Contents

About Language

1 The
1.1
1.2
1.3

1.4
1.5

Nature of Language
Communication . . . . . . . . . . . . . . . . . . .
Syntax and Semantics . . . . . . . . . . . . . . .
Natural Languages and Programming Languages
1.3.1 Structure . . . . . . . . . . . . . . . . . .
1.3.2 Redundancy . . . . . . . . . . . . . . . . .
1.3.3 Using Partial Information: Ambiguity and
1.3.4 Implicit Communication . . . . . . . . . .
1.3.5 Flexibility and Nuance . . . . . . . . . . .
1.3.6 Ability to Change and Evolve . . . . . . .
The Standardization Process . . . . . . . . . . .
1.4.1 Language Growth and Divergence . . . .
Nonstandard Compilers . . . . . . . . . . . . . .

2 Representation and Abstraction

2.1 What Is a Program? . . . . . . . . . . . . . . . .
2.2 Representation . . . . . . . . . . . . . . . . . . .
2.2.1 Semantic Intent . . . . . . . . . . . . . . .
2.2.2 Explicit versus Implicit Representation . .
2.2.3 Coherent versus Diffuse Representation .
2.3 Language Design . . . . . . . . . . . . . . . . . .
2.3.1 Competing Design Goals . . . . . . . . . .
2.3.2 The Power of Restrictions . . . . . . . . .
2.3.3 Principles for Evaluating a Design . . . .
2.4 Classifying Languages . . . . . . . . . . . . . . .
2.4.1 Language Families . . . . . . . . . . . . .
2.4.2 Languages Are More Alike than Different

iii

. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
Abstraction
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

3
4
5
6
6
7
8
9
10
10
11
12
12

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

17
17
20
21
22
22
25
25
27
30
41
41
49

iv

CONTENTS

3 Elements of Language
3.1 The Parts of Speech . . . . . . . . . . . . . . . . . .
3.1.1 Nouns . . . . . . . . . . . . . . . . . . . . . .
3.1.2 Pronouns: Pointers . . . . . . . . . . . . . . .
3.1.3 Adjectives: Data Types . . . . . . . . . . . .
3.1.4 Verbs . . . . . . . . . . . . . . . . . . . . . .
3.1.5 Prepositions and Conjunctions . . . . . . . .
3.2 The Metalanguage . . . . . . . . . . . . . . . . . . .
3.2.1 Words: Lexical Tokens . . . . . . . . . . . . .
3.2.2 Sentences: Statements . . . . . . . . . . . . .
3.2.3 Larger Program Units: Scope . . . . . . . .
3.2.4 Comments . . . . . . . . . . . . . . . . . . . .
3.2.5 Naming Parts of a Program . . . . . . . . .
3.2.6 Metawords That Let the Programmer Extend
4 Formal Description of Language
4.1 Foundations of Programming Languages . . . . . . .
4.2 Syntax . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.1 Extended BNF . . . . . . . . . . . . . . . . .
4.2.2 Syntax Diagrams . . . . . . . . . . . . . . . .
4.3 Semantics . . . . . . . . . . . . . . . . . . . . . . . .
4.3.1 The Meaning of a Program . . . . . . . . . .
4.3.2 Definition of Language Semantics . . . . . . .
4.3.3 The Abstract Machine . . . . . . . . . . . . .
4.3.4 Lambda Calculus: A Minimal Semantic Basis
4.4 Extending the Semantics of a Language . . . . . . .
4.4.1 Semantic Extension in FORTH . . . . . . .

II

. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
the Language
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

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

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

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

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

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

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

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

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

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

Describing Computation

5 Primitive Types
5.1 Primitive Hardware Types . . . . . . . . . . . . .
5.1.1 Bytes, Words, and Long Words . . . . . .
5.1.2 Character Codes . . . . . . . . . . . . . .
5.1.3 Numbers . . . . . . . . . . . . . . . . . .
5.2 Types in Programming Languages . . . . . . . .
5.2.1 Type Is an Abstraction . . . . . . . . . .
5.2.2 A Type Provides a Physical Description .
5.2.3 What Primitive Types Should a Language
5.2.4 Emulation . . . . . . . . . . . . . . . . . .

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

51
51
51
52
53
55
58
59
59
62
64
67
70
70

.
.
.
.
.
.
.
.
.
.
.

77
78
78
82
87
90
90
90
92
96
107
110

115
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
Support?
. . . . . .

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

117
. 118
. 118
. 118
. 120
. 126
. 126
. 127
. 130
. 133

CONTENTS
5.3

A Brief History of Type Declarations . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

5.3.1 Origins of Type Ideas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
5.3.2 Type Becomes a Definable Abstraction . . . . . . . . . . . . . . . . . . . . . . 137

6 Modeling Objects
6.1 Kinds of Objects . . . . . . . . . . . . . . . . . . . . . .
6.2 Placing a Value in a Storage Object . . . . . . . . . . .
6.2.1 Static Initialization . . . . . . . . . . . . . . . . .
6.2.2 Dynamically Changing the Contents of a Storage
6.2.3 Dereferencing . . . . . . . . . . . . . . . . . . . .
6.2.4 Pointer Assignment . . . . . . . . . . . . . . . .
6.2.5 The Semantics of Pointer Assignment . . . . . .
6.3 The Storage Model: Managing Storage Objects . . . . .
6.3.1 The Birth and Death of Storage Objects . . . . .
6.3.2 Dangling References . . . . . . . . . . . . . . . .

. . . .
. . . .
. . . .
Object
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

143
144
146
146
148
152
154
156
158
158
169

7 Names and Binding

7.1 The Problem with Names . . . . . . . . . . . . . . . . . . . . .
7.1.1 The Role of Names . . . . . . . . . . . . . . . . . . . . .
7.1.2 Definition Mechanisms: Declarations and Defaults . . .
7.1.3 Binding . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.1.4 Names and Objects: Not a One-to-One Correspondence
7.2 Binding a Name to a Constant . . . . . . . . . . . . . . . . . .
7.2.1 Implementations of Constants . . . . . . . . . . . . . . .
7.2.2 How Constant Is a Constant? . . . . . . . . . . . . . . .
7.3 Survey of Allocation and Binding . . . . . . . . . . . . . . . . .
7.4 The Scope of a Name . . . . . . . . . . . . . . . . . . . . . . . .
7.4.1 Naming Conflicts . . . . . . . . . . . . . . . . . . . . . .
7.4.2 Block Structure . . . . . . . . . . . . . . . . . . . . . . .
7.4.3 Recursive Bindings . . . . . . . . . . . . . . . . . . . . .
7.4.4 Visibility versus Lifetime. . . . . . . . . . . . . . . . . .
7.5 Implications for the Compiler / Interpreter . . . . . . . . . . .

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

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

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

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

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

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

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

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

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

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

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

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

175
175
176
178
180
185
186
190
191
191
193
193
195
198
200
205

8 Expressions and Evaluation

8.1 The Programming Environment . . . .
8.2 Sequence Control and Communication
8.2.1 Nesting . . . . . . . . . . . . .
8.2.2 Sequences of Statements . . . .
8.2.3 Interprocess Sequence Control .
8.3 Expression Syntax . . . . . . . . . . .
8.3.1 Functional Expression Syntax .

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

211
212
213
214
215
216
216
217

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

vi

CONTENTS

8.4

8.3.2 Operator Expressions . . . . . . .

8.3.3 Combinations of Parsing Rules . .
Function Evaluation . . . . . . . . . . . .
8.4.1 Order of Evaluation . . . . . . . .
8.4.2 Lazy or Strict Evaluation . . . . .
8.4.3 Order of Evaluation of Arguments

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

9 Functions and Parameters

9.1 Function Syntax . . . . . . . . . . . . . . . . . . .
9.1.1 Fixed versus Variable Argument Functions
9.1.2 Parameter Correspondence . . . . . . . . .
9.1.3 Indefinite-Length Parameter Lists . . . . .
9.2 What Does an Argument Mean? . . . . . . . . . .
9.2.1 Call-by-Value . . . . . . . . . . . . . . . . .
9.2.2 Call-by-Name . . . . . . . . . . . . . . . . .
9.2.3 Call-by-Reference . . . . . . . . . . . . . . .
9.2.4 Call-by-Return . . . . . . . . . . . . . . . .
9.2.5 Call-by-Value-and-Return . . . . . . . . . .
9.2.6 Call-by-Pointer . . . . . . . . . . . . . . . .
9.3 Higher-Order Functions . . . . . . . . . . . . . . .
9.3.1 Functional Arguments . . . . . . . . . . . .
9.3.2 Currying . . . . . . . . . . . . . . . . . . .
9.3.3 Returning Functions from Functions . . . .

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

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

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

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

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

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

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

10 Control Structures
10.1 Basic Control Structures . . . . . . . . . . . . . . . . . . . . . .
10.1.1 Normal Instruction Sequencing . . . . . . . . . . . . . .
10.1.2 Assemblers . . . . . . . . . . . . . . . . . . . . . . . . .
10.1.3 Sequence, Subroutine Call, IF, and WHILE Suffice . . .
10.1.4 Subroutine Call . . . . . . . . . . . . . . . . . . . . . . .
10.1.5 Jump and Conditional Jump . . . . . . . . . . . . . . .
10.1.6 Control Diagrams . . . . . . . . . . . . . . . . . . . . .
10.2 Conditional Control Structures . . . . . . . . . . . . . . . . . .
10.2.1 Conditional Expressions versus Conditional Statements
10.2.2 Conditional Branches: Simple Spaghetti . . . . . . . . .
10.2.3 Structured Conditionals . . . . . . . . . . . . . . . . . .
10.2.4 The Case Statement . . . . . . . . . . . . . . . . . . . .
10.3 Iteration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.3.1 The Infinite Loop . . . . . . . . . . . . . . . . . . . . . .
10.3.2 Conditional Loops . . . . . . . . . . . . . . . . . . . . .
10.3.3 The General Loop . . . . . . . . . . . . . . . . . . . . .

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

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

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

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

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

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

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

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

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

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

.
.
.
.
.
.

.
.
.
.
.
.

218
222
224
224
227
230

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

233
. 234
. 234
. 235
. 237
. 239
. 240
. 242
. 244
. 247
. 248
. 249
. 254
. 255
. 257
. 259

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

267
. 268
. 269
. 270
. 270
. 272
. 273
. 274
. 275
. 275
. 277
. 278
. 284
. 289
. 290
. 290
. 292

vii

CONTENTS
10.3.4 Counted Loops . . . . . . . . .
10.3.5 The Iteration Element . . . . .
10.4 Implicit Iteration . . . . . . . . . . . .
10.4.1 Iteration on Coherent Objects .
10.4.2 Backtracking . . . . . . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

293
298
301
301
303

11 Global Control
11.1 The GOTO Problem . . . . . . . . . . .
11.1.1 Faults Inherent in GOTO . . . . .
11.1.2 To GOTO or Not to GOTO . . .
11.1.3 Statement Labels . . . . . . . . .
11.2 Breaking Out . . . . . . . . . . . . . . .
11.2.1 Generalizing the BREAK . . . . .
11.3 Continuations . . . . . . . . . . . . . . .
11.4 Exception Processing . . . . . . . . . . .
11.4.1 What Is an Exception? . . . . .
11.4.2 The Steps in Exception Handling
11.4.3 Exception Handling in Ada . . .

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

309
310
310
313
315
317
320
321
327
327
328
331

III

Application Modeling

335

12 Functional Languages
12.1 Denotation versus Computation . . .
12.1.1 Denotation . . . . . . . . . .
12.2 The Functional Approach . . . . . .
12.2.1 Eliminating Assignment . . .
12.2.2 Recursion Can Replace WHILE
12.2.3 Sequences . . . . . . . . . . .
12.3 Miranda: A Functional Language . .
12.3.1 Data Structures . . . . . . . .
12.3.2 Operations and Expressions .
12.3.3 Function Definitions . . . . .
12.3.4 List Comprehensions . . . . .
12.3.5 Infinite Lists . . . . . . . . .
13 Logic Programming
13.1 Predicate Calculus
13.1.1 Formulas .
13.2 Proof Systems . . .
13.3 Models . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

337
338
339
341
342
344
347
350
351
351
352
355
358

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

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

.
.
.
.

361
. 362
. 362
. 365
. 367

viii

CONTENTS
13.4 Automatic Theorem Proving . . . . .
13.4.1 Resolution Theorem Provers .
13.5 Prolog . . . . . . . . . . . . . . . . . .
13.5.1 The Prolog Environment . . . .
13.5.2 Data Objects and Terms . . . .
13.5.3 Horn Clauses in Prolog . . . . .
13.5.4 The Prolog Deduction Process .
13.5.5 Functions and Computation . .
13.5.6 Cuts and the not Predicate .
13.5.7 Evaluation of Prolog . . . . . .

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

368
370
375
375
375
376
379
380
385
389

14 The Representation of Types

14.1 Programmer-Defined Types . . . . . . . . . . . . . . .
14.1.1 Representing Types within a Translator . . . .
14.1.2 Finite Types . . . . . . . . . . . . . . . . . . .
14.1.3 Constrained Types . . . . . . . . . . . . . . . .
14.1.4 Pointer Types . . . . . . . . . . . . . . . . . . .
14.2 Compound Types . . . . . . . . . . . . . . . . . . . . .
14.2.1 Arrays . . . . . . . . . . . . . . . . . . . . . . .
14.2.2 Strings . . . . . . . . . . . . . . . . . . . . . . .
14.2.3 Sets . . . . . . . . . . . . . . . . . . . . . . . .
14.2.4 Records . . . . . . . . . . . . . . . . . . . . . .
14.2.5 Union Types . . . . . . . . . . . . . . . . . . .
14.3 Operations on Compound Objects . . . . . . . . . . .
14.3.1 Creating Program Objects: Value Constructors
14.3.2 The Interaction of Dereferencing, Constructors,
14.4 Operations on Types . . . . . . . . . . . . . . . . . . .

. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
and Selectors
. . . . . . . .

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

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

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

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

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

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

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

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

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

393
394
394
396
397
398
400
400
406
408
412
419
422
422
423
430

15 The Semantics of Types

15.1 Semantic Description . . . . . . . . . . . . . . . . . .
15.1.1 Domains in Early Languages . . . . . . . . .
15.1.2 Domains in Typeless Languages . . . . . .
15.1.3 Domains in the 1970s . . . . . . . . . . . . .
15.1.4 Domains in the 1980s . . . . . . . . . . . . .
15.2 Type Checking . . . . . . . . . . . . . . . . . . . . .
15.2.1 Strong Typing . . . . . . . . . . . . . . . . .
15.2.2 Strong Typing and Data Abstraction . . . . .
15.3 Domain Identity: Different Domain/ Same Domain?
15.3.1 Internal and External Domains . . . . . . . .
15.3.2 Internally Merged Domains . . . . . . . . . .
15.3.3 Domain Mapping . . . . . . . . . . . . . . . .

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

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

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

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

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

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

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

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

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

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

435
436
436
437
441
444
444
445
446
448
448
449
450

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

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

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

.
.
.
.
.
.
.
.
.
.

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

.
.
.
.
.
.
.
.
.
.

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

.
.
.
.
.
.
.
.
.
.

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

.
.
.
.
.
.
.
.
.
.

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

.
.
.
.
.
.
.
.
.
.

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

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

ix

CONTENTS
15.4 Programmer-Defined Domains . . . . . . . . .
15.4.1 Type Description versus Type Name .
15.4.2 Type Constructors . . . . . . . . . . .
15.4.3 Types Defined by Mapping . . . . . .
15.5 Type Casts, Conversions, and Coercions . . .
15.5.1 Type Casts. . . . . . . . . . . . . . . .
15.5.2 Type Conversions . . . . . . . . . . .
15.5.3 Type Coercion . . . . . . . . . . . . .
15.6 Conversions and Casts in Common Languages
15.6.1 COBOL . . . . . . . . . . . . . . . . .
15.6.2 FORTRAN . . . . . . . . . . . . . . . .
15.6.3 Pascal . . . . . . . . . . . . . . . . . .
15.6.4 PL/1 . . . . . . . . . . . . . . . . . . .
15.6.5 C . . . . . . . . . . . . . . . . . . . . .
15.6.6 Ada Types and Treatment of Coercion
15.7 Evading the Type Matching Rules . . . . . .

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

452
452
453
454
459
460
465
466
470
470
470
471
472
472
475
479

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

489
490
492
497
500
501
505
506

17 Generics
17.1 Generics . . . . . . . . . . . . . . . . . . . . . . . . . .
17.1.1 What Is a Generic? . . . . . . . . . . . . . . . .
17.1.2 Implementations of Generics . . . . . . . . . .
17.1.3 Generics, Virtual Functions, and ADTs . . . .
17.1.4 Generic Functions . . . . . . . . . . . . . . . .
17.2 Limited Generic Behavior . . . . . . . . . . . . . . . .
17.2.1 Union Data Types . . . . . . . . . . . . . . . .
17.2.2 Overloaded Names . . . . . . . . . . . . . . . .
17.2.3 Fixed Set of Generic Definitions, with Coercion
17.2.4 Extending Predefined Operators . . . . . . . .
17.2.5 Flexible Arrays . . . . . . . . . . . . . . . . . .
17.3 Parameterized Generic Domains . . . . . . . . . . . .
17.3.1 Domains with Type Parameters . . . . . . . . .
17.3.2 Preprocessor Generics in C . . . . . . . . . . .

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

511
512
512
513
515
516
521
521
521
523
524
525
527
530
531

16 Modules and Object Classes

16.1 The Purpose of Modules . . . . . . . .
16.2 Modularity Through Files and Linking
16.3 Packages in Ada . . . . . . . . . . . . .
16.4 Object Classes. . . . . . . . . . . . . .
16.4.1 Classes in C++ . . . . . . . . .
16.4.2 Represented Domains . . . . .
16.4.3 Friends of Classes . . . . . . .

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

x
18 Dispatching with Inheritance
18.1 Representing Domain Relationships . . . . . . . . . . . .
18.1.1 The Mode Graph and the Dispatcher . . . . . . .
18.2 Subdomains and Class Hierarchies. . . . . . . . . . . . .
18.2.1 Subrange Types . . . . . . . . . . . . . . . . . .
18.2.2 Class Hierarchies . . . . . . . . . . . . . . . . . .
18.2.3 Virtual Functions in C++. . . . . . . . . . . . .
18.2.4 Function Inheritance . . . . . . . . . . . . . . . .
18.2.5 Programmer-Defined Conversions in C++ . . . .
18.3 Polymorphic Domains and Functions . . . . . . . . . . .
18.3.1 Polymorphic Functions . . . . . . . . . . . . . . .
18.3.2 Manual Domain Representation and Dispatching
18.3.3 Automating Ad Hoc Polymorphism . . . . . . .
18.3.4 Parameterized Domains . . . . . . . . . . . . . .
18.4 Can We Do More with Generics? . . . . . . . . . . . . .
18.4.1 Dispatching Using the Mode Graph . . . . . . .
18.4.2 Generics Create Some Hard Problems . . . . . .

CONTENTS

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

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

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

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

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

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

A Exhibits Listed by Topic

A.1 Languages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.1.1 Ada . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.1.2 APL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.1.3 C++ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.1.4 C and ANSI C . . . . . . . . . . . . . . . . . . . . . . . . . .
A.1.5 FORTH . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.1.6 FORTRAN . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.1.7 LISP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.1.8 Miranda . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.1.9 Pascal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.1.10 Prolog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.1.11 Scheme and T . . . . . . . . . . . . . . . . . . . . . . . . . .
A.1.12 Other Languages . . . . . . . . . . . . . . . . . . . . . . . .
A.2 Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.2.1 Application Modeling, Generics, and Polymorphic Domains
A.2.2 Control Structures . . . . . . . . . . . . . . . . . . . . . . .
A.2.3 Data Representation . . . . . . . . . . . . . . . . . . . . . .
A.2.4 History . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.2.5 Lambda Calculus . . . . . . . . . . . . . . . . . . . . . . . .
A.2.6 Language Design and Specification . . . . . . . . . . . . . .
A.2.7 Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.2.8 Translation, Interpretation, and Function Calls . . . . . . .

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

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

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

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

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

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

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

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

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

541
. 542
. 542
. 549
. 549
. 550
. 554
. 556
. 558
. 561
. 561
. 562
. 563
. 567
. 568
. 571
. 575

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

585
. 585
. 585
. 586
. 587
. 588
. 590
. 591
. 591
. 592
. 593
. 596
. 597
. 597
. 598
. 598
. 599
. 600
. 600
. 600
. 601
. 601
. 602

CONTENTS

xi

A.2.9 Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 602

xii

CONTENTS

Preface
This text is intended for a course in advanced programming languages or the structure of programming language and should be appropriate for students at the junior, senior, or masters level.
It should help the student understand the principles that underlie all languages and all language
implementations.
This is a comprehensive text which attempts to dissect language and explain how a language
is really built. The first eleven chapters cover the core material: language specification, objects,
expressions, control, and types. The more concrete aspects of each topic are presented first, followed by a discussion of implementation strategies and the related semantic issues. Later chapters
cover current topics, including modules, object-oriented programming, functional languages, and
concurrency constructs.
The emphasis throughout the text is on semantics and abstraction; the syntax and historical
development of languages are discussed in light of the underlying semantical concepts. Fundamental
principles of computation, communication, and good design are stated and are used to evaluate
various language constructs and to demonstrate that language designs are improving as these
principles become widely understood.
Examples are cited from many languages including Pascal, C, C++, FORTH, BASIC, LISP,
FORTRAN, Ada, COBOL, APL, Prolog, Turing, Miranda, and Haskell. All examples are annotated so
that a student who is unfamiliar with the language used can understand the meaning of the code
and see how it illustrates the principle.
It is the belief of the authors that the student who has a good grasp of the structure of computer
languages will have the tools to master new languages easily.
The specific goals of this book are to help students learn:
To reason clearly about programming languages.
To develop principles of communication so that we can evaluate the wisdom and utility of the
decisions made in the process of language design.
To break down language into its major components, and each component into small pieces so
that we can focus on competing alternatives.
To define a consistent and general set of terms for the components out of which programming
languages are built, and the concepts on which they are based.
xiii

xiv

CONTENTS
To use these terms to describe existing languages, and in so doing clarify the conflicting
terminology used by the language designers, and untangle the complexities inherent in so
many languages.
To see below the surface appearance of a language to its actual structure and descriptive
power.
To understand that many language features that commonly occur together are, in fact, independent and separable. To appreciate the advantages and disadvantages of each feature. To
suggest ways in which these basic building blocks can be recombined in new languages with
more desirable properties and fewer faults.
To see the similarities and differences that exist among languages students already know, and
to learn new ones.
To use the understanding so gained to suggest future trends in language design.

Acknowledgement
The authors are indebted to several people for their help and support during the years we have
worked on this project. First, we wish to thank our families for their uncomplaining patience and
understanding.
We thank Michael J. Fischer for his help in developing the sections on lambda calculus, functional languages and logic. and for working out several sophisticated code examples. In addition,
his assistance as software and hardware systems expert and TeX guru made this work possible.
Several reviewers read this work in detail and offered invaluable suggestions and corrections.
We thank these people for their help. Special thanks go to Robert Fischer and Roland Lieger for
reading beyond the call of duty and to Gary Walters for his advice and for the material he has
contributed.
Finally, we thank our students at the University of New Haven and at Sacred Heart University
for their feedback on the many versions of this book.
Parts of this manuscript were developed under a grant from Sacred Heart University.

Part I

About Language

Chapter 1

The Nature of Language

Overview
This chapter introduces the concept of the nature of language. The purpose of language
is communication. A set of symbols, understood by both sender and receiver, is combined according to a set of rules, its grammar or syntax. The semantics of the language
defines how each grammatically correct sentence is to be interpreted. Using English as
a model, language structures are studied and compared. The issue of standardization
of programming languages is examined. Nonstandard compilers are examples of the use
of deviations from an accepted standard.

This is a book about the structure of programming languages. (For simplicity, we shall use the
term language to mean programming language.) We will try to look beneath the individual
quirks of familiar languages and examine the essential properties of language itself. Several aspects
of language will be considered, including vocabulary, syntax rules, meaning (semantics), implementation problems, and extensibility. We will consider several programming languages, examining
the choices made by language designers that resulted in the strengths, weaknesses, and particular
character of each language. When possible, we will draw parallels between programming languages
and natural languages.
Different languages are like tools in a toolbox: although each language is capable of expressing
most algorithms, some are obviously more appropriate for certain applications than others. (You
can use a chisel to turn a screw, but it is not a good idea.) For example, it is commonly understood
that COBOL is good for business applications. This is true because COBOL provides a large
variety of symbols for controlling input and output formats, so that business reports may easily be
3

CHAPTER 1. THE NATURE OF LANGUAGE

made to fit printed forms. LISP is good for artificial intelligence applications because it supports
dynamically growing and shrinking data. We will consider how well each language models the
objects, actions, and relationships inherent in various classes of applications.
Rather than accept languages as whole packages, we will be asking:
What design decisions make each language different from the others?
Are the differences a result of minor syntactic rules, or is there an important underlying
semantic issue?
Is a controversial design decision necessary to make the language appropriate for its intended
use, or was the decision an accident of history?
Could different design decisions result in a language with more strengths and fewer weaknesses?
Are the good parts of different languages mutually exclusive, or could they be effectively
combined?
Can a language be extended to compensate for its weaknesses?

1.1

Communication

A natural language is a symbolic communication system that is commonly understood among a

group of people. Each language has a set of symbols that stand for objects, properties, actions,
abstractions, relations, and the like. A language must also have rules for combining these symbols.
A speaker can communicate an idea to a listener if and only if they have a common understanding
of enough symbols and rules. Communication is impaired when speaker and listener interpret a
symbol differently. In this case, either speaker and/or listener must use feedback to modify his or
her understanding of the symbols until commonality is actually achieved. This happens when we
learn a new word or a new meaning for an old word, or correct an error in our idea of the meaning
of a word.
English is for communication among people. Programs are written for both computers and
people to understand. Using a programming language requires a mutual understanding between a
person and a machine. This can be more difficult to achieve than understanding between people
because machines are so much more literal than human beings.
The meaning of symbols in natural language is usually defined by custom and learned by
experience and feedback. In contrast, programming languages are generally defined by an authority,
either an individual language designer or a committee. For a computer to understand a human
language, we must devise a method for translating both the syntax and semantics of the language
into machine code. Language designers build languages that they know how to translate, or that
they believe they can figure out how to translate.

1.2. SYNTAX AND SEMANTICS

On the other hand, if computers were the only audience for our programs we might be writing
code in a language that was trivially easy to transform into machine code. But a programmer must
be able to understand what he or she is writing, and a human cannot easily work at the level of
detail that machine language represents. So we use computer languages that are a compromise
between the needs of the speaker (programmer) and listener (computer). Declarations, types,
symbolic names, and the like are all concessions to a humans need to understand what someone
has written. The concession we make for computers is that we write programs in languages that can
be translated with relative ease into machine language. These languages have limited vocabulary
and limited syntax. Most belong to a class called context-free languages, which can be parsed easily
using a stack. Happily, as our skill at translation has increased, the variety and power of symbols
in our programming languages have also increased.
The language designer must define sets of rules and symbols that will be commonly understood
among both human and electronic users of the language. The meaning of these symbols is generally conveyed to people by the combination of a formal semantic description, analogy with other
languages, and examples. The meaning of symbols is conveyed to a computer by writing small
modules of machine code that define the action to be taken for each symbol. The rules of syntax
are conveyed to a computer by writing a compiler or interpreter.
To learn to use a new computer language effectively, a user must learn exactly what combinations
of symbols will be accepted by a compiler and what actions will be invoked for each symbol in the
language. This knowledge is the required common understanding. When the human communicates
with a machine, he must modify his own understanding until it matches the understanding of
the machine, which is embodied in the language translator. Occasionally the translator fails to
understand a phrase correctly, as specified by the official language definition. This happens
when there is an error in the translator. In this case the understanding of the translator must
be corrected by the language implementor.

1.2

Syntax and Semantics

The syntax of a language is a set of rules stating how language elements may be grammatically
combined. Syntax specifies how individual words may be written and the order in which words
may be placed within a sentence.
The semantics of a language define how each grammatically correct sentence is to be interpreted.
In a given language, the meaning of a sentence in a compiled language is the object code compiled
for that sentence. In an interpreted language, it is the internal representation of the program, which
is then evaluated. Semantic rules specify the meaning attached to each placement of a word in a
sentence, the meaning of omitting a sentence element, and the meaning of each individual word.
A speaker (or programmer) has an idea that he or she wishes to communicate. This idea is the
speakers semantic intent. The programmer must choose words that have the correct semantics so
that the listener (computer) can correctly interpret the speakers semantic intent.
All languages have syntax and semantics. Chapter 4 discusses formal mechanisms for expressing

CHAPTER 1. THE NATURE OF LANGUAGE

the syntax of a language. The rest of this book is primarily concerned with semantics, the semantics
of particular languages, and the semantic issues involved in programming.

1.3

Natural Languages and Programming Languages

We will often use comparisons with English to encourage you to examine language structures
intuitively, without preconceived ideas about what programming languages can or cannot do. The
objects and functions of a program correspond to the nouns and verbs of natural language. (We
will use the word functions to apply to functions, procedures, operators, and some commands.
Objects include variables, constants, records, and so on.)
There are a number of language traits that determine the character of a language. In this section
we compare the ways in which these traits are embodied in a natural language (English) and in
various programming languages. The differences between English and programming languages are
real, but not as great as they might at first seem. The differences are less extreme now than
they were ten years ago and will decrease as programming languages continue to evolve. Current
programming language research is directed toward:
Easing the constraints on the order in which statements must be given.
Increasing the uses of symbols with multiple definitions.
Permitting the programmer to talk about and use an object without knowing details of its
representation.
Facilitating the construction of libraries, thus increasing the number of words that can be
understood implicitly.
Increasing the ability of the language to express varied properties of the problem situation,
especially relationships among classes of objects.

1.3.1

Structure

Programs must conform to very strict structural rules. These govern the order of statements and
sections of code, and particular ways to begin, punctuate, and end every program. No deviation
from these rules is permitted by the language definition, and this is enforced by a compiler.
The structure of English is more flexible and more varied, but rules about the structure of
sentences and of larger units do exist. The overall structure of a textbook or a novel is tightly
controlled. Indeed, each kind of written material has some structure it must follow. In any situation
where the order of events is crucial, such as in a recipe, English sentences must be placed in the
correct sequence, just like the lines in a program.
Deviation from the rules of structure is permitted in informal speech, and understanding can
usually still be achieved. A human listener usually attempts to correct a speakers obvious errors.

1.3. NATURAL LANGUAGES AND PROGRAMMING LANGUAGES

For example, scrambled words can often be put in the right order. We can correct and understand
the sentence: I yesterday finished the assignment. Spoonerisms (exchanging the first letters of
nearby words, often humorously) can usually be understood. For example, I kee my sids was
obviously intended to mean I see my kids. A human uses common sense, context, and poorly
defined heuristics to identify and correct such errors.
Most programming language translators are notable for their intolerance of a programmers
omissions and errors. A compiler will identify an error when the input text fails to correspond
to the syntactic rules of the language (a syntax error) or when an object is used in the wrong
context (a type error). Most translators make some guesses about what the programmer really
meant, and try to continue with the translation, so that the programmer gets maximum feedback
from each attempt to compile the program. However, compilers can rarely correct anything more
than a trivial punctuation error. They commonly make faulty guesses which cause the generation
of heaps of irrelevant and confusing error comments.
Some compilers actually do attempt to correct the programmers errors by adding, changing,
respelling, or ignoring symbols so that the erroneous statement is made syntactically legal. If the
attempted correction causes trouble later, the compiler may return to the line with the error and try
a different correction. This effort has had some success. Errors such as misspellings and errors close
to the end of the code can often be corrected and enable a successful translation. Techniques have
been developed since the mid-1970s and are still being improved. Such error-correcting compilers
are uncommon because of the relatively great cost for added time and extra memory needed. Some
people feel that the added costs exceed the added utility.

1.3.2

Redundancy

The syntactic structure of English is highly redundant. The same information is often conveyed by
several words or word endings in a sentence. If required redundancy is absent, as in the sentence I
finishes the assignment tomorrow, we can identify that errors have occurred. The lack of agreement
between I and finishes is a syntactic error, and the disagreement of the verb tense (present)
with the meaning of tomorrow is a semantic error. [Exhibit 1.1]
A human uses the redundancy in the larger context to correct errors. For example, most people
would be able to understand that a single letter was omitted in the sentence The color of my coat
is back. Similarly, if a listener fails to comprehend a single word, she or he can usually use the
redundancy in the surrounding sentences to understand the message. If a speaker omits a word,
the listener can often supply it by using context.
Programming languages are also partly redundant, and the required redundancy serves as a
way to identify errors. For example, the first C declaration in Exhibit 1.2 contains two indications
of the intended data type of the variable named price: the type name, int, and the actual type,
float, of the initial value. These two indicators conflict, and a compiler can identify this as an
error. The second line contains an initializer whose length is longer than the declared size of the
array named table. This lack of agreement in number is an identifiable error.

CHAPTER 1. THE NATURE OF LANGUAGE

Exhibit 1.1. Redundancy in English.

The subject and verb of a sentence must agree in number. Either both must be singular or both
plural:
Correct: Mark likes the cake. Singular subject, singular verb.
Wrong:
Mark like the cake.
Singular subject, plural verb.
The verb tense must agree with any time words in the sentence:
Correct: I finished the work yesterday. Past tense, past time.
Wrong: I finish the work yesterday.
Present tense, past time.
Where categories are mentioned, words belonging to the correct categories must be used.
Correct: The color of my coat is black. Black is a color.
Wrong: The color of my coat is back. Back is not a color.
Sentences must supply consistent information throughout a paragraph. Pronouns refer to the
preceding noun. A pronoun must not suddenly be used to refer to a different noun.
Correct: The goalie is my son. He is the best. His name is Al.
Wrong:
The goalie is my son. He is the best. He is my father.
These errors in English have analogs in programming languages. The first error above is analogous
to using a nonarray variable with a subscript. The second and third errors are similar to type errors
in programming languages. The last error is analogous to faulty use of a pointer.

1.3.3

Using Partial Information: Ambiguity and Abstraction

English permits ambiguity, that is, words and phrases that have dual meanings. The listener must
disambiguate the sentence, using context, and determine the actual meaning (or meanings) of the
speaker.1
To a very limited extent, programming languages also permit ambiguity. Operators such as +
have two definitions in many languages, integer+integer and real+real. Object-oriented languages
permit programmer-defined procedure names with more than one meaning. Many languages are
block-structured. They permit the user to define contexts of limited scope, called blocks. The same
symbol can be given different meanings in different blocks. Context is used, as it is in English, to
disambiguate the meaning of the name.
1

A pun is a statement with two meanings, both intended by the speaker, where one meaning is usually funny.

Exhibit 1.2. Violations of redundancy rules in ANSI C.

int price = 20.98;
/* Declare and initialize variable.
int table[3] = {11, 12, 13, 14}; /* Declare and initialize an array.

*/
*/

1.3. NATURAL LANGUAGES AND PROGRAMMING LANGUAGES

The primary differences here are that context is defined very exactly in each programming
language and quite loosely in English, and that most programming languages permit only limited
ambiguity.
English supports abstraction, that is, the description of a quality apart from an instance. For
example, the word chair can be defined as a piece of furniture consisting of a seat, legs, and
back, and often arms, designed to accommodate one person. 2 This definition applies to many
kinds of chairs and conveys some but not all of a particular chairs properties. Older programming
languages do not support this kind of abstraction. They require that all an objects properties be
specified when the name for that object is defined.
Some current languages support very limited forms of abstraction. For example, Ada permits
names to be defined for generic objects, some of whose properties are left temporarily undefined.
Later, the generic definition must be instantiated by supplying actual definitions for those properties. The instantiation process produces fully specified code with no remaining abstractions which
can then be compiled in the normal way.
Smalltalk and C++ are current languages whose primary design goal was support for abstraction.
A Smalltalk declaration for a class chair would be parallel to the English definition. Languages
of the future will have more extensive ability to define and use partially specified objects.

1.3.4

Implicit Communication

English permits some things to be understood even if they are left unsaid. When we read between
the lines in an English paragraph, we are interpreting both explicit and implicit messages. Understanding of the explicit message is derived from the words of the sentence. The implicit message
is understood from the common experience of speaker and listener. People from different cultures
have trouble with implicit communication because they have inadequate common understanding.
Some things may be left implicit in programming languages also. Variable types in FORTRAN
and the type of the result of a function in the original Kernighan and Ritchie C may or may not
be defined explicitly. In these cases, as in English, the full meaning of such constructs is defined
by having a mutual understanding, between speaker and listener, about the meaning of things left
unspecified. A programmer learning a new language must learn its implicit assumptions, more
commonly called defaults.
Unfortunately, when a programmer relies on defaults to convey meaning, the compiler cannot
tell the difference between the purposeful use of a default and an accidental omission of an important
declaration. Many experienced programmers use explicit declarations rather than rely on defaults.
Stating information explicitly is less error prone and enables a compiler to give more helpful error
comments.
2

Cf. Morris [1969].

10

1.3.5

CHAPTER 1. THE NATURE OF LANGUAGE

Flexibility and Nuance

English is very flexible: there are often many ways to say something. Programming languages
have this same flexibility, as is demonstrated by the tremendous variety in the solutions handed in
for one student programming problem. As another example, APL provides at least three ways to
express the same simple conditional branch.
Alternate ways of saying something in English usually have slightly different meanings, and
subtlety and nuance are important. When different statement sequences in a programming language
express the same algorithm, we can say that they have the same meaning. However, they might still
differ in subtle ways, such as in the time and amount of memory required to execute the algorithm.
We can call such differences nuances.
The nuances of meaning in a program are of both theoretical and practical importance. We
are content when the work of a beginning programmer has the correct result (a way of measuring
its meaning). As programmers become more experienced, however, they become aware of the
subtle implications of alternative ways of saying the same thing. They will be able to produce a
program with the same meaning as the beginners program, but with superior clarity, efficiency,
and compactness.

1.3.6

Ability to Change and Evolve

Expressing an idea in any language, natural or artificial, can sometimes be difficult and awkward.
A person can become speechless when speaking English. Words can fail to express the strength
or complexity of the speakers feelings. Sometimes a large number of English words are required to
explain a new concept. Later, when the concept becomes well understood, a word or a few words
suffice.
English is constantly evolving. Old words become obsolete and new words and phrases are
added. Programming languages, happily, also evolve. Consider FORTRAN for example. The
original FORTRAN was a very limited language. For example, it did not support parameters and
did not have an IF...THEN...ELSE statement. Programmers who needed these things surely found
themselves speechless, and they had to express their logic in a wordy and awkward fashion. Useful
constructs were added to FORTRAN because of popular demand. As this happened, some of the
old FORTRAN words and methods became obsolete. While they have not been dropped from the
language yet, that may happen someday.
As applications of computers change, languages are extended to include words and concepts
appropriate for the new applications. An example is the introduction of words for sound generation
and graphics into Commodore BASIC when the Commodore-64 was introduced with sound and
graphics hardware.
One of the languages that evolves easily and constantly is FORTH. There are several public domain implementations, or dialects, used by many people and often modified to fit a users hardware
and application area. The modified dialect is then passed on to others. This process works like
the process for adding new meanings to English. New words are introduced and become common

1.4. THE STANDARDIZATION PROCESS

11

knowledge gradually as an increasing number of people learn and use them.

Translators for many dialects of BASIC, LISP, and FORTH are in common use. These languages
are not fully standardized. Many dialects of the original language emerge because implementors are
inspired to add or redesign language features. Programs written in one dialect must be modified to
be used by people whose computer understands a different dialect. When this happens we say
that a program is nonportable. The cost of rewriting programs makes nonstandardized programming languages unattractive to commercial users of computers. Lack of standardization can also
cause severe difficulties for programmers and publishers: the language specifications and reference
material must be relearned and rewritten for each new dialect.

1.4

The Standardization Process

Once a language is in widespread use, it becomes very important to have a complete and precise
definition of the language so that compatible implementations may be produced for a variety of
hardware and system environments. The standardization process was developed in response to
this need. A language standard is a formal definition of the syntax and semantics of a language.
It must be a complete, unambiguous statement of both. Language aspects that are defined must
be defined clearly, while aspects that go beyond the limits of the standard must be designated
clearly as undefined. A language translator that implements the standard must produce code
that conforms to all defined aspects of the standard, but for an undefined aspect, it is permitted
to produce any convenient translation.
The authority to define an unstandardized language or to change a language definition may
belong to the individual language designer, to the agency that sponsored the language design, or to
a committee of the American National Standards Institute (ANSI) or the International Standards
Organization (ISO). The FORTRAN standard was originated by ANSI, the Pascal standard by ISO.
The definition of Ada is controlled by the U.S. Department of Defense, which paid for the design
of Ada. New or experimental languages are usually controlled by their designers.
When a standards organization decides to sponsor a new standard for a language, it convenes
a committee of people from industry and academia who have a strong interest in and extensive
experience with that language. The standardization process is not easy or smooth. The committee
must decide which dialect, or combination of ideas from different dialects, will become the standard. Committee members come to this task with different notions of what is good or bad and
different priorities. Agreement at the outset is rare. The process may drag on for years as one or
two committee members fight for their pet features. This happened with the original ISO Pascal
standard, the ANSI C standard, and the new FORTRAN-90 standard.
After a standard is adopted by one standards organization (ISO or ANSI), the definition is
considered by the other. In the best of all worlds,the new standard would be accepted by the
second organization. For example, ANSI adopted the ISO standard for Pascal nearly unchanged.
However, smooth sailing is not always the rule. The new ANSI C standard is not acceptable to
some ISO committee members, and when ISO decides on a C standard,it may be substantially

12

CHAPTER 1. THE NATURE OF LANGUAGE

different from ANSI C.

The first standard for a language often clears up ambiguities, fixes some obvious defects, and
defines a better and more portable language. The ANSI C and ANSI LISP standards do all of
these things. Programmers writing new translators for this language must then conform to the
common standard, as far as it goes. Implementations may also include words and structures, called
extensions, that go beyond anything specified in the standard.

1.4.1

Language Growth and Divergence

After a number of years, language extensions accumulate and actual implementations diverge so
much that programs again become nonportable. This has happened now with Pascal. The standard
language is only minimally adequate for modern applications. For instance, it contains no support
for string processing or graphics. Further, it has design faults, such as an inadequate case statement, and design shortcomings, such as a lack of static variables, initialized variables, and support
for modular compilation. Virtually all implementations of Pascal for personal computers extend
the language. These extensions are similar in intent and function but differ in detail. A program
that uses the extensions is nonportable. One that doesnt use extensions is severely limited. We
all need a new Pascal standard.
When a standardized language has several divergent extensions in common use, the sponsoring
standards agency may convene a new committee to reexamine and restandardize the language. The
committee will consider the collection of extensions from various implementations and decide upon
a new standard, which usually includes all of the old standard as a subset.
Thus there is a constant tension between standardization and diversification. As our range of
applications and our knowledge of language and translation techniques increase, there is pressure
to extend our languages. Then the dialects in common use become diversified. When the diversity
becomes too costly, the language will be restandardized.

1.5

Nonstandard Compilers

It is common for compilers to deviate from the language standard. There are three major kinds of
deviations: extensions, intentional changes, and compiler bugs. The list of differences in Exhibit
1.3 was taken from the Introduction to the Turbo Pascal Reference Manual, Version 2.0. With
each new version of Turbo, this list has grown in size and complexity. Turbo Pascal version 5 is a
very different and much more extensive language than Standard Pascal.
An extension is a feature added to the standard, as string operations and graphics primitives are
often added to Pascal. Items marked with a + in Exhibit 1.3 are true extensions: they provide
processing capabilities for things that are not covered by the standard but do not change the basic
nature of the language.
Sometimes compiler writers believe that a language, as it is officially defined, is defective;
that is, some part of the design is too restrictive or too clumsy to use in a practical application
environment. In these cases the implementor often redefines the language, making it nonstandard

13

1.5. NONSTANDARD COMPILERS

Exhibit 1.3. Summary of Turbo Pascal deviations from the standard.

syntactic
extensions

semantic
extensions

semantic
changes
!
!
!

+
*
+
*
*
!
*
+
+
+

Absolute address variables

Bit/byte manipulation
Direct access to CPU memory and data ports
Dynamic strings
Free ordering of sections within declaration part
Full support of operating system facilities
In-line machine code generation
Include files
Logical operations on integers
Program chaining with common variables
Random access data files
Structured constants
Type conversion functions (to be used explicitly)

and incompatible with other translators. This is an intentional change. Items marked with a ! in
Exhibit 1.3 change the semantics of the language by circumventing semantic protection mechanisms
that are part of the standard. Items marked by a * are extensions and changes to the syntax of
the language that do not change the semantics but, if used, do make Turbo programs incompatible
with the standard.
A compiler bug occurs where, unknown to the compiler writer, the compiler implements different
semantics than those prescribed by the language standard. Examples of compiler bugs abound. One
Pascal compiler for the Commodore 64 required a semicolon after every statement. In contrast, the
Pascal standard requires semicolons only as separators between statements and forbids a semicolon
before an ELSE. A program written for this nonstandard compiler cannot be compiled by a standard
compiler and vice versa.
An example of a common bug is implementation of the mod operator. The easy way to
compute i mod j is to take the remainder after using integer division to calculate i/j. According
to the Pascal standard, quoted in Exhibit 1.4, 3 this computation method is correct if both i and
j are positive integers. If i is negative, though, the result must be adjusted by adding in the
modulus, j. The standard considers the operation to be an error if j is negative. Note that mod is
only the same as the mathematical remainder function if i >= 0 and j > 0.
Many compilers ignore this complexity, as shown in Exhibits 1.5 and 1.6. They simply perform
an integer division operation and return the result, regardless of the signs of i and j. For example,
in OSS Pascal for the Atari ST, the mod operator is defined in the usual nonstandard way. The OSS
3

Cooper [1983], page 3-1.

14

CHAPTER 1. THE NATURE OF LANGUAGE

Exhibit 1.4. The definition of mod in Standard Pascal.

The value of i mod j is the value of i(k*j) for an integer value k, such that 0 <= (i mod j) <
j. (That is, the value is always between 0 and j.)
The expression i mod j is an error if j is zero or negative.

Pascal reference manual (page 6-26) describes mod as follows:

The modulus is the remainder left over after integer division.
Compiling and testing a few simple expressions [Exhibit 1.5] substantiates this and shows how OSS
Pascal differs from the standard. Expression 2 gives a nonstandard answer. Expressions (3) through
(6) compile and run, but shouldnt. They are designated as errors in the standard, which requires
the modulus to be greater than 0. These errors are not detected by the OSS Pascal compiler or
run-time system, nor does the OSS Pascal reference manual state that they will not be detected,
as required by the standard.
In defense of this nonstandard implementation, one must note that this particular deviation is
common and the function it computes is probably more useful than the standard definition for mod.
The implementation of mod in Turbo Pascal is different, but also nonstandard, and may have
been an unintentional deviation. It was not included on the list of nonstandard language features.
[Exhibit 1.3] The author of this manual seems to have been unaware of this nonstandard nature of
mod and did not even describe it adequately. The partial information given in the Turbo reference
manual (pages 5152) is as follows:
mod is only defined for integers
its result is an integer
12 mod 5 = 2
Exhibit 1.5. The definition of mod in OSS Pascal for the Atari ST.
1.
2.
3.
4.
5.
6.

Expression
5 mod 2
-5 mod 2
5 mod -2
-5 mod -2
5 mod 0
-5 mod 0

OSS result
1
-1
1
-1
0
-1

Answer according to Pascal Standard

Correct.
Should be 1 (between 0 and the modulus-1).
Should be detected as an error.
Should be detected as an error.
Should be detected as an error.
Should be detected as an error.

15

1.5. NONSTANDARD COMPILERS

Exhibit 1.6. The definition of mod in Turbo Pascal for the IBM PC.
1.
2.
3.
4.
5.
6.

Expression
5 mod 2
-5 mod 2
5 mod -2
-5 mod -2
5 mod 0
-5 mod 0

Turbo result
1
-1
1
-1
Run-time error
Run-time error

Answer according to Pascal Standard

Correct.
Should be -1 + 2 = 1.
Should be an error.
Should be an error.
Correct.
Correct.

The reference manual for Turbo Pascal version 4.0 still does not include mod on the list of nonstandard features. However, it does give an adequate definition (p. 240) of the function it actually
computes for mod:
the mod operator returns the remainder from dividing the operands:
i mod j = i (i/j) j.
The sign of the result is the sign of i. An error occurs if j = 0.
Compiling and testing a few simple expressions [Exhibit 1.6] substantiates this definition. Expression 2 gives a nonstandard answer. Expressions (3) and (4) are designated as errors in the
standard, which requires the modulus to be greater than 0. These errors are not detected by the
Turbo compiler. Furthermore, its reference manual does not state that they will not be detected,
as required by the standard.
While Turbo Pascal will not compile a div or mod operation with 0 as a constant divisor, the
result of i mod 0 can be tested by setting a variable, j, to zero, then printing the results of i mod
j. This gives the results on lines (5) and (6).
Occasionally, deviations from the standard occur because an implementor believes that the
standard, although unambiguous, defined an item wrong; that is, some other definition would
have been more efficient or more useful. The version incorporated into the compiler is intended as
an improvement over the standard. Again, the implementation of mod provides an example here.
In many cases, the programmer who uses mod really wants the arithmetic remainder, and it seems
foolish for the compiler to insert extra lines of code in order to compute the unwanted standard
Pascal function. At least one Pascal compiler (for the Apollo workstation) provides a switch that
can be set either to compile the standard meaning of mod or to compile the easy and efficient
meaning. The person who wrote this compiler clearly believed that the standard was wrong to
include the version it did rather than the integer remainder function.
The implementation of input and output operations in Turbo Pascal version 2.0 provides another
example of a compiler writer who declined to implement the standard language because he believed
his own version was clearly superior. He explains this decision as follows: 4
4

Borland [1984], Appendix F.

16

CHAPTER 1. THE NATURE OF LANGUAGE

The standard procedures GET and PUT are not implemented. Instead, the READ and
WRITE procedures have been extended to handle all I/O needs. The reason for this is
threefold: Firstly, READ and WRITE gives much faster I/O, secondly variable space
overhead is reduced, as file buffer variables are not required, and thirdly the READ
and WRITE procedures are far more versatile and easier to understand than GET and
PUT.

The actual Turbo implementation of READ did not even measure up to the standard in a minimal
way, as it did not permit the programmer to read a line of input from the keyboard one character
at a time. (It is surely inefficient to do so but essential in some applications.) Someone who did not
know that this deviation was made on purpose would think that it was simply a compiler bug. This
situation provides an excellent example of the dangers of taking the law into your own hands.
Whether or not we agree with the requirements of a language standard, we must think carefully
before using nonstandard features. Every time we use a nonstandard feature or one that depends
on the particular bit-level implementation of the language, it makes a program harder to port from
one system to another and decreases its potential usefulness and potential lifetime. Programmers
who use nonstandard features in their code should segregate the nonstandard segments and
thoroughly document them.

Exercises
1. Define natural language. Define programming language. How are they different?
2. How are languages used to establish communication?
3. What is the syntax of a language? What are the semantics?
4. What are the traits that determine the character of a language?
5. How do these traits appear in programming languages?
6. What need led to standardization?
7. What is a standard for a language?
8. What does it mean when a language standard defines something to be undefined?
9. How does standardization lead to portability?
10. What three kinds of deviations are common in nonstandard compilers?
11. What are the advantages and disadvantages of using nonstandard language features?

Chapter 2

Representation and Abstraction

Overview
This chapter presents the concept of how real-world objects, actions, and changes in
the state of a process are represented through a programming language on a computer.
Programs can be viewed as either a set of instructions for the computer to execute
or as a model of some real-world process. Languages designed to support these views
will exhibit different properties. The language designer must establish a set of goals
for the language and then examine them for consistency, importance, and restrictions.
Principles for evaluating language design are presented. Classification of languages into
groups is by no means an easy task. Categories for classifying languages are discussed.

Representation may be explicit or implicit, coherent or diffused.

2.1

What Is a Program?

We can view a program two ways.

1. A program is a description of a set of actions that we want a computer to carry out. The
actions are the primitive operations of some real or abstract machine, and they are performed
using the primitive parts of a machine. Primitive actions include such things as copying
data from one machine register or a memory location to another, applying an operation to a
register, or activating an input or output device.

17

18

CHAPTER 2. REPRESENTATION AND ABSTRACTION

2. A program is a model of some process in the real or mathematical world. The programmer
must set up a correspondence between symbols in the program and real-world objects, and
between program functions and real-world processes. Executing a function represents a change
in the state of the world or finding a solution to a set of specifications about elements of that
world.

These two world-views are analogous to the way a builder and an architect view a house. The
builder is concerned with the method for achieving a finished house. It should be built efficiently
and the result should be structurally sound. The architect is concerned with the overall function
and form of the house. It should carry out the architects concepts and meet the clients needs.
The two world-views lead to very different conclusions about the properties that a programming
language should have. A language supporting world-view (1) provides ready access to every part
of the computer so that the programmer can prescribe in detail how the computer should go
about solving a given problem. The language of a builder contains words for each material and
construction method used. Similarly, a program construction language allows one to talk directly
about hardware registers, memory, data movement, I/O devices, and so forth. The distinction isnt
simply whether the language is low-level or high-level, for assembly language and C are both
designed with the builder in mind. Assembly language is, by definition, low-level, and C is not,
since it includes control structures, type definitions, support for modules, and the like. However,
C permits (and forces) a programmer to work with and be aware of the raw elements of the host
computer.
A language supporting world-view (2) must be able to deal with abstractions and provide
a means for expressing a model of the real-world objects and processes. An architect deals with
abstract concepts such as space, form, light, and functionality, and with more concrete units such as
walls and windows. Blueprints, drawn using a formal symbolic language, are used to represent and
communicate the plan. The builder understands the language of blueprints and chooses appropriate
methods to implement them.
The languages Smalltalk and Prolog were designed to permit the programmer to represent and
communicate a world-model easily. They free the programmer of concerns about the machine and
let him or her deal instead with abstract concepts. In Smalltalk the programmer defines classes
of objects and the processes relevant to these classes. If an abstract process is relevant to several
classes, the programmer can define how it is to be accomplished for each. In Prolog the programmer
represents the world using formulas of mathematical logic. In other languages, the programmer
may use procedures, type declarations, structured loops, and block structure. to represent and
describe the application. Writing a program becomes a process of representing objects, actions,
and changes in the state of the process being modeled [Exhibit 2.1].
The advantage of a builders language is that it permits the construction of efficient software
that makes effective use of the computer on which it runs. A disadvantage is that programs tailored
to a particular machine cannot be expected to be well suited to another machine and hence they
are not particularly portable.

2.1. WHAT IS A PROGRAM?

19

Exhibit 2.1. Modeling a charge account and relevant processes.

Objects: A program that does the accounting for a companys charge accounts must contain
representations for several kinds of real-world objects: accounts, payments, the current balance,
items charged, items returned, interest.
Actions: Each action to be represented involves objects from a specified class or classes. The
actions to be represented here include the following:
Credit a payment to an account.
Send a bill to the account owner.
Debit a purchase to an account.
Credit a return to an account.
Compute and debit the monthly interest due.
Changes of state: The current balance of an account, todays date, and the monthly payment
date for that account encode the state of the account. The balance may be positive, negative,
or zero, and a positive balance may be either ok or overdue. Purchases, returns, payments, and
monthly due dates and interest dates all cause a change in the state of the account.

Moreover, a programmer using such a language is forced to organize ideas at a burdensome level
of detail. Just as a builder must be concerned with numerous details such as building codes, lumber
dimensions, proper nailing patterns, and so forth, the program builder likewise deals with storage
allocation, byte alignment, calling sequences, word sizes, and other details which, while important
to the finished product, are largely unrelated to its form and function.
By way of contrast, an architects language frees one from concern about the underlying
machine and allows one to describe a process at a greater level of abstraction, omitting the minute
details. A great deal of discretion is left to the compiler designer in choosing methods to carry out
the specified actions. Two compilers for the same architects language often produce compiled code
of widely differing efficiency and storage requirements.
In fact, there is no necessary reason why there must be a compiler at all. One could use the
architects language to specify the form and function of the finished program and then turn the
job over to a program builder. However, the computer can do a fairly good job of automatically
producing a program for such languages, and the ability to have it do so gives the program architect
a powerful tool not available to the construction architectthe ability to rapidly prototype designs.
This is the power of the computer, and one of the aspects that makes the study of programming
language so fascinating!

20

CHAPTER 2. REPRESENTATION AND ABSTRACTION

Exhibit 2.2. Representations of a date in English.

Dates are abstract real-world objects. We represent them in English by specifying an era,
year, month, and day of the month. The era is usually omitted and the year is often omitted in
representations of dates, because they can be deduced from context. The month is often encoded
as an integer between 1 and 12.
Full, explicit representation: January 2, 1986 AD
Common representations:

2.2

January 2, 1986
Jan. 2, 86
Jan. 2
1-2-86
2 Jan 86
86-1-2

Representation

A representation of an object is a list of the relevant facts about that object in some language
[Exhibit 2.2]. A computer representation of an object is a mapping of the relevant facts about that
object, through a computer language, onto the parts of the machine.
Some languages support high-level or abstract representations, which specify the functional
properties of an object or the symbolic names and data types of the fields of the representation
[Exhibit 2.3]. A high-level representation will be mapped onto computer memory by a translator.
The actual number and order of bytes of storage that will be used to represent the object may vary
from translator to translator. In contrast, a computer representation is low level if it describes a
particular implementation of the object, such as the amount of storage that will be used, and the
position of each field in that storage area [Exhibit 2.4].
A computer representation of a process is a sequence of program definitions, specifications, or

Exhibit 2.3. High-level computer representations of a date.

An encoding of the last representation in Exhibit 2.2 is often used in programs. In a high-level
language the programmer might specify that a date will be represented by three integers, as in this
Pascal example:
TYPE date = RECORD year, month, day: integer END;
VAR BirthDate: date;
The programmer may now refer to this object and its components as:
BirthDate or
BirthDate.year or BirthDate.month or BirthDate.day

2.2. REPRESENTATION

21

Exhibit 2.4. A low-level computer representation of a date.

In a low-level language such as assembler or FORTH, the programmer specifies the exact number
of bytes of storage that must be allocated (or ALLOTted) to represent the date. In the FORTH
declaration below, the keyword VARIABLE causes 2 bytes to be allocated, and 4 more are explicitly
allocated using ALLOT. Then the programmer must manually define selection functions that access
the fields of the object by adding an appropriate offset to the base address.
VARIABLE birth_date 4 ALLOT
: year 0 + ;
( Year is first -- offset is zero bytes. )
: month 2 + ;
( Month starts two bytes from the beginning. )
: day 4 + ;
( Day is the fifth and sixth bytes. )
The variable named birth_date and its component fields can now be accessed by writing:
birth_date or
birth_date year or birth_date month or birth_date day

statements that can be performed on representations of objects from specified sets. We say that
the representation of a process is valid, or correct, if the transformed object representation still
corresponds to the transformed object in the real world.
We will consider three aspects of the quality of a representation: semantic intent, explicitness,
and coherence. Abstract representations have these qualities to a high degree; low-level representations often lack them.

2.2.1

Semantic Intent

A data object (variable, record, array, etc.) in a program has some intended meaning that is known
to the programmer but cannot be deduced with certainty from the data representation itself. This
intended meaning is the programmers semantic intent. For example, three 2-digit integers can
represent a womans measurements in inches or a date. We can only know the intended meaning
of a set of data if the programmer communicates, or declares, the context in which it should be
interpreted.
A program has semantic validity if it faithfully carries out the programmers explicitly declared
semantic intent. We will be examining mechanisms in various languages for expressing semantic
intent and ensuring that it is carried out. Most programming languages use a data type to encode
part of the semantic intent of an object. Before applying a function to a data object, the language
translator tests whether the function is defined for that object and, therefore, is meaningful in its
context. An attempt to apply a function to a data object of the wrong type is identified as a
semantic error. A type checking mechanism can thus help a programmer write semantically valid
(meaningful) programs.

22

CHAPTER 2. REPRESENTATION AND ABSTRACTION

Exhibit 2.5. The structure of a table expressed implicitly.

Pascal permits the construction and use of sorted tables, but the fact that the table is sorted
cannot be explicitly declared. We can deduce that the table is sorted by noting that a sort algorithm
is invoked, that a binary search algorithm is used, or that a sequential search algorithm is used
that can terminate a search unsuccessfully before reaching the end of the table.
The order of entries (whether ascending or descending) can be deduced by careful analysis of
three things:
The comparison operator used in a search (< or >)
The order of operands in relation to this operator
The result (true or false) which causes the search to terminate.
Deductions of this sort are beyond the realistic present and future abilities of language translators.

2.2.2

Explicit versus Implicit Representation

The structure of a data object can be reflected implicitly in a program, by the way the statements
are arranged [Exhibit 2.5], or it can be declared explicitly [Exhibit 2.6]. A language that can declare
more kinds of things explicitly is more expressive.
Information expressed explicitly in a program may be used by the language translator. For
example, if the COBOL programmer supplies a KEY clause, the processor will permit the programmer to use the efficient built-in binary search command, because the KEY clause specifies that the
file is sorted in order by that field. The less-efficient sequential search command must be used to
search any table that does not have a KEY clause.
A language that permits explicit communication of information must have a translator that
can identify, store, organize, and utilize that information. For example, if a language permits
programmers to define their own types, the translator needs to implement type tables (where type
descriptions are stored), new allocation methods that use these programmer-defined descriptions,
and more elaborate rules for type checking and type errors.
These translator mechanisms to identify, store, and interpret the programmers declarations
form the semantic basis of a language. Other mechanisms that are part of the semantic basis are
those which implement binding (Chapters 6 and 9), type checking and automatic type conversion
(Chapter 15), and module protection (Chapter 16).

2.2.3

Coherent versus Diffuse Representation

A representation is coherent if an external entity (object, idea, or process) is represented by a single

symbol in the program (a name or a pointer) so that it may be referenced and manipulated as a
unit [Exhibit 2.7]. A representation is diffuse if various parts of the representation are known by

2.2. REPRESENTATION

23

Exhibit 2.6. COBOL: The structure of a table expressed explicitly.

COBOL allows explicit declaration of sorted tables. The key field(s) and the order of entries
may be declared as in the following example. This table is intended to store the names of the
fifty states of the United States and their two-letter abbreviations. It is to be stored so that the
abbreviations are in alphabetical order.
01 state-table.
02 state-entry
OCCURS 50 TIMES
ASCENDING KEY state-abbrev
INDEXED BY state-index.
03 state-abbrev
PICTURE XX.
03 state-name
PICTURE X(20).
This table can be searched for a state abbreviation using the binary-search utility. A possible call
is:
SEARCH ALL state-entry
AT END PERFORM failed-search-process
WHEN st-abbrev (state-index) = search-key PERFORM found-process.
COBOL also permits the programmer to declare Pascal-like tables for which the sorting order and
key field are not explicitly declared. The SEARCH ALL command cannot be used to search such a
table; the programmer can only use the less efficient sequential search command.

different names, and no one name or symbol applies to the whole [Exhibits 2.8 and 2.9].
A representation is coherent if all the parts of the represented object can be named by one
symbol. This certainly does not imply that all the parts must be stored in consecutive (or contiguous) memory locations. Thus an object whose parts are connected by links or pointers can still be
coherent [Exhibit 2.10].
The older languages (FORTRAN, APL) support coherent representation of complex data objects

Exhibit 2.7. A stack represented coherently in Pascal.

A stack can be represented by an array and an integer index for the number of items currently in
the array. We can represent a stack coherently by grouping the two parts together into a record.
One parameter then suffices to pass this stack to a function.
TYPE stack = RECORD
store: ARRAY [1..max_stack] of stack_type;
top: 0..max_stack
END;

24

CHAPTER 2. REPRESENTATION AND ABSTRACTION

Exhibit 2.8. A stack represented diffusely in FORTRAN and Pascal.

A stack can be represented diffusely as an array of items and a separate index (an integer in the
range 0 to the size of the array).
FORTRAN: A diffuse representation is the only representation of a stack that is possible in
FORTRAN because the language does not support heterogeneous records. These declarations create
two objects which, taken together, comprise a stack of 100 real numbers:
REAL
STSTORE( 100 )
INTEGER STTOP
Pascal: A stack can be represented diffusely in Pascal. This code allocates the same amount of
storage as the coherent version in Exhibit 2.7, but two parameters are required to pass this stack
to a procedure.
TYPE stack_store = ARRAY [1..max_stack] of stack_type;
stack_top = 0..max_stack;

Exhibit 2.9. Addition represented diffusely or coherently.

FORTH: The operation of addition is represented diffusely because addition of single-length
integers, double-length integers, and mixed-length integers are named by three different symbols,
(+, D+, and M+) and no way is provided to refer to the general operation of integer +.
C: The operation of addition is represented coherently. Addition of single-length integers,
double-length integers, and mixed-length integers are all named +. The programmer may refer
to + for addition, without concern for the length of the integer operands.

Exhibit 2.10. A coherently but not contiguously represented object.

LISP: A linked tree structure or a LISP list is a coherent object because a single pointer to the
head of the list allows the entire list to be manipulated. A tree or a list is generally implemented
by using pointers to link storage cells at many memory locations.
C: A sentence may be represented as an array of words. One common representation is a
contiguous array of pointers, each of which points to a variable-length string of characters. These
strings are normally allocated separately and are not contiguous.

2.3. LANGUAGE DESIGN

25

only if the object can be represented by a homogeneous array of items of the same data type. 1 Where
an object has components represented by different types, separate variable names must be used.
COBOL and all the newer languages support coherent heterogeneous groupings of data. These are
called records in COBOL and Pascal, and structures in C.
The FORTRAN programmer can use a method called parallel arrays to model an array of
heterogeneous records. The programmer declares one array for each field of the record, then
uses a single index variable to refer to corresponding elements of the set of arrays. This diffuse
representation accomplishes the same goal as a Pascal array of records. However, an array of records
represents the problem more clearly and explicitly and is easier to use. For example, Pascal permits
an array of records to be passed as a single parameter to a function, whereas a set of parallel arrays
in FORTRAN would have to be passed as several parameters.
Some of the newest languages support coherence further by permitting a set of data representations to be grouped together with the functions that operate on them. Such a coherent grouping
is called a module in Modula-2, a cluster in CLU, a class in Smalltalk, and a package in
Ada.

2.3

Language Design

In this section we consider reasons why a language designer might choose to create an architects
language with a high degree of support for abstraction, or a builders language with extensive
control over low-level aspects of representation.

2.3.1

Competing Design Goals

Programming languages have evolved greatly since the late 1950s when the first high-level languages,
FORTRAN and COBOL, were implemented. Much of this evolution has been made possible by the
improvements in computer hardware: todays machines are inconceivably cheap, fast, and large (in
memory capacity) compared to the machines available in 1960. Although those old machines were
physically bulky and tremendously expensive, they were hardly more powerful than machines that
today are considered to be toys.
Along with changes in hardware technology came improvements in language translation techniques. Both syntax and semantics of the early languages were ad hoc and clumsy to translate.
Formal language theory and formal semantics affected language design in revolutionary ways and
have resulted in better languages with cleaner semantics and a more easily translatable syntax.
There are many aspects of a language that the user cannot modify or extend, such as the data
structuring facilities and the control structures. Unless a language system supports a preprocessor,
the language syntax, also, is fixed. If control structures and data definition facilities are not built
1

The EQUIVALENCE statement can be used to circumvent this weakness by defining the name of the coherent
object as an overlay on the storage occupied by the parts. This does not constitute adequate support for compound
heterogeneous objects.

26

CHAPTER 2. REPRESENTATION AND ABSTRACTION

in, they are not available. Decisions to include or exclude such features must, therefore, be made
carefully. A language designer must consider several aspects of a potential feature to decide whether
it supports or conflicts with the design goals.
During these thirty years of language development, a consensus has emerged about the importance of some language features, for example, type checking and structured conditionals. Most new
languages include these. On other issues, there has been and remains fundamental disagreement,
for instance, over the question of whether procedural or functional languages are better. No
single set of value judgments has yet emerged, because different languages have different goals and
different intended uses. The following are some potential language design goals:
Utility. Is a feature often useful? Can it do important things that cannot be done using other
features of the language?
Convenience. Does this feature help avoid excessive writing? Does this feature add or eliminate clutter in the code?
Efficiency. Is it easy or difficult to translate this feature? Is it possible to translate this
feature into efficient code? Will its use improve or degrade the performance of programs?
Portability. Will this feature be implementable on any machine?
Readability. Does this form of this feature make a program more readable? Will a programmer
other than the designer understand the intent easily? Or is it cryptic?
Modeling ability. Does this feature help make the meaning of a program clear? Will this
feature help the programmer model a problem more fully, more precisely, or more easily?
Simplicity. Is the language design as a whole simple, unified, and general, or is it full of
dozens of special-purpose features?
Semantic clarity. Does every legal program and expression have one defined, unambiguous,
meaning? Is the meaning constant during the course of program execution?
These goals are all obviously desirable, but they conflict with each other. For example, a simple
language cannot possibly include all useful features, and the more features included, the more
complicated the language is to learn, use, and implement. Ada illustrates this conflict. Ada was
designed for the Department of Defense as a language for embedded systems, to be used in all
systems development projects, on diverse kinds of hardware. Thus it necessarily reflects a high
value placed on items at the beginning and middle of the preceding list of design goals. The result
is a very large language with a long list of useful special features.
Some language researchers have taken as goals the fundamental properties of language shown
at the end of the list of design goals. Outstanding examples include Smalltalk, a superior language
for modeling objects and processes, and Miranda, which is a list-oriented functional language that
achieves both great simplicity and semantic clarity.

27

2.3. LANGUAGE DESIGN

Exhibit 2.11. A basic type not supported by Pascal.

Basic type implemented by most hardware: bit strings
Common lengths: 8, 16, and 32 bits (1, 2, and 4 bytes)
Operations built into most hardware
a right shift n places
a left shift n places
a and b
a or b
a exclusive or b
complement a

2.3.2

Symbol in C
a >> n
a << n
a & b
a | b
a ^ b
~ a

The Power of Restrictions

Every language imposes restrictions on the user, both by what it explicitly prohibits and by what
it simply doesnt provide. Whenever the underlying machine provides instructions or capabilities
that cannot be used in a user program, the programming language is imposing a restriction on the
user. For example, Pascal does not support the type bit string and does not have bit string
operators [Exhibit 2.11]. Thus Pascal restricts access to the bit-level implementations of objects.
The reader must not confuse logical operators with bitwise operators. Pascal supports the logical
(Boolean) data type and logical operators and, or, and not. Note that there is a difference between
these and the bitwise operators [Exhibit 2.12]. Bitwise operators apply the operation between every
corresponding pair of bits in the operands. Logical operators apply the operation to the operands
as a whole, with 00000000 normally being interpreted as False and anything else as True.
In general, restrictions might prevent writing the following two sorts of sentences:
1. Useless or meaningless sentences such as 3 := 72.9 + 0 a0 .
2. Sentences useful for modeling some problem, that could be written efficiently in assembly

Exhibit 2.12. Bitwise and logical operations.

The difference between bitwise and logical operations can
output from these operations in C:
Operation
Operands as bit strings Result
bitwise and
10111101 & 01000010
00000000
logical and
10111101 && 01000010 00000001
complement
logical not

01000010
! 01000010

10111101
00000000

be seen by comparing the input and

Explanation
no bit pairs match
both operands represent True
each bit is flipped
operand is True

28

CHAPTER 2. REPRESENTATION AND ABSTRACTION

Exhibit 2.13. Useful pointer operations supported in C.

In the C expressions below, p is a pointer used as an index for the array named ages. Initially, p
will store the machine address of the beginning of ages. To make p index the next element, p will
be incremented by the number of bytes in one element of ages. The array element is accessed by
dereferencing the pointer.
This code contains an error in logic, which is pointed out in the comments. It demonstrates one
semantic problem that Pascals restrictions were designed to prevent. The loop will be executed
too many times, and this run-time error will not be detected. Compare this to the similar Pascal
loop in Exhibit 2.14.
int ages[10];
...
p = ages;
end = &ages[10];
while (p <= end){
printf("%d \n", *p);
p++;
}

/* Array "ages" has subscripts 0..9.

/*
/*
/*
/*
/*

*/

Make p point at ages[0]. */

Compute address of eleventh array element. */
Loop through eleventh array address. */
Print array element in decimal integer format. */
Increment p to point at next element of ages. */

code but are prohibited.

A good example of a useful facility that some languages prohibit is explicit address manipulation.
This is supported in C [Exhibit 2.13]. The notation for pointer manipulation is convenient and is
generally used in preference to subscripting when the programmer wishes to process an array
sequentially.
In contrast, manipulation of addresses is restricted in Pascal to prevent the occurrence of meaningless and potentially dangerous dangling pointers (see Chapter 6). Address manipulation is
prohibited and address arithmetic is undefined in Pascal. Nothing comparable to the C code in
Exhibit 2.13 can be written in Pascal. A pointer cant be set to point at an array element, and it
cannot be incremented to index through an array.
This is a significant loss in Pascal because using a subscript involves a lot of computation: the
subscript must be checked against the minimum and maximum legal values, multiplied by the size
of an array element, and added to the base address of the array. Checking whether a pointer has
crossed an array boundary and using it to access an element could be done significantly faster.
Let us define flexibility to mean the absence of a restriction, and call a restriction good if it
prevents the writing of nonsense, and bad if it prevents writing useful things. Some restrictions
might have both good and bad aspects. A powerful language must have the flexibility to express a
wide variety of actionspreferably a variety that approaches the power of the underlying machine.
But power is not a synonym for flexibility. The most flexible of all languages is assembly
language, but assemblers lack the power to express a problem solution succinctly and clearly. A

29

2.3. LANGUAGE DESIGN

Exhibit 2.14. A meaningless operation prohibited in Pascal but not in C.

Subscripts are checked at run time in Pascal. Every subscript that is used must be within the
declared bounds.
VAR ages: array[0..9] of integer;
p: integer;
...
p := 0;
while p <= 10 do begin
/* Loop through last array subscript. */
writeln( ages[p] ); /* Print the array element. */
p:=p+1
/* Make p point at next element of array.
end;

*/

The last time around this loop the subscript, p, has a value that is out of range. This will be
detected, and a run-time error comment will be generated. The analogous C code in Exhibit 2.13
will run and print garbage on the last iteration. The logical error will not be detected, and no error
comment will be produced.

second kind of power is provided by sophisticated mechanisms in the semantic basis of a language
that let the programmer express a lot by saying a little. The type definition and type checking
facility in any modern language is a good example of a powerful mechanism.
A third kind of power can come from good restrictions that narrow the variety of things that
can be written. If a restriction can eliminate troublesome or meaningless sentences automatically,
then programmers will not have to check, explicitly, whether such meaningless sections occur in
their programs. Pascal programs rarely run wild and destroy memory. But C and FORTH programs,
with unrestricted pointers and no subscript bounds checking, often do so. A language should have
enough good restrictions so that the programmer and translator can easily distinguish between a
meaningful statement and nonsense.
For example, an attempt to access an element of an array with a subscript greater than the
largest array subscript is obviously meaningless in any language. The underlying machine hardware
permits one to FETCH and STORE information beyond the end of an array, but this can have no
possible useful meaning and is likely to foul up the further operation of the program. The semantics
of standard Pascal prescribe that the actual value of each subscript expression should be checked at
run time. An error comment is generated if the value is not within the declared array bounds. Thus,
all subscripting in Pascal is safe and cannot lead to destruction of other information [Exhibit 2.14].
No such array bounds check is done in C. Compare Exhibits 2.13 and 2.14. These two code
fragments do analogous things, but the logical error inherent in both will be trapped by Pascal
and ignored by C. In C, a FETCH operation with too large a subscript can supply nonsensical
information, and a STORE can destroy vital, unrelated information belonging to variables allocated
before or after the array. This situation was exploited to create the computer network worm that

30

CHAPTER 2. REPRESENTATION AND ABSTRACTION

invaded hundreds of computer systems in November 1988. It disabled these systems by flooding
their processing queues with duplicates of itself, preventing the processing of normal programs.
This escapade resulted in the arrest and conviction of the programmer.
Often, as seen in Exhibit 2.13, a single feature is both useful and dangerous. In that case,
a language designer has to make a value judgement about the relative importance of the feature
and the danger in that feature. If the designer considers the danger to outweigh the importance,
the restriction will be included, as Wirth included the pointer restrictions in Pascal. If the need
outweighs the danger, the restriction will not be included. In designing C, Kernighan and Ritchie
clearly felt that address manipulation was vital, and decided that the dangers of dangling pointers
would have to be avoided by careful programming, not by imposing general restrictions on pointers.

2.3.3

Principles for Evaluating a Design

In the remaining chapters of this book we will sometimes make value judgments about the particular
features that a language includes or excludes. These judgments will be based on a small set of
principles.
Principle of Frequency
The more frequently a language feature will be used, the more convenient its use should be, and
the more lucid its syntax should be. An infrequently used feature can be omitted from the core of
the language and/or be given a long name and less convenient syntax.
C provides us with examples of good and poor application of this principle. The core of the C
language does not include a lot of features that are found in the cores of many other languages. For
example, input/output routines and mathematical functions for scientific computation are not part
of the standard language. These are relegated to libraries, which can be searched if these features
are needed. There are two C libraries which are now well standardized, the math library and the
C library (which includes the I/O functions).
The omission of mathematical functions from C makes good sense because the intended use of
C was for systems programming, not scientific computation. Putting these functions in the math
library makes them available but less convenient. To use the math library, the loader must have
the library on its search path and the user must include a header file in the program which contains
type declarations for the math functions.
On the other hand, most application programs use the input-output functions, so they should
be maximally convenient. In C they arent; in order to use them a programmer must include the
appropriate header file containing I/O function and macro declarations, and other essential things.
Thus nearly every C application program starts with the instruction #include hstdio.hi. This
could be considered to be a poor design element, as it would cost relatively little to build these
definitions into the translator.

2.3. LANGUAGE DESIGN

31

Principle of Locality
A good language design enables and encourages, perhaps even enforces, locality of effects. The
further the effects of an action reach in time (elapsed during execution) or in space (measured in
pages of code), the more complex and harder it is to debug a program. The further an action has
influence, the harder it is to remember relevant details, and the more subtle errors seem to creep
into the code.
To achieve locality, the use of global variables should be minimized or eliminated and all transfers
of control should be short-range. A concise restatement of this principle, in practical terms is:
Keep the effects of everything confined to as local an area of the code as possible.
Here are some corollaries of the general principle, applied to lexical organization of a program
that will be debugged on-line, using an ordinary nonhierarchical text editor:
A control structure that wont fit on one screen is too long; shorten it by defining one or more
scopes as subroutines.
All variables should be defined within one screen of their use. This applies whether the users
screen is large or smallthe important thing is to be able to see an entire unit at one time.
If your subroutine wont fit on two screens, it is too long. Break it up.
Global Variables. Global variables provide a far more important example of the cost of nonlocality. A global variable can be changed or read anywhere within a program. Specifically, it can be
changed accidentally (because of a typographical error or a programmers absentmindedness) in a
part of the program that is far removed from the section in which it is (purposely) used.
This kind of error is hard to find. The apparent fault is in the section that is supposed to use
the variable, but if that section is examined in isolation, it will work properly. To find the cause
of the error, a programmer must trace the operation of the entire program. This is a tedious job.
The use of unnecessary global variables is, therefore, dangerous.
If the program were rewritten to declare this variable locally within the scope in which it is used,
the distant reference would promptly be identified as an error or as a reference to a semantically
distinct variable that happens to have the same name.
Among existing languages are those that provide only global variables, provide globals but
encourage use of locals and parameters, and provide only parameters.
Unrestricted use of global variables. A BASIC programmer cannot restrict a variable to
a local scope. This is part of the reason that BASIC is not used for large systems programs.

32

CHAPTER 2. REPRESENTATION AND ABSTRACTION

Use of global variables permitted but use of locals encouraged. Pascal and C are
block structured languages that make it easy to declare variables in the procedure in which they
are used.2 Their default method of parameter passing is call-by-value. Changing a local variable
or value parameter has only local effects. Programmers are encouraged to use local declarations,
but they can use global variables in place of both local variables and parameters.
Use of global variables prohibited. In the modern functional languages there are no global
variables. Actually, there are no variables at all, and parameter binding takes the place of assignment to variables. Assignment was excluded from this class of languages because it can have
nonlocal effects. The result is languages with elegant, clean semantics.
Principle of Lexical Coherence
Sections of code that logically belong together should be physically adjacent in the program. Sections of code that are not related should not be interleaved. It should be easy to tell where one
logical part of the program ends and another starts. A language design is good to the extent that
it permits, requires, or encourages lexical coherence.
This principle concerns only the surface syntax of the language and is, therefore, not as important as the other principles, which concern semantic power. Nonetheless, good human engineering
is important in a language, and lexical coherence is important to make a language usable and
readable.
Poor lexical coherence can be seen in many languages. In Pascal the declarations of local
variables for the main program must be near the top of the program module, and the code for main
must be at the bottom [Exhibit 2.15]. All the function and procedure definitions intervene. In a
program of ordinary size, several pages of code come between the use of a variable in main and its
definition.
Recently, hierarchical editors have been developed for Pascal. They allow the programmer to
hide a function definition under the function header. A program is thus divided into levels,
with the main program at the top level and its subroutines one level lower. If the subroutines have
subroutines, they are at level three, and so on. When the main program is on the screen, only
the top level code appears, and each function definition is replaced by a simple function header.
This brings the main programs body back into the lexical vicinity of its declarations. When the
programmer wishes to look at the function definition, simple editor commands will allow him to
descend to that level and return.
A similar lack of coherence can be seen in early versions of LISP. 3 LISP permits a programmer
to write a function call as a literal function, called a lambda expression, followed by its actual
arguments, as shown at the top of Exhibit 2.16. The dummy parameter names are separated from
the matching parameter values by an arbitrarily long function body.
2
3

Local declarations are explained fully in Chapter 6; parameters are discussed in Chapter 9, Section 9.2.
McCarthy et al. [1962].

2.3. LANGUAGE DESIGN

33

Exhibit 2.15. Poor lexical coherence for declarations and code in Pascal.
The parts of a Pascal program are arranged in the order required to permit one-pass compilation:
Constant declarations.
Type declarations.
Variable declarations.
Procedure and Function declarations.
Code.
Good programming style demands that most of the work of the program be done in subroutines,
and the part of the program devoted to subroutine definitions is often many pages long. The variable
declarations and code for the main program are, therefore, widely separated, producing poor lexical
coherence.

Exhibit 2.16. Syntax for lambda expressions in LISP.

The order of elements in the primitive syntax is:
((lambda
( hlist of dummy parameter namesi )
( hbody of the functioni ))
hlist of actual parameter valuesi)
The order of elements in the extended syntax is:
(let
( hlist of dummy name - actual value pairsi )
( hbody of the functioni ))

34

CHAPTER 2. REPRESENTATION AND ABSTRACTION

Exhibit 2.17. A LISP function call with poor coherence.

The following literal function is written in the primitive LISP syntax. It takes two parameters, x,
and y. It returns their product plus their difference. It is being called with the arguments 3.5 and
a + 2. Note that the parameter declarations and matching arguments are widely separated.
((lambda ( x
(+ (* x
(- x
3.5 (+ a 2)

y )
y)
y) ))
)

This lack of lexical coherence makes it awkward and error prone for a human to match up the
names with the values, as shown in Exhibit 2.17. The eye swims when interpreting this function
call, even though it is simple and the code section is short.
Newer versions of LISP, for example Common LISP, 4 offer an improved syntax with the same
semantics but better lexical coherence. Using the let syntax, dummy parameter names and actual
values are written in pairs at the top, followed by the code. This syntax is shown at the bottom of
Exhibit 2.16, and an example of its use is shown in Exhibit 2.18.
A third, and extreme, example of poor lexical coherence is provided by the syntax for function
definitions in SNOBOL. A SNOBOL IV function is defined by a function header of the following
form:
( 0 hnamei (hparameter listi) hlocal variable name listi 0 , 0 hentry labeli0 )
The code that defines the action of the subroutine can be anywhere within the program module,
and it starts at the line labeled hentry labeli. It does not even need to be all in the same place,
4

Kessler[1988], p. 59.

Exhibit 2.18. A LISP function call with good coherence.

The following function call is written in LISP using the extended let syntax. It is semantically
equivalent to the call in Exhibit 2.17.
(let ((x 3.5) (y (+ a 2) ))
((+ (* x y)
(- x y)) ))
Compare the ease of matching up parameter names and corresponding arguments here, with
the difficulty in Exhibit 2.17. The lexically coherent syntax is clearly better.

35

2.3. LANGUAGE DESIGN

Exhibit 2.19. Poor lexical coherence in SNOBOL.

SNOBOL has such poor lexical coherence that semantically unrelated lines can be interleaved,
and no clear indication exists of the beginning or end of any program segment. This program
converts English to Pig Latin. It is annotated below.
1.
2.
3.
4.
5.
6.
7.
8.
9.

PROC
MAIN
PIG1
LOOP

END

DEFINE(0PIG(X) Y, Z0 , 0 PIG10 )
OUTPUT = PIG(IN)
IN = INPUT
PIG = NULL
X SPAN(0 0 ) =
X BREAK(0 0 ) . Y SPAN(0 0 ) =
Y LEN(1) . Z =
PIG = PIG Y Z 0 AY0
OUTPUT = 0 .0

:(MAIN)
:F(END) S(PROC)
:F(RETURN)
:F(RETURN)
:(LOOP)

Program Notes. The main program begins on line 1, with the declaration of a header for a
subroutine named PIG. Line 1 directs that execution is to continue on the line named MAIN. The
subroutine declaration says that the subroutine PIG has one parameter, X, and two local variables,
Y and Z. The subroutine code starts on the line with the label PIG1.
Lines 2, 3, and 9 belong to the main program. They read a series of messages, translate each
to Pig Latin, write them out, and quit when a zero-length string is entered.
Lines 4 through 8 belong to the subroutine PIG. Line 4 initializes the answer to the null string.
Line 5 strips leading blanks off the parameter, X. Line 6 isolates the next word in X (if any), and
line 7 isolates its first letter. Finally, line 8 glues this word onto the output string with its letters
in a different order and loops back to line 6.

since each of its lines may be attached to the next by a GOTO.

Thus a main program and several subroutines could be interleaved. (We do admit that a
sane programmer would never do such a thing.) Exhibit 2.19 shows a SNOBOL program, with
subroutine, that translates an English sentence into Pig Latin. The line numbers are not part of
the program but are used to key it to the program notes that follow.
Principle of Distinct Representation
Each separate semantic object should be represented by a separate syntactic item. Where a single
syntactic item in the program is used for multiple semantic purposes, conflicts are bound to occur,
and one or both sets of semantics will be compromised. The line numbers in a BASIC program
provide a good example.
BASIC was the very first interactive programming language. It combined an on-line editor, a
file system, and an interpreter to make a language in which simple problems could be programmed

36

CHAPTER 2. REPRESENTATION AND ABSTRACTION

Exhibit 2.20. BASIC: GOTOs and statement ordering both use line numbers.
Line numbers in BASIC are used as targets of GOTO and also to define the proper sequence of the
statements; the editor accepts lines in any order and arranges them by line number. Thus the user
could type the following lines in any order and they would appear as follows:
2
4
6
8

SUM = SUM + A
PRINT SUM
IF A < 10 GO TO 2
STOP

Noticing that some statements have been left out, the programmer sees that three new lines
must be inserted. The shortsighted programmer has only left room to insert one line between each
pair, which is inadequate here, so he or she renumbers the old line 2 as 3 to make space for the
insertion. The result is:
1
2
3
4
5
6
8

LET SUM =
LET A = 1
SUM = SUM
PRINT SUM
LET A = A
IF A < 10
STOP

0
+ A
+ 1
GO TO 2

Notice that the loop formed by line 6 now returns to the wrong line, making an infinite loop.
Languages with separate line numbers and statement labels do not have this problem.

quickly. The inclusion of an editor posed a new problem: how could the programmer modify the
program and insert and delete lines? The answer chosen was to have the programmer number every
line, and have the editor arrange the lines in order by increasing line number.
BASIC was developed in the context of FORTRAN, which uses numeric line numbers as statement
labels. It was, therefore, natural for BASIC to merge the two ideas and use one mechanism, the
monotonically increasing line number, to serve purposes (1) and (2) below. When the language
was extended to include subroutines, symbolic names for them were not defined either. Rather, the
same line numbers were given a third use. Line numbers in BASIC are, therefore, multipurpose:
1. They define the correct order of lines in a program.
2. They are the targets of GOTOs and IFs.
3. They define the entry points of subroutines (the targets of GOSUB).
A conflict happens because inserting code into the program requires that line numbers change,
and GOTO requires that they stay constant. Because of this, adding lines to a program can be a

2.3. LANGUAGE DESIGN

37

complicated process. Normally, BASIC programmers leave regular gaps in the line numbers to allow
for inserting a few lines. However, if the gap in numbering between two successive lines is smaller
than the number of lines to be inserted, something will have to be renumbered. But since the targets
of GOTOs are not marked in any special way, renumbering implies searching the entire program for
GOTOs and GOSUBs that refer to any of the lines whose numbers have been changed. When found,
these numbers must be updated [Exhibit 2.20]. Some BASIC systems provide a renumbering utility,
others dont. In contrast, lines can be added almost anywhere in a C program with minimal local
adjustments.
Principle of Too Much Flexibility
A language feature is bad to the extent that it provides flexibility that is not useful to the programmer, but that is likely to cause syntactic or semantic errors.
For example, any line in a BASIC program can be the target of a GOTO or a GOSUB statement.
An explicit label declaration is not neededthe programmer simply refers to the line numbers used
to enter and edit the program. A careless or typographical error in a GOTO line number will not be
identified as a syntactic error.
Every programmer knows which lines are supposed to be the targets of GOTOs, and she or he
could easily identify or label them. But BASIC supplies no way to restrict GOTOs to the lines that
the programmer knows should be their targets. Thus the translator cannot help the programmer
ensure valid use of labels.
We would say that the ability to GOTO or GOSUB to any line in the program without writing
an explicit label declaration is excessively flexible: it saves the programmer the minor trouble of
declaring labels, but it leads to errors. If there were some way to restrict the set of target lines,
BASIC would be a better and more powerful language. Power comes from a translators ability
to identify and eliminate meaningless commands, as well as from a languages ability to express
aspects of a model.
Another example of useless flexibility can be seen in the way APL handles GOTO and statement
labels. APL provides only three control structures: the function call, sequential execution, and a
GOTO statement. A GOTO can only transfer control locally, within the current function definition.
All other control structures, including ordinary conditionals and loops, must be defined in terms of
the conditional GOTO.
As in BASIC, numeric line numbers are used both to determine the order of lines in a program
and as targets of the GOTO. But the problems in BASIC with insertions and renumbering are avoided
because, unlike BASIC, symbolic labels are supported. A programmer may write a symbolic label
on a line and refer to it in a GOTO, and this will have the correct semantics even if lines are inserted
and renumbering happens. During compilation of a function definition (the process that happens
when you leave the editor), the lines are renumbered. Each label is bound to a constant integer
value: the number of the line on which it is defined. References to the label in the code are replaced
by that constant, which from then on has exactly the same semantics as an integer. (Curiously,
constants are not otherwise supported by the language.)

38

CHAPTER 2. REPRESENTATION AND ABSTRACTION

Exhibit 2.21. Strange but legal GOTOs in APL.

The GOTO is written with a right-pointing arrow, and its target may be any expression. The
statements below are all legal in APL.
(x + 2) 6 Legal so long as the result is an integer.
347
An array of line numbers is given; control will be
transferred to the first.
0
N returns a vector of N numbers, ranging from 1 to
N . Thus, 0 returns a vector of length 0, which is the
null object. A branch to the null object is equivalent
to a no-op.

Semantic problems arise because the labels are translated into integer constants and may be
operated on using integer operations such as multiplication and division! Further, the APL GOTO is
completely unrestricted; it can name either a symbolic label or an integer line number, whether or
not that line number is defined in that subroutine. Use of an undefined line number is equivalent
to a function return. These semantics have been defined so that some interpretation is given no
matter what the result of the expression is [Exhibit 2.21].
Because the target of a GOTO may be computed and may depend on variables, any line of the
subroutine might potentially be its target. It is impossible at compile time to eliminate any line
from the list of potential targets. Thus, at compile time, the behavior of a piece of code may be
totally unpredictable.
APL aficionados love the flexibility of this GOTO. All sorts of permutations and selection may be
done on an array of labels to implement every conceivable variety of conditional branch. Dozens
of useful idioms, or phrases, such as the one in Exhibit 2.22, have been developed using this GOTO
and published for other APL programmers to use.
It is actually fun to work on and develop a new control structure idiom. Many language
designers, though, question the utility and wisdom of permitting and relying on such idiomatic
control structures. They must be deciphered to be understood, and the result of a mistake in
definition or use is a totally wrong and unpredictable branch. Even a simple conditional branch

Exhibit 2.22. A computed GOTO idiom in APL.

(NEG, EQ, POS) [ 2 + N ]
This is a three-way branch very similar to the previous example and analogous to the FORTRAN
arithmetic IF. The signum function, , returns -1 if N is negative, +1 if N is positive, and 0
otherwise. Two is added to the result of signum, and the answer is used to subscript a vector of
labels. One of the three branches is always taken.

2.3. LANGUAGE DESIGN

39

to the top of a loop can be written with four different idioms, all in common use. This makes it
difficult to learn to read someone elses code. Proofs of correctness are practically impossible.
We have shown that APLs totally unrestricted GOTO has the meaningless and useless flexibility
to branch to any line of the program, and that the lack of any other control structure necessitates
the use of cryptic idioms and produces programs with unpredictable behavior. These are severe
semantic defects! By the principle of Too Much Flexibility, this unrestricted GOTO is bad, and APL
would be a more powerful language with some form of restriction on the GOTO.
The Principle of Semantic Power
A programming language is powerful (for some application area) to the extent that it permits the
programmer to write a program easily that expresses the model, the whole model, and nothing but
the model. Thus a powerful language must support explicit communication of the model, possibly
by defining a general object and then specifying restrictions on it. A restriction imposed by the
language can support power at the price of flexibility that might be necessary for some applications.
On the other hand, a restriction imposed by the user expresses only the semantics that the user
wants to achieve and does not limit him or her in ways that obstruct programming.
The programmer should be able to specify a program that computes the correct results and
then be able to verify that it does so. All programs should terminate properly, not crash. Faulty
results from correct data should be provably impossible.
Part of a model is a description of the data that is expected. A powerful language should let the
programmer write data specifications in enough detail so that garbage in is detected and does
not cause garbage out.
The Principle of Portability
A portable program is one that can be compiled by many different compilers and run on different
hardware, and that will work correctly on all of them. If a program is portable, it will be more
useful to more people for more years. We live in times of constant change: we cannot expect to
have the same hardware or operating system available in different places or different years.
But portability limits flexibility. A portable program, by definition, cannot exploit the special
features of some hardware. It cannot rely on any particular bit-level representation of any object or
function; therefore, it cannot manipulate such things. One might want to do so to achieve efficiency
or to write low-level system programs.
Languages such as Standard Pascal that restrict access to pointers and to the bit-representations
of objects force the programmer to write portable code but may prohibit him or her from writing
efficient code for some applications.
Sometimes features are included in a language for historical reasons, even though the language
supports a different and better way to write the same thing. As languages develop, new features are
added that improve on old features. However, the old ones are seldom eliminated because upward
compatibility is important. We want to be able to recompile old programs on new versions of the

40

CHAPTER 2. REPRESENTATION AND ABSTRACTION

Exhibit 2.23. An archaic language feature in FORTRAN.

The arithmetic IF statement was the only conditional statement in the original version of FORTRAN. It is a three-way conditional GOTO based directly on the conditional jump instruction of
the IBM 704. An example of this statement is:
IF (J-1) 21, 76, 76
The expression in parentheses is evaluated first. If the result is negative, control goes to the first
label on the list (21). For a zero value, control goes to the second label (76), and for a positive
value, to the third (76). More often than not, in practical use, two of the labels are the same.
The arithmetic IF has been superseded by the modern block IF statement. Assume that
hblock 1i contains the statements that followed label 21 above, and hblock2i contains the statements
following statement 76. Then the following block IF statement is equivalent to the arithmetic
IF above:
IF J-1 .LT. 0 THEN
hblock 1i
ELSE
hblock 2i
ENDIF

translator. Old languages such as COBOL and FORTRAN have been through the process of change
and restandardization several times. Some features in these languages are completely archaic,
and programmers should be taught not to use them [Exhibit 2.23]. Many of these features have
elements that are inherently error prone, such as reliance on GOTOs. Moreover, they will eventually
be dropped from the language standard. At that point, any programs that use the archaic features
will require extensive modernization before they can be modified in any way.
Our answer to redundant and archaic language features is simple: dont use them. Find out
what constitutes modern style in a language and use it consistently. Clean programming habits
and consistent programming style produce error-free programs faster.
Another kind of redundancy is seen in Pascal, which provides two ways to delimit comments:
(* This is a comment. *) and { This is a comment. } The former way was provided, as part of the
standard, for systems that did not support the full ASCII character set. It will work in all Pascal
implementations and is thus more portable. The latter way, however, is considered more modern
and preferred by many authors. Some programmers use both: one form for comments, the other
to comment out blocks of code.
The language allows both kinds of comment delimiters to be used in a program. However,
mixing the delimiters is a likely source of errors because they are not interchangeable. A comment
must begin and end with the same kind of delimiter. Thus whatever conventions a programmer
chooses should be used consistently. The programmer must choose either the more portable way
or the more modern way, a true dilemma.

2.4. CLASSIFYING LANGUAGES

2.4

41

Classifying Languages

It is tempting to classify languages according to the most prominent feature of the language and to
believe that these features make each language group fundamentally different from other groups.
Such categorizations are always misleading because:
Languages in different categories are fundamentally more alike than they are different. Believing that surface differences are important gets in the way of communication among groups
of language users.
We tend to associate things that occur together in some early example of a language category.
We tend to believe that these things must always come together. This impedes progress in
language design.
Category names are used loosely. Nobody is completely sure what these names mean, and
which languages are or are not in any category.
Languages frequently belong to more than one category. Sorting them into disjoint classes
disguises real similarities among languages with different surface syntax.

2.4.1

Language Families

Students do need to understand commonly used terminology, and it is sometimes useful to discuss
a group of languages having some common property. With this in mind, let us look at some of the
language families that people talk about and try to give brief descriptions of the properties that
characterize each family. As you read this section, remember that these are not absolute, mutually
exclusive categories: categorizations are approximate and families overlap heavily. Examples are
listed for each group, and some languages are named several times.
Interactive Languages. An interactive language is enmeshed in a system that permits easy
alternation between program entry, translation, and execution of code. We say that it operates
using a REW cycle: the system Reads an expression, Evaluates it, and Writes the result on the
terminal, then waits for another input.
Programs in interactive languages are generally structured as a series of fairly short function
and object definitions. Translation happens when the end of a definition is typed in. Programs are
usually translated into some intermediate form, not into native machine code. This intermediate
form is then interpreted. Many interactive languages, such as FORTH and Miranda, use the term
compile to denote the translation of source code into the internal form.
Examples: APL, BASIC, FORTH, LISP, T, Scheme, dBASE, Miranda.
Structured Languages. Control structures are provided that allow one to write programs without using GOTO. Procedures with call-by-value parameters 5 are supported. Note that we call Pascal
5

See Chapter 9.

42

CHAPTER 2. REPRESENTATION AND ABSTRACTION

a structured language even though it contains a GOTO, because it is not necessary to use that GOTO
to write programs.
Examples: Pascal, C, FORTH, LISP, T, Scheme.
Strongly Typed Languages. Objects are named, and each name has a type. Every object
belongs to exactly one type (types are disjoint). The types of actual function arguments are
compared to the declared types of the dummy parameters during compilation. A mismatch in
types or in number of parameters will produce an error comment. Many strongly typed languages,
including Pascal, Ada, and ANSI C, include an escape hatchthat is, some mechanism by which
the normal type-checking process can be evaded.
Examples: FORTRAN 77, Pascal, Ada, ANSI C (but not the original C), ML, Miranda.
Object-oriented Languages. These are extensions or generalizations of the typed languages.
Objects are typed and carry their type identity with them at all times. Any given function may
have several definitions, which we will call methods. 6 Each method operates on a different type of
parameter and is associated with the type of its first parameter. The translator must dispatch each
function call by deciding which defining method to invoke for it. The method associated with the
type of the first parameter will be used, if it exists.
Object-oriented languages have nondisjoint types and function inheritance. The concept of
function inheritance was introduced by Simula and popularized by Smalltalk, the first language to
be called object-oriented. A type may be a subset of another type. The function dispatcher will
use this subset relationship in the dispatching process. It will select a function belonging to the
supertype when none is defined for the subtype.
Actually, many of these characteristics also apply to APL, an old language. It has objects that
carry type tags and functions with multiple definitions and automatic dispatching. It is not a full
object-oriented language because it lacks definable class hierarchies.
Examples: Simula, Smalltalk, T, C++. APL is object-oriented in a restricted sense.
Procedural Languages. A program is an ordered sequence of statements and procedure calls
that will be evaluated sequentially. Statements interact and communicate with each other through
variables. Storing a value in a variable destroys the value that was previously stored there. (This
is called destructive assignment.) Exhibit 2.24 is a diagram of the history of this language family.
Modern procedural languages also contain extensive functional elements. 7
Examples: Pascal, C, Ada, FORTRAN, BASIC, COBOL.
Functional Languages, Old Style. A program is a nested set of expressions and function calls.
Call-by-value parameter binding, not assignment, is the primary mechanism used to give names to
variables. Functions interact and communicate with each other through the parameter stack.
6
7

This is the term used in Smalltalk.

See Chapter 8.

43

2.4. CLASSIFYING LANGUAGES

Exhibit 2.24. The development of procedural languages.

Concepts and areas of concern are listed on the left. Single arrows show how these influenced
language design and how some languages influenced others. Dotted double arrows indicate that a
designer was strongly influenced by the bad features of an earlier language.
1950

Mathematical Notation

1950

Unit Record Equipment

Symbolic Names
Data Specification

Symbolic Assemblers (mid-1950s)

FORTRAN
(1956)

COBOL (1958)

ALGOL-58
MAD (1959)

Structured Control
1960

ALGOL-60

APL
(1962)
Nonalgorithmic
Specification
Interactive Use
Structured Data
Object-Oriented
Programming

RPG (1964)

CPL (1963)

BASIC (1964)

BCPL
(1967)

Simula (1967) PL/1 (1966)

ALGOL-68

1970

B (1970)

Pascal (1973)

Concurrency

1960

1970

C (1972)

Concurrent
Pascal (1975)

Data Abstraction

Modula
CLU
(1977)

1980
Smalltalk

1980

Ada (1982)
True BASIC (1980s)

1985

1985
C++

44

CHAPTER 2. REPRESENTATION AND ABSTRACTION

Exhibit 2.25. The development of functional languages.

LISP, McCarthy (1959)

1960

1960
INTERLISP (1974)
BB&N, Xerox
MAC LISP, MIT
(1968)

1970

Pure Functional Languages

ISWIM: Landin (1966)
LOGO -- recursion and
graphics for children
(early 1970s)

1970

Scheme, MIT (1975)

Lexical Scoping

1980

ML: Milner (1978)

FP: Backus (1978)

T, Yale (1982)
Object Oriented

1990

Franz LISP (1979)

Under Berleley UNIX

Common LISP (1985)

Lexical Scoping

1980

Miranda (1986)
Haskell (1988)
1990

Certain characteristics are commonly associated with functional languages. Most are interactive
and oriented toward the list data structure. Functions are objects that can be freely manipulated,
passed as parameters, composed, and so on. Permitting functions to be manipulated like other
objects gives a language tremendous power. Exhibit 2.25 is a diagram of the development of this
language family.
LISP and its modern lexically scoped descendants support destructive assignment and sequences
of expressions, which are evaluated in order. When these features are used, these languages become
procedural, like Pascal. These languages are, therefore, functional in the same sense that Pascal
is structured. It is never necessary to use the semantically messy GOTO in Pascal. Any semantics
that can be expressed with it can be expressed without it. Similarly, it is not necessary to use the
semantically messy destructive assignment in LISP, but it is used occasionally, to achieve efficiency,
when changing one part of a large data structure.
Examples: LISP, Common LISP, T, Scheme.

2.4. CLASSIFYING LANGUAGES

45

Functional Languages, New Style. Considerable work is now being done on developing functional languages in which sequences of statements, variables, and destructive assignment do not
exist at all. Values are passed from one part of a program to another by function calls and parameter
binding.
There is one fundamental difference between the old and new style functional languages. The
LISP-like languages use call-by-value parameters, and these new languages use call-by-need (lazy
evaluation).8 A parameter is not evaluated until it is needed, and its value is then kept for future
use. Call-by-need is an important semantic development, permitting the use of infinite lists,
which are objects that are part data and part program, where the program part is evaluated, as
needed, to produce the next item on the list.
The terminology used to talk about these new functional programming languages is sometimes
different from traditional programming terminology. A program is an unordered series of static
definitions of objects, types, and functions. In Miranda it isnt even called a program, it is
called a script. Executing a program is replaced by evaluating an expression or reducing
an expression to its normal form. In either case, though, computation happens.
Since pure functional programming is somewhat new, it has not reached its full development
yet. For example, efficient array handling has yet to be included. As the field progresses, we should
find languages that are less oriented to list processing and more appropriate for modeling nonlist
applications.
Examples: ML, Miranda, Haskell.
Parallel Languages. These contain multitasking primitives that permit a program to fork into
two or more asynchronous, communicating tasks that execute some series of computations in parallel. This class of languages is becoming increasingly important as highly parallel hardware develops.
Parallel languages are being developed as extensions of other kinds of languages. One of the
intended uses for them is to program highly parallel machines such as the HyperCube. There
is a great deal of interest in using such machines for massive numeric applications like weather
prediction and image processing. It is not surprising, therefore, that the language developed for
the HyperCube resembled a merger of the established number-oriented languages, FORTRAN and
APL.
There is also strong interest in parallel languages in the artificial intelligence community, where
many researchers are working on neural networks. Using parallelism is natural in such disciplines.
In many situations, a programmer wishes to evaluate several possible courses of action and choose
the first one to reach a goal. Some of the computations may be very long and others short, and
one cant predict which are which. One cannot, therefore, specify an optimal order in which to
evaluate the possibilities. The best way to express this is as a parallel computation: Evaluate
all these computations in parallel, and report to me when the first one terminates. List-oriented
parallel languages will surely develop for these applications.
8

Parameter passing is explained fully in Chapter 9.

46

CHAPTER 2. REPRESENTATION AND ABSTRACTION

Finally, the clean semantics of the assignment-free functional languages are significantly easier to
generalize to parallel execution, and new parallel languages will certainly be developed as extensions
of functional languages.
Examples: Co-Pascal, in a restricted sense. LINDA, OCCAM, FORTRAN-90.
Languages Specialized for Some Application. These languages all contain a complete generalpurpose programming language as their basis and, in addition, contain a set of specialized primitives
designed to make it convenient to process some particular data structure or problem area. Most
contain some sophisticated and powerful higher-level commands that would require great skill and
long labor to program in an unspecialized language like Pascal. An example is dBASE III which
contains a full programming language similar to BASIC and, in addition, powerful screen handling
and file management routines. The former expedites entry and display of information, the latter supports a complex indexed file structure in which key fields can be used to relate records in
different files.
Systems programming languages must contain primitives that let the programmer manipulate the bits and bytes of the underlying machine and should be heavily standardized and
widely available so that systems, once implemented, can be easily ported to other machines.
Examples: C, FORTH.
Business data processing languages must contain primitives that give fine and easy control over details of input, output, file handling, and precision of numbers. The standard floatingpoint representations are not adequate to provide this control, and some form of fixed-point numeric
representation must be provided. The kind of printer or screen output formatting provided in FORTRAN, C, and Pascal is too clumsy and does not provide enough flexibility. A better syntax and
more options must be provided. Similarly, a modern language for business data processing must
have a good facility for defining screens for interactive input. A major proportion of these languages
is devoted to I/O.
Higher-level commands should be included for common tasks such as table handling and sorting.
Finally, the language should provide good support for file handling, including primitives for handling
sequential, indexed, and random access files.
Examples: RPG (limited to sequential files), COBOL, Ada.
Data base languages contain extensive subsystems for handling internal files, and relationships among files. Note that this is quite independent of a good subsystem for screen and printer
I/O.
Examples: dBASE, Framework, Structured Query Language (SQL).
List processing languages contain primitive definitions for a linked list data type and the
important basic operations on lists. This structure has proven to be useful for artificial intelligence

2.4. CLASSIFYING LANGUAGES

47

programming.
Implementations must contain powerful operations for direct input and output of lists, routines
for allocation of dynamic heap storage, and a garbage collection routine for recovery of dynamically
allocated storage that is no longer accessible.
Examples: LISP, T, Scheme, Miranda.
Logic languages are interactive languages that use symbolic logic and set theory to model
computation. Prolog was the first logic language and is still the best known. Its dominant characteristics define the language class. A Prolog program is a series of statements about logical
relations that are used to establish a data base, interspersed with statements that query this data
base. To evaluate a query, Prolog searches that data base for any entries that satisfy all the constraints in the query. To do this, the translator invokes an elaborate expression evaluator which
performs an exhaustive search of the data base, with backtracking. Rules of logical deduction are
built into the evaluator.
Thus we can classify a logic language as an interactive data base language where both operations
and the data base itself are highly specialized for dealing with the language of symbolic logic and
set theory. Prolog is of particular interest in the artificial intelligence community, where deductive
reasoning on the basis of a set of known facts is basic to many undertakings.
Examples: HASL, FUNLOG, Templog (for temporal logic), Uniform (unifies LISP and Prolog),
Fresh (combines the functional language approach with logic programming), etc.
Array processing languages contain primitives for constructing and manipulating arrays
and matrices. Sophisticated control structures are built in for mapping simple operations onto
arrays, for composing and decomposing arrays, and for operating on whole arrays.
Examples: APL, APL-2, VisiCalc, and Lotus.
String processing languages contain primitives for input, output, and processing of character strings. Operations include searching for and extracting substrings specified by complex
patterns involving string functions. Pattern matching is a powerful higher-level operation that
may involve exhaustive search by backtracking. The well-known string processing languages are
SNOBOL and its modern descendant, ICON.
Typesetting languages were developed because computer typesetting is becoming an economically important task. Technical papers, books, and drawings are, increasingly, prepared for
print using a computer language. A document prepared in such a language is an unreadable mixture of commands and ordinary text. The commands handle files, set type fonts, position material,
and control indexing, footnotes, and glossaries. Drawings are specified in a language of their own,
then integrated with text. The entire finished product is output in a language that a laser printer
can handle. This book was prepared using the languages mentioned below, and a drafting package
named Easydraw whose output was converted to Postscript.

48

CHAPTER 2. REPRESENTATION AND ABSTRACTION

Examples: Postscript, TeX, LaTeX.

Command languages are little languages frequently created by extending a systems user
interface. First simple commands are provided; these are extended by permitting arguments and
variations. More useful commands are added. In many cases these command interfaces develop
their own syntax (usually ad hoc and fairly primitive) and truly extensive capabilities. For example,
entire books have been written about UNIX shell programming. Every UNIX system includes one
or several shells which accept, parse, and interpret commands. From these shells, the user may
call system utilities and other small systems such as grep, make, and flex. Each one has its own
syntax, switches, semantics, and defaults.
Command languages tend to be arcane. In many cases, little design effort goes into them
because their creators view them as simple interfaces, not as languages.
Fourth-generation Languages . This curious name was applied to diverse systems developed in the mid-1980s. Their common property was that they all contained some powerful new
control structures, statements, or functions by which you could invoke in a few words some useful
action that would take many lines to program in a language like Pascal. These languages were
considered, therefore, to be especially easy to learn and user friendly, and the natural accompaniments to fourth-generation hardware, or personal computers.
Lotus 1-2-3 and SuperCalc are good examples of fourth-generation languages. They contain a
long list of commands that are very useful for creating, editing, printing, and extracting information
from a two-dimensional data base called a spreadsheet, and subsystems for creating several kinds
of graphs from that data.
HyperCard is a data base system in which it is said that you can write complex applications
without writing a line of code. You construct the application with the mouse, not with the keyboard.
The designers of many fourth-generation languages viewed them as replacements for programming languages, not as new programming languages. The result is that their designs did not really
profit as much as they could have from thirty years of experience in language design. Like COBOL
and FORTRAN, these languages are ad hoc collections of useful operations.
The data base languages such as dBASE are also called fourth-generation languages, and
again their designers thought of them as replacements for computer languages. Unfortunately, these
languages do not eliminate the need for programming. Even with lots of special report-generating
features built in, users often want something a little different from the features provided. This
implies a need for a general-purpose language within the fourth-generation system in which users
can define their own routines. The general-purpose language included in dBASE is primitive and
lacks important control structures. Until the newest version, dBASE4, procedures did not even
have parameters, and when they were finally added, the implementation was unusual and clumsy.
The moral is that there is no free lunch. An adaptable system must contain a general-purpose
language to cover applications not supported by predefined features. The whole system will be
better if this general-purpose language is carefully designed.

2.4. CLASSIFYING LANGUAGES

2.4.2

49

Languages Are More Alike than Different

Viewing languages as belonging to language families tends to make us forget how similar all
languages are. This basic similarity happens because the purpose of all languages is to communicate
models from human to machine. All languages are influenced by the innate abilities and weaknesses
of human beings, and are constrained by the computers inability to handle irreducible ambiguity.
Most of the differences among languages arise from the specialized nature of the objects and tasks
to be communicated using a given language.
This book is not about any particular family of languages. It is primarily about the concepts
and mechanisms that underlie the design and implementation of all languages, and only secondarily
about the features that distinguish one family from another. Most of all, it is not about the myriad
variations in syntax used to represent the same semantics in different languages. The reader is
asked to try to forget syntax and focus on the underlying elements.

Exercises
1. What are the two ways to view a program?
2. How will languages supporting these views differ?
3. What is a computer representation of an object? A process?
4. Define semantic intent. Define semantic validity. What is their importance?
5. What is the difference between explicit and implicit representation? What are the implications
of each?
6. What is the difference between coherent and diffuse representation?
7. What are the advantages of coherent representation?
8. How can language design goals conflict? How can the designer resolve this problem?
9. How can restrictions imposed by the language designer both aid and hamper the programmer?
10. Why is the concept of locality of effect so important in programming language design?
11. What are the dangers involved when using global variables?
12. What is lexical coherence? Give an example of poor lexical coherence.
13. What is portability? Why does it limit flexibility?
14. Why is it difficult to classify languages according to their most salient characteristics?

50

CHAPTER 2. REPRESENTATION AND ABSTRACTION

15. What is a structured language? Strongly typed language? Object-oriented language? Parallel
language? Fourth-generation language?
16. Why are most languages more similar than they are different? From what causes do language
differences arise?
17. Discuss two aspects of a language design that make it hard to read, write, or use. Give an
example of each, drawn from a language with which you are familiar.
18. Choose three languages from the following list: Smalltalk, BASIC, APL, LISP, C, Pascal, Ada.
Describe one feature of each that causes some people to defend it as the best language for
some application. Choose features that are unusual and do not occur in many languages.

Chapter 3

Elements of Language

Overview
This chapter presents elements of language, drawing correlations between English parts
of speech and words in programming languages. Metalanguages allow languages to describe themselves. Basic structural units, words, sentences, paragraphs, and references,
are analogous to the lexical tokens, statements, scope, and comments of programming
languages.

Languages are made of words with their definitions, rules for combining the words into meaningful larger units, and metawords (words for referring to parts of the language). In this section
we examine how this is true both of English and of a variety of programming languages.

3.1
3.1.1

The Parts of Speech

Nouns

In natural languages nouns give us the ability to refer to objects. People invent names for objects so
that they may catalog them and communicate information about them. Likewise, names are used
for these purposes in programming languages, where they are given to program objects (functions,
memory locations, etc.). A variable declaration is a directive to a translator to set aside storage
to represent some real-world object, then give a name to that storage so that it may be accessed.
Names can also be given to constants, functions, and types in most languages.

51

52

CHAPTER 3. ELEMENTS OF LANGUAGE

First-Class Objects
One of the major trends throughout the thirty-five years of language design has been to strengthen
and broaden the concept of object. In the beginning, programmers dealt directly with machine
locations. Symbolic assemblers introduced the idea that these locations represented real-world data,
and could be named. Originally, each object had a name and corresponded to one storage location.
When arrays were introduced in FORTRAN and records in COBOL, these aggregates were viewed
as collections of objects, not as objects themselves.
Several years and several languages later, arrays and records began to achieve the status of firstclass objects that could be manipulated and processed as whole units. Languages from the early
seventies, such as Pascal and C, waffled on this point, permitting some whole-object operations on
aggregate objects but prohibiting others. Modern languages support aggregate-objects and permit
them to be constructed, initialized, assigned to, compared, passed as arguments, and returned as
results with the same ease as simple objects.
More recently, the functional object, that is, an executable piece of code, has begun to achieve
first-class status in some languages, which are known as functional languages. The type object
has been the last kind of object to achieve first-class status. A type-object describes the type of
other objects and is essential in a language that supports generic code.
Naming Objects
One of the complex aspects of programming languages that we will study in Chapter 6 involves the
correspondence of names to objects. There is considerable variation among languages in the ways
that names are used. In various languages a name can:
Exist without being attached, or bound, to an object (LISP).
Be bound simultaneously to different objects in different scopes (ALGOL, Pascal).
Be bound to different types of objects at different times (APL, LISP).
Be bound, through a pointer, to an object that no longer exists (C).
Conversely, in most languages, a single object can be bound to more than one name at a time,
producing an alias. This occurs when a formal parameter name is bound to an actual parameter
during a function call.
Finally, in many languages, the storage allocated for different objects and bound to different
names can overlap. Two different list heads may share the same tail section [Exhibit 3.1].

3.1.2

Pronouns: Pointers

Pronouns in natural languages correspond roughly to pointers in programming languages. Both

are used to refer to different objects (nouns) at different times, and both must be bound to (defined
to refer to) some object before becoming meaningful. The most important use of pointers in

53

3.1. THE PARTS OF SPEECH

Exhibit 3.1. Two overlapping objects (linked lists).

List1:

The

Only

List2:

Your

Time

Direction

From

Here

Is

Up.

programming languages is to label objects that are dynamically created. Because the number of
these objects is not known to the programmer before execution time, he cannot provide names for
them all, and pointers become the only way to reference them.
When a pointer is bound to an object, the address of that object is stored in space allocated
for the pointer, and the pointer refers indirectly to that object. This leads to the possibility that
the pointer might refer to an object that has died, or ceased to exist. Such a pointer is called
a dangling reference. Using a dangling reference is a programming error and must be guarded
against in some languages (e.g., C). In other languages (e.g., Pascal) this problem is minimized by
imposing severe restrictions on the use of pointers. (Dangling references are covered in Section
6.3.2.)

3.1.3

Adjectives: Data Types

In English, adjectives describe the size, shape, and general character of objects. They correspond,
in a programming language, to the many data type attributes that can be associated with an object
by a declaration or by a default. In some languages, a single attribute is declared that embodies
a set of properties including specifications for size, structure, and encoding [Exhibit 3.2]. In other
languages, these properties are independent and are listed separately, either in variable declarations
(as in COBOL) or in type declarations, as in Ada [Exhibit 3.3].
Some of the newer languages permit the programmer to define types that are related hierarchically in a tree structure. Each class of objects in the tree has well-defined properties. Each subclass
has properties of its own and also inherits all the properties of the classes above it in the hierarchy.
Exhibit 3.4 gives an example of such a type hierarchy in English. The root of this hierarchy is
the class vertebrate, which is characterized by having a backbone. All subclasses inherit this

Exhibit 3.2. Size and encoding bundled in C.

The line below declares a number that will be represented in the computer using floating-point
encoding. The actual number of bytes allocated is usually four, and the precision is approximately
seven digits. This declaration is the closest parallel in C to the Ada declaration in Exhibit 3.3.
float price;

54

CHAPTER 3. ELEMENTS OF LANGUAGE

Exhibit 3.3. Size and encoding specified separately in Ada.

The Ada declarations below create a new type named REAL and a REAL object, price.
The use of the keyword digits indicates that this type is to be derived from some predefined
type with floating-point encoding.
The number seven indicates that the resulting type must have at least seven decimal digits
of precision.
type REAL is digits 7;
price: REAL;

property. At the next level are birds, which have feathers, and mammals, which have hair. We
can, therefore, conclude that robins and chickens are feathered creatures, and that human beings
are hairy. Going down the tree, we see that roosters and hens inherit all properties of chickens,
including being good to eat. According to the tree, adults and children are both human (although
members of each subclass sometimes dispute this). Finally, at the leaf level, both male and female
subclasses exist, which inherit the properties of either adults or children.
Inheritance means that any function defined for a superclass also applies to all subclasses.
Thus if we know that constitutional rights are guaranteed for human beings, we can conclude that
girls have these rights.
Using an object-oriented language such as Smalltalk a programmer can implement types (classes)
with exactly this kind of hierarchical inheritance of type properties. (Chapter 17 deals with this
topic more fully.)

Exhibit 3.4. A type hierarchy with inheritance in English.

Vertebrate
Bird
Robin
Rooster

Chicken
Hen

Sheep
Ram

Mammal
Human Being
Adult
Child

Ewe
Man

Woman Boy

Girl

3.1. THE PARTS OF SPEECH

3.1.4

55

Verbs

In English, verbs are words for actions or states of being. Similarly, in programming languages,
we see action words such as RETURN, BREAK, STOP, GOTO, and :=. Procedure calls, function calls,
and arithmetic operators all direct that some action should happen, and are like action verbs.
Relational operators (=, >, etc.) denote states of beinghey ask questions about the state of some
program object or objects.
In semistandard terminology, a function is a program object that receives information through
a list of arguments, performs a prescribed computation on that information, calculates some answer, and returns that value to the calling program. In most languages function calls can be
embedded within the argument lists of other function calls, and within arithmetic expressions.
Function calls are usually denoted by writing the function name followed by an appropriate series
of arguments enclosed in parentheses. Expressions often contain more than one function call. In
this case each language defines (or explicitly leaves undefined) the order in which the calls will be
executed.1
A procedure is just like a function except that it does not return a value. Because no value
results from executing the procedure, the procedure call constitutes an entire program statement
and cannot be embedded in an expression or in the argument list of another call.
An operator is a predefined function whose name is often a special symbol such as +. Most
operators require either one or two arguments, which are called operands. Many languages support
infix notation for operators, in which the operator symbol is written between its two operands or
before or after its single operand. Rules of precedence and associativity [Chapter 8, Section 8.3.2]
govern the way that infix expressions are parsed, and parentheses are used, when necessary, to
modify the action of these rules.
We will use the word function as a generic word to refer to functions, operators, and procedures
when the distinctions among them are not important.
Some languages (e.g., FORTRAN, Pascal, and Ada) provide three different syntactic forms for
operators, functions, and procedures [Exhibit 3.5]. Other languages (e.g., LISP and APL) provide
only one [Exhibits 3.6 and 3.7]. To a great extent, this makes languages appear to be more
different in structure than they are. The first impression of a programmer upon seeing his or her
first LISP program is that LISP is full of parentheses, is cryptic, and has little in common with other
languages. Actually, various front ends, or preprocessors, have been written for LISP that permit
the programmer to write using a syntax that resembles ALGOL or Pascal. This kind of preprocessor
changes only cosmetic aspects of the language syntax. It does not add power or supply kinds of
statements that do not already exist. The LISP preprocessors do demonstrate that LISP and ALGOL
have very similar semantic capabilities.
1

This issue is discussed in Chapter 8.

56

CHAPTER 3. ELEMENTS OF LANGUAGE

Exhibit 3.5. Syntax for verbs in Pascal and Ada.

These languages, like most ALGOL-like languages, have three kinds of verbs, with distinct ways of
invoking each.
Functions: The function name is written followed by parentheses enclosing the list of arguments.
Arguments may themselves be function calls. The call must be embedded in some larger
statement, such as an assignment statement or procedure call. This is a call to a function
named push with an embedded call on the sin function.
Success := push(rs, sin(x));
Procedures: A procedure call constitutes an entire program statement. The procedure name is
written followed by parentheses enclosing the list of arguments, which may be function calls.
This is a call on Pascals output procedure, with an embedded function call:
Writeln (Success, sin(x));
Operators: An operator is written between its operands, and several operators may be combined
to form an expression. Operator-expressions are legal in any context in which function calls
are legal.
Success := push(rs, (x+y)/(x-y));

Exhibit 3.6. Syntax for verbs in LISP.

LISP has only one class of verb: functions. There are no procedures in LISP, as all functions return
a value. In a function call, the function name and the arguments are enclosed in parentheses (first
line below). Arithmetic operators are also written as function calls (second line).
(myfun arg1 arg2 arg3)
(+ B 1)

3.1. THE PARTS OF SPEECH

57

Exhibit 3.7. Syntax for verbs in APL.

APL provides a syntax for applying operators but not for function calls or procedure calls. Operators
come in three varieties: dyadic (having two arguments), monadic (having one argument), and niladic
(having no arguments).
Dyadic operators are written between their operands. Line [1] below shows + being used
to add the vector (5 3) to B and add that result to A. (APL expressions are evaluated rightto-left.) Variables A and B might be scalars or length two vectors. The result is a length two
vector.
Monadic operators are written to the left of their operands. Line [2] shows the monadic
operator |, or absolute value.
Line [3] shows a call on a niladic operator, the read-input operator, 2. The value read is
stored in A.
[1] A + 5 3 + B
[2] | A
[3] A 2
The programmer may define new functions but may not use more than two arguments for those
functions. Function calls are written using the syntax for operators. Thus a dyadic programmerdefined function named FUN would be called by writing:
A FUN B
When a function requires more than two arguments, they must be packed or encoded into two
bunches, sent to the function, then unpacked or decoded within the function. This is awkward and
not very elegant.

The Domain of a Verb

The definition of a verb in English always includes an indication of the domain of the verb, that is,
the nouns with which that verb can meaningfully be used. A dictionary provides this information,
either implicitly or explicitly, as part of the definition of each verb [Exhibit 3.8].
Similarly, the domain of a programming language verb is normally specified when it is defined.
This specification is part of the program in some languages, part of the documentation in others.
The domain of a function is defined in most languages by a function header, which is part of the
function definition. A header specifies the number of the objects required for the function to operate
and the formal names by which those parameters will be known. Languages that implement strong
typing also require the types of the parameters to be specified in the header. This information is
used to ensure that the function is applied meaningfully, to objects of the correct types [Exhibit
3.9].

58

CHAPTER 3. ELEMENTS OF LANGUAGE

Exhibit 3.8. The domain of a verb in English.

Verb: Cry
Definition for the verb cry,

paraphrased from the dictionary. a

1. To make inarticulate sobbing sounds expressing grief, sorrow, or pain.

2. To weep, or shed tears.
3. To shout, or shout out.
4. To utter a characteristic sound or call (used of an animal).
The domain is defined in definitions (1) through (3) by stating that the object/creature that cries
must be able to sob, express feelings, weep, or shout. Definition (4) explicitly states that the
domain is an animal. Thus all of the following things can cry: human beings (by definitions 1,
2, 3), geese (4), and baby dolls (2).
a

Morris [1969], p. 319.

The range of a function is the set of objects that may be the result of that function. This must
also be specified in the function header (as in Pascal) or by default (as in C) in languages that
implement type checking.

3.1.5

Prepositions and Conjunctions

In English we distinguish among the parts of speech used to denote time, position, conditionals, and
the relationship of phrases in a sentence. Each programming language contains a small number
of such words, used analogously to delimit phrases and denote choices and repetition (WHILE,
ELSE, BY, CASE, etc.). The exact words differ from language to language. Grammatical rules
state how these words may be combined with phrases and statements to form meaningful units.

Exhibit 3.9. The domains of some Pascal functions.

Predefined functions
chr
ord
trunc
A user-defined function header

Domains
An integer between 0 and 127.
Any object of an enumerated type.
A real number.

: FUNCTION search (N:name; L:list ):

list;

The domain of search is pairs of objects, one of type name, the other of type list. The
result of search is a list; its range is, therefore, the type list.

3.2. THE METALANGUAGE

59

By themselves these words have little meaning, and we will deal with them in Chapter 10, where
we examine control structures.

3.2

The Metalanguage

A language needs ways to denote its structural units and to refer to its own parts. English has
sentences, paragraphs, essays, and the like, each with lexical conventions that identify the unit and
mark its beginning and end. Natural languages are also able to refer to these units and to the
words that comprise the language, as in phrases such as the paragraph below, and USA is an
abbreviation for the United States of America. These parts of a language that permit it to talk
about itself are called a metalanguage. The metalanguage that accompanies most programming
languages consists of an assortment of syntactic delimiters, metawords, and ways to refer to structural units. We consider definitions of the basic structural units to be part of the metalanguage
also.

3.2.1

Words: Lexical Tokens

The smallest unit of any written language is the lexical tokenthe mark or series of marks that
denote one symbol or word in the language. To understand a communication, first the tokens must
be identified, then each one and their overall arrangement must be interpreted to arrive at the
meaning of the communication. Analogously, one must separate the sounds of a spoken sentence
into tokens before it can be comprehended. Sometimes it is a nontrivial task to separate the string
of written marks or spoken sounds into tokens, as anyone knows who has spent a day in a foreign
country.
This same process must be applied to computer programs. A human reader or a compiler
must first perform a lexical analysis of the code before beginning to understand the meaning. The
portion of the compiler that does this task is called the lexer.
In some languages lexical analysis is trivially simple. This is true in FORTH, which requires
every lexical token to be delimited (separated from every other token) by one or more spaces.
Assembly languages frequently define fixed columns for operation codes and require operands to
be separated by commas. Operating system command shells usually call for the use of spaces and
a half dozen punctuation marks which are tokens themselves and also delimit other tokens. Such
simple languages are easy to lexically analyze, or lex. Not all programming languages are so simple,
though, and we will examine the common lexical conventions and their effects on language.
The lexical rules of most languages define the lexical forms for a variety of token types:
Names (predefined and user-defined)
Special symbols
Numeric literals
Single-character literals

60

CHAPTER 3. ELEMENTS OF LANGUAGE

Multiple-character string literals

These rules are stated as part of the formal definition of every programming language. A lexer for a
language is commonly produced by feeding these rules to a program called a lexer generator, whose
output is a program (the lexer) that can perform lexical analysis on a source text string according
to the given rules. The lexer is the first phase of a compiler. Its role in the compiling process is
illustrated in Exhibit 4.3.
Much of the feeling and appearance of a language is a side effect of the rules for forming tokens.
The most common rules for delimiting tokens are stated below. They reflect the rules of Pascal, C,
and Ada.
Special symbols are characters or character strings that are nonalphabetic and nonnumeric.
Examples are ;, +, and :=. They are all predefined by the language syntax. No new
special symbols may be defined by the programmer.
Names must start with an alphabetic character and must not contain anything except letters,
digits, and (sometimes) the _ symbol.
Everything that starts with a letter is a name.
Names end with a space or a special symbol.
Special symbols generally alternate with names and literals. Where two special symbols or
two names are adjacent, they must be separated by a space.
Numeric literals start with a digit, a +, or a . They may contain digits, ., and E
(for exponent). Any other character ends the literal.
Single-character literals and multiple-character strings are enclosed in matching single or
double quotes. If, as in C, a single character has different semantics from a string of length
1, then single quotes may be used to delimit one and double quotes used for the other.
Note that spaces are used to delimit some but not all tokens. This permits the programmer
to write arithmetic expressions such as a*(b+c)/d the way a mathematician would write them.
If we insisted on a delimiter (such as a space) after every token, the expression would have to
be written a * ( b + c ) / d , which most programmers would consider to be onerous and
unnatural.
Spaces are required to delimit arithmetic operators in COBOL. The above expression in COBOL
would be written a * (b + c) / d. This awkwardness is one of the reasons that programmers
are uncomfortable using COBOL for numeric applications. The reason for this requirement is that
the - character is ambiguous: COBOLs lexical rules permit - to be used as a hyphen in
variable names, for example, hourly-rate-in. Long, descriptive variable names greatly enhance
the readability of programs.

3.2. THE METALANGUAGE

61

Hyphenated variable names have existed in COBOL from the beginning. When COBOL was
extended at a later date to permit the use of arithmetic expressions an ambiguity arose: the hyphen
character and the subtraction operator were the same character. One way to avoid this problem is
to use different characters for the two purposes. Modern languages use the - for subtraction and
the underbar, _, which has no other function in the language, to achieve readability.
As you can see, the rules for delimiting tokens can be complex, and they do have varied repercussions. The three important issues here are:
Code should be readable.
The language must be translatable and, preferably, easy to lex.
It is preferable to use the same conventions as are used in English and/or mathematical
notation.
The examples given show that a familiar, readable language may contain an ambiguous use of
symbols. A few language designers have chosen to sacrifice familiarity and readability altogether
in order to achieve lexical simplicity. LISP, APL, and FORTH all have simpler lexical and syntactic
rules, and all are considered unreadable by some programmers because of the conflict between their
prior experience and the lexical and syntactic forms of the language.
Let us examine the simple lexical rule in FORTH and its effects. In other languages the decision
was made to permit arithmetic expressions to be written without delimiters between the variable
names and the operators. A direct consequence is that special symbols (nonalphabetic, nonnumeric,
and nonunderbar) must be prohibited in variable names. It may seem natural to prohibit the use
of characters like + and ( in a name, but it is not at all necessary.
FORTH requires one or more space characters or carriage returns between every pair of tokens,
and because of this rule, it can permit special characters to be used in identifiers. It makes no
distinction between user-defined names and predefined tokens: either may contain any character
that can be typed and displayed. The string #$% could be used as a variable or function name if
the programmer so desired. The token ab* could never be confused with an arithmetic problem
because the corresponding arithmetic problem, a b * , contains three tokens separated by spaces.
Thus the programmer, having a much larger alphabet to use, is far freer to invent brief, meaningful
names. For example, one might use a+ to name a function that increments its argument (a
variable) by the value of a.
Lexical analysis is trivially easy in FORTH. Since its lexical rules treat all printing characters the
same way and do not distinguish between alphabetic characters and punctuation marks, FORTH
needs only three classes of lexical tokens:
Names (predefined or user-defined).
Numeric literals.
String literals. These can appear only after the string output command, which is ." (pronounced dot-quote). A string literal is terminated by the next " (pronounced quote).

62

CHAPTER 3. ELEMENTS OF LANGUAGE

These three token types correspond to semantically distinct classes of objects that the interpreter
handles in distinct ways. Names are to be looked up in the dictionary and executed. Numeric
literals are to be converted to binary and put on the stack. String literals are to be copied to the
output stream. The lexical rules of the language thus correspond directly to its semantics, and the
interpreter is very short and simple.
The effect of these lexical rules on people should also be noted. Although the rules are simple
and easy to learn, a programmer accustomed to the conventions in other languages has a hard time
learning to treat the space character as important.

3.2.2

Sentences: Statements

The earliest high-level languages reflected the linguistic idea of sentences: a FORTRAN or COBOL
program is a series of sentencelike statements. 2 COBOL statements even end in periods. Most
statements, like sentences, specify an action to perform and some object or objects on which to
perform the action. A language is called procedural, if a program is a sequence statements,
grouped into procedures, to be carried out using the objects specified.
In the late 1950s when FORTRAN and COBOL were developed, the punched card was the
dominant medium for communication from human to computer. Programs, commands to the
operating system, and data were all punched on cards. To compile and (one hoped) run a program,
the programmer constructed a deck usually consisting of:
An ID control card, specifying time limits for the compilation and run. 3
A control card requesting compilation, an object program listing, an error listing, and a
memory map.4
A series of cards containing the program.
A control card requesting loading and linking of the object program.
A control card requesting a run and a core dump 5 of the executable program.
2

Caution: In a discussion of formal grammars and parsing, the term sentence is often used to mean the entire
program, not just one statement.
3
Historical note: Some of the items on the control cards are hard to understand in todays environment. Limiting
the time that a job would be allowed to run (using a job time limit) was important then because computer time was
very costly. In 1962, time on the IBM 704 (a machine comparable in power to a Commodore 64) cost $600 per hour
at the University of Michigan. For comparison, Porterhouse steak cost about $1 per pound. Short time limits were
specified so that infinite loops would be terminated by the system as soon as possible.
4
The memory map listed all variable names and their memory addresses. The map, object listing, and core
(memory) dump together were indispensable aides to debugging. They permitted the programmer to reconstruct the
execution of the program manually.
5
Most debugging was done in those days by carefully analyzing the contents of a core (memory) dump. The kind
of trial and error debugging that we use today was impractical because turnaround time for a trial run was rarely less
than a few hours and sometimes was measured in days. In order to glean as much information as possible from each
run, the programmer would analyze the core dump using the memory maps produced by the compiler and linker.

63

3.2. THE METALANGUAGE

Exhibit 3.10. Field definitions in FORTRAN.

Columns
1
15
6
772
7380

Use in a FORTRAN program

A C or a * here indicates a comment line.
Statement labels
Statement continuation mark
The statement itself
Programmer ID and sequence numbers.

End of statement. At end of line, unless column 6 on the next card is punched to indicate a
continuation of the statement.
Indenting convention.

Start every statement in column 7. (Indenting is not generally used.)

A control card marking the beginning of the data.

A series of cards containing the data.
A JOB END control card marking the end of the deck.
Control cards had a special character in column 1 by which they could be recognized. Because
a deck of cards could easily become disordered by being dropped, columns 73 through 80 were
conventionally reserved for identification and sequence numbers. The JOB END card had a different
special mark. This made it easy for the operating system to abort remaining segments of a job
after a fatal error was discovered during compilation or linking.
Because punched cards were used for programs as well as data, the physical characteristics of
the card strongly influenced certain aspects of the early languages. The program statement, which
was the natural program unit, became tightly associated with the 80-column card, which was the
natural media unit. Many programmers wrote their code on printed coding forms, which looked
like graph paper with darker lines marking the fields. This helped keypunch operators type things
in the correct columns.
The designers of the FORTRAN language felt that most FORTRAN statements would fit on one
line and so chose to require that each statement be on a separate card. The occasional statement
that was too long to fit could be continued on another card by placing a character in column six of
the second card. Exhibit 3.10 lists the fields that were defined for a FORTRAN card.
COBOL was also an early fixed-format language, with similar but different fixed fields. Due to
the much longer variable names permitted in COBOL, and the wordier and more complex syntax,
The programmer would trace execution of the program step by step and compare the actual contents of each memory
location to what was supposed to be there. Needless to say, this was slow, difficult, and beyond the capabilities of
many people. Modern advances have made computing much more accessible.

64

CHAPTER 3. ELEMENTS OF LANGUAGE

many statements would not fit on one line. A convention that imitated English was introduced:
the end of each statement was marked by a period. A group of statements that would be executed
sequentially was called a paragraph, and each paragraph was given an alphanumeric label. Within
columns 1372, indenting was commonly used to clarify the meaning of the statements.
Two inventions in the late 1960s combined to make the use of punched cards for programs obsolete. The remote-access terminal and the on-line, disk-based file system made it both unnecessary
and impractical to use punched cards. Languages that were designed after this I/O revolution reflect the changes in the equipment used. Fixed fields disappeared, the use of indentation to clarify
program structure became universal, and a character such as ; was used to separate statements
or terminate each statement.6

3.2.3

Larger Program Units: Scope

English prepositions and conjunctions commonly control a single phrase or clause. When a larger
scope of influence is needed in English, we indicate that the word pertains to a paragraph. In programming languages, units that correspond to such paragraphs are called scopes and are commonly
marked by a pair of matched opening and closing marks. Exhibits 3.11, 3.13, and 3.15 show the
tremendous variety of indicators used to mark the beginning and end of a scope.
In FORTRAN the concept of scope was not well abstracted, and scopes were indicated in a
variety of ways, depending on the context. As new statement types were added to the language
over the years, new ways were introduced to indicate their scopes. FORTRAN uses five distinct
ways to delimit the scopes of the DATA statement, DO loop, implied DO loop, logical IF (true
action only), and block IF (true and false actions) [Exhibit 3.11]. This nonuniformity of syntax
does not occur in the newer languages.
Two different kinds of ways to end scopes are shown in Exhibit 3.11. The labeled statement
at the end of a DO scope ends a specific DO. Each DO statement specifies the statement label of the
line which terminates its scope. (Two DOs are allowed to name the same label, but that is not
relevant here.) We say that DO has a labeled scope. In contrast, all block IF statements are ended
by identical ENDIF lines. Thus an ENDIF could end any block IF statement. We say that block
IF statements have unlabeled scopes.
The rules of FORTRAN do not permit either DO scopes or block IF scopes to overlap partially.
That is, if the beginning of one of these scopes, say B, comes between the beginning and end of
another scope, say A, then the end of scope B must come before the end of scope A. Legal and
illegal nestings of labeled scopes are shown in Exhibit 3.12.
All languages designed since 1965 embody the abstraction scope. That is, the language
supplies a single way to delimit a paragraph, and that way is used uniformly wherever a scope is
needed in the syntax, for example, with THEN, ELSE, WHILE, DO, and so on. For many languages,
this is accomplished by having a single pair of symbols for begin-scope and end-scope, which are
6

Most languages did not use the . as a statement-mark because periods are used for several other purposes
(decimal points and record part selection), and any syntax becomes hard to translate when symbols become heavily
ambiguous.

