POINT
structure Sphere : SPHERE
end
For the purposes of this example, we have reduced geometry to two concepts, that of a point in space and that of a sphere.
Points and vectors are fundamental to representing geometry. They are
described by the following (abbreviated) signatures:
signature VECTOR =
sig
type vector
val zero : vector
val scale : real * vector -> vector

22.1 Combining Abstractions

175

val add : vector * vector -> vector

val dot : vector * vector -> real
end
signature POINT =
sig
structure Vector : VECTOR
type point
(* move a point along a vector *)
val translate : point * Vector.vector -> point
(* the vector from a to b *)
val ray : point * point -> Vector.vector
end
The vector operations support addition, scalar multiplication, and inner
product, and include a unit element for addition. The point operations
support translation of a point along a vector and the creation of a vector as
the difference of two points (i.e., the vector from the first to the second).
Spheres are implemented by a module implementing the following
(abbreviated) signature:
signature SPHERE =
sig
structure Vector : VECTOR
structure Point : POINT
type sphere
val sphere : Point.point * Vector.vector -> sphere
end
The operation sphere creates a sphere centered at a given point and with
the radius vector given.
These signatures are intentionally designed so that the dimension of
the space is not part of the specification. This allows us using the mechanisms to be introduced in chapter 23 to build packages that work in
an arbitrary dimension without requiring run-time conformance checks.
It is the structures, and not the signatures, that specify the dimension.
Two- and three-dimensional geometry are defined by structure bindings
like these:
structure Geom2D :> GEOMETRY = ...
structure Geom3D :> GEOMETRY = ...
R EVISED 11.02.11

D RAFT

V ERSION 1.2

22.1 Combining Abstractions

176

As a consequence of the use of opaque ascription, the types Geom2D.Point.point

and Geom3D.Point.point are distinct. This means that dimensional conformance is enforced by the type checker. For example, we cannot apply
Geom3D.Sphere.sphere to a point in two-space and a vector in three-space.
This is a good thing: the more static checking we have, the better off
we are. Closer inspection reveals that, unfortunately, we have too much
of a good thing. Suppose that p and q are two-dimensional points of type
Geom2D.Point.point. We might expect to be able to form a sphere centered at p with radius determined by the vector from p to q:
Geom2D.Sphere.sphere (p, Geom2D.Point.ray (p, q)).
But this expression is ill-typed! The reason is that the types Geom2D.Point.Vector
and Geom2D.Sphere.Vector.vector are also distinct from one another, which
is not at all what we intend.
What has gone wrong? The situation is quite subtle. In keeping with
the guidelines discussed in section 21.1, we have incorporated as substructures the structures on which a given structure depends. For example,
forming a ray from one point to another yields a vector, so an implementation of POINT depends on an implementation of VECTOR. Thus, POINT has
a substructure implementing VECTOR, and, similarly, SPHERE has substructures implementing VECTOR and POINT.
This leads to a proliferation of structures. Even in the very simplified geometry signature given above, we have two copies of the point
abstraction, and three copies of the vector abstraction! Since we used
opaque ascription to define the two- and three-dimensional implementations of the signature GEOMETRY, all of these abstractions are kept distinct
from one another, even though they may be implemented identically.
In a sense this is the correct state of affairs. The various copies of,
say, the vector abstraction might well be distinct from one another. In
the elided implementation of two-dimensional geometry, we might have
used completely incompatible notions of vector in each of the three places
where they are required. Of course, this may not be what is intended, but
(so far) there is nothing in the signature to prevent it. Hence, we are compelled
to keep these types distinct.
What is missing is the expression of the intention that the various copies
of vectors and points within the geometry abstraction be identical, so that
we can mix-and-match the vectors constructed in various components of
R EVISED 11.02.11

D RAFT

V ERSION 1.2

22.1 Combining Abstractions

177

the package. To support this it is necessary to constrain the implementation to use the same notion of vector throughout. This is achieved using
a type sharing constraint. The revised signatures for the geometry package
look like this:
signature SPHERE =
sig
structure Vector : VECTOR
structure Point : POINT
sharing type Point.Vector.vector = Vector.vector
type sphere
val sphere : Point.point * Vector.vector -> sphere
end
signature GEOMETRY =
sig
structure Point : POINT
structure Sphere : SPHERE
sharing type Point.point = Sphere.Point.point
and Point.Vector.vector = Sphere.Vector.vector
end
These equations specify that the two copies of the point abstraction, and
the three copies of the vector abstraction must coincide. In the presence
of the above sharing specification, the ill-typed expression above becomes
well-typed, since now the required type equation holds by explicit specification in the signature.
As a notational convenience we may use a structure sharing constraint
instead to express the same requirements:
signature SPHERE =
sig
structure Vector : VECTOR
structure Point : POINT
sharing Point.Vector = Vector
type sphere
val sphere : Point.point * Vector.vector -> sphere
end
signature GEOMETRY =
sig
R EVISED 11.02.11

D RAFT

V ERSION 1.2

22.1 Combining Abstractions

178

structure Point : POINT

structure Sphere : SPHERE
sharing Point = Sphere.Point
and Point.Vector = Sphere.Vector
end
Rather than specify the required sharing type-by-type, we can instead
specify it structure-by-structure, with the meaning that corresponding types
of shared structures are required to share. Since each structure in our example contains only one type, the effect of the structure sharing specification above is identical to the preceding type sharing specification.
Not only does the sharing specification ensure that the desired equations hold amongst the various components of an implementation of GEOMETRY,
but it also constrains the implementation to ensure that these types are the
same. It is easy to achieve this requirement by defining a single implementation of points and vectors that is re-used in the higher-level abstractions.
structure Vector3D : VECTOR = ...
structure Point3D : POINT =
struct
structure Vector : VECTOR = Vector3D
.
.
.
end
structure Sphere3D : SPHERE =
struct
structure Vector : VECTOR = Vector3D
structure Point : POINT = Point3D
.
.
.
end
structure Geom3D :> GEOMETRY =
struct
structure Point = Point3D
structure Sphere = Sphere3D
end
The required type sharing constraints are true by construction.
Had we instead replaced the above declaration of Geom3D by the following one, the type checker would reject it on the grounds that the reR EVISED 11.02.11

D RAFT

V ERSION 1.2

22.1 Combining Abstractions

179

quired sharing between Sphere.Point and Point does not hold, because
Sphere2D.Point is distinct from Point3D.
structure Geom3D :> GEOMETRY =
struct
structure Point = Point3D
structure Sphere = Sphere2D
end
It is natural to wonder whether it might be possible to restructure the
GEOMETRY signature so that the duplication of the point and vector components is avoided, thereby obviating the need for sharing specifications.
One can re-structure the code in this manner, but doing so would do violence to the overall structure of the program. This is why sharing specifications are so important.
Lets try to re-organize the signature GEOMETRY so that duplication of
the point and vector structures is avoided. One step is to eliminate the substructure Vector from SPHERE, replacing uses of Vector.vector by Vector.Point.vector.
signature SPHERE =
sig
structure Point : POINT
type sphere
val sphere :
Point.point * Point.Vector.vector -> sphere
end
After all, since the structure Point comes equipped with a notion of vector,
why not use it?
This cuts down the number of sharing specifications to one:
signature GEOMETRY =
sig
structure Point : POINT
structure Sphere : SPHERE
sharing Point = Sphere.Point
end
If we could further eliminate the substructure Point from the signature
SPHERE we would have only one copy of Point and no need for a sharing
specification.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

22.1 Combining Abstractions

180

But what would the signature SPHERE look like in this case?
signature SPHERE =
sig
type sphere
val sphere :
Point.point * Point.Vector.vector -> sphere
end
The problem now is that the signature SPHERE is no longer self-contained.
It makes reference to a structure Point, but which Point are we talking
about? Any commitment would tie the signature to a specific structure,
and hence a specific dimension, contrary to our intentions. Rather, the
notion of point must be a generic concept within SPHERE, and hence Point
must appear as a substructure. The substructure Point may be thought of
as a parameter of the signature SPHERE in the sense discussed earlier.
The only other move available to us is to eliminate the structure Point
from the signature GEOMETRY. This is indeed possible, and would eliminate
the need for any sharing specifications. But it only defers the problem,
rather than solving it. A full-scale geometry package would contain more
abstractions that involve points, so that there will still be copies in the
other abstractions. Sharing specifications would then be required to ensure that these copies are, in fact, identical.
Here is an example. Let us introduce another geometric abstraction,
the semi-space.
signature SEMI SPACE =
sig
structure Point : POINT
type semispace
val side : Point.point * semispace -> bool option
end
The function side determines (if possible) whether a given point lies in
one half of the semi-space or the other.
The expanded GEOMETRY signature would look like this (with the elimination of the Point structure in place).
signature EXTD GEOMETRY =
sig
R EVISED 11.02.11

D RAFT

V ERSION 1.2

22.2 Sample Code

181

structure Sphere : SPHERE

structure SemiSpace : SEMI SPACE
sharing Sphere.Point = SemiSpace.Point
end
By an argument similar to the one we gave for the signature SPHERE, we
cannot eliminate the substructure Point from the signature SEMI SPACE,
and hence we wind up with two copies. Therefore the sharing specification is required.
What is at issue here is a fundamental tension in the very notion of
modular programming. On the one hand we wish to separate modules
from one another so that they may be treated independently. This requires that the signatures of these modules be self-contained. Unbound
references to a structure such as Point ties that signature to a specific implementation, in violation of our desire to treat modules separately
from one another. On the other hand we wish to combine modules together
to form programs. Doing so requires that the composition be coherent,
which is achieved by the use of sharing specifications. What sharing specifications do for you is to provide an after the fact means of tying together
several different abstractions to form a coherent whole. This approach to
the problem of coherence is a unique and uniquely effective feature
of the ML module system.

22.2

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 23
Parameterization
To support code re-use it is useful to define generic, or parameterized, modules that leave unspecified some aspects of the implementation of a module. The unspecified parts may be instantiated to determine specific instances of the module. The common part is thereby implemented once and
shared among all instances.
In ML such generic modules are called functors. A functor is a modulelevel function that takes a structure as argument and yields a structure
as result. Instances are created by applying the functor to an argument
specifying the interpretation of the parameters.

23.1

Functor Bindings and Applications

Functors are defined using a functor binding. There are two forms, the
opaque and the transparent. A transparent functor has the form
functor funid(decs):sigexp = strexp
where the result signature, sigexp, is transparently ascribed; an opaque functor has the form
functor funid(decs):>sigexp = strexp
where the result signature is opaquely ascribed. A functor is a modulelevel function whose argument is a sequence of declarations, and whose
result is a structure.

23.1 Functor Bindings and Applications

183

Type checking of a functor consists of checking that the functor body

matches the ascribed result signature, given that the parameters of the
functor have the specified signatures. For a transparent functor the result signature is the augmented signature determined by matching the
principal signature of the body (relative to the assumptions governing the
parameters) against the given result signature. For opaque functors the
ascribed result signature is used as-is, without further augmentation by
type equations.
In chapter 21 we developed a modular implementation of dictionaries
in which the ordering of keys is made explicit as a substructure. The implementation of dictionaries is the same for each choice of order structure
on the keys. This can be expressed by a single functor binding that defines
the implementation of dictionaries parametrically in the choice of keys.
functor DictFun
(structure K : ORDERED) :>
DICT where type Key.t = K.t =
struct
structure Key : ORDERED = K
datatype a dict =
Empty |
Node of a dict * Key.t * a * a dict
val empty = Empty
fun insert (None, k, v) =
Node (Empty, k, v, Empty)
fun lookup (Empty, ) = NONE
| lookup (Node (dl, l, v, dr), k) =
if Key.lt(k, l) then
lookup (dl, k)
else if Key.lt (l, k) then
lookup (dr, k)
else
v
end
The functor DictFun takes as argument a single structure specifiying the
ordered key type as a structure implementing signature ORDERED. The opaquely
ascribed result signature is
R EVISED 11.02.11

D RAFT

V ERSION 1.2

23.1 Functor Bindings and Applications

184

DICT where type Key.t = K.t.

This signature holds the type a dict abstract, but specifies that the key
type is the type K.t passed as argument to the functor. The body of the
functor has been written so that the comparison operations are obtained
from the Key substructure. This ensures that the dictionary code is independent of the choice of key type and ordering.
Instances of functors are obtained by application. A functor application
has the form
funid(binds)
where binds is a sequence of bindings of the arguments of the functor.
The signature of a functor application is determined by the following
procedure. We assume we are given the signatures of the functor parameters, and also the true result signature of the functor (the given signature for opaque functors, the augmented signature for the transparent
functors).
1. For each argument, match the argument signature against the corresponding parameter signature of the functor. This determines an
augmentation of the parameter signature for each argument (as described in chapter 20).
2. For each reference to a type component of a functor parameter in
the result signature, propagate the type definitions of the augmented
parameter signature to the result signature.
The signature of the application determined by this procedure is then
opaquely ascribed to the application. This means that if a type is left abstract
in the result signature of a functor, that type is new in every instance of
that functor. This behavior is called generativity of the functor.1
Returning to the example of the dictionary functor, the three versions of
dictionaries considered in chapter 21 may be obtained by applying DictFun
to appropriate arguments.
structure LtIntDict = DictFun (structure K = LessInt)
structure LexStringDict = DictFun (structure K = LexString)
structure DivIntDict = DictFun (structure K = DivInt)
1 The

alternative, called applicativity, means that there is one abstract type shared by all
instances of that functor.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

23.2 Functors and Sharing Specifications

185

In each case the functor DictFun is instantiated by specifying a binding for

its argument stucture K. The argument structures are, as described in chapter 21, implementations of the signature ORDERED. They specify the type of
keys and the sense in which they are ordered.
The signatures for the structures LtIntDict, LexStringDict, and DivIntDict
are determined by instantiating the result signature of the functor DictFun
according to the above procedure. Consider the application of DictFun to
LtIntDict. The augmented signature resulting from matching the signature of LtIntDict against the parameter signature ORDERED is the signature
ORDERED where type t=int
Assigning this to the parameter K, we deduce that the type K.t is equivalent to int, and hence the result signature of DictFun is
DICT where type Key.t = int
so that IntLtDict.Key.t is equivalent to int, as desired. By a similar process we deduce that the signature of LexStringDict is
DICT where type Key.t = string
and that the signature of DivIntDict is
DICT where type Key.t = int.

23.2

Functors and Sharing Specifications

In chapter 22 we developed a signature of geometric primitives that contained sharing specifications to ensure that the constituent abstractions
may be combined properly. The signature GEOMETRY is defined as follows:
signature GEOMETRY =
sig
structure Point : POINT
structure Sphere : SPHERE
sharing Point = Sphere.Point
and Point.Vector = Sphere.Vector
and Sphere.Vector = Sphere.Point.Vector
end
R EVISED 11.02.11

D RAFT

V ERSION 1.2

23.2 Functors and Sharing Specifications

186

The sharing clauses ensure that the Point and Sphere components are
compatible with each other.
Since we expect to define vectors, points, and spheres of various dimensions, it makes sense to implement these as functors, according to the
following scheme:
functor PointFun
(structure V : VECTOR) : POINT = ...
functor SphereFun
(structure V : VECTOR
structure P : POINT) : SPHERE =
struct
structure Vector = V
structure Point = P
.
.
.
end
functor GeomFun
(structure P : POINT
structure S : SPHERE) : GEOMETRY =
struct
structure Point = P
structure Sphere = S
end
A two-dimensional geometry package may then be defined as follows:
structure Vector2D : VECTOR = ...
structure Point2D : POINT =
PointFun (structure V = Vector2D)
structure Sphere2D : SPHERE =
SphereFun (structure V = Vector2D and P = Point2D)
structure Geom2D : GEOMETRY =
GeomFun (structure P = Point2D and S = Sphere2D)
A three-dimensional version is defined similarly.
There is only one problem: the functors SphereFun and GeomFun are
not well-typed! The reason is that in both cases their result signatures
require type equations that are not true of their parameters! For example,
R EVISED 11.02.11

D RAFT

V ERSION 1.2

23.3 Avoiding Sharing Specifications

187

the signature SPHERE requires that Point.Vector be the same as Vector,

which is not satisfied by the body of SphereFun. For these to be true, the
structures P.Vector and V must be equivalent. This is not true in general,
because the functors might be applied to arguments for which this is false.
Similar problems plague the functor GeomFun. The solution is to include
sharing constraints in the parameter list of the functors, as follows:
functor SphereFun
(structure V : VECTOR
structure P : POINT
sharing P.Vector = V) : SPHERE =
struct
structure Vector = V
structure Point = P
.
.
.
end
functor GeomFun
(structure P : POINT
structure S : SPHERE
sharing P.Vector = S.Vector and P = S.Point) : GEOMETRY =
struct
structure Point = P
structure Sphere = S
end
These equations preclude instantiations for which the required equations
do not hold, and are sufficient to ensure that the requirements of the result
signatures of the functors are met.

