COMP 348 Principles of Programming Languages

Lecture Notes
Peter Grogono
These notes may be copied for students who are taking either
COMP 348 Principles of Programming Languages or COMP 6411
Comparative Study of Programming Languages.
c (Peter Grogono, 2002, 2003, 2004
Department of Computer Science and Software Engineering
Concordia University
Montreal, Quebec
CONTENTS ii
Contents
1 Introduction 1
1.1 How important are programming languages? . . . . . . . . . . . . . . . . . . . . . 1
1.2 Features of the Course . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2 Anomalies 3
3 The Procedural Paradigm 7
3.1 Early Days . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.2 FORTRAN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.3 Algol 60 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.4 COBOL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.5 PL/I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.6 Algol 68 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.7 Pascal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.8 Modula–2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.9 C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.10 Ada . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4 The Functional Paradigm 20
4.1 LISP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.1.1 Basic LISP Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.1.2 A LISP Interpreter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.1.3 Dynamic Binding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2 Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.3 SASL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.4 SML . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.5 Other Functional Languages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5 Haskell 34
5.1 Getting Started . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
5.1.1 Hugs Commands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.1.2 Literate Haskell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5.1.3 Defining Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.1.4 Syntax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.1.5 Tracing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.2 Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.2.1 Lists . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.2.2 Lazy Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
CONTENTS iii
5.2.3 Pattern Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.2.4 Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.2.5 Building Haskell Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.2.6 Guards . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.3 Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.3.1 Using Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.3.2 Failure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.3.3 Reporting Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.3.4 Standard Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.3.5 Standard Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.3.6 Defining New Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.3.7 Defining Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.3.8 Types with Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.4 Input and Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.4.1 String Formatting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.4.2 Interactive Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.5 Reasoning about programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
6 The Object Oriented Paradigm 77
6.1 Simula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.2 Smalltalk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.3 CLU . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
6.4 C++ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.5 Eiffel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.5.1 Programming by Contract . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.5.2 Repeated Inheritance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.5.3 Exception Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.6 Java . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.6.1 Portability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.6.2 Interfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.6.3 Exception Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.6.4 Concurrency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
6.7 Kevo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
6.8 Other OOPLs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.9 Issues in OOP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.9.1 Algorithms + Data Structures = Programs . . . . . . . . . . . . . . . . . . 93
6.9.2 Values and Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
6.9.3 Classes versus Prototypes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
CONTENTS iv
6.9.4 Types versus Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
6.9.5 Pure versus Hybrid Languages . . . . . . . . . . . . . . . . . . . . . . . . . 96
6.9.6 Closures versus Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.9.7 Inheritance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.10 A Critique of C++ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.11 Evaluation of OOP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
7 Backtracking Languages 106
7.1 Prolog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
7.2 Alma-0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
7.3 Other Backtracking Languages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
8 Structure 118
8.1 Block Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
8.2 Modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
8.2.1 Encapsulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
8.3 Control Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
8.3.1 Loop Structures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
8.3.2 Procedures and Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
8.3.3 Exceptions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
9 Names and Binding 125
9.1 Free and Bound Names . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
9.2 Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
9.3 Early and Late Binding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
9.4 What Can Be Named? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
9.5 What is a Variable Name? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
9.6 Polymorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
9.6.1 Ad Hoc Polymorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
9.6.2 Parametric Polymorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
9.6.3 Object Polymorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
9.7 Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
9.8 Scope and Extent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
9.8.1 Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
9.8.2 Are nested scopes useful? . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
9.8.3 Extent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
LIST OF FIGURES v
10 Abstraction 138
10.1 Abstraction as a Technical Term . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
10.1.1 Procedures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
10.1.2 Functions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
10.1.3 Data Types. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
10.1.4 Classes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
10.2 Computational Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
11 Implementation 144
11.1 Compiling and Interpreting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
11.1.1 Compiling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
11.1.2 Interpreting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
11.1.3 Mixed Implementation Strategies . . . . . . . . . . . . . . . . . . . . . . . . 146
11.1.4 Consequences for Language Design . . . . . . . . . . . . . . . . . . . . . . . 147
11.2 Garbage Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
12 Conclusion 150
References 151
A Abbreviations and Glossary 156
B Theoretical Issues 159
B.1 Syntactic and Lexical Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
B.2 Semantics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
B.3 Type Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
B.4 Regular Languages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
B.4.1 Tokens . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
B.4.2 Context Free Grammars . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
B.4.3 Control Structures and Data Structures . . . . . . . . . . . . . . . . . . . . 162
B.4.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
C Haskell Implementation 164
List of Figures
1 The list structure (A (B C)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2 A short session with Hugs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3 Hugs commands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4 Parameters for :set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5 A Literate Haskell program . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
LIST OF FIGURES vi
6 Pattern matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
7 Parsing precedence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
8 Operator precedence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
9 Standard Class Inheritance Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
10 The Class Frac . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
11 Trying out type Frac . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
12 Class BTree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
13 The fringe of a tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
14 Key-value tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
15 Proof of ancestor(pam, jim) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
16 Effect of cuts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
17 REs and Control Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
18 REs and Data Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
19 The graph for succ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
COMP 348 Principles of Programming Languages
1 Introduction
In order to understand why programming languages (PLs) are as they are today, and to predict
how they might develop in the future, we need to know something about how they evolved. We
consider early languages, but the main focus of the course is on contemporary and evolving PLs.
The treatment is fairly abstract. We not discuss details of syntax, and we try to avoid unproductive
“language wars”. Here is a useful epigram for the course:
C and Pascal are the same language. Dennis Ritchie.
We can summarize our goals as follows: if the course is successful, it should help us to find useful
answers to questions such as these.
What is a good programming language?
How should we choose an appropriate language for a particular task?
How should we approach the problem of designing a programming language?
1.1 How important are programming languages?
As the focus of practitioners and researchers has moved from programming to large-scale software
development, PLs no longer seem so central to Computer Science (CS). Brooks (1987) and others
have argued that, because coding occupies only a small fraction of the total time required for
project development, developments in programming languages cannot lead to dramatic increases
in software productivity.
The assumption behind this reasoning is that the choice of PL affects only the coding phase of the
project. For example, if coding requires 15% of total project time, we could deduce that reducing
coding time to zero would reduce the total project time by only 15%. Maintenance, however, is
widely acknowledged to be a major component of software development. Estimates of maintenance
time range from “at last 50%” (Lientz and Swanson 1980) to “between 65% and 75%” (McKee
1984) of the total time. Since the choice of PL can influence the ease of maintenance, these figures
suggest that the PL might have a significant effect on total development time.
Another factor that we should consider in studying PLs is their ubiquity: PLs are everywhere.
The “obvious” PLs are the major implementation languages: Ada, C++, COBOL, FOR-
TRAN, . . . .
PLs for the World Wide Web: Java, Javascript, . . . .
Scripting PLs: Javascript, Perl, Tcl/Tk, Python, . . . .
“Hidden” languages: spreadsheets, macro languages, input for complex applications, . . . .
The following scenario has occurred often in the history of programming. Developers realize that
an application requires a format for expressing input data. The format increases in complexity
until it becomes a miniature programming language. By the time that the developers realize this,
it is too late for them to redesign the format, even if they had principles that they could use to
help them. Consequently, the notation develops into a programming language with many of the
bad features of old, long-since rejected programming languages.
1
1 INTRODUCTION 2
There is an unfortunate tendency in Computer Science to re-invent language features without
carefully studying previous work. Brinch Hansen (1999) points out that, although safe and provably
correct methods of programming concurrent processes were developed by himself and Hoare in the
mid-seventies, Java does not make use of these methods and is insecure.
We conclude that PLs are important. However, their importance is often under-estimated, some-
times with unfortunate consequences.
1.2 Features of the Course
We will review a number of PLs during the course, but the coverage is very uneven. The time
devoted to a PL depends on its relative importance in the evolution of PLs rather than the extent
to which it was used or is used. For example, C and C++ are both widely-used languages but,
since they did not contribute profound ideas, they are discussed briefly. To provide insight into
functional programming, we consider Haskell in some depth.
With hindsight, we can recognize the mistakes and omissions of language designers and criticize
their PLs accordingly. These criticisms do not imply that the designers were stupid or incompetent.
On the contrary, we have chosen to study PLs whose designers had deep insights and made profound
contributions. Given the state of the art at the time when they designed the PLs, however, the
designers could hardly be expected to foresee the understanding that would come only with another
twenty or thirty years of experience.
We can echo Perlis (1978, pages 88–91) in saying:
[Algol 60] was more racehorse than camel . . . . a rounded work of art . . . . Rarely
has a construction so useful and elegant emerged as the output of a committee of 13
meeting for about 8 days . . . . Algol deserves our affection and appreciation.
and yet
[Algol 60] was a noble begin but never intended to be a satisfactory end.
The approach of the course is conceptual rather than technical. We do not dwell on fine lexical
or syntactic distinctions, semantic details, or legalistic issues. Instead, we attempt to focus on the
deep and influential ideas that have developed with PLs.
Exercises are provided mainly to provoke thought. In many cases, the solution to an exercise can
be found (implicitly!) elsewhere in these notes. If we spend time to think about a problem, we are
better prepared to understand and appreciate its solution later. Some of the exercises may appear
in revised form in the examination at the end of the course.
Many references are provided. The intention is not that you read them all but rather that, if you
are interested in a particular aspect of the course, you should be able to find out more about it by
reading the indicated material.
2 Anomalies
There are many ways of motivating the study of PLs. One way is to look at some anomalies:
programs that look as if they ought to work but don’t, or programs that look as if they shouldn’t
work but do.
Example 1: Assignments. Most PLs provide an assignment statement of the form v := E in
which v is a variable and E is an expression. The expression E yields a value. The variable v is not
necessarily a simple name: it may be an expression such as a[i] or r.f. In any case, it yields an
address into which the value of E is stored. In some PLs, we can use a function call in place of E
but not in place of v. For example, we cannot write a statement such as
top(stack) = 6;
to replace the top element of a stack. 2
Exercise 1. Explain why the asymmetry mentioned in Example 1 exists. Can you see any
problems that might be encountered in implementing a function like top? 2
Example 2: Parameters and Results. PLs are fussy about the kinds of objects that can be
passed as parameters to a function or returned as results by a function. In C, an array can be
passed to a function, but only by reference, and a function cannot return an array. In Pascal,
arrays can be passed by value or by reference to a function, but a function cannot return an array.
Most languages do not allow types to be passed as parameters.