65

3.2. THE METALANGUAGE

Exhibit 3.11. Scope delimiters in FORTRAN.

The following program contains an example of each linguistic unit that has an associated scope.
The line numbers at the left key the statements to the descriptions, in the table that follows, of
the scope indicators used.
1
INTEGER A, B, C(20), I
2
DATA A, B /31, 42/
3
READ* A, B, ( C(I), I=1,10)
4
DO 80 I= 1, 10
5
IF (C(I) .LT. 0) C(I+10)=0
6
IF (C(I) .LT. 100) THEN
7
C(I+10) = 2 * C(I)
8
ELSE
9
C(I+10) = C(I)/2
10
ENDIF
11 80 CONTINUE
12
END
Scope of
Dimension list
DATA values
Implied DO
Subscript list
DO loop
Logical IF
Block IF (true)
Block IF (false)

Begins at
( after array name
First /
( in I/O list
( after array name
Line following DO
After (hconditioni)
After THEN
After ELSEIF or ELSE

Ends at
The next )
Second /
I/O loop control
Matching )
Statement with DO label
End of line
ELSE, ELSEIF, or ENDIF
ELSE, ELSEIF, or ENDIF

Exhibit 3.12. Labeled scopes.

Correct Nesting:

Faulty Nesting:

Begin Scope A
Begin Scope B

Begin Scope A
Begin Scope B

End Scope B
End Scope A

End Scope A
End Scope B

Line #s
1
2
3
3
5 - 11
5
7
9

66

CHAPTER 3. ELEMENTS OF LANGUAGE

Exhibit 3.13. Tokens used to delimit program scopes.

Language
C
LISP
Pascal

PL/1

Beginning of Scope
{
(
BEGIN
RECORD
CASE
DO;
DO hloop controli;

End of Scope
}
)
END
END
END
END;
END;

used to delimit any kind of scope [Exhibit 3.13]. In these languages it is not possible to nest scopes
improperly because the compiler will always interpret the nesting in the legal way. A compiler will
match each end-scope to the nearest unmatched begin-scope. This design is attractive because it
produces a language that is simpler to learn and simpler to translate.
If an end-scope is omitted, the next one will be used to terminate the open scope regardless
of the programmers intent [Exhibit 3.14]. Thus an end-scope that was intended to terminate an
IF may instead be used to terminate a loop or a subprogram. A compiler error comment may
appear on the next line because the program element written there is in an illegal context, or error
comments may not appear until the translator reaches the end of the program and finds that the
wrong number of end-scopes was included. If an extra end-scope appears somewhere else, improper
nesting might not be detected at all.
Using one uniform end-scope indicator has the severe disadvantage that a nesting error may
not be identified as a syntactic error, but become a logical error which is harder to identify and
correct. The programmer has one fewer tool for communicating semantics to the compiler, and the
compiler has one fewer way to help the programmer achieve semantic validity. Many experienced
programmers use comments to indicate which end-scope belongs to each begin-scope. This practice
makes programs more readable and therefore easier to debug, but of course does not help the
compiler.
A third, intermediate way to handle scope delimiters occurs in Ada. Unlike Pascal, each kind of
scope has a distinct end-scope marker. Procedures and blocks and labeled loops have fully labeled
end-scopes. Unlike FORTRAN, a uniform syntax was introduced for delimiting and labeling scopes.
An end-scope marker is the word end followed by the word and label, if any, associated with the
beginning of the scope [Exhibit 3.15].
It is possible, in Ada, for the compiler to detect many (but not all) improperly nested scopes
and often to correctly deduce where an end-scope has been omitted. This is important, since a
misplaced or forgotten end-scope is one of the most common kinds of compile-time errors.
A good technique for avoiding errors with paired delimiters is to type the END marker when
the BEGIN is typed, and position the cursor between them. This is the idea behind the structured

67

3.2. THE METALANGUAGE

Exhibit 3.14. Nested unlabeled scopes in Pascal.

Begin Scope
Begin Scope
End Scope
End Scope

i := 0;
WHILE a <= 100 DO BEGIN
IF a mod 7 = 0 THEN BEGIN
i := i + 1;
writeln (i, a)
END
END

editors. When the programmer types the beginning of a multipart control unit, the editor inserts
all the keywords and scope markers necessary to complete that unit meaningfully. This prevents
beginners and forgetful experts from creating malformed scopes.

3.2.4

Comments

Footnotes and bibliographic citations in English permit us to convey general information about
the text. Analogously, comments, interspersed with program words, let us provide information
about a program that is not part of the program. With comments, as with statements, we have
the problem of identifying both the beginning and end of the unit. Older languages (COBOL,
FORTRAN) generally restrict comments to separate lines, begun by a specific comment mark in a
fixed position on the line [Exhibit 3.16]. This convention was natural when programs were typed
on punch cards. At the same time it is a severe restriction because it prohibits the use of brief
comments placed out of the way visually. It therefore limits the usefulness of comments to explain
obscure items that are embedded in the code.
The newer languages permit comments and code to be interspersed more freely. In these

Exhibit 3.15. Lexical scope delimiters in Ada.

Begin-scope markers
hblock_namei: hdeclarationsi BEGIN
PROCEDURE hproc_namei
LOOP
hlabeli:LOOP
CASE
IF hconditioni THEN or
ELSIF hconditioni THEN
ELSE

End-scope markers
END hblock_namei
END hproc_namei;
END LOOP;
END LOOP hlabeli;
END CASE
ELSIF, ELSE, or END IF
END IF

68

CHAPTER 3. ELEMENTS OF LANGUAGE

Exhibit 3.16. Comment lines in older languages.

In these languages comments must be placed on a separate line, below or above the code to
which they apply.
Language
Comment line is marked by
FORTRAN
A C in column 1
COBOL
A * in column 7
at the beginning of a line
original APL The lamp symbol:
BASIC
REM at the beginning of a line

languages, statements can be broken onto multiple lines and combined freely with short comments
in order to do a superior job of clarifying the intent of the programmer. Both the beginning and
end of a comment are marked [Exhibit 3.17]. Comments are permitted to appear anywhere within
a program, even in the middle of a statement.
A nearly universal convention is to place the code on the left part of the page and comments
on the right. Comments are used to document the semantic intent of variables, parameters, and
unusual program actions, and to clarify which end-scope marker is supposed to match each beginscope marker. Whole-line comments are used to mark and document the beginning of each program
module, greatly assisting the programmers eye in finding his or her way through the pages of code.
Some comments span several lines, in which case only the beginning of the first line and end of
the last line need begin- and end-comment marks. In spite of this, many programmers mark the
beginning and end of every line because it is aesthetically nicer and sets the comment apart from
code.
With all the advantages of these partial-line comments, one real disadvantage was introduced
by permitting begin-comment and end-comment marks to appear anywhere within the code. It is
not unusual for an end-comment mark to be omitted or typed incorrectly [Exhibit 3.18]. In this
case all the program statements up to the end of the next comment are taken to be part of the
nonterminated comment and are simply swallowed up by the comment.

Exhibit 3.17. Comment beginning and end delimiters.

These languages permit a comment and program code to be placed on the same line. Both the
beginning and end of the comment is marked.
Language
C
PL/1
Pascal
FORTH

Comments are delimited by

/* . . . */
/* . . . */
(* . . . *) or { . . . }
( ...)

3.2. THE METALANGUAGE

69

Exhibit 3.18. An incomplete comment swallowing an instruction.

The following Pascal code appears to be ok at first glance, but because of the mistyped end-comment
mark, the computation for tot_age will be omitted. The result will be a list of family members
with the wrong average age!
A person_list is an array of person cells, each containing a name and an age.
PROCEDURE average_age(p: person_list);
VAR famsize, tot_age, k:integer;
BEGIN
readln(famsize); (* Get the number of family members to process.*)
tot_age := 0;
FOR k := 1 TO famsize DO BEGIN
writeln( p[k].name ); (* Start with oldest family member. * )
tot_age := tot_age + p[k].age; (* Sum ages for average. *)
END;
writeln(0Average age of family = 0 , tot_age/famsize)
END;

The translator may not ever detect this violation of the programmers intent. If the next
comment is relatively near, and no end-scope markers are swallowed up by the comment, the
program may compile with no errors but run strangely. This can be a very difficult error to debug,
since the program looks correct but its behavior is inconsistent with its appearance! Eventually the
programmer will decide that he or she has clearly written a correct instruction that the compiler
seems to have ignored. Since compilers do not just ignore code, this does not make sense. Finally
the programmer notices that the end-comment mark that should be at the end of some prior line
is missing.
This problem is an example of the cost of over-generality. Limiting comments to separate lines
was too restrictive, that is, not general enough. Permitting them to begin and end anywhere
on a line, though, is more general than is needed or desired. Even in languages that permit
this, comments usually occupy either a full line or the right end of a line. A more desirable
implementation of comments would match the comment-scope and comment placement rules with
the actual conventions that most programmers use, which are:
whole-line comments
partial-line comments placed on the right side of the page
multiple-line comments
Thus comments should be permitted to occur on the right end of any line, but they might as well
be terminated by the end of the line. Permitting multiple-line comments to be written is important,

70

CHAPTER 3. ELEMENTS OF LANGUAGE

Exhibit 3.19. Comments terminating at end of line.

These languages permit comments to occupy entire lines or the right end of any line. Comments
start with the comment-begin mark listed and extend to the carriage return on the right end of the
line. (TeX is a text processing language for typesetting mathematical text and formulas. It was
used to produce this book.)
Language
Ada
LISP
TeX
UNIX command shell
C++

Comment-begin mark
-; (This varies among implementations.)
%
#
//

but it is not a big burden to mark the beginning of every comment line, as many programmers do
anyway to improve the appearance of their programs. The payoff for accepting this small restriction
is that the end-of-line mark can be used as a comment-end mark. Since programmers do not forget
to put carriage returns in their programs, comments can no longer swallow up entire chunks of
code. Some languages that have adopted this convention are listed in Exhibit 3.19.
Some languages support two kinds of comment delimiters. This permits the programmer to
use the partial-line variety to delimit explanatory comments. The second kind of delimiter (with
matched begin-comment and end-comment symbols) is reserved for use during debugging, when
the programmer often wants to comment out, temporarily, large sections of code.

3.2.5

Naming Parts of a Program

In order to refer to the parts of a program, we need meta-words for those parts and for whatever
actions are permitted. For example, C permits parts of a program to be stored in separate files and
brought into the compiler together by using #include hfile_namei. The file name is a metaword
denoting a section of the program, and #include is a metaword for the action of combining it
with another section.
Most procedural languages provide a GOTO instruction which transfers control to a specific
labeled statement somewhere in the program. The statement label, whether symbolic or numeric,
is thus a metaword that refers to a part of the program. Since the role of statement labels cannot
be fully understood apart from the control structures that use them, labels are discussed with the
GOTO command in Section 11.1.

3.2.6

Metawords That Let the Programmer Extend the Language

There are several levels on which a language may be extended. One might extend:
The list of defined words (nouns, verbs, adjectives).

3.2. THE METALANGUAGE

71

The syntax but not the semantics, thus providing alternative ways of writing the same meanings one could write without the extension.
The actual semantics of the language, with a corresponding extension either of the syntax or
of the list of defined words recognized by the compiler.
Languages that permit the third kind of extension are rare because extending the semantics
requires changing the translator to handle a new category of objects. Semantic extension is discussed
in the next chapter.
Extending the Vocabulary
Every declaration extends the language in the sense that it permits a compiler to understand
new words. Normally we are only permitted to declare a few kinds of things: nouns (variables,
constants, file names), verbs (functions and procedures), and sometimes adjectives (type names)
and metawords (labels). We cannot normally declare new syntactic words or new words such as
array. The compiler maintains one combined list or several separate lists of these definitions.
This list is usually called the symbol table, but it is actually called the dictionary in FORTH.
New symbols added to this list always belong to some previously defined syntactic category with
semantics defined by the compiler.
Each category of symbol that can be declared must have its own keyword or syntactic marker
by which the compiler can recognize that a definition of a new symbol follows. Words such as TYPE,
CONST, and PROCEDURE in Pascal and INTEGER and FUNCTION in FORTRAN are metawords that
mean, in part, extend the language by putting the symbols that follow into the symbol table.
As compiler technology has developed and languages have become bigger and more sophisticated, more kinds of declarable symbols have been added to languages. The original BASIC
permitted no declarations: all two-letter variable names could be used without declaration, and
no other symbols, even subroutine names, could be defined. The newest versions of BASIC permit
use of longer variable names, names for subroutines, and symbolic labels. FORTRAN, developed in
19541958, permitted declaration of names for variables and functions. FORTRAN 77 also permits
declaration of names for constants and COMMON blocks. ALGOL-68 supported type declarations
as a separate abstraction, not as part of some data object. Pascal, published in 1971, brought type
declarations into widespread use. Modula, a newer language devised by the author of Pascal, permits declaration and naming of semantically separate modules. Ada, one of the newest languages
in commercial use, permits declaration of several things missing in Pascal, including the range
and precision of real variables, support for concurrent tasks, and program modules called generic
packages which contain data and function declarations with type parameters.

72

CHAPTER 3. ELEMENTS OF LANGUAGE

Exhibit 3.20. Definition of a simple macro in C.

In C, macro definitions start with the word #define, followed by the macro name. The string to
the right of the macro name defines the meaning of the name.
The #define statements below make the apparent syntax of C more like Pascal. They permit the
faithful Pascal programmer to use the familiar scoping words BEGIN and END in a C program. (These
words are not normally part of the C language.) During preprocessing, BEGIN will be replaced by
{ and END will be replaced by }.
#define BEGIN {
#define END }

Syntactic Extension without Semantic Extension

Some languages contain a macro facility (in C, it is part of the preprocessor). 7 This permits the
programmer to define short names for frequently used expressions. A macro definition consists of
a name and a string of characters that becomes the meaning of the name [Exhibit 3.20]. To use a
macro, the programmer writes its name, like a shorthand notation, in the program wherever that
string of characters is to be inserted [Exhibit 3.21].
A preprocessor scans the source program, searching for macro names, before the program is
7

The C preprocessor supports various compiler directives as well as a general macro facility.

Exhibit 3.21. Use of a simple macro in C.

Macro Calls. The simple macros defined in Exhibit 3.20 are called in the following code
fragment. Unfortunately, the new scope symbols, BEGIN and END, and the old ones, { and },
are now interchangeable. Our programmer can write the following code, defining two well-nested
scopes. It would work, but it isnt pretty or clear.
BEGIN

x = y+2;
if (x < 100) { x += k; y = 0; END
else x = 0; }

Macro Expansion. During macro expansion the macro call is replaced by the defining string.
The C translator never sees the word BEGIN.
{
x = y+2;
if (x < 100) { x += k; y = 0; }
else x = 0; }

3.2. THE METALANGUAGE

73

parsed. These macro names are replaced by the defining strings. The expanded program is then
parsed and compiled. Thus the preprocessor commands and macro calls form a separate, primitive,
language. They are identified, expanded, and eliminated before the parser for the main language
even begins its work.
The syntax for a macro language, even one with macro parameters, is always simple. However,
piggy-backing a macro language on top of a general programming language causes some complications. The source code will be processed by two translators, and their relationship must be made
clear. Issues such as the relationship of macro calls to comments or quoted strings must be settled.
In C, preprocessor commands and macro definitions start with a # in column 1. 8 This distinguishes them from source code intended for the compiler. Custom (but not compiler rules) dictates
that macro names be typed in uppercase characters and program identifiers in lowercase. Case
does not matter to the translator, but this custom helps the programmer read the code.
Macro calls are harder to identify than macro definitions, since they may be embedded anywhere
in the code, including within a macro definition. Macro names, like program identifiers, are variablelength strings that need to be identified and separated from other symbols. Lexical analysis must,
therefore, be done before macro expansion. Since the result of expansion is a source string, lexical
analysis must be done again after expansion. Since macro definitions may contain macro calls, the
result of macro expansion must be rescanned for more macro calls. Control must thus pass back
and forth between the lexer and the macro facility. The lexical rules for the preprocessor language
are necessarily the same as the rules for the main language.
In the original definition of C, the relationship among the lexer, preprocessor, and parser was
not completely defined. Existing C translators thus do different things with macros, and all are
correct by the language definition. Some C translators simply insert the expanded macro text
back into the source text without inserting any blanks or delimiters. The effect is that characters
outside a macro can become adjacent to characters produced by the macro expansion. The program
line containing the expanded macro is then sent back to the lexer. When the lexer processes this,
it forms a single symbol from the two character strings. This gluing action can produce strange
and unexpected results.
The ANSI standard for C has clarified this situation. It states that no symbol can bridge a macro
boundary. Lexical analysis on the original source string is done, and symbols are identified, before
macro expansion. The source string that defines the macro can also be lexed before expansion,
since characters in it can never be joined with characters outside it. These rules clean up a
messy situation. The result of expanding a macro still must be rescanned for more macro calls, but
it does not need to be re-lexed. The definition and call of a macro within a macro are illustrated
in Exhibits 3.22 and 3.23.
A general macro facility also permits the use of parameters in macro definitions [Exhibit 3.24].
In a call, macro arguments are easily parsed, since they are enclosed in parentheses and follow the
macro name [Exhibit 3.25]. To expand a macro, formal parameter names must be identified in the
definition of the macro. To do this, the tokens in the macro definition must first be identified. Any
8

Newer C translators permit the # to be anywhere on the line as long as it is the first nonblank character.

74

CHAPTER 3. ELEMENTS OF LANGUAGE

Exhibit 3.22. A nest of macros in C.

The macros defined here are named PI and PRINTX. PRINTX expands into a call on the library
function that does formatted output, printf. The first parameter for printf must be a format
string, the other parameters are expressions denoting items to be printed. Within the format, a
% field defines the type and field width for each item on the I/O list. The \t prints out a tab
character.
#define PI
#define PRINTX

3.1415927
printf("Pi times x = %8.5f\t", PI * x)

Exhibit 3.23. Use of the simple PRINTX macro.

The macro named PRINTX is used below in a for loop.
for (x=1; x<=3; x++) PRINTX;
Before compilation begins, the macro name is replaced by the string printf("Pi times x =
%8.5f\t", PI * x), giving a string that still contains a macro call:
for (x=1; x<=3; x++) printf("Pi times x = %8.5f\t", PI * x);
This string is re-scanned, and the call on the macro PI is expanded, producing macro-free source
code. The compiler then compiles the statement:
for (x=1; x<=3; x++) printf("Pi times x = %8.5f\t", 3.1415927 * x);
At run time, this code causes x to be initialized to 1 before the loop is executed. On each
iteration of the loop, the value of x is compared to 3. If x does not exceed 3, the words Pi times
x = are printed, followed by the value of 3.1415927 * x as a floating-point number with five
decimal places (%8.5f), followed by a tab character (\t). The counter x is then incremented. The
loop is terminated when x exceeds 3. Thus a line with five fields is printed, as follows:
Pi times x = 3.14159
Pi times x = 6.28319
Pi times x = 9.42477

3.2. THE METALANGUAGE

75

Exhibit 3.24. A macro with parameters in C.

The macro defined here is named PRINT. It is similar to the PRINTX macro in Exhibit 3.22, but
it has a parameter.
#define PRINT(yy) printf(#yy " = %d\t", yy)
The definition for PRINT is written in ANSI C. References to macro parameters that occur
within quoted strings are not recognized by the preprocessor. However, the # symbol in a macro
definition causes the parameter following it to be converted to a quoted string. Adjacent strings
are concatenated by the translator. Using both these facts, we are able to insert a parameter value
into a quoted format string.

token that matches a parameter name is replaced by the corresponding argument string. Finally,
the entire string of characters, with parameter substitutions, replaces the macro call.
The original definition of C did not clearly define whether tokens were identified before or
after macro parameters were processed. This is important because a comment or a quoted string
looks like many words but forms a single program token. If a preprocessor searches for parameter
names before identifying tokens, quoted strings will be searched and parameter substitution will
happen within them. Many C translators work this way; others identify tokens first. The ANSI
C standard clarifies this situation. It decrees that tokenization will be done uniformly before
parameter substitution.
Macro names are syntactic extensions. They are words that may be written in the program
and will be recognized by the compiler. Unlike variable declarations they may stand for arbitrarily
complex items, and they may expand into strings that are not even syntactically legal units when
used alone. Macros can be used to shorten code with repetitive elements, to redefine the compiler
words such as BEGIN, or to give symbolic names to constants. What they do not do is extend the

Exhibit 3.25. Use of the print macro with parameters.

The macro named PRINT is used here, with different variables supplied as parameters each time.
PRINT(x); PRINT(y); PRINT(z);
These macro calls will be expanded and produce the following compilable code:
printf("x = %d\t", x);
printf("y = %d\t", y);
printf("z = %d\t", z);
Assume that at run time the variables x, y, and z contain the values 1, 3, and 10, respectively.
Then executing this code will cause one line to be printed, as follows:
x = 1
y = 3
z = 10

76

CHAPTER 3. ELEMENTS OF LANGUAGE

semantics of the language. Since all macro calls must be expanded into compilable code, anything
written with a macro call could also be written without it. No power is added to the language
by a macro facility.

Exercises
1. Why are function calls considered verbs?
2. What is the domain of a verb? Define the domain and range of a function.
3. What is a data type? Inheritance?
4. What is a metalanguage?
5. What is a lexical token? How are lexical tokens formed? Use a language with which you are
familiar as an example. What are delimiters?
6. How are programming language statements analogous to sentences?
7. What is the scope of a programming language unit? How is it usually denoted?
8. How is it possible to improperly nest scopes? How can this be avoided by designers of
programming languages?
9. What is the purpose of a comment? How are comments traditionally handled within programs? What is the advantage of using a carriage return as a comment delimiter?
10. The language C++ is an extension of C which supports generic functions and type checking.
For the most part, C++ is C with additions to implement things that the C++ designers
believed are important and missing from C. One of the additions is a second way to denote
a comment. In C, a comment can be placed almost anywhere in the code and is delimited
at both ends. In this program fragment two comments and an assignment statement are
intermingled:
x=y*z

/* Add the product of y and z */+x;

/* to x.

*/

C++ supports this form but also a new form which must be placed on the right end of the
line and is only delimited at the beginning by //:
x=y*z + x // Add the product of y and z to x.
Briefly explain why the original comment syntax was so inadequate that a new form was
needed.
11. How can we extend a language through its vocabulary? Its syntax?
12. What is a macro? How is it used within a program?

Chapter 4

Formal Description of Language

Overview
The syntax of a language is its grammatical rules. These are usually defined through
EBNF (Extended Backus-Naur Form) and/or syntax diagrams, both discussed in this
chapter. The meaning of a program is represented by p-code (portable code) or by a
computation tree. The language syntax defines the computation tree that corresponds
to each legal source program.
Semantics are the rules for interpreting the meaning of programming language statements. The semantic specification of a language defines how each computation tree
is to be implemented on a machine so that it retains its meaning. Being always concerned with the portability of code, we define the semantics of a language in terms
of an implementation-independent model. One such model, the abstract machine, is
composed of a program environment, shared environment, stack, and streams. The
semantic basis of a language means the specific version of the machine that defines
the language, together with the internal data structures and interpretation procedures
that implement the abstract semantics. Lambda calculus is an example of a minimal
semantic basis.
A language may be extended primarily through its vocabulary and occasionally through
its syntax, as in EL/1, or through its semantics, as in FORTH.

77

78

4.1

CHAPTER 4. FORMAL DESCRIPTION OF LANGUAGE

Foundations of Programming Languages

Formal methods have played a critical role in the development of modern programming languages.
Formal methods were not available in the mid-1950s when the first higher-level programming languages were being created. The most notable of these efforts was FORTRAN, which survives (in
greatly expanded form) to this day. Even though the syntax and semantics of the early FORTRAN
were primitive by todays standards, the complexity of the language was at the limit of what could
be handled by the methods then available. It was quickly realized that ad hoc methods are severely
limited in what they can achieve, and a more systematic approach would be needed to handle
languages of greater expressive power and correspondingly greater complexity.
Contemporaneously with the implementation of the FORTRAN language and compiler, a new
language, ALGOL, was being defined using a new formal approach for the specification of syntax
and semantics. Even though it required several more years of research before people learned how to
compile ALGOL efficiently, the language itself had tremendous influence on the design of subsequent
programming languages. Concepts such as block structure (cf. Chapter 7) and delayed evaluation
of function parameters (cf. Chapter 8), introduced in ALGOL, have reappeared in many subsequent
modern programming languages.
ALGOL was the first programming language whose syntax was formally described. A notation
called BNF, for Backus-Naur Form, was invented for the purpose. BNF turned out to be equivalent
in expressive power to context-free grammars, developed by the linguist Noam Chomsky for describing natural language, but the BNF notation turned out to be easier for people, so variations on
it are still used in describing most programming languages. An attempt was made to give a rigorous English-language specification of the semantics of ALGOL. Nevertheless, the underlying model
was not well understood at the time, and ALGOL appeared at first to be difficult or impossible to
implement efficiently.
Syntax and semantic interpretations were specified informally for early languages. Then, motivated by the new need to describe programming languages, formal language theory flourished.
Some of the major developments in the foundations of computer science are shown in Exhibit 4.1.
Formal syntax and parsing methods grew from work on automata theory and linguistics [Exhibit
4.1]. Formal methods of semantic specification [Exhibit 4.2] grew from early work on logic and
computability and were especially influenced by Churchs work on the lambda calculus. In this
chapter, we give a brief introduction to some of the formal tools that have been important to the
development of modern-day programming languages.

4.2

Syntax

The rules for constructing a well-formed sentence (statement) out of words, a paragraph (module)
out of sentences, and an essay (program) out of paragraphs are the syntax of the language. The
syntax definitions for most programming languages take several pages of text. A few are very
short, a few very long. There is at least one language (ALGOL-68) in which the syntax rules that

79

4.2. SYNTAX

Exhibit 4.1. Foundations of computer science.

Giuseppe Peano
Set theory (1895)
1900

1900

1910

Alfred North Whitehead

Bertrand Russel
Symbolic logic (1910)

1910

1920

Automated mathematics
Post
Incompleteness theorem, Goedel (1931)

1920

Recursive function theory

Church, Rosser (1930s)

1930

1930

Post systems

Computability theory
Turing (1936)

1940

1950

1940
Formal language theory
Chomsky

Information theory
Shannon

Electronics

1950
Switching theory

1960

Formal syntactic definition

Backus and Naur (1960)

1970

Knuth: Parsing methods,

compiler theory
Compiler compilers
EL/1: Extensible syntax

1980

1990

Automata theory
Rabin, Scott
Complexity theory
Hartmanis, Blum
Computational
cryptography (1976)
Diffie, Hellman
Public key system
Rivest, Shamir,
Adelman (1978)

1960

1970

Randomized
algorithms

1980

1990

80

CHAPTER 4. FORMAL DESCRIPTION OF LANGUAGE

Exhibit 4.2. Formal semantic specification.

1930 Post systems

1940

Recursive function theory

Church, Rosser (1930s)

Computability theory
Turing (1936)

1940

Lambda calculus
Church (1941)

1950

1960

1970

1950
Program correctness
and verification (1960s)
Referential transparency, Strachey
Formal semantic definition
SECD machine, Landin (1964)
Vienna definition of PL/1 (1967)
Denotational semantics (1971)
Scott, Strachey
Milner: Type theory (1978)

1980

1990

1930

Functional languages:
ML
Miranda
Haskell

1960

1970
Concurrency, Dijkstra (1968)
Hoare: CSP (1978)
Distributed computing (1978)
Lamport
Collaborative computing
1988

1980

1990

determine whether or not a statement should compile are so complicated that only an expert can
understand them.
It is usual to define the syntax of a programming language in a formal language. A variety
of formalisms have been introduced over the years for this purpose. We present two of the most
common here: Extended Backus-Naur Form (EBNF) and syntax diagrams.
An EBNF language definition can be translated by a program called a parser generator 1 into
a program called a parser [Exhibit 4.3]. 2 A parser reads the users source code programs and
determines the syntactic category (part of speech) of every source symbol and combination of
1

The old term was compiler compiler. This led to the name of the UNIX parser generator, yacc, which stands
for yet another compiler compiler.
2
A parser generator can only handle grammars for context-free languages. Defining this language class is beyond
the scope of this book. Note, though, that the grammars published for most programming languages are context free.

81

4.2. SYNTAX

Exhibit 4.3. The compiler is produced from the language definition.

In the following diagram, programs are represented by rectangles and data by circles. The lexer
and parser can be automatically generated from the lexical specifications and syntax of a contextfree language by a parser generator and its companion lexer generator. This is represented by the
vertical arrows in the diagram. The lexer and parser are the output data of these generation steps.
A code generator requires more hand work: the compiler writer must construct an assembly
code translation, for every syntax rule in the grammar, which encodes the semantics of that rule
in the target machine language.
The lexer, parser, and code generator are programs that together comprise the compiler. The
compilation process is represented by the horizontal chain in the diagram.
EBNF
rules for
Pascal
syntax

Lexical
rules for
Pascal

Lexer
Generator

Lexer
for
Pascal

Hand-coded
semantic
interpretation
for each
rule

Parser Generator
also known as
Compiler Compiler

Pascal Compiler
Source
code for
a Pascal
program

Lexer
for
Pascal

Parser
for
Pascal

Parsed,
intermediate
form of
program

Pascal
Code
Generator

Object
code for
program

symbols. Its output is the list of the symbols defined in the program and a parse tree, which
specifies the role that each source symbol is serving, much like a sentence diagram of an English
sentence. The parser forms the heart of any compiler or interpreter for the language.
The study of formal language theory and parsing has strongly affected language design. Older
languages were not devised with modern parsing methods in mind. Their syntax was usually
developed ad hoc. Consequently, a syntax definition for such a language, for example FORTRAN,
is lengthy and full of special cases. By todays standards these languages are also relatively slow
and difficult to parse.
Newer languages are designed to be parsed easily by efficient algorithms. The syntax for Pascal
is brief and elegant. Pascal compilers are small, as compilers go, and can be implemented on
personal computers. The standard LISP translator 3 is only fifteen pages long!
3

Griss and Hearn [1981].

82

CHAPTER 4. FORMAL DESCRIPTION OF LANGUAGE

4.2.1

Extended BNF

Backus-Naur Form, or BNF, is a formal language developed by Backus and Naur for describing
programming language syntax. It gained widespread influence when it was used to define ALGOL in
the early 1960s. The original BNF formalism has since been extended and streamlined; a generally
accepted version, named Extended BNF, is presented here.
An EBNF grammar consists of:
A starting symbol.
A set of terminal symbols, which are the keywords and syntactic markers of the language
being defined.
A set of nonterminal symbols, which correspond to the syntactic categories and kinds of
statements of the language.
A series of rules, called productions, that specify how each nonterminal symbol may be
expanded into a phrase containing terminals and nonterminals. Every nonterminal has one
production rule, which may contain alternatives.
The Syntax of EBNF
The syntax for EBNF itself is not altogether standardized; several minor variations exist. We define
a commonly used version here.
The starting symbol must be defined. One nonterminal is designated as the starting symbol.
Terminal symbols will be written in boldface and enclosed in single quotes.
Nonterminal symbols will be written in regular type and enclosed in hangle bracketsi.
Production rules. The nonterminal being defined is written at the left, followed by a ::= sign
(which we will pronounce as goes to). After this is the string, with options, which defines
the nonterminal. The definition extends up to but does not include the . that marks the
end of the production. When a nonterminal is expanded it is replaced by this defining phrase.
Blank spaces between the ::= and the . are ignored.
Alternatives are separated by vertical bars. Parentheses may be used to indicate grouping.
For example, the rule
s ::= ( a | bc ) d .
indicates that an s may be replaced by an ad or a bcd.
An optional syntactic element is a something-or-nothing alternativeit may be included or
not included as needs demand. This is indicated by enclosing the optional element in square
brackets, as follows:

83

4.2. SYNTAX
s ::= [a] d .
This formula indicates that an s may be replaced by an ad or simply by a d.

An unspecified number of repetitions (zero or more) of a syntactic unit is indicated by enclosing the unit in curly brackets. For example, the rule
s ::= {a}d .
indicates that an s may be replaced by a d, an ad, an aad, or a string of any number of
as followed by a single d. A frequently occurring pattern is the following:
s ::= t{t}
This means that s may be replaced by one or more copies of t.
Recursive rules. Recursive production rules are permitted. For example, this rule is directly
recursive because its right side contains a reference to itself:
s ::= asz | w .
This expands into a single w, surrounded on the left and right by any number of matched pairs
of a and z: awz, aawzz, aaawzzz, etc.
Tail recursion is a special kind of recursion in which the recursive reference is the last symbol
in the string. Tail recursion has the same effect as a loop. This production is tail recursive:
s ::= as | b .
This expands into a string of any number of as followed by a b.
Mutually recursive rules are also permitted. For example, this pair of rules is mutually recursive
because each rule refers to the other:
s ::= at | b .
t ::= bs | a .
A single s could expand into any of the following: b, aa, abb, abaa, ababb, ababaa, etc.
Combinations of alternatives, optional elements, recursions, and repetitions often occur in a
production, as follows:
s ::= {a | b} [c] d .
This rule indicates that an s may be replaced by any of the following: d, ad, bd, cd, acd, bcd,
aad, abd, aacd, abcd, bd, bad, bbd, bcd, bacd, bbcd, and many more.
Using EBNF
To illustrate the EBNF rules, we give part of the syntax for Pascal, taken from the ISO standard
[Exhibit 4.4]. The first few rules of the grammar are given, followed by several rules from the
middle of the grammar which define what a statement is. The complete set of EBNF grammar
rules cannot be given here because it is too long. 4 Following are brief explanations of the meaning
4

It occupies nearly six pages in Cooper [1983].

84

CHAPTER 4. FORMAL DESCRIPTION OF LANGUAGE

Exhibit 4.4. EBNF production rules for parts of Pascal.

program ::= hprogram-headingi ; hprogram-blocki .

program-heading ::= program hidentifieri [ ( hprogram-parametersi ) ].

program-parameters ::= hidentifier-listi .
identifier-list ::= hidentifieri { , hidentifieri } .
program-block ::= hblocki .
block ::= hlabel-declaration-parti hconstant-declaration-parti
htype-declaration-parti hvariable-declaration-parti
hprocedure-and-function-declaration-parti hstatement-parti .
variable-declaration-part ::= [ var { hidentifier-listi :

htypenamei ; } ] .

statement-part ::= compound statement .

compound-statement ::= begin hstatement-sequencei end .
statement-sequence ::= hstatementi { ; hstatementi } .
statement ::= [ hlabeli :

( hsimple-statementi | hstructured-statementi ).

simple-statement ::= hempty-statementi | hassignment-statementi |

hprocedure-statementi | hgoto-statementi .
structured-statement ::= hcompound-statementi | hconditional-statementi |
hrepetitive-statementi | hwith-statementi .

of these rules.
The production for the starting symbol states that a program consists of a heading, a semicolon, a block and a period. The semicolon and period are terminal symbols and will form
part of the finished program. The symbols program-heading and program-block are
nonterminals and need further expansion.
The program-heading starts with the terminal symbol program, which is followed by the
name of the program and an optional, parenthesized list of parameters, used for file names.
The program parameters, if they are used, are just a list of identifiers, that is, a series of one
or more identifiers separated by commas.
The program block consists of a series of declarations followed by a single compound statement.
The production for compound statement forms an indirectly recursive cycle with the rules
for statement sequence, and statement. That is, a statement can be a structured statement,

85

4.2. SYNTAX

which can be a compound statement, which contains a statement-sequence, which contains a

statement, completing the cycle.
The rule for statement contains an optional label field and the choice between simplestatement and structured-statement.
The rules for simple-statement and structured-statement define all of Pascals control structures.
Generating a Program. To generate a program (or part of a program) using a grammar, one
starts with the specified starting symbol and expands it according to its production rule. The
starting symbol is replaced by the string of symbols from the right side of its production rule.
If the rule contains alternatives, one may use whichever option seems appropriate. The resulting
expansion will contain other nonterminal symbols which then must be expanded also. When all
the nonterminals have been expanded, the result is a grammatically correct program.
We illustrate this derivation process by using the EBNF grammar for ISO Standard Pascal to
generate a ridiculously simple program named little. Parts, but not all, of this grammar are
given in Exhibit 4.4.5
The starting symbol is hprogrami. Wherever possible, more than one nonterminal symbol is
reduced on each line, in order to shorten the derivation.
hprogrami
hprogram-headingi ; hprogram-blocki .
program hidentifieri ; hblocki .
program little ; hlabel-declaration-parti hconstant-declaration-parti
hvariable-declaration-parti hprocedure-and-function-declaration-parti
hstatement-parti .
program little ; var hvariable-declarationi ; hcompound-statementi .
program little ; var hidentifier-listi :
begin hstatement-sequencei end .

htype-denoteri ;

program little ; var hidentifieri : htype-denoteri ;

begin hstatementi ; hstatementi end .
program little ; var x : integer ;
begin hsimple-statementi ; hsimple-statementi end .
program little ; var x : integer ;
begin hassignment-statementi ; hprocedure-statementi end .
program little ; var x : integer ;
begin hvariable-accessi := hexpressioni ;
hprocedure-identifieri ( hwriteln-parameter-listi ) end .
5

The complete grammar can be found in Cooper [1983], pp 15358.

86

CHAPTER 4. FORMAL DESCRIPTION OF LANGUAGE

program little ; var x : integer ;
begin hentire-variablei := hsimple-expressioni ;
writeln (hwrite-parameteri ) end .
program little ; var x : integer ; begin hvariable-identifieri:= htermi ;
writeln ( hexpressioni ) end .
program little ; var x : integer ; begin hidentifieri := hfactori ;
writeln ( hsimple-expressioni ) end.
program little ; var x : integer ; begin x := hunsigned-constanti ;
writeln ( htermi ) end .
program little ; var x : integer ; begin x := hunsigned-numberi ;
writeln ( hfactori ) end .
program little ; var x : integer ; begin x := hunsigned-integeri ;
writeln ( hvariable-accessi ) end .
program little ; var x : integer ; begin x := 17 ;
writeln ( hentire-variablei ) end .
program little ; var x : integer ; begin x := 17 ;
writeln ( hvariable-identifieri ) end .
program little ; var x : integer ; begin x := 17 ; writeln ( hidentifieri ) end .
program little ; var x : integer ; begin x := 17 ; writeln ( x ) end .

Parsing a Program. The process of syntactic analysis is the inverse of this generation process.
Syntactic analysis starts with source code. The parsing routines of a compiler determine how the
source code corresponds to the grammar. The output from the parse is a tree-representation of the
grammatical structure of the code called a parse tree.
There are several methods of syntactic analysis, which are usually studied in a compiler course
and are beyond the scope of this book. The two broad categories of parsing algorithms are called
bottom-up and top-down. In top-down parsing, the parser starts with the grammars starting
symbol and tries, at each step, to generate the next part of the source code string. A brief description
of a bottom-up method should serve to illustrate the parsing process. In a bottom-up parse,
the parser searches the source code for a string which occurs as one alternative on the right side
of some production rule. Ambiguity is resolved by looking ahead k input symbols. The matching
string is replaced by the nonterminal on the left of that rule. By repeating this process, the program
is eventually reduced, phrase by phrase, back to the starting symbol. Exhibit 4.5 illustrates the
steps in forming a parse tree for the body of the program named little.
All syntactically correct programs can be reduced in this manner. If a compiler cannot do the
reduction successfully, there is some error in the source code and the compiler produces an error

4.2. SYNTAX

87

comment containing some guess about what kind of syntactic error was made. These guesses are
usually close to being correct when the error is discovered near where it was made. Their usefulness
decreases rapidly as the compiler works on and on through the source code without discovering the
error, as often happens.

4.2.2

Syntax Diagrams

Syntax diagrams were developed by Niklaus Wirth to define the syntax of Pascal. They are also
called railroad diagrams, because of their curving, branching shapes. This is the form in which
Pascal syntax is usually presented in textbooks. Syntax diagrams and EBNF can express exactly
the same class of languages, but they are used for different purposes. Syntax diagrams provide a
graphic, two-dimensional way to communicate a grammar, so they are used to make grammatical
relationships easier for human beings to grasp.
EBNF is used to write a grammar that will be the input to a parser generator. Corresponding to
each production is code for the semantic action that the compiler should take when that production
is parsed. The rules of an EBNF syntax are often more broken up than seems necessary, in order to
provide hooks for all the semantic actions that a compiler must perform. When a grammar for
the same language is presented as syntax diagrams, several EBNF productions are often condensed
into one diagram, making the entire grammar shorter, less roundabout, and easier to comprehend.
A Wirth syntax diagram definition has the same elements as an EBNF grammar, as follows:
A starting symbol.
Terminal symbols, written in boldface but without quotes, sometimes also enclosed in round
or oval boxes.
Nonterminal symbols, written in regular type.
Production rules are written using arrows (as in a flow chart) to indicate alternatives, options,
and indefinite repetition. Each rule starts with a nonterminal symbol written at the left and
ends where the arrow ends on the right.
Nonterminal symbols are like subroutine calls. To expand one, you go to the correct diagram,
follow the arrows through the diagram until it ends, and return to the calling point to finish the
calling production. Branch points correspond to alternatives and indicate that any appropriate
choice can be made. Repetition is encoded by backward-pointing arrows which form explicit loops.
Direct and indirect recursion are both allowed.
Syntax diagrams are given in Exhibits 4.6 and 4.7, which correspond exactly to the EBNF
grammar fragments in Exhibit 4.4.
In spite of the simplicity and visual appeal of syntax diagrams, though, the official definition
of Pascal grammar is written in EBNF, not syntax diagrams. EBNF is a better input language
for a parser generator and provides a clearer basis for a formal definition of the semantics of the
language.
Revision 1.8 1992/06/09 17:15:02 fischer

88

CHAPTER 4. FORMAL DESCRIPTION OF LANGUAGE

Exhibit 4.5. Parsing a simple Pascal program.

We perform a bottom-up parse of part of the program named little, using standard Pascal
syntax, part of which is shown in Exhibit 4.4. Starting with the expression at the top, we identify
a single token or a consecutive series of tokens that correspond to the right side of a syntactic rule.
This series is then reduced, or replaced by the left side of that rule. The final reduction is shown
at the bottom of the diagram.

begin

:=

17

writeln
procedureidentifier

identifier

unsignedinteger

variableidentifier

unsignednumber

variableidentifier

entirevariable

unsignedconstant

entirevariable

variableaccess

factor

variableaccess

term

identifier

factor

simpleexpression
expression

term
simpleexpression

assignmentstatement

expression

simplestatement

writeparameter

statement

writelnparameterlist
procedurestatement
simplestatement
statement
statementsequence
compoundstatement

end

89

4.2. SYNTAX

Exhibit 4.6. Syntax diagram for program.

This diagram corresponds to the EBNF productions for program, program-heading, programparameters, and identifier list. The starting symbol is program .

,
program

program

identifier

identifier

;
.

program-block

Exhibit 4.7. Syntax diagrams for statement.

These diagrams correspond to the EBNF productions for statement, simple-statement, structuredstatement, compound-statement, and statement-sequence.

statement

label

:
assignment-statement
procedure-call-statement
goto-statement
compound-statement
if-statement
case-statement
with-statement
while-statement
repeat-statement
for-statement
;

compound-statement

begin

statement

end

90

CHAPTER 4. FORMAL DESCRIPTION OF LANGUAGE

Exhibit 4.8. Translation: from source to machine code.

The object code was generated by OSS Pascal for the Atari ST.
Source code:
P-code tree:
begin
;
x := 17;
y:= x+1
:=
:=
end
Object code:
moveq #17,d0
move d0,x
addq #1,d0
move d0,y

4.3
4.3.1

17

+
x

Semantics
The Meaning of a Program

A modern language translator converts a program from its source form into a tree representation.
This tree representation is sometimes called p-code, a shortening of portable code, because it is
completely independent of hardware. This tree represents the structure of the program. The formal
syntax of the language defines the kinds of nodes in the tree and how they may be combined. In
this tree, the nodes represent objects and computations, and the structure of the tree represents
the (partial) order in which the computations must be done. If any part of this tree is undefined
or missing, the tree may have no meaning.
The formal semantics defines the meaning of this tree and, therefore, the meaning of the program. A language implementor must determine how to convert this tree to machine code for a
specific machine so that his or her translation will have the same meaning as that defined by the
formal semantics. This two-step approach is used because the conversion from source text to tree
form can be the same for all implementations of a language. Only the second step, code generation,
is hardware-dependent [Exhibit 4.8].

4.3.2

Definition of Language Semantics

The rules for interpreting the meaning of statements in a language are the semantics of the language.
In order for a language to be meaningful and useful, the language designers, compiler writers,
and programmers must share a common understanding of those semantics. If no single semantic
standard exists, or no common understanding of the standard exists, various compiler writers will
implement the language differently, and a programmers knowledge of the language will not be
transferable from one implementation to another. This is indeed the situation with both BASIC

4.3. SEMANTICS

91

and LISP; many incompatible versions exist.

Knowing the full syntax of a programming language is enough to permit an experienced person
to make a guess about the semantics, but such a guess is at best rough, and it is likely to be wrong
in many details and in some major ways. This is because highly similar syntactic forms in similar
languages often have different semantics.
The syntax of a programming language needs only to describe all strings of symbols that comprise legal programs. To define the semantics, one must either define the results of some real or
abstract computer executing the program, or write a complete set of mathematical formulas that
axiomatize the operation of the program and the expected results. Either way, the definition must
be complete, precise, correct, and nonambiguous. Neither kind of definition is easy to make.
The semantics of a language must thus define a highly varied set of things, including but not
limited to:
What is the correct interpretation of every statement type?
What do you mean when you write a name?
What happens during a function call?
In what order are computations done?
Are there syntactically legal expressions that are not meaningful?
In what ways does a compiler writer have freedom?
To what extent must all compilers produce code that computes the same answers?
In general, answering such questions takes many more pages than defining the syntax of a
language. For example, syntax diagrams for Pascal can be printed in eight pages, three of which
also contain extensive semantic information. 6 In contrast, a complete semantic description of Pascal,
at a level that can be understood by a well-educated person, takes 142 pages. 7 Part of the reason
for this difference is the dissimilarity between the meta-languages in which syntax and semantics
are defined.
The semantics of natural languages are communicated to learners by a combination of examples
and attempts to describe the meaning. The examples are required because an English description
of semantics will lack precision and be as ambiguous as English. Similarly, English alone is not
adequate to define the semantics of a programming language because it is too vague and too
ambiguous to define highly complex things in such a way that no doubt remains about their
meaning.
Just as it is possible to create a formal system such as EBNF to define language syntax, it is
possible to create a formal system to define programming language semantics. 8 There is a major
6

Dale and Lilly [1985], pages A1A8.

Cooper [1983].
8
Historical note: The Vienna Definition of PL/1 defined a new language for expressing semantics and defined
the semantics of PL/1 in it. ALGOL-68 also had its own, impenetrable, formal language that tried to eliminate most
of the need for a semantic definition by including semantics in the syntax. The result was a book-length syntax.
7

92

CHAPTER 4. FORMAL DESCRIPTION OF LANGUAGE

difference, though. The languages used to express syntax are relatively easy to learn and can be
mastered by any student with a little effort. The languages used to express semantics are very
difficult to read and extremely difficult to write.
The primary use for a formal semantic definition is to establish a single, unambiguous standard
for the semantics of the language, to which all other semantic descriptions must conform. It defines
all details of the meaning of the language being described and provides a precise answer to any
question about details of the language, even details that were never considered by the language
designer or semantics writer. Precision and completeness are more important for this purpose than
readability, and formal semantic definitions are not easy to read.
A definition which only experts can read can serve as a standard to determine whether a compiler
implements the standard language, but it is not really adequate for general use. Someone must
study the definition and provide additional explanatory material so that educated nonexperts can
understand it. Following is a quote from Coopers Preface 9 which colorfully expresses the role of
his book in providing a usable definition of Pascal semantics:
The purpose of this manual is to provide a correct, comprehensive, and comprehensible
reference for Pascal. Although the official Standard promulgated by the International
Standards Organization (ISO) is correct by definition, the precision and terseness
required by a formal standard makes it quite difficult to understand. This book is
aimed at students and implementors with merely human powers of understanding, and
only a modest capacity for fasting and prayer in the search for the syntax or semantics
of a domain-type or variant selector.
Coopers book includes the definitions from the ISO standard and provides added explanatory
material and examples. Compiler writers and textbook authors, in turn, can (but too many do not)
use books such as Standard Pascal to ensure that their translations, explanations, and examples
are correct.

4.3.3

The Abstract Machine

In order to make language definitions portable and not dependent on the properties of any particular
hardware, the semantics of a computation tree must be defined in terms of an abstract model of
a computer, rather than some specific hardware. Such a model has elements that represent the
computer hardware, plus a facility for defining and using symbols. It forms a bridge between the
needs of the human and computer. On one hand, it can represent symbolic computation, and on
the other hand, the elements of the model are chosen so that they can be easily implemented on
real hardware.
We describe an abstract machine here which we will use to discuss the semantics of many languages. It has five elements: the program environment, the stack, streams, the shared environment,
and the control.
9

Cooper [1983], p. ix.

4.3. SEMANTICS

93

This abstract machine resembles both the abstract machine underlying FORTH 10 and the SECD
machine that Landin used to formalize the semantics of LISP. 11 Landins SECD machine also has
a stack and a control. Its environment component is our program environment, and our streams
replace Landins dump.
The FORTH model contains a dictionary which implements our program environment. FORTH
has two stacks (for parameters and return addresses) which together implement our stack, except
that no facility is provided for parameter names or local names. 12 The FORTH system defines input
and output from files (our streams) and how a stream may be attached to a program. Finally,
FORTH has an interpreter and a compiler which together define our control element.
Our abstract machine has one element, the shared environment, not present in either the FORTH
model or the SECD machine, as those models did not directly support multitasking.
Program Environment. This environment is the context internal to the program. It includes
global definitions and dynamically allocated storage that can be reached through global objects.
It is the part of the abstract machine that supports communication between any nonhierarchically
nested modules in a single program. Each function, F, exists in some symbolic context. Names are
defined outside of F for objects and other functions. If these names are in Fs program environment,
they are known to F and permit F to refer to those objects and call those functions.
The program environment is implemented by a symbol table (oblist in LISP, dictionary in
FORTH). When a symbol is defined, its name is placed in the symbol table, which connects each
name to its meaning. Predefined symbols are also part of the environment. The meaning of a name
is stored in some memory location, either when the name is defined or later. Either this space itself
(as in FORTH) or a pointer to it (as in LISP) is kept adjacent to the name in the symbol table.
Depending on the language, the meaning may be stored into the space by binding and initialization
and/or it may be changed by assignment.
Shared Environment. This is the context provided by the operating system or program development shell. It is the part of the abstract machine that supports communication between a
program and the outside world. A model for a language that supports multitasking must include
this element to enable communication between tasks. Shared objects are in the environment of two
or more tasks but do not belong to any of them.
Objects that can be directly accessed by the separate, asynchronous tasks that form a job are
part of the shared environment. Intertask messages are examples.
The Stack. The stack is the part of the computation model that supports communication between
the enclosing and enclosed function calls that form an expression. It is a segmented structure of
10

Brodie [1987], Chapter 9.

Landin [1964].
12
The dictionary in FORTH 83 is structured as a list of independent vocabularies, giving some support for local
names.
11

94

CHAPTER 4. FORMAL DESCRIPTION OF LANGUAGE

theoretically unlimited size. The top stack segment, or frame, provides a local environment and
temporary objects for the currently active function. This local environment consists of local names
for objects outside the function (parameters) and for objects inside the function (local variables).
Local environments for several functions can exist simultaneously and will not interfere with each
other. Suspension of one function in order to execute another is possible, with later reactivation of
the first in the same state as when it was suspended.
The stack is implemented by a stack. A stack pointer is used to point at the stack frame (local
environment) for the current function, which points back to a prior frame. A frame for a function
F is created above the prior frame upon entry to F, and is destroyed when F exits. Storage for
function parameters and a function return address are allocated in this frame and initialized (and
possibly later removed) by the calling program.
Upon entry to F, the names of its parameters are added to the local environment by binding
them to the stack locations that were set up by the calling program. The local symbols defined
in F are also added to the environment and bound to additional locations allocated in Fs stack
frame. The symbol table is managed in such a way as to permit these names to be removed from
the environment upon function exit.
Streams. Streams are one medium of communication between different tasks that are parts of
a job. A program exists in the larger context of a computer system and its files. The abstract
machine, therefore, must reflect mass storage and ways of achieving data input and output. A
stream is a model of a sequential file, as seen by a program. It is a sequence, in time, of data
objects, which can be either read or written. Symbolic names for streams and for the files to which
they are bound must be part of the program environment.
The concept of a stream is actually more general than the concept of a sequential file. Suppose
two tasks are running concurrently on a computer system, and the output stream of one becomes
the input stream of the other. A small buffer to hold the output until it is reprocessed can be
enough to implement both streams.
Control. The control section of the abstract model implements the semantic rules of the language
that define the order in which the pieces of the abstract computation tree will be evaluated. It
defines how execution of statements and functions is to begin, proceed, and end, including the
details of sequencing, conditional execution, repetition, and function evaluation. (Chapter 8 deals
with expressions and function evaluation, and Chapter 10 deals with control statements.)
Three kinds of control patterns exist: functional, sequential, and asynchronous. 13 These patterns are supported in various combinations in different languages. Each kind of control pattern is
associated with its own form of communication, as diagrammed in Exhibit 4.9.
Functional control elements communicate with each other by putting parameters on the stack
and leaving results in a return register. In the diagram, functions F1, F2, and F3 are all part of
Process_1 and have associated stack frames on the stack for Process_1. When F3 is entered, its
13

Developed fully in Chapter 8.

95

4.3. SEMANTICS

Exhibit 4.9. Communication between modules.

Program
Stack
F3 g: ?
F2 t: 10

Program
Environment

Shared
Environment

Global variables
Dynamic storage

O.S. storage:
messages, etc.

f: ?

P2: ON

Streams
Out

In

Pipe

F1 s: 30
return: 6
Hierarchical
Execution
F1 Parameter s;
return(s/5);
F2

Parameter t;
f:=F1(t*3);

Sequential
Execution

Local g;
Call F2(10);
F3 g := 2*f+1;
Signal( P2 );

Concurrent
Execution

Wait_for (P2).
Pipe output to
Process 3.

Process 2

Sequential
Execution

Pipe input from

Process 2.

Process 3

Process 1

stack frame is created. Then when F3 calls F2 and F2 calls F1, frames for F2 and F1 are created on
the stack. The frame for F1, indicated by a <, is the current frame. Parameters are initialized
during the function-calling process. When F1 returns it will return a 6 to F2.
Functions within the same process share access to global variables in the program environment
for that process. Sequential constructs in these functions communicate by assigning values to these
variables. Function F2 communicates with F3, and sequential statements in F3 communicate with
each other through the global variable named f in the program environment. F1 will return the
value 6 to F2, which will assign it to a global variable, f. This variable is accessible to F3, which
will use its value to compute g.
Concurrent tasks communicate through the shared environment. Process_1 and Process_2
share asynchronous, concurrent execution and synchronize their operations through signals left in
the shared environment.
Sequential tasks communicate through streams. The output from Process_2 becomes the input
for Process_3. To implement this, the operating system has connected their output and input

96

CHAPTER 4. FORMAL DESCRIPTION OF LANGUAGE

streams through an operating system pipe. This pipe could be implemented either by conveying
the data values to Process_3 as soon as they are produced by Process_2 or by storing the output
in a buffer or a file, then reading it back when the stream is closed.
A Semantic Basis. The formal semantic definition of a language must include specific definitions
of the details of the abstract machine that implements its semantics. Different language models
include and exclude different elements of our abstract machine. Many languages do not support a
shared environment. The new functional languages do not support a program environment, except
for predefined symbols. The control elements, in particular, differ greatly from one language to the
next.
We define the term semantic basis of a language to mean the specific version of the abstract
machine that defines the language, together with the internal data structures and interpretation
procedures that implement the abstract semantics. Layered on top of the semantic basis is the
syntax of the language, which specifies the particular keywords, symbols, and order of elements to
be used to denote each semantic unit it supports.
The semantic basis of a language must define the kinds of objects that are supported, the
primitive actions, and the control structures by which the objects and actions are linked together,
and the ways that the language may be extended by new definitions. The features included in a
semantic basis completely determine the power of a language; items left out cannot be defined by
the programmer or added by using macros. Where two different semantic units provide roughly
the same power, the choice of which to include determines the character of the language and the
style of programs that will be written in it. Thus a wise language designer gives careful thought to
the semantic basis before beginning to define syntax.

4.3.4

Lambda Calculus: A Minimal Semantic Basis

It is perhaps surprising that a very small set of semantic primitives, excluding goto and assignment,
can form an adequate semantic basis for a language. This was proven theoretically by Churchs
work on lambda calculus.14
Lambda calculus is not a programming language and is not directly concerned with computers.
It has no programs or objects or execution as we understand them. It is a symbolic, logical system
in which formulas are written as strings of symbols and manipulated according to logical rules.
We need to be knowledgeable about lambda calculus for three reasons. First, it is a complete
system: Church has shown that it is capable of representing any computable function. Thus any
language that can implement or emulate lambda calculus is also complete.
Second, lambda calculus gives us a starting point by defining a minimal semantic basis for
computation that is mathematically clean. As we examine real computer languages we want to
distinguish between necessary features, nice features (extras), nonfeatures (things that the language
14

Church [1941].

97

4.3. SEMANTICS

Exhibit 4.10. Lambda calculus formulas.

Formulas
x
(x.((yy)x))
(z.(y(z.z)))
((z.(zy))x)

Comments
Any variable is a formula.
Lambda expressions are formulas.
The body of this lambda expression is an application.
Why is this formula an application?

would be better off without), and missing features which limit the power of the language. The
lambda calculus gives us a starting point for deciding which features are necessary or missing.
Finally, an extended version of lambda calculus forms the semantic basis for the modern functional languages. The Miranda compiler translates Miranda code into tree structures which can
then be interpreted by an augmented lambda calculus interpreter. Lambda calculus has taken on
new importance because of the recent research on functional languages. These languages come
exceedingly close to capturing the essence of lambda calculus in a real, translatable, executable
computer language. Understanding the original formal system gives us some grasp of how these
languages differ from C, Pascal, and LISP, and supplies some reason for the aspects of functional
languages that seem strange at first.
Symbols, Functions, and Formulas
There are two kinds of symbols in lambda calculus:
A single-character symbol, such as y, used to name a parameter and called a variable.
Punctuation symbols (, ), ., and .
These symbols can be combined into strings to form formulas according to three simple rules:
1. A variable is a formula.
2. If y is a variable and F is a formula, then (y.F ) is a formula, which is called a lambda
expression; y is said to be the parameter of the lambda expression, and F is its body.
3. If F and G are formulas, then (F G) is a formula, which is called an application.
Thus every lambda calculus formula is of one of three types: a variable, a lambda expression,
or an application. Examples of formulas are given in Exhibit 4.10.
Lambda calculus differs from programming languages in that its programs and its semantic
domain are the same. Formulas can be thought of as programs or as the data upon which programs
operate. A lambda expression is like a function: it specifies a parameter name and has a body that
usually refers to that parameter.15 An application whose first formula is a lambda expression is like
15

The syntax defined here supports only one-argument functions. There is a common variant which permits
multiargument functions. This form can be mechanically converted to the single-argument syntax.

98

CHAPTER 4. FORMAL DESCRIPTION OF LANGUAGE

Exhibit 4.11. Lambda calculus names and symbols.

Formulas
x, y, z, etc.
G = (x.(y(yx)))
H = (GG)

Comments
Single lowercase letters are variables.
A symbolic name may be defined to stand for a formula.
Previously defined names may be used in describing formulas.

a function callthe function represented by the lambda expression is called with the second formula
as an argument. Thus ((x.F )G) intuitively means to call the function (x.F ) with argument G.
However, not all formulas can be interpreted as programs. Formulas such as (xx) or (y(x.z)) do
not specify a computation; they can be thought of as data.
In order to talk about lambda formulas, we will often give them symbolic names. To avoid
confusing our names, which we use to talk about formulas, with variables, which are formulas,
we use uppercase letters when naming formulas. As a shorthand for the statement, let F be the
formula (x.(yx)), we will write simply F = (x.(yx)). If we then write a phrase like, the formula
(F z) is an application, the formula we are talking about is ((x.(yx))z). In general, wherever F
appears, it should be replaced by its definition. Since names are just a shorthand for formulas, a
circular definition such as F = (x.(yF )) is meaningless. Examples of symbols and definitions
are shown in Exhibit 4.11.
As another shorthand, when talking about formulas, we may omit unnecessary parentheses.
Thus we may write x.y instead of (x.y). In general, there may be more than one way to insert
parentheses to make a meaningful formula. For example, x.yx might mean either (x.(yx)) or
((x.y)x). We use the rules that the body of a lambda expression extends as far to the right as
possible, and sequences associate to the left. Thus, in the above example, the body of the lambda
expression is yx, so the fully parenthesized form is (x.(yx)). Examples of these rules are given in
Exhibit 4.12.
Free and Bound Variables. A parameter name is a purely local name. It binds all occurrences
of that name on the right side of the lambda expression. A symbol on the right side of a lambda

Exhibit 4.12. Omitting parentheses when writing lambda calculus formulas.

Shorthand
f xy
x.y.x
x.xy.y
(x.(xx))(zw)
x.y.yzw

Meaning
((f x)y)
(x.(y.x))
(x.(x(y.y)))
((x.(xx))(zw))
(x.(y.((yz)w)))

4.3. SEMANTICS

99

Exhibit 4.13. Lambda expressions for TRUE and FALSE.

Expressions
T = x.y.x

F = x.y.y

Comments
The symbol T represents the logical value TRUE. You should read the definition of T as follows: T is a function of parameters x and y. Its body ignores
y and returns x. (We say the argument y is dropped.)
F names the lambda expression which represents FALSE.

expression is bound if it occurs as a parameter, immediately following the symbol , on the left side
of the same expression or of an enclosing expression. The scope of a binding is the entire right side
of the expression. In Exhibit 4.14, the x defines a local name and binds all occurrences of x in
the expression. We say that each bound occurrence of x refers to the particular x that binds it.
An occurrence of a variable x in F is free if x is not bound. Thus the occurrence of p in
(y.py) is free, but the occurrence of y in that same formula is bound (to y). In the formula
(x(x.((x.x)x))), the variable x occurs five times. The second and third occurrences are bindings;
the other three occurrences are uses. The first occurrence is free, since it does not lie within the
scope of any x-expression. The fourth occurrence is bound to the third occurrence, and the fifth
occurrence is bound to the second occurrence.
These binding rules are the familiar scoping rules of block-structured programming languages
such as Pascal. The operator x declares a new instance of x. All occurrences of x within its scope
refer to that instance, unless x is redeclared by a nested x. In other words, an occurrence of a
variable is always bound to the innermost enclosing block in which x is declared.
Representing Computation
Church invented a way to use lambda formulas to represent computation. He assigned interpretations to certain formulas, making them represent the basic elements of computation. (Some, but
not all, lambda expressions have useful interpretations.) The formulas shown in this chapter are
some of the most basic in Churchs system, including formulas that represent truth values [Exhibit
4.13], the integers [Exhibit 4.15], and simple computations on them [Exhibit 4.16]. More advanced
formulas are able to represent recursion. As you work through these examples the purpose and
mechanics of these basic definitions should become clearer.
Now that we know what lambda calculus formulas are, we need to talk about what they do.
Evaluation rules allow one formula to be transformed to another. A formula which cannot be
transformed further is said to be in normal form. The meaning of a formula is its normal form, if it
has one; otherwise, the formula is undefined. An undefined formula corresponds to a nonterminating
computation. Exhibit 4.14 dissects an expression and looks at its parts.

100

CHAPTER 4. FORMAL DESCRIPTION OF LANGUAGE

Exhibit 4.14. Dissection of a lambda expression.

A lambda expression, with name:
Useful interpretation:

2 = x.y.x(xy)

the number two

Breakdown of elements
2=
Declares the symbol 2 to be a name for the following expression.
x.
The function header names the parameter, x. Everything that follows this . is
the expression body.
y.x(xy) The body of the original expression is another expression with a parameter named
y. Parameter names are purely arbitrary; this expression would still have the
same meaning if it were rewritten with a different parameter name, as in: q.x(xq)
x(xy)
This is the body of the inner expression. It contains a reference to the parameter
y and also references to the parameter x from the enclosing expression.

Reduction. Consider a lambda expression which represents a function. At the abstract level,
the meaning, or semantics, of the expression is the mathematical function that it computes when
applied to an argument. Intuitively, we want to be able to freely replace an expression by a simpler
expression that has the same meaning. The rules for beta and eta reduction permit us to do so.
The main evaluation rule for lambda calculus is called beta reduction and it corresponds to the
action of calling a function on its argument. A beta reducible expression is an application whose left
part is a lambda expression. We also use the term beta redex as a shortening of reducible expression. When a lambda expression is applied to an argument, the argument formula is substituted
for the bound variable in the body of the expression. The result is a new formula.
A second reduction rule is called eta reduction. Eta reduction lets us eliminate one level of
binding in an expression of the form x.f (x). In words, this is a special case in which the lambda
argument is used only once, at the end of the body of the expression, and the rest of the body is
a lambda expression applied to this parameter. If we apply such an expression to an argument,
one beta reduction step will result in the simpler form f (x). Eta reduction lets us make this
transformation without supplying an argument. Specifically, eta reduction permits us to replace
any expression of the form x.f (x), where f represents a function, by the single symbol f .
After a reduction step, the new formula may still contain a redex. In that case, a second
reduction step may be done. When the result does not contain a beta-redex or eta-redex, the
reduction process is complete. We say such a formula is in normal form.
Many lambda expressions contain nested expressions. When such an expression is fully parenthesized it is clear which arguments belong to which function. When parentheses are omitted,
remember that function application associates to the left; that is, the leftmost argument is substituted first for the parameter in the outermost expression.
We now describe in more detail how reduction works. When we reduce a formula (or subformula)
of the form H = ((x.F )G), we replace H by the formula F 0 , where F 0 is obtained from F by

101

4.3. SEMANTICS

Exhibit 4.15. Lambda calculus formulas that represent numbers.

0=
1=
2=

x.y.y
x.y.xy
x.y.x(xy)

The formula for zero has no occurrences of its first parameter in its body. Note that it is the
same as the formula for F . Zero and False are also represented identically in many programming
languages.
The formula for the integer one has a single x in its body, followed by a y. The formula for two
has two xs. The number n will be represented by a formula in which the first parameter occurs n
times in succession.

substituting G for each reference to x in F . Note that if F contains another binding x, the
references to that binding are not replaced. For example, ((x.xy)(zw)) reduces to ((zw)y) and
((x.x(x.(xy)))(zz)) reduces to (zz)(x.(xy)).
When an expression containing an unbound symbol is used as an argument to another lambda
expression, special care must be taken. Any occurrence of a variable in the argument that was
free before the substitution must remain free after the substitution. It is not permitted for a
variable to be captured by an unrelated during substitution. For example, it is not permitted
to apply the reduction rule to the formula ((x.(y.x))(zy)), since y is free in (zy), but after
substitution, that occurrence of y would not be free in (y.(zy)). To avoid this problem, the
parameter must be renamed, and all of its bound occurrences must be changed to the new name.
Thus ((x.(y.x))(zy)) could be rewritten as ((x.(w.x))(zy)), after which the reduction step
would be legal.
Examples of Formulas and Their Reductions
The formulas T and F in Exhibit 4.13 accomplish the equivalent of branching by manipulating
their parameters. They take the place of the conditional statement in a programming language.
T (true) returns its first argument and discards the second. Thus it corresponds to the IF..THEN
statement which evaluates the THEN clause when the condition is true. Similarly, the formula F
(false) corresponds to the IF..ELSE clause. It returns its second parameter just as an IF statement
evaluates the second, or ELSE clause, when the condition is false.
The successor function, S, applied to any integer, gives us the next integer. Exhibit 4.16 shows
the lambda formula that computes this function. Given any formula for a number n, it returns
the formula for n + 1. The function ZeroP (zero predicate) tests whether its argument is equal to
the formula for zero. If so, the result is T , if not, F . Exhibit 4.17 shows how we would call S and
ZeroP . The process of carrying out these computations will be explained later.
Church was able to show that lambda calculus can represent all computation, by representing

102

CHAPTER 4. FORMAL DESCRIPTION OF LANGUAGE

Exhibit 4.16. Basic arithmetic functions.

The successor function for integers. Given the formula for any integer, n, this formula adds
one x and returns the formula for the next larger integer.
S = n.(x.y.nx(xy))
Zero predicate.

This function returns T if the argument = 0 and F otherwise.

ZeroP = n.n(x.F )T

numbers, conditional evaluation, and recursion. Crucial to the power of his system is that there is
no distinction between objects and functions. In fact, objects, in the sense of data objects, were
not defined at all. Expressions called normal forms take their place as concrete things that exist
and can be tested for identity. A formula is in normal form if it contains no redexes. Not all formulas
have a normal form; some may be reduced infinitely many times. These formulas, therefore, do not
represent objects. They are the analog of infinite recursions in computer languages.
For example, let us define the symbol twin to be a lambda expression that duplicates its
parameter:
twin = x.xx
The function twin can be applied to itself as an argument. The application looks like this:
(twin twin)
The preceding line shows this application symbolically. Now we rewrite this formula with the

Exhibit 4.17. Some lambda applications.

An application consists of a function followed by an argument. The first three applications
listed here use the number symbols defined in Exhibit 4.15 and the function symbols defined in
Exhibit 4.16. These three applications are evaluated step-by-step in Exhibits 4.18, 4.19, and 4.20.
(S 1)
Apply the successor function to the function 1.
(ZeroP 0) Apply ZeroP to 0 (Does 0 = zero?)
(ZeroP 1) Does 1 = zero?
((GH) x) Apply formula G to formula H, and apply the result to x.
The last application has the same meaning when written without the parentheses: GHx.

103

4.3. SEMANTICS

name of the function replaced by its definition. Parentheses are used, for clarity, to separate
expressions:
((x.xx)(twin))
This formula contains a redex and so it is not in normal form. When we apply the reduction
rule, the function, x.xx, makes two copies of its parameter, giving:
(twin twin)
Thus the result of reduction is the same as the formula we started with! Clearly, a normal form
can never be reached.
Higher-Order Functions
If lambda calculus were a programming language, we would say that it treats functions as firstclass objects and supports higher-order functions. This means that functions may take functions
as parameters and return functions as results. With this potential we can do some highly powerful
things.
We can define a lambda expression, F , to be the composition of two other expressions, say G
and H. (This means that F is the expression produced by applying G to the result of H.) This
cannot be done in most programming languages. C, for example, permits you to execute a function
G on the result of executing H. But C does not let you write a function that takes two functional
parameters, G and H, and returns a function, F , that will later accept some argument and apply
first H to it and then apply G to the result.
A formula that implements recursion can be defined as the composition of two higher-order
functions. Thus lambda calculus does not need to have recursion built in; it can be defined
within the system. In contrast, recursion is, and must be, built into C and Pascal.
A language with higher-order functions also permits one to curry a function. G is a currying
of F if G has one fewer parameter than F and computes its result by calling F with a constant
in place of the omitted parameter. Currying, combined with generic dispatching, 16 is one way to
implement functions with optional arguments.
Evaluation / Reduction
Any model of computation must represent action as well as objects. Actions are represented in
the lambda calculus by applying the reduction rule, which requires applying the renaming and
substitution rules.
To reduce a formula, F , one finds a subformula, S, anywhere within F , that is reducible. To be
reducible, S must consist of a lambda expression, L, followed by an argument, A. The reduction
process then consists of two steps: renaming and substitution.
16

See Chapter 18.

104

CHAPTER 4. FORMAL DESCRIPTION OF LANGUAGE

Exhibit 4.18. Reducing (S 1).

Compute the Successor of 1. The answer should be 2. For clarity, the formula for one has been
written using p and q instead of x and y. (This is, of course, permitted. The symbols that are used
for bound variables may be renamed any time.)
Write out S.
(n.(x.y.nx(xy))1)
Substitute 1 for n, reduce.
(x.y.1x(xy))
Write out the definition of 1.
(x.y.(p.q.pq)x(xy))
Substitute x for p, and reduce.
(x.y.(q.xq)(xy))
(x.y.x(xy))
Substitute (xy) for q, reduce.
The answer is the formula for 2, which is, indeed, the successor of 1.

Renaming. Renaming is required only if unbound symbols occur in A. They must not have the
same name as Ls parameter. If such a name conflict occurs, the parameter in L must be renamed
so that the unbound symbol will not be captured by Ls parameter. The new name may be any
symbol whatsoever. The formula for L is simply rewritten with the new symbol in place of the old
one.
Substitution. After renaming, each parameter reference on the right side of L is replaced by a
copy of the entire argument-expression, and the resulting string replaces the subexpression S. The
, the dummy parameter, and the . are dropped.
Exhibits 4.18, 4.19, and 4.20 illustrate the reduction process. Three simple formulas are given
and reduced until they are in normal form. The comments on the left in these exhibits document
each choice of redex and the corresponding substitution process. The following explanations are
given so that you may develop some intuition about how these functions work.
Successor. Intuitively, the successor function must take a numeric argument (a nest of two
lambda expressions) and insert an additional copy of the outermost parameter into the middle of
the formula. This is accomplished as follows:
On the first reduction step, the formula for S embeds its argument, n, in the middle of a
nested lambda expression. The symbols x and y in the formula for S are bound by the
lambdas at the left. We rename the bound variables in the formula for n to avoid confusion;
during the reduction process, this p and q will be eliminated.
The formula for n now forms a redex with the x in the tail end of the formula for S. Reducing
this puts as many copies of x into the result as there were copies of p in n. Remember, we
want to end up with exactly one additional copy of x.
This added x comes from the (xy) at the right of the formula for S. The result of the preceding

105

4.3. SEMANTICS

Exhibit 4.19. Reducing (ZeroP 0).

Apply ZeroP to 0, that is, determine whether 0 equals zero. The answer should be T .
Write out ZeroP followed by 0.
((n.n(x.F )T )0)
Substitute 0 for n in the body of ZeroP and reduce.
(0(x.F )T )
Write out the formula for zero.
((x.y.y)(x.F )T )
Substitute (x.F ) for x, and reduce.
((y.y)T )
Substitute T for y, reduce.
T
So 0 does equal 0. Note that the argument, (x.F ), was dropped in the fourth step because the
parameter, x, was not referenced in the body of the function.

reduction forms a redex with this (xy). When we reduce, this final x is sandwiched between
the other xs and the y, as desired.
Essentially, the y in a number is a growth bud that permits any number of xs to be appended
to the string. It would be easy, now, to write a definition for the function plus2.
Zero predicate. Remember, 0 and F are represented by the same formula. Thus the zero
predicate must turn F into T and any other numeric formula into F . (The behavior of ZeroP on
nonnumeric arguments is undefined. Applying ZeroP to a nonnumber is like a type error.) Briefly,
the mechanics of this computation work as follows:
An integer is represented by a formula that is a nest of two lambda expressions.
ZeroP takes its argument, n, and appends two expressions, x.F and T , to n. These two

Exhibit 4.20. Reducing (ZeroP 1).

Write out ZeroP followed by 1.
Substitute 1 for n, reduce.
Write out the formula for 1.
Substitute (x.F ) for x, reduce.
Substitute T for y, reduce.
Substitute T for x and reduce.

((n.n(x.F )T )1)
(1(x.F )T )
((x.y.xy)(x.F )T )
((y.(x.F )y)T )
((x.F )T )
F

On the last line, the parameter x does not appear in the body of the function, so the argument, T ,
is simply dropped. So 1 does not equal 0.
Applying ZeroP to any nonzero number would give the same result, but involve one more
reduction step for each x in the formula.

106

CHAPTER 4. FORMAL DESCRIPTION OF LANGUAGE

Exhibit 4.21. A formula with three redexes.

Assume that P 3 (which adds 3 to its argument) and (which computes the product of two
arguments) have already been defined. (They can be built up out of the successor function.) Then
the formula
( (P 3 4) (P 3 9))
has three reducible expressions: (P 3 4), (P 3 9), and ( (P 3 4) (P 3 9)).

expressions form arguments for the two lambda expressions in n. The entire unit forms two
nested applications.
We reduce the outermost lambda expression first, using the argument x.F . If n is 0, this
argument is discarded because the formula for zero does not contain a reference to its
parameter. For nonzero arguments, this expression is kept.
The inner expression (from the original argument, n) forms an application with the argument
T . If n was zero, this reduces immediately to T . If n was nonzero, there is one more reduction
step and the result is F .
The Order of Reductions
Not every expression has a normal form; some can be reduced forever. But if a normal form exists
it can always be reached by some chain of reductions. When each lambda expression in a formula
is nested fully within another, only one order of reduction is possiblefrom the outside in. But it
is possible to have a formula with two reducible lambda expressions at the same level, side by side
[Exhibit 4.21]. Further, whatever redex you select next, the normal form can still be reached. Put
informally, you cannot back yourself into a corner from which you cannot escape. This important
result is named the Church-Rosser Theorem after the logicians who formally proved it.
Some expressions that do have normal forms contain subexpressions that cannot be reduced to
normal form. This seems like a contradiction until you realize that, in the process of evaluation,
whole sections of a formula may be discarded. For example, in a conditional structure, either
the then part or the else part will be skipped. The computation enclosing the conditional can
still terminate successfully, even if the part that is skipped contains an infinite computation.
By the Church-Rosser theorem, a normal form, if it exists, can be reached by reducing subformulas in any order until there are no reducible subformulas left. However, although you cannot get
blocked in reducing such an expression, you can waste an infinite amount of effort if you persist
in reducing a nonterminating part of the formula. Since any subformula may be discarded by a
conditional, and never need to be evaluated, it is wiser to postpone evaluating a sub-expression until it is needed. If, eventually, a non-terminating sub-formula must be evaluated, then the formula
has no normal form. If, on the other hand, it is discarded, the formula in which this infinite

4.4. EXTENDING THE SEMANTICS OF A LANGUAGE

107

computation was embedded can still be computed (reduced to normal form).

A further theorem proves that if a normal form can be reached, then it can be reached using
the outside-in order of evaluation. That is, at each step the outermost possible redex is chosen.
(The formulas in Exhibits 4.20, 4.19, and 4.18 were all reduced in outside-in order.) This order is
called the normal order of evaluation in lambda calculus and corresponds to call-by-name reduction
order in a programming language.17 It may not be a unique order, since sometimes the outermost
formula is not reducible, but may contain more than one redex side-by-side. In that case, either
may be reduced first.
The Relevancy of Lambda Calculus
Lambda calculus has been proven to be a fully general way to symbolize any computable formula.
Its semantic basis contains representations of objects (normal forms) and functions ( expressions).
Because functions are objects, and higher-order functions can be constructed, the system is able to
represent conditional branching, function composition, and recursion. Computation is represented
by the process of reduction, which is defined by the rules for renaming, parameter substitution,
and formula rewriting.
Although lambda calculus is a formal logical system for manipulating formulas and symbols, it
provides a model of computation that can be and has been used as a starting point for defining
programming languages. LISP was originally designed to be an implementation of lambda calculus,
but it did not capture the outside-in evaluation semantics.

4.4

Extending the Semantics of a Language

Let us define an extension to be a set of definitions which augment a language with an entirely new
facility that can be used in the same way that preexisting facilities are used. Some of the earliest
languages were not very extensible at all. The original FORTRAN allowed variables to be defined
but not types or functions (in a general sense). Function definitions were limited to one line. All
modern languages are extensible in many ways. Any time we define a new object, a new function,
or a new data type, we are extending the language. Each such definition extends the list of words
that are meaningful and adds new expressive power. Pascal, LISP, and the like. are extensible in
this sense: by building up a vocabulary of defined functions and/or procedures, we ultimately write
programs in a language that is much more extensive and powerful than the bare language provided
by the compiler.
Historically, we have seen that extensibility depends on uniform, general treatment of a language
feature. Any time a translator is designed to recognize a specific, fixed set of keywords or defined
symbols, that portion of the language is not extensible. The earliest BASIC was not extensible at all;
even variable names were all predefined (only two-letter names were permitted). FORTRAN, one
of the earliest computer languages, can help us see how the design of a language and a translator
17

See chapter 9, Section 9.2.

108

CHAPTER 4. FORMAL DESCRIPTION OF LANGUAGE

can create barriers to extensibility. We will look at types and functions in early FORTRAN and
contrast them to the extension facilities in more modern languages.
Early FORTRAN supported a list of predefined mathematical functions. The translator recognized calls on those predefined functions, but users could not define their own. This probably
happened because the designers/implementors of FORTRAN provided a static, closed list of function names instead of simply permitting a list that could grow. The mechanics of translating a
function call are also simpler if only one- and two-argument functions have to be supported, rather
than argument lists of unlimited size.
In contrast, consider early LISP. Functions were considered basic (as lambda expressions are
basic in lambda calculus), and the user was expected to define many of them. The language as a
whole was designed to accept and translate a series of definitions and enter each into an extensible
table of defined functions. The syntax for function calls was completely simple and modeled after
lambda calculus, which was known to be completely general. LISP was actually easier to translate
than FORTRAN.
Consider type extensions. In FORTRAN, there were two recognized data types, real and integer.
These were hard wired into the language: variables whose names started with letters I through
N were integers, all other variables were real. On the implementation level, FORTRAN parsers
were written to look at each variable name and deduce the type from it. This was certainly a
convenient system, since it made declarations unnecessary, but it was not extensible. The system fell
apart when FORTRAN was extended to support alphabetic data and double-precision arithmetic.
In contrast, look at Pascal. Pascal has four primitive data types and several ways to build new
simple and aggregate types out of the primitive types. The language has a clear notion of what
a type is, and when a new type is or is not constructed. Each time the programmer uses a type
constructor, a new type is added to the list of defined types. Thereafter, the programmer may use
the new type name in exactly the same ways that primitive type names may be used.
Although Pascal types are extensible, there are predefined, nonextensible relationships among
the predefined types, just as there are in FORTRAN. Integers may be converted to reals, and vice
versa, under specific, predefined circumstances. These conversion relationships are nonextensible;
the triggering circumstances cannot be modified, and similar conversion relationships for other types
cannot be defined. Object-oriented languages carry type-extensibility one step farther, permitting
the programmer to define relationships between types and extend the set of situations in which a
conversion will take place. This is accomplished, in C++ for example, by introducing the notion of
a constructor function, which builds a value of the target type out of components of the original
type. The programmer may define her or his own constructors. The translator will use those
constructors to avoid a type error under specified circumstances, by converting an argument of the
original type to one of the target type.
In all the cases described here, extension is accomplished by allowing the programmer to define
new examples of a semantic category that already exists in the translator. To enable extension, a
new syntax is provided for defining new instances of existing categories. However, the programmer
writes the same syntax for using an extension as for using a predefined facility. Old categories are
extended; entirely new things are not added. Some languages, those with macro facilities, allow

4.4. EXTENDING THE SEMANTICS OF A LANGUAGE

109

the programmer to extend the language by supplying new notation for existing facilities. However,
very few languages support additions or changes to the basic syntactic structure or the semantic
basis of the language. Changing the syntactic structure would involve changing the parser, which is
normally fixed. Changing the semantic basis would involve adding new kinds of tables or procedures
to the translator to implement the new semantics.
What would it mean to extend the syntactic structure of a language? Consider the break
instruction in C and the EXIT in Ada. These highly useful statements enable controlled exits from
the middle of loops. Pascal does not have a similar statement, and an exit from the middle of a
loop can be done only with a GOTO. But the GOTO lacks the safely controlled semantics of break
and EXIT. Because it is so useful, EXIT is sometimes added to Pascal as a nonstandard extension.
Doing this involves extending the parsing phase of the compiler to recognize a new keyword and
modifying the code generation phase to generate a branch from the middle of a loop to the first
statement after the loop. Of course, a programmer cannot extend a Pascal compiler like this. It
can only be done when the compiler is being written.
The ANSI C dialect and the language C++ are both semantic extensions of C. ANSI C extended
the original language by adding type checking for function calls and some coherent operations on
structured data. C++ adds, in addition, semantically protected modules (classes), virtual functions,
and polymorphic domains. This kind of semantic extension is implemented by changing the compiler
and having it do work of a different nature than is done by an old C compiler. These extensions
mentioned required modifying the process of translating a function call, adding new information to
the symbol table, implementing new restrictions on visibility, and adding type checking and type
conversion algorithms.
The code and tables of a compiler are normally off-limits to the ordinary language user. In
most languages, a programmer cannot access or change the compilers tables. The languages EL/1,
FORTH, and T break this rule; EL/118 permitted additions to the compilers syntactic tables, with
accompanying semantic extensions, and FORTH permits access to the entire compiler, including
the symbol table and the semantic interpretation mechanisms.
EL/1 (Extensible Language 1) actually permitted the programmer to supply new EBNF syntax
rules and their associated interpretations. The translator included a preprocessor and a compiler
generator which combined the user-supplied syntax rules with the built-in ones and produced a
compiler for the extended language. The semantic interpretations for the new syntactic rules,
supplied by the user, were then used in the code generation phase.
A very similar thing can be done in T. T is a semantic extension of Scheme which includes data
structuring primitives, object classes, and a macro preprocessor which can be used to extend the
syntax of the language. Each preprocessor symbol is defined by a well-formed T expression. With
these tools, extensions can be constructed that are not possible in C, Pascal, or Scheme. We could,
for example, use the macro facility to define the syntax for a for loop expression and define the
semantics to be a complex combination of initializations, statement executions, increments, and
result-value construction.
18

Wegbreit [1970].

110

CHAPTER 4. FORMAL DESCRIPTION OF LANGUAGE

4.4.1

Semantic Extension in FORTH

We use FORTH to demonstrate the kind of extension that can be implemented by changing the
parser and semantic interpretation mechanisms of a translator. Two kinds of limited semantic
extension are possible in FORTH:
We may add new kinds of information to the symbol table, with accompanying extensions to
the interpreter.
We may modify the parser to translate new control structures.
We shall give an example of each kind of extension below. In both cases, the extension is
accomplished by using knowledge of the actual implementation of the compiler and accessing tables
that would (in most compilers) be protected from user tampering. FORTH has several unusual
features that make it possible to do this kind of extension.
First, like LISP, FORTH is a small, simple language with a totally simple structure. FORTH
books explain the internal structure of the language and details of the operation of the compiler
and interpreter. Second, the designers of FORTH anticipated the desire to extend the rather
rudimentary language and included extension primitives, the words CREATE and DOES>, that
denote a compiler extension, and the internal data structures to implement them.
Finally, FORTH is an interpretive language. The compiler produces an efficient intermediate
representation of the code, not native machine code. Control changes from the interpreter to the
compiler when the interpreter reaches the : at the beginning of a definition, and switches back
to the interpreter when the compiler reaches the ; at the end of the definition. Words are also
included that permit one to suspend a compilation in the middle, interpret some code, and return
to the compilation. Thus variable declarations, ordinary function definitions, segments of code to
be interpreted, and extensions to the compiler can be freely intermixed. The only requirement is
that everything be defined before it is used.
New Types. Unextended, FORTH has three semantic categories, or data types, for items in the
dictionary (symbol table): constant, variable, and function. By using the words CREATE and DOES>
inside what otherwise looks like a normal function definition, more types can be added. CREATE
enters the name of the new type category into the dictionary. Following it must be FORTH code
for any compile-time actions that must be taken to allocate and/or initialize the storage for this
new type. This compile-time section is terminated by the DOES>, which marks this partial entry
as a new semantic category. Finally, the definition includes FORTH code for the semantic routine
that should be executed at run time when items in this category are referenced [Exhibit 4.22].
Having added a type, the FORTH interpreter can be extended to check the type of a function
parameter and dispatch (or execute) one of several function methods, depending on the type. New
data types are additional examples of a category that was built into the language. However, type
checking was not built into FORTH in any way. When we implement type checking, we add a
semantic mechanism to the language that did not previously exist. This is true semantic extension.

111

4.4. EXTENDING THE SEMANTICS OF A LANGUAGE

Exhibit 4.22. Definition in FORTH of the semantics for arrays.

0
1
2
3
4
5
6
7
8
9
10
11
12

2by3array
create
2 , 3 ,
12 allot
does>
rangecheck

(
(
(
(
(
(
(
(
(
(

2by3array box
10 1 2 box !

( Declare and allocate an array variable named box.

( Store the number 10 in box[1,2]. )

linearsub

The ":" marks the beginning of a definition. )

Compile time actions for type declarator 2by3array. )
Store dimensions in the dictionary with the object. )
Allocate 12 bytes for 6 short integers. )
Run time actions to do a subscripted fetch. )
Function call to check that both subscripts are )
within the legal range. )
Function call to compute the effective memory )
address, given base address of array and subscripts.)
End of data type definition. )
)

Program Notes
Comments are enclosed in parentheses.
The definition of the new type declarator goes from line 0 to line 9.
, stores the prior number in the dictionary.
Lines 5 and 7 are calls on the functions rangecheck and linearsub, which the programmer
must define and compile before this can be compiled. Linearsub must leave its result, the
desired memory address, on the stack.
Line 11 declares a 2by3array variable named box. When this line is compiled, the code on
lines 2 and 3 is run to allocate and initialize storage for the new array variable.
Line 12 puts the value 10 on the stack, then the subscripts 1 and 2. When the interpreter
processes the reference to box, the semantic routine for 2by3array (lines 58) is executed. This
checks that the subscripts are within range, then computes a memory address and leaves it
on the stack.
Finally, that address is used by ! to store the 10 that was put on the stack earlier. ! is
the assignment operation. It expects a value and an address to be on the stack and stores
the value in that address.

112

CHAPTER 4. FORMAL DESCRIPTION OF LANGUAGE

Adding a new control structure. CREATE and DOES> provide semantic extension without
corresponding syntactic extension. They permit us to extend the data structuring capabilities of
the language but not to add things like new loops that would require modifying the syntax. To the
extent that the FORTH compilers code is open and documented, though, the clever programmer
can even extend the syntax in a limited way. We have code that adds a BREAK instruction to exit
from the middle of a FIG FORTH loop. This code uses a compiler variable that contains the address
of the end of the loop during the process of compiling the loop. The code for BREAK cannot be
added to FORTH 83. Many compiler variables that were documented in FIG FORTH are kept secret
in the newer FORTH 83. These machine- and implementation-dependent things were taken out of
the language documentation in order to increase the portability of programs written in FORTH,
and the portability of the FORTH translator itself. Providing no documentation about the internal
operation of the compiler prevents the syntax from being extended.

Exercises
1. Briefly define EBNF and syntax diagrams. How are they used, and why are they necessary?
2. Describe the compilation process from source code to object code.
3. Consider the following EBNF syntax. Rewrite this grammar as syntax diagrams.
sneech ::=

* |
( ( hsneechi ) ) |
[ hbanderi ] * hsneechi

bander ::=

{ +$+ | # } | ( % hbanderi )

4. Which of the following sentences are not legal according to the syntax for sneeches, given
in question 3? Why?
a. (*)
f. #####**
b. (+$+*)
g. (+$+#
c. *
h. +$+#*
d. *****
i. *+$+#
e. %%%**
j. %#*+$+**
5. Rewrite the following syntax diagrams as an EBNF grammar.

113

4.4. EXTENDING THE SEMANTICS OF A LANGUAGE

#
blit

[
slit

grit

grit

blit

Y
B

slit
N

slit

grit

6. What is the difference between a terminal and nonterminal symbol in EBNF?

7. What is a production? How are alternatives denoted in EBNF? Repetitions?
8. Using the production rules in Exhibit 4.4, generate the program called easy which has two
variables: a, an integer, and b, a real. The program initializes a to 5. Then b gets the result
of multiplying a by 2. Finally, the value of b is written to the screen followed by a new line.
9. What are the EBNF productions for the conditional statement in Pascal? Show the corresponding syntax diagrams from a standard Pascal reference.
10. Show the syntax diagram for the For statement in Pascal. List several details of the meaning
of the For that are not defined by the syntax diagram.
11. What are semantics?
12. What is the difference between a program environment and a shared environment?
13. What is a stream?
14. Why is lambda calculus relevant in a study of programming language?
15. Show the result of substituting u for x in the following applications. Rename bound variables
where necessary.
a. ((x.y.x)u)
b. ((x.y.z)u)
c. ((x.u.ux)u)
d. ((x.x.ux)u)

114

CHAPTER 4. FORMAL DESCRIPTION OF LANGUAGE

16. Each item below is a lambda application. We have used a lot of parentheses to help you parse
the expressions. Reduce each formula, until no redex remains. One of the items requires
renaming of a bound variable.
a. ((x.y.x(xy))(pq)q)
b. ((x.y.y)(pq)q)
c. ((z.(y.yz))(x.xy))
d. ((x.y.y(xy))(p.pp)q)
17. Verify the following equality. Start with the left-hand side and substitute the formula for
twice. Then reduce the formula until it is in normal form. This may look like a circular
reduction, but the formula reaches normal form after eight reduction steps.
Let twice = f.x.f (f x).
Show that twice twice gz = g(g(g(gz))).
Hints: Write out the formula for twice only when you are using it as a function; keep arguments
in symbolic form. Each time you write out twice, use new names for the bound variables. Be
careful of the parentheses. Remember that function application associates to the left.
18. Show that 3 is the successor of 2, using the lambda calculus representations defined for
integers.
19. Define the function plus2 using a lambda formula. Demonstrate that your formula works
by applying it to the formula for 1.
20. Construct a lambda formula to express the following conditional expression. (Assume that
x is a Boolean value, T or F .) Verify the correctness of your expression by applying it to T
and F and reducing to get 0 or 2.
If x is true then return 0 else return 2.
21. How do EL/1 and FORTH allow the semantics of the languages to be extended?

Part II

Describing Computation

115

Chapter 5

Primitive Types

Overview
This chapter explains the concept of types within programming languages and the
hardware that supports these types. Computer memory is an array of bits usually
grouped into addressable 8-bit segments called bytes. Words are groups of bytes, usually
2, 4, and sometimes 8 bytes long. All data types in programming languages must be
mapped onto the bytes and words of the machine. Logical computer instructions operate
on bytes and words, but other instructions operate on objects that are represented by
codes which are superimposed on bit strings. Common codes include ASCII, EBCDIC,
binary integer, packed decimal and floating point.
A data type is an abstraction: a description of a set of properties independent of any
specific object that has those properties. A previously defined type is referred to by a
type name. A type description identifies the parts of a nonprimitive type. A specific
type is a homogeneous set of objects, while a generic type is a set that includes objects
of more than one specific type. Each type is a set of objects with an associated set of
functions. A type defines the representation for program objects. Several attributes are
defined by the type of an object, including encoding, size, and structure.
Every language supports a set of primitive data types. Usually these include integer,
real, Boolean, and character or string. A language standard determines the minimum
set of primitive types that the language compiler must implement. Hardware characteristics influence which types a language designer chooses to make primitive. If the
hardware does not support a required type, that type may have to be emulated, that
is, implemented in the software.
Type declarations have a long history, going back to the earliest languages which sup117

118

CHAPTER 5. PRIMITIVE TYPES

ported primitive types built into the instruction set of the host machine. By the late
1960s, types were recognized as abstractions. Type declarations defined value constructors, selectors, and implicit type predicates. In the 1970s, with the emergence of Ada,
types were treated as objects in a limited way. There were cleaner type compatibility
rules, support for type portability, and explicit constraint on the values of a type. Recent research includes the issues of type hierarchies with inherited properties and the
implementation of nonhomogeneous types in a semantically sound way.

5.1

Primitive Hardware Types

A translator must map a programmers objects and operations onto the storage and instructions
provided by the computer hardware. To understand the primitive types supported by languages,
one should also understand the hardware behind those types.

5.1.1

Bytes, Words, and Long Words

Computer memory is a very long array of bits, normally organized into groups. 1 Each group has
an address, used to store and fetch data from it. Modern machines usually have 8-bit bytes and are
byte addressable. Bytes are grouped into longer 2- and 4-byte units called words and long words.
Some machines have a few hardware instructions that support double word, or 8-byte, operations.
Bytes and words form the basis for all representation and all computation. They are the
primitive data type onto which all other data types must be mapped.
Computer instruction sets include some instructions that operate on raw, uninterpreted bytes
or words. These are called logical instructions. They include right and left shifts, and bitwise
complement, and, or, and exor operations.
Most computer instructions, though, are intended to operate on objects other than bit strings,
such as numbers or characters. All objects must be represented by bit strings, but they have
semantics above and beyond the bits that represent them. These objects are represented by codes
that are superimposed on the bit strings. Common encodings include ASCII, EBCDIC, binary
integer, packed decimal, and floating point.

5.1.2

Character Codes

Long before IBM built computers, it built unit record equipment, which processed data recorded
on punched cards. Keypunches were used to produce this data, and line printers could copy cards
to fanfold paper. Tabulating machines were used to process the data. These had plug boards,
1

The Burroughs memory of the B1700/B1800 series of computers was an undivided string of bits that was actually
bit-addressable.

5.1. PRIMITIVE HARDWARE TYPES

119

on which a skilled person could build programs by using plug-in cables to connect holes that
represented card columns to holes that represented functions such as + and . 2
Punched cards were in common use before computers were invented and quite naturally became
the common input medium for computers. The Hollerith character code, used for punched cards,
was adapted for use in computers, and called Binary Coded Decimal, or BCD. Hollerith code was
a decimal code. It used one column with twelve punch positions to represent each digit. (These
positions were interpreted as +, , 0...9.) Alphabetic letters were represented as pairs of punches,
one in the zone area (+, , 0), and one in the digit area (1..9). This gives 27 combinations,
which is one too many for our alphabet, and the 0:1 punch combination was not used for any letter.
The tradition was that this combination was omitted from the alphabet because the two closely
spaced punches made it physically weak. However, this combination was used to represent /.
Thus the alphabet had a nonalpha character in its middle.
The entire Hollerith character set had no more than 64 codes. Letters and digits accounted for
36 of these; the rest were other punctuation and control characters and were represented by double
or triple punches. The BCD code used sixty four 6-bit codes to represent this character set.
The BCD character code was reflected in various ways in the computer hardware of the 1950s
and early 1960s. It has always been practical to make word size a multiple of the character code
size. Hardware was built with word lengths of 24, 36, 48, and 60 bits (making 4, 6, 8, and 10
characters per word). Floating-point encoding was invented for the IBM 704; its 36-bit words were
long enough to provide adequate range and precision.
Software, also, showed the effects of this character code. FORTRAN was designed around this
severely limited character set (uppercase only, very few available punctuation symbols). FORTRAN
identifiers were limited to six characters because that is what would fit into one machine word on
the IBM 704 machine. COBOL implemented numeric data input formats that were exactly like
Hollerith code. If you wanted to input the number 371, you punched only three columns, and put
the sign over the rightmost, giving the number code 37J. The number +372 was encoded
as 37B. This wholly archaic code is still in use in COBOL today and is increasingly difficult to
explain and justify to students.
Dissatisfaction with 6-bit character codes was rampant; sixty four characters are just not enough.
People, reasonably, wanted to use both upper- and lower-case letters, and language designers felt
unreasonably restricted by the small set of punctuation and mathematical symbols that BCD
provided. Two separate efforts in the early 1960s produced two new codes, EBCDIC (Extended
BCD Interchange Code) and ASCII (American Standard Code for Information Interchange).
EBCDIC was produced and championed by IBM. It was an 8-bit code, but many of the 256
possible bit combinations were not assigned any interpretation. Upper- and lowercase characters
were included, with ample punctuation and control characters. This code was an extension of
BCD; the old BCD characters were mapped into EBCDIC in a systematic way. Certainly, that
made compatibility with old equipment less of a problem.
2

These were archaic in the early 1960s, but a few early computer science students had the privilege of learning to
use them.

120

CHAPTER 5. PRIMITIVE TYPES

Unfortunately, the EBCDIC code was not a sensible code because the collating sequence
was not normal alphabetical order. 3 Numbers were greater than letters, and like BCD, alphabetic
characters were intermingled with nonalphabetic characters.
ASCII code grew out of the old teletype code. It uses seven bits, allowing 128 characters. Upperand lowercase letters, numerals, many mathematical symbols, a variety of useful control characters,
and an escape are supported. The escape could be used to form 2-character codes for added
items.4 ASCII is a sensible code; it follows the well-established English rules for alphabetization.
It has now virtually replaced EBCDIC, even on IBM equipment.
An extended 8-bit version of ASCII code is now becoming common. It uses the additional
128 characters for the accented and umlauted European characters, some graphic characters, and
several Greek letters and symbols used in mathematics. Hardware intended for the international
market supports extended ASCII.

5.1.3

Numbers

We take integers and floating-point numbers for granted, but they are not the only ways, or even
the only common and useful ways, to represent data.
In the late 1950s and early 1960s, machines were designed to be either scientific computers or
business computers. The memory of a scientific computer was structured as a sequence of words
(commonly 36 bits per word) and its instruction set performed binary arithmetic. Instructions were
fixed length and occupied one word of memory.
Packed Decimal
The memory of a business computer was a series of BCD bytes with an extra bit used to mark the
beginning of each variable-length word. Objects and instructions were variable length. Numbers
were represented as a series of decimal (BCD) digits. Arithmetic was done in base ten, not base
two.
The distinction between scientific and business computers profoundly affected the design of programming languages. COBOL, a business language, was oriented toward variable-length objects
and supported base ten arithmetic. In contrast, FORTRAN was a scientific language. Its data
objects were one word long, or arrays of one-word objects, and computation was done either in
binary or in floating point. Characters were not even a supported data type.
In 1964, IBM introduced a family of computers with innovative architecture, intended to serve
both the business and scientific communities. 5 The memory of the IBM 360 was byte-addressable.
The hardware had general-purpose registers to manipulate byte, half-word (2-byte), and word
(4-byte) sized objects, plus four 8-byte registers for floating-point computation. The instruction
3
The collating sequence of a code is the order determined by the < relationship. To print out a character code
in collation order, start with the code 00000000, print it as a character, then add 1 and repeat, until you reach the
largest code in the character set.
4
It is often used for fancy I/O device control codes, such as reverse video on.
5
Gorsline [1986], p. 317.

5.1. PRIMITIVE HARDWARE TYPES

121

Exhibit 5.1. Packed-decimal encoding.

A number is a string of digits. The sign may be omitted if the number is positive, or
represented as the first or last field of the string.
Each decimal digit is represented by 4 bits, and pairs of digits are packed into each byte. The
string may be padded on the left to make the length even.
The code 0000 represents the digit 0, and 1001 represents 9. The six remaining possible
bit patterns, 1010 . . . 1111, do not represent legal digits.

set supported computation on binary integers, floating point, and integers represented in packed
decimal with a trailing sign [Exhibit 5.1].
Many contemporary machines support packed decimal computation. Although the details of
packed-decimal representations vary somewhat from machine to machine, the necessary few instructions are included in the Intel chips (IBM PC), the Motorola chips (Apollo workstations, Macintosh,
Atari ST), and the Data General MV machines.
Packed-decimal encoding is usually used to implement decimal fixed-point arithmetic. A decimal
fixed-point number has two integer fields, one representing the magnitude, the other the scale (the
position of the decimal point). The scale factors must be taken into account for every arithmetic
operation. For instance, numbers must be adjusted to have the same scale factors before they can
be added or subtracted. Languages such as Ada and COBOL, which support fixed-point arithmetic,
do this adjustment for the programmer. 6
Base two arithmetic is convenient and fast for computers, but it cannot represent most base ten
fractions exactly. Furthermore, almost all input and output is done using base ten character strings.
These strings must be converted to/from binary during input/output. The ASCII to floating point
conversion routines are complex and slow.
Arithmetic is slower with packed decimal than with binary integers because packed-decimal
arithmetic is inherently more complex. Input and output conversions are much faster; a packeddecimal number consists of the last 4 bits of each ASCII or EBCDIC digit, packed two digits per
byte. Arithmetic is done in fixed point; a specified number of digits of precision is maintained, and
numbers are rounded or truncated after every computation step to the required precision. Control
over rounding is easy, and no accuracy is lost in changing the base of fractions.
In a data processing environment, packed decimal is often an ideal representation for numbers.
Most business applications do more input and output than computation. Some, such as banking
6

Unfortunately, the Ada standard does not require that fixed-point declarations be implemented by decimal fixedpoint arithmetic! It is permissible in Ada to approximate decimal fixed-point computation using numbers represented
in binary, not base ten, encoding!

122

CHAPTER 5. PRIMITIVE TYPES

Exhibit 5.2. Representable values for signed and unsigned integers.

These are the smallest and largest integer values representable on a twos complement machine.
Type
Signed

Length
4 bytes
2 bytes
1 byte

Minimum
-2,147,483,648
-32,768
-128

Maximum
2,147,483,647
32,767
127

Unsigned

4 bytes
2 bytes
1 byte

0
0
0

4,294,967,295
65,535
255

and insurance computations, require total control of precision and rounding during computation
in order to meet legal standards. For these applications, binary encoding for integers and floatingpoint encoding for reals is simply not appropriate.
Binary Integers
Binary numbers are built into modern computers. However, there are several ways that binary
numbers can be represented. They can be different lengths (2- and 4-byte lengths being the most
common), and be signed or unsigned. If the numbers are signed, the negative values might be
represented in several ways.
Unsigned integers are more appropriate than signed numbers for an application that simply
does not deal with negative numbers, for example, a variable representing a machine address or a
population count. Signed and unsigned numbers of the same length can represent exactly the same
number of integers; only the range of representable numbers is different [Exhibit 5.2]. On a modern
twos complement machine, unsigned arithmetic is implemented by the same machine instructions
as signed arithmetic.
Some languages, for example C, support both signed and unsigned integers as primitive types.
Others, for example Pascal and LISP, support only signed integers. Having unsigned integer as a
primitive type is not usually necessary. Any integer that can be represented as an unsigned can
also be represented as a signed number that is one bit longer. There are only a few situations in
which this single bit makes a difference:
An application where main storage must be conserved and a 1-byte or 2-byte integer could
be used, but only if no bit is wasted on the sign.
An application where very large machine addresses or very large numbers must be represented
as integers, and every bit of a long integer is necessary to represent the full range of possible
values.

5.1. PRIMITIVE HARDWARE TYPES

123

The intended application area for the language involves extensive use of the natural numbers
(as opposed to the integers). By using type unsigned we can constrain a value to be
nonnegative, thereby increasing the explicitness of the representation and the robustness of
the program.
Unsigned will probably be included as a primitive type in any language whose intended
applications fit one of these descriptions. C was intended for systems programming, in which access
to all of a machines capabilities is important, and so supports unsigned as a primitive type. 7
Signed Binary Integers The arithmetic instructions of a computer define the encoding used for
numbers. The ADD 1 instruction determines the order of the bit patterns that represent the integers.
Most computers count in binary, and thus support binary integer encoding. Most compilers use
this encoding to represent integers. Although this is not the only way to represent the integers,
binary is a straightforward representation that is easy for humans to learn and understand, and it
is reasonably cheap and fast to implement in hardware. 8
Large machines support both word and long word integers; very small ones may only support
byte or word sized integers. On such machines, a compiler writer must use the short word instructions to emulate arithmetic on longer numbers that are required by the language standard.
For example, the instruction set on the Commodore 64 supported only byte arithmetic, but Pascal translators for the Commodore implemented 2-byte integers. Adding a pair of 2-byte integers
required several instructions; each half was added separately and then the carry was propagated.
Negative Numbers One early binary integer representation was sign and magnitude. The leftmost bit was interpreted as the sign, and the rest of the bits as the magnitude of the number. The
representations for +5 and 5 differed only in one bit. This representation is simple and appealing
to humans, but not terrific for a computer. An implementation of arithmetic on sign-and-magnitude
numbers required a complex circuit to propagate carries during addition, and another one to do
borrowing during subtraction.
CPU circuitry has always been costly, and eventually designers realized that it could be made less
complex and cheaper by using complement notation for negative numbers. Instead of implementing
+ and , a complement machine could use + and negate. Subtraction is equivalent to
negation followed by addition. Negation is trivially easy in ones complement representationjust
flip the bits. Thus 00000001 represented the integer 1 and 11111110 represented the integer
negative one. A carry off the left end of the word was added back in on the right. The biggest
drawback of this system is that zero has two representations, 0000000 (or +0) and 11111111
(or 0).
A further insight occurred in the early 1960s: complement arithmetic could be further simplified by using twos complement instead of ones complement. To find the twos complement of a
7

We must also note that the primitive type byte or bitstring is lacking in C, and unsigned is used instead.
While this is semantically unattractive, it works.
8
Other kinds of codes have better error correction properties or make carrying easier.

124

CHAPTER 5. PRIMITIVE TYPES

Exhibit 5.3. IEEE floating-point formats.

Format Name
Short real
Long real
Temporary real

Length
4 bytes
8 bytes
10 bytes

Sign
31
63
79

Bit fields in representation

Exponent
Mantissa
3023
220, with implicit leading 1
6252
510, with implicit leading 1
7864
630, explicit leading 1

number, complement the bits and add one. The twos complement of 00000001 (representing 1)
is 11111111 (representing 1). Twos complement representation has two good properties that
are missing in ones complement: there is a unique representation for zero, 00000000, and carries
off the left end of a sum can simply be ignored. Twos complement encoding for integers has now
become almost universal.
Floating Point
Many hardware representations of floating-point numbers have been used in computers. Before the
advent of ASCII code, when characters were 6 bits long, machine words were often 36 or 48 bits
long. (It has always been convenient to design a machines word length to be a multiple of the byte
length.) Thirty-six bits is enough to store a floating-point number with a good range of exponents
and about eight decimal digits of precision. Forty-eight or more bits allows excellent precision.
However, word size now is almost always 32 bits, which is a little too small.
In order to gain the maximum accuracy and reasonable uniformity among machines, the IEEE
has developed a standard for floating-point representation and computation. In this discussion,
we focus primarily on this standard. The IEEE standard covers all aspects of floating pointthe
use of bits, error control, and processor register requirements. It sets a high standard for quality.
Several modern chips, including the IBM 8087 coprocessor, have been modeled after it.
To understand floats, you need to know both the format and the semantics of the representation.
A floating-point number, N , has two parts, an exponent, e, and a mantissa, m. Both parts are
signed numbers, in some base. If the base of the exponent is b, then N = m b e .
The IEEE standard supports floats of three lengths: 4, 8, and 10 bytes. Let us number the bits
of a float starting with bit 0 on the right end. The standard prescribes the float formats shown in
Exhibit 5.3. The third format defines the form of the CPU register to be used during computation.
Exhibit 5.4 shows how a few numbers are represented according to this standard.
The sign bit, always at the left end, is the sign of the entire number. A 1 is always used for
negative, 0 for a positive number.
The exponent is a signed number, often represented in bias notation. A constant, called the bias,
is added to the actual exponent so that all exponent values are represented by unsigned positive
numbers. In the case of bias 128, this is like twos complement with the sign bit reversed.
The advantage of a bias representation is that, if an ordinary logical comparison is made,

125

5.1. PRIMITIVE HARDWARE TYPES

Exhibit 5.4. Floating point on the SparcStation.

The SparcStation is a new RISC workstation built by Sun Microsystems. It has a floating-point
coprocessor modeled after the IEEE standard. Using the Sun C compiler to explore its floatingpoint representation, we find that the C float type is implemented by a field with the IEEE 4-byte
encoding. A few numbers are shown here with their representations printed in hex notation and in
binary with the implied 1-bit shown to the left of the mantissa.
Binary Representation
Exponent Mantissa
00000000 0.0000000 00000000 00000000

Decimal
0.00

Hex
00000000

Sign
0

0.25
0.50
1.00
-1.00

3E800000
3F000000
3F800000
BF800000

0
0
0
1

01111101
01111110
01111111
01111111

1.0000000
1.0000000
1.0000000
1.0000000

00000000
00000000
00000000
00000000

00000000
00000000
00000000
00000000

10.00
5.00
2.50
1.25

41200000
40A00000
40200000
3FA00000

0
0
0
0

10000010
10000001
10000000
01111111

1.0100000
1.0100000
1.0100000
1.0100000

00000000
00000000
00000000
00000000

00000000
00000000
00000000
00000000

positive numbers are greater than negative numbers. Absolutely no special provision needs to be
made for the sign of the number. With 8 bits in the exponent, 00000000 represents the smallest
possible negative exponent, and 11111111 is the largest positive exponent. 10000000 generally
represents an exponent of either zero or one. When interpreted as a binary integer, 10000000
is 128. If this represents an exponent of zero, we say that the notation is bias 128, because
128 0 = 128. When 10000000 represents an exponent of one, we say the notation is bias 127,
because 128 1 = 127.
In the IEEE standard, the exponent is represented in bias 127 notation, and the exponent
10000000 represents +1. This can be seen easily in Exhibit 5.4. The representation for 2.50 has
an exponent of 10000000. The binary point in 1.010000 must be moved one place to the right
to arrive at 10.1, the binary representation of 2.50. Thus 10000000 represents +1.
Floating-point hardware performs float operations in a very long register, much longer than the
24 bits that can be stored in a float. To maintain as many bits of precision as possible, the mantissa
is normalized after every operation. This means that the leading 0 bits are shifted to the left
until the leftmost bit is a 1. Then when you store the number, all bits after the twenty-fourth
are truncated (discarded). A normalized mantissa always starts with a 1 bit, therefore this bit
has no information value and can be regenerated by the hardware when needed. So only bits 224
of the mantissa are stored, in bits 220 of the float number.
The mantissa is a binary fraction with an implied binary point. In the IEEE standard, the
point is between the implied 1 bit and the rest of the mantissa. Some representations place the

126

CHAPTER 5. PRIMITIVE TYPES

Exhibit 5.5. A hierarchy of abstractions.

Electronic Device
Computer

Minicomputer

Microcomputer
Apple IIE

TRS-80

Microphone

IBM PC

Amana

Microwave Oven
Litton

DG MV8000

binary point to the left of the implied 1 bit. These interpretations give the same precision but
different ranges of representable numbers.

5.2
5.2.1

Types in Programming Languages

Type Is an Abstraction

An abstraction is the description of a property independent from any particular object which has
that property. Natural languages contain words that form hierarchies of increasing degrees of
abstraction, such as TRS-80, microcomputer, computer, and electronic device [Exhibit
5.5]. TRS-80 is itself an abstraction, like a type, describing a set of real objects, all alike. Most
programming language development since the early 1980s has been aimed at increasing the ability
to express and use abstractions within a program. This work has included the development of
abstract data types, generic functions, and object-oriented programming. We consider these topics
briefly here and more extensively later.
A data type is an abstraction: it is the common property of a set of similar data objects. This
property is used to define a representation for these objects in a program. Objects are said to
have or to be of that type to which they belong. Types can be primitive, defined by the
system implementor, or they can be programmer defined. We refer to a previously defined type by
using a type name. A type declaration defines a type name and associates a type description with
it, which identifies the parts of a nonprimitive type [Exhibit 5.6]. The terms type and data type
are often used loosely; they can refer to the type name, the type description, or the set of objects
belonging to the type.
If all objects in a type have the same size, structure, and semantic intent, we call the type
concrete or specific. A specific type is a homogeneous set of objects. All the primitive types in
Pascal are specific types, as are Pascal arrays, sets, and ordinary records made out of these basic
types. A variant record in Pascal is not a specific type, since it contains elements with different
structures and meanings.

5.2. TYPES IN PROGRAMMING LANGUAGES

127

Exhibit 5.6. Type, type name, and type description.

Real-world objects: a set of rectangular boxes.
Type: We will represent a box by three real numbers: its length, width, and depth.
Type name declared in Pascal: TYPE box_type =
Possible type descriptions in Pascal:
ARRAY [1..3] OF real
RECORD length, width, depth:

real END

A generic domain is a set that includes objects of more than one concrete type [Exhibit 5.7].
A specific type that is included in a generic domain is called an instance or species of the generic
domain, as diagrammed in Exhibit 5.8. Chapters 15 and 17 explore the subjects of type abstraction
and generic domains.

5.2.2

A Type Provides a Physical Description

The properties of a type are used to map its elements onto the computers memory. Let us focus
on the different attributes that are part of the type of an object. These include encoding, size, and
structure.
Exhibit 5.7. Specific types and generic domains.
Specific types:

Integer arrays of length 5

Character arrays of length 10
Real numbers
Integer numbers

Generic domains:
Intarray: The set of integer arrays, of all lengths.
Number: All representations on which you can do arithmetic, including floating point, integer,
packed decimal, etc.

128

CHAPTER 5. PRIMITIVE TYPES

Exhibit 5.8. Specific types are instances of generic domains.

The generic domain Number has several specific subtypes, including Real, Integer, and Complex.
Objects (variables) have been declared that belong to these types. Objects named V, W,
and X belong to type Real; objects J and K belong to type Integer, and C belongs to type
Complex.
All six objects also belong to the generic domain Number.
Number
Real
V

Complex
C

Integer
J

Encoding. The instruction set of each machine includes instructions that do useful things on
certain encodings (bit-level formats) of data. For example, the Data General MV8000 has instructions that perform addition if applied to numbers encoded with 4-bits per decimal digit. Because
of this built-in encoding, numbers can be conveniently represented in packed-decimal encoding
in Data General COBOL. Where an encoding must be implemented that is not directly supported
by the hardware, the implementation tends to be inefficient.
Size. The size of an object can be described in terms of hardware quantities such as words or
bytes, or in terms of something meaningful to the programmer, such as the range of values or the
number of significant digits an object may take on.
Structure. An object is either simple or it is compound. A simple object has one part with no
subparts. No operators exist within a language that permit the programmer to decompose simple
objects. In a language that has integer as a simple type, integer is generally undecomposable.
In standard Pascal, integers are simple objects, as are reals, Booleans, and characters. In various
Pascal extensions, though, an integer can be decomposed into a series of bytes. In these dialects
integer is not a simple type. Primitive types may or may not be simple. In both cases, integer
is a primitive type; that is, it is a predefined part of the language.
A compound object is constructed of an ordered series of fields of specific types. A list of these
fields describes the structure of the object. If the fields of the compound object all have the same
type, it is a homogeneous compound. These are commonly called array, vector, matrix,

5.2. TYPES IN PROGRAMMING LANGUAGES

129

or string. The dimensions of an array and its base type (the type of its elements) define its
structure.
If the fields of a compound object have different types, it is a heterogeneous compound. These
are commonly called records or structures. An ordered list of the types of each field of a record
defines its structure.
The distinctions among structure, encoding, and size are seen most clearly in COBOL, where
these three properties are specified separately by the programmer.
Structure in COBOL. The internal structure of each data object is defined by listing its fields
and subfields, in order. The subfields of a field are listed immediately following the field and given
higher level numbers to indicate that they are subordinate to it.
Encoding in COBOL. Character data has only one encoding: the character code built into the
machine hardware. Depending on the compiler, several encodings may be provided for numbers,
with DISPLAY being the default. The COBOL programmer may specify the encoding in a USAGE
clause. In Data General COBOL, the programmer can choose from the following set:
DISPLAY
COMPUTATIONAL
COMP-2
COMP-3

ASCII or EBCDIC characters

binary fixed point
packed binary-coded-decimal fixed point
floating point

Double-precision encoding is also provided in some COBOL implementations. Each encoding

has inherent advantages, which must be understood by the programmer. Input and output require
operands of DISPLAY usage. Arithmetic can be done on all usages except DISPLAY. The most
efficient numeric I/O conversion is between DISPLAY and COMP-2. The most efficient arithmetic is
done in COMPUTATIONAL.
Conversion from one encoding to another is performed automatically when required in COBOL.
If a numeric variable does not have the default usage, DISPLAY, conversion is performed during the
input and output processes, as in most languages. If a numeric variable represented in DISPLAY
usage is used in an arithmetic statement, it will be converted to packed decimal. (This conversion
is fast and efficient). The arithmetic will be done in packed-decimal encoding, and the result will
be converted back to display usage if it is stored in a DISPLAY variable.
Size in COBOL. Size is defined by supplying a PICTURE clause for every field that has no subfields [Exhibit 5.9]. The PICTURE illustrates the largest number of significant characters or decimal
digits that will ever be needed to represent the field. Note that the programmer describes the size
of the object being represented, not the size, in bytes, of the representation. Different amounts of
storage could be allocated for equal size specifications with different encoding specifications.
At the other extreme from COBOL, the language BASIC permits the programmer to specify
only whether the object will encode numeric or alphanumeric objects, and to declare the structure
of arrays (number of dimensions and size of each). The encoding is chosen by the translator and
hidden from the user. Thus BASIC is simpler to use. It frees the programmer from concern about
the appropriateness of the encoding. At the same time, it provides no easy or efficient control over

130

CHAPTER 5. PRIMITIVE TYPES

Exhibit 5.9. Size and encoding specifications in COBOL.

Three simple variables are defined, named PRICE, DISCOUNT, and ITEM.
01 PRICE
PICTURE 999V99.
01 DISCOUNT PICTURE V999 USAGE COMPUTATIONAL.
01 ITEM
PICTURE XXXX.
PRICE has a numeric-character encoding, indicated by the 9s in the PICTURE clause and the
absence of a USAGE clause. The size of this variable is defined by the number of 9s given, and
decimal position is marked by the V. In this case, the number has two decimal places and is less
than or equal to 999.99.
DISCOUNT has binary fixed-point encoding (because of the USAGE clause). Its size is three decimal
digits, with a leading decimal point.
ITEM has alphanumeric encoding, indicated by the Xs in its PICTURE. Its size is four characters.
Any alphanumeric value of four or fewer characters can be stored in this variable.

precision and rounding. BASIC is thus a better tool for the beginner, but a clumsy tool for the
professional.

5.2.3

What Primitive Types Should a Language Support?

The usual set of primitive data types in a language includes integer, real, Boolean, and character
or string. However, Ada has many more and BASIC has fewer.
A language standard determines the minimum set of primitive types that must be implemented
by a compiler. Choosing this set is the job of the language designer. A language implementor may
choose to support additional types, however. For example, Turbo Pascal supports a type string
that is not required by the standard. The string type is a language extension.
The decision to make a type primitive in a computer language is motivated by hardware
characteristics and the intended uses of the language. Compromises must often be made. A language designer must decide to include or exclude a type from the primitive category by considering
the cost of implementing and using it [Exhibit 5.10] as opposed to the cost of not implementing it
[Exhibit 5.11].
Types that are not primitive sometimes cannot be implemented efficiently, or even implemented
at all, by the user. For example, the ANSI C standard does not support packed-decimal numbers. A
user could write his or her own packed decimal routines in C. To achieve adequate precision the user
would probably map them onto integers, not floats. Masking, base 10 addition and multiplication,
carrying, and the like could be implemented. However, the lack of efficiency in the finished product
would be distressing, especially when you consider that many machines provide efficient hardware
instructions to do this operation.
If users are expected to need a certain type frequently, the language is improved by making
that type primitive. Packed decimal is not a primitive type in C because the intended usage of C

5.2. TYPES IN PROGRAMMING LANGUAGES

131

Exhibit 5.10. Costs of implementing a primitive type.

Every added feature complicates both the language syntax and semantics. Both require added
documentation. If every useful feature were supported, the language would become immense
and unwieldy.
Standardization could become more difficult, as there is one more item about which committee
members could disagree. This could be an especially severe problem if a type is complex, its
primitive operations are extensive, or it is unclear what the ideal representation should be.
The compiler and/or library and/or run-time system become more complex, harder to debug,
and consume more memory. Literals, input and output routines, and basic functions must be
defined for every new primitive type.
If typical hardware does not provide instructions to handle the type, it may be costly and
inefficient to implement it. Perhaps programmers should not be encouraged to use inefficient
types.

Exhibit 5.11. Costs of omitting a primitive type.

Inefficiency: failing to include an operation that is supported by the hardware leads to a huge
increase in execution time.
Language structure may be inadequate to support the type as a user extension, as Pascal
cannot support variable-length strings or bit fields with bitwise operators.
Some built-in functions such as READ, WRITE, assignment, and comparison are generic in
nature. They work on all primitive types but not necessarily on all user-defined types. If
these functions cannot be extended, a user type can never be as convenient or easy to use as
a primitive type.
Primitive types have primitive syntax for writing literals. Literal syntax is often not extensible
to user-defined types.

132

CHAPTER 5. PRIMITIVE TYPES

was for systems programming, not business applications. In this case, the cost of not implementing
the type is low, and the cost of implementing it is increased clutter in the language.
As another example, consider the string type in Pascal. It was almost certainly a mistake to
omit a string manipulation package from the standard language. Alphabetic data is very common,
and many programs use string data. The Pascal standard recognizes that strings exist but does
not provide a reasonable set of string manipulation primitives. The standard defines a string
to be any object that is declared as a packed array[1..n] of char, where n is an integer
> 1. String output is provided by Write and Writeln. String comparison and assignment are
supported, but only for strings of equal length. Length adjustment, concatenation, and substrings
are not supported, and Read cannot handle strings at all. A programmer using Standard Pascal
must read alphabetic fields one character at a time and store each character into a character array.
Virtually all implementations of Pascal extend the language to include a full string type with
reasonable operations. Unfortunately, these extensions have minor differences and are incompatible
with each other. Thus there are two kinds of costs associated with omitting strings from standard
Pascal:
1. User implementations of string functions are required. These execute less efficiently than
system implementations could.
2. Because programmers use strings all the time, many compilers are extended to support a
string type and some string functions. Using these extensions makes programs less portable
because the details of the extensions vary from compiler to compiler.
Including strings in the language makes a language more complex. Both the syntax and semantic definitions become longer and require more extensive documentation. The minimal compiler
implementation is bigger. In the case of Pascal and strings, none of these reasons justify the
omission.
When language designers do decide to include a primitive type, they must extend the language
syntax for declarations, but they have some choices about how to include the operations on that
type. The meanings of operators such as < are usually extended to operate on elements of the
new type. New operators may also be added. Any specific function for the new type may be
omitted, added to the language core, or included in a library. The latter approach becomes more
and more attractive as the number of different primitive types and functions increases. A modular
design makes the language core simpler and smaller, and the library features do not add complexity
or consume space unless they are needed.
For example, exponentiation is a primitive operation that is important for much scientific computation. Pascal, C, and FORTRAN all support floating-point encoding but have very unequal
support for exponentiation. In Pascal, exponentiation in base 10 is not supported by the standard
at all; it must be programmed using the natural logarithm and exponentiation functions (ln
and exp). In C, an exponentiation function, pow, is included in the mathematics library along
with the trigonometric functions. In contrast, FORTRANs intended application was scientific com-

5.3. A BRIEF HISTORY OF TYPE DECLARATIONS

133

putation, and the FORTRAN language includes an exponentiation operator, **, as part of the
language core.

5.2.4

Emulation

The types required by a language definition may or may not be supported by the hardware of
machines for which that language is implemented. For example, Pascal requires the type real,
but floating-point hardware is not included on many personal computers. In such situations, data
structures and operations for that type must be implemented in software. Another example: fixedpoint arithmetic is part of Ada. This is no problem on hardware that supports packed-decimal
encoding, but on a strictly binary machine, an Ada translator must use a software emulation or
approximation of fixed-point arithmetic.
The representation for an emulated primitive type is a compromise. On the one hand, it should
be as efficient as possible for the architecture of the machine. On the other hand, it should conform
as closely as possible to the typical hardware implementation so that programs are portable. The
hardware version and the emulation should give the same answers!
When floating point is emulated, the exponent is sometimes represented as a 1-byte integer, and
the mantissa is represented by 4 or more bytes with an implied binary point at the left end. This
produces an easily manipulated object with good precision. A minimum of shifting and masking is
needed when this representation is used. However, it sometimes does not produce the same answers
as a 4-byte hardware implementation.
Other software emulations try to conform more closely to the hardware. Accurate emulation
of floating-point hardware is more difficult and slower, but has the advantage that a program will
give the same answers with or without a coprocessor. A good software emulation should try to
imitate the IEEE hardware standard as closely as possible without sacrificing acceptable efficiency
[Exhibit 5.12].

5.3

A Brief History of Type Declarations

The ways for combining individual data items into structured aggregates form an important part
of the semantic basis of any language.

5.3.1

Origins of Type Ideas

Types Were Based on the Hardware. The primitive types supported by the earliest languages
were the ones built into the instruction set of the host machine. Some aggregates of these types
were also supported; the kinds of aggregates differed from language to language, depending on both
the underlying hardware and the intended application area. In these old languages, there was an
intimate connection between the hardware and the language.
For example, FORTRAN, designed for numeric computation, was first implemented on the IBM
704. This was the first machine to support floating-point arithmetic. So FORTRAN supported one-

134

CHAPTER 5. PRIMITIVE TYPES

Exhibit 5.12. An emulation of floating point.

This is a brief description of the software emulation of floating point used by the Mark Williams
C compiler for the Atari ST (Motorola 68000 chip). Note that it is very similar to, but not quite
like, the IEEE standard shown in Exhibits 5.3 and 5.4.
Bit
Bits
Bits

31:
3023:
220:

Sign
Characteristic, base 2, bias 128
Normalized base 2 mantissa, implied high-order 1, binary point immediately to the left of the implied 1.
Binary Representation
Exponent Mantissa
00000000 .00000000 00000000 00000000

Decimal
0.00

Hex
00000000

Sign
0

0.25
0.50
1.00
-1.00

3F800000
40000000
40800000
C0800000

0
0
0
1

01111111
10000000
10000001
10000001

.10000000
.10000000
.10000000
.10000000

00000000
00000000
00000000
00000000

00000000
00000000
00000000
00000000

10.00
5.00
2.50
1.25

42200000
41A00000
41200000
40A00000

0
0
0
0

10000100
10000011
10000010
10000001

.10100000
.10100000
.10100000
.10100000

00000000
00000000
00000000
00000000

00000000
00000000
00000000
00000000

word representations of integers and floating-point numbers. The 704 hardware had index registers
that were used for accessing elements of an arrayso FORTRAN supported arrays.
COBOL was used to process business transactions and was implemented on byte-oriented business machines. It supported aggregate variables in the form of records and tables, represented
as variable-length strings of characters. One could read or write entire COBOL records. This corresponded directly to the hardware operation of reading or writing one tape record. One could
extract a field of a record. This corresponded to a hardware-level load register from memory
instruction. The capabilities of the language were the capabilities of the underlying hardware.
Type was not a separate idea in COBOL. A structured variable was not an example of a
structured typeit was an independent object, not related to other objects. The structured variable
as a whole was named, as were all of its fields, subfields, and sub-subfields. To refer to a subfield,
the programmer did not need to start with the name of the whole object and give the complete
pathname to that subfield; it could be referred to directly if its name was unambiguous.
FORTRAN supported arrays, and COBOL supported both arrays (called tables) and records.
It would be wrong, though, to say that they supported array or record types, because the structure
of these aggregates was not abstracted from the individual examples of that structure. One could
use a record in COBOL, and even pass it to a subroutine, but one could not talk about the type of

5.3. A BRIEF HISTORY OF TYPE DECLARATIONS

135

Exhibit 5.13. Declaration and use of a record in COBOL.

We declare a three-level record to store data about a father. If other variables were needed to
store information about other family members, lines two through six would have to be repeated.
COBOL provides no way to create a set of uniformly structured variables. Field names could be
the same or different for a second family member. The prevailing style is to make them different
by using a prefix, as in F-FIRST below.
1

FATHER.
2
NAME.
3
LAST
3
F-FIRST
3
F-MID-INIT
2
F-AGE

PIC
PIC
PIC
PIC

X(20).
X(20).
X.
99.

Assume that FATHER is the only variable with a field named F-FIRST, and that MOTHER also has
a field named LAST. Then we could store information in FATHER thus:
MOVE Charles TO F-FIRST. MOVE Brown TO LAST IN FATHER.
Note that the second line gives just enough information to unambiguously identify the field desired;
it does not specify a full pathname.

that record. Each record object had a structure, but that structure had no name and no existence
apart from the object [Exhibit 5.13].
LISP Introduced Type Predicates. LISP was the earliest high-level language to support dynamic storage allocation, and it pioneered garbage collection as a storage management technique.
In the original implementation of LISP, its primitive types, atom and list, were drawn directly from
the machine hardware of the IBM 704. An atom was a number or an identifier. A list was
a pointer to either an atom or a cell. A cell was a pair of lists, implemented by a single machine word. The 36-bit machine instruction word had four fields: operation code, address, index,
and decrement. The address and decrement fields could both contain a machine address, and the
hardware instruction set included instructions to fetch and store these fields.
Here again we see a close relationship between the language and the underlying hardware. This
two-address machine word was used to build the two-pointer LISP cell. The three fundamental
LISP functions, CAR, CDR, and CONS, were based directly on the hardware structure. CAR extracted
the address field of the cell, and CDR extracted the decrement field. (Note that the A in CAR and
the D in CDR came from address and decrement.) CONS constructed a cell dynamically and
returned a pointer to it. This cell was initialized to point at the two arguments of CONS.
Note that all LISP allocations were a fixed sizeone word. Only one word was ever allocated at
a time. However, the two pointers in a cell could be used to link cells together into tree structures

136

CHAPTER 5. PRIMITIVE TYPES

of indefinite size and shape.

The concept of type was more fully developed in LISP than in FORTRAN or COBOL. Types
were recognized as qualities that could exist separately from objects, and LISP supported type
predicates, functions that could test the type of an argument at run time. Predicates were provided
for the types atom and list. These were essential for processing tree structures whose size and
shape could vary dynamically.
SNOBOL: Definable Patterns. SNOBOL was another language of the early 1960s. It was
designed for text processing and was the first high-level language that had dynamically allocated
strings as a primitive type. This was an important step forward, since strings (unlike arrays,
records, and list cells) are inherently variable-sized objects.
New storage management techniques had to be developed to handle variable-sized objects.
Variables were implemented as pointers to numbers or strings, which were stored in dynamically
allocated space. Dynamic binding, not assignment, was used to associate a value with a variable.
Storage objects were created to hold the results of computations and bound to an identifier. They
died when that identifier was reused for the result of another computation. A storage compaction
technique was needed to reclaim dead storage objects periodically. The simplest such technique is
called storage compaction. It involves identifying all live storage objects and moving them to one
end of memory. The rest of the memory then becomes available for reuse.
A second new data type was introduced by SNOBOL: the pattern. Patterns were the first
primitive data type that did not correspond at all to the computer hardware. A pattern is a
string of characters interspersed with wild cards and function calls. The language included a
highly powerful pattern matching operation that would compare a string to a pattern and identify
a substring that matched the pattern. During the matching process, the wild cards would be
matched against first one substring and then another, until the entire pattern matched or the
string was exhausted.9
COBOL permitted the programmer to define objects with complex structured data types but
not to refer to those types. LISP provided type predicates but restricted the user to a few primitive
types. PL/1 went further than either: its LIKE attribute permitted the programmer to refer to a
complex user-defined type.
PL/1 was developed in the mid-1960s for the IBM 360 series of machines. It was intended to be
the universal language that would satisfy the needs of both business and scientific communities.
For this reason, features of other popular languages were merged into one large, conglomerate
design. Arithmetic expressions resembled FORTRAN, and pointers permitted the programmer to
construct dynamically changing tree structures. Declarations for records and arrays were very much
like COBOL declarations.
Types could not be declared separately but were created as a side effect of declaring a structured
variable. Once a type was created, though, more objects of the same type could be declared by
saying they were LIKE the first object [Exhibit 5.14].
9

Compare this to the pattern matching built into Prolog, Chapter 10, Section 10.4.

5.3. A BRIEF HISTORY OF TYPE DECLARATIONS

137

Exhibit 5.14. Using the LIKE attribute in PL/1.

The design of PL/1 was strongly influenced by COBOL. This influence is most obvious in the
declaration and handling of structures. Here we declare a record like the one in Exhibit 5.13. We
go beyond the capabilities of COBOL, though, by declaring a second variable, MOTHER, of the same
structured type.
DCL 1

DCL 1

FATHER,
2
NAME,
3
LAST
CHAR (20),
3
FIRST
CHAR (20),
3
MID-INIT CHAR (1),
2
F-AGE
PIC 99;
MOTHER LIKE FATHER;

To create unambiguous references, field names of both MOTHER and FATHER must be qualified
by using the variable name:
MOTHER.LAST = FATHER.LAST;

5.3.2

Type Becomes a Definable Abstraction

By the late 1960s, types were recognized as abstractionsthings that could exist apart from any
instances or objects. The fundamental idea, developed by C. Strachey and T. Standish, is that a
type is a set of constructors (to create instances), selectors (to extract parts of a structured type),
and a predicate (to test type identity). Languages began to provide ways to define, name, and use
types to create homogeneous sets of objects. ALGOL-68 and Simula were developed during these
years.
Simula pioneered the idea that a type definition could be grouped together with the functions
that operate on that type, and objects belonging to the type, to form a class. Thus Simula was
the first language to support type modules and was a forerunner of the modern object-oriented
languages. 10
ALGOL-68 contained type declarations and very carefully designed type compatibility rules.
The type declarations defined constructors (specifications by which structured variables could be
allocated), selectors (subscripts for arrays and part names for records), and implicit type predicates.
Type identity was the basis for extensive and carefully designed type checking and compatibility
rules. Some kinds of type conversions were recognized to be (usually) semantically valid, and so
were supported. Other type relationships were seen as invalid. The definition of the language
was immensely complex, partly because of the type extension and compatibility rules, and partly
because the design goal was super-generality and power.
10

See Chapter 17 for a discussion of object-oriented languages.

138

CHAPTER 5. PRIMITIVE TYPES

Reactions to Complexity: C and Pascal

Two languages, C and Pascal, were developed at this time as reactions against the overwhelming
size and complexity of PL/1 and ALGOL-68. These were designed to achieve the maximum amount
of power with the minimum amount of complexity.
C moved backwards with respect to type abstractions. The designers valued simplicity and
flexibility of the language more than its ability to support semantic validity. They adopted type
declarations as a way to define classes of structured objects but omitted almost all use of types
to control semantics. C supported record types (structs and unions) with part selectors and
arrays with subscripting. Record types were full abstractions; they could be named, and the names
used to create instances, declare parameters, and select subfields. Arrays, however, were not fully
abstracted, independent types; array types could not be named and did not have an identity that
was distinct from the type of the array elements.
The purpose of type declarations in C was to define the constructors and selectors for a new
type. The declaration supplied information to the compiler that enabled it to allocate and access
compound objects efficiently. The field names in a record were translated into offsets from the
beginning of the object. The size of the base type of an array became a multiplier, to be applied to
subscripts, producing an offset. At run time, when the program selected a field of the compound,
the offset was added to the address of the beginning of the compound, giving an effective address.
At this period of history, types were not generally used as vehicles for expressing semantic
intent. Except for computing address offsets, there were very few contexts in early C in which the
type of an object made a difference in the code the translator generated. 11 Type checking was
minimal or nonexistent. Thus C type declarations did not define type predicates. Type identity
was not, in general, important. The programmer could not test it directly, as was possible in LISP,
nor was it checked by the compiler before performing function calls, as it is in Pascal. 12
Niklaus Wirth, who participated in the ALGOL-68 committee for some time, designed Pascal to
prove that a language could be simple, powerful, and semantically sound at the same time. 13 Pascal
retained both the type declarations and type checking rules of ALGOL-68 and achieved simplicity
by omitting ALGOL-68s extensive type conversion rules. The resulting language is more restrictive
than C, but it is far easier to understand and less error prone.
Ada: The Last ALGOL-Like Language?
In the late 1960s, the U.S. Department of Defense (DoD) realized that the lack of a common
computer language among its installations was becoming a major problem. By 1968, small-scale
research efforts were being funded to develop a core language that could be extended in various
directions to meet the needs of different DoD groups. Design goals included generality of the core
language, extensibility, and reasonable efficiency.
11

The exception was automatic conversions between numeric types in mixed expressions.
Type checking and the semantic uses of types are discussed at length in Chapter 15.
13
It is said that he never dreamed that Pascal would achieve such widespread use as a teaching language.
12

5.3. A BRIEF HISTORY OF TYPE DECLARATIONS

139

In the early 1970s, DoD decided to strictly limit the number of languages in use and to begin
design of one common language. A set of requirements for this new language were developed by
analyzing the needs of various DoD groups using computers. Finalized in 1976, these requirements
specified that the new language must support modern software engineering methods, provide superior error checking, and support real-time applications. After careful consideration, it was decided
that no existing language met these criteria.
Proposals were sought in 1977 for an ALGOL-like language design that would support reliable,
maintainable, and efficient programs. Four proposals were selected, from seventeen submitted, for
further development. One of these prototype languages was selected in 1979 and named Ada. Major
changes were made, and a proposed language standard was published in 1980.
Ada took several major steps forward in the area of data types. These included
Cleaner type compatibility rules.
Explicit constraints on the values of a type.
Support for type portability.
Types treated as objects, in a limited way.
Ada was based on Pascal and has similar type compatibility rules. These rules are an important
aid to achieving reliable programs. However, Pascal is an old language, and its compatibility rules
have holes; some things are compatible that, intuitively, should not be. Ada partially rectified
these problems.
The idea of explicit constraints on the values belonging to a type was present in Pascal in the
subrange types. In Ada, this idea is generalized; more kinds of constraints may be explicitly stated.
These constraints are automatically checked at run time when a value is stored in a constrained
variable.14
In the older languages, the range of values belonging to a type often depended on the hardware
on which a program ran. A program, debugged on one computer, often ran incorrectly on another.
By providing a means to specify constraints, Ada lets the programmer explicitly state data characteristics so that appropriate-sized storage objects may be created regardless of the default data
type sizes on any given machine. A programmer can increase the portability of code substantially
by using constrained types.
A Pascal type can be used to declare a parameter, but it cannot be a parameter. Ada carries
the abstraction of types one step further. Ada supports modules called generic packages. These are
collections of declarations for types, data, and functions which depend on type parameters and/or
integer parameters. Each type declaration in a generic package defines a generic type, or a family
of types, and must be instantiated, or expanded with specific parameters, to produce a specific type
declaration during the first phase of compilation. 15 Thus although the use of type as parameters
is restricted to precompile time, Ada types are objects in a restricted sense.
14
15

To achieve efficiency, Ada permits this checking to be turned off after a program is considered fully debugged.
See Chapter 17.