23.3

Avoiding Sharing Specifications

As with sharing specifications in signatures, it is natural to wonder whether

they can be avoided in functor parameters. Once again, the answer is
yes, but doing so does violence to the structure of your program. The
chief virtue of sharing specifications is that they express directly and concisely the required relationships without requiring that these relationships
be anticipated when defining the signatures of the parameters. This greatly
R EVISED 11.02.11

D RAFT

V ERSION 1.2

23.3 Avoiding Sharing Specifications

188

facilitates re-use of off-the-shelf code, for which it is impossible to assume

any sharing relationships that one may wish to impose in an application.
To see what happens, lets consider the best-case scenario from chapter 22 in which we have minimized sharing specifications to one tying together the sphere and semispace components. That is, were to implement
the following signature:
signature EXTD GEOMETRY =
sig
structure Sphere : SPHERE
structure SemiSpace : SEMI SPACE
sharing Sphere.Point = SemiSpace.Point
end
The implementation is a functor of the form
functor ExtdGeomFun
(structure Sp : SPHERE
structure Ss : SEMI SPACE
sharing Sphere.Point = SemiSpace.Point) =
struct
structure Sphere = Sp
structure SemiSpace = Ss
end
To eliminate the sharing equation in the functor parameter, we must
arrange that the sharing equation in the signature EXTD GEOMETRY holds.
Simply dropping the sharing specification will not do, because then there
is no reason to believe that it will hold as required in the signature. A
natural move is to factor out the implementation of POINT, and use it to
ensure that the required equation is true of the functor body. There are
two methods for doing this, each with disadvtanges compared to the use
of sharing specifications.
One is to make the desired equation true by construction. Rather than
take implementations of SPHERE and SEMI SPACE as arguments, the functor
ExtdGeomFun takes only an implementation of POINT, then creates in the
functor body appropriate implementations of spheres and semi-spaces.
functor SphereFun
(structure P : POINT) : SPHERE =
R EVISED 11.02.11

D RAFT

V ERSION 1.2

23.3 Avoiding Sharing Specifications

189

struct
structure Vector = P.Vector
structure Point = P
.
.
.
end
functor SemiSpaceFun
(structure P : POINT) : SEMI SPACE =
struct
.
.
.
end
functor ExtdGeomFun1
(structure P : POINT) : GEOMETRY =
struct
structure Sphere =
SphereFun (structure P = Point)
structure SemiSpace =
SemiSpaceFun (structure P = Point)
end
The problems with this solution are these:
The body of ExtdGeomFun1 makes use of the functors SphereFun and
SemiSpaceFun. In effect we are limiting the geometry functor to arguments that are built from these specific functors, and no other. This is
a significant loss of generality that is otherwise present in the functor ExtdGeomFun, which may be applied to any implementations of
SPHERE and SEMI SPACE.
The functor ExtdGeomFun1 must have as parameter the common element(s) of the components of its body, which is then used to build
up the appropriate substructures in a manner consistent with the required sharing. This approach does not scale well when many abstractions are layered atop one another. We must reconstruct the
entire hierarchy, starting with the components that are conceptually
furthest away as arguments.
There is no inherent reason why ExtdGeomFun1 must take an implementation of POINT as argument. It does so only so that it can reconR EVISED 11.02.11

D RAFT

V ERSION 1.2

23.3 Avoiding Sharing Specifications

190

struct the hierarchy so as to satisfy the sharing requirements of the

result signature.
Another approach is to factor out the common component, and use
this to constrain the arguments to the functor to ensure that the possible
arguments are limited to situations for which the required sharing holds.
functor ExtdGeomFun2
(structure P : POINT
structure Sp : SPHERE where Point = P
structure Ss : SEMI SPACE where Point = P) =
struct
structure Sphere = Sp
structure SemiSpace = Ss
end
Now the required sharing requirements are met, but it is also clear that
this approach has no particular advantages over just using a sharing specification. It has the disadvantage of requiring a third argument, whose
only role is to make it possible to express the required sharing. An application of this functor must provide not only implementations of SPHERE
and SEMI SPACE, but also an implementation of POINT that is used to build
these!
A slightly more sophisticated version of this solution is as follows:
functor ExtdGeomFun3
(structure Sp : SPHERE
structure Ss : SEMI SPACE where Point = Sp.Point) =
struct
structure Sphere = Sp
structure SemiSpace = Ss
end
The extra parameter to the functor has been eliminated by choosing one
of the components as a representative and insisting that the others be
compatible with it by using a where clause.2
2 Officially,

we must write where type Point.point = Sp.Point.point, but many

compilers accept the syntax above.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

23.4 Sample Code

191

This solution has all of the advantages of the direct use of sharing specifications, and no further disadvantages. However, we are forced to violate
arbitrarily the inherent symmetry of the situation. We could just as well
have written
functor ExtdGeomFun4
(structure Ss : SEMI SPACE
structure Sp : SPHERE where Point = Sp.Point) =
struct
structure Sphere = Sp
structure SemiSpace = Ss
end
without changing the meaning.
Here is the point: sharing specifications allow a symmetric situation to be
treated in a symmetric manner. The compiler breaks the symmetry by choosing representatives arbitrarily in the manner illustrated above. Sharing
specifications off-load the burden of making such tedious (because arbitrary) decisions to the compiler, rather than imposing it on the programmer.

23.4

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Part IV
Programming Techniques

193
In this part of the book we will explore the use of Standard ML to build
elegant, reliable, and efficient programs. The discussion takes the form of
a series of worked examples illustrating various techniques for building
programs.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 24
Specifications and Correctness
The most important tools for getting programs right are specification and
verification. In this Chapter we review the main ideas in preparation for
their subsequent use in the rest of the book.

24.1

Specifications

A specification is a description of the behavior of a piece of code. Specifications take many forms:
Typing. A type specification describes the form of the value of an
expression, without saying anything about the value itself.
Effect Behavior. An effect specification resembles a type specification,
but instead of describing the value of an expression, it describes the
effects it may engender when evaluated.
Input-Output Behavior. An input-output specification is a mathematical formula, usually an implication, that describes the output of a
function for all inputs satisfying some assumptions.
Time and Space Complexity. A complexity specification states the time
or space required to evaluate an expression. The specification is most
often stated asymptotically in terms of the number of execution steps
or the size of a data structure.

24.1 Specifications

195

Equational Specifications. An equational specification states that one

code fragment is equivalent to another. This means that wherever
one is used we may replace it with the other without changing the
observable behavior of any program in which they occur.
This list by no means exhausts the possibilities. What specifications have
in common, however, is a descriptive, or declarative, flavor, rather than a
prescriptive, or operational, one. A specification states what a piece of code
does, not how it does it.
Good programmers use specifications to state precisely and concisely
their intentions when writing a piece of code. The code is written to solve
a problem; the specification records for future reference what problem the
code was intended to solve. The very act of formulating a precise specification of a piece of code is often the key to finding a good solution to
a tricky problem. The guiding principle is this: if you are unable to write
a clear specification, you do not understand the problem well enough to solve it
correctly.
The greatest difficulty in using specifications is in knowing what to
say. A common misunderstanding is that a given piece of code has one
specification stating all there is to know about it. Rather you should see
a specification as describing one of perhaps many properties of interest.
In ML every program comes with at least one specification, its type. But
we may also consider other specifications, according to interest and need.
Sometimes only relatively weak properties are important the function
square always yields a non-negative result. Other times stronger properties are needed the function fib applied to n yields the nth and n-1st
Fibonacci number. Its a matter of taste and experience to know what to
say and how best to say it.
Recall the following two ML functions from chapter 7:
fun
|
|
fun
|
|

fib 0 = 1
fib 1 = 1
fib n = fib (n-1) + fib (n-2)
fib 0 = (1, 0)
fib 1 = (1, 1)
fib n =
let
val (a, b) = fib (n-1)

R EVISED 11.02.11

D RAFT

V ERSION 1.2

24.2 Correctness Proofs

196

in
(a+b, a)
end
Here are some specifications pertaining to these functions:
Type specifications:
val fib : int -> int
val fib : int -> int * int
Effect specifications:
The application fib n may raise the exception Overflow.
The application fib n may raise the exception Overflow.
Input-output specifications:
If n 0, then fib n evaluates to the nth Fibonacci number.
If n 0, then fib n evaluates the nth and n 1st Fibonacci
number, in that order.
Time complexity specifications:
If n 0, then fib n terminates in O(2n ) steps.
If n 0, then fib n terminates in O(n) steps.
Equivalence specification:
For all n 0, fib n is equivalent to #1(fib n)

24.2

Correctness Proofs

A program satisfies, or meets, a specification iff its execution behavior is

as described by the specification. Verification is the process of checking
that a program satisfies a specification. This takes the form of proving a
mathematical theorem (by hand or with machine assistance) stating that
the program implements a specification.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

24.2 Correctness Proofs

197

There are many misunderstandings in the literature about specification

and verification. It is worthwhile to take time out to address some of them
here.
It is often said that a program is correct if it meets a specification. The
verification of this fact is then called a correctness proof. While there is nothing wrong with this usage, it invites misinterpretation. As we remarked
in section 24.1, there is in general no single preferred specification of a
piece of code. Verification is always relative to a specification, and hence
so is correctness. In particular, a program can be correct with respect to
one specification, and incorrect with respect to another. Consequently,
a program is never inherently correct or incorrect; it is only so relative to a
particular specification.
A related misunderstanding is the inference that a program works
from the fact that it is correct in this sense. Specifications usually make
assumptions about the context in which the program is used. If these assumptions are not satisfied, then all bets are off nothing can be said
about its behavior. For this reason it is entirely possible that a correct program will malfunction when used, not because the specification is not met,
but rather because the specification is not relevant to the context in which
the code is used.
Another common misconception about specifications is that they can
always be implemented as run-time checks.1 There are at least two fallacies here:
1. The specification is stated at the level of the source code. It may
not even be possible to test it at run-time. See chapter 32 for further
discussion of this point.
2. The specification may not be mechanically checkable. For example,
we might specify that a function f of type int->int always yields a
non-negative result. But there is no way to implement this as a runtime check this is an undecidable, or uncomputable, problem.
For this reason specifications are strictly more powerful than run-time
checks. Correspondingly, it takes more work than a mere conditional
1 This

misconception is encouraged by the C assert macro, which introduces an executable test that a certain computable condition holds. This is a fine thing, but from
this many people draw the conclusion that assertions (specifications) are simply boolean
tests. This is false.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

24.2 Correctness Proofs

198

branch to ensure that a program satisfies a specification.

Finally, it is important to note that specification, implementation, and
verification go hand-in-hand. It is unrealistic to propose to verify that an
arbitrary piece of code satisfies an arbitrary specification. Fundamental
computability and complexity results make clear that we can never succeed in such an endeavor. Fortunately, it is also completely artificial. In
practice we specify, code, and verify simultaneously, with each activity informing the other. If the verification breaks down, re-consider the code
or the specification (or both). If the code is difficult to write, look for insights from the specification and verification. If the specification is complex, rough out the code and proof to see what it should be. There is no
magic bullet, but these are some of the most important, and useful, tools
for building elegant, robust programs.
Verifications of specifications take many different forms.
Type specifications are verified automatically by the ML compiler.
If you put type annotations on a function stating the types of the
parameters or results of that function, the correctness of these annotations is ensured by the compiler. For example, we may state the
intended typing of fib as follows:
fun fib (n:int):int =
case n
of 0 => 1
| 1 => 1
| n => fib (n-1) + fib (n-2)
This ensures that the type of fib is int->int.
Effect specifications must be checked by hand. These generally state
that a piece of code may raise one or more exceptions. If an exception is not mentioned in such a specification, we cannot conclude that
the code does not raise it, only that we have no information. Notice
that a handler serves to eliminate a may raise specification. For example, the following function may not raise the exception Overflow,
even though fib might:
fun protected fib n =
(fib n) handle Overflow => 0
R EVISED 11.02.11

D RAFT

V ERSION 1.2

24.3 Enforcement and Compliance

199

A handler makes it possible to specify that an exception may not be

raised by a given expression (because the handler traps it).
Input-output specifications require proof, typically using some form
of induction. For example, in chapter 7 we proved that fib n yields
the nth Fibonacci number by complete induction on n.
Complexity specifications are often verified by solving a recurrence
describing the execution time of a program. In the case of fib we
may read off the following recurrence:
T (0) = 1
T (1) = 1
T ( n + 2) = T ( n ) + T ( n + 1) + 1
Solving this recurrence yields the proof that T (n) = O(2n ).
Equivalence specifications also require proof. Since equivalence of
expressions must account for all possible uses of them, these proofs
are, in general, very tricky. One method that often works is induction plus hand-simulation. For example, it is not hard to prove by
induction on n that fib n is equivalent to #1(fib n). First, plug in
n = 0 and n = 1, and calculate using the definitions. Then assume
the result for n and n + 1, and consider n + 2, once again calculating
based on the definitions, to obtain the result.

24.3

Enforcement and Compliance

Most specifications have the form of an implication: if certain conditions

are met, then the program behaves a certain way. For example, type specifications are conditional on the types of its free variables. For example,
if x has type int, then x+1 has type int. Input-output specifications are
characteristically of this form. For example, if x is non-negative, then so is
x+1.
Just as in ordinary mathematical reasoning, if the premises of such
a specification are not true, then all bets are off nothing can be said
about the behavior of the code. The premises of the specification are preconditions that must be met in order for the program to behave in the manner described by the conclusion, or post-condition, of the specification. This
R EVISED 11.02.11

D RAFT

V ERSION 1.2

24.3 Enforcement and Compliance

200

means that the pre-conditions impose obligations on the caller, the user
of the code, in order for the callee, the code itself, to be well-behaved. A
conditional specification is a contract between the caller and the callee: if
the caller meets the pre-conditions, the caller promises to fulfill the postcondition.
In the case of type specifications the compiler enforces this obligation
by ruling out as ill-typed any attempt to use a piece of code in a context that does not fulfill its typing assumptions. Returning to the example
above, if one attempts to use the expression x+1 in a context where x is not
an integer, one can hardly expect that x+1 will yield an integer. Therefore
it is rejected by the type checker as a violation of the stated assumptions
governing the types of its free variables.
What about specifications that are not mechanically enforced? For example, if x is negative, then we cannot infer anything about x+1 from the
specification given above.2 To make use of the specification in reasoning
about its used in a larger program, it is essential that this pre-condition be
met in the context of its use.
Lacking mechanical enforcement of these obligations, it is all too easy
to neglect them when writing code. Many programming mistakes can be
traced to violation of assumptions made by the callee that are not met by
the caller.3 What can be done about this?
A standard method, called bullet-proofing, is to augment the callee with
run-time checks that ensure that its pre-conditions are met, raising an exception if they are not. For example, we might write a bullet-proofed
version of fib that ensures that its argument is non-negative as follows:
local
exception PreCond
fun unchecked fib
| unchecked fib
| unchecked fib
unchecked fib
in

0 = 1
1 = 1
n =
(n-1) + unchecked fib (n-2)

2 There

are obviously other specifications that carry more information, but were only
concerned here with the one given. Moreover, if f is an unknown function, then we will,
in general, only have the specification, and not the code, to reason about.
3 Sadly, these assumptions are often unstated and can only be culled from the code
with great effort, if at all.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

24.3 Enforcement and Compliance

201

fun checked fib n =

if n < 0 then
raise PreCond
else
unchecked fib n
end
It is worth noting that we have structured this program to take the
pre-condition check out of the loop. It would be poor practice to define
checked fib as follows:
fun bad checked fib n =
if n < 0 then
raise PreCond
else
case n
of 0 => 1
| 1 => 1
| n => bad checked fib (n-1) + bad checked fib (n-2)
Once we know that the initial argument is non-negative, it is assured that
recursive calls also satisfy this requirement, provided that youve done the
inductive reasoning to validate the specification of the function.
However, bullet-proofing in this form has several drawbacks. First, it
imposes the overhead of checking on all callers, even those that have ensured that the desired pre-condition is true. In truth the run-time overhead
is minor; the real overhead is requiring that the implementor of the callee
take the trouble to impose the checks.
Second, and far more importantly, bullet-proofing only applies to specifications that can be checked at run-time. As we remarked earlier, not all
specifications are amenable to run-time checks. For these cases there is
no question of inserting run-time checks to enforce the pre-condition. For
example, we may wish to impose the requirement that a function argument of type int->int always yields a non-negative result. There is no
run-time check for this condition we cannot write a function nonneg of
type (int->int)->bool that determines whether or not a function f always yields a non-negative result. In chapter 32 we will consider the use
of data abstraction to enforce at compile time specifications that may not
be checked at run-time.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 25
Induction and Recursion
This chapter is concerned with the close relationship between recursion
and induction in programming. If a function is recursively-defined, an inductive proof is required to show that it meets a specification of its behavior. The motto is
when programming recursively, think inductively.
Doing so significantly reduces the time spent debugging, and often leads
to more efficient, robust, and elegant programs.