It would be useful, for example, to write a procedure such as
void sort (type t, t a[]) ¦. . . .¦
and then call sort(float, scores). 2
Exercise 2. Suppose you added types as parameters to a language. What other changes would
you have to make to the language? 2
Exercise 3. C++ provides templates. Is this equivalent to having types as parameters? 2
Example 3: Data Structures and Files. Conceptually, there is not a great deal of difference
between a data structure and a file. The important difference is the more or less accidental feature
that data structures usually live in memory and files usually live in some kind of external storage
medium, such as a disk. Yet PLs treat data structures and files in completely different ways. 2
Exercise 4. Explain why most PLs treat data structures and files differently. 2
Example 4: Arrays. In PLs such as C and Pascal, the size of an array is fixed at compile-time.
In other PLs, even ones that are older than C and Pascal, such as Algol 60, the size of an array
is not fixed until run-time. What difference does this make to the implementor of the PL and the
programmers who use it?
Suppose that we have a two-dimensional array a. We access a component of a by writing a[i,j]
(except in C, where we have to write a[i][j]). We can access a row of the array by writing a[i].
Why can’t we access a column by writing a[,j]?
3
2 ANOMALIES 4
It is easy to declare rectangular arrays in most PLs. Why is there usually no provision for triangular,
symmetric, banded, or sparse arrays? 2
Exercise 5. Suggest ways of declaring arrays of the kinds mentioned in Example 4. 2
Example 5: Data Structures and Functions. Pascal and C programs contain declarations
for functions and data. Both languages also provide a data aggregation — record in Pascal and
struct in C. A record looks rather like a small program: it contains data declarations, but it
cannot contain function declarations. 2
Exercise 6. Suggest some applications for records with both data and function components. 2
Example 6: Memory Management. Most C/C++ programmers learn quite quickly that it
is not a good idea to write functions like this one.
char *money (int amount)
¦
char buffer[16];
sprintf(buffer, "%d.%d", amount / 100, amount % 100);
return buffer;
¦
Writing C functions that return strings is awkward and, although experienced C programmers have
ways of avoiding the problem, it remains a defect of both C and C++. 2
Exercise 7. What is wrong with the function money in Example 6? Why is it difficult for a C
function to return a string (i.e., char *)? 2
Example 7: Factoring. In algebra, we are accustomed to factoring expressions. For example,
when we see the expression Ax + Bx we are strongly tempted to write it in the form (A + B)x.
Sometimes this is possible in PLs: we can certainly rewrite
z = A * x + B * x;
as
z = (A + B) * x;
Most PLs allow us to write something like this:
if (x > PI) y = sin(x) else y = cos(x)
Only a few PLs allow the shorter form:
y = if (x > PI) then sin(x) else cos(x);
Very few PLs recognize the even shorter form:
y = (if (x > PI) then sin else cos)(x);
2
Example 8: Returning Functions. Some PLs, such as Pascal, allow nested functions, whereas
others, such as C, do not. Suppose that we were allowed to nest functions in C, and we defined
the function shown in Listing 1.
2 ANOMALIES 5
Listing 1: Returning a function
typedef int intint(int);
intint addk (int k)
¦
int f (int n)
¦
return n + k;
¦
return f;
¦
The idea is that addk returns a function that “adds k to its argument”. We would use it as in the
following program, which should print “10”:
int add6 (int) = addk(6);
printf("%d", add6(4));
2
Exercise 8. “It would be fairly easy to change C so that a function could return a function.” True
or false? (Note that you can represent the function itself by a pointer to its first instruction.) 2
Example 9: Functions as Values. Example 8 is just one of several oddities about the way
programs handle functions. Some PLs do not permit this sequence:
if (x > PI) f = sin else f = cos;
f(x);
2
Exercise 9. Why are statements like this permitted in C (with appropriate syntax) but not in
Pascal? 2
Example 10: Functions that work in more than one way. The function Add(x,y,z)
attempts to satisfy the constraint x +y = z. For instance:
Add(2,3,n) assigns 5 to n.
Add(i,3,5) assigns 2 to i.
2
Example 11: Logic computation. It would be convenient if we could encode logical assertions
such as
∀ i . 0 < i < N ⇒ a[i − 1] < a[i]
in the form
all (int i = 1; i < N; i++)
a[i-a] < a[i]
and
∃ i • 0 ≤ i < N ∧ a[i] = k
2 ANOMALIES 6
Listing 2: The Factorial Function
typedef TYPE . . . .
TYPE fac (int n, TYPE f)
¦
if (n <= 1)
return 1;
else
return n * f(n-1, f);
¦
printf("%d", fac(3, fac));
in the form
exists (int i = 0; i < N; i++)
a[i] == k
2
Example 12: Implicit Looping. It is a well-known fact that we can rewrite programs that
use loops as programs that use recursive functions. It is less well-known that we can eliminate the
recursion as well: see Listing 2.
This program computes factorials using neither loops nor recursion:
fac(3, fac) = 3 fac(2, fac)
= 3 2 fac(1, fac)
= 3 2 1
= 6
In early versions of Pascal, functions like this actually worked. When type-checking was improved,
however, they would no longer compile. 2
Exercise 10. Can you complete the definition of TYPE in Example 2? 2
The point of these examples is to show that PLs affect the way we think. It would not occur to
most programmers to write programs in the style of the preceding sections because most PLs do
not permit these constructions. Working with low-level PLs leads to the belief that “this is all that
computers can do”.
Garbage collection provides an example. Many programmers have struggled for years writing
programs that require explicit deallocation of memory. Memory “leaks” are one of the most
common forms of error in large programs and they can do serious damage. (One of the Apollo
missions almost failed because of a memory leak that was corrected shortly before memory was
exhausted, when the rocket was already in lunar orbit.) These programmers may be unaware of
the existence of PLs that provide garbage collection, or they may believe (because they have been
told) that garbage collection has an unacceptably high overhead.
3 The Procedural Paradigm
The introduction of the von Neumann architecture was a crucial step in the development of elec-
tronic computers. The basic idea is that instructions can be encoded as data and stored in the
memory of the computer. The first consequence of this idea is that changing and modifying the
stored program is simple and efficient. In fact, changes can take place at electronic speeds, a very
different situation from earlier computers that were programmed by plugging wires into panels.
The second, and ultimately more far-reaching, consequence is that computers can process programs
themselves, under program control. In particular, a computer can translate a program from one
notation to another. Thus the stored-program concept led to the development of programming
“languages”.
The history of PLs, like the history of any technology, is complex. There are advances and set-
backs; ideas that enter the mainstream and ideas that end up in a backwater; even ideas that are
submerged for a while and later surface in an unexpected place.
With the benefit of hindsight, we can identify several strands in the evolution of PLs. These
strands are commonly called “paradigms” and, in this course, we survey the paradigms separately
although their development was interleaved.
Sources for this section include (Wexelblat 1981; Williams and Campbell-Kelly 1989; Bergin and
Gibson 1996).
3.1 Early Days
The first PLs evolved from machine code. The first programs used numbers to refer to machine
addresses. One of the first additions to programming notation was the use of symbolic names
rather than numbers to represent addresses.
Briefly, it enables the programmer to refer to any word in a programme by means of a
label or tag attached to it arbitrarily by the programmer, instead of by its address in
the store. Thus, for example, a number appearing in the calculation might be labelled
‘a3’. The programmer could then write simply ‘A a3’ to denote the operation of adding
this number into the accumulator, without having to specify just where the number is
located in the machine. (Mutch and Gill 1954)
The key point in this quotation is the phrase “instead of by its address in the store”. Instead of
writing
Location Order
100 A 104
101 A 2
102 T 104
103 H 24
104 C 50
105 T 104
the programmer would write
7
3 THE PROCEDURAL PARADIGM 8
A a3
A 2
T a3
H 24
a3) C 50
T a3
systematically replacing the address 104 by the symbol a3 and omitting the explicit addresses.
This establishes the principle that a variable name stands for a memory location, a principle that
influenced the subsequent development of PLs and is now known — perhaps inappropriately — as
value semantics.
The importance and subroutines and subroutine libraries was recognized before high-level pro-
gramming languages had been developed, as the following quotation shows.
The following advantages arise from the use of such a library:
1. It simplifies the task of preparing problems for the machine;
2. It enables routines to be more readily understood by other users, as conventions
are standardized and the units of a routine are much larger, being subroutines
instead of individual orders;
3. Library subroutines may be used directly, without detailed coding and punching;
4. Library subroutines are known to be correct, thus greatly reducing the overall
chance of error in a complete routine, and making it much easier to locate errors.
. . . . Another difficulty arises from the fact that, although it is desirable to have
subroutines available to cover all possible requirements, it is also undesirable to allow
the size of the resulting library to increase unduly. However, a subroutine can be made
more versatile by the use of parameters associated with it, thus reducing the total size
of the library.
We may divide the parameters associated with subroutines into two classes.
EXTERNAL parameters, i.e. parameters which are fixed throughout the solution of a
problem and arise solely from the use of the library;
INTERNAL parameters, i.e. parameters which vary during the solution of the problem.
. . . . Subroutines may be divided into two types, which we have called OPEN and
CLOSED. An open subroutine is one which is included in the routine as it stands
whereas a closed subroutine is placed in an arbitrary position in the store and can be
called into use by any part of the main routine. (Wheeler 1951)
Exercise 11. This quotation introduces a theme that has continued with variations to the present
day. Find in it the origins of concern for:
correctness;
maintenance;
encapsulation;
parameterization;
genericity;
reuse;
space/time trade-offs.
3 THE PROCEDURAL PARADIGM 9
2
Machine code is a sequence of “orders” or “instructions” that the computer is expected to execute.
The style of programming that this viewpoint developed became known as the “imperative” or
“procedural” programming paradigm. In these notes, we use the term “procedural” rather than
“imperative” because programs resemble “procedures” (in the English, non-technical sense) or
recipes rather than “commands”. Confusingly, the individual steps of procedural PLs, such as
Pascal and C, are often called “statements”, although in logic a “statement” is a sentence that is
either true or false.
By default, the commands of a procedural program are executed sequentially. Procedural PLs
provide various ways of escaping from the sequence. The earliest mechanisms were the “jump”
command, which transferred control to another part of the program, and the “jump and store link”
command, which transferred control but also stored a “link” to which control would be returned
after executing a subroutine.
The data structures of these early languages were usually rather simple: typically primitive values
(integers and floats) were provided, along with single- and multi-dimensioned arrays of primitive
values.
3.2 FORTRAN
FORTRAN was introduced in 1957 at IBM by a team led by John Backus. The “Preliminary
Report” describes the goal of the FORTRAN project:
The IBM Mathematical Formula Translation System or briefly, FORTRAN, will com-
prise a large set of programs to enable the IBM 704 to accept a concise formulation of a
problem in terms of a mathematical notation and to produce automatically a high-speed
704 program for the solution of the problem. (Quoted in (Sammet 1969).)