140

CHAPTER 5. PRIMITIVE TYPES

Recent Developments
Since the early 1980s, data type research has been directed toward implementing abstract data
types, type hierarchies with inherited properties, and implementing nonhomogeneous types in a
semantically sound way. These issues are covered in Chapter 17.

Exercises
1. Define: bit, byte, word, long word, double word.
2. What is unique about logical instructions?
3. How are objects represented in the computer? Explain.
4. What is the purpose of computer codes? Explain.
5. What were the dissatisfactions with early computer codes?
6. What was the difference between the memory of a business and a scientific computer?
7. What is packed decimal? How is it used?
8. What is an unsigned number? How is it represented in memory?
9. What is sign and magnitude representation? Ones complement? Twos complement?
10. How are negative numbers represented in modern computers?
11. What is a floating-point number? What are the problems associated with the representation
of floating-point numbers?
12. Name the computer that you use. For each number below, give the representation (in binary)
that is used on your computer.
a.
b.
c.
d.
e.

The largest 32-bit integer

The largest positive floating-point number
Negative one-half, in floating-point
The smallest positive float (closest to zero)
The negative floating-point number with the greatest magnitude

13. Even though FORTH does not contain semantic mechanisms that implement subscripting,
array bounds, or variables with multiple slots, it can be said that FORTH has arrays.
Explain.
14. What is a data type? Type declaration? Type description?

5.3. A BRIEF HISTORY OF TYPE DECLARATIONS

141

15. Compare the way that the type of an object is represented in Pascal and in APL. Point out
similarities and differences.
16. What is a specific data type? Generic data type?
17. Explain the three attributes of a data type: encoding, size, and structure.
18. What determines the set of primitive data types associated with a language?
19. What is the usual set of primitive types associated with a language?
20. What is type emulation? Why is it needed?
21. How were types supported by the earliest languages? Give a specific example.
22. How is type represented in COBOL? LISP? SNOBOL?
23. Compare the pattern matching in SNOBOL to the database search in Prolog.
24. What is type checking? Type compatibility?
25. What are value constructors? Selectors?
26. What new ideas did Simula pioneer?
27. Why was Ada developed?
28. What major steps in the area of data typing were used in Ada?

142

CHAPTER 5. PRIMITIVE TYPES

Chapter 6

Modeling Objects

Overview
This chapter creates a framework for describing the semantics and implementation of
objects so that the semantics actually used in any language can be understood and the
advantages and drawbacks of the various implementations can be evaluated.
We assume the reader is familiar with the use of objects such as variables, constants,
pointers, strings, arrays, and records. When we survey the popular programming languages, we see a great deal of commonality in the semantics of these things in all languages. There are also important differences, sometimes subtle, that cause languages
to feel different, or require utterly different strategies for use. A program object embodies a real-world object within a program. The program object is stored in a storage
object, a collection of contiguous memory cells. Variables are storage objects that store
pure values; pointer variables store references.
Initialization and assignment are two processes that place a value in a storage object.
Initialization stores a program object in the storage object when the storage object is
created. Assignment may be destructive or coherent. Extracting the contents from a
storage object is known as dereferencing. Assignment and dereferencing of pointer variables usually yield references to ordinary variables rather than pure values. Managing
computer memory involves creating, destroying, and keeping storage objects available.
Three strategies are static storage, stack storage, and heap storage.

143

144

CHAPTER 6. MODELING OBJECTS

Exhibit 6.1. Representing objects.

External object: a length of 2" by 4" lumber.
Program object: a 32-bit floating-point value.
Storage object: a memory location with four consecutive bytes reserved for this number.
External object: a charge account.
Program object: a collection of values representing a customers name, address,
account number, billing date, and current balance.
Storage object: a series of consecutive memory locations totaling 100 bytes.

6.1

Kinds of Objects

A program is a means of modeling processes and objects that are external to the computer. External
objects might be numbers, insurance policies, alien invaders for a video game, or industrial robots.
Each one may be modeled in diverse ways. We set up the model through declarations, allocation
commands, the use of names, and the manipulation of pointers. Through these, we create objects
in our programs, give them form, and describe their intended meaning. These objects are then
manipulated by the functions and operators of a language.
We start by making a distinction between the memory location in which data is stored and the
data itself. The ways of getting data into and out of locations are explored.
A program object is the embodiment of an object in the program. It may represent an external
object, such as a number or a record, in which case it is called a pure value. It may also represent
part of the computer system itself, such as a memory location, a file, or a printer. During execution,
the program manipulates its program objects as a means of simulating meaningful processes on
the external objects or controlling its own internal operations. It produces usable information from
observed and derived facts about the program objects.
A program commonly deals with many external objects, each being represented by a pure value
program object [Exhibit 6.1]. While all the external objects exist at once, their representing program
objects can be passed through the computer sequentially and so do not have to be simultaneously
present. For example, an accounting program deals with many accounts. Representations of these
accounts are put in some sequence on an input medium and become program objects one at a time.
In order to manipulate program objects, the program must generally store all or part of a
program object in memory. It uses a storage object for this purpose. A storage object is a collection
of contiguous memory cells (bits, bytes, etc.) in which a program object, called its value or contents,
can be stored.1
A reference is the memory address of a storage object and is the handle by which the object
1

A storage object sometimes encompasses more cells than are needed to store the value. These cells, commonly
added to achieve word alignment, are called padding.

145

6.1. KINDS OF OBJECTS

Exhibit 6.2. Values, variables, and pointers.

The relationship between storage objects and program objects is illustrated. Boxes represent
storage objects, letters represent pure values of type character, small circles (o) from which arrows
emerge represent references, and dotted lines represent dereferencing.

A reference to a
pointer variable

102
102

A pointer variable: 136

contains

A reference to
a variable

136

136

A variable array 3 of char :

C
A
T

contains

C
A
T

A pure value array 3 of char

is accessed. In older terminology, a pure value is called an r-value or right-hand-value, because it

can occur to the right of an assignment operator. A reference is called an l-value or left-hand-value.
A reference is created when a storage object is allocated. This reference is itself a program object
and may be stored in another storage object for later use [Exhibit 6.2]. A program must possess a
reference to a program object in order to use that object.
The allocation process sets aside an area of unused computer memory to make a new storage
object. The process is essentially the same whether it is being carried out by an interpreter, which
does the allocation at run time when the command is interpreted, or by a compiler, which deals
with addresses of the storage objects that will be allocated at some future time when the program
is executed.
Allocation procedures are usually part of the implementation of a language, not part of the
language definition, so the actual allocation process often differs from one translator to the next,
as well as from one language to the next. Typically, though, the allocation process will include the
following actions:
1. The translator must determine N , the number of bytes of memory that are needed. In some
languages the programmer communicates this information by specifying the data type of the
new object. Size, in bytes, is calculated by the translator and stored as part of the definition
of a type. In lower-level languages the programmer specifies the allocation size explicitly.
2. A segment of free storage is located with length L N . A reference to the first location in
this segment is saved.
3. The address of the beginning of the free storage area is incremented by N , thus removing N

146

CHAPTER 6. MODELING OBJECTS

bytes from free storage.

4. If an initial value was defined, it is stored in the new storage object.

5. The address, or reference, saved in step 2 is returned as the result of the allocation process.
It is the means by which the program is able to find the new storage object.
A variable is a storage object in which a pure value may be stored. Pure values and the
variables in which they are stored have the same size and structure and are considered to be the
same data type in many languages. We distinguish between them here because they have very
different semantics. Operations you can perform with variables are to allocate and deallocate them
and to fetch values from and store values into them. In contrast, pure values can be combined and
manipulated with operators and functions, but not allocated and deallocated.
A pointer variable is a storage object, or part of a storage object, in which a reference may
be stored. (Often this term will be shortened to pointer.) Pointers are used to create storage
structures such as game trees and linked lists and are an important means of modeling external
objects.

6.2
6.2.1

Placing a Value in a Storage Object

Static Initialization

A storage object receives a value by one of two processes: initialization or assignment. Until a value
is stored in a storage object, it is said to contain garbage, or to have an undefined value. (When we
wish to indicate an undefined value we will write ?.)
Using an undefined value is a commonly made semantic error which generally cannot be detected
by a language translator. For this reason some translators initialize all variables to zero, which is the
most commonly useful initial value, or to some distinctive bit pattern, so that the semantic error can
be more easily detected. It is poor programming practice to depend on such automatic initialization,
however. Different translators for the same language may implement different initialization policies,
and the program that depends on a particular policy is not portable.
Initialization stores a program object in the storage object when the storage object is created.
Many languages permit the programmer to include an initializing clause in an object declaration.
Typical declaration forms are shown in Exhibits 6.3 and 6.4. In each exhibit, declarations are given
for an integer variable, a character string, and an array of real numbers, and initial values are
declared for each.
Initializing compound objects, such as arrays and records, is restricted or not allowed in some
languages. Two problems are involved here: how to denote a structured value, and how to implement initialization of dynamically allocated structured objects. The FORTRAN and C examples
[Exhibits 6.3 and 6.4] illustrate two approaches to defining the structure of the initializer.
In FORTRAN, the programmer writes an explicit loop or nest of loops which specify the order
in which the fields of an array will be initialized and then provides a series of constants that will

6.2. PLACING A VALUE IN A STORAGE OBJECT

147

Exhibit 6.3. Initial value declarations in FORTRAN.

CHARACTER*3 EOFLAG
DIMENSION A (8)
DATA EOFLAG, ISUM / NO , 0/, ( A(I), I=1,8) / 8.2, 2.6, 3.1, 17.0, 4 * 0.0 /
Notes:
In FORTRAN, simple integers and reals may be declared implicitly. Explicit declarations
must be given for arrays and strings.
Initial values are given in separate DATA declarations which must follow the statements that
declare the storage objects.
A single DATA statement can initialize a list of objects. It must contain exactly as many
initial values as fields to be initialized. Initial values may be repeated by using a repeat count
with a *.
An array may be initialized by giving a loop-controlling expression.

Exhibit 6.4. Initial value declarations in C.

static
static

char end_of_file_flag [ ] = "no ";

int isum = 0;
float a[8] = {8.2, 2.6, 3.1, 17.0};

Notes:
In C an initial value may be given as part of a variable declaration.
Static arrays can be initialized by listing the correct number of values for the array enclosed
in brackets. (The property static is explained in Section 6.3.)
The programmer may omit the array length specifier from the declaration, as in the top line,
and the length of the storage object will be deduced from the length of the initial value list.
If too few initializers are given to fill an array, remaining elements are initialized to zero.

148

CHAPTER 6. MODELING OBJECTS

evaluate to the desired initial values. A repetition count can be specified when several fields are
to be initialized to the same value. Part or all of an array may be initialized this way. This is a
powerful and flexible method, but it does complicate the syntax and semantics of the language.
Contrast this to a C initializer. Its structure is denoted very simply by enclosing the initial
values in brackets, which can be nested to denote a type whose fields are themselves structured
types. The same simple syntax serves to initialize both records and arrays. Initializers can be
constants or constant expressions; that is, expressions that can be evaluated at compile time. In
some ways, this is not as flexible a syntax as FORTRAN provides. If any field of a C object is
initialized, then all fields will be initialized. If the same nonzero value is to be placed in several
fields, it must be written several times. The one shortcut available is that, if the initializer has too
few fields, the remaining fields will default to an initial value of zero.
It is likely that the designers of C felt that FORTRAN initializers are too flexiblethat they
provide unnecessary flexibility, at the cost of unnecessary complication. Applying something akin
to the principle of Too Much Flexibility, they chose to include the simpler, but still very useful,
form in C.
All data storage in FORTRAN is created and initialized at load time. A translator can evaluate
the constant expressions in an initializer and generate store instructions to place the resulting
values into storage when the program code is loaded. Modern languages, though, support dynamic
allocation of local variables in stack frames. (These are called automatic variables in C.) The
initialization process for automatic variables is more complex than for static variables.
Suppose a function F contains a declaration and initializations for a local array, V . This array
cannot be initialized at load time because it does not yet exist. The translator must evaluate the
initializing expressions, store the values somewhere, and generate a series of store instructions to
be executed every time F is called. These copy precomputed initial values into the newly allocated
area. This process was considered complex enough that the original definition of C simply did not
permit initialization of automatic arrays. ANSI C, however, supports this useful facility.

6.2.2

Dynamically Changing the Contents of a Storage Object

Destructive Assignment.
In many languages, one storage object can be used to store different program objects at different
times. Assignment is an operation that stores a program object into an existing storage object and
thus permits the programmer to change the value of a storage object dynamically. This operation
is sometimes called destructive assignment because the previous contents of the storage object are
lost. The storage object now represents a different external object, and we say that its meaning
has changed.
Functional languages are an important current research topic. The goal of this research is to
build a language with a clean, simple semantic model. Destructive assignment is a problem because
it causes a change in the meaning of the symbol that names the storage object. It complicates a
formal semantic model considerably to have to deal with symbols that mean different things at

149

6.2. PLACING A VALUE IN A STORAGE OBJECT

Exhibit 6.5. Initializing and copying a compound object in Pascal.

Pascal declarations are given below for a record type named person and for two personvariables, a and b. In Pascal, compound objects cannot be initialized coherently, so three assignments are used to store a record-value into b. On the other hand, records can be assigned coherently,
as shown in the last line, which copies the information from b to a.
TYPE person = RECORD age, weight:
VAR a, b : person;
BEGIN
b.age := 10;
b.weight := 70;
b.sex := M;
a := b;
...
END;

integer; sex:

char END;

different times.
In a functional language, parameter binding is used in place of destructive assignment to associate names with objects. At the point that a Pascal programmer would store a computed value in
a variable, the functional programmer passes that value as an argument to a function. The actions
following the assignment in the Pascal program, and depending on it, would form the body of the
function. A series of Pascal statements with assignment gets turned outside in and becomes a
nest of function calls with parameter bindings. 2 This approach produces an attractive, semantically clean language because the parameter name has the same meaning from procedure entry to
procedure exit.
Coherent Assignment. An array or a record is a compound object: a whole made up of parts
which are objects themselves. Some but not all programming languages permit coherent assignment
of compound objects. In such languages an entire compound variable is considered to be a single
storage object, and the programmer can refer to the compound object as a whole and assign
compound values to it [Exhibits 6.5 and 6.7].
In COBOL any kind of object could be copied coherently. It is even possible to use one coherent
READ statement to load an entire data table from a file into memory. In most older languages,
though, assignment can only be performed on simple (single-word) objects. An array or a record
is considered to be a collection of simple objects, not a coherent large object. The abstract process
of placing a compound program object into its proper storage object must be accomplished by a
series of assignment commands that store its individual simple components.
2

A deeply nested expression can look like a rats nest of parentheses; deep nesting is avoided by making many
short function definitions.

150

CHAPTER 6. MODELING OBJECTS

Exhibit 6.6. Initializing and copying a compound object in K&R C.

A record type named person is defined, and two person-variables, a and b, are declared. The
variable b is initialized by the declaration and copied into a by the assignment statements.
The property static causes the variable to be allocated in the program environment rather
than on the stack, so that it can be initialized at load time. K&R C did not support initialization
of dynamically allocated structured objects.
typedef struct {int age, weight; char sex;} person;
static person a, b = {10, 70, M};
{

a.age = b.age;
a.weight = b.weight;
a.sex = b.sex;
...}

An example of the lack of coherent assignment can be seen in the original Kernighan and Ritchie
definition of C. Coherent assignment was not supported; to copy a record required one assignment
statement for each field in the record. Thus three assignments would be required to copy the
information from b to a in Exhibit 6.6. However, coherent initialization of record variables was
supported, and b could be initialized coherently.
Even in languages that support coherent compound assignment, the programmer is generally
permitted to assign a value to one part of the compound without changing the others. In such
situations, care must always be taken to ensure that a compound storage object is not left containing
parts of two different program objects!

Exhibit 6.7. Initializing and copying a compound object in ANSI C.

This example is written in ANSI C, which is newer than both K&R C and Pascal. The difference
between this and the clumsier versions in Exhibits 6.5 and 6.6 reflects the growing understanding
that coherent representations and operations are important.
The type and object declarations are the same in both versions of C, as are initializations. But
compound objects can be assigned coherently in ANSI C, so only one assignment is required to copy
the information from b to a. Further, dynamically allocated (automatic) structs may be initialized
in ANSI C.
typedef struct {int age, weight; char sex;} person;
person a, b = {10, 70, M};
{ a = b; ...}

151

6.2. PLACING A VALUE IN A STORAGE OBJECT

Exhibit 6.8. Languages where assignment is a statement.

A yes in the third column indicates that compound objects (such as arrays and records) may
be assigned coherently, as a single action. A yes in the fourth column indicates that one ASSIGN
statement may be used to store a value in several storage objects.
Language
COBOL

FORTRAN
ALGOL
PL/1
FORTH
Pascal
Ada

Assignment Symbol
MOVE
= (in a COMPUTE statement)
ADD, SUBTRACT, MULTIPLY, DIVIDE
=
:=
=
!
:=
:=

Compound
Assignment?
yes
no
no
no
no
yes
no
yes
yes

Multiple
Assignment?
yes
yes
yes
no
no
yes
no
no
no

Assignment Statements versus Assignment as a Function. Assignment is invoked either

by writing an explicit ASSIGN operator or by calling a READ routine. In either case, two objects
are involved, a reference and a value. The reference is usually written on the left of the ASSIGN
operator or as the parameter to a READ routine, and the value is written on the right of the
ASSIGN or is supplied from an input medium.
Assignment is one of a very small number of operations that require a reference as an argument.
(Others are binding, dereference, subscript, and selection of a field of a record.) The purpose of an
assignment is to modify the information in the computers memory, not to compute a new value. It is
the only operation that modifies the value of existing storage objects. For this reason, ASSIGN and
READ occur in many languages as statement types or procedures rather than as functions. Exhibit
6.8 lists the symbols and semantics for the ASSIGN statements in several common programming
languages.
In other languages, ASSIGN is a function that returns a result and may, therefore, be included
in the middle of an expression. Exhibit 6.9 shows ASSIGN functions in common programming
languages. LISP returns the reference as the result of an assignment. C returns the value, so that
it may be assigned to another storage object in the same expression or may be used further in
computing the value of an enclosing expression. Exhibit 6.10 demonstrates how one assignment
can be nested within another.
When ASSIGN returns a value, as in C, a single expression may be written which assigns that
value to several storage objects. We call this multiple assignment. While this facility is not essential,
it is often useful, especially when several variables need to be zeroed out at once. The same end is
achieved in other languages, such as COBOL, by introducing an additional syntactic rule to allow

152

CHAPTER 6. MODELING OBJECTS

Exhibit 6.9. Languages where assignment is a function.

A yes in the third column indicates that compound objects (such as arrays and records) may
be assigned coherently, as a single action.
Language
LISP
APL
C (1973)
C (ANSI)

Assignment Symbol
replaca, replacd
(also used for binding)
=
=

Compound
Assignment?
some versions
yes
no
yes

Result
Returned
reference
value
value
value

an ASSIGN statement to list references to several storage objects, all of which will receive the single
value provided.

6.2.3

Dereferencing

Dereferencing is the act of extracting the contents from a storage object. It is performed by the
FETCH operation, which takes a reference to a storage object and returns its value. When a pointer
variable is dereferenced, the result is another reference. This could be a reference to a variable,
which itself could be dereferenced to get a pure value, or it could be a reference to another pointer,
and so forth.
Whereas ASSIGN is always written explicitly in a language, its inverse, FETCH, is often invoked
implicitly, simply by using the name of a storage object. Many languages (e.g., FORTRAN, Pascal,
C, COBOL, BASIC, LISP) automatically dereference a storage object in any context where a program

Exhibit 6.10. Assignment as a function in C.

An array length is defined as a constant at the top of the program to facilitate modifications.
Then the array ar is declared to have 100 elements, with subscripts from 0 to 99. Two integers
are declared and set to useful numbers: num_elements holds the number of elements in the array,
and high_sub holds the subscript of the last element.
#define MAXLENGTH 100
float ar[ MAXLENGTH ];
int high_sub, num_elements;
high_sub = (num_elements = MAXLENGTH) - 1;
The last line contains two assignments. The constant MAXLENGTH is stored into the variable
num_elements, and it is also returned as the result of the assignment function. This value is then
decremented by one, and the result is stored in high_sub.

6.2. PLACING A VALUE IN A STORAGE OBJECT

153

Exhibit 6.11. Dereferencing by context in Pascal.

We analyze the dereferences triggered by evaluating this expression:
xarray[ point_1.number ] := eval_function( point_2 ) ;
Assume the objects referenced have the following types:
xarray:
point_1, point_2:
eval_function:

An array of unspecified type.

Pointer to a record with a field called number.
A function taking one pointer parameter and returning something of
the correct type to be stored in xarray.

A variety of dereference contexts occur. Contexts (1), (3), and (4) occur together on the left,
as do contexts (2) and (5) on the right.
Reference
xarray
point_1

point_2

Is it dereferenced here?
No, it is on the left of a := operator.
Yes, explicitly, by the operator. Although this is part of a subscript expression, explicit dereference must be used because pointer variable names
are not dereferenced in a pointer expression.
You cannot tell from this amount of context. It will not be dereferenced if
the function definition specifies that it is a VAR parameter. If VAR is not
specified, it will be automatically dereferenced.

object is required. Thus a variable name written in a program sometimes means a reference and
sometimes a pure value, depending on context. This introduces complexity into a language. You
cannot just see a symbol, as in lambda calculus, and know what it means. You must first examine
where it is in the program and how it is used. To define the dereferencing rules of a language,
contexts must be enumerated and described. The commonly important contexts are:
1. The left-hand side of an assignment operator.
2. The right-hand side of an assignment operator.
3. Part of a subscript expression.
4. A pointer expression.
5. A parameter in a function or procedure call.
Note that these contexts are not mutually exclusive but can occur in a confusing variety of combinations, as shown in Exhibit 6.11. Many other combinations of dereferencing contexts are, of
course, possible.
Whether or not a reference is dereferenced in each context varies among languages. In context
(1) dereferencing is never done, as a reference is required for an ASSIGN operation. But when a
subscript expression (3) occurs in context (1), dereferencing will happen within the subscript part

154

CHAPTER 6. MODELING OBJECTS

Exhibit 6.12. Explicit dereferencing in FIG FORTH.

All FORTH expressions are written in postfix form, so you should read and interpret the operators from left to right. The FETCH operator is @. It is written following a reference and extracts
the contents from the corresponding storage object.
On lines 1 and 2, variables named XX and Y are declared and initialized to 13 and 0, respectively.
Line 3 dereferences the variable XX and multiplies its value by 2. The result is stored in Y, which
is not dereferenced because a reference is needed for assignment.
1
13 VARIABLE XX
2
0 VARIABLE Y
3
XX @ 2 * Y ! ( Same as Y = XX * 2 in FORTRAN. )
The expression XX 2 * would multiply the address, rather than the contents, of the storage
object named XX by 2.

of the expression (the subscripted variable itself will not be dereferenced). In contexts (2) and (3)
most languages will automatically dereference, as long as the situation does not also involve context
(4).
In context (4) languages generally do not dereference automatically. They either provide an
explicit FETCH operator or combine dereferencing with other functions. Examples of FETCH operators are the Pascal and C *. Examples of combined operators are -> in C, which
dereferences a pointer and then returns a reference to a selected part of the resulting record, and
car and cdr in LISP, which select a part of a record and then dereference it.
In context (5), there is no uniformity at all among languages. The particular choices and
mechanisms used in various languages are discussed fully in Chapter 8, Section 8.4, and Chapter
9, Section 9.2.
There are also languages in which storage objects are never automatically dereferenced, the
most common being FORTH. In such languages the dereference command must be written explicitly
using a dereference operator (@ in FORTH) [Exhibit 6.12]. The great benefit of requiring explicit
dereference is simplicity. A variable name always means the same thing: a reference. Considering
the kind of complexity (demonstrated above) that is inherent in deriving the meaning of a reference
from context, it is easy to understand the appeal of FORTHs simple method. The drawback of
requiring explicit dereference is that an additional symbol must be written before most uses of a
variable name, adding visual clutter to the program and becoming another likely source of error
because dereference symbols are easily forgotten.

6.2.4

Pointer Assignment

Pointer assignment is ordinary assignment where the required reference is a reference to a pointer
variable and the value is itself a reference, usually to an ordinary variable. Languages that support
pointer variables also provide a run-time allocation function that returns a reference to the newly

155

6.2. PLACING A VALUE IN A STORAGE OBJECT

Exhibit 6.13. Pointer assignments in Pascal.

We assume the initial state of storage shown in Exhibit 6.14.
TYPE list = cell;
cell = RECORD value:char; link:list END;
VAR P1, P2, P3: list;
Code
P2 := P1;
P3 := P1 .link;

Comments
Dereference P1 and store its value in P2.
Dereference P1, select its link field, which is a pointer variable, and
dereference it. Store the resulting reference in P3.

P1, P2, and P3 all share storage now. We can refer to the field containing the % as
P2.link.value or as P3.value. Note that a pointer must be explicitly dereferenced, using , before accessing a field of the object to which it points.

allocated storage. This reference is then assigned to a pointer variable, which is often part of a
compound storage object. Pointer assignment allows a programmer to create and link together
simple storage objects into complex, dynamically changing structures of unlimited size.
Multiple pointers may be attached to an object by pointer assignment. The program object
of a pointer is a reference to another storage object. When the pointer assignment P2 := P1 is
executed, the program object P1, which is a reference to some object, Cell1, is copied into the
storage object of P2, thus creating an additional pointer to Cell1 and enabling P2 as well as P1
to refer to Cell1. Thus two objects now store references to one storage object, and we say they
share storage dynamically. This is illustrated in Exhibits 6.13 and 6.14.
While such sharing is obviously useful, it creates a complex situation in which the contents of
the storage structure attached to a name may change without executing an assignment to that
name. This makes pointer programs hard to debug and makes mathematical proofs of correctness
very hard to construct. Many programmers find it impossible to construct correct pointer programs

Exhibit 6.14. Pointer structures sharing storage.

Storage, diagrammed before and after the pointer assignments in Exhibit 6.13.

Before:
P1

After:

P2

P3

?
$

P1

P2

P3

?
%

&

&

156

CHAPTER 6. MODELING OBJECTS

Exhibit 6.15. Pointer assignments with dereference in C.

The right side of the assignment is dereferenced if it evaluates to a structure or a simple object.
typedef struct { int age; float weight; } body;
body s;
/* A variable of type body. */
body *ps, *qs;
/* Two pointers to bodies. */
int k;
/* An integer variable. */
int *p, *q;
/* Two pointers to integers. */
p = &k;

/* Store the address of k in p, that is, make p point at k.

Note: the & operator prevents automatic dereferencing. */

ps = &s;

/* Make ps point at s.

q = p;

/* Dereference p to get the address stored in p, and store

that address in q, making q point at the same thing as p.

qs = ps;

*/

/* Make qs point at the same thing as ps.

k
p
q

*/

*/

s
17

ps

37

105.2

qs

without making diagrams of their storage objects and pointer variables.

6.2.5

The Semantics of Pointer Assignment

There are two likely ways in which a pointer assignment could be interpreted: with and without
automatic dereferencing of the right-hand side. Pascal does dereference, as is shown in Exhibit
6.13. In such a language the statement Q := P is legal if P and Q are both pointers. This makes Q
point at whatever P is pointing at. The assignment P := K is illegal if P is a pointer and K is an
integer. Exhibit 6.15 shows several pointer assignments in C where the right side is dereferenced.
In a hypothetical language, := could be defined such that the assignment p := k would
be legal and would make p point at k. In this case, pointer assignment is interpreted without
dereferencing the right side. In such a language we could create a chain of pointers as follows:
k := 5.4;
-- k is type float.
p := k;
-- p must be type pointer to float.
q := p;
-- q must be type pointer to pointer to float.
These assignments, taken together, would construct a pointer structure like this:

157

6.2. PLACING A VALUE IN A STORAGE OBJECT

Exhibit 6.16. Pointer assignment without dereference in C.

We declare an integer variable, k; integer pointers, p1 and p2; an array of five integers, a; a
function that returns an integer, f; and a pointer to a function that returns an integer, p3.
The right side of a C assignment is not dereferenced if it refers to an array or a function.
int k, *p1, *p2, a[5], f(), *p3();
p1 = &k;
/* Make p1 point at k. */
p2 = a;
/* Make p2 point at the array. Note absence of "&".*/
p2 = &a[0];
/* Make p2 point at the address of the zeroth element of the
array. This has the same effect as the line above. */
p3 = f;
/* Store a reference to the function f in pointer p3.
Note that f is not dereferenced.*/

p1

p3

Machine code

a
p2
a[0] a[1] a[2] a[3] a[4]
Note that p2 = &a; is syntactically incorrect because the name of an array means the address
of its zeroth element. One must either omit the & or supply a subscript.

K
5.4

Exhibit 6.16 shows pointer assignments in C which set pointers to an array and a function. In
these contexts, in C, the right side will not be dereferenced.
While either interpretation of pointer assignment could make sense, we would expect to see
either one or the other used consistently in a language. One of the unusual and confusing facets of
C is that the semantics of pointer assignment depends on the type of the expression on the right.
If it denotes a simple object (such as an integer or a pointer) or an object defined as a struct,
automatic dereferencing is used [Exhibit 6.15]. If the right-hand object is an array or a function,
the second meaning, without dereferencing, is implemented [Exhibit 6.16].

158

6.3

CHAPTER 6. MODELING OBJECTS

The Storage Model: Managing Storage Objects

The differences among languages are easier to understand when the underlying mechanisms are
known. A key part of any translator is managing the computer memory; storage objects must be
created, kept available, and destroyed when appropriate. Three storage management strategies are
in common use with all three present in some translators, but only one in others. These are static
storage and two kinds of dynamic storage: stack storage and heap storage.

6.3.1

The Birth and Death of Storage Objects

A storage object is born when it is allocated, and it dies when it is no longer available for use by
the program. The lifetime, or extent, of a storage object is the span of time from its birth to its
death. An object that lives until the program is terminated is immortal. Most objects, however, die
during program execution. It is a semantic error to attempt to reference a storage object after it
has died. The run-time system will typically reuse the formerly occupied storage for other purposes,
so references to a dead object will yield unpredictable results.
Deallocation is the recycling process by which dead storage objects are destroyed, and the
storage locations they occupied are made available for reuse by the allocation process. Deallocation
happens sometime, often not immediately, after death.
All live objects must be simultaneously in the computers virtual memory. Real computers have
limited memory, so it is important that the lifetimes of objects correspond to the period of time
during which they are actually needed by the program. By having an object die when it is no
longer useful, we can recycle the storage it formerly occupied. This enables a program to use a
larger number of storage objects than would otherwise fit into memory.
Static Storage Objects
A compiler plans what storage objects will be allocated to a program at load time, and when the
object code will be copied into computer memory, linked, and made ready to run. Such objects
are allocated before execution begins and are immortal. These are called static storage objects
because they stay there, unmoved, throughout execution. Static allocation is often accompanied
by initialization. The compiler chooses run-time locations for the static objects and can easily put
initial values for these locations into the object code.
The number of static storage objects in a program is fixed throughout execution and is equal to
the number of static names the programmer has used. Global variables are static in any language.
Some languages (for example, COBOL) have only static objects, while others (for example, Pascal)
have no static storage except for globals. Still others (ALGOL, C) permit the programmer to declare
that a nonglobal object is to be static. In ALGOL, this is done by specifying the attribute OWN
as part of a variable declaration. In C, the keyword static is used for this attribute.
A language with only static storage is limiting. It cannot support recursion, because storage
must be allocated and exist simultaneously for the parameters of a dynamically variable number of
calls on any recursive function.

6.3. THE STORAGE MODEL: MANAGING STORAGE OBJECTS

159

A language that limits static storage to global variables is also limiting. Many complex applications can be best modeled by a set of semi-independent functions. Each one of these performs
some simple well-defined task such as filling a buffer with data or printing out data eight columns
per line. Each routine needs to maintain its own data structures and buffer pointers. Ideally, these
are private structures, protected from all other routines. These pointers cannot be ordinary local
variables, since the current position on the line must be remembered from one call to the next, and
dynamically allocated variables are deallocated between calls. On the other hand, these pointers
should not be global, because global storage is subject to accidental tampering by unrelated routines. The best solution is to declare these as static local storage, which simultaneously provides
both continuity and protection.
Finally, the unnecessary use of static objects, either global or local, is unwise because they are
immortal. Using them limits the amount of storage that can be recycled, thereby increasing the
overall storage requirements of a program.
Dynamic Storage Objects
Storage objects that are born during execution are called dynamic. The number of dynamic storage
objects often depends on the input data, so the storage for them cannot be planned by the compiler
in advance but must be allocated at run time. The process of choosing where in memory to allocate
storage objects is called memory management. A memory manager must be sure that two storage
objects that are alive at the same time never occupy the same place in memory. It should also try
to use memory efficiently so that the program will run with as small an amount of physical memory
as possible.
Memory management is a very difficult task to do well, and no single scheme is best in all
circumstances. The job is considerably simplified if the memory manager knows something in
advance about the lifetimes of its storage objects. For this reason, languages typically provide
several different kinds of dynamic storage objects which have different lifetime patterns.
The simplest pattern is a totally unrestricted lifetime. Such an object can be born and die at
any time under explicit control of the programmer. Nothing can be predicted about the lifetimes
of these objects, which are generally stored in an area of memory called the heap. Whenever a new
one is born, the storage manager tries to find a sufficiently large unused area of heap memory to
contain it. Whenever the storage manager learns of the death of a heap object, it takes note of the
fact that the memory is no longer in use.
There are many problems in recycling memory. First of all, the blocks in use may be scattered
about the heap, leaving many small unused holes instead of one large area. If no hole is large
enough for a new storage object, then the new object cannot be created, even though the total size
of all of the holes is more than adequate. This situation is called memory fragmentation.
Second, a memory manager must keep track of the holes so that they can be located when
needed. A third problem is that two or more adjacent small holes should be combined into one
larger one. Different heap memory managers solve some or all of these problems in different ways.
We will talk about some of them later in this chapter.

160

CHAPTER 6. MODELING OBJECTS

Because of the difficulty in managing a heap, it is desirable to use simpler, more efficient but
restricted memory managers whenever possible. One particularly common pattern of lifetimes is
called nested lifetimes. In this pattern, any two objects with different lifetimes that exist at the
same time have well-nested lifetimes; that is, the lifetime of one is completely contained within the
lifetime of the other. This pattern arises from block structure and procedure calls.
Storage for local block variables and procedure parameters only needs to exist while that block
or procedure is active. We say that a block is active when control resides within it or within some
procedure called from it. A storage object belonging to a block can be born when the block begins
and die when the block ends, so its lifetime coincides with the time that the block is active. Blocks
can be nested, meaning that a block B that starts within a block A finishes before A does. It
follows that the lifetimes of any storage objects created by B are contained within the lifetimes of
objects created by A.
Dynamic Stack Storage
Storage for objects with nested lifetimes can be managed very simply using a stack, frequently
called the run-time stack. This is an area of memory, like the heap, on which storage objects are
allocated and deallocated. Since, in the world of nested lifetime objects, younger objects always die
before older ones, objects can always be allocated and deallocated from the top of the stack. For
such objects, allocation and deallocation are very simple processes. The storage manager maintains
a stack allocation pointer which indicates the first unused location on the stack. When a program
block is entered, this pointer is incremented by the number of bytes required for the new storage
object(s) during the allocation process. Deallocation is accomplished at block exit time by simply
decrementing the stack allocation pointer by the same number of bytes. This returns the newly
freed storage to the storage pool, where it will be reused.
In languages that support both heap and stack storage objects, the stack objects should be
used wherever possible because their lifetime is tied to the code that uses them, and the birth and
death processes are very efficient and automatic. (This is the reason that stack-allocated objects
are called auto in C.)
Storage managers typically use stack storage for a variety of purposes. When control enters a
new program block, a structure called a stack frame, or activation record, is created on the top of
the stack. The area past the end of the current stack frame is used for temporary buffers and for
storing intermediate results while calculating long arithmetic expressions [Exhibit 6.17, right side].
A stack frame3 includes several items: parameters, local variables, the return address, and the
return value (if the block is a function body). It also contains two pointers, called the static link
and dynamic link [Exhibit 6.17, left side].
Let us define the lexical parent of a block to be that block which encloses it on the program
listing. The lexical parent of the outermost block or blocks is the system. A lexical ancestor is a
parent or the parent of a parent, and so on. The static link points to the stack frame of the current
3

The rest of this section explains the structure of the stack for a lexically scoped, block-structured language.

161

6.3. THE STORAGE MODEL: MANAGING STORAGE OBJECTS

Exhibit 6.17. The structure of the run-time stack.

The order of the parts within a stack frame is arbitrary, as is the relationship of global storage and
program code to the stack. The diagrams indicate a functional arrangement of the necessary kinds
of information.
A single stack frame:
The program and stack at run time:

Program Code
Global and
static storage

Parameters
Return Address
Dynamic Link
Static Link
Return Value
Local Variables

Top of stack

Stack frame for

oldest block
Stack frames for
other blocks
Stack frame for
newest block

Top of stack

Temporary
locations

blocks lexical parent. At run time, these links form a chain that leads back through the stack
frames for all the blocks that lexically enclose the current block. Since the location of a lexical
ancestors frame is not predictable at compile time, the chain of static links must be followed to
locate a storage object that was allocated by an ancestor. This is, of course, not as efficient as
finding a local object, and it is one good reason to use parameters or local variables wherever
possible.
The dynamic parent of a block is the block which called it during the course of execution and to
which it must return at block exit time. The dynamic link points to the stack frame of the current
blocks dynamic parent. This link is used to pop the stack at block exit time.
The static and dynamic links are created when the stack frame is allocated at run time. During
this process, several things are entered into the locations just past the end of the current frame.
This process uses (and increments) the local-allocation pointer which points to the first free location
on the stack. Before beginning the call process, this pointer is saved. The saved value will be used
later to pop the stack. The sequence of events is as follows:

162

CHAPTER 6. MODELING OBJECTS

1. The calling program puts the argument values on the stack using the local-allocation pointer.
Typically, the last argument in the function call is loaded on the stack first, followed by the
second-last, and so on. The first argument ends up at the top of the stack.
2. The return address is written at the top of the stack, above the first argument.
3. The current top-of-stack pointer is copied to the top of the stack. This will become the new
dynamic link field. The address of this location is stored into the top-of-stack pointer.
4. The static link for the new frame is written on the stack. This is the same as either the static
link or the dynamic link of the calling block. Code is generated at compile time to copy the
appropriate link.
5. The local allocation pointer is incremented by enough locations to store the return value and
the local variables. If the locals have initializers, those values are also copied.
6. Control is transferred to the subroutine.
At block exit time, the stack frame must be deallocated. In our model, the return value is in
the frame (rather than in a register), so the frame must be deallocated by the calling program.
To do this, the value in the dynamic link field of the subroutines frame is copied back into the
top-of-stack pointer, and the local-allocation pointer is restored to its value prior to loading the
arguments onto the stack.
Stack storage enables the implementation of recursive functions by permitting new storage objects to be allocated for parameters and local variables each time the function is invoked. An
unlimited number of storage objects which correspond to each parameter or local name in the
recursive function can exist at the same time: one set for every time a recursive block has been
entered but not exited [Exhibits 6.18 and 6.19]. Each time a recursive procedure exits, the corresponding stack frame is deallocated, and when the original recursive call returns to the calling
program, the last of these frames dies. The number of storage objects simultaneously in existence
for a recursive program is limited only by the program logic and the amount of storage available
for stack allocation, not by the number of declared identifiers in the program.
Dynamic Heap Storage
There are situations in which heap storage must be used because the birth or death patterns
associated with stack storage are too restrictive. These include cases in which the size or number
of storage objects needed is not known at block entry time and situations in which an object must
outlive the block in which it was created.
Heap Allocation. Heap allocation is invoked by an explicit allocation command, which we will
call ALLOC. Such commands can occur anywhere in a program, unlike local variable declarations
which are restricted to the beginning of blocks. Thus a heap-allocated object can be born at any

6.3. THE STORAGE MODEL: MANAGING STORAGE OBJECTS

163

Exhibit 6.18. A recursive function in Pascal.

The following (foolish) recursive function multiplies jj inputs together. Exhibit 6.19 traces an
execution of this function.
FUNCTION product (jj: integer):integer;
VAR kk: integer;
BEGIN
IF jj <= 0 THEN product := 1
ELSE BEGIN
readln(kk);
product := kk * product(jj-1);
END
END;

Exhibit 6.19. Stack frames for recursive calls.

If the function product in Exhibit 6.18, were called with the parameter 2, two recursions would
happen. Assume the inputs 25 and 7 were supplied. Just before returning from the second recursion
the stack would contain three stack frames as diagrammed. The ? in a stack location indicates
an undefined value.
Points at stack frame from lexical parent

SL

jj : 2
return : ?
kk : 25
jj : 1
return : ?
kk : 7
jj : 0
return : 1
kk : ?

Temporary
locations

Key:
SL is a Static Link
DL is a Dynamic Link

DL
Stack frame from original call
DL
Stack frame from first recursive call
DL
Stack frame from second recursive call
Top of stack

164

CHAPTER 6. MODELING OBJECTS

Exhibit 6.20. Dynamic allocation in FORTH.

Allocation: HERE hexpressioni ALLOT
Storage can be allocated dynamically in the dictionary, which stores the symbol table and
all global objects. The programmer is given access to the top-of-dictionary pointer through
the system variable named HERE. The code above puts the current value of HERE on the
stack. Then it evaluates the expression, which must produce an integer, N. Finally, ALLOT
adds N bytes to the dictionary pointer. The address of the newly allocated area is left on the
stack; the user must store it in a pointer variable.
Deallocation: Users must write their own storage management routines if they wish to free
and reuse dynamic storage.

time. ALLOC reserves storage in the heap and returns a reference to the new storage object. The
allocation process for heap storage is somewhat more complicated than that for stack storage, since
there may be two places to look in the heap for available storage. Initially there is only a large,
empty area with an associated allocation pointer which is incremented (like the stack pointer) when
storage is allocated from that area. After some objects have died, there may also be a freelist which
contains references to these formerly used locations. Clearly, items on the freelist might be scattered
all over the heap and be of quite varied sizes. The memory manager must contain algorithms to
keep track of the sizes and merge adjacent free areas, and these algorithms must be fast to avoid
degrading the performance of the system.
An ALLOC command takes some indication of the required size of the new object and finds
and reserves that much memory. Either it returns a reference to the memory location, or it stores
that reference in a pointer variable which thereafter gives access to the new object. The ways these
actions are incorporated into current languages are truly varied [Exhibits 6.20 through 6.25]. The
new storage object is used, later, by dereferencing this pointer, and it remains alive as long as the
pointer or some copy of the pointer points at it.

Exhibit 6.21. Dynamic allocation in LISP.

Allocation: (cons hexpr1i hexpr2i)
This allocates a new list cell and returns a reference to it. The left field of the cell is initialized
to the result of evaluating hexpr1i and its right field to hexpr2i.
Deallocation: Most LISP systems rely on garbage collection.

6.3. THE STORAGE MODEL: MANAGING STORAGE OBJECTS

165

Dead Heap Objects. We call a dead heap object garbage. Management of dead heap objects
is very different from stack management. A heap object dies either when the last reference to the
object is destroyed (let us call this a natural death) or when it is explicitly killed using a KILL
command.
The run-time system of a language translator must manage heap storage allocation just as it
manages stack frame allocation. However, when an object dies a natural death, both the programmer and the run-time system may be unaware of that death. A hole containing garbage is left in
the heap at an unknown location.
A KILL command takes a reference to a storage object, kills the storage object, and puts the reference onto the freelist, where it can be recycled. In languages that implement KILL, programmers
who use extensive amounts of dynamic storage are strongly urged to keep track of their objects
and KILL them when they are no longer useful. (In general, this will be well before the object dies
a natural death.) It is only through an explicit KILL command that the system can reclaim the
storage.
Recycling a dead heap cell is more complex than recycling stack cells. The system cannot
simply decrement the heap allocation pointer because, in general, dead objects are in the middle
of the heap, not at the end. A data structure called a freelist is generally used to link together the
recycled cells and provide a pool of cells available for future reuse. Conceptually, a freelist is just
a list of reclaimed and reusable storage objects. However, that is not a simple thing to implement
efficiently. The objects might all be interchangeable, or they might be of differing sizes. In any case,
they are probably scattered all over heap memory. The language designer or system implementor
must decide how to organize the freelist to maximize its benefits and minimize bookkeeping.
Ignore dead cells. The easiest implementation of KILL is to ignore it! Although this seems
to be a misimplementation of a language, it has been done. The Pascal reference manual for the
Data General MV8000 explicitly states that the Dispose command is implemented as a no-op.
This compiler runs under the AOS-VS operating system, which is a time-shared, paged, virtual
memory system.
The philosophy of the compiler writer was that most programs dont gobble up huge amounts
of storage, and those that do can be paged. Old, dead storage objects will eventually be paged out.
If all objects on the page have died, that page will never again be brought into memory. Thus the
compiler depends on the storage management routines of the operating system to deep-six the
garbage. This can work very well if objects with similar birth times have similar useful lifetimes.
If not, each of many pages might end up holding a few scattered objects, vastly increasing the
memory requirements of the process and degrading the performance of the entire system.
Keep one freelist. One possibility is to maintain a single list which links together all free
areas. To do this, each area on the list must have at least enough bytes to store the size of the area
and a link. (On most hardware that means 8 bytes.) Areas smaller than this are not reclaimable.
Many or most C and Pascal compilers work this way.

166

CHAPTER 6. MODELING OBJECTS

Exhibit 6.22. Dynamic allocation in C.

Allocation: In the commands that follow, T and basetype are types, N is an integer. The
malloc function allocates one object of type T or size N bytes. calloc allocates an array of
N objects of type basetype and initializes the entire array to zero. Both malloc and calloc
return a reference to the new storage object. The programmer must cast that reference to
the desired pointer type and assign it to some pointer variable.
malloc(sizeof(T))
malloc(N)
calloc(N, sizeof(basetype))
Deallocation: free(ptr);
ptr must be a pointer to a heap object that was previously allocated using malloc or calloc.
That object is linked onto a freelist and becomes available for reuse.

A compiler could treat this 8-byte minimum object size in three ways. It could refuse to allocate
anything smaller than 8 bytes; a request for a smaller area would be increased to this minimum.
This is not as wasteful as it might seem. Those extra bytes often have to be allocated anyway
because many machines require every object to start on a word or long-word boundary (a byte
address that is divisible by 2 or 4).
Alternately, the compiler could refuse to reclaim anything smaller than the minimum. If a tiny
object were freed, its bytes would simply be left as a hole in the heap. The philosophy here is that
tiny objects are probably not worth bothering about. It takes a very large number of dead tiny
objects to fill up a modern memory.
A fragmentation problem can occur with these methods for handling variable-sized dead objects.

Exhibit 6.23. Dynamic allocation in Pascal.

Allocation: New(hPtrNamei);
The hPtrNamei must be declared as a pointer to some type, say T. A new cell is allocated of
type T, and the resulting reference is stored in the pointer variable.
Deallocation: Dispose(hPtrNamei);
The object pointed at by PtrName is put onto the freelist.

6.3. THE STORAGE MODEL: MANAGING STORAGE OBJECTS

167

With many rounds of allocation and deallocation, the average size of the objects can decrease, and
the freelist may end up containing a huge number of tiny, worthless areas. If adjacent areas are not
glued together, one can end up with most of the memory free but no single area large enough to
allocate a large object.
Joining adjacent areas is quick and easy, but one must first identify them. Ordinarily this would
require keeping the freelist sorted in order of address and searching it each time an object is freed.
This is certainly time-consuming, and the system designer must decide whether the time or the
space is more valuable.
One final implementation of variable-sized deallocation addresses this problem. In this version,
each allocation request results in an 8-byte header plus the number of bytes requested, rounded up
to the nearest word boundary. At first this seems very wasteful, but using the extra space permits
a more satisfactory implementation of the deallocation process.
The 8-byte header contains two pointers that are used to create a doubly linked circular list of
dynamically allocated areas. One bit somewhere in the header is set to indicate whether the area
is currently in use or free. The areas are arranged on this list in order of memory address. Areas
that are adjacent in memory are adjacent in the list. Disposing of a dead object is very efficient
with this implementation: one only needs to set this bit to indicate free. Then if either of the
neighboring areas is also free, the two can be combined into one larger area.
When a request is made for more storage, the list can be scanned for a free cell that is large
enough to satisfy the new request. Scanning the list from the beginning every time would be very
slow, since many areas that are in use would have to be bypassed before finding the first free area.
But a scanning pointer can be kept pointing just past the most recently allocated block, and the
search for a free area can thus start at the end of the in-use area. By the time the scanner comes
back around to the beginning of the list, many of the old cells will have been freed. Thus we have a
typical time/space trade-off. By allocating extra space we can reduce memory-management time.
Keep several freelists. A final strategy for managing free storage is to maintain one freelist
for every size or type of storage object that can be freed. Thus all cells on the list are interchangeable, and their order doesnt matter. This simplifies reallocation, avoids the need for identifying
adjacent areas, and, in general, is simpler and easier to implement. This reallocation strategy is
used by Ada and Turing [Exhibits 6.24 and 6.25].
One of the problems with heap-allocated objects is in knowing when to kill them. It is all too
easy to forget to kill an object at the end of its useful lifetime or to accidentally kill it too soon.
This situation is complicated by the way in which pointer structures may share storage. A storage
object could be shared by two data structures, one of which is no longer useful and apparently
should be killed, while the other is still in use and must not be killed. If we KILL this structure we
create a dangling pointer which will eventually cause trouble. Identifying such situations is difficult
and error prone, but omitting KILL instructions can increase a programs storage requirements
beyond what is readily available.
For this reason, some languages, such as LISP, automate the process of recycling dead heap

168

CHAPTER 6. MODELING OBJECTS

Exhibit 6.24. Dynamic allocation in Ada.

Allocation: NEW htypei

(hexpressioni)

This allocates an object of the type requested. If the optional expression is supplied, it is
evaluated and the result is used to initialize the new storage object. NEW is a function that
returns a reference to the new object. The programmer must assign this reference to a variable
of an ACCESS type (a pointer).
Deallocation: Explicit deallocation is not generally used in Ada. In most Ada implementations, the dynamically allocated cells in a linked structure are automatically deallocated
when the stack frame containing the pointer to the beginning of the structure is deallocated.
Some Ada implementations contain full garbage collectors, like LISP.
When it is necessary to recycle cells explicitly, a programmer may use a generic package named
Unchecked_Deallocation.4 This package must be instantiated (expanded, like a macro, at
compile time) for each type of cell that is to be deallocated. Each instantiation produces a
procedure, for which the programmer supplies a name, that puts that kind of cell on a freelist.
(Different cell types go on different freelists.) Use of this facility is discouraged because it
may lead to dangling pointers.

Exhibit 6.25. Dynamic allocation in Turing.

Allocation: new hcollectioni, hptri
To dynamically allocate cells of a particular type, the programmer must explicitly declare
that the type forms a collection. The new command allocates one cell from the desired
collection and stores the reference in hptri.
Deallocation: free hcollectioni, hptri
The object pointed at by ptr is returned to the collection it came from, where it will be
available for reuse. The pointer object ptr is set to nil.

6.3. THE STORAGE MODEL: MANAGING STORAGE OBJECTS

169

objects. In cases where a KILL command does not exist or is not used, heap objects still die, but
the memory manager is not aware of the deaths when they happen. To actually recycle these dead
heap cells requires a nontrivial mechanism called a garbage collector, which is invoked to recycle
the dead storage objects when the heap becomes full or nearly full.
A garbage collector looks through storage, locating and marking all the live objects. It can tell
that an object is alive if it is static or stack-allocated or if there is a pointer to it from some other
live object anywhere in storage. The garbage collector then puts references to all of the unmarked,
dead areas on the freelist, where the allocator will look for reusable cells.
While this scheme offers a lot of advantages, it is still incumbent on the programmer to destroy
references to objects that are no longer needed. Furthermore, garbage collection is slow and costly.
On the positive side, the garbage collector needs to be run only when the supply of free storage
is low, which is an infrequent problem with large, modern memories. Thus garbage collection has
become a practical solution to the storage management problem.

6.3.2

Dangling References

The case of a name or a pointer that refers to a dead object is problematical. This can happen
with heap storage where the programming language provides an explicit KILL command. The
programmer could allocate a heap-object, copy the resulting reference several times, then KILL the
object and one of its references. The other references will still exist and point at garbage. These
pointers are called dangling references or dangling pointers.
This situation could also arise if the program is able to store references to stack-allocated
objects. Assume that a reference to a stack-allocated variable declared in an inner block could be
stored in a pointer from an outer block. During the lifetime of the inner block, this can make good
sense. When storage for the inner block is deallocated, though, the reference stored in the outer
block becomes garbage. If it were then used, it would be an undefined reference.
Initially, a dangling reference points at the value of the deallocated variable. Later, when the
storage is reused for another block, the address will contain useful information that is not relevant
to the pointer. Thus the pointer provides a way of accessing and modifying some random piece of
storage.
Serious errors can be caused by the accidental use of a dangling reference. Because the storage
belonging to any inner block might be affected, the symptoms of this kind of error are varied and
confusing. The apparent error happens at a point in the program that is distant from the block
containing the dangling reference. If the inner blocks are modified, the symptoms may change; the
part that was malfunctioning may start to work, and some other part may suddenly malfunction.
This kind of error is extremely difficult to trace to its cause and debug.
Because of the potential severe problems involved, pointers into the stack are completely prohibited in Pascal. Pascal was designed to be simple and as foolproof as possible. The designers
opinion is that all programmers are occasionally fools, and the language should provide as much
protection as possible without prohibiting useful things.
Pascal completely prevents dangling pointers that point into the stack by prohibiting all pointers

170

CHAPTER 6. MODELING OBJECTS

to stack-allocated objects. The use of Pascal pointers is thus restricted to heap storage. Linked
lists and trees, which require the use of pointers, are allocated in the heap. Simple variables and
arrays can be allocated on the stack. Address arithmetic is not defined. Although this seems like
a severe restriction, its primary bad effect is that subscripts must be used to process arrays, rather
than the more efficient indexing methods which use pointers and address arithmetic.
In contrast, the use of pointers is not at all restricted in C. The & operator can be used freely
and lets the programmer point at any object, including stack objects that have been deallocated.
When control leaves an inner block, and its stack frame is deallocated, any pointer that points into
that block will contain garbage. (An example of such code and corresponding diagrams are given
in Exhibits 9.25 and 9.26.)
A language, such as C, which permits unrestricted use of addresses must either forgo the use of
an execution stack or cope with the problem of dangling references. Allocation of parameters and
local variables on the execution stack is a simple and efficient method of providing dynamically
expandable storage, which is necessary to support recursion. Alternatives to using a stack exist
but have high run-time overhead.
The other possibility is to permit the programmer to create dangling references and make it
the programmers responsibility to avoid using them meaninglessly. A higher level of programming
skill is then required because misuse of pointers is always possible. A high premium is placed on
developing clean, structured methods for handling pointers.
One design principle behind the original C was that a systems programmer does not need a
foolproof language but does need free access to all the objects in his or her programs. Permitting
free use of pointers was also important in the original C because it lacked other important features.
Since structured objects could not be passed coherently to and from subroutines, any subroutine
that worked on a structure had to communicate with its calling program using a pointer to the
structure.
In the new ANSI C, this weakness is changed but not eliminated. It permits coherent assignment
of structures, but not arrays. Similarly, structures may be passed to and from functions without
using pointers, but an array parameter is always passed by pointer. Thus if C had the same
restriction on pointers that Pascal has, the language would be much less powerful, perhaps not
even usable.
How, then, can Pascal avoid the need to have pointers to stack objects? It has two facilities
that are missing in C:
Compound stack-allocated objects are coherent. They can be operated on, assigned, compared, and passed as parameters coherently.
References to objects can be passed as parameters by using the VAR parameter declarator.
Unfortunately, returning a compound value from a function is not permitted in the standard
language and must be accomplished by storing the answer in a VAR parameter.
Most standard algorithms and data structures can be coded easily within these restrictions, using
the fact that compound objects are coherent. Some experts assert that Pascal is a better language

6.3. THE STORAGE MODEL: MANAGING STORAGE OBJECTS

171

for these applications because the programmer does not need to exercise as much care.
On the other hand, there are, occasionally, situations in which the pointer restrictions in Pascal
prevent the programmer from coding an algorithm at all, and others in which the code would have
been much more efficient if the programmer had used pointers to stack objects. One might say
that C is a better language for these applications.

Exercises
1. Define and explain the relationship among: external object, program object, storage object,
and pointer object.
2. What is the difference between a variable and a pointer variable?
3. By what two means can a value be placed in a storage object? What is the difference between
the two processes?
4. What is the difference between destructive assignment and coherent assignment?
5. What is multiple assignment? How is it used?
6. Why do languages that implement assignment as a function allow the programmer more
flexibility than those that implement assignment as a statement?
7. Define dereferencing. Which abstract function implements it?
8. Choose a programming language with which you are familiar. Give a sample of code in which
dereferencing is explicit. Give another in which it is implicit.
9. What language contexts must be enumerated in order to define the implicit dereferencing
rules in a language?
10. Give some examples of FETCH operators which are used to dereference pointer expressions.
11. How does a pointer assignment differ from an ordinary assignment?
12. Given p1 pointing at this list of integers, write legal pointer assignments for p2 and p3.

p1

50

p2

25

p3

62

13. In the C language, what are the meanings of & and * in pointer notation?

172

CHAPTER 6. MODELING OBJECTS

14. In C, what is the meaning of the name of an array? Do you need an & when assigning a
pointer to an array? Why or why not?
15. What is the lifetime of a storage object?
16. What is the difference between a static storage object and a dynamic one?
17. Why is a language that only supports static storage items limiting?
18. What is memory management? Why is it important?
19. What is a run-time stack? A stack-frame? A heap?
20. What are the purposes of the static and dynamic links in a stack frame?
21. Name two languages in which local variables can be declared and are allocated each time a
subroutine is entered. Give examples of local variable declarations.
22. Name a language in which local variables cannot be defined.
23. What is static local storage? In what ways is it better than global storage and ordinary local
storage? Give an example, in some language, of a declaration that creates static local storage.
24. Suppose a language allows initial values to be specified for local variables, for example, the
following declarations which define X and initialize it to 50:
Language
Ada
C

Declaration
X: integer := 50;
int X=50;

When and how often does initialization happen in the following two cases?
a. X is an ordinary local variable.
b. X is a static local variable.
25. Explain the differences in lifetime, accessibility, and creation time between:
a. An ordinary local variable.
b. A static local variable.
c. A global variable.
See page ??.
26. Name a language that does not support dynamic storage at all. (All storage is static.) Explain
two ways in which this limits the power of the language.

6.3. THE STORAGE MODEL: MANAGING STORAGE OBJECTS

173

27. What is the purpose of an ALLOC command? What is garbage? A freelist? What is the
function of a KILL command?
28. Give examples in LISP and C of expressions that allocate nonstack (heap) storage dynamically.
29. Explain the reallocation strategy used by Turing. What are its advantages?
30. What is a dangling reference, and what problems can be caused by it?
31. Name a language in which pointers exist but can only point at dynamically allocated heap
objects, not at objects allocated in the stack.
32. Name a language in which a pointer can point at any variable and pointer arithmetic is
possible. Give an example of code.
33. Write a paragraph discussing the following questions: In what sense does FORTH or assembly
language have pointers? What can be done with them? Are there any restrictions? Name
two common errors that can occur with this kind of pointers.
34. Choose three languages from the list: APL, Ada, Pascal, C, FORTH, and assembler. What
restrictions are there on the use of pointers in each? What effect do these restrictions have
on flexibility and ease of use of the language? What effect do they have on the safety of
code?

174

CHAPTER 6. MODELING OBJECTS

Chapter 7

Names and Binding

Overview
This chapter discusses the definition, implementation, and semantics of names. The
meaning of a symbol or name in a programming language is its binding. In most
languages, all sorts of entities can have names. Binding creates an association between
a name and its storage object. Binding can be static; the name is bound to the object
when the object is allocated and remains bound throughout the program. Or, binding
can be dynamic. In this case, a name can be bound to one object, unbound, and rebound
to another within the run of a program. Constant declarations bind a symbolic name
to a literal value.
The scope of a name is that part of the program where a name is known by the translator.
Naming conflicts occur when some name is accidentally used more than once within a
linear program. Modularity and block structure allow the programmer to limit the scope
of a name to a block and all its nested blocks. It is the job of the interpreter or compiler
to determine the proper meaning of ambiguous names according to the semantics of the
language.

7.1

The Problem with Names

This section concerns the ways that we define symbols, or names, in a programming language, give
those names meaning (or meanings), and interpret references to names. In lambda calculus this
issue is very simple; every name acquires a unique meaning in one of two ways:
175

176

CHAPTER 7. NAMES AND BINDING

1. Some names are defined, by declaration, to stand for formulas.

2. Parameter names acquire meaning during a reduction step. When a lambda expression is
applied to an argument, that argument becomes the meaning of the parameter name.
Lambda calculus is referentially transparent: wherever a name appears in an expression, the defining
formula can be substituted without changing the meaning of the expression. The reverse is also true;
the meaning of an expression does not change if a defined name is substituted for a subexpression
which matches its defining formula.
Thus lambda calculus makes a simple one-to-one correspondence between names and meanings. Most programming languages, however, are not so simple. This section tries to explain and
straighten out all the myriad ways in which real languages complicate the naming problem.

7.1.1

The Role of Names

We use names to talk about objects in a computer. In simplest terms, a name is a string of
characters that a programmer can write in a program. Different languages have different rules for
the construction of names, but intuitively, a name is just like a word in Englisha string of letters.
A name must be given a meaning before it can be used. The meaning of a name is its binding, and
we say the name is bound to that meaning.
While objects can be created dynamically in most languages, names cannot. Names are written
in the program, and the text of the program does not change when the program is executed. Bindings, though, change when objects are created or destroyed. They attach the changing collection
of objects to the fixed collection of names.
Naming would need little explanation if languages followed a one object-one name rule. However, the situation is not so simple. Languages permit a bewildering mismatch between the number
of objects that exist and the number of names in the program. On the one hand, an object can have
no name, one name, or multiple names bound to it. On the other hand, a name can be bound to
no object (a dangling pointer), one object (the usual case), or several objects (a parameter name in
a recursive function). This complexity comes about because of block structure, parameters, recursion, pointers, alias commands, KILL commands, and explicit binding commands. In this section,
we explore the way names are used in writing programs and the binding mechanisms provided by
various languages.
Symbolic names are not necessary for a computer to execute a program: compilers commonly
remove names altogether and replace them by references. Nor are names necessary for a person
to write a program: the earliest method for programming computers, writing absolute machine
code, did not use names. Nonsymbolic programming requires considerable skill and extraordinary
attention to detail, and symbolic assemblers, which permit the programmer to define names for
locations, were a great leap forward because names help the programmer write correct code. It is
much easier for a human to remember a hundred names than a hundred machine addresses.
In addition, names have an important semantic aspect that is appreciated by experienced programmers. A program that uses well-chosen names that are related to the meaning of the objects

7.1. THE PROBLEM WITH NAMES

177

Exhibit 7.1. Bad names.

You may enjoy the challenge of figuring out what this function does without using diagrams.
At least one highly experienced Pascal programmer failed on his first try.
TYPE
my_type = x_type;
x_type = RECORD next: char; prior: my_type END;
FUNCTION store_it (temp: my_type, jxq:
VAR jqx, first: my_type;
BEGIN
first := temp;
jqx := temp.prior;
WHILE (jqx <> NIL) DO BEGIN
IF jqx.next = jxq
THEN jqx := NIL
ELSE BEGIN

char):

my_type;

first := jqx;
jqx := jqx.prior
END
END;
store_it := first;
END;

being named is much easier to debug than a program with randomly chosen, excessively general, or
overused names like J and temp. A program that uses names inappropriately can be terrible
to debug, since the human working on the program can be misled by a name and fail to connect
the name with an observed error. It is even harder for another programmer, unfamiliar with the
program, to maintain that code.
The function definition in Exhibit 7.1 was written with names purposely chosen to disguise and
confuse its purpose. The English semantics of every name used are wrong for the usage of the
corresponding object. A compiler would have no trouble making sense of this clear, concise code,
but a human being will be hindered by ideas of what names are supposed to mean and will have
trouble understanding the code. You may enjoy trying to decode it before reading further.
Several naming sins occur in Exhibit 7.1:
Two names were used that have subtle differences: jxq, jqx.
Nonsuggestive names were used: temp, my_type, x_type, jxq.
Suggestive names were inappropriately used: store_it names a search routine that does no

178

CHAPTER 7. NAMES AND BINDING

Exhibit 7.2. Good names.

TYPE
list_type = cell_type;
cell_type = RECORD value:

char; next:

list_type END;

FUNCTION search (letter_list: list_type, search_key: char):

VAR scanner, follower: list_type;
BEGIN
follower := letter_list;
scanner := letter_list.next;
WHILE (scanner <> NIL) DO BEGIN
IF scanner.value = search_key
THEN scanner := NIL
ELSE BEGIN
follower := scanner;
scanner := scanner.next
END
END;
search := follower;
END;

list_type;

storing. Next names a value field, rather than the traditional pointer. Prior names a
pointer field pointing at the next item in the list.
A name was used that seemed appropriate on first use but did not reflect the actual usage
of the object: first started as a pointer to the first thing in the list, but it is actually a
scanning pointer.
A list of good name substitutions for the program in Exhibit 7.1 is:
my_type = list_type prior = next
jxq = search_key
x-type = cell_type
next = value
jqx = scanner
store_it = search
temp = letter_list first = follower
Rewritten with the names changed, the purpose of this code should be immediately apparent.
Try reading the code in Exhibit 7.2. Anyone familiar with Pascal and with list processing should
understand this code readily.

7.1.2

Definition Mechanisms: Declarations and Defaults

All sorts of entities can have names in most languages: objects and files (nouns), functions and
procedures (verbs), types (adjectives), and more. Depending on the rules of the language, the

7.1. THE PROBLEM WITH NAMES

179

Exhibit 7.3. Predefined names in Pascal.

This is a list of all the names that are predefined in UCSD Pascal.
types
constants
files
functions
procedures

integer, real, Boolean, char, text

NIL, TRUE, FALSE, MAXINT
input, output
odd, eof, eoln, abs, sqr, sqrt, sin, cos, arctan,
ln, exp, trunc, round, ord, chr, succ, pred
read, readln, write, writeln, get, put, rewrite,
reset, page, new, dispose, pack, unpack

programmer might or might not be permitted to use the same name for entities in different classes.
As in English, there must be some way to give meaning to a name and some way to find the meaning
of a name when it is used. Declarations, defaults, and the language definition itself are the means
used in programming languages to give meaning to names.
The symbol table is a data structure maintained by every translator that is analogous to a
dictionary. It stores names and their definitions during translation. 1
A name must be defined before it can be used. In some languages this happens the first time
it is used; in others all names must be explicitly declared. A declaration is a statement that causes
the translator to add a new name to its list of defined names in the symbol table.
Many functions, types, and constants are named by the language designer, and their definitions
are built into all implementations of the language. We call these primitive symbols [Exhibits 7.3
and 7.4]. These names are not like other reserved words. They do not occur in the syntax that
defines the language, and the programmer may define more names in the same categories. They
are a necessary part of a language definition because they provide a basic catalog of symbols in
terms of which all other symbols must be defined.
1

This structure has also been called the environment or dictionary.

Exhibit 7.4. Predefined names in C.

These are the names that are predefined in C:
types
int, long, short, unsigned, float, double, char
constants

NULL (TRUE, FALSE, and EOF are also defined in many versions of
hstdio.hi, the header file for the standard I/O package.)

functions

Every C implementation has a library, which contains I/O functions, numeric

functions, and the like. The libraries are fairly well standardized from one
implementation to another and are far too extensive to list here.

180

CHAPTER 7. NAMES AND BINDING

In many interpreted languages the programmer is not required to declare types, because allocation decisions do not have to be made in advance of execution, and at execution time, the type
of a datum can often be determined by examining the datum itself. Names are generally added
to the symbol table the first time they are mentioned. Thus names are typeless. These languages
are sometimes called typeless because types are not declared and not stored in the symbol table
with the names.
Objects, on the other hand, are never typeless. Every storage object has a fixed size, and size
is one aspect of type. Every program object has a defined encoding, another aspect of type. In a
typeless language, the type of an object must still be recorded. Since the type is not stored with
the name, it must be encoded somehow as part of the object itself.
In a compiled language, the type of each name must be supplied either by a declaration or by
a default so that the compiler can know how many bytes to allocate for the associated storage
object. The type is stored in the symbol table with the name, and remains unchanged throughout
the rest of translation. Pascal requires that a type be declared for each name. FORTRAN permits
the programmer to write explicit declarations, but if an identifier does not appear in a declaration,
a default type will be used which depends on the first letter of the name. The original C permitted
function return types and parameter types, but not variable types, to be declared as integer by
default.

7.1.3

Binding

In compiled languages, a name exists only in the symbol table at translation time and objects
exist only at run time. Names are gone before objects are created; they are not part of objects.
In interpreted languages, names and objects coexist. In both cases, a name acquires one or more
meanings during the course of translation, by a process called binding.
Binding creates an association between a name (in the symbol table) and a storage object (an
area of memory). We can picture a binding as a pointer from the name to the storage object. A
binding differs from an ordinary pointer, though, because it reaches from the systems storage area
into the programmers area. Moreover, in compiled languages, the binding spans time as well as
space. At compile time it holds the location where an object will someday be allocated. Finally,
bindings are unlike pointers because the translator automatically dereferences bindings but does
not dereference pointers. We represent bindings in our diagrams as arrows (like pointers) but drawn
in boldface, because they are not ordinary pointers.
Binding is invoked by the translator whenever a declaration is processed but can also be invoked
explicitly by the programmer in many languages. A binding is static if it never changes during the
lifetime of the program. Otherwise it is said to be dynamic. At any time a name might be bound
to a particular object to which it refers, or it might be unbound, in which case it refers to nothing
and is said to be undefined, or it might be multiply bound to different objects in different scopes.
Names of variables, types, pure values, and functions are identified and recorded in the symbol
table during translation of a program. Another column of the symbol table records the bindings.
Like allocation, binding can be static, block structured, or dynamic.

7.1. THE PROBLEM WITH NAMES

181

Exhibit 7.5. A symbol table with static bindings.

Symbol Table
type
real

name binding

Run-Time Memory
storage object

length

array[1..4] of integer ages

Typed Languages / Static Binding

Most of the familiar languages (COBOL, FORTRAN, ALGOL, C, Pascal, Ada) belong to a class called
typed languages. In these languages each name defined in a program unit has a fixed data type
associated with it, and often declared with it.
In the oldest and simplest of these languages, such as symbolic assemblers and COBOL, name
binding is static. A name is bound to an object when the object is allocated and remains bound to
the same storage object until the end of the program. In such languages, when a meaning is given
to a name, that name retains the meaning throughout the program.
Static binding occurs in typed languages that are non-block structured. In a static language,
there is no concept of a program block enclosed within another program block, producing a local
program scope in which a name could be redefined. 2 A static binding associates a name with a
storage object of fixed type and size at a fixed memory address.
Static binding can be implemented simply by using three columns in the symbol table to store
the name, type, and binding [Exhibit 7.5]. We can describe this kind of symbol table as flatit
has the form of a simple one-dimensional list of entries, where each entry has three fields.
Each declaration (explicit or default) specifies a name and a type. It causes the compiler to
select and set aside an area of storage appropriate for an object of that type. Although this storage
will not exist until run time, its address can be computed at compile time and stored in the symbol
table as the binding for the name. Note, in Exhibit 7.5, that the run-time memory contains only
the storage object; the symbol table no longer needs to be present. It was used to generate machine
code and discarded at the end of translation. 3
A Typed Language with Dynamic Binding
FORTH is an interactive, interpretive language embedded in a program development system. A
complete system contains an editor, an assembler, an interpreter, and a compiler. This compiler
does not generate machine code, rather, it lexes and parses function definitions and produces an
2

However, additional names can be bound to a COBOL object, by using REDEFINES, as explained in Section 7.1.4.
Some translators are embedded in systems that provide a symbolic debugger. These systems must keep the
symbol table and load it along with the object code for the program.
3

182

CHAPTER 7. NAMES AND BINDING

Exhibit 7.6. Names, types, and bindings in FIG FORTH.

This dictionary segment contains two words, an integer and an array of four integers. The righthand column has 4 bytes of memory per line in the diagram.
name :
link :
type :
body :
name :
link :
type :
body :

6len
gth
?
int

(Length of name followed by name.)

(Pointer to previous word in dictionary.)
(Pointer to run-time code for integer variables.)
(4 bytes, properly called the "parameter field".)

4age
s
length
int
(Pointer to run-time code for integer variables.)
(16 bytes of storage, enough for four variables.)

intermediate program form that can be interpreted efficiently.

FORTH is a typed language. Its symbol table, called the dictionary, is only a little more
complex than the simple, flat symbol table used for a static language. The dictionary is organized
into several vocabularies, each containing words for a different subsystem. Each vocabulary is
implemented by a simple, flat symbol table. Unlike COBOL and assembler, though, FORTH is an
interactive language system. A user wishing to create a new application subsystem is permitted
to create a new vocabulary or to add to an existing vocabulary. The user may alternate between
defining objects and functions, and executing those functions. The dictionary may thus grow
throughout a session.
The dictionary contains an entry for each defined item [Exhibit 7.6]. Function names, variable
names, and constant names are all called words. Entries for all the primitive words are loaded
into the dictionary when you enter the FORTH system. A dictionary entry is created for a userdefined word when a declaration is processed, and it will remain in the dictionary until the user
gives the command to FORGET the symbol. The FORTH dictionary is stack-structured; new items
are added at the top of the stack and can be defined in terms of anything below them on the stack.
Each entry has four fields:
The name field holds the name of the word, stored as a string whose first byte contains the
length of the name.
The link field is used to organize the dictionary into a data structure that can be searched
efficiently. Searching must be done during translation, when the definition of a function refers
to a symbol, or at run time when the interpreter evaluates a symbolic expression interactively.
The implementation of the link field and its position relative to the name field varies among

7.1. THE PROBLEM WITH NAMES

183

different versions and implementations of FORTH. Exhibit 7.6 shows the relationships defined
for FIG FORTH.
The code field is the functional equivalent of a type field. It identifies, uniquely, the kind of
object this word represents (function, variable, constant, or programmer-defined type).
The parameter field, or body, contains the specific meaning of the word. For constants, it is
a pure value. For variables it is a storage object. For functions, it contains code that can be
interpreted.
FORTH maintains a rudimentary sort of type information in the code field of each dictionary
entry. This field is a pointer to a run-time routine that determines the semantics of the name.
It is actually a pointer to some code which will be run whenever this word is used at run time.
This code defines the interpretation method for objects of this type. Thus constants, variables, and
user-defined types can be interpreted differently. Initially only the types function, variable,
and constant are built in, but others can be added. When a new type declarator is defined, two
pieces of code are given: one to allocate and initialize enough storage for an object of the new type,
and a second to interpret run-time references to the name of an object of this type. A pointer
to this second piece of code becomes the unique identifier for the new type, and also becomes the
contents of the code field for all objects declared with the new type.
FORTH differs from the simple static languages in one important way: it permits the user to
redefine a word that is already in the dictionary. The translator will provide a warning message, but
accept the redefinition. Henceforth the new definition will be used to compile any new functions,
but the old one will be used to interpret any previously compiled functions. This opens up the
possibility of redefining primitive symbols. The new definition can call the original definition and,
in addition, do more elaborate processing. The simple relationship between a name and its meaning
no longer holds at all.
The FORGET command is an unusual feature that has no counterpart in most language translators. It does not just remove one item from the dictionary, it pops the entire dictionary stack
back to the entry before its argument, forgetting everything that has been defined since! This is a
rudimentary form of symbol table management which does not have either the same purpose or the
same power as the stack-structured symbol tables used to implement block structure. A FORTH
programmer alternates between compiling parts of his or her code and testing them. FORGET lets the
programmer erase the results of part of a compilation, correct an error in that part, and recompile
just one part. Thus FORGET is an important program-development tool.
Typed Languages / Block Structured Binding. The connection between a name and its
meaning is further complicated by block structure. FORTH permits a new definition of a name to
be given, and it will permanently replace the old version (unless it is explicitly forgotten). A
block structured language permits this same kind of redefinition, but such a language will restore
the original definition after exit from the block containing the redefinition. Block structure and the
semantic mechanisms that implement it are taken up in Section 7.4.

184

CHAPTER 7. NAMES AND BINDING

Exhibit 7.7. A symbol table with dynamic binding.

Symbol Table

Run-Time Memory

name binding

storage object : type tag

length

17.3

: real

Explicit Dynamic Binding. Fully dynamic binding is available only in interpreted languages or
ones such as LISP with simple, uniform type structures. In such a language, types can be associated
with objects, not names, and are stored with the object in memory, rather than in the symbol table.
The symbol table has only two columns, the name and its current binding. The type must be stored
with or encoded into the object in memory, or discarded altogether as in assembly language. This
is illustrated in Exhibit 7.7.
With fully dynamic binding, a name can be unbound from one object and rebound to any
other at any time, even in the middle of a block, by explicit programmer command. In such a
language, the type of the object bound to a name may change dramatically, and these languages
are sometimes called typeless because no definite type is associated permanently with a name.
SNOBOL and APL are examples of this language class.
These typeless languages nevertheless commonly do implement objects of different types. For
example, in APL there are two basic types, number and character. These types are implemented
by attaching a type tag to the object itself, rather than to the name in the symbol table [Exhibit
7.8]. The symbol table contains only the name and the binding, and the programmer is permitted
to bind a name to any object. Thus at different times a name may be bound to storage areas of
different sizes, each with an associated type tag.
In such languages, binding often serves the same purpose as does assignment in Pascal and is
often mistaken for assignment. The essential difference is that assignment does not change the
storage object to which a name is bound, but changes the program object which is the contents of

Exhibit 7.8. Names, types, and binding in APL.

APL is a typeless language and so has no permanent association of types with names. Rather,
a type tag is associated with each storage object, and the combination may be bound to any name.
Symbol Table
name binding
length
ages

Run-Time Memory
storage object : type tag
: scalar number
: array of 4 numbers

7.1. THE PROBLEM WITH NAMES

185

Exhibit 7.9. Dynamic binding in APL.

On two successive executions of the input statement
Q2
one could legally supply the following inputs: 0 aeiouy0 (which is a character array) and 37.1 (which
is a number). Thus we would get first the following binding:
Symbol Table
name binding
Q

Run-Time Memory
storage object : type tag
a

u y

: array of 6 characters

and second the following binding:

Symbol Table
name binding
Q

Run-Time Memory
storage object : type tag
37.1

: scalar number

that storage object. Binding results in a different storage object being bound to a name.
The APL input command is: 2. Executing 2 causes a pure value of some type to be accepted
as input from the users terminal. The input may be a single number or character (called a scalar)
or it may be an array (or vector) of any length of either base type. The type of the value is
determined, storage is allocated and initialized to this value, and a reference to this storage is
returned. The operator binds this new object to the name on its left. Exhibit 7.9 uses the 2
operator to illustrate the dynamic nature of binding in APL.

7.1.4

Names and Objects: Not a One-to-One Correspondence

We have seen that although the number of names in a program cannot change dynamically, many
applications require that the storage in use expand dynamically. We, therefore, must be able to
refer to storage locations that do not have unique names. This need is met by using addresses,
or pointers, rather than names to refer to storage, and by binding the same name, recursively, to
objects in different stack frames.
Conversely, there are times when it is convenient to bind multiple names to a single storage
object or a part of an object. This is often done in order to associate a second data type with that
object, or to package logically independent objects together into a group to facilitate their handling.
(A full discussion of this topic is in Chapter 15.) The names given to this kind of declaration in
some common languages are as follows:

186

CHAPTER 7. NAMES AND BINDING

Exhibit 7.10. Sharing through EQUIVALENCE in FORTRAN.

DIMENSION P1(3)
EQUIVALENCE (P1, P2), (P1(2), P3)
The DIMENSION statement causes an array of length 3 to be allocated and bound to P1. The
EQUIVALENCE statement binds a second name to the base address of the array P1, and a third name
to a segment of that array, as illustrated below. P1, P2, and P3 now share storage.
P1 P2 P3

FORTRAN
COBOL
Pascal
C

EQUIVALENCE declaration
REDEFINES clause in a declaration
Variant record with no tag field
Union data type

When multiple-name binding is used, storage is not allocated for the second name, but it is
bound to the same address as the first and serves as a second way to refer to the same storage
object [Exhibit 7.10]. If used carelessly, this can cause bizarre misinterpretations of data.
On the other hand, misinterpreting the data on purpose can be an easy way to compute a
pseudo-random function. Exhibit 7.11 shows a variant record used as the input to a hashing
function. The data is really a character string, but the hashing function is told to construe it as
an integer and square it. The result is semantically meaningless but can be used to access a hash
table. A diagram of this dual-type object is shown in Exhibit 7.12.

7.2

Binding a Name to a Constant

A constant declaration permits the programmer to attach a symbolic name to a literal value.
Although giving a name to a constant does not change the meaning of a program, a judicious
choice of names can clarify the programmers intended meaning to other programmers. Defining
the name once and then using it many times in place of a frequently used value also makes it easier
to modify a program: only the definition needs to be changed to change all occurrences of that
value.
Many but not all languages permit the programmer to name constants. Some place severe
restrictions on the kinds of constants that can be named and/or on the ways initial values can be
specified. For example, Pascal requires that the value of a constant be a literal value of a primitive
type. Constant expressions may not be used as initial values [Exhibit 7.13] even though they could
easily be evaluated by the compiler, as they are in FORTRAN 77 [Exhibit 7.14].

7.2. BINDING A NAME TO A CONSTANT

187

Exhibit 7.11. A variant record used to compute a hash index.

We define a type that provides storage for 4 bytes, which can be interpreted either as a single
integer or an array of four characters. This type is useful in writing a hash function that takes a
character array as its parameter and uses integer operations to compute a hash address. Diagrams
of the variant-record object are shown in Exhibit 7.12.
TYPE twokinds = 1..2;
hashkey = RECORD CASE twokinds OF
1: (char_ar: ARRAY [1..4] of char);
2: (int_way: integer)
END;
FUNCTION hash (k: hashkey): integer;
BEGIN
hash := (k.int_way * k.int_way) MOD table_size
END;

Exhibit 7.12. Diagrams of a Pascal variant record.

The storage object for k in Exhibit 7.11 would have the following structure:

k.char_ar

k.int_way
This object has dual semantics. If the actual parameter were the string ABCD, the two
interpretations of the program object would be as follows:

k.char_ar
(in hexadecimal notation)
41

42

43

44

k.char_ar (shown as ASCII)

A

41424344
k.int_way
(The hexdecimal integer
41,424,344 equals the
decimal integer 1,094,861,636.)

188

CHAPTER 7. NAMES AND BINDING

Exhibit 7.13. Constant declarations in Pascal.

CONST pi = 3.1416;
max_length = 500;
max_index = 499; { Equals max_length - 1 }
A real constant and two integer constants are declared. As is often the case, the last one depends
on the second, but the dependency cannot be made explicit in Pascal and is indicated only by a
comment.

In Ada, constant declarations do not have to precede the rest of the program. A constant
declaration may come after variable declarations and depend on the initial values given to those
variables [Exhibit 7.15].
FORTH, like Ada, permits the declaration of a constant at any point during a program. A
constant definition consists of an expression, the word CONSTANT, and the name of the constant.
The expression will be evaluated first, and its value will be bound to the name. FORTH, unlike

Exhibit 7.14. Constant declarations in FORTRAN 77.

PARAMETER (PI = 3.1416, MAXLEN = 500, MAXIX = MAXLEN-1)
The same three constants are created as in the Pascal example of Exhibit 7.13. In FORTRAN it
is possible to use a constant expression to specify the value of a constant. PARAMETER declarations
must precede the executable code.

Exhibit 7.15. Mixed variable and constant initializations in Ada.

pi: CONSTANT float := 3.1416;
m: integer;
n: integer := 5;
max_length: CONSTANT integer := n * 100;
max_index: CONSTANT integer := max_length - 1;
The real constant pi and two integer variables are declared. The second variable is initialized
to 5. The integer constants max_length and max_index are declared and bound, respectively, to
500 and 499. These are legal defining expressions because n is previously initialized. The following
constant declaration would not be legal because the value of m is not defined:
max_width: CONSTANT integer := m - 1; -- Not legal.

7.2. BINDING A NAME TO A CONSTANT

189

Exhibit 7.16. Variable and constant declarations in FIG FORTH.

100 VARIABLE offset
500 CONSTANT max_length
max_length offset + CONSTANT max_index
A variable named offset is defined and initialized to 100. A constant named max_length is
defined and initialized to 500. A second constant named max_index is initialized to the sum of
offset plus max_length.

Ada, is interactive. This declaration is translated at run time, not at a prior compile time. Thus
there is no need to restrict the initializing expression, which may depend on the current (run-time)
value of any name [Exhibit 7.16]. The scope of the constant name is all of the program that follows
the declaration. 4
The original K&R C contained no special defining word for constants, but a semantics that
meets the same needs as the constant declaration in Pascal can be achieved by using the compiletime preprocessor that is part of C. Each #define statement is a macro definition and is expanded
by the preprocessor, replacing the defined names by the defining strings before the beginning of
actual compilation. [It is conventional, in C, to use capital letters for macro (constant) names and
lowercase for other identifiers.]
There is an important difference between a using a macro name and using a true constant. In
a constant declaration, if the constant value is defined by an expression, that expression will be
evaluated once, at compile time. In a macro definition, the expression will be evaluated at run
time every time the constant name is used. Of course this is inefficient, but it also leaves open the
possibility that the wrong answer might be computed because of unintended interactions between
the macro expansion and its context.
In Exhibit 7.17, the parentheses around the expression on the third line prevent unintended
interaction between the - operator in the expanded macro and other operators in the surrounding
program. Assume the parentheses were omitted, and consider this call on the constant MAX_INDEX:
totsize = MAX_INDEX * 10
This macro call would be expanded before parsing, yielding:
totsize = 500 - 1 * 10
Thus the value of totsize would be 490, not the intended 4990.
For two reasons, then, a macro-preprocessor does not wholly take the place of a constant
declarator in a language. First, it can lead to unnecessary run-time calculations unless the compiler
performs optimization on constant expressions. Second, macros can be tricky and misleading to
use. True constants are efficient and simple. Thus a const declaration, with semantics similar
to that in Ada, was added to the language by the ANSI C standard. The last line in Exhibit 7.17
4

Scope is explained in Section 7.4.

190

CHAPTER 7. NAMES AND BINDING

Exhibit 7.17. Constant declarations in C.

#define PI 3.1416
#define MAX_LENGTH 500
#define MAX_INDEX (MAX_LENGTH - 1)
const int max_index = MAX_LENGTH - 1;

/* ANSI C only */

shows the declaration of a constant integer. The initializing expression will be evaluated once, when
storage for max_index is allocated, and the result will be stored in the allocated object. Thereafter,
the value cannot be changed.

7.2.1

Implementations of Constants

In spite of the differences in syntax, the semantics of constants are nearly identical in FORTH,
FORTRAN, and Pascal.5 In all three languages, a name is entered into the symbol table, a value is
provided or calculated and bound to it at compile time, and that binding does not change thereafter.
These semantics can be implemented in two ways. One alternative is to evaluate the defining
expression for the constant before compilation, and substitute the result for the constants name in
the source code. There can be no possibility of accidentally assigning a new value to the constant,
because constant names are eliminated from the program before compile time. In this case, it might
not even be necessary to allocate any run-time storage for the constant. Instead of compiling the
code to fetch from a memory address (like a variable), the compiler might make the constant into
a part of the compiled code by generating a load immediate instruction.
The second implementation for constants permits their names to exist in symbolic form during
translation. This is necessary if the constant name is to have a block structured scope. The
translator allocates a run-time storage location for the constant, evaluates the defining expression,
and stores the result in the allocated space. This is the same way a variable would be initialized.
In this case, some mechanism must be embedded in the translator to prevent the programmer from
changing the constant by an assignment or read statement. We can say that this implementation
provides an initialized read only variable (IROV).
Constants are implemented as IROVs in FORTH. The storage set aside for constants and variables is the same, and the initialization process is the same. Different semantic behavior is achieved
at run time by associating different semantic routines with constants and variables. The translator
puts a constants value on the stack when its name is referenced. But when a variable name is
referenced, the storage address is placed on the stack and must be explicitly dereferenced to obtain
the value of the variable.
5

Except that the scope of definition of the constant name cannot be restricted to an inner block in FORTH or
FORTRAN.

7.3. SURVEY OF ALLOCATION AND BINDING

7.2.2

191

How Constant Is a Constant?

The IROV implementation of constants leads to a generalization that is found in Ada and ANSI C.
We see that the constant names in Pascal can have local scope. (That is, a constant name declared
in an inner block or subroutine is only known in that block.) It would be logical to evaluate the
initializing expression at block entry time, rather than at compile time.
If we defer evaluation and binding for a constant until block entry time, constants can be
created and initialized when local variables are allocated. They are essentially initialized read-only
variables. An initializing expression can contain references to parameters and variables in outer
blocks as well as to literal constants. With this interpretation, the defining expression is reevaluated
each time the block is reentered, leading to the situation that a constant might not be constant!
(Such a local constant would remain constant for the lifetime of its block, but the name might be
bound to a different constant value the next time the block was entered.)
The primary advantage of such a block entry constant is that a constant can be calculated based
on input parameters, and yet the language translator guarantees the integrity of the value: it can
not be accidentally changed. The cost is more time spent during block entry.

7.3

Survey of Allocation and Binding

Early High-Level Languages.

The designers of the earliest high-level languages expected the language users to want to represent
certain types of external objects. They made it easy for the programmer to represent those types.
COBOL records are a natural representation of a business transaction, being very much like the
paper representation of those transactions. FORTRAN makes the representation of numbers and
vectors natural. At this historical stage of language development, there was a close relationship
among (1) external objects being represented, (2) internal names for those objects, and (3) the
storage objects that implemented them. There is no question of what constitutes an object in these
languages: it is the combination of all of the above.
Allocation is static, and binding is static except for parameter binding. Each set of similar
external objects (policies, dimensions, etc.) is represented by a typed identifier. If the external
object has parts and subparts, each part in the representation can be named and referenced. When
an identifier is defined by the programmer, a storage object is allocated and bound to that identifier
for the life of the program.
Additional names can be bound to a storage object through REDEFINES or EQUIVALENCE
statements. Additional names can also be bound to objects temporarily by parameter binding
during a call on a subprogram.
Interactive Languages. An interesting and different set of choices was made in BASIC. Name
binding is totally static: subprograms do not have parameters, so unlike FORTRAN, there is no
need to deviate from totally static binding. On the other hand, allocation of storage for string

192

CHAPTER 7. NAMES AND BINDING

variables is dynamic. This is done by statically binding the string identifier to a pointer object and
dynamically binding the pointer to a string object in string storage.
Block Structured Languages. Concern over the difficulty of debugging large programs was
beginning to develop in the late 1950s when ALGOL was designed. ALGOL was a remarkably
advanced language for those years: it incorporated a clean design, adequate control structures, and
recursion.
Block structure, local variables, and parameter binding were devised to ease the problems
of name conflicts. With these new semantic mechanisms, the same identifier could be used to
correspond to different external objects in different program blocks. This eased the writing of large
programs by permitting the programmer to disregard or forget that an identifier had already been
used once, so long as the previous use was in an irrelevant context. Block structure is also needed
to support recursion.
Dynamic Languages.
With LISP, we see a radically different approach to modeling external objects and processes. LISP
objects are implemented by linked structures, called lists of storage objects called cells. A list is a
pointer to a cell or to a simple object called an atom. A cell is a pair of lists. Identifiers do not
have data types, but can be bound to any object at any time. Storage is allocated dynamically
for parameters and also by calling the allocation function CONS. (CONS takes two parameters. It
constructs a cell and initializes its two fields to those parameters.)
We see that the relationship between external objects and storage objects exhibited by the
block structured languages breaks down completely in LISP. External objects are represented by
lists, which can point at lists that represent smaller external objects. Big objects are made out
of references to smaller objects, which are true objects, not subfields. This is in contrast to the
COBOL concept that big objects may have many small parts, but these parts are not, literally, at
the same level.
Variable names exist in LISP, corresponding roughly to FORTRAN object names. However, the
number of variable names is not limited to the number of names introduced by the programmer;
names can be generated dynamically by a program. Variables may be bound to LISP objects,
but the binding is dynamic, not static, so that one variable name may refer to different program
objects with different storage amounts and locations at different times. Since storage objects may
be dynamically created and pointer assignment is provided, the size of a single program object is
bounded only by available storage.
APL, another early interactive language, has dynamic allocation and binding for parameter
names and local variables in functions. The basic objects in APL are arrays, either numeric or
alphabetic. A tremendous variety of array and matrix operations are predefined. These can produce
results of any size and shape, for which storage is allocated dynamically. The resulting references
are bound dynamically to untyped identifiers. Global names can be used without declaration.

7.4. THE SCOPE OF A NAME

193

Local variable names, parameter names, and a local name for the functions return value need to
be declared in the function header to distinguish them from global names.
Smalltalk is a relatively new language that extends the LISP concept of object by having all
objects belong to classes, either system-defined or programmer-defined. With this extension, objects
must now contain a class (or type) indicator as well as private storage. Again, big objects are made
out of smaller ones. It is significant that, as in APL, this type information is attached to the
storage object rather than to an identifier. It is this property that has earned the label objectoriented language for Smalltalk.
Combining Static and Dynamic Objects. In Pascal, ALGOL-like block structured allocation
and binding exist side-by-side with LISP-like dynamic creation and pointer assignment, although
the two facilities are not closely merged. Pascal programs tend to use either one allocation scheme
or the other, with little mixing, because you cannot point at stack-allocated objects.
In C, this restriction on what a pointer may point at is lifted. The address of any storage object
and any part of that object is available within the program and can be stored in a pointer. Pointers
may, therefore, point at any object.
Identifier binding is block structured in both C and Pascal. An identifier is bound to a new
storage object at block entry and remains bound to the same object until it is deallocated. Even
though objects can be dynamically allocated from heap storage, these are bound to pointers rather
than identifiers.
Data types are associated with identifiers, as in FORTRAN, not with objects, as in Smalltalk and
APL. Pointers have a declared base type and should only be bound to objects of the appropriate
type. C will give warning errors but still compile if this rule is breached. Pascal will fail to compile.

7.4
7.4.1

The Scope of a Name

Naming Conflicts

When symbolic programming languages were new, each identifier stood for one storage object, and
there was no need to distinguish between identifiers and internal names. Very soon though, people
began to compile programs that included subroutines written by other people. They began to
encounter naming conflicts when they would accidentally use some name that was used internally
by the subroutine.
The first approach to avoiding such name conflicts was to assign a unique prefix to all of the
names used in the subroutine. For example, a SIN subroutine that needed a local variable COUNT
might use the name SINCOUNT. The programmer needed only to avoid using names with that
prefix in other subroutines. To simplify the programmers task, some early symbolic languages
contained a PREFIX statement that would automatically prefix the name, so the programmer
could write COUNT and the translator would convert it to SINCOUNT.
As programs became longer, naming conflicts became a problem even within the same program.
A programmer working on one part of the program could not easily remember all of the names

194

CHAPTER 7. NAMES AND BINDING

used in other parts of the program, and accidental duplication of names could lead to bugs that
were very difficult to locate.
The concepts of modularity and scope were developed to alleviate this problem. A program is
written as many manageable-sized modules, each with a particular well-defined purpose. Modules
interact with each other only in well-defined ways. Names defined within a module, unless explicitly
exported, are not visible outside of the module. This introduced the new concept of the scope
of a name, which is the portion of code in which the name is meaningful. Scoped names could be
reused for new objects in different scopes.
The scope of a name is that part of the program in which the name is known and will be
recognized by the translator. Scope can be global, in which case the name is known throughout the
program, or it can be local, meaning that it is only known within that program block in which it
was defined. In programs with nested scopes, a name, N , may also be relatively global by being
declared in a block that is neither innermost nor outermost. In programs that are compiled in
many modules, names can also have external scope, which means that the name is known to all
modules.
Along with scoping came the need to distinguish between two concepts that until then had been
synonymousidentifiers and complete names. Identifiers are the symbols that the programmer
writes, and complete names are the things that receive bindings. The scope is added to the identifier
the programmer writes in order to obtain the complete name. Thus if COUNT is defined in two
modules SIN and SORT, then the complete names generated by the translator might be written
SIN.COUNT and SORT.COUNT. Only one declaration for an identifier may occur per scope, and
thus the translator can form a unique name by sequentially numbering the scopes it encounters
and concatenating a scope number with the identifier that the programmer used. An attempt to
redefine an identifier in the same scope is a translation-time semantic error.
In most modern languages, however, the programmer does not give names to all scopes. Nevertheless, the same rules apply. The identifier COUNT, defined in two different scopes s1 and s2,
denotes two different names. When programs are translated, the complete name is formed automatically by the translator. There is no need for the programmer to write the complete names, so
complete names are not a part of the syntax for programming languages. However, we still need
a way to write a complete name in this chapter, so we will write the name as a pair of a scope
number and an identifier, for example, (s1.COUNT) and (s2.COUNT).
The programmer refers to (s1.COUNT) when writing the identifier COUNT within scope s1
and has no way to refer to this variable elsewhere. This limitation is intentional and is what gives
scope its power. By providing no way to write the complete name (s1.y) outside of the scope s1, we
are making accidental references to (s1.y) impossible. The information stored in a named variable
is only visible or accessible within the scope of the name. It is hidden from the rest of the
program.
In most languages, each subroutine definition comprises a scope, and thus parameters can be
given dummy names. It is immaterial whether the identifiers used for dummy names are the
same as or different from names in the calling program, for the translator will generate different
complete names for them.

7.4. THE SCOPE OF A NAME

195

Language design came full-circle with the invention of object-oriented languages, where programmers, again, have reason to write complete names. Each object class has a name and defines
a scope. Within the definition of a class, the programmer uses simple names to refer to the class
members (objects and functions). Some class members are private, and are accessible only within
the class definition. However, a class may also have public members, which can be used by other
parts of the program.6 When this is done, a double colon symbol, called the scope-resolution
operator is written between the class name and the member name to form a complete name.

7.4.2

Block Structure

The idea of block structure, introduced in ALGOL-60, permitted the programmer to define nested
modules called blocks. In a block structured language, the programmer can introduce additional
blocks at quite arbitrary places, wherever they are convenient. Identifiers declared within a block
are translated to complete names whose scope is that block and all nested blocks. A well-written
program uses many short blocks, and names are declared in the smallest possible block. Each block
corresponds to a scope. Thus several sets of declarations might be relevant to a particular use of a
name in the program.
This situation can arise in C because a new block may by opened anywhere, using {, and
names may be defined at the beginning of any block.
In Pascal, blocks cannot be opened and closed at arbitrary places. But each function or procedure body is a block, and it may be nested inside other function or procedure definitions. 7 Thus we
arrive at nested scopes by a different path. The simple program in Exhibit 7.18 has three scopes
which are outlined by boxes. One scope is created by the main program, and two by functions
within it. Let us call the scopes A, B, and C. Scope A is the lexical parent of scopes B and C.
Nested blocks reintroduce ambiguity into the naming rules that scopes were invented to avoid.
We said previously that although an identifier might refer to different names in different scopes, the
identifier together with the scope in which it was contained was enough to disambiguate it. Now,
however, we see that scopes can be nested, so an instance of an identifier can appear simultaneously
in the several scopes of its enclosing blocks. Once again we are faced with an ambiguity. We must
determine which scope identifier to use in forming the complete name.
To resolve this ambiguity, we introduce yet another conceptthat of a defining occurrence of
an identifier. Each identifier is given meaning by a declaration. 8 At that point, the identifier, with
its type, is entered into the symbol table. All other occurrences of that identifier are nondefining
occurrences and are called uses. Uses are found in executable statements and initial value clauses.
An identifier can be redefined by a new declaration in a nested block, and this introduces a
new complete name. The scope of each complete name is the smallest scope that contains its
declaration. In order to translate a use of an identifier, the compiler must decide which complete
6

Object oriented scope and referencing rules are explained in Chapter 16, Section 16.4 and in Chapter 18, Section
18.2.
7
All function definitions in C are at the top level. They may not be nested within each other.
8
In some languages a name is defined by default when the translator sees the identifier for the first time.

196

CHAPTER 7. NAMES AND BINDING

Exhibit 7.18. Nested scopes in Pascal.

PROGRAM A;
VAR x, y: integer;

scope A

FUNCTION B( d: integer ): integer;

BEGIN B := x+d END;

scope B

FUNCTION C( d: integer ): real;

VAR x: real;
BEGIN x := 5.1; C := x+B(d+1) END;

scope C

BEGIN x:=3; y:= 0; writeln( C(y) ) END.

name corresponds to it. The rule used is called lexical scoping: 9 The complete name to which a
use, U, refers is the one produced by the nearest declaration for U. The nearest declaration is the
one in the smallest enclosing block.
Exhibit 7.18 has a short program with a nest of three scopes, defined by the main program,
(scope A) and the two subroutines (scopes B and C). The object identifiers in use are x, y and
d. Each parameter or local variable declaration defines a complete name, thus the complete names
formed from these identifiers are: (A.x), (A.y), (B.d), (C.x), and (C.d). Subroutine names
are visible in the scope of the enclosing block (so that they may be called from that block). Thus
the complete names of the subroutines are A.B and A.C. This program is rewritten using complete
names in Exhibit 7.19. Exhibit 7.20 shows the contents of the stack during execution of function B.
A straightforward implementation of lexical scoping works as follows. When a block or subprogram is entered, storage for its parameters and local variables is allocated in a stack frame on the
run-time stack. The stack frame also includes the return address for the subprogram and a pointer
to the stack frames of the lexical and dynamic parents of the block. The pointer to the lexical
parent is called the static link, and the pointer to the dynamic parent is called the dynamic link.
The static links are the means of implementing complete names, and they provide an easy way
to describe lexical scoping. To find the correct complete name for a use, U , the compiler must start
at the stack frame for the block that contains U . If there is no declaration for U in that stack
frame, start following the chain of static links backward through the stack. The first declaration
for U that you find is the relevant one. Following the static link can bypass a variable number of
stack frames that were allocated for blocks called by the lexical parent of the current block.
Once the compiler determines which scope defines U , the chain of stack frames can be short9

The word lexical is used because the scope of a name is a static property determined by how the program is
laid out on the listing.

7.4. THE SCOPE OF A NAME

197

Exhibit 7.19. Complete names in Pascal.

PROGRAM A;
VAR (A.x), (A.y): integer;
FUNCTION B( (B.d): integer ): integer;
BEGIN B := (A.x)+(B.d) END;
FUNCTION C( C.d: integer ): real;
VAR C.x: real;
BEGIN (C.x):=5.1; A.C:= (C.x)+A.B((C.d)+1) END;
BEGIN (A.x):=3; (A.y):= 0; writeln(A.C(A.y)) END.

Since x is not redefined in scope B, the x referred to in that scope is the one in the lexically enclosing
scope, which is scope A.

circuited so that run-time references will be more efficient. If the definition is found at a global
level, the address to which it is bound is static and may simply be compiled into the code. If
the definition of U is neither local nor global, but occurs in some scope between local and global,
extra storage can be allocated in the local stack frame for a pointer to the defining reference. At
block entry time the actual address of the proper binding can be determined by tracing through
the frames by the above method, then copied into this pointer area. A use of a relatively global
variable then becomes like a use of a VAR parameter, causing an automatic dereferencing of the
pointer. (VAR parameters are explained in Chapter 9, Section 9.2.)
A final, confusing aspect of nested scopes must be mastered: the phenomenon of masking. Redeclaration of an identifier locally will mask any definitions of the same identifier that are relatively
global to it. By this we mean that the object created by the declaration in the enclosing block cannot be accessed within the enclosed block because every use of that identifier in the enclosed block
will be associated with the object created in the smaller scope. In the Pascal example [Exhibits
7.18 and 7.19], the object (A.y) is accessible within function C because it was declared globally.
But the object (A.x) cannot be accessed within C because the local declaration for x will swallow
up all references to x.
A Block Structured Symbol Table.
Let us say that a complete name is born when the compiler begins translating the block in which
it is defined. It dies at the end of that block. The symbol table must include only names that are
alive, since dead names must not be used to interpret any references.

198

CHAPTER 7. NAMES AND BINDING

Exhibit 7.20. A stack diagram for nested scopes.

This diagram shows the contents of the stack during execution of the program in Exhibit 7.18, at
the point that the main program has called function c and c has called function b, but b has not
yet returned.

Frame for block c

Frame for block b

Key:
SL is a Static Link
DL is a Dynamic Link

x: 3
y: 0

Global storage
- block a

SL

d: 0

DL

x: 5.1
d: 1

DL

Top of stack

In a block structured language, a name becomes visible when it is born and invisible again
when it dies. However, a living name also becomes temporarily invisible when it is masked by a
declaration in an inner block, and it becomes visible again when the masking variable dies.
Block structured binding cannot be implemented with a flat symbol table (that is, a onedimensional list) because bindings for invisible names must be retained, and bindings for dead
names must be discarded. Thus the symbol table grows and shrinks, and we can describe it as
stack-structured. For each name declared in a block, a type and binding are pushed onto the
stack for that name when the compiler begins translating the block and are popped off when the
translator reaches the lexical end of the block. There are two ways to organize the stackeither
the bindings for the block can be inserted as a group, or each variable name can have an associated
stack of bindings. Exhibit 7.21 shows a program in which several variables are declared in more
than one block. Note that a function name is in the scope of its enclosing block, not the block
created by the function definition. Exhibit 7.22 shows this stack-structured symbol table with a
stack of bindings for each multiply defined name.

7.4.3

Recursive Bindings

The number of names in a particular program is determined by the programmer and does not
change after the program is translated. However, any language with recursion permits creation
of an unlimited number of stack-allocated objects at run time. On each invocation of a recursive
function, new storage is allocated, in a new stack frame, for parameters and local variables. This
new storage is bound to the parameter and local names and remains bound to them until control
exits from the subprogram. We say that each stack frame corresponds to one dynamic scope.
Thus each name declared in a recursive function must be bound simultaneously to several objects
in different dynamic scopes. If a subprogram has called itself five times and not yet returned once,

7.4. THE SCOPE OF A NAME

199

Exhibit 7.21. A Pascal program demonstrating block structure.

This program contains a function within a function. During translation of the innermost function block, block III, three sets of bindings are recorded in the symbol table. This is diagrammed
in Exhibit 7.22. However, only the bindings for the innermost block are active, or visible, at this
time.
program demo_bind;
I
const a=10;
var b, c: real;
function d (a:real):

real;

function c (b:integer):
begin c:=b/2 end;

II
real;

III

begin d := a+c(b) end;

begin writeln(a+trunc(d(c))) end.

Exhibit 7.22. A symbol table with block structured bindings.

We depict a block structured symbol table using columns for the objects type, block identifier,
name, and binding. The symbol table entry for each name is a stack of types and bindings. These
are pushed onto and popped off of the symbol stacks at the beginning and end of translation of the
relevant program block.
Symbol Table
type

block name binding

real variable

II

integer constant

integer variable

III b

real variable

function(integer): real

II

real variable

function(real): real

Run-Time Memory
storage object

10

A piece of executable code

A piece of executable code

200

CHAPTER 7. NAMES AND BINDING

there are six sets of storage objects simultaneously bound to the local names (one for the original
call and one for each of the five recursive calls).
Recursively bound names are ambiguous, and this ambiguity is not like the static ambiguity
introduced by block structure. In a block structured program, a name may be simultaneously
within the scope of several definitions [Exhibit 7.18]. But we are able to identify and resolve the
resulting ambiguity at compile time by identifying the lexical scope of each name and forming a
unique complete name from the block identifier and object identifier.
This method for disambiguation will not work for the dynamic, run-time ambiguity caused by
recursion because the recursive invocations all come from the same lexical block. A single symbol
declaration, such as the declaration for jj or kk in Exhibit 6.18, produces more than one allocation
and binding. The disambiguation rule that applies here is the rule for dynamic scoping: the most
recent active binding is used. To find that binding, start searching in the current stack frame. If the
name is not defined there, follow the dynamic link back to the stack frame of the calling function,
and so on. All frames on the stack will be examined in order, until the definition of the identifier
is found.
Lexical versus Dynamic Scoping. The rule for dynamic scoping is used in place of the lexical
scoping rule in non-block-structured languages such as APL and the older forms of LISP. In any
language where both recursion and nested scopes are supported (such as Pascal or C), both lexical
and dynamic scoping rules must be used. The rule for lexical scoping determines the mapping from
multiply defined identifiers onto complete names. Then the rule for dynamic scoping defines the
meaning of names in recursive scopes.
For simple situations, lexical and dynamic scoping produce the same result. However, in general, they do not. The difference is illustrated by the interpretation of the global variable x in
Exhibits 7.23 (LISP) and 7.25 (Pascal). LISP uses dynamic scoping, as diagrammed in Exhibit 7.24,
and Pascal uses lexical scoping, as diagrammed in Exhibit 7.26 The functions defined in the two
languages are identical except for the scoping rules used to interpret them, but they compute and
print different answers.

7.4.4

Visibility versus Lifetime.

The scope of a name is not necessarily the same as the lifetime of the object to which it is bound.
This mismatch of scope and lifetime can happen in several ways. The most familiar is by using a
reference (VAR) parameter in a subroutine call. During the procedure call the dummy parameter
name is bound to an actual parameter. Storage for a calling block has a longer lifetime than the
called block, since the called block will exit first. Thus, within the procedure, the parameter name
refers to an object with permanence greater than the scope of the name.
In languages with dynamic allocation, such as Pascal and C, list and tree structures can be built
out of heap-allocated objects. As each new cell of the structure is allocated, a reference to it is
stored in its predecessor. A reference to the first cell of the structure must be stored in a named,
stack-allocated, pointer variable. Frequently the scope of these pointers is less than global, while

7.4. THE SCOPE OF A NAME

201

Exhibit 7.23. Dynamic scoping in LISP.

Program Notes
Define a function named A,
with two parameters, x and y.
The function prints the result of calling a function named C
on the parameter y.

(defun A
(x y)
(printc (C y)))
(defun B
(d)
(+ d x))

Define a function named B,

with one parameter, d.
The function returns the result of adding d and x. Since x is
not defined locally, it is a global reference .

(defun C
(d)
(let (x 5.1))
(+ x (B (+ d 1))))

Define a function named C,

with one parameter, d.
Set local variable x to value 5.1.
Return the result of adding x to the result of executing function B on the parameter d incremented by 1.

(A 3 0)

Call function A with parameters 3 and 0. It will print the

number 11.2.

Exhibit 7.24. Diagram of bindings in dynamic scoping.

Stack frames (with dynamic links) are shown for the stage of execution at which all the functions
in Exhibit 7.23 have been called and none has returned. The name x is used but not defined in
function B, so its meaning is determined by following the dynamic links backward until the definition
of x in function C is encountered.
In function A :

In function C :

In function B :

5.1

Dynamic parent: LISP system

Dynamic parent: A
Dynamic parent: C, so x
refers to the binding in C.

202

CHAPTER 7. NAMES AND BINDING

Exhibit 7.25. Lexically scoped functions in Pascal.

PROGRAM a;
VAR x, y: integer;
FUNCTION b(d: integer):integer;
BEGIN b:= x+d END;
FUNCTION c(d: integer):real;
VAR x: real;
BEGIN x:= 5.1; c:= x + b(d+1) END;
BEGIN x:= 3; y:= 0; writeln(c(y)) END.
We define a main program, named a, and two functions, b and c. The program is the lexical parent
of both functions. When this is executed, a calls c which calls b. The number 9.1 will be printed.

Exhibit 7.26. Lexical scoping in Pascal.

At one point during execution, the main program has called function c which has called function b,
so stack frames for all three coexist. Program a is the dynamic parent of c, which is the dynamic
parent of b.
Pascal is a lexically scoped language, so one follows the static link from the current stack frame
to find the meaning of nonlocal names. The x referred to in function b is not the object in cs
frame, but the one in as frame, because a is the lexical parent.
In program a :

In function c :

In function b

5.1

Dynamic parent: system

Lexical parent: system
Dynamic parent: a
Lexical parent: a
Dynamic parent: c
Lexical parent: a

7.4. THE SCOPE OF A NAME

203

heap storage is permanent. Thus the head-pointer points at an object that could live longer than
itself and could be passed outward by copying its reference into a pointer with a longer lifetime.
Nonhierarchical Sharing. A serious shortcoming of lexical scoping is its strict hierarchical
nature: names defined within a block cannot be exported and made known outside the block. In a
language that is limited to hierarchical scoping, there are useful protection and sharing mechanisms
that cannot be defined or emulated. Examples of such data facilities are named common storage
in FORTRAN and packages in Ada. Such facilities can be superimposed on lexical scoping only
by expanding a languages semantic basis to include more than one kind of scoping.
An Ada package has two classes of names, some which are known outside the scope of the
package, some known only within the package. 10
In FORTRAN, access to a storage object may be shared by several subroutines. To do this
you place the objects name in a COMMON statement and include identical COMMON declarations in
all subroutines that need access to the object. The user can create and name several independent
COMMON areas. These named COMMON areas can be used to set up a nonhierarchical sharing structure
[Exhibit 7.27]. This sharing structure cannot be built in a strictly block structured language, since
block structure and lexical scoping either permits no sharing of data at all, or indiscriminate sharing
among all subroutines at the same level.
A COMMON area does not belong to any single subroutine, but is allocated in the program
environment area where it can be accessed by any subroutine containing the right declaration. The
declaration for a COMMON statement supplies names (with associated types) which will be bound to
successive locations in the common area.
It is the programmers job to ensure that all subroutines that share a common area contain
compatible declarations. The compiler will not check that. If the order or types of the names
declared in two subroutines differ, everything will compile and link, but compute nonsense. On the
other hand, the particular names declared in different subroutines are completely arbitrary and can
be different.
An example of the use of COMMON to achieve nonhierarchical sharing is shown in Exhibit 7.27.
There are three common areas, and matching common declarations are included in the pair of
subroutines that share each area. Subroutines POINT and BLANK share the storage area named
POINTBLANK which contains the variables X and Y. Likewise, subroutines POINT and CHECK share
area CHECKPOINT and BLANK and CHECK share BLANKCHECK. The common areas and bindings created
by these definitions are diagrammed in Exhibit 7.28. Note that CHECK and POINT call the variables
in area CHECKPOINT by different names, but the data types of the components do match.
Static Local Storage. ALGOL OWN variables and C variables with the static storage class are
examples of another kind of mismatch between lifetime and scope. These variables are statically
allocated and are immortal like global variables, but the scope of the names of these objects is
block structured. Each time a block is entered in which a static variable is declared, the variable
10

See Chapter 16 where this topic is fully developed.

204

CHAPTER 7. NAMES AND BINDING

Exhibit 7.27. Named common in FORTRAN.

Just the COMMON declarations are given here for three subroutines that share storage in a nonhierarchical fashion.
SUBROUTINE POINT
COMMON /POINTBLANK/ X, Y(100)
COMMON /CHECKPOINT/ JJ, KK, LL
SUBROUTINE BLANK
COMMON /POINTBLANK/ X, Y(100)
COMMON /BLANKCHECK/ F(5), G
SUBROUTINE CHECK
COMMON /BLANKCHECK/ F(5), G
COMMON /CHECKPOINT/ J, K, L

Exhibit 7.28. A diagram of common storage in FORTRAN.

This is a storage diagram of the common areas created in Exhibit 7.27. An arrow represents a
binding.
X
Y
F
G

BLANK

...

Area POINTBLANK :
X
Y
JJ
KK
LL

POINT
Area CHECKPOINT :

CHECK
F
G
J
K
L

Area BLANKCHECK :

7.5. IMPLICATIONS FOR THE COMPILER / INTERPRETER

205

name becomes known again and refers to the static object, which becomes accessible. The run-time
system ensures that the name always refers to the same storage address. At block exit time the
name becomes undefined, and outside the block, no name refers to the object. The object is then
invisible and remains inaccessible and unchangeable until its block is reentered. 11
This is an important facility. It permits the value of a variable to be retained between executions
of a subprogram and yet protected from external tampering. The need for this is apparent if you
consider an output buffering procedure which must remember how full it left the buffer in order
to know where to store the next value. The buffer pointer must be retained between calls on
the output function, yet it should be hidden from all other parts of the program to ensure
unhindered operation. The ability to hide information is now recognized as a cornerstone of
good programming practice.
In a language such as Pascal, which does not support static local objects, the buffer pointer
would have to be declared globally (all globals are static). This works, but it increases the scope
of the pointer to the entire program and creates the possibility that the value of the pointer can be
accidentally changed by some remote and irrelevant section of code.

7.5

Implications for the Compiler / Interpreter

A language translator must determine the proper meaning of an ambiguous name according to
the semantics defined for that language. We need to make a distinction here between the way
interpreters and compilers handle the task.
One of the essential differences between interpreters and compilers stems from the fact that
more information is often available at run time than at compile time. An interpreter can query
the user about the required amount of data on each run and allocate arrays exactly as long as
are needed. A compiler must allocate storage before actual data values are known, and therefore,
perhaps, before actual storage requirements are known. The programmer using a compiler must,
therefore, establish array lengths adequate for the maximum size data set the program will ever
process. If a program is to process data sets whose size is highly variable, its array lengths will be
excessive on most runs, and considerable storage will be allocated that is never used or needed.
Typed languages can be translated by a compiler. Type declarations permit a compiler to
anticipate how much storage will be needed in the future to hold program objects. Commands can
be compiled to increment and decrement the stack allocation pointer by the appropriate amounts
when control enters and leaves program blocks at run time. Typeless languages cannot be compiled
in the same sense; the source code can be parsed and symbol tables can be set up, but storage
management must be done dynamically because the size of the storage objects that will be needed
is not predictable.
In an interpreter, translation and execution proceed at the same time. Dynamic allocation
with dynamic binding is the natural and easy method to implement. The symbol table exists
during execution, and binding of allocation address to symbol is done when the allocation happens.
11

An extended example of the use of static storage in C is given in Chapter 16.

206

CHAPTER 7. NAMES AND BINDING

Each symbolic reference is interpreted just before it is executed, in the context of the results of
prior computations, and thus the most recent binding of a symbol is used. The path of execution
of any program may depend on inputs and conditionals and, therefore, be impossible to predict.
Consequently, dynamic binding may bind a name differently from run to run, causing inconsistent
interpretations of global references and errors that are hard to track down or identify.
In a compiler, on the other hand, all symbolic names are discarded at the end of compilation,
before execution begins (except when a symbolic debugger is in use). All binding of symbolic name
to storage location is done within the compiler, even for local variables that become allocated in the
middle of execution. For locals, the compiler determines the address, relative to the current stack
frame, that they will occupy in the future when they become allocated. Thus, for an ambiguous
name, a compiler must maintain a stack of bindings that is parallel to the stack of allocation areas
that will exist in the future when the program is executed. A binding is pushed onto the stack at
the time the compiler begins to translate a subprogram and is popped off it when the end of the
subprogram is compiled. The binding used to interpret a symbolic reference in the program is the
one on the top of the stack. This translation scheme implements lexical scoping: the binding used
to interpret a name is the one in the smallest enclosing block in which the symbol is defined. Lexical
scoping is the only method that can be implemented by a compiler. A compiler can determine how
definitions are nested, but it cannot guess the order in which they will be executed.
Lexical scoping is considered to be superior for two reasons. First, the binding that will be used
is always predictable, and, therefore, programs with lexical scoping are easier to debug. Second,
there are languages, for example LISP, for which it is common to write both interpreters and
compilers. The interpreter is used while a program is being developed and debugged, then the
compiler is used to produce code that executes faster. Since dynamic and lexical scoping produce
different semantics, it has often been the case that a program, fully debugged under an interpreter
using dynamic scoping, will not work when it is compiled and will have to be debugged again.
The scoping discipline affects the meaning of a program and, therefore, should be part of a formal
definition of any programming language.
Interpreting Block Structure. Block structure is implemented using stack allocation and block
structured binding. The following is a description of the operation of block structure in an interpreted language:
1. Storage objects are allocated on the run-time stack whenever a block is entered. One object
is allocated for each locally declared name or parameter. Local variables are usually not
initialized. Parameter storage is initialized to the actual parameter values or references to
variable parameters.
2. The new storage object is bound to the local name, giving the name an additional binding.
If the block is a recursive procedure, the name could have bindings created in the enclosing
blocks, in which case more than one storage object is bound to the same name, and the
meaning of the name is the most recent (dynamic) binding. If control leaves the block and
then reenters it, the name will be bound to a different storage object.

7.5. IMPLICATIONS FOR THE COMPILER / INTERPRETER

207

3. Local bindings remain static until control leaves the block at block exit time. A given name
refers to the same storage object from block entry until block exit. At block exit, all local
storage objects are deallocated and local identifiers revert to their prior bindings (possibly
undefined). The freed storage is made available for reuse by popping the run-time stack.
The number of storage objects on the stack changes only when blocks are entered or exited,
and then only changes by the number of names (locals + parameters) declared in that block.
Compiling Block Structure. The implementation of block structure in a compiled language
is not quite the same as the implementation in an interpreted language. The difference is that
the compiler deals with storage that will be allocated rather than storage that is allocated. The
address at which a stack frame will start is completely unpredictable at compile time. However,
the number, size, and order of the items in each future stack frame is known to the compiler.
A common frame-management scheme is to have a central core in each stack frame that
contains everything except parameters and local variables. A calling program puts parameters on
the stack and increments the stack pointer (which is kept in a machine register which we will call
the SP ). Then the calling program fills in the static link, dynamic link, and return address in the
core area of the stack frame. It puts a pointer to the dynamic link field of the new frame into a
machine register. (Let us call this register the FP, for frame pointer.) The contents of the new
dynamic link is the current value of FP. Finally, the program branches to the subroutine.
The subroutine can now find its stack frame because the address is in the FP. It can use the SP
to find the top of the stack. The first action of the subroutine is to increment the SP to allocate
space for local variables. The subprogram code is compiled to refer to its parameters and local
variables using addresses that are relative to the FP. The last parameter might be FP 2, and
the first might be FP 12. The local variables will be in locations such as FP + 4 and FP + 20.
Stack locations past the SP are available for use as a scratch pad and will be used for temporary
storage during calculations.
At block exit, several things happen:
The return value, if any, of the subprogram is loaded into a machine register. (It is often
called the return register.)
The dynamic link field is copied back into the FP.
The SP is decremented by the size of the stack frame.
A branch is taken to the return address.
This erases all trace of the subprogram.

Exercises
1. What is a name? What is the meaning of a name? Do languages follow a one object-one
name rule? Explain.

208

CHAPTER 7. NAMES AND BINDING

2. Do compilers need names to execute a program? Explain. What is the semantic aspect of a
name? Why is this important for human interpretation of a program?
3. What is the role of the symbol table? How are new names added to it?
4. In APL, why are names typeless?
5. What are primitive symbols? Why are they necessary?
6. How do names relate to objects in a compiled language? In an interpreted one?
7. What is binding? How is it invoked?
8. What do we call storage that is allocated once at load time and remains until program exit,
but whose usage is restricted to one function?
9. Define and contrast static binding, dynamic binding, and block structured binding.
10. In modern functional languages, assignment is not supported, and in APL, it is rarely used.
Dynamic binding takes the place of assignment. Explain how this is implemented.
11. In block structured languages, when is a name visible? Invisible?
12. This question and the next seven involve the following skeletal program written in C. How
many name scopes are defined in this program skeleton? Draw a box around each block and
label these blocks with the letters A, B, C, etc.
int x, y, z;
fun1()
{
int j, k, l;
{
int m, n, x;
...
}
}
fun3()
{
int y, z, k;
...
}
main()
{ ... }
13. Name a block that is nested within another block.
14. In which scopes are the global variables x, y, and z all accessible?
15. What variables are accessible to the main function?

7.5. IMPLICATIONS FOR THE COMPILER / INTERPRETER

209

16. Are k in fun1 and k in fun3 the same variable? Why or why not?
17. Are the global variables y and z accessible in fun3? Why or why not?
18. In which blocks does j refer to a global variable? In which blocks is j a local variable?
19. In which block or blocks can we use both m and k?
20. Name a language in which storage can be shared by two functions, neither of which is enclosed
within the other, but kept private from all other functions and from the main program.
Explain how to accomplish this goal in your language.
21. How are FORTRANs named common storage and Adas packages examples of nonhierarchical data facilities?
22. What is the function of a constant declaration? Why is it useful to the programmer?
23. What is the difference between defining a true constant whose meaning is a constant expression
and using a macro to define a symbolic name for that expression?
24. How are constants implemented in these languages: Pascal, Ada, FORTRAN, FORTH?
25. What is a naming conflict?
26. How did the concept of scope and modularity resolve the problem of naming conflicts? What
is the difference between an identifier and a complete name?
27. What is a block structure? What are nested blocks? What is the scope of the complete name
of identifiers declared within a block?
28. Consider a language in which all variables must be declared with fully specific types. Is this
language more likely to be interpreted or compiled? Why?
29. Why is the scope of a name not equal to its lifetime? Explain.
30. Explain how the compiler and interpreter each determine the semantics of an ambiguous
name.

210

CHAPTER 7. NAMES AND BINDING

Chapter 8

Expressions and Evaluation

Overview
This chapter introduces the concept of the programming environment and the role of
expressions in a program. Programs are executed in an environment which is provided
by the operating system or the translator. An editor, linker, file system, and compiler
form the environment in which the programmer can enter and run programs. Interactive language systems, such as APL, FORTH, Prolog, and Smalltalk among others, are
embedded in subsystems which replace the operating system in forming the programdevelopment environment. The top-level control structure for these subsystems is the
Read-Evaluate-Write cycle.
The order of computation within a program is controlled in one of three basic ways:
nesting, sequencing, or signaling other processes through the shared environment or
streams which are managed by the operating system.
Within a program, an expression is a nest of function calls or operators that returns
a value. Binary operators written between their arguments are used in infix syntax.
Unary and binary operators written before a single argument or a pair of arguments
are used in prefix syntax. In postfix syntax, operators (unary or binary) follow their
arguments. Parse trees can be developed from expressions that include infix, prefix and
postfix operators. Rules for precedence, associativity, and parenthesization determine
which operands belong to which operators.
The rules that define order of evaluation for elements of a function call are as follows:
Inside-out: Evaluate every argument before beginning to evaluate the function.

211

212

CHAPTER 8. EXPRESSIONS AND EVALUATION

Outside-in: Start evaluating the function body. When (and if) a parameter is
used in the body, evaluate the corresponding argument. There are two variants,
as follows:
1. Call-by-name: An argument expression is evaluated each time the corresponding parameter is used. This is inefficient and may result in different values for
the argument at different times.
2. Call-by-need: An argument expression is evaluated the first time its parameter
is used, and the result is stored for future use.
Inside-out evaluation has been used for most block-structured languages; outside-in
has recently been efficiently implemented, in the form of call-by-need, for functional
languages. Some languages, however, use a mixture of strategies. Call-by-value-andreturn is used in order to facilitate concurrent, multitasked operations.

8.1

The Programming Environment

Each program module is executed in some environment. In these days that environment is rarely
the bare machine; almost all machines run under an operating system (OS). The OS forms an
interface between the hardware, the user, and the users program, and it creates the environment
in which the users program runs.
Compiled Language Systems
When working with a compiler, a programmer either works in the environment provided by the OS
or uses a development shell, often provided with the translator, to tailor the OS environment to
current needs. An editor, linker, file system, and the compiler itself are included in this environment.
Together they enable the programmer to enter, compile, and link/load programs. The programmer
can give OS commands to execute his or her own program or others, and when execution is done,
control returns to the OS or to the shell.
If the OS supports multitasking (making it an M-OS) there may be other programs, or tasks,
in this context, and the M-OS will mediate between them. It supervises messages, controls signals,
and manages shared memory. A task might run indefinitely, putting itself to sleep when it needs
information from another task and waking up in response to an interrupt generated when that
other task supplies the information.

8.2. SEQUENCE CONTROL AND COMMUNICATION

213

Interactive Language Systems

The category of interactive languages includes APL, FORTH, the functional languages, Prolog,
Smalltalk, and many others. These languages are embedded in subsystems that take the place of
the OS in forming the users program development environment. Many of these contain independent
file systems and editors, and a separate interactive control cycle.
Access to the enclosing OS is almost always restricted. For example, just one language is
generally supported, with no possibility of interaction with subprograms written in a different
language. Also, even if the OS supports multitasking, the language subsystem may not give access
to it. It also may not support use of an outside text editor and may restrict file system access,
allowing fewer kinds of operations than the OS.
These language subsystems implement a top-level control structure known as the Read-EvaluateWrite (REW) cycle. When the programmer enters this environment she or he sees a prompt on the
terminal and the interpreter is in control, waiting for programmer input. The programmer enters
expressions which the interpreter reads and evaluates. It then writes the value of the expression, if
any, and a prompt on the users terminal. The user initiates all actual execution.
Three kinds of items are entered to form a complete program: (1) definitions of objects (namevalue pairs), (2) definitions of functions, with parameter names, and (3) expressions or function
calls with actual arguments. Items of type (1) are handled by the interpreter. Names are entered,
with their value bindings, into the program environment.
For items of type (2), the interpreter will call a half-compiler, which reads the expression, lexes
it, parses it, and converts it to an internal form which represents the semantics of the expression in
a form that is convenient for later evaluation. This form might be a computation tree or a linear
list of actions in postfix order. (A full compiler would also generate native machine code.)
For items of type (3), the interpreter evaluates the expression and leaves the result on the stack.
During execution, this expression may call functions that are defined in its program environment.
After returning from the top-level call, the value left on the stack is generally printed out by the
REW cycle for inspection by the programmer.

8.2

Sequence Control and Communication

A program specifies a set of actions that must be done with a set of representations of objects. In
some cases the order of these actions is not important, but generally order matters, particularly
when interaction with the outside environment (user or file system) is involved. The words the
programmer writes in the program to invoke this ordered series of actions must be arranged in a
way that specifies this order, or the programmer might specify that the order is not important. In
any case, the connection between the way the programmer writes things and the order of resulting
outputs must generally be predictable, easy to control, and stable.
Three basic ways have been developed for specifying the order in which computation will happen:
Nesting of expressions, lexically or symbolically.

214

CHAPTER 8. EXPRESSIONS AND EVALUATION

Sequencing of statements and procedure calls.
Signaling to other tasks.

8.2.1

Nesting

Individual symbols are the lowest-level construct in any programming language. These are formed
into expressions (see Section 8.3) which are contained in larger expressions and, in the end, organized
into some larger unit. In some languages these units are statements, written in sequence. In others
they are function definitions, written sequentially or in nested form.
Lexical Nesting by Parenthesization
Each function call is a function name with expressions as arguments, which might themselves be
function calls. When a call does contain another function call, we say that it is a nested expression.
A nested expression can be evaluated using only a stack for intermediate results, and without
use of assignment or nonstack variables. 1 Here we consider in detail how the stack works. Before a
function call can be evaluated, its argument expressions must be evaluated. Evaluation of arguments
thus proceeds inward through the nest of function calls, until there are no more enclosed expressions.
As each argument expression is evaluated, its value is left at the top of the stack. When all
arguments for a function call have been evaluated, the function can be called. The values on the
stack will be bound to parameter names within the function. The function will run, compute an
answer, and leave it on the stack for later use by an enclosing expression.
Parameter binding enables communication of information in and out through the hierarchical
nest of function calls. During execution, the position on the stack where an expressions results will
be stored corresponds directly to the position in which the programmer wrote the subexpression in
the program, and to the position in the program of the expression that will use the results.
Symbolic Nesting by Declaration
A function call with function calls in its parameter list, which, in turn, have embedded function
calls can become quite incomprehensible. This is so even if indentation and spacing are used to
clarify the structure of the whole unit.
Declarations are used (among other reasons) to reduce the apparent nesting depth of expressions. A local variable declaration permits the programmer to define a name for the result of a
subexpression. A subroutine declaration permits one to isolate and name some part of the program
code. These defined names can then be used to build routines with a lower apparent degree of
nesting, which are easier for a human to cope with than a deeply nested program. Modularizing a
program by defining local names and short functions helps a programmer create a correct program
more easily.
1
All the familiar older languages use a stack. Some of the modern functional languages use a more complex data
structure that avoids evaluating arguments that are not actually used in the function code.

8.2. SEQUENCE CONTROL AND COMMUNICATION

215

Programs in functional languages tend to be highly modular because of this. Since parameter
binding, not assignment, is used to give names to computed values, a new function must be defined
whenever the programmer wishes to name an intermediate result. Further, when a nested expression has many nesting levels, humans tend to become confused about the meaning of the code.
Subroutines are used to minimize the depth of nesting in a single program unit.

8.2.2

Sequences of Statements

A statement is an expression that leaves no return value on the stack and thus cannot be used as the
argument to a function. Statements are executed for their side effects, not for their return values.
Side effects include changing the value in a storage object and changing the status of anything in
the program environment, the shared environment, or the streams. To be useful, a statement must
use one of these channels to affect the world, since it does not change the program stack. Either it
must alter program memory or produce input/output activity.
A procedural language is one in which a program is a sequence of statements. Procedural
languages commonly contain statement forms to do the following tasks:
Assignment statement: Changes the value of a variable.
Binding statement: Enters a named value into the symbol table.
Procedure call: A call on a user-defined function that returns no result.
Input statement: Changes the values of some variables.
Output statement: Changes the state of the external world.
Control statement: A compound of other statements which implements a loop or conditional structure.
Procedural and functional languages both contain expressions. Within an expression, one function communicates its result directly to the enclosing function without using the program environment. This is done by taking arguments from the program stack and leaving results on it.
In functional languages, the expression is the outermost control structure, and the stack is
the only form of communication between the parts of a program. Modern functional languages
do not support data objects in the program environment. In contrast, in procedural languages
(ALGOL, FORTRAN, Pascal, APL, etc.) all expressions eventually become part of some statement,
and a program body is a series of statements, written top to bottom in the order they are to be
executed. Communication between one statement and another can go through variables in the
program environment; each statement finds its inputs in a set of variables and leaves its results in
variables for possible later use [Exhibit 4.9].

216

8.2.3

CHAPTER 8. EXPRESSIONS AND EVALUATION

Interprocess Sequence Control

When all computers had small memories and single, slow processors, it was adequate to design
languages for writing independent programs. Such a program would follow one control path during
execution, which was independent of all other programs. Its memory was its own, and it could
interact with the world only through input and output. But as multiprocessor systems and networks
have become common, we have had to expand our area of consideration from a single, isolated
program (or process or task) to a group of asynchronous processes that are actively interacting and
communicating.
A single process, no matter how large, can communicate through global variables and parameters. If that process were to be compiled in several modules, each module could have external
symbols (variables and functions) that would be defined in other modules. In this case, the linking
loader would connect each external reference from one module with the definition of that object
or function in another module, creating a single, connected unit that could communicate through
variables and subroutine calls.
Communication for two separate interacting processes is quite different. No compiler or linking
loader links up external references in one process to symbols defined in the other. In a modern
system, one process may not know the location of another process with which it is interacting.
All communication must take place through the operating system or systems which host the two
processes, or through the file system. In terms of our formal model, communication must be through
the shared environment or through streams, and both of these areas of storage are managed by the
OS. The specific kind of communication possible depends on the operating system host.
One common mechanism is message passing. A message is a string of bytes that is written
into an OS buffer and either broadcast to all processes or addressed to one particular process
whose process-ID is known to the sender. To send a message, a program calls the operating
systems message sending routine. To receive a message, a process makes a corresponding call to
the operating system, signaling that it is ready to process a message either from anyone or from a
particular sender. The message is then copied into a buffer belonging to the receiver.
Another common communication mechanism is the semaphore. This is a storage location in
protected OS memory, accessible to a program only through system calls. Semaphores come in
many varieties, but they basically function as pigeonholes that contain one of two signals meaning
either wait until Im ready or go ahead. These are commonly used to control and synchronize
access to files and buffers that are shared among interacting processes.

8.3

Expression Syntax

An expression is a nest of function calls (or operators, in some languages) that returns a value. The
syntax for writing an expression varies among languages. Exhibit 8.1 shows examples of function
calls drawn from several languages which illustrate the various kinds of syntax in common use. In
all cases, the action represented is adding one to an integer, A, and returning the result.

217

8.3. EXPRESSION SYNTAX

Exhibit 8.1. Syntax for expressions.
Lambda calculus
An application

( S A )

Compute the successor of A.

An expression
An expression

( + A 1 )
( 1+ A )

Add the value of A and 1.

Add 1 to the value of A.

Pascal
A function call
Use of an operator

succ( A )
A + 1

Compute the successor of A.

Add the value of A and 1.

LISP

APL
An expression
A + 1
Add the value of A and 1.
(APL has no other syntax for function calls.)

8.3.1

Functional Expression Syntax

In the most basic syntax, one writes an expression by writing a function name, with its actual
arguments, enclosed in parentheses. In lambda calculus and in LISP the function name is inside
the left paren [Exhibit 8.1]; in most languages it is outside the left paren. We will call the first
variant lambda calculus function-call syntax and the latter variant normal function-call syntax.
In modern languages, an actual argument is an expression. It is evaluated, in the context of
the calling program, either before control is transferred to the function or during evaluation of
the function. The value of each argument expression is used to initialize the corresponding formal
parameter.
There are two ways that the correspondence between arguments and parameters can be specified: positionally or by keyword. With positional notation, which is supported in virtually every
language, the first argument value initializes the first parameter name, the second argument corresponds to the second name, and so on. To call a function positionally one must know the number
and order of the parameters, and know the type and semantics of each, but not its declared name.
A few languages support a second parameter correspondence mechanism, sometimes called
correspondence-by-keyword, or named parameter association. To use named association, the argument expression in the function call is preceded by the dummy parameter name as defined in
the procedure definition. This mechanism has the advantage that the arguments in a call may be
written in any order, and it is useful for implementing functions with optional arguments (Section
9.1.1). It is syntactically more complex, though, than positional correspondence, and more complex to implement. Further, the parameter names, which are otherwise arbitrary and local to the
function, must be made known to all users.
Some confusion of terminology has developed among different programming languages. In LISP
+ is called a function and is written with lambda calculus function-call syntax. In Pascal it

218

CHAPTER 8. EXPRESSIONS AND EVALUATION

is called an operator and written with infix operator syntax; that is, each operator is written
between its operands. In APL + is called a function but written in infix operator syntax.
Operators are just a syntactic variant of functions; the underlying semantics are the same.
Both denote an action which, when applied to some objects (the arguments), will produce another
object (the function result). Some languages include only the function call syntax; some include
only the operator syntax; some include both. There are two minor differences between operators
and functions. First, functions may have any number of arguments, while operators are almost
universally limited to one or two. Second, in languages that include both functions and operators,
it is generally not possible to define new operators; all newly defined verbs are functions.

8.3.2

Operator Expressions

An operator with two arguments is called binary or dyadic. It is generally called using infix operator
syntax; that is, the operator is written between its arguments. A single-argument operator, known
as unary or monadic, is generally called using prefix syntax ; that is, the operator is written before
its argument. Some languages also have monadic postfix operators, which are written after the
operand. C has a diverse collection and even has one two-part operator with three operands!
Any expression specifies a computation tree, which is the meaning of the source expression. The
task of a compiler (specifically, the parsing phase of a compiler) is to take the source syntax of an
expression and produce the computation tree, or parse tree, that it denotes.
For normal function-call syntax, producing a parse tree is an easy task. Each function call
corresponds to a node in the tree, and each argument corresponds to a child of that node. If some
argument position contains a nested function call, then that child is itself a node with children
[Exhibit 8.2].
Developing a parse tree from an expression that includes a combination of infix, prefix, and
postfix operators is much more complex. Two sets of rules (the rules for precedence and the
rules for associativity), plus explicit parenthesization, determine which operand belongs to which
operator.
Parenthesization
Languages that have infix operators permit the programmer to use explicit parentheses to surround
an operator and its one or two operands. Like an expression written using nested function calls, a
fully parenthesized infix expression explicitly and unambiguously specifies which operands belong
to each operator. Exhibit 8.2 shows the parse trees derived from a LISP expression that is a nest of
function calls, and the analogous Pascal expression, written with operators. The trees are identical
except for the order of the operator and the first operand.
Postfix and Prefix: Unambiguous Unparenthesized Expression Syntax
The parentheses in LISP, or in a fully parenthesized infix expression, group each function or operator
with its operands. This permits one to draw a parse tree with no further information and no

219

8.3. EXPRESSION SYNTAX

Exhibit 8.2. Nested function calls and parenthesized infix expressions.