25.1

Exponentiation

Lets start with a very simple series of examples, all involving the computation of the integer exponential function. Our first example is to compute
2n for integers n 0. We seek to define the function exp of type int->int
satisfying the specification
if n 0, then exp n evaluates to 2n .
The precondition, or assumption, is that the argument n is non-negative. The
postcondition, or guarantee, is that the result of applying exp to n is the number 2n . The caller is required to establish the precondition before applying
exp; in exchange, the caller may assume that the result is 2n .
Heres the code:
fun exp 0 = 1
| exp n = 2 * exp (n-1)

25.1 Exponentiation

203

Does this function satisfy the specification? It does, and we can prove
this by induction on n. If n = 0, then exp n evaluates to 1 (as you can
see from the first line of its definition), which is, of course, 20 . Otherwise,
assume that exp is correct for n 1 0, and consider the value of exp
n. From the second line of its definition we can see that this is the value
of 2 p, where p is the value of exp (n 1). Inductively, p 2n1 , so
2 p = 2 2n1 = 2n , as desired. Notice that we need not consider
arguments n < 0 since the precondition of the specification requires that
this be so. We must, however, ensure that each recursive call satisfies this
requirement in order to apply the inductive hypothesis.
That was pretty simple. Now let us consider the running time of exp
expressed as a function of n. Assuming that arithmetic operations are executed in constant time, then we can read off a recurrence describing its
execution time as follows:
T (0) = O (1)
T ( n + 1) = O (1) + T ( n )
We are interested in solving a recurrence by finding a closed-form expression for it. In this case the solution is easily obtained:
T (n) = O(n)
Thus we have a linear time algorithm for computing the integer exponential function.
What about space? This is a much more subtle issue than time because it is much more difficult in a high-level language such as ML to see
where the space is used. Based on our earlier discussions of recursion and
iteration we can argue informally that the definition of exp given above
requires space given by the following recurrence:
S (0) = O (1)
S ( n + 1) = O (1) + S ( n )
The justification is that the implementation requires a constant amount of
storage to record the pending multiplication that must be performed upon
completion of the recursive call.
Solving this simple recurrence yields the equation
S(n) = O(n)
R EVISED 11.02.11

D RAFT

V ERSION 1.2

25.1 Exponentiation

204

expressing that exp is also a linear space algorithm for the integer exponential function.
Can we do better? Yes, on both counts! Heres how. Rather than count
down by ones, multiplying by two at each stage, we use successive squaring to achieve logarithmic time and space requirements. The idea is that
if the exponent is even, we square the result of raising 2 to half the given
power; otherwise, we reduce the exponent by one and double the result,
ensuring that the next exponent will be even. Heres the code:
fun
fun
fun
|

square (n:int) = n*n

double (n:int) = n+n
fast exp 0 = 1
fast exp n =
if n mod 2 = 0 then
square (fast exp (n div 2))
else
double (fast exp (n-1))

Its specification is precisely the same as before. Does this code satisfy
the specification? Yes, and we can prove this by using complete induction,
a form of mathematical induction in which we may prove that n > 0 has
a desired property by assuming not only that the predecessor has it, but
that all preceding numbers have it, and arguing that therefore n must have
it. Heres how its done. For n = 0 the argument is exactly as before.
Suppose, then, that n > 0. If nis even, the value of exp n is the result
of squaring the value of exp (n 2). Inductively this value is 2(n2) , so
squaring it yields 2(ndiv2) 2(n2) = 22(n2) = 2n , as required. If, on
the other hand, n is odd, the value is the result of doubling exp (n 1).
Inductively the latter value is 2(n1) , so doubling it yields 2n , as required.
Heres a recurrence governing the running time of fast exp as a function of its argument:
T (0) = O (1)
T (2n) = O(1) + T (n)
T (2n + 1) = O(1) + T (2n)
= O (1) + T ( n )
Solving this recurrence using standard techniques yields the solution
T (n) = O(lgn)
R EVISED 11.02.11

D RAFT

V ERSION 1.2

25.1 Exponentiation

205

You should convince yourself that fast exp also requires logarithmic space
usage.
Can we do better? Well, its not possible to improve the time requirement (at least not asymptotically), but we can reduce the space required to
O(1) by putting the function into iterative (tail recursive) form. However,
this may not be achieved in this case by simply adding an accumulator argument, without also increasing the running time! The obvious approach
is to attempt to satisfy the specification
if n 0, then skinny fast exp (n, a) evaluates to 2n a.
Heres some code that achieves this specification:
fun skinny fast exp (0, a) = a
| skinny fast exp (n, a) =
if n mod 2 = 0 then
skinny fast exp (n div 2,
skinny fast exp (n div 2, a))
else
skinny fast exp (n-1, 2*a)
It is easy to see that this code works properly for n = 0 and for n > 0
when n is odd, but what if n > 0 is even? Then by induction we compute
2(n2) 2(n2 ) a by two recursive calls to skinny fast exp.
This yields the desired result, but what is the running time? Heres a
recurrence to describe its running time as a function of n:
T (0) = 1
T (2n) = O(1) + 2T (n)
T (2n + 1) = O(1) + T (2n)
= O(1) + 2T (n)
Here again we have a standard recurrence whose solution is
T ( n ) = O ( n ).
Can we do better? The key is to recall the following important fact:
2(2n) = (22 )n = 4n .
We can achieve a logarithmic time and exponential space bound by a
change of base. Heres the specification:
R EVISED 11.02.11

D RAFT

V ERSION 1.2

25.1 Exponentiation

206

if n 0, then gen skinny fast exp (b, n, a) evaluates to bn

a.
Heres the code:
fun gen skinny fast exp (b, 0, a) = a
| gen skinny fast exp (b, n, a) =
if n mod 2 = 0 then
gen skinny fast exp (b*b, n div 2, a)
else
gen skinny fast exp (b, n - 1, b * a)
Lets check its correctness by complete induction on n. The base case
is obvious. Assume the specification for arguments smaller than n > 0.
If n is even, then by induction the result is (b b)(n2) a = bn a, and
if n is odd, we obtain inductively the result b(n1) b a = bn a. This
completes the proof.
The trick to achieving an efficient implementation of the exponential
function was to compute a more general function that can be implemented
using less time and space. Since this is a trick employed by the implementor of the exponential function, it is important to insulate the client from
it. This is easily achieved by using a local declaration to hide the generalized function, making available to the caller a function satisfying the
original specification. Heres what the code looks like in this case:
local
fun gen skinny fast exp (b, 0, a) =
| gen skinny fast exp (b, n, a) = ...
in
fun exp n = gen skinny fast exp (2, n, 1)
end
(The ellided code is the same as above.) The point here is to see how
induction and recursion go hand-in-hand, and how we used induction
not only to verify programs after-the-fact, but, more importantly, to help
discover the program in the first place. If the verification is performed
simultaneously with the coding, it is far more likely that the proof will go
through and the program will work the first time you run it.
It is important to notice the correspondence between strengthening
the specification by adding additional assumptions (and guarantees) and
R EVISED 11.02.11

D RAFT

V ERSION 1.2

25.2 The GCD Algorithm

207

adding accumulator arguments. What we observe is the apparent paradox that it is often easier to do something (superficially) harder! In terms
of proving, it is often easier to push through an inductive argument for a
stronger specification, precisely because we get to assume the result as the
inductive hypothesis when arguing the inductive step(s). We are limited
only by the requirement that the specification be proved outright at the
base case(s); no inductive assumption is available to help us along here. In
terms of programming, it is often easier to compute a more complicated
function involving accumulator arguments, precisely because we get to
exploit the accumulator when making recursive calls. We are limited only
by the requirement that the result be defined outright for the base case(s);
no recursive calls are available to help us along here.

25.2

The GCD Algorithm

Lets consider a more complicated example, the computation of the greatest common divisor of a pair of non-negative integers. Recall that m is
a divisor of n, written m|n, iff n is a multiple of m, which is to say that
there is some k 0 such that n = k m. The greatest common divisor of
non-negative integers m and n is the largest p such that p|m and p|n. (By
convention the g.c.d. of 0 and 0 is taken to be 0.) Heres the specification
of the gcdfunction:
if m, n 0, then gcd(m,n) evaluates to the g.c.d. of m and n.
Euclids algorithm for computing the g.c.d. of m andn is defined by
complete induction on the product mn. Heres the algorithm, written in
ML:
fun gcd (m:int, 0):int = m
| gcd (0, n:int):int = n
| gcd (m:int, n:int):int =
if m>n then
gcd (m mod n, n)
else
gcd (m, n mod m)
Why is this algorithm correct? We may prove that gcd satisfies the
specification by complete induction on the product m n. If m n is zero,
R EVISED 11.02.11

D RAFT

V ERSION 1.2

25.2 The GCD Algorithm

208

then either mor n is zero, in which case the answer is, correctly, the other
number. Otherwise the product is positive, and we proceed according to
whether m > n or m n. Suppose that m > n. Observe that mmodn =
m (m n) n, so that (mmodn) n = m n (m n)n2 < m n,
so that by induction we return the g.c.d. of mmodn and n. It remains to
show that this is the g.c.d. of m and n. If d divides both mmodn and n,
then k d = (mmodn) = (m (m n) n)and l d = n for some nonnegative k and l. Consequently, k d = m (m n) l d, so m =
(k + (m n) l ) d, which is to say that d divides m. Now if d is any
other divisor of m and n, then it is also a divisor of (mmodn) and n, so
d > d0 . That is, d is the g.c.d. of m and n. The other case, m n, follows
similarly. This completes the proof.
At this point you may well be thinking that all this inductive reasoning
is surely helpful, but its no replacement for good old-fashioned bulletproofing conditional tests inserted at critical junctures to ensure that
key invariants do indeed hold at execution time. Sure, you may be thinking, these checks have a run-time cost, but they can be turned off once
the code is in production, and anyway the cost is minimal compared to,
say, the time required to read and write from disk. Its hard to complain
about this attitude, provided that sufficiently cheap checks can be put into
place and provided that you know where to put them to maximize their
effectiveness. For example, theres no use checking i > 0 at the start of the
then clause of a test for i > 0. Barring compiler bugs, it cant possibly be
anything other than the case at that point in the program. Or it may be
possible to insert a check whose computation is more expensive (or more
complicated) than the one were trying to perform, in which case were
defeating the purpose by including them!
This raises the question of where should we put such checks, and what
checks should be included to help ensure the correct operation (or, at least,
graceful malfunction) of our programs? This is an instance of the general
problem of writing self-checking programs. Well illustrate the idea by elaborating on the g.c.d. example a bit further. Suppose we wish to write a
self-checking g.c.d. algorithm that computes the g.c.d., and then checks
the result to ensure that it really is the greatest common divisor of the two
given non-negative integers before returning it as result. The code might
look something like this:
exception GCD ERROR
R EVISED 11.02.11

D RAFT

V ERSION 1.2

25.2 The GCD Algorithm

209

fun checked gcd (m, n) =

let
val d = gcd (m, n)
in
if m mod d = 0 andalso
n mod d = 0 andalso ???
then
d
else
raise GCD ERROR
end
Its obviously no problem to check that putative g.c.d., d, is in fact a common divisor of mand n, but how do we check that its the greatest common
divisor? Well, one choice is to simply try all numbers between d and the
smaller of m and n to ensure that no intervening number is also a divisor, refuting the maximality of d. But this is clearly so inefficient as to be
impractical. But theres a better way, which, it has to be emphasized, relies on the kind of mathematical reasoning weve been considering right
along. Heres an important fact:
d is the g.c.d. of m and n iff d divides both m and n and can be written
as a linear combination of m and n.
That is, d is the g.c.d. of mand n iff m = k d for some k 0, n = l d for
some l 0, and d = a m + b n for some integers (possibly negative!)
a and b. Well prove this constructively by giving a program to compute
not only the g.c.d. d of m and n, but also the coefficients a and b such that
d = a m + b n. Heres the specification:
if m, n 0, then ggcd (m, n) evaluates to (d, a, b) such that
d divides m, d divides n, and d = a m + b n; consequently, d is
the g.c.d. of m and n.
And heres the code to compute it:
fun ggcd (0, n) = (n, 0, 1)
| ggcd (m, 0) = (m, 1, 0)
| ggcd (m, n) =
if m>n then
R EVISED 11.02.11

D RAFT

V ERSION 1.2

25.2 The GCD Algorithm

210

let
val (d, a, b) = ggcd (m mod n, n)
in
(d, a, b - a * (m div n))
end
else
let
val (d, a, b) = ggcd (m, n mod m)
in
(d, a - b*(n div m), b)
end
We may easily check that this code satisfies the specification by induction
on the product m n. If m n = 0, then either m or n is 0, in which case
the result follows immediately. Otherwise assume the result for smaller
products, and show it for m n > 0. Suppose m > n; the other case
is handled analogously. Inductively we obtain d, a, and b such that d is
the g.c.d. of mmodn and n, and hence is the g.c.d. of m and n, and d =
a (mmodn) + b n. Since mmodn = m (m n) n, it follows that
d = a m + (b a (m n)) n, from which the result follows.
Now we can write a self-checking g.c.d. as follows:
exception GCD ERROR
fun checked gcd (m, n) =
let
val (d, a, b) = ggcd (m, n)
in
if m mod d = 0 andalso
n mod d = 0 andalso d = a*m+b*n
then
d
else
raise GCD ERROR
end
This algorithm takes no more time (asymptotically) than the original, and,
moreover, ensures that the result is correct. This illustrates the power of
the interplay between mathematical reasoning methods such as induction
and number theory and programming methods such as bulletproofing to
achieve robust, reliable, and, what is more important, elegant programs.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

25.3 Sample Code

25.3

211

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 26
Structural Induction
The importance of induction and recursion are not limited to functions
defined over the integers. Rather, the familiar concept of mathematical induction over the natural numbers is an instance of the more general notion
of structural induction over values of an inductively-defined type. Rather than
develop a general treatment of inductively-defined types, we will rely on a
few examples to illustrate the point. Lets begin by considering the natural
numbers as an inductively defined type.

26.1

Natural Numbers

The set of natural numbers, N, may be thought of as the smallest set containing 0 and closed under the formation of successors. In other words, n
is an element of N iff either n = 0 or n = m + 1 for some m in N. Still
another way of saying it is to define N by the following clauses:
1. 0 is an element of N.
2. If m is an element of N, then so is m + 1.
3. Nothing else is an element of N.
(The third clause is sometimes called the extremal clause; it ensures that
we are talking about N and not just some superset of it.) All of these
definitions are equivalent ways of saying the same thing.

26.1 Natural Numbers

213

Since N is inductively defined, we may prove properties of the natural

numbers by structural induction, which in this case is just ordinary mathematical induction. Specifically, to prove that a property P(n) holds of
every n in N, it suffices to demonstrate the following facts:
1. Show that P(0) holds.
2. Assuming that P(m) holds, show that P(m + 1) holds.
The pattern of reasoning follows exactly the structure of the inductive definition the base case is handled outright, and the inductive step is handled by assuming the property for the predecessor and show that it holds
for the numbers.
The principal of structural induction also licenses the definition of functions by structural recursion. To define a function f with domain N, it suffices to proceed as follows:
1. Give the value of f (0).
2. Give the value of f (m + 1) in terms of the value of f (m).
Given this information, there is a unique function f with domain N
satisfying these requirements. Specifically, we may show by induction on
n 0 that the value of f is uniquely determined on all values m n. If
n = 0, this is obvious since f (0) is defined by the first clause. If n = m + 1,
then by induction the value of f is determined for all values k m. But the
value of f at n is determined as a function of f (m), and hence is uniquely
determined. Thus f is uniquely determined for all values of n in N, as was
to be shown.
The natural numbers, viewed as an inductively-defined type, may be
represented in ML using a datatype declaration, as follows:
datatype nat = Zero | Succ of nat
The constructors correspond one-for-one with the clauses of the inductive definition. The extremal clause is implicit in the datatype declaration
since the given constructors are assumed to be all the ways of building
values of type nat. This assumption forms the basis for exhaustiveness
checking for clausal function definitions.
(You may object that this definition of the type nat amounts to a unary
(base 1) representation of natural numbers, an unnatural and space-wasting
R EVISED 11.02.11

D RAFT

V ERSION 1.2

26.2 Lists

214

representation. This is indeed true; in practice the natural numbers are

represented as non-negative machine integers to avoid excessive overhead. Note, however, that this representation places a fixed upper bound
on the size of numbers, whereas the unary representation does not. Hybrid representations that enjoy the benefits of both are, of course, possible
and occasionally used when enormous numbers are required.)
Functions defined by structural recursion are naturally represented by
clausal function definitions, as in the following example:
fun
|
fun
|

double Zero = Zero

double (Succ n) = Succ (Succ (double n))
exp Zero = Succ(Zero)
exp (Succ n) = double (exp n)

The type checker ensures that we have covered all cases, but it does not ensure that the pattern of structural recursion is strictly followed we may
accidentally define f (m + 1) in terms of itself or some f (k ) where k > m,
breaking the pattern. The reason this is admitted is that the ML compiler
cannot always follow our reasoning: we may have a clever algorithm in
mind that isnt easily expressed by a simple structural induction. To avoid
restricting the programmer, the language assumes the best and allows any
form of definition.
Using the principle of structure induction for the natural numbers, we
may prove properties of functions defined over the naturals. For example,
we may easily prove by structural induction over the type nat that for
every n N, exp n evaluates to a positive number. (In previous chapters
we carried out proofs of more interesting program properties.)