This suggests that the IBM team’s goal was to eliminate programming! The following quotation
seems to confirm this:
If it were possible for the 704 to code problems for itself and produce as good programs
as human coders (but without the errors), it was clear that large benefits could be
achieved. (Backus 1957)
It is interesting to note that, 20 years later, Backus (1978) criticized FORTRAN and similar
languages as “lacking useful mathematical properties”. He saw the assignment statement as a
source of inefficiency: “the von Neumann bottleneck”. The solution, however, was very similar
to the solution he advocated in 1957 — programming must become more like mathematics: “we
should be focusing on the form and content of the overall result”.
Although FORTRAN did not eliminate programming, it was a major step towards the elimination
of assembly language coding. The designers focused on efficient implementation rather than elegant
language design, knowing that acceptance depended on the high performance of compiled programs.
FORTRAN has value semantics. Variable names stand for memory addresses that are determined
when the program is loaded.
The major achievements of FORTRAN are:
efficient compilation;
3 THE PROCEDURAL PARADIGM 10
separate compilation (programs can be presented to the compiler as separate subroutines, but
the compiler does not check for consistency between components);
demonstration that high-level programming, with automatic translation to machine code, is
feasible.
The principal limitations of FORTRAN are:
Flat, uniform structure. There is no concept of nesting in FORTRAN. A program consists
of a sequence of subroutines and a main program. Variables are either global or local to
subroutines. In other words, FORTRAN programs are rather similar to assembly language
programs: the main difference is that a typical line of FORTRAN describes evaluating an
expression and storing its value in memory whereas a typical line of assembly language
specifies a machine instruction (or a small group of instructions in the case of a macro).
Limited control structures. The control structures of FORTRAN are IF, DO, and GOTO. Since
there are no compound statements, labels provide the only indication that a sequence of
statements form a group.
Unsafe memory allocation. FORTRAN borrows the concept of COMMON storage from assembly
language program. This enables different parts of a program to share regions of memory, but
the compiler does not check for consistent usage of these regions. One program component
might use a region of memory to store an array of integers, and another might assume that
the same region contains reals. To conserve precious memory, FORTRAN also provides the
EQUIVALENCE statement, which allows variables with different names and types to share a
region of memory.
No recursion. FORTRAN allocates all data, including the parameters and local variables of
subroutines, statically. Recursion is forbidden because only one instance of a subroutine can
be active at one time.
Exercise 12. The FORTRAN 1966 Standard stated that a FORTRAN implementation may
allow recursion but is not required to do so. How would you interpret this statement if you were:
writing a FORTRAN program?
writing a FORTRAN compiler?
2
3.3 Algol 60
During the late fifties, most of the development of PLs was coming from industry. IBM domi-
nated, with COBOL, FORTRAN, and FLPL (FORTRAN List Processing Language), all designed
for the IBM 704. Algol 60 (Naur et al. 1960; Naur 1978; Perlis 1978) was designed by an interna-
tional committee, partly to provide a PL that was independent of any particular company and its
computers. The committee included both John Backus (chief designer of FORTRAN) and John
McCarthy (designer of LISP).
The goal was a “universal programming language”. In one sense, Algol was a failure: few complete,
high-quality compilers were written and the language was not widely used (although it was used
more in Europe than in North America). In another sense, Algol was a huge success: it became the
3 THE PROCEDURAL PARADIGM 11
Listing 3: An Algol Block
begin
integer x;
begin
function f(x) begin ... end;
integer x;
real y;
x := 2;
y := 3.14159;
end;
x := 1;
end
standard language for describing algorithms. For the better part of 30 years, the ACM required
submissions to the algorithm collection to be written in Algol.
The major innovations of Algol are discussed below.
Block Structure. Algol programs are recursively structured. A program is a block. A block
consists of declarations and statements. There are various kinds of statement; in particular,
one kind of statement is a block. A variable or function name declared in a block can
be accessed only within the block: thus Algol introduced nested scopes. The recursive
structure of programs means that large programs can be constructed from small programs.
In the Algol block shown in Listing 3, the two assignments to x refer to two different variables.
The run-time entity corresponding to a block is called an activation record (AR). The AR
is created on entry to the block and destroyed after the statements of the block have been
executed. The syntax of Algol ensures that blocks are fully nested; this in turn means that
ARs can be allocated on a stack. Block structure and stacked ARs have been incorporated
into almost every language since Algol.
Dynamic Arrays. The designers of Algol realized that it was relatively simple to allow the size of
an array to be determined at run-time. The compiler statically allocates space for a pointer
and an integer (collectively called a “dope vector”) on the stack. At run-time, when the size
of the array is known, the appropriate amount of space is allocated on the stack and the
components of the “dope vector” are initialized. The following code works fine in Algol 60.
procedure average (n); integer n;
begin
real array a[1:n];
. . . .
end;
Despite the simplicity of the implementation, successor PLs such as C and Pascal dropped
this useful feature.
Call By Name. The default method of passing parameters in Algol was “call by name” and it
was described by a rather complicated “copy rule”. The essence of the copy rule is that the
program behaves as if the text of the formal parameter in the function is replaced by the
text of the actual parameter. The complications arise because it may be necessary to rename
some of the variables during the textual substitution. The usual implementation strategy was
to translate the actual parameter into a procedure with no arguments (called a “thunk”);
3 THE PROCEDURAL PARADIGM 12
Listing 4: Call by name
procedure count (n); integer n;
begin
n := n + 1
end
Listing 5: A General Sum Function
integer procedure sum (max, i, val); integer max, i, val;
begin
integer s;
s := 0;
for i := 1 until n do
s := s + val;
sum := s
end
each occurrence of the formal parameter inside the function was translated into a call to this
function.
The mechanism seems strange today because few modern languages use it. However, the
Algol committee had several valid reasons for introducing it.
Call by name enables procedures to alter their actual parameters. If the procedure count
is defined as in Listing 4, the statement
count(widgets)
has the same effect as the statement
begin
widgets := widgets + 1
end
The other parameter passing mechanism provided by Algol, call by value, does not allow
a procedure to alter the value of its actual parameters in the calling environment: the
parameter behaves like an initialized local variable.
Call by name provides control structure abstraction. The procedure in Listing 5 provides
a form of abstraction of a for loop. The first parameter specifies the number of iterations,
the second is the loop index, and the third is the loop body. The statement
sum(3, i, a[i])
computes a[1]+a[2]+a[3].
Call by name evaluates the actual parameter exactly as often as it is accessed. (This is
in contrast with call by value, where the parameter is usually evaluated exactly once, on
entry to the procedure.) For example, if we declare the procedure try as in Listing 6,
it is safe to call try(x > 0, 1.0/x), because, if x ≤ 0, the expression 1.0/x will not be
evaluated.
Own Variables. A variable in an Algol procedure can be declared own. The effect is that the
variable has local scope (it can be accessed only by the statements within the procedure)
but global extent (its lifetime is the execution of the entire program).
Exercise 13. Why does i appear in the parameter list of Sum? 2
3 THE PROCEDURAL PARADIGM 13
Listing 6: Using call by name
real procedure try (b, x); boolean b; real x;
begin
try := if b then x else 0.0
end
Exercise 14. Discuss the initialization of own variables. 2
Algol 60 and most of its successors, like FORTRAN, has a value semantics. A variable name stands
for a memory addresses that is determined when the block containing the variable declaration is
entered at run-time.
With hindsight, we can see that Algol made important contributions but also missed some very
interesting opportunities.
An Algol block without statements is, in effect, a record. Yet Algol 60 does not provide
records.
The local data of an AR is destroyed after the statements of the AR have been executed. If
the data was retained rather than destroyed, Algol would be a language with modules.
An Algol block consists of declarations followed by statements. Suppose that declarations and
statements could be interleaved in a block. In the following block, D denotes a sequence of
declarations and S denotes a sequence of statements.
begin
D
1
S
1
D
2
S
2
end
A natural interpretation would be that S
1
and S
2
are executed concurrently.
Own variables were in fact rather problematic in Algol, for various reasons including the diffi-
culty of reliably initializing them (see Example 14). But the concept was important: it is the
separation of scope and extent that ultimately leads to objects.
The call by name mechanism was a first step towards the important idea that functions can
be treated as values. The actual parameter in an Algol call, assuming the default calling
mechanism, is actually a parameterless procedure, as mentioned above. Applying this idea
consistently throughout the language would have led to high order functions and paved the
way to functional programming.
The Algol committee knew what they were doing, however. They knew that incorporating the
“missed opportunities” described above would have led to significant implementation problems. In
particular, since they believed that the stack discipline obtained with nested blocks was crucial for
efficiency, anything that jeopardized it was not acceptable.
Algol 60 was simple and powerful, but not quite powerful enough. The dominant trend after Algol
was towards languages of increasing complexity, such as PL/I and Algol 68. Before discussing
these, we take a brief look at COBOL.
3 THE PROCEDURAL PARADIGM 14
3.4 COBOL
COBOL (Sammett 1978) introduced structured data and implicit type conversion. When COBOL
was introduced, “programming” was more or less synonymous with “numerical computation”.
COBOL introduced “data processing”, where data meant large numbers of characters. The data
division of a COBOL program contained descriptions of the data to be processed.
Another important innovation of COBOL was a new approach to data types. The problem of type
conversion had not arisen previously because only a small number of types were provided by the
PL. COBOL introduced many new types, in the sense that data could have various degrees of
precision, and different representations as text. The choice made by the designers of COBOL was
radical: type conversion should be automatic.
The assignment statement in COBOL has several forms, including
MOVE X to Y.
If X and Y have different types, the COBOL compiler will attempt to find a conversion from one
type to the other. In most PLs of the time, a single statement translated into a small number of
machine instructions. In COBOL, a single statement could generate a large amount of machine
code.
Example 13: Automatic conversion in COBOL. The Data Division of a COBOL program
might contain these declarations:
77 SALARY PICTURE 99999, USAGE IS COMPUTATIONAL.
77 SALREP PICTURE $$$,$$9.99
The first indicates that SALARY is to be stored in a form suitable for computation (probably, but
not necessarily, binary form) and the second provides a format for reporting salaries as amounts
in dollars. (Only one dollar symbol will be printed, immediately before the first significant digit).
The Procedure Division would probably contain a statement like
MOVE SALARY TO SALREP.
which implicitly requires the conversion from binary to character form, with appropriate formatting.