An expression written in LISP, using functional notation, and in Pascal, using fully parenthesized
infix notation. Both denote the same computation tree.

(* a (+ (- b c) a))

(a * ((b - c) + a))

LISP: a nested expression.

Pascal: fully parenthesized infix notation.

knowledge of the meaning of the operator. In a language where each operator has a fixed number
of operands, these parentheses are nice but not necessary. They guide the eye and the parser, and
permit us to check whether each operator has been written with the correct number of operands.
There are other ways, however, to indicate which operands belong to which operator.
We can actually eliminate the parentheses if we are willing to restrict each operator to exactly
one meaning and require that its number of operands be fixed. (This eliminates the possibility of
an ambiguous operator, such as in FORTRAN, which exists in both unary and binary forms.)
With this restriction, there are two forms for expressions, called prefix order and postfix order,
that unambiguously specify the computation tree without the use of parentheses or precedence. In
a prefix expression, each operator is written before its operands, like a LISP expression without the
parentheses [Exhibit 8.3].
In postfix order,2 an operator is written after its operands. The FORTH language and Postscript 3
use postfix notation exclusively. This order corresponds to the actual order of machine instructions
needed to compute the value of an expression using a stack machine. The evaluation process is
illustrated in Exhibit 8.4, and works like this:
2
3

Also known as reverse Polish notation.

A language for controlling graphic output devices. Adobe Systems [1985].

Exhibit 8.3. Unambiguous ways to write an arithmetic expression.

This is how the expression (a * ((b - c) + a)) would be written in prefix and postfix notations.
Prefix :
* a + - b c a This is like LISP, but without parentheses. The scope of each
operator is the two operands or subexpressions following it.
Postfix : a b c - a + * This is used in FORTH. The scope of each operator is the two
operands or subexpressions preceding it.

220

CHAPTER 8. EXPRESSIONS AND EVALUATION

Exhibit 8.4. Evaluation of a postfix expression using a stack.

We evaluate the expression a b c - a + *
from Exhibit 8.3. Assume the stack is initially
empty and the names a, b, and c are bound to constant values: a = 5, b = 4, c = 7.
Evaluation Step Stack Contents (top is at right)
a
5
b
5, 4
c
5, 4, 7
5, -3
a
5, -3, 5
+
5, 2
10
*

Evaluate the expression left-to-right.

As each symbol is encountered, put its current value on the stack.
For each operator, take two values off the stack and execute the operation. Put the result
back on the stack.
One might ask why LISP uses both prefix order and parentheses if prefix order is unambiguous
without the parentheses. The parentheses in LISP permit its functions and operators to take
a varying number of operands [Exhibit 8.5, line 6]. Several primitive LISP functions can have
variable-length argument lists, including one at the core of the language. The LISP conditional
expression, COND, is a variable-argument function.
Note the similarity between form 4, prefix notation, and form 5, written in LISP. If functions
in LISP were restricted to a fixed number of parameters, the parentheses would not be necessary.
But LISP has variable-argument functions, including +. In form 6, the expression is rewritten using
the variable-argument form of +, and parentheses are used to delimit the scope of the +. It can be
cogently argued that this is the clearest way to write this expression.
Form 6, in LISP, strongly resembles form 7, in COBOL, which adds a variable number of items
to the last item. Reserved words are used in COBOL instead of parentheses to bracket the variablelength parameter list, but the principle is the same; it is possible to parse a language that has
variable-length parameter lists as long as the lists are bracketed somehow.
Precedence
The early computer languages FORTRAN and ALGOL-60 were designed for use by scientists and
engineers who were accustomed to traditional mathematical notation for expressions. 4 Engineers
wanted a natural way to write formulas in their programs; that is, they wanted the formulas
4

This is evident from their names. FORTRAN was derived from FORmula TRANslator, and ALGOL from ALGOrithmic Language.

221

8.3. EXPRESSION SYNTAX

Exhibit 8.5. Bracketing used to permit variable-argument expressions.

Seven forms are given of an expression which sums four items. In the first four forms the + operator
must be written three times, because it is built into these languages that the + operator takes
two operands.
Forms 5 and 6 show an analogous expression written in LISP. Form 5 has two operands for each
+, but version 6 takes advantage of the fact that + can accept a variable number of arguments.
Form 7 is a statement, not a function: it modifies the value of C rather than returning a
value. Note that it uses the reserved words ADD and TO, rather than parentheses, to bracket the
variable-length list.
1.
2.
3.
4.
5.
6.
7.

Fully parenthesized infix:

Infix without parentheses:
Postfix notation:
Prefix notation:
Use of + in binary form:
Multiple arguments for +:
A statement that modifies C:

(( (a + b) + 3) + c)
a + b + 3 + c
a b + 3 + c +
+ + + a b 3 c
(+ (+ (+ a b) 3) c )
(+ a b 3 c)
ADD A B 3 TO C.

Pascal or C
Pascal or C
FORTH
LISP
LISP
COBOL

in their computer programs to look as much as possible like formulas on paper. Engineers and
mathematicians have never considered full parenthesization, prefix order, or postfix order to be
natural ways to express a formula. The familiar way of writing expressions using operator
precedence was used in FORTRAN and ALGOL-60 because it had been in use in mathematics for
many years. Since then precedence expressions have been implemented in all their descendents.
Without parentheses or additional parsing rules, infix notation is ambiguous. The symbols
written on the page do not indicate which operands belong with each operator. Consider the
expression in Exhibit 8.6, where op1 and op2 are two unspecified binary operators. Two parses of
this expression are possible and correspond to the two trees shown. The rule of operator precedence
was adopted from mathematics to disambiguate the nonparenthesized parts of infix expressions.
The precedence rule determines which parse is correct, and thus whether the scope of op2 includes
the result of op1 or vice versa.
In a precedence language, each operator is assigned a precedence level. These levels can be
arbitrary integers; only their order matters. The standard precedence rule can now be stated:
If precop1 > precop2 then the meaning is parse tree I.
Else if precop1 < precop2 then the meaning is parse tree II.
Else precop1 = precop2 . Use the rule of associativity to determine which parse is correct.

222

CHAPTER 8. EXPRESSIONS AND EVALUATION

Exhibit 8.6. Infix notation is ambiguous without precedence.

Source expression: A op1 B op2 C
Possible parse trees:
(I)

(II)

op1

op2

op1

op2

Associativity
Associativity is used as a tie-breaking rule in languages with precedence. It is also used in APL as
the only added parsing rule to disambiguate infix expressions. Associativity governs the choice of
parse trees when there are consecutive binary operators with equal precedence. Associated with
each precedence class is an associativity direction, either left-to-right or right-to-left. All operators
with the same precedence must have the same associativity. This direction is used as follows:
If the associativity of op1 and op2 is left-to-right then the meaning of the expression is parse
tree I.
Else (the associativity of op1 and op2 is right-to-left) the meaning of the expression is parse
tree II.
The associativity of each operator is determined by the language designer and is usually chosen
to seem natural to mathematicians. For example, in C, the associativity of - (subtraction) is leftto-right, but the associativity of = (assignment) is right-to-left. The programmer can write: X= Y=
Z-1-X to compute the value ((Z-1)-X) and store the answer in both Y and X.
In APL there were so many operators that the language designer evidently felt that establishing precedence classes for them would cause more confusion than help. Therefore, although
APL functions are written with infix notation, only an associativity rule, right-to-left, is used for
disambiguation.

8.3.3

Combinations of Parsing Rules

Existing languages use varying combinations of rules to define the meaning of an expression. Very
different effects can be achieved by combining these few simple rules, as illustrated by Exhibits
8.7, 8.8, and 8.9. Exhibit 8.7 gives brief summaries of the parsing rules in some of the languages
previously mentioned. Exhibit 8.8 shows how the arithmetic expression ((b c)/((a + 1) (b 2)))
would be written in each language using as few parentheses as possible. Exhibit 8.9 shows the

8.3. EXPRESSION SYNTAX

223

Exhibit 8.7. Varying rules for parsing expressions.

Summaries of the parsing rules are given below for several languages.
FORTH: postfix notation, no precedence, no associativity, no parentheses. Operators are
executed left-to-right. Parameter lists are not delimited at all.
APL: infix notation for all functions, no precedence, right-to-left associativity. Parentheses
may be used to override associativity.
Pascal: infix notation for operators, with precedence and left-to-right associativity. Parentheses may be used to override precedence or associativity.
C: infix, prefix, and postfix operators, with precedence. Left-to-right or right-to-left associativity, depending on the operator. Parentheses may be used to override precedence or
associativity.

Exhibit 8.8. Varying syntax for the use of operators.

Each expression below illustrates the rules summarized in Exhibit 8.7 and will translate into
the computation tree in Exhibit 8.9. A, B, and C are integer variables.
FORTH:

B @ C @ * A @ 1 + B @ 2 - * /

Note: A variable name in FORTH is interpreted to mean the address of the variable. The
symbol @ must be used explicitly to fetch the value from that address.
APL:
Pascal:

(B C) (A + 1) B 2
B * C / ((A + 1) * (B - 2))

Relevant precedence values, with highest precedence listed first: ( *, / ) > ( +, - ).

224

CHAPTER 8. EXPRESSIONS AND EVALUATION

Exhibit 8.9. A computation tree.

This tree represents the computation ((b c)/((a + 1) (b 2))). Each leaf represents the value
of an integer variable, and each interior node represents an operation to be applied to the values
returned by its children. The expressions in Exhibit 8.8 will all translate into this computation
tree.
/
*
B

*
C
A

+
1 B

computation tree into which all of these expressions will be translated. Note the similarities and
contrasts in these languages, especially the contrast in placement of parentheses.
The language syntax, parentheses, precedence, and associativity all help the translator to parse
a program and produce a parse tree. After the parse tree is produced, it is interpreted, or code is
generated that will later be executed to evaluate the parse tree. It is perhaps surprising, but true,
that the same parse tree can have very different meanings in different languages. There are still
semantic issues to be resolved after parsing that determine the meaning of the expression.

8.4

Function Evaluation

A function call (like a lambda calculus application) consists of a literal function or a function name,
with formulas representing actual arguments, and an indication that the function is to be applied to
those arguments. In many languages, the actual arguments may contain function calls, producing
a hierarchy, or nest, of calls. The order and method of evaluation of such a nest depends on the
programming language and sometimes also on the compiler writer.

8.4.1

Order of Evaluation

Two very different evaluation rules are in use which govern the order of evaluation of the elements
of a function call:
1. Inside-out evaluation: Evaluate every actual argument before beginning to evaluate the function itself. In a given language, arguments might be evaluated left-to-right, right-to-left, or
in an order determined by the compiler writer.

8.4. FUNCTION EVALUATION

225

Exhibit 8.10. The substitution rule.

The substitution rule of lambda calculus defines the semantics of arguments. It is stated here
using the familiar terms function, argument, and parameter instead of the usual lambda
calculus terminology.
Let P be a parameter for function F .
Let Exp, an expression, be the corresponding actual argument.
Declarations in F must not capture unbound variables in Exp. If some variable, V , defined
in F has the same name as an unbound symbol in P , rename it with a unique name, say V 2,
and replace all V s in F by V 2.
Now the meaning of F (Exp) is found by replacing every occurrence of P in F by Exp.

2. Outside-in evaluation: Start evaluating the function call first. The first time the value of an
argument is needed, evaluate that argument and remember the answer in case that argument
is used again.
Rule (1) has been used for LISP, Pascal, ALGOL-60, C, and most other languages designed since
block structure was developed for ALGOL-60. Implementations of these languages are based on the
use of a stack to store arguments and local variables.
Recently, efficient implementation methods for rule (2) have been developed, leading to a new
class of languages called functional languages (Miranda, Haskell, etc.). Let us examine the relationship between evaluation order, stacks, and block structure.
The substitution rule of lambda calculus says that if an expression, E, contains a nonprimitive
symbol, S, the meaning of the E is the same as if each occurrence of the symbol S were replaced
by its definition. This rule applies in the following two situations:
S is a symbol whose meaning is defined by the user in a declaration.
S is a parameter of E, and S is given a meaning by applying E to a list of arguments. (The
meaning of S is the expression that forms the argument corresponding to S.)
This seems to be an obvious definition and clearly something that should be part of the implementation of any programming language. But implementing it directly causes some trouble for
two reasons. Let Exp be the meaning of S. Exp may contain unbound symbols, that is, symbols
whose meaning is not defined within Exp. We say there is a name conflict if one of these unbound
symbols also occurs in E. If we simply substitute Exp for S in E, these unbound symbols will be
captured and given spurious meanings by bindings declared in E. Thus symbols involved in name
conflicts must be renamed. The substitution rule is stated formally in Exhibit 8.10.

226

CHAPTER 8. EXPRESSIONS AND EVALUATION

The need for renaming complicates the substitution rule. A second problem is presented by
recursion. A recursive definition would lead to an infinite number of substitutions if you actually
tried to physically substitute the definition of S into the expression E.
At this point we can take a shortcut. To the extent that every copy of S is identical, a single
copy is able to emulate multiple copies. Thus we only need to copy the variable parts of S. We
are thereby led directly to making a distinction between pure code, which never varies, and
data, which does. Multiple copies of pure code may be implemented efficiently by making
one copy and accessing it with a jump to subroutine instruction. On the other hand, a correct
implementation of the substitution rule implies that multiple copies of data and argument values
must exist simultaneously.
The oldest computer languages, FORTRAN and COBOL, use only static storage, with one storage
location for each variable or parameter. For this reason neither language can support recursion.
When ALGOL-60 was first designed no one knew how to implement it efficiently, because an efficient
implementation for multiple copies of local variables had not been invented.
Within a few years, however, a significant insight was gained. No matter how many copies of
the variables for a function S exist, the real computer has only one processor and, therefore, can
be actively using only one copy at a time. All that is necessary to implement the substitution
rule efficiently is to make one copy (the one needed at the moment) of Ss variables conveniently
available and bind them to the local names defined in S. When a copy of S is fully evaluated we no
longer need its corresponding variables. They can be forgotten, and the names in S can be bound
to some other copy. All is well so long as the correct copy of the variables always gets bound to
the names.
When function evaluation rule (1) is used, the groups of variables are allocated, used, and
deallocated in a last-in-first-out manner. The only set of variables needed at the moment are
those for the current function. They can be created at the beginning of function execution and
forgotten at the end. The stack of our abstract machine is able to perform all the required allocation,
binding, and deallocation operations efficiently. But to use a stack like this we must always evaluate
the arguments to a function before evaluating the function. At first sight this restriction seems
to prohibit nothing of importance, and the evaluation order it requires is intuitively appealing.
Actually, it is an important and fundamental restriction, as we shall see.
The evaluation rule of lambda calculus specifies that the order of evaluation of parts of an
expression is undefined, but that parameters must be evaluated if and when they are used (if not
earlier). It can be shown that anything that can be computed using this unconstrained evaluation
order can also be computed using outside-in order, but some computable things cannot be computed
using inside-out order.
The outside-in evaluation rule is: evaluate as much of an expression as possible before using the
substitution rule, then replace a symbolic parameter by its definition. This strategy has one very
nice property: a parameter that is never used does not ever have to be evaluated. This can happen
when the only references to that parameter are in a part of the program that is skipped over by a
conditional.

8.4. FUNCTION EVALUATION

227

Outside-in evaluation also has one awkward property: the simple stack implementation for
parameters wont work. A different parameter evaluation method named call-by-need has been
developed which produces an efficient implementation of outside-in evaluation. It avoids both
the unnecessary evaluation of unused parameters that is inherent in call-by-value, and the repeated
evaluation of the same parameter inherent in call-by-name. It is often referred to as lazy evaluation,
because all unnecessary parameter evaluation is avoided.
In call-by-need evaluation, every function is evaluated until the point that it refers to one of
its parameters, P . That parameter is then evaluated, and its value is saved. Evaluation then
proceeds until the next parameter reference. All future references to P in the same function will
be interpreted to have this saved value.
A programmer can exploit the lazy nature of this evaluation order by using, as arguments,
expressions that are erroneous or nonterminating under some conditions. The programmer must
then use these parameters in guarded ways, checking for the dangerous conditions and being sure
not to evaluate the argument if they occur. Lazy evaluation can also be used to build a useful kind
of data structure called an infinite list. These lists have a head section that is like an ordinary list,
and a tail that is a function that can be evaluated to produce the next item on the list. Evaluating
it N times extends the list by N elements.
An if..then..else conditional is actually a function of three arguments, but a very special
kind of function. The intent and meaning of a conditional is that either the second argument or
the third, but never both, is evaluated. Thus, by its very nature, a conditional, must be evaluated
outside-in. All languages use a variant of lazy evaluation for conditional statements and expressions.
Thus a programmer can emulate lazy evaluation in any language by using conditionals liberally.
Exhibit 12.14 shows the use of if in Pascal to emulate the lazy evaluation built into Miranda.

8.4.2

Lazy or Strict Evaluation

The most important issue concerning order of evaluation is whether an expression will be evaluated outside-in or inside-out. The modern functional languages are the only ones that apply either
rule consistently; they use outside-in evaluation order. Other languages use a mixture of strategies, inside-out for most things, but outside-in for conditionals and sometimes also for Boolean
expressions.
Ordinary inside-out evaluation is called strict evaluation. An evaluator is strict if it evaluates
all the arguments to a function before beginning to evaluate a function. Its opposite is lazy evaluation. Evaluation is lazy if subexpressions are evaluated only when necessary; if the outcome of
an expression does not depend on evaluation of a subexpression, the subexpression is skipped. In
modern functional languages, all function calls are interpreted using call-by-need, a kind of lazy
evaluation. Each argument expression is evaluated the first time the function body refers to the
corresponding parameter. The resulting value is bound to the parameter name and is available for
future use. If an argument is not used, it will not be evaluated.
In general, to evaluate an arithmetic expression one must first evaluate all its subexpressions.
(For example, to calculate a + b one must first know the values of a and b.) But Boolean expres-

228

CHAPTER 8. EXPRESSIONS AND EVALUATION

Exhibit 8.11. Lazy evaluation of AND.

Value of L
T
T
F
F

Value of R
T
F
T
F

Value of L AND R
T
F
F
F

We see that if L is FALSE, the value of the expression is always FALSE, and R does not need
to be evaluated.

sions are different. Exhibits 8.11 and 8.12 show that it is sometimes possible to know the outcome
of a logical AND or OR operation after evaluating only one of its subexpressions. Thus evaluation
of the second subexpression can be skipped and execution time can be saved. Some languages,
for example, MAD and C, use lazy evaluation for Boolean expressions (also called short circuit
evaluation) because it is more efficient.
Boolean expressions in many other languages, for example, Pascal, are evaluated inside-out.
Thus parts of a program that could, under some circumstances, cause run-time errors must be enclosed inside a control statement (an IF or a WHILE) which checks the error condition and determines
whether or not to execute the code.
For example, in Exhibit 8.13, an extra control statement must be used to check for a subscriptout-of-bounds error before executing the subscripted expression. There are two possible reasons for
leaving any search loop: either the key item was found, or the array to be searched was exhausted.
We would like to test both conditions in the following WHILE statement:
WHILE (scan < 101) and (a[scan] <> key) ...
But this would cause the program to bomb when scan exceeded 100, and a[101] was tested.
This happens because Pascal uses inside-out evaluation order, and both comparisons are made before

Exhibit 8.12. Lazy evaluation of OR.

Value of L
T
T
F
F

Value of R
T
F
T
F

Value of L OR R
T
T
T
F

We see that if L is TRUE, the value of the expression is always TRUE, and R does not need to
be evaluated.

229

8.4. FUNCTION EVALUATION

Exhibit 8.13. Extra code needed with inside-out evaluation.
TYPE array_type = array[0..100] of char;
FUNCTION search (a:array_type, key:char):integer;
VAR done_flag: boolean;
scan: -1 .. 101;

(* An integer in the range -1 to 101.

BEGIN
done := FALSE;
(* Flag to control loop exit.
scan := 0;
WHILE ( scan < 101) and not done DO
IF a[scan] = key
THEN done := TRUE
(* Leave loop next time. *)
ELSE scan := scan + 1;
IF scan = 101 THEN search := -1 ELSE search := scan;
END

*)

*)

performing the WHILE test.

To code this function in Pascal, the two tests must be written separately. An extra boolean
variable is then introduced to effect exit from the loop when the second condition becomes true.
We see that the Pascal version requires an extra control statement and makes exit from a simple
loop awkward.
Compare this to the parallel form in Exhibit 8.14 written in C. C uses lazy evaluation for Boolean
expressions, with arguments evaluated in a left-to-right order. Further evaluation is aborted as soon
as the outcome of the expression is determined. The expression that controls the while statement
can, therefore, encompass both the subscript-out-of-range test and the search test. The error
condition is tested first. If the subscript is too large, the rest of the expression (including the part
that would malfunction) is not evaluated. This greatly simplifies construction of the loop.

Exhibit 8.14. Simpler control structures with outside-in evaluation.

int search (char a[], char key);
{
int scan;
scan = 0;
while ( scan < 101 && a[scan] != key ) ++scan;
if (scan == 101) return -1; else return scan;
}

230

CHAPTER 8. EXPRESSIONS AND EVALUATION

Exhibit 8.15. Expressions whose values are indeterminate in C.

int x, z, a[10];
int throwaway (int b, int c) { return c };
x = 3;
x = 3;
x = 3;

8.4.3

z = x * ++x;
z = throwaway(++x, x);
a[x] = x++;

/* z could be 9 or 12. */
/* z could be 3 or 4. */
/* 3 is stored in either a[3] or a[4].

*/

Order of Evaluation of Arguments

One final design decision is whether to specify that arguments are evaluated left-to-right or rightto-left, or leave the order unspecified. This makes no difference when inside-out evaluation is used
and the subexpressions have no side effects. But when expressions can produce output or modify
memory, the order of evaluation can determine the outcome of the program.
In the modern functional languages, this order is unspecified. These languages do not implement
destructive assignment, and, therefore, there is no problem with side effects except where output
is concerned. For these situations, the languages provide a way, called strict evaluation, to specify
an ordered evaluation of all subexpressions.
In languages such as Pascal where functions and operators cannot have side effects, there is
also no problem. All side effects (assignment, input, output) are restricted to statements, whose
order is clearly specified. It is impossible to tell whether a Pascal compiler evaluates expressions
left-to-right or right-to-left.
Finally, there are languages such as C and APL where expressions can have side effects. Assignment, input, and output are all expressions in these languages. In addition, C has an increment
operator. The language definition in these cases must clearly specify what the evaluation order is.
APL does specify right-to-left.
In C, though, the decision was left to the language implementor, and both right-to-left and
left-to-right evaluation are permitted and considered to conform to the standard. It is, therefore,
incumbent upon a C programmer to avoid using any variable in an expression whose value is changed
by a preceding or following subexpression with a side effect [Exhibit 8.15]. The result of such an
expression is called indeterminate; that is, it may vary from compiler to compiler, even on the same
hardware.

Exercises
1. What is the programming environment? How is it provided?
2. What is the Read-Evaluate-Write cycle? How is it implemented?
3. What are the three basic methods that specify order of computation?

8.4. FUNCTION EVALUATION

231

4. Explain how an expression can be nested.

5. How is a stack used in evaluating a nested expression?
6. What is an expression? How does the role of expressions differ in procedural and functional
languages?
7. What is message passing, and how is it accomplished between two separate interacting
processes?
8. Parameters allow program modules to communicate with each other. Name at least three
mechanisms or data structures that must be a part of the semantic basis of a language to
implement parameter passing for recursive subroutines.
9. Is there a difference between operators and functions? Explain.
10. Define infix, prefix, and postfix operator syntax.
11. What is the role of parentheses in an expression?
12. How do prefix and postfix order for expressions eliminate the need for parentheses?
13. Using a stack, evaluate the following postfix expression. Assume the stack is initially empty
and that x=2, y=3, and z=4. Show the contents of the stack after each push or pop.
x y + z - x z * 14. Draw a computation tree for the following infix expression:
((x/y) * (((3*z) + (2 + y)) +7))
15. Rewrite the expression of question 14 with the minimum number of parentheses, using the
precedence rules in C or Pascal.
16. Draw a computation tree for the following prefix expression:
* / x y + 3 * * z 2 + y 7
17. Explain the two evaluation rules which govern order of evaluation of elements of a function
call.
18. What is the difference between call-by-need and call-by-name?
19. What is automatic dereferencing? Give a specific example.
20. In FORTRAN, call-by-reference is the only mechanism for passing parameters. Why is this
now considered poor design?

232

CHAPTER 8. EXPRESSIONS AND EVALUATION

21. When should call-by-reference be used instead of call-by-value? Give two distinct situations.
22. In APL, parameter names are declared but their types are not specified. What is the purpose of the parameter declarations? How can communication work without parameter type
specifications?

Chapter 9

Functions and Parameters

Overview
A function call consists of a name and actual arguments. These correspond to the
declared function name and its formal parameters. Functions may be defined with a
fixed or variable number of arguments, depending upon the language. Missing or extra
parameters are handled by the translator of the particular language.
Arguments are passed to a function by the function call. They are usually matched
up with parameter names positionally; the nth argument in the function call becomes
the value of the nth parameter name in the function definition. Correspondence-bykeyword is also used. In this method, the dummy parameter name is written next to
the argument expression in the function call. This parameter passing mechanism is very
useful for functions that expect variable-length argument lists.
After the argument is passed, it is interpreted by the receiving function. In ALGOL-60,
parameter interpretation mechanisms were limited to call-by-name and call-by-value.
Call-by-reference was used in early FORTRAN. Call-by-need and call-by-pointer have
since been devised as methods for passing and interpreting parameters.
A higher-order function is one that takes a function as an argument and/or returns a
function as a result. Pure functional languages fully support higher-order functions.
Flexible code and mapping are two common applications of functional arguments.
Currying is a way of looking at a function of two or more arguments, so that it is
considered to be a higher-order function of one argument that returns a function.
Closure is an operation that binds the free variables of an expression. It creates and returns a function. The most common application of closure is partial parameterization.
233

234

9.1
9.1.1

CHAPTER 9. FUNCTIONS AND PARAMETERS

Function Syntax
Fixed versus Variable Argument Functions

Many languages require that a function be defined with a fixed number of parameters of specified
types. The required parameters and types for predefined functions are specified in the syntax of
the language. The elements required for a control statement are similarly defined by the language
syntax.
The parser requires every statement and function call to be well-formed: it counts the arguments, checks their types, and binds them to the defined dummy parameter names. If either the
number or the type of arguments is wrong, the program is not well-formed and the translator gives
an error comment. If some element is misplaced or missing, the translator cannot, in general, know
the intended meaning of the program. For example, if an END statement is missing, it is impossible
in many languages for a translator to guess, accurately, where it was supposed to be.
Because the parser rejects programs that are not well-formed, being well-formed can be seen
as a syntactic requirement. But the syntax merely reflects semantic requirements: the code that
defines the semantics of these statements and functions is not prepared to handle a variable number
or type of argument.
A function uses a certain number of parameters; that is, it refers to their names. There are
several ways in which there might be a mismatch between the number of arguments passed and
the number of parameters that are actually used, as follows:
Unused arguments: Parameters might be named and passed but not referred to anywhere in
the function. In a procedural language, an unused argument might as well be omitted.
In functional languages, however, there are situations in which this is useful. (Consider
the lambda calculus functions T and F , defined in Section 4.3.4, which each discard one
parameter.)
Optional parameters: A parameter might be named and used, but no corresponding argument
passed. This can be meaningful if default values are defined for missing arguments, and the
argument-parameter correspondence is unambiguously specified.
Indefinite-length argument lists: Parameters might be passed but not named. These can be
useful if some way is supplied by the language for referring to them.
Optional parameters. Several modern languages support optional parameters. If the number
of actual arguments passed is smaller than the number of parameters called for, some parameters
cannot be given a meaning in the normal way. If this is permitted, the meaning of such an omission
must be defined by default or explicitly within the function. Every parameter must have a meaning

235

9.1. FUNCTION SYNTAX

Exhibit 9.1. An Ada function with a default value for one parameter.
Imagine that this function is a small routine inside a chefs program for figuring out ingredient
quantities when the quantity of a recipe must be changed. The usually case is to double the recipe,
so the program has an optional parameter with a default value of 2.0.
function

CONVERT (
)
begin

quantity:real,
proportion: real := 2.0
return real is
return quantity*proportion

end CONVERT;

Legal and meaningful calls on CONVERT:

CONVERT(3.75)
CONVERT(3.75, .5)

-- Double the amount, result is 7.5.

-- Halve the amount, result is 1.675.

if and when it is actually used in a computation. One solution, used in Ada, is that parameters
may be defined to have default values. If an actual argument is supplied, its value is used for that
parameter, otherwise the default value is used [Exhibit 9.1].

9.1.2

Parameter Correspondence

When calling a function with optional parameters, arguments may (or may not) be omitted. This
complicates the problem of matching up argument values and parameter names, and makes the
simple positional correspondence syntax inadequate. The problem can be handled in three ways:
Permit the user to use adjacent commas to hold the place for an omitted argument.
Require all arguments that are supplied to precede the ones that are omitted.
Use correspondence-by-keyword for all arguments after the first one that is omitted, or for
all optional arguments.
All three ways have been used. For example, command line interpreters commonly use the first
and third conventions, and Ada supports the second and third. The Ada CONVERT function from
Exhibit 9.1 had one optional parameter. Because it was the last parameter, we could use positional
argument correspondence to call the CONVERT function.
If a function has several optional parameters, however, which could be included or omitted
independently, positional correspondence cannot be used. Exhibit 9.2 shows how correspondenceby-keyword can be used to specify the correct parameter values. The rule in Ada is that all positional
arguments in a call must precede all keyword arguments.
Pros and Cons. Whether or not a language should support optional parameters reduces to the
usual question of values. Adding syntax and semantic mechanisms to support correspondence-bykeyword does complicate a language and its compiler. Sometimes the effect of optional parameters

236

CHAPTER 9. FUNCTIONS AND PARAMETERS

Exhibit 9.2. Correspondence-by-keyword in Ada.

Imagine that the function Wash_Cycle is built into the controller of an automated washing
machine. It takes parameters that describe the character of a load of laundry and executes an
appropriate wash cycle. Only the first parameter, the load size, is required; default values are
defined for all others. (We show a possible function header and omit the function body.) Assume
that cloth_type, soil_type, and error_type are previously defined enumerated types.
FUNCTION Wash_Cycle
(
Weight: IN integer;
Fabric: IN clothtype := cottonknit;
Soil_level: IN soiltype := average;
Dark: IN Boolean := False;
Delicate: IN Boolean := False;
) RETURN error_type IS ...
Following are several calls on Wash_Cycle that would process different kinds of loads. Note the
mixed use of positional and keyword correspondence.
Result:= Wash_Cycle(10);
-- A large load of T-shirts.
Result:= Wash_Cycle(4, denim, Dark => True);
-- A small load of jeans.
Result:= Wash_Cycle(8, Soil_level => filthy, );
-- Lots of dirty socks.
Result:= Wash_Cycle(Weight => 2, Delicate => True, Fabric => wool);
-- Two winter sweaters.

can be achieved in some other way, without adding special syntax or semantics. For example, the
functional languages support higher-level functions (functions that can return functions as their
result). In these languages, a closure is the function that results from binding some (or perhaps
all) of the arguments of another function. A function and a set of closures made from that function
are like an Ada function with default values for some parameters.
Using optional parameters has a mixed effect on semantic validity. When a programmer omits
an optional parameter, the code is less explicit and more concise. This may or may not be a wise
trade-off; explicit code is usually easier to debug and easier to modify. Being forced to write out all
the arguments every time prevents a programmer from accidentally omitting one that was needed.
On the other hand, if a function has a dozen parameters that are usually called with the same
values, writing them all out every time hides the variable parts of the program and makes it harder
to comprehend.

237

9.1. FUNCTION SYNTAX

Exhibit 9.3. A FORTH function with a variable number of arguments.

:

SumN
0
swap
0 DO
+
LOOP
;

9.1.3

(
(
(
(
(
(
(
(
(

Sums N numbers. N must be on top of parameter stack, )

with numbers to be summed below it. )
Put an initial sum (zero) on top of the stack. )
Exchange positions of N and initial sum. N now on top.)
Do loop removes 0 and N from stack, and will iterate )
from 0 through N-1. )
+ adds top two items on stack and leaves result there. )
Terminates loop that began with DO. )
End of definition. Answer is on top of stack. )

Indefinite-Length Parameter Lists

Some languages permit a function call to pass more arguments than the number of parameter
names defined in the function. Such indefinite-length parameter lists are useful in functions that
apply some uniform process to a variable number of objects, such as read and write in Pascal.
A Pascal programmer cannot write such a function in Pascal, because all Pascal function calls
must be well-formed. The syntax of Pascal treats read and write as special cases. In contrast,
C and FORTH do not require function calls to be well-formed. A FORTH function definition does
not specify either parameter names or the number of parameters. Parameters are left explicitly
on the parameter stack before calling a function, and the function manipulates the stack directly,
removing parameters if and when they are needed. Thus a function can be written, such as SumN
in Exhibit 9.3, that reads its first parameter(s) and uses it (them) to determine how many optional
parameters to read.
Exhibit 9.4 shows legal and meaningful calls on SumN. In every case, the number of arguments
provided equals the number consumed by the function, and removed from the stack.
Compare the meaningful calls in Exhibit 9.4 with the first meaningless one in Exhibit 9.5. If
this call were made beginning with an empty stack, the result would be a run-time stack underflow.
If the stack was nonempty before this call, a value left there for some other purpose would be
consumed, and the inevitable stack underflow would be postponed indefinitely.

Exhibit 9.4. Meaningful calls on a variable-argument function in FORTH.

Note: Execution is strictly left-to-right.
21 4 55 62
32 5
32 5

4 SumN
2 SumN
2 SumN 5 *

(
(
(
(

Sum the four numbers 21, 4, 55, and 62. )

Sum the two numbers 32 and 5. )
First sum the two numbers 32 and 5, then )
multiply the result by 5. )

238

CHAPTER 9. FUNCTIONS AND PARAMETERS

Exhibit 9.5. Meaningless calls on a variable-argument function in FORTH.

32 5

3 SumN

( Consumes one more parameter than was supplied, )

( causing immediate or eventual stack underflow. )

21 4 55 62

3 SumN

( Fails to consume a parameter, leaving garbage )

( on the stack. )

Too many arguments are supplied in the second meaningless call. This leaves garbage (the extra
argument) on the stack. This garbage may or may not interfere with future program operation and
can lead to a run-time error. In any case, leaving garbage lying around is undesirable.
Note that FORTH is a postfix language and has no syntactic markers for the beginning and
end of a parameter list. Moreover, there is no way to tell how many parameters will be used by a
function without executing it on specific data. It is impossible, therefore, for a FORTH translator
to help the programmer achieve semantic validity by checking that she or he has called his functions
in meaningful ways.
FORTHs stack operations are extremely low-level. Higher-level languages do not normally let
the programmer manipulate the translators internal data structures because the semantics of this
kind of operation are unclear and unpredictable at compile time. If the number of parameters
actually supplied is fewer than the number needed, the program will crash. If too many parameters
are supplied, junk will build up on the stack.
Although C functions do specify how many parameters they expect, it is possible to call a C
function with a different number of parameters. In general, it would not be very useful to pass
extra parameters to a function, but that is how the printf function works [Exhibit 9.6]. It uses
the number of arguments specified by the information in its first argument (which is the format
specification) to determine how many more arguments to process.

Exhibit 9.6. A call on a variable-argument function in C.

The call on printf below has three argumentsa format string and two numeric items. The
format, in quotes, contains two % fields; each one directs that an integer is to be converted to
ASCII and written to the output stream.
printf ("The sum is %d and the average is %d\n", sum, sum/count);
This call will execute correctly because the number (and types) of conversions specified in the
format matches the rest of the argument list. The format string is not actually checked at all at
compile time. Rather, it is passed on to the run-time system to be interpreted when printf is
called. A mismatch in number or type between the format fields and the actual arguments will
cause a run-time error or simply produce garbage.

9.2. WHAT DOES AN ARGUMENT MEAN?

239

C will translate programs containing calls with missing or extra parameters, and such programs
will behave much like similar programs in FORTH. All is well if the program actually uses the
number of parameters that are supplied. If too few are supplied, the program will either access
garbage, causing unpredictable behavior, or it will stop with a stack underflow.
Supplying too many parameters will not cause C to malfunction. In current C implementations,
parameters are removed from the stack by the calling program, not by the subroutine. This produces
a longer calling sequence but avoids creating a serious problem by passing too many parameters.
Excess parameters will sit on the stack unused until the subroutine exits, but then they will be
removed. Junk will not pile up on the stack.
Pros and Cons. The examples cited previously help us answer the question, To what extent is
it good and/or bad to permit argument lists of indefinite length?
Putting syntactic markers around parameter lists in a program makes them easier for humans
to work with, and thus is good. Once these markers are required by a language, it is easy for a
translator to check that the right number of parameters is used in every call. Doing so prevents
the programmer from writing nonsense that will cause programs to malfunction or crash. Thus a
parameter checking mechanism in the translator can be a powerful aid to writing correct programs.
On the other hand, there is a real cost involved. Once in a while it is very useful to write a
function, such as printf, that accepts a variable number of parameters. This can be done by the
programmer in C but not in a language such as Pascal that requires well-formed function calls. The
Pascal I/O procedures do accept a variable number of parameters, but they are predefined and no
functions like them can be defined by the programmer.
Whether or not a restriction is severe depends partly on whether the useful things it prohibits can
be achieved some other way. For example, a function in a list-oriented language might accomplish
the same end more simply by taking a single argument that is a variable-length list.
Thus the question of whether to require function calls to be well-formed can be reduced to
making a set of value judgements about the relative importance of (1) semantic validity and (2)
unrestricted flexibility. The answer also depends on what other mechanisms are included in the
language.
Most language designers placed a higher value on communicating and preserving semantic intent
than the designers of FORTH and C, who valued flexibility more highly. Neither of these languages
enforces the very helpful restriction that a function must be called with a semantically meaningful
number of parameters. Both are considered to be systems programming languages, and both
provide nearly full access to the machine hardware. Both permit programs to be written that can
crash easily and in spectacular ways.

9.2

What Does an Argument Mean?

Chapter 8, Section 8.4.1, discusses the order in which arguments are evaluated. Here we examine
how an argument is passed to a function and interpreted within the function.

240

CHAPTER 9. FUNCTIONS AND PARAMETERS

Exhibit 9.7. A procedure using call-by-value in Pascal.

PROCEDURE test1 (J2, A2: integer; P2:list);
BEGIN
writeln (J2, A2, P2.value);
J2 := J2 + 1;
P2 := P2.next;
writeln (J2, A2, P2.value)
END;

ALGOL-60, the first formally defined higher-level language, was defined with inside-out expression evaluation and both call-by-value and call-by-name parameter passing mechanisms. Since then,
other call mechanisms have been devised: call-by-need, call-by-reference, call-by-value-and-return,
and call-by-pointer. The syntax and the semantics of these call mechanisms differ in important
ways, which we explore in this section.

9.2.1

Call-by-Value

Call-by-value is the simplest and cleanest parameter passing mechanism, both in terms of its semantics and its implementation. In call-by-value, an expression written as an actual parameter is
evaluated before beginning the execution of a subprogram. The resulting value is written into the
stack area reserved for the subprogram and bound to the corresponding formal parameter name
within the subprogram. Exhibit 9.7 shows a Pascal procedure named test1 which has call-by-value
parameters. Exhibit 9.8 shows the relevant storage allocated for both the calling program and for
test1 at the beginning of execution of the call:
test1( J1, A1[J1], P1.next );
Exhibit 9.9 shows the result of executing test1, just after the procedure return and after
deallocation of the stack frame for test1. Two lines of output are produced by this call:
1 30 %
2 30 $
Assigning a value to the dummy parameter affects only the stack storage for the subroutine; it
does not change anything in the storage area belonging to the calling program. Note, in Exhibit
9.9, that nothing in the stack frame of the calling program has been modified. The call-by-value
mechanism is, thus, a powerful tool for limiting unintended side effects and making a program
modular. Call-by-value guarantees that no subroutine can mess up the values of its nonpointer
arguments in the calling context. For this reason, it is the most useful of the mechanisms for passing
parameters and should be used wherever it can do the job.
Passing a pointer, by value, to a subroutine will not permit the pointer itself to be changed, but
it permits modification of storage owned by the calling program and accessible through the pointer.

241

9.2. WHAT DOES AN ARGUMENT MEAN?

Exhibit 9.8. Result of parameter binding by value in Pascal.

Double arrows represent bindings; plain arrows represent pointers. The symbol marks the
current stack frame.

Stack frame for

calling program
J1
1
A1

Stack frame
for test1
J2
1

30 45 26

30

P1

A2
P2

Exhibit 9.9. Result of execution with value parameters.

Deallocated
Stack frame for
calling program
J1
1
A1
P1

Stack frame
for test1
2
J2

30 45 26

30
*

A2
P2

242

CHAPTER 9. FUNCTIONS AND PARAMETERS

In Exhibit 9.7 it would have been possible to modify contents of the list by assigning a new value
to P2.value or P2.next. This is demonstrated in Section 9.2.6. A pointer argument, therefore,
partially breaches the protection inherent in call-by-value.

9.2.2

Call-by-Name

In the vocabulary of programming languages, lambda calculus uses a call-by-name parameter interpretation rule. This is defined formally in Exhibit 8.10. Briefly put, each occurrence of a
call-by-name formal parameter in an expression is to be replaced by the entire actual argument
expression written in the function application or call.
Two function passing mechanisms were implemented in ALGOL-60 because it was recognized
that call-by-value parameters do not permit information to be passed back from a function to the
calling program.1 While a single value may be returned as the value of a function, this is often not
adequate. Many common applications, such as the swap routine, require passing back more than
one datum.
The common swap subroutine takes two arguments, both variables, and uses a temporary
storage object to exchange their values. Since the values of two variables are changed, this cannot
be done with a function return value. What is actually necessary is to pass the two storage objects
(not just the values in those objects) into the subroutine. The values in these objects can then be
interchanged. We can pass a storage object by passing its name, its address, or a pointer pointing
to it.
ALGOL-60 was developed at the same time as LISP, when lambda calculus was strongly influencing programming language design. This was before reference parameter binding was well
understood and accepted, and before the concept of a storage object was well understood. It was
not surprising, then, that the ALGOL-60 designers included call-by-name as a means of implementing swap and similar procedures.
In a call-by-name system, entire formulas are passed as arguments, to be evaluated within the
called function, but using the symbols defined in the calling program. Thus one could pass a
variable name to a subroutine, and inside the subroutine it would evaluate to the variables address
and provide access to that variables storage object.
Symbols in an argument expression are evaluated in the context of the calling program. This
means that if a symbol, S, occurs in an argument expression and is redefined as a local symbol in
the subprogram, the global meaning (not the local meaning) is used in evaluating the argument.
To implement this rule, name conflicts must be eliminated. This requires lambda-calculus-like
renaming [Exhibit 8.10], which is somewhat hard to implement. 2
Call-by-name semantics has two more nasty properties. Suppose that the definition of a function, F , contains three references to the parameter P , and that we call F with an expression, E,
as the actual parameter corresponding to P . If we use call-by-value or call-by-reference (defined
in Section 9.2.3), then E will be evaluated once, before evaluating the function, and its value will
1
2

This is true unless the language supports pointer types. ALGOL-60 did not have pointers.
The difficulty of translating call-by-name contributed to the unpopularity of ALGOL-60 in this country.

243

9.2. WHAT DOES AN ARGUMENT MEAN?

Exhibit 9.10. Parameter binding by name.

In our imaginary Pascal-like language, only the procedure header would differ from the version in
Exhibit 9.7. The header for the call-by-name version of test would be:
PROCEDURE test2 (NAME J2, A2:

integer; NAME P2:list);

Execution of the procedure call

test2( J1, A1[J1], P1.next );
is traced in Exhibits 9.11 and 9.12.

be bound to P . When F is evaluated, that single value will be used three times. But if we use
call-by-name, then the original expression, E, will be evaluated every time the code of F refers to P .
In this example, E will be evaluated three times. Of course, this is not very efficient, particularly
if E is a large expression.
Worse yet, if the process of evaluating E has a side effect (such as producing output or changing
the value of a global variable), the side effect will happen more than once, and the results of the
first time might affect the value of the expression the other times or might produce extra output.
Thus call-by-name on a parameter with a side effect is semantically messy as well as inefficient and
hard to implement.
Pascal does not support parameter binding by name. In order to produce an example of binding
by name that can be compared to the binding examples in Exhibits 9.7 and 9.13 we must use an
imaginary extension of Pascal that includes the keyword NAME, placed before a dummy parameter,
to indicate that the parameter is to be passed by name. Exhibit 9.10 shows a procedure, test2,
written in this imaginary language. Exhibit 9.11 shows the bindings created for test2 and the
relevant storage allocated for the calling program, just after parameter binding and before procedure
execution.
The result, just after the procedure return, of executing test2 is shown in Exhibit 9.12. The
bindings for the subroutine have been undone. Note that, unlike test1, test2 modifies the linked
listit unlinks one cell. The value of J1 is also changed. The following two lines of output are
produced by this call. The second line differs from the call-by-value results because the argument
expression A1[J1] was evaluated a second time after the value of J1 had been changed.
1 30 %
2 45 $
Call-by-name proved to be a mistake and a headache. It was awkward to implement, inefficient
to execute, and more general than necessary to implement routines such as swap. Since the mid1960s a better way (call-by-reference) has been used to implement the desired semantics. Ordinary
call-by-name has not been incorporated in a language since ALGOL-60, as a primary function call
mechanism. Macro languages, such as the C preprocessor, use something akin to call-by-name to
interpret parameters when a macro definition is expanded. Argument strings are substituted bodily
for occurrences of the parameters. This can result in multiple copies of the argument string. If the

244

CHAPTER 9. FUNCTIONS AND PARAMETERS

Exhibit 9.11. Result of parameter binding by name.

Stack frame for
calling program
J1
A1

Bindings
for test2

1
30

J2 means J1
45

26

P1

A2 means A1[J1]
*

P2 means P1^.next

argument string is an expression, it will be compiled several times and executed several times.

9.2.3

Call-by-Reference

The call-by-reference parameter passing mechanism is also called call-by-address and, in Pascal,
VAR parameter passing. A reference parameter is used to pass a storage object to a subprogram.
An actual argument must be a variable name or an expression whose value is an address.
To implement a reference parameter, the compiler allocates enough storage for a pointer in the
subprograms stack area and binds the dummy parameter name to this stack location. A pointer
to the actual parameter is stored in this space. During procedure execution, the dummy parameter
name is indirectly bound to the arguments storage object (which belongs to the calling program).

Exhibit 9.12. Result of execution with by-name binding.

Unbound
Stack frame for
calling program
J1
2
A1
P1

Former Test2
bindings
J2 means J1

30 45 26

A2 means A1[J1]
*

P2 means P1^.next

245

9.2. WHAT DOES AN ARGUMENT MEAN?

Exhibit 9.13. A procedure using call-by-reference in Pascal.

Only the procedure header differs from the version in Exhibit 9.7. The keyword VAR indicates
call-by-reference.
PROCEDURE test3 (VAR J2, A2:

integer; VAR P2:list);

Execution of the following call on test3 is traced in Exhibits 9.14 and 9.15:
test3( J1, A1[J1], P1.next );

The value in this storage object may be changed by assigning a value to the dummy name.
When the attribute VAR is used to declare a dummy parameter name in Pascal, the parameter is
generally translated as call-by-reference. 3 Exhibit 9.13 shows the function header for a by-reference
version of the test procedure. Exhibit 9.14 shows the relevant storage allocated for both the calling
program and for test3, after parameter binding but before procedure execution.
In this implementation the stack contents are the same as if a pointer argument had been passed
by-value. But the code that the programmer writes in a subroutine is different because the pointer
which implements the reference parameter is a binding, not an ordinary pointer.
A reference parameter binding is a true binding; it is transparent to the programmer. When
using call-by-reference, the translator automatically dereferences every occurrence of a by-reference
parameter. A naive user does not even realize that the binding pointer exists. In the stack diagram
in Exhibit 9.14, this extra automatic dereference is indicated by the double binding arrows from
the slots allocated for the parameters to the actual arguments. Thus we can say that the reference
parameter is indirectly bound to the argument. As with any binding, the programmer cannot change
the binding by executing an assignment.
The results of call-by-reference differ from both call-by-value and call-by-name. Exhibit 9.15
shows the stack just after the procedure return, and after deallocation of the stack frame for test3.
Note that two things have been modified in the storage diagram: the value of J1 and a cell in the
linked list. These changes are like the ones produced by test2, using call-by-name. But the output
produced differs from the test2 output, and is like the test1 output:
1 30 %
2 30 $
All parameters are passed by reference in FORTRAN, as are all array parameters in C. Call-byreference is optional in Pascal and is used to implement output parameters in Ada.
In languages that have both call-by-value and call-by-reference, the former is the preferred
parameter mechanism for most purposes because it provides greater protection for the calling
program and makes the subprogram a more thoroughly isolated module. Call-by-reference is used
in two situations: when the result of the subprogram must be returned through the parameter list
and when copying the entire actual argument would be intolerably inefficient (e.g., when it is a
3

Call-by-value-and-return is the other possible implementation for a VAR parameter. See Section 9.2.5.

246

CHAPTER 9. FUNCTIONS AND PARAMETERS

Exhibit 9.14. VAR parameter binding in Pascal.

Stack frame for

calling program
J1
1
A1

Stack frame
for test3
J2

30 45 26

P1

A2
*

P2

Exhibit 9.15. Result of execution with a VAR parameter.

Deallocated
Stack frame for
calling program
J1
A1
P1

Stack frame
for test3
J2

2
30 45 26

A2
*

P2

9.2. WHAT DOES AN ARGUMENT MEAN?

247

long array).

9.2.4

Call-by-Return

In practice, Pascal programmers use the call-by-reference mechanism for two reasons:
1. So that a value may be returned.
2. To avoid consuming the execution time and stack space necessary to pass a very large argument by value.
Using a VAR parameter for the first reason is semantically sound and in keeping with the intent
of the language designer. Using a VAR parameter for the second reason is not semantically sound,
even though it may be necessary to achieve acceptable performance from a program. Call-by-value
semantics is the only sound semantics for a parameter that does not carry information out of the
subroutine.
This illustrates a defect in the design of Pascal; the only efficient mechanism for passing large
arguments (call-by-reference) is tied to the semantics appropriate for return values. In Ada, this
design defect was partially corrected by permitting the programmer to specify the desired semantics
and letting the translator choose an appropriate mechanism. Thus an Ada parameter has a declared
mode. The mode is declared to be in if the parameter carries information into the subroutine or
out if it carries information back to the calling program. A two-way parameter is declared to be
in out.
The mode in is like a value parameter except that, within the subroutine, it has the semantics of
a constant and a new value cannot be assigned to it. This is more restrictive than call-by-value, since
most languages permit a value parameter to be used as a local variable and receive assignments.
For this reason, we will refer to this mode as call-by-constant. This can be implemented by using
call-by-value and prohibiting the use of the parameter name on the left side of an assignment.
The Ada mode out is also referred to as call-by-return. A call-by-return parameter is writeonly; it carries information out of the subprogram but not into the subprogram. Within the
subprogram, an out parameter can be implemented by a local variable. The program is allowed to
store information in this variable but not fetch information from it. When the function returns,
the final value of the out parameter is stored in the location specified by corresponding argument.
For output parameters, call-by-return is semantically cleaner than call-by-reference because access
to the location in the calling program is write-only and can happen only at function return time.
Call-by-return is therefore preferred.
The Ada mode in out corresponds to the VAR parameter mechanism in Pascal. Unfortunately,
the Ada standard does not fully specify the semantics that must accompany an in out parameter. A
compiler is permitted to implement either call-by-reference or call-by-value-and-return. A program
that depends on the difference between these two calling mechanisms is only partially defined. As
seen in the next section, this is a truly unfortunate shortcoming in a language that was designed
to support multitasking!

248

CHAPTER 9. FUNCTIONS AND PARAMETERS

9.2.5

Call-by-Value-and-Return

We can combine call-by-value and call-by-return to achieve two-way communication between called
and caller, with no accessing restrictions inside the subroutine. At first look, call-by-value-andreturn seems more complex than call-by-reference, and it seems to offer no advantages. While this
is true in an isolated program on a simple machine, it does not hold in more complex environments.
Two factors, hardware complexity and concurrency, make call-by-value-and-return the preferred
mechanism.
Consider a machine with a partitioned memory, or a program that is being executed on a
networked system. In such hardware environments, nearby memory is fast (and therefore cheap)
to access, and more distant memory costs more. A function that executes using only local storage
will be more efficient than one that accesses data in another memory segment. When call-by-value
is used, the data is copied from the callers memory into the local memory. Although this copying
operation takes time, much more time may be saved by avoiding out-of-segment references to the
parameter. Call-by-value is simply more efficient once the argument has been copied. This is true
on any partitioned architecture, even an IBM PC, although the savings may be minor. However,
for a program on a network that is processing data stored at a remote location, the savings would
be substantial, and the difference in response time would be important.
The difference between call-by-reference and call-by-value-and-return is very important in an
application where two or more processes concurrently access the same data object. Suppose that
processes P1 and P2 both have received argument ARG as a reference parameter. Using call-byvalue-and-return, the answer is always one of three things:
If the first process to begin, P1, ends before the second one, P2, starts, the final value in
ARG is just the result of ordinary, sequential execution.
If P1 starts, then P2 starts, then P1 returns, then P2 returns, the result of P1s execution is
wiped out. The value in ARG is the answer from P2, as if P1 had never been called.
If P1 starts, then P2 starts, then P2 returns, then P1 returns, the result of P2s execution is
wiped out. The value in ARG is the answer from P1, as if P2 had never been called.
In any case, the final result is some value that was correctly and meaningfully computed. It may
seem very undesirable to have the possibility that the result of calling a procedure could be totally
wiped out. However, that situation is better than the alternative; call-by-reference can cause real
trouble in a concurrent application. If both P1 and P2 read and modify ARG several times, and if
their actions are interspersed, the final value of ARG can be completely unpredictable. 4 Worst of
all, the value left in ARG could be different from any legal value that could be computed by either
process alone, or both processes executed in either sequence.
The modern call-by-value-and-return semantics should be used for Ada because Ada was specifically designed to be used:
On any machine architecture.
4

Most textbooks on operating systems cover this problem in a chapter on concurrency.

9.2. WHAT DOES AN ARGUMENT MEAN?

249

Exhibit 9.16. A swap routine using reference parameters in Pascal.

The VAR keyword is used before the parameter name to indicate a reference parameter. Assuming
that j and k are integers, an appropriate call would be: swap1 (j, k);
PROCEDURE swap1 (VAR a, b: integer);
VAR t:integer;
BEGIN
t:=a; a:=b; b:=t
END;

On networks of machines.
With concurrent, multitasked applications.

9.2.6

Call-by-Pointer

Call-by-pointer is a subcase of call-by-value. The contents of the pointer variable in the calling
program is an address. During parameter passing this address is copied into the corresponding
pointer variable in the stack frame for the subroutine. Now two pointers, one in each stack frame,
point at the same object. Because the argument is the address of a storage object belonging to
the calling program, it can be used to return a value from the subroutine. Unlike a by-reference
parameter, though, a pointer parameter must be explicitly dereferenced within the subprogram.
Three versions of a swap subroutine are given here in two languages. These illustrate the similarities and differences between call-by-reference and call-by-pointer. In all three, the programmer
indicates that a variable address, not a value, is passed into the subroutine so that the swapped
values can be passed outward. Exhibit 9.16 shows a procedure named swap1 which uses call-byreference. Exhibit 9.17 shows the storage allocated for execution of swap1.
The subprogram stack frames for call-by-pointer [Exhibit 9.21] and call-by-reference contain
the same information, but that information has a different meaning because of the extra automatic
dereference that accompanies call-by-reference. When call-by-pointer is used to implement the
same process, the programmer must write an explicit dereference symbol each time the pointer
parameter is used, as shown in Exhibits 9.18 and 9.20.
Contrast Exhibit 9.17 to Exhibit 9.21. Note the double binding arrows in the former, and the
single pointer arrows in the latter. These semantic differences account for and correspond to the
absence of explicit dereference symbols in Exhibit 9.16 and their presence ( in Pascal and *
in C) in Exhibits 9.18 and 9.20.
Moreover, although a by-reference binding cannot be changed, the address stored in the pointer
parameter can be changed by an assignment statement, as shown on the last line of Exhibit 9.22.
Use of call-by-pointer is severely restricted in Pascal. Since Pascal pointers can never point
at objects on the stack, one can only use call-by-pointer to process dynamically allocated storage

250

CHAPTER 9. FUNCTIONS AND PARAMETERS

Exhibit 9.17. The implementation of reference parameters in Pascal.

Stack diagrams for swap1 are shown here, just before the code in swap1 is executed and after return
from the subroutine. The arrowhead at the top of the stack frame indicates the current stack frame.

stack frame
for main

stack frame
for swap1

stack frame
for main

stack frame
for swap1

t
At entry to swap1

3
After swap1

Deallocated

Exhibit 9.18. A swap routine using pointer parameters in Pascal.

The pointer parameters are passed by value and explicitly dereferenced within the subroutine.
TYPE int_ptr = integer;
VAR jp, kp: int_ptr;
PROCEDURE swap2 (a, b: int_ptr);
VAR t:integer;
BEGIN
t:=a ; a :=b ; b :=t
END;
Assume that the following code has been executed. The last line is an appropriate call on swap2.
Execution of this call is traced in Exhibit 9.19.
NEW(jp); jp := 3;
NEW(kp); kp := 5;
swap2 (jp, kp);

251

9.2. WHAT DOES AN ARGUMENT MEAN?

Exhibit 9.19. Storage for swap2.

This diagram shows the storage allocated for swap2 and the calling program, just after exit from
the call on swap2. Note that the values of pointers jp and kp were copied into swap2s stack frame,
and provide access to objects that were dynamically allocated by the calling program.

stack frame
for main

dynamic
storage

stack frame
for swap2

jp

kp

b
3

After swap2

Deallocated

objects. This accounts for the difference between Exhibit 9.19 and Exhibit 9.24. C does not restrict
use of pointers; the programmer can use the & operator to get the address of (a pointer to) any
variable. Thus call-by-pointer can be used in C to pass an argument allocated in either the stack
or the heap.
Two versions of the swap routine can be written in Pascal, with the use of the second restricted
to heap-allocated objects. In contrast, only one version is possible in C because C does not support
call-by-reference. In its place, call-by-value is used to pass the value of a pointer (an address).
We call this parameter passing method call-by-pointer [Exhibit 9.20]. In this version of the swap
routine, the type of the dummy parameters is int * (pointer-to-integer). The parameters must
be explicitly dereferenced (using the * symbol), just as is used in the Pascal swap2 routine.
Two distinct kinds of calls on this one function are possible. In the first, shown in Exhibit 9.21,
we use the & operator to pass the addresses of the two integers whose values are to be swapped.

Exhibit 9.20. Pointer parameters in C.

void swap3 (int *a, int *b)
{ int t;
t = *a; *a = *b; *b = t;
}

252

CHAPTER 9. FUNCTIONS AND PARAMETERS

Exhibit 9.21. Call-by-pointer using & in C.

main(){
int j=3, k=5;
swap3 (&j, &k);
}

/* Declare and initialize two integers.

*/

The stack diagram on the left shows execution of this call just after creation of the stack frame for
swap3. The diagram on the right shows storage after exit from swap3.

stack frame
for main

stack frame
for swap3

stack frame
for main

stack frame
for swap3

t
At entry to swap3, first call

3
After swap3

Deallocated

This is not possible in Pascal, as Pascal does not have an & (address of) operator.
Note that only the values in the variables j and k are swapped; the pointers a and b point at
the same variables throughout. A reader who is confused by this should note, in Exhibit 9.22, that
an assignment statement that changes a pointer is quite different from an assignment that changes
a value.
The second call, shown in Exhibit 9.23, corresponds to the call on swap2 in Pascal. In the main
program, we initialize pointers to the two integer variables and pass those pointers (by value, of
course) to swap3. The result is the same as the first call: the stack frame for swap3 contains two
pointers, initialized at block entry to the addresses of j and k. The only difference between these

Exhibit 9.22. Assignments using pointers in C.

int j=3, k=5;
int *p= &j;

/* j and k are integer variables.

/* Make the pointer p point at j.

*/
*/

*p = k;
p = &k;

/* Variable j now contains a 5. */

/* Make p point at k instead of j. */

253

9.2. WHAT DOES AN ARGUMENT MEAN?

Exhibit 9.23. Call-by-pointer using pointer variables in C.

main(){
int j=3, k=5;
int *jp = &j, *kp = &k;

/* Declare and initialize two integers.

/* Make pointers point at the integers.

*/
*/

swap3 (jp, kp);

}
The stack diagram in Exhibit 9.24 shows execution of this call just after creation of the stack frame
for swap3.

two kinds of call-by-pointer is the presence or absence of the pointer variables in main. If these
pointers are useful for other reasons, the second call would be stylistically better. Otherwise the
first call should be used.
There are two important practical differences between reference parameters and pointer parameters. The first has already been noted: pointers must be explicitly dereferenced. Experience
has shown that a common error is using one too few or one too many dereference symbols. The
call-by-reference mechanism is inherently less error prone.
Second, once you are permitted to point freely at items allocated on the stack, it is possible to
create dangling pointers. This is why the address of operator was omitted from Pascal. Wirths

Exhibit 9.24. Stack diagram for call-by-pointer.

This diagram shows execution of the code from Exhibit 9.23.

stack frame
for main

stack frame
for swap3

jp
kp
After exit from second call on swap3

Deallocated

254

CHAPTER 9. FUNCTIONS AND PARAMETERS

Exhibit 9.25. Creation of two dangling pointers in C.

int * dangle (int ** ppp) /* The parameter is the address of a pointer.
{
int p=5;
int m=21;
*ppp = &p;
return &m;
main()
{
int k = 17;
int *pm, *pk = &k;
pm = dangle(&pk);
}

*/

/* Dereference ppp to get the pointer whose address

was passed. Make it point at local p. */
/* A second dangling pointer. */ }

/* pk is a pointer pointing at k. */
/* Upon return, pk and pm point at */
/* deallocated spaces. */

Exhibit 9.26 shows storage before and after returning from the call on dangle.

intent was that Pascal should prevent the programmer from making many semantic errors by
providing only semantically safe operators.
There are two simple ways to create a dangling pointer in a language such as C that permits an
address of operation to be applied to any object. First, it can be done by passing an argument
which is the address of a pointer to an object. Assume pk is a pointer to an object. Then &pk is
the address of a pointer. By passing &pk to a subprogram we permit the subprogram to change
the value of pk. Inside the subprogram, pk can be set to point to a local variable. Second, a
dangling reference may be created by returning the address of a local variable from a function.
Upon returning from the function in either case, the local variable will be deallocated and the
calling program will point to a storage object that no longer exists. This is illustrated in Exhibits
9.25 and 9.26.

9.3

Higher-Order Functions

A higher-order function is one that takes a function as an argument or returns a function as its
result. Lambda calculus has higher-order functionst makes no distinction between a functional
object and a nonfunctional object. The pure functional languages, which are modeled after lambda
calculus, give full support to higher-order functions, and this is the basis of much of their power.
Other languages give partial support or no support, because of the difficulties involved. In this
section we look at some of the applications of higher-order functions and consider some of the
difficulties involved in their implementation.

255

9.3. HIGHER-ORDER FUNCTIONS

Exhibit 9.26. Creating a dangling pointer.

The stack before execution of dangle, showing parameter bindings and local allocation.

Stack frame for main

k
17
pk
pm

Stack frame for dangle

ppp
5

21

The stack, after return from dangle, showing results of execution.

Stack frame for main

k
17
pk
pm

Stack frame for dangle

ppp
5

21

Deallocated

9.3.1

Functional Arguments

It is not too difficult and it is very useful to support functions as arguments to other functions.
Let us consider two of the more common applications of functional arguments, flexible code and
mapping functions.
Flexible Code. Sometimes a complex general process can have several variants. For example,
consider a spreadsheet that consists of one line of data per student and one column for each assignment or exam, in which the students grade is recorded. In the process of assigning grades,
an instructor will sort that data many times in many ways. Sometimes the sort will be on an
alphabetical column, sometimes on a numeric column. Sometimes ascending order is needed, sometimes descending order. A grading program (or spreadsheet used to do the grading) must contain
a general sort procedure, which takes the comparison function as its argument, along with the
function to select the correct column as the sort key. For this application, four different comparison
functions might be given, numeric <, alphabetic <, numeric >, or alphabetic >. By using
functional arguments, one sort routine can handle all four jobs.

256
Exhibit 9.27.
Code
A1101
+/ A
+\A
/ A
A
\A

CHAPTER 9. FUNCTIONS AND PARAMETERS

Mapping functions in APL.

Operation
Results
Bind A to an array of four Boolean values.
plus reduce
3 (Count true values.)
plus scan
1223
logical-or reduce
1 (At least one true exists.)
logical-and reduce 0 (Some value is false.)
logical-and scan
1 1 0 0 (Find first false.)

Consider another example program for finding the roots of an arbitrary mathematical function. One way to program this is to write and compile the code to calculate the function (or many
functions) and pass that function as an argument to the general root-finding procedure. This permits the root-finder to be kept in a library and used for many applications without modification.
The general graphing facilities of packages such as Lotus must also be able to take an arbitrary
function as its argument and operate on it.
Mapping Functions. Frequently, we have a monadic or dyadic operation that we wish to apply
to a list or array of data objects. A mapping function defines a particular pattern for extending
the operation to process multiple data values instead of one or two. It takes two arguments,
a function and a list of data values, and applies the function to the data values in the defined
pattern. Examples of mapping functions in APL and LISP are given in Exhibits 9.27 and 9.28.
The APL reduce (/) and scan (\) operations both take a dyadic functional argument
and an array argument. Reduce computes a running operation. The result is the same as if the
functional argument were written between each pair of values from the data array, and the entire
expression were then evaluated. Thus if the function argument is +, reduce totals the array. The
scan operation works similarly, except that its answer is the same length as the input array and
contains all the partial sums. These operations are fundamental to giving APL its power and are
also essential operations in modern parallel algorithms.
There are several mapping functions in LISP, that iterate a functional argument over a list
argument or arguments. In all cases, the number of list arguments must match the number of
parameters in the functional argument; a monadic function needs one list, a dyadic function needs
two. Let us use the function mapcar as an example; mapcar goes down the argument lists, selecting
the next value from each list argument at each step and applying the functional argument to those
values. The result becomes the next item in the result list. A null argument produces a null result,
and a list of any other length maps onto another, transformed list of equal length. Iteration stops
when the end of the shortest list-argument is reached. If the functional argument returns a list,
the final output is a list of lists [Exhibit 9.28].
The function mapcdr is similar to mapcar, except that the parameter of the functional argument
must be of type list, and at each step, the function is applied to the entire list, and then the head

257

9.3. HIGHER-ORDER FUNCTIONS

Exhibit 9.28. Some applications of mapcar in LISP.

Assume x is the list (4 9 16 25), y is (1 2 3 4), z is NIL, and w is ((3 4 5) (2 1) (7 9 3 6)), which
is a list of lists. The quote mark before the function name causes the actual, executable function
(not the result of evaluating it) to be sent to mapcar.
Expression
(mapcar +1 x)
(mapcar + x y)
(mapcar +1 z)
(mapcar (lambda (a b)
(- (* 3 a) b) ))
x y)
(mapcdr caar w)

Comments
Increment each element.
Add corresponding elements.
The empty list is no problem.
Apply the function ((3*a)-b)
to corresponding elements
of lists x and y.
The first item of each sublist.

List Returned
(5 10 17 26)
(5 11 19 29)
NIL

(11 25 45 71)

of the list is discarded. (The remaining list argument is one value shorter after each iteration.) The
result of each step is appended to the output list.
Implementation. A functional argument can be passed to another function easily and efficiently
by passing a pointer to the function, not the function itself. The only implementation problem
concerns type checkingthe type of the functional argument, that is, the types of all its arguments
and the type of its result, must be known before code can be compiled to call that function. This
accounts for one major difference between the functional languages and Pascal or ANSI C. All
support functional parameters, but the functional languages are interpreted and do not require the
declaration of the full type of a functional argument since its type can be deduced, when necessary,
from the functions definition.
A short Pascal function with a functional argument is shown in Exhibit 9.29. It is a very limited
version of APLs reduce operator. The capabilities of this code are much more restricted than the
reduce operation in APL because Pascal is a strongly typed, compiled language, and APL is not.
5 Note that the functional parameter, f, is declared with a full header, including dummy parameter
names x and y which are meaningless and never used.

9.3.2

Currying

In functional languages,6 every function of two or more arguments is considered to be a higher-order

function of one argument that returns another function. This way of looking at functions is called
currying7 The returned function represents a function of one fewer argument which is the same as
5

A similar version of reduce with some type flexibility is given in C in Exhibit 16.4.
We will use Miranda syntax, because it is simple and elegant, to illustrate the properties common to all functional
languages.
7
Named after Haskell Curry, an early logician who used this device extensively in his work on semantics.
6

258

CHAPTER 9. FUNCTIONS AND PARAMETERS

Exhibit 9.29. A Pascal version of reduce.

This function will apply any dyadic integer function to reduce an array of ten integers.
Type row_type = array[0..9] of integer;
Function Reduce( Function f(x:integer, y:integer):integer;
ar: row_type):integer;
Var sum, k: integer;
Begin
sum := ar[0];
For k=1 to 9 do
sum := f(sum, ar[k]);
Reduce := sum;
End;

the original function in which the first argument has been fixed to the given value. This pattern
is repeated until all but one of the arguments have been used, and the final function returns a
data-object.
An example should help make this clear. Consider the Miranda function pythag a b = sqrt
(a * a + b * b), which uses the Pythagorean theorem to compute the length of the hypotenuse
of a right triangle whose base and height have lengths a and b respectively. Pythag is considered by
Miranda to be a function that takes one numeric argument and returns a function. That function
takes a numeric argument and returns a number. Thus the result returned by pythag 3 is a
function that finds the hypotenuse of any right triangle whose base has length 3. The following
Miranda expressions all return the length of the hypotenuse (= 5) of a right triangle with base 3
and height 4:
pythag 3 4
(pythag 3) 4
f 4 if we have previously defined f = pythag 3
In the last expression above, we have given a temporary name f to the function returned by
pythag 3. f is itself a legitimate function and can be used like any other function. For example,
the expression
(f 4) + (f 7)
computes the sum of the lengths of the hypotenuses of two right triangles, one with base 3 and
height 4 and the other with base 3 and height 7. See if you can figure out what is computed by
this Miranda expression:
f (f 4) where f = pythag 3

259

9.3. HIGHER-ORDER FUNCTIONS

Exhibit 9.30. The type of reduce in Miranda.

reduce ::
(num num num )
[num]
num
num

The function reduce takes three arguments,

a binary operator,
a list of numbers,
and an identity element.
It returns a number.

Notes:
Type num means number and is a primitive type in Miranda.
A function type is denoted by use of parentheses.
We denote the type of a list with base type T as [T].

The notation for expressing a curried function type is:

hfunction namei ::hargument typei hresult typei
The type of a higher-order function is written by repeated uses of the notation. The
sign is right associative, as shown below, permitting us to write the type of an ordinary numeric
function without using parentheses. These two type expressions are equivalent and denote the type
of myfunc:
myfunc ::
myfunc ::

num num num

num (num num)

We can use this notation to write the type of a function that takes a function as a parameter. For
example, Exhibit 9.30 shows the type of a Miranda version of the reduce function. (The function
itself is shown and discussed in Chapter 12, Exhibit 12.15.) This version of reduce takes a binary
arithmetic operator, a list of numbers (possibly null), and an identity value, and returns a number.
The identity value must be appropriate for the given operator; it is used to initialize the summing
process and is returned as the value of reduce when the list argument is a null list.

9.3.3

Returning Functions from Functions

C supports functional arguments, in the form of pointers to functions, and C++ carries this further
to permit the predefined operators to be named and used as arguments also. Both languages
permit the programmer to return a pointer to a function as the result of a function. However,
a C or C++ function cannot do one thing that distinguishes functional languages from all other
languages: create a function within a function and return it as the result of the function. At first
sight, this facility might seem esoteric and not very useful; it is, however, a powerful and concise

260

CHAPTER 9. FUNCTIONS AND PARAMETERS

Exhibit 9.31. Partially closing a function to make new functions.

We define
by closing
plus
plus

a simple function of two arguments in Miranda and produce two new functions from it
its first argument. Assume the symbol x is bound in the enclosing context.
:: num (num num )
a b = a + b

add1 :: num num

add1 = plus 1
plusx :: num num
plusx = plus (x * 10)

way to achieve several practical ends, including composition functions, closures, and infinite lists.
We discuss closures here and infinite lists in Chapter 12.
Closures
Closure is an operation that can be applied to an expression. We close an expression by binding
each of its free variables to the value of another expression. A closure creates a function, which is
returned as the result of the closure process. The most useful and easily understood application of
closure is known as partial parameterization. In this process we take a function of n arguments,
bind one to a locally defined value, and produce a function of n 1 arguments.
For example, we can partially parameterize + by closing its first argument; the result is a
one-argument function that adds a constant to its argument. If the closed argument is bound to 1,
the result is commonly known as +1, or increment. Exhibit 9.31 shows the types and definitions,
in Miranda, for two functions produced by closing the first argument of plus. The result is always a
function of one argument, whose type is the same as the type of plus, with the first clause omitted.
An Application of Closures. As an example of the possible use of closures, consider the currency conversion problem. An international bank has a general currency-conversion function,
Convert, that takes a conversion rate and an amount of currency and returns an amount in a
different currency. The bank has a second program that is run each morning. It reads the current
conversion rates for all currencies the bank handles, then creates two closures of the convert program for each rate. For example, after reading the fact that one mark = .6 dollars, two closures
would be created:
MarksToDollars = Convert 1.67
DollarsToMarks = Convert .6
The resulting closures would be accessed through a menu offering dozens of specialized conversion
routines. These could be called up during the business day, as needed.

9.3. HIGHER-ORDER FUNCTIONS

261

Exhibit 9.32. We cant make closures in Pascal.

The function MySum calls the function Reduce from Exhibit 9.29 with the first argument bound to
a specific operation.
Function MyOp(x: integer, y:integer):integer;
Begin
MyOp := abs(x) + abs(y)
End;
Function MySum(ar: row_type):integer;
Begin
MySum := Reduce( MyOp, ar)
End;

Implementation of Closures. Closure is a useful device for taking general, library code modules
and tailoring them to the needs of a particular environment. Although it produces a new function,
it does not require translating new code; the function code for the closure is just like the code of an
existing function. Thus we can implement closure using a record containing bindings for symbols,
together with a pointer to a function.
When we close a function with a constant value, as in the function add1 [Exhibit 9.31], there
is no question about what the meaning is or how to construct the closure. In this case the closure
would contain a pointer to the function plus and the information that the symbol a in plus is
bound to 1.
However, a problem arises when we close plus with an expression such as x * 10: what should
be recorded in the closure? Should that expression be evaluated and the result stored in the closure,
or should the expression itself become part of the closure? The answer is, we can still be lazy. We
can safely defer evaluation of the expression until the value is needed. There is no need to evaluate
the expression when we make a closure, because the result of evaluating the expression will always
be the same. Each symbol used in the expression has one meaning from the time it becomes defined
until the end of its lifetime; assignment doesnt exist and cant change that meaning. So we defer
evaluation, and compute x * 10 when it is used, if it is used.
The symbol x may have some other meaning in the surrounding program, and the closure, which
is a function, might be passed around and executed in that other context. Whatever happens,
though, when plusx is used, x must mean whatever it did when plus was closed. To make this
work, the closure must contain the current binding for x, and the data-part of the closure must
have a lifetime longer than the lifetime of the block that created the closure.
Except for the lifetime problem, the data-part of a closure is identical to the stack frame that
would be created for a function activation; it contains the local context in terms of which local
symbols must be interpreted. The difference between a closure and a stack frame is that stack
frames have nested lifetimes and closures do not. A stack frame is deallocated before deallocating
the stack frame for the block that created it, but a closure must outlive that block. For this reason,
closure is a process that can only be done in functional languages or in other languages that have
a special closure facility.
The student must get a clear understanding of the difference between a closure and the functions

262

CHAPTER 9. FUNCTIONS AND PARAMETERS

that can be written in traditional languages. The Pascal programmer can write (and then compile)
a function that calls another function with a constant parameter. For example, MySum, defined in
Exhibit 9.32, calls the function Reduce, from Exhibit 9.29 with its first argument bound to a locally
defined function, MyOp. A function such as MySum is not like a closure for two reasons:
This code must be compiled before it becomes a function! The binding of f in Reduce to
MyOp is created by the compiler, not at run time.
The lifetime and visibility of MySum are limited. It can be used only within the block in which
MyOp and Mysum are defined, and it will die when that block dies.
Most things that can be done with closures, and with higher-order functions in general, can also
be done in Pascal. To emulate a closure, the programmer would write a program with the necessary
bindings in it at a global level, then compile the code. When executed, the result will work just
like a closure. However, the ability to create closures during execution gives a more concise and
efficient way to accomplish the job.

Exercises
1. What is the difference between an argument and a parameter?
2. What is a well-formed function call? Explain the following statement: K&R C is a language
in which function calls do not need to be well-formed. Is this an advantage or disadvantage
for C programmers?
3. In Ada, a function may be called with fewer arguments than the number of parameters in the
function definition. Explain how. Why is this useful?
4. In FORTH, function definitions do not include parameter declarations. This permits the
programmer to write a function that uses different numbers of arguments on each call. Explain
one benefit and one drawback of this design.
5. How does call-by-value afford protection to the calling program?
6. Why is a call-by-value method of parameter passing ineffective for a swap routine that intends
to swap values in the calling program?
7. Why does Pascal refuse to bind a VAR parameter to a constant?
8. Why does Pascal need VAR parameters, but C functions can communicate as well using only
value parameters?
9. The attributes in, out, and in out that are specified in Ada parameter declarations
have a basic good property that Pascal cannot achieve with its declarations of VAR or value
parameter passing. What is it? (Note: The fact that Ada permits optional parameters is not
relevant here.)

9.3. HIGHER-ORDER FUNCTIONS

263

10. Explain why a language that supports multitasking will be likely to support call-by-valueand-return rather than call-by-reference.
11. A short program is given below in Pascal with one procedure definition. Show the storage
allocated for the main program and the stack frame (activation record) created for the call on
the subroutine. Note that one parameter is a reference (VAR) parameter, and one is a value
parameter.
Program ParameterDemo;
Var I: integer;
A: array[1..3] of integer;
Procedure Strange (Var X: integer; Y: integer);
Var temp: integer;
Begin temp:=X; X:=X+Y; Y:=temp End;
Begin (* main program *)
For I := 1 to 3 Do A[I] := 10-I;
I := 2;
Writeln ( Initial values , I, A[1], A[2], A[3]);
Strange( A[I], I );
Writeln ( Final values , I, A[1], A[2], A[3]);
End.
12. Use the stack frame diagram from question 11 to trace execution of this program. Write values
in the storage locations, and show how the values change during execution of the program.
Show the output produced.
13. Change both of the parameters of the procedure named Strange, in question 11, to VAR
parameters. Answer the questions 11 and 12 again showing the storage setup, data changes,
and output.
14. What is lazy evaluation? Why is it used?
15. What is call-by-pointer? Why is it a subcase of call-by-value and different from call-byreference?
16. Explain the differences between call-by-reference (as in a Pascal VAR parameter) and call-bypointer, with respect to the following:
a. What is written in the procedure call.
b. What is written in the procedure header.
c. How the parameter is used within the procedure body.
17. What is a dangling pointer?

264

CHAPTER 9. FUNCTIONS AND PARAMETERS

18. Write a short piece of Pascal or C code that creates a dangling pointer.
19. Fill in answers to the twelve questions in the following chart. To name times, use the phrases
load time, block entry time, block exit time, program termination, any time, or
other (please explain other). To label accessible locations, use the phrases anywhere,
anywhere except where masked, in the declaring block.
Type of variable
global variable
static local variable
ordinary local variable
dynamic heap variable

Time of
creation?
1.
4.
7.
10.

Time of
death?
2.
5.
8.
11.

Where is it
accessible?
3.
6.
9.
12.

20. What is printed by the following Pascal program? Make diagrams of storage for both the main
program and subroutine, showing the initial contents and changes in contents. Distinguish
between VAR and value parameters in your diagram. This is trickier than tricky. You are
being tested on comprehension of name scoping and VAR and value parameters.
PROGRAM trick (INPUT, OUTPUT);
VAR J, K, L: INTEGER;
FUNCTION F(J: INTEGER; VAR L: INTEGER):CHAR;
VAR K: INTEGER;
BEGIN
K := L + J;
L := K;
J := K;
IF K > 10 THEN F := Y ELSE F := N
END;
BEGIN
(* main program *)
J := 3; K := 15; L := 4;
WRITELN( F(L, J) );
WRITELN (J, K, L)
END.
21. What is a higher-order function? Explain two uses for higher-order functions.
22. Name two languages that do not support higher-order functions in any way. Explain how you
found the information to give the answer.
23. Explain what a closure is.
24. C gives some support for higher-order functions: pointers to functions may be passed as
parameters and returned from functions. Explain what kind of support is missing in Cthat

9.3. HIGHER-ORDER FUNCTIONS

265

is, what can a programmer do with higher-order functions in a functional language that C
does not support?
25. Express the type of the function search from Chapter 8 Exhibit 8.14 in Pascal type notation,
as if it were to be passed as an argument to another function.
26. Express the type of the function search from Chapter 8 Exhibit 8.14 in curried notation.

266

CHAPTER 9. FUNCTIONS AND PARAMETERS

Chapter 10

Control Structures

Overview
This chapter treats the subject of control structures. A control structure is a language
feature which defines the order of evaluation of expressions or larger units. Control
structures exist in a variety of forms, the lowest level being the primitive control instructions which are defined by the hardware. These include instruction sequencing,
conditional and unconditional jumps, and subroutine calls. Control structures in highlevel programming languages developed out of and are implemented in terms of these
instructions.
Above the expression level, there are four kinds of control structures. These are subroutine call with parameters, statement sequence, conditional execution, and repetition. In
order to be useful, programs must be able to perform a section of code for one particular set of inputs and a different set of actions for another. Functional languages have
only control expressions: that is, expressions that contain one or more conditions, and
for each condition, an expression to evaluate when the condition is true. Procedural
languages, on the other hand, have both control expressions and control statements.
Conditional forms are basic and absolutely essential in a programming language because
they are inherently outside-in control structures. The conditional lets us test whether
it is safe to proceed with a computation before doing so. Various conditional forms are
discussed, ranging from very primitive to very general. The generalized conditional is
shown to be the most flexible. and the CASE statement the most efficient.
The simplest decision structure is the conditional IF hconditioni GOTO hlabeli. Unfortunately, its usage often leads to spaghetti code because the GOTO may lead anywhere in
the program. This problem is largely resolved by using structured conditionals such as
267

268

CHAPTER 10. CONTROL STRUCTURES

If..Then..Elseif..Else. The CASE statement is an efficient structured conditional
control statement with multiple branches. There are three ways to implement the CASE
statement to achieve greater efficiency than a structured conditional. Many languages
support the CASE; several of these (including the forms in Pascal, C, and early COBOL)
have defects which make the CASE statement cumbersome to use.
Iterative control structures are used to process homogeneous data structures such as
arrays, lists, and files. The most primitive loop is the infinite loop, which is a onein/zero-out structure. The two simplest finite loops, formed with a conditional GOTO,
are the repeat loop (with its test at the end of the loop) and the while loop (with a test
at the top). These correspond to the structured REPEAT and WHILE statements found
in most languages. Ada, Turing, and FORTH have combined these forms into a more
general loop in which the loop termination test may appear anywhere in the loop.
If the length of a data structure is known, automatic processing can be performed by a
counted loop. The FOR loop has two parts:
A control element, containing an initial value for the loop variable, a goal, and an
increment.
A scope, containing a sequence of statements to be iterated.
Most languages restrict the expressions that may be used in the control element, and
most prohibit changing the loop variable within the scope. The iteration element is a
generalization of the counted loop with no restrictions on the expressions within the
control element or use of the loop variable. It is not, strictly, a counted loop, and it
cannot be implemented as efficiently.
Some languages support implicit iteration. Their basic data structure is a set, array,
or list of some sort, and they support a variety of iterative functions to process an
entire compound object and return a simple or compound result. Implicitly iterative
expressions add great power to a language and permit many complex operations to be
expressed briefly and clearly.

10.1

Basic Control Structures

A control structure is a language feature defined in the semantics of the language (not in syntax
only) that defines the order of evaluation of expressions or larger program units.
The primitive control instructions defined by the hardware in all computers are listed below.
The control structures in higher languages developed out of these and are implemented in terms
of them. Execution starts when the computer is turned on or receives a RESET signal. A boot

269

10.1. BASIC CONTROL STRUCTURES

Exhibit 10.1. Diagram of a sequence of statements.

Statement 1
Statement 2
...
Statement n

sequence is built into the computer hardware which causes actual execution to start somewhere, in
a location defined by the hardware. In modern computers, the boot sequence initializes whatever
is necessary, loads the operating system (the OS), then transfers control to it. The OS then
accepts a series of commands and carries them out by loading and transferring control to other
system programs or to application programs. The transfer of control to a program is represented
graphically by the arrow leading into the top of its control diagram.

10.1.1

Normal Instruction Sequencing

Normally, instructions are executed in the order that they are loaded into memory. This is carried
out by the instruction cycle of the machine. A typical version of the cycle is as follows:
1. Instruction fetch: Load the CPUs instruction register from the memory address stored in
the instruction counter register (IC).
2. Instruction counter increment: Add 1 to the address stored in the IC, in preparation for
fetching the next instruction.
3. Decode the instruction: Connect all the appropriate registers and logic circuits. If the instruction references memory, load the memory address register with the address.
4. Execute: Do the current instruction. If it references memory, do the fetch or store. If it is an
arithmetic instruction, perform the operation.
Machine instructions are executed in the order in which they are loaded into memory unless
that order is changed by a jump instruction. A sequence of instructions that does not include jumps
will be diagrammed as a simple box, as in Exhibit 10.1.
Sequences are a basic control structure in procedural languages. The code section of a program
in a procedural language is a sequence of statements; execution starts at the first and progresses to
the next until the end is reached. Programmers are trained to analyze their problems as ordered
sequences of steps to be performed on a set of objects. The pure functional languages do not
support statement sequences. Rather, sequences are eliminated by use of nested function calls and
lazy evaluation. Chapter 12 discusses this approach to programming.

270

10.1.2

CHAPTER 10. CONTROL STRUCTURES

Assemblers

A symbolic assembly language is built on the semantic basis provided by the raw machine, machine
language, a symbol table, and ways to define variables, labels, and functions. A macro assembler
provides source code macros in addition. This semantic basis is completely flexible: using it you
can express any action that your machine can do. But it is not a very powerful semantic basis
because it lacks any way to convey the programmers semantic intent to the translator, and the
translator has no mechanisms for ensuring that this intent is carried out.
Macros ease the task of writing code but do not extend the semantic basis, since all macros are
expanded, that is, replaced by their definitions, before translating the program.
Assemblers impose no restrictions, good or bad, on the programmer; they permit the programmer to write good code or totally meaningless code. Higher-level languages impose many kinds of
helpful restrictions, including some that can be checked at compile time and others that must be
checked at load-link time or at run time. These restrictions help the translator identify meaningless
code.
Two of the primitive semantic features supported by assemblers cause a lot of difficulty in debugging and in proving that a program is correct. These are (1) the GOTO, and (2) destructive
assignment. The GOTO introduces complex data-dependent flow patterns which cannot be predicted at compile time, and assignment to variables introduces dependency on time and the order
of evaluation into the meaning of an expression. Both make proofs of correctness hard to construct.
In the rest of this chapter, we examine aspects of control structures that are supported by
higher-level programming languages. Two kinds of control structures above the expression level
are basic: subroutine call with parameters and conditional execution (discussed in Section 10.2).
Procedural languages have two more basic kinds of control: execution of a sequence of statements
and repetition (discussed in Section 10.3). Many languages, especially older ones, provide a fifth
kind of control structure: a GOTO [Section 11.1], which is more or less restricted depending on
the language. Finally, some languages contain primitives for communication between concurrent,
asynchronous processes, such as the task names and accept statement in Ada. 1

10.1.3

Sequence, Subroutine Call, IF, and WHILE Suffice

In 1968 Dijkstra2 wrote an important and provocative paper that advocated that the GOTO
statement should be dropped from programming languages because it was the cause of far too
many program errors. This position was based on earlier work by Jacopini 3 which showed that any
flow diagram can be written, without arbitrary GOTOs, in terms of conditionals, while loops, and
sequences [Exhibit 10.2].
Dijkstra did not claim that programs limited to these control structures would be maximally
1

Concurrency primitives are beyond the scope of this book.

Dijkstras ideas were published first in Europe: Dijkstra [1965]. His ideas became known in the United States
after he published a second short paper: Dijkstra [1968].
3
Cf. B
ohm and Jacopini [1966].
2

271

10.1. BASIC CONTROL STRUCTURES

Exhibit 10.2. A sufficient set of basic control structures.

These are the traditional flow diagrams for Dijkstras three control structures.

The WHILE loop

Condition

False

The IF ... THEN ... ELSE

True

Condition

The sequence

False

True
Scope
of Loop

Scope of
True action

Scope of
False action

efficient, merely that they could be written, and written more rapidly and with less likelihood of
errors. His evidence was drawn from real industrial experience and from personal observation.
Each of these control structures is comprised of a group of boxes with connecting arrows. Each
group is a one-in/one-out structure; that is, on a flowchart, one line comes into the control structure
and one line goes out of it. This is an important property: programs limited to these formation
patterns are easier to debug because the effects of any transfer of control are confined to a small
and defined part of the algorithm.
Restricting control to one-in/one-out structures also makes mathematical proofs about the
correctness of the program much easier to construct. Such a program can be decomposed into a
series of short one-in/one-out sections, and correctness conditions can be written that relate the
program state at the top of a section to the state at the bottom. Next, the correctness of each
section can be proved independently. Finally, these short proofs can be combined into a whole,
yielding a proof of correctness of the entire program whose structure matches the program structure.
The term structured programming has no single definition but is generally used to mean a topdown style of program development. The first step is to state the high-level purpose, requirements,
and general process of a program. Each process step is then refined by elaborating the definition
and adding more detail. In the end, procedures and functions are written to correspond to each
portion of the problem definition, and these must be written using only the one-in/one-out control
structures provided by the programming language.
Pascal is the structured language most widely used now for instruction. It contains recursive
functions and procedures, the three control structures just discussed (conditionals, WHILE loops,
sequences), two additional loop control structures (FOR and REPEAT loops) [Exhibits 10.27 and
10.23], and a badly designed multibranching conditional structure (CASE) [Exhibit 10.18]. These

272

CHAPTER 10. CONTROL STRUCTURES

Exhibit 10.3. Diagram of a subroutine call with parameters.

The left diagram represents a procedure call (a function call statement); the right diagram shows
three function calls nested within an expression.
Function
Name
Parameter 1
...
Parameter n

c :=

Foo
a
Goo
3

Goo
17

extra control structures are not necessary in a complete language, but they are certainly nice to
have. Pascal also contains a GOTO instruction, but students are often not told that it exists and are
seldom taught how or why to use it.

10.1.4

Subroutine Call

This instruction, also called jump to subroutine, saves the current value of the IC, then stores the
entry address of the subroutine in the IC. Subroutine return restores the saved address and brings
control back to the instruction after the call. A simple subroutine call box is shown on the left
in Exhibit 10.3. This is diagrammed as a two-part box, divided horizontally, with the subroutine
name written above the line and the parameters written below [Exhibit 10.3]. Where a subroutine
call is nested within another call or within a statement, these boxes can be used to show the nesting
structure. For example, the box on the right in Exhibit 10.3 represents the following nested function
calls:
c := Foo(a, Goo(3), b) + Goo(17);
All modern programming languages provide the ability to define and call subroutines with
parameters. Without subroutines, it is nearly impossible to write and debug a large program. The
programmer must break a large problem into conceptually manageable pieces. Subroutines are the
programmers tool for implementing each part in such a way that the interfaces are clearly defined
and the interactions among parts are limited. This methodology, of decomposing a problem into
clearly defined parts with limited and clearly defined interactions, then building a program by
implementing each, is called top-down programming. It is now recognized to be the basis of good
program design.
To understand why there is no real substitute for subroutines with parameters, let us look at a
pair of languages that did not, initially, support them: BASIC and dBASE.

273

10.1. BASIC CONTROL STRUCTURES

Exhibit 10.4. Diagram of the computers primitive GOTO instruction.

Box before the GOTO

GOTO

Another box representing a labelled

statement elsewhere in the program.

BASIC was originally defined4 lacking a crucial feature. It had a GOSUB xxx statement which
did a subroutine jump to statement number xxx, and a RETURN statement which returned control
to the calling program, but parameters and results had to be passed through global variables. The
programmer could emulate the normal parameter passing mechanism by storing each parameter
value in a global variable, calling the subroutine, and using the same global variable name inside
the subroutine. There were no dummy parameter names. The subroutine would likewise store
its results in global variables. Thus what is easy in most languages became cumbersome in BASIC,
and the programmer had to be careful not to use the same name twice for two different purposes.
Worse, the subroutines actions and those of the main program were not well isolated from each
other.
BASIC was recently revised and updated by the author of the original BASIC language, John
Kemeny. The new version is called True BASIC 5 , perhaps to distinguish it both from the original
language and from the dozens of nonstandardized extended BASIC implementations that were
developed. This new version upgrades the original language extensively, but perhaps the most
important improvement is the inclusion of subroutines with parameters.
dBASE is a popular data base language for personal computers that includes extensive procedures and functions for forming data bases and handling and extracting data from data base files.
In addition, it includes a general programming facility similar to BASIC which can be used to create
fancy input and output programs. This general programming facility was added to dBASE in pieces,
as it became clear that data base primitives alone were not enough. Flexibility in handling input,
output, and some processing and statistics gathering can only be achieved with a general-purpose
language. The most recent version of dBASE was extended to include a parameter facility for subroutines. With this extension dBASE has become a general, interpretive, procedural programming
language specialized for handling data bases.

10.1.5

Jump and Conditional Jump

Jump and conditional jump instructions change the next instruction to be executed by storing a
new address into the IC. They differ from a jump to subroutine instruction in that they do not save
or restore the address from which the jump originates. The GOTO is an unconstrained transfer of
control to some other part of the program. It is diagrammed as an arrow, as in Exhibit 10.4.
4
5

Dartmouth [1965].
Kemeny [1985].

274

CHAPTER 10. CONTROL STRUCTURES

Exhibit 10.5. Diagram of the primitive conditional.

Prior statement
Condition

Next statement executed

for TRUE condition.

Next statement executed

for FALSE condition.

A test or condition will be diagrammed as a trapezoid [Exhibit 10.5]. Control enters at the top
of this box. Control leaves to the right through the diagonal side (along the horizontal arrow) if the
condition is true. It leaves through the bottom (the downward arrow) for a false condition. The
arrows are not normally written on the diagram and are shown here and in subsequent diagrams
only to emphasize the flow of control.

10.1.6

Control Diagrams

In order to visualize and compare the wide range of control statements in different languages, we
need a language-independent representation for each basic kind of control. Flowcharts provide
a graphic representation of control flow but cannot represent the difference between GOTO and
structured transfers of control. We define a new kind of diagram, called a control diagram, that can
represent the limited branches inherent in modern control structures. 6 Simple diagrams are defined
for each of the primitive control structures. These diagrams will be elaborated and extended to
develop higher-level control structures.
Modern control structures are one-in/one-out units, that is, control can enter the unit at exactly
one place and leave the unit at exactly one place. Within a control structure, control may jump
around, skip sections, and form repetitive loops.
Our control diagrams represent this control flow implicitly and/or explicitly. Each control
diagram has a frame with one entry arrow and one exit arrow. Smaller boxes may be nested within
the outer frame. Control flows downward, implicitly, unless an arrow or a branch (indicated by a
box section with a diagonal side) directs it elsewhere.
The control diagram for a sequence of statements was shown in Exhibit 10.1. The single entry
is indicated by an arrow along the top edge of the sequence box. When control enters this box it
progresses downward through any enclosed statements until it reaches the bottom line, then exits
on the single exit arrow.
The control diagram for a subroutine call was shown in Exhibit 10.3. When control enters a
subroutine call box, the arguments are evaluated and bound to the parameter names; then, control
6

These diagrams are modeled after, but are not the same as, Nassi and Schneidermans system.

10.2. CONDITIONAL CONTROL STRUCTURES

275

Exhibit 10.6. Syntax and uses of a conditional expression in C.

Syntax: hconditioni ? hexpressioni : hexpressioni
Two examples of usage:
a = b<c ? b : c;
/* Store the smaller value in a. */
printf("%d\n", b<c ? b : c); /* Print whichever is smaller, b or c.

*/

goes to the subroutine code and eventually returns to the bottom of the subroutine call box. Then
it leaves by the exit arrow.
Control diagrams for conditionals, loops, and limited control transfers will be presented with
these control structures in the remaining sections of this chapter.

10.2

Conditional Control Structures

All programming languages must contain some control structure that permits conditional execution
of a section of code or selection among two or more alternative sections of code. This is obviously
fundamental to programming: a program that performs the same actions for all inputs is of limited
use.
The kinds of control structures included in a language are, in part, determined by whether the
language is functional (nested), or a mixture of functional and procedural (sequential). Functional
languages contain only expressionsthey have no statements, and, therefore, all control structures
in these languages are control expressions. Procedural languages all have control statements, and
some also include control expressions. Let us explore the difference between control expressions
and control statements by looking at conditionals.

10.2.1

Conditional Expressions versus Conditional Statements

The conditional expression is the control structure most intimately associated with the functional
languages. It is an expression that contains one or more conditions and, for each condition, an
expression to evaluate when the condition is true. A final expression is included to evaluate when
all the conditions are false. The conditional expression returns, as its result, the value of whichever
expression was evaluated. The simplest conditional expression has the general form:
if hconditioni then hexpression 1i else hexpression 2i
For example, the value of the following expression is B or C, whichever is smaller:
if B < C then B else C;
Because a conditional expression returns a result, it may be embedded in a larger expression.
Exhibit 10.6 shows the syntax and usage of the conditional expression in C.
Contrast this to the conditional statement, which returns no result. A simple conditional statement has the basic form:

276

CHAPTER 10. CONTROL STRUCTURES

Exhibit 10.7. Syntax and uses of a conditional statement in C.

Syntax: if (hconditioni) hstatementi else hstatementi;
Usage: these statements are equivalent to the conditional expressions in Exhibit 10.6.
if (b<c) a=b; else a=c;
/* Store the smaller value in a. */
if (b<c) printf("%d\n", b); else printf("%d\n", c);

if hconditioni then hstatement 1i else hstatement 2i;

No result is returned, and the conditional statement is used by including entire statements in its
clauses and by placing the entire unit in a sequence of statements. To be useful, the included
statements must either do I/O or store their results in a set of variables, as illustrated by Exhibit
10.7.
It is not necessary for a language to provide both a conditional expression and a conditional
statement, although some do. Anything that can be written using a conditional expression can also
be written (more redundantly) using a conditional statement, statement sequencing, assignment,
and a temporary variable, V. To do this, the conditional expression is first extracted from its context
and its THEN and ELSE clauses modified to store their results in V:
if hconditioni then V=hexpression 1i else V=hexpression 2i
Then the outer expression, in which the conditional expression was embedded, is written with V
substituted for the conditional expression. Exhibit 10.7 shows the result of applying this process
to the expressions in Exhibit 10.6.
The opposite transformation is much more complex. Rewriting an entire conditional statement
as an expression may not be possible to do mechanically if it contains several statements in the
THEN and ELSE clauses. In the simple case, the transformation is done by factoring out the
common portion of the THEN and ELSE clauses. If what remains of the clauses can be written
as two expressions, hTexpi and hFexpi, and both return the same type result, we can finish the
transformation. The common part extracted from the THEN and ELSE clauses is written once
(instead of twice), with a hole left where hTexpi and hFexpi had been. Then the new conditional
expression is formed by writing
if horiginal conditioni then hTexpi else hFexpi
and is embedded in this hole.
The Effect of Conditional Expressions on Programming Style. A Pascal program cannot
be written in nested, LISP-like style because Pascal does not support conditional expressions. The
IF in Pascal is a statementit does not return a result and therefore may not be used in an argument
to a function. The IF statement must communicate through the the program environment and not
through parameters on the stack. In contrast, the LISP conditional, COND, is an expression that
returns a result which can be used in an enclosing expression. LISP conditionals can, therefore,

10.2. CONDITIONAL CONTROL STRUCTURES

277

communicate with the rest of the program using parameters. It is this difference, not the presence
or absence of many parentheses in the syntax, that is the root of the difference between the LISP
and Pascal programming styles.
Most languages are not purely sequential or purely nested. LISP is primarily a nested language,
but it also contains a rplaca (destructive assignment) function and the progn control structure.
These enable the programmer to use variables and to list a series of expressions to be executed
sequentially. Further, all the major sequential languages permit nested function calls, stack allocation, and parameter binding, and some, for example C, contain a conditional expression.
In spite of this extensive overlap in the actual semantic features supported by languages in the
two classes, LISP programs tend to be written as nests of function calls, and C programs tend to
be written as sequences of statements.
In the late 1970s several people decided that all those parentheses in LISP were bad. They
wrote front-end processors which accepted programs written in a much more Algol-like or C-like
syntax and translated them to ordinary LISP. These preprocessors supported arithmetic expressions
with infix operators and conditionals that started with IF. But the programs written for these frontends still used the modular, nested programming style, common with LISP, with little or no use of
variables.
On the other hand, consider C. Theoretically a C programmer could write using a LISP-like
style, but most do not. One can only speculate on the reasons for this. A few possibilities come to
mind:
1. The loop control structures in C are useful. C programmers often use arrays and files. Looping
statements are a natural way to process these data structures. But loops in C are statements,
not functions. If a programmer uses them, he or she cannot write in a purely nested style. In
contrast, LISP programmers often use list- and tree-structured data. Recursion, with nesting,
is a natural way to process these structures.
2. A programmer using C expects to write in a sequential rather than nested style, and a programmer writing in a LISP variant expects to write in a nested style. The expected style
dominates the program even though either style could be used in either language.
3. Sequential processing may be more natural for humans than highly nested function calls.
Breaking the code into statements and storing intermediate results in variables helps the
programmer keep things straight in his or her mind.
It is likely that some combination of these reasons works to keep C programmers writing programs that are more heavily procedural than the typical LISP program.

10.2.2

Conditional Branches: Simple Spaghetti

The simplest, most primitive decision structure is the simple Conditional GOTO, or IF-GO, diagrammed in Exhibit 10.8. This statement exactly implements the conditional branch instruction

278

CHAPTER 10. CONTROL STRUCTURES

Exhibit 10.8. Syntax and diagram of the simple IF-GO.

Possible source syntax:
Diagram

IF hconditioni GOTO hlabeli;

(Keywords vary.)

:
IF
condition

TRUE
THEN GOTO <label>

FALSE
ELSE (continue in sequence)

built into the hardware of most computers, and it was the only conditional control structure in the
original BASIC.
The IF-GO is often used to create spaghetti code, or code that contains many IF-GO instructions with overlapping scopes that form intertwined threads of code. Spaghetti code is easy to
write but hard to debug. It has poor lexical coherence and poor locality of effects. Visually, the
spaghetti is the GOTO arrow in the diagram, which may end anywhere in the program.
The IF-GO causes poor lexical coherence because it permits the THEN and ELSE clauses to be
placed at widely separated places in the program. Further, if statement labels are distant and not
in numeric or alphabetic order, they can be hard to locate in a large program. These problems
can, of course, have greater or smaller impact depending on the style of the programmer. The
lexical coherence problem is reduced considerably by using the IF-GO to emulate a structured
IF...THEN...ELSE [Exhibit 10.9]. The important semantic problems created by use of IF-GO are
discussed in Chapter 11, Section 11.1.

10.2.3

Structured Conditionals

Conditional Diagrams. Conditionals will always be diagrammed by boxes (or box sections)
with one diagonal side. Normally, control flows through that diagonal side when the condition is
true, but this default can be overridden by drawing a horizontal arrow labeled FALSE or F
that crosses the diagonal side. The other condition (normally FALSE) causes control to continue
flowing downward. Sometimes the diagrams are explicitly labeled for emphasis, even when the
default defines the desired direction of control flow.
Structured conditionals are one-in/one-out units, and the corresponding control diagrams have
a single entry at the top and a single exit somewhere along the bottom. The conditional is built out
of a series of condition clauses, which are rectangles divided in the middle by a diagonal line. The
condition is written to the left of the diagonal line, the actions to be performed when the condition

279

10.2. CONDITIONAL CONTROL STRUCTURES

Exhibit 10.9. The IF in BASIC.

Two control skeletons are shown here. On the left side, the THEN clause is placed at the end
of the program, far distant from the related IF and ELSE clauses. On the right, the THEN clause is
placed directly after its related ELSE clause. The number of GOTO instructions and labels required
for both cases is the same, but the right version has shorter spaghetti strands and, therefore, better
lexical coherence.

200

300

Poor lexical coherence

IF hconditioni GO 300
helse clausei
...
hrest of programi
...
END
hthen clausei
...
GO 200

100
200

Better lexical coherence

IF hconditioni GO 100
helse clausei
...
GO 200
hthen clausei
...
hrest of programi
...
END

is true are written on the right.

The jump around the ELSE clause or clauses is diagrammed by using a control frame along the
right and bottom edges of the box. Just as control structures replace GOTOs, these frames replace
long twisting control transfer arrows. Control enters a frame by way of a horizontal arrow after
completing a TRUE clause, then flows down to the bottom and out by the exit arrow. If control
reaches the final ELSE clause, it then goes downward into the frame and, again, out at the bottom
[Exhibit 10.10].
The Basic Forms. Any of the forms of the structured conditional can be implemented either as
a conditional expression (as in LISP) or as a conditional statement (as in Pascal). They differ from
the simple conditional branch statement by being one-in/one-out structures. Contrast the diagram
of the conditional branch in Exhibit 10.8 to the diagram of the structured IF-THEN conditional
on the left in Exhibit 10.10. The difference is that a surrounding frame has become a part of the
control structure, delimiting the scope of the TRUE clause. The conditional provides two ways to
enter this frame, but there is only one way to leave it: downward.
A more general form of this same control structure also includes a scope of actions for the FALSE
case, giving the full IF...THEN...ELSE control structure on the right in Exhibit 10.10. Use of this
control diagram is illustrated in Exhibit 10.11.
All modern procedural languages provide a conditional control structure with at least two
clauses, although the keywords and punctuation vary from language to language. The greatest
difference is the way the scopes of the THEN and ELSE are delimited. There are two possibilities:

280

CHAPTER 10. CONTROL STRUCTURES

Exhibit 10.10. Diagrams of the structured conditional.

In both diagrams, the GOTO of the simple IF-GO has been replaced by a block of statements, to
be executed when the condition is true. After executing this block control passes to the right into
the frame of the control diagram. If the condition is false, control passes downward into the frame.
In the diagram on the right, a box section is added for the ELSE clause. If the condition is
FALSE, control flows into this section, and from there downward into the frame.
This is a one-in/one-out structure. The second exit of the simple conditional branch has been
captured by the frame of this structure. However control reaches the frame, it leaves the frame
by the single downward exit.

IF
condition
FALSE

THEN
True actions

IF
condition
FALSE

THEN
True actions
TRUE

TRUE

ELSE
False actions
ENDIF

ENDIF

Exhibit 10.11. A structured conditional statement in Pascal.

A Pascal conditional statement is shown and diagrammed.
if index <= max then begin
writeln( name[index] );
IF
THEN
index := index + 1 end
index <= max
writeln( name[index] );
else
index := index + 1;
FALSE
TRUE
writeln(Index out of range);
ELSE
writeln( Index out of range );

ENDIF

10.2. CONDITIONAL CONTROL STRUCTURES

281

1. The scopes of the THEN and ELSE are each restricted to a single statement. If more than
one action is required, a compound statement may be used. This is a series of statements
delimited by BEGIN...END or equivalent bracketing symbols. This kind of syntax is used in
Pascal and C.
2. An explicit ENDIF is provided. The series of statements between the THEN and the ENDIF (or
the optional ELSE) is the TRUE-scope. The series of statements between the ELSE and the
ENDIF is the FALSE-scope. This syntax is used in FORTRAN, Ada, and Turing.
The advantage of choice 1 is that it has fewer keywords and the language syntax is simpler. A
disadvantage is that the syntax is ambiguous if one IF is nested inside another, and only one has
an ELSE clause. An extra rule in the grammar forces this lone ELSE to be parsed with the nearer
IF.
The advantages of choice 2 are that the program does not become filled with BEGIN...END pairs
which can interfere with indenting and can be a source of clutter. Also, the ambiguity problem
for nested conditionals is solved by the ENDIF. If ENDIF is placed before the lone ELSE, that ELSE
parses with the outer IF; otherwise, it goes with the inner IF.
Finally, this IF...THEN...ELSE control structure can be extended again to permit a series
of conditions to be tested [Exhibit 10.12]. This produces the most general and flexible possible
conditional control structure. Exhibit 10.13 shows a use of the general conditional in LISP. In other
languages, conditional clauses after the first are often denoted by the keyword ELSIF or ELSEIF, as
shown in Exhibit 10.14.
Some languages support the IF...THEN...ELSE but not the ELSEIF clause. To write a generalized conditional in these languages, one writes a series of conditionals, nesting each successive IF
inside the prior ELSE clause. The Pascal code fragment in Exhibit 10.15 does the same thing as
the LISP and Ada versions in Exhibits 10.13 and 10.14. Note the similarity to Ada code. The same
Pascal statement is shown indented in two ways. The version on the left shows the actual nesting
depth of the IFs. The version on the right reflects the semantics of the situation, a choice among
several parallel cases.
A Short History of Conditionals. At about the same time as the simple IF...THEN...ELSE
was introduced in ALGOL-60, the generalized conditional came into use in MAD 7 and in LISP.8 The
generalized form was later included in PL/1 9 but then was omitted from many new languages (e.g.,
C and Pascal) which instead included the simpler IF...THEN...ELSE from ALGOL-60.
The original FORTRAN did not include any structured conditionals at all. All forms of IF
used GOTOs to transfer control. The ANSI committee in charge of revising the FORTRAN standard
7
Michigan Algorithm Decoder, Arden, Galler, and Graham [1963]. This language was based on the preliminary
ALGOL report (1958) and implemented at the University of Michigan in about 1959.
8
LISt Processing language, McCarthy et al. [1962]. This language was implemented at M.I.T. in the late 1950s.
9
Programming Language 1, implemented by IBM in 1967 and intended to be the language that would fill everyones
needs. B. Galler, one of the designers of MAD, was on the committee that designed PL/1.

282

CHAPTER 10. CONTROL STRUCTURES

Exhibit 10.12. Diagram of the generalized structured conditional.

This is a one-in/one-out structure. Any number of conditions may be listed in series, each
associated with a scope of actions to be executed if that condition is TRUE. The conditions are
tested in the order written, and the actions for the first true condition are executed. If no condition
is true, the FALSE scope at the bottom is executed. After executing a scope, control passes into
the frame.
IF
condition 1
ELSE IF
condition 2
...

THEN
True actions 1
THEN
True actions 2

ELSE IF
condition n

THEN
True actions n
ELSE
False actions
ENDIF

resisted all urging to include a structured conditional in FORTRAN II and even in FORTRAN IV.
Many nonstandard extensions of FORTRAN IV included IF...THEN...ELSE...ENDIF, though, and
finally, with FORTRAN 77, the full structured IF with ELSEIF was included because of popular
demand. This illustrates both the trouble with design-by-committee, and the slowness with which
good new ideas are accepted by people accustomed to doing things in other ways.
The syntax of a language is certainly simplified (and translation is therefore made easier) by
omitting the ELSEIF. The ELSEIF adds no power to the language, since the same flow of control
can be created using the simple IF...THEN...ELSE control structure where the scope of each ELSE
is another IF statement.
Opinions about what is most important in language design vary from community to community

Exhibit 10.13. Use of the general conditional expression cond in LISP.

This call on the print function takes one piece of information, an age, and prints a word describing that age. Its control diagram is the same as the diagram in Exhibit 10.14.
(print (cond ((< age 5) ( "young child"))
((< age 13)( "child"))
((< age 18)( "adolescent"))
( T
( "adult"))))

283

10.2. CONDITIONAL CONTROL STRUCTURES

Exhibit 10.14. Use of the general conditional statement in Ada.

I/O in Ada is too complex to be explained here. Please simply assume that the user has defined
a procedure named pr that takes one string argument, and prints it, followed by a carriage return.
This code fragment then does the same thing as the LISP version in Exhibit 10.13. Its diagram is
on the right.
IF age < 5 THEN
pr("young child");
IF
THEN
ELSIF age < 13 THEN
age < 5
pr( "young child" );
pr("child");
ELSE IF
THEN
ELSIF age < 18 THEN
age < 13
pr( "child" );
pr("adolescent");
THEN
ELSE IF
pr( "adolescent" );
ELSE
age < 18
pr("adult");
ELSE
pr( "adult" );
END IF;
ENDIF

Exhibit 10.15. Using Pascal to emulate ELSEIF.

Indentation shows nesting depth.
Each IF is within the prior ELSE.
IF age < 5 THEN
writeln(young child)
ELSE IF age < 13 THEN
writeln(child)
ELSE IF age < 18 THEN
writeln(adolescent)
ELSE
writeln(adult);

Left-justified ELSE IF shows semantic

structurea series of equal choices.
IF age < 5 THEN
writeln(young child)
ELSE IF age < 13 THEN
writeln(child)
ELSE IF age < 18 THEN
writeln(adolescent)
ELSE
writeln(adult);

284

CHAPTER 10. CONTROL STRUCTURES

and change often. In this case, there is no agreement whether a good language should include
an ELSEIF clause. The ELSEIF statement has some nice properties:
It is a good reflection of the semantics of a series of parallel choices, a commonly occurring
control pattern.
It combines well with the ENDIF to delimit a series of conditional clauses without need for
BEGIN...END bracketing, which clutters the code.
A series of IFs has each IF nested within the ELSE clause of the preceding IF. This causes
many programmers to indent each IF more than the prior one, as shown in the left column of
Exhibit 10.15. In contrast, the ELSEIF statement encourages programmers to indent scopes
uniformly, in a manner consistent with their status as equally important alternatives.
Thus factors relating to human engineering are the primary reasons for including the more complex
ELSEIF syntax in a language design. It is not theoretically necessary or even helpful.

10.2.4

The Case Statement

While the generalized conditional is maximally flexible, it is not always maximally efficient or
convenient to write. Consider the situation in which one of a set of N actions must be taken
depending on the value of a single expression [Exhibit 10.13]. It is redundant to write that expression
N 1 times as part of N conditions, and it is inefficient to evaluate the expression N 1 times,
especially if it is a long expression. Some efficiency can be gained by evaluating the test expression
once and storing the answer in a variable, but that still requires fetching that value N 1 times.
In assembly language, a programmer would write only one fetch.
This need was met in nonstructured procedural languages by the computed GOTO statement.
APL, COBOL, and FORTRAN all contain versions of a computed GOTO. In this control structure,
an expression is evaluated and its result is used to select one statement label from a list of labels.
Control then goes to that label.
Essentials of the CASE Statement. What is needed here is a structured conditional control
statement with multiple branches that implements this common control pattern in an efficient
way. The single expression can be evaluated once at the top of the control structure and its value
loaded into a machine register where it can then be used to index a transfer vector or be compared
efficiently to a series of constants to determine which set of actions to execute. This general CASE
structure is illustrated in Exhibit 10.16.
This is the control structure implemented by the CASE in Ada. It is a complex control structure
whose semantics are defined by the following list of rules:
1. Each set of case labels is a series of constants or constant expressions of the same discrete
type as the result of the test expression.

10.2. CONDITIONAL CONTROL STRUCTURES

285

Exhibit 10.16. The CASE statement.

Evaluate the test expression.
Compare result to case labels.
GOTO the matching label.
Case Labels 1

Actions 1

Case Labels 2

Actions 2

...
Case Labels n

Actions n

Actions for other values

2. The actions executed are the ones following the case label that is equal to the value of the
test expression. The same constant may not be included in two sets of labels.
3. If no label equals the test value, the other values clause is executed.
4. After executing a set of actions, control passes to the right, into the frame, and from there
to the single exit.
Possible variations of the general CASE structure involve the rules for the type of value returned
by the test expression and the rules for labeling cases. The value must, in any case, be a discrete
value. Real numbers such as 3.91 would not make meaningful case labels because of the approximate
nature of real arithmetic.
Some languages require that every possible value of the test expression occur exactly once as
a case label. If a type with a large number of possible values, such as integer, is permitted, some
way must be provided to avoid listing all possible values. This might be done by using subranges
of the type as case labels or by using an OTHERS clause to cover nonexceptional cases.
There are three good implementations of the CASE structure. In all three, the test expression is
evaluated once at the top and loaded into a machine register. An easy implementation is possible
if the machine instruction set includes a computed goto instruction, and the range of values of the
control expression is small and matches the range that the hardware is built to handle.
The second implementation, a transfer vector, is exceptionally fast but possibly requires a lot of
space. It uses the register containing the test expression value to index an array of code addresses.
The compiler generates this array by first sorting the case labels into ascending order, then storing
the address of the object code for a case action in the array position corresponding to the value

286

CHAPTER 10. CONTROL STRUCTURES

of its case label. To execute case N , the translator generates a simple indirect GOTO through the
N th position of the label array. With this implementation, every possible value in the type of the
test expression must occur as a case label, and the transfer array will be as long as the number of
possible case labels.
A third implementation of the case is possible if one of the above conditions does not hold. If
a possible case label may be omitted or a very large type (like integer) is permitted for the case
labels, the CASE must be translated into a series of conditional GOTOs. A series of case labels can
be compared very efficiently to the register containing the test value, and a GOTO taken to the first
label matching the value in the register. This is not as fast as use of a transfer vector, but it is a
big improvement over the code that would be generated (by a nonoptimizing compiler) for a series
of IF statements.
Defective Versions of the CASE Statement.
COBOLs CASE. The CASE statement was introduced into COBOL a long time ago when the
programming community did not understand that GOTOs cause unending trouble. Unlike the modern
CASE, the COBOL CASE is a multiway conditional GOTO, and a program containing a CASE becomes
one massive nest of GOTOs. Procedure calls and GOTOs interact in a curious way in COBOL; the
same block of code can be entered either way. If a block is entered by a procedure call, control
returns to the caller at block end. But if it is entered by a GOTO, control continues on to the next
lexical program element. Using the CASE in such an environment makes a program even harder to
debug than it would be in FORTRAN or in BASIC. Control can shoot off in any direction if the
programmer is not careful.
A good CASE statement is particularly useful in modern data processing applications where online data entry is expedited by menu-driven programs. The general CASE is ideal for menu handling.
COBOL 8510 introduced a new statement type, called EVALUATE which implements the general case
structure [Exhibit 10.17]. It is an especially nice version of the general control structure because it
allows range expressions to be used as case labels.
The introduction of the EVALUATE statement in COBOL fixes an old defect and illustrates the
way languages can grow and adapt to an increasingly sophisticated public. It also illustrates one
negative aspect of language growth and restandardization: the proliferation of keywords. Rather
than redefine the keyword CASE to have a modern semantics, a new keyword was introduced. The
standardization committee had no choice about introducing a new keyword: if they redefined the
old one, all existing programs that used CASE would become obsolete overnight. In the business
world this is certainly unacceptable. Many companies have programs in use that were written ten
years ago and are simply modified every time the companys needs change. Reworking all those
programs to eliminate the CASE GOTOs is not reasonable. The old CASE statement must now be
categorized as an archaic feature that should never be used. Perhaps in twenty years it can be
dropped from the language.
10

A good presentation of the changes in the language can be found in Stern and Stern [1988].

287

10.2. CONDITIONAL CONTROL STRUCTURES

Exhibit 10.17. The COBOL EVALUATE statement.

General Syntax
EVALUATE hexpressioni
WHEN hcondition-1i haction-1i
WHEN hcondition-2i haction-2i
...
WHEN hcondition-ni haction-ni
WHEN OTHER haction-otheri
END EVALUATE

Code Sample

EVALUATE EXAM-SCORE
WHEN 90 THRU 100 PERFORM A-ROUTINE
WHEN 80 THRU 89 PERFORM B-ROUTINE
...
WHEN 0 THRU 59 PERFORM F-ROUTINE
WHEN OTHER PERFORM ERROR-ROUTINE
END EVALUATE

Pascals CASE. Perhaps because the CASE statement is inherently complex, some languages contain
variants of it that can be considered to be faulty. That is, they fail to include some important part
of the general structure.
Experience has shown that a CASE conditional structure needs to include an OTHERS clause
which is to be executed if none of the case labels matches the test value. For example, a very
common application for a CASE statement is to handle character data. In such situations, it is often
desirable to provide special handling for a few special characters, and one or two routines to be
executed for most of the other possible characters. The potential efficiency of a CASE statement
makes it attractive in implementing this situation. With an OTHERS clause, this structure requires
writing out case labels only for cases that require special handling. But without an OTHERS clause,
case labels would have to be written and tested for all 128 possible ASCII characters. This is messy
and cumbersome, and most programmers will end up using a series of half a dozen IFs instead of
writing a CASE statement with 128 case labels.
Standard Pascal fails to provide an OTHERS clause for its CASE statement. It has the logical
structure shown in Exhibit 10.18. Because it lacks this clause, the Pascal CASE statement is rarely
appropriate for real situations. Many Pascal implementations are extended to contain an OTHERS
clause. Unfortunately, there is no single standard way to add such an extension, and the extensions
often have slight syntactic differences, making a program containing a CASE nonportable.
If the value of the test expression is not included among the case labels, it is an error according
to the ISO Pascal standard. (That is, it is a violation of strict Pascal semantics.) In this situation
the person who builds a translator has several choices:
Detect the error at compile time and give an error comment.
Detect the error at run time and halt execution.
Make clear in the documentation that the error will not be detected.
Handling a case error in the third way makes it legal (by the standard) to implement one sensible
default: if no case label matches, do nothing. Occasionally this is even what the programmer wants

288

CHAPTER 10. CONTROL STRUCTURES

Exhibit 10.18. The Pascal CASE with no default clause.

Evaluate the test expression.
Compare result to the case labels.
GOTO the matching label.
Case Labels 1

Actions 1

Case Labels 2

Actions 2

...
Case Labels n

Actions n

to do. But failing to detect a language error is just not a good way to get around the defects in
the language. Ideally, this defect will be corrected in the next Pascal standard.
Cs switch Statement. The switch statement in C does have an OTHERS clause, marked by the
keyword default. However, C omits a different part of the general CASE control structure: the
exit into the frame after completion of the CASE actions. The general CASE is like a series of simple
conditionals, each with an accompanying scope. The C switch is not. It is like a computed GOTO,
where all target statements are within the scope of the switch statement. Restricting the targets
to the scope of the switch is a great improvement over the semantics of a computed GOTO, since
it enforces locality of effects and lexical coherence. Nonetheless, the C switch is less than a true
CASE structure.
The switch differs from the general CASE structure because once C goes to a case label and
completes its corresponding actions, it keeps on going, doing, in turn, all the actions for the
succeeding cases! [See Exhibit10.19, first column.] This is rarely, if ever, what the programmer
wants. The switch lacks the frame on the right which is part of the CASE structure.
The C switch would be unusable except for the existence of the break statement, which can be
used to attach a frame to any C control structure. Executing a break statement sends control into
the frame and thus immediately terminates the switch statement in which it is embedded. Thus
the break is like a structured GOTO out of the control structure.
C forces the programmer to explicitly write the break which should be an integral part of the
CASE control structure. This does provide some flexibility. The programmer may choose to leave
the control structure (break) or continue with the actions for the other cases. But this is a case of
too much, error prone flexibility. It is a very rare situation in which one would want to continue
with all the rest of the case actions.
One can only guess at the reasoning of the C designers when they defined the switch this

289

10.3. ITERATION

Exhibit 10.19. The C switch statement without and with break.

switch( <text expression> ) {
Evaluate the test expression.
Compare result to case labels.
GOTO the matching label.
case <label 1>: <statement-sequence>
case <label 2>: <statement-sequence>
...

switch( <text expression> ) {

Evaluate the test expression.
Compare result to case labels.
GOTO the matching label.
case <label 1> <statement-sequence>
break;
case <label 2> <statement-sequence>
break:
...
case <label n> <statement-sequence>
break;
default:
Actions for other values

case <label n>: <statement-sequence>

default:

Actions for other values

}

way. Since the break exists, all the internal mechanisms must be included in the translator to
implement the general CASE structure. The reasoning that the combination of switch and break is
more flexible than the general CASE is specious, since an added control command dont-break
could just as well be combined with the CASE to produce the same flexibility as switch plus break,
and be less error prone in the common case. Perhaps this strangeness is simply an accident of
history, and a reflection of the fact that C is a twenty-year-old language. In any case, by the
principle of Too Much Flexibility, we can say that it is a design error.

10.3

Iteration

Iterative control structures are used to process homogeneous aggregate data structures such as
arrays, lists, and files. In procedural languages, iteration is accomplished through explicit loop
statements which do not return values although they may modify variables or arrays in memory.
In functional languages, iteration is done through functions in which repetition is implicit, such
as the LISP map functions. These often return lists that were lengthened by one element on each
iteration.
Diagramming Loops. Iterative control structures are one-in/one-out units, and the corresponding control diagrams still must have a single entry at the top and a single exit somewhere along
the bottom. The nonsequential transfers of control in a structured loop are diagrammed by using
a control frame around parts of the box. Control enters the frame by falling through from the
bottom of the scope, then continues in a clockwise direction, to the left across the bottom, and

290

CHAPTER 10. CONTROL STRUCTURES

Exhibit 10.20. The simplest loop: an infinite loop.

Primitive Loop
Label:
Scope

Structured Loop
Begin

Repeat

upward on the left. Control reenters the scope along a horizontal arrow, near the upper left, that
leads into some enclosed box section. Control leaves the loop structure at the bottom along a
vertical exit arrow.

10.3.1

The Infinite Loop

The most primitive loop is formed by placing a statement label at the top of the desired loop scope
and writing a branch to that label at the bottom of the scope. Its diagram is, therefore, formed
by joining the diagrams of a labeled sequence of statements to the diagram of a GOTO [Exhibit
10.20, left]. Far from being useless, infinite loops are now used extensively to implement processes
that interact by using the I/O interrupt system. This kind of infinite loop, though, formed by an
unconstrained GOTO, has been replaced in modern languages by the structured infinite loop [Exhibit
10.20, right].
In the structured form, the loop scope is delimited by keywords (Begin and Repeat are used
here), and the branch statement is generated by the translator. As shown in Exhibit 10.20, this is
a one-in/zero-out structure! The loop scope is framed on the bottom and left, and the only exit
from the frame is back into the scope at the top. When used with an exit statement, the exit adds
a frame and exit arrow on the right. The result is equivalent to the diagram in Exhibit 10.25.

10.3.2

Conditional Loops

Two different simple conditional loops can be formed with a single conditional GOTO. The first form
is the same as the primitive infinite loop except that the GOTO at the bottom of the scope is a
conditional GOTO [Exhibit 10.21, left]. This is frequently called a REPEAT LOOP.
In the second form, usually called a WHILE LOOP, the conditional branch is placed at the top
of the scope [Exhibit 10.21, right]. This form is more frequently useful. Loops are used to process
files, lists, and arrays, and it is usually necessary to prepare for the possibility of empty files and
zero-length data structures.
Corresponding to the two kinds of conditional GOTO loops are the two structured loops shown
in Exhibit 10.22. Their translations into machine code will be the same, but the structured form
automatically generates a semantically correct branch instruction. In the structured loop, the

291

10.3. ITERATION

Exhibit 10.21. The simplest finite loops, formed with a GOTO.

Repeat Until loop

Label:
Scope
IF
condition
False

Do While loop
False

While
condition
True

True
Scope
Label:

programmer marks the beginning and end of the loop scope with the exit test and a keyword or
special character. The compiler then generates the GOTO that forms the loop.
A WHILE loop has the exit test at the top of the scope, a REPEAT loop at the bottom. The WHILE
loop was one of the control structures recommended by Dijkstra to be used in eliminating GOTOs.
Structured loops are easier and faster to translate than the GOTO versions, 11 and they are less error
prone in the hands of any programmer.
The REPEAT is not a very useful loop form because the scope is always executed at least once.
In almost all situations some case can arise in which the scope should be skipped. Often these cases
correspond to faulty input data. The WHILE makes it easier to check for these errors. For example,
the REPEAT loop in Exhibit 10.23 will malfunction if the input, n, is less than 1.
With either of these loop forms, the loop test may have a positive or a negative sense; that is, it
may loop on TRUE and exit on FALSE, or loop on FALSE and exit on TRUE. The former type is called
a While test and the latter type is called an Until test. This is a detail that must be checked when
11

See Chapter 11, Section 11.1.

Exhibit 10.22. Structured conditional loops.

Repeat loop

While loop

Repeat

While

True

Condition

Scope

F
a
Scope
l
s
End While e
True

While condition
False

292

CHAPTER 10. CONTROL STRUCTURES

Exhibit 10.23. REPEAT and WHILE loops in Pascal.

These loops both print out the first n elements of the array a. The repeat loop will malfunction
unless n >= 1.
Repeat loop:

While loop:

Readln(n);
i:= 1;
Repeat
writeln(i, a[i]);
i := i + 1
Until i=n;

Readln(n);
i:= 1;
While i<=n Do Begin
writeln(i, a[i]);
i := i + 1
End;

learning a new language. Exhibit 10.23 shows samples of the REPEAT and WHILE loops in Pascal.
Note that the REPEAT loop ends with an Until test.
The WHILE loop is an appropriate basis for an iterative expression as well as for an iterative
statement. It is built into LISPs family of map functions and into iterative constructs in modern
functional languages such as Miranda. This topic is taken up more fully in Section 10.4.

10.3.3

The General Loop

Restricting loops to a single exit at either the top or the bottom of the scope is artificial. It leads
to awkward programming practices such as the priming reada read instruction given to bring
in the first line of data before entering the main processing loop [Exhibit 11.5]. The main loop then
must end with a duplicate of that read instruction. There is no reason why a loop should not have
its exit test at any arbitrary place within its scope, as in the FORTH program in Exhibit 10.24.
Let us define the general loop [Exhibit 10.25] to be a structured loop that merges the two common
loop forms (REPEAT and WHILE) into a single more flexible looping form. The loop termination test
in the general loop is not constrained to occur at the beginning or end of the loop. The beginning
and end of the scope will be bracketed by the keywords LOOP and END LOOP, and statements of the
scope may come both before and after the termination test. The termination test could be either a
While test or an Until test, or the language might provide both forms, giving the programmer an
option.
This very nice loop form is supported in FORTH, Turing, and Ada [Exhibit 10.26]. It is likely
to become more widely used since REPEAT and WHILE loops are simply special cases of this general
structure, and language translation is simplified by reducing the number of special cases. Providing
the extra flexibility costs little or nothing.

293

10.3. ITERATION

Exhibit 10.24. The general loop in FORTH 83.

Assume getn has been defined to be a function that reads an integer from the keyboard and
leaves it on the stack.
VARIABLE sum
: read&add
( Print a prompt and read a value. )
( If it is positive, sum it and repeat, else quit. )
0 sum !
( Initialize sum to zero. )
BEGIN
." Please enter a number and type a CR."
getn 0 >
( True if input is a positive number. )
WHILE
( Exit loop if prior expression is not True. )
sum @ + sum ! ( Fetch sum, add input, store back in sum. )
REPEAT
sum @ .
( Fetch value of sum and print it. )
;
( End of function definition. )

10.3.4

Counted Loops

Finally, we can add some mechanism to any of the loop forms (REPEAT, WHILE, or LOOP) to automate
sequential processing of a data structure whose length is known. We call the result a counted loop
because the number of iterations is fixed at the beginning of the loop, and the loop exits when the
iteration count reaches that number. Counted loops differ from general loops because the number
of repetitions of a general loop may depend on the result of computations within the loop.
There are nearly as many variations in the details of the counted loop as there are languages
that support it. In general, though, a variable, called the loop variable, must somehow be defined.
It is initialized at the top of the loop, incremented at the bottom of the loop, and tested by a loop
test. Within the scope of the loop, the value of the loop variable is available and is often used as

Exhibit 10.25. The general loop.

Loop
Scope A
While test
Scope B
End Loop

True

False

294

CHAPTER 10. CONTROL STRUCTURES

Exhibit 10.26. The general loop in Ada.

This function does the same thing as the FORTH function in Exhibit 10.24. Assume that
write_prompt is a programmer-defined procedure that writes a message to the screen, and
get_number is a function that reads an integer from the keyboard and returns it.
FUNCTION read_add RETURN integer IS
n: integer;
sum: integer := 0
-- Initialize sum to zero.
LOOP
write_prompt("Please enter a number and type a CR.");
n:= get_number ;
EXIT WHEN n<=0 ;
sum := sum + n;
END LOOP;
return sum;
END read_add;

a subscript. When the value of the loop variable exceeds the preset limit, the loop is terminated.
An additional clean-up step may or may not be done at loop exit time.
Exhibit 10.27 gives a diagram of a typical counted loop with the loop test at the top, like a
WHILE loop. This form is implemented in several languages, including Pascal, COBOL, FORTH, and
BASIC.12
The nature and implementation of the loop variable has evolved through the years. FORTRAN
and COBOL had only static storage, and loop variables were ordinary global variables. PL/1, an
old language, permits the loop variable to be declared in any scope surrounding the loop. As the
years passed, the special nature of the loop variable was better recognized, and various special
restrictions were applied to it.
Pascal, for example, requires that the loop variable be declared locally in the block that contains
the counted loop. Some languages restrict the use of the loop variable. Pascal and all versions of
FORTRAN say that the programmer must not assign a value to the loop variable within the loop.
This is done for efficiency. During execution of the loop, the loop variable can be temporarily
implemented by a machine register, instead of or in addition to an ordinary memory location. This
makes it very fast to test, increment, and use the loop variable. The typical loop refers to its loop
variable many times. Thus many machine cycles are saved by not fetching and storing this value
every time it is referenced or updated.
FORTH predeclares a local identifier, I, to name the loop variable. (If two loops are nested, I is
the index for the inner one and J for the outer one.) The loop variable is kept on top of the return
12

The for loop in C is not a counted loop by our definition; it is dealt with in the next section.

295

10.3. ITERATION

Exhibit 10.27. The counted loop.

Syntax in various languages
FORTH:
10 0 DO hscopei LOOP
Pascal:
For J:=1 to 10 Do Begin
hstatement sequencei
End;
BASIC:
FOR J=1 TO 10
hsequence of linesi
NEXT J

For
Initialize counter.
Does counter meet or exceed goal?
False
Scope of loop.
Increment counter.

T
r
u
e

End For

COBOL:
PERFORM hparagraph-namei
VARYING J FROM 1 BY 1
UNTIL J > 10.
stack, rather than in an ordinary storage object, and it is deallocated automatically at loop exit.
Two newer languages, Ada and Turing, go further in this direction. They open a local declaration
scope at the beginning of the counted loop and automatically declare the loop variable name as a
local variable in that scope. The loop variable is used within the loop and deallocated when the
loop terminates. This arrangement has one real advantage: a loop variable, by its nature, has only
local meaning, and that meaning is obvious from the context. (This is reflected in the fact that
programmers still like to use single-letter names for loop variables.) Declaring the loop variable as
a local variable within the loop increases the modularity of the program and minimizes chances of
accidental error.
A question now arises about the meaning of the loop variable after exit from the loop. If the
loop variable was local to the loop, the answer is clear and easy to implement: since the loop
variable was deallocated at loop exit time, its value is undefined. That means that it is completely
implementation-dependent, and the programmer should never attempt to use the loop variable
after a loop exit. This makes it easy for a compiler writer to implement the languageshe or he
can use a register for the loop variable, to gain execution efficiency, and does not need to store the
final value back into a memory location.
In most languages the loop variable is not local to the loop. However, there is still a question
about its meaning after loop exit. Does it contain its original value? Its value on the final time
through the loop? A value that is one greater than the goal? In FORTRAN II, the loop variable
was kept in a machine index register and its value was officially undefined after loop exit. The ISO
Pascal standard also states that the value is undefined after loop exit.

296

CHAPTER 10. CONTROL STRUCTURES

This design decision has been called a victory of religious fervor over common sense. 13 It is
easy to define in a formal semantic definition and easy to implement efficiently. However, the programmer often needs to know the value of the index after an abortive loop exit. If that information
is not available, the programmer is forced to introduce yet another variable and assign the index
value to it inside the loop. This is surely neither efficient nor desirable.
FORTRAN 77 illustrates a more practical design. The language standard defines what the index
value will be after an exit. The index value may still be stored in a machine register during the
loop, but if so, an epilogue, or clean-up step, must be executed before leaving the loop normally
or by a GOTO. The epilogue will copy the current value of the loop variable back into its ordinary
memory location. The loop variable may, thus, have either of two values after exit. If the loop
exited normally after completing N iterations, the loop variable equals the initial value plus N
times the step size. (This is at least one greater than the goal value). If a GOTO is used to leave the
loop, the loop variable will contain the value current at the time of exit.
The counted loop requires several arguments: the size and sign of the increment, the initial and
terminal values for the loop variable, the beginning and end of the scope to be iterated, and the
exact termination condition. Different languages provide more or less flexibility in specifying these
things and have different rules about the times at which each element is evaluated. Some ways
in which counted loop statements differ from language to language are trivial syntactic details;
others are more major semantic differences that determine the power and flexibility of the control
structure.
FORTRAN II is an example of a language that has differed in a major semantic way from most.
It placed the execution of the scope before the loop test so that every loop was executed at least
once. That is now recognized to be a mistake, and all modern languages place the test at the top,
before the loop scope.
In all languages, control must enter a counted loop through the initialization code (you cannot
GOTO into a counted loop). But there are differences in the rules about what happens at loop entry
and loop exit, and the ways that control is allowed leave a loop. Further, treatment of the loop
variable, increment, and goal expressions can be different.
All languages increment or decrement the loop variable after executing the scope and before
retesting the exit condition. Most of the time, programmers write loops that increment or decrement
the value of a variable by 1, so Pascal and Ada restrict the increment to be either +1 or 1. This
shortens the language syntax a little but is an unnecessary and annoying restriction.
Some languages exit when the loop variable meets the goal, some when the goal is exceeded,
and some permit the programmer to write his or her own test using relational operators. Among
the latter group, some exit when the test expression evaluates to TRUE and others when it evaluates
to FALSE.
If the language allows the programmer to write an expression (rather than a constant) as the
goal, it must also specify when that expression will be evaluated. The original FORTRAN, designed
for efficiency, permitted the goal to be a constant or simple variable name. The value was fetched
13

B. Kernighan, informal comment.

297

10.3. ITERATION

Exhibit 10.28. Pascal computes a tripcount.

Assume i and k have been declared to be local integer variables in the program that contains
this code fragment. It is not clear from looking at this loop whether it will produce three lines
of output or an infinite number. To know the answer you must know that Pascal does compute a
tripcount before entering the loop.
k := 3;
FOR i := 1 TO k DO BEGIN
writeln(i, k);
k := k + 1
END;
writeln(i);

{ A tripcount of three is computed here.

{ Changing k does not change the tripcount.

{ Answer is unpredictable; i is undefined.

}
}

once and loaded into a register. Changing the value of the goal variable within the loop did not
affect loop termination. Modern FORTRAN permits the goal to be specified by an expression, but
the meaning remains the same: it is evaluated only once, before entering the loop. At that time,
the tripcount, or number of loop repetitions, is fixed. Changes to values of variables within the loop
cannot affect the tripcount.
This restriction does not apply to the PL/1 FOR loop. Its goal is specified by an arbitrary
expression that is reevaluated every time around the loop. If the values of some of its variables
change, the loop termination condition changes. This is also true in COBOL.
Three things must be fixed before loop entry in order to determine a tripcount: the initial value
of the loop variable, the goal value, and the increment value. If any one of these can be changed
within the loop, the number of iterations is unpredictable. In Pascal, all three are fixed. The
increment value is always +1 or 1. The language standard states that the loop variable may not
be modified within the loop. The goal expression is evaluated only once, at the top of the loop.
These three together determine the tripcount [Exhibit 10.28].
A small change in the definition of the semantics of a counted loop can radically affect its
behavior. If either the goal value or the increment value can be changed inside the loop, the
tripcount cannot be determined ahead of time. FORTH has two counted loop forms, DO...LOOP
and DO...+LOOP. In both of these, the beginning and ending index values are computed before the
loop. The DO...LOOP has a constant increment value of +1, and so has a fixed tripcount. But the
DO...+LOOP uses whatever value is left on the top of the stack as the increment value, producing
completely unpredictable behavior.
Clearly, in order to understand the semantics of the counted loop in any particular language,
all of these matters must be clarified. The counted loop is thus a very interesting statement type.
While its general intent is obvious, knowing its actual meaning requires knowledge of many small
details. Programmers changing to a language with different details make lots of errors writing
counted loops.