26.2

Lists

Generalizing a bit, we may think of the type a list as inductively defined by the following clauses:
1. nil is a value of type a list
2. If h is a value of type a, and t is a value of type a list, then h::t
is a value of type a list.
3. Nothing else is a value of type a list.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

26.3 Trees

215

This definition licenses the following principle of structural induction

over lists. To prove that a property P holds of all lists l, it suffices to proceed as follows:
1. Show that P holds for nil.
2. Show that P holds for h::t, assuming that P holds for t.
Similarly, we may define functions by structural recursion over lists as
follows:
1. Define the function for nil.
2. Define the function for h::t in terms of its value for t.
The clauses of the inductive definition of lists correspond to the following (built-in) datatype declaration in ML:
datatype a list = nil | :: of a * a list
(We are neglecting the fact that :: is regarded as an infix operator.)
The principle of structural recursion may be applied to define the reverse function as follows:
fun reverse nil = nil
| reverse (h::t) = reverse t @ [h]
There is one clause for each constructor, and the value of reverse for h::t is
defined in terms of its value for t. (We have ignored questions of time and
space efficiency to avoid obscuring the induction principle underlying the
definition of reverse.)
Using the principle of structural induction over lists, we may prove
that reverse l evaluates to the reversal of l. First, we show that reverse
nil yields nil, as indeed it does and ought to. Second, we assume that
reverse t yields the reversal of t, and argue that reverse (h::t) yields
the reversal of h::t, as indeed it does since it returns reverse (t @ [h]).

26.3

Trees

Generalizing even further, we can introduce new inductively-defined types

such as 2-3 trees in which interior nodes are either binary (have two children) or ternary (have three children). Heres a definition of 2-3 trees in
ML:
R EVISED 11.02.11

D RAFT

V ERSION 1.2

26.4 Generalizations and Limitations

216

datatype a twth tree =

Empty |
Bin of a * a twth tree * a twth tree |
Ter of a * a twth tree * a twth tree * a twth tree
How might one define the size of a value of this type? Your first
thought should be to write down a template like the following:
fun size Empty = ???
| size (Bin ( , t1, t2)) = ???
| size (Ter ( , t1, t2, t3)) = ???
We have one clause per constructor, and will fill in the ellided expressions
to complete the definition. In many cases (including this one) the function
is defined by structural recursion. Heres the complete definition:
fun size Empty = 0
| size (Bin ( , t1, t2)) =
1 + size t1 + size t2
| size (Ter ( , t1, t2, t3)) =
1 + size t1 + size t2 + size t3
Obviously this function computes the number of nodes in the tree, as you
can readily verify by structural induction over the type a twth tree.

26.4

Generalizations and Limitations

Does this pattern apply to every datatype declaration? Yes and no. No
matter what the form of the declaration it always makes sense to define a
function over it by a clausal function definition with one clause per constructor. Such a definition is guaranteed to be exhaustive (cover all cases),
and serves as a valuable guide to structuring your code. (It is especially
valuable if you change the datatype declaration, because then the compiler
will inform you of what clauses need to be added or removed from functions defined over that type in order to restore it to a sensible definition.)
The slogan is:
To define functions over a datatype, use a clausal definition
with one clause per constructor
R EVISED 11.02.11

D RAFT

V ERSION 1.2

26.5 Abstracting Induction

217

The catch is that not every datatype declaration supports a principle

of structural induction because it is not always clear what constitutes the
predecessor(s) of a constructed value. For example, the declaration
datatype D = Int of int | Fun of D->D
is problematic because a value of the form Fun( f ) is not constructed directly from another value of type D, and hence it is not clear what to regard
as its predecessor. In practice this sort of definition comes up only rarely;
in most cases datatypes are naturally viewed as inductively defined.

26.5

Abstracting Induction

It is interesting to observe that the pattern of structural recursion may be

directly codified in ML as a higher-order function. Specifically, we may
associate with each inductively-defined type a higher-order function that
takes as arguments values that determine the base case(s) and step case(s)
of the definition, and defines a function by structural induction based on
these arguments. An example will illustrate the point. The pattern of
structural induction over the type nat may be codified by the following
function:
fun nat rec base step =
let
fun loop Zero = base
| loop (Succ n) = step (loop n)
in
loop
end
This function has the type a -> (a -> a) -> nat -> a.
Given the first two arguments, nat rec yields a function of type nat
-> a whose behavior is determined at the base case by the first argument
and at the inductive step by the second. Heres an example of the use of
nat rec to define the exponential function:
val double =
nat rec Zero (fn result => Succ (Succ result))
val exp =
nat rec (Succ Zero) double
R EVISED 11.02.11

D RAFT

V ERSION 1.2

26.5 Abstracting Induction

218

Note well the pattern! The arguments to nat rec are

1. The value for Zero.
2. The value for Succ n defined in terms of its value for n.
Similarly, the pattern of list recursion may be captured by the following
functional:
fun list recursion base step =
let
fun loop nil = base
| loop (h::t) = step (h, loop t)
in
loop
end
The type of the function list recursion is
a -> (b * a -> a) -> b list -> a
It may be instantiated to define the reverse function as follows:
val reverse = list recursion nil (fn (h, t) => t @ [h])
Finally, the principle of structural recursion for values of type a twth tree
is given as follows:
fun twth rec base bin step ter step =
let
fun loop Empty = base
| loop (Bin (v, t1, t2)) =
bin step (v, loop t1, loop t2)
| loop (Ter (v, t1, t2, t3)) =
ter step (v, loop t1, loop t2, loop t3)
in
loop
end
Notice that we have two inductive steps, one for each form of node.
The type of twth rec is
R EVISED 11.02.11

D RAFT

V ERSION 1.2

26.6 Sample Code

219

a -> (b * a * a -> a) -> (b * a * a * a -> a) -> b twth tree -> a

We may instantiate it to define the function size as follows:
val size =
twth rec 0
(fn ( , s1, s2)) => 1+s1+s2)
(fn ( , s1, s2, s3)) => 1+s1+s2+s3)
Summarizing, the principle of structural induction over a recursive
datatype is naturally codified in ML using pattern matching and higherorder functions. Whenever youre programming with a datatype, you
should use the techniques outlined in this chapter to structure your code.

26.6

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 27
Proof-Directed Debugging
In this chapter well put specification and verification techniques to work
in devising a regular expression matcher. The code is similar to that sketched
in chapter 1, but we will use verification techiques to detect and correct a
subtle error that may not be immediately apparent from inspecting or even
testing the code. We call this process proof-directed debugging.
The first task is to devise a precise specification of the regular expression matcher. This is a difficult problem in itself. We then attempt to verify
that the matching program developed in chapter 1 satisfies this specification. The proof attempt breaks down. Careful examination of the failure
reveals a counterexample to the specification the program does not satisfy it. We then consider how best to resolve the problem, not by change of
implementation, but instead by change of specification.

27.1

Regular Expressions and Languages

Before we begin work on the matcher, let us first define the set of regular expressions and their meaning as a set of strings. The set of regular
expressions is given by the following grammar:
r ::= 0 | 1 | a | r1 r2 | r1 + r2 | r
Here a ranges over a given alphabet, a set of primitive letters that may
be used in a regular expression. A string is a finite sequence of letters of
the alphabet. We write for the null string, the empty sequence of letters.
We write s1 s2 for the concatenation of the strings s1 and s2 , the string s

27.1 Regular Expressions and Languages

221

consisting of the letters in s1 followed by those in s2 . The length of a string

is the number of letters in it. We do no distinguish between a character and
the unit-length string consisting solely of that character. Thus we write a s
for the extension of s with the letter a at the front.
A language is a set of strings. Every regular expression r stands for a
particular language L(r ), the language of r, which is defined by induction
on the structure of r as follows:
L(0)
L(1)
L (a)
L (r1 r2 )
L (r1 + r2 )
L (r )

=
=
=
=
=
=

0
1
{a}
L (r1 ) L (r2 )
L (r1 ) + L (r2 )
L (r )

This definition employs the following operations on languages:

0
1
L1 + L2
L1 L2
L (0)
L ( i +1)
L

=
=
=
=
=
=
=

{}
L1 L2
{ s1 s2 | s1 L1 , s2 L2 }
1
L L (i )
S
(i )
i 0 L

An important fact about L is that it is the smallest language L0 such

that 1 + L L0 L0 . Spelled out, this means two things:
1. 1 + L L L , which is to say that
(a) L , and
(b) if s L and s0 L , then s s0 L .
2. If 1 + L L0 L0 , then L L0 .
This means that L is the smallest language (with respect to language containment) that contains the null string and is closed under concatenation
on the left by L.
Lets prove that this is the case. First, since L(0) = 1, it follows immediately that L . Second, if l L and l 0 L , then l 0 L(i) for some i 0,
R EVISED 11.02.11

D RAFT

V ERSION 1.2

27.2 Specifying the Matcher

222

and hence l l 0 L(i+1) by definition of concatentation of languages. This

completes the first step. Now suppose that L0 is such that 1 + L L0 L0 .
We are to show that L L0 . We show by induction on i 0 that L(i) L0 ,
from which the result follows immediately. If i = 0, then it suffices to show
that L0 . But this follows from the assumption that 1 + L L0 L0 , which
implies that 1 L0 . To show that L(i+1) L0 , we observe that, by definition, L(i+1) = L L(i) . By induction L(i) L0 , and hence L L(i) L0 , since
L L0 L0 by assumption.
Having proved that L is the smallest language L0 such that 1 + L L0
0
L , it is not hard to prove that L satisfies the recurrence L = 1 + L L .
We just proved the right to left containment. For the converse, it suffices
to observe that 1 + L (1 + L L ) 1 + L L , for then the result follows by
minimality the result. This is easily established by a simple case analysis.
Exercise 1
Give a full proof of the fact that L = 1 + L L .
Finally, a word about implementation. We will assume in what follows
that the alphabet is given by the type char of characters, that strings are
elements of the type string, and that regular expressions are defined by
the following datatype declaration:
datatype regexp =
Zero | One | Char of char |
Plus of regexp * regexp |
Times of regexp * regexp |
Star of regexp
We will also work with lists of characters, values of type char list, using
ML notation for primitive operations on lists such as concatenation and
extension. Occasionally we will abuse notation and not distinguish (in the
informal discussion) between a string and a list of characters. In particular
we will speak of a character list as being a member of a language, when in
fact we mean that the corresponding string is a member of that language.

27.2

Specifying the Matcher

Let us begin by devising a specification for the regular expression matcher.

As a first cut we write down a type specification. We seek to define a funcR EVISED 11.02.11

D RAFT

V ERSION 1.2

27.2 Specifying the Matcher

223

tion match of type regexp -> string -> bool that determines whether or
not a given string matches a given regular expression. More precisely, we
wish to satisfy the following specification:
For every regular expression r and every string s, match r s
terminates, and evaluates to true iff s L(r ).
We saw in chapter 1 that a natural way to define the procedure match
is to use a technique called continuation passing. We defined an auxiliary
function match is with the type
regexp -> char list -> (char list -> bool) -> bool
that takes a regular expression, a list of characters (essentially a string, but
in a form suitable for incremental processing), and a continuation, and
yields a boolean. The idea is that match is takes a regular expression r, a
character list cs, and a continuation k, and determines whether or not some
initial segment of cs matches r, passing the remaining characters cs0 to k in
the case that there is such an initial segment, and yields false otherwise.
Put more precisely,
For every regular expression r, character list cs, and continuation k, if cs = cs0 @cs00 with cs0 L(r ) and k cs00 evaluates to true, then match is r cs k evaluates true; otherwise,
match is r cs k evaluates to false.
Unfortunately, this specification is too strong to ever be satisfied by any
program! Can you see why? The difficulty is that if k is not guaranteed to
terminate for all inputs, then there is no way that match is can behave as
required. For example, if there is no input on which k terminates, the specification requires that match is return false. It should be intuitively clear
that we can never implement such a function. Instead, we must restrict
attention to total continuations, those that always terminate with true or
false on any input. This leads to the following revised specification:
For every regular expression r, character list cs, and total continuation k, if cs = cs0 cs00 with cs0 L(r ) and k cs00 evaluates to true, then match is r cs k evaluates to true; otherwise, match is r cs k evaluates to false.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

27.2 Specifying the Matcher

224

Observe that this specification makes use of an implicit existential quantification. Written out in full, we might say For all . . . , if there exists cs0
and cs00 such that cs = cs0 cs00 with . . . , then . . . . This observation makes
clear that we must search for a suitable splitting of cs into two parts such
that the first part is in L(r ) and the second is accepted by k. There may,
in general, be many ways to partition the input to as to satisfy both of
these requirements; we need only find one such way. Note, however, that
if cs = cs0 @cs00 with cs0 L(r ) but k cs00 yielding false, we must reject
this partitioning and search for another. In other words we cannot simply
accept any partitioning whose initial segment matches r, but rather only
those that also induce k to accept its corresponding final segment. We may
return false only if there is no such splitting, not merely if a particular
splitting fails to work.
Suppose for the moment that match is satisfies this specification. Does
it follow that match satisfies the original specification? Recall that the function match is defined as follows:
fun match r s =
match is r
(String.explode s)
(fn nil => true | false)
Notice that the initial continuation is indeed total, and that it yields true
(accepts) iff it is applied to the null string. Therefore match satisfies the following property obtained from the specification of mathc is by plugging
in the initial continuation:
For every regular expression r and string s, if s L(r ), then
match r s evaluates to true, and otherwise match r s evaluates to false.
This is precisely the property that we desire for match. Thus match is correct (satisfies its specification) if match is is correct.
So far so good. But does match is satisfy its specification? If so, we are
done. How might we check this? Recall the definition of match is given
in the overview:
fun match is Zero k = false
| match is One cs k = k cs
| match is (Char c) nil k = false
R EVISED 11.02.11

D RAFT

V ERSION 1.2

27.2 Specifying the Matcher

225

| match is (Char c) (d::cs) k =

if c=d then k cs else false
| match is (Times (r1, r2)) cs k =
match is r1 cs (fn cs => match is r2 cs k)
| match is (Plus (r1, r2)) cs k =
match is r1 cs k orelse match is r2 cs k
| match is (Star r) cs k =
k cs orelse
match is r cs (fn cs => match is (Star r) cs k)
Since match is is defined by a recursive analysis of the regular expression
r, it is natural to proceed by induction on the structure of r. That is, we treat
the specification as a conjecture about match is, and attempt to prove it by
structural induction on r.
We first consider the three base cases. Suppose that r is 0. Then no
string is in L(r ), so match is must return false, which indeed it does.
Suppose that r is 1. Since the null string is an initial segment of every
string, and the null string is in L(1), we must yield true iff k cs yields true,
and false otherwise. This is precisely how match is is defined. Suppose
that r is a. Then to succeed cs must have the form a cs0 with k cs0 evaluating
to true; otherwise we must fail. The code for match is checks that cs has
the required form and, if so, passes cs0 to k to determine the outcome, and
otherwise yields false. Thus match is behaves correctly for each of the
three base cases.
We now consider the three inductive steps. For r = r1 + r2 , we observe
that some initial segment of cs matches r and causes k to accept the corresponding final segment of cs iff either some initial segment matches r1 and
drives k to accept the rest or some initial segment matches r2 and drives k
to accept the rest. By induction match is works as specified for r1 and r2 ,
which is sufficient to justify the correctness of match is for r = r1 + r2 .
For r = r1 r2 , the proof is slightly more complicated. By induction
match is behaves according to the specification if it is applied to either
r1 or r2 , provided that the continuation argument is total. Note that the
continuation k0 given by fn cs => match is r2 cs k is total, since by
induction the inner recursive call to match is always terminates. Suppose
that there exists a partitioning cs = cs0 @cs00 with cs0 L(r )and k cs00 evaluating to true. Then cs0 = cs10 cs20 with cs10 L(r1 ) and cs20 L(r2 ), by
definition of L(r1 r2 ). Consequently, match is r2 (cs20 cs00 ) k evaluates to
R EVISED 11.02.11

D RAFT

V ERSION 1.2

27.2 Specifying the Matcher

226

true, and hence match is r1 cs10 cs20 cs00 k0 evaluates to true, as required.
If, however, no such partitioning exists, then one of three situations occurs:
1. either no initial segment of cs matches r1 , in which case the outer
recursive call yields false, as required, or
2. for every initial segment matching r1 , no initial segment of the corresponding final segment matches r2 , in which case the inner recursive
call yields false on every call, and hence the outer call yields false,
as required, or
3. every pair of successive initial segments of cs matching r1 and r2
successively results in k evaluating to false, in which case the inner
recursive call always yields false, and hence the continuation k0 always yields false, and hence the outer recursive call yields false,
as required.
Be sure you understand the reasoning involved here, it is quite tricky to
get right!
We seem to be on track, with one more case to consider, r = r1 . This
case would appear to be a combination of the preceding two cases for alternation and concatenation, with a similar argument sufficing to establish correctness. But there is a snag: the second recursive call to match is
leaves the regular expression unchanged! Consequently we cannot apply
the inductive hypothesis to establish that it behaves correctly in this case,
and the obvious proof attempt breaks down.
What to do? A moments thought suggests that we proceed by an inner induction on the length of the string, based on the idea that if some
initial segment of cs matches L(r ), then either that initial segment is the
null string (base case), or cs = cs0 @cs00 with cs0 L(r1 ) and cs00 L(r )
(induction step). We then handle the base case directly, and handle the
inductive case by assuming that match is behaves correctly for cs00 and
showing that it behaves correctly for cs. But there is a flaw in this argument the string cs00 need not be shorter than cs in the case that cs0 is the
null string! In that case the inductive hypothesis does not apply, and we
are once again unable to complete the proof.
This time we can use the failure of the proof to obtain a counterexample to the specification! For if r = 1 , for example, then match is r cs k
does not terminate! In general if r = r1 with L(r1 ), then match is r
R EVISED 11.02.11