2
Exercise 15. Despite significant advances in the design and implementation of PLs, it remains
true that FORTRAN is widely used for “number crunching” and COBOL is widely used for data
processing. Can you explain why? 2
3.5 PL/I
During the early 60s, the dominant languages were Algol, COBOL, FORTRAN. The continuing
desire for a “universal language” that would be applicable to a wide variety of problem domains led
IBM to propose a new programming language (originally called NPL but changed, after objections
from the UK’s National Physical Laboratory, to PL/I) that would combine the best features of
these three languages. Insiders at the time referred to the new language as “CobAlgoltran”.
The design principles of PL/I (Radin 1978) included:
the language should contain the features necessary for all kinds of programming;
a programmer could learn a subset of the language, suitable for a particular application,
without having to learn the entire language.
3 THE PROCEDURAL PARADIGM 15
An important lesson of PL/I is that these design goals are doomed to failure. A programmer who
has learned a “subset” of PL/I is likely, like all programmers, to make a mistake. With luck, the
compiler will detect the error and provide a diagnostic message that is incomprehensible to the
programmer because it refers to a part of the language outside the learned subset. More probably,
the compiler will not detect the error and the program will behave in a way that is inexplicable to
the programmer, again because it is outside the learned subset.
PL/I extends the automatic type conversion facilities of COBOL to an extreme degree. For exam-
ple, the expression (Gelernter and Jagannathan 1990)
(’57’ || 8) + 17
is evaluated as follows:
1. Convert the integer 8 to the string ’8’.
2. Concatenate the strings ’57’ and ’8’, obtaining ’578’.
3. Convert the string ’578’ to the integer 578.
4. Add 17 to 578, obtaining 595.
5. Convert the integer 595 to the string ’595’.
The compiler’s policy, on encountering an assignment x = E, might be paraphrased as: “Do
everything possible to compile this statement; as far as possible, avoid issuing any diagnostic
message that would tell the programmer what is happening”.
PL/I did introduce some important new features into PLs. They were not all well-designed, but
their existence encouraged others to produce better designs.
Every variable has a storage class: static, automatic, based, or controlled. Some of
these were later incorporated into C.
An object associated with a based variable x requires explicit allocation and is placed on
the heap rather than the stack. Since we can execute the statement allocatex as often as
necessary, based variables provide a form of template.
PL/I provides a wide range of programmer-defined types. Types, however, could not be named.
PL/I provided a simple, and not very safe, form of exception handling. Statements of the
following form are allowed anywhere in the program:
ON condition
BEGIN;
. . . .
END;
If the condition (which might be OVERFLOW, PRINTER OUT OF PAPER, etc.) becomes TRUE,
control is transferred to whichever ON statement for that condition was most recently executed.
After the statements between BEGIN and END (the handler) have been executed, control returns
to the statement that raised the exception or, if the handler contains a GOTO statement, to the
target of that statement.
Exercise 16. Discuss potential problems of the PL/I exception handling mechanism. 2
3 THE PROCEDURAL PARADIGM 16
3.6 Algol 68
Whereas Algol 60 is a simple and expressive language, its successor Algol 68 (van Wijngaarden
et al. 1975; Lindsey 1996) is much more complex. The main design principle of Algol 68 was
orthogonality: the language was to be defined using a number of basic concepts that could be
combined in arbitrary ways. Although it is true that lack of orthogonality can be a nuisance in
PLs, it does not necessarily follow that orthogonality is always a good thing.
The important features introduced by Algol 68 include the following.
The language was described in a formal notation that specified the complete syntax and
semantics of the language (van Wijngaarden et al. 1975). The fact that the Report was very
hard to understand may have contributed to the slow acceptance of the language.
Operator overloading: programmers can provide new definitions for standard operators such
as “+”. Even the priority of these operators can be altered.
Algol 68 has a very uniform notation for declarations and other entities. For example, Algol
68 uses the same syntax (mode name = expression) for types, constants, variables, and func-
tions. This implies that, for all these entities, there must be forms of expression that yield
appropriate values.
In a collateral clause of the form (x, y, z), the expressions x, y, and z can be evaluated in
any order, or concurrently. In a function call f(x, y, z), the argument list is a collateral clause.
Collateral clauses provide a good, and early, example of the idea that a PL specification should
intentionally leave some implementation details undefined. In this example, the Algol 68 report
does not specify the order of evaluation of the expressions in a collateral clause. This gives
the implementor freedom to use any order of evaluation and hence, perhaps, to optimize.
The operator ref stands for “reference” and means, roughly, “use the address rather than the
value”. This single keyword introduces call by reference, pointers, dynamic data structures,
and other features to the language. It appears in C in the form of the operators “*” and “&”.
A large vocabulary of PL terms, some of which have become part of the culture (cast, coercion,
narrowing, . . . .) and some of which have not (mode, weak context, voiding,. . . .).
Like Algol 60, Algol 68 was not widely used, although it was popular for a while in various parts
of Europe. The ideas that Algol 68 introduced, however, have been widely imitated.
Exercise 17. Algol 68 has a rule that requires, for an assignment x := E, the lifetime of the
variable x must be less than or equal to the lifetime of the object obtained by evaluating E.
Explain the motivation for this rule. To what extent can it be checked by the compiler? 2
3.7 Pascal
Pascal was designed by Wirth (1996) as a reaction to the complexity of Algol 68, PL/I, and other
languages that were becoming popular in the late 60s. Wirth made extensive use of the ideas of
Dijkstra and Hoare (later published as (Dahl, Dijkstra, and Hoare 1972)), especially Hoare’s ideas
of data structuring. The important contributions of Pascal included the following.
Pascal demonstrated that a PL could be simple yet powerful.
The type system of Pascal was based on primitives (integer, real, bool, . . . .) and mech-
anisms for building structured types (array, record, file, set, . . . .). Thus data types in
Pascal form a recursive hierarchy just as blocks do in Algol 60.
3 THE PROCEDURAL PARADIGM 17
Pascal provides no implicit type conversions other than subrange to integer and integer
to real. All other type conversions are explicit (even when no action is required) and the
compiler checks type correctness.
Pascal was designed to match Wirth’s (1971) ideas of program development by stepwise re-
finement. Pascal is a kind of “fill in the blanks” language in which all programs have a similar
structure, determined by the relatively strict syntax. Programmers are expected to start with
a complete but skeletal “program” and flesh it out in a series of refinement steps, each of
which makes certain decisions and adds new details. The monolithic structure that this idea
imposes on programs is a drawback of Pascal because it prevents independent compilation of
components.
Pascal was a failure because it was too simple. Because of the perceived missing features, supersets
were developed and, inevitably, these became incompatible. The first version of “Standard Pascal”
was almost useless as a practical programming language and the Revised Standard described a
usable language but appeared only after most people had lost interest in Pascal.
Like Algol 60, Pascal missed important opportunities. The record type was a useful innovation
(although very similar to the Algol 68 struct) but allowed data only. Allowing functions in a
record declaration would have paved the way to modular and even object oriented programming.
Nevertheless, Pascal had a strong influence on many later languages. Its most important innova-
tions were probably the combination of simplicity, data type declarations, and static type checking.
Exercise 18. List some of the “missing features” of Pascal. 2
Exercise 19. It is well-known that the biggest loop-hole in Pascal’s type structure was the variant
record. How serious do you think this problem was? 2
3.8 Modula–2
Wirth (1982) followed Pascal with Modula–2, which inherits Pascal’s strengths and, to some ex-
tent, removes Pascal’s weaknesses. The important contribution of Modula–2 was, of course, the
introduction of modules. (Wirth’s first design, Modula, was never completed. Modula–2 was the
product of a sabbatical year in California, where Wirth worked with the designers of Mesa, another
early modular language.)
A module in Modula–2 has an interface and an implementation. The interface provides infor-
mation about the use of the module to both the programmer and the compiler. The implementation
contains the “secret” information about the module. This design has the unfortunate consequence
that some information that should be secret must be put into the interface. For example, the
compiler must know the size of the object in order to declare an instance of it. This implies that
the size must be deducible from the interface which implies, in turn, that the interface must contain
the representation of the object. (The same problem appears again in C++.)
Modula–2 provides a limited escape from this dilemma: a programmer can define an “opaque” type
with a hidden representation. In this case, the interface contains only a pointer to the instance
and the representation can be placed in the implementation module.
The important features of Modula–2 are:
Modules with separated interface and implementation descriptions (based on Mesa).
Coroutines.
3 THE PROCEDURAL PARADIGM 18
3.9 C
C is a very pragmatic PL. Ritchie (Ritchie 1996) designed it for a particular task — systems
programming — for which it has been widely used. The enormous success of C is partly accidental.
UNIX, after Bell released it to universities, became popular, with good reason. Since UNIX
depended heavily on C, the spread of UNIX inevitably led to the spread of C.
C is based on a small number of primitive concepts. For example, arrays are defined in terms
of pointers and pointer arithmetic. This is both the strength and weakness of C. The number of
concepts is small, but C does not provide real support for arrays, strings, or boolean operations.
C is a low-level language by comparison with the other PLs discussed in this section. It is designed
to be easy to compile and to produce efficient object code. The compiler is assumed to be rather
unsophisticated (a reasonable assumption for a compiler running on a PDP/11 in the late sixties)
and in need of hints such as register. C is notable for its concise syntax. Some syntactic features
are inherited from Algol 68 (for example, += and other assignment operators) and others are unique
to C and C++ (for example, postfix and prefix ++ and --).
3.10 Ada
Ada (Whitaker 1996) represents the last major effort in procedural language design. It is a large
and complex language that combines then-known programming features with little attempt at
consolidation. It was the first widely-used language to provide full support for concurrency, with
interactions checked by the compiler, but this aspect of the language proved hard to implement.
Ada provides templates for procedures, record types, generic packages, and task types. The corre-
sponding objects are: blocks and records (representable in the language); and packages and tasks
(not representable in the language). It is not clear why four distinct mechanisms are required
(Gelernter and Jagannathan 1990). The syntactic differences suggest that the designers did not
look for similarities between these constructs. A procedure definition looks like this:
procedure procname ( parameters ) is
body
A record type looks like this:
type recordtype ( parameters ) is
body
The parameters of a record type are optional. If present, they have a different form than the
parameters of procedures.
A generic package looks like this:
generic ( parameters ) package packagename is
package description
The parameters can be types or values. For example, the template
generic
max: integer;
type element is private;
package Stack is
. . . .
might be instantiated by a declaration such as
package intStack is new Stack(20, integer)
3 THE PROCEDURAL PARADIGM 19
Finally, a task template looks like this (no parameters are allowed):
task type templatename is
task description
Of course, programmers hardly notice syntactic differences of this kind: they learn the correct
incantation and recite it without thinking. But it is disturbing that the language designers
apparently did not consider passible relationships between these four kinds of declaration. Changing
the syntax would be a minor improvement, but uncovering deep semantic similarities might have a
significant impact on the language as a whole, just as the identity declaration of Algol 68 suggested
new and interesting possibilities.