298

CHAPTER 10. CONTROL STRUCTURES

Exhibit 10.29. The iteration element.

Syntax: Iterate (hai;hbi;hci) hscopei;
Iterate
Do all the initializations in <a>.
Evaluate expressions in <b>, in sequence.
Test result of last expression.
F
a
Scope of loop.
True
l
Evaluate all expressions in <c>.
s
e
End Scope

10.3.5

The Iteration Element

The most general form of the FOR loop is called the iteration element. 14 It was developed in 1963 and
was incorporated into the MAD language at that time. Dennis Ritchie, who used MAD at Harvard,
later incorporated it into his own language, C. The iteration element generalizes the semantics of
each part of the FOR statement and is equally implementable as a loop statement and as a loop
expression. It has two sections: a control element and a scope, diagrammed in Exhibit 10.29. The
control element contains:
a. A list of arbitrary expressions to be evaluated before entering the loop (initializations). Control
passes to expression (b) after executing these expressions.
b. An arbitrary expression whose result must be interpretable as a truth value. This is evaluated
before executing the loop scope. Iteration continues until this value is FALSE. At that point
control passes out of the iteration element.
c. A list of expressions to be executed after the scope of the loop (increments). Control passes to
expression (b) after executing these expressions.
The scope begins at the end of the control element. In an iteration statement, the scope contains
a single statement or a compound statement, consisting of a sequence of statements delimited by
begin-scope and end-scope symbols. Executing the scope results in a side effect such as input,
output, or assignment. In an iteration expression, a program object is computed each time the
scope is executed. These program objects are collected into an aggregate (a list or an array) that
becomes the result of the entire iteration expression.
14

Galler and Fischer [1965].

299

10.3. ITERATION

Exhibit 10.30. Uses of the iteration element in C.

/* Read a file of subscripts and print corresponding array elements.
for( ; scanf("%d",&k)!=EOF; )
if (k>=0 && k<=max) printf("%d\n", ar[k]);

*/

/* Read N, then read and sum N numbers. */

for (sum=count=0, scanf("%d",&N); count<N; count++)
{ scanf("%d", &k); sum+=k;}
/* Sum the numbers in a linked list. */
for (current=list, sum=0;
current!=NULL;
sum+=current->value, current=current->next);
The notation current->value on the last line means to dereference the pointer current and select
the value field of the resulting cell. It is equivalent, in C, to (*current).value.

Iteration in C.

The C for statement implements the iteration element. The syntax for for is:

for(hexpressioni; hexpressioni; hexpressioni) hstatementi

Exhibit 10.30 shows examples of the use of the for. For these loops, assume identifiers have
been declared as follows:
A cell has two fields, named value and next.
current and list are pointers to cells.
k and N are integer variables.
ar is an array of integers with subscripts 0. . . max.
scanf stores the data in the address given as an argument and returns an error/success code
which will equal EOF when end-of-file is found.
The biggest difference between the iteration element and ordinary FOR loops is the totally
unrestricted nature of the pieces of the control element. This leads to all sorts of unusual constructs,
like the last loop in Exhibit 10.30 in which all of the work of the loop is done by the control element,
and the scope is empty.
When using a normal FOR statement, there is generally one way to set up the initialization,
goal, increment, and scope to achieve the desired result. When using the iteration element, there
can be several ways to do the same thing [Exhibit 10.31]. This flexibility confuses programmers
who are accustomed to restricted forms of the FOR. Many want to be told what the right way is,
but, of course, there is no single right way. At best we can give rough guidance:

300

CHAPTER 10. CONTROL STRUCTURES

Exhibit 10.31. Many ways to write the same iteration.

These for loops do the same thing as the last loop in Exhibit 10.30. (Note: When two C
expressions are connected by a , they become the syntactic equivalent of a single expression.)
sum=0;
for (current=list; current!=NULL; current=current->next)
sum+=current->value;
/* One action in each slot. */
for (current=list, sum=0; current!=NULL; current=current->next)
sum+=current->value;
/* The usual way to write it in C */
current=list, sum=0;
for ( ; current!=NULL; )
sum+=current->value, current=current->next;
current=list, sum=0;
/* Very bad style. */
for ( ; ; ) if (current==NULL) break;
else sum+=current->value, current=current->next;
The following two loops do the same process and work correctly on nonempty lists. Both will
cause a run-time error if current is the NULL list.
for (current=list, sum=0;
(current=current->next)!=NULL;
sum+=current->value);
/* Ugly but efficient.

*/

current=list, sum=0;
/* A total misuse of the for loop.
for ( ; current=current->next; sum+=current->value )
if (current==NULL) break;

*/

Put all the initializations relevant to the loop in part (a). Dont initialize loop variables or
accumulators earlier, for instance, in the declarations. (This follows from the principle of
Local Effects.) Dont put an action here that is repeated in some other loop part.
Part (b) should contain the loop test. If some action, such as reading data, needs to be done
before the test on every iteration (including the first), put it there also. If you have no loop
test here, use an explicit infinite loop.
Part (c) normally contains the statement that changes the loop variable, in preparation for
the test. If more than one linear data structure is being processed, increment the current
element pointers for all of them here.
The loop body should contain any actions that are left, such as output, function calls, and

10.4. IMPLICIT ITERATION

301

Exhibit 10.32. Equivalent loop forms in C.

while ( hconditioni ) hscopei;
for ( ; hconditioni ; ) hscopei;

control statements.
Because the initialization, test, and increment parts of the iteration element are arbitrary expressions, evaluated every time around the loop, super-efficient implementations are not feasible.
The FORTRAN implementation (where the loop variable is kept in an index register) makes both
incrementation and use as a subscript very fast. This cannot be done if the increment clause is an
arbitrary list of expressions to evaluate.
On the other hand, this iteration element has total flexibility and no other forms of iteration are
necessary. It could serve as the only iterative form in the language. C does have a while statement,
but it is equivalent to a degenerate for loop [Exhibit 10.32].

10.4

Implicit Iteration

Some languages support particularly powerful operations on aggregates of data in which iteration
is implicit. In the modern functional languages this is the only kind of iteration supported, and it
plays an important role. To be useful, an iterative construct in a functional language must return
a value, since using destructive assignment to modify the contents of an array is prohibited and
there is no other way to communicate the results of the iteration.

10.4.1

Iteration on Coherent Objects

Implicit iteration is one of the most striking features of APL. Most operators in APL are defined
for numbers but are applied iteratively to their arguments if the arguments are arrays or matrices.
Operators defined for arrays are iterated if the arguments are matrices. How iteration is performed
on an APL object depends on its number of dimensions and on the operation being iterated.
Compare the first few examples in Exhibit 10.33 to the LISP code in Exhibit 9.28, and compare
the last line to the C code in Exhibit 10.34. (The APL operator = performs comparison, not
assignment.)
When coupled with dynamic storage allocation, automatic implicit iteration is tremendously
powerful. Code that would fill many lines in a language with explicit loops can often be written in
one line of APL. Compare the last expression in Exhibit 10.33 to its (partial) C equivalent in Exhibit
10.34. The difference in length illustrates the power of implicit looping when used to process all
the elements in a data aggregate.

302

CHAPTER 10. CONTROL STRUCTURES

Exhibit 10.33. Use of implicit looping in APL.

5 ANSWER A FUNC B
[1] ANSWER (3 A)- B
5

Define a two-argument function named FUNC.

Return 3 times A minus B.

Bind X and Y to integer vectors.

X 4 9 16 25

Y 1 2 3 4
A 3 by 4 matrix.
Z (3 4) 1 2 3 4 1 4 9 16 1 8 27 64

Expression
X + 1
X + Y
X FUNC Y
(X + 1) = X + Y
Z = 1
+/ Z=W

Value Computed
(5 10 17 26)
(5 11 19 29)
(11 25 45 71)
(1 0 0 0)
((1 0 0 0)(1 0 0 0)(1 0 0 0))
If W is a scalar, the result is a vector containing the number
of occurrences of W in each row of Z. If W is the same shape as
Z, the result is a vector whose length is the number of rows
in Z, containing the number of positions in each row where
corresponding elements of W and Z match.

Exhibit 10.34. Explicit looping in ANSI C takes more writing.

int j, k, count[3];
int w;
/* Assume w is initialized before entering loops. */
int z[3][4] = {{1, 2, 3, 4}, {1, 4, 9, 16}, {1, 8, 27, 64}};
Given these declarations, the double loop and printf statement below do more or less the same
thing as the final line of APL code in Exhibit 10.33.
for(j=0; j<3; j++)
{
for (k=count[j]=0; k<4; k++)
if ( z[j][k] == w ) count[j]++;
}
printf("%d %d %d\n", count[0], count[1], count[2] );

10.4. IMPLICIT ITERATION

303

Exhibit 10.35. Implicit looping in dBMAN.

Assume the current record type has fields named lastname, sex, hours, grosspay, and
withheld, and that the current file contains records of this type for each employee. Statistics
can be gathered on subsets of the file selected by simple statements which access each record in the
file and test or process it. Examples of dBMAN commands follow:
1. display all for lastname = 0 Jones0 and sex = 0 M0
2. copy to lowpay.dat all for grosspay < 10000
3. sum grosspay, withheld to grosstot, wheldtot all for hours > 0
Line 1 finds and displays all male employees named Jones. Line 2 creates a new file that contains
copies of some of the records in the current file. Line 3 sums the current gross pay and withholding
fields for all employees who have logged some working hours.

A more important benefit of implicit looping is its flexibility. Any size or shape of arrays can
be accommodated, so long as every operator is called with arguments whose shapes make sense
when combined. APL defines iteration patterns for scalar/scalar, scalar/vector, scalar/matrix,
vector/vector, vector/matrix, and so on through higher and higher dimensional objects. The APL
code that counts the occurrences of W in Z (+/ Z=W ) will work on any shape matrix, and it also will
provide a meaningful answer if W is a simple integer. Different C code would have to be written for
those cases.
The primary data aggregate type supported by functional languages is the list, and the implicit
iterative functions in these languages process lists. In APL the data aggregate type is the array,
and operators are iterated over arrays. A data base language such as dBASE also makes extensive
use of implied iteration to process all data elements of its basic aggregate type, which is file.
dBMAN, a dBASE look-alike for the Atari computer, supports many operations that test or
manipulate each record in a file, in turn, possibly producing a new file as the result. Examples of
such operations are display and sum [Exhibit 10.35].
The term fourth-generation language, sometimes applied to data base languages, recognizes
the great ease with which large volumes of data can be manipulated using implicitly iterative
statements.

10.4.2

Backtracking

A few languages exist that not only support implicit iteration, but combine it with a trial-and-error
backtracking algorithm to enable the programmer to write pattern matching expressions. Prolog
will search a data base of relations and extract items, one at a time, that match a pattern specified
by the programmer. SNOBOL and its successor, ICON, search text strings for the first section that

304

CHAPTER 10. CONTROL STRUCTURES

Exhibit 10.36. A data base of Prolog relationships.

1.
2.
3.
4.
5.
6.

likes(Sara,John)
likes(Jana,Mike)
likes(Mary,Dave)
likes(Beth,Sean)
likes(Mike,Jana)
likes(Dave,Mary)

7. likes(John,Jana)
8. likes(Sean,Mary)
9. does(Mike,skating)
10. does(Jana,swimming)
11. does(Sara,skating)
12. does(Dave,skating)

13.
14.
15.
16.

does(John,skating)
does(Sean,swimming)
does(Mary,skating)
does(Beth,swimming)

matches a specified pattern (these patterns can be very complex).

In all three languages the semantics of the search command requires that all possibilities must
be examined before returning a FAIL result. In some cases this can be done only by making a trial
match, then continuing the matching process. If the process later fails, then the system must back
up to the point it made the trial match, discard that possibility, and search for a different trial
match. This process is called backtracking. The implementation of backtracking uses a stack to
store pointers to the positions at which each trial match was made. It is a well-understood recursive
process but nontrivial to write and potentially very slow to execute.
We can illustrate this process by a Prolog example. Prolog maintains a data base of facts and
relations. It searches this data base in response to queries, looking for a set of facts that match
the users requirements. Each query is one or a series of patterns to be matched, in order. As each
pattern is matched, any variables in the pattern become bound to the corresponding values in the
fact, and a marker is left pointing to that fact. If some later pattern cannot be matched, Prolog
will come back to this marker and try to find another match. All searches start at the top of the
data base and move sequentially through it. If the end of the data base is reached and a matching
fact is not found, Prolog returns the answer no.
Exhibit 10.36 contains a brief data base about a group of teenagers, their current crushes, and
their favored activities. The numbers listed are not part of the data base but are for the purpose of
tracing the pattern matching operation. The reader should keep in mind that this data base and
the query we will trace are both very simple. Prolog can express far more complex relationships
with multiple, nested dependencies.
The owner of this data base is looking for a compatible couple to share an afternoon of sport
with himself and his date. Exhibit 10.37 gives his query, which contains four goals to be satisfied
together. When and if a match is found, Prolog will tell him the girl, the guy, and the activity he
wants. The chart below the query traces the steps Prolog would follow in trying to match all the
patterns.
String processing and artificial intelligence application programs commonly need to perform
exhaustive searches. Thus the implicit iteration plus backtracking provided in ICON and Prolog
make them very appropriate and very powerful languages for these application areas.

305

10.4. IMPLICIT ITERATION

Exhibit 10.37. Answering a Prolog query.

Query:

?- likes(X,Y), likes(Y,X), does(X,Z), does(Y, Z)

Marker Bind Bind Marker Marker
Bind
for
X
Y
for
for
Z
Goal 1
to
to
Goal 2 Goal 3
to
1

Sara

John

Jana

Mike

Marker
for
Goal 4

7? no
5? yes
10

swimming
9? no

no
no
3

Mary

Dave
6? yes
15

skating
12? yes

Answer:

Next Step
Try goal 1.
Try goal 2.
Retry goal 1.
Try goal 2.
Try goal 3.
Try goal 4.
Retry goal 3.
Retry goal 2.
Retry goal 1.
Try goal 2.
Try goal 3.
Try goal 4.
Solution found.

X=Mary, Y=Dave, Z=skating

Exercises
1. What is a primitive control instruction?
2. What is the difference between a jump and a conditional jump?
3. What is the purpose of using subroutines in a program?
4. Why is sequence considered to be a basic control structure in procedural languages but not
in functional ones?
5. What is a one-in-one-out unit? Give a specific example.
6. Of what does a sufficient set of basic control structures consist, according to Dijkstra? Why
are these structures easier to debug than programs that use the unconstrained GOTO?
7. What is the difference between a conditional statement and a conditional expression? Give
an example of each.

306

CHAPTER 10. CONTROL STRUCTURES

8. How are structured conditionals different from simple conditional branch statements?
9. In what two ways are THEN and ELSE clauses delimited in modern procedural languages?
10. What are the advantages of each method discussed in question 9?
11. What is a CASE statement?
12. Why can a CASE statement be implemented more efficiently than a series of conditionals?
13. Why is an infinite loop considered a one-in-zero-out control structure?
14. What are the two simple conditional loops that can be formed? How are they implemented
as structured loops?
15. What is a general loop? How is it different from the other structured loops? What are its
advantages?
16. What is a counted loop? When is it used?
17. What is the loop variable? How has it evolved through the years?
18. What happens to the loop variable upon exit from the loop?
19. What is a tripcount? How do changes to the loop variable affect the tripcount?
20. What is an iteration element? How does it differ from a counted loop? (Both have been
called FOR loops.)
21. To understand exactly how a for loop operates you must know several details about what
happens in what order and under what conditions. Make a list of questions which you must
answer to interpret the meaning of a for.
22. The for loop in Pascal is superficially similar to the for loop in C but has very different
semantics. List three ways in which the Pascal loop is restricted when compared to the C
loop.
23. Diagram all five of the following FORTH loops, and write the keywords in appropriate places
on each diagram.
a. begin...again
b. begin...while...repeat
c. begin...until
d. do...loop
e. do...+loop

10.4. IMPLICIT ITERATION

307

24. What is implicit iteration? What is its purpose?

25. Name two programming languages that support iteration without the use of explicit control
structures such as FOR, WHILE, GOTO, or recursion. What kind of iteration is supported?
26. Why is implicit iteration often combined with backtracking? How is it used?

308

CHAPTER 10. CONTROL STRUCTURES

Chapter 11

Global Control

Overview
This chapter deals with several kinds of nonlocal control structures that have come into
common use. They range in sophistication from the ubiquitous GOTO to continuations,
a form of control transfer based on higher-order functions. Between these extremes are
some very useful and practical constructs: BREAK, labeled loops, and exception handlers.
The GOTO instruction is supported in many languages but completely absent in others.
There are three faults inherent in the GOTO: bad effects on translation, bad effects on
proofs of correctness, and bad human engineering properties. In spite of the fact that
GOTO is out of favor, most procedural languages provide a GOTO instruction because
sometimes it is the most efficient and practical way to implement a control pattern.
In the late 1960s, there was considerable controversy about using GOTO within a program.
Dijkstra advocated that programs should be limited to a set of three basic control
structures: sequence, conditional, and while loop. Each is a one-in/one-out structure
making it possible to debug a program by debugging a series of shorter sections of code.
Non-local transfers of control can be important when a programmer needs to abort
execution of a loop or recover from an error. In some languages, this can only be
done by using a GOTO. In more modern procedural languages, the break instruction and
exception handlers have been introduced to meet this need. In functional languages,
non-local transfers of control can be accomplished by using continuations.

309

310

11.1

CHAPTER 11. GLOBAL CONTROL

The GOTO Problem

GOTO is ancient, simple, and efficient; it mirrors the primitive jump instruction built into all machine
hardware. To understand why we need modern forms of control transfer, the student must first
understand what is wrong with the GOTO.
When the GOTO was first introduced into higher-level languages, it was thought to be both basic
and necessary. It was considered basic because it directly reflected the branch instructions of all
computers. It was considered necessary because it was used to compensate for the nonexistence
of an important semantic mechanism that was not well understood until the late 1960s. Early
languages such as assembly language, APL, FORTRAN, FORTRAN II, COBOL, and BASIC did not
embody the abstraction scope. Labeled statements and/or the GOTO were used in place of beginscope and end-scope markers. APL, FORTRAN II, and BASIC provided no other way to delimit
the required scopes. COBOL was different. It had a more complex system with a murky kind of
semantics that caused or permitted a tremendous variety of subtle errors.
Since then, there has been considerable argument about the wisdom of using GOTO in a program,
or even including GOTO in a programming language. More foolish assertions have been made about
this issue than about almost any other programming language feature. Use of any kind of GOTO
instruction does cause some problems. We need to examine how this can lead to trouble, why some
programmers want to use such a flawed control structure, and whether the need could be eliminated
by better language design.

11.1.1

Faults Inherent in GOTO

We can divide the faults of the GOTO control structure roughly into three categories: bad effects on
translation, bad effects on proofs of correctness, and bad human engineering properties.
Bad Effects on the Translation Process. GOTOs are harder to translate than structured control
statements. Backward and forward branches require different techniques. A backward branch is
used to form loops. To translate it, the location of each label in the object code of a program must
be remembered until the end of the code has been reached, since any instruction at any distance
could branch back to it. Thus the global symbol table becomes cluttered if many labels are used.
Forward branches are used to implement IF control structures. One conditional branch is
needed at the location of the test, and an unconditional branch is needed to skip around the ELSE
clause, if it exists. A label is needed just after the last line in the scope of a conditional. These
statements are true whether applied to the code a programmer would write in BASIC or the code
produced by a translator for a Pascal IF. But a forward branch cannot be translated when it is first
encountered, since the location of the target label in the object code is not known at that time.
In the case of a forward GOTO, the location of the forward branch itself must be remembered,
and later, when a label is found, the list of incomplete forward branches must be searched and any
branch that refers to that label must be patched up and deleted from the list. Again, in a program
with a lot of labels, this is slow and inefficient. Translation of the forward branches in an IF is

11.1. THE GOTO PROBLEM

311

easier and much more efficient. The locations of the forward branches can be kept on a stack. Each
ELSE and ENDIF can patch up the branch on the top of the stack. (An ELSE also places an item on
the stack, to be patched by the ENDIF.) No searching is necessary, and since exactly one end-scope
exists for each begin-scope, the stack can be popped when the top item is used.
Correctness is Hard to Prove with GOTOs. The bad mathematical properties of GOTO are
all related to modularity. Correctness proofs start by breaking the program into one-in/one-out
sections and proving that whenever some correctness property is true at the beginning of a section,
it is also true at the end. Such proofs become much easier when the size of each section is small.
But a program cannot be divided between a GOTO and its target label. Thus the more that GOTOs
are used, and the more tangled they become, the harder it is to prove that the program does its
intended job.
Some pragmatists are not deterred by this reasoning. They believe that mathematical proofs
of correctness are seldom, if ever, useful, and that proofs are not used in real situations. Unfortunately, the same properties that make a program a bad candidate for a mathematical proof make
it hard for a human to understand and debug. Proof and debugging have similar goals and face
similar problems.
Some of the GOTOs might be data-dependent conditional branches. With these, the flow of
control cannot be predicted by looking at the program, and the number of possible control paths
that must be checked is doubled by each IF. With many such branches we get a combinatorial
explosion of possible control paths, and complete debugging of a long program becomes highly
unlikely.
The best policy for a programmer who must use GOTO because the chosen language lacks an
essential control structure is to use GOTO only to emulate properly nested, one-in/one-out control
structures. All forward and backward branches are kept short. This takes self-discipline and does
not make translation any easier, but it avoids the debugging problems.
Bad Human Factors Inherent in the Best Use of GOTOs. When scopes are defined by labels
and GOTO there is often poor visual correspondence between the true structure of the program and
its apparent structure, represented by the order of lines on the page. FORTRAN II and BASIC
had only an IF...GO, not a full IF...THEN...ELSE structure. The latter can be emulated using
GOTO, but this has the effect of putting the ELSE clause under the IF (where control goes when the
condition is false) and putting the THEN clause at some remote program location, often the end.
The end of such a program tends to be a series of stray clauses from a bunch of unrelated IFs.
Things that belong together are not placed together in the source code, and the code has poor
lexical coherence [Exhibit 10.9, the IF in BASIC].
Writing such code is error prone because the programmer will often write an IF and simply
forget to write the code for the remote section. Debugging such code is also more difficult because
it requires lots of paging back and forth. Patching or modifying code after it is written usually
introduces even more remote sections. Eventually the flow of control can become simply too

312

CHAPTER 11. GLOBAL CONTROL

confusing to follow.
Why Spaghetti is Unpalatable. The most difficult situation for either proof or debugging is
one in which the scopes of different IFs or loops overlap partially, in a non-well-nested manner,
and unconditional GOTOs thread in and out through these merged scopes. In this situation there
may be a very large number of ways that control could reach a particular statement. This kind of
program is called spaghetti code because of its tangled nature. Spaghetti code is easy to write but
tricky to debug. The outstanding characteristic of spaghetti code is that virtually everything has
global effects.
If one simply sits down and writes out lines of code as they occur to the conscious mind, adding
afterthoughts, corrections, and improvements as needed, the result will probably be spaghetti code,
even if a structured language is used. The lack of clarity is caused by too many IF statements asking redundant questions about obscurely named global variables. Disorganized thinking produces
disorganized code, and disorganized code is hard to understand and hard to debug.
The bugs in most programs can be eliminated by cleaning up the code. The truth of this has
been demonstrated over and over by beginning students, coming to a professor with programs
whose bugs they cannot locate, let alone fix. The professor can look at such a program, fail to see
the error right away, and suggest that a few messy parts be cleaned up and rewritten, keeping all
routines short and all variables local. Having no other recourse, the student follows these orders.
The student comes back an hour later with a working program. The bug never was found, but it
disappeared when good hygiene was applied to the program. The irony is that the programmer
accomplished in an hour of clean-up what had not been achieved in a whole evening of debugging
through patches and modifications.
The performance of many spaghetti programs can be improved by cleaning them up. For small
jobs and quick answers, the stream-of-consciousness programming method is usable and may even
be fastest in the long run. For large or production jobs, it is disastrous. The resulting program
is usually loose: it is longer, more complex, and less efficient than a well-thought-out, wellstructured program.
A spaghetti program has a short useful lifetime and poor portability. In a spaghetti program
with ill-defined scopes and global variables, it is difficult to define or understand how the parts of
the program interact. It is, therefore, difficult to modify or extend the program in even minor ways
without introducing errors into parts that formerly worked.
It is for these reasons that the business and defense communities have given their full backing
to structured system design and structured programming techniques. Any program is expensive
to develop, and one with a short lifetime is too expensive. Most managers do not understand
programming, and they are at the mercy of their staff unless they can establish some sort of
criteria for the right way to do things. To this end, they often require or forbid the use of certain
language features, require a certain style of programming, and require extensive documentation of
every step of the program development process.
Rigid rules about methodology and style often do stand in the way of modernization. New

11.1. THE GOTO PROBLEM

313

Exhibit 11.1. Some languages with no GOTO and no labels.

Icon:

Has ordinary WHILE and UNTIL loops and an IF...THEN...ELSE control structure.

Prolog:

Uses the REW cycle and nested functions like LISP.

FORTH: Has structured loops and conditionals only. An assembler is an integral part of a FORTH
system, but even the FORTH assembly language does not provide an arbitrary GOTO; it has only
limited-distance forward and backward jumps.
Turing:

Has a structured conditional and structured loops, with break.

and better languages and methods are not accepted until a new generation of managers takes over.
But firm application of these rules protects a manager against unscrupulous programmers like one
professional who bragged to a class that he had tenure at his company because his program
code was both crucial to operations and undecipherable by anyone else. Perhaps to defend against
programmers like this, one local company has a program that turns spaghetti code written in
COBOL into modern structured COBOL code.

11.1.2

To GOTO or Not to GOTO

Against the GOTO. On the one hand it is amply clear that excessive use of GOTO and highly
unrestricted GOTO instructions make programs error prone, hard to read, hard to debug, and perhaps
impossible to prove correct by mathematical methods. The simple existence of a label in a
program slows down translation. For these reasons, the computing community has largely adopted
Dijkstras position that the use of GOTO is bad, and that structured programming, without GOTO, is
better.
Many of the newer computer languages contain no GOTO at all [Exhibit 11.1]. The omission of
GOTO is particularly interesting in the cases where a new language is basically a revision of an old
one, as Icon is a revision of SNOBOL. Both are general languages, designed by R. Griswold, with
particular adaptations and many functions included to make string processing easy.
SNOBOL, designed in the early 1960s, is built around an unusual one-in/three-out control
structure. Each line of a SNOBOL-4 program invokes a series of function calls, name bindings,
and pattern matches. Depending on the success or failure of this series of actions, processing can
continue at any of three lines [Exhibit 11.2]. Even a short program can become a rats nest of
intertwining GOTOs.
Icon is the modern successor to SNOBOL-4. The string-processing orientation remains, but the
curious GOTO structure is gone. It is replaced by structured control statements, including WHILE,

314

CHAPTER 11. GLOBAL CONTROL

Exhibit 11.2. The unique control structure in SNOBOL-4.

SNOBOL-4 was built on a unique one-in/three-out control pattern. Control flowed in at the left
side of a line. Function calls, pattern matches, and explicit bindings were executed in left-to-right
order. At the right end of the line, the process produced either a success or failure result,
which determined the next line to be executed:

Yes
Series of function
calls, pattern
matches, and Succeed?
explicit bindings
No

Yes
No
No
F label defined?
Yes
S label defined?

S label
Next sequential line
F label

UNTIL, and IF...THEN...ELSE. Icon programs look more like C than like SNOBOL.
In Favor of the GOTO. In spite of the fact that GOTO is currently out of favor, most procedural
languages do provide a GOTO instruction. This is because using a GOTO is sometimes the only efficient
or practical way to implement an essential control pattern. Articles appear regularly in the popular
computing magazines hotly debating the merits of the GOTO instruction and exhibiting programs
that cannot be written efficiently (in languages such as Pascal) without it.
It is important to understand that if the control structures of a language reflect the commonly
occurring process control patterns, then the programmer will be happy to use them instead of GOTO.
Some languages are clearly deficient in control structures. For example, FORTRAN lacks a WHILE
loop and a CASE statement, and BASIC lacks a WHILE loop, an IF...THEN...ELSE conditional, and
procedure parameters. Textbooks for beginning programmers in these languages have been rewritten to preach the use of structured programming, even though adequate one-in/one-out control
structures are not provided by the language.
Authors of textbooks for both languages advocate achieving structure by restricting the
use of labels and GOTO to emulate the standard IF and WHILE control structures in Exhibit 10.2.
Unfortunately, this produces FORTRAN code that is awkward, cluttered, and unmotivated by any
real need. Dijkstras control structures are not the only possible good ones and are not the most
appropriate to superimpose upon a GOTO language.
Pascal contains all of Dijkstras control structures, but these form a very limited tool bag for the
expert programmer. One might suggest that limiting the programmer to WHILE, FOR, REPEAT,
IF...THEN...ELSE, and CASE is like limiting a skilled carpenter to the use of a screwdriver, hammer,
wrench, saw, and a broken chisel. Perhaps the programmer can do the job required, but not in a
very pretty way. These five control structures satisfy programming needs much of the time, but
there are many common situations in which they fall short. Pascal programmers who use GOTO do
so primarily when they wish to implement the following four control structures efficiently:

11.1. THE GOTO PROBLEM

315

An OTHERWISE clause in the case statement.

A loop statement whose exit test can be placed anywhere in the scope [Exhibit 11.4].
A break statement [Exhibit 11.9] to be used in loops, or a two-exit loop, returning a success
or abort result.
A way to terminate a nest of loops all at once after an error [Exhibit 11.10, Ada loop labels].
As long as programmers work with languages such as FORTRAN and Pascal, GOTO will be used to
emulate important missing control structures, and we will continue to hear arguments from naive
programmers about the value and indispensability of GOTO. In a properly designed language, GOTO
has few, if any, uses.
Ada has the control primitives, missing in Pascal, to handle the needs described. Consequently,
there is little demand for the GOTO in Ada. It is not possible to give a list of all the control structures
that might sometime be useful, but we can provide enough structures to satisfy 99.9% of the needs,
and make 99.9% of the uses of GOTO unnecessary. It remains a design issue whether a language should
provide GOTO to allow the programmer to write efficient code for that remaining .1% of situations.

11.1.3

Statement Labels

A discussion of GOTO must cover the target of the GOTOa statement label. Programming languages
have used a variety of methods to define labels, including label declarations, simple use, and merging
labels with source code line numbers.
Line Numbers Used as Labels. Some interactive languages, notably BASIC and APL, took a
shortcut approach to statement labels: the same line numbers that define the sequence of statements
were used as the targets for GOTO. Like all shortcuts, this caused real semantic trouble.
First, it deprives the programmer of any way to distinguish lines that form the top of a program
scope from lines that are simply in the middle of some sequence of commands. 1 A programmer
normally knows which lines of code should and should not become the target of a GOTO. Even a
programmer who chooses to create spaghetti code with tangled scopes knows where each thread
should and should not begin. A language that provides no way to communicate this information is
not offering the programmer a good vehicle for modeling a problem.
Second, as discussed in Chapter 2, Section 2.3.3, line number/labels are a compromise because
the semantics needed for line numbers are different from those needed for labels. If the semantics
defined for them are appropriate for use by the on-line editor, then the semantics are inappropriate
for defining flow-of-control, and vice versa.
1

Discussed more fully in Chapter 2, Section 2.3.3.

316

CHAPTER 11. GLOBAL CONTROL

Exhibit 11.3. The form and definition of labels.

Language
Ada

Form
alphanumeric

APL

alphanumeric

COBOL

alphanumeric

FORTRAN

numeric

LISP

alphanumeric

Pascal

numeric

PL/1

alphanumeric

SNOBOL-4

alphanumeric

Rules for making and using labels

Place << label_name >> before any statement. Use
in the GOTO statement.
Place label followed by : before any statement. Use
in statement.
Place beginning of label in columns 811 (ordinary
statements cannot start before column 12). Use in
the GOTO or PERFORM statements.
Place in columns 15 at the beginning of any executable statement. Use in the GOTO statement.
Labels and go expressions are legal only within
a prog or a do body. The label is defined by
an expression that forms part of the body of the
prog. The expression (go label_name) within a
prog body causes transfer of control to the expression after label_name.
Declare label at top of program and place it, followed
by :, before any statement. Use in GOTO statement.
Place label followed by : before any statement. Use
in GOTO statement.
Place label and : at the left side of any statement.
Use labels by placing their names in the succeed or
fail field of any statement.

Defined Labels. Most languages avoid these problems by having a programmer simply write
a label at the beginning of any line. In Pascal, the label must also be declared at the top of
the program. (This permits Pascal to be compiled easily in one pass through the source code.
Languages without declared labels either require two passes or some complicated patching-up of
the generated code at the end of the first pass.) Exhibit 11.3 summarizes the rules for defining and
using labels in several common languages.
Conclusions. The uses and misuses of GOTO and statement labels have been considered, along
with a variety of ways to minimize the use of GOTO. These include the obvious waysinclusion of
structured loops and conditionals in the language, and ways that have only recently been appreciated, including providing more flexible loop control structures and a structured BREAK statement
which replaces the unstructured GOTO.

11.2. BREAKING OUT

317

Exhibit 11.4. An input loop with an exit in the middle.

A simple interactive read loop is shown here in FORTRAN, demonstrating a common situation
in which the loop exit is placed at a position other than the top or bottom:
100 PRINT (0 Please type your name.0)
READ *, (NAME)
IF (NAME .EQ. 0 Quit0 ) 200
PRINT (0 Hello 0 , Name, 0 Nice to see you.0 )
GOTO 100
200 PRINT (0 Goodbye.0)
Compare this code to the Pascal version in Exhibit 11.5.

11.2

Breaking Out

Many loops have two possible reasons for termination, not one. The basic search loop is a good
example: control will leave the loop after all data items have been searched and the key item is not
found anywhere. This is called a normal exit, or FINISH. On the other hand, if the item is in the
data base, a search loop should terminate as soon as it is found, before examining all the remaining
items. This is called an abortive exit, or ABORT. In a search loop, a normal FINISH is associated
with failure and an ABORT with success.
As another example, an input loop will ABORT if a premature end-of-file is encountered and will
FINISH if the number of input items meets expectations. In this case, the FINISH is associated
with success and the ABORT with failure.
A loop syntax that provides two ways to leave the loop is important. It is applicable to the
large number of situations in which a repeated process can either FINISH or ABORT. To write such
a loop using GOTO is easy. The normal termination condition can be placed in the loop control
element; the ABORT condition is placed in an IF-GO within the loop [Exhibit 11.4].
To accomplish the same thing without using IF-GO is more complex. In the case of an input
loop, this is generally handled by using a priming read [Exhibit 11.5]. The prompt and input
statements are written twice, once just before the loop, guaranteeing that the loop variable will
be initialized, and once at the end of the loop, to prepare for the next test. This is correct and
executes efficiently, but it is clumsy to write.
An inexperienced programmer thinks that he can solve this problem by using an UNTIL loop
instead of a WHILE loop. Unfortunately, this just leads to a different kind of repetition: the loop exit
condition has to be tested within the loop (as well as at the end) in order to avoid processing nondata
when the end-of-file condition occurs [Exhibit 11.6]. This is even less elegant than the priming read,
since it introduces an extra level of nested logic and it causes unnecessary duplication of work on
every iteration.
Sometimes two exit conditions can be combined into a single loop control expression using AND.

318

CHAPTER 11. GLOBAL CONTROL

Exhibit 11.5. The Pascal input loop with a redundant priming read.
Writeln (0 Please type your name.0 );
Readln (Name);
While Name <> 0 Quit0 do begin
Writeln (0 Hello 0 , Name, 0 Nice to see you.0 );
Writeln (0 Please type your name.0 );
Readln (Name);
end;
Writeln (0 Goodbye.0);
Compare this code to the FORTRAN version in Exhibit 11.4.

In other cases this is not possible, either because some action must be done between testing the
conditions, or because the second condition would cause a run-time error whenever the first is true
[Exhibit 11.7].
The WHILE loop in Exhibit 11.7 translates into the following steps:
1. (k<=max) is evaluated (call the result t1).
2. (a[k]>0) is evaluated (call the result t2).
3. (t1 and t2) is evaluated (call the result t3.)
4. If t3 is false the loop exits, otherwise sum = sum + a[k] and k := k + 1 are executed.
5. Control returns to step 1.
This code segment will bomb with a subscript out of range error at the end of summing
any array whose elements are all positive. What happens is this: when (1) is false, (2) is evaluated
anyway, and since k is too large, this causes an error. (In fact, that is why test (1) was included.)
To program such a loop correctly in Pascal one must put condition (1) in the loop control
element and condition (2) in a separate statement inside the loop. A superfluous Boolean variable

Exhibit 11.6. The Pascal input loop with an extra test.

Repeat
Writeln (0 Please type your name.0 );
Readln (Name);
If Name <> 0 Quit0 Then Writeln (0 Hello 0 , Name, 0 Nice to see you.0 );
Until Name = 0 Quit0 ;
Writeln (0 Goodbye.0);

319

11.2. BREAKING OUT

Exhibit 11.7. Some Pascal code that will stop with a run-time error.
Assume k is an integer, and a is an array with subscripts from 1 to max. This code sums the
items in a up to but not including the first negative or zero element.
k := 1; sum := 0;
While ( k<=max ) and ( a[k]>0 ) Do Begin
sum := sum + a[k];
k := k + 1
End;

is declared, and an IF statement within the loop sets it to TRUE if the ABORT condition happens.
Then the test in the loop control element is modified to test this flag also. This will cause an exit
from the loop at the beginning of the iteration after the ABORT condition was discovered. After loop
exit, some other mechanism must be found to determine what condition caused the exit and take
appropriate action. This awkward program idiom is illustrated in Chapter 8, Exhibit 8.13, where
a flag named done is introduced to permit an abortive loop exit.
In more complex situations a Pascal programmer might use several Boolean flags, set in several
places, and combined into a complex expression for the exit condition. This is logical spaghetti
just as surely as if we used many GOTOs to create this control pattern. Our goal is to eliminate the
need for spaghetti in our programs. 2 To do this we need a structured way to exit from the middle
of a loop. Such an exit mechanism is provided by the BREAK statement (or its equivalent) in many
languages. The BREAK is simply a structured GOTO to the end of the control unit. It may be placed
anywhere within a loop, and it transfers control immediately to the statement following the loop.
Its control diagram [Exhibit 11.8] is an immediate exit into the frame of the statement. (A frame
is added if the control structure does not already have one.)
2

See Section 11.1.

Exhibit 11.8. Structured conditional loops with BREAK.

BREAK exits to the right.

A loop with a conditional BREAK.

Loop

Scope A
Break

Scope A
If <condition> Then
Else <Scope B>

Scope B
End Loop

Break

320

CHAPTER 11. GLOBAL CONTROL

Exhibit 11.9. The input loop implemented using break in C.

char *name;
/* This is a string variable. */
for(;;)
/* An explicit endless loop. */
{
printf ("Please type your name. ");
scanf ("%s", name);
if (name[1] == \0) break; /* Exit if no name was entered. */
printf ("Hello %s, Nice to see you.\n", name);
}
printf ("Goodbye.");

C3 and the newer procedural languages, Ada and Turing, all include some form of the BREAK
statement for aborting a loop in a structured manner. In Ada and Turing it is called EXIT. This
is shown on a control diagram as a frame along the right side of the box which is connected to
the bottom frame, from which control leaves the entire diagram. Control enters this frame along a
horizontal arrow wherever a break occurs in the program.
Exhibit 11.9 gives one final version of our print routine, written in C using break, which causes
immediate exit from a loop. By putting a break instruction inside an if, and embedding the whole
thing in an infinite loop, we can achieve the same thing as in the original FORTRAN example,
simply and efficiently.

11.2.1

Generalizing the BREAK

In addition to statement labels, for use with the GOTO instruction, Ada also has loop labels, which
are used to create double-exit loops and to allow immediate, one-step exit from a nest of loops.
The simple BREAK statement takes control immediately out of the enclosing control structure, and
into the control scope that surrounds it. The Ada EXIT is a generalization of this idea; it can cause
an immediate exit from an entire nest of loops (or any part of a nest) by naming the loop label of
the outermost loop. It can be used for any kind of loop (FOR, WHILE, or infinite) 4 [Exhibit 11.10].
We will call this control structure the FAR_EXIT. The canonical example of an application for the
FAR_EXIT is a search loop over a two-dimensional data structure such as a matrix or a list of lists.
A double loop is used to process the structure, and when the key item is found, control must leave
both loops.
Implementing this exit actually requires more than a simple GOTO. When an Ada FOR loop is
entered, stack space is allocated for the loop variable. This obviously must be deallocated upon exit
3

C also has a continue statement, which can be used inside any loop. It transfers control immediately to the top
of the loop and begins the next iteration.
4
Ada is the only major language that lets you name a loop scope so that you can exit from it.

11.3. CONTINUATIONS

321

Exhibit 11.10. Loop labels in Ada.

In the following code, a dictionary is represented as a linked list of words, and each word is
represented as a linked list of characters. The double loop searches a dictionary for the first word
containing the character 0 #0 . Upon loop exit, word_pointer and char_pointer either point to the
required word and the 0 #0 , or to null, if no such word was found. These variables must be defined
outside the outer_loop.
outer_loop:

WHILE dictionary_pointer /= null

word_pointer := dictionary_pointer.word;
char_pointer := word_pointer;
WHILE char_pointer /= null
IF char_pointer.letter = 0 #0
THEN EXIT outer_loop;
ELSE char_pointer := char_pointer.next
END IF;
END LOOP;
END LOOP outer_loop;
-- Control comes here from the EXIT above.

from the loop. A FAR_EXIT can be placed at any nesting depth. When executed, it will terminate
all loops in the nest, up to and including the labeled loop, and any other loops that were placed
within that outer loop. The obvious implementation of this action is to scan backwards through
the stack, deallocating everything up to and including the storage for the labeled loop. This kind
of sudden exit can be quite useful for error handling but must be accompanied by some way to pass
information outward, to the calling program, about the reason for the sudden exit. Variables that
are global to the outer loop are used for this purpose in Ada.

11.3

Continuations

All the nonlocal control structures discussed so far have been defined in terms of sequences of statements. Functional languages do not have such sequences but must supply some way to accomplish
the same ends in a functionally acceptable manner. Programmers who use functional languages
still must write search loops and make sudden error-exits. The continuation is a functional object
that meets this need.
Briefly, a continuation is a function which acts like the remainder of the program which
is

still to be executed. For example, suppose a programmer writes a program to compute 2 and
then print out the answer. We can think of this program as having two parts:
1. The job of the square root function is to compute a real number which, when multiplied with

322

CHAPTER 11. GLOBAL CONTROL

Exhibit 11.11. Packaging a continuation in Scheme.

(define (makecc)
(let ((x (call-with-current-continuation
(lambda (cc) (set! gcc cc) ()) )))
(display "returned value: ")
(display x) ))

itself, equals 2.
2. The print routine waits to receive a number from the square root function, at which time it
prints the number it receives on the users screen.
Step 2the part of the program that will make use of the answer provided by part 1is called
the continuation of the program. You can think of the term continuation just as a fancy name
for the rest of the program still to be executed.
The concept of a continuation exists in all programming languages and is of concern to the
language implementor, but only languages that support higher-order functions are able to give
the programmer explicit access to the continuation. This is because the continuation acts like a
function. It waits to receive a value and then performs an action or computation which depends on
that value. Manipulating a continuation is like manipulating a function. If functions are already
first-class objects, then continuations can be treated like any other function. If not, then special
language primitives would be required. It is not surprising that most languages elect to omit such
facilities.
Even in a language that supports higher-order functions, a primitive is needed to package
up the current continuation in a form that the program can use. Packaging the continuation is
like establishing a checkpointeverything must be saved that is needed in order to safely restart
execution at a later time. This typically includes the program counter, the stack, the environment,
and so forth. Saving the current continuation can be quite an expensive operation indeed!
We will use Scheme syntax to illustrate the formation, semantics, and use of continuations. In
Scheme, this primitive is called call-with-current-continuation, which is often shortened to call/cc.
It is a higher-order function which takes another function, f , as its single argument. It creates a oneargument function, cc, from the current continuation and then calls f with argument cc. Function
f is an arbitrary program, usually written by the programmer, which may or may not make use
of the function cc. If cc is ever called, the current environment is discarded and replaced with the
one saved when cc was created, and the argument to cc is passed back as the value returned by
call/cc.
In Exhibit 11.11, we show how to use call/cc to store the current continuation in a global
variable, gcc. The argument given to call/cc is the nameless function defined by (lambda
(cc) (set! gcc cc) ()). call/cc applies this lambda expression to the current continuation,

11.3. CONTINUATIONS

323

which results in the local symbol cc (in the body of the lambda expression) being bound to the
current continuation. This lambda expression doesnt do muchit just assigns its argument to
a global variable named gcc, which exports the continuation into the surrounding environment.
Finally, the lambda expression returns a null list to call/cc, which passes the value on to its caller,
the let expression, which binds the symbol x to the return value, in this case, the null list. The
display statements cause the comment returned value and the value of x to be printed on the
terminal.
After returning from call/cc, the name gcc can be used, any time, to call the saved continuation. If the user types in the command (gcc 3), she or he will see the response returned
value: 3. The reason this happens is that gcc returns control to the previous control point
where call/cc was about to return with the null list, but this time call/cc returns 3 instead.
Thus x gets bound to 3. Execution now continues as before, printing the comment and the current
value of x.
This example just demonstrates how a continuation may be formed and called; it does not do
anything very useful. However, the call/cc is a powerful control structure which can be used to
emulate other control structures, including WHILE loops and BREAK. This ability to break out of
a piece of code is the most common use for call/cc. If one thinks of execution over time, the
continuation is created when you execute the call/ccCALL/CC, just before f is called, and the
continuation consists of the code that will be executed after f returns. The execution of f itself
is not included in the continuation; thus calling the continuation from within f has the effect of
causing control to jump ahead to the point it would be if f were allowed to finish normally.
In the Scheme code in Exhibit 11.12, the function named search implements the search procedure by making a continuation which is used to break out of the search process when the search
key is found in the data list. Search creates a continuation, binds it to the symbol cc, then calls
scan with cc as an argument. Scan performs a recursive search and uses its continuation-argument
(which it calls foundf) to break out when the key is found. Scan returns the rest of the list starting
with the key.
The continuation requires an argument, so when it is called by scan, the answer found by scan
is passed to the continuation, which in turn passes it on to the calling environment. A transcript
of a test of search is shown in Exhibit 11.13. You can see that calling the continuation really does
abort the recursive search.
In the search example, the continuation was used to do a forward jump, skipping over work
that would otherwise have been done. However, there is nothing in the definition of a continuation
that restricts its use to jumping forward in time. It can also be used to return to a previously
visited point of the computation. Suppose, in the above example, that cc(w) were called, not while
executing f but while executing cc itself. This seems circular, and indeed it is! This causes cc to
start over again, but this second execution is given argument w, whereas the first execution might
well have been with some other argument. (If not, the program will be in an infinite loop!)
The next two examples show how call/cc can be used to emulate WHILE. Although Scheme has
an iterative control structure, it is not a pure functional construct. Call/cc can be used instead,
in order to make Scheme truly behave like a pure functional language.

324

CHAPTER 11. GLOBAL CONTROL

Exhibit 11.12. Call/cc used in Scheme to break out of a loop.

; Search a list, break when found. Returns tail of list, starting
; with the atom equal to the key.
(define (search key ls)
(call-with-current-continuation
; label break point
(lambda (cc) (scan key ls cc)))) ; calling cc causes break
; Perform actual search. Called like search, but "foundf" is a
; functional argument that is applied to the part of the list
; starting with the first occurrence of the search key.
(define (scan key ls foundf)
(if (null? ls)
; if end of list
()
; return failure token
(begin
; else
(display "examining element ")
; document progress
(display (car ls))
(newline)
(if (eq? key (car ls))
; test list head
(foundf ls)
; break out if found
(scan key (cdr ls) foundf))))) ; else recursively scan tail
A transcript of a test run is shown in Exhibit 11.13

Exhibit 11.14 shows a simple counting loop, implemented using a continuation. At the top of
the routine, a continuation is formed and bound to the name cc. The continuation marks the
spot to return to later. (This is analogous to what a compiler must do to translate the beginning of
a WHILE loop.) The result of call/cc is a list containing this continuation and a number. The code
analogous to the body of the loop starts with let; the first thing it does is to give names to the
items in the list that was returned by call/cc; the continuation is named gcc and the number is
named n. Now, this number is tested. If it exceeds the limit, a ! is displayed and execution ends
because there is nothing more to do. If n is less than the limit, it is displayed and the continuation
is called (with n incremented) to continue the looping process. Test results are shown at the bottom
of the exhibit; the command (count 5) does, indeed, count to 5.
Exhibit 11.15 shows another function: one that computes factorial. Compare this to the Pascal
version of factorial in Exhibit 11.16.
When used in this manner, the continuation becomes like a WHILE statement: it returns repeatedly to a fixed starting point, each time bringing with it a value that can be tested and can
potentially stop the repetitions. Each time the continuation is called, the new context replaces the
old context, just as the new value of the WHILE loop variable replaces the old value. Here is a wolf

11.3. CONTINUATIONS

325

Exhibit 11.13. Testing the break routine.

This is a transcript of a test run of the code from Exhibit 11.12, translated by MIT Scheme. The
characters 1 ]=> are Schemes prompt.
; Define some test data
1 ]=> (define boys (tom dick harry bob ted))
1 ]=> (search dick boys)
examining element tom
examining element dick
;Value:

(dick harry bob ted)

1 ]=> (search john boys)

examining element tom
examining element dick
examining element harry
examining element bob
examining element ted
;Value:

()

Exhibit 11.14. Using call/cc to make a loop.

; (count n) prints the numbers 1...n ! on a line.
(define (count limit)
(let ((x (call-with-current-continuation
(lambda (cc) (cons cc 1)))))
(let ((gcc (car x))
(n (cdr x)))
(cond
((> n limit) (display "!"))
((begin
(display n)
(display " ")
(gcc (cons gcc (1+ n)))))))))
1 ]=> (count 5)
1 2 3 4 5 !
;No value

326

CHAPTER 11. GLOBAL CONTROL

Exhibit 11.15. A Pascal-like implementation of factorial, using a continuation.

(define (fact n)
(let ((x (call-with-current-continuation
(lambda (cc) (list cc 1 1)))))
(let ((gcc (car x))
(k
(cadr x))
(p
(caddr x)))
(cond
((> k n) p)
((gcc (list gcc (1+ k) (* k p))))))))
1 ]=> (fact 5)
;Value: 120

in sheeps clothing: destructive assignment disguised as a higher-order function in a form that is

politically correct for a functional language.
This kind of backing-up behavior is appropriate in several real-life situations. The Prolog
interpreter backs up whenever the tentative instantiation it has been trying fails. A transaction
processing system backs up if the transaction it is in the middle of processing is forced to abort.
An operating system backs up the I/O system when a disk read error occurs and tries the read
again.

Exhibit 11.16. The factorial function in Pascal.

function fact (n: integer) : integer;
var
temp: integer;
i: integer;
begin
temp := 1;
for i := 1 to n do temp := temp * i;
fact := temp
end;

11.4. EXCEPTION PROCESSING

11.4

Exception Processing

11.4.1

What Is an Exception?

327

An exception is a situation that makes it impossible for a program to proceed with normal processing. It is an unpredictable situation with which the program must cope. Exceptions happen
frequently enough so that a robust system cannot ignore them, but seldom enough so that they are
unusual. There are three major ways in which exceptions arise:
1. A hardware error occurs and triggers a hardware interrupt signal. Examples of hardware
errors are highly varied and well known to all programmers. These include: an attempt
to divide by zero, numeric overflow during an arithmetic operation, an attempt to access
protected memory or a nonexistent address, a disk-read error, I/O device not ready, and
others.
2. A system software module identifies an error situation. Examples include: a nonnumeric
character entered in a numeric input field, an attempt to use a subscript outside of defined
array bounds, an attempt to store an out-of-range value in a subrange-type variable, an
attempt to open a file that does not exist, an attempt to write a file on a device that has no
available space, an end-of-input file, and so on.
3. A user function identifies a logically impossible or inconsistent situation. Examples, of course,
are specific to the application. Some might be: a request for processing on a data item that
is not in the data base, illegal data value in input stream, and the like.
A program that ignores exceptions invites disaster. For example, arithmetic overflow can be
ignored in COBOL, and will be, unless the programmer explicitly tests for it. The result is that
the COBOL program continues its normal computations using a meaningless bit pattern that looks
like data and is likely to print out a meaningless answer. Ignoring an end-of-file condition in a C
program will generally result in an infinite loop processing the last data item repeatedly. The worst
possible thing to do about an exception is pretend that nothing is wrong and keep processing.
A second possible response to an exception is to terminate the program immediately and return
control to the system. This action is what the Pascal standard requires as a response to a subrange
error, subscript error, or arithmetic overflow. While this may be an acceptable way to manage
faulty student programs, abrupt termination is not acceptable for a production system. At the
very least, relevant information about the state of the system and the cause of the error should be
made available.
In the old days, a sudden termination of this sort would usually be followed by an octal or
hexadecimal dump. The programmer could then manually decode and trace the action of the
program and perhaps find the cause of the sudden termination. A more modern response, typical
in Pascal systems, is the call traceback, which dumps the chain of dynamic links from the runtime stack. Unfortunately, a dump is a primitive tool for tracing errors, and a call traceback

328

CHAPTER 11. GLOBAL CONTROL

offers too little information. Further, even if the programmer can decode these communications,
the end user, who must handle the error, generally cannot.
Robust systems must be able to recover from errors. Ignoring exceptions and sudden death
are both unacceptable alternatives. Exceptions must be identified, if possible, and their effects
controlled and limited as much as possible. Often, if the exception is handled at the right level,
corrective action can be taken and the program can finish normally. For example, if a user of an
interactive program types an illegal input, it should be identified, and the user should have an
opportunity to retype the data. Sometimes the length of the code to handle exceptions exceeds the
length of the code to process normal data.

11.4.2

The Steps in Exception Handling

There are four steps in processing an exception condition:

Detect the exception.

Pass control to the unit that is able to take action.

Handle the exception by taking constructive action.

Continue with normal execution after correcting the problem.

The last two steps are straightforward and can be done with ordinary program code, so long as
the language provides support for the first two steps. Unfortunately, support for steps 1 and 2 is
spotty and limited in most current languages.
Detecting Hardware and Software Exceptions. Computer hardware detects many kinds of
errors automatically and generates hardware interrupts to signal that an exception has happened.
An interrupt is processed by the operating system which generally sets a status flag in response.
These status flags can be read by application programs at run time, but many languages do not
provide language primitives to do so. Thus the warning provided by the hardware and the operating
system is simply ignored.
Even in languages as early as COBOL, the importance of error containment was recognized.
Arithmetic commands could contain an optional ON SIZE ERROR clause which provided a way to
send control to an error routine. After processing the error, control could return to the next step
and the program could continue. PL/1 was another early language with some support for exception
handling. It has defined names for system-generated hardware and software exceptions. The user
can define handlers for these exceptions using an ON CONDITION clause. In case of an error, control
would come to a user-defined routine, giving the programmer an opportunity to print appropriate
error comments. For most errors, however, the programmer cannot prevent program termination
after printing comments.
One of the glaring omissions in standard Pascal is any way to test hardware status codes.
Arithmetic overflow, division by zero, and the like cannot be either detected or handled within the
user program. Because error detection and recovery is an essential part of a production program,

11.4. EXCEPTION PROCESSING

329

most commercial Pascal implementations extend the language by introducing predicates which test
the system status flags.
Some languages use error codes, returned by system subroutines, to provide the programmer
with information about exceptions. This is a common way to handle I/O errors. For example,
in the standard C library, input, output, file-open, and dynamic memory-allocate commands all
return error codes for exceptional conditions. The users program may ignore these codes, which is
risky, or take action.
Finally, some languages provide a general exception-handling control structure, along with a
list of predefined exception names. The programmer may test for these names in the program, and
the control structure delivers control to user-defined exception-handling code. In Section 11.4.3 we
examine this exception facility in Ada.
Passing Control. The user often knows what can and should be done to recover from or correct
an exception error. However, a good program is written as a series of modules; the exception may
be discovered in a low-level module that is called from many places. That module often cannot take
constructive corrective action because it does not have access to important data objects. Intelligent
processing can be done only in a context where the meaning of the data objects is known and the
severity of the exception can be evaluated. For example, in an interactive program, an invalid
input exception may be detected in a deeply nested subroutine, but only the main interactive
module can interact with the user.
Knowledge of the exception must be passed back from the routine that discovered the exception
to the point at which it can be handled gracefully. There are two ways that control can be transferred in a language with no special exception control mechanism: by using GOTO or a continuation,
or by propagating an error code back up a chain of nested subroutine returns.
GOTO is not a good solution. Although it passes control, it cannot pass information about
the exception with that control. Any such information has to be passed through global variables.
Further, most languages have restrictions about the target of a GOTO; for example, control cannot
pass into the middle of a loop or a conditional, and it can only go to a label that is visible in the
scope of the GOTO. The result is that exception handlers end up being defined at the top level, so
that they are accessible. After taking action, it is then difficult to resume normal operation.
In a functional language, continuations can be used to handle exceptions in much the same way
as a GOTO can be used.
Passing an error code back up a chain of calls can be done in any language, and the return
value can encode information about the exact nature of the exception. We call this process error
propagation. Control can be passed backwards to the right level for handling the problem. One
difficulty of this approach is that the error handling code must be intermixed with the code for
normal processing, making the logic for both cases harder to write and comprehend. A more
serious difficulty is that the error propagation code must be present in every subroutine in the
chain between the detecting module and the handling module. These intermediate routines have
no interest in the error; they did not cause it and dont know how to handle it, yet they have to

330

CHAPTER 11. GLOBAL CONTROL

Exhibit 11.17. I/O with error checking in C.

void silly()
{
int age, code;
char * name;
FILE * infile, outfile;
if ( !(infile = fopen("myinput", "r"))
{ fprintf(stderr, "Cannot open input file.\n"); exit(1);}
if ( !(outfile = fopen("myout", "w"))
{ fprintf(stderr, "Cannot open output file.\n"); exit(1);}
if ( fscanf(infile, "%s%d", name, &age) <2)
fprintf(stderr, "Bad data record ignored\n");
if (fprintf(outfile, "Hi %s; I hear you are %d years old!\n", name, age)
== EOF)
{ fprintf(stderr, "Cannot write to file myout!\n"); exit(1
}

The system call exit(1) causes the program to abort gracefully, flushing all buffers and closing all
open files.

pass it on. This greatly clutters the code and obscures its logic. The code itself is very error prone.
For example, a well-written C program that uses the standard I/O library will check the code
returned by every I/O statement and take corrective action if the result differs from the expected
result. In Exhibits 11.17 and 11.18, two versions of an I/O routine are givena robust version and a
clear version. The error-trapping code makes the robust version much longer and more obscure; the
actual I/O operations are buried in a mass of exception code. This demonstrates the importance
of an exception control structure in a modern language.

Exhibit 11.18. I/O without error checking in C.

void silly()
{
int age, code;
char * name;
FILE * infile, outfile;
infile = fopen("myinput", "r");
outfile = fopen("myout", "w");
fscanf(infile, "%s%d", name, &age);
fprintf(outfile, "Hi %s, I hear you are %d years old!\n", name, age);
}

331

11.4. EXCEPTION PROCESSING

11.4.3

Exception Handling in Ada

The facility for exception handling in Ada provides a real contrast to the do-it-yourself versions
we have discussed. Ada provides excellent support for detection of both hardware and software
generated exceptions. It provides reasonable, clean solutions to the problems of passing control,
handling exceptions, and restarting normal processing.
The exception facility includes four kinds of statements:
A declaration form for user-defined exception names.

A command to signal that an exception has happened.

A syntactic word, EXCEPTION, to separate the regular program code from the exception
handlers.

A syntax for defining a procedure to handle one kind of exception.

Exhibit 11.19 shows samples of each kind of exception statement. In the following discussion, we
examine the usage of each part of the exception system.
User-defined Exceptions. Descriptions of the Ada language and the standard library packages
list the names and semantics of the exceptions that the Ada system might generate. In addition,
programmers can define their own exceptions, which will be handled the same way as those defined
by the system. A simple declaration defines a name as an exception:
hexception namei :

EXCEPTION;

To signal a user-defined exception, we use the RAISE statement:

RAISE hexception namei;
Exception Handlers. An exception handler is a body of code labeled by the name of an exception. A program can have handlers for user-defined exceptions and/or for system-defined exceptions.
A programmer writing code that must be reliable needs to define handlers for any of those exceptions that might be raised by that job. Any block of code (BEGIN...END) can have local exception
handlers, which are declared at the bottom of the block. The syntax for defining an exception
handler is:
BEGIN
hcode for processing normal casei
EXCEPTION
WHEN hexception namei =>
haction for that exceptioni
WHEN ...
WHEN OTHERS =>
END

332

CHAPTER 11. GLOBAL CONTROL

Exhibit 11.19. Handling an integer overflow exception in Ada.

The simple program Circles accepts interactive input and calls procedure One_Circle to process
it. One_Circle calls an output routine, Print_Circle, which is omitted here for brevity. If
an arithmetic overflow occurs during execution of One_Circle, normal processing of that data is
aborted: Print_Circle is not called, and the normal prompt is not printed. Instead, the user is
prompted to reenter the data, and processing proceeds with the new data.
WITH Text_IO; USE Text_IO; PACKAGE Int_IO IS NEW Integer_IO(integer);
-- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -PROCEDURE Circles IS
Diameter: Integer;
BEGIN
Put ("Please enter an integer diameter.");
LOOP
BEGIN
Int_IO.Get(Diameter);
EXIT WHEN Diameter<=0;
One_Circle(Diameter);
Put ("Please enter another diameter (0 to quit): ");
EXCEPTION
WHEN Diameter_Too_Big_Error =>
Put("Invalid input--diameter too large. Please reenter:
END
END LOOP;
Put ("Processing finished.");
Put New_line;
END Circles;
-- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -PROCEDURE One_Circle (Diam: Integer) IS
Radius, Circumference, Area : Integer;
BEGIN
Circumference := Diam * 22 / 7;
Radius := Diam / 2;
Area := Radius * Radius * 22 / 7;
Print_Circle(Diameter, Radius, Circumference, Area);
EXCEPTION
WHEN Numeric_Error =>
RAISE Diameter_Too_Big_Error;
END One_Circle;

");

11.4. EXCEPTION PROCESSING

333

The exception portion of the block can contain as many WHEN clauses as needed, and their order
does not matter. (These clauses must be mutually exclusive, though; one exception name cannot
be used to trigger two clauses.) A program that wishes to be sure to catch all exceptions, even
unexpected ones, can include a handler labeled WHEN OTHERS =>. In the Circles example, both
the main program and the subprogram have one exception handler.
Passing Control and Resuming Normal Processing. The code of a handler is translated in
the context of its enclosing block; that is, variable bindings and visibility are the same as for the
rest of the code in the block. The block BEGIN...END also defines the scope of the abort action
associated with each exception: all code is skipped from the point at which the exception is raised
until the end of the block, and the exception handler is executed in its place. If an exception is
raised during execution of a subroutine call, all statments in the block after the call will be skipped.
Raising an exception starts an abort action. If there is no local handler for the exception,
the current block is aborted and its frame is removed from the run-time stack. The exception
is then raised in the routine that called the aborted routine. If that routine has a handler, it
is executed. Otherwise, it is aborted and the exception is passed backward another level. This
aborting continues, removing one stack frame each time, until a stack frame is found that contains
a handler for the exception. The handler is executed and stops the popping. If a program has
no handler for an exception, all frames will be popped, the main procedure will be aborted, and
control will return to the system.
Raising and Propagating Exceptions. In procedure One_Circle, the computation of Area
is likely to overflow with relatively small inputs. 5 The code to check for this overflow is not
placed immediately after the exception-prone statement; rather, it is separated and put into the
EXCEPTION section of the block. Thus the flow of control for normal processing is clear and uninterrupted. If arithmetic overflow does happen, the system will raise the Numeric_Error exception
flag and abort the block from the point of * 22. Control will pass directly to the statement RAISE
Diameter_Too_Big_Error, then leave the enclosing block, which, in this case, is the body of the
One_Circle procedure.
The command RAISE Diameter_Too_Big_Error renames the exception condition, an action
which seems pointless in a short example. However, this can be an important way to communicate
information. In a large program, there might be many situations in which a Numeric_Error could
be generated, and many different potential responses. Renaming the exception provides more
information to the receiver about the cause of the problem and the context in which it arose and
makes it possible to take more specific corrective action.
The act of raising the new exception sends control back to procedure Circles with the information that the diameter was too large. This procedure takes corrective action by asking the user
to reenter the data. At this point, the exception handling is complete. The exception has been
5

On our system, integers are 2 bytes, and overflow occurs for diameter = 205.)

334

CHAPTER 11. GLOBAL CONTROL

caught, and it is not passed up to the next level. The problem has no further repercussions;
normal processing is resumed.

Exercises
1. What is spaghetti code? What causes it?
2. What are the three faults inherent in the GOTO control structure?
3. Why was GOTO initially used in the development of programming languages?
4. Name two programming languages in which any line of the program may be the target of a
GOTO, whether or not that line has a declared statement label.
5. Name two programming languages that do not support GOTO at all.
6. What are the current arguments against using GOTOs? For using GOTOs?
7. What is a statement label? Why is it necessary?
8. What are the two ways to leave a loop? Give a specific example of each.
9. Explain the difference between BREAK and GOTO. Why is BREAK a superior control structure?
10. Explain why Ada has loop labels in addition to an EXIT statement.
11. Explain why labels in FORTRAN are more semantically sound than labels in BASIC.
12. Briefly, what is a continuation?
13. Why are continuations interesting?
14. Describe two applications of continuations.
15. How do exceptions arise? What is the result of ignoring them?
16. Compare exception handling in Pascal and in C.
17. Explain why exceptions are connected to nonlocal transfers of control.
18. Compare I/O error handling in C and in Ada. Comment on lexical coherence and human
engineering.
19. Compare an exception name declaration in Ada to a label declaration in Pascal. Comment
on purpose and likely implementation.
20. Why is an exception control structure superior to a GOTO for containing and recovering from
errors?

Part III

Application Modeling

335

Chapter 12

Functional Languages

Overview
This chapter presents the goals and attractions of functional programming. Functional
languages have developed as semantically clean ways to denote objects, functions, and
function application.
The concept of denotation, as opposed to computation, is introduced. Functional constructs denote objects, in contrast to procedural constructs which specify methods for
constructing objects.
Development of functional languages started with lambda calculus, which was extended
to denote types, literals, and operations for primitive data domains. Lazy evaluation
and the absence of destructive assignment were also carried over from lambda calculus
into the modern functional languages.
Functional languages have a minimum of explicit control structures. Explicit loops are
replaced by tail recursion and implicit iterative constructs. Sequences of statements can
be replaced by lazy evaluation, nested context blocks, or nested procedure calls because
there is no destructive assignment.
Miranda is presented as an example of a functional language because of its generality, power, and simple, attractive syntax. List comprehensions and infinite lists are
explained, and applications are given.

337

338

CHAPTER 12. FUNCTIONAL LANGUAGES

12.1

Denotation versus Computation

During the 1960s, several researchers began work on proving things about programs. Efforts were
made to prove that:
A program was correct.
Two programs with different code computed the same answers when given the same inputs.
One program was faster than another.
A given program would always terminate.
While these are abstract goals, they are all, really, the same as the practical goal of getting the
program debugged.
Several difficult problems emerged from this work. One was the problem of specification: before
one can prove that a program is correct, one must specify the meaning of correct, formally and
unambiguously. Formal systems for specifying the meaning of a program were developed, and they
looked suspiciously like programming languages.
Researchers began to analyze why it is often harder to prove things about programs written in
traditional languages than it is to prove theorems about mathematics. Two aspects of traditional
languages emerged as sources of trouble because they are very difficult to model in a mathematical
system: mutability and sequencing.
Mutability. Mathematics is a pure denotational system; a symbol, once defined in a given context, means one thing. The truth or falsity of a mathematical statement is a permanent property
of the statement; it does not change over time. But programming languages have variables, whose
contents can be changed at any time by a remote part of the program. Thus a mathematical proof
about a program that uses assignment is necessarily nonmodular and must consider all parts of the
program at once.
Programming language functions are not like mathematical functions because of the mutability
of variables. In mathematics, a function is a mapping from one set of values (the domain) onto
another (the range). To use a function you supply an argument from the domain, and the function
returns an element from the range. If the function is used twice with the same argument, the
element returned will be the same. We can thus say that a mathematical function f denotes a set
of ordered pairs (x, f (x)).
It is easy to write a program that does not correspond to any mathematical function. Exhibit
12.1 shows a definition of a C function that, if called twice with the same argument, returns different
answers. This function increments a static local variable each time it is called, and returns the sum
of the argument and the value of that variable. Although this code uses a static variable, which is
not supported by many languages, the same function could be written in almost any language by
using a global variable instead.

339

12.1. DENOTATION VERSUS COMPUTATION

Exhibit 12.1. A nonfunctional function in C.

int my_func(int x)
{
static int z = 0;
return x+(z++);
}

/* Load-time initialization. */
/* Value of z increases after each call.

*/

Sequencing. When a mathematician develops a theorem, she or he defines symbols, then writes
down facts that relate those symbols. The order of those facts is unimportant, so long as all the
symbols used are defined, and it certainly does not matter where each fact is written on the paper!
A proof is a static thingits parts just are, they do not act.
In contrast, a program is a description of a dynamic process. It prescribes a sequence of actions
to be taken, not a collection of facts. Sometimes the order of a pair of program statements does not
matter, but other times it does. Some statements are in the scope of a loop and are evaluated many
times, each time producing a different result. To fully describe a program during execution (which
we call a process), we need not only the program code, but a copy of the stack, the values of global
and static variables, the current values of the computers program counter and other hardware
registers, and a copy of the input.

12.1.1

Denotation

Functional languages are an attempt to move away from all this complexity and toward a more
mathematical way of specifying computation. In a functional language, each expression of the
language denotes an object. Objects are pure values and may be primitive data elements such
as integers, functions, or higher-order functions. Thus each expression can be thought of as a
description of a static, mathematical entity rather than of a dynamic computation.
Some examples will help make this clear. The expression 3 denotes the integer 3, as do the
expressions 2 + 1 and 5 2. The meanings of these expressions are fixed for all time and do not
change depending on what has happened elsewhere in the program. In functional languages, these
same properties are extended to functions. (Remember that a function is a set of ordered pairs.)
The factorial function, fact(n), takes an integer argument n and returns a number n!, which is
the product of the first n natural numbers. This mathematical function might be described in a
functional way by a pair of recurrence equations, for example,
fact 0 = 1
fact n = n fact (n 1) if n 1
The first equation describes the value of fact for argument 0. The second equation describes
how the value of fact n is related to the value of fact(n 1). It take some effort to show that
there is a unique function (on the natural numbers) that satisfies these equations and hence that

340

CHAPTER 12. FUNCTIONAL LANGUAGES

the equations do in fact specify a function, but it is clear from the form of the equations that
whatever the meaning is, it does not depend on the history of execution up to this point. We say
that this pair of equations denote the mathematical function n!.
Denotational semantics is a way of describing the meaning of a programming language construct
by giving the mathematical object that it denotes. In the above examples, the meaning of 2+1 is the
integer 3, and the meaning of the two equations specifying fact is the factorial function. Contrast
this style of description with the Pascal program for computing factorial shown in Exhibit 11.16.
What this Pascal expression denotes is the process of successively multiplying temp by the
integers 1, 2, 3, . . . , n. In this case, the result of that process happens to be functional, and the
program does indeed compute the factorial function, but it is not obvious without careful inspection
of the code that the program in fact computes a function.
Functions are denoted by -expressions. For example, consider the function G denoted by the
expression
Let G be f.x.( if x = 0 then 1 else x f (x 1)).
G denotes a function of two arguments, f , a function, and x, an integer. Suppose p is the function
p(n) = 3n. Then Gp5 denotes the value of the expression
if 5 = 0 then 1 else 5 p(5 1)
which is equal to 5 p(4) = 5 12 = 60.
Fixed Points. The function G above has a curious property: namely, if h is the mathematical
function factorial(n), then G(h) is also the factorial function. We say that the factorial function is a
fixed point of G because h = G(h). Moreover, it is the least fixed point in the sense that any function
h0 satisfying the equation h0 = G(h0 ) agrees with factorial at all of the points where factorial is
defined (the nonnegative integers). We can use this fact and the least fixed point operator, Y ,
to write an expression that denotes the factorial function, namely Y G, or equivalently,
Y f.x.( if x = 0 then 1 else x f (x 1))
Intuitively, Y f just means, in the following expression, f refers recursively to the function being
defined.
A curiousity of lambda calculus is that Y itself can be denoted by a (rather complicated) lambda
expression. Hence Y does not have to be added to the language as a special primitive for defining
recursive functions. In practice, it almost always is added for efficiency reasons. Rather than
replace Y by the equivalent lambda expression, it is much more efficient in the implementation to
allow pointers to functions; then Y f . . . simply binds f to the function being defined by the x . . .
expression. Lambda calculus is a sufficient semantic basis for expressing all computations because
it is able to express the Y operator as a formula.
The fact that pure lambda calculus lacks types is not an oversight; it is necessary in order to
define Y . The reason is that Y works with any function as its first argument, not just with functions

12.2. THE FUNCTIONAL APPROACH

341

of a specific type.1 Imposing a type structure on lambda calculus, as many functional languages
do, weakens the lambda calculus and makes it necessary to introduce additional constructs to allow
for recursive definitions. Convenience of use also dictates the addition of other constructs. Still, it
is elegant that so simple a system as the untyped lambda calculus could be the basis of a powerful
programming language!

12.2

The Functional Approach

Abstractly a functional language is simply a system of notation for denoting primitive values and
functions. Lambda calculus, which was already proven to be a fully general formal system in which
any computable function could be defined, was used as a notation for early functional languages.
However, lambda calculus has two serious shortcomings as a language for programmers: it lacks
any concept of types and it is hopelessly clumsy to use.
Nevertheless, lambda calculus has been taken as the starting point for the development of most
real functional languages. The earliest practical language in this family was LISP, which now has
two important descendants, Common LISP and Scheme.
Another class of languages has been developed recently, which we will call pure functional languages. These languages follow lambda calculus and functional principles much more closely. This
family includes ML, Miranda, and Haskell. An extended version of lambda calculus was developed
which includes types and operators and literals for the underlying data domains. This system forms
the semantic basis upon which the pure functional languages are built. Functional programs are
parsed and translated into computation trees which represent lambda applications. At execution
time, the lambda evaluator crawls over these trees, binding arguments, evaluating expressions as
needed, and replacing expression trees by constant values wherever possible.
Pure functional languages come extremely close to implementing the semantics that Church
defined for lambda calculus. They share many traits with the older functional languages, specificly, functions are first-class objects, conditional expressions are used, rather than conditional
statements, and repetition is done implicitly or with recursion, not by explicit loops. The key
differences between the newer and the older functional languages are:
The prohibition on destructive assignment.
The complete absence of statement sequences.
The use of lazy parameter evaluation.
We will consider each of these issues and show how denotational constructs can take the place of
procedural language elements.
1

The type of a function is defined by the types of its argument and return value.

342

CHAPTER 12. FUNCTIONAL LANGUAGES

12.2.1

Eliminating Assignment

Destructive assignment causes semantic problems because it creates a time- and sequence-dependent
change in the meaning of a name. Prohibiting destructive assignment does not imply that a
programmer cannot give a name to a value, merely that a name, once bound to a value, must
remain bound to the same value throughout the scope of the name. Thus the meaning of a name
is constant, and not dependent on the order of execution. This greatly simplifies the task of
understanding the program and proving that it is correct.
In procedural languages, assignment is used for several purposes:
1. To store the result of computing a subexpression.
2. In conjunction with conditional statements.
3. With a loop variable, to mark the progress of a loop.
4. To build complex data structures.
To show that assignment can be eliminated, we must show how each of these things can be accomplished, with reasonable efficiency and style, in a functional way.
Programmers store the results of subexpressions goal (1) for many good reasons. When an
expression is complex, a local name can be used to break out and name a portion of it, so that
the structure and meaning of the entire expression can be made clearer. Local names are also used
in Pascal for efficiency; rather than write a common subexpression several times, the subexpression
is evaluated first, and its result is assigned to a local variable. The local name is then used in
further computations. Both uses of local variables can reduce the complexity of expressions and
make a Pascal program easier to understand and debug. It is easy to devise a functional method
to meet this need. We simply need to let the programmer open a local context block in which to
define local names for expressions. Miranda provides three ways to define a symbol:
As a global symbol, by declaring the name with an =, just as symbols are defined in lambda
calculus.
As a parameter name. The name is bound to an argument expression during a function call.
As a local symbol, by using a where clause.
A Miranda where clause defines one or more local names and gives an expression for the meaning
of each. The entire group of definitions forms a local context and can be implemented by a stack
frame. Where clauses can also be nested, producing block structured name-binding.
Goal (2) is satisfied by using a conditional expression, which returns a value, instead of a
conditional statement. The value returned by the conditional expression can be bound to a name
by parameter binding or by a where clause.
Extensive support for list data structures, including list constructors, whole-list operations with
implied iteration, and expressions that denote complex lists all combine to minimize the need for
loops, goal (3). Sequences of data values are represented as lists, and operations that process such

343

12.2. THE FUNCTIONAL APPROACH

Exhibit 12.2. Primitive operations for arrays in APL.

APL (Array Processing Language) was designed to make working with arrays easy. It contains
more kinds of array primitives than other languages, including operations such as generalized crosssection, concatenation, lamination, and whole-array arithmetic. It gives excellent support for all
sorts of array manipulation.
In these examples, the variables named ar and br are 3 by 3 arrays.
Array Primitive
Exhibit a literal array
Fetch dimensions of array
Create and initialize an array
Subscript (random access)
Cross section
Change one array item
Prepare for sequential access
Access current item
Whole-matrix operations
Concatenation on dimension n
Lamination (makes more dimensions)

Example in APL
17 18 19
ar
ar 3 3 0
ar[1;3]
ar[1 2; 2]
ar[1,1] 17
ix 1
ar[ix; iy]
ar + 1
ar ,[1] br
ar ,[.5] br

sequences are written by mapping simple operations over lists. This device eliminates the need
for most explicit loops and many statement sequences. The remaining repetitive processes are
programmed using recursion. These topics are discussed at length in Sections 12.2.2 and 12.3.
Goal (4) is the only real difficulty. Prohibiting destructive assignment means that a data
structure cannot be modified after it is created. For small data structures this is no problem, as
a new copy of the structure can be created that incorporates any desired change. This, in fact, is
how call-by-value works: objects are copied during the function call, and the new copies are bound
to the parameter names within the function. Most list processing and a lot of processing on records
and arrays can be done efficiently this way. The exception is when a large data structure, such as
a 1000 element array, is used as a parameter and slightly modified within the subroutine. Copying
it all in order to modify one element might produce unacceptable inefficiency.
Efficient implementation of arrays in a functional language is a current research topic. Work
centers on the question of whether the original and modified versions of the array are both used.
If not, then the language system can perhaps avoid the copy operation. It can do a destructive
assignment to an array element on the sly and pretend that it didnt. This trick will succeed
if the program never again asks to see the original version of the array. Because of the efficiency
problem with large arrays, existing pure functional languages are all list-oriented.
When (and if) ways are found to circumvent these problems, careful design efforts will be
needed to integrate arrays into functional languages. APL is an array-processing language that

344

CHAPTER 12. FUNCTIONAL LANGUAGES

has many things in common with functional languages and lets the programmer use a functional
programming style to do complex combinations of whole-array operations. To support this style,
APL has a large and complex set of array operations [Exhibit 12.2]. The set of primitive composition
and decomposition operations needed to make arrays usable without constant use of assignment is
large compared to the set provided for list handling in Miranda [Exhibit 12.11]. Todays list-oriented
functional languages are small and simple; tomorrows languages, with arrays added, will be larger
and more complex.

12.2.2

Recursion Can Replace WHILE

In Chapter 10, we discussed the fact that all programs can be written in terms of three control
structures: statement sequence, IF...THEN...ELSE, and the WHILE loop. That is, these three
control structures are sufficient for all programming needs. However, they are not necessary; both
WHILE and sequence can be eliminated. An alternative set of necessary control structures is:
IF...THEN...ELSE
recursive procedure call
Repetition of sections of code is certainly fundamental to programming; problems that do not
involve repetition are usually solved by hand, not by computer. But repetition does not have to be
achieved with an explicit loop. It is possible to implement repetition without a GOTO or any kind
of looping statement, by using tail recursion.
A tail recursion is a recursive call to a function, in such a position within the function that it will
always be the last thing executed before the function return. Frequently, this will be at the lexical
end of that function, but it could also be nested within a control structure [Exhibit 12.3]. Tail
recursion produces the same result as a simple infinite loop or simple backwards GOTO. Combined
with a conditional, tail recursion can emulate a WHILE or FOR loop. The parallel between tail
recursive functions and loops is so straightforward that either form can be mechanically translated
to the other form, either by a programmer or by a compiler.
To use recursion to emulate a loop, one writes the body of the recursive function as an IF
statement that tests the loop-termination condition. The answer returned by the function, if any,
can be returned by the THEN clause, and the body of the loop is written as the scope of the ELSE
clause. (A procedure that returns no value can put the loop body in the THEN clause and eliminate
the ELSE clause.) The loop variable in the loop version is replaced by a parameter in the recursive
version. On each recursive call, the parameter is increased or decreased by the value of the loops
increment expression.
The iterative and recursive mechanisms that implement repetition are quite different. In an
iterative implementation, the contents of a single machine address, associated with the loop variable is changed during each repetition, and execution continues until the value of that variable
reaches some goal. Exhibit 12.4 shows an iterative procedure to print the contents of part of an
array. One storage object is allocated for the parameter count, and the value in that location is
changed once for each item printed out.

345

12.2. THE FUNCTIONAL APPROACH

Exhibit 12.3. Recursive implementation of repetition.

The following procedure prints the values of an integer array. We would call print_out( LL,
N ); to print out the first N items of array LL. The outer procedure is a shell that initializes a
parameter for the inner, recursive procedure, print_out_recursive. The recursive procedure will
be called N times and will create and initialize one copy of the parameter count each time. No copy
of count is ever incremented. After the last item is printed, all N calls on print_out_recursive
return, and all N copies of the parameters are deallocated.
Procedure print_out (list:

list_type; list_size:

integer);

Procedure print_out_recursive (count:integer);

Begin
writeln( list [count] );
if ( count < list_size ) then print_out_recursive (count+1);
End;
Begin
print_out_recursive (1);
End;

In a tail-recursive implementation of repetition, each recursion causes allocation of additional

storage on the stack for the current value(s) of the parameter(s) and for the function return address.
New stack frames are piled on top of old ones. The recursive descent ends when the value of some
parameter reaches a goal; thus this parameter must be different in each successive stack frame.
When the value in the newest stack frame reaches the goal, repetition stops, and all the accumulated
stack locations are freed, one frame at a time. Each parameter value is only used during the cycle
that placed it on the stack. Values in locations below the top of the stack are never used again,
and all are popped off together when the termination condition eventually happens.
A recursive version and an iterative version of a print_out procedure are given in Exhibits
12.3 and 12.4. Both routines are written in Pascal, and the implementations described are those
produced by a typical Pascal compiler. These routines print out the elements of an array starting
at the subscript count. The array, called list, has list_size elements. (To work correctly,
list_size must be >= 1.) Let N be the number of items to be printed; N = list_size count
+1. The appropriate calls for the two inner procedures, to print out all list items, are identical:
print_out_recursive(1);
print_out_iterative(1);
The stack allocations for both ashown in Exhibit 12.5.
Repetition through recursion does use more storage space and more overhead time than repetition through iteration. Stack overflow could limit the number of repetitions. But coupled with
some kind of conditional control structure and some form of sequential execution, recursion forms

346

CHAPTER 12. FUNCTIONAL LANGUAGES

Exhibit 12.4. Iterative implementation of repetition.

The following procedure uses iteration to print the values of an integer array. We would call
print_out( LL, N ); to print out the first N items of array LL. The outer procedure is a shell
that initializes a parameter for the inner, iterative procedure, print_out_iterative. The iterative
procedure will be called once. It creates one copy of the parameter count, which is incremented N
times.
Procedure print_out (list:

list_type; list_size:

integer);

Procedure print_out_iterative (count:integer);

Begin
While (count <= list_size) Do Begin
writeln( list [count] );
count := count + 1
End
End;
Begin
print_out_iterative (1);
End;

Exhibit 12.5. Stack diagrams for a tail recursion and a loop.

Assume that the global variable list is this array of four real numbers:

list

10.0 17.0

0.1

28.9

The stacks for print_out_recursive(1) and print_out_iterative(1) are shown just as the
last item on the list is being processed, but before the function return.

list : 10.0
: 17.0
: 0.1
: 28.9

list : 10.0
: 17.0
: 0.1
: 28.9
count: 4

count: 1
count: 2

Note: this storage object previously

held the values 1, 2, and 3.

count: 3
count: 4

Top of stack

Top of stack

Recursive Implementation

Iterative Implementation

12.2. THE FUNCTIONAL APPROACH

347

an adequate semantic basis for a programming language; explicit loops are not needed. Further,
even when the programmer writes repetition using recursion, a translator can recognize tail recursion and generate iterative, not recursive, code. Translators for many languages descended from
LISP do exactly that.
Whether or not to include iterative control structures in a language thus depends on some value
judgements. The questions are:
Is the iterative syntactic form better, more natural, or more appropriate than the recursive
form for communicating some problems?
Is iterative syntax better enough to warrant including nonnecessary statements in the language syntax?
Will programmers be likely to write better code with the extra iterative forms?
If the answers are yes, iterative control structures should be included in a language. If not,
they are excess baggage. The older LISP-like functional languages included explicit map functions
which implement iteration over lists. Newer functional languages include extensive collections of
list operators that do implied iteration.

12.2.3

Sequences

Traditional languages define the flow of control in terms of sequences of statements (control statements, procedure calls, and assignments). These languages also support nested expressions and
function calls; some even include conditional expressions. The programmer often has a choice
about whether to use heavily nested code, or to break that code into a sequence of statements.
Writing a sequence of statements is often the simplest, clearest, and most direct way to express
an idea. We have found by experience that pulling out the parts of a highly nested expression
and writing them as a separate steps often clarifies the meaning of a program. It also permits the
programmer to do a better job of attaching comments to obscure parts of the code.
Lambda calculus has shown us that statement sequences are really unnecessary in a theoretical
sense. Everything can be done by nesting functions and function calls. However, lists of statements
do provide the programmer with a second tool, in addition to function definitions, for breaking
up code into comprehensible units. Modern functional languages meet the need for separating out
parts of the code without supporting statement sequences at all. In this section, we examine how
this is accomplished.
Nesting Replaces Assignment Sequences. A sequence of assignment statements that does
not assign two different values to the same variable and, therefore, does not destroy information,
can be emulated by a nest of functions with parameter bindings and local symbol definitions taking
the place of the assignments. This is true in traditional languages as well as in functional languages.

348

CHAPTER 12. FUNCTIONAL LANGUAGES

Exhibit 12.6. A sequence of statements in Pascal.

c := a + b;
s := sin(c*c);
write (a, b, c, s);

Exhibit 12.6 shows a simple series of assignment and output statements, written in Pascal. In such
a sequence, each line forms the context in which the next line is interpreted.
To form nested contexts in a functional language, one can use a nest of function calls and/or
block declarations. Exhibit 12.7 shows how the Pascal assignment sequence might be rewritten in
LISP using symbols defined in nested blocks. The first assignment in the Pascal sequence forms the
outermost context block in LISP, creating the context for the rest of the nest. The final write in
Pascal is innermost in the LISP version, so that its symbols are interpreted in the context of all the
surrounding bindings.
Lazy Evaluation Replaces Sequences. In a Pascal program, the programmer must write statements physically in the same order that they will be evaluated. If statement B depends on the result
of statement A, then A must come first. In contrast, when lazy evaluation is used, the sequence in
which expressions are evaluated can be completely different from their order on the written page,
because no expression is evaluated until its result is needed. When we look at this situation one
way, we see that we have no direct control over the order of evaluation. When looked at from the
other side, though, we see that we do not have to worry about that order!
Miranda provides the keyword where to permit the programmer to define and initialize local
symbols. Local names and their defining expressions are written at the end of a Miranda definition,
following where. These expressions are evaluated using lazy evaluation, an important difference
from both Pascal and LISP. This frees the programmer from concern about the order in which the
definitions are written within the where clause. Exhibit 12.8 defines a function called Answers that
is analogous to the Pascal function in Exhibit 12.6. In Answers, the symbols s and c are both
defined, but s, which depends on c, is defined first. This causes no problem. When s is used, its
definition will be evaluated; in the middle of that process, where s refers to c, the expression for c
will be evaluated. The order of the definitions on the page is immaterial.

Exhibit 12.7. A LISP function nest emulating a sequence.

(print (let ((c (+ a b)))
(let ((s (sin (* c c))))
(list a b c s))))

12.2. THE FUNCTIONAL APPROACH

349

Exhibit 12.8. Defining local symbols in Miranda.

This code is more or less like the Pascal and LISP examples in Exhibits 12.6 and 12.7. The function
Answers returns a tuple of values, which will be printed if the function is called from the top level.
Answers a b = (a b c s)
where
s = sin(c * c)
c = a+b

Note, also, that the explicit nesting present in LISP is gone. Miranda does not require each
symbol to be defined before it is used. The set of where clauses, taken together, form a context
for each other and for the function code. This is different from LISP semantics, which require that
each clause be explicitly nested within the context formed by the definitions of all its symbols.
Guarded Expressions. In traditional languages, all statements must be written in one sequence
or another. Frequently, though, the sequence of a short group of statements does not matter, so long
as all are executed before the next part of the program. Some newer procedural languages, such
as Concurrent Pascal2 provide a parallel execution control structure, consisting of scope markers
enclosing a series of statements. The semantics of parallel execution are that the statements in
the scope may be executed in any sequence, or all at the same time, on different processors. No
statement in such a scope computes a value that is referenced anywhere else in the group, and no
two statements compute the same value.
This concept is carried over into the functional languages in the form of guarded expressions
and where clauses. A guarded expression is a generalized conditional control structure in which the
condition clauses are evaluated in parallel. It is very much similar to the generalized conditional in
LISP. However, there is a critically important difference: the clauses of a LISP cond are evaluated
sequentially, and the first true condition triggers evaluation of its corresponding clause. In contrast,
a guarded expression is evaluated in parallel . This does not affect the outcome if only one guard
is true. However, if two (or more) guards are true, either corresponding clause might be evaluated;
a nondeterministic choice is made by the translator. Thus the value of a LISP conditional with
more than one true clause is defined and deterministic because all clauses following the first true
clause are ignored. In contrast, the value of a Miranda expression with two or more true guards is
undefined. (Using such an expression is therefore risky.)
Output Sequences. The use of conditional expressions eliminates the need for conditional statements. Using recursion we can eliminate explicit loops. Using symbol definitions, nested contexts,
and lazy evaluation we can eliminate sequences of assignments. Having done these things, the only
remaining kinds of statements in traditional languages are those related to input and output.
2

Brinch Hansen [1975].

350

CHAPTER 12. FUNCTIONAL LANGUAGES

Exhibit 12.9. A simple Miranda script.

Miranda
sq n = n * n
z = sq x / sq y

Lambda Calculus
sq = n. * n n
z = / (sq x) (sq y)

The premise of lazy evaluation is that no expression is ever evaluated unless its result is needed.
For most computations this works nicely, but it is not well suited to producing output. By definition,
it is the side effect of the output process that is wanted, not the value returned. One device is to
permit the programmer to specify that strict evaluation should be used. With strict evaluation all
arguments are evaluated, even if they would be skipped using the rule for lazy evaluation. This
permits the programmer to cause the evaluation of an expression whose result is not used by the
rest of the program, and whose only effect is output.
Another way to produce an output sequence is to use the sequencing built into data objects in
place of a sequence of statements. The compound data types, lists (sequences of values, all of the
same type) and tuples (nonhomogeneous sequences of values), are both ordered sequences of values.
Both can be constructed dynamically. Since structured objects are first-class objects in functional
languages, any kind of output supported by the language (interactive output or file output) can
also be used for lists and tuples.

12.3

Miranda: A Functional Language

Miranda is a simple and elegant functional language that illustrates both the principles and the
attractiveness of the functional approach. 3 This section deals with the basic elements of Miranda;
several advanced features are discussed in Chapter 18.
A program in Miranda is a collection of definitions, called a script, that define and name objects.
Like lambda calculus, each definition has (on the left) the symbol being defined and (on the right)
several clauses that comprise its definition. An object can be a simple data type, a list, a tuple, or
a function. A very simple script which defines and calls the square function is shown in Exhibit
12.9.
A function call is written by writing the function name followed by one or more arguments,
which can be simple items, lists, or tuples. No parentheses or delimiters of any kind are used to
enclose the arguments. Note the similarity between the definition and calls on the function sq,
shown on the left in Exhibit 12.9, and the corresponding definitions, expressed in lambda calculus,
on the right. Note that Miranda uses infix notation for ordinary arithmetic operators and uses
end-of-line rather than a character such as ; to end the expression.
The primitive types in Miranda are characters (type char, written in single quotes), truth values
(type bool, values True and False), and numbers, which include both integers and reals (type num).
3

Turner [1986].

12.3. MIRANDA: A FUNCTIONAL LANGUAGE

351

The basic data structures in Miranda are the list and the tuple. A list is an ordered sequence of
homogeneously typed values and is written by enclosing the list values in square brackets. A string
is a list of characters, and a literal string can be written as a list of characters, or as characters
between double quote marks. A tuple is a nonhomogeneous sequence of values, and it is written
by enclosing the values in parentheses.

12.3.1

Data Structures

Tuples
Miranda supports two compound data structures, tuples 4 and lists. A tuple is simply a list of objects
whose types do not have to be alike. It could be implemented by a record data type or by a list of
pointers to the members of the tuple. In Miranda we denote a literal tuple by enclosing a series of
values in parentheses. The argument list of a function is also a tuple, although it is written with
no punctuation at all.
Tuples are used in Miranda to define enumerated data types and to form complex data structures,
just as records are used in traditional languages. The methods for defining and accessing tuple types
involve a powerful pattern-matching mechanism. They are dealt with in Chapter 18.
Lists
Miranda lists can be denoted in several ways. The simplest list is a sequence of data items enclosed in
square brackets [Exhibit 12.10]. However, this notation is cumbersome for long lists and impossible
for infinite lists, so Miranda also provides shorthand ways to denote lists. We can denote a list of
consecutive values by giving the first and last value. Thus [1..5] means the same thing as [1,2,3,4,5].
This notation can be extended to denote any arithmetic progression of values. The first two values
are written, followed by .. and the last value. This is enough information to define the list; the
difference between the first value and second value is taken as the step size for the progression. An
infinite list, such as all the numbers greater than 10, is denoted by giving just the first value or
the first two values. A list may also be denoted by a powerful kind of expression known as a list
comprehension, which is discussed below.

12.3.2

Operations and Expressions

The usual arithmetic operators (+, -, *. /, div, mod) are defined with the usual precedence
and are written in infix notation. Miranda has some basic operators for building and decomposing
lists, as well as several powerful operations that work on list arguments. Exhibit 12.11 shows the
use of the five basic list operators, which are:
4

Tuple rhymes with scruple, and it comes from the mathematical term n-tuple, which comes from the
family of words quadruple, quintuple, etc.

352

CHAPTER 12. FUNCTIONAL LANGUAGES

Exhibit 12.10. Denoting lists in Miranda.

[]
The null list.
[1..n]
The integers from one to n.
odd_numbers = [1,3..100]
Odd numbers from one to one hundred.
evens = [ 10,12 .. 100]
The even numbers between 10 and 100.
eleven_up = [11 .. ]
All numbers greater than 10.
evens_up = [12,14 ..]
All even numbers greater than 10.
week_days = ["Mon","Tue","Wed","Thur","Fri"]
A list of 5 strings is bound to week_days.

L1 ++ L2
item : List
List ! n
L1 -- L2
# List

12.3.3

Concatenate list L2 onto the end of L1.

Add the item to the head of the List.
Select the nth item from the List.
Remove values in L2 from list of values in L1.
Return the number of items on the List.

Function Definitions

A function definition is called an equation because it does look like a mathematical equation.
The simplest functions can be written on one line, like the function sq in Exhibit 12.9. The function
name is written, followed by a dummy parameter list, followed by = and an expression which is
written in terms of the dummy parameters. If the definition requires more than one line of code,
succeeding lines must be indented.
In Miranda, guarded expressions are used for conditional control. A guarded expression consists
of a series of lines, each with an expression on the left, followed by a comma and a Boolean guard
on the right. The interpretation is this: The guard expressions are evaluated in parallel. If the
result of any guard is true, the corresponding expression is evaluated and returned. If more than

Exhibit 12.11. Using list operations in Miranda.

Expression
week_days = ["Mon","Tue","Wed","Thur","Fri"]
# week_days
days = week_days ++ ["Sat","Sun"]
0:[1,2,3]
week_days ! 2
week_day -- "Fri"

Result
5
Make a list of 7 days.
[0,1,2,3]
"Wed" (Zero-based subscript.)
["Mon","Tue","Wed","Thur"]

353

12.3. MIRANDA: A FUNCTIONAL LANGUAGE

Exhibit 12.12. A guarded expression.
gcd a b = gcd (a-b) b,
= gcd a (b-a),
= a,

a>b
a<b
a=b

one guard is true, any one of the corresponding values might be returned. An otherwise clause
is permitted at the end.
A recursive definition of the greatest common divisor algorithm, implemented using a guarded
expression, is shown in Exhibit 12.12. This shows the proper use of guardsexactly one guard is
true, and the set of guards explicitly cover all possibilities.
In a functional language, binding must take the place of assignment. Of course, a function call
offers an opportunity for evaluating expressions and binding the results to parameter names. However, a nest of many function calls can be completely unreadable and difficult for the programmer
to construct. Miranda provides parameter binding plus a second way to bind a value to a name:
the where clause. Within a function definition, where is used to define a local name for the value
of an expression. The Miranda programmer uses where in the same way that a Pascal programmer
uses assignment, to give a name to a subexpression that will be used more than once. Where clauses
may be nested to arbitrary depth, producing block structured name binding. Indentation of inner
blocks is compulsory, because layout information is used by the parser.
Compare the Miranda code in Exhibit 12.13 to the more-or-less equivalent Pascal code in Exhibit
12.14. Both functions compute the two roots of a quadratic equation, but the Pascal code is twice
as long and far more complex. This is due to several factors:
Guarded expressions are syntactically briefer than nested if...then...else statements.
Miranda frees the programmer from concern about evaluation order. The local symbols are

Exhibit 12.13. Where takes the place of assignment.

quadroot a b c = error "complex roots",
= [term1],
= [term1+term2, term1-term2],
where
delta = b*b - 4*a*c
radix = sqrt delta
term1 = -b/(2*a)
term2 = radix/(2*a)

delta < 0
delta = 0
delta > 0

354

CHAPTER 12. FUNCTIONAL LANGUAGES

Exhibit 12.14. Assignment used to name subexpressions in Pascal.

type root_list= ^root_type;
type root_type = record root:

real;

next:

root_list

function quadroot(a, b, c: real): root_list;

var delta, radix, term2: real;
roots, temp: root_list;
begin
roots := NIL;
delta := b*b - 4*a*c;
if delta < 0
then writeln(Error: complex roots)
else begin
new(roots);
roots^.root:= -b/(2*a);
if delta = 0
then roots^.next := NIL
else begin
radix := sqrt(delta);
term2 := radix/(2*a);
new (temp);
roots^.next:= temp;
temp^.root := roots^.root - term2;
roots^.root := roots^.root + term2;
temp^.next := NIL
end
end;
quadroot := roots
end;

end;

12.3. MIRANDA: A FUNCTIONAL LANGUAGE

355

Exhibit 12.15. The reduce function in Miranda.

The reduce function may be defined in Miranda to take three parameters: a function, a list, and
a number. The third argument is the identity value for the given function; its value is used to
initialize the summing process and is returned as the value of reduce when the list argument is a
null list.
reduce f [] n = n
reduce f (a:x) n = f a (reduce x n)
The type of reduce, as discussed in Chapter 9, Section 9.3, is:
reduce :: (num num num )
[num]
num
num

defined in a group at the bottom. They will be evaluated if needed and in whatever order
they might be needed. The order in which the where clauses are written is immaterial.
The Pascal code uses nested conditionals to implement lazy evaluation of the local variables.
In Miranda, lazy evaluation just happensthe programmer does not have to force it by
carefully ordering and nesting the code. If a subexpression is not needed, its where clause is
not evaluated.
The goal of implementing denotational philosophy forced Mirandas designers to provide an
assignment-free method for returning a list. The result is that Miranda lets the programmer
denote variable-length lists with ease; the programmer simply writes a list of expressions
within brackets. In Pascal, however, the list must be constructed and initialized piece by
painful piece.
Higher-order Functions
Like all functional languages, Miranda supports higher-order functions. Thus Miranda programs
often use partial parameterization through closures and functional parameters. A Miranda version
of the reduce function is given in Exhibit 12.15; compare its simplicity to the Pascal version in
Exhibit 9.29.

12.3.4

List Comprehensions

Loops that use explicit loop variables are inherently nonfunctional; both FOR and WHILE loops
depend on statement sequences and assignment. In contrast, a comprehension is a pure functional construct which provides a way to specify iteration without using assignment or sequences

356

CHAPTER 12. FUNCTIONAL LANGUAGES

Exhibit 12.16. The syntax for a list comprehension.

ZF expr ::=
qual-list ::=
generator ::=
filter ::=
body ::=

[ hbodyi | hqual-listi ]
generator { generator | filter }
hvariablei hlist-expressioni
hBoolean expressioni
hlist expressioni

of statements.
A list comprehension is an expression that denotes a finite or infinite list with an arbitrarily
complex structure.5 A comprehension has two parts: on the left is an expression. and on the
right are a series of clauses called a qualifier list, consisting of generators and filters separated by
semicolons. The qualifier list specifies an ordered series of values to bind to the free variable(s) in the
expression. The syntax for list comprehensions is shown in Exhibit 12.16 and an example is shown
in Exhibit 12.17. The clauses that form the qualifier list consist of one or more list-expressions
called generators followed by zero or more Boolean expressions called filters. 6
In many ways a comprehension is like a FOR loop control structure. The expression is like the
body of the loop, its free variable(s) are like the loop control variable(s), and the generators and
filters specify a series of values for the free variable, much as a loop control element specifies a
series of values for the loop variable. A comprehension differs from a loop because a comprehension
returns a list as its result, whereas a loop does not return a value. A comprehension with one
generator is like a single loop; with two generators (for two free variables) a comprehension is like
a pair of nested loops.
For simplicity, let us first examine simple comprehensions with just one generator. A generator
is used to produce and name an ordered set of values. It specifies a symbol and a list expression.
The list expression generates a series of values which are bound, in sequence, to the symbol. Then
the symbol with its value is passed through all of the filters; if any filter evaluates to FALSE, the
value is discarded and the next value is generated. If the value passes all the tests, it is substituted
for its variable in the expression on the left side of the comprehension and used to compute the
5

This terminology is from Haskell; the term used in Miranda is Zermelo Frankel expression, or ZF expression.
This form was named after two early logicians who developed set theory.
6
The use of generators and filters goes back to an early low-level language named IPL/5, Newell [1961]. Notation
similar to that shown here was used by Schwarz in his set-processing language, SetL.

Exhibit 12.17. A list comprehension.

Expression: [ n+2 | n [1..10] ]
Read this as: the list of all values for n+2 such that n takes on the list of values 1 through 10.

357

12.3. MIRANDA: A FUNCTIONAL LANGUAGE

Exhibit 12.18. Some list comprehensions.
Expression
1 [ n*2 | n [2 .. 10] ]
2 [ n*n | n [1, 2 .. ]]
3
4

[ x+y | x [1..3]; y [3, 7] ]

[ x*y | x [1..3]; y [1..3]; x>=y ]

Result
[4, 8, 12, 16, 20]
All perfect squares;
1, 4, 9, etc.).
[4, 5, 6, 8, 9, 10 ]
[1, 2, 4, 3, 6, 9]

next value in the output list.

Let us look at a series of list comprehensions to see how generators and filters work. Line 1 in
Exhibit 12.18 is a very simple example. A single generator is used to produce a list of even numbers
from 2 to 10. This list is fed into the expression on the left, producing a list of multiples of 4
from 4 to 20. In line 2, the result is an infinite list of square numbers.
The combination of generators and filters can specify a list whose structure is considerably more
complex than the control element of a FOR loop, because a generator can start with any existing
list, and any number of filters can be applied to the elements on that list. To use a comprehension,
the programmer specifies a list of values that contain all the ones of interest and perhaps some
other, unwanted values. Then filters are appended to specify subsets of values to remove from the
result. Several filters may be needed if the values to be excluded form a complex pattern.
A comprehension with two free variables needs two generators. In this case, the generators are
executed in a nested fashion; the second generator runs through all of its values for each value
generated by the first. The number of value-pairs generated is the product of the lengths of the
two generator lists. In line 3, two lists are generated and the final result is the pairwise sum of
those lists. Even though two generators are nested here, the final result is a linear list, not a
multidimensional structure. A filter may test just one member of the value-pair, or it may combine
them in a test [Exhibit 12.18, line 4]. Thus the length of the output from a comprehension can be
anything from zero up to the product of the lengths of all the generated lists.
When list comprehensions are used in a program, the result is often a concise, clear statement
of the algorithm. For example, consider the quicksort program in Exhibit 12.19. The function
sort takes a list as its argument and returns a sorted list. Line 1 defines the action of sort

Exhibit 12.19. Quicksort in Miranda.

1.
2.
3.
4.

sort [] = []
sort (pivot:rest) = sort [ y | y rest; b pivot ]
++ [pivot] ++
sort [ z | z rest; b > pivot ]

358

CHAPTER 12. FUNCTIONAL LANGUAGES

on a null-list argumentit simply returns the null list. This is the base case for the recursive
definition of sort. Lines 2 through 4 define sort on all other lists. The fact that the argument
must be a list is denoted by the parentheses between the function name and the = sign which
terminates the header. (Remember, parentheses are not used to enclose parameter lists.) Within
these parentheses, the : notation is used to give local names to the head and tail of the argument.
Thus the name pivot will be bound to the first element, or head, of the list, and rest will be
bound to the rest (or tail) of the list.
The sort function works by separating its argument (a list) into three sections: the first element,
named pivot, one new list containing all elements from the rest of the list that are less than or
equal to the pivot (line 2), and a second new list containing all elements that are greater than the
pivot (line 4). The result returned by sort is these three parts, concatenated (lines 2 through 4)
in the correct order, after the sublists have been recursively sorted.
The program differs from a quicksort program written in a procedural language because it does
not define the procedure for finding all the elements of the list that are less than the first element,
pivot. Instead, a list comprehension is used to denote those elements and supply them as an
argument to a recursive call on the sort function. Implementation of the actual process of finding
those elements is left to the system.
This definition of quicksort is certainly brief and is relatively easy to understand. Further, a
proof of its correctness is straightforward because the program is simply a statement of what it
means to be sorted. However, note the inefficiency involved in processing each element of the
input list twice, once to find small elements, and again to find large ones!

12.3.5

Infinite Lists

Some of the lists denoted in Exhibits 12.10 and 12.18 are infinite lists; that is, they go on forever.
In this section we examine how an infinite list can be represented in a real computer, and why we
might want to use one.
Think of an infinite list as having two parts, a finite head, consisting of values that have already
been computed, and an infinite tail, consisting of code for generating the rest of the list. We will
be lazy about constructing this listnew values will only be computed when they are needed. The
program may access this list like any other, for example, by using subscripts. When the program
needs a value at the head of the list, the desired value is found and returned.
The infinite tail of the list will not be represented as data (obviously). Instead, it will be a
functional object. When the program needs a value on the tail of the list, the function on the tail is
used to extend the head of the list as follows. When it is called, it computes the next list item and
constructs a new functional object for computing the rest of the list. The value and the function
become the new tail of the list, replacing the old functional object. (The list has now grown by one
data item.) This process is repeated until the required list item is generated.
These functional objects can be implemented by closures, consisting of a context that remembers
the current position in the list, and a function that, given the current position, can calculate the
next list item. The new functional object that is created is just a new closure, whose context is the

12.3. MIRANDA: A FUNCTIONAL LANGUAGE

359

Exhibit 12.20. A memoized function.

pow10 0
pow10 (n+1)

= 1
= 10 * powlist !

powlist = [ pow10(x)

x <- [0..]

new position on the list.

Memo Functions.
A generally useful device that can be used in any language is called memoizing. We do it to increase
efficiency when a function is called repeatedly on the same inputs. For example, consider the
process of ASCII-to-floating-point number conversion, which must be done each time a floatingpoint number is read from the input stream. The last step in the conversion is to divide the
intermediate result by a power of 10 that corresponds to the number of decimal places in the input.
We can write a function pow10(d) that returns 10 to the power d. If many numbers are to be read,
though, we would expect to need the same result many times, and computing powers is a slow job.
The obvious answer is to make a table of powers of 10 ahead of time, then just use a subscript to
extract the right power. We can conceive of this in a lazy way. How many powers of 10 do we
need? That isnt really known ahead of time. Should we just calculate a bunch and hope that our
table is long enough? What happens if it isnt? Infinite lists provide a straightforward solution to
this problem. We can use an infinite list to memoize our Pow10 function. When a particular power
of 10 is needed for the first time, the infinite list will be extended that far (and no farther). If a
power is called for that is already on the list, it will be returned directly. Exhibit 12.20 shows a
version of this function memoized by using an infinite list.
Input and Output
Input and output are obvious necessities for any programming language, including a functional
language. However, both are inherently time-dependent. The usual concept of input, that a storage
object takes on a series of values brought in by executing READ requests, is not consistent with the
functional approach. However, infinite lists provide a functionally acceptable way to model I/O. 7
The input stream is modeled as an infinite list of values. Each time you poke the input stream, a
new value is appended to the list of inputs and can then be accessed by a program.
We can also model the output stream as an infinite list. It is important, though, that real
output starts to happen before the entire list is evaluated. This can be guaranteed, even in a lazy
system, by having the output function return a success or fail response and then testing that
7

Continuations can also be used to implement I/O.

360

CHAPTER 12. FUNCTIONAL LANGUAGES

response. The act of testing forces the output function to be evaluated, which has the side effect
of making the output happen.

Exercises
1. Explain what it means to denote a function as opposed to writing code for a function.
2. Define mutability and explain why it is undesirable.
3. Explain the sense in which a Pascal or C programmer must over-sequence code.
4. How does lazy evaluation eliminate much of the concern about sequencing?
5. How does recursion differ from iteration?
6. Name three mechanisms or data structures that must be part of the semantic basis of a
language that supports recursion.
7. FORTRAN supports local variables and parameters in functions. What semantic mechanism
is missing that makes FORTRAN unable to support recursion?
8. What is tail recursion?
9. How can recursion be used to eliminate the need for the WHILE construct?
10. What is the minimal set of necessary control structures? Name a language that contains
these alone.
11. What is the difference between lazy and strict evaluation?
12. What is a guarded expression? How can it be misused?
13. What is the difference between a tuple and a list in Miranda?
14. What is an infinite list? How can it be used to model I/O?
15. What is a list comprehension? Qualifier? Filter?
16. How does a list comprehension differ from a traditional loop?
17. What is memoizing? Why is it used?
18. Write a Miranda script for Euclids greatest common divisor algorithm. Compare your code
to the gcd programs in Prolog [Exhibit 13.13] and C [Exhibit 13.14].

Chapter 13

Logic Programming

Overview
This chapter presents the goals and applications of logic programming and the historical
and theoretical context in which it was developed.
In a logic language, the programmer defines predicates and writes logical formulas involving these predicates, constants, and variables. In this way, a data base of facts and
relations is constructed.
Central to logic programming are the logical operation of implication and the process of
inference. The programmer states rules which the language interpreter may use to infer
new facts from those in the data base. This inference is done by a powerful reasoning
technique called resolution.
Prolog is introduced as an example of a logic language. It was the first and is still
widely known and used today. The relationship between deduction in an axiomatized
system and computation is explored by comparing short Prolog programs to analogous
programs in a procedural language.

In a procedural language, we define constants and variables and accept input data. We describe
computation by defining functions and/or procedures which call preexisting functions and operators
to compute new values from existing data. Assignment or binding is used to store the new data
for future use. Because a program consists of a series of control units which will be executed
sequentially, the order of computation is largely determined by the programmer. Within each
control unit are other control units and function calls which largely (but not completely) define the
sequence in which the actions and computations will be done.
361

362

CHAPTER 13. LOGIC PROGRAMMING

In a logic language, we also define constants and variables. However, computation is done by
finding data objects that satisfy a set of constraints. The programmer states facts, or relationships,
about data objects, as well as inference rules by which conclusions may be drawn about those
objects. A query may then be writtena statement whose truth or falsity the programmer does
not know. The language interpreter attempts to prove that statement from the facts and rules
previously provided.
The programmer does not define either the nature of the computational process or the exact
order in which computations will happen. Those aspects are defined by the theorem-proving mechanisms built into the logic language interpreter. For this reason, logic languages are often described
as declarative rather than imperative.
The traditional languages based on arithmetic, assignment, and sequential computation force
programmers to over-specify the sequencing of their nonsequential algorithms and nonordered data.
For some applicationsfor example, language recognition, artificial intelligence, data base query
languages, and expert systemsit is easier and more comfortable to model the application declaratively. Interest in logic programming languages is, therefore, especially strong in these areas.

13.1

Predicate Calculus

The predicate calculus is a formal logical system for symbolizing facts about data structures and
reasoning about them. It is also the basis on which logic programming languages are built. Before
one can understand logic programming languages, one must be familiar with the symbols and
concepts of logic. We review them here briefly.

13.1.1

Formulas

The components of the first-order predicate calculus are symbols for constants, variables, logical
operations, functions, predicates, and quantifiers. Symbols may be combined to make formulas.
Constants
A constant denotes one specific object, which belongs to some specific domain, or set, of real
or mathematical objects. This domain is also called a universe of discourse. Normally, several
domains are involved in a single logical system, and different symbols are defined for each. In this
discussion, we generally take the domain to be the set of numeric integers, and we will denote the
integers by ordinary base-10 numerals. Names of other constants will be written using lowercase
letters (including the underscore).
Variables
A variable ranges over a particular domain and represents some unspecified object from that domain. During a deduction, a variable may be instantiated, that is, bound to one constant object

363

13.1. PREDICATE CALCULUS

Exhibit 13.1. Predicates and arity.

Predicate
odd(X)
father(F,S)
divide(N,D,Q,R)

Arity
1
2
4

Intent
X is an odd number.
F is the father of S.
N divided by D gives quotient Q and remainder R.

Instantiated predicate
odd(2)
father(David, Solomon)
divide(23, 3, 7, 2)
divide(23, 3, 7, N)

Truth value
False.
True.
True.
Neither true nor false until we instantiate N.

from its domain. In general, several variables will be simultaneously instantiated, each to a different
value, and we use the term instantiate in this more general sense. We will use uppercase letters
to name variables.
Functions
Symbols are used to name functions and may have parameters. A function name written with
the appropriate collection of arguments denotes the result of executing that function on those
arguments and is called a term. The arguments to functions may be constants, variables, or other
terms. We will use lowercase letters (with the underscore) to name functions.
Predicates
A predicate is a symbol that denotes some property that an object, or several objects, might have.
Thus if gummy and biggerthan are predicates and j and k are objects, then gummy(k) is an assertion
that k has the property gummy, and biggerthan(j,k) asserts that j is biggerthan k. We will use
lowercase letters to name predicates.
The arguments to a predicate are terms. A predicate applied to constants, or to variable-free
terms, is called a sentence or a proposition. It has a truth value (true or false), depending on
whether or not the predicate is true of the objects denoted by its arguments. A predicate may also
be applied to variables, or to terms containing variables, but it does not form a sentence until all
those variables are instantiated. If the instantiated predicate is true, then we say the predicate is
satisfied by the instantiation.
A predicate may have one, two, or more arguments. [Exhibit 13.1] We use the word arity to
denote the number of arguments, and describe a predicate with arity N as an N -ary predicate.
Thus, the predicate divide in Exhibit 13.1 has arity 4 and is a 4-ary predicate.

364

CHAPTER 13. LOGIC PROGRAMMING

Exhibit 13.2. Quantified predicates.

1.
2.
3.
4.
5.
6.

Quantified predicate
X odd(X)
X odd(X)
X (X=2*Y+1 odd(X))
X Y divide(X, 3, Y, 0)
X Y divide(X, 3, Y, 0)
X Y divide(X, 3, Y, 0)

Truth value
false.
true, satisfied by X=3.
true.
false for X=2.
true, satisfied by X=12, Y=4.
false, but harder to prove.

Quantifiers
Variables let us denote arbitrary objects in a domain; quantifiers let us denote sets of objects.
There are two kinds of quantifiers: existential and universal. The existential quantifier is written
X and read as there exists an X such that, and the universal quantifier is written Y and read
as for all Y. Quantifiers are written at the beginning of a formula and bind all occurrences of
the quantified variable within the formula. The formula to which the quantifier is applied is called
the scope. Thus, in Exhibit 13.2, both occurrences of the variable X on the third line are bound by
the quantifier X. The scope of the quantifier is X = 2 * Y + 1 odd(X).
A quantified formula containing no free variables is a sentence and has a truth value. An
existentially quantified formula is true if the scope of the quantifier(s) can be satisfied, that is, if
there is at least one way to instantiate the variables (each from its proper domain) so that the
resulting proposition is true. It is therefore easy to show that the quantified predicates on lines 2
and 5 in Exhibit 13.2 are truewe simply find suitable values for the variables.
Proving the truth of a universally quantified predicate is more difficult, especially if the universe
of discourse is large or infinite. One must demonstrate that the scope is true for all objects in the

Exhibit 13.3. Quantified formulas.

1.
2.
3.

Quantified predicate
X odd(X)
X odd(X)
X X=2*Y+1 odd(X)

Negation
X not odd(X)
X not odd(X)
X not(X=2*Y+1 odd(X))
X not( not(X=2*Y+1) or odd(X) )
X (X=2*Y+1) and not odd(X))

4.
5.
6.

X Y divide(X, 3, Y, 0)
X Y divide(X, 3, Y, 0)
X Y divide(X, 3, Y, 0)

X Y not divide(X, 3, Y, 0)
X Y not divide(X, 3, Y, 0)
X Y not divide(X, 3, Y, 0)

365

13.2. PROOF SYSTEMS

Exhibit 13.4. Semantics of not, or, and, and implies.

p
true
true
false
false

q
true
false
true
false

not p
false
true

p or q
true
true
true
false

p and q
true
false
false
false

q or not p
p implies q
true
false
true
true

universe. On the other hand, the negation of a universally quantified predicate is an existentially
quantified predicate, as shown in Exhibit 13.3. Thus it is easy to show that a universally quantified
statement is false; one just needs to find a single instantiation of the variables that makes the scope
false, as shown on line 4 of Exhibit 13.2.
Logical Operations
Formulas can be combined to build bigger formulas using logical operations. Only a small number
of logical operations, and, or, and not, are needed. The semantics of these operations are
defined by the truth tables in Exhibit 13.4. Other logical operations such as implies are often
used and can be defined in terms of the three basic operations.

13.2

Proof Systems

When we use the predicate calculus to represent some problem area, we start by defining symbols
to represent objects, functions, and predicates from our external domains [Exhibit 13.5]. Then we
write a set of sentences in terms of these symbols that represent facts about our system. These
formulas, called axioms, usually capture only a portion of the logicians semantic intent: that
portion relevant to the current problem. Other aspects of semantic intent may be stated in the
form of comments.
A proof system is a set of valid deduction rules that can be applied to axioms to deduce theorems.
A deduction is an application of one of the valid inference rules of the proof system. In a simple
deductive step, we derive a conclusion from the axioms and previously deduced theorems according

Exhibit 13.5. Logical symbols and sentences.

Constants: 17, gnu, apple_tree
Variables:
X, YY
Functions: successor(X), product(Y,Z)
Predicates: even(X), even(successor(17)), cousin(Y,Z), remainder(X, Y, Z)

366

CHAPTER 13. LOGIC PROGRAMMING

to that rule. A proof is a series of sentences, ending with the sentence to be proved, such that each
sentence is either an axiom or a valid deduction from some subset of the previous sentences in the
proof. If there are N sentences in a proof, we will call it an N -step proof. A refutation is a proof
of the negation of a sentence, which is the same as a proof that the sentence is inconsistent with
the given axioms.
Two proofs of the same proposition can differ drastically in length and complexity. A logician
constructs a proof or a refutation using insight, pattern recognition, and a bit of luck. With more
insight (and luck) a proof is clear and brief; with less it can be tedious and murky. Even the
clumsiest proof, however, takes some skill to construct.
A sentence that can be deduced from the axioms is called a valid sentence. Any valid sentence
can also be called a theorem.1 The set of all theorems derivable from a set of axioms is called a
theory. We say that a theory is consistent if it does not contain contradictory theorems.
Classical Logic
Logicians have invented several different valid proof systems with very different deduction rules.
One such system, often called classical logic, has been in use for centuries. It is based on the ancient
deduction rule called modus ponens (the way of the bridge), which is the primary rule used to
reason about conditional sentences. A conditional sentence is of the form:
C1 C2
where the arrow symbolizes the logical operation implies, and C1 and C2 are sentences. C1 C2
is read C1 implies C2; C1 is called the premise and C2 is called the conclusion. Modus ponens
says,
If the truth of the premise can be established, then one can infer that the conclusion is
also true.
In symbols, this lets one derive the theorem C2 if both the conditional sentence C1 C2 and the
sentence C1 have been previously proved (or are axioms).
Other proof rules are needed in order to deal with quantified sentences and with formulas
containing free variables. For example, the rule of generalization says that if a formula C(X)
containing a free variable X can be proved (meaning that C is true no matter how X is instantiated),
then one can infer the sentence XC(X).
Clausal Logic and Resolution
Clausal logic is an alternative to classical logic. It is based on a powerful deduction rule called
resolution which is described in Section 13.4.1. Resolution often allows much shorter proofs because
one resolution inference can take the place of many simple modus ponens inferences.
1

In other contexts this word is used only for particularly interesting or useful valid sentences.

367

13.3. MODELS

Exhibit 13.6. A simple theory: the Theory of Ugliness.

Axioms for the Theory of Ugliness:
1. X ugly(X)
not ugly(X+1)
2. X not ugly(X+1) ugly(X)
3. 2 = 1 + 1
Theorems from the Theory of Ugliness:
1. X ugly(X)
ugly(X+2)
2. X ugly(X+2)
ugly(X)

Prolog itself is a proof system which uses clausal logic and resolution. Clausal logic generally
writes implies using the left arrow, so that the conditional sentence C1 C2 becomes
C2 C1
and is read C2 if C1.

13.3

Models

Recall that a theory is a set of formulas derivable from a set of axioms. Given a theory, we can try
to find a model for it. A model is not something that can be written or discussed within a formal
system. Rather, it is one possible interpretation of the semantic intent of a theory and is presented
in a metalanguage. To make a model, we find interpretations for each constant, function, and
predicate symbol in the axioms, so that the axioms are all true statements about those interpretations. Specifically, we must define each universe of discourse and show the correspondence between
members of a universe, or domain, and the constants in the axioms. The predicate symbols in the
axioms must also be assigned to properties of objects in these domains, and the function symbols
must be assigned to functions that operate on objects from the specified domains. An inconsistent
theory has no models because it is impossible to find interpretations that simultaneously satisfy
contradictory theorems.
Because a set of axioms only partially defines the functions and predicates of a theory, one
theory may have several models. For example, consider the very simple theory in Exhibit 13.6,
which we will call the Theory of Ugliness. To find a model for this theory we must find a domain
of discourse which contains interpretations for the constant symbols 1 and 2, for the function
symbol +, and for the predicate symbol ugly. This is easily done.
Let our domain of discourse be the integers. Let 1 represent the integer one and let 2
represent the integer two.
Let + represent the function of addition on the integers.

368

CHAPTER 13. LOGIC PROGRAMMING

Now there are two possible interpretations for ugly: the integer predicate odd or the
integer predicate even.

We can now make two models for the Theory of Ugliness. In one model, all odd numbers are
ugly, even numbers are not. In the other model, all even numbers are ugly, odd numbers are
not. Both models are completely consistent with the axioms, yet the axioms do not let us identify
even one ugly number with certainty. Nevertheless, we can reason about ugly numbers and prove,
for example, that if X is ugly, then so is X+2. (This is the theorem mentioned in Exhibit 13.6.)
The fact that a set of axioms can have more than one model is a very important matter in logic
programming applications; we do not have to axiomatize all aspects of a model in order to reason
about it.

13.4

Automatic Theorem Proving

Symbolic logic was first formulated by Frege 2 , at the end of the nineteenth century. This work was
extended during the next fifty years by several pioneers. Certain properties of logical theories have
been central to this work. The first is completeness: a theory is complete if every sentence that is
true in all models of the theory can be proved from its axioms by using its deductive rules. We say
that such a sentence is a logical consequence of the axioms.
A second property, first studied by Hilbert, is decidability. A theory is decidable if one can decide
(that is, prove) whether any given sentence is or is not valid in that theory. This is equivalent to
saying that, for any sentence, either it or its negation can be proved. A theory can be complete but
not decidable if there are sentences that are true in some models but not in others. Such sentences
are not logical consequences of the axioms. For example, in the Theory of Ugliness, the sentence
ugly(3) is true in one model, where the predicate ugly corresponds to the integer property odd, but
not in the alternate model where ugly represents even. Thus ugly(3) is not a logical consequence
of the theory, and neither ugly(3) nor not ugly(3) can be proved from the axioms.
By 1940, the logical foundations of mathematics had been thoroughly established and explored.
The central theorems are as follows:
1. The predicate calculus is complete. That is, the proof techniques of the predicate calculus
can be used to develop a formal proof for every logically true proposition. (Goedel, Herbrand,
1930.)
2. If a proposition can be proved at all, then it can be proved in a direct way, without requiring
insight. That is, a mechanical proof method exists. (Gentzen, 1936.)
3. There is not, and cannot be, an algorithm that correctly identifies propositions that are not
true and cannot be proved. Even though a proof procedure exists for true propositions, a
proof may take an indefinitely long time to produce by this procedure. There is no time at
2

Frege [1879].

13.4. AUTOMATIC THEOREM PROVING

369

which we can conclude that a proof will not be discovered in the future. As a consequence,
Hilberts decision problem is unsolvable for the predicate calculus. (Church, Turing, 1936.)
Since the invention of electronic computers, there has been interest in automating the proof
process. At first this was only of academic interest. More recently, however, the community has
come to understand that deduction and computation are analogous processesindeed, computations can be written as deductions, and vice versa. Since 1940, logicians have been working on the
problem of automatic theorem proving programs:
We know that we can devise an algorithm (or write a program) that will take a set of axioms
and enumerate and prove all possible propositions that can be derived from those axioms.
Of more interest, we would like an automatic system which, given a single interesting proposition, will prove it.
We would like this system to be efficient enough to be useful.
We would like a language in which to express propositions, and a language processor that
will find proofs for these propositions, if they exist.
As a consequence of the completeness result, we know that it is theoretically possible to build
a program that can prove any theorem that is a logical consequence of a set of axioms. However, if
we simply generate and prove propositions systematically, most of them will be simple variations
of similar propositions and any interesting theorems would be buried in an avalanche of trivia.
Moreover, the process would take forever. This is analogous to the old question of the monkeys
and the typewriters: if you put 1000 monkeys in front of 1000 typewriters, and they all randomly
type, day and night, will they ever produce any great literature? The answer is yes: eventually the
random process will produce a poem, or more, but the valuable information will be lost in a sea of
garbage. Finding the great stuff would be impractical.
Once we restrict our attention to proving a single proposition, the main problem is efficiency.
If we start with the axioms and make every inference possible, a few might be relevant to our
chosen proposition, but most are not, and we have no automatable way to tell which. Further,
any deduction we make can lead to several other deductions, creating multiple, branching lines of
reasoning. Visualize all possible deductions from a set of axioms as a tree structure. The width
and bushiness of the tree are determined by the number of alternative deductions possible at each
stage. A sentence at level one of the tree represents an axiom. A sentence at level two is a theorem
with a two-step proof, and the sentences at level N are theorems with N step proofs.
If we could build an automatic theorem-prover to start with a set of axioms, make every possible inference, and follow every possible line of reasoning, our proof method would be complete.
However, that would create an impossibly bushy and impossibly deep tree-structure of lines of
deduction. We say that there is a combinatorial explosion of possible lines of reasoning, because of
the huge number of different combinations of inferences that we might make.

370

CHAPTER 13. LOGIC PROGRAMMING

We could try to generate proofs by constructing this tree in a breadth-first manner; that is,
make all two-step deductions before starting on the three-step proofs. If we do so, we build a very
bushy tree. By the time we have made all ten-step deductions, we could have so many possibilities
that remembering and extending them all becomes unmanageable. And ten-step proofs are material
for student exercises! Hundred-step proofs would be practically impossible. On the other hand, we
could try to generate proofs in a depth-first manner; that is, we could follow one line of reasoning
to its logical conclusion before exploring other lines. However, using this approach, we could get
trapped in an infinite, circular line of reasoning, and our proof process would never terminate.
Some propositions which can be easily proved by a human with insight are very difficult to
prove by constructing the entire reasoning tree. For example, Exhibit 13.7 states a very uninteresting proposition: The number 65534 is a good number. This is proved from axioms that the
number zero is good and conditional rules about good numbers that use the arithmetic operations
of addition, multiplication, and exponentiation to the base two. The complete line-of-reasoning
tree for this foolish proposition would be immensely bushy, since four or five rules can be applied
at each step and, usually, only one of them leads to this twenty-one-step proof or to a similar short
proof. Moreover, one branch of the tree is at least 32,767 lines long. Finally, rules 3 and 4 can be
applied alternately to create an infinite chain of deductions that makes no progress. Whether we
search this tree breadth first or depth first, any systematic exploration of it will become mired in
masses of uninteresting inferences. Admittedly, propositions like this are pathological cases and are
of no real interest to anyone. However, they illustrate that, in order to be truly useful, a general
proof algorithm must be able to avoid trouble when given such a proposition to prove.

13.4.1

Resolution Theorem Provers

Early in the history of automatic theorem proving, it was clearly understood that a brute-force,
enumerate-them-all approach could never be practical. A usable automatic theorem-prover would
have to use techniques that were different from the step-by-step deductions that humans were
accustomed to making. Two breakthroughs happened. First, it was shown that a highly restricted
form of the predicate calculus, called clausal logic, was still complete. In clausal logic, all sentences
are a series of predicates, possibly negated, connected by or operators.
Second, a new proof technique named resolution was invented. The goal of resolution was to
collapse many simple modus-ponens-like deduction steps into one large inference, thereby eliminating whole sections of the unmanageable deduction tree. Resolution operated on clausal sentences
and relied on a pattern-matching algorithm called unification to identify sections of a proposition
that could be simplified.
The process of resolution on two clauses is analogous to algebraic simplification; it is a way to
derive a new, simpler fact from a pair of related facts. First, unification is used to identify portions
of the two clauses that are related and can be simplified by a cut operation. Specifically, given
the two assertions:
A (B or C)

13.4. AUTOMATIC THEOREM PROVING

371

Exhibit 13.7. A proof that 65,534 is a good number.

Let the predicate good(X) mean that the number X is good.
Let the predicate powtwo(Y,Z) mean that 2 to the power Y is Z.
Prolog syntax is used below to write the propositions. The comma in axiom 6 should be
interpreted as the logical operator and.
Axioms.
1. good(0).
2. powtwo(0,1).
3. good(A-2)
4. good(A+2)
5. good(A*2)
6. good(A)
7. powtwo(P+1,A*2)

good(A).
good(A).
good(A).
powtwo(B,A), good(B).
powtwo(P,A).

The shortest proof of the proposition that 65,534 is a good number:

8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.

good(2)
good(4)
powtwo(1,2)
powtwo(2,4)
powtwo(3,8)
powtwo(4,16)
good(16)
powtwo(5,32)
powtwo(6,64)
powtwo(7,128)
powtwo(8,256)
powtwo(9,512)
powtwo(10,1024)
powtwo(11,2048)
powtwo(12,4096)
powtwo(13,8192)
powtwo(14,16384)
powtwo(15,32768)
powtwo(16,65536)
good(65536)
good(65534)

by
by
by
by
by
by
by
by
by
by
by
by
by
by
by
by
by
by
by
by
by

rule
rule
rule
rule
rule
rule
rule
rule
rule
rule
rule
rule
rule
rule
rule
rule
rule
rule
rule
rule
rule

4,
5,
7,
7,
7,
7,
6,
7,
7,
7,
7,
7,
7,
7,
7,
7,
7,
7,
7,
6,
3,

A=0, with line 1.

A=2, with line 8.
P=0, A=1, with rule 2.
P=1, A=2, with line 10
P=2, A=4.
P=3, A=8.
A=16, B=4.
P=4, A=16.
P=5, A=32.
P=6, A=64.
P=7, A=128.
P=8, A=256.
P=9, A=512.
P=9, A=1024.
P=9, A=2048.
P=9, A=4096.
P=9, A=8192.
P=9, A=16384.
P=9, A=32768.
A=65536, B=16, with lines 14, 26.
A=65536, with line 27.

372

CHAPTER 13. LOGIC PROGRAMMING

and
(C or D) E
then resolution lets us infer that
(A or D) (B or E)
Before making this inference, though, we must establish the correspondence between the C parts of
the two formulas. Since these subformulas might be quite complicated and involve many predicates
with variables and constants, unification (the process of establishing the correspondence or showing
that no correspondence exists) is not a trivial matter.
A resolution deduction is a sequence of resolution inferences. A resolution proof is a resolution
deduction whose conclusion is false; it establishes that the premises, taken together, form a
contradiction. A resolution proof can be used to prove that the negation of any single premise can
be deduced from the other premises by ordinary logical deduction.
The original resolution methods were better than the brute-force method of enumerating and
testing all possible lines of reasoning but were still not practical or useful in real applications.
For example, the proposition in Exhibit 13.7 would cause an early resolution-theorem-prover to
become mired in irrelevant deductions. Since then, it has been shown that we can restrict the
kinds of propositions even further, in ways that make the resolution more efficient. Happily, these
restrictions affect only the way a proposition is stated; they do not affect the completeness of the
system.
Some interesting theorems are easily derived using ordinary resolution. However, the computation for most theorems is very extensive and is still not of any practical use. A more advanced
method, named hyperresolution, was developed that operates on Horn clauses, which are even more
highly restricted sentences. It is this form of resolution on which logic programming languages are
built.
Implementing Resolution Efficiently
Horn Clauses. A Horn clause is a clause with at most one unnegated predicate symbol. As
shown in Exhibit 13.8, this is equivalent to saying that a sentence may contain a single implies
operator, whose premise is several predicates connected by and and whose conclusion contains
only one predicate symbol.
Unification Most of the time spent in finding a resolution proof is actually spent in identifying
related phrases in two sentences that can be simplified, or cut. A hyperresolution inference depends
on the result of a unification procedure. Hyperresolution says that, from the premises
A1 and A2 and . . . and An
and
(B1 and B2 and . . . and Bn) D

373

13.4. AUTOMATIC THEOREM PROVING

Exhibit 13.8. Ways to write a Horn clause.

We show that a Horn clause can be rewritten as an implication whose premise is a series of predicates
connected by and operators and whose conclusion is a single predicate. Let each capital letter in
the following derivation stand for a predicate symbol with its required arguments.
A Horn clause is a clause with at most one nonnegated predicate, such as:
not P or not Q or not R or S or not T
Because or is associative and commutative, we can move all the negated predicates to the right:
S or not P or not Q or not R or not T
Using DeMorgans law, this expression can be transformed to the form:
S or not(P and Q and R and T )
The operator implies is defined in terms of not and or, and can be written with either a
right-pointing arrow or a left-pointing arrow. Thus the two expressions
BA

AB

are defined to mean

B or not A
We use the left-pointing implies arrow here to conform to the order of terms in Prolog syntax.
Now we can rewrite the Horn clause as follows:
S (P and Q and R and T )

we can infer D, where all variables in D have been instantiated with expressions that unify the set
{{A1,B1}, {A2,B2},. . . ,{An,Bn}}.
The unification algorithm finds a set of substitutions for the variables in the two formulas so
that, after making the substitutions, the two clauses are equal. This is a kind of pattern-matching
process. Exhibit 13.9 shows examples of simple unification problems and their results.
We use the unification algorithm to determine whether or not a conditional Horn clause, C,
can be used in a resolution, given the current set of unconditional clauses in the data base. To
perform a hyperresolution step, we need to find a set of unconditional clauses that cover all the
predicate symbols in C, and a single set of instantiations for all the variables in these clauses that
unifies each unconditional clause with the corresponding predicate. Let us describe the process of
unifying an unconditional clause whose predicate symbol is P , and a single predicate symbol, Q,

374

CHAPTER 13. LOGIC PROGRAMMING

Exhibit 13.9. Unifying clauses.

1.
2.
3.
4.

Formula 1
pp(a,b)
pp(Q(c,d),Z)
pp(X,Y) and qq(Y,a)
qq(pp, gg(X,Y), X, Y)

Formula 2
pp(X,Y)
pp(Y,e)
pp(a,b) and qq(Z,W)
qq(Y, Z, hh(U, kk), U)

Most General Unification

X=a, Y=b
Y=Q(c,d), Z=e
W=a, X=a, Y=b, Z=b
U=pp, X=hh(pp,kk)
Y=pp, Z= gg( hh(pp,kk), pp)

in the conditional clause C as follows:

A cluster is a set of terms that have been mapped into one equivalence class and must be
unified.
To begin the unification process, map P and Q into a cluster.
1. Choose any cluster. Look at all the terms in the cluster. (In unifying M items, there
will be M terms.)
2. If one term is a constant and if another term is a predicate or a constant that does
not match, no unification is possible and the process fails. If all terms are matching
constants, this branch is successfully unified. Otherwise proceed.
3. If two or more terms are predicates and they are different, no unification is possible and
the process fails. Otherwise, proceed, and let N be the arity of the cluster.
4. We know we are merging a combination of one constant or one predicate with one or more
variables. Merge all terms in the cluster into a single unified super-term. Temporarily,
the super-term will inherit all the arguments from the predicate terms in the cluster, so
that it will have M first arguments, M second arguments, ..., and M N th arguments.
5. We must replace each collection of M N th arguments that was created in step 4 by a
unified term. For j from 1 to N , look at all the N th arguments that belong to this superterm. (If we are unifying M terms, there will be M arguments to consider.) Make a new
cluster by mapping together all the terms that are the jth arguments of the super-term.
During this process, we may find that one of the terms to be mapped was mapped to a
different cluster by a previous operation. In that case, the two clusters are mapped into
one.
6. Replace the cluster we have been processing by the super-term created in step 4 with
its arguments replaced by the argument clusters from step 5.
7. Repeat steps 1 through 6 until there are no remaining clusters in the formula.
The process will end when it has descended through all levels of argument nesting.

13.5. PROLOG

13.5

Prolog

13.5.1

The Prolog Environment

375

Prolog is an interactive language for expressing predicates, relations, and axioms. The programmer
may state facts or make queries about those facts. In response to a query, Prolog will attempt to
find one instantiation of the querys variables that makes the formula true.
In addition to the language for writing logical assertions and queries, a Prolog system includes
a metalanguage for managing the data base of facts and rules. Large data bases can be entered
either from a file or from the keyboard by calling the input function consult(filename). Once in
the system, parts of the data base can be listed (by using listing(predicate_name)) or edited
(by using retract(Z)), and the data base can be extended (by using asserta(X) or assertz(X)).
After entering a data base, the programmer may interactively enter queries, test hypotheses,
and assert more facts. In response to a query, Prolog executes its proof-procedure and searches for
a set of instantiations that satisfies a querys variables. If this search succeeds, Prolog responds
by writing out the instantiations that were found. The programmer can indicate that she or he
is satisfied (by typing a carriage return), and Prolog responds by saying yes and giving a new
prompt. Alternatively, the programmer may ask Prolog to search for additional instantiations (by
typing a ; and a carriage return). This search process continues until the programmer decides she
or he has seen enough or the entire data base has been searched and no further set of instantiations
can be found, at which time Prolog responds by writing no and a new prompt.