D RAFT

V ERSION 1.2

27.2 Specifying the Matcher

227

cs k fails to terminate. In other words, match is does not satisfy the specification we have given for it. Our conjecture is false!
Our failure to establish that match is satisfies its specification lead to a
counterexample that refuted our conjecture and uncovered a genuine bug
in the program the matcher may not terminate for some inputs. What
to do? One approach is to explicitly check for looping behavior during
matching by ensuring that each recursive calls matches some non-empty
initial segment of the string. This will work, but at the expense of cluttering the code and imposing additional run-time overhead. You should
write out a version of the matcher that works this way, and check that it
indeed satisfies the specification weve given above.
An alternative is to observe that the proof goes through under the additional assumption that no iterated regular expression matches the null

string. Call a regular expression r standard iff whenever r 0 occurs within

r, the null string is not an element of L(r 0 ). It is easy to check that the proof
given above goes through under the assumption that the regular expression r is standard.
This says that the matcher works correctly for standard regular expressions. But what about the non-standard ones? The key observation is that
every regular expression is equivalent to one in standard form. By equivalent
we mean accepting the same language. For example, the regular expressions r + 0 and r are easily seen to be equivalent. Using this observation
we may avoid the need to consider non-standard regular expressions. Instead we can pre-process the regular expression to put it into standard
form, then call the matcher on the standardized regular expression.
The required pre-processing is based on the following definitions. We
will associate with each regular expression r two standard regular expressions (r ) and r with the following properties:
1. L((r )) = 1 iff L(r ) and L((r )) = 0 otherwise.
2. L(r ) = L(r ) \ 1.
With these equations in mind, we see that every regular expression r may
be written in the form (r ) + r , which is in standard form.
The function mapping regular expressions to regular expressions is

R EVISED 11.02.11

D RAFT

V ERSION 1.2

27.3 Sample Code

228

defined by induction on regular expressions by the following equations:

(0)
(1)
(a)
(r1 + r2 )
(r1 r2 )
(r )

=
=
=
=
=
=

0
1
0
(r1 ) (r2 )
(r1 ) (r2 )
1

Here we define 0 1 = 1 0 = 1 1 = 1 and 0 0 = 0 and 0 1 =

1 0 = 0 0 = 0 and 1 1 = 1.
Exercise 2
Show that L((r )) = 1 iff L(r ).
The definition of r is given by induction on the structure of r by the
following equations:
0
1
a
(r1 + r2 )
(r1 r2 )
(r )

=
=
=
=
=
=

0
0
0
r1 + r2
(r1 ) r2 + r1 (r2 ) + r1 r2

(r ) + r

The only tricky case is the one for concatenation, which must take account
of the possibility that r1 or r2 accepts the null string.
Exercise 3
Show that L(r ) = L(r ) \ 1.

27.3

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 28
Persistent and Ephemeral Data
Structures
This chapter is concerned with persistent and ephemeral abstract types. The
distinction is best explained in terms of the logical future of a value. Whenever a value of an abstract type is created it may be subsequently acted
upon by the operations of the type (and, since the type is abstract, by no
other operations). Each of these operations may yield (other) values of that
abstract type, which may themselves be handed off to further operations
of the type. Ultimately a value of some other type, say a string or an integer, is obtained as an observable outcome of the succession of operations
on the abstract value. The sequence of operations performed on a value of
an abstract type constitutes a logical future of that type a computation
that starts with that value and ends with a value of some observable type.
We say that a type is ephemeral iff every value of that type has at most one
logical future, which is to say that it is handed off from one operation of
the type to another until an observable value is obtained from it. This is
the normal case in familiar imperative programming languages because
in such languages the operations of an abstract type destructively modify
the value upon which they operate; its original state is irretrievably lost by
the performance of an operation. It is therefore inherent in the imperative
programming model that a value have at most one logical future. In contrast, values of an abstract type in functional languages such as ML may
have many different logical futures, precisely because the operations do
not destroy the value upon which they operate, but rather create fresh
values of that type to yield as results. Such values are said to be persistent

230
because they persist after application of an operation of the type, and in
fact may serve as arguments to further operations of that type.
Some examples will help to clarify the distinction. The primitive list
types of ML are persistent because the performance of an operation such
as consing, appending, or reversing a list does not destroy the original
list. This leads naturally to the idea of multiple logical futures for a given
value, as illustrated by the following code sequence:
(* original list *)
val l = [1,2,3]
val m1 = hd l
(* first future of l *)
val n1 = rev (tl m1)
(* second future of l *)
val m2 = l @ [4,5,6]
Notice that the original list value, [1,2,3], has two distinct logical futures,
one in which we remove its head, then reverse the tail, and the other in
which we append the list [4,5,6] to it. The ability to easily handle multiple logical futures for a data structure is a tremendous source of flexibility
and expressive power, alleviating the need to perform tedious bookkeeping to manage versions or copies of a data structure to be passed to
different operations.
The prototypical ephemeral data structure in ML is the reference cell.
Performing an assignment operation on a reference cell changes it irrevocably; the original contents of the cell are lost, even if we keep a handle on
it.
val r = ref 0
(* original cell *)
val s = r
val = (s := 1)
val x = !r
(* 1! *)
Notice that the contents of (the cell bound to) r changes as a result of performing an assignment to the underlying cell. There is only one future
for this cell; a reference to its original binding does not yield its original
contents.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

231
More elaborate forms of ephemeral data structures are certainly possible. For example, the following declaration defines a type of lists whose
tails are mutable. It is therefore a singly-linked list, one whose predecessor
relation may be changed dynamically by assignment:
datatype a mutable list =
Nil |
Cons of a * a mutable list ref
Values of this type are ephemeral in the sense that some operations
on values of this type are destructive, and hence are irreversible (so to
speak!). For example, heres an implementation of a destructive reversal
of a mutable list. Given a mutable list l, this function reverses the links in
the cell so that the elements occur in reverse order of their occurrence in l.
local
fun ipr (Nil, a) = a
| ipr (this as (Cons ( , r as ref next)), a) =
ipr (next, (r := a; this))
in
(* destructively reverse a list *)
fun inplace reverse l = ipr (l, Nil)
end
As you can see, the code is quite tricky to understand! The idea is the
same as the iterative reverse function for pure lists, except that we re-use
the nodes of the original list, rather than generate new ones, when moving
elements onto the accumulator argument.
The distinction between ephemeral and persistent data structures is
essentially the distinction between functional (effect-free) and imperative
(effect-ful) programming functional data structures are persistent; imperative data structures are ephemeral. However, this characterization is
oversimplified in two respects. First, it is possible to implement a persistent data structure that exploits mutable storage. Such a use of mutation
is an example of what is called a benign effect because for all practical purposes the data structure is purely functional (i.e., persistent), but is in
fact implemented using mutable storage. As we will see later the exploitation of benign effects is crucial for building efficient implementations of
persistent data structures. Second, it is possible for a persistent data type
R EVISED 11.02.11

D RAFT

V ERSION 1.2

28.1 Persistent Queues

232

to be used in such a way that persistence is not exploited rather, every value of the type has at most one future in the program. Such a type
is said to be single-threaded, reflecting the linear, as opposed to branching, structure of the future uses of values of that type. The significance
of a single-threaded type is that it may as well have been implemented as
an ephemeral data structure (e.g., by having observable effects on values)
without changing the behavior of the program.

28.1

Persistent Queues

Here is a signature of persistent queues:

signature QUEUE = sig
type a queue
exception Empty
val empty : a queue
val insert : a * a queue -> a queue
val remove : a queue -> a * a queue
end
This signature describes a structure providing a representation type for
queues, together with operations to create an empty queue, insert an element onto the back of the queue, and to remove an element from the front
of the queue. It also provides an exception that is raised in response to
an attempt to remove an element from the empty queue. Notice that removing an element from a queue yields both the element at the front of
the queue, and the queue resulting from removing that element. This is
a direct reflection of the persistence of queues implemented by this signature; the original queue remains available as an argument to further queue
operations.
By a sequence of queue operations we shall mean a succession of uses
of empty, insert, and remove operations in such a way that the queue
argument of one operation is obtained as a result of the immediately preceding queue operation. Thus a sequence of queue operations represents
a single-threaded time-line in the life of a queue value. Here is an example
of a sequence of queue operations:
val q0 : int queue = empty
R EVISED 11.02.11

D RAFT

V ERSION 1.2

28.1 Persistent Queues

val
val
val
val

q1 =
q2 =
(h1,
(h2,

insert (1, q0)

insert (2, q1)
q3) = remove q2
q4) = remove q3

233

(* h1 = 1, q3 = q1 *)
(* h2 = 2, q4 = q0 *)

By contrast the following operations do not form a single thread, but

rather a branching development of the queues lifetime:
val
val
val
val
val
val

q0 :
q1 =
q2 =
(h1,
(h2,
(h2,

int queue = empty

insert (1, q0)
insert (2, q0)
q3) = remove q1
q4) = remove q3
q4) = remove q2

(*
(*
(*
(*

NB: q0, not q1! *)

h1 = 1, q3 = q0 *)
raise Empty *)
h2 = 2,, q4 = q0 *)

In the remainder of this section we will be concerned with single-threaded

sequences of queue operations.
How might we implement the signature QUEUE? The most obvious approach is to represent the queue as a list with, say, the head element of
the list representing the back (most recently enqueued element) of the
queue. With this representation enqueueing is a constant-time operation,
but dequeuing requires time proportional to the number of elements in
the queue. Thus in the worst case a sequence of n enqueue and dequeue
operations will take time O(n2 ), which is clearly excessive. We can make
dequeue simpler, at the expense of enqueue, by regarding the head of the
list as the front of the queue, but the time bound for n operations remains the same in the worst case.
Can we do better? A well-known trick achieves an O(n) worst-case
performance for any sequence of n operations, which means that each operation takes O(1) steps if we amortize the cost over the entire sequence.
Notice that this is a worst-case bound for the sequence, yielding an amortized
bound for each operation of the sequence. This means that some operations
may be relatively expensive, but, in compensation, many operations will
be cheap.
How is this achieved? By combining the two naive solutions sketched
above. The idea is to represent the queue by two lists, one for the back
half consisting of recently inserted elements in the order of arrival, and
one for the front half consisting of soon-to-be-removed elements in reverse order of arrival (i.e., in order of removal). We speak loosely of the
R EVISED 11.02.11

D RAFT

V ERSION 1.2

28.1 Persistent Queues

234

halves of the queue, because we will not, in general, maintain an even

split of elements between the front and the back lists. Rather, we will arrange things so that the following representation invariants holds true:
1. The elements of the queue listed in order of removal are the elements
of the front followed by the elements of the back in reverse order.
2. The front is empty only if the back is empty.
These invariants are maintained by using a smart constructor that
creates a queue from two lists representing the back and front parts of the
queue. This constructor ensures that the representation invariant holds
by ensuring that condition (2) is always true of the resulting queue. The
constructor proceeds by a case analysis on the back and front parts of the
queue. If the front list is non-empty, or both the front and back are empty,
the resulting queue consists of the back and front parts as given. If the
front is empty and the back is non-empty, the queue constructor yields
the queue consisting of an empty back part and a front part equal to the
reversal of the given back part. Observe that this is sufficient to ensure
that the representation invariant holds of the resulting queue in all cases.
Observe also that the smart constructor either runs in constant time, or in
time proportional to the length of the back part, according to whether the
front part is empty or not.
Insertion of an element into a queue is achieved by consing the element onto the back of the queue, then calling the queue constructor to
ensure that the result is in conformance with the representation invariant. Thus an insert can either take constant time, or time proportional to
the size of the back of the queue, depending on whether the front part is
empty. Removal of an element from a queue requires a case analysis. If
the front is empty, then by condition (2) the queue is empty, so we raise an
exception. If the front is non-empty, we simply return the head element
together with the queue created from the original back part and the front
part with the head element removed. Here again the time required is either constant or proportional to the size of the back of the queue, according
to whether the front part becomes empty after the removal. Notice that if
an insertion or removal requires a reversal of k elements, then the next k
operations are constant-time. This is the fundamental insight as to why
we achieve O(n) time complexity over any sequence of n operations. (We
will give a more rigorous analysis shortly.)
R EVISED 11.02.11

D RAFT

V ERSION 1.2

28.2 Amortized Analysis

235

Heres the implementation of this idea in ML:

structure Queue :> QUEUE = struct
type a queue = a list * a list
fun make queue (q as (nil, nil)) = q
| make queue (bs, nil) = (nil, rev bs)
| make queue (q as (bs, fs)) = q
val empty = make queue (nil, nil)
fun insert (x, (back,front)) =
make queue (x::back, front)
exception Empty
fun remove ( , nil) = raise Empty
| remove (bs, f::fs) = (f, make queue (bs, fs))
end
Notice that we call the smart constructor make queue whenever we
wish to return a queue to ensure that the representation invariant holds.
Consequently, some queue operations are more expensive than others, according to whether or not the queue needs to be reorganized to satisfy the
representation invariant. However, each such reorganization makes a corresponding number of subsequent queue operations cheap (constanttime), so the overall effort required evens out in the end to constant-time
per operation. More precisely, the running time of a sequence of n queue
operations is now O(n), rather than O(n2 ), as it was in the naive implementation. Consequently, each operation takes O(1) (constant) time on
average, i.e., when the total effort is evenly apportioned among each of
the operations in the sequence. Note that this is a worst-case time bound
for each operation, amortized over the entire sequence, not an average-case time
bound based on assumptions about the distribution of the operations.

28.2

Amortized Analysis

How can we prove this claim? First we given an informal argument, then
we tighten it up with a more rigorous analysis. We are to account for the
total work performed by a sequence of n operations by showing that any
sequence of n operations can be executed in cn steps for some constant
c. Dividing by n, we obtain the result that each operations takes c steps
when amortized over the entire sequence. The key is to observe first that
R EVISED 11.02.11

D RAFT

V ERSION 1.2

28.2 Amortized Analysis

236

the work required to execute a sequence of queue operations may be apportioned to the elements themselves, then that only a constant amount of
work is expended on each element. The life of a queue element may be
divided into three stages: its arrival in the queue, its transit time in the
queue, and its departure from the queue. In the worst case each element
passes through each of these stages (but may die young, never participating in the second or third stage). Arrival requires constant time to
add the element to the back of the queue. Transit consists of being moved
from the back to the front by a reversal, which takes constant time per
element on the back. Departure takes constant time to pattern match and
extract the element. Thus at worst we require three steps per element to account for the entire effort expended to perform a sequence of queue operations. This is in fact a conservative upper bound in the sense that we may
need less than 3n steps for the sequence, but asymptotically the bound is
optimal we cannot do better than constant time per operation! (You
might reasonably wonder whether there is a worst-case, non-amortized
constant-time implementation of persistent queues. The answer is yes,
but the code is far more complicated than the simple implementation we
are sketching here.)
This argument can be made rigorous as follows. The general idea is to
introduce the notion of a charge scheme that provides an upper bound on
the actual cost of executing a sequence of operations. An upper bound on
the charge will then provide an upper bound on the actual cost. Let T (n)
be the cumulative time required (in the worst case) to execute a sequence
of n queue operations. We will introduce a charge function, C (n), representing the cumulative charge for executing a sequence of n operations and
show that T (n) C (n) = O(n). It is convenient to express this in terms
of a function R(n) = C (n) T (n) representing the cumulative residual, or
overcharge, which is the amount that the charge for n operations exceeds
the actual cost of executing them. We will arrange things so that R(n) 0
and that C (n) = O(n), from which the result follows immediately.
Down to specifics. By charging 2 for each insert operation and 1 for
each remove, it follows that C (n) 2n for any sequence of n inserts and
removes. Thus C (n) = O(n). After any sequence of n 0 operations
have been performed, the queue contains 0 b n elements on the back
half and 0 f n elements on the front half. We claim that for every
n 0, R(n) = b. We prove this by induction on n 0. The condition
clearly holds after performing 0 operations, since T (0) = 0, C (0) = 0,
R EVISED 11.02.11

D RAFT

V ERSION 1.2

28.2 Amortized Analysis

237

and hence R(0) = C (0) T (0) = 0. Consider the n + 1st operation. If

it is an insert, and f > 0, T (n + 1) = T (n) + 1, C (n + 1) = C (n) + 2,
and hence R(n + 1) = R(n) + 1 = b + 1. This is correct because an insert
operation adds one element to the back of the queue. If, on the other hand,
f = 0, then T (n + 1) = T (n) + b + 2 (charging one for the cons and one
for creating the new pair of lists), C (n + 1) = C (n) + 2, so R(n + 1) =
R(n) + 2 b 2 = b + 2 b 2 = 0. This is correct because the back is
now empty; we have used the residual overcharge to pay for the cost of the
reversal. If the n + 1st operation is a remove, and f > 0, then T (n + 1) =
T (n) + 1 and C (n + 1) = C (n) + 1 and hence R(n + 1) = R(n) = b. This
is correct because the remove doesnt disturb the back in this case. Finally,
if we are performing a remove with f = 0, then T (n + 1) = T (n) + b + 1,
C (n + 1) = C (n) + 1, and hence R(n + 1) = R(n) b = b b = 0. Here
again we use of the residual overcharge to pay for the reversal of the back
to the front. The result follows immediately since R(n) = b 0, and hence
C ( n ) T ( n ).
It is instructive to examine where this solution breaks down in the
multi-threaded case (i.e., where persistence is fully exploited). Suppose
that we perform a sequence of n insert operations on the empty queue,
resulting in a queue with n elements on the back and none on the front.
Call this queue q. Let us suppose that we have n independent futures
for q, each of which removes an element from it, for a total of 2n operations. How much time do these 2n operations take? Since each independent future must reverse all n elements onto the front of the queue before
performing the removal, the entire collection of 2n operations takes n + n2
steps, or O(n) steps per operation, breaking the amortized constant-time
bound just derived for a single-threaded sequence of queue operations.
Can we recover a constant-time amortized cost in the persistent case? We
can, provided that we share the cost of the reversal among all futures of
q as soon as one performs the reversal, they all enjoy the benefit of its
having been done. This may be achieved by using a benign side effect to
cache the result of the reversal in a reference cell that is shared among all
uses of the queue. We will return to this once we introduce memoization
and lazy evaluation.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

28.3 Sample Code

28.3

238

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 29
Options, Exceptions, and
Continuations
In this chapter we discuss the close relationships between option types,
exceptions, and continuations. They each provide the means for handling
failure to produce a value in a computation. Option types provide the
means of explicitly indicating in the type of a function the possibility that
it may fail to yield a normal result. The result type of the function forces
the caller to dispatch explicitly on whether or not it returned a normal
value. Exceptions provide the means of implicitly signalling failure to return a normal result value, without sacrificing the requirement that an application of such a function cannot ignore failure to yield a value. Continuations provide another means of handling failure by providing a function
to invoke in the case that normal return is impossible.