Exercise 20. Propose a uniform style for Ada declarations. 2
4 The Functional Paradigm
Procedural programming is based on instructions (“do something”) but, inevitably, procedural PLs
also provide expressions (“calculate something”). The key insight of functional programming (FP)
is that everything can be done with expressions: the commands are unnecessary.
This point of view has a solid foundation in theory. Turing (1936) introduced an abstract model of
“programming”, now known as the Turing machine. Kleene (1936) and Church (1941) introduced
the theory of recursive functions. The two theories were later shown (by Kleene) to be equivalent:
each had the same computational power. Other theories, such as Post production systems, were
shown to have the same power. This important theoretical result shows that FP is not a complete
waste of time but it does not tell us whether FP is useful or practical. To decide that, we must
look at the functional programming languages (FPLs) that have actually been implemented.
Most functional language support high order functions. Roughly, a high order function is a
function that takes another function as a parameter or returns a function. More precisely:
A zeroth order expression contains only variables and constants.
Afirst order expression may also contain function invocations, but the results and parameters
of functions are variables and constants (that is, zeroth order expressions).
In general, in an n-th order expression, the results and parameters of functions are (n−1)-th
order expressions.
A high order expression is an n-th order expression with n ≥ 2.
The same conventions apply in logic with “function” replaced by “function or predicate”. In
first-order logic, quantifiers can bind variables only; in a high order logic, quantifiers can bind
predicates.
4.1 LISP
Anyone could learn LISP in one day, except that if they already knew FORTRAN, it
would take three days. Marvin Minsky
Functional programming was introduced in 1958 in the form of LISP by John McCarthy. The
following account of the development of LISP is based on McCarthy’s (1978) history.
The important early decisions in the design of LISP were:
to provide list processing (which already existed in languages such as Information Processing
Language (IPL) and FORTRAN List Processing Language (FLPL));
to use a prefix notation (emphasizing the operator rather than the operands of an expression);
to use the concept of “function” as widely as possible (cons for list construction; car and cdr
for extracting list components; cond for conditional, etc.);
to provide higher order functions and hence a notation for functions (based on Church’s (1941)
λ-notation);
to avoid the need for explicit erasure of unused list structures.
McCarthy (1960) wanted a language with a solid mathematical foundation and decided that recur-
sive function theory was more appropriate for this purpose than the then-popular Turing machine
model. He considered it important that LISP expressions should obey the usual mathematical laws
allowing replacement of expressions and:
20
4 THE FUNCTIONAL PARADIGM 21
s s s s
s s s s
E E E
NIL
E E
NIL
c
A
c
c
B
c
C
Figure 1: The list structure (A (B C))
Another way to show that LISP was neater than Turing machines was to write a
universal LISP function and show that it is briefer and more comprehensible than
the description of a universal Turing machine. This was the LISP function eval [e, a],
which computes the value of a LISP expression e, the second argument a being a list
of assignments of values to variables. . . . Writing eval required inventing a notation
for representing LISP functions as LISP data, and such a notation was devised for the
purpose of the paper with no thought that it would be used to express LISP programs
in practice. (McCarthy 1978)
After the paper was written, McCarthy’s graduate student S. R. Russel noticed that eval could be
used as an interpreter for LISP and hand-coded it, thereby producing the first LISP interpreter.
Soon afterwards, Timothy Hart and Michael Levin wrote a LISP compiler in LISP; this is probably
the first instance of a compiler written in the language that it compiled.
The function application f(x, y) is written in LISP as (f x y). The function name always comes
first: a + b is written in LISP as (+ a b). All expressions are enclosed in parentheses and can be
nested to arbitrary depth.
There is a simple relationship between the text of an expression and its representation in memory.
An atom is a simple object such as a name or a number. A list is a data structure composed of
cons-cells (so called because they are constructed by the function cons); each cons-cell has two
pointers and each pointer points either to another cons-cell or to an atom. Figure 1 shows the list
structure corresponding to the expression (A (B C)). Each box represents a cons-cell. There are
two lists, each with two elements, and each terminated with NIL. The diagram is simplified in that
the atoms A, B, C, and NIL would themselves be list structures in an actual LISP system.
The function cons constructs a list from its head and tail: (cons head tail). The value of
(car list) is the head of the list and the value of (cdr list) is the tail of the list. Thus:
(car (cons head tail)) → head
(cdr (cons head tail)) → tail
The names car and cdr originated in IBM 704 hardware; they are abbreviations for “contents
of address register” (the top 18 bits of a 36-bit word) and “contents of decrement register” (the
bottom 18 bits).
It is easy to translate between list expressions and the corresponding data structures. There
is a function eval (mentioned in the quotation above) that evaluates a stored list expression.
Consequently, it is straightforward to build languages and systems “on top of” LISP and LISP is
often used in this way.
4 THE FUNCTIONAL PARADIGM 22
It is interesting to note that the close relationship between code and data in LISP mimics the von
Neumann architecture at a higher level of abstraction.
LISP was the first in a long line of functional programming (FP) languages. Its principal contri-
butions are listed below.
Names. In procedural PLs, a name denotes a storage location (value semantics). In LISP, a name
is a reference to an object, not a location (reference semantics). In the Algol sequence
int n;
n := 2;
n := 3;
the declaration int n; assigns a name to a location, or “box”, that can contain an integer.
The next two statements put different values, first 2 then 3, into that box. In the LISP
sequence
(progn
(setq x (car structure))
(setq x (cdr structure)))
x becomes a reference first to (car structure) and then to (cdr structure). The two
objects have different memory addresses. A consequence of the use of names as references to
objects is that eventually there will be objects for which there are no references: these objects
are “garbage” and must be automatically reclaimed if the interpreter is not to run out of
memory. The alternative — requiring the programmer to explicitly deallocate old cells —
would add considerable complexity to the task of writing LISP programs. Nevertheless, the
decision to include automatic garbage collection (in 1958!) was courageous and influential.
A PL in which variable names are references to objects in memory is said to have reference
semantics. All FPLs and most OOPLs have reference semantics.
Note that reference semantics is not the same as “pointers” in languages such as Pascal and
C. A pointer variable stands for a location in memory and therefore has value semantics; it
just so happens that the location is used to store the address of another object.
Lambda. LISP uses “lambda expressions”, based on Church’s λ-calculus, to denote functions.
For example, the function that squares its argument is written
(lambda (x) (* x x))
by analogy to Church’s f = λx . x
2
. We can apply a lambda expression to an argument to
obtain the value of a function application. For example, the expression
((lambda (x) (* x x)) 4)
yields the value 16.
However, the lambda expression itself cannot be evaluated. Consequently, LISP had to resort
to programming tricks to make higher order functions work. For example, if we want to pass
the squaring function as an argument to another function, we must wrap it up in a “special
form” called function:
(f (function (lambda (x) (* x x))) . . . .)
Similar complexities arise when a function returns another function as a result.
Dynamic Scoping. Dynamic scoping was an “accidental” feature of LISP: it arose as a side-effect
of the implementation of the look-up table for variable values used by the interpreter. The
C-like program in Listing 7 illustrates the difference between static and dynamic scoping. In
4 THE FUNCTIONAL PARADIGM 23
Listing 7: Static and Dynamic Binding
int x = 4; // 1
void f()
¦
printf("%d", x);
¦
void main ()
¦
int x = 7; // 2
f();
¦
C, the variable x in the body of the function f is a use of the global variable x defined in the
first line of the program. Since the value of this variable is 4, the program prints 4. (Do not
confuse dynamic scoping with dynamic binding!)
A LISP interpreter constructs its environment as it interprets. The environment behaves like
a stack (last in, first out). The initial environment is empty, which we denote by '`. After
interpreting the LISP equivalent of the line commented with “1”, the environment contains
the global binding for x: 'x = 4`. When the interpreter evaluates the function main, it inserts
the local x into the environment, obtaining 'x = 7, x = 4`. The interpreter then evaluates the
call f(); when it encounters x in the body of f, it uses the first value of x in the environment
and prints 7.
Although dynamic scoping is natural for an interpreter, it is inefficient for a compiler. In-
terpreters are slow anyway, and the overhead of searching a linear list for a variable value
just makes them slightly slower still. A compiler, however, has more efficient ways of ac-
cessing variables, and forcing it to maintain a linear list would be unacceptably inefficient.
Consequently, early LISP systems had an unfortunate discrepancy: the interpreters used dy-
namic scoping and the compilers used static scoping. Some programs gave one answer when
interpreted and another answer when compiled!
Exercise 21. Describe a situation in which dynamic scoping is useful. 2
Interpretation. LISP was the first major language to be interpreted. Originally, the LISP inter-
preter behaved as a calculator: it evaluated expressions entered by the user, but its internal
state did not change. It was not long before a form for defining functions was introduced
(originally called define, later changed to defun) to enable users to add their own functions
to the list of built-in functions.
A LISP program has no real structure. On paper, a program is a list of function definitions;
the functions may invoke one another with either direct or indirect recursion. At run-time, a
program is the same list of functions, translated into internal form, added to the interpreter.
The current dialect of LISP is called Common LISP (Steele et al. 1990). It is a much larger and
more complex language than the original LISP and includes many features of Scheme (described
below). Common LISP provides static scoping with dynamic scoping as an option.
Exercise 22. The LISP interpreter, written in LISP, contains expressions such as the one shown
in Listing 8. We might paraphrase this as: “if the car of the expression that we are currently
evaluating is car, the value of the expression is obtained by taking the car of the cadr (that is,
the second term) of the expression”. How much can you learn about a language by reading an
interpreter written in the language? What can you not learn? 2
4 THE FUNCTIONAL PARADIGM 24
Listing 8: Defining car
(cond
((eq (car expr) ’car)
(car (cadr expr)) )
. . . .
4.1.1 Basic LISP Functions
The principal — and only — data structure in classical LISP is the list. Lists are written (a b c).
Lists also represent programs: the list (f x y) stands for the function f applied to the arguments
x and y.
The following functions are provided for lists.
cons builds a list, given its head and tail.
first (originally called car) returns the head (first component) of a list.
rest (originally called cdr) returns the tail of a list.
second (originally called cadr, short for car cdr) returns the second element of a list.
null is a predicate that returns true if its argument is the empty list.
atom is a predicate that return true if its argument is not a list — and must therefore be an
“atom”, that is, a variable name or other primitive object.