13.5.2

Data Objects and Terms

The basic components of a Prolog program are objects (constants, variables, structures, lists),
predicates, operators, functions, and rules.
Constants: Integers are predefined and represented in base 10. User-defined constants, or
atoms, have names that start with lowercase letters. Remember that a Prolog program, like
a theory, may have many models (or none) and a Prolog atom may correspond to different
objects in different models.
Variables: All variable names start with uppercase letters. All occurrences of the same
variable in a rule will be bound to the same object. The symbol _ can be used in place of
a variable name in cases where the program does not need to refer to this same object twice.
If the _ symbol occurs several times in a rule, it may be bound to several different objects.
Structures: Compound data types (record types) may be defined by specifying a functor (the
type name) and a list of components (fields), in parentheses:
htype namei( hcomponenti, . . . , hcomponenti )
Lists may be denoted by using square brackets. The empty brackets [] denote the empty
(null) list. The notation [a,b,c] denotes a list of three terms whose head is a and whose

376

CHAPTER 13. LOGIC PROGRAMMING

last item is c. The notation [A|X] denotes a list of length one or greater with head A and
tail, or remainder, X, which may be a list, a single atom, or null. This notation is often used
in function definitions to give local names (A and X) to the parts of a list.
This bracket notation is a sugaring of a less attractive basic syntax; [Head|Tail] is a
syntactic variant of .(Head,Tail) and [x,y,z] is a variant of .(x, .(y, .(z, []))).
We will call both forms list specifications. A list specification can denote a fully defined list,
such as [a, b, c], a fully variable list, such as [Head|Tail], or a list with a defined head
and a variable tail, such as [a,b|Z]. The term open list refers to a list with a variable tail.
During the proof process, an open list may be unified (matched) with any list that has the
same constant as its head. The selector functions head(x) and tail(x) are defined to return
the head and tail of a list argument.
Predicates =, \=, <, =<, >, and =>are defined. Users may introduce their own predicate
symbols, which are written in lowercase. Again, remember that the semantics of these symbols
remains largely in the programmers mind. Within Prolog they have no meaning except that
captured by the axioms, or rules, the programmer supplies.
The integer operators +, -, *, /, and mod are defined and may be written either in standard
functional notation: +(3,X), or in infix form: 3+X. When infix form is used, the operators
have the usual precedence. These operators may be used in an is formula, thus: X is
(A+B) mod C. When Prolog processes such a formula it performs the computation using
the current bindings of A, B, and C, and binds the result to X. An error is generated if any
variable on the right of the is is not bound to a number.

A term is a predicate symbol or operator or function symbol, together with a parenthesized list
of the right number of arguments.

13.5.3

Horn Clauses in Prolog

Deductive Rules
Rules are the axioms that form the basis for deductions in Prolog. A rule represents a statement
of the form
If a set of premises are all true,
then we may infer that a given conclusion is also true.
The conclusion is written on the left, followed by a :- sign, and the premises are written on the
right, separated by commas. The variables in a rule are implicitly, universally quantified. A rule is
also called a conditional clause. Examples of rules are given in Exhibit 13.10.
A fact is an assertion that some object or objects satisfy a given predicate. Formally, a fact is a
rule which has a conclusion but no conditions, and it is, therefore, sometimes called an unconditional
clause. Examples of facts are given in Exhibit 13.10. The programmer adds a new fact to the data

377

13.5. PROLOG

Exhibit 13.10. Prolog rules and facts.

Conditional clauses
pretty(X) :- artwork(X).
(rules)
pretty(X) :- color(X,red), flower(X).
watchout(X) :- sharp(X,_).
Unconditional clauses
(facts)

color(rose, red).
sharp(rose, stem).
sharp(holly, leaf).
flower(rose).
flower(violet).
artwork(painting(Monet, haystack_at_Giverny)).

base by asserting it. If X is a fact, then the predicate asserta(X) appends the fact X to the
beginning of the data base, and the predicate assertz(X) appends it to the end.
Prolog rules and facts are Horn clauses. The conditions of a rule are a conjunction of terms,
even though we write commas between the terms rather than and operators. Each single rule
represents the and of a series of conditions. The conclusion is a single positive term, and the
:- represents a sign. An underscore is a wild card. Thus a Prolog rule is a Horn clause written
in the form shown on the last line of Exhibit 13.8. Prolog facts are unconditional Horn clauses.
Typically, one predicate may have several rules defined for it [Exhibit 13.13]. A list of rules
for one predicate represents the or of those conditionsonly one rule will be used to interpret a
given call on that predicate, and the one used will be the first one that can be satisfied. Thus we
can use a sequence of rules to express a generalized conditional semantic structure.
Rules may be recursive and thus may be used to implement repetition in an algorithm. A
recursive predicate must, of course, have at least two rules (the base case and the recursive step).

Exhibit 13.11. Prolog queries.

?- pretty(rose).
yes
?- pretty(Y).
Y=painting(Monet, haystack_at_Giverny).
Y=rose
no
?- pretty(W),sharp(W,Z).
W=rose Z=stem
no

378

CHAPTER 13. LOGIC PROGRAMMING

Exhibit 13.12. Interpreting Prolog rules.

Rules 2 and 3 from the gcd algorithm [Exhibit 13.13] are dissected here. In both rules, A and B are
input parameters and D is an output parameter.
gcd(A, B, D) :- (A<B), gcd(B, A, D).
This right side is the analog of a Pascal IF...THEN statement. Prolog interprets it as follows: test
whether (A<B) is true. If so, call predicate gcd with arguments (B, A, D). If that succeeds, D will
be bound to the answer.
gcd(A, B, D) := (A>B), (B>0), R is A mod B, gcd(B, R, D).
This performs the equivalent of two Pascal if statements, an assignment, and a function call. Prolog
interprets it thus: test whether (A>B) is true. If so, test (B>0). If both conditions hold, calculate
A mod B and bind to R. Finally, call predicate gcd with arguments (B, R, D). If all succeed, D will
be bound to the answer.

Queries
A query in Prolog is a request to Prolog to prove a theorem. Because the question posed is the goal
of the proof process, a query is also called a goal. Syntactically, a query is a list of terms, separated
by commas [Exhibit 13.11]. Semantically, the commas represent and operators. If the query has
no variables, Prolog will attempt to prove it from the rules and facts previously given. Thus, in
Exhibit 13.11, the query ?- pretty(rose) can be proved from the second rule taken together with
the first and fourth facts in Exhibit 13.10.
If the query does contain variables, the Prolog theorem-prover attempts to find a set of instantiations that satisfy the query. All variables in the query are, implicitly, quantified by ; that
is, a query asks whether any set of objects exists that can satisfy the clause. Thus to respond to
the query ?- pretty(Y), Prolog tries to find some object, Y, that has the property pretty.
It will begin by using rule 1, the first rule given for the predicate pretty, and will combine
this with fact 5, instantiating Y to a painting, and producing the output Y = painting(Monet,
haystack_at_Giverny). If the programmer types ; to ask for another instantiation, Prolog will
continue to search its data base and find the second instantiation, Y = rose. If the search is
continued again, no further instantiations succeed, and the answer no is printed.
A query may contain a series of predicate terms which will always be processed in order, left to
right, and processing will be aborted if any one predicate cannot be satisfied [Exhibit 13.12. This
built-in conditional sequencing is used where if statements and sequences of assignments would
be used in a procedural language.

13.5. PROLOG

13.5.4

379

The Prolog Deduction Process

A query establishes a goal for the deductive process, in the form of a conjunction of terms. To
satisfy this goal, Prolog first takes each individual term in the goal, in order, as a subgoal, and
recursively attempts to satisfy that subgoal.
The subgoal is a predicate with arguments. Prolog begins by finding, in the data base, the rules
for that predicate which have the right arity for the subgoal, and then starts with the first rule. It
attempts to unify the head (conclusion) of the rule with the subgoal; if there is a conflict because
constant terms do not match, Prolog will go on to the next rule for that predicate. If the head can
be unified with the subgoal, it means that this rule is, potentially, applicable.
Prolog then tries to satisfy each of the condition terms on the right side of that rule. To do
this, it searches for a set of facts that cover the terms of the goal. If this process succeeds, the
rule is applied to the goal, the goals variables are instantiated to the objects discovered by the
unification algorithm, and the set of instantiations is returned to the calling context. At the top
level, the instantiations are printed out.
We can summarize the operation of the Prolog proof process as follows:
1. Work on the leftmost subgoal first.
2. Select the first applicable rule.
3. Search the facts in the order they appear in the data base.
Careful attention to the order in which goals are written, rules are given, and facts are asserted
can improve the performance of the proof system substantially. It makes good sense, when defining
a multirule predicate, to make the first rule the one that will be used most frequently, if that is
possible. For example, in the gcd algorithm [Exhibit 13.13], the rule that ends the recursion must
be written before rule 2, to prevent rule 2 from being invoked with a zero divisor, which would
generate an error. However, rules 2 and 3 could be written in either order, and rule 3 is placed last
because it will be executed at most once, to swap the arguments on the first recursion, in the case
that the first argument is smaller than the second.
The programmer must also be careful of the order of the conditions within a rule. Since these
conditions are taken, in order, as subgoals, and variables become instantiated as the subgoals
succeed, a condition that is intended to instantiate a variable must precede all conditions that
expect that variable to be instantiated. This is like saying, in a procedural language, that variables
must be initialized before they are used. Further, because the conditions are tested in order, a
series of conditions is very much like a nested conditional. Sometimes a condition is simple to test;
sometimes satisfying it can involve a great deal of computation. Where two conditions must both
be true before proceeding with a rule, it is prudent to write the simple one first, so that, if it is
false, it will abort the rule before making the costly test. A simple condition can thus guard a
complex or difficult condition.

380

CHAPTER 13. LOGIC PROGRAMMING

Exhibit 13.13. Euclids algorithm in Prolog.

These three rules axiomatize Euclids greatest common divisor algorithm. The numbers on the left
are for reference only; they are not part of the Prolog code.
1. | gcd(A, 0, A).
2. | gcd(A, B, D) :- (A>B), (B>0), R is A mod B, gcd(B, R, D).
3. | gcd(A, B, D) :- (A<B), gcd(B, A, D).

Backtracking
The recursive descent from goal to subgoal ends successfully when a subgoal is reached that corresponds to a known fact in the data base. The instantiations used to satisfy the subgoal are passed
back up and are used to instantiate the variables in the goal.
The recursive descent ends in failure if the fail predicate is encountered while processing a
rule, or if the entire data base has been searched and no relevant information was found. Failure
of a goal does not mean that the goal is false, only that it cannot be proven from the facts given.
When a subgoal fails, control backtracks, that is, it passes back up to the level of the goal above,
and Prolog attempts to find a different set of instantiations that satisfy the original goal. In terms
of the tree of possible proofs, this amounts to backing up one path until an unexplored branching
point is found, then going down the new branch of the tree. Exhibit 10.36, at the end of Chapter
10, illustrates this recursive backtracking.

13.5.5

Functions and Computation

A programmer trying to solve a problem or model a system starts with a set of inputs, in some
given form, and wishes to derive from them a set of outputs. In a language like C, this is done by
defining functions to carry out a series of calculations and manipulations on the data that produce
a result in the desired form. We call this procedural programming.
We might view the same problem in a different way. Instead of specifying the method to reach
the desired output, we could describe the output desired, in a declarative, axiomatic language.
Prolog is a language for axiomatizing a desired result. The process of computation, in a C program,
is replaced by the process of proof in Prolog, where a proof consists of finding a data object that
satisfies the formula for the result. Thus ordinary computation, as well as data base searches, can
be expressed in Prolog.
Computation
In place of a C function that returns an answer, a Prolog programmer writes a predicate with one
additional parameter, which will be used for returning the answer. It is customary to write the
output parameter or parameters last. Instead of instructions to perform a computation, the Prolog
programmer writes axioms to verify that the answer is correct. During the proof process, the

13.5. PROLOG

381

Exhibit 13.14. Euclids algorithm in C.

int gcd(int A, int B)
{
int D;
if (B >A) D=B, B=A, A=D; /* swap */
do D=(A % B), A=B, B=D;
while (D >0);
return A;
}

output parameter names become bound to values that have been verified, and these values are then
returned to the calling environment. If the predicate involves an actual arithmetic computation,
one of its rules will contain an is clause, which directs Prolog to actually evaluate an expression
and bind its result to a variable so that it can be returned.
An example of ordinary computation expressed as a Prolog predicate is the implementation of
Euclids algorithm for calculating the greatest common divisor of A and B [Exhibit 13.13]. This
recursive definition has three rules. The first rule stops the recursion when the gcd has been found.
The second rule does the work. It is a recursive axiom that says that the gcd of two numbers, A
and B, is also the gcd of B and (A mod B), and the gcd of every remainder calculated. The third
rule is invoked only when the second argument is larger than the first, and it simply reorders the
arguments so that the larger one is first.
Let us compare the Prolog gcd program with the same algorithm expressed in C [Exhibit 13.14].
We see that the C program requires more syntactic detailssuch things as type declarations for
the parameters and an explicit return statement. As in Miranda, Prolog variables are untyped,
but objects are typed and their types are deduced by the system. Return values are implemented
as output parameters in Prolog, and the value is returned by Prologs action (during unification) of
instantiating the output parameter.
The three Prolog rules for gcd are exactly echoed in the remaining three lines of the C program.
Rule one corresponds to the whileboth stop the recursion/iteration when the answer has been
calculated. Rule two corresponds to the doboth perform one mod operation and shift the arguments to the left for the next recursion/iteration. Finally, rule three corresponds to the ifboth
just swap the inputs. Thus we see that an axiomatic expression of an algorithm can look and work
very much like a procedural formulation.
Sorting
Algorithms, such as sorts, that manipulate data can also be expressed axiomatically. An axiomatization of quicksort is shown in Exhibit 13.15. The quicksort function is expressed as a recursive
function with three rules. Rules 4 and 5 are the base cases for the recursion, and they stop it when

382

CHAPTER 13. LOGIC PROGRAMMING

Exhibit 13.15. Quicksort in Prolog.

1. |
2. |
|
3. |
|

split(_,[],[],[]).
split(Pivot, [Head|Tail], [Head|Sm], Lg) :Head<Pivot, split(Pivot, Tail, Sm, Lg).
split(Pivot, [Head|Tail], Sm, [Head|Lg]) :Pivot<Head, split(Pivot, Tail, Sm, Lg).

4. |
5. |
6. |
|
|
|
|

quicksort([],[]).
quicksort([Head|[]], Head).
quicksort([Pivot|Unsorted], AllSorted) :split( Pivot, Unsorted, Small, Large),
quicksort(Small, SmSorted),
quicksort(Large, LgSorted),
append(SmSorted, [Pivot|LgSorted], AllSorted).

the list to be sorted has zero elements or one element. Rule 6 does most of the work. It separates
the argument (a list) into three parts (smaller values, pivot, and larger values), sorts the parts, then
appends them together in sorted order. Quicksort calls a subroutine, named split, that actually
does the separation.
The task of split is to separate the list into a sublist of small values and a sublist of large
values by comparing each list element to the Pivot. The input arguments are Pivot, the first value
on the sublist that is being sorted, and Unsorted, the rest of that sublist. The value of Pivot
will be compared to each element of Unsorted, and the elements of Unsorted will be divided into
two lists, those smaller than Pivot and those that are larger. The last two arguments, Small and
Large, are output parameters that split will use to return the sorted Small and Large sublists.
Split is implemented recursively, by three rules. Rule 1 is the base case for the recursion; if
the argument is the null list, it will return immediately with null lists as its output parameters.
Each time split is called with a list of at least one item, either the second rule or the third will be
used. In both cases, the symbol Head is bound to the first item remaining unsorted list, and Tail
is bound to whatever remains. On the next call, the first item of Tail will be peeled off. This will
continue until the remaining tail is null, at which time the first rule for split will terminate the
recursion.
The second rule for split is applied whenever the first item on the Unsorted list is smaller
than the Pivot. This first item is peeled off and remembered, then split calls itself recursively
to process the items remaining on the unsorted list. When this recursive call returns, all items
smaller than the Pivot will be on the list Sm, and all larger ones will be on the list Lg. Since the
Head element was smaller than Pivot, it belongs on the Sm list, and so it is appended to the front
of that list. The extended Sm list and the Lg list are then returned to the calling routine.

13.5. PROLOG

383

Exhibit 13.16. Quicksort in C.

Notes on this code are given in Exhibit 13.17.
int * split( int *first, int *last )
{
int *small;
/* small is a left-to-right scanner.
int *large;
/* large is a right-to-left scanner.
int swap, pivot = *first;
int *scan;

*/
*/

for( small=first, large=last+1; ; )

{
while (* ++small < pivot); /* Scan until large item is found. */
while (* --large > pivot); /* Scan until small item is found. */
if (small >= large) break; /* Quit if scanners have crossed. */
swap =*large; *large=*small; *small=swap;
}
*first = *large; *large = pivot;
return large;
/* This marks the split point. */
}
void quicksort(int * first, int * last)
{
int * split_point;
if (last<=first) return;
/* Only one item -- no action needed.
split_point = split (first, last);
quicksort (first, split_point-1);
quicksort (split_point+1, last);

*/

The third rule for split is applied whenever the first item on the Unsorted list is larger than
the Pivot. Its operation is just like the second rule, except that, because the Head element is bigger
than the Pivot, it is concatenated to the Lg list.
Let us look at how the Prolog proof system is being used here to implement a sort procedure, and
compare it to a quicksort program written in a procedural language. In C, the code for quicksort
uses recursion, explicit iteration, explicit conditionals, comparisons, pointers, and assignments [Exhibits 13.16 and 13.17]. In contrast, the Prolog code uses only recursion and comparison. How
is the rest of the work done? First, note that in both languages, the main quicksort routine is
recursive. However, the split routine is tail-recursive in Prolog but iterative in C, for efficiency.
The act of calling a routine recursively looks just the same in the two languages, and it has the
same semantic effect.
The data structures used are different; the Prolog version sorts a list, while the C version sorts

384

CHAPTER 13. LOGIC PROGRAMMING

Exhibit 13.17. Notes on the C quicksort.

1. The parameters first and last must be pointers to the beginning and end of an array of
integers. On exit from quicksort, the values in this array are in sorted order. Pointers,
rather than subscripts, are used to index the array.
2. A sentinel value equal to maxint must follow the last data element to stop the left-to-right
scanning pointer in the case that the pivot value is the largest value in the array.
3. The inner while loops identify the leftmost and rightmost elements that are stored in the
wrong part of the array. These two elements are then swapped.
4. After exit from the for loop, all items to the left of small are small (less than or equal to
pivot value) and all items to the right of large are large. The large scanner points to a
small element, and small either points to the same thing or points to the large element on
larges right.
5. Before returning, the pivot element is swapped into the middle, between the small and large
elements. This is its final resting place. The small and large areas remain to be sorted by
future recursive calls on quicksort.

an array. The action of moving through the elements of the list is accomplished in Prolog by
recursively binding a local name, Tail, to the list with its head removed. This recursive binding
takes the place of the increment (++) and decrement (--) operations in C, both of which have
assignment as a side effect.
Binding and concatenation are used in Prolog in place of the assignment operations in C. As
Prologs split pulls each item off the unsorted list, it binds the item to a local variable name,
Head. Since the function is called recursively, once for each element on the unsorted list, enough
bindings are created to hold all the list elements. As each call to split returns, the bound item is
appended to the list returned by the lower level. In contrast, the C version uses iteration instead
of recursion to perform the split operation. New storage is not created; rather, assignment is used
to swap pairs of values in the original storage area.
The sequential execution of C statements is echoed exactly in the sequential application of the
subgoals in each Prolog rule. Thus the main routine, quicksort, looks almost the same in the two
languages.
Finally, in the C code, explicit if statements are used to end recursion in quicksort and to end
the split operation, and while statements are used to determine whether an element belongs in
the small or the large part of the array. In Prolog all this is accomplished by the elaborate pattern
matching algorithm (unification) that is built into the proof system. Prolog uses unification to

385

13.5. PROLOG

select which rule to apply when split is called, which determines whether the next step will add
an element to the small list or the large list, or end the recursion.

13.5.6

Cuts and the not Predicate

Two major theoretical results have had a strong bearing on Prolog: clausal logic is complete but
not decidable. So although every true clausal theorem can be proved, no effective procedure can
ever exist that will always produce a proof and terminate in a finite amount of time. This means
that if the Prolog proof system relied on resolution alone, a programmer might not know of not
knowing whether a given query would ever be answered. Prolog does have a way, called a cut, to
control the proof process so that a programmer can avoid being trapped in lengthy deductions that
seem likely to be fruitless. However, when the cut operation is used for this purpose in Prolog, it
destroys the completeness of the proof system and leaves open the possibility that a provable goal
might fail.
Cuts
A cut is written as ! and may appear as one of the conditions in a rule. 3 Informally, a cut prunes
off one branch of the proof-search tree by telling the proof system to abandon a chain of reasoning
under certain conditions. In some ways it is analogous to a break instruction.
Perhaps the best way to think of a cut is to imagine that it is a barrier, placed by the programmer
in a rule, to stop fruitless backtracking. Consider a rule with several terms:
P : Q, R, S, !, T, U, V.
In trying to satisfy this rule, the proof system starts by searching for a unification of conditions Q,
R, and S. Backtracking might occur several times during this search, and control might go back as
far as condition Q. If Prolog fails to satisfy this part of the rule, it will go on to try the next rule
for P . However, if the conditions Q, R, and S on the left are eventually satisfied, control passes
through the cut to the conditions T , U , and V on the right. At this point, all variable bindings for
conditions Q, R, and S are frozen, or committed, and the information that would permit Prolog to
backtrack back through these conditions is discarded.
The proof system now begins to try to find a unification for conditions T , U , and V that is
consistent with the frozen bindings. Again, a great deal of backtracking can happen among these
clauses, and, perhaps, some unification of the whole rule may be found. In this case, the rule
succeeds, and a unification is returned. If (during backtracking) control ever returns to the cut, it
means that the attempt to unify conditions T , U , and V with the frozen bindings has failed. At
this point, the two pruning actions of the cut take place:
1. Instead of returning to reinstantiate the left part of the rule, the entire rule fails immediately.
3

Do not confuse this meaning of the term cut with the meaning of cut in a resolution step of clausal logic.

386

CHAPTER 13. LOGIC PROGRAMMING

Exhibit 13.18. Cutting off a search.

These predicates axiomatize what it means to be on academic probation at a hypothetical university.
Loosely, the requirement is that the closer a student is to graduation, the closer the students grade
point average must be to the minimum gpa for graduation, which is 2.00. (We use the integer form,
200, rather than the decimal form, 2.00, in this code.) Each student is represented in the data base
by a fact of the form shown in Exhibit 13.19.
1. |
year(S, Y) :- student(S,Y,_).
2. |
gpa( S, G) :- student(S,_,G).
3.
4.
5.
6.
7.
8.

|
|
|
|
|
|

probation(S,
probation(S,
probation(S,
probation(S,
probation(S,
probation(S,

X):-year(S,fr), !, gpa(S,X), X<150.

X):-year(S,so1),!, gpa(S,X), X<160.
X):-year(S,so2),!, gpa(S,X), X<170.
X):-year(S,ju1),!, gpa(S,X), X<180.
X):-year(S,ju2),!, gpa(S,X), X<190.
X):-year(S,se), gpa(S,X), X<200.

2. The goal that caused this rule to be processed also fails and no more attempts are made to
satisfy the predicate, even if there are more, untried rules for it.
Safe Cuts. A safe cut is one that cannot possibly cause a provable goal to fail. These are used for
the sake of efficiency, in situations where the conditions that guard the various rules for a predicate
are mutually exclusive. This is illustrated by the code in Exhibit 13.18, which determines whether
a student should be put on academic probation. (Students are listed in Exhibit 13.19.) The grade
point average needed to avoid probation becomes higher each year. Thus one rule is included in
the probation predicate for each grade-point level involved.
If we ask probation(tal), the query will fail, because tal is a fine student. In the process
of answering this query, Prolog will have looked at all the rules for the predicate probation and
will have failed to satisfy the year(S,Y) term on all but the last rule. In this case, the cuts
in the prior rules have no effect. Similarly, the cuts have no effect during processing of the query
year(S,se), probation(S,G), and Prolog will return with the instantiation S=rae, G=195
indicating that rae is, indeed, in trouble.

Exhibit 13.19. Data for the probation predicate.

student(ali, so1, 195).
student(jess, fr, 142).
student(mark, so1, 152).
student(rae, se, 195).

student(dale, ju2, 189).

student(ken, fr, 199).
student(nan, so2, 170).
student(sal, ju2, 298).

student(jan,
student(les,
student(pat,
student(tal,

fr, 372).
so2, 315).
ju1, 175).
se, 400).

13.5. PROLOG

387

Exhibit 13.20. The not in Prolog.

Definition of not:
1.
2.

|
|

not(X) :- X, !, fail.
not(_).

The condition not(X) succeeds if X fails, and fails if X succeeds.

Similarly, there is no inefficiency problem if we ask probation(jess,G). Prolog tries rule 3

first, finds that it describes jess well, and succeeds with the instantiation G=142. No backtracking
ever happens.
However, if we ask probation(jan,G)., Prolog finds that rule 3 for probation should be
processed because jan is a freshman. It then finds jans gpa, compares it to the minimum for
freshmen, and fails. This failure initiates backtracking, and control backs up to the cut without
finding any other instantiations for X. If the cut were not there, backtracking would continue; rule
3 would then fail, and rules 4 through 8 would all be tried and fail on the year condition. All this
work is nonproductive because a student who is a freshman cannot simultaneously fall into any of
the other categories. The presence of the cut in rule 3 acts like an else clause and eliminates the
fruitless testing of all the other rules. In a chain of exclusive conditions, therefore, each rule except
the last should be written with a cut.
Cuts Implement the Operator not. Perhaps the reader has noted the conspicuous absence of
not in any of the examples given. One of the restrictions placed on both rules and queries is that
the conditions must all be nonnegative. This is required in order to permit the use of resolution as
a proof system.
Of course, several of the built-in predicates have a negative form alsofor example, we have
both = and \=. The Prolog language does include a built-in not operator which can be applied
to a condition term. However, not is problematical and the results obtained by using it can be
misleading. Consider what happens during processing of not, whose definition is shown in Exhibit
13.20:
If X is false, rule 1 fails before passing through the cut.
In that case, rule 2 is used. This rule always succeeds, no matter what its argument is, and
it instantiates nothing.
If X is true, control passes through the cut in rule 1 and comes to the fail. This term always
fails, initiating backtracking.
When the backtracking reaches the cut, rule 1 fails, and since rule 2 is not used because of
the cut, the predicate fails.
The result is that not(X) succeeds if X fails and fails if X succeeds.

388

CHAPTER 13. LOGIC PROGRAMMING

Exhibit 13.21. The trouble with not.

Suppose we have the simple predicate test:
test(S, T) :- S = T.
The problems with not are illustrated by this transcript of a test run:
?- test(3, 5).
no
?- test(5, 5).
yes
?- not( test(5, 5) ).
no
?- test(X, 3), R is X+2.
X=3
R=5.
?- not( not( test(X, 3))), R is X+2.
!error in arithmetic expression: not a number.
The responses to the first four queries are exactly what one expects. However, the final one is a
surprise, since the query is the double-not of the preceeding query, but the response is different!

The disturbing fact about not is that the results of evaluating the predicates P and not(not(P))
are not necessarily the same, and the results of evaluating the predicates P and not(P) are sometimes
the same! This is illustrated by the simple example in Exhibit 13.21. This difficulty is caused by
the fact that not is implemented using a cut, and when we backtrack to a cut, the goal fails. Thus
we see that the last query in Exhibit 13.21 fails, even though it seems that it should not.
One last difficulty with not arises from the fact that clausal logic is not decidable. In practical
terms, this means that sometimes we can prove a theorem, T , sometimes we can prove its negation,
not T , and sometimes we cannot prove either! In the last case, the theorem is true in some models
of the theory but not in others, and we certainly cannot conclude that T is false just because we
cannot prove that T is true. However, when Prolog cannot prove that T is true, the predicate
not(T ) succeeds. This is so, even if the reason for the failure of T is that the data base contains
too little relevant data or that the programmer made a program-sequencing error! Obviously, we
must be extremely careful when using not.
Unsafe Cuts. Sometimes a programmer might decide to use a cut when implementing some
heuristic part of a computation. In this case, the programmer knows that there is some possibility
that the desired solution may lie on that part of the proof-tree that is being cut off. Use of the
cut for such purposes is considered to be impure, since it destroys the completeness of the proof
system. However, in many artificial intelligence and optimization problems, it may not be possible,

13.5. PROLOG

389

because of time constraints, to fully explore the proof-tree. Pruning the tree might eliminate the
best solution, and it might even eliminate the only solution, but if the heuristics are skillfully chosen,
these unhappy outcomes can be largely avoided. On the positive side, skillfully used heuristic cuts
can dramatically speed up processing and reduce the memory requirements of a computation to be
within practical limits.

13.5.7

Evaluation of Prolog

Prolog will not be appropriate in applications where efficiency is a major consideration. Its performance is limited by its interactive, interpretive nature, and by the lack of destructive assignment
operations. All computation is done through parameter binding, and parameter binding is done
by the complex and relatively slow unification algorithm. Further, the programmer has only limited control over execution order. Ordinary applications that can be programmed directly in a
procedural language such as C will be more efficient than the same applications in Prolog.
However, Prolog is a very attractive language for applications in which the programmer does
not know how to organize the data or the computational process. It can be used to express some
kinds of information much more directly than the common procedural languages, and it makes it
possible to integrate procedural program elements with nonprocedural ones.
Finally, because order of evaluation is left largely unspecified, Prolog is useful for applications
in which parallel evaluation is required. It will be an appropriate language for implementation on
computers with highly parallel architecture, such as the Connection Machine.

Exercises
1. What is the difference between computation and deduction? In what sense are they analogous
processes?
2. What is a universe of discourse?
3. What does it mean to instantiate a variable? How are variables instantiated in the process
of deduction?
4. What is a term? A predicate? A sentence? Make clear what the differences are among these
concepts.
5. Give an example of a 3-ary predicate.
6. Given the predicates in Exhibit 13.1, are the following instantiated predicates true or false?
a. odd(13)
b. odd(7+3)
c. divide(45,5,9,0)

390

CHAPTER 13. LOGIC PROGRAMMING

d. divide(56,5,11,3)
e. father(Isaac, Abraham)

7. What is the difference between the meaning of X and X ?

8. Which of the following quantified predicates is true? What is the negation of each?
a.
b.
c.
d.
e.

Y even(Y)
X even(X)
X X=X+0
X Y X=Y+1
X Y X=Y*1

9. Define and explain the relationship between axioms and theories.

10. How is it possible for a theory to have more than one model?
11. What is a proof? A refutation?
12. Write an example of a Horn clause in the form of a disjunction of terms. Write a logical
sentence that is not a Horn clause.
13. If a theory has only unconditional axioms (no conditional axioms), the theory has very few
theorems and all the proofs are quite boring. Explain.
14. Write examples of conditional and unconditional Horn clauses. Write a Prolog sentence that
corresponds to each.
15. What are the elements of the Prolog programming environment?
16. A Prolog rule specifies a set of constraints that must be simultaneously true to draw a conclusion. Explain how each part of a rule fits this description.
17. Given the data base in Exhibit 13.10, how would Prolog respond to the following queries?
(a)
(b)
(c)
(d)

a.
b.
c.
d.

pretty(violet).
watchout(holly).
watchout(X),pretty(X).
watchout(X),color(X,green).

18. Prolog cannot always prove a true assertion and it cannot always disprove a false assertion.
Explain.
19. In Prolog, a fact is a single nonnegated term, such as pretty(mary). The same term, or a
similar one such as pretty(pat) or pretty(X), can be a query. However, a term that is a
fact and one that is a query have different meanings. Explain.

391

13.5. PROLOG
20. Given these facts and rules:
facts
flower(crocus, spring, white).
flower(violet, spring, blue).
flower(iris, summer, blue).
flower(rose, summer, red).
flower(marigold, summer, orange).
tree(holly)

rules
color(F,C) :- flower(F,S,C).
color(T,green) :- tree(T).
pretty(F) :- color(F,red).
pretty(F) :- color(F,blue).
grows(X) :- tree(X).
grows(X) :- flower(X,Y,Z).

Find all sets of instantiations that unify each of the following pairs of clauses:
(a) a. flower(F,spring,Y), pretty(F).
(b) b. flower(F,summer,Y), pretty(F).
(c) c. grows(X), pretty(X).
(d) d. flower(Z,Y,orange), pretty(Z).
21. Discuss the difference in efficiency between the C quicksort [Exhibit 13.16] and the Prolog
version [Exhibit 13.15].
22. Compare and contrast the operation and efficiency of the Miranda gcd algorithm [Exhibit
12.12] and the Prolog version [Exhibit 13.13].
23. Compare the clarity of code in the C quicksort [Exhibit 13.16] and the Prolog version [Exhibit
13.15]. Which do you think is easier to understand? Why?
24. What are the syntactic similarities and differences between the Miranda quicksort script [Exhibit 12.19] and the Prolog quicksort [Exhibit 13.15]?
25. Compare and contrast the operation and efficiency of the Miranda quicksort script [Exhibit
12.19] and the Prolog quicksort [Exhibit 13.15]?

392

CHAPTER 13. LOGIC PROGRAMMING

Chapter 14

The Representation of Types

Overview
We began the discussion of types in Chapter 5 by the primitive types supported by
languages and their connection with the hardware types supported by typical computers.
In this chapter we consider the implementation of both primitive and user-defined types.
Finally, we continue in Chapter 15 with a discussion of the semantics of types.
To augment the primitive types, most languages permit types to be defined by the
programmer. A type definition enables the programmer to define the physical properties
of a new type and to name it. A type object represents this type information inside the
translator. It is composed of three parts: a name, a type, and a body of information.
Various types of types include: primitive, array, record, and enumerated. Each kind of
type has its own declaration syntax and corresponds to a distinct type of type object
within the translator.
We consider simple types (enumerated types, constrained types, pointer types) compound types (arrays, strings, sets, records), and union types (free and discriminated).
Operations which can be performed on compound objects include value construction
the combination of a set of components into a single compound object, selectionwhich
enables the programmer to reference a part of the compound object, and dereferencing
which maps a reference into a program object or value. Modern programming languages
implement some or all of these to various degrees.

393

394

CHAPTER 14. THE REPRESENTATION OF TYPES

14.1

Programmer-Defined Types

A data type is a set of objects with an associated set of functions that permit us to manipulate
the objects. To use a type we must be able to represent those objects and functions in our source
code and in the computer. A type definition gives us ways to do both. It supplies all information
about the physical properties of the new type, and gives names to the type itself, and, if it is a
structured type, to its parts. From this information, a translator can build constructors, selectors,
and predicates that let us create and manipulate objects of the new type.
In this section, we examine the ways in which types can be represented in the source code and
within a translator.1 A new type representation may be defined by:
Listing all its members.
Placing restrictions on an existing type.
Combining elements from existing types into an aggregate.
We will look at typical forms for these kinds of type declarations and see how a translator uses the
information they provide.

14.1.1

Representing Types within a Translator

Whether a type is primitive or defined by the programmer, the properties of the type are stored by
the translator when the type is defined. For the time being, let us imagine that all this information
is stored in one place, and let us use the term type-object to refer to the collection of information that
the translator stores about a type. In reality, depending on the language being implemented, the
type information could be stored coherently in a type-object, or it could be scattered throughout
various translator tables. Moreover, the amount and nature of the information stored varies from
language to language, even for similar data types.
There are various types of types: primitive types, array types, record types, enumerated types,
and others. For each type of type, a different collection of information must be supplied by the
programmer and stored by the translator. Thus a type-object can be seen as a three-part entity,
having a name (usually2 ), a type, and a body of information, which we will refer to as the body.
The type of a type tells us how to interpret the body of the type, and the body of the type tells
us how to interpret objects of that type. Every kind of type declaration supported by a language
corresponds to a distinct type of type within the translator.
A data object also consists of a name, a type, and a body of information. The name is kept in
the symbol table, and the type, represented by a pointer to a type-object, may be attached either
to the name or to the body, depending on the language. Let us diagram objects as shown in Exhibit
1

Chapter 15 covers the semantics of types and the ways in which distinct meanings, or semantics, can be given to
types with the same representation. In this section we are concerned primarily with the representation of types, not
the semantics of similarly represented types.
2
In some languages, it is possible to create a type with no name.

395

14.1. PROGRAMMER-DEFINED TYPES

Exhibit 14.1. Object diagrams.

objects name
type body

14.1. The object is represented by a T shaped figure, with the objects name (if there is one)
written on top, the type on the left, and the body on the right. The type of an object is always a
pointer to some type-object. The body of a variable is a storage object and will be diagrammed as
a box. The body of a constant or a type is a series of data values.
The body of a type must contain enough information to support whatever semantics are defined
for the type. The nature and amount of this information varies among the different types of types,
and from language to language. Examples are: the base type of a pointer or array type, the
dimension of an array, and the order, types, and names of the fields of a record type. In addition,
a type body might contain extra or redundant information, designed to make use of the type more
efficient. One such redundant fact might be the total size, in bytes, of an object of this type. While
this might be calculated from other information, it is used frequently and should be kept handy.
We will define (somewhat arbitrarily) general and useful type-objects for each type of type
as we consider it. Although examples of type declarations are drawn from several languages, we
should emphasize that the type-objects diagrammed do not necessarily reflect existing translator
mechanisms for those languages. However, they would be appropriate as part of the semantic basis
of a new object-oriented language with similar features.
Type-objects for the primitive types are defined by the compiler itself. We will diagram a
primitive type body by listing only the size, in bytes, and denote the rest of the types definition
as a code definition. An integer variable named year is diagrammed, with its type, in Exhibit
14.2.
Exhibit 14.2. An integer object and its type.

integer
2 bytes
code definition
primitive
type

year
1935

396

CHAPTER 14. THE REPRESENTATION OF TYPES

14.1.2

Finite Types

When a type is semantically unrelated to existing types and has a small number of members, it
is practical to define it by simply listing, or enumerating, identifiers for the members of the type.
These are called the type constants.
If a language supports enumerated types, it must also provide some functions that are automatically defined for every enumerated type. Comparison for equality must be supported. Other
useful and common functions include comparison for inequality, successor, and predecessor. Input
and output routines are sometimes supplied.
The programmer supplies only identifiers for the type constants; the translator must create an
encoding for them. The obvious encoding is to use the integers, in order, starting with zero, to
represent the type constants. This encoding makes it easy to implement comparison and successor
functions; the corresponding integer functions are simply carried over to the new type. The standards for both Pascal and ANSI C explicitly state that the enumerated type will be represented by
integers.3
As is common when languages are compared, type declarations that are syntactically very
similar can create types with widely varying semantics. C is a language at one extreme. The
enumerated type declaration is no more and no less than a convenient way to define identifiers
for integer constants. An enumerated type, in C, is implemented by type int and has semantics
identical to int. This is consistent with the use of types in C primarily to allocate and access
data objects, rather than as a vehicle for semantics. Enumerated type constants are represented
as integers and are considered to be integers. Characters and truth values, the two primitive
enumerated types, are integers in C.
The treatment of enumerated types in Pascal is like C, with one very important exception: the
enumerated type is semantically distinct from the type integer, which is used to represent it. Type
integer is incompatible with enumerated types in Pascal. This means that you cannot, for example,
multiply two values of an enumerated type in Pascal, as you could in C! Nor can you mix values of
two different enumerated types in an expression.
There are some functions that are predefined for any enumerated type in Pascal. These are:
Succ (successor), Pred (predecessor), Ord (conversion to type integer), assignment, and all the
comparison operators. An enumeration constant or variable may also be used as a subscript.
These functions are enough to make enumerated types useful, but not enough to make their use
convenient. Unfortunately, Pascal lacks convenient means to read in and print out enumerated
constants. Thus every program that uses an enumerated type must contain code to perform input
and output conversions. Often this takes the form of two rather tedious CASE statements.
We can only speculate why enumerated I/O is not supported in Pascal. One possibility is that
the idea was so new when Pascal was designed that Wirth did not realize that omitting I/O would
limit the uses of enumerated types. A more likely explanation is the quest for simplicity. Pascal
I/O is very simple and less flexible than the formatted I/O functions provided by many languages.
Pascal provides less control over detail than FORTRAN, C, or APL. Adding enumerated I/O would
3

C permits programmers to use the default codes or assign their own integer codes.

14.1. PROGRAMMER-DEFINED TYPES

397

Exhibit 14.3. A type-object for an enumerated type.

TYPE warm = (red, maroon, magenta, pink, coral);

name
size
list of identifiers
enum
type

warm
1 byte
"red","maroon","magenta","pink","coral"
enum
type

have complicated a simple and elegant design, and may have been considered too unimportant to
justify this cost.
To provide I/O for enumerated types, a language definition would have to include format descriptors for type constants. A translator would need the constant identifiers available at run time,
so that inputs could be encoded and outputs decoded, automatically. Thus a type-object for an
enumerated type would need to contain a list of pointers to the names (represented as character
strings) [Exhibit 14.3].
Sending type constants to the output stream is easy enough, but some uniform way would be
needed to recognize them in the input stream. All of these things are easy enough to define and
implement, and support for input and output would make enumerated types considerably more
useful.
An implementation of enumerated type I/O would increase the complexity of the compilers
I/O system and increase the amount of system code that would be included with a program at
run time. In a program that used enumerated types, there would be a corresponding decrease in
programmer-generated code, because programs would not need to manually encode and decode the
enumerated constants. Overall, the costs seem modest and the benefits real.

14.1.3

Constrained Types

Some languages permit a programmer to define a new type by applying constraints to an existing
type. This is a powerful tool for expressing semantic intent. If a language supports constrained
types, the run-time system for that language must check each value of the type to ensure that it
obeys the constraint. The type-object must therefore include the base type and the limiting values.
Pascal and Ada provide subrange type declarations which define a new type whose members
are a consecutive subset of an existing simple type. The programmer specifies the initial and final
values that will belong to the new type. (The base type of these values is implicit.) Exhibit 14.4
shows a type-object for a Pascal subrange type.
Values of the constrained type and the underlying base type are compatible and may be combined in operations. Functions for either type may be applied to the other. However, a computed
value is checked (at run time) before assigning it to a constrained variable; any violation of the

398

CHAPTER 14. THE REPRESENTATION OF TYPES

Exhibit 14.4. A constrained type in Pascal.

TYPE age = 0..150;

name
size
base type
subrange limit values
type

age
2 bytes
subrange 0,150
type

integer

constraints causes a run-time error.

C does not support subrange types. This is consistent with the general philosophy that C types
are used to define storage configurations, not semantics. A constraint has no effect on the size or
the encoding of the representation; it is relevant only to the semantics of values of that type. (A
value may be acceptable or not in the constrained type, while all representable values are legal for
the base type.)

14.1.4

Pointer Types

All pointers represent machine addresses, so the semantics of a pointer type do not depend on
the semantics of its base type. (The base type of a pointer is the type of the object to which the
pointer points.) Moreover, the size of a pointer is quite independent of the base type, so no storage
layout information is needed in the body of a pointer type except the size of pointers on the target
machine. Any pointer value will physically fit into any pointer variable.
This innate relationship among pointer types is exploited in FORTH. It supports pointers in
the sense that the address of any object can be obtained and stored in an integer variable. Pointer
arithmetic (implemented in terms of integer arithmetic) is possible, and integer variables may
be declared and used for pointer values. However, no distinction is made between pointers with
different base types.4
Most languages, however, require the programmer to declare the base type of every pointer.
Therefore, we will diagram a pointer type as shown in Exhibit 14.5. Information about the base
type of a pointer is not needed for allocation or for access. It is required only by languages whose
compilers use that type information to determine the semantics of the program. Even among
languages that require declaration of a pointers base type, that type information may be used in
the following different ways:
1. To compile the correct meaning of ambiguous operators, such as +.
2. To determine the number of bytes to fetch when a pointer is dereferenced and used in an
expression.
4

Indeed, no distinction is made between pointers and integers!

399

14.1. PROGRAMMER-DEFINED TYPES

Exhibit 14.5. A C pointer and its type-object.

We declare and initialize an integer pointer, pk, to point at an integer, k, which is initialized to the
value 5.
int k=5, *pk=&k;

pk
4 bytes
pointer
type

k
integer
2 bytes
code definition
primitive
type

3. To determine the type of the result of a dereference, so that it may check for type errors.
4. To dynamically allocate space for an object of the base type.
Similar declarations can have very different meanings in different languages. For example, in C,
the base type is used only for purposes 1 and 2. In the following list, assume p is a pointer whose
base type is BT:
1. Arithmetic on pointers is defined in terms of the base type. The expression p+1 will cause p
to be incremented by the length of an object of type BT.
2. The meaning of dereference and fetch depends on the base type. The expression *p +1 will
cause a fetch operation from the memory location stored in p and increment that value by 1.
The number of bytes fetched will depend on BT, which must be a simple, primitive type.
3. In general, C does not perform type checking.
4. The type of the pointer is not used for allocating base-type objects. To allocate a base-type
object dynamically, one refers directly to the name of the base type, not to the pointer.
In contrast, Pascal uses the same information for the last three purposes. In the following list,
assume p is a pointer whose base type is BT, and b is an object of type BT:
1. There are no operations on pointers whose meaning depends on the base type. Dereferencing
is the only pointer operation that is supported, and its meaning is the same for all pointers.
2. Any type of pointer may be dereferenced and used in an assignment. The statement b :=
p; copies n bytes, where n is the size of BT.

400

CHAPTER 14. THE REPRESENTATION OF TYPES

3. Pascal is a fully type-checked language. In the expression b + p * 2, the base type of p

is checked to ensure that it is appropriate for the surrounding expression.
4. The argument to a call on NEW (which performs dynamic allocation in Pascal) must be a
pointer, p. An object is allocated of ps base type, and the resulting reference is stored in p.

14.2

Compound Types

14.2.1

Arrays

Modern languages permit simple types to be combined to form compound, or structured, types.
These combinations can be positional (arrays, strings) or not (sets), and homogeneous (arrays,
strings, sets) or heterogeneous (records). They can have fixed size (arrays, records) or variable size
(strings, sets). Let us look at typical type declarations, type-objects, and accessing methods for
these compound types.
An array is a fixed-length sequence of elements of a single type, called the base type. The length
of the sequence is called its dimension. Elements of an array are accessed by position, and the
position values of the first and last elements are called the array bounds. In theory, array positions
and bounds could be values of any simple, discrete type, called the index type. Possible index types
include the integers and any type that is implemented by mapping onto the integers. In practice,
various languages are more or less restrictive about the index types that can be used. For simplicity,
in this discussion, let us presume that integers are used for the index type.
To access an individual element of an array, we append a subscript to the array name. This
is an expression (usually enclosed in parentheses or square brackets) whose value falls within the
bounds of the array. The subscript, base address, and base type of the array are used to compute
the effective address, or address of the desired component. This computation is simplest and most
efficient if the language requires zero-based subscripting. In this system, if an array has dimension
D, its bounds are 0 . . . D 1. The effective address formulas for zero-based and arbitrary-based
subscripting are shown in Exhibit 14.6.
In the old days, when all computer memory was a series of words (not bytes) and all numbers
were one word long, arrays of numbers were implemented very simply and efficiently. The arrays
base address was loaded into the computers memory address register and the subscript was loaded
into an index register. Then the computers indexed-fetch or indexed-store instruction dynamically
computed the desired effective address.
Multidimensional Arrays
Some older languages, such as MAD, FORTRAN, and APL, specifically supported arrays of two or
more dimensions. (Following APL terminology, let us use the word rank to mean the number of
dimensions of an array.) The rank was limited, in FORTRAN, to the number of index registers
available on the host computer, since each subscript was kept in a register for efficiencys sake.

401

14.2. COMPOUND TYPES

Exhibit 14.6. Formulas for the effective address computation.

Let ba be the base address of the array A, and let size be the number of addressable units (bytes)
required to store a value of the base type of A. Then the effective address corresponding to A[s1]
is:
Zero-based subscript: ba + size s1
Other-based subscript: ba + size (s1 lower bound of A)
For a multidimensional array, with zero-based subscripting and declared dimensions d1, d2, . . . , dn,
the effective address corresponding to A[s1, s2, . . . sn] is:
ba + size ((((s1 d2 + s2) d3 + s3) . . .) dn + sn)

However, computing an effective address for an array of rank n cannot be done simply by adding
together all the subscripts; it requires n 1 multiplications and additions, plus one index operation
[Exhibit 14.6]. An example of a three-dimensional subscript computation for Pascal arrays is shown
in Exhibit 14.7.
Support for multidimensional arrays is important in languages that are intended for use by
scientists and engineers who work regularly with matrices. In these older languages, special syntax

Exhibit 14.7. Effective address computation for Pascal arrays.

Here we give declarations for two arrays of rank 3. Assume that storage for Matrix starts at
location 1000 and storage for Box starts at location 2000. In this implementation, the size of a real
is 4 bytes and the size of an integer is two bytes.
VAR Matrix: array[0..1] of array [0..2] of array [0..4] of real;
Box: array[1..2] of array [5..7] of array [-2..2] of integer;
Matrix has zero-based subscripts, so we use the simpler form of the formula to compute the
effective address corresponding to Matrix[1][2][3].
1000 + 4 * (((1*3 +3)*5 + 3)) = 1000 + 4 * 33 = 1132
Box does not have zero-based subscripts, so the lower bound for each dimension must
be subtracted from the corresponding subscript. Here we compute the effective address for
Box[2][7][-1]:
2000 + 2 *(( (2-1)*3 + (7-5) )*5 + (-1--2) ) =
2000 + 2 *( 5*5 + 1 )
=
2000 + 2*26
=

2052

402

CHAPTER 14. THE REPRESENTATION OF TYPES

Exhibit 14.8. A three-dimensional array in FORTRAN.

REAL AR3
DIMENSION AR3[5, 2, 4]
* Store a number in one element of the array.
AR3[2,2,1] = 17.0

and semantics were provided to support higher-rank arrays [Exhibit 14.8].

Type declarations were invented in the late 1960s, and this changed the way that higher-rank
arrays were implemented. Languages designed since then have array type declarations that define
arrays in terms of an arbitrary base type. Thus a two-dimensional array became, simply, an array
of arrays. Special syntax and special semantic rules were no longer needed to handle higher-rank
arrays; they could be handled by iterating the type definitions and type accessing syntax for rank
one arrays, as shown in Exhibit 14.9, item B.
This is a real simplification in the language. The syntax is simpler because the brackets written
to declare and use subscripts need to delimit only a single expression, not a list of expressions. The
semantics is simpler because the complex formula for computing a higher-rank effective address is
replaced by iterating the simple formula for computing a one-dimensional effective address. The
size of the various base types replaces the declared dimensions in the formula [Exhibit 14.10]. C,
was designed with streamlined simplicity in mind. It supports only zero-based subscripting and
only one-dimensional arrays. A higher-rank array is defined and referenced as an array of arrays.
Pascal, on the other hand, made an interesting concession to custom. Pascal has type declarations and supports arrays with an arbitrary base type, so it does not need to have FORTRAN-like
array notation. However, programmers were used to the multisubscript FORTRAN notation shown
in Exhibit 14.8. So Wirth compromised. The semantic basis of Pascal is like C: it supports only

Exhibit 14.9. Basic and sugared notation for arrays in Pascal.

A. Sugared notation:
VAR Sugared: array[1..3, 1..10, 0..3] of char;
c := Sugared[1, 5, 2];
B. Equivalent array declared using basic notation:
VAR Plain: array[1..3] of array [1..10] of array [0..3] of char;
c := Plain[1][5][2];

403

14.2. COMPOUND TYPES

Exhibit 14.10. Effective address computation for a four-level type.

Assume that storage for the variable named Plain from Exhibit 14.9 starts at address 1000.
Let us compute the effective address for Plain[1][5][2].
1.
2.
3.
4.

Type
char
array [0..3] of (1)
array [1..10] of (2)
array [1..3] of (3)

Dimension
4
10
3

Total Size
1 byte
4 bytes
40 bytes
120 bytes

Effective address = base address of Plain +

(1-1)*40 bytes + (5-1)* 4 bytes + (2-0)*1 byte =
1000 + 0 + 16 + 2 = 1018

one-dimensional arrays. But the syntax of Pascal was extended to include FORTRAN-like subscript
notation, with all subscripts (optionally) included between a single pair of brackets. This extra
declaration form is called syntactic sugar because it is unnecessary but makes the language sweeter
and more attractive for many users. The multidimensional notation that the programmer writes
is converted by the parser to the basic notation [Exhibit 14.9]. Of course, the programmer may
always choose to write directly in the basic notation. A Pascal code generator contains semantic
interpretation routines for the one-dimensional notation only.
Ada Arrays and Slices. In the treatment of arrays, as in many other ways, Ada is like a greatly
complicated (and more capable) version of Pascal. Like Pascal, Ada provides syntactic forms for
declaring both arrays of arrays and multidimensional arrays. Unlike Pascal, though, the two forms
are not equivalent! Ada provides an operation called slicing 5 for arrays of arrays that cannot be
applied to multidimensional arrays.
A slice of an array is the contents of a consecutive series of locations, denoted by the array
name with a range of subscripts. If Ar is an array with bounds 1. . . 10, then the slice from positions
3 through 5 is denoted thus:
Ar(3..5)
A slice can be used in an assignment or a function call. Slicing provides a nice way to do some
array operations coherently that would need to be written with a loop in Pascal or C. However, Ada
slices are not as flexible or as powerful a tool as the APL subscripting facility, described in Section
14.3.2.
5

This term was introduced by ALGOL-68.

404

CHAPTER 14. THE REPRESENTATION OF TYPES

Exhibit 14.11. Diagrams of array type-objects.

The type declaration is given using Ada syntax.
type Matrix is array(-5..5, 1..5) of float;

type name
total size
index type
array bounds
type
dimension
element size
base type

matrix
220 bytes
integer
array -5, 5
type
11 slots
20 bytes

20 bytes
array
type

integer
1, 5
5 slots
4 bytes @
float

Type-Objects for Arrays

We see that the type-object for an array type in a modern language only needs to have information
about one dimension. For that dimension it must store the total size, the index type, the bounds,
and a pointer to the base type. For efficient use in various computations, it might also contain the
dimension and the size of the base type. Exhibit 14.11 shows diagrams of two array type-objects,
representing an array of arrays in a language with non-zero-based subscripts. Exhibit 14.12 shows
the simpler type representation possible with zero-based subscripts. One number, the dimension,
takes the place of three: dimension, lower bound, and upper bound.
Arrays with Undeclared Bounds. In certain situations it makes sense to declare an array but
not specify the array bounds. This is permitted in some languages if either:
1. The omitted information can be deduced from the context, or
2. The omitted information is not needed by the compiler.

Exhibit 14.12. Diagram of a zero-based array type-object.

The type declaration is given using C syntax.
typedef float [5] vector;

type name
total size
index type
array dimension
type
element size
base type

vector
20 bytes
array
type

integer

5 slots
4 bytes @
float

14.2. COMPOUND TYPES

405

Where structured initializers may be given for arrays, and a particular lower bound is either
required or is the default, as in C and Ada, the upper bound of an array can be deduced from an
initializer if one is given as part of the declaration. In this situation, omitting the bounds from
the declaration reduces the redundancy and eliminates the chance that the programmer might
miscalculate the dimension needed. However, it also eliminates the chance that the compiler can
detect a faulty initializer. (This is another example of the general rule that redundancy is a pain
to write into a program, but can serve as a double check on the typists accuracy.)
An example of the second principle is seen in C. An array parameter may be declared with an
unspecified dimension because C does not use the dimension information for parameters. All C
arrays are passed by reference, so only a pointer is allocated for the array in the functions stack
frame. Further, C does not perform bounds checks of any sort. Thus the dimension of an array
parameter is immaterial within the function. Within the C subroutine, only the base type of an
array is ever needed.
Contrast this to Pascal which does pass arrays by value, if desired, and does check array bounds.
Both of these facilities can be used to make programs more reliable and easier to debug; however,
this security comes at a high price. The bounds of a Pascal array parameter must be declared and
fixed at compile time. Therefore each function can accept, as arguments, only one fixed size of
array. In C, any length array of the appropriate base type can be passed to a function. Thus the
Pascal programmer must edit and recompile subroutines if the length of the data array changes,
whereas the C programmer only needs to change the declaration of the data array in one routine.
Semantic Protection with Arrays. We have seen how array type-objects are used to allocate
array storage objects and access individual elements. Some languages, such as C, use the type
information for these purposes only. This same information can also be used to identify run-time
errors, as it is in Pascal and Ada. These languages compare the value of the subscript expression to
the declared array bounds, and halt with an error comment if the bounds are violated. Of course,
this slows down execution. However, this cost is probably justified. Semidebugged C programs
often run wild and erase the contents of memory, forcing the user to reboot the workstation.
Pascal programs seldom do.
Bounds checking is most useful during program development and debugging. Omitting the
checks during this phase would be penny wise and pound foolish. Although the checking code
consumes execution time and memory space, a single error caught pays for many tests. However,
one might assert that finished production programs should not incur the run-time overhead of
bounds checking because fully debugged programs just dont run wild. (Of course, one can debate
whether any program is ever fully debugged.) Most programs perform well with these checks
included. Occasionally, a bounds check in an inner loop can cause a significant slowdown and is
therefore undesirable.
The Ada language includes a way to turn off unwanted checking, called a suppress PRAGMA. If
a compiler honors a suppression request (it is not required to do so), type checking is suppressed
throughout the block that contains the pragma declaration. The design of Ada encourages very

406

CHAPTER 14. THE REPRESENTATION OF TYPES

limited and selected use of suppression pragmas. A separate declaration is needed for each kind of
check that is to be suppressed. The programmer is urged to use these only in fully debugged code
that is unacceptably slow, and then to place the suppression declaration in the smallest block that
contains the slow code section. A PRAGMA declaration to suppress bounds checking has this format:
PRAGMA Suppress (Index_Check);

14.2.2

Strings

A string is a variable-length array of characters. In many languages (for example, Pascal and Ada)
the type string is treated as a special case of an array, with an integer lower bound of 1 and a
base type of character. This treatment does not really capture the essence of strings. Strings are
different from general arrays because:
The length of a string is, in general, unknown at compile time.
Strings of many different lengths are commonly used together.
A string variable should be able to contain any string.
A string function should be able to work on any string.
Strings have special semantics, such as the rule for alphabetic-order comparison of strings
with unlike lengths.
The strings supported by Ada and Pascal are not variable-length objects. A string variable
has a declared length and can only contain string values that are short enough to fit within this
maximum. Strings shorter than that must be manually padded with blanks. Using fixed-length
arrays to represent variable-length strings doesnt really work.
A good representation for strings must embody their essential variable-length nature, use storage
efficiently, provide for efficient processing, and have reasonable error-recovery properties. There are
two representations for string values that meet these criteria: the counted string and the terminated
string.
In a counted string the first byte (in subscript position 0) is an unsigned integer which gives the
length of the string. String functions must calculate the length of a newly constructed string and
store it in the first byte. With this representation, strings can be processed easily using for loops.
The error recovery potential is not ideal, but it is adequate. If a program error happens and the
count byte is changed to a garbage value, any loop that uses the garbage will still terminate after
no more than 255 iterations. No runaway loop in a string function will be infinite and force the
user to reboot a machine.
A terminated string contains only characters, but following the last data character there is a
terminator, probably the null character (ASCII code 0) (strings cannot contain the terminator).
Functions must be written to add a terminator to the end of any newly constructed string. Strings
are normally processed by looping while the data is nonnull. This representation for strings works

407

14.2. COMPOUND TYPES

Exhibit 14.13. A type-object for a null terminated string.

char *str1 = "This is a string literal.";

str1
4 bytes
pointer
type

variable
array
type

integer
1, variable
1 byte @
char

well for error recovery. If a program error happens and a null terminator gets wiped out, it is
highly likely that another zero byte is somewhere nearby in memory and will terminate the runaway
operation soon.
Both of these representations make efficient use of storage and support efficient string processing.
The null terminated string, used in C, seems to have slightly better error recovery properties.
Exhibit 14.13 shows appropriate type-objects for strings.
Once the representation of a string value is decided, we can deal with the question of string
variables. In a compiled language, storage for variables in run-time stack frames is laid out at
compile time. The size of each allocation is fixed then and is not variable. This conflicts with the
variable-length nature of strings. One good solution is to use a dynamically allocated string storage
area, and let a string variable be a pointer into this area. This is how BASIC and SNOBOL IV are
implemented. All the string functions take pointer arguments and return pointer results. Storage
for newly created strings is taken automatically from the system-managed string store.
C implements a version of strings that is half way between the fully functioning string store and
the fixed-length strings of Pascal. Like BASIC, a string in C is a pointer to an array of characters.
Unlike BASIC, the semantic basis of C does not include a string store, and the string functions do
not allocate new storage. C functions such as strcat", or string concatenate, require that one
argument be a pointer to a preallocated storage object long enough to store the result. The effect
of this mixed approach is that Cs extensive and powerful string library is less easy to use than the
string facilities in SNOBOL.
There is a good reason why C does not supply automatically managed dynamic storage: simplicity and execution efficiency were both design goals in C. Managing dynamic storage is not simple,
and it requires some sort of string compaction or garbage collection facility. These facilities are
all slow and costly. Using a compaction algorithm becomes necessary when available storage has
all been used and is now occupied largely by dead strings. (A dead string is one that is no longer
bound to any live name or pointer.)
Rather than include a complex dynamic storage management system, the designers of C chose
to let programmers implement whatever portion of such a system might really be needed. One

408

CHAPTER 14. THE REPRESENTATION OF TYPES

standard technique used in C string programs is to implement a dynamic string store similar to
that which is built into BASIC.

14.2.3

Sets

The set types supported by Pascal are variants of Boolean arrays. A set value is like a packed
array of Booleans, where each index position, or slot, represents one constant of a subrange type
or enumerated type, called the base type of the set.
The number of constants in the base type determines the number of Booleans in the set value;
the first Boolean in the set value corresponds to the first constant in the base type, and so on. A
set with one member is represented by a Boolean array with one element which is TRUE and all the
others FALSE. A set with several members is represented by an array with several TRUE elements
[Exhibit 14.14].
The in operation in Pascal is the selector function for a set. The operands of in are a value
of the base type, v, and a set value, S. The result of the expression v in S is the Boolean value
stored in the position of S that corresponds to v. For example, in Exhibit 14.14, common is a set
variable of type combo. We would write IF green IN common to find out whether green is
a common color.
The standard mathematical set operations union, intersection, and difference are denoted by
the Pascal operators +, *, and -, respectively. These take operands of a specific set type
and return a result of the same set type. These operators are implemented by applying Boolean
operations to corresponding elements of the two operand arrays. Union is implemented by or, and
intersection by and. Set difference (-) is implemented by the composition of complement and
and.
A Boolean value can be represented by a single bit. In an efficient representation of a Boolean
array, we pack the bits into a series of bytes, so that every bit is used. By packing the Boolean
array, we are able to use the efficient bitwise logical operations to implement set operations. These
operations let you turn on and turn off bits singly, or several at a time. Turning a bit on
corresponds to adding a member to a set. The diagram on the left in Exhibit 14.15 shows an
appropriate type-object for a Pascal set type.
Although sets are included in Pascal, most languages do not have a corresponding set type
constructor. This is not surprising, since both the semantics and the representation for sets are
somewhat complex. One of the design goals for Pascal was simplicity, and it is interesting to ask
why Wirth included this nonsimple type of type in his design. The reason was a combination of
three factors: completeness, validity, and simplicity.
Completeness. A powerful language should contain constructs that support standard mathematical notation. It should also reflect as many as possible of the capabilities of the typical modern
computer. Including a set type in Pascal, with the representation described, completes the language in two ways.

409

14.2. COMPOUND TYPES

Exhibit 14.14. A set type in Pascal.

We declare an enumerated type, color, and a set type, combo, whose base type is color. Some
variables of type combo are declared and initialized. The Boolean array representations of these
variables are diagrammed below.
TYPE

VAR

color = (red, pink, orange, yellow, green, blue, violet,

magenta, brown, black, gray, white);
combo = set of color;
tree, daffodil, iris: combo;
palette, common: combo;

daffodil := [white, yellow, orange, green];

iris := [white, yellow, blue, violet, green];
tree := [white, gray, brown, black, green];
palette := iris + daffodil;
{ + means set union. }
common := iris * tree * daffodil;
{ * means set intersection.

In the following representations, a 0 represents a FALSE value, and a 1 represents a TRUE value.
Cardinality of base type (color): 12
Number of elements in a combo value: 12
Size of each element: 1 bit
Representation of daffodil: 001110000001
Representation of iris: 000111100001
Representation of palette: 001111100001
Representation of common: 000010000001

The bitwise operations on a computer are vital for applications such as number conversion, using
hardware switches, and packing and unpacking data values. With sets, Pascal provides access to
the bitwise hardware instructions; that is, it provides a type constructor and some operators that
translate into the bitwise operations.
Sets are a standard mathematical notion that can be used to model quite a variety applications.
By including an efficient implementation of sets and set operations, Pascal supports an important
mathematical notation.
Why, then, does C not support sets? The design goals for C were significantly different from
those for Pascal. The fact that mathematicians use sets was of little importance in designing a
language for systems programming. The need for access to the bitwise operations was compelling,
but it was met in a different, and simpler, way. C includes the bitwise operators & (and), |

410

CHAPTER 14. THE REPRESENTATION OF TYPES

Exhibit 14.15. Type-objects for set types.

The diagram on the left is a type-object for a Pascal set type. On the right is a type-object
appropriate for an Ada-like implementation of sets as Boolean arrays. We use the type-object
defined, above, for zero-based arrays. The symbols used in these type-objects are defined thus:
E, an enumerated type, is the base type of this set type.
N is the number of enumeration constants in E.
L is the smallest integer greater than or equal to (N/8).

Pascal:
set type name
L bytes
E
set
type

Ada:
type name
L bytes
E
N elements
packed
1 bit @
Boolean
array
type

(or), ^ (exor), and ~ (complement), which can be applied to any type of value.
If a C programmer wanted to implement set semantics he or she would define a set value as an
array of unsigned integers, where each integer represents 16 or 32 index positions in the set. To
implement the set operations, the programmer would simply iterate the primitive bitwise operators
as many times as necessary to process each byte of the array. This is a simple and straightforward
implementation. There is no particular need to make sets a primitive type constructor.
Validity. Support for semantic validity was one of Wirths most important design goals. All
primitive operations that he included were semantically valid for some primitive type. Strong type
checking ensured that validity was maintained by every operation.
Wirth would have considered the semantics for the bitwise operators in C to be unacceptable.
An unsigned integer type should be used to store unsigned integers, not bit strings. The bitwise
logical operators have no semantic validity when applied to integers. Basically, they can only be
meaningfully applied to packed arrays of Booleans, and that is what Pascal supports.
But semantic validity and completeness were also of primary importance in the design of Ada,
and Ada does not have a set type constructor. It does provide an easy way to implement the set
semantics, though. In Ada, the type one-dimensional array of Booleans is given special semantics.
The logical operators and, or, xor, and not may be applied to one-dimensional Boolean arrays.
The relational operators =, <=, and the like may be applied to any one-dimensional array. Thus
sets can be easily implemented as Boolean arrays. The diagram on the right in Exhibit 14.15
shows an appropriate type-object for this set implementation. Exhibit 14.16 shows a set of Ada

14.2. COMPOUND TYPES

411

Exhibit 14.16. Implementation of sets in Ada.

We declare an enumerated type, color, and a set type, combo, whose base type is color. Some
variables of type combo are declared and initialized. The Boolean-array representations of these
variables are diagrammed below.
type
type

color is (red, pink, orange, yellow, green, blue, violet,

magenta, brown, black, gray, white);
combo is array(red..white) of boolean;

T: constant boolean := TRUE;

F: constant boolean := FALSE;
tree: constant combo := (F,F,F,F,T,F,F,F,T,T,T,T);
daffodil: constant combo := (F,F,T,T,T,F,F,F,F,F,F,T);
iris: constant combo := (F,F,F,T,T,T,T,F,F,F,F,T);
palette, common: combo;
palette := iris or daffodil;
common := iris and tree and daffodil;

type declarations to implement types analogous to the Pascal color example from Exhibit 14.14.
Exhibit 14.17 shows how corresponding set selection operations would be written in Pascal and
Ada.
Simplicity. Wirth intended Pascal to be a powerful but simple and minimal language. Highly
complex semantic mechanisms did not fit this purpose. He also wished to have a clean, general
design with few special cases and few restricted, special-purpose capabilities. 6
A language with suitably powerful array operations would not need a special type of type to
6

Note that the special semantics for strings in Pascal is an example of what Wirth wished to avoid. String types
break many of the type rules that govern the rest of the language.

Exhibit 14.17. Selecting a member of a set.

In Pascal we select a member of a set using in:
if green in common then ...
In Ada we use a subscript to select a set member:
if common(green) then ...

412

CHAPTER 14. THE REPRESENTATION OF TYPES

implement the semantics of Pascal sets, as shown by the array-of-Booleans implementation in Ada.
APL also supports coherent array operations which make the Boolean-array implementation of sets
easy and straightforward.
But the powerful, general array operations in APL do not meet the criteria of simplicity and
minimality. The solution in Ada (including primitive bitwise operators that are defined only for
one-dimensional Boolean arrays, and coherent array operations that apply only to one-dimensional
scalar arrays) is too nongeneral and special-purpose to meet Pascals design goals. While Pascal
sets are not a simple type of type, they are less complex than APL array operations, more elegant
than the Ada solution, and more valid than the C approach.

14.2.4

Records

A record is a compound object consisting of an ordered series of components of heterogeneous types.

These objects are usually implemented by contiguous blocks of memory, in which the fields of a
record object are stored in the order specified by the programmer. 7 A modern language permits
the programmer to manipulate a record object coherently or to access its components and process
them individually.
Type-Objects for Records
The type-object for a record will contain all the information needed to allocate and access the
record. For allocation, only the total size of a record storage object is needed. Intuitively, a record
object is no more and no less than the sum of its parts. However, in real implementations, the
storage object for a record may contain more bytes than are needed to store the components, and
the extra bytes will generally contain garbage. This happens because many computers require all
numbers to be aligned on word or long-word boundaries. With word alignment, all objects start at
even byte addresses; with long-word alignment, all addresses are evenly divisible by 4. Padding is
inserted into a record type if the size of some field in a record declaration would cause the following
field to break alignment rules. The amount of padding included (if any) depends on the hardware.
Thus all we can say about the size of a record is that it is at least as great as the sum of the sizes
of its parts.
To access a record component, a compiler needs to know its offset, that is, the number of bytes
between the beginning of the record object and the beginning of the desired field. Each offset is
the sum of the sizes of all preceding fields, plus any preceding padding bytes. When the compiler
processes the list of component types in a record declaration it calculates these offsets and binds
each one to the corresponding record part name. We will represent all this information (offsets,
part names, component types) as part of the type-object for a record [Exhibit 14.18].
7

This order is by no means necessary unless a language permits the programmer to circumvent the usual accessing
methods and access the records representation directly.

413

14.2. COMPOUND TYPES

Exhibit 14.18. A type-object for a record.

struct personcell{char* name; char sex; float salary; personcell* next};
We would diagram the type personcell as follows:

type name
total size
number of fields
record field offsets
type
field names
component types

personcell
14 bytes
4 fields
record 0, 4, 6, 10
type
name, sex, salary, next
char *
float
char
personcell *

Exhibit 14.19. Defining a record type in FORTH.

We define functions to allocate and access a record type named personcell. Objects of this type
will be allocated dynamically and attached to linked lists.
: new_personcell
here 10 allot;
( Allocate 10 bytes; leave address on stack.
)
( Selector functions for the record type personcell. All expect a
)
( single argument which is a pointer to a record. All leave on
)
( the stack a reference to one component.
)
: .name ;
( Zero offset for first field.
)
: .sex
2 + ;
( Assume string pointers take 2 bytes.
)
: .salary 4 +;
( One byte for sex, one for padding.
)
: .next
8 +;
( Leave four bytes for double-length integer.
)
Assume that the variable employee is a pointer to a personcell record. If we wish to store
a number in the salary field of this employees record, we write the following expression. (The
operator D! performs a double-length assignment.)
89700 employee @ .salary D!

414

CHAPTER 14. THE REPRESENTATION OF TYPES

Part Selectors
The part names for a record type are selector functions. Like functions, each one takes an argument
(a reference to a record), performs an action (increments that reference by an offset amount), and
returns a reference to another object (one component). The nature of part names can be seen most
clearly in FORTH. FORTH has no special provision for records, but it permits the programmer
to allocate storage manually, manipulate references to objects, and do address arithmetic. When
a FORTH programmer wants to use records he or she defines an allocation function and a set
of selector functions [Exhibit 14.19]. These selectors are used very much like Pascal part names;
compare the FORTH expression from the last line of the exhibit to the equivalent expressions in
Pascal and C:
FORTH:
Pascal:
C:

89700 employee @ .salary D!

employee.salary := 89700
employee->salary = 89700

So we see that part names can be defined as functions in a language that gives access to lowlevel information. Why, then, do most languages provide special declarations? There are several
reasons: convenience, economy, portability, and semantic safety.
Convenience. The part name function definitions in FORTH are simple and brief, but even
so, they are somewhat of a nuisance to write out. The record type declaration is clearly a more
convenient way to convey this information. A type declaration conveys all the relevant information
concisely.
Economy. Implementing selectors as ordinary functions is overkill. Associating a selector name
with an offset takes very little storage, and putting these pairs in the type-object gives an extremely
brief way to represent the information. Moreover, we do not need the full generality of functions
for this purpose. Selectors are constant functions that are applied and expanded by the compiler.
Compiled code contains address arithmetic, not function calls.
Most languages permit the same part name to be used in multiple record definitions. Storing
the part name/offset mappings in the type-object is an easy way to implement the required kind
of ambiguity; it allows the same part name to be bound to different offsets in different types. It
is interesting to note that very old C compilers did not work like this. They probably stored the
part names in the symbol table with the function names and other identifiers. The result was that
each part name could only be used to mean one offset. A part name could be used in more than
one struct declaration, but each use had to correspond to the same offset amount! This has been
modernized by the ANSI C standard.
Portability. To define the FORTH selectors we had to know exactly how many bytes would be
required to implement each primitive type. But this varies from machine to machine. Whatever
constants we use for the offsets, they will be wrong for some FORTH implementations.

14.2. COMPOUND TYPES

415

C provides the same access to low-level information as does FORTH, so one could define record
selectors the same way in C. However, this would be foolish, because then the user would have to
worry about inserting padding bytes and accommodating varying sizes for the primitive types. If
the user writes a struct declaration, the compiler takes care of all this.
Semantic Safety. The part selectors defined in FORTH are accompanied by absolutely no checking. If an offset was one too large, the boundary between successive parts would be violated. A
program that tried to access that component would get the last byte(s) of it and the first byte
of the next component! These bytes do not form a valid object of any sort. Moreover, a FORTH
function is not restricted to use with the correct type of argument; any selector could be used on
an entirely inappropriate type of record, producing a garbage answer.
Pathnames
Let us define the term pathname to mean the sequence of identifiers, starting with the name of an
object and continuing with a series of selectors (part names or subscripts). A pathname designates
a particular field, f n, of a particular object, Ob.
When a compiler translates a pathname, it uses the series of selectors to compute the address
of the specified component, or the effective address. Initially, the effective address is set to the
base address, that is, the address of the first location in Ob. The current type-object is set to the
type-object of Ob, and the current selector is set to the first selector in the pathname.
The compiler looks for the current selector in the current type-object. This field has an associated type t1, and an offset amount, n1. For records, n1 is listed explicitly in the type-object.
For arrays, n1 is found by multiplying the component size (from the type-object) by the array
index minus the arrays lower bound. The compiler then adds n1 to the effective address, sets the
current type-object to t1, and goes on to the next selector in the pathname. This process is iterated
until all selectors are used, and a final offset amount is calculated. A pathname that contains no
variable subscripts can be processed entirely at compile time. Otherwise, the constant portions of
the computation are done at compile time, and the rest must be deferred until run time.
Partial Pathnames. When part names are simply entered into the symbol table, as they are in
COBOL, PL/1, and old versions of C, the programmer must be careful about using the same part
name in two different record types. We eliminate this problem by storing the record part names in
the type-objects. This makes them into local names, that is, names whose meaning within the type
is quite independent of any other meanings in other contexts. Localizing names makes it easier to
write correct code. However, one feature present in older languages was lost by this change.
In PL/1, the programmer could refer to a sub-subfield without specifying a full pathname;
the programmer wrote only the name of the object itself and the name of the sub-subfield. This
kind of short-cut naming can shorten and simplify code, especially where a record type contains a
structured component, and so on, for several levels.

416

CHAPTER 14. THE REPRESENTATION OF TYPES

Pascal has one statement type that partially compensates for this loss of convenience. A with
statement allows the programmer to establish a local context, within which the initial portion of a
pathname can be omitted. The form of a with statement is as follows:
with hpartial pathnamei do hscopei
When the compiler begins to translate a with statement, it evaluates the effective address,
Ea_pp, for the partial pathname. (If there are variable fields in the pathname, code is generated
to complete this process at run time.) During compilation of the with scope, Ea_pp is the starting
point for computing effective addresses of components, and the current type-object is set to the
type of the last field in the partial pathname. Within the scope, all references to field names defined
for this type are legal and will be interpreted as offsets added to Ea_pp.
At run time, any variable fields in Ea_pp are evaluated once, when control enters the with
block. Within the block, only the tail section of each pathname must be specified and evaluated,
saving both execution time and space [Exhibit 14.20].
The principle here is valid and important: when doing several operations with one part of a large
compound object, it is more efficient (in several ways) to mark the beginning of that component
and make local references relative to that mark. Looking at other programming languages, we see
more general ways to solve the same problem.
A C programmer does not need a special statement type to accomplish this; she or he simply
sets a pointer to the desired component (at entry to the block) and makes references within the
block relative to that pointer [Exhibit 14.21]. The with statement is included in Pascal because
Cs simple solution is not available. (Recall that, in Pascal, pointers to stack-allocated variables are
prohibited.) Moreover, the Pascal solution is semantically cleaner because, in C, the pointer is not
constrained to be constant within the local scope.
LISP incorporates what might be called the right way to solve the problem. Using let, the
programmer can create a new block with a local symbol and bind that symbol to any object. (The
programmer would bind the new symbol to a pointer to the Ea_pp.) C-style references relative to
this pointer could then be used within the local scope.
Records Are Much Like Arrays
There are many similarities between records and arrays; both are compound types with a series of
parts, accessed by selector functions. Their type-objects contain similar information, except that
an array type-object is simpler because all fields of an array are the same size and type. Where
an array type-object contains one piece of information about a component, a record type-object
contains a list.
However, the familiar syntax for subscript (with parentheses or brackets) is markedly different
from the syntax for record part selection (with a dot). This is partly an accident of history; neither
subscript notation nor dot notation is engraved in stone. A language designer could choose to use
either syntax, as shown in Exhibit 14.22.
There is a more important difference between arrays and records than the syntax used for part
selection. That is the fact that the traditional array selectors are numeric and record selectors are

14.2. COMPOUND TYPES

417

Exhibit 14.20. Partial pathnames using with in Pascal.

With is used here to simplify the source code and reduce execution time for access to one part of a
complex data structure, a stack of student records. When the with block is entered, the effective
address of the top student on the stack is calculated. All references to the field names defined for
StudentType will be interpreted relative to this address.
Type NameType =
packed array[1..20] of char;
StudentType = record LastName, FirstName: NameType;
Sex: char;
Id: integer
end;
ClassType =
record { A class is a stack of students. }
Top: integer;
Member: array[1..50] of StudentType;
end;
Var

N: Integer;
CS101: ClassType;
Stu: StudentType;

...
With CS101.Member[ CS101.Top ] Do Begin
Writeln( Id, FirstName, LastName );
If Sex=F Then FemaleTotal := FemaleTotal + 1
Else MaleTotal := MaleTotal + 1
End;

not. Let us consider the twin questions:

1. Why dont we use symbolic names to select array components?
2. Why dont we use integers to select record components?
The essence of an array is that its components are semantically uniform. They represent objects
from the same domain, and we expect to apply the same operations to each, in turn. The number
of array elements and position of a particular element in the array is of secondary importance, at
most. Because these elements have uniform meaning to the programmer, a uniform way to name
them is needed Giving them numbers is a good solution. Giving individual names to a series of
similar objects would be silly.
In contrast, the components of a record are not semantically uniform. They represent different
aspects of the same object, not separate objects. Even when components are the same type, they

418

CHAPTER 14. THE REPRESENTATION OF TYPES

Exhibit 14.21. Using a pointer in C to emulate with.

typedef struct

char last_name[20], first_name[20];

char sex;
int id;
} student_type;
typedef struct {
int top;
student_type member[50];
} class_type; /* A class is a stack of students. */
...
{
int female_total, male_total;
class_type cs101;
student_type *stu;
...
stu = &cs101.member[ cs101.top ];
printf("%d %s %s\n", stu->id, stu->first_name, stu->last_name );
if(stu->sex==F) ++female_total; else ++male_total;
...
}

Exhibit 14.22. Alternative syntax for part selection.

Let Student be a StudentType, as declared in Exhibit 14.20, and let Class be an array of
StudentType. There could be many clear and unambiguous ways to denote part selection. A few
possibilities are listed here.
Traditional syntax
Possible syntax
Uniform syntax
Functional syntax

Array Selection
Class[5]
Class.5
Class@5
Subscript(Class,5)

Record Selection
Student.FirstName
Student[FirstName]
Student@FirstName
Select (Student, FirstName)

14.2. COMPOUND TYPES

419

represent different concepts and can easily be given different names. Defining symbolic names for
record components is, therefore, natural and functional.
However, constant numeric selectors could be used for records. Allowing numeric (as well as
symbolic) selection could have some advantages, especially for writing library or utility routines.
Variable numeric selectors cannot be used in a strongly typed compiled language because the type
of the result of every expression must be determined at compile time. Suppose R is a record variable
with five components. Then a compiler can determine the type of R[3] by looking at the third entry
in the list of component types in the type-object for R. But a compiler cannot know the type of
R[i], where i is a variable; it could be any one of the component types of the record. Thus a variable
cant be used to select a record part in a statically type-checked language.
Some languages are interpreted, not compiled. These languages eliminate the restriction that
the type of the result of an expression must be known ahead of time. If such a language also
permitted the programmer to use the information in a records type-object, some very useful and
powerful routines could be written. For example, we could write a general debugging package with
a routine that could print out and label the components of any record. Such a routine would use
the type-object to find out how many components were in the record, then execute a loop that
used the field name and field type information in the type-object to print out the records value.
This kind of code is polymorphic, that is, the type of a function argument is tested at run time
in order to know how to execute the function. 8

14.2.5

Union Types

Two kinds of types are called unions because the semantics of a given storage object can vary. There
are actually two kinds of unions, with very different semantic properties, known as free unions and
discriminated unions.
A free union is a semantically unsound type. An object of a free union type has two or more
possible semantic interpretations, and there is no field, either in the object itself or in its associated
type-object, that defines which set of semantics is currently valid. A free union type declaration
is not a way to build a compound type out of simpler types. Rather, it is a way to create objects
with ambiguous semantics. A free union type-object is shown in Exhibit 14.23. Its form is like
the type-object for a record. Its semantics differ from a record in that only enough space for the
longest field is allocated, and all fields have an offset of 0 bytes (the offsets, therefore, do not need
to be part of the type-object.) Many languages, including Pascal and C, support free union types.
These will be dealt with fully in Chapter 15, Section 15.7, after the discussion of type checking.
A discriminated union is, in theory 9 , a semantically sound type, because the alternative semantic interpretations are controlled by a field that contains a case specifier. The general form
of a discriminated union is: Common partkeyset of variant parts. The Pascal variant record
declaration is an example of a discriminated union type constructor. It defines a record type with
three sections: an initial section with fields that are common to all variants, a tag field that is a
8
9

We will cover polymorphic code in Chapter 17.

We say in theory because the implementation in Pascal is faulty and semantically unsafe.

420

CHAPTER 14. THE REPRESENTATION OF TYPES

Exhibit 14.23. A type-object for a free union type.

typedef union{double dd; int kk; char *str} strange_type;

type name
strange_type
nondiscriminated
union type

total size
number of variants
field names
types of variants

8 bytes
3 variants
dd, kk, str

nondiscriminated
union type

double

int

char*

Exhibit 14.24. A type-object for a discriminated union type.

A discriminated union type has two parts. The first part is essentially an ordinary record ending
with the discriminant field; the second part is a free union. This is an appropriate type-object for
the type declared in Exhibit 14.25.

type name
total size
number of nonvariant fields
discriminated field offsets
union type
nonvariant field names
component types

Person
10 bytes
2 nonvariant fields
discriminated 0, 6, 8
born, sex
union type
date

nondiscriminated
union type

total size
number of variants
field names
types of variants

nondiscriminated
union type

gender
2 bytes
3 variants
Bearded, Children, null
Boolean

Integer

null

14.2. COMPOUND TYPES

421

Exhibit 14.25. The discriminated variant record in Ada.

type Gender is (Male, Female, Unknown);
type Person(Sex:Gender:=Unknown) is
record
Born: Date;
case Sex is
when Male => Bearded: Boolean;
when Female => Children: Integer;
when Unknown => null;
end case;
end record;
Sam:

Person;

-- A variable of type Person.

code for the variant currently stored in the variable, and a set of variant parts, each having any
number of unrelated parts [Exhibit 14.24]. The clear intent of variant types is to conserve storage
by permitting mutually exclusive information fields to occupy the same positions in the record.
This reduces the total amount of storage needed for variant objects to the amount needed for the
longest variant. We can say that the variant fields share storage.
If storage is plentiful, there is no need for this type of type. The same semantics can be achieved
by using an ordinary record which contains all possible fields and a tag field to say which subset
of the fields is currently meaningful. Storage for all fields would then exist all the time, and some
subset would contain meaningless garbage.
There are some curious, perhaps faulty, aspects of Pascals variant records. The tag field controls
which set of field names is defined at any given time and should always correspond to the information
actually stored in the record. To assign a value to a variant record, the variant tag must be stored
first. This causes the corresponding field names to become defined. However, nothing forces a
programmer to finish the job. A value of one variant type, with its tag, could be stored in the
variable. Then the tag could be changed. At this point, the tag label does not match the contents
of the variable, and if the variable is used, the bits will be given an invalid interpretation! The
rules of Pascal allow this to happen, and thus the variant record is a gaping hole in Pascals type
system.
Ada also supports discriminated variant records that are similar to Pascals, but with an important difference that corrects the loophole in Pascals semantic rules. Assignment is restricted
so that a variant storage object can never contain a case specifier for one variant and a value of
another.
In Exhibit 14.25, we declare a discriminated union type with one common field (Born), a
discriminant named Sex whose default value is Unknown, and three sets of variant fields, labeled

422

CHAPTER 14. THE REPRESENTATION OF TYPES

Exhibit 14.26. Legal and illegal uses of a discriminated variant record.

Sam := (Male, (1970,Jan,3), FALSE);
Sam.Sex := Female;
Sam.Children := 2;

-- This is ok.
-- Prohibited.
-- Undefined, Sam is Male.

by the elements of type Gender. We may construct a value of type Person by supplying a list of
relevant data items, thus:
(Male, (1970,Jan,3), FALSE)
We may assign such a value to a Person variable. But we may not assign values to the tag
field or the variant field independently, These rules forcibly maintain the consistency between the
tag field and the information stored. Let Sam be a variable of type Person. Then the whole-record
assignment on the first line of Exhibit 14.26 is permitted, but the partial-record assignment on
the second line is not. The third assignment is not permitted because the field name Children is
not defined for objects with the tag Male. These constraints in Ada close the loopholes present in
Pascal.

14.3

Operations on Compound Objects

14.3.1

Creating Program Objects: Value Constructors

A program object of a primitive type is created initially by executing a function or by writing a literal
in a program. For example, the result of performing an addition operation is a program object.
Arithmetic expressions and most functions create program objects as their results. (Sometimes
these program objects are references). These values are kept on a run-time stack.
Programming language conventions have developed to allow the writing of literal values of all
the primitive types. Most languages have distinguishable syntactic forms for types real, integer,
and character, and sometimes for Boolean. Additional primitive types, such as short integer,
packed decimal, and byte, are sometimes supported. Sometimes the same program object may be
denoted by more than one literal expression. For example, C lets the programmer write an octal,
hexadecimal, or decimal literal to denote an integer value.
A language that permits a programmer to define new types should provide some way to write
literals for these types. Pascal does not do this; the programmer cannot, therefore, define a constant
of a user-defined type. C permits such a literal value to be written as an initializer, in a declaration,
but not in any other context. This is a wholly unnecessary restriction.
A language could include a constructor function which we will call MAKE to solve this problem.
The first argument of MAKE is a type. Following this is a variable number of literals or expressions
appropriate in number and composition for program objects of the specified type [Exhibit 14.27].
MAKE takes the separate pure values in this list, bundles them into an object of the given type, and

14.3. OPERATIONS ON COMPOUND OBJECTS

423

Exhibit 14.27. Making a record object from its components.

Assume that the type complex-pair has been defined as a record containing real and imaginary
components, and that imaginary numbers are made out of reals. Then a complex-pair would be
constructed by executing the following:
MAKE( complex-pair, 0.0, MAKE( imag, 5.6) );

returns this value as its result. It is important to note that this result should be a coherent object,
temporarily residing on the run-time stack. It can, therefore, be dealt with or manipulated as a
whole even if it is a record or an array.
Another possible approach with the same semantics but slightly different syntax was used
in Ada. Each defined type name becomes a constructor function automatically. To construct a
program object of the new type, the new type name is written preceding the appropriate series of
components, listed in parentheses [Exhibit 14.28].

14.3.2

The Interaction of Dereferencing, Constructors, and Selectors

Dereferencing maps a reference into a program object, or value. This reference/value relationship is complicated by the introduction of compound data-objects (arrays and records), selection
functions, and constructors. A selection function takes, as parameters, a compound object and a
part specification. It returns a reference to the specified part of the compound [Exhibit 14.29]. A
constructor combines a set of components into a single compound object.
The selection functions in many familiar languages (e.g., FORTRAN and Pascal) take references
to compound objects as parameters and return references to simple objects as results. The simple
reference is then dereferenced if the context requires. Simple selection functions are combined to
build the essential accessing functions for abstract data types such as stack in Exhibit 14.30.
A constructor creates a compound program object from a set of pure values on the run-time

Exhibit 14.28. Making a record object from its components in Ada.

type imag is new float;
type complex_pair is
record rp: float;
ip: imag
end record;

-- Floats will be used to represent type imag.

-- A complex pair is a float and an imag.

complex_pair(0.0, imag(5.6)) -- Make an float into an imag, then

-- use it with another float to make
-- a complex pair.

424

CHAPTER 14. THE REPRESENTATION OF TYPES

Exhibit 14.29. Pascal selection functions.

In Pascal, the part names associated with record types are selection functions. We define these
functions when we make a type declaration. The following declaration creates two new selection
functions, named top (line a) and store (line b).
Subscript is a selection function that is predefined for all arrays, but depends on the declared
array bounds. Line b defines the legal range of subscripts for the store component of a stack,
and fully determines the meaning of subscript on this kind of object.
TYPE stack = RECORD
top: integer;
store: array [1..100] of StackItem
END;

{a}
{b}

Exhibit 14.30. Using selection functions in Pascal.

We combine the selection functions subscript, top, and store to build a pop function for a
stack.
a. The parameter to the pop function is the address of the beginning of a stack storage object.
(The keyword VAR indicates that a Pascal parameter is a reference.)
b. All three selection functions are composed, to arrive at the address of the element at the top of
the stack. This address is dereferenced because it appears on the right side of an assignment.
This yields a program object which is later returned by the function.
c. A different stack component, the top-of-stack index, is then selected twice. The right-hand
occurrence is dereferenced and decremented. The resulting value is stored in the address
given on the left (the same location), modifying the value of the compound object.
FUNCTION pop(VAR S:stack):StackItem;
BEGIN
pop := S.store[ S.top ];
S.top := S.top - 1
END;

{a}
{b}
{c}

425

14.3. OPERATIONS ON COMPOUND OBJECTS

Exhibit 14.31. Value constructors in APL.

Using a literal or a variable name causes the corresponding value to be placed on the stack.
Assume N is bound to a numeric value; it will be automatically dereferenced and placed on
the stack between the other two numbers.
We use the concatenate operator, comma, to construct a one-dimensional numeric array, or
vector, from the three simple values on the stack.
2, N, 5.1

stack. It can leave the resulting compound object on the stack or allocate a variable of the correct
shape and initialize it to the given set of pure values. We call the former a value constructor and
the latter a reference constructor.
Several languages give partial support to value constructors; for example, C permits them in
initializing expressions and Ada supports a similar notation for compound literal expressions. APL,
however, is one of the few languages that supports explicit, run-time value construction with no
restrictions. Use of APLs value constructor, comma, is shown in Exhibit 14.31. The reshape
function, , is a combined type cast and constructor; its use is illustrated in Exhibit 14.35.
Reference construction implies dynamic allocation and the use of pointers. It is not supported
at all in many common languages, but it has a basic and central importance in LISP. The LISP
cons function is a reference constructor which allocates a new cell and initializes it to the two
arguments on the stack.
The possible relationships among references, program objects, compound objects, and their
components are captured in Exhibit 14.32. Each arrow represents a type of function, pointing from

Exhibit 14.32. Selection and dereferencing.

Compound
Reference
Reference
Selector
Simple
Reference

Compound Dereference
Allocation,
Construction
+ Initialization
Selection +
Dereference

Compound
Program Object

Value
Constructor
Reference
Constructor

Simple Dereference

Value
Selector

Simple
Program Object

426

CHAPTER 14. THE REPRESENTATION OF TYPES

Exhibit 14.33. Pascal selection and dereferencing.

Compound
Reference
Part of array or (a)
record as a
VAR parameter
or on left side
of an assignment.
Simple
Reference

(d) Call-by-value with array or record.

Assignment with array or record.
Comparison for = with record.

(c)

Compound
Program Object

Part of array or record in any

other context.

(b) Call-by-value with simple variable.

Simple
Program Object

Simple variable in an expression.

the type of its argument to the type of its result. Thus the arrow for value selector starts at
compound program object and ends at simple program object, and the arrow for reference
constructor starts at simple program object and goes to compound reference. No arrow in
the graph leads from a collection of simple references to a compound reference. This is because
programs deal with real storage. Whereas compound storage objects must occupy contiguous
blocks of memory, a set of individual storage locations normally are not allocated contiguously
and, therefore, cannot be combined directly into a coherent compound object.
Selection and Compound Objects in Pascal. Most languages do not support all of these kinds
of selection and dereferencing functions. Pascal incorporates a limited subset of these possible
function types as diagrammed in Exhibit 14.33. Moreover, Pascal only supports some of these
function types in a highly restricted fashion. (The letters in the diagram key it to the following
explanation of these restrictions.)
a: Reference selectors. Like most common languages, Pascal provides selection functions that
return references when used on the left side of an assignment statement. This permits a single
part of a compound storage object to be changed without manipulating the entire compound.
Subscripting an array (or selection of one field of a record) returns a reference to the selected
part.
b: Simple dereference. Dereferencing is automatic when a reference appears to the right of an
:=, or in an expression, or as an actual argument corresponding to a value parameter.
c: Selection with dereference. The result of a subscript or selection operation is a simple reference. Like any reference, this is automatically dereferenced unless it is on the left side
of an assignment operator or is used as an argument to a function with a VAR parameter.
Thus a reference to part of an object cannot be obtained and stored or manipulated by the
programmer.

14.3. OPERATIONS ON COMPOUND OBJECTS

427

d: Compound dereference. Any Pascal variable, even a compound variable, may be passed as a
value parameter. This causes the variable to be dereferenced. Otherwise, Pascals support for
compound dereference is severely restricted. Record variables will be dereferenced when they
are compared for equality, or when the value of one record variable is copied into another.
Array variables can be copied but not compared.
Pascal does not support the other three kinds of functions at all. It provides dynamic storage
allocation and pointers, but not reference constructors. A series of separate operations is required
to allocate a new compound storage object and initialize its fields. Nothing like the LISP cons is
supported.
Value constructors are not supported at all in Pascal, not even those which construct compound
literal values as in Ada and C. A compound value can only be constructed by storing its components,
one at a time, into a compound storage object. Compound values may not even be returned from
functions in the standard language.
With no value constructors, value selectors are not needed. All compound values are created,
piece by piece, by storing components in a compound variable. Reference selection may then be
used to decompose the value, if needed.
Selection and Compound Objects in APL. The power of APL comes from its generalized
ability to manipulate compound program objects. Here we will look at the selectors and constructors
in APL and note how they differ from those in most languages.
Many APL operations, including the selection functions, can result in compound pure values
being left on the stack for further processing by other operators. All the selection operators can
be used as value selectors, to select some portion of a compound value that was just computed. A
complex expression might perform several computation steps, each time creating, on the stack, an
answer of a new size or new number of dimensions. This is in stark contrast to Pascal, where every
compound value must be created by a series of assignments, and the size and shape of every result
must be known at compile time.
APL Selectors. In FORTRAN, C, and Pascal, selection functions operate only on variables.
The question of selecting a part of a compound program object never occurs because these languages
do not support either literals of compound types or functions that return objects of compound types.
In contrast, APL supports compound program objects and selection functions on both variables
and program objects, and APL incorporates a much more extensive and less restricted subset of
the possible function types from Exhibit 14.32.
APL compounds are limited to arrays (records do not exist in APL), but the programmer can do
many more things with arrays than in Pascal [Exhibit 14.34]. APL gives full support to compound
dereferencing (arrow d), value selectors (arrow e), and value constructors (arrow f).
Simple dereferencing (arrow b) works similarly in Pascal and APL; anything on the right side
of an assignment arrow, and any variable name in any other kind of expression, is automatically
dereferenced. Both languages prohibit storing and calculating with a reference to part of an object.

428

CHAPTER 14. THE REPRESENTATION OF TYPES

Exhibit 14.34. APL selection and dereferencing.

Use of array name in expression

Compound
Reference

(d) or on right of assignment.

(f)

(a)

Use of subscript on left side

of an assignment.
Simple
Reference

Compound
Program Object

(c)

catenate
Subscript, selection, etc. in expression iota
or on right side of assignment
rho

(e)

subscript,
selection,
compression,
etc.

Simple
Program Object

(b) Use of simple variable name in an

expression or on right of assignment.

Compound dereferencing (arrow d), though, is much less restricted in APL than in Pascal. Compound variables, like simple variables, can be used in most expressions. Using a compound variable
name as an operand causes compound dereferencing and leaves a compound object on the stack.
APL is like Pascal where reference selection (arrow a) is concerned. Selection results in a
reference in APL only when subscripts are used to the left of an assignment arrow; selection is
followed by dereferencing (arrow c) for all other uses of subscripts. Unlike Pascal, though, APL lets
the programmer write a subscript expression that denotes any subset of the rows and columns of a
matrix. All of these locations will be selected and receive assignments or be dereferenced [Exhibit
14.36].
Unlike Pascal, subscript is not the only selection function that operates on arrays; there are
several selection operators that take an array or matrix operand and an operand that specifies
which part of the array is to be selected, and compute a new array of a different shape. These are
listed in Exhibit 14.37.
Exhibit 14.35. Constructing matrices in APL.
These lines create two literal matrices and bind them to the identifiers M1 and M2. Diagrams of
these matrices are shown below.
M1 3 5 11 12 13 14 15 21 22 23 24 25 31 32 33 34 35
M2 4 4 0
M1:

11
21
31

12
22
32

13
23
33

14
24
34

15
25
35

M2:

0
0
0
0

0
0
0
0

0
0
0
0

0
0
0
0

429

14.3. OPERATIONS ON COMPOUND OBJECTS

Exhibit 14.36. Selecting a submatrix in APL.

Reference selection and two operators that perform value selection are illustrated below. These
operations use the matrix values M1 and M2 created in Exhibit 14.35. The first line selects two
characters, "no", from the string "random" and binds the resulting value to M3. The second line
selects six values from M1 (rows 1 and 3, columns 2, 4, and 5) and assigns them to six selected
locations in M2 (rows 1 and 4, columns 1, 2, and 3). The result is diagrammed.
M3 0 0 1 0 1 0 / "random"
M2[1 4; 1 2 3] M1[ 1 3; 2 4 5]
Result of value selection:
Result of reference selection
and assignment of a compound
value to parts of M2:

12
32

14
34

15
35

12
0
0
32

14
0
0
34

15
0
0
35

0
0
0
0

Exhibit 14.37. APL selection functions.

We give the most basic meaning of APLs selector functions. The / has additional meanings
when used in other contexts.
Symbol
Name
Semantics
NV
take
Select N values from V and include them in the answer. If N
is positive, take the first N values; if N is negative, take the
last N.
NV
drop
Eliminate N values from V and include the rest of V in the
answer. If N is positive, drop the first N values; if N is negative, drop the last N.
V1 / V2 select
The two array operands must have the same length, and the
value in each position of V1 should be a positive integer, N.
Then N copies of the value in the corresponding position of
V2 will be selected and included in the result.
[...]
subscript Works for matrices of any number of dimensions. An array
element is selected if all of its indices (row, column, etc.) are
included in the subscript list.

430

CHAPTER 14. THE REPRESENTATION OF TYPES

Exhibit 14.38. Using value constructors in APL.

We give the most basic meaning of APLs three constructors. All three symbols have additional
meanings when used in other contexts.
Symbol Name
Semantics
N
iota
Create an array value consisting of the numbers 1 through N.
A1 , A2 catenate If A1 and A2 are arrays, form a new array value by concatenating the elements of A1 and A2.
DA
reshape D is an array of dimensions, and A is an array of values.
Form a new matrix, whose shape is specified by D, containing
the values in A. If there are too few values in A, use them
cyclically until the new matrix is filled up.

APL Constructors. A value constructor takes two or more values, on the stack, and combines
them into a compound value, which it leaves on the stack. APL relies everywhere on value constructors. All APL operators, when given compound arguments, can construct compound values as their
results. An array literal is denoted simply by writing a series of simple literal numbers (without
any delimiters or punctuation) and an array may be input by typing a similar series of numbers. In
addition, there are two operators whose sole purpose is to construct array values [Exhibit 14.38],
and one operator that can be used to cast a one-dimensional array into any multidimensional
shape.
Because APL supports compound selectors, compound dereferencing, and value constructors in
very general ways, compound objects can be processed as easily as simple objects. APL combines
these unusual and powerful data-handling facilities with implicit iteration 10 . The result is that
many algorithms can be written succinctly or as one liners in APL that would require several
lines of code and explicit loops in most other languages.

14.4

Operations on Types

Often a systems programmer wants to design and write library functions or general implementations
of common, useful algorithms. He or she is faced with a type-declaration dilemma:
The actual size and structure of an argument must generally be known to process that structure correctly.
But a generally applicable program should be usable for many variations of a data structure.
For example, a package of stack functions should work on stacks implemented by arrays of
any length or base type.
10

Chapter 10.4

431

14.4. OPERATIONS ON TYPES

Exhibit 14.39. The lone type operation in C.

Assume that T1 is a defined type name. Line (a) declares two pointers of type T1*, or pointerto-T1. Line (b) allocates one storage object large enough to hold a value of type T1. The result of
malloc must be explicitly cast to type T1* before it stored in the pointer variable, pt1. Line (c)
allocates an array of n such objects and stores the reference in pt2.
T1 *pt1, *pt2;
pt1 = (T1*)(malloc(sizeof T1));
pt2 = (T1*)(calloc(n, sizeof T1));

/* a */
/* b */
/* c */

Binding time becomes a problem in a compiled language. Data types for all objects must be
bound (fully specified and fixed) before code can be compiled; allocation and selection functions
depend on this information. In Ada the binding time problem is solved by putting the library code
in a generic package, which is a code schema with one or more type-parameters. To use such a
package you first instantiate the package with particular type-arguments. This binds the types and
creates fully specified code which can then be compiled normally.
But there is still a problem. Code often depends on some particular property of a type, such
as the size of objects of that type or the number of elements in an array. The code in the body of
a generic package needs to have access to this kind of information about the type-arguments used
in the instantiation process. Even when we write library functions for an interpreted language, we
need type predicates that can test the types of arguments and conditionally execute the appropriate
code.
Type-objects provide a clean solution to these problems. They are objects and can be passed
as arguments, just like data objects. A pointer to a type-object serves as a unique identifier for the
type, providing an easy implementation for a type-predicate. A type-object has a body that stores
specific information about the properties of objects of that type. To find the attributes of a type,
we need the ability to access the information stored in its type-object. Thus we need selectors for
the type-objects.
The actual selectors that are meaningful for a type-object depend, obviously, on the type of the
type. The type pointer in a type-object provides a way to find out what selectors are appropriate.
Finally, the language definition must list the selector functions that maybe used for each type of
type.
C provides us with one example of a type operation. The sizeof operator in C may be applied
to any type or any object. It accesses the type-object and returns the value of the size field.
The sizeof operator is necessary in C to provide program portability. It is most commonly
used in conjunction with the dynamic allocation functions, malloc and calloc, which require
the programmer to tell the allocator how many bytes to allocate [Exhibit 14.39]. Since the size of
any structure can vary from machine to machine, an operator was provided that would return the
size of a type in the current implementation.

432

CHAPTER 14. THE REPRESENTATION OF TYPES

Exhibit 14.40. Some of the type operations supported by Ada.

Attribute
Meaning
For any type:
storage_size Total number of storage units needed
size
Number of bits needed for type or allocated for object
For integer types, subranges, and enumerated types:
first
Minimum value in the type
last
Maximum value in the type
For floating point types:
digits
Number of decimals in mantissa of representation
emax
Largest exponent value
epsilon
Difference between two successive representable values
large
Largest positive value
For array types:
first(N)
last(N)

Lower bound for N th index position

Upper bound for N th index position

Ada supports generic packages and, therefore, must provide a variety of type-selectors. These
are referred to as attributes, and they give a package access to many kinds of important information.
Some, but by no means all, of these attributes are listed in Exhibit 14.40. The first one is the analog
of Cs sizeof. The last selectors listed apply to any array type and return the bounds of the N th
dimension. These type-selectors let the library programmer write code that can process any type
array.
Many of these attribute selectors define the limits inherent in an implementation of the basic types. Note the similarity between the third selector, last, and Pascals implementationdependent constant maxint. The type-objects for the basic types in Ada are more complex than
those we diagrammed, since they must also contain the limit information. The ability to get this
kind of information about an implementation can be an important ingredient in writing reliable,
portable code.

Exercises
1. What is the difference between a primitive and a programmer-defined type?
2. What is a type object? What are its components?
3. The process of type-checking in Pascal is very simple and fast, even if the types being compared
are complex and nested. Explain. How is this related to type objects?

14.4. OPERATIONS ON TYPES

433

4. What is a type constant? Why is it used?

5. What is a pointer type? What information is necessary to represent a pointer type? Explain
two ways that this information might be used by the compiler.
6. What is a compound type?
7. What information is necessary to represent an array type?
8. What is an index type? What is its role in an array?
9. How is an effective address computed for a one-dimensional array?
10. How was the implementation of multidimensional arrays simplified in the late 1960s?
11. What is a slice of an array?
12. Compare the highly flexible arrays in APL and the much more restricted ones in Pascal.
Comment on: selection functions, coherent operations, use as parameters, and returning
arrays as function results.
13. What is bounds checking? What is the advantage of having it in a language? The disadvantage?
14. What is a string? A counted string? A terminated string?
15. What is the difference in representation between a C string and a Pascal string?
16. Why was the set type included in Pascal?
17. What are the similarities between an array and a record?
18. What are the differences between an array and a record? Explain why a representation of a
record type is more complex than a representation of an array type.
19. Explain why the fields of a record cannot be accessed by subscript (like elements of an array)
in a strongly typed language.
20. How can a record component be accessed?
21. What is a pathname?
22. What is the difference between a free union and a discriminated union?
23. How has Ada closed the loophole found in Pascals variant record?
24. How are literal objects of primitive types created in a program?

434

CHAPTER 14. THE REPRESENTATION OF TYPES

25. How are literal objects of user-defined types created in Pascal? C?

26. What is a selection function? What does it return?
27. How does a value constructor create a compound object?
28. Why is binding time a problem in compiled languages?
29. How can we find the attributes of a type?
30. Give an example of a type operation in C and Ada.

Chapter 15

The Semantics of Types

Overview
In this chapter we explore the semantics of types. In addition to describing physical
properties of objects, types can be used to define the domain of a function and thereby
control its application. There are several different approaches to domains and domain
or type checking.
We consider the use of types and the development of type checking from the earliest
languages to modern strongly-typed languages. Types are properties of data objects
and they describe data size, structure, and encoding. Types may also be associated
with identifiers to restrict the domain of objects to which an identifier may be bound.
Type restrictions can be enforced at compile time or at run time.
In early languages, types were used primarily for storage allocation, storage access, and
to control application of predefined functions. Two types were considered compatible
if they described the same storage format. This definition of compatibility became
inadequate when newer languages began using types to carry semantic information.
The translators for the newer strongly-typed languages use domain or type checking to
ensure the semantic validity of function calls. A major research problem has been to
find semantically meaningful extensions of the type compatibility rules.
A distinction is made between external domains, those within a programmers application area, and internal domains, the semantic groupings of objects or types identified by
the language translator. While older languages had a fixed number of distinct domains,
type constructors in modern languages permit programmer-defined domains.
Type casts cause a change in the semantics of an objectthe domain label is changed,
but the bits are not. A conversion is a change in the physical properties of an object
435

436

CHAPTER 15. THE SEMANTICS OF TYPES

size, encoding, or reference level. A coercion is a conversion or cast that is automatically
invoked by the translator.
While type checking aids the programmer, it creates an inflexible environment. A
programmer who needs to convert from integer to floating point or to compute a hash
index, for example, would be hampered by strong type checking. Modern languages
often provide escape hatches so that the experienced programmer can evade the type
matching rules.

15.1

Semantic Description

Types can be used to embody both the physical properties and the semantic properties of objects.
In Chapter 5, we explored the use of types to describe physical properties. In this chapter we
look at the other use of types: to define the domain of a function and thereby to control function
applicability. A domain is the set of objects over which a function is defined. Objects in this set
must share common physical properties (size, structure) and semantic properties (encoding, intent).
Functions defined for a domain depend on the common properties of its elements. If the size
or structure of a function parameter differed from what the function expected, the results of the
computation would probably be wrong. Similarly you would get nonsense if the meaning of an
actual parameter was different from the meaning of objects for which the function was designed.
Data type definitions were developed as a way to specify the size and structure of variables, so that
translators could allocate appropriate amounts of storage. Type checking is a way the translator
can use the same information to help the programmer eliminate errors and inconsistencies in code.
Checking was minimal in early languages, but it has become more sophisticated through the
years. In this section we look briefly at several different approaches to domains and domain checking.
We consider the typing rules in a series of languages, from very old to fairly new.

15.1.1

Domains in Early Languages

Assemblers, the earliest computer languages, had implicit domains rather than explicit domains:
addresses, integers, indices, and the like [Exhibit 15.1]. The programmer used elements of these
domains, but their properties and relationships were part of the programming lore, not part of
the language. All assembly language objects were represented by storage locations (or blocks of
locations), each location big enough to store an integer. The type of the object (integer, address,
or index) was not part of the programit existed only in the programmers mind. Computers
typically had some instructions intended to do useful things on each domain. However, the language
translator had no way of knowing whether a variable represented a true integer, an address, or an
index. Thus the translator could not ensure meaningful use of instructions. Some higher-level

15.1. SEMANTIC DESCRIPTION

437

Exhibit 15.1. Operations defined for assembly language domains.

Domain Name
machine address
integer
index

Operations Defined
Goto, fetch a value, store a value.
Arithmetic and comparison operations.
Load index register, add to base address of an array.

languages, such as FORTH, also use a single domain to represent integers, addresses, and indices,
and provide no way to ensure their appropriate use.
Early Typed Languages. Languages designed in the 1950s, such as FORTRAN and ALGOL,
embodied fixed sets of primitive, or predefined, domains. The language and the translator made
distinctions among these domains. The domain of each variable was defined (by default in FORTRAN, by declaration in ALGOL) and became permanently associated with the variable identifier.
The original FORTRAN was a very primitive language. It did not have variable declarations;
the type of a variable was derived by default from the first letter of its name. Names were restricted
to six letters; this is so short that meaningful names were hard to devise. The domains integer
and real were supported as unrelated numeric types that could not be used in combination.
The ALGOL domain structure was much richer than FORTRAN. Integers and reals became
related domains [Exhibit 15.2]. Automatic type conversions were introduced so that values of either
numeric type could be mixed in arithmetic statements. The external domain real was commonly
represented internally by single machine words in floating-point encoding. 1 Integers were also
commonly represented as single machine words in binary sign and magnitude encoding. With this
representation, the internal domain integer is not a strict subset of the internal domain floating
point. Some numbers can be represented exactly both ways, some cannot. Very large integers
have too many digits of accuracy and can only be approximated in floating-point representation.
Nonintegral floating-point numbers can only be approximated in integer representation.
ALGOL introduced Boolean as a distinct domain. Boolean values were produced by comparison
operators and used by conditional statements. All variables were declared, and the declaration
was used for both allocation and type checking. ALGOL had too many primitive types to make
FORTRAN-style defaults useful.

15.1.2

Domains in Typeless Languages

Some languages, which are called dynamically typed languages 2 , support dynamic allocation of
storage objects whenever they are needed to contain an input value or the result of a computation. 3
Identifiers are typeless in the sense that types are not permanently attached to them. Rather, a
1

These were 36 bits long on the IBM 704.

The less precise term typeless languages has also been used.
3
These languages are usually interpreted, not compiled.

438

CHAPTER 15. THE SEMANTICS OF TYPES

Exhibit 15.2. Domain relationships in ALGOL.

Unrelated. The domains Boolean and integer were independent. No relationship existed
between integer values and Boolean values.
Intersection. The external domain integer is a subset of the external domain real; all integers
are reals, and some reals are integers. However, in a typical 4-byte implementation, the domains
integer and real intersect, but neither is a subset of the other.
Numbers Not Exactly Representable as Integers
5,000,000,000 Larger than maximum 4-byte integer.
Has a fractional part.
3.25
Numbers Exactly Representable as Integer and Float
2,147,481,880 No bits in low-order byte of this integer.
3
No bits in high-order byte of this integer.
Number Not Exactly Representable as a Float
2,101,111,111 Integer has 1 bits in both high-and low-order bytes.
Number Exactly Representable as Neither
1/3
Has fractional part that is infinite when expressed in binary.

type tag is attached to each storage object when it is created. The storage object is then bound
dynamically to an identifier. Thus the type that is indirectly associated with an identifier may
change dynamically. Examples of such languages are LISP, APL, and SNOBOL.
In these languages, the programmer does not declare types for variables and function parameters.
However, all those variables have types, which are necessary to describe the data size, structure,
and encoding. These languages all incorporate the various data encodings supported by typical
computer instruction sets (character, integer, floating point, bitstring). Further, the domain of an
object is usually tested, at run time, before a primitive operation is applied to it. Exhibit 15.3 gives
examples of domain-checking in APL, a dynamically typed language. A run-time error comment is
generated if the domain is not appropriate.
Types are not used to control function applicability in these languages. Any programmer-defined
function can be applied to any object. The language does not support the concept that a function
might be meaningful for some arguments but not for others. However, even though the domains of
arguments to programmer-defined functions are not checked, many (but not all) semantic errors are
detected when the program eventually calls a primitive operation with an inappropriate parameter.
In Exhibit 15.4 we define a simple APL function named DEMO that accepts a parameter, S, passed
by value. At function exit, the value of R will be returned as the value of the function. Meaningful
and meaningless calls on DEMO (and their results) are shown in Exhibit 15.5. The domain mismatch

15.1. SEMANTIC DESCRIPTION

439

Exhibit 15.3. Domains are checked by APL primitive operators.

The domain of an argument is checked by primitive operators. Use of an operand belonging to the
wrong domain results in an error comment.
Programmer writes
A 2.11 34.2 17 18.1
2 A [3]
2 A [0 N 0 ]

APLs output
17
Domain Error

Note below
(a)
(b)
(c)

a. An array of four numbers is created and bound to the variable A.

b. 2 means to output the value of the expression on the right side of the arrow. The value
of the third element of the array A is selected and printed.
c. A subscript in APL must belong to the domain integer. A character may not be used as a
subscript.

Exhibit 15.4. A simple APL function.

Note that the domain of the programmer-defined function is not declared and therefore cannot be
checked.
5 R DEMO S
[1] A 2.11 34.2 4.8 18.1
[2] R A[S]
5
Line 1 We define a length-4 array.
Line 2 The parameter, S, is used to subscript A. If the subscript is in the range 1..4, the corresponding item in the array will be selected and bound to R, the local name for the result.

440

CHAPTER 15. THE SEMANTICS OF TYPES

Exhibit 15.5. APL does not check domains of programmer-defined functions.

Programmer writes
2 DEMO 3
2 DEMO N
2 DEMO 5

APLs response
4.8
Domain Error, line 2 of DEMO
Index error, line 2 of DEMO

Notes
(a)
(b)
(c)

a. A correct function call. A[3] is returned.

b. The domain is checked on a call to a primitive function.
c. The subscript range is also checked. 5 is too large for array A.

in the second call is not detected when control enters DEMO. However, it is detected during the
subscript computation, because subscript is a primitive function.
Dynamically typed languages often supply domain predicates for the primitive domains, so that
programmers may write their own domain checks. The most general kind of domain predicate is a
Boolean function of two parameters; let us call it IN. One parameter is an object, Ob, the other is
a domain name, D. The predicate IN(Ob, D) returns TRUE if D is the domain of Ob.
Some languages do not permit a domain name to be used as a parameter. Such a language
might supply a separate domain predicate for each domain. In this case the domain name is made
part of the predicate name, and each predicate tests whether its single argument belongs to a
specific domain. This is well illustrated by the domain predicates in LISP [Exhibit 15.6].
If a dynamically typed language supports domain predicates, a programmer can do manual
domain checking within functions. A function will accept, as arguments, objects of any domain.
Within the function the programmer writes conditional statements that test the domain and take
one of several branches. Conditionals can be set up that will emulate the checking done automatically in Pascal. Thus, whatever operations are ultimately applied to the data, they are sure to be

Exhibit 15.6. Types and domain checking in LISP.

The information given here is for the dialect Common LISP.
Automatic domain checking: primitive functions check the domains of their arguments.
Primitive domains: number (integer), symbol (the name of an object), atom (any nonlist entity),
list (a sequence of lists and/or atoms, delimited by parentheses).
Primitive domain predicates: numberp, symbolp, atom, listp, null (true for empty lists). Example:
(numberp s) returns T (for true) if s is a number, and F otherwise.

15.1. SEMANTIC DESCRIPTION

441

appropriate for the data encoding. Domain testing becomes another form of data-validity checking,
similar to checking for an absurd data value or a table index that is out of range.

15.1.3

Domains in the 1970s

Early C: Domains Checked for Primitive Operators Only.

Type compatibility is a complex question in C. First, we must distinguish among various versions
of C, especially between the old semistandard version of C which is described in the Kernighan
and Ritchie book (we refer to this as K&R C) and ANSI C. There are far-reaching differences in the
typing and type checking rules between the oldest versions and ANSI C. Here we discuss both K&R
C and the ANSI standard.
Second, some, but not all, implementations of K&R C have an accompanying program named
LINT, which may be used by the programmer to examine programs and locate Pascal-style type
errors. A LINT program defines many things as errors that the accompanying C compiler would let
go, and thus LINT defines a language with different semantics than C, but the same syntax. LINT
implements type constructors and domain identity checking similar to that in Pascal.
Third, many recent non-ANSI compilers generate nonfatal warning messages for Pascal-style
type errors, but compile the code anyway. It amounts to a semantic quibble whether the warning
messages or the generated code defines the semantics of the language.
Finally, type compatibility for array and record types in K&R C was almost a moot point. There
was nothing that you could do coherently with a compound object except store its address into a
pointer. Thus the question of type checking for function parameters was reduced to a question of
what could be done with pointers.
A rough generalization about early C translators is that types were used primarily for storage
allocation and access, as described in Chapter 5. Types were declared so that the compiler could
know the number, position, and encoding of the fields of an object. This is essential information for
producing object code. The types of arguments to primitive functions and operators were checked.
But, as in APL, arguments to programmer-defined functions were not checked at all. A function
call could have the wrong number, wrong type, or wrong sequence of arguments, and the violation
would not be detected.
Further, the rules for pointers created a gaping hole in the type structure. A pointer, declared
to point at one type, could be set to point at an object of a different type. Older C compilers would
not comment on this type violation. In newer compilers, this would trigger a warning message,
but object code will be generated anyway. In Exhibit 15.7, the pointer r is made to point at an
object whose type does not match rs declared type. Then a field is selected from rs object, and
interpreted as if it were the proper type for r, which is untrue. This technique can be used to
fool the compiler into compiling code that would cause a type error if it were executed without
the pointers. Thus the language definition and the compiler provide no sure control over whether
or not a pointer is pointing at an object of the appropriate type.

442

CHAPTER 15. THE SEMANTICS OF TYPES

Exhibit 15.7. Pointer assignments in K&R C.

Two types named funny and runny are defined, and are used to declare two pointers (named f
and r) and an ordinary structured variable (named ff). Note that these types are structurally
different. The following statements will compile without causing a fatal error. Recent non-ANSI C
compilers are likely to produce a warning message for the last assignment.
struct runny {int a; int b;} *r;
struct funny {int a; float c;} *f, ff= {2, 3.1};
f = & ff;
r = f;

/* This makes f point at the variable ff. */

/* This sets r to point at the same thing as f,
which is, of course, the wrong type for r. */
After executing these lines, storage will contain:

f
ff
r

3.1

The first expression, below, is normal. The second field of a funny object is selected. It is
expected to be, and is, a float. The second expression, though, selects the second field of a runny
object, which is expected to be an integer, but is actually a float that will be interpreted as an
integer.
f>c Select 2nd part of fs struct; float result is interpreted as float.
r>b Select 2nd part of rs struct; float result is interpreted as int.
Programmer-defined Domains
Time and experience have influenced design philosophy. The historical trend has been to give
domains an increasingly important role in programming languages because they are a great aid
to producing semantically sound programs. In older languages, all domains were predefined, and
their relationships (if any) were predefined and not modifiable. No domain checking was done
that distinguished between objects of similarly represented domains. For example, in C (which
was developed by Ritchie in about 1972), integers and truth values are the same domain. More
important, the types of the arguments to programmer-defined functions were not checked.
In newer languages, such as Pascal (released by Jensen and Wirth in 1974), ways were provided
to define new domains, and more elaborate domain-matching rules were implemented. In addition
to the usual record, array, and pointer types, Pascal included several new kinds of type definitions:
subrange type, enumerated type,4 set type, and type mapping [Exhibit 15.8]. Domain relationships
began to become important. In Pascal, these were pre-defined and unmodifiable. Let us consider
the four types of relationships in Pascal.
4

Enumerated types were not supported by early versions of C.

15.1. SEMANTIC DESCRIPTION

443

Exhibit 15.8. Diagrams of the domain relationships in Pascal.

TYPE index = 0..100;
age
= integer;
Boolean, integer, and real are primitive types.
The domain Boolean is semantically unrelated to integer, even though it is defined by mapping
onto a subset of the integers.
Integer and real intersect; that is, some reals can be converted to integers; most or all integers
(depending on the implementation) are also representable as reals.
The domain index is a subset of the domain integer, and fully compatible.
The domain age is merged with the domain integer. There is no distinction.
integer
age

integer

Merger

Subset

integer
real

index

Boolean

Intersection

integer

Unrelated

Intersection. The domains real and integer intersect in Pascal as they did in ALGOL. Set
domains can intersect with other set domains over the same base type.
Unrelated. Domains created by array or record or pointer type declarations are unrelated to
any other domain. Domains defined by enumeration (including the primitive domains char and
Boolean) are unrelated semantically to any other domain. However, all enumerated types are
represented by integers and can be converted to the domain integer by explicit use of the primitive
function ord.
Subset. The Pascal subrange declaration narrows a domain to a compatible smaller domain
with the same representation. Variables of a subrange type are restricted to storing values in the
subrange. Attempting to store a value outside this range results in a run-time error and program
termination. Using a subrange variable incurs run-time overhead due to automatic range-checking
every time a value is stored in it! While the protection provided by subranges is valuable, especially

444

CHAPTER 15. THE SEMANTICS OF TYPES

during program debugging, the cost is great. Further, because a range error causes a Pascal program
to bomb, subranges cannot be used for ordinary input validation. They are primarily useful for
trapping subscript values that have unintentionally become too large or too small.
Merger. Compatibility rules were very loose in the original K&R C; they are far tighter in Pascal.
For example, truth values, integers, and addresses comprise different domains, and a pointer to a
real is incompatible with a pointer to an integer. The compiler checks the domains of all function
arguments and requires that the formal and actual parameters have the same domain identity (not
just the same representation, as in C).
However, even in Pascal, the relationships among types, structures, and domains are not easily
stated. In Pascal, we may have merged domains, that is, more than one type name associated with
one domain [Exhibit 15.18], and more than one domain associated with one structural description
[Exhibit 15.14]. The number of distinct domains is equal neither to the number of names nor to
the number of different structures defined.

15.1.4

Domains in the 1980s

Languages newer than Pascal embody a clearer view of the nature and uses of domains. For example,
Ada gives the programmer more control over the relationships among new domains (see Section
15.3).
In C, the programmer specifies the physical size of variables by declaring them to be long
int, short int, or int (meaning either one). In contrast, Ada encourages the programmer
to declare domains abstractly, by specifying characteristics (range, precision) of the values to be represented, rather than concretely, by specifying the number of machine locations needed. Programs
containing abstract specifications are more easily ported to machines with diverse architectures.
The run-time overhead incurred in Pascal for range checking is also incurred in Ada. However,
the Ada designers provided a means (the pragma) to turn this checking off. The programmer can
have the dual advantages of automatic checking during debugging and efficiency when a program
enters production use.
Finally, there are languages that let the programmer define a domain with more than one
representation. Having multiply represented domains is basic to the modern functional languages
and the object-oriented languages. 5

15.2

Type Checking

Translators use domain checking (often called type checking) to ensure the semantic validity of
function calls. In a domain-checked language, the domain of each formal parameter of a function is
declared to be a specific or generic data type. When the programmer tries to apply a function to a set
of actual parameters, the translator verifies that the type of each actual parameter is appropriate;
5

These are discussed in Chapter 17.

15.2. TYPE CHECKING

445

the type of the actual parameter must be contained in the domain of the corresponding formal
parameter. If this is true we say the actual and formal parameters are type compatible, or that their
types match.
If any parameter fails this test, the function call will not be carried out and an error comment
will be generated. This process can be done dynamically, rejecting meaningless function calls at
run time, or statically, detecting type errors at compile time. The actual rules for determining
whether two types match can be confusingly complex and vary greatly from language to language.
The rest of this chapter discusses and contrasts the compatibility rules for a few languages.

15.2.1

Strong Typing

The term strongly typed language means, roughly, a language that checks the semantic validity of
all function calls. There is considerable confusion, though, about the exact meaning of this term.
In this section we will go through a series of increasingly sophisticated definitions of strong typing.
Some early authors called a language typed (abbreviated as ST#1) if the programmer used
declarations to specify the types of variables. By this definition, APL and BASIC are not ST#1
and C and Pascal are. This definition classes languages such as BASIC, (where the data type is
implicitly declared by the form of the variable name) with languages such as APL which do not
associate types with identifiers at all. A better definition (ST#2) is: a language is typed if there
is a type associated with each variable name, and only objects of that type can be stored in the
variable. APL is not ST#2 and BASIC, C and Pascal are.
It became clear that the important distinction was not whether every object had a declared
type, but whether those types were used to support semantic validity. A new name and a third,
stricter definition evolved: A strongly typed language (ST#3) is one in which the compiler enforces
these type compatibility rules:
All objects (variables, values, and formal parameters) are divided into sets called types, usually
labeled by type names. Types can be built in or programmer defined. Each object belongs
to exactly one type.
A type may have two or more variants. The variant to which an object belongs can be
ascertained by a program at run time. Perhaps it is encoded by a tag field associated with
the object.
A variable can only store values of the same type.
In a legal function call, the type of each actual argument must match the type of the corresponding formal parameter. The exact definition of compatibility is language-dependent.
Roughly, though, two types are usually compatible if they are the same type or are overlapping
subsets of the same type.
If an argument belongs to a type with variants, the programmer must explicitly test which
variant is present and write code that handles each variant appropriately.

446

CHAPTER 15. THE SEMANTICS OF TYPES

APL is not ST#3. Neither is the original Kernighan and Ritchie version of C, because calls to
user-defined functions are not type-checked. Pascal, Ada, and ANSI C are ST#3 except for some
escape hatches. These are Pascals nondiscriminated variant record and the analogous union
data type in C, as well as ANSI Cs continued support of K&R Cs original function definition syntax
and semantics. BASIC is ST#3, in a rather uninteresting way, because it will never let you apply
numeric operators to strings or vice versa.
When using the method of top-down programming, a programmer begins by working with a
very general description of a problem and then making several passes over that description. On
each pass the programmer specifies more details of both the data representation and the method for
computation. In the middle of this process an abstract solution for the problem has been described.
The algorithms and data are defined in a rough fashion but nothing is specified fully yet. Finally,
after several passes, the programmer arrives at a fully specified algorithm which can be coded.
Languages that are ST#3 offer limited support for top-down programming, primarily because
they require the full specification of data representations too early in the process. Functions can
only be written to accept parameters of fully specified types. In order to define functions over
abstract data types, a language must permit postponement of the time at which representation
details need to be pinned down.

15.2.2

Strong Typing and Data Abstraction

An abstract data type, or ADT, is a combination of one or more abstract domains together with a set
of functions that operate on them, and related data items that characterize the domain. An ADT
is an abstraction. We can define an abstraction in English by specifying its required properties.
For example, we can define a tire as a resilient covering for the rim of a wheel. This definition is
a generic definition because it includes tires of many sizes, shapes, and materials. Similarly, we
can use English to define an ADT; to define stack we define the semantics of push and pop
when applied to a stack object.
Only part of this information can be expressed explicitly within a programming language: the
intent and semantic properties of the ADT must remain implicit. For example, the fact that the
push and pop operations for the stack ADT implement a last-in-first-out accessing pattern is not
expressible in any way except a comment. What we can express in a program are abstract functions
(function headers without accompanying bodies) and abstract domains (domain names without
accompanying data type definitions). For example, Exhibit 15.9 defines the ADT array_stack in
an Ada-like language. An array_stack is an array-based data structure capable of storing multiple
data items, together with the functions Push and Pop that implement a LIFO accessing pattern,
the function Top that returns a copy of the most recently pushed item, and predicates Empty and
Full.
We can make a specific implementation of an ADT in an ST#3 language by defining a specific
data type to represent each abstract domain in the ADT, and a specific function to carry out
each abstract function. For example, to make a specific implementation of a stack, we first define
a specific data type to represent the stack. Functions push and pop are then defined that

15.2. TYPE CHECKING

447

Exhibit 15.9. Definition of the ADT array_stack.

We define the abstract data type array_stack by declaring a parameterized generic domain
and listing the relevant abstract functions. We use syntax modeled after Ada, but extended by
several constructs that are not supported by Ada.
ADT array_stack Begin
ParamType Stack(
T:type;
-- Type of data objects to be stored.
L:integer);
-- Maximum number of simultaneous objects.
Function Top(
S: in Stack)
-- Stack itself is unchanged.
return T;
-- Returns copy of most recently pushed data.
Function Pop(
S: in out Stack)
-- Top item is removed from S.
return T;
-- Returns most recently pushed data.
Function Push(
item: in T,
-- Store item at top of stack.
S: in out Stack)
return Boolean;
-- Return true if push executed successfully.
Function Full(
S: in Stack )
-- Stack itself is unchanged.
return Boolean;
-- Return TRUE if S.store is full.
Function Empty(
S: in Stack )
-- Stack itself is unchanged.
return Boolean;
-- Return TRUE if stack contains no data.
End ADT array_stack;

operate on that type to carry out the semantics defined for the ADT operations. We are able to
define a stack of 100 characters in Pascal (to be represented by an array [1..100] of char), or a stack
of integers (to be represented by a linked list of integers). However, we cannot define the abstract
data type itself in these languages.
This is because, by definition, an ST#3 language has disjoint (nonoverlapping) types. 6 There is
no general provision for talking about groups of types or defining a function that would apply to
more than one type, or to a partially unspecified type. A function definition in an ST#3 language
specifies the types of its arguments, so every ST#3 function depends on a previously defined specific
type and cannot be written in a general way to operate on a generic domain. Thus a program in
Pascal that uses two kinds of stacks requires two sets of type declarations and two separate but
nearly identical definitions of each stack function.
6

Variant records and unions are inadequate support for generic types.

448

CHAPTER 15. THE SEMANTICS OF TYPES

Many Pascal students have asked why they could not make one set of definitions for a stack
and its operations, then have it automatically applied to stacks of any kind of object. The best
that can be done in Pascal is to write a set of stack definitions and store them in a source code
library. To use the stack code, the source code must be edited to change the type definitions, then
compiled. The language does not support compiled modules with type flexibility, and it does not
even provide an automated way to do the editing job.
When Ada was developed, a facility was included to automate this tailoring process. A set of
definitions called a generic package can be defined in Ada, which contains the type declarations
and function definitions for an abstract data type such as stack. These definitions are written in
terms of a type parameter. Later, the package must be instantiated with a concrete type during
the first phase of compilation. This process generates ordinary program code that is ST#3, and it
is then compiled. Generic packages for standard algorithms on commonly useful data structures
can be included in libraries. These are called in and instantiated when needed. The need to write
and debug new code is minimized. However, the end product is the same as if the programmer had
written fully concrete code in the first place. Thus Adas generic facility adds real convenience but
not significant flexibility to the language.
Chapter 17 describes Ada generic packages and examines the support for data abstraction in
more modern languages such as Miranda and C++, which permit the user to define generic types
and functions over these types. Chapter 17 also presents a fourth definition of strong typing
which admits automatic type checking for these nonhomogeneous types.

15.3

Domain Identity: Different Domain/ Same Domain?

15.3.1

Internal and External Domains

There is generally not a one-to-one relationship between type names and distinct domains. Most
languages use type names to define domain membership, but a translator may implement two type
names as the same domain or as distinguishable domains. It is possible in some languages to define
two different type names which have identical semantics and, therefore, denote the same domain,
and also to define different semantics for structurally identical types, so that they become distinct
domains.
We must first distinguish between external domains and internal domains. External domains
are those that occur in the programmers application area. Internal domains are the semantic
groupings of objects or types that are recognized and maintained by a language translator. These
are not necessarily the same thing. Two types, A and B, form distinguishable internal domains
if the translator implements different semantics for them. Operationally that means that some
function is defined differently for type A than it is for type B. Two distinguishable domains are
independent if no function defined for one is applicable to objects from the other, and vice versa. If
this property holds in one direction but not in the other, we say the domains are semiindependent.
If the implementations of two external domains have the same internal semantics, we say they are
internally merged.

15.3. DOMAIN IDENTITY: DIFFERENT DOMAIN/ SAME DOMAIN?

449

Exhibit 15.10. Merging the domains truth value and integer.

Neither C nor FORTH distinguishes between integers and truth values. There are many situations
in which this is convenient. Programmers using C and FORTH commonly use the tricks below
to shorten their code. (The examples are written in both languages.)
Using a truth value as an integer. Your process compares two input streams, item by item,
and gives a TRUE answer if the items are not equal. You wish to count the number of nonmatching
item pairs. This is easy: just add up the TRUE values. (TRUE is represented by 1 in C and by 1
in FORTH.)
C: difference_count = difference_count + (item1 != item2);
FORTH: item1 item2 = NOT difference_count @ + difference_count !
Using an integer as a truth value. You wish to terminate a loop when the input variable, an
integer, equals zero. Easy! Remember that 0 is the representation for FALSE, and write:
C: while (input_variable) { hprocessi } ;
FORTH: BEGIN input_variable @ WHILE hprocessi REPEAT

15.3.2

Internally Merged Domains

Some languages maintain no semantic distinction among domains that have the same internal
representation. For example, in FORTH and C, integers, truth values, and characters are represented
identically. In these languages, an integer is a truth value is a character, and a truth value or a
character is an integer. A value belonging to any one of these external domains can be used in a
context appropriate for any other one. An integer operation may be applied at will to a truth value
or a character. No conversion process is needed to go between these domains.
This is one of the very convenient aspects of C: the translator lets the programmer decide
whether it is meaningful to use a truth value as if it were an integer, and thus does not prevent the
programmer from exploiting an implicit relationship between domains [Exhibit 15.10].
Of course, one could also claim that this kind of code is obscure and ought to be well documented
if it is written at all. In fact, C and FORTH programmers use these tricks frequently, and the very
commonness of such code reduces the difficulty of understanding it.
A more serious cost associated with internally merged domains is that the translator has no way
of knowing which objects belong to which external domain, and so cannot help the programmer
avoid unintended and meaningless operations. This is illustrated by the very strange C code in
Exhibit 15.11. Let us trace the execution of this odd expression:
The ASCII codes for a and b are compared. a is not greater than b, so the answer
is FALSE, which is represented by 0 (4 bytes).
The character c (one byte) is interpreted as an integer. Its ASCII code, 99, must be
lengthened to 4 bytes to match the length of the integer 5.

450

CHAPTER 15. THE SEMANTICS OF TYPES

Exhibit 15.11. Many misinterpretations!

The / and the + are nonsensical operations. They would be flagged as errors in Pascal,
but they are accepted in C because integers, characters, and truth values belong to one internally
merged domain.

if ( ( a > b ) / ( 5 + c) ) x=1; else t=1;

integer 63
FALSE = 0

68

Result is that t becomes 1.

0 = FALSE
FALSE selects the else clause.

We add 5 to 99 giving 104 (4 bytes).

The truth value, 0, from item (1) is interpreted as an integer and divided by 104, giving 0 (4
bytes).
This integer 0 is now interpreted as the truth value FALSE, and used in the if test, selecting
the else clause.
Finally, we store 1 in the variable t.
This example illustrates the variety of ways that C values can be interpreted, but it is so
artificial that it does not provide a convincing example of the cost of internally merged domains.
However, this kind of domain merging does frequently lead to errors, especially for programmers
who commonly use more than one language. Exhibit 15.12 illustrates a common and galling error
that most former Pascal programmers make when they begin to write code in C.
As a general rule, the language that has distinct internal domains representing distinct external
domains is easier to learn and easier to use. A compiler with full type checking is a powerful ally
in the battle to debug a large program. The price paid for this assistance is that the programmer
must explicitly indicate domain conversions.

15.3.3

Domain Mapping

An existing domain may be used to implement a new domain by mapping the elements of the new
domain onto the elements of the old. This produces two domains with a common implementation
but dissimilar semantic intent. For example, one could represent the imaginary numbers by mapping
each imaginary onto the corresponding real and leaving implicit the fact that each one represents
that real multiplied by i.
When a new domain, D 0 , is implemented by mapping it onto an old domain, D, we say that
every element of D 0 is represented by an element from D. In this case, the object from D 0 can be

15.3. DOMAIN IDENTITY: DIFFERENT DOMAIN/ SAME DOMAIN?

451

Exhibit 15.12. The bane of the former Pascal programmer.

The Pascal comparison operator is =, but C uses == for comparison and = for assignment. Intending to repeat a process many times, as long as the variable named a remains equal
to zero, the absentminded programmer writes:
do hprocess to be repeatedi while ( a=0 );
Unfortunately, this process will be executed exactly once. The first time the expression a=0
is evaluated, the value zero will be stored in the variable a, and also returned as the value of the
assignment expression. The while test interprets the 0 as FALSE and terminates the loop.
Moreover, when the perplexed programmer looks at the value in a, she or he will see the number
zero and be unable to understand why the loop did not repeat!

converted into an object from D, or vice versa, by changing only the domain identifier attached
to the object. The physical form of the value does not need to be changed. The same domain,
D, may be used to implement other domains also, making a many-to-one relationship between
implemented domains and an implementing domain [Exhibit 15.13].
Although a D-object is structurally identical to a D 0 -object, different sets of functions are
probably appropriate for D and D 0 . For example, bit strings are sometimes used to represent both
integers and arrays of switches. Division is meaningful for integers but not for switches, and masking
operations may be meaningful for arrays of switches but not for integers. Concatenation might be
a meaningful operation for bit strings that are neither integers nor switches. Thus the semantic
intent of the objects domain, not just its physical representation, determines what operations are
meaningful. The intent should be considered before applying a function to the object.
Mapped Domains with Distinct Identities
The primitive types in Pascal demonstrate domain mapping where all domains have separate identities. Several Pascal primitive types are actually represented in the computer as integers [Exhibit
15.13]. Pascal also permits the programmer to define an enumerated type. To do this, the

Exhibit 15.13. Domains mapped onto integers in Pascal.