29.1

The n-Queens Problem

We will explore the trade-offs between these three approaches by considering three different implementations of the n-queens problem: find a way
to place n queens on an n n chessboard in such a way that no two queens
attack one another. The general strategy is to place queens in successive
columns in such a way that it is not attacked by a previously placed queen.
Unfortunately its not possible to do this in one pass; we may find that we
can safely place k < n queens on the board, only to discover that there is
no way to place the next one. To find a solution we must reconsider earlier

29.1 The n-Queens Problem

240

decisions, and work forward from there. If all possible reconsiderations of

all previous decisions all lead to failure, then the problem is unsolvable.
For example, there is no safe placement of three queens on a 3x3 chessboard. This trial-and-error approach to solving the n-queens problem is
called backtracking search.
A solution to the n-queens problem consists of an n n chessboard
with n queens safely placed on it. The following signature defines a chessboard abstraction:
signature BOARD =
sig
type board
val new : int -> board
val complete : board -> bool
val place : board * int -> board
val safe : board * int -> bool
val size : board -> int
val positions : board -> (int * int) list
end
The operation new creates a new board of a given dimension n 0. The
operation complete checks whether the board contains a complete safe
placement of n queens. The function safe checks whether it is safe to
place a queen at row i in the next free column of a board B. The operation
place puts a queen at row i in the next available column of the board.
The function size returns the size of a board, and the function positions
returns the coordinates of the queens on the board.
The board abstraction may be implemented as follows:
structure Board :> BOARD =
struct
(* rep: size, next free column, number placed, placements
inv: size>=0, 1<=next free<=size,
length(placements) = number placed
*)
type board = int * int * int * (int * int) list
fun new n = (n, 1, 0, nil)
fun size (n, , , ) = n
R EVISED 11.02.11

D RAFT

V ERSION 1.2

29.2 Solution Using Options

241

fun complete (n, , k, ) = (k=n)

fun positions ( , , , qs) = qs
fun place ((n, i, k, qs),j) =
(n, i+1, k+1, (i,j)::qs)
fun threatens ((i,j), (i,j)) =
i=i orelse j=j orelse
i+j = i+j orelse
i-j = i-j
fun conflicts (q, nil) =
false
| conflicts (q, q::qs) =
threatens (q, q) orelse conflicts (q, qs)
fun safe (( , i, , qs), j) =
not (conflicts ((i,j), qs))
end
The representation type contains redundant information in order to make
the individual operations more efficient. The representation invariant ensures that the components of the representation are properly related to one
another (e.g., the claimed number of placements is indeed the length of the
list of placed queens, and so on.)
Our goal is to define a function
val queens : int -> Board.board option
such that if n 0, then queens n evaluates either to NONE if there is no safe
placement of n queens on an n n board, or to SOME B otherwise, with B a
complete board containing a safe placement of n queens. We will consider
three different solutions, one using option types, one using exceptions,
and one using a failure continuation.

29.2

Solution Using Options

Heres a solution based on option types:

(* addqueen bd yields SOME bd with bd a
complete safe placement extending bd,
R EVISED 11.02.11

D RAFT

V ERSION 1.2

29.3 Solution Using Exceptions

242

if one exists, and yields NONE otherwise

*)
fun addqueen bd =
let
fun try j =
if j > Board.size bd then
NONE
else if Board.safe (bd, j) then
case addqueen (Board.place (bd, j))
of NONE => try (j+1)
| r as (SOME bd) => r
else
try (j+1)
in
if Board.complete bd then
SOME bd
else
try 1
end
fun queens n = addqueen (Board.new n)
The characteristic feature of this solution is that we must explicitly check
the result of each recursive call to addqueen to determine whether a safe
placement is possible from that position. If so, we simply return it; if not,
we must reconsider the placement of a queen in row j of the next available
column. If no placement is possible in the current column, the function
yields NONE, which forces reconsideration of the placement of a queen in
the preceding row. Eventually we either find a safe placement, or yield
NONE indicating that no solution is possible.

29.3

Solution Using Exceptions

The explicit check on the result of each recursive call can be replaced by
the use of exceptions. Rather than have addqueen return a value of type
Board.board option, we instead have it return a value of type Board.board,
if possible, and otherwise raise an exception indicating failure. The case
analysis on the result is replaced by a use of an exception handler. Heres
the code:
R EVISED 11.02.11

D RAFT

V ERSION 1.2

29.3 Solution Using Exceptions

243

exception Fail
(* addqueen bd yields bd, where bd is a complete safe
placement extending bd, if one exists, and raises Fail otherwise
*)
fun addqueen bd =
let
fun try j =
if j > Board.size bd then
raise Fail
else if Board.safe (bd, j) then
addqueen (Board.place (bd, j))
handle Fail => try (j+1)
else
try (j+1)
in
if Board.complete bd then
bd
else
try 1
end
fun queens n =
SOME (addqueen (Board.new n))
handle Fail => NONE
The main difference between this solution and the previous one is that
both calls to addqueen must handle the possibility that it raises the exception Fail. In the outermost call this corresponds to a complete failure to
find a safe placement, which means that queens must return NONE. If a safe
placement is indeed found, it is wrapped with the constructor SOME to indicate success. In the recursive call within try, an exception handler is
required to handle the possibility of there being no safe placement starting in the current position. This check corresponds directly to the case
analysis required in the solution based on option types.
What are the trade-offs between the two solutions?
1. The solution based on option types makes explicit in the type of
the function addqueen the possibility of failure. This forces the programmer to explicitly test for failure using a case analysis on the result of the call. The type checker will ensure that one cannot use a
R EVISED 11.02.11

D RAFT

V ERSION 1.2

29.4 Solution Using Continuations

244

Board.board option where a Board.board is expected. The solution

based on exceptions does not explicitly indicate failure in its type.
However, the programmer is nevertheless forced to handle the failure, for otherwise an uncaught exception error would be raised at
run-time, rather than compile-time.
2. The solution based on option types requires an explicit case analysis
on the result of each recursive call. If most results are successful,
the check is redundant and therefore excessively costly. The solution
based on exceptions is free of this overhead: it is biased towards the
normal case of returning a board, rather than the failure case
of not returning a board at all. The implementation of exceptions
ensures that the use of a handler is more efficient than an explicit
case analysis in the case that failure is rare compared to success.
For the n-queens problem it is not clear which solution is preferable. In
general, if efficiency is paramount, we tend to prefer exceptions if failure
is a rarity, and to prefer options if failure is relatively common. If, on
the other hand, static checking is paramount, then it is advantageous to
use options since the type checker will enforce the requirement that the
programmer check for failure, rather than having the error arise only at
run-time.

29.4

Solution Using Continuations

We turn now to a third solution based on continuation-passing. The idea is

quite simple: an exception handler is essentially a function that we invoke
when we reach a blind alley. Ordinarily we achieve this invocation by
raising an exception and relying on the caller to catch it and pass control
to the handler. But we can, if we wish, pass the handler around as an
additional argument, the failure continuation of the computation. Heres
how its done in the case of the n-queens problem:
(* addqueen bd yields bd, where bd is a complete safe
placement extending bd, if one exists, and otherwise,
yields the value of fc ()
*)
fun addqueen (bd, fc) =
R EVISED 11.02.11

D RAFT

V ERSION 1.2

29.4 Solution Using Continuations

245

let
fun try j =
if j > Board.size bd then
fc ()
else if Board.safe (bd, j) then
addqueen
(Board.place (bd, j),
fn () => try (j+1))
else
try (j+1)
in
if Board.complete bd then
SOME bd
else
try 1
end
fun queens n =
addqueen (Board.new n, fn () => NONE)
Here again the differences are small, but significant. The initial continuation simply yields NONE, reflecting the ultimate failure to find a safe placement. On a recursive call we pass to addqueen a continuation that resumes
search at the next row of the current column. Should we exceed the number of rows on the board, we invoke the failure continuation of the most
recent call to addqueen.
The solution based on continuations is very close to the solution based
on exceptions, both in form and in terms of efficiency. Which is preferable?
Here again there is no easy answer, we can only offer general advice. First
off, as weve seen in the case of regular expression matching, failure continuations are more powerful than exceptions; there is no obvious way to
replace the use of a failure continuation with a use of exceptions in the
matcher. However, in the case that exceptions would suffice, it is generally preferable to use them since one may then avoid passing an explicit
failure continuation. More significantly, the compiler ensures that an uncaught exception aborts the program gracefully, whereas failure to invoke
a continuation is not in itself a run-time fault. Using the right tool for the
right job makes life easier.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

29.5 Sample Code

29.5

246

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 30
Higher-Order Functions
Higher-order functions those that take functions as arguments or return functions as results are powerful tools for building programs. An
interesting application of higher-order functions is to implement infinite
sequences of values as (total) functions from the natural numbers (nonnegative integers) to the type of values of the sequence. We will develop a
small package of operations for creating and manipulating sequences, all
of which are higher-order functions since they take sequences (functions!)
as arguments and/or return them as results. A natural way to define many
sequences is by recursion, or self-reference. Since sequences are functions,
we may use recursive function definitions to define such sequences. Alternatively, we may think of such a sequence as arising from a loopback
or feedback construct. We will explore both approaches.
Sequences may be used to simulate digital circuits by thinking of a
wire as a sequence of bits developing over time. The ith value of the
sequence corresponds to the signal on the wire at time i. For simplicity
we will assume a perfect waveform: the signal is always either high or
low (or is undefined); we will not attempt to model electronic effects such
as attenuation or noise. Combinational logic elements (such as and gates
or inverters) are operations on wires: they take in one or more wires as
input and yield one or more wires as results. Digital logic elements (such
as flip-flops) are obtained from combinational logic elements by feedback,
or recursion a flip-flop is a recursively-defined wire!

30.1 Infinite Sequences

30.1

248

Infinite Sequences

Let us begin by developing a sequence package. Here is a suitable signature defining the type of sequences:
signature SEQUENCE =
sig
type a seq = int -> a
(* constant sequence *)
val constantly : a -> a seq
(* alternating values *)
val alternately : a * a -> a seq
(* insert at front *)
val insert : a * a seq -> a seq
val map : (a -> b) -> a seq -> b seq
val zip : a seq * b seq -> (a * b) seq
val unzip : (a * b) seq -> a seq * b seq
(* fair merge *)
val merge : (a * a) seq -> a seq
val stretch : int -> a seq -> a seq
val shrink : int -> a seq -> a seq
val take : int -> a seq -> a list
val drop : int -> a seq -> a seq
val shift : a seq -> a seq
val loopback : (a seq -> a seq) -> a seq
end
Observe that we expose the representation of sequences as functions. This
is done to simplify the definition of recursive sequences as recursive functions. Alternatively we could have hidden the representation type, at the
expense of making it a bit more awkward to define recursive sequences.
In the absence of this exposure of representation, recursive sequences may
only be built using the loopback operation which constructs a recursive
sequence by looping back the output of a sequence transformer to its
input. Most of the other operations of the signature are adaptations of
familiar operations on lists. Two exceptions to this rule are the functions
stretch and shrink that dilate and contract the sequence by a given time
R EVISED 11.02.11

D RAFT

V ERSION 1.2

30.1 Infinite Sequences

249

parameter if a sequence is expanded by k, its value at i is the value of

the original sequence at i/k, and dually for shrinking.
Heres an implementation of sequences as functions.
structure Sequence :> SEQUENCE =
struct
type a seq = int -> a
fun constantly c n = c
fun alternately (c,d) n =
if n mod 2 = 0 then c else d
fun insert (x, s) 0 = x
| insert (x, s) n = s (n-1)
fun map f s = f o s
fun zip (s1, s2) n = (s1 n, s2 n)
fun unzip (s : (a * b) seq) =
(map #1 s, map #2 s)
fun merge (s1, s2) n =
(if n mod 2 = 0 then s1 else s2) (n div 2)
fun stretch k s n = s (n div k)
fun shrink k s n = s (n * k)
fun drop k s n = s (n+k)
fun shift s = drop 1 s
fun take 0 = nil
| take n s = s 0 :: take (n-1) (shift s)
fun loopback loop n = loop (loopback loop) n
end
Most of this implementation is entirely straightforward, given the ease
with which we may manipulate higher-order functions in ML. The only
tricky function is loopback, which must arrange that the output of the
function loop is looped back to its input. This is achieved by a simple
recursive definition of a sequence whose value at n is the value at n of the
sequence resulting from applying the loop to this very sequence.
The sensibility of this definition of loopback relies on two separate
ideas. First, notice that we may not simplify the definition of loopback
as follows:

R EVISED 11.02.11

D RAFT

V ERSION 1.2

30.1 Infinite Sequences

250

(* bad definition *)
fun loopback loop = loop (loopback loop)
The reason is that any application of loopback will immediately loop forever! In contrast, the original definition is arranged so that application of
loopback immediately returns a function. This may be made more apparent by writing it in the following form, which is entirely equivalent to the
definition given above:
fun loopback loop =
fn n => loop (loopback loop) n
This format makes it clear that loopback immediately returns a function
when applied to a loop functional.
Second, for an application of loopback to a loop to make sense, it must
be the case that the loop returns a sequence without touching the argument sequence (i.e., without applying the argument to a natural number). Otherwise accessing the sequence resulting from an application of
loopback would immediately loop forever. Some examples will help to
illustrate the point.
First, lets build a few sequences without using the loopback function,
just to get familiar with using sequences:
val
val
val
fun

evens : int seq = fn n => 2*n

odds : int seq = fn n => 2*n+1
nats : int seq = merge (evens, odds)
fibs n =
(insert
(1, insert
(1, map (op +)
(zip (drop 1 fibs, fibs)))))(n)

We may inspect the sequence using take and drop, as follows:

take 10 nats
take 5 (drop 5 nats)
take 5 fibs

(* [0,1,2,3,4,5,6,7,8,9] *)
(* [5,6,7,8,9] *)
(* [1,1,2,3,5] *)

Now lets consider an alternative definition of fibs that uses the loopback
operation:
R EVISED 11.02.11

D RAFT

V ERSION 1.2

30.2 Circuit Simulation

251

fun fibs loop s =

insert (1, insert (1,
map (op +) (zip (drop 1 s, s))))
val fibs = loopback fibs loop;
The definition of fibs loop is exactly like the original definition of fibs,
except that the reference to fibs itself is replaced by a reference to the
argument s. Notice that the application of fibs loop to an argument s
does not inspect the argument s!
One way to understand loopback is that it solves a system of equations
for an unknown sequence. In the case of the second definition of fibs, we
are solving the following system of equations for f :
f0 = 1
f1 = 1
f ( n + 2) = f ( n + 1) + f ( n )
These equations are derived by inspecting the definitions of insert, map,
zip, and drop given earlier. It is obvious that the solution is the Fibonacci
sequence; this is precisely the sequence obtained by applying loopback to
fibs loop.
Heres an example of a loop that, when looped back, yields an undefined sequence any attempt to access it results in an infinite loop:
fun bad loop s n = s n + 1
val bad = loopback bad loop
(* infinite loop! *)
val = bad 0
In this example we are, in effect, trying to solve the equation sn = sn + 1
for s, which has no solution (except the totally undefined sequence). The
problem is that the next element of the output is defined in terms of the
next element itself, rather than in terms of previous elements. Consequently, no solution exists.