A form looks like a function but does not evaluate its argument. The important forms are:
quote takes a single argument and does not evaluate it.
cond is the conditional construct has the form
(cond (p
1
e
1
) (p
2
e
2
) ... (t e))
end works like this: if p
1
is true, return e
1
; if p
2
is true, return e
2
; . . . ; else return e.
def is used to define functions:
(def name (lambda parameters body ))
4.1.2 A LISP Interpreter
The following interpreter is based on McCarthy’s LISP 1.5 Evaluator from Lisp 1.5 Programmer’s
Manual by McCarthy et al., MIT Press 1962.
An environment is a list of name/value pairs. The function pairs builds an environment from
a list of names and a list of expressions.
(def pairs (lambda (names exps env)
(cond
((null names) env)
(t (cons
(cons (first name) (first exps))
(pairs (rest names) (rest exps) env) )) ) ))
The function lookup finds the value of a name in a table. The table is represented as a list of
pairs.
4 THE FUNCTIONAL PARADIGM 25
(def lookup (lambda (name table)
(cond
((eq name (first (first table))) (rest (first table)))
(t (lookup name (rest table))) ) ))
The heart of the interpreter is the function eval which evaluates an expression exp in an environ-
ment env.
(def eval (lambda (exp env)
(cond
((null exp) nil)
((atom exp) (lookup exp env))
((eq (first exp) (quote quote)) (second exp))
((eq (first exp) (quote cond)) (evcon (second exp) env) )
(t (apply (first exp) (evlist (rest exp) env) env)) ) ))
The function evcon is used to evaluate cond forms; it takes a list of pairs of the form (cnd exp)
and returns the value of the expression corresponding to the first true cnd.
(def evcon (lambda (tests env)
(cond
((null tests) nil)
(t
(cond
((eval (first (first tests)) env) (eval (second (first tests)) env))
(t (evcon (rest tests) env)) ) ) ) ))
The function evlist evaluates a list of expressions in an environment. It is a straightforward
recursion.
(def evlist (lambda (exps env)
(cond
((null exps) nil)
(t (cons (eval (first exps) env)
(evlist (rest exps) env) )) ) ))
The function apply applies a function fun to arguments args in the environment env. The cases
it considers are:
Built-in functions: first, second, cons, and other built-in functions not considered here.
Any other function with an atomic name: this is assumed to be a user-defined function, and
lookup is used to find the lambda form corresponding to the name.
A function which is a lambda form.
(def apply (lambda (fun args env)
(cond
((eq fun (quote first)) (first (first args)))
((eq fun (quote second)) (second (first args)))
((eq fun (quote cons)) (cons (first args) (second args)))
((atom fun) (apply (eval fun env) args env))
((eq (first fun) (quote lambda))
(eval (third fun) (pairs (second fun) args env)) ) ) ))
4 THE FUNCTIONAL PARADIGM 26
4.1.3 Dynamic Binding
The classical LISP interpreter was implemented while McCarthy was still designing the language.
It contains an important defect that was not discovered for some time. James Slagle defined a
function like this
(def testr (lambda (x p f u)
(cond
((p x) (f x))
((atom x) (u))
(t (testr (rest x) p f (lambda () (testr (first x) p f u)))) ) ))
and it did not give the correct result.
We use a simpler example to explain the problem. Suppose we define:
(def show (lambda ()
x ))
Calling this function generates an error because x is not defined. (The interpreter above does not
incorporate error detection and would simply fail with this function.) However, we can wrap show
inside another function:
(def try (lambda (x)
(show) ))
We then find that (try t) evaluates to t. When show is called, the binding (x.t) is in the
environment, and so that is the value that show returns.
In other words, the value of a variable in LISP depends on the dynamic behaviour of the program
(which functions have been called) rather than on the static text of the program.
We say that LISP uses dynamic binding whereas most other languages, including Scheme,
Haskell, and C
++
, use static binding.
Correcting the LISP interpreter to provide static binding is not difficult: it requires a slightly
more complicated data structure for environments (instead of having a simple list of name/value
pairs, we effectively build a list of activation records). However, the dynamic binding problem
was not discovered until LISP was well-entrenched: it would be almost 20 years before Guy Steele
introduced Scheme, with static binding, in 1978.
People wrote LISP compilers, but it is hard to write a compiler with dynamic binding. Conse-
quently, there were many LISP systems that provided dynamic binding during interpretation and
static binding for compiled programs!
4.2 Scheme
Scheme was designed by Guy L. Steele Jr. and Gerald Jay Sussman (1975). It is very similar to
LISP in both syntax and semantics, but it corrects some of the errors of LISP and is both simpler
and more consistent.
The starting point of Scheme was an attempt by Steele and Sussman to understand Carl Hewitt’s
theory of actors as a model of computation. The model was object oriented and influenced by
Smalltalk (see Section 6.2). Steele and Sussman implemented the actor model using a small LISP
4 THE FUNCTIONAL PARADIGM 27
Listing 9: Factorial with functions
(define factorial
(lambda (n) (if (= n 0)
1
(* n (factorial (- n 1))))))
Listing 10: Factorial with actors
(define actorial
(alpha (n c) (if (= n 0)
(c 1)
(actorial (- n 1) (alpha (f) (c (* f n)))))))
interpreter. The interpreter provided lexical scoping, a lambda operation for creating functions, and
an alpha operation for creating actors. For example, the factorial function could be represented
either as a function, as in Listing 9, or as an actor, as in Listing 10. Implementing the interpreter
brought an odd fact to light: the interpreter’s code for handling alpha was identical to the code
for handling lambda! This indicated that closures — the objects created by evaluating lambda —
were useful for both high order functional programming and object oriented programming (Steele
1996).
LISP ducks the question “what is a function?” It provides lambda notation for functions, but a
lambda expression can only be applied to arguments, not evaluated itself. Scheme provides an
answer to this question: the value of a function is a closure. Thus in Scheme we can write both
(define num 6)
which binds the value 6 to the name num and
(define square (lambda (x) (* x x)))
which binds the squaring function to the name square. (Scheme actually provides an abbreviated
form of this definition, to spare programmers the trouble of writing lambda all the time, but the
form shown is accepted by the Scheme compiler and we use it in these notes.)
Since the Scheme interpreter accepts a series of definitions, as in LISP, it is important to understand
the effect of the following sequence:
(define n 4)
(define f (lambda () n))
(define n 7)
(f)
The final expression calls the function f that has just been defined. The value returned is the
value of n, but which value, 4 or 7? If this sequence could be written in LISP, the result would
be 7, which is the value of n in the environment when (f) is evaluated. Scheme, however, uses
static scoping. The closure created by evaluating the definition of f includes all name bindings
in effect at the time of definition. Consequently, a Scheme interpreter yields 4 as the value of
(f). The answer to the question “what is a function?” is “a function is an expression (the body
of the function) and an environment containing the values of all variables accessible at the point
of definition”.
Closures in Scheme are ordinary values. They can be passed as arguments and returned by func-
tions. (Both are possible in LISP, but awkward because they require special forms.)
4 THE FUNCTIONAL PARADIGM 28
Listing 11: Differentiating in Scheme
(define derive (lambda (f dx)
(lambda (x)
(/ (- (f (+ x dx)) (f x)) dx))))
(define square (lambda (x) (* x x)))
(define Dsq (derive sq 0.001))
Listing 12: Banking in Scheme
(define (make-account balance)
(define (withdraw amount)
(if (>= balance amount)
(sequence (set! balance (- balance amount))
balance)
("Insufficient funds"))
(define (deposit amount)
(set! balance (+ balance amount))
balance)
(define (dispatch m)
(cond
((eq? m ’withdraw) withdraw)
((eq? m ’deposit) deposit)
(else (error "Unrecognized transaction" m))))
dispatch)
Example 14: Differentiating. Differentiation is a function that maps functions to functions.
Approximately (Abelson and Sussman 1985):
Df(x) =
f(x + dx) −f(x)
dx
We define the Scheme functions shown in Listing 11. After these definitions have been evaluated,
Dsq is, effectively, the function
f(x) =
(x + 0.001)
2
− x
2
0.001
= 2x +
We can apply Dsq like this: (->” is the Scheme prompt):
-> (Dsq 3)
6.001
2
Scheme avoids the problem of incompatibility between interpretation and compilation by being
statically scoped, whether it is interpreted or compiled. The interpreter uses a more elaborate data
structure for storing values of local variables to obtain the effect of static scoping. In addition to
full support for high order functions, Scheme introduced continuations.
Although Scheme is primarily a functional language, side-effects are allowed. In particular, set!
changes the value of a variable. (The “!” is a reminder that set! has side-effects.)
4 THE FUNCTIONAL PARADIGM 29
Listing 13: Using the account
-> ((acc ’withdraw) 50)
50
-> ((acc ’withdraw) 100)
Insufficient funds
-> ((acc ’deposit) 100)
150
Example 15: Banking in Scheme. The function shown in Listing 12 shows how side-effects
can be used to define an object with changing state (Abelson and Sussman 1985, page 173). The
following dialog shows how banking works in Scheme. The first step is to create an account.
-> (define acc (make-account 100))
The value of acc is a closure consisting of the function dispatch together with an environment in
which balance = 100. The function dispatch takes a single argument which must be withdraw
or deposit; it returns one of the functions withdraw or deposit which, in turn, takes an amount
as argument. The value returned is the new balance. Listing 13 shows some simple applications of
acc. The quote sign (’) is required to prevent the evaluation of withdraw or deposit. 2
This example demonstrates that with higher order functions and control of state (by side-effects)
we can obtain a form of OOP. The limitation of this approach, when compared to OOPLs such as
Simula and Smalltalk (described in Section 6) is that we can define only one function at a time.
This function must be used to dispatch messages to other, local, functions.
4.3 SASL
SASL (St. Andrew’s Symbolic Language) was introduced by David Turner (1976). It has an Algol-
like syntax and is of interest because the compiler translates the source code into a combinator
expression which is then processed by graph reduction (Turner 1979). Turner subsequently designed
KRC (Kent Recursive Calculator) (1981) and Miranda (1985), all of which are implemented with
combinator reduction.
Combinator reduction implements call by name (the default method for passing parameters in
Algol 60) but with an optimization. If the parameter is not needed in the function, it is not
evaluated, as in Algol 60. If it is needed one or more times, it is evaluated exactly once. Since
SASL expressions do not have side-effects, evaluating an expression more than once will always
give the same result. Thus combinator reduction is (in this sense) the most efficient way to pass
parameters to functions. Evaluating an expression only when it is needed, and never more than
once, is called call by need or lazy evaluation.
The following examples use SASL notation. The expression x::xs denotes a list with first element
(head) x and remaining elements (tail) xs. The definition
nums(n) = n::nums(n+1)
apparently defines an infinite list:
nums(0) = 0::nums(1) = 0::1::nums(2) = . . . .