Semantic Intent
integer numbers
alphabetic characters
machine addresses
truth values (false, true)
enumerated type

Typical Implementations
full-word bit strings
full-word integers (8-bit if packed)
integers 0 .. virtual memory size
integers (0=false, 1=true)
integers 0 .. cardinality of the type

452

CHAPTER 15. THE SEMANTICS OF TYPES

programmer enumerates names for the elements of a finite type, and the type is implemented by
mapping the elements, in the order given, onto the integers.
The computer hardware itself defines the mappings for integers, addresses, and characters. Integers are defined by the operation of the machine language add, negate, and comparison instructions.
Addresses are mapped onto the integers by the memory mapping hardware of the machine; one
computes the next machine address in a sequence by using the integer add 1 machine instruction.
The mapping from characters to bit strings is defined by the I/O devices, most of which implement
a common character code such as ASCII.
Truth values are mapped onto bit strings by a language implementor. False is virtually always
represented as a string of 0 bits. Anything nonzero is therefore taken to be true. When a true
value must be generated by the translator, most languages generate the string 00000001 with a
number of leading zeros appropriate to fill a memory location. (Some translators generate a string
of all 1 bits.)
Pascal defines primitive functions and procedures for these primitive domains. Although the
representations of these domains are structurally compatible with each other, the primitive functions
only accept arguments from the defined domain. A Pascal translator checks the declared data type
of each argument against the declared domain of a function and enforces what is, ideally, the
programmers semantic intent. If the programmer wants to do some operation that would violate
the type rules, he or she must first explicitly convert the argument to a different type.

15.4

Programmer-Defined Domains

15.4.1

Type Description versus Type Name

In the historical progression from ALGOL, through C, Pascal, and Ada, to C++ and ML, types have
become the basis for increasingly powerful semantic mechanisms and have been given increasingly
clear semantics. Languages have supported programmer-defined types since the 1960s. During
this time the relation between the type description, the type name, and the types semantics have
varied greatly. In the older languages, such as K&R C, a type name was no more and no less than
a shorthand notation for the type description. All types with the same description were merged
into one internal domain. In such a language two structurally identical objects belong to the same
domain even if they were declared using different type names.
A more modern approach is to give meaning to the type name over and above the meaning of
the type description. In such a language the programmer may use the type name to help express
semantics. Thus the same type description may be associated with more than one type name to
express semantically different types that happen to have the same representation [Exhibit 15.14].
The compiler can then use these differentiated type names to help the programmer achieve semantic
validity. In this situation, type checking can catch errors that checking only for structural properties
cannot detect.

15.4. PROGRAMMER-DEFINED DOMAINS

453

Exhibit 15.14. Two types with the same description.

In writing a graphics program one might use two related but distinct types:
A set of points in the Cartesian plane
A set of points in the polar coordinate plane
Both of these types are normally represented by a pair of real numbers. In the Pascal type declarations below, both are given the same structure. The two type names will not be synonyms, since
functions defined for one cannot be applied to the other.
TYPE cart_point = ARRAY [1..2] OF real;
polar_point = ARRAY [1..2] OF real;

15.4.2

Type Constructors

Old languages had a fixed number of distinct domains. In modern languages, type constructors
are provided to permit the programmer to define new domains. Each new domain may be, and
normally is, bound to a new type name. The new type name can be used to declare objects that
belong to the domain and to declare the domain of function parameters.
A type constructor is a keyword or syntactic construct whose use creates a new domain. Not
all ways of declaring new types are type constructors. Type constructors are chosen by a language
designer and vary greatly among languages. For example, array and are type constructors in
Pascal but the analogous [hdimensioni] and * in C do not construct domains. However, struct
in C and the analogous record in Pascal are both type constructors.
Constructed Domains and Type Checking
When a programmer uses a type constructor in a type declaration, the declared type name is
bound to the newly formed domain. Thereafter, other program statements can declare objects and
function parameters in that domain by referring to the type name directly or indirectly [Exhibit
15.15].
In an ST#3 language, a domain, D, created by a type constructor is independent; that is, it
is functionally incompatible with all existing or future domains. 7 No functions may be applied to
objects from D unless they are explicitly defined for D, and functions defined for D may not be
applied to objects from any other domain. An object is type-compatible with a formal parameter
if and only if both were declared to belong to the same domain.
Every occurrence of a type constructor constructs a different domain. If types D1 and D2
are declarations with identical type definitions containing a type constructor, D1 and D2 are
incompatible. In some languages, a type constructor may be used in a variable or parameter
7

In more modern languages, this incompatibility may be modified by rules for domain inheritance.

454

CHAPTER 15. THE SEMANTICS OF TYPES

Exhibit 15.15. A type constructor in Pascal.

The keyword array is a type constructor in Pascal; each use of array builds a new domain. Here
we construct a new domain and bind the name BoxDimensions to it. We use this type name to
define a function, vol, that operates on BoxDimensions.
Giving a new name to a type does not construct a domain in Pascal. The type name Tank
names the same domain as BoxDimensions.
TYPE
VAR

BoxDimensions = array [1..3] of Real;

Tank = BoxDimensions;
p, q: BoxDimensions;
t: Tank;

FUNCTION vol(d: BoxDimensions): real;

BEGIN vol := d[1] * d[2] * d[3] END;
The function vol may be legally applied to variables p and q. The types match because p, q,
and d were all declared to be BoxDimensions. This function may also be applied to t, since it was
declared using the same instance of array as was BoxDimensions.

declaration. This constructs a domain that is incompatible with everything because the new domain
has no name, and other parts of the program are unable to refer to it. The only variables that can
ever belong to this domain are those created by the same declaration.
Thus the variables p and q in Exhibit 15.16 have the same (unnamed) domain. Variables r and
s have the same domain, structurally identical to the domain of p and q, but semantically distinct
because they were declared with different uses of the type constructor array. No other variables
or parameters can belong to the same domain as p and q, or as r and s. Specifically, variable v
does not belong to the same domain as any of these other variables.
An object and a parameter belonging to different domains are incompatible, even if they have the
same structural description. A type constructor, therefore, is never used to declare the parameters
of a function: no variable could ever match the type of the parameter, and the function could
never be called. In Exhibit 15.16, the function NoGood can never be called because the type of its
parameter, t, can never match the type of any variable. Specifically, t is not in the same domain
as p, r, or v, but belongs to a fourth domain with distinct semantics.

15.4.3

Types Defined by Mapping

Non-Independent Mapped Types

In some languages, a type defined by mapping does not construct a new, independent domain.
Rather, the new type name is an alternative way to refer to an existing domain. The typedef
declaration in C is an example. It defines the new type name as a synonym for its definition, which

15.4. PROGRAMMER-DEFINED DOMAINS

455

Exhibit 15.16. Incompatible domains in Pascal.

Four incompatible domains are constructed here, by using the type constructor array four
times. Each of the identifiers p, r, v, and t belongs to a different domain. The function NoGood
cannot ever be called because its parameter, t, is not in the same domain as any other object.
TYPE

VitalStats = array [1..3] of real;

VAR

p, q: array [1..3] of real;

r, s: array [1..3] of real;
v :
VitalStats;

FUNCTION NoGood(t: array[1..3] of real):real;

BEGIN NoGood := t[1] + t[2] + t[3] END;

is often a structural description [Exhibit 15.17]. Although typedef does not create a new domain,
a typedef declaration is useful in C because the syntax for using typedef names is clearer and
more convenient than the syntax for using their descriptions.
C was one of the earliest languages developed that permitted the programmer to declare new
types. At that time, the relationships among the name of a type, its representation, and its
semantics were only partially understood. This is probably why C has fewer type constructors than
newer languages, and why typedef is not a type constructor. 8
8

The typedef declaration was added to C after the language had been in use for some years.

Exhibit 15.17. Merged domains defined using typedef in C.

A point in a plane can be represented as a pair of real numbers. We define a record type using the
C type constructor struct. The identifier pt is called a type tag and can be used to refer to the
constructed domain.
We use typedef here to map three type names onto the domain pt: polar_point, xy_point,
and twisted_point. We now have four names by which we can refer to the same domain, and
they are used, below, to declare a group of type-compatible variables: pp, xp, tp, and sp.
typedef struct pt {float c1, c2;} polar_point, xy_point;
typedef pt twisted_point;
polar_point pp;
xy_point xp;
twisted_point tp;
struct pt sp;

456

CHAPTER 15. THE SEMANTICS OF TYPES

Exhibit 15.18. Merged domains in Pascal.

We declare two new type names, (LengthInFeet and LengthInMeters) by mapping them onto
the domain Real. These two types are synonyms for each other and for Real.
TYPE

LengthInFeet = Real;
LengthInMeters = Real;

Pascal was developed at about the same time as C. Although Pascal does support several type
constructors (array, record, , and enumeration), it does not permit the programmer to
define a mapped domain with distinct semantics. A new type, D 0 , can be mapped onto an old type,
D, as shown in Exhibit 15.18. Objects belonging to D 0 are compatible with functions defined for D,
and objects of type D are compatible with functions defined for D 0 . The two external domains are
merged into one internal domain. Exhibit 15.19 shows examples of function calls on three Pascal
types that belong to the same domain.
In this situation, Pascal does not help the programmer to make the semantic distinction between
variables that represent reals and lengths, nor between different kinds of lengths. The type names
are full synonyms in all contexts, and the semantics of the two external domains are merged. The
fact that the type definitions in Exhibit 15.18 are not type constructors can be attributed to the
lack of full understanding of domains and mapping in the early 1970s.

Exhibit 15.19. Function calls with merged domains in Pascal.

Three variables are declared, one of each type defined in Exhibit 15.18. A new function,
FeetToMeters, is defined for lengths. It accepts a length, in feet, and converts it to the corresponding length, in meters.
VAR r:
f:
m:

Real;
LengthInFeet;
LengthInMeters;

FUNCTION FeetToMeters( f:LengthInFeet):LengthInMeters;

BEGIN FeetToMeters := f * 12 * 0.00254 END;
These variables are fully type-compatible; they belong to the same internal domain. Real
functions, such as * and /, and length functions, such as FeetToMeters, can legally be applied
to any combination of lengths and reals. The following function calls and assignments are all legal,
although those in the second row are semantically invalid.
r := FeetToMeters(f);
f := FeetToMeters(m);

r := f;
r := m * 5.0;

f := 3.89;
m := f / 2.0;

15.4. PROGRAMMER-DEFINED DOMAINS

457

Exhibit 15.20. Creating a compatible type name by mapping.

The new type name length is declared as a synonym for real or float.
Ada
subtype length is float;
Pascal TYPE length = real;
C
typedef float length;

The connection among domains, type names, and representations in Pascal and C is confusing.
We may have one domain associated with more than one type name, and more than one domain
associated with one structural description, as in Exhibit 15.14. The number of distinct domains is
equal neither to the number of type names nor the number of different structures defined.
Mapped Types that Form Distinct Domains
The confusing connection among domains, type names, and representations was cleaned up in
Ada. The type structure in Ada is richer than either Pascal or C. Ada permits the programmer
to declare a type D 0 , mapped onto D, but choose whether D and D 0 will be synonyms or name
distinct domains. Ada provides two different declaration forms for using an old type to represent a
new one.
The first form, marked by the keyword subtype, creates a new name for part (or all) of
the old domain [Exhibit 15.20]. The names become synonyms as they would in the corresponding
Pascal [Exhibit 15.18] and C [Exhibit 15.17] declarations.
The second Ada form, marked by the keyword new, creates a derived type [Exhibit 15.21]. A
derived type is a new domain that is at least partially distinct from the old domain.
An issue arises with using a new domain that was created by mapping. If the new domain
were completely incompatible with existing domains, no functions would be defined on it, and
you couldnt define any! We must use existing function definitions to define the set of operations
appropriate for the new domain.
The primitive operations of subscript, part selection, and dereferencing are defined for all domains that are constructed using the type constructors array, record, or . Given an object,
O, of a new domain, we can use these primitives to extract parts of O that belong to old domains.
Then we can use the functions defined for the old domains to implement the primitive functions for
the new domain. We need a similar way to go from new new domain to old for mapped domains.

Exhibit 15.21. An Ada derived type is a new domain.

The type name new_type will represent a domain that is only partially compatible with the
domain named old_type.
type new_type is new old_type;

458

CHAPTER 15. THE SEMANTICS OF TYPES

Exhibit 15.22. Type casts in Ada.

The domain tonnage is derived from the domain float. Casts are predefined in Ada for derived
types. Two casts are used here in order to permit mixed-domain arithmetic.
type tonnage is new float;
t1, t2: tonnage;
ff : float1;

-- Two variables of type tonnage,

-- and one of type float.

ff := 3.2;
t1 := tonnage( float(t2) * ff );

This could be in the form of a general casting operation, REP, that would take an object of a
new mapped domain and extract the object that represents it in the old domain. REP would be
a compile-time operation, changing only the domain label and not the representation. An inverse
cast, MAKE, would also be needed to relabel values computed in the old domain as elements of the
new domain. REP and MAKE are analogous to the Pascal functions ord and chr, but they describe
the general relationship between a mapped domain and its representation, rather than the specific
relationship between characters and integers [Exhibit 15.22].
Ada supports type casts between representing and represented types. A cast is called by putting
the target domain name in front of the value to be relabeled. REP and MAKE casts are both written
this way.
Derived types could have been defined in Ada as completely independent new domains, but they
were not. One-way compatibility was retained. We summarize the compatibility rules here, for a
new type T derived from an old type R:
Literals of type T are written exactly like R literals.
Functions predefined for type R can be applied to type T.
Functions defined for type T cannot be applied to type R.
A value of type R can be explicitly cast to type T, and vice versa.
Such a cast must be done before values of types R and T can be mixed in an expression or
an assignment statement.
These rules implement one-way semantic protection. Objects of new_type can use functions for
old_type, but not vice versa [Exhibit 15.23]. The advantage of this partial compatibility is that
the basic definitions for a new ADT can be a bit shorter, because explicit casts are not required.
Literal values, also, dont need to be cast to the new type. The costs of this compatibility are that
the Ada programmer must learn a complex set of type compatibility rules, and that the compiler is

15.5. TYPE CASTS, CONVERSIONS, AND COERCIONS

459

Exhibit 15.23. Domain compatibility rules in Ada.

Two new types are defined to be represented by floats. Length is a synonym for float and distance
is semantically distinct. Variables are defined of the three types.
subtype length is float;
type distance is new float;
flo: float;
len: length;
dis: distance;
The following expressions are all legal (put is an output
flo := 8.1;
len := 8.1;
flo := flo+2.0;
len := len+2.0;
put(flo);
put(len);
flo := flo * flo;
len := len * len;
len := flo;
len := len + flo;

procedure predefined for floats).

dis := 8.1;
dis := dis+2.0;
put(dis);
dis := dis * dis;

The following statements have type errors.

dis := flo;
An explicit type cast, as in: dis := distance(flo); must be done
before the assignment.
dis := dis + flo;

Mixed type arithmetic is not allowed. One operand must be cast

before the operation.

unable to detect half of the unintentional type errors. Further, after the basic functions for a new
domain have been defined, those functions can and should be used exclusively in defining further
operations. Compatibility with the old type is no longer needed or desirable, but it is still permitted
by Ada. This seems to be a real defect in Adas type system.

15.5

Type Casts, Conversions, and Coercions

In the previous section we mentioned the topic of type casts briefly [Exhibit 15.23]. Here we explore
the nature of casts and also examine other kinds of conversion processes. We examine what happens
to the physical type and the semantic properties of the converted object during the conversion. In
Section 15.6, the type compatibility and conversion rules in a number of common languages are
described in detail.
Conversion processes can be classified into two categories: conversions and casts. Processes in
both categories change some property of their parameter. Both might be called for explicitly or
invoked automatically. The words cast, conversion, and coercion are all in common use, but

460

CHAPTER 15. THE SEMANTICS OF TYPES

Exhibit 15.24. A hierarchy of mapped domains.

The following set of domains are all represented internally as floats, but they have different
operations defined.
Dimensions (but not distances or ages) can be multiplied to compute areas and volumes.
Distances (but not ages) can be added to compute a total distance.
Neither dimensions nor distances can be added to floats.
Ages can be subtracted from each other to yield a float (age difference) and added to floats
to produce another age.
The domains length and age were defined by mapping directly onto float. Domains dimension and distance were defined by mapping onto length. We have thus created this tree of
domains:
Dimension

Distance
Length

Age
Float

have fuzzy meanings.9 We use all three terms, and give them distinct meanings. We define the
terms briefly here and explain them at length in the following sections. A type cast is a change in
semantic labeling involving one domain that is mapped onto another. A conversion is a change in
the size, encoding, or reference level of an argument. A coercion is a conversion or a cast that is
invoked automatically by the translator.

15.5.1

Type Casts.

Section 15.3.3 dealt with the domain relationship called domain mapping, in which a new domain
is implemented by setting up a correspondence between its members and members of some existing
domain. Typically, the old and new domains are semantically unrelated. Several domains, D 10 ,
D20 , etc. may all be mapped onto one representing domain, D. Mapped domains can thus form a
tree-structured hierarchy [Exhibit 15.24].
A type cast is a conversion between mapped domains. It is a curious thing: it leaves the bits of
a value unchanged but alters its domain label, thereby changing the semantics of the object. Type
casts convert values in a represented domain, D 0 , to or from values in the representing domain, D.
9

The word cast is used in books about C to include all explicitly written casts and conversions.

15.5. TYPE CASTS, CONVERSIONS, AND COERCIONS

461

Exhibit 15.25. Pascal casts for integer-mapped domains.

Represented
Domain D 0
enumerated type
character
addresses
truth value

Representing
Domain D
integers
integers
integers
integers

Casting Functions
D 0 to D
D to D 0
ord
Not primitive
ord
chr
Prohibited
Prohibited
ord
Not primitive

For example, binary integers are used to represent characters in Pascal. The function chr, which
takes an integer and returns the corresponding character, is a cast; so is its inverse, ord [Exhibit
15.25].
The operation of casting happens entirely at compile time. A cast actually does nothing to its
argument except relabel the argument value with a different type-object. The purpose of a cast
is to communicate to the compiler that it is meaningful to use an object in what would appear to
be the wrong domain context. This prevents the compiler from generating a type-error comment.
The compiler does not generate run-time code for a cast operationthe physical representation of
the cast argument is already appropriate for the target domain and does not need to be changed.
Once the basic functions and data structures of an ADT are defined, an application program
works within the defined domains and rarely needs casts. However, the ability to cast a value
between the represented domain and the representing domain can be very important in defining
the basic ADT functions. For example, with the mapped domains imaginary and real, one
needs to use real operations in order to define imaginary arithmetic. To do this one needs to
cast the imaginary operands to real, do real arithmetic, then (in some cases) cast the result
back to imaginary [Exhibit 15.26].

Exhibit 15.26. Using casts in a mapped domain.

Using Ada, we define imaginary numbers by mapping them onto the reals. (The real domain is
called float in Ada. Addition and multiplication of imaginary numbers is then defined in terms
of arithmetic on the reals. Explicit casts are used here to clarify the semantics, even where they
could be omitted.
type imag is new float;
function "+" (q,r: imag) return imag is
begin return imag( float(q) + float(r) ) end "+";
function "*" (q,r: imag) return float is
begin return -1 * float(q) * float(r) end "*";

462

CHAPTER 15. THE SEMANTICS OF TYPES

Exhibit 15.27. Implementing a cast in Pascal.

If we accept the semantics that 0 represents FALSE and all nonzero integers represent TRUE, the
following code converts an integer to a truth value:
FUNCTION IntToTv (k: integer): Boolean;
BEGIN if k = 0 then IntToTv := FALSE else IntToTv := TRUE END;

Pascal places limitations on casting which make the language clumsy or inappropriate for systems programming. Certain important type casts are not supported at all. In one such case, we
can write a simple function that converts integers to truth values [Exhibit 15.27]. But the casts
for address to integer and integer to address are prohibited altogether and cannot be implemented
within the defined semantics of the language. Thus any kind of address arithmetic is prohibited in
Pascal, causing some systems programmers to avoid the language.
We say that a cast is a promotion if it moves from domain D to domain D 0 , like chr. A
demotion, such as ord, casts a value from D 0 to D.
A promotion cast adds a layer of semantics to the object that it did not formerly have. With
demotion casts, there is a loss of semantic information. The cast strips off the semantics of domain
D 0 , like an extra suit of clothes, exposing the underlying semantics of the domain D. If D was itself
a represented domain, another demotion cast would strip off another layer of semantics.
Casts Are Essential. Both demotion and promotion casts are essential during the bootstrapping
process that creates the functions for a new domain. Demotions permit operations defined for the
old domain to be used on demoted members of the new domain. One must demote a value and
operate on the underlying representation since there are, initially, no functions for the new domain.
A promotion cast must be applied to the result of the computation to lift it back to the level of
the new domain [Exhibit 15.28].
After this bootstrapping process is finished, elements of the new domain can and should be
manipulated only by functions defined for the new domain. It would be desirable then to seal off
the mapping relationship and prohibit further demotion and promotion casts.
Casters Beware! Casts are inherently dangerous operations; changing the meaning of an object should not be done lightly. It is legitimately done only in the process of implementing new
domains, and in systems programming environments where the programmer is forced to deal with
the underlying representations of objects in order to achieve acceptable efficiency.
Many languages use type coercion to change the type of an argument to the type expected by an
operator. This is semantically meaningful if the two domains have related semantics, like integers
and reals. But a type cast, by its nature, relabels a value with the identity of a domain that is
usually unrelated, and thus changes the meaning of the value. A programmer writing an explicit
cast presumably knows what he or she is doing and is taking responsibility for the meaning of the

15.5. TYPE CASTS, CONVERSIONS, AND COERCIONS

463

Exhibit 15.28. Using casts in Ada.

Here we declare two mapped domains,
domains:
type length is new integer; -type area is new integer;
-k: integer := 15;
-x: length;
-a: area;
--

length and area, and variables from the new and old
Length and area form new domains.
Area and length are independent domains.
Declare and initialize integer k.
x is a variable in the domain "length".
a is a variable in the domain "area".

A type cast is called for by writing the name of the target type followed by parentheses enclosing
an element from the original type. We may create a length-value out of an integer by using a
promotion cast:
x := length(i)
A demotion cast must be used to define basic operations for lengths. Here we define an additional
method for the primitive operator *.
function "*"( x, y: length) return area is
-- Note a
begin
return area(integer(x) * integer(y));
-- Notes b, c
end "*";
a. The function returns a value of type area.
b. The demotion cast, is used twice, integer(x) and integer(y), so that the meaning of length
multiplication may be defined in terms of multiplication on the underlying domain, integer.
c. The promotion cast, area(z), is used to promote the underlying integer representation to the
appropriate mapped domain. Note that length is not the appropriate domain.

464

CHAPTER 15. THE SEMANTICS OF TYPES

Exhibit 15.29. Mangled meaning in PL/1.

Let us examine the automatic conversions triggered by the expression: a <= b <= c

IF ( a<= b <= c )

either TRUE = 1
or FALSE = 0

THEN x=1; ELSE t=1;

Result is that x becomes 1.

always TRUE for c > 0

TRUE selects the THEN clause.

1. Compare a to b; the answer is a truth value, represented as a bit.

2. Demote the domain of the truth value to the domain bit string.
3. Promote the domain of the bit string to the domain integer, to match the integer c and the
definition of <=.
4. Promote the length of the integer to the length of c.
5. Compare the resulting integer to c. As long as c is positive, the result will always be TRUE.
6. Use this TRUE in the if test, selecting the then clause.
7. Store 1 in the variable x.

result. A flexible language must support explicit casts. However, a compiler has no understanding
of the meaning of anything. It certainly cannot discern those contexts in which it is appropriate to
change the meaning of an object.
A complication of promotion casts makes them particularly dangerous to invoke automatically.
Many domains can be mapped onto a single representing domain. When a demotion cast is used
on any of these mapped domains, it exposes the semantics of its single underlying domain. But
a promotion cast might go in any of several directions. If a promotion cast is done by a compiler
that cannot understand the code, it may utterly distort the meaning of a value. A combination
of automatic demotion and promotion casts is likely to produce total nonsense, as in the PL/1
example in Exhibit 15.29
Consider what has happened here. The second item relates a mapped domain to its underlying
representation and maps truth values onto length-1 bit strings; 1B represents TRUE and 0B
represents FALSE. These domains do not have the same semantic intent. When a demotion cast
is made, the semantic intent of the mapped domain is lost. In this example, when FALSE or TRUE
is represented as a single bit and demoted to type bit string, it causes the original intent to be

15.5. TYPE CASTS, CONVERSIONS, AND COERCIONS

465

forgotten. It is no longer possible to tell that this bit string originally represented a truth value.
In the third and fourth steps, the single bit is promoted to integer and lengthened to the size
of the integer c by padding it with leading zero bits. Lengthening is a semantically safe operation,
but promotion is not. In the fifth step, the former truth value is used as an integer, even though
this meaning is wholly inappropriate. The nonsense result of a <= b <= c is a result of the
automatic invocation of semantics-changing casts, a demotion followed by a promotion.

15.5.2

Type Conversions

A type conversion changes one or more of the physical properties of the object: its size, its encoding,
or its reference level. Examples of encoding conversion routines are the FORTRAN functions INT
and NINT, which change a number from floating-point encoding to binary integer encoding, using
truncation and rounding, respectively. Changing a short integer to a long integer, or vice versa, is
a size conversion. A reference conversion substitutes a value where an address or pointer is given
(by fetching) or an address where a value is given.
We are often concerned whether a conversion preserves all the information in a value, or whether
some information is lost. We use the term conservative for conversions in which all of the information
originally present is retained in the converted form, and information-losing for conversions that do
not preserve all information.
Size Conversion
Size conversion changes the number of bytes used to represent a value without changing the values
encoding or its semantics. Some modern machines support integers of three or four lengths, and
floats of two lengths. Many languages reflect this hardware structure by supporting types with the
same encoding and different sizes. This produces a flexible language that can be used to achieve
both time and space efficiency. However, when operands of different sizes are mixed in an operation,
their sizes must generally be adjusted to match.
For example, C provides integers of at least three lengths, 1 byte, 2 bytes and 4 bytes. The 1byte size, called char, and the 2-byte size, called short int or short, are needed to achieve
acceptable storage efficiency. Moreover, short integers correspond to the hardware capabilities of
personal computers, and so can be faster to fetch and faster in arithmetic operations than longs.
Four-byte integers, called long int or simply long, are needed to store many pointers and
numbers greater than 32,767.
We call size conversions which lengthen the representation promotions, and those which shorten
it demotions. Size promotions are always conservative; demotions lose information if the argument
is large. Size adjustment is probably the safest and most useful kind of conversion, and the easiest
to implement. Size adjustment is the only kind of automatic conversion that is generally supported
in Ada.

466

CHAPTER 15. THE SEMANTICS OF TYPES

Exhibit 15.30. Reference level in C.

We use C to define and initialize several objects that have integer encoding, long length, and various
ref-levels.
#define SAMPLE 50000L
/* "L" makes this a long integer, ref-level 0. */
long int k
= SAMPLE, /* Ref-level 1: a variable. */
*pk
= &k,
/* Ref-level 2: a pointer. */
**ppk = &pk;
/* Ref-level 3: a pointer to a pointer. */

Encoding Conversion
When an external domain has multiple internal encodings, an encoding conversion maps one encoding of an object onto its nearest counterpart in the other encoding scheme. The ideal is that
unconverted and converted forms are both representations of the same external object. This is
more or less true for the conversions integer-to-floating point and floating point-to-integer that are
predefined in many languages. If the machine implementation of floating point uses more bytes
than implementation of integers in the same machine, the conversion from integer to floating point
is conservative; all integers can be represented exactly in a long enough floating-point encoding.
In the other direction, floating point-to-integer conversion is defined to give as good a mapping
as possible, although information is necessarily lost if the number has a fractional part or a large
exponent. The answer returned from a conversion can be either the integer value nearest to the real
value (conversion by rounding) or the first integer value closer to zero than the real value (conversion
by truncation). Some languages (for example, FORTRAN and Pascal), provide primitives for both
versions, others provide only one. For example, Ada) supports only conversion by rounding.
Reference Conversion
Let us define the reference level, or ref-level of an argument to be the number of times a dereference
operation must be applied to arrive at a pure value. Thus the ref-level of 3 is 0, and the ref-level
of a pointer pointing to a variable that contains this 3 is 2 [Exhibit 15.30]. A reference conversion
takes an argument of one reference level and returns a result with a difference reference level (but
the same encoding and size).
A reference demotion takes a storage object and returns a program object. A reference promotion takes a program object and returns a storage object. Demotions are used all the time;
promotions are rare.

15.5.3

Type Coercion

Conversion processes can be invoked two ways: explicitly or automatically. A call on a conversion
function or a type cast that is written in the source code is an explicit conversion process. A
coercion or automatic conversion is one that is invoked by the compiler but does not appear in the

15.5. TYPE CASTS, CONVERSIONS, AND COERCIONS

467

Exhibit 15.31. Coercions and explicit conversions in Pascal.

Line (a) contains two explicit type casts. The function ord casts its argument from the domain
char to the underlying domain integer.
The addition in line (b) will trigger a coercion from integer encoding to real (floating point)
encoding. The compiler will invoke the encoding conversion because no machine instruction can
add an integer to a real without conversion. Converting the value makes sense of the source code.
The right sides of lines (a) and (b) trigger reference coercion, to extract values from the variables
c, Number, and DigitValue.
Var c: char;
Number: real
DigitValue: integer;
DigitValue := ord(c)-ord(0);
Number := Number*10.0 + DigitValue;

{ a }
{ b }

These two lines could appear in an input conversion routine that manually converts numbers in
an ASCII input stream to floating-point encoding. (This is not the best way to do the job, though.)

source code. Language definitions generally specify a set of possible coercions, which is a subset of
the conversions that can be applied explicitly.
A translator will attempt to coerce an argument when the source code does not make sense
as written, because the domain of some argument fails to match the domain of an operator or
function applied to it. If one (or a series) of the permitted coercions will make the argument fit the
context, it (they) will be invoked [Exhibit 15.31]. Coercions are used in many languages to change
the encoding, the size, and/or the reference level of an argument.
When a coercion must be done because of mixed-type operands, a translator could theoretically
choose to convert either argument to the type of the other. However, it is usually true that
one direction is conservative and the other is not. A conservative conversion process is always
preferred over the inverse information-losing process because it is unlikely to distort the meaning
of the argument. This is particularly important when the translator (which does not and cannot
understand the meaning of the code) is invoking the conversion. A programmer might find that the
nonconservative conversion does exactly what is needed when it discards some information. But
the translator has no way of knowing all of a programmers intent, and it must always play it
safe by choosing a conservative transformation when possible.
Size Coercion
When a language supports multiple sizes of types with the same encoding, keeping track of the
size of objects and manually converting between sizes is burdensome and distracting. It becomes
more burdensome as the number of related types grows. Such languages typically use size coercion

468

CHAPTER 15. THE SEMANTICS OF TYPES

Exhibit 15.32. COBOL fixed-point size specifications.

PICTURE Type Denoted
999
A three-digit integer.
9V99
A three-digit number with two decimal places.
999PP
A number with two implied trailing zeros, in the range from 100 to 99,900.
PP999
A number with a leading decimal point followed by two implied zeros,
in the range from .00999 down to .00001
(promotion) in any context where the argument sizes are mixed. Size demotion is not a conservative
operation, so arguments are only coerced to the smaller size when a long value must be stored in
a small storage object.
For example, numbers may be declared in COBOL to be from one to many digits long, with
a decimal point anywhere, either within these digits or any number of places before or after the
string of digits. A PICTURE clause which conveys all this information is included in each variable declaration [Exhibit 15.32]. COBOLs type flexibility is actually important in many business
applications where the number of decimal places maintained before rounding is vital. But in order
to use two numbers of different sizes in an arithmetic expression, one number must generally be
converted to the size of the other. COBOL does this automatically; it would be a distracting and
awkward process for a programmer to manage.
Finally, a modern language must support the use and manipulation of character strings, either
as part of the language or as part of a standard library. The string facility must conveniently provide
for reading strings of varying lengths, storing them in variables, and for operating on strings of
different lengths.
Storage can be handled in two ways. Dynamically typed languages (APL, SNOBOL) and BASIC
store strings in a dynamically allocated string storage area and implement string variables by
binding the variable name to a string value. This permits string length to vary as needed. Typed
languages (C, Pascal, FORTRAN, Ada) preallocate storage for all variables and allocate fixed-size
areas to hold strings. Strings stored in these areas are adjusted in length by truncation, marking
the right end with a string termination symbol or padding the right end with blanks.
A language (or string library) should handle input, output, padding (lengthening), and truncation (shortening) operations automatically. Unhappily, many languages do not. Pascal, for example,
forces programmers to write their own string input routines and manually pad or truncate strings
to make them fit into the fixed-length spaces.
Encoding Coercion
Encoding coercions are important when a program deals with semantically related domains with
different encodings. FORTRAN originally did not implement automatic conversions in arithmetic
expressions. (That is, mixed type expressions were prohibited.) This prohibition was dropped in
FORTRAN 77 because of popular demand. Experience showed that it was a terrible nuisance for

15.5. TYPE CASTS, CONVERSIONS, AND COERCIONS

469

programmers to concern themselves with explicitly converting a 1 to a 1.0 for a computation or a

4.0 to a 4 in order to use it as a subscript.
APL and BASIC go farther. They dont even require a programmer to make a distinction
between integer numbers and real numbers but choose an appropriate encoding dynamically and
coerce numbers when necessary.
Integer-to-real and real-to-integer conversions are both applied automatically in many languages
(FORTRAN, C) but under different circumstances. The conservative conversion is always used in
cases where arithmetic is done on operands of mixed type. This implements the intuitive requirement that information, and therefore meaning, should be preserved during all operations in order
that the ultimate result be meaningful and as accurate as possible.
The information-losing conversion is only applied automatically when absolutely necessary, that
is, when the programmer orders that a floating-point number be stored in an integer location. If
the programmer did this purposely, she or he is likely to be using this assignment command as an
easy way to invoke the truncation function. If the programmer did not realize this was a mixed
type assignment, the conversion is likely to result in a program error.
One might claim that such an information-losing conversion function should never be automatically invoked and that a programmer should explicitly invoke an information-losing conversion if
one is needed. This is, in fact, the rule adopted by Pascal. Ada goes even furtherit does not, in
general, invoke encoding coercions of any kind. 10
Reference Coercion
A reference coercion happens when a program object is supplied in a context that requires a storage
object, or vice versa, and the translator makes good the discrepancy. Automatic dereferencing,
invoked by language translators in most languages, is a downward coercion; an address is supplied
in a context that requires a value [Exhibit 15.35]. The translator fetches the required value from
the given address. In simple contexts, such as the arithmetic statements in Exhibit 15.31, reference
coercion is a real convenience and eliminates the explicit dereferences that would otherwise clutter
the code. However, in linked list programs, which contain many situations in which automatic and
explicit dereferencing must be mixed, reference coercion leads to endless confusion. It requires great
care to write the correct combination of explicit and implicit dereferences in complex situations.
The last two lines of code in Exhibit 15.35 illustrate the nonintuitive nature of expressions that
combine explicit dereference (*) and reference coercion.
As seen in Chapter 6, languages exist that do not do automatic dereferencing. However, it is
so omnipresent that programming students tend to accept it as the only normal convention and
forget that a conversion process is happening.
The opposite coercion, from value to reference, is rare, since it amounts to creating a nameless
variable and storing a value in it. Under restricted circumstances it is done automatically in
FORTRAN. All function parameters in FORTRAN are passed by reference, but a programmer is
10
Curiously, there is one exception to this in Ada; a integer value will be coerced to a real type if it is multiplied
by a real value.

470

CHAPTER 15. THE SEMANTICS OF TYPES

permitted to call a function with a pure value as the argument. In this situation FORTRAN must
coerce the value to an address by storing it somewhere and passing the address to the function.
Although this is the extent of reference coercions in the familiar languages, other languages (EL1,
Aleph) have been designed which apply a second reference coercion automatically when needed to
fit the context. We are all familiar with the automatic dereferencing of variables to get values.
These languages will dereference a pointer to get a value, if it fits the context.

15.6

Conversions and Casts in Common Languages

This section gives a description of the type transformations that are implemented in COBOL and
in five languages from the ALGOL family (FORTRAN, C, PL/1, Pascal, Ada). In all except Ada, if
a conversion is conservative it will be performed automatically when operands of mixed type are
used in an expression. A nonconservative conversion will only be performed when the programmer
directs that a value be stored in a variable of the lesser type. Ada has more restrictive conversion
rules.

15.6.1

COBOL

Both size and encoding conversions are performed automatically whenever they are necessary to
perform a specified operation. In practice this means almost constantly, for the following reasons:
first, objects of many sizes are usually declared; second, the encoding ASCII-string is used for most
numeric variables, but the encoding packed-decimal or binary-integer must be used for arithmetic.
Thus the language processor is constantly involved in adjusting sizes and changing encodings. The
conversions it uses are normally conservative. An information-losing size conversion will only be
performed when the programmer directs that a value be stored in a variable too small to hold it,
in which case the value will be truncated.
Because COBOL does use automatic size demotion, it is possible to lose the high-order digit(s) of
a number by attempting to store it in a field that is too short. When this happens in an arithmetic
statement, an error indicator is turned on. The programmer may (or may not) choose to test the
overflow indicator in the program. In a simple assignment statement (MOVE), though, no indication
of error is given. The programmer must be careful to avoid this, as the language does not provide
a reasonable level of error checking.

15.6.2

FORTRAN

The standard language supports a variety of types. In addition to the usual CHARACTER, LOGICAL,
INTEGER, and REAL, there are two numeric types, DOUBLE (double-precision floating point) and
COMPLEX. No conversions or casts are defined for the type LOGICAL.
Two explicit casts are provided as standard functions in FORTRAN 77. These are never applied
automatically.

15.6. CONVERSIONS AND CASTS IN COMMON LANGUAGES

ICHAR(ch)
CHAR(i)

471

Demotes a character to the underlying ASCII code.

Promotes an integer in the range 0..127 to a character.

Only two primitive types share both semantics and encoding; thus only one pair of size conversion functions is built into the language. The conversion marked * is conservative.
*

DBLE(r)
REAL(d)

Promotes the size from real to double precision.

Demotes the size from double precision to real.

Encoding conversions are more numerous because three distinct numeric types are available,
integer, floating point (real and double), and complex. The conversions marked * are usually
conservative.

*
*
*

15.6.3

INT(r)
NINT(r)
REAL(i)
REAL(c)
DBLE(i)
CMPLX(x)
CMPLX(d)

Converts
Converts
Converts
Converts
Converts
Converts
Converts

real, double, or complex to integer by truncation.

a real or double to an integer by rounding.
an integer to single-precision floating point.
complex to real by discarding the imaginary part.
from integer to double-precision floating point.
from integer or real to complex.
from double precision to complex.

Pascal

The standard language supports a few encodings that can occur in types of different sizes: character
strings can be declared to be of any length, and character strings and Boolean arrays can be either
packed or unpacked. String lengths are adjusted for comparison and assignment as necessary.
Character strings and Boolean arrays are packed and unpacked automatically, when necessary, in
a way that is quite transparent to the programmer.
Standard Pascal provides two types, integer and real, that implement domains with related
semantics. The conversion functions accessible to the programmer are:
round(r) converts a real to an integer by rounding.
trunc(r) converts a real to an integer by truncation.
No function is provided to explicitly convert an integer to a real. Integer literals are also considered to be real literals, and the values of integer variables are converted automatically when necessary. Reals are never automatically converted to integers, because that is usually an informationlosing conversion. This conversion must always be done explicitly.
Implementations of Pascal for microcomputers typically support two more types with integer encoding, a second size (so that both 2-byte and 4-byte integers are provided) and the type unsigned
integer (frequently used to implement bit masks). Conversions are provided that promote short
integers to longer ones. These are conservative operations and are applied automatically in some
common implementations. Casts from unsigned to integer may also be performed automatically.
When these are combined with the normally safe length changes, they can result in unintended

472

CHAPTER 15. THE SEMANTICS OF TYPES

changes of meaning as is typical with automatically applied casts. It is in these areas that Turbo
Pascal deviates most strongly from the standard language. Standard Pascals semantic protection is
lost in order to gain access to the underlying representation and convenience in handling mixtures
of the three integer-encoded types.

15.6.4

PL/1

In order to understand Adas restrictions on automatic type conversion, it helps to understand the
effects of the generalized conversion rules in PL/1. Extensive, generalized, automatic conversion is
supported, so that a PL/1 translator will convert any type to any type. The language designers
perceived meaningful domain relationships between several pairs of primitive types (realinteger;
bitstringanything; numeric character stringnumber; bittruth value; shorter objectlonger
object of same encoding). Conversions were defined for all of these relationships and are invoked
automatically by the translator, singly or in series, whenever the programmer codes a mixed-type
operation. Because the conversion from any type to any type is defined, expressions that are
syntactically illegal in many other languages become legal in PL/1 and produce some surprising or
nonsensical results [Exhibit 15.29].
Since any PL/1 object is type compatible (by conversion) with any other object, the language
and translator cannot use domain checking to help the programmer achieve semantic validity. This
is a great loss, and it far exceeds the value of automatic conversion mechanism. Coercions are
convenient but not necessary, since explicit conversions could be used instead.

15.6.5

Basic Types in C. C implements a large variety of types semantically related to integers. Integers
may come in three lengths, short int (2 bytes in the ANSI C standard), long int (4 bytes), and
char (1 byte). The type name int refers to either long or short, whichever is more efficient for
the hardware.11 The type name char means 1-byte integer; the semantics of chars are not
differentiated from the semantics of integers. Integer operations can be applied to chars and vice
versa. The external domains character and 1-byte integer are thus merged internally in C.
An integer of any of the three lengths may have a sign or not; unsigned objects are created by
including the type modifier unsigned in the declaration. Taken together then, there are six integer
domains with eight type names.
The unsigned and signed types form distinguishable but not independent domains. In some
cases different code will be generated for values of the two varieties. In particular, the methods for
promoting the length of signed and unsigned integers are different (details are given below). On the
other hand, the machine instruction for integer + will be used by C to translate x + y whether
11

In K&R C, the lengths of short and long integers were not fixed. The only rule was that shorts could not be longer
than integers and longs could not be shorter.

15.6. CONVERSIONS AND CASTS IN COMMON LANGUAGES

473

these numbers are signed or unsigned. This will normally carry out the programmers intent. 12 A
general rule for the safe use of these types is to use signed integers for numeric computation and
unsigned for everything else (especially bit masks and addresses). Because unsigned types directly
reflect the bit-string nature of computer storage, we considered them to be the basic domain and
assert that signed values are mapped onto the unsigned. Thus we consider an int-to-unsigned cast
to be a demotion, and unsigned-to-int a promotion.
Floating-point numbers of two lengths are implemented: float (usually 4 bytes) and double
(usually 8 bytes).
Truth values exist and are mapped onto the integers. They have no separate type name but are
created by comparison operators and used by conditionals, as in any language. FALSE is represented
by 0, TRUE by any nonzero integer. A 1 is generated when the system must create a TRUE value.
Truth values do not form a distinguishable internal domain, as their semantics are not differentiated
from the semantics of the underlying type, int.
Casts in C. C does not distinguish between conversions and casts, although neither the processes
nor their semantics are similar. Both are called casts in reference books, which are likely to
explain that a cast performs a type conversion. All conversions and casts may be explicitly
invoked using the same syntax; the name of the target type is written in parentheses before the
name or expression denoting the value that is to be changed [Exhibit 15.33].
Size Conversions in C. Size promotions may be done explicitly and are also done automatically
when operands of mixed lengths are combined. Also, short integers (int and unsigned) are promoted
to integer, and floats are promoted to double 13 when they are passed as parameters to a function.
Within the function, parameters may be declared using the promoted or the original, nonpromoted,
type. When the latter is done, the argument will be automatically demoted again. Characters are
promoted and demoted at the convenience of the translator, whenever they are manipulated.
An unsigned value (integer or character) is promoted by padding it with 0 bits. A signed integer
is promoted by sign-extension; that is, by padding the high-order end with copies of the high-order
bit of the value. This retains the sign of the object and its absolute value if it is an integer. It is
thus the semantically correct way to promote a signed integer. Signed characters may be promoted
by either method, depending on the translator. Values are demoted by truncating the high-order
end, an inherently risky operation.
Encoding Conversions in C. Two encoding conversions are implemented to convert floatingpoint numbers to integers and vice versa. These may be invoked automatically, or explicitly using
the casting syntax in Exhibit 15.33. The conservative conversion, from integer to floating point,
12

A nice aspect of twos complement encoding for negative numbers is that no special provision needs to be made
for the sign during an addition operation.
13
Non-ANSI only.

474

CHAPTER 15. THE SEMANTICS OF TYPES

Exhibit 15.33. Syntax and semantics for casts and conversions in C.

int i;
unsigned u;
char c;
float f, g;
(int)u

(unsigned)i

(int)c
(float)i
(int)f
(int)(f+g)
i = (int)f+g

Promote an unsigned to an int, no change in bit pattern. The result

is undefined if the unsigned value is larger than the maximum int
value. However, the operation will almost certainly be carried out
by doing nothing! The high-order bit will simply be reinterpreted as
a sign.
Demote an int to an unsigned, no change in bit pattern. The result is
undefined for negative integers. However, the bits of a negative number will probably be unchanged, and the sign bit will be reinterpreted
as a large positive power of 2.
A size promotion from one byte to int (2 or 4 bytes).
Convert from integer encoding to a floating point.
Convert encoding from float to integer by truncation.
Convert the sum from float to integer by truncation.
Convert the float f to an integer by truncation, then convert it back
to a float in order to add it to g. The result is a float; convert it to
an int by truncation and store it in i.

is applied automatically when mixed type operands are used in an expression. This conversion
might be automatically applied in combination with promoting the length of one of the operands.
Coercion in C. One of the anomalies of C syntax can be best understood in terms of reference
coercion. A pointer may be set to point at a variable by assigning the address of the variable to
the pointer. One indicates that the address, not the value, of the variable is to be used by writing
an & before the variable name. This inhibits the reference coercion that normally would have
been applied to the variable. Pointers may also point to arrays and functions, but in these cases
no reference coercion ever happens, and the & must not be written [Exhibit 15.34].
Pointers, explicit dereferencing, reference coercion, and inhibited coercion are all combined in
Exhibit 15.35. Arithmetic operators in C operate on numeric values, not on addresses. A commonly
useful complex data structure involves a pointer to an array of pointers which in turn index an
array of data objects. Here we declare such a structure and show some code that uses the structure.
Note that in C, adding two values is a legal operation but adding two addresses is not. Adding an
integer to a pointer which represents an array index is also legal.

15.6. CONVERSIONS AND CASTS IN COMMON LANGUAGES

475

Exhibit 15.34. Coercion complications in C.

Here we declare a variety of variables and pointers. The C syntax for setting a pointer to point at
an object is shown on the fifth line. The & operator means address of. Note that we do not
use & on the last two lines.
int k, *kk, *aa;
Declare an integer and two pointers to integers.
int a[5];
An array of 5 integers.
int f();
This declares a function f (to be defined elsewhere) that returns an
integer.
int *ff();
A pointer to a function such as f.
kk = &k;
aa = &a[0];
aa = a;
ff = f;

Pointer kk is set to point at k, an integer variable.

Pointer aa points at the first element of a, an integer.
This also makes aa point at the beginning of array a, because the
name of an array is coerced to mean its first element.
Set ff to point at the beginning of function f.

Comments on the right in Exhibit 15.35 document the dereferences; each assignment statement
is echoed, with code letters replacing the variables. The code E marks each explicit dereference,
I marks each inhibited dereference, N marks contexts in which no dereference takes place, and
C marks each variable where C applies reference coercion. Where multiple dereferences happen,
a list of code letters is used; these should be read left-to-right.
Note that most ordinary expressions are marked by C. For example, in the expression (index1
< index2), both pointer operands represent storage objects and both are coerced to pointer values
which are then compared.
The C language does not provide for handling array values coherently, nor for handling pure
values of function types. When the programmer writes an assignment involving an array or a function, the translator makes sense of the request by inhibiting the automatic coercion (dereference)
that normally would have taken place on the right side of an assignment statement. This is a
double negative situation: automatic inhibition of an automatic coercion results in no action at all.
The result is that the address originally given is assigned to the pointer.
The expression on the last line of Exhibit 15.35 is daunting; its meaning can best be ascertained
by making careful diagrams. This illustrates the complexity inherent in the C approach to references
and coercion!

15.6.6

Ada Types and Treatment of Coercion

Basic Types in the Standard Type Definition Package. As with C, a profusion of type
definitions is included in the standard Ada package. These implement four external domains: truth
value, character, integer, and real. The numeric domains are both implemented by several primitive
types.

476

CHAPTER 15. THE SEMANTICS OF TYPES

Exhibit 15.35. Explicit dereference and reference coercion in C.

int
int
int
int
int

j, k, m;
value_array[100];
*scanner, *temp;
*index_array[10];
**index1, **index2;

/*
/*
/*
/*
/*

Integers. */
An array of integers. */
Pointers to integers. */
An array of pointers to integers.
Pointers to pointers to integers.

/* Initialize three pointers. */

scanner=value_array;
/* N = N */
index1=index_array,
/* N = N */
index2=&index_array[10];
/* N = I */
/* Make elements of the index array point at every tenth value. */
while (index1 < index2)
/* C < C */
/* Compare the values of the pointers, not the values they point at!
{
index1++; scanner+=10;
/* N = C + 1 ; N = C + 10 */
*index1 = scanner;
/* E = C */
}
/* Input two integers between 0 and 9; set pointers to those slots.
scanf("%d%d", &j, &k);
/* I, I */
index1=&index_array[j];
/* Set a pointer to jth index. */
index2=&index_array[k];
/* N = I */
/* Swap two pointers in index_array. */
temp=*index1;
/* N = E,C */
*index1=*index2;
/* E = E,C */
*index2=temp;
/* E = C /*
/* Add an int value to 3 times another int. */
m= **index1 + *((*index2)+j)*3; /* N = E, E, C + E((E, C) + C) N */

*/
*/

*/

*/

15.6. CONVERSIONS AND CASTS IN COMMON LANGUAGES

477

Exhibit 15.36. Ada integer type specification.

Here are two type declarations that will map two new domains onto some primitive integer type
with an appropriate length.
type line_count
type fathom

is range 0..66;
is range -5_000 ..

0;

If we declare an object of one of these constrained types, we can be sure of the range of values
that can ever be stored it it. The variable line_no can hold only values between 0 and 66.
line_no:

line_count;

In contrast, these declarations for objects of the primitive integer types define variables whose
range of values can vary from one implementation to another.
count:
population:
index:

integer;
long_integer;
short_integer;

All implementations support type integer, whose length, as in C, may vary from implementation to implementation. Types natural (like unsigned, range 0..max_representable), positive
(like natural but excludes 0), long_integer, and short_integer might also be implemented.
If the programmer simply specifies the range of values that she or he intends to store in a variable,
the translator will choose an appropriate size for the variable. Note the similarity to COBOL. The
practical reason for using this facility is illustrated by Exhibit 15.36. The range of values that can
be stored in the variable line_no is fixed and does not depend on the implementation. In contrast,
the ranges of the other three variables depend on the translator. Using the type declaration with a
range clause, therefore, produces more portable code and is considered better programming style.
An attempt to store an out-of-range number, such as 85, into line_no would cause the run-time
error CONSTRAINT_ERROR (called an exception in Ada).
As with the integer types, reals are implemented by a group of types whose lengths depend on
the implementation. These come in two varieties, floating-point types (like reals in Pascal and floats
in C) and fixed-point types (as in COBOL). The primitive type names are float, long_float, and
short_float [Exhibit 15.37].
Truth values are supported: the Ada type Boolean is defined as the enumerated type (FALSE,
TRUE). No physical or semantic relationship exists between Boolean and any other type and, therefore, no conversions or casts are defined for Booleans.
Characters, specifically the ASCII characters, are the other predefined enumerated type. The
mapping between the ASCII character sequence and the integers 0..127 is defined, but no functions
are defined to cast characters to integers or vice versa.

478

CHAPTER 15. THE SEMANTICS OF TYPES

Exhibit 15.37. Ada real type specification.

Use of the keyword digits indicates that the new type is to be a floating-point type; delta
indicates fixed point. The new type will be mapped onto a primitive real type with an appropriate
length and encoding. The precision of a 4-byte floating-point value is about 7 digits; 8-byte floats
provide about 16 digits of precision.
Floating-Point Declarations.
type mass
is digits 15;
-- Needs double length.
type pressure is digits 7
range 0.0 .. 25.0; -- Single length ok.
For fixed-point numbers, the declaration completely determines the minimum size of the representation, as does a COBOL PICTURE clause. The delta clause specifies the required number of
decimal places of accuracy, and the range clause shows the number of places needed to the left of
the decimal.
Fixed-Point Declarations:
type dollars is delta 0.01 range 0.0 .. 10_000.0;
type voltage is delta 0.1
range -12.0 .. 24.0;
The variable dollars can take on values between 0 and 10,000. These values will be binary
approximations to these values 0.01, 0.02, etc. Successive elements in the approximation must not
differ by more than .01.

Ada Conversions and Casts Ada is the only language we have considered here whose designers
made careful distinctions among the various ways an object of one type may be changed into an
object of another type.
Ada provides type casts that move between a programmer-defined mapped type and its underlying type. These are for the purpose of defining the semantics of the mapped domain and are
never applied automatically. No casts are defined that move between predefined mapped types and
their representations.
Explicit conversions are not needed between different size objects of the same encoding. The
safe conservative size promotions are used freely and automatically, without any involvement
on the part of the programmer. Indeed, when using the recommended programming style the
programmer may not even be aware which size of a primitive type is used to implement his or her
objects. Ada will promote the size of the smaller object automatically when objects of different
sizes are mixed in an expression.
Representation conversion functions are defined for all numeric types to convert among the
three basic numeric encodings, integer, fixed point, and floating point. A conversion is invoked by
writing the name of the target type like a function name, and writing the value to be converted as
the argument of the function. Example: integer(123.9).

15.7. EVADING THE TYPE MATCHING RULES

479

Exhibit 15.38. Mixed type addition in Ada.

We extend the operator + to operate on one integer and one float by supplying an additional
computation method for +.
function "+" (x: integer; y:float) return float is
begin
return (float(x) + y)
end "+";

These representation conversions are available for explicit programmer use, but they are never
used for automatic coercion. This greatly simplifies the problem of maintaining protected semantic
domains. If conversions and casts are not applied automatically, the programmers semantic intent
can not accidentally be violated. The cost of this simplicity is inconvenience to the programmer,
since mixed-type arithmetic is, therefore, not predefined. The Ada programmer must either use
explicit type casts to do mixed-type arithmetic, or explicitly define each operator for each desired
combination of mismatched argument types [Exhibit 15.38]. At best, this is a nuisance; at worst,
it causes time and space inefficiency to translate and store the code for these simple and repetitive
function bodies.
Non-Extensible Domain Relationships
We have seen that many languages have predefined domain relationships and will use these relationships when they invoke automatic conversions. PL/1 is the most permissive of the group, as it
will automatically convert anything to anything. Ada is the most discerning of the group, and it
will invoke only casts and length conversions automatically.
None of these languages permit programmers to limit or control application of a built-in relationship, and none permit them to define domain relationships of their own that will be automatically
invoked. For example, a programmer could define a new representation for numbers and write a
type conversion routine to convert integers to the new representation. But none of these translators
will use the conversion automatically, the way the integer-to-real conversion is used.
Chapter 17 explores more modern and sophisticated type systems that permit the programmer
to define more kinds of domain relationships.

15.7

Evading the Type Matching Rules

Type checking is an immense aid to the programmer. The more strict the type rules are, the
fewer stupid mistakes and oversights will go undetected. The cost of this protection, though,
is inflexibility. Occasionally a programmer needs to perform an operation that breaks the rules.

480

CHAPTER 15. THE SEMANTICS OF TYPES

Applications that can use such flexibility include conversion of integer to floating point and efficient
computation of a hash index.
Users do not normally write integer-to-floating-point conversion routines because compilers
normally supply them as primitive functions. A systems programmer, though, needs a language
that will permit him or her to deal with hardware-dependent number representations. The bits
of an integer value must be tested, shifted, and masked during the conversion process. The value
starts out as an integer, ends up as a float, and is nothing recognizable in between. A systems
programming language needs the flexibility to refer to this object as both an integer and a float.
Let us say that the number being converted must have a dual type.
Another example of the use of dual-typed objects is hashing. The intent of a hash function is to
generate a random-looking but repeatable integer within a specified range. Think of the argument
to the hash function as a bit pattern. Hashing should scramble these bits so that the various inputs
generate uniformly distributed integers as outputs. An easy way to do this is to apply an operation
that is normally meaningless for the data and take the resulting scrambled bits. For example,
adding the right and left halves of a character string using integer addition would randomize its
bits in a repeatable, efficient, and possibly useful way.
Languages that are ST#3 (all objects are typed, types do not overlap, and all function calls are
type checked) do not let a program do such operations in a straightforward way. Many, however,
provide declaration forms that function as an escape hatch by which the programmer can get
around the normal typing restrictions. By using such an escape hatch, a programmer declares that a
normally meaningless operation is meaningful in a program. These declarations bind multiple names
(with different types) to a single storage object or to part of an object. This gives the programmer
a way to circumvent restrictions imposed by the type checking system of the translator.
When doing this kind of operation, it is up to the programmer to make sure that the result is
semantically valid. Further, any time a program depends on a particular underlying representation,
and exploits that representation in a computation, it becomes nonportable code. Different compilers
and different machines use different representations. An obvious example is the order of two bytes
in the representation of a short integer. On an IBM PC, they will be arranged low-order, highorder. But on a Macintosh, they will be high-order, low-order. Any code that depends on the order
of these bytes is nonportable.
Examples are given here in several languages of declarations that can bind dual types to a single
object.
FORTRAN. The EQUIVALENCE declaration is an entirely unrestricted way to map one object onto
another. No restrictions are placed on the relative sizes of these objects, or on the relative positions
of their beginning bytes. EQUIVALENCE is used in a very restricted way in Exhibit 15.39 to associate
two types with one variable. Lines (a) and (b) of Exhibit 15.39 declare four names: a real, an
integer, a one-dimensional integer array, and a two-dimensional integer array. In line (c) we use
EQUIVALENCE to map the real variable Z and the integer variable INTZ onto the same storage
location.

15.7. EVADING THE TYPE MATCHING RULES

481

Exhibit 15.39. EQUIVALENCE in FORTRAN.

EQUIVALENCE may be used to bind an additional name and type to any variable or any part of a
subscripted variable. (The letters on the right relate each declaration to its explanation.)
REAL Z
(a)
INTEGER INTZ, M1(100), M2(4,25)
(b)
EQUIVALENCE (INTZ, Z)
(c)
EQUIVALENCE (M1(1), M2(1,1))
(d)
The following pairs of lines refer to the same storage locations:
X = Z
Copy the dual-type value into a real variable.
K = INTZ - 1
Interpret the dual-type value as an integer and subtract 1.
K = M1(29)
Copy an array item into K.
K = M2(1,8)
Does the same thing as the line above.
Note that the subscript in the last line is (1,8) rather than (2,4) because FORTRAN arrays are
stored in column-major order, not row-major order like most other languages.

Line (d), similarly, maps the two arrays onto the same space. These two arrays have a different
number of dimensions but an equal total number of elements. We now have two names for the
same array, allowing us to refer to it using either linear or two-dimensional subscripts.
COBOL. The REDEFINES declaration is like EQUIVALENCE except that the original object and the
redefined type are required to be the same size. The redefinition must immediately follow the
original declaration.
Pascal. In FORTRAN, we declare two variables, then say, as an afterthought, that they are one
and the same object. This serves the purpose of attaching two types to an object but certainly
does not implement the abstraction dual type. In Pascal, we can explicitly declare a type with
two meanings, then use it to declare variables.
This is done with a nondiscriminated variant record. The common portion of the variant record
and the tag field are omitted altogether, producing a record that consists entirely of two or more
variant parts. Outside of the context of Pascal, this kind of dual type is called a free union type.
It is impossible to tell, at run time, what the semantics of a non-discriminated variant object is
supposed to be. A program may use an object as first one type, then another, without using any
conversions or casts.
The examples used above for FORTRAN EQUIVALENCE are rewritten in Exhibit 15.40 in Pascal as
nondiscriminated variant records. Analogous type declarations are given for an integer-real variable
(line b) and an array that can be accessed using either one or two subscripts (line c). The Pascal
variant record syntax requires declaration of an enumerated type (line a) with the correct number
of variants.

482

CHAPTER 15. THE SEMANTICS OF TYPES

Exhibit 15.40. Variant records in Pascal.

TYPE TwoVariants = 1..2;
IntReal = RECORD CASE TwoVariants OF
1: (IntName: integer);
2: (RealName: real );
END;
OneDim =
TwoDim =
DualDim =

ARRAY [1..100] OF integer;

ARRAY [1..4, 1..25] OF integer;
RECORD CASE TwoVariants OF
1: ( vector: OneDim );
2: ( matrix: TwoDim );
END;

{a}
{b}

{c}

Exhibit 15.41 shows examples of the use of these records to implement dual-type objects. We
declare an IntReal variable, R, and a DualDim array, A. When R is used in a context that requires
a real number, we refer to R.RealName. In an integer context we write R.IntName. Similarly,
A.vector[3] and A.matrix[1,3] refer to the same location.
C Has Two Escape Routes. In C, the union type constructor builds a free union type, like the
nondiscriminated variant record in Pascal. The syntax for the union is similar to but simpler than
the Pascal syntax for variant records. Exhibit 15.42 shows how the union type constructor can be
used in C to accomplish the same goal as the FORTRAN code in Exhibit 15.39 and the Pascal code
in Exhibit 15.41. Type declarations are given (line 1) for an integer-real variable and (line 4) for
an array that can be accessed using either one or two subscripts.

Exhibit 15.41. Use of a Pascal nondiscriminated variant record.

VAR R:
A:
X:
K:

IntReal;
DualDim;
Real;
Integer;

X = R.RealName
K = R.IntName - 1

Copy the dual-type value into a real variable.

Reinterpret dual-type value as integer; subtract 1.

K = A.vector[29]
K = A.matrix[2,4]

Copy the second array item into K.

Does the same thing as the line above.

15.7. EVADING THE TYPE MATCHING RULES

483

Exhibit 15.42. Union data type in C.

typedef union {

long int_name;
float real_name;} int_real;

int_real r;

/* 1 */
/* 2 */

typedef int one_dim [100];

typedef int two_dim [25][4];
typedef union { one_dim vector;
two_dim matrix;} dual_dim ;
dual_dim a;

/* 3 */
/* 4 */

In line 1, type int_real is defined to be either a long integer or a floating-point number.

Enough storage will be allocated to store whichever variant is longer. (They are often the
same length.)
Line 2 creates an object, r, of type int_real. To access it as a long integer we use the name
r.int_name. To access it as a float we use the name r.real_name.
Line 3 defines a type named dual_dim that can be either a one- or two-dimensional array. In
either case, it has 100 integer elements.
Line 4 declares an object, a of type dual_dim. We may access this array with either one or
two subscripts, thus: a.vector[k] or a.matrix[m][n].
Interestingly, the union type constructor is not the only way, or the easiest way, to create a
dual-type object in C. This same end can be achieved by making two pointers (an integer pointer
and a float pointer) point at the same object [Exhibit 15.43]. When the program needs an int, it
can access the object through the integer pointer. To get a float, it can use the float pointer. 14
This is demonstrated in Exhibit 15.43. Here we allocate one storage object, named k, and make
two pointers, plong and pfloat, point at it. Both pointers are initialized to point at k.
Any pointer in C can be cast to any other pointer type. This causes no change except a relabeling
of the domain of the pointer. Specifically, a pointer cast does not convert the representation of the
object to which the pointer points. 15
Ada Closes the Loopholes. The designers of Ada put a high value on designing a semantically
safe language, that would be wholly and completely ST#3. Free union types and unrestricted
14

This trick is the basis of the object-oriented function dispatching in C++.

The explicit cast on the initialization is not necessary at all in many older C implementations, and it is used only
to avoid a warning message in newer compilers. The code would compile, run, and produce the same answer without
the cast.
15

484

CHAPTER 15. THE SEMANTICS OF TYPES

Exhibit 15.43. Using pointers in C to attach two types to an object.

In C, long integers and floats are usually the same length, so we use these two types to create a
dual-type object, k. Then we compute a hash function by referring to k first as a float, and then as
an int. We store a float value into k using pflot, and take it out using plong. This does not cause
a representation conversion. Applying integer division to the misinterpreted value will not cause a
representation conversion either. The result is scrambled bits; it is complete nonsense to divide a
float value by an integer. Thus we have hashed the original number.
#define TABLESIZE 1000
long int j, k;
long int *plong=&k;
float *pflot = (float*)(&k);
*pflot = 3.1;
j = *plong % TABLESIZE;

/* Put a floating-point value into k. */

/* Apply mod to k to calculate a value between
zero and the size of the hash table. */

The programmer who uses this coding trick must do so carefully; any code that casts the type of
a pointer, or depends on the representations of two types to be the same length, is nonportable.

casts are ways to evade the constraints imposed by strong typing, and they are not permitted in
Ada.
Ada does support type casts, specifically, those necessary casts between types that are related
by mapping. Type conversions are also supported between pairs of numeric types such as integer
and real. These conversions must be invoked explicitly, using the same syntax as for type casts.
But, unlike C, Ada does not support unrestricted casts from any type to any type. Thus the pointer
trick will not work in Ada.
This surely increases both the readability and the portability of Ada programs. Type changes
must be explicitly stated, eliminating confusion about intent. By eliminating free union types,
access to the actual bit-level implementation of Ada objects is shut off, forcing programmers to
write implementation-independent code. The cost is that flexibility to do some useful things is lost.
The advantage is more reliable code.

Exercises
1. What is a domain?
2. How were implicit domains represented in early computer languages?
3. What is a predefined domain? Which ones were supported by the original FORTRAN? What
determined the domain of each variable?

15.7. EVADING THE TYPE MATCHING RULES

485

4. How did ALGOL enrich the domain structure of previous languages?

5. Why is typeless sometimes used as a synonym for dynamically typed language?
6. When is the domain of a variable tested in a dynamically typed language? Why is it tested?
Why is it tested at that time?
7. Can function applicability be controlled in dynamically typed languages? Explain.
8. Why can early C be considered a weakly typed language?
9. Why were domains given increasingly important roles as programming languages developed?
10. What is a subrange declaration? Why is it useful?
11. How has Ada enhanced the use of domains beyond Pascal?
12. What is type checking?
13. Briefly define strong typing, and explain how it can be, at the same time, a boon and a
burden.
14. How has the definition of strongly typed language evolved?
15. What is an ADT? Why cant we define an ADT in C or Pascal?
16. What is an external domain? Internal domain?
17. What are internally merged domains?
18. What are the problems associated with internally merged domains?
19. When a new domain has been mapped onto an old one, is the semantic intent of the two
domains identical? Why or why not?
20. Name two domains whose semantics are totally different, even though they are both sometimes
represented as integers. Give an example of two values that might be represented by the same
integer.
21. Why is Pascal considered a strongly typed language? Explain.
22. What is a type constructor? Give an example. How does a type constructor permit the
programmer to define new domains?
23. Why are some domains which have the same structural description considered incompatible
in Pascal?
24. What is the role of a typedef declaration in C?

486

CHAPTER 15. THE SEMANTICS OF TYPES

25. How does Ada improve upon the type structure found in C and Pascal?
26. What is a type cast? Conversion? Coercion? Give an example of each in a language that is
familiar to you.
27. For each of the following kinds of type conversions, say whether the original and converted
objects represent the same thing and whether the representation (bit pattern) is changed.
a. integer to real
b. The result of an addition used as a truth value
c. long integer to short integer
28. Give and explain an example of automatic type conversion destroying the semantics of an
object. The example may be drawn from any language that uses automatic conversion, but
please say what language you are using.
29. What is an encoding conversion? Size conversion? Reference conversion?
30. Name two domains that contain some elements that have the same meaning. As an example,
show a pair of values that correspond.
31. Explain the following statement: A conversion involves a run-time computation, but a cast
happens entirely at compile time.
32. A data value in one domain is changed to a corresponding value with the same (or similar)
meaning in a second domain. Is this a conversion or a cast?
33. Consider the Pascal statements given below. What semantic action is implicit in this fragment?
jj: integer;
ff: real;
begin
ff := 3.0
jj := 1;
ff := ff + jj;
end
34. Would the following code in Ada (analogous to the Pascal code in question 33) be legal? Why
or why not?
rr: float := 3.0;
jj: integer := 1;
begin
rr := rr + jj;
end

15.7. EVADING THE TYPE MATCHING RULES

487

35. What is the difference between a promotion and demotion cast?

36. Why are casts considered dangerous operations?
37. Consider these equivalent expressions, where b is an integer:
In APL: 8 > b > 3
In C or Pascal: 3 < b < 8
The Pascal expression causes a compile-time error. What is it?
38. The APL and C expressions in question 37 will compile and run without errors. Evaluate the
expression in either language using the value b = 2. What semantic distortion is committed
by the translator that causes the result to be different from the programmers apparent
semantic intent? Why does it not produce a type error in these languages?
39. APL objects retain their type tags on the stack, and data types are checked by all primitive
operators (unlike FORTH). Nonetheless, because statement labels are mapped onto integers,
APL will sometimes execute code that is semantic nonsense and would be identified as a type
error in other languages. Give an example of code that is legal in APL but that computes
nonsense. Explain briefly what semantic error is being made. (Note: The fact that identifiers
do not have types is not a semantic error.)
40. Explain what the semantic problem can be when a language applies two-way automatic
conversion between a user-defined data type and the primitive data type used to represent it.
41. Explain or give an example of one way that C will permit a programmer to make a semantically
absurd calculation.
42. Describe or give an example of a situation (other than that described in answer to questions
37 and 38) that would cause a compile-time error in Pascal but that would be legal at compile
time in C.
43. Describe or give an example of a situation that would cause a run-time error comment in
Pascal but that would not be considered to be an error in C.
44. Give an example of an error that Pascal cannot identify at compile time but will identify at
run time. Why cant this error be identified at compile time?
45. What is meant by a conservative conversion?
46. When does a translator attempt to coerce an argument?
47. Why is size coercion essential in the use and efficiency of computer languages?
48. When are encoding coercions used?

488

CHAPTER 15. THE SEMANTICS OF TYPES

49. In what sense is automatic dereferencing a downward coercion?

50. Why is it sometimes necessary to evade type compatibility rules?
51. Give an example of a loophole in a strongly typed language.
52. How have Ada designers created a semantically safe language?
53. Because FORTH objects lose their types when they are put on the stack, FORTH permits the
programmer to violate the semantic intent of these objects, either by accident or on purpose.
Explain and give an example of such a violation.
54. Happy Hacker is using a version of FORTH in which integers occupy 4 bytes. He has extended
the semantics of the language to include a new type declarator, REALVAR, for real (floatingpoint encoding) variables, each taking 4 bytes. Happy wants code to compute the same
function as in the Pascal code in question 33, and has written this:
1 VARIABLE JJ
3 REALVAR FF
FF @ JJ @ + FF !
Would this code work? If so, what answer would be left in FF? If not, why not?

Chapter 16

Modules and Object Classes

Overview
For clarity and easy debugging, a small program should be written as a collection of
functions. However, use of subprograms alone is inadequate for a large, complex system.
Encapsulated modules were developed as tools for managing team efforts and complex
tasks. An encapsulated module is bigger than a function definition, yet smaller than a
program. These modules facilitate the creation of code libraries, separate compilation,
the grouping of logically related elements for the implementation of abstract data types,
the sharing of public data, and the protection of private information.
Separately compiled C files, Ada packages, and object classes in C++ are three strategies for achieving modularity. Before module-related facilities were included within
programming languages, files and separate compilation were used to achieve the goals
of modularity. In a modularized C program, an object or function in one file can be
shared by a program in another file by using an extern specification, and kept private through the use of the keyword static. The operating systems linking loader
completed the connections among the programs parts. An automated method, the
makefile, was invented to make linking UNIX applications easier and more foolproof.
Packages in Ada provide the framework to group the variables and functions of an
abstract data type. A package consists of a header which declares shared data and a
body which defines private symbols and the ADT functions. Unlike C, Adas goals of
modularity are achieved from within the language.
In C++, the class serves a similar role as the package in Ada. Classes contain both
functions and data. C++ also allows classes to be constructed in a hierarchical fashion
so that they are related to each other and can inherit data and functions. Within a
489

490

CHAPTER 16. MODULES AND OBJECT CLASSES

Exhibit 16.1. Terms for modules.

C
Ada
Modula
CLU
C++, Simula, Smalltalk

a separately compiled file

package
module
cluster
class

class, the definition of a function is called a method. The C++ class is a template
that must be instantiated by the use of the class name in a declaration. The result is
an object whose fields will be initialized if the class contains an initialization function
called a constructor.
Some relationships among classes are difficult to achieve using only public and private
parts. There are times when we would like parts shared with some modules but not all.
C++ provides a solution for this problem: a class or a function may be declared to be
a friend of another class. Access is then shared by the parent class and the friend class.

16.1

The Purpose of Modules

It is universally accepted by computer scientists that programs should be designed and implemented
in a modular fashion. For small programs, this merely means that a program is written as a
collection of function calls. Each function definition must be short, its purpose clearly defined, and
its interface clearly declared and documented. Global variables are avoided in order to minimize
unwanted interaction between functions and side effects.
When writing large systems programs, though, this basic modular methodology is helpful but
inadequate. The complexity of a large system is so great that it must be organized as something
other than a list of thousands of function definitions. Another level of program structure is needed so
that things that belong together can be grouped and isolated from all possible outside interference.
The semantic bases of many modern languages have been extended to support a unit that is larger
than a single function but smaller than a program. Such units are called by varied names [Exhibit
16.1] but have similar purposes and properties, which we examine here. Sample modules are given
in Ada and in C++.
Encapsulated modules were developed in response to several needs that have been mentioned
earlier. These are:
To facilitate separate compilation and creation of code libraries.

16.1. THE PURPOSE OF MODULES

491

To enable the logical grouping of type declarations and relevant representation-dependent

functions.
To support implementation-independent abstract data types.
To implement private information (static local data).
To implement controlled, nonhierarchical sharing of data.
Grouping Logically Related Elements. A module provides a framework for grouping together
a set of related functions and is thus an ideal way to implement an ADT. Within the module, the
details of the implementation and functions are defined that access specific fields in the representation. The module must include all representation-dependent functions that are needed to use
the ADT domains. These can be written, compiled, and tested in relative isolation. The ADT
functions often call each other and might use information that is private within the module. Their
interdependence is documented by their inclusion in the same module. The array_stack, defined
in Exhibit 15.9, is a good example; in Section 16.3 we give an implementation of this ADT in Ada.
Sharing Public Facilities. To be useful, a module must have some functions or data objects
that are public and can be accessed by other parts of the program. These symbols are made
known to other modules, on demand. If a program, P , wishes to use a module, M , then P must
contain a command to include M . Then within P , the public symbols in M can be referenced.
Protecting Private Information. An encapsulated module contains declarations for types,
functions, and data objects. These declarations are separated into two groups, of public and private
symbols. Public symbols may be accessed by other modules and form the interface between a
module and the rest of the world. In an ADT implementation, the ADT functions would be
declared to be public symbols, so that other parts of a program could call them. In addition, the
module might contain private functions which are called by the ADT functions but cannot be called
from outside the module.
The array_stack ADT is particularly simple and does not require any locally defined types.
To implement more complex ADTs, however, we must be able to make private type declarations.
Having the possibility of internally defined types also enables the representation of the ADT to
be changed easily, simply by changing the private type declaration. The effects of that change
are limited to the scope of the module definition. Once the modules functions are redefined to
work on the new representation, all code that depends on those functions would automatically
work. For example, suppose a module were used to define a stack, and the representation chosen
initially was an array. Later, it became clear that a linked list representation would fit the data
characteristics better. To make this change, the data type of the stack storage and definitions of
the stack functions (all within the module) would have to be changed. However, when the outer
program called push or pop, the call would function as expected and store or return a value. The
caller would never know that the representation of the ADT had been changed.

492

CHAPTER 16. MODULES AND OBJECT CLASSES

A module must provide for private data. Many large systems have modules that operate as
coroutines, acting asynchronously to carry out a set of related concurrent tasks. For example, an
application that uses dynamic storage allocation will often have a separate memory management
module which gets storage from the operating system in efficient quantities and manages lists of
new and freed cells. I/O buffering and caching are done by coroutines. To support coroutines we
need private data objects with a static lifetime that can be shared among all the functions in the
module. These must be created and initialized when the module is first entered.
Often, a group of private data declarations can be used in a module in place of a single object
of a record type. In Pascal, for example, a stack is usually implemented as a record consisting of
an array of data slots and an integer top-of-stack subscript. We declare a record data type with
these parts so that a stack can be passed around as one coherent argument. However, when we
implement a stack module, the stack-data-object is internal to the module, and within the module
it is global to all the modules functions. It is not passed anywhere as a parameter. In such cases,
there is no real advantage to gathering the two parts of the stack into one record.
The data parts of the ADT would be declared in the private part of the module. Private symbols
can be used only within the module; we say they are hidden within it. In an ordinary function,
the local variables are similarly hidden. However, local variables cannot do all that is needed.
By concealing the data-parts of the ADT, we force the programmer to access them only through
the ADT functions. The nature, type, extent, and value of all private data is concealed from the
program outside the module. This organization forces the programmer to use an implementationindependent coding style, making the entire package easier to debug and modify.
Implementation Strategies.
that are in common use:

We will examine three basic strategies for achieving modularity

1. Cs separately compiled files.

2. Ada packages.
3. Object classes in C++.

16.2

Modularity Through Files and Linking

Before module-related facilities were included within programming languages, files and separate
compilation were used to achieve the goals of modularity. This was not an altogether satisfactory
solution, since it relied on the operating system environment, specifically the linker, to complete
the connections among the parts of the program. The structure of the application as a whole was
not expressed anywhere within the program code! Instead, the user determined this structure by
telling the system linker which modules to include.
Giving complex commands to a linker is error prone and cumbersome, so an automated method,
the makefile, was invented for linking UNIX applications. Currently, several systems support some
kind of make facility. The example we give and explain here is written in ANSI C and UNIX.

493

16.2. MODULARITY THROUGH FILES AND LINKING

Exhibit 16.2. Header file: modules.h.

/* Set this constant to the number of data items in your array.
#define LEN 4

*/

/* Set these definitions to the desired data type and format specifier.
/* NUMBER must be defined to be a standard numeric type. */
#define D "%ld"
typedef long int NUMBER;

*/

extern NUMBER reduce(NUMBER (*)(), NUMBER)

A modularized program in C consists of four or more files. We list their purposes here, then
examine an example of each written in C. The files are as follows:
A makefile, which represents the entire program. It lists the components and describes how
each depends on the others. The contents of this file are not C code. Rather, they are
operating system commands, interpreted by the systems make facility. When you make
a program, any necessary compilation and linking is done, according to the instructions given
in the makefile. The result is a machine-code module that is ready to load and run.
A header file, containing declarations and definitions of symbols that must be shared by two
or more code modules in order to attain consistency.
A main module, containing the main program, where execution begins. Any declarations in
the main program are local to it and are not shared by other modules.
One or more modules containing functions that are called by main. These share the header
file with main, but are compiled separately and may contain private declarations of any sort.
These modules often are used for coroutines that handle buffering and storage management.
Library Packages
We present a short interactive C application consisting of four files to show how the parts of an
application are written and combined. The header file [Exhibit 16.2] must be included in the other
files when they are compiled; note the #include statements at the top of Exhibits 16.3 and 16.4.
A header file provides a way to coordinate the assumptions made in the code modules. In this
example, the code modules both need to know the length and base type of the arrays that will be
processed. Using the particular constant definitions in Exhibit 16.2, we would work with arrays of
four long integers and print the answers out in a %ld (long decimal) format field. However, we
could change these three declarations to handle any other length and base type. For example, to
work with arrays of ten floating-point numbers, we would edit the header file to say:

494

CHAPTER 16. MODULES AND OBJECT CLASSES

Exhibit 16.3. Main module: sumup.c.

#include <stdio.h>
#include "modules.h"
/* This constant is the number of array operations that are defined.
#define OPS 6
NUMBER
NUMBER
NUMBER
NUMBER
NUMBER
NUMBER

fi_plus(a, b)
fi_times (a, b)
fl_or(a, b)
fl_and(a, b)
fb_or(a, b)
fb_and(a, b)

NUMBER
NUMBER
NUMBER
NUMBER
NUMBER
NUMBER

a,
a,
a,
a,
a,
a,

b;
b;
b;
b;
b;
b;

{return
{return
{return
{return
{return
{return

a
a
a
a
a
a

*/

+ b;}
* b;}
|| b;}
&& b;}
| b;}
& b;}

static NUMBER (*op_ar[OPS])() = {fi_plus,fl_or,fb_or,fi_times,fl_and,fb_and};

static char *label[OPS] = {"plus ", "logical or ", "bitwise or ",
"times ", "logical and", "bitwise and"};
main()
{ /* Header file must define NUMBER
NUMBER ar[LEN];
NUMBER *ar_last = &ar[LEN-1];
NUMBER *p;
int k;

to
/*
/*
/*
/*

be a standard numeric type. */

Declare an array of numbers. */
Mark the last slot of the array. */
A scanning pointer for the array. */
An index for the function array. */

puts("This program demonstrates some whole-array operations.\n\n");

do {
printf ("Please enter %d integers separated by spaces.\n", LEN);
p=ar;
/* Start input at head of array. */
while( p<=ar_last && scanf(D, p)==1 ) ++p;
if (p <= ar_last)
printf("Premature EOF or conversion error; job terminated.\n"),
exit(1);
while (getchar()!=\n);
/* Flush rest of input line. */
putchar (\n);
/* Prepare for output. */
/* Apply each operation in the op array to the number array.
for(k=0; k<OPS; k++)
printf ("%s "D"\n", label[k], reduce(op_ar[k], ar));
printf("\nDo you want to enter more data?
}
while (getchar() == y);
}

(y/n)\n");

*/

495

16.2. MODULARITY THROUGH FILES AND LINKING

Exhibit 16.4. Subroutine module: reduce.c.

#include "modules.h"

/* a */

NUMBER reduce(NUMBER (*op)(), NUMBER ar)

{
int k;
NUMBER sum;

/* b */

sum = ar[0];
for(k=1; k<LEN; k++) sum = (*op)(sum, ar[k]);
return sum;

/* c */
/* d */

#define LEN 10
#define D "%f"
typedef float NUMBER;
Note the extern declaration for the function reduce. When this line is included in the main
module, it tells the compiler two things:
The reduce function is defined in another module.
It takes two arguments, a pointer to a function and a number, and returns a number.
This information permits the compiler to compile correct and meaningful calls to a function it has
never seen.
/subsubsectionKeeping Private Information
In a modularized C program, a programmer can create both shared functions and private
functions by judicial use of header files and static declarations. The modifier static is the
opposite of extern. The keyword extern is used to denote an object or function that is to be
shared by other modules, while static denotes a private item. A global symbol (function or data
object) that is not declared to be either one is extern.
The main module, shown in Exhibit 16.3 is intended only as a demonstration of how the parts
of a modularized program work together. 1 It sets up an array of integers, then sends the array to
reduce to be processed, in turn, by each of six dyadic integer functions. Finally, it prints all the
answers and queries the user about more data.
Sharing Information
The subroutine module shown in Exhibit 16.4 contains the function reduce. 2 This performs a
running operation, using whatever dyadic function is passed to it as an argument. For example,
1
2

This main program is for demonstration purposesit does not do anything that is particularly useful.
Compare the syntax for functional parameters in C to the Pascal syntax shown in Exhibit 9.29.

496

CHAPTER 16. MODULES AND OBJECT CLASSES

Exhibit 16.5. The makefile for the sumup program.

sumup: main.o reduce.o
cc -o sumup main.o reduce.o
main.o: main.c modules.h
cc -o main.o main.c
reduce.o: reduce.c modules.h
cc -o reduce.o reduce.c

if the argument operation is +, reduce computes the running sum: a[0] + a[1] + a[2] + ...
+ a[n]. If the argument function is *, the answer is the running product.
Note that this module includes the header file (line a ). Thus the local variable, sum, (line c),
will be type long or short or integer or float, whichever is selected in the header file. The important
thing is that it will be the same type as the array elements created in main, and an appropriate
type for the function argument. Each time the header file is modified, both the subroutine module
and the main module, which depend on it, must be recompiled.
The type declaration NUMBER (*op)(), in the header of reduce, declares op to be a pointer
to a function that returns a NUMBER. We can refer to this function through the pointer by writing
its name with a dereference operator: (*op). Thus line d calls the function which was passed
to reduce as an argument. The arguments to that function are the current partial sum and the
current array element.
Defining the Application
The #include commands in the two code modules express the relationship among the three files,
but not in a coherent way. By reading all the files, a user could deduce the relationship, but
that relationship is not presented in a single place or in a way that is convenient to process.
Although the problem is small for a program with only two code modules and a header, it can
become considerable with a large many-module application. The UNIX makefile makes the
relationships and dependencies explicit. A makefile is a file of executable system commands; the
example shown here contains UNIX commands to compile and link a program.
The makefile is the root of a tree of files that comprise the application, and it is the basis of an
automated version-control system which serves two purposes:
The most recent version of every file involved will be used to construct the executable module.
No object module will be sent to the linker if a corresponding source or header file has been
modified after the last compilation for that module.
If any step in the compiling and linking process is unnecessary, it will be skipped. No compiling
or linking operation will ever be redone unless the files on which it depends have been modified.

16.3. PACKAGES IN ADA

497

A makefile explicitly defines how the final application depends on separately compiled object modules, and how each of these depends on source code and header files. In our example [Exhibit 16.5],
the first line tells the linker that the finished program will be called sumup, and that it can be
produced by linking two user-created object files, main.o and reduce.o. (The object files that
belong to the C library do not need to be listed.) The third and fifth lines define the source files
on which the two user-defined object modules depend.
The second line of the makefile contains a call to link the object files that comprise this application, and to produce an executable file named sumup. The fourth and sixth lines contain calls
on the compiler to produce the necessary object files. 3 These lines are invoked only when needed.
The make facility checks the date-last-modified on source files, object files, and executable files. If
a linked module already exists, and none of its component object files have been modified or need
modification, the linker can return immediately without wasting the effort to redo a job that is
already done. An object file needs to be recompiled if any of the source files on which it depends
have been changed since the creation time of the object file. These dependencies are defined by
the last two lines in the makefile, where we see that the object file main.o depends on main.c and
modules.h, and, similarly, reduce.o depends on a code module and a header module. If one of
these three files were modified, the relevant object module or modules would be recompiled.
To use a makefile, that is, to compile and link the program, the programmer simply says make,
or make sumup. (The longer form is needed if the users file directory contains makefiles for more
than one application.) Any source module that has been edited since the last call on make will
be recompiled, then the object modules will be linked, and the result stored as an executable file
named sumup. To run the compiled program, the programmer will say sumup.
Makefiles, separate compilation, and declarations for external and static symbols thus combine
to provide a language/system that achieves the goals of grouping, sharing, and protection.

16.3

Packages in Ada

Grouping Related Elements

Packages in Ada provide a framework under which to group together the variables and functions
of an ADT, uninterrupted by unrelated variables and code. A package has much better lexical
coherence than a comparable set of definitions in Pascal. An Ada package has a header, which declares all the externally visible symbols, and a body, which can include type definitions, static local
variables, definitions for the publicly accessible functions, and private local functions. We will show
how an Ada package definition can be used to implement the ADT array_stack [Exhibit 16.6]. 4
This package contains definitions for a stack representation and the representation-dependent stack
functions, PUSH, POP and TOP.
3
These two lines are not strictly necessary because they only state explicitly what the system would do by default
if no compile commands were given.
4
In Chapter 17 we show how a package can be embedded within a generic framework which allows for varying the
base type of this ADT.

498

CHAPTER 16. MODULES AND OBJECT CLASSES

Exhibit 16.6. Defining a stack package in Ada.

-- The package header declares the externally visible symbols of the ADT.
package STACK is
function PUSH(k: integer) return boolean;
function POP return integer;
function TOP return integer;
end STACK;
-- The package body defines private symbols and the ADT functions.
-- The private variables cannot be referenced outside this package.
package body STACK is
STKLEN: constant:= 10;
stk: array(1..STKLEN)of integer;
-- static local storage.
tos: integer range -1..STKLEN;
-- Negative 1 is an error code.
-- These functions can be called from outside the package because they
-- have been declared as external symbols in the package header.
function PUSH(k: integer) is
begin
if tos<STKLEN then
tos:=tos+1; stk(tos):= k;
return true;
else return false;
end if
end PUSH;
function TOP return integer is
begin
if tos >0 then return stk(tos);
else return SYSTEM.MIN_INT;
-- Use value from the package SYSTEM.
end if;
end TOP;
function POP return integer is
ans: integer;
begin
ans:= TOP; tos:=tos-1;
return ans;
end POP;

-- Remove & return top stack item.

-- This code is run at load time, and initializes the package data.
begin
tos:=0;
end STACK;

16.3. PACKAGES IN ADA

499

Sharing Public Facilities

Any symbol declared in the header of an Ada package can be shared by other modules. In our
example, this includes the functions PUSH, POP, and TOP. To use an Ada package, the programmer
writes a with command within the main program [Exhibit 16.7]. Having done this, one can refer
to a part of that module by its complete name: the module name followed by the function or data
name (see the left side of the exhibit). Using complete names avoids a possible problem caused by
using the same identifier for parts of different modules.
Because packages are used extensively in Ada programs, this syntax for naming the parts of a
package becomes annoying. Ada provides an alternate syntax that can usually be used and is more
convenient. Using the use command causes all the public names in a package to be included with
the local symbols in the calling program, eliminating the need to use the module name in every
reference (see the right side of exhibit). 5
By selectively including modules where they are needed, we can construct a nonhierarchical
sharing structure. Each part of a program can be given access to exactly those other parts it
needs. The possibility of many kinds of unintentional interactions between functions and data can
be minimized, and the resulting software becomes easier to debug and maintain.
Private Information
Any symbol declared in the package body but not listed in the header is private and can only be
accessed from within the package. In our Ada example, the stack-array and the stack-pointer are
declared as pieces of private data, and the length of the stack is a private constant [Exhibit 16.6,
lines 10 to 12]. The stack-pointer (tos) and the array (stk) are not grouped into a record-object,
as they would be in Pascal, because there is no advantage to that grouping. The two variables
are global within the package and invisible outside it, so they will never be passed anywhere as
arguments.
Private variables declared in a package are static; that is, the value computed during one
activation of a stack function remains until the next stack function is called. Package variables
are not deallocated when control leaves the package, and they are not reinitialized when control
returns. This permits the programmer to declare protected, local objects that can be used globally.
This system is a tremendous improvement over the way similar problems must be handled in Pascal.
In a Pascal program, data structures that are to be used throughout the program must be declared
globally. They are usually initialized by the main program at the beginning of program execution.
Pascal provides no way to protect the data structures from the rest of the program, or to ensure
that they are accessed only through the proper functions.
Many variables and most data structures must be initialized when they are allocated. If a data
structure is local and private within a package, the package itself must initialize it. The code on
the last three lines of this Ada package serves this purpose [Exhibit 16.6]. At load time, before
5

Much like the with command in Pascal.

500

CHAPTER 16. MODULES AND OBJECT CLASSES

Exhibit 16.7. Using an Ada package.

Basic Syntax:
with STACK;
procedure MyJob is
...
STACK.push(MyValue);

Sugared Syntax:
with STACK;
use STACK;
procedure MyJob is
...
push(MyValue);

beginning to execute the main program, storage is allocated for all private variables in all included
packages. The main initializing code in each of these packages is then executed.
Comparison to C
The goals of grouping, sharing, and privacy are achieved much the same way in an Ada package as
they are in a C module. The primary differences between the two systems is that Ada packages have
been brought entirely within the Ada language, whereas the C system relies on using a command
file that is outside the C language.

16.4

Object Classes.

In object-oriented languages, the class serves much the same roles as the package does in Ada: it
organizes things that belong together and allows for private information. In addition, the classes
can be constructed so that there are hierarchies of classes, all related to each other. 6 The bestknown object-oriented languages are Simula, (the first object-oriented language), Smalltalk, and
C++. In all of these, a class may contain data fields and functions, which we will call members of
the class. Like Ada packages, classes may have private and public members.
Terminology varies somewhat among these languages. The term method was invented for
Smalltalk,7 and it means one definition, in one class, for a function. This term is useful, and
we will use it whenever we need to distinguish between an entire function and a single definition for
that function. A function name represents a conceptual process. A class function might be (and
usually is) given different (but related) method definitions in various parts of a class hierarchy. 8
Terminology used for function definitions and calls is also variable. Simula has procedures, C++
has functions, and Smalltalk has messages. In Simula and C++, we speak of calling a procedure or
function, while in Smalltalk we send a message to an object. In this discussion, we will use the
terms function and function call.
6

See Chapter 18.

This term is not generally used in describing Simula and C++.
8
Virtual functions are covered in Chapter 18.
7

501

16.4. OBJECT CLASSES.

Exhibit 16.8. Defining a class in C++.

class char_stack {
int size;
char *tos, *end;
char* s;
public:
char_stack(int sz)
{
s=new char[size=sz];
tos=s;
end=tos+(size-1);
}

// Here are the private parts of the class.

//
//
//
//
//

char_stack() { delete s; }
int push(char c)
{
return ( tos<=end ?
char pop()
{
return ( tos>s ?

All remaining symbols are public.

This is the classs constructor.
Allocate an array of sz chars.
Initialize top to point at first char.
Mark last slot in stack.

// This object destructor is used when

// control leaves the objects scope.

// Return 0 if stack is full.

(*tos++ =c) : 0 ); }

*--tos :

// Return null char if stack is empty.

\0 ); }

char top();
};

// End of class char_stack.

char char_stack::top()
{
return ( tos>s ? *tos :

16.4.1

// Return top but dont pop.

\0 ); }

Classes in C++

For the rest of this discussion, we will focus attention primarily on the semantics and syntax of
C++, since it promises to become the most widely used object-oriented language.
A C++ class has the same elements as an Ada package, and it is declared with similar syntax. 9
A sample class, named char_stack, is defined in Exhibit 16.8. It has four private data-members
and four public function-members.
9
In C++, a struct data type is also a kind of class. It may contain function declarations, and any struct
declaration actually defines a class. However, it is a class with no private parts. The difference between a struct
variable and a class variable is the accessing restrictions on the private data and functions. We will limit further
discussion to classes declared as class.

502

CHAPTER 16. MODULES AND OBJECT CLASSES

Exhibit 16.9. Instantiating a C++ class.

main ()
{
char c;
char_stack stk1(100);
char_stack stk2(10);
...
stk3 = new char_stack(20);
stk1.push(#);
stk2.push(%);
c = stk2.pop();
}

// We can instantiate a class more than once,

// and get multiple copies of the objects.

// We must specify which "push" to use.

Instances and Naming

A C++ class is a template, like a type declaration in Pascal, and must be instantiated to create
objects. To instantiate a class, you use the class name in a declaration, as you would use a Pascal
type name. The result of instantiation is an object, which is represented by a record-type storageobject with one field for each of the class variables. Those fields will be initialized if the class
contains an initialization function.
When we instantiate a class, we bind the resulting object to a name. The objects name is used
to refer to both data fields and the function members. For example, in Exhibit 16.9, the object
stk1 is an instance of the C++ class char_stack, and it has the function-members stk1.push and
stk1.pop and the data members stk1.size, stk1.top, stk1.s, and stk1.end.
Within the context of the class, the members may be referenced by using a simple name. A
public member can be referenced outside the class using a dotted notation:
hclass_namei.hfunction_namei ( hargument_listi)
For example, the last three lines in Exhibit 16.9 call functions defined in the class char_stack.
The Implicit Argument. When we call a function that is a class member, we must specify
which instance of the function we are calling. To do this, we write an object name followed by the
member function name. (In Smalltalk, these are separated by spaces, while Simula and C++ use
dotted notation, as in Exhibit 16.9.) The data members of an object do not need to be passed
as arguments to the function members: each function may operate on the data fields of its own
instance. Thus the record of data-members is an implicit argument to all of its function members.
As a result of having an implicit argument, each class function needs one fewer explicit argument.
However, the name of the object that includes both the function instance and the implicit argument
must be specified as part of the function name. To a programmer accustomed to traditional
languages, such as Pascal, this looks like writing the name of the first argument on the left side of

16.4. OBJECT CLASSES.

503

a function name instead of on the right side with the rest of the argument list. For example, in
Pascal, the name stk1_stack would be written as part of the argument list to push or pop, but in
an object-oriented language it is written as part of the name of the function [Exhibit 16.9].
Having an implicit argument leads to one difficulty. Unlike an explicit argument, the syntax
provides no way to give a local name to this record. Yet some functions, particularly recursive
ones, must refer explicitly to the implicit argument. This problem is solved by a keyword that can
be used to refer to any implicit argument. In Smalltalk, the keyword is self. In C++, the keyword
this refers to a pointer to the implied argument.
Implementation
Classes contain both functions and data. The functions are declared within the class definition and
may be either public or private. Theoretically, a copy of each class function is made each time
the class is instantiated. The newly instantiated functions are part of the new class-object and are
the only functions that have access to its private data fields. In practice, though, there is no need
to duplicate the code of the class functions for every class instance. Thus an instance of a class can
be implemented by a record variable with one field for each class data-member. This differs from
an ordinary record-type variable because the access to the private fields is restricted to member
functions.10 Initialization is done by a member function called a constructor, which must have the
same name as the class.
Constructors and Initialization. Each class may contain a set of constructor functions. 11 A
class variable may be created in a declaration or by using the new operator, which creates a storage
object dynamically. In both cases, a constructor function is called to initialize the new object.
Like Ada (see Exhibit 14.28), C++ permits the programmer to call a constructor function in an
expression to create a pure value of the class out of its components. 12 This is often done inside
a function to create a return value of the new type, and it is also used in declarations to create
appropriate initializers.
Constructors are ordinary functions, with or without parameters. The only restriction is that
all the constructors in a class must have different types of argument lists. 13 If a constructor has
parameters, an argument list must be supplied in a declaration, a call on new, or in an explicit
call on the constructor. Exhibit 16.9 creates three instances of the class char_stack with different
arguments and binds them to the names stk1, stk2, and stk3. The class char_stack must have
a definition for its constructor which takes an integer argument. This function is freely definable
and can perform any appropriate initialization.
10
A Simula class has an associated main procedure, like that in an Ada package, which is called when a class
object is constructed to initialize it.
11
In current terminology, the name of the constructor function may be overloaded.
12
Review Chapter 14, Section 14.3.1.
13
The power of C++ constructors goes well beyond initialization, and will be further developed in Chapter 17.

504

CHAPTER 16. MODULES AND OBJECT CLASSES

A C++ class may also have a destructor function, which does the opposite of the constructor
function. If a destructor is defined for the class, it is invoked automatically at block exit to dispose
of any class instances that were created at block entry or during execution of the block. This
permits automatic recovery of storage locations that were created using new, a major improvement
over C and Pascal.
Function Declarations and Definitions. A function is a class member if it is declared within
the class. Remember that a function declaration is simply a header linestating the return type,
the function name, and the argument types. Compare the C++ code in Exhibit 16.8 to the Ada
code in Exhibit 16.6. Ada requires the programmer to write each function declaration twice: once
to declare whether it is public or private, and once with the definition. C++ syntax is simpler.
It permits the function definition either to follow the declaration immediately, or to be given
elsewhere. Practical considerations determine whether the definition should be with or separated
from the declaration. Definitions written with the declaration are expanded as an in-line code; that
is, the translator replaces each function call, in line, by a new copy of the compiled code for the
function. All the time and space overhead of stack frames, jump-to-subroutine, and subroutine
return is avoided.
If a member function is declared but not defined in a class, its definition must eventually be
given; often, this is done just after the end of the class. Defining a method outside the class does
not affect the status of the method as a class member. The method still has members access rights.
The difference is practical, not semantic; functions defined this way are compiled separately using
the ordinary function call and return mechanism.
Giving a definition outside its class has one affect on syntax. Because some functions have
multiple defining methods, we must specify the full name of the method when we define it, so that
the compiler can tell with which class the method belongs. To denote the full name of a method
we use the scope resolution operator, ::. It can be used in dyadic or monadic form:
hclass_namei::hfunction_namei( hargument_listi)
::hfunction_namei( hargument_listi)
Thus the full name of the method for pop in the class char_stack is char_stack::pop, and the
constructor function for the class is char_stack::char_stack.
Denoting a Single Method The meaning of the scope resolution operator, char_stack::push,
is easily confused with the dotted notation used in a function call,stk1.push, but they do not
mean the same thing. The first notation denotes a particular method which is a member of a
particular class. Every reference to a method outside its defining class must be written using ::.
Methods in programmer-defined classes are denoted using the dyadic ::, and globally defined
methods are denoted by using :: in its monadic form.
In contrast, the dotted notation is used refer to a function, not one of its methods. It is
important to be able to denote a single function method, as opposed to the entire function, which

16.4. OBJECT CLASSES.

505

Exhibit 16.10. A represented domain in C++.

class imag
{
im: float;
operator float() { return im; }
public:
imag(){ im=0.0; }
imag(float f){ im=f; }
friend imag operator+(imag a, imag b){imag( float(a)+float(b) ); }
friend float operator*(imag a, imag b){ -1*(float(a) * float(b)); }
} // Operators + and * are friends of this class.

is a collection of methods. A function call must be dispatched, 14 while a method may simply be
called directly. The dotted reference stk1.push means Start with the object stk1. Run the
function dispatcher to select the most appropriate method for the function push, then call that
method with stk1 as its implied argument. If no method for push is defined in the class of its
implied argument, some other method, higher in the class hierarchy, will be dispatched.
The most common reason for using a method name instead of a function name is that one
method for a function is being defined in terms of another. The philosophy of object-oriented
programming dictates that we should be able to use a single function name for both old and new
methods, so long as they implement a common external function. For example, a customized
printing function might be defined for a user-defined type by making several calls on one of the
predefined print functions. Within the definition of the new print method, we must be able to
denote the old method by its method name, in order to avoid a hopeless circular definition.

16.4.2

Represented Domains

Chapter 15, Section 15.4.3, discusses the problems inherent in using one domain to represent
another. Briefly, the language must allow some connection between the domains in order to permit
the programmer to define functions intrinsic to the new domain. However, once these are defined,
all further compatibility between the domains is undesirable. In Pascal, this unwanted compatibility
is unavoidable. In Ada, one-way protection is provided, as described in Exhibits 15.22 and 15.23.
In C++ we have more control over the situation.
A language with classes provides a different and much more satisfactory approach to this problem. Where in Ada, we would define a new represented domain, in C++ we would define a class
with one private data member from the old domain. The class imag in Exhibit 16.10 is an analog
of the Ada type imag defined in Exhibit 15.26. The information provided by the programmer and
14
The distinction cannot be fully understood until the material in Chapter 17 is mastered. However, we give a
brief explanation here.

506

CHAPTER 16. MODULES AND OBJECT CLASSES

even the syntax are similar in the two languages. The important difference is the semantics.
No casts or compatibilities are predefined for C++ classes, so the programmer is not stuck with
unwanted ones. In return, the two casts that are needed to define arithmetic for imaginary numbers
must be defined explicitly. The first is the constructor imag(float), which changes a float value
to type imag.15 Because these two functions are private, the relationship between types float
and imag is completely hidden from the outside world, and the semantics of type imag are fully
protected. In contrast, the semantics of the Ada type in Exhibit 15.26 make the cast imag(float)
public, and make the cast float() automatic. (That is, an imag is acceptable in any context that
requires a float.)
The other cast needed in our package is one that changes an imag number to its underlying
float representation so that we may operate on it. This is the operator function named float,
defined in Exhibit 16.10. 16 In C and C++ syntax, a cast is (syntactically) a prefix operator, not
a function, and it is called by placing the name of the cast to the left of the expression which is its
operand. Since the name of a cast is a type name enclosed in parentheses, this looks like a function
call with the parentheses around the wrong thing. The designer of C++ gave the option of calling
any operator, including a cast, using either operator syntax or standard function call syntax. Thus
we can call operator float() either of these two ways:
float(17); float (x+2)
(float) 17; (float)(x+2)

16.4.3

Friends of Classes

The class mechanism lets the programmer impose restrictions on type compatibility, object visibility, and access to parts of objects. These restrictions are immensely powerful and can be a huge
help in achieving semantically valid programs. However, when one uses a restricting mechanism, it
is often difficult to achieve exactly the right degree of restrictions.
This is true of classes, also. There can be relationships among classes that are important but
hard to capture using only public and private parts. This gives us tree-structured sharing among
classes. However, sometimes graph-structured sharing is needed. We could say that, sometimes,
we want semiprivate parts: parts that can be shared with some other program modules but not
with all.
We use classes to make programs more modular and more reliable. The important semantic
mechanism for achieving these goals is accessing restrictions. The data type of a class is hidden
within the class, and other functions are forced to work through the accessing functions defined for
the class. By using this methodology, we can easily change the representation of a class without
affecting the correctness of other parts of the program. However, each function call has a cost
in time and space efficiency. Although the cost of a single function call is small, these costs add
up. The cost of doing everything through function calls can be great. When the access operation
15
16

Conversion functions are treated more fully in Chapter 18, Section 18.1.
The word operator in lines 3, 7, and 8 is a keyword.

16.4. OBJECT CLASSES.

507

itself is trivial, like the pop function in Exhibit 16.8, the cost of calling the function can exceed the
cost of executing it. Thus applying modular methodology throughout a program could introduce
unacceptable inefficiency.
Two solutions are provided in C++ for this problem: friends and in-line code. A method may
be declared to produce in-line code, like a macro, rather than a separately compiled code module. If
the method is actually defined between the begin-class and end-class brackets, it will be expanded
in-line. If it is declared inside the class but defined elsewhere, it will be compiled as a separate
code module. The semantics and syntax are otherwise the same, as is the ability to access the
private members of the class. With short functions, such as push and pop, in-line code is preferred,
because it is more time- and space-efficient than separate compilation. For long functions, in-line
code would still be faster, but it could make the compiled program much longer if there are two
or more calls on the function. Thus in-line compilation is a bad idea for moderate-length and long
functions, and separate compilation is more commonly used.
The friend mechanism can also be used to avoid the inefficiency involved in constant calls
to trivial functions. A class will share its private members with friend functions, but not with
other functions. Friendship is given, not taken. A class declares who its friends are; a function
cannot declare which classes are its friends. The friendship relationship can be declared functionby-function, or an entire class can be declared to be a friend, which means that all its functions are
friends.
For example, assume that we have defined two classes, forest and tree. A forest is to be
a collection of trees, in this case an array. In general though, whether it is a set, an array, or a list,
it would be hidden within the class forest [Exhibit 16.11]. Both constructors for class tree are
called by function forest::grow_tree; this would be permitted even if forest were not a friend,
because these constructors are public in class tree. However, grow_tree also sets the value field
of the new tree node it creates. To do this, it uses the member name value, which is private to
class tree. This is permitted because class forest is a friend of class tree.
Comparisons
The class mechanism in C++ differs from the Ada package mechanism in many ways. Among the
most important are the use of classes to form type hierarchies and govern function inheritance and
automatic type conversion, as explained in Chapter 17. These facilities grow directly from the fact
that the C++ class is an extension of the facility for defining record types.
The Ada package has one property, though, that is lacking in the class. A package can contain
objects, functions, and more than one type declaration. Thus if we were writing a package, both
types tree and forest would be written in the same package. Having done this, the right of forests
to access tree parts would be established. A back door mechanism, such as the friend mechanism,
is not needed to establish a relationship of trust between the two types.

508

CHAPTER 16. MODULES AND OBJECT CLASSES

Exhibit 16.11. Friend classes.

class tree
{
tree * l_son;
tree * r_son;
char value;
public:
tree(){ l_son= r_son= NULL; value= \0; }
tree(tree *a, tree *b){ l_son =a; r_son =b; value= \0; }
friend class forest;
}
class forest
{
int n;
tree woods[100];
public:
forest(){ n=0; }
forest(tree * t)
{
if( n==100 ) cout <<"Forest is full.";
else woods[n++] = t;
}
void grow_tree(int n, char c)
{
tree * t = new tree;
t->value = c;
forest[n] = tree(t, forest[n]);
}
}

Exercises
1. Why should programs be modularized?
2. What are the dangers in using global variables?
3. What is the purpose of encapsulating a module?
4. How can a module share data and yet protect private information?
5. How do private type declarations make an ADT representation invisible to the programmer?
6. How does C use files and separate compilation to achieve the goals of modularity?

16.4. OBJECT CLASSES.

509

7. What is the role of the makefile facility? A header file? The main module?
8. What is the role of extern and static in C declarations?
9. How did Ada and C++ extend the goals of grouping, sharing, and protection to their languages?
10. How does Ada distinguish between public and private symbols?
11. What is a method in Smalltalk?
12. A C++ class is a template. How is it instantiated? Initialized?
13. Explain two ways in which constructor function might be treated differently than other functions in a C++ class.
14. What is an implicit argument? How is it used in C++?
15. What is the role of a destructor function?
16. What are the differences in the declaration of a function in C++ and Ada syntax?
17. What is the role of the scope resolution operator?
18. How does C++ give the programmer more control than Ada over semantics and compatibility
when using one domain to represent another?
19. What is the role of a friend class?
20. What is in-line code? When is its usage preferred? Not preferred?
21. Contrast classes in C++ and packages in Ada.

510

CHAPTER 16. MODULES AND OBJECT CLASSES

Chapter 17

Generics

Overview
Generic domains are domains which include more than one specific type of object.
They are used to express abstractions, to make code reusable, and to support top-down
programming. There are four kinds of generic domains: parameterized, subdomains,
polymorphic, and ad-hoc.
Several issues must be considered when implementing generic domains. They have
been considered in the support for generics that has been built into various languages.
The solutions vary in flexibility, preservation of semantic validity, binding time, and
efficiency. Approaches to this problem include: overloaded operators, flexible arrays,
polymorphic types, parameterized domains with instantiation, class hierarchies with
inheritance, and declarable domain relationships.
A corresponding tool for expressing procedural abstractions is the virtual function. A
virtual function is a function name, header, and description but no code. An ADT can
be represented in a programming language by a generic domain together with virtual
function declarations.
A generic function is a single abstract operation defined over a generic domain. Most
languages have a few primitive generic functions for all primitive types, which are automatically extended to user-defined types. However, user-defined generic functions create
semantic problems. The code that implements any process must be appropriate for the
type of its arguments; what works for one type does not for another. The translator,
therefore, must be able to handle multiple definitions for a generic function, and it must
decide which method to use for each function call. This process is known as dispatching.
An ad-hoc generic domain has subdomains that are related by semantics rather than by
511

512

CHAPTER 17. GENERICS

structure. Definitions of a function over two specific domains with an ad-hoc relationship
may look different if they depend on representation. However, they perform the same
semantic action on objects of the different specific domains.
Many older languages support some predefined generics for forms of nonhomogeneous
domains. These include union data types, overloaded names, fixed sets of generic definitions with coercion, extending predefined operators, flexible arrays, parameterized
generic domains, domains with type parameters, preprocessors, and generic packages.
These are far more limited than support for generic types found in object-oriented
languages because final binding of the type parameters happens at precompile time.

17.1

Generics

17.1.1

What Is a Generic?

Chapter 15 discussed domains, type checking, and type conversions in several familiar strongly
typed languages. In these languages, types are used to define domains such that each new type
constructed, with a few exceptions, defines a distinct domain incompatible with other domains,
and each domain contains only one type of object. In this chapter we examine the ways to declare
and use generic domains, that is, domains that include more than one type of object. We see
how generic domains can be used to express abstractions, to make code reusable, and to support
top-down programming.
We say the domain of a function is specific if every argument to the function is defined to be
a single specific type. The properties of a specific domain are fully defined by the specific type
definition used to define the function, and the function makes use of these properties in order to
perform meaningful computations. The opposite of specific is generic. A function domain is
generic if some argument can be of two or more different specific types. There are four ways in
which a domain, D, may be generic:
D may be defined by a type expression with components of an unspecified type and/or an
array with unspecified array bounds. We call such a type expression a parameterized type.
Several languages provide some support for this kind of generics. The domain it defines is a
parameterized generic domain.1
D may have a subdomain, D 0 . This occurs when D is defined as a type for which subtypes
have been declared. For example, the Pascal subrange type 1..10 is a subtype of the type
integer.
1

Implementations of generics over parameterized domains are discussed in Section 17.3.

513

17.1. GENERICS

Exhibit 17.1. Generic domains and functions on them.

Domain
number
matrix of numbers
character string
set of ?
stack of ?

Generic Functions on the Domain

+, -, *, /
inverse, transpose, +
concatenation, substring
union, intersection
push, pop, top

D may be defined by a polymorphic type. A polymorphic type is a single type with internal
variability, like a discriminated union.
D is an ad hoc generic domain if it includes objects of more than one specific type, such
that all of the specific types are representations of the same external domain. The semantic
relationships among these species are important, but the species are related in an ad hoc
manner, not by structure or representation. For example, the generic domain number
includes at least two specific domains integer and real in most languages, and the generic
domain tree_node can be defined to include specific types leaf_node and interior_node.
Ad hoc generics are a topic of current research interest.
Consider the domains listed in Exhibit 17.1. The domains number and matrix of numbers
are generic because they include both integers and reals. Matrix and string are generic because
their dimensions are not specified, so these types include matrices of all shapes and strings of all
lengths. Set is generic because the base type of the set is not specified and, therefore, sets of
reals, characters, cells, and any other type. are all included. Stack is generic for the same reason,
and also because the size and structure of stacks can vary.

17.1.2

Implementations of Generics

There are several ways that generic and polymorphic functions may be supported within a programming language, but all approaches must provide answers for the same syntactic and semantic
problems. The issues to be considered include:
What kind of generic and/or polymorphic domains may we define? Parameterized domains?
Structurally related domains? Domains with a small number of variants? Unlimited ad hoc
generics?
How and under what conditions may we define a generic or polymorphic function? May we
define functions over generic domains?
How do we translate calls on functions with more than one specific method? What information
is considered by the translator when it chooses a method? Can this choice be deferred until
run time or is it always made at compile time?

514

CHAPTER 17. GENERICS

If we permit run-time type variability, is it possible to compile code with reasonable efficiency?
How do generic domains and generic functions interact with type coercion?
What does it mean to call a function which has a generic formal parameter? [See Exhibit
17.3] What type matching rules should apply? How can we implement such calls?

Support for generics has been achieved to different degrees for domains with subdomains, parameterized domains, and ad hoc generic domains. The solutions provided by various languages
vary in their flexibility, their ability to preserve semantic validity, their time of binding, and the
inefficiency inherent in the implementation. We will examine several approaches to this problem in
some detail.
Overloaded Operators. A few generic arithmetic operators are built into the language. A
programmer can define new methods for these operators but cannot define any new operators or
functions that are generic in the same way. Adas operators are an example and are discussed in
Section 17.2.
Flexible Arrays. A programmer can use array parameters whose bounds are not defined at
compile time. Either the array length must be passed as a separate parameter, or the code for the
function must work correctly for all possible values of the array length. Several older languages
support flexible arrays, including FORTRAN, C, and Pascal. These are covered in Section 17.2.
Parameterized Domains with Instantiation. The generic module is kept in the form of
parameterized source code. A preprocessor is used to instantiate the parameterized code with
actual type arguments and generate fully specific, ordinary code which is bound to a unique name
and then compiled. The programmer uses the unique name, not the generic name, in the code.
This kind of generic module is supported by Ada and can be implemented in C using macros. It is
discussed in Section 17.3.
Class Hierarchies with Inheritance. Objects belong to classes, as do function methods. The
symbol table is available at run time, and the class of the actual arguments to a function can be
examined. One name is given to an abstract function, and many methods can be defined for that
name, so long as each method is defined in a different class. A function method has one implied
argument, and the translator uses the class of that argument to determine which method to dispatch
for each function call. This kind of support for generics is present in Simula, Smalltalk, and C++.
Because generic dispatching must sometimes be postponed until runtime, these languages are at
least partly interpreted. This is discussed in Chapter 18, Section 18.2.

17.1. GENERICS

515

Polymorphic Types. The type of each symbol can be determined at run time. Functions are
polymorphic, and each part of a function definition is controlled by a predicate. These predicates
may be used to test the type or the value of the actual argument for each call. The translator
will use a powerful pattern-matching facility to determine which predicate is true and select the
corresponding action. Miranda and Haskell support this kind of generics. They are discussed in
Chapter 18, Section 18.3.
Declarable Domain Relationships. Generic function calls are interpreted, not compiled. The
type information for generic objects is kept available at run time. One name is given to an abstract
function and is used to refer to all of its methods. A generalized graph of relationships between
domains can be created or declared, and is used to dispatch functions. In a generic function call,
the arguments might be of any specific type which is a subdomain of the generic domain, according
to the graph. A function call with generic parameters translates into code that will look at the
argument types at run time and execute the function method that is most appropriate for those
types, using coercion if necessary. Examples of this approach to dispatching are drawn from C++
and Miranda and are discussed in Chapter 18, Section 18.4.
The extent to which a language supports generics is an important issue. The greater the extent,
the more flexible and adaptable a program can be at run time. The smaller the extent, the more
difficult it is to create reusable code and code libraries. Generic domains are supported to a minimal
extent in most languages, but the kind of support provided, the restrictions, and whether support
extends to programmer-defined domains and functions vary dramatically from one language to
another. Traditional typed languages such as Pascal usually have some predefined generic domains
with predefined relationships, but they lack any way for programmers to define their own generic
domains or domain relationships. For example, a general matrix multiplication function cannot be
defined in Pascal or FORTRAN because there is no way in these languages to define the generic
domain matrix of integers or reals.
Programmer-defined subdomains and domains with integer parameters are relatively easy to
implement and are supported in many languages. Domains with type parameters are more difficult
to implement in a strongly typed, compiled language because ordinary type checking requires that
the specific type of each function argument be known (or deduced) at compile time. Ad hoc generic
domains are the most difficult to implement because there is no structural relationship among the
variants that can be exploited.

17.1.3

Generics, Virtual Functions, and ADTs

A generic domain is an abstraction that represents the shared structural and/or semantic properties of its subdomains. Thus a language that permits programmers to declare their own generic
domains provides a powerful tool with which to express data abstractions. A corresponding tool for
expressing procedural abstractions is the abstract function, or virtual function. This is a function
name, a function header (or prototype), and a description of the intent of the function, with no
accompanying code. The header defines the number and types of the function parameters and re-

516

CHAPTER 17. GENERICS

turn value. The accompanying description defines the intended semantics of each, and any relevant
assumptions or restrictions.
An abstract data type can be represented in a programming language by a generic domain
together with a collection of virtual function declarations. A virtual function would be included
for each essential representation-dependent process to be performed on members of the domain.
This set of virtual functions explicitly describes the common behavior of all objects of the generic
domain and is, thus, a description of the semantics of the generic domain. The virtual functions
would be guaranteed to exist for every subdomain and would form the basis for defining all
functions on the generic domain. In a few languages, a programmer is able to explicitly declare
and name a generic domain, D, then use a single method to define a generic function, GF , over D.
The function header for GF would refer to D, and the code for GF could call the virtual functions
defined for D and rely only upon properties that are common to all specific domains contained in
D.
A realization of a generic domain is some specific, fully defined type that implements the semantics of that domain. All realizations of a generic domain must have methods defined for all of
the domains virtual functions. A program could contain several realizations of the same generic
domain. For example, in Pascal, integer and real are both realizations of the domain number.
The translator and linker must guarantee that, by run time, realizations of all promised functions
for all subdomains must exist.

17.1.4

Generic Functions

A generic function is a single abstract operation that is defined over a generic domain. In most
languages, FETCH, STORE, and comparison for equality are primitive generic functions that are
defined for all primitive types. Most languages developed since the early 1970s extend these primitive generics automatically to user-defined types. Let us call these three functions the universal
generic functions. Other functions that are commonly generic are READ, WRITE, and the arithmetic
operators.
A generic function creates a semantic problem. In order for a process to be meaningful, the
code that implements the process must be appropriate for the type of its arguments. In general,
code that is meaningful for one type is not meaningful for other types. We must, therefore, ask
how a function can meaningfully process data from a generic domain. In the older languages and
their translators, this was done in an ad hoc manner. With a limited number of generic functions
built into the language, as in Pascal and Ada, the translator can treat these functions as special
cases, look at the types of the actual arguments, and generate specific code that is appropriate
for them. On the other hand, if the user is permitted to define his or her own generic functions,
the translators type-checking algorithms must be extended to form some general mechanism for
handling generic functions. Such mechanisms are the subject of this chapter.
Depending on the language, a user-defined generic function may be defined by one block of code
whose parameter is from a generic domain, or by several independent blocks of code which have
been declared to share a common name. When a function is defined by more than one body of

517

17.1. GENERICS

Exhibit 17.2. A generic function as a collection of methods.

In this diagram a method is represented by a shield-shaped box, with its domain written across its
top. A function is a set of methods with a common name. Note that the function + in FORTRAN
77 is much more extensive than the function of the same name in Pascal.

"+" in Pascal

"+" in FORTRAN-77

integer, integer

integer, integer

complex, complex

real, real

real, real

double, double

code, each of the separately defined bodies is called a method for the function, and each method
must be declared with arguments from a different combination of domains. We define the domain
of such a function to be the union of the domains of all its methods.
There are three ways that a generic function may be defined by a single block of code: the
processing method can be independent of the argument type (as in comparison for equality), the
code can be polymorphic (see Chapter 18, Section 18.3), or the code can bypass the problem of
processing a generic argument by passing the argument on in a call to some other generic function
[Exhibit 17.4]. Functions that operate over ad hoc generic domains are either polymorphic or are
defined as a collection of methods [Exhibit 17.2].
Dispatching a Function Method
It is more difficult for a translator to handle function names with multiple methods than names
that represent one block of code. The process of deciding which method to use for a call is known
as dispatching the call, and it is carried out by a part of the translator called the dispatcher.
The dispatcher uses information about the function arguments to choose an appropriate method.
Dispatching could be (and is, in some languages) done on the basis of the number of arguments,
the types of one or all arguments, or the actual value of the arguments.
Some languages will try to coerce a function argument if the types of the actual arguments do
not match any of the defined methods. Depending on the language, the dispatcher might use a
subtype relationship (Pascal or Ada), possible type conversions (Pascal, FORTRAN), or hierarchical
relationships among the domains (C++, Smalltalk). For example, if the function Square in Exhibit
17.3 were called with an integer argument, the hierarchical relationship between the domains
integer and number could be used to dispatch the function call. If none of the defined methods is

518

CHAPTER 17. GENERICS

Exhibit 17.3. A function with generic formal parameters.

Assume that number is a generic type which includes types integer and real. If we were to
define the function Square on numbers (using a Pascal-like syntax), the function header would be
similar to this:
Function Square( a:

number):number;

Assume that the integer variables k and m have been declared, and we call the function thus:
k := Square(m);
We would like this call to be legal. The formal parameter, a, is generic, and the actual argument
has a specific type which is a subtype of that generic. We would like the subtype integer to inherit
the definition for Square that was supplied for its generic supertype.

appropriate for the context, and no coercion is permitted that can make the arguments appropriate
for some method, the function call contains a type error.
In Chapter 18 we consider several very different implementations of programmer-definable
generic functions, with very different rules for dispatching. Some of these rules, such as those in
Miranda, are quite general and require a sophisticated dispatcher. Others, such as the hierarchical
dispatching rule in Simula and Smalltalk are more limited and can be implemented simply.
Functions over Generic Domains
Some languages permit the programmer to define a generic function with an argument from a
generic domain and only one defining method [Exhibit 17.4]. When such a function is called,
the arguments type will be some specific type that is included in the generic domain. Such a
method cannot operate directly on its parameters, but must either test its argument and take an
appropriate branch (polymorphic behavior) or perform its process by calling other generic functions
defined over the same domain. The lower-level function could, in turn, pass the argument on to
another generic function. However, a chain of such generic calls must ultimately end with a call
on a bottom-level polymorphic function or a generic function that is defined by a set of specific
methods. This bottom-level function will include one body of code for each representation of the
generic data type that the programmer is using. In the prettyprint example, the chain of calls
will end with a call on some body for print_node that is defined to format and print out the data
from one specific type of node.
Generic Functions in Ad hoc Domains
An ad hoc generic domain has subdomains that are related by their semantics rather than by
having some common structure. Definitions of a function over two specific domains with an ad hoc
relationship may look quite different if they depend on the representation. The most that can be

17.1. GENERICS

519

Exhibit 17.4. A generic function defined by a single method.

Assume that you are working with binary trees in a hypothetical language that permits definition of generic procedures. Prettyprint is a procedure that prints a generic data type called
binary_tree.
Prettyprint calls the explicitly generic functions print_node, go_to_left_son, and
go_to_right_son, but it does not do any operations that depend on how the tree is represented.
The programmer must define one method for each of these three functions for every kind of
tree node being used. Once this is done, one method suffices to define prettyprint for the entire
generic domain binary_tree, no matter how many ways that domain might be implemented.

said is that they perform the same semantic action on objects of the different specific domains. For
example, consider the ADT Stack, which includes the two generic subdomains array_stack and
linked_list_stack. The type declarations that are appropriate for implementing an array_stack
and a linked_list_stack are shown in Exhibits 17.5 and 17.6. These sets of definitions have little
in common except the names. Specifically, the selector functions defined in one bear no systematic
relationship to the selectors defined for the other.
The code for push on an array-represented stack [Exhibit 17.5] bears no easily described rela-

Exhibit 17.5. Pascal definition of push on an array_stack.

Here are type declarations and the function push for a stack of 100 reals:
CONST len = 100;
TYPE real_stack= RECORD
store: ARRAY [1..len] OF real;
top: 0..len END;
VAR rs: real_stack;
value: real;
FUNCTION push (VAR rs:real_stack; r:real): boolean;
BEGIN
IF rs.top = len THEN push := false (* error flag *)
ELSE BEGIN
push := true;
rs.top := rs.top + 1;
rs.store [rs.top] := r
END (* IF rs.top *)
END; (* push *)

520

CHAPTER 17. GENERICS

Exhibit 17.6. Pascal definition of push for a list_stack.

Compare the code below for the function push on a linked-list implementation of a stack to the
code in Exhibits 17.5 and 17.15 for the same generic operation on an array implementation of a
stack.
TYPE list_stack = stack_cell;
stack_cell = RECORD
val : real;
next: list_stack END;
VAR rs: list_stack;
value: real;
FUNCTION push (VAR rs:list_stack; r:real): boolean;
VAR t:list_stack;
BEGIN
push := true;
(* stack cannot overflow *)
t := rs;
(* save current stack pointer *)
NEW(rs);
(* append new cell to top of stack *)
rs.val := r;
(* initialize new cell *)
rs.next := t
END (* push *)

tionship to the code for push on a linked-list-represented stack [Exhibit 17.6], even though they
carry out the same abstract operation. Representation-dependent functions such as push and pop
must be defined by separate methods for the species of an ad hoc generic domain. However, the
common intent of the subdomains is known to the programmer, and the common functionality can
be expressed by providing appropriate definitions for the stack operations. Further functions can
then be defined in a representation-independent way for the generic domain stack, in terms of
these declared abstract functions.
When two species of a generic domain are related by their intent, it is possible to define conversion functions that map the elements of one onto the elements of the other and vice versa. This
mapping is done in such a way that corresponding elements are alternative representations of the
same external object. When such a mapping is applied, the physical form of the program object is
changed but the semantic intent is preserved. These conversion functions are unlike other functions.
They do not merely operate on elements of the generic domain; they actually define the semantics
of the generic relationship.

17.2. LIMITED GENERIC BEHAVIOR

17.2

521

Limited Generic Behavior

Working with domains that have several natural representations becomes difficult in a strongly
typed language when the representations must interact, unless the language permits some deviation from the one-type-one-semantics rule. For this reason, many older languages support some
predefined generics and/or permit some flexibility in the types of array arguments. In this section
we explore ways in which limited forms of nonhomogeneous domains can be and have traditionally
been supported.

17.2.1

Union Data Types

Generic domains arise naturally in many algorithms. For example, consider the data structure
called a balanced tree, or B-tree. This is a branching index structure in which internal nodes
point to other nodes and leaf nodes point to potentially large data records. Node, then, has two
varieties, internal and leaf, and the programmer working with these trees must mix the two
representations in one structure. What is needed is the ability to define a type node that has
two independent realizations, or a type pointer-to-node that can point at either a data record or
another node.
Union data types have been used to implement balanced trees, but they are not a good solution
for two reasons. First is the possible violation of the semantic intent unless the union is discriminated, in which case the translator guarantees that the discriminant tag is always correct, and the
type checker always checks this tag. With fewer restrictions, a function could be applied to the
wrong variant. Second, union data types are not a practical solution when the size of the more
frequently used representation is much smaller than the size of the other, because of the excessive
amount of storage that would be consumed by the allocated but unused portions of every record.
B-trees are an example of a programmer-defined domain for which union data types are not a good
implementation.
The C language is popular partly because it permits a direct implementation of this kind of
pointer structure. Any C pointer type can be cast to any other pointer type, and such a cast can
be used to make a pointer to a leaf node assignment-compatible with a pointer to an interior node.
Since types are not checked when pointer assignments are done, there is no problem having pointers
in the same tree point at two types of nodes. Programmers then become responsible for doing their
own type checking on the nodes. This kind of flexibility is not available, though, in languages that
provide the semantic protection of thorough type checking.

17.2.2

Overloaded Names

The several methods of a generic function implement a single external operation over different
species of one generic domain. In an ideal language, we would have one name per generic function,
whether the function was predefined (like =) or defined by the programmer. For example, no
matter how many instances of the stack ADT we make, we would like to be able to refer to the

522

CHAPTER 17. GENERICS

Exhibit 17.7. An overloaded function, +.

Many different abstract processes have been called +, all of them dyadic, associative, and
resembling addition in some way:
Addition: Compute the arithmetic sum of two numbers.
Logical OR: Result is true if either argument, or both arguments, are true.
Set union: Join two sets together and eliminate duplicates.
String concatenation: Attach one string to the end of another.
+ is an overloaded generic function because these processes do not have the same semantic
intent, and their domains are unrelated. At the same time, ordinary addition is defined over the
generic domain number, which includes integers, reals, etc. This subset of + that deals with
addition forms a pure generic function.

stack functions as push and pop. It is evident that older languages, such as Pascal are far
from ideal! If we defined two kinds of stacks in a Pascal program, we would have to define the
stack functions twice and give them different names. In modern languages, we are permitted to
overload a function name by defining multiple methods for one function. An overloaded name is
a name with two or more bindings in the same scope.
Overloaded names, unfortunately, are not restricted to use with generic functions. If a language
supports name overloading, nothing can stop a programmer from overusing it. An overloaded name
could be used to denote semantically unrelated processes on unrelated types as easily as to denote
a true generic function.
Ada supports the traditional generic arithmetic operators and extensible name overloading. 2
The semantics of these operators can be extended by adding new methods to the existing predefined
set. This is called overloading because each operator name denotes several specific methods: a
heavy load for one name. However, no distinction is made in Ada between methods that carry
out the semantics of arithmetic and those whose semantics are unrelated. Both kinds of methods
may be added to the primitive functions. For example, we might use overloading to extend Adas
+ function to include methods for integer + float and string concatenate string, as
suggested in Exhibit 17.7. The first method belongs in the same generic family as the primitive
methods. The second one is semantically unrelated and does not. The language permits any
existing arithmetic operator to be overloaded, but new ones may not be defined.
Using the arithmetic function name + to symbolize something unrelated to addition (such
as concatenation) is a questionable practice. We want to make a clear distinction between generic
2

A short clarification of terminology seems necessary here to avoid confusion. Two kinds of generic objects are
supported in Ada. In Ada references, the term generic refers to parameterized source-code modules, which are
covered later in this chapter, and the term overloaded refers to operator names which have ambiguous meanings.

17.2. LIMITED GENERIC BEHAVIOR

523

functions that implement a single abstract function and those that include methods for a mishmash
of abstract functions. Hereafter, we will use the term generic function to mean a function name
that represents a single abstract function over a generic domain. The term overloaded function
will refer to a single name used to represent any collection of function methods, possibly having
unrelated meanings. Pure generic functions have clean semantics and obey the principle of distinct
representation (see Chapter 2.3); overloaded functions do not.

17.2.3

Fixed Set of Generic Definitions, with Coercion

Number is the best-known and most-used generic domain. It was recognized early in the history of
computing that integers and reals (and in COBOL, BCD strings) are all representations of numbers,
and that the programmer must be able to work with all kinds of numbers without worrying about
type conversions. Thus the generic domain number has been built into most older languages. To
say that this generic mode is built-in, we mean that the programmer is permitted to use objects
belonging to both primitive types as if they belonged to the same domain. More precisely:
Two or more representations of the external domain number are defined in the language as
primitive types.
Conversion functions are predefined that allow one- or two-way compatibility between those
primitive types.
The language syntax allows objects of the two types to be used together in contexts that were
defined only for homogeneous pairs.
The compiler will coerce an argument of one type to make it appropriate for a context that
requires the other.
In FORTRAN 77, C, and Pascal, the functions +, -, *, and / are generic functions,
defined over the implicit generic domain number [Exhibit 17.2]. However, languages designed
before the mid-1970s rarely permit the programmer to define new generic functions or even to
define new methods for existing generic functions.
When generic functions are built into a language, they often interact with rules for type coercion.
In the diagrams of + for both Pascal and FORTRAN, note the absence of a method for adding an
integer and a real. These languages permit the programmer to write arithmetic expressions that
involve a mixture of real and integer operands. However, these computations are not implemented
by separately defined methods. Rather, they are implemented by a combination of the method
for real addition and a coercion function which converts an integer to a real. The conversion from
integer to real is meaningful because both domains can be used to represent the same external
objects (whole numbers). These conversion functions can thus be invoked, without a change in the
semantics of the argument, when they are needed to carry out the computation.
In contrast, Ada also supports the implicit generic domain number but not mixed-type arithmetic. The arithmetic operators are generic, and predefined methods for + include:

524

CHAPTER 17. GENERICS

Exhibit 17.8. Definition of new methods for an Ada operator.

We can implement mixed-type addition by adding new methods to the predefined set. The new
method works by explicitly converting one of the operands to be compatible with the type of the
other, creating operands of the right types for a previously defined method. The first definition
below defines the semantics for 3 + 5.2 but not for 5.2 + 3. The second definition is also
required to permit + to be called with integer and float operands in either order.
function "+" (x: integer, y:float) return float is
begin return float(x) + y end "+";
function "+" (x: float, y:integer) return float is
begin return x + float(y) end "+";

function "+" (x,y: integer) return integer;

function "+" (x,y: float ) return float ;
However, Ada does not support representation coercion of any sort, so addition of an integer and
a float is not predefined. It is possible for a programmer to define a limited number of mixed-type
operations in Ada. This is explained more fully in Section 17.2.4.
A more elaborate built-in generic function is the COBOL assignment command, MOVE. Several
definitions of MOVE exist which allow the large number of numeric and nonnumeric data types to be
used together. MOVE is also invoked automatically in the process of performing arithmetic, input,
and output.

17.2.4

Extending Predefined Operators

As in most languages, the arithmetic operators in Ada are generic, and the intended meaning of
each operator is determined by looking at the types of its operands and selecting the method
defined for those types. This very limited generic facility was included in Ada, as in most common
languages, for convenience, so that the programmer could use more than one representation of
numbers without needing to learn unfamiliar unique names for the operator methods.
Unlike many languages, however, Ada does not provide automatic type coercion to make a
set of operands conform to the types declared in one of the available methods. Thus mixed-type
operations are not predefined. To compensate for this lack, Ada permits more methods to be
loaded onto the existing arithmetic operators 3 [Exhibit 17.8]. Each new method must involve a
new combination of operand types. As a happy side effect, the overloading mechanism permits us
to extend the basic arithmetic operators to work on programmer-defined types [Exhibit 17.9]. A
separate function must be explicitly provided for every combination of operator and operand types
that the programmer wishes to use. This can lead to quite a lot of definitions if more than one
3

A similar facility was provided by the MAD compiler, Arden, Galler, and Graham [1963].

17.2. LIMITED GENERIC BEHAVIOR

525

Exhibit 17.9. Extensions of Ada operators for new types.

This exhibit further develops the material in Exhibit 17.8. We extend * to operate on the
mapped domain float and extend + to add two vectors.
type mass is new float;
type vector is array (1..5) of integer;
function "*" (x: integer; y:mass) return mass is
begin return mass (float(x) * float(y)) end "*";
function "+" (x,y: vector) return vector is
z: vector;
begin
for k in 1..5 loop z(k) := x(k) + y(k) end loop;
return z;
end "+";

operator or more than two types of operands are involved. To write a large number of these would
indeed be tedious!
Examples of new method definitions are given in Exhibit 17.9. The first declaration extends
* to work on an integer and a new type represented by float. The domain structure built up
here begins to be complicated, demonstrating a two-step relationship within a generic domain. To
multiply a mass by an integer requires a demotion cast (mass-to-float) and a conversion (integerto-float). Finally, a promotion cast (float-to-mass) is required to correctly label the type of the
result.
The second function defined in Exhibit 17.9 extends the + operator to a compound type,
vector.4

17.2.5

Flexible Arrays

It is not difficult to code a function that can process any length array of a given base type. Processing
of flexible arrays is supported in some very old languages, for example both FORTRAN [Exhibit
17.10] and C [Exhibit 17.11]. These functions work because the method for summing the elements
of a vector involves a process that is repeated for each vector element. The length of the vector is
passed as a separate argument and used to stop repetition.
Flexible array parameters are implemented very simply and very similarly in C and FORTRAN.
Array arguments are passed by reference in both languages. This avoids the need to allocate storage
space for the argument in the functions stack frame, and the need to copy the contents of the array
4

This is possible because Ada supports coherent passage of compound objects as parameters. An explicit declaration of k is not required; the for loop implicitly declares the loop variable.

526

CHAPTER 17. GENERICS

Exhibit 17.10. An implicit parameterized domain in FORTRAN.

The letters at the right refer to the notes in the text; they are not part of the program.

10

REAL FUNCTION SUMUP (AR, N)

REAL AR(N)
SUMUP=0.0
DO 10 I=1,N
SUMUP = SUMUP + AR(I)
CONTINUE
END

(a)
(b)
(c)
(d)
(e)

into that stack frame. Neither language supports automatic run-time bounds checks for arrays. The
lack of prior knowledge of the size of the array causes no problem for the compiler, since no compiletime decisions are based on it. To ensure correct processing, a flexible array argument must have
a recognizable terminating value, or the array length must be passed as a separate argument.
In FORTRAN, the dimensions of an array parameter do not need to be declared, so long as they
are also passed as parameters or declared in a COMMON statement as global variables. Processing
can proceed correctly so long as the array length is known at run time. Exhibit 17.10 shows how
a flexible array might be processed in FORTRAN. Following are the program notes:
a. SUMUP is defined as a function of two parameters that returns a real.
b. The parameter AR is defined as a real array of unknown length.
c. This line defines the scope of a loop (up to statement #10) and directs that it should be executed
N times, with I taking on the values 1 to N.
d. This is not a recursive call. FORTRAN does not support recursion. Within a function, the
function name serves as a local variable.
e. The final value assigned to SUMUP is returned at END.
Similarly, in C, arrays are passed by reference and subscript bounds are not checked at all. With
multidimensional arrays, all dimensions except the first must be known at compile time in order
to translate subscript expressions. However, the first (left-hand) dimension is not used in subscript
computations. Thus the first dimension of an array parameter does not need to be declared in the
function.
In Exhibit 17.11 we define a C version of the FORTRAN function in Exhibit 17.10. The following
notes are keyed to the comments in the code. Line a declares the types of the formal parameters.
Note that the length of array ar is not specified. Line b declares a local floating-point variable,
sum, to be used as an accumulator. Lines c and d comprise a for loop that will be repeated n
times, with variable i taking on the subscripts 0 to n 1. This loop adds the i-th array element
to the sum. Line e returns from the function with the final value of sum.

527

17.3. PARAMETERIZED GENERIC DOMAINS

Exhibit 17.11. A parameterized domain in C: array of floats.

real sumup (float ar[], int n)
{
int i;
float sum;
for ( sum=0.0, i=0; i<n; i++ )
sum = sum + ar[i];
return sum;
}

/* a */
/*
/*
/*
/*

b
c
d
e

*/
*/
*/
*/

These are examples of implicit, not explicit, parameterized generic domains. However, they
demonstrate that an implementation of integer-parameterized domains can be straightforward. A
function with a flexible array parameter is, technically, generic, because it can accept arguments
of many specific types. However, it does not cause the compilation and interpretation problems
inherent in type-parameterized generics. The code for a flexible array argument can be compiled
and type checked (but not bounds checked) because all the types involved are determined at compile
time.5
Arrays with Explicitly Parameterized Bounds
The inclusion (or exclusion) of flexible array parameters in Pascal caused a great deal of dissension
among the twenty members of the International Standards Organization committee that developed
the standard for Pascal. In the end, array arguments with flexible bounds, called conformant
arrays, were included in ISO Level 1 Standard Pascal, but not in the ANSI standard. American
implementations of Pascal, therefore, generally lack this important facility.
One of the issues the committee argued about was run-time bounds checks, which are normally
part of Pascal semantics. Rather than expecting the programmer to pass this information separately
or implicitly, the syntax for a conformant array parameter indicates that the bounds are passed as
part of the argument. These bounds are named in the formal parameter list and can be accessed
within the code. Exhibit 17.12 shows how the function to sum a vector would be written in ISO
Pascal. The conformant array parameter can be passed as either a value or a VAR parameter.