30.2

Circuit Simulation

With these ideas in mind, we may apply the sequence package to build
an implementation of digital circuits. Lets start with wires, which are
represented as sequences of levels:
R EVISED 11.02.11

D RAFT

V ERSION 1.2

30.2 Circuit Simulation

252

datatype level = High | Low | Undef

type wire = level seq
type pair = (level * level) seq
val Zero : wire = constantly Low
val One : wire = constantly High
(* clock pulse with given duration of each pulse *)
fun clock (freq:int):wire =
stretch freq (alternately (Low, High))
We include the undefined level to account for propagation delays and
settling times in circuit elements.
Combinational logic elements (gates) may be defined as follows. We introduce an explicit unit time propagation delay for each gate the output
is undefined initially, and is then determined as a function of its inputs. As
we build up layers of circuit elements, it takes longer and longer (proportional to the length of the longest path through the circuit) for the output
to settle, exactly as in real life.
(* apply two functions in parallel *)
infixr **;
fun (f ** g) (x, y) = (f x, g y)
(* hardware logical and *)
fun logical and (Low, ) = Low
| logical and ( , Low) = Low
| logical and (High, High) = High
| logical and = Undef
fun logical not Undef = Undef
| logical not High = Low
| logical not Low = High
fun logical nop l = l
(* a nor b = not a and not b *)
val logical nor =
logical and o (logical not ** logical not)
type unary gate = wire -> wire
type binary gate = pair -> wire
fun gate f w 0 = Undef
(* logic gate with unit propagation delay *)
R EVISED 11.02.11

D RAFT

V ERSION 1.2

30.2 Circuit Simulation

|
val
val
val

253

gate f w i = f (w (i-1))
(* unit delay *)
delay : unary gate = gate logical nop
inverter : unary gate = gate logical not
nor gate : binary gate = gate logical nor

It is a good exercise to build a one-bit adder out of these elements, then

to string them together to form an n-bit ripple-carry adder. Be sure to
present the inputs to the adder with sufficient pulse widths to ensure that
the circuit has time to settle!
Combining these basic logic elements with recursive definitions allows
us to define digital logic elements such as the RS flip-flop. The propagation delay inherent in our definition of a gate is fundamental to ensuring
that the behavior of the flip-flop is well-defined! This is consistent with
real life flip-flops depend on the existence of a hardware propagation
delay for their proper functioning. Note also that presentation of illegal
inputs (such as setting both the R and the S leads high results in metastable
behavior of the flip-flop, here as in real life. Finally, observe that the flipflop exhibits a momentary glitch in its output before settling, exactly as
in hardware.1
fun RS ff (S : wire, R : wire) =
let
fun X n = nor gate (zip (S, Y))(n)
and Y n = nor gate (zip (X, R))(n)
in
Y
end
(* generate a pulse of bs n wide, followed by w *)
fun pulse b 0 w i = w i
| pulse b n w 0 = b
| pulse b n w i = pulse b (n-1) w (i-1)
val S = pulse Low 2 (pulse High 2 Zero);
val R = pulse Low 6 (pulse High 2 Zero);
val Q = RS ff (S, R);
val = take 20 Q;
val X = RS ff (S, S);
(* unstable! *)
1 All

of these behaviors may be observed by using take and drop to inspect the values
on the circuit.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

30.3 Regular Expression Matching, Revisited

val

254

= take 20 X;

It is a good exercise to derive a system of equations governing the RS flipflop from the definition weve given here, using the implementation of the
sequence operations given above. Observe that the delays arising from
the combinational logic elements ensure that a solution exists by ensuring
that the next element of the output refers only the previous elements,
and not the current element.
Finally, we consider a variant implementation of an RS flip-flop using
the loopback operation:
fun loopback2 (f : wire * wire -> wire * wire) =
unzip (loopback (zip o f o unzip))
fun RS ff (S : wire, R : wire) =
let
fun RS loop (X, Y) =
(nor gate (zip (S, Y)),
nor gate (zip (X, R)))
in
loopback2 RS loop
end
Here we must define a binary loopback function to implement the flipflop. This is achieved by reducing binary loopback to unary loopback by
composing with zip and unzip.

30.3

Regular Expression Matching, Revisited

The regular expression matcher introduced in Chapter 1 may be elegantly

reformulated using higher-order functions. The code for the matcher may
be restructured to separate the high-level control structure of the matcher
from the details of how that control structure is implemented. A matcher
is a function that, given a string, matches an initial segment of that string.
A match combinator is a higher-order function that combines zero or more
matchers to form a compound matcher with the following behavior:
1. FAIL: always fails to match;
2. NULL: always matches the null string;
R EVISED 11.02.11

D RAFT

V ERSION 1.2

30.3 Regular Expression Matching, Revisited

255

3. LITERALLY c: match exactly the character c at the start of the input

string, yielding the remainder of the string;
4. m1 OR m2 : match m1 on a prefix of the input, if possible; if not, match
m2 on a prefix of the same input; otherwise fail;
5. m1 THEN m2 : match m1 on a prefix of the input, and then match m2
on the remaining suffix.
6. REPEATEDLY m: match m zero or more times on a prefix of the input.
Using these combinators we may present the regular expression matcher
in a particularly concise form:
fun match Zero = FAIL
| match One = NULL
| match (Char c) = LITERALLY c
| match (Plus (r1, r2)) = match r1 OR match r2
| match (Times (r1, r2)) = match r1 THEN match r2
| match (Star r) = REPEATEDLY (match r)
fun accepts regexp =
let
val matcher = match regexp
val done = fn nil => true | => false
in
fn str => matcher (String.explode string) done
end
The auxiliary function, match, transforms each regular expression into a
matcher, using the combinators just described.
The definition of the function match illustrates the concept of staged
computation. The function match is defined by induction on the structure
of the regular expression. It transforms each regular expression into a
matcher for that regular expression, using the combinators to build up the
matcher by inductive analysis of the regular expression. The definition of
accepts ensures that the regular expression is analyzed once, yielding a
matcher that may be applied many times to candidate strings, determining
whether or not they match the given regular expression.
It remains to define the matching combinators. A matcher is a function
of type
R EVISED 11.02.11

D RAFT

V ERSION 1.2

30.4 Sample Code

256

char list -> (char list -> bool) -> bool

that is given a character list and a continuation, and yields either true or
false (please see Chapter 1 to review how the matcher works). The match
combinators are defined as follows:
fun FAIL cs k = false
fun NULL cs k = k cs
fun LITERALLY c cs k =
(case cs of nil => false | c::cs => (c=c) andalso (k cs))
fun OR (m1, m2) cs k = m1 cs k orelse m2 cs k
infix 8 OR
fun THEN (m1, m2) cs k = m1 cs (fn cs => m2 cs k)
infix 9 THEN
fun REPEATEDLY m cs k =
let
fun mstar cs = k cs orelse m cs mstar
in
mstar cs
end
The details of the control flow, which is managed through the use of continuations, is packaged into the match combinators, so that the matcher
itself is as perspicuous as possible.

30.4

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 31
Memoization
In this chapter we will discuss memoization, a programming technique for
cacheing the results of previous computations so that they can be quickly
retrieved without repeated effort. Memoization is fundamental to the implementation of lazy data structures, either by hand or using the provisions of the SML/NJ compiler.

31.1

Cacheing Results

We begin with a discussion of memoization to increase the efficiency of

computing a recursively-defined function whose pattern of recursion involves a substantial amount of redundant computation. The problem is
to compute the number of ways to parenthesize an expression consisting
of a sequence of n multiplications as a function of n. For example, the
expression 2 3 4 5 can be parenthesized in 5 ways:

((2 3) 4) 5, (2 (3 4)) 5, (2 3) (4 5), 2 (3 (4 5)), 2 ((3 4) 5).

A simple recurrence expresses the number of ways of parenthesizing a
sequence of n multiplications:
fun
|
fun
|

sum
sum
p 1
p n

f
f
=
=

0 = 0
n = (f n) + sum f (n-1)
1
sum (fn k => (p k) * (p (n-k))) (n-1)

31.1 Cacheing Results

258

where sum f ncomputes the sum of values of a function f (k ) with 1 k

n. This program is extremely inefficient because of the redundancy in the
pattern of the recursive calls.
What can we do about this problem? One solution is to be clever and
solve the recurrence. As it happens this recurrence has a closed-form solution (the Catalan numbers). But in many cases there is no known closed
form, and something else must be done to cut down the overhead. In this
case a simple cacheing technique proves effective. The idea is to maintain
a table of values of the function that is filled in whenever the function is
applied. If the function is called on an argument n, the table is consulted
to see whether the value has already been computed; if so, it is simply
returned. If not, we compute the value and store it in the table for future
use. This ensures that no redundant computations are performed. We will
maintain the table as an array so that its entries can be accessed in constant
time. The penalty is that the array has a fixed size, so we can only record
the values of the function at some pre-determined set of arguments. Once
we exceed the bounds of the table, we must compute the value the hard
way. An alternative is to use a dictionary (e.g., a balanced binary search
tree) which has no a priori size limitation, but which takes logarithmic time
to perform a lookup. For simplicity well use a solution based on arrays.
Heres the code to implement a memoized version of the parenthesization function:
local
val
val
in
fun
|
and

limit = 100
memopad = Array.array (100, NONE)
p 1 = 1
p n = sum (fn k => (p k) * (p (n-k))) (n-1)
p n =
if n < limit then
case Array.sub of
SOME r => r
| NONE =>
let
val r = p n
in

R EVISED 11.02.11

D RAFT

V ERSION 1.2

31.2 Laziness

259
Array.update (memopad, n, SOME r);
r
end

else
p n
end
The main idea is to modify the original definition so that the recursive
calls consult and update the memopad. The exported version of the
function is the one that refers to the memo pad. Notice that the definitions
of p and p are mutually recursive!

31.2

Laziness

Lazy evaluation is a combination of delayed evaluation and memoization.

Delayed evaluation is implemented using thunks, functions of type unit
-> a. To delay the evaluation of an expression exp of typea, simply write
fn () => exp. This is a value of type unit -> a; the expression exp is effectively frozen until the function is applied. To thaw the expression,
simply apply the thunk to the null tuple, (). Heres a simple example:
val thunk =
fn () => print "hello"
val = thunk ()

(* nothing printed *)
(* prints hello *)

While this example is especially simple-minded, remarkable effects can

be achieved by combining delayed evaluation with memoization. To do
so, we will consider the following signature of suspensions:
signature SUSP =
sig
type a susp
val force : a susp -> a
val delay : (unit -> a) -> a susp
end
The function delay takes a suspended computation (in the form of a
thunk) and yields a suspension. Its job is to memoize the suspension
R EVISED 11.02.11

D RAFT

V ERSION 1.2

31.2 Laziness

260

so that the suspended computation is evaluated at most once once the

result is computed, the value is stored in a reference cell so that subsequent
forces are fast. The implementation is slick. Heres the code to do it:
structure Susp :> SUSP =
struct
type a susp = unit -> a
fun force t = t ()
fun delay (t : a susp) =
let
exception Impossible
val memo : a susp ref =
ref (fn () => raise Impossible)
fun t () =
let val r = t ()
in memo := (fn () => r); r end
in
memo := t;
fn () => (!memo)()
end
end
Its worth discussing this code in detail because it is rather tricky. Suspensions are just thunks; force simply applies the suspension to the null
tuple to force its evaluation. What about delay? When applied, delay allocates a reference cell containing a thunk that, if forced, raises an internal
exception. This can never happen for reasons that will become apparent in
a moment; it is merely a placeholder with which we initialize the reference
cell. We then define another thunk t that, when forced, does three things:
1. It forces the thunk t to obtain its value r.
2. It replaces the contents of the memopad with the constant function
that immediately returns r.
3. It returns r as result.
We then assign t to the memo pad (hence obliterating the placeholder),
and return a thunk dt that, when forced, simply forces the contents of the
memo pad. Whenever dt is forced, it immediately forces the contents of
R EVISED 11.02.11

D RAFT

V ERSION 1.2

31.3 Lazy Data Types in SML/NJ

261

the memo pad. However, the contents of the memo pad changes as a result
of forcing it so that subsequent forces exhibit different behavior. Specifically, the first time dt is forced, it forces the thunk t, which then forces
t its value r, zaps the memo pad, and returns r. The second time dt is
forced, it forces the contents of the memo pad, as before, but this time the
it contains the constant function that immediately returns r. Altogether
we have ensured that t is forced at most once by using a form of selfmodifying code.
Heres an example to illustrate the effect of delaying a thunk:
val t = Susp.delay (fn () => print "hello")
(* prints hello *)
val = Susp.force t
val = Susp.force t
(* silent *)
Notice that hello is printed once, not twice! The reason is that the suspended computation is evaluated at most once, so the message is printed
at most once on the screen.

31.3

Lazy Data Types in SML/NJ

The lazy datatype declaration1

datatype lazy a stream = Cons of a * a stream
expands into the following pair of type declarations
datatype a stream! = Cons of a * a stream
withtype a stream = a stream! Susp.susp
The first defines the type of stream values, the result of forcing a stream
computation, the second defines the type of stream computations, which
are suspensions yielding stream values. Thus streams are represented
by suspended (unevaluated, memoized) computations of stream values,
which are formed by applying the constructor Cons to a value and another
stream.
The value constructor Cons, when used to build a stream, automatically
suspends computation. This is achieved by regarding Cons e as shorthand
for Cons (Susp.susp (fn () => e). When used in a pattern, the value
constructor Cons induces a use of force. For example, the binding
1 Please

see chapter 15 for a description of the SML/NJ lazy data type mechanism.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

31.3 Lazy Data Types in SML/NJ

262

val Cons (h, t) = e

becomes
val Cons (h, t) = Susp.force e
which forces the right-hand side before performing pattern matching.
A similar transformation applies to non-lazy function definitions the
argument is forced before pattern matching commences. Thus the eager
tail function
fun stl (Cons ( , t)) = t
expands into
fun stl! (Cons ( , t)) = t
and stl s = stl! (Susp.force s)
which forces the argument as soon as it is applied.
On the other hand, lazy function definitions defer pattern matching
until the result is forced. Thus the lazy tail function
fun lstl (Cons ( , t)) = t
expands into
fun lstl! (Cons ( , t)) =
t
and lstl s =
Susp.delay (fn () => lstl! (Susp.force s))
which a suspension that, when forced, performs the pattern match.
Finally, the recursive stream definition
val rec lazy ones = Cons (1, ones)
expands into the following recursive function definition:
val rec ones = Susp.delay (fn () => Cons (1, ones))
Unfortunately this is not quite legal in SML since the right-hand side
involves an application of a a function to another function. This can either
be provided by extending SML to admit such definitions, or by extending
the Susp package to include an operation for building recursive suspensions such as this one. Since it is an interesting exercise in itself, well
explore the latter alternative.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

31.4 Recursive Suspensions

31.4

263

Recursive Suspensions

We seek to add a function to the Susp package with signature

val loopback : (a susp -> a susp) -> a susp
that, when applied to a function f mapping suspensions to suspensions,
yields a suspension s whose behavior is the same as f (s), the application
of f to the resulting suspension. In the above example the function in
question is
fun ones loop s = Susp.delay (fn () => Cons (1, s))
We use loopback to define ones as follows:
val ones = Susp.loopback ones loop
The idea is that ones should be equivalent to Susp.delay (fn () => Cons
(1, ones)), as in the original definition and which is the result of evaluating Susp.loopback ones loop, assuming Susp.loopback is implemented
properly.
How is loopback implemented? We use a technique known as backpatching. Heres the code
fun loopback f =
let
exception Circular
val r = ref (fn () => raise Circular)
val t = fn () => (!r)()
in
r := f t ; t
end
First we allocate a reference cell which is initialized to a placeholder
that, if forced, raises the exception Circular. Then we define a thunk that,
when forced, forces the contents of this reference cell. This will be the
return value of loopback. But before returning, we assign to the reference
cell the result of applying the given function to the result thunk. This ties
the knot to ensure that the output is looped back to the input. Observe
that if the loop function touches its input suspension before yielding an
output suspension, the exception Circular will be raised.
Here is the code for this chapter.
R EVISED 11.02.11

D RAFT

V ERSION 1.2

31.5 Sample Code

31.5

264

Sample Code

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 32
Data Abstraction
An abstract data type (ADT) is a type equipped with a set of operations for
manipulating values of that type. An ADT is implemented by providing
a representation type for the values of the ADT and an implementation for
the operations defined on values of the representation type. What makes
an ADT abstract is that the representation type is hidden from clients of the
ADT. Consequently, the only operations that may be performed on a value
of the ADT are the given ones. This ensures that the representation may be
changed without affecting the behavior of the client since the representation is hidden from it, the client cannot depend on it. This also facilitates
the implementation of efficient data structures by imposing a condition,
called a representation invariant, on the representation that is preserved by
the operations of the type. Each operation that takes a value of the ADT as
argument may assume that the representation invariant holds. In compensation each operation that yields a value of the ADT as result must guarantee that the representation invariant holds of it. If the operations of the
ADT preserve the representation invariant, then it must truly be invariant
no other code in the system could possibly disrupt it. Put another way,
any violation of the representation invariant may be localized to the implementation of one of the operations. This significantly reduces the time
required to find an error in a program.