The function second returns the second element of a list. In SASL, we can define it like this:
second (x::y::xs) = y
Although nums(0) is an “infinite” list, we can find its second element in SASL:
4 THE FUNCTIONAL PARADIGM 30
second(nums(0)) = second(0::nums(1)) = second(0::1::nums(2)) = 1
This works because SASL evaluates a parameter only when its value is needed for the calculation
to proceed. In this example, as soon as the argument of second is in the form 0::1::. . . ., the
required result is known.
Call by need is the only method of passing arguments in SASL but it occurs as a special case in
other languages. If we consider if as a function, so that the expression
if P then X else Y
is a fancy way of writing if(P,X,Y), then we see that if must use call by need for its second and
third arguments. If it did not, we would not be able to write expressions such as
if x = 0 then 1 else 1/x
In C, the functions && (AND) and [[ (OR) are defined as follows:
X && Y ≡ if X then Y else false
X [[ Y ≡ if X then true else Y
These definitions provide the effect of lazy evaluation and allow us to write expressions such as
if (p != NULL && p->f > 0) . . . .
4.4 SML
SML (Milner, Tofte, and Harper 1990; Milner and Tofte 1991) was designed as a “metalanguage”
(ML) for reasoning about programs as part of the Edinburgh Logic for Computable Functions
(LCF) project. The language survived after the rest of the project was abandoned and became
“standard” ML, or SML. The distinguishing feature of SML is that it is statically typed in the
sense of Section B.3 and that most types can be inferred by the compiler.
In the following example, the programmer defines the factorial function and SML responds with its
type. The programmer then tests the factorial function with argument 6; SML assigns the result
to the variable it, which can be used in the next interaction if desired. SML is run interactively,
and prompts with “-”.
- fun fac n = if n = 0 then 1 else n * fac(n-1);
val fac = fn : int -> int
- fac 6;
val it = 720 : int
SML also allows function declaration by cases, as in the following alternative declaration of the
factorial function:
- fun fac 0 = 1
= | fac n = n * fac(n-1);
val fac = fn : int -> int
- fac 6;
val it = 720 : int
Since SML recognizes that the first line of this declaration is incomplete, it changes the prompt to
“=” on the second line. The vertical bar “|” indicates that we are declaring another “case” of the
declaration.
Each case of a declaration by cases includes a pattern. In the declaration of fac, there are
two patterns. The first, 0, is a constant pattern, and matches only itself. The second, \tt n,
4 THE FUNCTIONAL PARADIGM 31
Listing 14: Function composition
- infix o;
- fun (f o g) x = g (f x);
val o = fn : (’a -> ’b) * (’b -> ’c) -> ’a -> ’c
- val quad = sq o sq;
val quad = fn : real -> real
- quad 3.0;
val it = 81.0 : real
Listing 15: Finding factors
- fun hasfactor f n = n mod f = 0;
val hasfactor fn : int -> int -> bool
- hasfactor 3 9;
val it = true : bool
is a variable pattern, and matches any value of the appropriate type. Note that the definition
fun sq x = x * x; would fail because SML cannot decide whether the type of x is int or real.
- fun sq x:real = x * x;
val sq = fn : real -> real
- sq 17.0;
val it = 289.0 : real
We can pass functions as arguments to other functions. The function o (intended to resemble the
small circle that mathematicians use to denote functional composition) is built-in, but even if it
wasn’t, we could easily declare it and use it to build the fourth power function, as in Listing 14. The
symbols ’a, ’b, and ’c are type names; they indicate that SML has recognized o as a polymorphic
function.
The function hasfactor defined in Listing 15 returns true if its first argument is a factor of
its second argument. All functions in SML have exactly one argument. It might appear that
hasfactor has two arguments, but this is not the case. The declaration of hasfactor introduces
two functions, as shown in Listing 16. Functions like hasfactor take their arguments one at a
time. Applying the first argument, as in hasfactor 2, yields a new function. The trick of applying
one argument at a time is called “currying”, after the American logician Haskell Curry. It may be
helpful to consider the types involved:
hasfactor : int -> int -> bool
hasfactor 2 : int -> bool
hasfactor 2 6 : bool
The following brief discussion, adapted from (
˚
Ake Wikstr¨om 1987), shows how functions can be
used to build a programmer’s toolkit. The functions here are for list manipulation, which is a
widely used example but not the only way in which a FPL can be used. We start with a list
generator, defined as an infix operator.
Listing 16: A function with one argument
- val even = hasfactor 2;
val even = fn : int -> bool;
- even 6;
val it = true : bool
4 THE FUNCTIONAL PARADIGM 32
Listing 17: Sums and products
- fun sum [] = 0
= | sum (x::xs) = x + sum xs;
val sum = fn : int list -> int
- fun prod [] = 1
= | prod (x::xs) = x * prod xs;
val prod = fn : int list -> int
- sum (1 -- 5);
val it = 15 : int
- prod (1 -- 5);
val it = 120 : int
Listing 18: Using reduce
- fun sum xs = reduce add 0 xs;
val sum = fn : int list -> list
- fun prod xs = reduce mul 1 xs;
val prod = fn : int list -> list
- infix --;
- fun (m -- n) = if m < n then m :: (m+1 -- n) else [];
val -- = fun : int * int -> int list
- 1 -- 5;
val it = [1,2,3,4,5] : int list
The functions sum and prod in Listing 17 compute the sum and product, respectively, of a list of
integers. We note that sum and prod have a similar form. This suggests that we can abstract the
common features into a function reduce that takes a binary function, a value for the empty list,
and a list. We can use reduce to obtain one-line definitions of sum and prod, as in Listing 18. The
idea of processing a list by recursion has been captured in the definition of reduce.
- fun reduce f u [] = u
= | reduce f u (x::xs) = f x (reduce f u xs);
val reduce = fn : (’a -> ’b -> ’b) -> ’b -> ’a list -> ’b
We can also define a sorting function as
- val sort = reduce insert nil;
val sort = fn : int list -> int list
where insert is the function that inserts a number into an ordered list, preserving the ordering:
- fun insert x:int [] = x::[]
| insert x (y::ys) = if x <= y then x::y::ys else y::insert x ys;
val insert = fn : int -> int list -> int list
4.5 Other Functional Languages
We have not discussed FP, the notation introduced by Backus (1978). Although Backus’s speech
(and subsequent publication) turned many people in the direction of functional programming, his
ideas for the design of FPLs were not widely adopted.
Since 1978, FPLs have split into two streams: those that provide mainly “eager evaluation” (that
is, call by value), following LISP and Scheme, and those that provide mainly “lazy evaluation”,
4 THE FUNCTIONAL PARADIGM 33
following SASL. The most important language in the first category is still SML, while Haskell
appears to be on the way to becoming the dominant FPL for teaching and other applications.
5 Haskell
COMP 348 is about the principles of programming languages. In order to understand the prin-
ciples, you need to know more than one programming language. Even this is not enough: you
really need to know more than one style of programming. Accordingly, the course begins with an
introduction to a language that is quite different to languages that you may already be familiar
with such as C
++
, Java, and Basic.
Haskell is a functional programming language in the sense of Section 4. Although it has many
of the features that we associate with other languages, such as variables, expressions, and data
structures, its foundation is the mathematical concept of a function.
You can find information about Haskell at www.haskell.org. Haskell gets its name from the
American logician Haskell Curry (1900–82 — see http://www.haskell.org/bio.html).
5.1 Getting Started
There are several Haskell compilers and all of them are free. One of the first Haskell implementa-
tions was called “Gofer”; we will be using a system derived from Gofer and called HUGS (Haskell
Users’ Gofer System). At Concordia, Hugs has been installed on both Solaris8 machines and
linux hosts in pkg/hugs98. Executables are in /site/bin and can be run without any special
preparation. Documentation is in pkg/hugs98/docs.
If you want to run Hugs on your own PC, download it from http://www.haskell.org/hugs;
installation is straightforward.
Hugs provides an interactive programming environment. It is based on a “read/eval/print” loop:
the system reads an expression, evaluates it, and prints the value. Figure 2 shows Hugs working out
2+2. The prompt Prelude> indicates that the only module that has been loaded is the “prelude”.
As well as expressions, the environment recognizes various commands with the prefix ”:”. For
instance, if you enter :browse (or :b for short), Hugs will display the functions and structures
defined in the current environment.
Haskell computes with integers and floating-point numbers. Integers have unlimited size and
conversion are performed as required.
Prelude> 1234567890987654321 / 1234567
1.0e+012 :: Double
Prelude> 1234567890987654321 ‘div‘ 1234567
1000000721700 :: Integer
Strings are built into Haskell. A string literal is enclosed in double quotes, as in C
++
. Haskell uses
the same escape sequences C
++
(e.g., ’\n’, etc.). However, Haskell strings are not terminated by
’\0’. The infix operator ++ concatenates strings.
Prelude> "Hallo" ++ " " ++ "everyone!"
"Hallo everyone!" :: [Char]
The basic data structure of almost all functional programming languages, including Haskell, is
the list. In Haskell, lists are enclosed in square brackets with items separated by commas. The
concatenation operator for lists is ++. This is the same as the concatenation operator for strings
— Haskell considers a string to be simply a list of characters.
Prelude> [1,2,3] ++ [4,5,6]
[1,2,3,4,5,6] :: [Integer]
The prelude provides a number of basic operations on lists.
34
5 HASKELL 35
__ __ __ __ ____ ___ _____________________________________________
|| || || || || || ||__ Hugs 98: Based on the Haskell 98 standard
||___|| ||__|| ||__|| __|| Copyright (c) 1994-2001
||---|| ___|| World Wide Web: http://haskell.org/hugs
|| || Report bugs to: hugs-bugs@haskell.org
|| || Version: December 2001 _____________________________________________
Haskell 98 mode: Restart with command line option -98 to enable extensions
Reading file "C:\Program Files\Hugs98\lib\Prelude.hs":
Hugs session for:
C:\Program Files\Hugs98\lib\Prelude.hs
Type :? for help
Prelude> 2+2
4 :: Integer
Prelude>
Figure 2: A short session with Hugs
Prelude> [1..20]
[1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20] :: [Integer]
Prelude> sum [1..20]
210 :: Integer
Prelude> product [1..20]
2432902008176640000 :: Integer
Prelude> take 5 [1..20]
[1,2,3,4,5] :: [Integer]
Prelude> drop 10 [1..20]
[11,12,13,14,15,16,17,18,19,20] :: [Integer]
Prelude> filter even [1..20]
[2,4,6,8,10,12,14,16,18,20] :: [Integer]
Prelude> take 5 "Peter Piper pickled a pepper"
"Peter" :: [Char]
Evaluating expressions is fun for a few minutes, but it quickly gets boring. To do anything inter-
esting, you have to write your own programs — which, in Haskell, means mainly writing functions
(although later we will also introduce new data types).