17.3

Parameterized Generic Domains

Varying the bounds or base type of an array or the type of a field in a record creates a collection of
closely related types. If we write a type expression with dummy parameters in place of one or more
of its fields, the result is a parameterized type expression. Consider the domain which is the union
5
Automatic bounds checks cannot be done in C since the array bounds are not specified at compile time and are
not supplied at run time in a form that the run-time system can access.

528

CHAPTER 17. GENERICS

Exhibit 17.12. A Pascal function with a conformant array parameter.

Compare this ISO Pascal function definition to the C example in Exhibit 17.11. Note that the
bounds of the Pascal array are automatically passed as parameters to the function; they are named
and can be used in the code.
Function SumUp(VAR ar[ Lower..Upper ] of real):real;
Var i: integer;
Sum: real;
Begin
Sum := 0.0;
For i:= Lower to Upper do
Sum := Sum + Ar[i];
SumUp := Sum
End;
Here are some data declarations and two appropriate calls on the SumUp function.
Var Scores: array[1..3] of real;
Charges: array[1..1000] of real;
Answer := SumUp(Scores);
Answer := SumUp(Charges);

Exhibit 17.13. A parameterized type.

We define a type with two parameters, an integer and a type, then instantiate it to produce
two specific subtypes. (The syntax used is an extension of Pascal.)
Intent: A buffer for items of type TT
Parameterized type expression:
TYPE BUF(m:integer, TT: type) = array [0..m] of TT
Specific instances derived by instantiation:
Call
Buf(5, real)
Buf(10, integer)

Resulting specific type

array [0..5] of real
array [0..10] of integer

529

17.3. PARAMETERIZED GENERIC DOMAINS

Exhibit 17.14. A parameterized linked list type.

Intent: A linked list cell.
Parameterized type expressions:
Type cell_ptr = cell;
Type cell(DD:type) = record data:
Specific instances:
Call
cell(integer)
cell(mytype)

DD; next:

cell_ptr end;

Resulting specific type

record data:integer; next:cell_ptr end
record data:mytype; next:cell_ptr end

of all the domains associated with a parameterized type, for all possible values of the parameters.
We call this a parameterized domain. Exhibit 17.13 shows a parameterized type and two domains
which have been derived from it by instantiation.
Exhibit 17.14 shows a mutually recursive pair of types which include a parameterized type and
a pointer type which depends on it. The two domains formed by these types are both generic.
Type cell has explicit parameters. Type cell_ptr is defined in terms of cell and thus inherits
the generic nature of cell.
A parameterized type expression is not compilable code as it stands. Rather, it is a template
from which code can be created by instantiation. In an instantiation call, the programmer supplies actual, specific arguments. When this call is processed, the arguments are substituted for the
dummy parameters in the type expression, according to the lambda calculus substitution rule. The
result is an ordinary type expression. Instantiation happens during an early phase of compilation,
before parsing. To use a parameterized generic type, the programmer would create a parameterized source code module containing the parameterized type expression(s) and related function
definitions.
With a type-parameterized generic, the processing method is not quite independent of the type
of an argument. For example, the code for pop and push on a stack of reals is almost the same as
the code for a stack of integers, except that a different number of bytes of information will need
to be fetched, stored, or returned to the calling program. Similarly, matrix multiplication is the
same process whether operating on a matrix of reals or integers, except that appropriate methods
for the * and + functions must be used.
If a function were defined for several species of one type-parameterized domain, the source code
for the various definitions would very likely be identical except for the types of the parameters.
(Compare the code for push in Exhibits 17.5 and 17.15.) The object code, however, is not identical,
because the compiler uses the type declarations to compile appropriate methods for the built-in
generic functions such as fetch, store, =, *, and +. Thus the types of objects must be known
before the code can be fully compiled.
When a programmer writes specific type declarations and code that uses them, we say that the

530

CHAPTER 17. GENERICS

Exhibit 17.15.

Pascal definition of push on a stack of characters.

Here are type declarations and the function push for a stack of 15 characters. Note that the code
between BEGIN and END is the same as in Exhibit 17.5; only the type declarations have changed.
CONST len = 15;
TYPE char_stack = RECORD
store: ARRAY [1..len] OF char;
top: 0..len END;
VAR rs: char_stack;
value: char;
FUNCTION push (VAR rs:char_stack; r:char):boolean;
BEGIN
IF rs.top = len THEN push := false (* error flag *)
ELSE BEGIN
push := true;
rs.top := rs.top + 1;
rs.store [rs.top] := r
END (* IF rs.top *)
END; (* push *)

data types are bound at source time. This is the only programming style supported by languages
in the Pascal family, and it does not permit use of user-defined generic domains.

17.3.1

Domains with Type Parameters

We should distinguish between an integer-parameterized domain (a flexible array) and a typeparameterized domain, which has a type parameter like the type cell in Exhibit 17.14. Few
existing languages permit component types to be parameterized, while several languages permit
functions to be written with implicit or explicit flexible arrays.
An ADT is a collection of types, functions, and objects that express an abstract process on
abstract data. When Pascal programmers wish to use an ADT such as stack or linked list,
they start by writing type declarations for their own data and for the ADT; they then write out
definitions for the ADT functions, making adjustments to the function headers so that they are
compatible with their own types. If two variants on the generic domain are needed, each part of
the ADT must be written out twice.
Rewriting and re-debugging the functions for a common ADT is tedious but not difficult. The
code for each new variety of stack is so similar to the code for other varieties that code can be copied
out of a reference book with minimal modification. This leads us to ask whether it is possible to
automate the process of coding up a new version of an ADT. We would like to keep the ADT code

17.3. PARAMETERIZED GENERIC DOMAINS

531

in a library, so that it does not need to be reentered manually each time it is used.
Many ADTs can be expressed as sets of parameterized type definitions and functions with
parameters of those types. Definitions of this sort can easily be kept in a library, but we need a
way to relate the type parameters in the library modules to the types in a users program. Also,
the function headers for the ADT must be made compatible with the users own data types.
The easiest way to achieve these goals is by adding a preprocessor to the language translator.
The ADT is coded as a parameterized module that contains type and function declarations written
in terms of the generic parameters. These symbols will be bound to specific meanings at precompile
time by appropriate preprocessor commands. A library of parameterized source code modules
could be made available to programmers to include and instantiate, as needed. Before compiling,
the preprocessor is used to supply specific meanings for the generic parameters. The source code
package is then expanded, like a macro, using these symbol definitions. All parameterized type
declarations, function definitions, and object declarations in the package are expanded to form
normal nonparameterized instantiations.

17.3.2

Preprocessor Generics in C

Exhibit 17.16 shows how one might use C to write code for a generic sorted-linked-list type.
Such a package would contain definitions for several linked-list functions, including insert_item,
delete_item, and several others. For simplicity, only one of these functions is shown here. (The
identifiers written in uppercase are symbols that will be defined and instantiated by the preprocessor.)
The generic function find_item was written in terms of the generic type names and the generic
functions GREATHAN and EQUALS, which must be defined before expansion. The functions of dereference (*), selection (->), and assignment (=) are also used, but these are defined by the nongeneric
part of the type definition. Nothing in the function code depends on the particular type to which
ELEMENT will be bound; an ELEMENT could be a single character or a lengthy record.
To use the generic list package in Exhibit 17.16, the user must include in the program definitions
of ELEMENT and of the associated type-dependent operations. In all, this is several lines of routine
and tedious code. A convenient way to automate the inclusion process is to put the code for
several commonly useful types into a header file and use conditional compilation to include only
the appropriate set of definitions.
A header file is given in Exhibit 17.17 that would permit the generic linked-list module to be
used for linked lists of character strings or of integers. If some other type of data were to be used,
a group of similar definitions would have to be written for it and included in another conditional
clause. To use this header file, the programmer must type one of the following pairs of lines at the
top of the code:
For lists of strings
#include "lists.h"
#define ALPHA 1

For lists of integers

#include "lists.h"
#define INTEGER 1

532

CHAPTER 17. GENERICS

Exhibit 17.16. Definition of a generic domain in C.

Below are type declarations for a generic type sorted linked list and one function whose
domain is this type. Each list cell consists of one data item and a pointer to the next list cell. The
generic symbols are written in uppercase letters and will be defined as macros by the programmer
and expanded by the C preprocessor before actual compilation. Exhibit 17.18 shows the result of
expanding this code with one set of symbol definitions shown in Exhibit 17.17.
typedef struct cell {ELEMENT data; struct cell *next;} CELL;
typedef CELL * LIST;
/*-------------------------------------------------------------*/
int find_item(ELEMENT find, LIST head, LIST *scan, LIST *prior)
{
scan = head;
prior = NULL;
while ( GREATHAN( find, (*scan)->data))
{
*prior = *scan;
*scan = (*scan)->next;
}
return EQUALS( find, (*scan)->data );
}

The value of the #define symbol is tested by the preprocessors #if command, which triggers
inclusion or exclusion of the dependent definitions. During preprocessing, the generic package is
converted to a set of ordinary declarations and definitions which define and process a specific type.
Exhibit 17.18 shows the ordinary code that results from expanding these generic definitions for
type argument char*.
Ada Generic Packages
We can implement an ADT in Ada as a generic package. A package is a module that contains type,
data, and/or function declarations. 6 An Ada generic is a parameterized template which can be
instantiated to create a subprogram or a package. We will examine a simple generic subprogram
first, then tackle the problem of combining a generic with a package.
Parameters to a generic subprogram can be types, array lengths, or function names. The generic
parameters are used as follows:
The integer parameters to the package are used as array bounds in parameterized type expressions within the package.
6

Ada packages were described in Chapter 16.

17.3. PARAMETERIZED GENERIC DOMAINS

533

Exhibit 17.17. Using a generic type in C.

This is a header file that would permit the generic linked-list module to be used for linked lists
of character strings or of integers.
/* The function "strcmp" performs alphabetic comparison of two strings, */
/* "strcpy" copies a string value into a string variable. */
/* \377 is the octal code for the largest 1-byte character. */
#if ALPHA
#define
#define
#define
#define
#define
#define
#elif SHORT
#define
#define
#define
#define
#define
#define
#endif

ELEMENT char *
LESSTHAN(x, y)
GREATHAN(x, y)
EQUALS(x, y)
ASSIGN(x, y)
MAXVAL

( (strcmp((x),(y)) <0 )
( (strcmp((x),(y)) >0 )
( (strcmp((x),(y)) == 0 ))
( strcpy((x),(y)) )
\377 /* Octal for char code 255.*/

ELEMENT short int

LESSTHAN(x, y) ( (x) <(y) )
GREATHAN(x, y) ( (x) >(y) )
EQUALS(x, y)
( (x)==(y) )
ASSIGN(x, y)
( (x)=(y) )
MAXVAL
0x7fff

The type parameters are also used in parameterized type expressions. If the data structure
within the package uses two interdependent types, such as cell and cell_pointer, both
must be passed as parameters.
A generic must have one functional parameter for each representation-dependent function or
operation that is used in the body of the generic.
Exhibit 17.19 gives an Ada generic definition for addition on length-three numeric vectors. Such
a definition might be part of a generic package and be included in a library. Let us examine each
part of the generic code in Exhibit 17.19 so that we may see how these parts work together.
We begin definition of a generic subprogram by declaring the generic parameters (written here
in capital letters). Line (a) declares a dummy type name, NUMBER. The predefined generic domain
named private is used for any type parameter for which assignment and tests for equality are
defined. Thus, this generic definition can be instantiated with arrays of integer, float, or any
other type which permits assignment and tests for equality.

534

CHAPTER 17. GENERICS

Exhibit 17.18. Specific code results from preprocessing.

This function is the result of including the header file in Exhibit 17.17 with the symbol ALPHA
defined to be 1, and expanding the parameterized code in Exhibit 17.16.
typedef struct cell {char * data; struct cell *next;} CELL;
typedef CELL * LIST;
/*-------------------------------------------------------------*/
int find_item(char * find, LIST head, LIST *scan, LIST *prior)
{
scan = head;
prior = NULL;
while ( strcmp( (find),((*scan)->data) ) > 0 ) {
prior = *scan;
scan = (*scan)->next;
}
return (strcmp( (find), ((*scan)->data) ) == 0);
}

Exhibit 17.19. A parameterized domain in Ada: vector of numbers.

Step 1 is definition of the generic template:
generic
type NUMBER is private;
type VECTOR is array (1 .. 3) of NUMBER;
with function PLUS (x,y: NUMBER) return NUMBER;
procedure vector_add (v1, v2: in VECTOR; v3: out VECTOR);

-----

(a)
(b)
(c)
(d)

procedure vector_add (v1, v2: in VECTOR; v3: out VECTOR) is

-- (e)
begin
for J in 1..3 loop v3(J) := PLUS( v1(J), v2(J) ) end loop;
end vector_add;

17.3. PARAMETERIZED GENERIC DOMAINS

535

Ada provides a very small collection of predefined generic domains. In addition to private,
generic domains are defined for integer types, floating-point types, fixed-point types, finite types,
and pointer types, and for the completely general type with no known properties. Ada programmers
are not allowed to define their own generic domains; each generic type parameter must be declared
to belong to one of the predefined domains. The programmer, therefore, selects one that comes
closest to the properties of his or her own applications domains. We use the type NUMBER in this
example, which is a super-domain of all the numeric generic domains that are predefined. We,
therefore, declare NUMBER as private.
A second type parameter, VECTOR, is declared on line (b). Even though VECTOR is defined in
terms of NUMBER, it must be passed as a separate parameter. This certainly makes the preprocessor
easier to implement but is annoying for the programmer.
The code of the generic routine cannot just use any operations except assignment and equalitycomparison, since these are the only functions that are guaranteed for the domain private. Often
this is no problem, as the code of many functions relies only on these two basic operations. For
example, only assignment is needed to define the push and pop functions for stacks. However,
the code for vector_add uses both assignment and + for NUMBERs. To do so, + must be declared
as a parameter (line c).
The generic procedure header must be declared before the actual definition is given. Line (d)
uses the parameter names to declare the function named vector_add. There are three parameters,
two vectors to add up, and one to receive the answer. On line (e) we finally begin definition of the
actual procedure. The syntax is the same as an ordinary nongeneric procedure, except that it uses
the generic type parameter(s).
A generic procedure or package must be instantiated before the code can be compiled. Prior to
instantiation, the programmer must write a specific type definition for each type parameter and a
specific function method for each of the packages functional parameters. Finally, the programmer
must write an instantiation command using these predefined specific objects as arguments. The
instantiation call is given as part of a declaration for the name that will be bound to the instantiated
procedure or package. This name must be a new, unique name. To instantiate a generic procedure
or package we write:
procedure hprocedure namei is new hgeneric namei ( hspecific argument listi);
package hpackage namei is new hgeneric namei ( hspecific argument listi);
During the instantiation process, the compiler substitutes the actual arguments supplied by
the programmer for the generic parameter names and expands the template into ordinary, fully
specific, compilable source code. This is then compiled and bound to a unique function name
supplied by the programmer. This name refers to exactly one procedure or package and is not
generic or ambiguous at all. If the programmer wishes to instantiate the same package twice, she
or he must supply two names for the results.
To instantiate the vector_add template we must first define any new specific types and functions
needed for the parameters PLUS, NUMBER, and VECTOR. For example, to create a vector of floats, type
float and floating-point addition are already defined, but we need to declare a new type consisting

536

CHAPTER 17. GENERICS

Exhibit 17.20. Instantiating a generic template in Ada.

type real_vec is array (1 .. 3) of float;
type int_vec is array (1 .. 3) of integer;

-- (f)

procedure int_vector_add
-- (g)
is new vector_add(NUMBER=>integer, VECTOR=>int_vec, PLUS =>"+");
procedure real_vector_add
-- (h)
is new vector_add(NUMBER=>float, VECTOR=>real_vec, PLUS =>"+");

of an array of three floats. Then we write an is new declaration containing a call on the template
with these actual arguments.
In Exhibit 17.20, we define two length-three numeric array typesint_vec and real_vec. We
use each of these types to instantiate the generic from Exhibit 17.19. In line (g) we instantiate
the vector_add template. We indicate that the type parameter NUMBER is to be replaced by the
type integer, VECTOR is to become int_vec, and PLUS is the standard +. The translator will
make these substitutions and generate ordinary, compilable, nongeneric code for a procedure named
int_vector_add.
We repeat the instantiation process in line (h) with types float and real_vec, producing
the nongeneric procedure named real_vector_add. Now we have two ordinary procedures that
are alike except for their names and the types of their parameters. The object code will contain
translations of both copies. Note that no definition of the function argument was given; when this
is done it defaults to the definition (if any) for that symbol in the context surrounding the generic
definition.
A Generic Package. We have described both packages and generics in Ada. The normal way
to use both generics and packages is by combining thema package definition is placed inside a
generic declaration, producing a generic package. Such a package has parameterized public and
private parts that can be instantiated with various component types and array lengths to produce
code tailored to an individual application. The elements that must be present in an Ada generic
definition are:
1. The generic parameter declarations.
2. The package header, containing declarations of the public symbols
3. The package body, containing definitions of public and private symbols.
As an example, we show in Exhibit 17.21 how the code for the stack package in Exhibit 16.6
can be generalized to a parameterized generic module. First, the entire package is nested within a
generic declaration. Then all references to the base type of the stack and the length of the stack
are replaced by references to the generic parameters.

537

17.3. PARAMETERIZED GENERIC DOMAINS

Exhibit 17.21. Declaration of a generic stack package in Ada.

generic
MAX: POSITIVE;
-- A positive integer.
type ELEMENT is private;
-- Type of a generic parameter.
package STACK is
-- Declare functions in package.
function PUSH(X: ELEMENT);
function POP return ELEMENT;
function TOP return ELEMENT;
package body STACK is
-- Functions, variables for package.
stk:array(1..MAX) of ELEMENT;
tos: INTEGER range 0..MAX;
function PUSH(X: ELEMENT) is
begin
if tos<MAX then
tos:=tos+1; stk(tos):= X;
return true;
else return false;
end if
end PUSH;
function TOP return ELEMENT is
begin
if tos >0 then return stk(tos);
else raise STK_FULL;
-- Exception error condition.
end TOP;

end if;

function POP return ELEMENT is

ans: ELEMENT;
begin
ans:= TOP; tos:=tos-1;
return ans;
end POP;
begin
tos:=0;
end STACK;

-- Initialization for this package.

538

CHAPTER 17. GENERICS

Exhibit 17.22. Instances of the stack package.

In writing a program to do a precedence parse and code generation for arithmetic expressions, one
needs to use two stacks: a stack of tokens and a stack of expression trees. Here we instantiate
the generic stack package from Exhibit 17.21 twice to create these two stacks. For simplicity here,
assume that tokens are represented by single characters, and that the function named do_error is
the programmers own error handler and is defined previously.
type TOKEN is CHARACTER;
type TREE is access CELL;
type CELL is record
data: TOKEN data;
lson: TREE;
rson: TREE;
end record;
declare
package tok_s is new STACK(20, TOKEN);
package tree_s is new STACK(60, TREE);
begin
if not tok_s.PUSH(#)
-- Push beginning-of-line symbol.
then do_error("Stack full.") -- Check error return.
endif;
...

We can instantiate this package, as shown in Exhibit 17.22, to create multiple stack packages,
each with its own storage and its own copy of the code for each function. Within this code, calls
on PUSH, POP, and TOP will be compiled with meanings for the fetch and store operations that are
appropriate for the actual type of the type-argument. Note that we give a name to each specific
package in the instantiation command. When we call the PUSH that belongs to a package, the name
of the package is used, with .PUSH, to denote the proper method for PUSH. The equivalent of the
stack package in Exhibit 16.6 could be created by instantiating STACK with the parameters (10,
INTEGER).
Evaluation
An abstract data type (ADT) is a collection of types, functions, and, possibly, objects that express
a generic process on generic data. An ADT can be defined in C by using the preprocessor and in Ada

17.3. PARAMETERIZED GENERIC DOMAINS

539

by using a generic package. Adas generic preprocessor is very much like the C macro preprocessor. 7
Thus the capabilities of Ada generics are very similar to those in C.
These preprocessor generics are relatively easy to implement and make real progress toward
the goal of creating practical, flexible code libraries. However, they are far more limited than the
support for generic types provided in various object-oriented languages. The primary limitation is
that final binding of the type parameters happens at precompile time. Thus all type flexibility is
lost before compilation begins. If a program deals with a domain that has two or more representations, there must be two or more instantiations of the package, and two or more names for those
instantiations.
It is much more difficult to support code with run-time generic flexibility. Most languages
which do so are translated by interpreters, not compilers. To support runtime generics, the type of
each object must be known at run-time, and the translator must include a sophisticated methoddispatcher which examines those types at run time and selects an appropriate method. We will
discuss run-time dispatching in Chapter 18.

Exercises
1. Contrast generic domains and specific domains.
2. Describe the four ways in which a domain may be considered generic.
3. What are the problems involved in the implementation of generics?
4. What is a virtual function? What is its role?
5. What are the universal generic functions? Explain.
6. What semantic problems are created by the generic function?
7. What is a method? What is dispatching the call?
8. Why must functions such as push and pop be defined by separate methods for the species
of an ad hoc generic domain?
9. Why is generic behavior considered limited in older languages?
10. What is a union data type? What are its limitations?
11. What is an overloaded name? What are the dangers involved in overloading?
12. How do built-in generic functions interact with rules for type coercion?
7

There is one major difference in name scoping. Unbound names in a C macro will be interpreted in the context
of the macro call. In contrast, unbound names in an Ada generic function will be interpreted in the lexical context of
the module in which the generic was defined, not the module containing the instantiation request.

540

CHAPTER 17. GENERICS

13. How do Adas predefined operators limit mixed-type operations? How does Ada compensate?
14. What is a flexible array? Why is a function with a flexible array parameter considered generic?
15. What is a parameterized domain? A parameterized type expression?
16. What is the difference between an integer-parameterized domain and a type-parameterized
domain? Why do some languages support the former but not the latter?
17. What is the role of the preprocessor in relating the types in library definitions to the types
in a userss program?
18. What is a generic package in Ada? How are generic packages related to ADTs?

Chapter 18

Dispatching with Inheritance

Overview
Types have traditionally been used for the two disparate purposes of describing storage
and describing semantics, and this dual use has led to compromises in the power of
the language to express type relationships. One approach to the conflict generated by
the two needs is to eliminate or subordinate one use. If type checking for function
applicability is not done, the language never stands in the way of the programmer,
never prevents programmers from doing things they know are sensible or exploiting
relationships they know exist. This approach is taken in C. Types are used at the
bottom level to allocate, access, and encode but are not used to describe the semantics
of classes of objects.
On the other hand, we have languages such as Smalltalk where all objects are represented as pointers. In Smalltalk the type (class) is used primarily to control function
application, and data representation is simplified into near uniformity.
Future languages will separate the storage mapping of an object and its domain membership, permitting them to be defined separately. Doing so will enable us to retain
the safety of type checking within a language that can exploit domain relationships. A
mode graph gives us a systematic way of representing domain relationships and type
conversions.
The goal is a language in which the programmer can manipulate abstractions of both
objects and functions. This requires that we be able to describe objects and related
classes of objects, actions, and variants of those actions. Then the translator must use
that information to dispatch functions and coerce arguments.
Implementation of these mechanisms opens up new problems and leads to the need to
541

542

CHAPTER 18. DISPATCHING WITH INHERITANCE

define parameter domains using type expressions with constraints, rather than simple
type constants. The semantic basis of a generic language must include a general typededuction or type-calculation system.

18.1

Representing Domain Relationships

In the familiar strongly typed languages, each type declaration creates a new domain, and it is
not possible to define two structurally dissimilar types to represent related external domains, then
further declare and use the relationship between them. In these languages, types are used both
to define the physical representation of an external domain and to define the semantic properties
(function applicability) of program objects. Some domain relationships are built into programming
languages. For example, most languages have a few conversion routines that can be used by the
compiler, when needed, to make sense of the code. However, most languages do not provide a way
to express relationships among user-defined domains.
Sometimes a programmers newly defined domains are unrelated to other domains used in a
program. In this case it is not a burden to define a new set of functions for each newly defined
domain. However, a programmers domains are often related and share common properties with
each other. Some functions might operate correctly on objects from a variety of related external
domains. In the traditional strongly typed language, the programmer must represent all related
external domains as variants of one type if they are to be processed by the same set of functions.
Generic functions and domains have been developed to solve this dilemma: they permit each
variation of a domain to be defined as its own subdomain, and then the relationships among the
subdomains are separately declared and used by the translator.
The class hierarchies in the common object-oriented languages (Smalltalk, C++) are treestructured. The only class relationship supported is the one created by deriving one class from
another, and each class can only be derived from one parent. Generalized graphs of classes are not
possible in these languages because the objects of a derived class are represented by adding fields
to the record type used to represent the parent class. But it is possible to use virtual functions, not
class derivation, as the basis for declaring class relationships, and once we do this, we can build a
generalized graph of classes.

18.1.1

The Mode Graph and the Dispatcher

In this discussion, we will use a structure called a mode graph to discuss how both domain relationships and dispatching can be generalized. We call one node in the graph a mode, and diagram it
as a round-cornered box, like the box labeled Number in Exhibit 18.1. The contents of a mode
box vary according to the type of the mode: generic mode boxes contain a list of virtual functions,
specific mode boxes contain a type object. Methods defined for a mode are sometimes listed near

543

18.1. REPRESENTING DOMAIN RELATIONSHIPS

Exhibit 18.1. Mode graph for the generic domain Number.

Number
Promises:
Const:
+(Number, Number) One
-(Number)
Zero
*(Number, Number)
/(Number, Number)

complex

record 8 bytes
rp, ip
real, imag

IS

IS
integer

real
primitive 4 bytes

primitive 2 bytes

CONV
i_to_r

imag
represented

Methods:
+(complex, complex)
-(complex):complex
*(complex, complex)
/(complex, complex)
One
Zero

4 bytes Methods:
+(imag, imag)
real
*(imag, imag)
*(imag, real)
-(imag)

REP

it. We will consider typed modes, parameterized modes, subdomain modes (or submodes), generic
modes, and representation modes. As we discuss modes, we will point out the aspects of this mode
graph that have been incorporated into modern languages.
The dispatcher is a process built into a generic programming language that uses the information
in the mode graph to dispatch methods for function calls. You may visualize it as a mouse that
crawls along a maze of tunnels (the mode graph) looking for a method definition that will satisfy
the call. Depending on the language, the dispatcher has more or less information available to it,
and has more or less freedom to move from one domain in the graph to another, related domain.
The job of dispatching potentially has two components: work that can be done at compile
time because all relevant information is available then, and work that must be deferred until run
time. We will distinguish these two job components by talking about the compile-time dispatcher
and the run-time dispatcher. Traditional languages have only a primitive compile-time dispatcher.
Modern languages, including object-oriented languages and functional languages, have an extensible
compile-time dispatcher and a run-time dispatcher.
Typed modes. A typed mode is used in the mode graph to represent nonpolymorphic types,
such as integer or a record type. Every typed mode has an associated type object. 1 Actual objects
all belong to some typed mode, and all function translation and dispatching must start with the
typed mode of the argument.2
1
2

Review Chapter 14 if necessary.

Just as object-oriented dispatching starts with the class of the implied argument.

544

CHAPTER 18. DISPATCHING WITH INHERITANCE

Generic Modes and Promises. A generic mode represents an abstract domain. Its meaning
is defined by a set of virtual functions, rather than by a representation. A generic mode is similar
to a C++ base class but differs in two important respects. First, a C++ class has an associated
representation, from which other representations may be derived; further, derivation is the only way
to create an additional representation of a class. In contrast, a generic mode has no representation
when it is first created, but typed modes and other generic modes may be attached to it later.
The typed modes are representations of the generic. They might be structurally related, like the
instances of a parameterized domain, or might have an arbitrary, ad hoc relationship to each other,
like integer has to real.
In an object-oriented class hierarchy, both the fields of the representation record and the functions that operate on the record are inherited. It is this inherited relationship that guarantees
that the code will be semantically meaningful when it is executed. In a mode graph, the fields
of one mode are not inherited by the related modes, and other means must be provided to assure
that the operation of the program will be semantically meaningful. Virtual functions can provide
this assurance. The box for each generic mode will contain a set of virtual function declarations
which we will call promises. The promises of a mode will be used, like virtual functions in C++,
to describe the implementation-dependent processes that characterize the semantics of the mode.
Any functions defined for the generic mode itself must be written only in terms of its promises.
Exhibit 18.1 shows a diagram for a generic mode Number. In it, promises state that numbers must
have four defined arithmetic operations and also have constants defined for the arithmetic identity
values, one and zero.
The generic facility in Ada has six predefined modes which can be used to declare the types of
arguments to generic packages. For example, the mode private promises only that assignment
and test for equality will be defined. Some other Ada modes are range <>, meaning some integer
type, and digits <>, meaning some floating-point type. In our terminology, short integers,
unsigned integers, long integers, and subranges of integers are all submodes of the predefined Ada
mode range <>. In turn, range <> is a submode of the mode (<>) (which includes any
discrete type).
Ada would be a better language if it provided a way for a programmer to declare new modes
with different promises. Having only six modes available, as in Ada, is severely limiting. If the
properties of a programmers domain do not happen to match one of the predefined modes, the
programmer must use some other mode with fewer promises, or limited private which has no
promises at all. If a function is needed which should be defined for the desired a bstract mode,
that function must be passed as an argument in the call that instantiates the package. This is a
nuisance, at best, and half defeats the purpose of having generics.
Submodes. A mode graph can be used to represent a variety of domain relationships by introducing varied kinds of links from one mode to another. The links corresponding to Ada modes and
to the derived from relationship in C++ are called submode links and are labeled by the word
IS. A submode is one representation of its base mode and inherits function methods from it. In

18.1. REPRESENTING DOMAIN RELATIONSHIPS

545

Exhibit 18.1, the modes integer and real are shown as submodes of Number.
An ad hoc generic domain includes two or more specific domains that are related by their meaning rather than by their structure. Physically, the two domains might be represented differently,
but logically, they have a common intent and common functionality. An IS link expresses the
subdomain relationship that exists between a generic domain and its representations.
In some languages, operations are defined only for specific domains, and the IS relationship is
not exploited. In contrast, consider APL, whose designers expected programmers to have infrequent
concern about the representation of numbers and frequent use for the semantic relationship among
the representations in use. The generic domain Number, not one of its specific representations,
was made primitive. Further, the relationship between the generic domain and its subdomain
was made explicit and flexible.3 Many APL operations are defined for numbers, rather than for
either integers or reals. Others, which require the semantics of a discrete type, are defined only for
integers.
Declaring and Using Mode Relationships. The set of promises on a mode provide a semantically sound criterion for being a submode. Thus if modes, promises, and links were all declarable,
we could declare M 0 to be a submode of M as soon as all promised methods were defined for
M 0 . In a generalized graph structure, a mode might have several supermodes, so these methods
might be defined directly for M 0 , or might be inherited by M 0 down some other submode link. In
our example, there is no problem with declaring that integer and real are submodes of Number.
But before the submode link for complex can be declared, the promised arithmetic functions and
constants must be defined. Once this is done, we can declare that complex IS Number. Now the
submode complex will inherit the functions defined for Number, just as the predefined submodes
integer and real do.
The promises defined for a generic mode guarantee that the promised functions will be defined
for every specific instance of the generic mode. This permits us to write representation-independent
functions whose domain is the generic mode. When one of those functions must access a representation, it calls a promised function.
When we declare that a specific mode IS a submode of a generic, all functions defined on the
generic mode are inherited by the submode. That means that a programmer can call a function
F, defined for the generic mode G, on an argument from the submode, S. If F calls one of the
promised functions, the dispatcher must find the appropriate method to carry out the promise. To
do this, it must look for and dispatch the method that is applicable to the submode S. Often, this
dispatching can be done at compile time, but in the most general case, the mode S is not known
until run time, and dispatching must be done at the last minute.
Instantiation Modes. Parameterized generic domains were discussed in Chapter 17, Section
17.3. The C and Ada implementations discussed there both require instantiation of the generic
3

In APL, a number is an integer if and only if it differs from an integral value by no more than the currently
defined comparison tolerance.

546

CHAPTER 18. DISPATCHING WITH INHERITANCE

Exhibit 18.2. Diagram of an instantiated mode.

VECTOR(T: type)
arraytype

3*sizeof T bytes
1..3
sizeof T bytes @
T

BIND
T is complex

VECTOR(complex)
arraytype

24 bytes
1..3
8 bytes @
complex

mode and its dependent functions at compile time. The result is that a set of specific types
and functions are generated and compiled, and the fact that these were derived from a generic
is forgotten. The mode graph can be used to represent the relationship between a parameterized
generic domain and an instantiation of that domain. We will diagram this relationship as a BIND
link, labeled by the bindings for the generic parameters that were specified in the instantiation call.
A BIND link is illustrated in Exhibit 18.2. We will explore the question of delaying binding until
run time and consider the advantages and problems caused by that delay.
Conversions and Coercions. Not every one-argument function is a conversion or a cast. Casts
arise only from mapped domains and are simply part of the strong typing system of the language.
Conversions are defined ad hoc, and only the programmer can know which of his or her functions
are legitimate conversions. The important aspect of a conversion is that it must not change the
meaning of an object. Defining a conversion can be done in any reasonable language, but most
languages do not provide a way to tell the compiler that the function is, indeed, a conversion.
C++ permits the programmer to define and declare conversion functions (called constructors),
and languages that are now at the stage of research will incorporate more general provisions for
declaring conversion functions. Once the compiler knows about the conversion, its dispatcher can
use it to coerce arguments, just as the built-in conversions are used.
Without some provision for coercion, domains with multiple representations are unwieldy and
almost impractical to use. We can demonstrate the problem in Ada. Ada permits the programmer
to define new domains, then define new methods for the existing generic operators that operate on
the new domains. The definitions of + and * for the new domain imag are an example [Exhibit
17.9]. But even though we can extend the intrinsic operators, it is a tedious job to achieve the ease
of use of mixed-type arithmetic that is built into Pascal, FORTRAN, and C, because Ada lacks any
provision for type coercion. To implement mixed-type arithmetic, the programmer must include
definitions for every combination of operator and operand types desired [Exhibit 18.3]. Thus if F
binary functions are to be defined over a generic domain with R representations, the number of
method definitions needed would be F R 2 !
FORTRAN actually supports these four numeric types and supports mixed-type arithmetic on

547

18.1. REPRESENTING DOMAIN RELATIONSHIPS

Exhibit 18.3. Exponential explosion of method definitions.

Assume that a programmer is using four representations of the domain Number: integer, float,
double, and complex. To implement full mixed-type arithmetic for these four types, without type
coercion, 16 method definitions must exist for each operator. Thus for + we would need:
+(int,
+(int,
+(int,
+(int,

int)
float)
double)
complex)

+(float,
+(float,
+(float,
+(float,

int)
float)
double)
complex)

+(double,
+(double,
+(double,
+(double,

int)
float)
double)
complex)

+(complex,
+(complex,
+(complex,
+(complex,

int)
float)
double)
complex)

For four arithmetic operators, we would need 64 methods!

all combinations of operands. However, FORTRAN does not have 64 method definitions for these
functions. It has only 16 methods (four each for four operators), plus the conversion functions
listed in Exhibit 18.4.
Type coercion between alternate representations of the same external domain reduces the number of function method definitions necessary. That is why PL/1 provided coercions from any basic
type to any other, directly or through a series of changes. This coercion facility was completely
general. Unfortunately, it masked real type errors by changing the type of an erroneous argument
to whatever would seem to work. Sometimes this was beneficial, but sometimes it made semantic
hash out of the data. It is obviously impossible to create a fixed set of conversions that will handle
all mode relationships for all time in a semantically valid manner. When PL/1 tried to anticipate
all future conversion needs, and used these conversions to coerce arguments, it resulted, all too
often, in converting an object to something else with an unrelated meaning. Worse, coercion errors
were especially difficult to track down because the semantically erroneous action was not caused
by anything explicitly written in the program, but by some subtle and poorly understood type
conversion built into the system.
The only adequate solution is to permit programmers to define their own conversion functions
and to declare them in such a way that the compiler can use them for coercion. It would then
be the programmers responsibility to define only semantically valid conversions. We will diagram
a conversion function, F , that converts from mode M 1 to mode M 2, as a link from M 1 to M 2,
labeled with CONV and the function name. In Exhibit 18.1, a conversion link is shown from integer

Exhibit 18.4. Coercions built into FORTRAN.

Result Type
integer
real
double

Functions Defined
INT(REAL), NINT(REAL), INT(DOUBLE)
REAL(INT), REAL(DOUBLE), ABS(CMPLX)
DBLE(INT), DBLE(REAL)

548

CHAPTER 18. DISPATCHING WITH INHERITANCE

to real, for a function named i_to_r. This means that i_to_r can be used explicitly or through
coercion to convert an integer to a real.
Represented Modes, Casts, and Coercions. Both conversion functions and casting functions
take a single argument from one domain and return a value from a different domain. When the
argument and result are both representations of the same object in the same external domain, the
function is a conversion.4 When the argument type was defined by mapping it onto the result type,
or vice versa, the function is a cast.
A cast represents the relationships between two domains. Casts are used, in some languages,
by function dispatchers. Thus we want to represent them in our mode graph. A cast relationship
is created when the programmer specifies that one mode is to be represented by another. We show
it in the mode graph by noting that the type of a mode is a represented type. For example, there
is a REP link from imag to real in Exhibit 18.1, because real numbers are being used to represent
type imaginary.
Although definition and use of a conversion is often similar in syntax to a cast, the difference is
very important: a conversion preserves semantics, a cast changes them. Therefore, a conversion is
safe for coercion, a cast is not. A cast should never be applied unless the programmer explicitly
directs it, and even then, the uses should be restricted to the private parts of classes or their
equivalent.
Dispatching
We can view a function as a collection of methods, and dispatching as the problem of choosing the
best method for each call. Simply put, the best method is the one whose parameter domains make
the best match for the types of the actual arguments. In older languages, only a few predefined
operators had more than one defining method, and the best match was easy to define; a method
is best if:
Its parameter types match the argument types exactly, or,
No method matches the argument types exactly but for some method, every nonmatching
argument can be coerced, by a built-in conversion function, to match exactly.
The dispatching problem is much more complex in a modern language. First, user-defined
functions, as well as predefined operators, have multiple defining methods, so the dispatcher must
be written in a more general way. Second, user domains have declarable relationships to other
user domains; where a subdomain or an instantiation relationship exists, it should be used by the
dispatcher to find methods that can be inherited. Third, user-defined conversion functions must be
considered and integrated, as coercions, into the dispatching process.
A mode graph gives us a systematic way to represent domain relationships and type conversions
and, thus, forms a framework in which we can look at the dispatching problem. We will use the
4

Review Chapter 15, Section 15.5, if necessary.

18.2. SUBDOMAINS AND CLASS HIERARCHIES.

549

Exhibit 18.5. Defining a subdomain in Pascal.

TYPE fingers = 1..10;
VAR f1, f2: fingers;
...
readln(f1);
f2 := f1 + 5;

{ A subrange type declaration.}

{ Integer functions are defined for fingers. }

{ A legal computation on subdomain fingers. }

The type declaration for finger defines a subdomain of the integers; integer operations such
as + and readln are defined over this subdomain.

concept of a mode graph to discuss dispatching in C++, Miranda, and Aleph, which is a newly
developed language that provides full support for generics.

18.2

Subdomains and Class Hierarchies.

18.2.1

Subrange Types

A domain D 0 is a subdomain of a superdomain D if all the elements of D 0 also belong to the

domain D. Many programming languages allow the programmer to declare a new domain that is
a subdomain. In Pascal and Ada, the domain corresponding to the subrange type is a subdomain
of the domain formed by the base type [Exhibit 18.5].
The domain formed by a subrange type is the easiest kind of subdomain to implement because
elements of the subrange type and the base type have the same representation. Values from the
subrange type may be freely stored in base-type variables; base-type values may be stored in
subrange variables if they are within the defined limits of the subrange. Subrange and base-type
values may be compared directly for equality.
Functions defined for the superdomain are inherited by the subdomain; that is, any function
defined for D is also applicable to objects of subdomain D 0 . In mathematics, the integers are a
subdomain of the reals which are, in turn, a subdomain of the complex numbers. Thus complex
operations are applicable to reals and real operations are applicable to integers. With Ada and
Pascal subrange types, functions defined for the base type are inherited by (applicable to) objects
in the subdomain, but not vice versa.
Any function defined for the base type could accept and process arguments from the subrange
type. We say that the subrange type can inherit these functions from the base type. No special
checking or conversion is needed to make a subrange-type argument appropriate for the basetype function. Conversely, functions defined for a subrange type can process any values from the
base type that are within meaningful limits. Thus the base type can inherit functions from the
subdomain, and those functions will be meaningful part of the time.

550

CHAPTER 18. DISPATCHING WITH INHERITANCE

Exhibit 18.6. Diagram of a domain hierarchy.

Domains are diagrammed above their subdomains. As we travel down the links, more and more
specific information can be known about each subdomain.

employee

virtual:
compute_pay
print_list

hourly

method:
compute_pay

professional

salaried

officer

method:
compute_pay

manager

methods:
print_list
do_payroll

To make subdomains and these simple inheritance relationships work, a translator must include
run-time range checks to ensure that every program object stored in a subrange variable or processed
by a subrange function does actually belong to the subdomain.

18.2.2

Class Hierarchies

Smalltalk and the other object-oriented languages provide much broader support for subdomains.
They allow the programmer to define a new subdomain whose representation is semantically related
to the superdomain but is not structurally identical to it. These languages exploit the hierarchically
related nature of many external domains. Start with a classification system, such as that shown in
Exhibit 18.6. Lines connect each domain (above) to its subdomains (below). Less detail is known
about or relevant to the broad class at the top of the hierarchy, but as we move down the subdomain
links, more and more details can be specified. Each category in the diagram has its own specific
characteristics and inherits characteristics from the categories above it. These domains and their
relationships are described in Exhibit 18.7.
Altogether, we have defined five kinds of data records, where each one must contain all the
information of the domain above it, plus more. For example, a manager record must contain
fields for an annual salary, name, social security #, two codes, and a link, because a manager is a
salaried employee and is an employee. However, a managers record must also contain a pointer
to the list of employees managed. Thus although the record types that might implement these
domains are structurally different, they have semantic commonalities. A function that processes
employee data could also process manager data correctly if the data fields of the two types were
represented in a compatible order.

551

18.2. SUBDOMAINS AND CLASS HIERARCHIES.

Exhibit 18.7. A group of related domains.

Here are descriptions of the fields in the records for two kinds of employees. (These records
have a link field so that they can be organized into lists.) In a representation of these domains, the
record describing the superdomain (employee) is a subrecord of that describing the subdomain.
Domain
Employee

Superdomain

Information to be recorded
Name, social security #,
dept code, job code,
pointer to next employee

Functions defined
Print employee list
Print paycheck
Enter new employee

Hourly_emp

Employee

Pay rate, hours worked,

overtime, vacation days,
group number

Compute pay
Enter time card
Change data

Salaried_emp

Employee

Annual salary,
vacation days used

Compute pay
Enter vacation

Manager

Salaried_emp

List of employees managed

Print list of people managed

Add new employee to list
Do payroll

Professional
Officer

Salaried_emp
Salaried_emp

Group number
List of managers managed,
permitted to sign checks?

Permission predicate

An object-oriented language, such as Smalltalk or C++, permits the programmer to declare and
use this kind of domain/subdomain relationship to define a hierarchy of domains. C++ classes
were introduced in Chapter 16 as an implementation of abstract data types. Classes are also used
to implement generic domains with subdomains and function inheritance.
We will use C++ syntax and rules to explain how these class hierarchies work. Given a class,
G, we can derive a subclass G0 from G, such that G0 has all the members (data and/or functions)
of G, plus more. The C++ syntax for declaring a derived class is :
class

hnew-class-namei : [public] hbase-class-namei {

hnew-private-membersi

(a)
(b)

hnew-public-membersi

(c)

public:
};
When we use a derived-class declaration, a new class, G 0 , will be constructed that is a subclass
of the specified base class, G. If the keyword public is used, (line a) all the public parts of G
will also be public parts in G0 ; if omitted, the public parts of G will become private parts of G 0 .
If G has private members, an instance of G 0 will contain fields for these parts, but they are not
visible to the functions defined for G 0 and can only be accessed by functions defined for G (unless

552

CHAPTER 18. DISPATCHING WITH INHERITANCE

a friend declaration is included in G). This visibility rule supports modular data-hiding, which
is so important for achieving reliable systems.
In section (b), the private members of G 0 are listed. Included here are data fields that exist for
the subclass G0 but not for the superclass G, and functions (if any) that will be used only locally,
by other functions defined for G0 . Finally, additional public members of G 0 are listed in section (c).
These must include all basic functions needed to operate on the subclass. Additional classes may
be derived, similarly, from either G or G 0 .
Exhibit 18.8 shows the C++ declarations that would implement part of the domain hierarchy
from Exhibit 18.7. To finish implementation of this hierarchy, we need to supply definitions for the
remaining class, officer, and definitions for all the functions declared in all the classes. Definitions
for these functions can be placed anywhere in the program, but good style dictates that they should
be placed just after the class declaration. The functions that must be defined are:
employee::employee
employee::print_paycheck
employee::print_list
employee::compute_pay
salaried::compute_pay
salaried::take_vacation
hourly::hourly
hourly::compute_pay
hourly::record_time_card
manager::manager
manager::add_employee
manager::print_list
manager::do_payroll
Although these method definitions are outside the class declaration, the function methods themselves were declared within the class, and are class members with full access privileges. The placement of the function definition (before or after the end of the class declaration) makes no difference
in the semanticsthe only difference is practical; definitions inside the class are expanded in-line,
like macros, those outside are compiled as subroutines. C++ functions can be overloaded, that is,
a function name can have several defining methods, belonging to several classes. For this reason,
the full name of a function method must be used when we define it outside its class. Thus we must
write hourly::compute_pay() and salaried::compute_pay, not simply compute_pay.
Representation and Visibility. A program object is created by instantiating a class, and this
action is triggered either by a declaration or a new command. The result is a record with one
field for each of the variables (both public and private) within the class. A derived class will be
represented by the same type of record with fields added on the end. The names of the class
variables are like the names of the fields of a record, except that more complexity is involved
because of information-hiding.
In a Pascal record, all the fields have the same visibility; that is, if one part is accessible
in a given context, then all parts are accessible. But in a C++ class, the fields corresponding to
private members have visibility restricted to the class functions, while public members have broader
visibility. When we make a derived class, fields for private members of the parent class will be part
of the record and take up space, but the names of those fields will not be known within the derived
class, and those fields will be inaccessible to new functions. However, the public functions of the
base class become members of the derived class and can be used to manipulate these fields.

18.2. SUBDOMAINS AND CLASS HIERARCHIES.

553

Exhibit 18.8. Deriving classes in C++.

Below are declarations for some of the classes described in Exhibit 18.7. Exhibit 18.6 is a
diagram of the relationships among these classes. This code is discussed throughout the rest of this
chapter. We presume that card and pay_data are previously defined classes.
class employee {
char name[], soc_sec[13], *dept_code, *job_code;
public:
employee * link;
employee(int)
// The constructor; argument is name length.
void print_paycheck;
virtual pay_data compute_pay;
virtual void print_list;

};

void employee::print_empl()
{
cout << "Name: " << name << "\n\t" << soc_sec
<< "\tDept: " << dept_code << "\tJob: " << job_code
<< "\n";
}
// End of class employee.

class salaried : public employee {

int annual_salary, vacation_used;
public:
pay_data compute_pay();
void take_vacation();
};
// End of class salaried.
class manager : salaried {
employee* staff;
public:
void add_employee();
void print_list();
void do_payroll();
manager();
};

//The constructor function.

// End of class manager,

class hourly : public employee {

float pay_rate, hours_worked, overtime;
int vacation_used;
public:
pay_data compute_pay();
void record_time_card(card*);
hourly();
};
// End of class hourly.

554

CHAPTER 18. DISPATCHING WITH INHERITANCE

Limiting Visibility. It is possible to use a derived class to modify the visibility of members of an
existing class. Let us say that class G has some private members and some public members. There
is no way that deriving a class from G can affect the visibility of its private membersthey may
be accessed only within G. But the public members of G may be made private in a class derived
from G. To do this, omit the word public from the header of the derived class declaration.
Public members of the base class then become private members of the derived class. We say that
a derived module is opaque when it completely hides its base module. A half-and-half, or semiopaque, situation may also be achieved. When the keyword public is omitted, you may list the
names of selected members from the base class in the public part of the derived class. This does
not create an additional field in the derived class; it simply controls the visibility of the base-class
field. The syntax is:
class

hnew-class-namei : hbase-class-namei {
hnew-private-membersi

public:
hbase_class_namei::hmember_namei;
hnew-public-membersi
};
Type compatibility. If G0 has a public base class G, then an object of type pointer-to-G 0 can
be stored in a pointer variable of type G (pointer-to-G) without use of explicit type conversion.
For example, the class employee has a field named link of type employee*. Every instance of
class manager and hourly also has this field, because manager and hourly were derived publicly
from employee, and link is public in employee. The link field is able to store addresses of type
hourly* and manager* because these classes were derived from employee. Thus we may make a
linked list of hourly employees and store the head in a manager record.
This is a crucially important issue. The idea of a class hierarchy is that all variants of a base
class are semantically related. Even more, an instance of a derived class is an instance of the
base class, with some added information. Compatibility of pointer types throughout the levels of
a hierarchy is essential to implement this underlying semantic notion.

18.2.3

Virtual Functions in C++.

C++ contains a simple kind of support for virtual functions. A base class may contain a virtual
declaration, such as the print_list function in Exhibit 18.8. The intent is that a virtual function should be used where the method for carrying out some abstract process is implementationdependent. Let us use the term virtual class to mean a class that contains a virtual function or one
that is derived from another virtual class. All the classes in the employee example are virtual
classes because the base class contains one virtual function.
A virtual function must be defined for the base class in which it is declared, but it may be
redefined for any derived classes that need a different method. In our example [Exhibit 18.9] the
method for computing a paycheck is different for salaried and hourly people, and depends on data

18.2. SUBDOMAINS AND CLASS HIERARCHIES.

555

Exhibit 18.9. Calling a C++ function with two defining methods.

We define output function methods for the classes in the employee hierarchy established in
Exhibit 18.6. The method in the derived class calls both of the methods in the base class.
void employee::print_list() // Does not print head of list.
{
employee * scan;
for (scan=link; scan != NULL; scan=scan->link) print_empl();
}
void manager::print_list()
{
cout << "\n\nManager: ";
print_empl();
cout << "\nEmployees Supervised:\n";
employee::print_list();
}

that is specific to these classes. The function compute_pay is, therefore, declared as a virtual
function and defined separately for the two classes. Another application of virtual functions is to
allow a method to do some actions specific to its local class, then call the general function from
a higher class to complete the job. In our example, the print_list method for employee prints
some headings, then calls the base method, using its full name, employee::print_list.
If another class is derived from a derived class, the classes form a hierarchy. A virtual function
may be defined at several levels of a hierarchy. In this case, several methods for a virtual function
may all be applicable to instances of a class at the bottom of the hierarchy. In our employee
example, two methods are defined for print_list (in class manager and the base class, employee).
Both could be applied to an instance of class manager, but only the method in the base class could
be applied to instances of the classes salaried and employee.
If a function is called using its full name (class_name::function_name) there is no ambiguity,
and the call is translated in the ordinary manner. However, if a function is called without using its
class name, the call is ambiguous and must be resolved by the dispatcher. The same dispatch rule
can be used here as is used in the simple case: start at the bottom of the class hierarchy and look
up the hierarchy tree until a method is found for the function. Thus if the programmer wishes to
use a method that is not the closest one to its objects class, the full, qualified name of that method
must be written.
It is up to the compiler and the linker to make the necessary connections between the virtual
function and its methods, and to ensure that all those methods are consistently typed. Similarly,
it is up to the translator to dispatch the correct method for a call on a virtual function. This is
complicated because the method being called might not even exist when the call is compiled, and
the actual dispatching must be done at link time or at run time.

556

CHAPTER 18. DISPATCHING WITH INHERITANCE

To understand why dispatching must be delayed, consider the do_payroll function defined
for class manager in our employee class hierarchy [Exhibit 18.7]. Assume that do_payroll
calls print_paycheck to print out a paycheck for each employee on the managers staff list. By
inheritance, we can use this function to print a paycheck for any employee, and we can call it
from the class manager. However, assume that print_paycheck calls upon the virtual function
compute_pay to get the data needed for a check. Now, compute_pay is defined differently for hourly
and salaried employees, and a managers staff list will generally contain both kinds of employees.
Thus at run time a single function call written in the print_paycheck function must sometimes
dispatch hourly::compute_pay and sometimes salaried::compute_pay.
This decision cannot be made until each list element is processed. Thus the type of the argument
to print_paycheck must be examined at run time. Happily, once this type is known, the simple
dispatching rule still works. In traditional compilers (such as the typical C compiler) the type
information from the declarations is put into the symbol table, used at compile time, then discarded.
Little or no type information is carried over to run time. In order to dispatch a virtual function,
though, it is necessary to have this information at run time. Some version of the translators type
objects must exist then, and each object must include a type pointer.
For this reason, objects belonging to virtual classes in C++ are compiled differently. A type
field (a pointer to the type object for the objects class) is made part of every object in a virtual
class. The dispatcher examines this type field and selects the right method for an object at run
time. Thus we incur space overhead when we use a virtual class, and time overhead when we call a
virtual function.5 The benefits of virtual functions, however, outweigh these costs. Using function
inheritance and virtual functions, we can:
Avoid proliferation of nearly identical names for implementations of the same external process.
Avoid writing duplicate copies of code for related classes.
Write simple code with highly flexible run-time behavior.
Extend the benefits of strong typing to nonhomogeneous domains.
Create class hierarchies that are easy to extend when new, related data types must be defined.

18.2.4

Function Inheritance

The real importance of the class hierarchy is that it defines a system of classes with related semantics
on which a system of function inheritance can operate. Briefly, functions defined in any class become
members of every derived class and may be applied to instances of the derived class and called as
if they were local. We say that the derived class inherits the function from its base class.
Every function call has an implied argument, which we will call the object of the call. To
translate a call, the compiler must find and dispatch the correct method for that object. The
dispatching rule is this:
5

This overhead is incurred only for virtual classes. It does not reduce the efficiency of operations on nonvirtual
classes.

18.2. SUBDOMAINS AND CLASS HIERARCHIES.

557

If a method is defined in the same class as the object of the call, the translator will dispatch
that method.
Otherwise, move up the class hierarchy, one level at a time, looking for function-members
with the correct name. Dispatch the first method for the function that is encountered.
Unless the function is declared to be virtual in the base class, there should be exactly one
applicable method in the hierarchy. If no appropriate method is found, there is a type error.
Constructors in Derived Classes. A constructor function is atypical, since we do not ordinarily
call it explicitly. Whenever the translator allocates space for a new class instance, it calls the
constructor function for that class to initialize the new storage. However, to instantiate a derived
class which has a constructor, the translator must execute two constructor functionsfirst, the
constructor for the base class, then the constructor for the derived class. Thus the derived class
must inherit the constructor from its base class.
This leads to a real problem, since many constructor functions have parameters. The header line
of each constructor definition specifies what parameters it needs, but how can the constructor for
the derived class convey the right arguments to the base constructor? The solution is to expand the
syntax of the language to allow the programmer to supply argument lists for a series of constructor
functions. Arguments for the constructor of the derived class are given in the instantiation call.
Arguments for the constructor of the base class (or classes) are given in the definition of the
constructor for the derived class. The syntax is:
hderived_namei::hderived_namei ( hargument_list_for_derived_classi) :
( hargument_list for base class constructori )
hinitializers for base class membersi
{ hbody of derived constructori }
In the definition of the derived constructor, the programmer can use the base-classs argument
list to pass on the derived-class arguments or to supply the base class constructor with constant
arguments. If there are several nested derived classes which have constructors, a list of argument
lists (separated by commas) is given, with the list for the most basic class first. The constructor for
the base class will be executed before the constructor for each class derived from it. (If both classes
have destructors, the destructor for the derived class will be executed before that of the base class.)
Strengths and Limitations. The classes in C++ are immensely powerful and, when used well,
can reduce both programming errors and tedious, repetitive coding work. This language is a giant
step beyond standard Pascal. Several major semantic ideas are covered here that go beyond the
traditional languages. They give C++ much of its power. These are:
Controllable public/private data and functions.
Declarable class relationships.

558

CHAPTER 18. DISPATCHING WITH INHERITANCE

Functions with multiple defining methods.
Function inheritance.

By themselves, these semantic mechanisms are somewhat limited. We briefly list these limitations
below and consider solutions in Section 18.4.
These mechanisms do not address the problem of generic domains with representations that
are related in an ad hoc manner. The only domain relationship that can be modeled is that
between a record type and a longer record type with the same initial parts.
All hierarchies and all function inheritance are totally tree structured; no class can have
two parent-classes. However, some external generic domains naturally form graph-structured
relationships in which a domain inherits properties from two different directions.
During dispatching, the types of the arguments to a call are not considered. Only the class
of the object of the call is used to select a method.
C++ is still a superset of C, and, therefore, you can get around all the visibility rules of C++
by using pointers, which are semantically insecure types in C! The problem is that a pointer
to any type can be cast to any other pointer type. This can be used to gain access to private
data.

18.2.5

Programmer-Defined Conversions in C++

With some limitations, the C++ programmer is able to define conversions that the dispatcher
can use. A class definition normally contains an ordinary constructor function, which is invoked
automatically when a new storage object for the class is allocated. In addition, the class may
contain one or more one-argument methods with the same name as the constructor. Each one is
taken to be a conversion function and must create a value of the type of the class. As with any C++
constructor, the new value is returned by assigning values to some or all of the class membersan
explicit return statement is not used in a constructor. The programmer must make sure that any
constructor with one argument is a semantically valid conversion function from the type of the
argument to the type of the class, because the compiler will use these methods, whenever needed,
for coercion.
C++ can use constructor functions to coerce arguments in simple cases. In the class complex of
Exhibit 18.10, four constructor methods are defined. The first one constructs the complex number
zero and will be used to initialize every complex variable for which no other initializer is specified.
In the example, the variable cx is initialized using this method.
The second method constructs a complex value from real and imaginary components [Exhibit
18.11]. This will be used primarily to define the member functions for the class; it defines the
relationship between a complex number and its components. It can also be used to construct
initializers for complex objects, as in line (b), and to create complex objects during execution, as
in line (d).

18.2. SUBDOMAINS AND CLASS HIERARCHIES.

559

Exhibit 18.10. Conversion functions in C++.

class complex
{
rp: float;
ip: imag;
public:
complex(){ rp=0.0; ip=imag(0.0); }
complex(float f, imag i){rp=f; ip=i;}
complex(float f){rp=f; ip=0;}
complex(imag i){rp=0; ip=i;}
complex operator-() { return complex(-rp, -ip); }
friend complex operator+( complex c1, complex c2 );
friend complex operator*( complex c1, complex c2 );
friend complex operator-( complex c1 );
}
complex operator+( complex c1, complex c2 )
{
return complex( c1.rp+c2.rp, c1.ip+c2.ip ); }
complex operator*( complex c1, complex c2 )
{
return complex( c1.rp*c2.rp + c1.ip*c2.ip, c1.rp*c2.ip + c1.ip*c2.rp ); }

The last two methods are semantically valid conversion functions from types float and imag
to type complex. They can be called explicitly, but they will also be used by the translator to
coerce arguments in calls on complex operations. Lines (a) and (d) illustrate contexts where an
argument will be coerced. In line (a), the float number -1.6 is used as an initializer for an instance
of class imag (from Exhibit 16.10). This triggers a call on the conversion from float to imag that
was defined in class imag.
In line (d), the operator + is called to add a complex number and a float. This situation is
somewhat more complex. The dispatcher looks at this call on + and must find an appropriate
method for it. The dispatching rule in C++ is given in Exhibit 18.12. By this rule, the method
complex operator+(complex, complex) is finally selected, and the second operand, fl, is coerced
to type complex.
Programmer-Defined Casts in C++. In C++, as in C, no distinction is made between conversions and casts. We have seen that a C++ class can be used to implement a mapped domain with
semantic protection against misuse. But casts are required to implement the basic functions for the
new mapped domain, and casting is not automatically defined for such a class. In this situation,
the facility for programmer-defined conversions can be used to define an upward cast. To specify

560

CHAPTER 18. DISPATCHING WITH INHERITANCE

Exhibit 18.11. Coercion with a programmer-defined function.

main()
{
float fl=2.5;
imag im=-1.6;
complex cx;
complex dx = complex( 1.0, imag(1.0) );
cx = new complex( .1, im );
cx = cx + fl;

//(a)
// Initialized to zero.
//(b)
//(c)
//(d)

the downward cast, from the class type to a representing type, an operator function may be used.
This defines the target-type name as a conversion function that can be called explicitly using the
normal syntax for casts. Both kinds of programmer-defined casts are illustrated in Exhibit 16.10,
where their use is required to define the basic functions for the new mapped class.
It is important that the two casts were defined as private functions. This means that, within the
class, the representation of a class object can be manipulated, as it must be, to make the necessary
calculations. However, outside the class, the relationship between the two domains is completely
unknown.
Exhibit 18.12. Dispatch rules for C++ overloaded functions.
1. Look for a method whose operand types match the argument types exactly.
2. If none is found, look for a method such that the argument types can be made to match the
parameter types by using no more than one predefined conversion function on each operand.
3. If none is found, and one or more operand belongs to a defined class, look for a method such
that the argument types can be made to match the parameter types by using no more than
one user-defined conversion function per operand.
4. If more than one possible way to do user-defined conversions is found, or if none is found, the
dispatch fails.

18.3. POLYMORPHIC DOMAINS AND FUNCTIONS

18.3

Polymorphic Domains and Functions

18.3.1

Polymorphic Functions

561

A polymorphic type is a single type definition that includes two or more alternative specific type
declarations. A polymorphic object is the representation of one of these species along with some
form of discriminant field that encodes which species is present.
It is important to understand the differences between polymorphic functions and the generic
arithmetic operators in languages such as Pascal and Ada. The generic nature of the Pascal operator
extends only until compile time; the compiler then chooses a specific method to implement the
operator. Thereafter Pascal code is fully specific. A polymorphic function does run-time type
checking, when necessary, and a language with good support for polymorphism will do this checking
automatically. For example, APL is an array-oriented polymorphic language which tags each data
object with a type field that describes the number and extent of its dimensions. The programmer
may write an APL function, say F un, to operate on a pair of numbers. The programmer may then
call it with a simple pair of numbers or with two equal-length arrays of numbers. The shapes of the
actual arguments are tested automatically, and the translator executes appropriate code. It will
apply F un once if given a simple pair of numbers, but apply it repeatedly to each corresponding
pair of elements if given two arrays.
The Pascal variant record is an early attempt to implement limited polymorphic types. It
permits the programmer to specify that objects of the type may have a variety of different sizes
and structures. However, it is not a very satisfactory implementation of polymorphism for two
reasons:
When creating a polymorphic object, storage is allocated for the largest variant, even when
the value stored there is much smaller.
The theoretical semantics of variant records are not enforced by the translator, and run-time
tests of the discriminant field must be coded manually.
More modern languages have extended this idea and addressed the problems.
A polymorphic function is a function that accepts arguments of a polymorphic type. It tests
the argument at run time to determine which variant is present, then executes code appropriate
for that argument. These run-time type tests might be automatically generated by the translator
or explicitly written by the programmer.
Polymorphism becomes very important in functional languages because they support higherorder functions. A higher-order function, often called a functional, may take functions as parameters
and/or produce a function as its result. Some familiar functionals are functional composition and
reduce [Exhibit 12.15]. But, in a typed language, functionals are almost useless unless they are
polymorphic. Thus higher-order functions have been a strong motivating force in the development
of polymorphism in functional languages.
Even a language with little or no support for domain checking may be extended to perform
polymorphic domain checking manually. To accomplish this the programmer would explicitly attach

562

CHAPTER 18. DISPATCHING WITH INHERITANCE

Exhibit 18.13. Manual polymorphic dispatching in Ada.

In older languages, we use manual type-testing to write code to process a polymorphic type. The
Ada procedure below might be written to print out an item belonging to the polymorphic type
Person that was defined in Exhibit 14.25. In this type, the field named Sex is the discriminant
tag. We explicitly test the contents of this field and dispatch one of three specific functions.
PROCEDURE Print_Person ( Who : Person ) IS
BEGIN
CASE Who^.Sex IS
WHEN Male =>
THEN Print_Male(Who);
WHEN Female =>
THEN Print_Fem(Who);
WHEN OTHERS =>
Sex_Error(Who);
END CASE
END Print_Person;

some type information to every data object, then explicitly include code in the function definitions
to test this type field.

18.3.2

Manual Domain Representation and Dispatching

Before looking at a modern implementation of polymorphic domains, let us see how polymorphism
might be implemented manually in an older language. An ad hoc polymorphic domain is the
discriminated union of two or more specific types. If an object OB belongs to an ad hoc polymorphic
domain, PD, then its value at run time may belong to any specific type included in PD. It is not
possible to predict, at compile time, which specific type this will be. However, at run time, the
specific type of any object is known.
In some languages, for example, Pascal, an ad hoc polymorphic domain may be implemented
as a variant record with a tag field. Exhibits 14.25 and 14.24 show a polymorphic type declared
in Ada. However, neither Ada nor Pascal supports the run-time type checking that is necessary to
ensure valid use of such domains. To process polymorphic objects, we must write code, such as
that in Exhibit 18.13, which explicitly tests the tag field and branches appropriately.
Explicit Domain Testing versus Strong Typing. Achieving semantic validity is the purpose
of both strong typing and explicitly testing the discriminant tags on a polymorphic type. However,
there are some major differences between the results achieved by the two systems.
On the practical side, explicit type-testing can get to be cumbersome. It forces the programmer
to write out, in every function definition, the instructions to check the discriminant tag of the
argument and produce an error comment if there is a mismatch. A strongly typed language handles
this kind of domain checking automatically for simple types; the programmer needs only to declare

18.3. POLYMORPHIC DOMAINS AND FUNCTIONS

563

a domain name for each object and parameter.

Second, polymorphic type checking both permits and requires domain testing to be postponed
until run time. On the one hand, simple strong-type-checking, which happens at compile time, is
significantly more efficient if the program is ever used to process a lot of data. Run-time checking
is time consuming. On the other hand, run-time domain checking is inherently more powerful and
flexible. It permits the domain-checking system (or the programmer) to make finer distinctions
between appropriate and inappropriate arguments. A run-time test can depend on the actual data
values, not just on the declared type of the argument. For example, in Pascal, strong-typing ensures
that a list-processing function will always get a list argument. In Miranda, automated polymorphic
domain testing can distinguish lists from nonlists but can also be used to distinguish lists containing
data from null lists.

18.3.3

Automating Ad Hoc Polymorphism

Polymorphic Domains
The modern functional languages have more advanced support for generics than any existing languages in the ALGOL family. This support comes at two levels:
User-defined domains may be ad hoc polymorphic. A domain may have several alternative
representations, like a discriminated union type. Functions can be written that test the
discriminant automatically and dispatch the appropriate code.
A generic type may be defined with a type parameter, and functions may be defined with
generic parameters of this sort. (These are covered in Section 18.4.)
We use Miranda to illustrate these powerful general definition and dispatching methods. A type
definition is written using the symbol ::=, which we read as is type. A polymorphic type
declaration consists of a series of clauses separated by the or symbol, |, which correspond to
the alternatives of a discriminated union type in Ada or Pascal. Each clause has a discriminant
name followed by a tuple of type specifiers, much like a simple Pascal record declaration. The
discriminant names can be used either as constructors, for making an object of that type out of
appropriate components, or as type predicates, when defining a function for the type.
In Exhibit 18.14 we use Miranda to define tree, a polymorphic type. The notes for this example
follow:
a. We define the type tree. A tree may be a leaf, which is an integer, or a node which is a tuple of
two trees. This declaration defines the discriminant tags Leaf and Node, which we use later
as object constructors and as type predicates.
b. We construct a tree of the first form, using the constructor Leaf. Note that the tag name is
part of the constructed objectthe result of this line is that the name leaf1 is bound to a
two-tuple consisting of the tag Leaf and the number 3.

564

CHAPTER 18. DISPATCHING WITH INHERITANCE

Exhibit 18.14. A tree type in Miranda.

We define an ad hoc polymorphic type named tree. This also defines the constructor/predicate
names Leaf and Node for the two variants of type tree. The last two lines construct two trees
named leaf1 and tree1. The letters on the right key the code to the notes in the text.
tree ::= Leaf integer | Node tree tree

(a)

leaf1 = (Leaf 3)
tree1 = (Node leaf1 (Node (Leaf 17) (Leaf 49))

(b)
(c)

c. We construct two leaves, as in the line above, and use them immediately to construct one node.
That node is, in turn, combined with the previously constructed leaf, leaf1, to make a tree
named tree1. The structure of this tree is shown in Exhibit 18.15.
Unlike Pascal, part names for the fields of a tuple are not supplied in the type declaration, and
the new type name may be referenced recursively in the declaration. The lack of defined part names
requires explanation. In Miranda, the type of an object is not part of the object, as in APL, nor is
it part of a name, as in Pascal. Types are deduced from the structure of objects. Field names, in
turn, are not part of the type definition. When an object is passed as an argument, a sophisticated
type-deduction algorithm6 is used to check whether it belongs to the domain of the function. The
tuple- or list-structure of the argument is checked, along with the types of its simple components.
As part of this checking process, dummy parameter names, defined in the function header, are
associated with each field of the argument. These are temporary and local, like parameter names
in traditional languages, and are used to access the fields of the argument within that function.
Thus the two parts of a pair may be called left and right within one function body and
head and tail in another.
6

This algorithm is derived from the work of Hindley and Milner. Cf. Milner [1978].

Exhibit 18.15. A tree with polymorphic nodes.

tree1
leaf1
leaf1

After line b.

17

49

Middle of line c.

3
17

After line c.

49

565

18.3. POLYMORPHIC DOMAINS AND FUNCTIONS

Exhibit 18.16. Polymorphic dispatching in Miranda.

We define and call a polymorphic function named max-tree with methods to process both kinds of
trees (nodes and leaves). The letters on the right key the example to the explanation in the text.
max-tree (Leaf ldata)
max-tree (Node n1 n2)

Evaluating

=
=
=
=

ldata
max1, max1>max2
max2, max2>max1
max1, otherwise
where
max1 = ( max-tree n1 )
max2 = ( max-tree n2 )

( max-tree leaf1 ) yields 3, and evaluating

(a)
(b)
(e)
(c)
(d)

( max-tree tree1 ) yields 49.

Miranda permits us to write functions over ad hoc polymorphic domains, like tree, that will
be domain-checked and dispatched automatically. A function definition is written as a series of
methods, each one defined for a different subdomain. The domain predicates are called patterns
and are written on the left. A pattern can specify an argument type, a discriminant tag (called a
constructor in Miranda), and/or data values. Following the pattern is an = and the appropriate
computation method for that particular subdomain. In Exhibit 18.16, we define a simple function
that processes the tree type defined in Exhibit 18.14. Following are line-by-line notes for the code
in the exhibit:
a. On this line and the next, the discriminant tags defined by the type declaration are used as
domain predicates (left), controlling the choice of computation methods (right). Line (a)
defines the function max-tree for any argument which matches the type named Leaf.
Within the body of the method, the single component of a Leaf will be called ldata. The
value of the expression on the right will be returned as the value of the function. This code
states that the maximum of a Leaf is the number stored in it.
b. The remaining code defines max-tree for arguments that are tuples with two fields. These fields
are named n1 and n2 within the method body.
c. The where clause defines a local context for this method, containing local names max1 and
max2. When the outside-in, lazy evaluation process reaches the first reference to max1 or
max2, the expression following the name in the where section is evaluated and the result is
bound to the local name. This value remains bound to the name throughout evaluation of the
block; subsequent references to the name refer to the value computed for the first reference.
d. The function max-tree is called, recursively, with the left field subtree as its argument. The
maximum value in the left subtree is bound to the name max1.

566

CHAPTER 18. DISPATCHING WITH INHERITANCE

Exhibit 18.17. Dispatching a Miranda call using data values.

sum-list []
sum-list (a:ls)

= 0
= a + sum-list ls

We recursively sum a list by adding the value of the first list item to the sum of the rest of the
list. If list1 = [3, 4, 17, 9, 5] and we evaluate sum-list list1 the result is 38.

e. This guarded expression compares the two maxima of the left and right subtrees and returns
the greater.
This automated run-time dispatching of function methods is implemented by a very general
and powerful pattern-matcher. It is not limited to checking for subtypes of a polymorphic domain,
but can also perform run-time checks involving the values of data objects. For instance, Exhibit
18.17 has one method defined for null lists and another method defined for lists with at least one
component.
Exhibit 18.18 summarizes and gives examples of the kinds of patterns that the Miranda programmer may use to define methods. Dispatching a function call is done by a run-time case analysis,
examining each pattern in turn, until one is found that matches the structure and/or value of the
actual argument. When a match is found, the dummy names used to write the pattern are bound
to the parts of the argument, and these bindings are used within the function body. Exhibit 18.19
shows the bindings that would happen for some calls on the functions sum-list [Exhibit 18.17],
max-tree [Exhibit 18.16], and pow10 [Exhibit 12.20].
The important principle here is that the dispatch is done at run time, distinguishing it from
anything that can be done with nonunion types in languages such as Pascal and Ada. It is true
that an Ada programmer can emulate run-time dispatching by using a discriminated union type
and coding the dispatching process manually, as in Exhibit 18.13. However, manual type checking
is never as convenient or as safe as checking that is built into the semantic basis of the language.
The Miranda function definitions using pattern matching are far clearer and more elegant than
corresponding Ada code.

Exhibit 18.18. Pattern construction in Miranda.

Pattern type
constant
number greater than 0
null list
nonnull list
tuple

Example
factorial 0 = 1
factorial (n+1) = (n+1)*factorial n
sum-list [] = 0
sum-list[a:x] = a + sum-list[x]
max-tree (Node n1 n2) = ...

18.3. POLYMORPHIC DOMAINS AND FUNCTIONS

567

Exhibit 18.19. Binding during the pattern match.

Data object definitions:
type
name value
tuple leaf1 = (Leaf 3)
tuple tree1 = (Node leaf1 (Node (Leaf 17) (Leaf 49))
list
list1 = [ 1, 3..11 ]
list
list2 = []
integer numb1 = 17
The functions below were defined in Exhibits 12.15, 12.20, 18.16, and 18.17.
Call
Bindings
max-tree leaf1
ldata is 3
n1 is leaf1, n2 is (Node (Leaf 17) (Leaf 49))
max-tree tree1
max-tree numb1
Pattern match fails.
max-tree list1
Pattern match fails.
pow10 numb1
n is 16.
a is 1, ls is [ 3, 5..11 ].
sum-list list1
Pattern matches null list, no binding necessary.
sum-list list2
reduce + list1 0 f is +, a is 1, x is [ 3, 5..11 ], and n is 0.

18.3.4

Parameterized Domains

Miranda is a strongly typed language. Every object has a type, even though it is not declared. The
type is deduced from the structure of the object. Functions, also, have types which are deduced
from the function code. If a function is called with an argument whose type is inconsistent with the
function definition, a compile-time error comment is generated. The programmer may also choose
to declare the type of a function, in which case a compile-time error is produced if the declared
type and the deduced type do not match.
Names may be defined for types and used as a notational convenience. The names carry no
semantic meaning in themselves but are simply a shorthand for the structural description. 7 To
define a type name, you use the == sign:
htypenamei == htype expressioni
Miranda supports domains with type parameters and provides a notation for talking about a
type parameter. Ordinary types are expressed structurally, as shown in Exhibit 18.20. The type
of a list is denoted by writing the type of its elements inside list brackets; the type of a tuple is
written as a tuple of types. The type of a function is written in curried notation, starting with the
type of the first argument and ending with the type of the result.
Type parameters are denoted by strings of asterisks. If a type expression has one type parameter,
7

See the discussion on type identity in Chapter15, Section 15.4.

568

CHAPTER 18. DISPATCHING WITH INHERITANCE

Exhibit 18.20. Miranda type expressions.

Assume T, T1, etc. are types. Then we can write the following type expressions:
Expression
Interpretation
[T]
A list with elements of type T.
[[T]]
A list of lists of elements of type T.
(T, T1, T2)
A tuple of three elements of types T, T1, and T2.
T T1
A function with argument type T and result type T1.
* *
A function whose argument and result have the same type.
[*] *
A function that returns an element of the same type as the
base type of its list argument.
* ** ** A function of two arguments whose result is the same type
as the second argument.

we write *; for an expression with three distinct types, we write *, **, and ***. If a type
expression uses the same type parameter symbol twice, it means that both occurrences must be
replaced by the same argument. For example, the type of the subscript function is:
!

::

[*] num *

Abstract Data Types. Miranda supports code modules analogous to the generic packages in
Ada, except that Mirandas type parameters are not bound at compile time, so it is capable of
true run-time generic behavior. A Miranda script may contain the directive to %include the
script in some other file, or a directive to %export some locally defined symbols. Scripts may
have parameters, just as in Ada, and a parameterized script must be instantiated by supplying
arguments in the %include command.
An ADT, called an abstype, is declared by specifying the ADT interface (public symbols)
followed by the definitions of those symbols (private part). Exhibit 18.21 declares a stack ADT
with a single type parameter for the base type of the stack. The stack itself will be represented
by a list. The rules for type compatibility and access to private parts are very similar to those
in C++ classes; within the scope of the declaration, the type stack * is considered to be just
the same as the type [*], and the implementation equations may access it using the ordinary
list operations. Outside the abstype declaration, though, a stack may only be accessed using the
declared functions.

18.4

Can We Do More with Generics?

We have looked at a variety of languages that support generic and/or polymorphic functions. It is
appropriate, now, to look back and compare these facilities, to analyze strengths and weaknesses,
and to ask whether a language could do more.

18.4. CAN WE DO MORE WITH GENERICS?

569

Exhibit 18.21. The ADT stack in Miranda.

abstype stack *
with empty::stack *
isempty::stack * bool
push :: * stack * stack *
pop :: stack * stack *
top :: stack * *
stack * ==[*]
empty
= []
isempty s = (s=[])
push a s = (a:s)
pop (a:s) = s
top (a:s) = a

||
||
||
||
||
||

A stack is a list of anything.

A literal stack, to start things off.
A null list represents an empty stack.
Append the new item to the head of the list.
Remove list head and discard; return changed stack.
Return the list head, dont change the stack.

Binding time, Flexibility, and Efficiency

The binding time for type variables is the first and biggest difference among the languages we have
studied. Ada, C++, and Miranda all support some sort of generic behavior. However, Adas generics
are much simpler, more efficient, and less flexible. Ada does not need a dispatcher at all, because
the programmer binds specific names to instantiations of generics, and uses those specific names
when writing the code. The result can be fully compiled, and is easy to compile, but it does not
permit any run-time variability at all. For example, variable-length strings are quite clumsy to
manage in Ada.
Contrast Ada to Miranda. Variable-length lists are used everywhere in Miranda programs, and
polymorphic tuple types exist at run time. The idea of type is defined in Miranda so that a single
type includes more than one specific representation. The Miranda compile-time dispatcher verifies
that a function is defined for the type but does not worry about which representation of the type
will be present later. The run-time-dispatcher examines the actual arguments and dispatches the
particular method that matches the argument. Many Miranda functions do repetitive processes
on data structures built from polymorphic tuple types. The run-time dispatching is essential to
support this, since the method selected will differ from one call to the next, depending on the shape
of the data structure.
Mirandas run-time variability makes a language that is flexible and easy to use; strong typing
ceases to be a barrier between the programmer and his or her work, and becomes a pure asset.
The cost, however, is execution speed. Repeating the pattern match for every repetition of every
function is a great deal slower than doing it once at compile time.
The C++ compiler takes a middle course. It makes a distinction between function calls for
which complete information is known at compile time and those where some run-time variability

570

CHAPTER 18. DISPATCHING WITH INHERITANCE

can possibly exist. In the former case, fixed code is compiled, as it would be in C. In the latter case,
the run-time dispatcher is called. (Review the discussion of virtual functions in Section 18.2.3.)
Using a mixed strategy like this seems to be the best strategy, although it is also the most difficult
to implement.
The real problem with a mixed strategy is deciding what can and what cannot be known or
deduced at compile time. C++ solves this problem simply by deciding that any computation that
involves a virtual function will be dispatched at run time. Ordinary functions will be dispatched
at compile time. This works to achieve reasonable efficiency only if the use of virtual functions is
relatively unusual.
Defining and Representing a Generic Function
In a traditional language, we know what a function is; it is one body of code, with one type. This
situation becomes complicated when we deal with generic functions, and we need to decide what a
function is and what can be done with functions.
We need to distinguish between the situation in which a function name is overloaded, that is,
used for two unrelated methods, or truly generic. A function is truly generic if the translator knows
about more than one method, knows the methods are related, and uses that relationship in the
dispatching process.
In comparing generic languages, we see different approaches to the question of what is a
function? In Miranda, a function can have several defining clauses, giving it polymorphic behavior,
but everything relevant to one function is defined in one place. Miranda functions are first-class
objects that can be passed to other functions and created dynamically. When we pass a functional
argument, this entire polymorphic unit is passed.
In contrast, C++ virtual functions are defined in bits and pieces, with each method inside
a different class. A constructor function, on the other hand, may have multiple definitions in the
same class. Finally, built-in functions (casts, arithmetic operators) have several definitions and are
not included in a class at all. Methods for these functions are defined piece by piece, as needed. It
is not necessary to know about or edit prior definitions of a function in order to extend the function
to handle a new submode type; the new method is simply included where it is needed.
Both constructors and virtual functions are true generic functions, not simply overloaded names.
The constructors in a class all bear a semantic relationship to each other; they all return an object
of the type of the class. The methods for a virtual function also are semantically related, even
though they are defined at different times and may be written by different people and compiled in
different code modules. In Miranda, this would not be possible; the methods for a virtual function
would all have to be collected and written in one place. Thus the C++ approach for defining a
generic function has a definite advantage for writing large systems.
A generic function is a collection of methods, each one defined over a different subdomain of a
generic domain. Some languages (such as Miranda) require the methods to be lexically grouped on
the page, in other languages they are semantically grouped by the translator (like Adas arithmetic
operators), in yet others they are tucked into classes and accessed through the class hierarchy so

18.4. CAN WE DO MORE WITH GENERICS?

571

that the methods are never connected together or treated as a unit (like virtual functions in an
object-oriented language).
Operations on Generic Functions
Let us consider a generic function to be a list (not ordered) of methods. You may picture it as a
linked list of code modules. There are three operations that we would like to define for a generic
function: passing it as an argument, dispatching it, and executing one of its methods.
When passing a generic as an argument, do we pass the entire unit (including all the methods),
as in Miranda? Or do we pass a single method? Passing an entire function is only meaningful
if we have a run-time dispatcher. In Ada, we sometimes pass a function as an argument when a
package is instantiated. If we wish to send + as an instantiation parameter, Ada permits us to
name the entire function as the argument; it does not require that we denote a single method for
+ [Exhibit 17.20]. An Ada instantiation argument is passed at precompile time and is substituted
in the package before Adas compile-time dispatcher works on the code. This is not the same as
passing an entire function at run time. Functional parameters in C must be single methods, because
no run-time dispatcher exists to handle a whole functions. Miranda scripts can pass entire generic
units at run time because all dispatching is done at run time.
Then consider dispatching; the dispatcher selects one method for each generic function call, if
an appropriate method exists. The compiler must keep the methods for a function in some sort of
data structure so that it can search all possibilities during dispatching.
A class hierarchy forms a convenient data structure for organizing the methods of a function.
The dispatching algorithm is easy to write if each method is attached to a class and no class has
more than one method for the same function. The algorithm searches for a method starting at the
class of the first (implied) argument and crawls up the class hierarchy tree until it either finds a
method or comes to the base class. In C++, the base class is required to have a method for each
virtual function, guaranteeing that the dispatch will never fail at run time. (This restriction is not
necessary, though.)
Handling methods that are not attached to classes is more difficult, as is handling functions
that have multiple definitions within one class. To make sense of this, the dispatcher must examine
a list of methods, looking at the type of each method. If the type of each argument matches the
declared type of the corresponding parameter, the method may be dispatched. We need to define
precisely what match means, and what happens if no match is found, or if more than one is
found.

18.4.1

Dispatching Using the Mode Graph

Consider a generalized dispatching algorithm in which generic functions are represented as lists of
methods, and the dispatcher can move around the entire mode graph, including submode links,
binding links, and conversion links. This kind of dispatching algorithm is used in two current

572

CHAPTER 18. DISPATCHING WITH INHERITANCE

research languages, Haskell8 and Aleph9 In the rest of this section, we will refer to this generalized
algorithm as the Dispatcher, with a capital letter.
Dispatching starts with a particular function (a list of methods) and a particular context for the
function call. The context is formed from the types of the actual arguments and (possibly) the type
of object that the function must return. The Dispatcher examines each method for the function,
in turn, and either eliminates it, selects it, or puts it on a short list for future consideration.
In an object-oriented dispatch, the only part of the context that is considered is the type of
the implied argument. This is easy to implement, but it is limited and can handle only simple
dispatches. Even in the examples we have seen, it is inadequate to handle the problem of two
constructors in one class. A more general and symmetric dispatcher would consider the types of
all the arguments, as C++ does with constructors, and as the Miranda dispatcher does with all
functions. The Dispatcher will consider each argument, in turn, and select a method only if all
arguments match or can be coerced to match. Functions such as constructors, which can have
different methods for different combinations of arguments, will be dispatched correctly.
Matching. Dispatching one argument reduces to the problem of finding a path through the graph
from the mode containing the type of the argument to the declared mode of the parameter. The
dispatcher will start at the argument mode, search up IS links and BIND links and across CONV
links, trying to reach the target mode. Unlike C++ dispatching, the Dispatcher is not limited to
using a single CONV link, and it can use CONV links in combination with the other links.
Using conversions freely brings up the question of semantic validity. What is there to prevent a
string of conversions from distorting the meaning of the argument beyond recognition? This issue is
the reason that we must distinguish carefully between casts and semantics-preserving conversions.
A cast is a semantically invalid operation, done only to access the underlying representation for
bootstrapping purposes. Casts should never be used freely and never to coerce arguments. Any
chain of conversions that goes through two casts is almost certainly semantically invalid. This puts
a burden on the programmer to distinguish which functions represent casts, and which are valid
conversions, then leave the casts out of the mode graph. So long as all links in the mode graph are
individually valid, the combination of those links should also be valid.
With a class hierarchy, there is only one way to travel from one class to another: up the tree
toward the base class. In contrast, a mode graph is an arbitrary graph; one mode can have two or
more IS links leading out of it and also have CONV links. We can consider CONV links to be more
costly than the other links. The dispatcher must therefore consider more than one path from the
starting point to the goal, and dispatching becomes a process of finding the shortest path through
a weighted graph, not just traveling up a tree structure.
Where the types of objects are fixed at compile time, all of this work can be done by the
compile-time Dispatcher, which then replaces the ordinary type checker. The Dispatcher identifies
a type error when the compile-time dispatch fails to find any applicable methods. Where the
8
9

Hudak et al. [1992].

A. Fischer and M. Fischer [1973], A. Fischer [1985], and A. Fischer and R. Fischer [1992].

573

18.4. CAN WE DO MORE WITH GENERICS?

Exhibit 18.22. Mode graph for points on a plane.

Point
Promises:
Read

carts

IS

Methods:
record 8 bytes
Read(carts)
2
Write(carts)
0, 4
Translate(carts, real, real)
x, y
real, real

Const:
Zero
One

IS

polar

Methods:
record 8 bytes
Read(polar)
2
Rotate(polar, real)
0, 4
r, theta
real, real

CONV c_to_p
CONV p_to_c

types of objects may retain some run-time variability, the Dispatcher can still identify total type
mismatches and eliminate many methods as possibilities. Sometimes, though, more than one
method is potentially applicable; in this case, the compile-time Dispatcher must return a short list
of potential methods, to be further culled by the run-time Dispatcher.
Generics Make Some Problems Easy. Let us define a sample generic mode and show how
dispatching would work on it. Assume we are writing a graphics program, in which we are concerned with representing and manipulating points on the plane. There are two good methods for
representing points: in Cartesian coordinates, as (x, y) pairs, or in polar coordinates, as (r, theta)
pairs. Selector functions x and y are defined for type carts, and r and theta are defined for type
polar. Note that selectors are ordinary functions and can be called using ordinary function syntax,
even though they are defined as part of the type declaration. This program will need to do many
operations on points, including rotation, translation, input, and output. Either representation can
be converted to the other using trigonometric functions or square roots.
Translation, that is, moving the point up, down, or sideways, is easy in Cartesian coordinates
but difficult in polar. Rotation, that is, moving a point in a circular arc around the origin, is easy in
polar but not Cartesian. Output must be in Cartesian, because the terminal screen uses Cartesian
coordinates. Some input is in polar, other in Cartesian. The program must use both representations,
and the programmer would like to have both available but not worry about representation all the
time.
The mode graph for these types is shown in Exhibit 18.22. The generic mode, point, promises
the function read and has two typed submodes, carts and polar. Any point data structures

574

CHAPTER 18. DISPATCHING WITH INHERITANCE

Exhibit 18.23. Generic calls.

VAR p1:
c1:

polar;
carts;

Read(c1);
Read(p1);
If theta(p1)
> theta(c1)
then Write(p1)
else Write(c1);

~~
~~
~~
~~
~~
~~

No problem: Read is defined for type carts,

and also for type polar.
The selector theta returns a real.
The Cartesian argument is coerced to polar.
The polar argument is coerced to Cartesian.
Write is defined for Cartesian.

and functions in the surrounding program are declared in terms of the generic mode, point.
Note that both submodes are represented by pairs of reals, but neither type must ever be cast
to the other, because their semantics are totally different. A package for ADT point will contain
private type declarations for the two point representations and public definitions for the ADT
functions. There would also be two data objects (or constant functions) for each type, named Zero
and One.
Exhibit 18.23 illustrates the power of this generic language to simplify the programmers task. 10
Two variables are declared and data values are read for them. The Dispatcher examines the methods
defined for Read and chooses the appropriate ones for the two Read commands. The If statement
then applies the selector function theta to both points, causing the point c1 to be coerced before
the selection can be done. One of the Write statements also coerces its argument.
Everything in this example can be dispatched at compile time. Even though Read is a virtual
function, it is only used with arguments whose type is fixed and known at compile time, permitting
the Dispatcher to identify the correct methods and necessary conversions. Contrast this situation
to the problem in Section 18.2.3 which illustrates the need for run-time dispatching. In that
application, cells of differing specific types were linked together in one list by pointers with a
generic base type. In that case, we could do compile-time dispatching on pointer operations but
we still needed run-time dispatching for operations on base-type objects.
This small example is a little artificial, but it illustrates that the programmer using a generic
language is, indeed, freed from constant concern about which representation is being used at the
moment. In a real program, there would be some reason why a particular point would be represented
one way or another, and the time spent performing conversions would be useful. The gain, over
traditional languages, is that the programmer can think in terms of the semantics of points, not the
semantics of carts and polar. The resulting code denotes the job of computation more clearly,
uncluttered by constant nuts-and-bolts conversion commands.
10
The syntax used here is Pascal-like, so that readers will understand it readily. However, you could certainly not
write this code in Pascal!

575

18.4. CAN WE DO MORE WITH GENERICS?

Exhibit 18.24. Using generic pointers.

Number * P, Q, R;

4.3

18.4.2

17

82

Generics Create Some Hard Problems

Dispatching Pointer Arguments

An interesting question arises when we declare a type that is a pointer to a generic type. For
example, let P, Q, and R be of type pointer to Number (see Exhibit 18.1) and assume that
they have been initialized to point at values as shown in Exhibit 18.24. Now assume we dereference
the pointers and add the results by saying P^+ Q^ or Q^+ R^. At first glance, it seems that these
additions should be straightforward; the promises guarantee that + will be defined for all Numbers,
and P, Q, and R have the same type.
Looking again, though, we see that the things that P, Q, and R point to may or may not
have the same type! Here is a situation in which dispatching must be delayed until run time.
The compile-time Dispatcher must make a short list of all the methods for + defined on submodes of Number, and the particular method to be used must be selected at the last moment.
In our example, the method +(real,real):real will be selected for the first addition and
+(integer,integer):integer will be selected for the second.
A second problem involves possible run-time type incompatibilities. The promises for Number
guarantee that every submode will have a definition for +. However, + is a dyadic function, and
these definitions all require that the two arguments to + have the same specific type. No methods
at all are defined for mismatched operands, like P^+ Q^ in the example. These are supposed to be
handled by the coercion mechanism, and in this case, they would be. The integer operand will be
converted to real by the CONV function i_to_r. But without the conversion links between the
submodes, the computation could not be done, and the dispatch would fail. To avoid run-time
errors, the programmer must provide enough CONV links to define the relationships among all the
submodes of a generic mode.
Using Generic Definitions
One of the attractions of generic dispatching is that functions defined high in the mode graph,
for high-level generic modes, can be inherited by all the submodes below them. Inheritance works
much as it does in a class hierarchy. But when we use a high-level generic method definition, there
is a problem with knowing the type of the result. The method definition can only define the result
type in terms of the generic mode to which it is attached. However, when this definition is used,

576

CHAPTER 18. DISPATCHING WITH INHERITANCE

Exhibit 18.25. What type is returned?

k: integer;
r: real;
v: vector;
...
r := v ! k ;

Parameterized mode:
Array(N:integer, T: type)
arraytype

N*sizeof T byt es
1..N
sizeof T bytes @
T

VECTOR
arraytype
BIND
N is 3
T is real

12 bytes
1..3
4 bytes @
real

it is used with a specific type object, and we would like to know the specific type of the result, not
just its generic category.
If we lose track of the specific type that results from an operation, no more compile-time
dispatching can be done with that result; all further dispatches for the entire expression must be
done at run time if they involve the result of the generic. Unless there is a way to keep track of the
specific types involved, generics become too impractical to use extensively, defeating their purpose.
Let us look at one example of this problem. Assume an object belongs to a typed mode, Vector,
which was derived from a parameterized mode, Array(N,T). Vector is connected to Array by a
BIND link that gives the bindings N=3, T=real [Exhibit 18.25]. The generic function Subscript
(written with the symbol !) is defined for mode Array and has the type:
!

::

Array(N, Any) integer Any

where mode Any is the predefined supermode of all modes.

What type is returned when we call Subscript on a Vector? The Subscript function was
defined to return mode Any. We know, though, that it will return a real in this case, and it is
important for the compile-time dispatcher to be able to use that fact to compile Vector expressions.
This is one instance of a very general and difficult problem that arises with parameterized
generic modes. We must be able to keep and pass on the specific attributes of the submodes if
generic functions defined for supermodes are to be useful. Of course, Subscript is predefined in
all languages on all array types and is treated as a special case; the programmer does not have to
define it. However, a powerful generic language should permit a programmer to redefine subscript
or to define other functions of this general nature.
To address this specific/generic problem, a language and its translator must have three things:
1. A general type notation that lets the programmer specify type constraints when she or he
uses generic modes to declare a function type. For example, the programmer must be able to
state that one argument must be the same type as another, or must have a specific structural
relationship to another arguments type.
2. Both the dispatcher and the programmer must have access to the information on the BIND
links and in the type objects of typed submodes. Type operations such as type of , value
of binding, and dimension of must be supported.

18.4. CAN WE DO MORE WITH GENERICS?

577

3. The dispatcher must perform type calculations, using these type operations. The purpose of
these calculations is twofold: to guarantee that any specified type constraints are obeyed, and
to deduce the most specific information available about the return types.
In the discussion of Miranda, in Section 18.3, we saw the use of the symbols *, **, etc.
to denote mode variables, and mode expressions such as [*] to denote a list of a generic type.
These asterisk symbols and expressions serve purpose (1), above; they permit the programmer to
specify constraints (for example, is the same type as and is a list of base type the same as) in
a generic function declaration.
Requirement (2), above, is partially met in Ada by providing many predefined type operations
which, essentially, let the programmer (or the system) access most of the fields of a type object.
Using this information, a programmer-defined function could do things that are built in and nonextensible in traditional languages, such as bounds checks, variations on the subscript function,
and the like.
Miranda accomplishes part of goal (2) by permitting the programmer to specify patterns for use
by the dispatcher. What seems to be missing is the ability to access the instantiation parameters
for parameterized modes. (Ada does not have this problem because all instantiation is done before
the program is compiled, and a program has full knowledge of the instantiation parameters used.)
Requirement (3) is invisible to the programmer; it does not affect the language syntax or the
list of defined operations. However, it affects the semantics in a major way. A dispatcher that is
able to do this will permit much more general use of generics and still be able to maintain strong
typing.
Mechanisms like these form part of the semantic basis of Haskell and Aleph, and are sure to play
a central role in languages of the future.

Exercises
1. Why cant we use two structurally dissimilar types to represent one external domain in a
traditional strongly typed language?
2. How do generic functions solve this problem?
3. What is a generic mode? How is a generic mode different from a C++ base class?
4. How do Adas predefined modes limit the programmer?
5. What are submode links? Promises? Supermodes?
6. Why are domains with multiple representations less practical to use than domains with a
single representation?
7. What is the danger of providing a general coercion facility?

578

CHAPTER 18. DISPATCHING WITH INHERITANCE

8. Why should the dispatchers coercion facility care whether a function is a conversion or a
cast?
9. What is a best match in dispatching in older languages?
10. Why is dispatching more complex in modern languages?
11. A subrange type can inherit functions from a base type. Explain.
12. How does a class hierarchy provide broader support for subdomains? Be specific.
13. What is the difference between the handling of a C++ function definition placed inside a class
and outside it?
14. When is a derived module considered opaque? How is this accomplished?
15. Why is the compatibility of pointer types throughout the levels of a hierarchy essential?
16. What is a virtual class? Why is it necessary?
17. Why must dispatching of virtual classes be delayed until link or run time? How is this
accomplished in C++?
18. What are the benefits of virtual functions? The costs?
19. What is the dispatching rule for function inheritance in C++?
20. Why is the definition of a constructor function for a derived class more complicated than for
a base class?
21. What is a polymorphic type? Object? Function?
22. How was polymorphism implemented in older languages?
23. What are the differences between explicit domain testing and strong typing?
24. How do modern functional languages support generics?
25. How are types of objects deduced in Miranda? Explain.
26. In Miranda, what is a pattern? A constructor?

Bibliography
[1] Abelson, H., et al. The revised revised report on Scheme or an UnCommon Lisp. AI Memo
848, MIT Artificial Intelligence Laboratory, Cambridge, Massachusetts, August 1985.
[2] Abelson, H., and G. J. Sussman. Structure and Interpretation of Computer Programs. The
MIT Press, Cambridge, Massachusetts, 1985.
[3] Adobe Systems Incorporated. PostScript Language Reference Manual. Addison-Wesley, Reading, Massachusetts, 2nd edition, 1990.
[4] Allen, J. R., M. E. Burke, and J. F. Johnson. Thinking About TLC-Logo. CBS College
Publishing, New York, 1983.
[5] Arden, B., B. Galler, and R. Graham. The Michigan Algorithm Decoder. Computing
Center, University of Michigan, 1963.
[6] Barnes, J. G. P. Programming in Ada. Addison-Wesley, Wokingham, England, 3rd edition,
1989.
hm, C., and G. Jacopini. Flow diagrams, Turing machines, and languages with only two
[7] Bo
formation rules. Communications of the ACM, 9(5):366371, May 1966.
[8] Booch, Grady. Software Engineering with Ada. Benjamin/Cummings, Menlo Park, California, 1987.
[9] Borland International, Scotts Valley, California. Turbo Pascal Reference Manual, Version 2.0,
1984.
[10] Brinch Hansen, P. The programming language concurrent Pascal. IEEE Transactions on
Software Engineering, 1:199206, 1975.
[11] Brodie, L. Starting FORTH. Prentice Hall, Englewood Cliffs, New Jersey, 2nd edition, 1987.
[12] Brown, J. A., S. Pakin, and R. P. Polivka. APL2 at a Glance. Prentice Hall, Englewood
Cliffs, New Jersey, 1988.

579

580

BIBLIOGRAPHY

[13] Church, A. The calculi of lambda conversion. Annals of Mathematical Studies, 6, 1941.
[14] Clocksin, W. F., and C. S. Mellish. Programming in Prolog. Springer-Verlag, Heidelberg,
1981.
[15] Cooper, D. Standard Pascal User Reference Manual. W. W. Norton and Company, New
York, 1983.
[16] Dale, N., and S. Lilly. Pascal Plus Data Structures, Algorithms, and Advanced Programming. D. C. Heath, Lexington, Massachusetts, 1985.
[17] Dartmouth College Computation Center, Dartmouth, New Hampshire. A Manual for BASIC,
1965.
[18] Dijkstra, E. W. Programming considered as a human activity. In Proceedings of IFIP
Congress, pages 213217, Amsterdam, Netherlands, 1965. North-Holland.
[19] Dijkstra, E. W. Go To statement considered harmful. Communications of the ACM,
11(3):147148, 1968.
[20] Fasel, J. H., and P. Hudak. A gentle introduction to Haskell. ACM SIGPLAN Notices,
27(5):T1T53, May 1992.
[21] Fischer, A. E. A Semantic Basis for a Programming Language with Declarable Mode Relationships. PhD thesis, Harvard University, 1985.
[22] Fischer, A. E., and M. J. Fischer. Mode modules as representations of domains. In
Proceedings of the ACM SIGACT/SIGPLAN Symposium on Principles of Programming Languages, pages 139143, October 1973.
[23] Fischer, A. E., and R. P. Fischer. Aleph implementation, 1992. Work in progress.
[24] Frege, G. Begriffsschrift: A Formula Language, Modeled upon That of Arithmetic, for Pure
Thought. Halle, 1879.
[25] Galler, B. A., and M. J. Fischer. The iteration element. Communications of the ACM,
8(6):349, 1965.
[26] Goldberg, A., and D. Robson. Smalltalk-80: The Language and Its Implementation.
Addison-Wesley, Reading, Massachusetts, 1983.
[27] Gorsline, G. W. Computer Organization. Prentice Hall, Englewood Cliffs, New Jersey, 2nd
edition, 1986.
[28] Griss, M. L., and A. C. Hearn. A portable LISP compiler. Software Practice and Experience, 11:541605, 1981.

BIBLIOGRAPHY

581

[29] Griswold, R. E., and M. T. Griswold. The Icon Programming Language. Prentice Hall,
Englewood Cliffs, New Jersey, 1983.
[30] Griswold, R. E., J. F. Poage, and I. P. Polonsky. The SNOBOL 4 Programming
Language. Prentice Hall, Englewood Cliffs, New Jersey, 2nd edition, 1971.
[31] Harbison, S. P., and G. L. Steele Jr. A C Reference Manual. Prentice Hall, Englewood
Cliffs, New Jersey, 3rd edition, 1991.
[32] Hindley, J. R. The principal type-scheme of an object in combinatory logic. Transactions
of the AMS, 146:2960, 1969.
[33] Holt, R. C., P. A. Matthews, J. A. Rosselet, and J. R. Cordy. The Turing Programming Language. Prentice Hall, Englewood Cliffs, New Jersey, 1988.
[34] Horowitz, E. Programming Languages: A Grand Tour. Computer Science Press, Rockville,
Maryland, 2nd edition, 1985.
[35] Hudak, P., et al. Report on the programming language Haskell: A non-strict, purely
functional language, version 1.2. ACM SIGPLAN Notices, 27(5):R1R164, May 1992.
R
[36] INMOS Limited. occam
2 Reference Manual. Prentice Hall, Hempel Hempstead, England,
1988.

[37] Kemeny, J. G. True BASIC. Addison Wesley, Reading, Massachusetts, 1985.

[38] Kernighan, B. W. Why Pascal is not my favorite programming language. In A. Feuer
and N. Gehani, editors, Comparing and Assessing Programming Languages: ADA, C, Pascal,
pages 170186. Prentice Hall, Englewood Cliffs, New Jersey, 1984.
[39] Kernighan, B. W., and D. M. Ritchie. The C Programming Language. Prentice Hall,
Englewood Cliffs, New Jersey, 2nd edition, 1988.
[40] Kessler, R. LISP, Objects, and Symbolic Programming. Scott, Foresman and Company,
Glenview, Illinois, 1988.
[41] Landin, P. J. The mechanical evaluation of expressions. Computer Journal, 6(4):308320,
1964.
[42] McCarthy, J., et al. LISP 1.5 Programmers Manual. M.I.T. Press, Cambridge, Massachusetts, 2nd edition, 1962.
[43] Meissner, L. P., and E. I. Organick. FORTRAN 77 Featuring Structured Programming.
Addison-Wesley, Reading, Massachusetts, 1980.
[44] Milner, R. A theory of type polymorphism in programming. Journal of Computer and
System Sciences, 17:348375, 1978.

582

BIBLIOGRAPHY

[45] Morris, W., editor. The American Heritage Dictionary of the English Language. Houghton
Mifflin Co., Boston, 1969.
[46] Naur, P. Report on the algorithmic language ALGOL 60. Communications of the ACM,
3:299314, 1960.
[47] Naur, P., and M. Woodger. Revised report on the algorithmic language ALGOL 60.
Communications of the ACM, 6:120, 1963.
[48] Newell, A., editor. Information Processing Language-V Manual. Prentice Hall, Englewood
Cliffs, New Jersey, 1961.
[49] Oracle Corporation, Belmont, California. Introduction to SQL, 1987.
[50] Perlis, A. J., and S. Rugaber. The APL idiom list. Research Report 87, Yale University,
Department of Computer Science, April 1977.
[51] Peyton Jones, S. L. The Implementation of Functional Programming Languages. PrenticeHall, 1987.
[52] Polivka, R. P., and S. Pakin. APL: The Language and Its Usage. Prentice Hall, Englewood
Cliffs, New Jersey, 1975.
[53] Pountain, D., and D. May. A Tutorial Introduction to occam Programming. BSP Professional Books, London, 1988.
[54] Reynolds, J. C. Gedankena simple typeless language based on the principle of completeness and the reference concept. Communications of the ACM, 13:308319, May 1970.
[55] Robinson, J. A. Logic and logic programming. Communications of the ACM, pages 4065,
March 1992.
[56] Rockey, C. J. Structured PL/1 Programming. Wm. C. Brown, Dubuque, Iowa, 2nd edition,
1985.
[57] Sammet, J. E. Programming Languages: History and Fundamentals. Prentice Hall, Englewood Cliffs, New Jersey, 1969.
[58] Sethi, R. Programming Languages: Concepts and Constructs. Addison-Wesley, Reading,
Massachusetts, 1989.
[59] Slade, S. The T Programming Language. Prentice Hall, Englewood Cliffs, New Jersey, 1987.
[60] Standish, T. PPLan extensible language that failed. ACM SIGPLAN Notices, pages
144145, December 1971.

583

BIBLIOGRAPHY
[61] Stern, N., and R. Stern. Structured COBOL Programming. Wiley, New York, 1988.
[62] Stroustrup, B.
sachusetts, 1986.

The C++ Programming Language.

Addison-Wesley, Reading, Mas-

[63] Turner, David. An overview of Miranda. ACM SIGPLAN Notices, December 1986.
[64] United States Department of Defense. Reference Manual for the Ada Programming Language:
Proposed Standard Document, July 1980.
[65] van Wijngaarden, A., et al. Report on the algorithmic language ALGOL 68. Numerische
Mathematik, 14:79218, 1969.
[66] Wegbreit, B. Studies in Extensible Programming Languages. PhD thesis, Harvard University,
1970.
[67] Wegbreit, B. The treatment of data types in EL/1. Communications of the ACM, 17(5):251
264, May 1974.
[68] Wilensky, R. LISPcraft. W. W. Norton and Company, New York, 1984.
[69] Wirth, N. The design of a Pascal compiler. Software Practice and Experience, 1:309333,
1971.
[70] Wirth, N. The programming language Pascal. Acta Informatica, 1:3563, 1971.
[71] Wirth, N. Modula: A language for modular multiprogramming. Technical Report TR18,
ETH Institute for Informatics, March 1976.

584

BIBLIOGRAPHY

Appendix A

Exhibits Listed by Topic

A.1

Languages

A work of this scope can only scratch the surface of the many languages that it mentions. For
further information, the reader is invited to consult the works cited under each language heading.
Full references are listed in the bibliography. Only one or two references are listed for each language;
for many languages several excellent books are available.

A.1.1

Ada

Suggested references: Barnes [1989], Booth [1987].

3.3 Size and encoding specified separately in Ada.
3.5 Syntax for verbs in Pascal and Ada.
3.15 Lexical scope delimiters in Ada.
3.19 Comments terminating at end of line.
6.24 Dynamic allocation in Ada.
7.15 Mixed variable and constant initializations in Ada.
9.1 An Ada function with a default value for one parameter.
9.2 Correspondence-by-keyword in Ada.
10.14 Use of the general conditional statement in Ada.
10.26 The general loop in Ada.
11.10 Loop labels in Ada.
11.19 Handling an integer overflow exception.
14.11 Diagrams of array type-objects.
585

586

APPENDIX A. EXHIBITS LISTED BY TOPIC

14.16 Implementation of sets in Ada.

14.17 Selecting a member of a set.
14.25 The discriminated variant record in Ada.
14.26 Legal and illegal uses of a discriminated variant record.
14.28 Making a record object from its components in Ada.
14.40 Some of the type operations supported by Ada.
15.21 An Ada derived type is a new domain.
15.22 Type casts in Ada.
15.23 Domain compatibility rules in Ada.
15.26 Using casts in a mapped domain.
15.28 Using casts in Ada.
15.36 Ada integer type specification.
15.37 Ada real type specification.
15.38 Mixed type addition in Ada.
16.6 Defining a stack package in Ada.
16.7 Using an Ada package.
17.8 Definition of new methods for an Ada operator
17.9 Extensions of Ada operators for new types.
17.19 A parameterized domain in Ada: vector of numbers.
17.20 Instantiating a generic template in Ada.
17.21 Declaration of a generic stack package in Ada.
17.22 Instances of the stack package.
18.13 Manual polymorphic dispatching in Ada.

A.1.2

APL

Suggested references: Polivka and Pakin [1975]. For APL2, see Brown, Pakin, and Polivka [1988].
2.21 Strange but legal GOTOs in APL.
2.22 A computed GOTO idiom in APL.
3.7 Syntax for verbs in APL.
3.16 Comment lines in older languages.
6.9 Languages where assignment is a function.

A.1. LANGUAGES
7.8 Names, types, and binding in APL.
7.9 Dynamic binding in APL.
8.1 Syntax for expressions.
8.7 Varying rules for parsing expressions.
8.8 Varying syntax for the use of operators.
9.27 Mapping functions in APL.
10.33 Use of implicit looping in APL.
12.2 Primitive operations for arrays in APL.
14.31 Value constructors in APL.
14.34 APL selection and dereferencing.
14.35 Constructing matrices in APL.
14.36 Selecting a submatrix in APL.
14.37 APL selection functions.
14.38 Using value constructors in APL.
15.3 Domains are checked by APL primitive operators.
15.4 A simple APL function.
15.5 APL does not check domains of programmer-defined functions.

A.1.3

C++

Suggested reference: Stroustrup [1986].

3.19 Comments terminating at end of line.
16.8 Defining a class in C++.
16.9 Instantiating a C++ class.
16.10 A represented domain in C++.
16.11 Friend classes.
18.8 Deriving classes in C++.
18.9 Calling a C++ function with two defining methods.
18.10 Conversion functions in C++.
18.11 Coercion with a programmer-defined function.
18.12 Dispatch rules for C++ overloaded functions.

587

588

A.1.4

APPENDIX A. EXHIBITS LISTED BY TOPIC

C and ANSI C

Suggested references: Kernighan and Ritchie [1988], Harbison and Steele [1991].
1.2 Violations of redundancy rules in ANSI C.
2.9 Addition represented diffusely or coherently.
2.10 A coherently but not contiguously represented object.
2.12 Bitwise and logical operations.
2.13 Useful pointer operations supported in C.
3.2 Size and encoding bundled in C.
3.13 Tokens used to delimit program scopes.
3.17 Comment beginning and end delimiters.
3.20 Definition of a simple macro in C.
3.21 Use of a simple macro in C.
3.22 A nest of macros in C.
3.23 Use of the simple PRINTX macro.
3.24 A macro with parameters in C.
3.25 Use of the print macro with parameters.
6.4 Initial value declarations in C.
6.6 Initializing and copying a compound object in K& R C.
6.7 Initializing and copying a compound object in ANSI C.
6.9 Languages where assignment is a function.
6.10 Assignment as a function in C.
6.15 Pointer assignments with dereference in C.
6.16 Pointer assignment without dereference in C.
6.22 Dynamic allocation in C.
7.4 Predefined names in C.
7.17 Constant declarations in C.
8.7 Varying rules for parsing expressions.
8.15 Expressions whose values are indeterminate in C.
9.6 A call on a variable-argument function in C.
9.20 Pointer parameters in C.

A.1. LANGUAGES
9.21 Call-by-pointer using & in C.
9.22 Assignments using pointers in C.
9.23 Call-by-pointer using pointer variables in C.
9.24 Stack diagram for call-by-pointer.
9.25 Creation of two dangling pointers in C.
9.26 Creating a dangling pointer.
10.6 Syntax and uses of a conditional expression in C.
10.7 Syntax and uses of a conditional statement in C.
10.19 The C switch statement without and with break.
10.30 Uses of the iteration element in C.
10.31 Many ways to write the same iteration.
10.32 Equivalent loop forms in C.
10.34 Explicit looping in ANSI C takes more writing.
11.9 The input loop implemented using break in C.
11.17 I/O with error checking in C.
11.18 I/O without error checking in C.
12.1 A nonfunctional function in C.
13.14 Euclids algorithm in C.
13.16 Quicksort in C.
13.17 Notes on the C quicksort.
14.5 A C pointer and its type-object.
14.12 Diagram of a zero-based array type-object.
14.13 A type-object for a null terminated string.
14.21 Using a pointer in C to emulate with.
14.39 The lone type operation in C.
15.7 Pointer assignments in K& R C.
15.10 Merging the domains truth value and integer.
15.11 Many misinterpretations!
15.12 The bane of the former Pascal programmer.
15.17 Merged domains defined using typedef in C.
15.30 Reference level in C.

589

590

APPENDIX A. EXHIBITS LISTED BY TOPIC

15.33 Syntax and semantics for casts and conversions in C.

15.34 Coercion complications in C.
15.35 Explicit dereference and reference coercion in C.
15.42 Union data type in C.
15.43 Using pointers in C to attach two types to an object.
16.2 Header file: modules.h.
16.3 Main module: sumup.c.
16.4 Subroutine module: reduce.c.
16.5 The makefile for the sumup program.
17.11 A parameterized domain in C: array of floats.
17.16 Definition of a generic domain in C.
17.17 Using a generic type in C.
17.18 Specific code results from preprocessing.

A.1.5

FORTH

Suggested reference: Brodie [1987].

2.4 A low-level computer representation of a date.
2.9 Addition represented diffusely or coherently.
3.17 Comment beginning and end delimiters.
4.22 Definition in FORTH of the semantics for arrays.
6.12 Explicit dereferencing in FIG FORTH.
6.20 Dynamic allocation in FORTH.
7.6 Names, types, and bindings in FIG FORTH.
7.16 Variable and constant declarations in FIG FORTH.
8.7 Varying rules for parsing expressions.
8.8 Varying syntax for the use of operators.
9.3 A FORTH function with a variable number of arguments.
9.4 Meaningful calls on a variable-argument function in FORTH.
9.5 Meaningless calls on a variable-argument function in FORTH.
10.24 The general loop in FORTH-83.
11.1 Some Languages with no GOTO and no labels.

A.1. LANGUAGES
14.19 Defining a record type in FORTH.
15.10 Merging the domains truth value and integer.

A.1.6

FORTRAN

Suggested reference: Meissner and Organick [1980]

2.8 A stack represented diffusely in FORTRAN and Pascal.
2.23 An archaic language feature in FORTRAN.
3.10 Field definitions in FORTRAN.
3.11 Scope delimiters in FORTRAN.
3.16 Comment lines in older languages.
6.3 Initial value declarations in FORTRAN.
7.10 Sharing through EQUIVALENCE in FORTRAN.
7.14 Constant declarations in FORTRAN 77.
7.27 Named common in FORTRAN.
7.28 A diagram of common storage in FORTRAN.
11.4 An input loop with an exit in the middle.
14.8 A three dimensional array in FORTRAN.
15.39 EQUIVALENCE in FORTRAN.
17.10 An implicit parameterized domain in FORTRAN.
18.4 Coercions built into FORTRAN.

A.1.7

LISP

Suggested references: Kessler [1988], Wilensky [1984].

2.10 A coherently but not contiguously represented object.
2.16 Syntax for lambda expressions in LISP.
2.17 A LISP function call with poor coherence.
2.18 A LISP function call with good coherence.
3.6 Syntax for verbs in LISP.
3.13 Tokens used to delimit program scopes.
3.19 Comments terminating at end of line.
6.9 Languages where assignment is a function.

591

592

APPENDIX A. EXHIBITS LISTED BY TOPIC

6.21 Dynamic allocation in LISP.

7.23 Dynamic scoping in LISP.
7.24 Diagram of bindings in dynamic scoping.
8.1 Syntax for expressions.
9.28 Some applications of mapcar in LISP.
10.13 Use of the general conditional expression, cond in LISP.
12.7 A LISP function nest emulating a sequence.
15.6 Types and domain checking in LISP.

A.1.8

Miranda

Suggested references: Peyton-Jones [1987], Turner [1986].

9.30 The type of reduce in Miranda.
9.31 Partially closing a function to make new functions.
12.8 Defining local symbols in Miranda.
12.9 A simple Miranda script.
12.10 Denoting lists in Miranda.
12.11 Using list operations in Miranda.
12.12 A guarded expression.
12.13 Where takes the place of assignment.
12.15 The reduce function in Miranda.
12.16 The syntax for a list comprehension.
12.17 A list comprehension.
12.18 Some list comprehensions.
12.19 Quicksort in Miranda.
12.20 A memoized function.
18.14 A tree type in Miranda.
18.15 A tree with polymorphic nodes.
18.16 Polymorphic dispatching in Miranda.
18.17 Dispatching a Miranda call using data values.
18.18 Pattern construction in Miranda.
18.19 Binding during the pattern match.

A.1. LANGUAGES
18.20 Miranda type expressions.
18.21 The ADT stack in Miranda.
18.23 Generic calls.

A.1.9

Pascal

Suggested reference: Cooper [1983].

1.3 Summary of Turbo Pascal deviations from the standard.
1.4 The definition of mod in Standard Pascal.
1.5 The definition of mod in OSS Pascal for the Atari ST.
1.6 The definition of mod in Turbo Pascal for the IBM PC.
2.3 High-level computer representations of a date.
2.5 The structure of a table expressed implicitly.
2.7 A stack represented coherently in Pascal.
2.8 A stack represented diffusely in FORTRAN and Pascal.
2.11 A basic type not supported by Pascal.
2.14 A meaningless operation prohibited in Pascal but not in C.
2.15 Poor lexical coherence for declarations and code in Pascal.
3.5 Syntax for verbs in Pascal and Ada.
3.9 The domains of some Pascal functions.
3.13 Tokens used to delimit program scopes.
3.14 Nested unlabeled scopes in Pascal.
3.17 Comment beginning and end delimiters.
3.18 An incomplete comment swallowing an instruction.
4.4 EBNF production rules for parts of Pascal.
4.5 Parsing a simple Pascal program.
4.6 Syntax diagram for program.
4.7 Syntax diagrams for statement.
6.5 Initializing and copying a compound object in Pascal.
6.11 Dereferencing by context in Pascal.
6.13 Pointer assignments in Pascal.
6.14 Pointer structures sharing storage.

593

594

APPENDIX A. EXHIBITS LISTED BY TOPIC

6.18 A recursive function in Pascal.

6.23 Dynamic allocation in Pascal.
7.1 Bad names.
7.2 Good names.
7.3 Predefined names in Pascal.
7.11 A variant record used to compute a hash index.
7.12 Diagrams of a Pascal variant record.
7.13 Constant declarations in Pascal.
7.18 Nested scopes in Pascal.
7.19 Complete names in Pascal.
7.20 A stack diagram for nested scopes.
7.21 A Pascal program demonstrating block structure.
7.22 A symbol table with block-structured bindings.
7.25 Lexically scoped functions in Pascal.
7.26 Lexical scoping in Pascal.
8.1 Syntax for expressions.
8.7 Varying rules for parsing expressions.
8.8 Varying syntax for the use of operators.
9.7 A procedure using call-by-value in Pascal.
9.8 Result of parameter binding by value in Pascal.
9.9 Result of execution with value parameters.
9.10 Parameter binding by name.
9.13 A procedure using call-by-reference in Pascal.
9.14 VAR parameter binding in Pascal.
9.15 Result of execution with a VAR parameter.
9.16 A swap routine using reference parameters in Pascal.
9.17 The implementation of reference parameters in Pascal.
9.18 A swap routine using pointer parameters in Pascal.
9.19 Storage for swap2.
9.29 A Pascal version of reduce.
9.32 We cant make closures in Pascal.

A.1. LANGUAGES
10.11 A structured conditional statement in Pascal.
10.15 Using Pascal to emulate ELSEIF.
10.18 The Pascal CASE with no default clause.
10.23 REPEAT and WHILE loops in Pascal.
10.28 Pascal computes a tripcount.
11.5 The Pascal input loop with a redundant priming read.
11.6 The Pascal input loop with an extra test.
11.7 Some Pascal code that will stop with a run-time error.
11.16 The factorial function in Pascal.
12.3 Recursive implementation of repetition.
12.4 Iterative implementation of repetition.
12.6 A sequence of statements in Pascal.
12.14 Assignment used to name subexpressions in Pascal.
14.4 A constrained type in Pascal.
14.7 Effective address computation for Pascal arrays.
14.9 Basic and sugared notation for arrays in Pascal.
14.10 Effective address computation for a four-level type.
14.14 A set type in Pascal.
14.15 Type-objects for set types.
14.17 Selecting a member of a set.
14.20 Partial pathnames using with in Pascal.
14.29 Pascal selection functions.
14.30 Using selection functions in Pascal.
14.33 Pascal selection and dereferencing.
15.8 Diagrams of the domain relationships in Pascal.
15.13 Domains mapped onto integers in Pascal.
15.14 Two types with the same description.
15.15 A type constructor in Pascal.
15.16 Incompatible domains in Pascal.
15.18 Merged domains in Pascal.
15.19 Function calls with merged domains in Pascal.

595

596

APPENDIX A. EXHIBITS LISTED BY TOPIC

15.25 Pascal casts for integer-mapped domains.

15.27 Implementing a cast in Pascal.
15.31 Coercions and explicit conversions in Pascal.
15.40 Variant records in Pascal.
15.41 Use of a Pascal nondiscriminated variant record.
17.5 Pascal definition of push on an array_stack.
17.6 Pascal definition of push for a list_stack.
17.12 A Pascal function with a conformant array parameter.
17.13 A parameterized type.
17.14 A parameterized linked list type.
17.15 Pascal definition of push on a stack of characters.
18.5 Defining a subdomain in Pascal.

A.1.10

Prolog

Suggested reference: Clocksin and C. S. Mellish [1981].

10.36 A data base of Prolog relationships.
10.37 Answering a Prolog query.
11.1 Some Languages with no GOTO and no labels.
13.10 Prolog rules and facts.
13.11 Prolog queries.
13.12 Interpreting Prolog rules.
13.13 Euclids algorithm in Prolog.
13.15 Quicksort in Prolog.
13.18 Cutting off a search.
13.19 Data for the probation predicate.
13.20 The not in Prolog.
13.21 The trouble with not.

A.1. LANGUAGES

A.1.11

597

Scheme and T

Suggested references: Abelson et al. [1985], Abelson and Sussman [1985], Slade [1987].
11.11 Packaging a continuation in Scheme.
11.12 Call/cc used in Scheme to break out of a loop.
11.13 Testing the break routine.
11.14 Using call/cc to make a loop.
11.15 A Pascal-like implementation of factorial, using a continuation.

A.1.12

Other Languages

Suggested references: ALGOL 60Naur and Woodger [1963]; BASICKemeny [1985]; COBOL
Stern and Stern [1988]; HaskellFasel and Hudak [1992]; IconGriswold and Griswold [1983];
LOGOAllen, Burke, and Johnson [1983]; PL/1Rockey [1985]; PostScriptAdobe Systems [1990];
SNOBOL 4Griswold, Poage, Polonsky [1971]; Smalltalk-80Goldberg and Robson [1983]; Turing
Holt et al. [1988].
2.6 COBOL: The structure of a table expressed explicitly.
2.19 Poor lexical coherence in SNOBOL.
2.20 BASIC: GOTOs and statement ordering both use line numbers.
3.13 Tokens used to delimit program scopes. (PL/1)
3.16 Comment lines in older languages.
3.17 Comment beginning and end delimiters. (PL/1)
5.9 Size and encoding specifications in COBOL.
5.13 Declaration and use of a record in COBOL.
5.14 Using the LIKE attribute in PL/1.
6.25 Dynamic allocation in Turing.
10.9 The IF in BASIC.
10.17 The COBOL EVALUATE statement.
10.35 Implicit looping in dBMAN.
11.1 Some Languages with no GOTO and no labels. (Icon, Prolog, Turing)
11.2 The unique control structure in SNOBOL-4.
15.2 Domain relationships in ALGOL.
15.29 Mangled meaning in PL/1.
15.32 COBOL fixed point size specifications.

598

A.2
A.2.1

APPENDIX A. EXHIBITS LISTED BY TOPIC

Concepts
Application Modeling, Generics, and Polymorphic Domains

5.7 Specific types and generic domains.

5.8 Specific types are instances of generic domains.
15.1 Operations defined for assembly language domains.
15.9 Definition of the ADT array_stack.
15.14 Two types with the same description.
15.16 Incompatible domains in Pascal.
15.20 Creating a compatible type name by mapping.
15.24 A hierarchy of mapped domains.
15.26 Using casts in a mapped domain.
17.1 Generic domains and functions on them.
17.2 A generic function as a collection of methods.
17.3 A function with generic formal parameters.
17.4 A generic function defined by a single method.
17.7 An overloaded function, +.
17.13 A parameterized type.
17.14 A parameterized linked list type.
18.1 Mode graph for the generic domain Number.
18.2 Diagram of an instantiated mode.
18.3 Exponential explosion of method definitions.
18.6 Diagram of a domain hierarchy.
18.7 A group of related domains.
18.22 Mode graph for points on a plane.
18.23 Generic calls.
18.24 Using generic pointers.
18.25 What type is returned?

A.2. CONCEPTS

A.2.2

Control Structures

10.1 Diagram of a sequence of statements.

10.2 A sufficient set of basic control structures.
10.3 Diagram of a subroutine call with parameters.
10.4 Diagram of the computers primitive GOTO instruction.
10.5 Diagram of the primitive conditional.
10.6 Syntax and uses of a conditional expression in C.
10.7 Syntax and uses of a conditional statement in C.
10.8 Syntax and diagram of the simple IF-GO.
10.10 Diagrams of the structured conditional.
10.12 Diagram of the generalized structured conditional.
10.16 The CASE statement.
10.20 The simplest loop: an infinite loop.
10.21 The simplest finite loops, formed with a GOTO.
10.22 Structured conditional loops.
10.25 The general loop.
10.27 The counted loop.
10.29 The iteration element.
11.4 An input loop with an exit in the middle.
11.8 Structured conditional loops with BREAK.
11.9 The input loop implemented using break in C.
11.11 Packaging a continuation in Scheme.
11.12 Call/cc used in Scheme to break out of a loop.
11.13 Testing the break routine.
11.14 Using call/cc to make a loop.
11.15 A Pascal-like implementation of factorial, using a continuation.
12.3 Recursive implementation of repetition.
12.4 Iterative implementation of repetition.
12.5 Stack diagrams for a tail recursion and a loop.

599

600

A.2.3

APPENDIX A. EXHIBITS LISTED BY TOPIC

Data Representation

2.1 Modeling a charge account and relevant processes.

2.2 Representations of a date in English.
3.1 Two overlapping objects (linked lists).
5.1 Packed-decimal encoding.
5.2 Representable values for signed and unsigned integers.
5.3 IEEE floating-point formats.
5.4 Floating point on the SparcStation.
5.5 A hierarchy of abstractions.
5.6 Type, type name, and type description.
5.12 An emulation of floating point.
6.1 Representing objects.
6.2 Values, variables, and pointers.
6.14 Pointer structures sharing storage.
7.1 Bad names.
7.2 Good names.

A.2.4

History

2.24 The development of procedural languages.

2.25 The development of functional languages.
4.1 Foundations of computer science.
4.2 Formal semantic specification.

A.2.5

Lambda Calculus

4.10 Lambda calculus formulas.

4.11 Lambda calculus names and symbols.
4.12 Omitting parentheses when writing lambda calculus formulas.
4.13 Lambda expressions for TRUE and FALSE.
4.14 Dissection of a lambda expression.
4.15 Lambda calculus formulas that represent numbers.
4.16 Basic arithmetic functions.

A.2. CONCEPTS
4.17 Some lambda applications.
4.18 Reducing (S 1).
4.19 Reducing (ZeroP 0).
4.20 Reducing (ZeroP 1).
4.21 A formula with three redexes.

A.2.6

Language Design and Specification

1.1 Redundancy in English.

3.4 A type hierarchy with inheritance in English.
3.8 The domain of a verb in English.
3.12 Labeled scopes.
5.10 Costs of implementing a primitive type.
5.11 Costs of omitting a primitive type.
8.3 Unambiguous ways to write an arithmetic expression.
8.5 Bracketing used to permit variable-argument expressions.
8.6 Infix notation is ambiguous without precedence.

A.2.7

Logic

2.12 Bitwise and logical operations.

13.1 Predicates and arity.
13.2 Quantified predicates.
13.3 Quantified formulas.
13.4 Semantics of not, or, and, and implies.
13.5 Logical symbols and sentences.
13.6 A simple theory: the Theory of Ugliness.
13.7 A proof that 32,769 is a good number.
13.8 Ways to write a Horn clause.
13.9 Unifying clauses.

601

602

A.2.8

APPENDIX A. EXHIBITS LISTED BY TOPIC

Translation, Interpretation, and Function Calls

4.3 The compiler is produced from the language definition.

4.8 Translation: from source to machine code.
4.9 Communication between modules.
6.17 The structure of the run-time stack.
6.19 Stack frames for recursive calls.
7.5 A symbol table with static bindings.
7.7 A symbol table with dynamic binding.
7.20 A stack diagram for nested scopes.
7.22 A symbol table with block-structured bindings.
7.24 Diagram of bindings in dynamic scoping.
8.2 Nested function calls and parenthesized infix expressions.
8.4 Evaluation of a postfix expression using a stack.
8.9 A computation tree.
8.10 The substitution rule.
8.11 Lazy evaluation of AND.
8.12 Lazy evaluation of OR.
8.13 Extra code needed with inside-out evaluation.
8.14 Simpler control structures with outside-in evaluation.
9.10 Parameter binding by name.
9.11 Result of parameter binding by name.
9.12 Result of execution with by-name binding.
9.19 Storage for swap2.
9.26 Creating a dangling pointer.

A.2.9

Types

14.1 Object diagrams.

14.2 An integer object and its type.
14.3 A type-object for an enumerated type.
14.6 Formulas for the effective address computation.
14.10 Effective address computation for a four-level type.

A.2. CONCEPTS
14.11 Diagrams of array type-objects.
14.12 Diagram of a zero-based array type-object.
14.13 A type-object for a null terminated string.
14.15 Type-objects for set types.
14.18 A type-object for a record.
14.22 Alternative syntax for part selection.
14.23 A type-object for a free union type.
14.24 A type-object for a discriminated union type.
14.26 Legal and illegal uses of a discriminated variant record.
14.27 Making a record object from its components.
14.31 Value constructors in APL.
14.32 Selection and dereferencing.

603

Index
abortive exit 317
abstract data type 446
abstract function 515
abstraction 126
abstraction 9
activation record 160
active block 160
ad hoc generic 513
ad hoc polymorphic domain 562
Ada derived type 457
Ada generic 532
additional binding 206
ADT 446
alias 52
allocation process 145
ambiguity 8
application 97
arity 363
array bounds 400
array dimension 400
ASCII 119
assertion, Prolog 377
assignment 148
atom, Prolog 375
automatic conversion 466
axioms 365
backtrack, Prolog 380
bad restriction 28
base type of a pointer 398
base type of a set 408
base type of an array 400
BCD 119
beta redex 100

beta reducible expression 100

binary operator 218
binding a name 180
binding at source time 530
binding time 431
birth of name 197
birth of object 158
bitwise operators 27
block exit time 207
block structure 195
block 195
block 8
BNF 82
bound variable 99
business computer 120
call-by-address 244
call-by-constant 247
call-by-name order 107
call-by-name 242
call-by-need 227
call-by-pointer 251
call-by-return 247
call-by-value-and-return 248
call/cc 322
call-with-current-continuation 322
captured reference 225
checkpoint 322
class member 500
clausal logic 366
closure 236
closure 260
coherent assignment 149
coherent representation 22
604

605

INDEX
coherent representation 23
combinatorial explosion 369
compiler bug 13
complete name 194
complete system 96
complete 368
comprehension 355
computation tree 218
computer representation 20
conclusion 366
concrete type 126
conditional expression 275
conditional sentence 366
conditional statement 275
conformant arrays 527
conservative conversion 465
constructor 423
contents of an object 144
context free languages 5
continuation 321
control diagram 274
control frame 279
control frame 289
control structure 268
counted loop 293
curry a function 103
currying 257
cut operation, for resolution 370
cut, Prolog 385
dangling reference 169
data hiding 506
data type 126
deallocation 158
death of an object 53
death of name 197
death of object 158
decidable 368
declaration 179
deduction 365
defaults 9
defining occurrence 195

demotion cast 462

demotion 465
denotational semantics 340
denotes 339
dereference 152
destructive assignment 148
dictionary 179
diffuse representation 22
dispatch a call 517
dispatch 42
dispatcher 517
dispatcher 543
distinguishable domains 448
domain checking 444
domain mapping 450
domain 436
dual type 480
dyadic operator 218
dynamic binding 180
dynamic binding 205
dynamic link 161
dynamic link 196
dynamic parent 161
dynamic scope 198
dynamic scoping 200
dynamic storage 159
EBCDIC 119
encapsulate 490
encoding conversion 465
encoding conversion 466
enumerated type 396
environment 179
error propagation 329
evaluation, inside-out 224
evaluation, outside-in 225
exception handler 331
exception 327
expansion of a non-terminal 82
explicit representation 22
expressive language 22
extension 107

606
extension 12
extent of an object 158
external domain 448
external object 144
external scope 194
first class object 103
first class objects 52
first-order predicate calculus 362
fixed point number 121
fixed point 121
fixed point 340
flexible language 10
flexible language 28
formal language 80
free union 481
free variable 99
freelist 164
front end processor 277
function domain 57
function method 517
function range 58
functional composition 103
functional language 148
functional language 45
functional 561
functor 375
garbage collection 47
garbage collector 169
garbage 146
garbage 165
general loop 292
generate a program 85
generator 356
generic domain 127
generic domain 512
generic function 516
generic function 523
generic object 9
generic package 448
generic package 532
generic packages 431

INDEX
global scope 194
goal 378
good restriction 28
guarded expression 349
heap allocation 162
heap storage 159
hidden data 492
higher-order function 103
higher-order function 254
Horn clause 372
hyperresolution 372
identifier 194
immortal object 158
implicit representation 22
independent domains 448
indeterminate result 230
index type 400
indirectly bound 245
infinite list 351
information-losing conversion 465
inherit 549
inheritance 42
inheritance 54
inherited function 556
inherited 549
initialization 146
in-line code 507
instance of a generic domain 127
instantiate a package 431
instantiate, in C++ 502
instantiate 168
instantiate 362
instantiated type 529
instantiation 9
internal domain 448
internally merged domains 448
invisible name 198
IROV 190
iteration element 298
labeled scope 64
lambda calculus formulas 97

607

INDEX
lambda calculus variable 97
lambda evaluation 226
lambda expression body 97
lambda expression parameter 97
lambda expression 97
lambda function call syntax 217
lambda reduction rule 103
lambda substitution 225
language extension 12
lazy evaluation 227
lazy evaluation 227
lex 59
lexer generator 60
lexer 59
lexical analysis 59
lexical ancestor 160
lexical delimiter 59
lexical parent 160
lexical scoping 196
lexical token 59
lifetime of object 158
LISP atom 192
LISP lambda expression 32
list comprehension 356
list specification 376
list 192
local scope 194
logic language 47
logical consequence 368
logical operators 27
loop variable 293
makefile 492
masked definition 197
meaning of a name 176
meaning of a program 90
meaning of an object 148
meaning of code 5
meaning of expression 218
memory fragmentation 159
memory management 159
message 500

metalanguage 59
method 42
method 500
mode graph 542
mode 542
modularity 194
module 490
monadic operator 218
multiple assignment 151
multiply bound name 180
name binding 176
name conflict 225
name refers to 180
naming conflict 193
N -ary 363
necessary control structures 344
nested lifetimes 160
nonstandardized language 11
normal exit 317
normal form 100
normal form 102
normal form 99
normal function call-syntax 217
normal order 107
N -step proof 366
nuance 10
object 502
object-oriented language 193
object-oriented language 42
one-in/one-out 271
opaque derived mode 554
open list 376
operands 55
overloaded name 522
parallel arrays 25
parallel evaluation 349
parallel language 45
parameter mode 247
parameterized generic domain 512
parameterized domain 529
parameterized module 531

608
parameterized type expression 527
parse tree 218
parse tree 81
parse tree 86
parser generator 80
parser 80
partial parameterization 260
p-code 90
pointer assignment 154
pointer variable 146
polymorphic object 561
polymorphic type 513
polymorphic type 561
portable code 90
portable program 39
postfix order 219
powerful language 28
predicate 363
prefix order 219
prefix syntax 218
premise 366
priming read 317
primitive control 268
primitive 437
private symbol 491
procedural language 42
process 339
program object 144
promises 544
promotion cast 462
promotion 465
proof system 365
proof 366
proposition 363
public symbol 491
pure functional language 341
pure value 144
qualifier list 356
rank of an array 400
recurrence equation 339
redex 100

INDEX
reduction 100
reference constructor 425
reference conversion 466
reference 144
refutation 366
relatively global scope 194
representation of an object 20
represented by 450
resolution deduction 372
resolution proof 372
resolution 366
resolution 370
rule, Prolog 376
run-time dispatcher 543
run-time stack 160
satisfy 363
scientific computer 120
scope of a name 194
scope of quantifier 364
scope 64
scope-resolution operator 195
selection function 423
semantic basis 22
semantic basis 96
semantic intent 21
semantic intent 5
semantic rules 5
semantic validity 21
semiindependent domains 448
sentence 363
short circuit evaluation 228
simple object 128
size conversion 465
size conversion 465
size 128
slice of an array 403
spaghetti code 312
species of a generic domain 127
specific domain 512
specific type 126
specification 338

609

INDEX
stack frame 160
static binding 180
static binding 181
static link 160
static link 196
static storage 158
storage compaction 136
storage object 144
stream 94
strict evaluation 227
strict evaluation 230
strict evaluation 350
strongly typed language 445
strongly typed language 445
structure of a program 90
structured editor 67
structured programming 271
subdomain 549
subroutine call 272
subroutine return 272
superdomain 549
symbol table 179
syntactic analysis 86
syntactic category 80
syntactic sugar 403
syntax 5
tail recursion 344
term 363
terminated string 406
theorem 366
theory 366
tripcount 297
tuple 351
type cast 460
type cast 460
type checking 444
type coercion 460
type coercion 466
type compatible 445
type constructor 453
type conversion 460

type conversion 465

type declaration 126
type description 126
type error 445
type match 445
type name 126
type of a type 394
type predicate 136
type predicate 431
type tag 455
type 137
type-object 394
unary operator 218
unbound name 180
unbound symbol 225
undefined name 180
undefined value 146
undefined value 295
unification 370
union types 419
unit record equipment 118
universe of discourse 362
unlabeled scope 64
Until test 291
upward compatibility 39
use of an identifier 195
use of parameter 234
valid representation 21
valid sentence 366
value constructor 425
value of an object 144
VAR parameter passing 244
variable declaration 51
variable 146
virtual class 554
virtual function 515
visibility 552
visible name 198
While test 291
word alignment 412
zero-based subscripting 400

610
ZF expression 356

INDEX