32.1 Dictionaries

32.1

266

Dictionaries

To make these ideas concrete we will consider the abstract data type of
dictionaries. A dictionary is a mapping from keys to values. For simplicity we take keys to be strings, but it is possible to define a dictionary for
any ordered type; the values associated with keys are completely arbitrary.
Viewed as an ADT, a dictionary is a type a dict of dictionaries mapping
strings to values of type a together with empty, insert, and lookup operations that create a new dictionary, insert a value with a given key, and
retrieve the value associated with a key (if any). In short a dictionary is an
implementation of the following signature:
signature DICT =
sig
type key = string
type a entry = key * a
type a dict
exception Lookup of key
val empty : a dict
val insert : a dict * a entry -> a dict
val lookup : a dict * key -> a
end
Notice that the type a dict is not specified in the signature, whereas the
types key and a entry are defined to be string and string * a, respectively.

32.2

Binary Search Trees

A simple implementation of a dictionary is a binary search tree. A binary

search tree is a binary tree with values of an ordered type at the nodes
arranged in such a way that for every node in the tree, the value at that
node is greater than the value at any node in the left child of that node,
and smaller than the value at any node in the right child. It follows immediately that no two nodes in a binary search tree are labelled with the same
value. The binary search tree property is an example of a representation
invariant on an underlying data structure. The underlying structure is a

R EVISED 11.02.11

D RAFT

V ERSION 1.2

32.2 Binary Search Trees

267

binary tree with values at the nodes; the representation invariant isolates
a set of structures satisfying some additional, more stringent, conditions.
We may use a binary search tree to implement a dictionary as follows:
structure BinarySearchTree :> DICT =
struct
type key = string
type a entry = key * a
(* Rep invariant: a tree is a binary search tree *)
datatype a tree =
Empty |
Node of a tree * a entry * a tree
type a dict = a tree
exception Lookup of key
val empty = Empty
fun insert (Empty, entry) =
Node (Empty, entry, Empty)
| insert (n as Node (l, e as (k, ), r), e as (k, )) =
(case String.compare (k, k)
of LESS => Node (insert (l, e), e, r)
| GREATER => Node (l, e, insert (r, e))
| EQUAL => n)
fun lookup (Empty, k) = raise (Lookup k)
| lookup (Node (l, (k, v), r), k) =
(case String.compare (k, k)
of EQUAL => v
| LESS => lookup (l, k)
| GREATER => lookup (r, k))
end
Notice that empty is defined to be a valid binary search tree, that insert
yields a binary search tree if its argument is one, and that lookup relies
on its argument being a binary search tree (if not, it might fail to find a
key that in fact occurs in the tree!). The structure BinarySearchTree is
sealed with the signature DICT to ensure that the representation type is
held abstract.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

32.3 Balanced Binary Search Trees

32.3

268

Balanced Binary Search Trees

The difficulty with binary search trees is that they may become unbalanced. In particular if we insert keys in ascending order, the representation is essentially just a list! The left child of each node is empty; the
right child is the rest of the dictionary. Consequently, it takes O(n) time in
the worse case to perform a lookup on a dictionary containing n elements.
Such a tree is said to be unbalanced because the children of a node have
widely varying heights. Were it to be the case that the children of every
node had roughly equal height, then the lookup would take O(lg n) time,
a considerable improvement.
Can we do better? Many approaches have been suggested. One that we
will consider here is an instance of what is called a self-adjusting tree, called
a red-black tree (the reason for the name will be apparent shortly). The general idea of a self-adjusting tree is that operations on the tree may cause a
reorganization of its structure to ensure that some invariant is maintained.
In our case we will arrange things so that the tree is self-balancing, meaning that the children of any node have roughly the same height. As we just
remarked, this ensures that lookup is efficient.
How is this achieved? By imposing a clever representation invariant on
the binary search tree, called the red-black tree condition. A red-black tree
is a binary search tree in which every node is colored either red or black
(with the empty tree being regarded as black) and such that the following
properties hold:
1. The children of a red node are black.
2. For any node in the tree, the number of black nodes on any two paths
from that node to a leaf is the same. This number is called the black
height of the node.
These two conditions ensure that a red-black tree is a balanced binary
search tree. Heres why. First, observe that a red-black tree of black height
h has at least 2h 1 nodes. We may prove this by induction on the structure
of the red-black tree. The empty tree has black-height 1 (since we consider
it to be black), which is at least 21 1, as required. Suppose we have a red
node. The black height of both children must be h, hence each has at most
2h 1 nodes, yielding a total of 2 (2h 1) + 1 = 2h+1 1 nodes, which is
at least 2h 1. If, on the other hand, we have a black node, then the black
R EVISED 11.02.11

D RAFT

V ERSION 1.2

32.3 Balanced Binary Search Trees

269

height of both children is h 1, and each have at most 2h1 1 nodes, for
a total of 2 (2h1 1) + 1 = 2h 1 nodes. Now, observe that a red-black
tree of height h with n nodes has black height at least h/2, and hence has at
least 2h/2 1 nodes. Consequently, lg(n + 1) h/2, so h 2 lg(n + 1).
In other words, its height is logarithmic in the number of nodes, which
implies that the tree is height balanced.
To ensure logarithmic behavior, all we have to do is to maintain the redblack invariant. The empty tree is a red-black tree, so the only question is
how to perform an insert operation. First, we insert the entry as usual for
a binary search tree, with the fresh node starting out colored red. In doing
so we do not disturb the black height condition, but we might introduce a
red-red violation, a situation in which a red node has a red child. We then
remove the red-red violation by propagating it upwards towards the root
by a constant-time transformation on the tree (one of several possibilities,
which well discuss shortly). These transformations either eliminate the
red-red violation outright, or, in logarithmic time, push the violation to
the root where it is neatly resolved by recoloring the root black (which
preserves the black-height invariant!).
The violation is propagated upwards by one of four rotations. We will
maintain the invariant that there is at most one red-red violation in the
tree. The insertion may or may not create such a violation, and each propagation step will preserve this invariant. It follows that the parent of a
red-red violation must be black. Consequently, the situation must look
like this. This diagram represents four distinct situations, according to
whether the uppermost red node is a left or right child of the black node,
and whether the red child of the red node is itself a left or right child. In
each case the red-red violation is propagated upwards by transforming it
to look like this. Notice that by making the uppermost node red we may be
introducing a red-red violation further up the tree (since the black nodes
parent might have been red), and that we are preserving the black-height
invariant since the great-grand-children of the black node in the original
situation will appear as children of the two black nodes in the re-organized
situation. Notice as well that the binary search tree conditions are also preserved by this transformation. As a limiting case if the red-red violation is
propagated to the root of the entire tree, we re-color the root black, which
preserves the black-height condition, and we are done re-balancing the
tree.
Lets look in detail at two of the four cases of removing a red-red vioR EVISED 11.02.11

D RAFT

V ERSION 1.2

32.3 Balanced Binary Search Trees

270

lation, those in which the uppermost red node is the left child of the black
node; the other two cases are handled symmetrically. If the situation looks
like this, we reorganize the tree to look like this. You should check that the
black-height and binary search tree invariants are preserved by this transformation. Similarly, if the situation looks like this, then we reorganize the
tree to look like this (precisely as before). Once again, the black-height and
binary search tree invariants are preserved by this transformation, and the
red-red violation is pushed further up the tree.
Here is the ML code to implement dictionaries using a red-black tree.
Notice that the tree rotations are neatly expressed using pattern matching.
structure RedBlackTree :> DICT =
struct
type key = string
type a entry = string * a
(* Inv: binary search tree + red-black conditions *)
datatype a dict =
Empty |
Red of a entry * a dict * a dict |
Black of a entry * a dict * a dict
val empty = Empty
exception Lookup of key
fun lookup (dict, key) =
let
fun lk (Empty) = raise (Lookup key)
| lk (Red tree) = lk tree
| lk (Black tree) = lk tree
and lk ((key1, datum1), left, right) =
(case String.compare(key,key1)
of EQUAL => datum1
| LESS => lk left
| GREATER => lk right)
in
lk dict
end
fun restoreLeft
(Black (z, Red (y, Red (x, d1, d2), d3), d4)) =
Red (y, Black (x, d1, d2), Black (z, d3, d4))
| restoreLeft

R EVISED 11.02.11

D RAFT

V ERSION 1.2

32.3 Balanced Binary Search Trees

271

(Black (z, Red (x, d1, Red (y, d2, d3)), d4)) =
Red (y, Black (x, d1, d2), Black (z, d3, d4))
| restoreLeft dict = dict
fun restoreRight
(Black (x, d1, Red (y, d2, Red
Red (y, Black (x, d1, d2), Black
| restoreRight
(Black (x, d1, Red (z, Red (y,
Red (y, Black (x, d1, d2), Black
| restoreRight dict = dict

(z, d3, d4)))) =

(z, d3, d4))
d2, d3), d4))) =
(z, d3, d4))

fun insert (dict, entry as (key, datum)) =

let
(* val ins : a dict->a dict insert entry *)
(* ins (Red ) may have red-red at root *)
(* ins (Black ) or ins (Empty) is red/black *)
(* ins preserves black height *)
fun ins (Empty) = Red (entry, Empty, Empty)
| ins (Red (entry1 as (key1, datum1), left, right)) =
(case String.compare (key, key1)
of EQUAL => Red (entry, left, right)
| LESS => Red (entry1, ins left, right)
| GREATER => Red (entry1, left, ins right))
| ins (Black (entry1 as (key1, datum1), left, right)) =
(case String.compare (key, key1)
of EQUAL => Black (entry, left, right)
| LESS => restoreLeft (Black (entry1, ins left, right))
| GREATER => restoreRight (Black (entry1, left, ins right)))
in
case ins dict
of Red (t as ( , Red , )) => Black t (* re-color *)
| Red (t as ( , , Red )) => Black t (* re-color *)
| dict => dict
end
end

It is worthwhile to contemplate the role played by the red-black invariant

in ensuring the correctness of the implementation and the time complexity
of the operations.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

32.4 Abstraction vs. Run-Time Checking

32.4

272

Abstraction vs. Run-Time Checking

You might wonder whether we could equally well use run-time checks to
enforce representation invariants. The idea would be to introduce a debug flag that, when set, causes the operations of the dictionary to check
that the representation invariant holds of their arguments and results. In
the case of a binary search tree this is surely possible, but at considerable
expense since the time required to check the binary search tree invariant is
proportional to the size of the binary search tree itself, whereas an insert
(for example) can be performed in logarithmic time. But wouldnt we turn
off the debug flag before shipping the production copy of the code? Yes,
indeed, but then the benefits of checking are lost for the code we care about
most! By using the type system to enforce abstraction, we can confine the
possible violations of the representation invariant to the dictionary package itself, and, moreover, we need not turn off the check for production
code because there is no run-time penalty for doing so.
A more subtle point is that it may not always be possible to enforce data
abstraction at run-time. Efficiency considerations aside, you might think
that we can always replace static localization of representation errors by
dynamic checks for violations of them. But this is false! One reason is that
the representation invariant might not be computable. As an example,
consider an abstract type of total functions on the integers, those that are
guaranteed to terminate when called, without performing any I/O or having any other computational effect. Another is that no run-time check can
be defined that ensures that a given integer-valued function is total. Yet
we can define an abstract type of total functions that, while not admitting
every possible total function on the integers as values, provides a useful
set of such functions as elements of a structure. By using these specified
operations to create a total function, we are in effect encoding a proof of
totality in the program itself.
Heres a sketch of such a package:
signature TIF = sig
type tif
val apply : tif -> (int -> int)
val id : tif
val compose : tif * tif -> tif
val double : tif
R EVISED 11.02.11

D RAFT

V ERSION 1.2

32.5 Sample Code

273

.
.
.
end
structure Tif :> TIF = struct
type tif = int->int
fun apply t n = t n
fun id x = x
fun compose (f, g) = f o g
fun double x = 2 * x
.
.
.
end
Should the application of such some value of type Tif.tif fail to terminate, we know where to look for the error. No run-time check can assure
us that an arbitrary integer function is in fact total.
Another reason why a run-time check to enforce data abstraction is impossible is that it may not be possible to tell from looking at a given value
whether or not it is a legitimate value of the abstact type. Heres an example. In many operating systems processes are named by integer-value
process identifiers. Using the process identifier we may send messages to
the process, cause it to terminate, or perform any number of other operations on it. The thing to notice here is that any integer at all is a possible
process identifier; we cannot tell by looking at the integer whether it is
indeed valid. No run-time check on the value will reveal whether a given
integer is a real or bogus process identifier. The only way to know is
to consider the history of how that integer came into being, and what
operations were performed on it. Using the abstraction mechanisms just
described, we can enforce the requirement that a value of type pid, whose
underlying representation is int, is indeed a process identifier. You are
invited to imagine how this might be achieved in ML.

32.5

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Chapter 33
Representation Independence and
ADT Correctness
This chapter is concerned with proving correctness of ADT implementations by exhibiting a simulation relation between a reference implementation (taken, or known, to be correct) and a candidate implementation
(whose correctness is to be established). The methodology generalizes
Hoares notion of abstraction functions to an arbitrary relation, and relies
on Reynolds notion of parametricity to conclude that related implementations engender the same observable behavior in all clients.

33.1

Sample Code

Here is the code for this chapter.

Chapter 34
Modularity and Reuse
1. Naming conventions.
2. Exploiting structural subtyping (type t convention).
3. Impedance-matching functors.

34.1

Sample Code

Here is the code for this chapter.

Chapter 35
Dynamic Typing and Dynamic
Dispatch
This chapter is concerned with dynamic typing in a statically typed language. It is commonly thought that there is an opposition between
statically-typed languages (such as Standard ML) and dynamically-typed
languages (such as Scheme). In fact, dynamically typed languages are a
special case of statically-typed languages! We will demonstrate this by
exhibiting a faithful representation of Scheme inside of ML.

35.1

Sample Code

Here is the code for this chapter.

Chapter 36
Concurrency
In this chapter we consider some fundamental techniques for concurrent
programming using CML.

36.1

Sample Code

Here is the code for this chapter.

Part V
Appendices

Appendix A
The Standard ML Basis Library
The Standard ML Basis Library is a collection of modules providing a basic
collection of abstract types that are shared by all implementations of Standard ML. All of the primitive types of Standard ML are defined in structures in the Standard Basis. It also defines a variety of other commonlyused abstract types.
Most implementations of Standard ML include module libraries implementing a wide variety of services. These libraries are usually not portable
across implementations, particularly not those that are concerned with the
internals of the compiler or its interaction with the host computer system.
Please refer to the documentation of your compiler for information on its
libraries.

Appendix B
Compilation Management
All program development environments provide tools to support building
systems out of collections of separately-developed modules. These tools
usually provide services such as:
1. Source code management such as version and revision control.
2. Separate compilation and linking to support simultaneous development
and to reduce build times.
3. Libraries of re-usable modules with consistent conventions for identifying modules and their components.
4. Release management for building and disseminating systems for general use.
Different languages, and different vendors, support these activities in different ways. Some rely on generic tools, such as the familiar Unix tools,
others provide proprietary tools, commonly known as IDEs (integrated
development environments).
Most implementations of Standard ML rely on a combination of generic
program development tools and tools specific to that implementation of
the language. Rather than attempt to summarize all of the known implementations, we will instead consider the SML/NJ Compilation Manager
(CM) as a representative program development framework for ML. Other
compilers provide similar tools; please consult your compilers documentation for details of how to use them.

B.1 Overview of CM

281

B.1

Overview of CM

B.2

Building Systems with CM

B.3

Sample Code

Here is the code for this chapter.

R EVISED 11.02.11

D RAFT

V ERSION 1.2

Appendix C
Sample Programs
A number of example programs illustrating the concepts discussed in the
preceding chapters are available in the Sample Code directory on the worldwide web.

Revision History
Revision

Date

Author(s)

Description

1.0
1.1
1.2

21.01.11
10.02.11
11.02.11

RH
RH
RH

Created
Expanded HOFs techniques
Added matching combinators to
higher-order function techniques

Bibliography
[1] Emden R. Gansner and John H. Reppy, editors. The Standard ML Basis
Library. Cambridge University Press, 2000.
[2] Peter Lee. Standard ML at Carnegie Mellon. Available within CMU at
http://www.cs.cmu.edu/afs/cs/local/sml/common/smlguide.
[3] Robin Milner, Mads Tofte, Robert Harper, and David MacQueen. The
Definition of Standard ML (Revised). MIT Press, 1997.