The environment does not allow you to introduce any new definitions:
Prelude> n = 99
ERROR - Syntax error in input (unexpected ‘=’)
Fortunately, it does allow you read definitions from a file. The command is :load and it is followed
by a path name. Here is how the definitions for this lecture are loaded. Note that the prompt
changes to reflect the newly loaded module:
Prelude> :load "e:/courses/comp348/src/lec1.hs"
Reading file "e:/courses/comp348/src/lec1.hs":
5 HASKELL 36
Command Brief description Note
:? List all Hugs commands
:! 'command` Execute a shell command
:quit Quit the current Hugs session
:load 'filename` . . . Load Haskell source from a file (1)
:also 'filename` . . . Load additional files (1)
:reload Repeat last load command (1)
:browse 'module` . . . Display functions exported by module(s)
:set 'options` Set flags and environment parameters (see Figure 4)
:cd 'directory` Change working directory
:edit 'file` Start the editor (2)
:find 'name` Start the editor with the cursor on the object named (2)
:gc Force garbage collection
:info 'name` . . . Display information about files, classes, etc.
:module 'module` Change module
:names 'pattern` Find names that match the pattern (2)
:project 'project file` Load a project (obsolete) (3)
:type 'expression` Print the type of 'expression` without evaluating it
:version Display Hugs version
Figure 3: Hugs commands
Hugs session for:
C:\Program Files\Hugs98\lib\Prelude.hs
e:/courses/comp348/src/lec1.hs
The file lec1.hs looks like this (the “....” indicates material that is discussed below):
module Lec1 where
fc1 n = if n == 0 then 1 else n * fc1(n - 1)
....
This code introduces a Haskell module called Lec1 that contains a number of definitions. The
first definition introduces a function called fc1.
In the examples that follow, we will show Haskell code as it appears in the source file. When we
need to show the effect of evaluating an expression, we will include the prompt, as in the examples
above.
5.1.1 Hugs Commands
Before describing the Haskell language, we digress briefly to explain the commands provided by the
Hugs environment. Figure 3 provides a brief summary of each command. Note that any command
can be abbreviated: for example, you can enter :b instead of :browse. Figure 4 shows options
for the :set command; these options can also be specified on the command line. The WinHugs
environment has icons for most of these commands.
Notes for Figure 3
5 HASKELL 37
±s Print # of reductions and cells after evaluation
±t Print type after evaluation
±f Terminate evaluation after first error
±g Print cells recovered by garbage collector
±l Use literate modules as default
±e Warn about errors in literate modules
±. Print dots to show progress (not recommended!)
±q Print nothing to show progress
±w Always show which modules are loaded
±k Show “kind” errors in full
±u Use show to display results
±i Chase imports when loading modules
p'string` Set prompt to 'string`
r'string` Set repeat last expression string to 'string`
P'string` Set search path for modules to 'string`
E'string` Use editor setting given by 'string`
F'string` Set preprocessor filter to 'string`
h'num` Set heap size to 'num` bytes (has no effect on current session)
c'num` Set constraint cutoff limit to 'num`
Figure 4: Parameters for :set. The symbol ± indicates a toggle: +x turns x on and -x turns it
off. The other commands take a string or a number as argument, as shown.
1. :load removes all definitions (except those belonging to the Prelude) and loads new defini-
tions from the named file(s). :also is similar but it does not remove previous definitions.
:reload repeats the previous :load command.
One easy way to develop a Hugs program is to:
Write the first version in a file fun.hs
Leaving the editor running, load this file into Hugs and try the functions: :l fun.hs
Make corrections in the editor window
Reload the file in the Hugs window: :r
2. The editor commands assume that Hugs has been linked to an editor. For example, you may
like to use emacs with Hugs. :edit suspends Hugs and starts the editor; when you close
the editor, Hugs resumes. :find is similar, but it positions the editor cursor at the chosen
object.
The default editor for Hugs in Windows is notepad. You may find it more convenient to ignore
these commands and simply run Hugs and your favourite editor as independent processes.
In this case, all you have to do is save the corrected file in the editor window and enter :r in
Hugs to reload the modified source code.
3. The project command reads a project file containing information about the files that Hugs
should load. Recent versions of Hugs use a smarter method of loading projects, called “import
chasing”, that make the :project command unnecessary in most cases.
You can obtain the information in Figure 4 by entering :set without any arguments. The output
also includes the current settings which, for a Windows session, will look something like this:
5 HASKELL 38
Current settings: +tfewuiAR -sgl.qQkI -h250000 -p"%s> " -r$$ -c40
Search path : -P{Hugs}\lib;{Hugs}\lib\hugs;{Hugs}\lib\exts;{Hugs}\lib\win32
Project Path :
Editor setting : -EC:\WINNT\notepad.exe
Preprocessor : -F
Compatibility : Haskell 98 (+98)
5.1.2 Literate Haskell
Donald Knuth proposed the idea that programs should be readable documents with detailed ex-
planations and commentary built in, rather than just code listings with occasional — and often
useless — comments. A program written in this way is called a literate program and the art of
writing such programs is called literate programming.
The first step towards literate programming is to make “comments” very easy to write and read.
A consequence of this step is that the code itself becomes slightly harder to write and read. The
overhead for code in literate Haskell (at least the Hugs version) is relatively light. Here are the
rules:
• Literate Haskell source files use the extension .lhs instead of .hs.
• A line beginning with > is interpreted as Haskell code.
• The lines before and after lines containing code must be either blank or lines containing code
(but see the e switch below).
• Any other lines are ignored by the compiler.
The check for blank lines can be suppressed by turning off the e switch in Hugs. Here’s an example
to illustrate how this works. Suppose that the source file vector.lhs contains this text:
Vector is an instance of the type class Show.
This enables easy conversion of vectors to strings.
> instance Show Vector where
> show (V x y z) = "(" ++ show x ++ ", " ++ show y ++ ", " ++ show z ++ ")"
By default, the e switch in the compiler is on (as if we had entered :set +e). When Hugs loads
this file, it will report:
Prelude> :l e:/courses/comp348/src/vector.lhs
Reading file "e:/courses/comp348/src/vector.lhs":
Parsing
ERROR "e:/courses/comp348/src/vector.lhs":14 - Program line next to comment
There are two ways to fix this problem. One is to turn off the e switch (its default setting is +e):
Prelude> :set -e
and then reload the source file — it will load without errors. The other way is to put blank lines
in the source file:
5 HASKELL 39
Vector is an instance of the type class Show.
This enables easy conversion of vectors to strings.
> instance Show Vector where
> show (V x y z) = "(" ++ show x ++ ", " ++ show y ++ ", " ++ show z ++ ")"
In general, it is a good idea to put blank lines in the source file anyway, because they make the
code stand out more clearly.
Figure 5 shows a small but complete example of a literate Haskell program.
5.1.3 Defining Functions
Since Haskell is a functional language, defining functions is fundamental and there are, in fact,
several ways to define functions. We will start by considering the favourite function of functional
programmers, the factorial function:
0 ! = 1
n! = 1 2 3 . . . n for n ≥ 1
The first point to note is that, since everything is a value in Haskell, there is no need for a return
statement: you write just the value that you want returned. Our first definition of the factorial
function uses recursion:
fc1 n = if n == 0 then 1 else n * fc1(n - 1)
This definition works fine, but it is not the preferred style for Haskell. Many functions, including
this one, are defined by cases. The cases for factorial are n = 0 and n > 0. Haskell has a case
expression (analogous to switch in C
++
and Java) and we use it here for the second definition of
factorial:
fc2 n = case n of
0 -> 1
n -> n * fc2(n - 1)
This is the first definition that is written on more than one line. It can be written on one line but, to
make the syntax legal, there must be a semicolon between the two case clauses (see Section 5.1.4):
fc2 n = case n of 0 -> 1 ; n -> n * fc2(n - 1)
Even this, however, is not idiomatic Haskell. The preferred form of definition is a set of equa-
tions:
fc3 0 = 1
fc3 n = n * fc3(n - 1)
Although we have written these equations together, they are in fact independent assertions for
Haskell. If we wrote the first but not the second, fc3 0 would evaluate correctly to 1, but fc3 4
would cause an error.
There are also “closed form” expressions for some functions. We can compute n! by constructing
the list [1,2,3,...,n] and then using the Prelude function product to multiply all of the list
items:
fc4 n = product [1..n]
This works correctly when n = 0 because the “product” of the empty list is defined to be 1:
Lec1> product []
5 HASKELL 40
Author: Peter Grogono
Date: 5 October 2003
Description: Example of a Literate Haskell source file
> module Vector where
This is a module for vectors in 3D Euclidean space.
We define V to be the constructor for vectors.
> data Vector = V Double Double Double
Vector is an instance of the type class Show.
This enables easy conversion of vectors to strings.
> instance Show Vector where
> show (V x y z) = "(" ++ show x ++ ", " ++ show y ++ ", " ++ show z ++ ")"
Vector is an equality type. Vectors are equal if all of their components
are equal. It might be an improvement to consider vectors to be equal if
their components are approximately equal.
> instance Eq Vector where
> V x y z == V x’ y’ z’ = x == x’ && y == y’ && z == z’
Vector is a numeric class. The following definitions provide sums,
differences, and outer (cross) products of vectors.
> instance Num Vector where
> V x y z + V x’ y’ z’ = V (x + x’) (y + y’) (z + z’)
> V x y z - V x’ y’ z’ = V (x - x’) (y - y’) (z - z’)
> V x y z * V x’ y’ z’ =
> V (y * z’ - y’ * z) (z * x’ - z’ * x) (x * y’ - x’ * y)
> negate (V x y z) = V (negate x) (negate y) (negate z)
A vector can be scaled by a real number. Haskell does not allow us to
overload ’*’ and ’/’ for vector/scalar operations.
> scale (V x y z) d = V (x * d) (y * d) (z * d)
These auxiliary functions define the dot product and length of a vector.
The expression ’norm v’ yields a unit vector with the same direction as v.
> dot (V x y z) (V x’ y’ z’) = x * x’ + y * y’ + z * z’
> len v = sqrt(dot v v)
> norm v = if l == 0
> then error "Norm of zero vector"
> else scale v (1/l)
> where l = len v
Figure 5: A Literate Haskell program
5 HASKELL 41
1 :: Integer
To illustrate some other styles of function definition, we will use an even simpler function: inc
adds 1 to its argument. The “obvious” definition is a single equation:
inc1 n = n + 1
Purists have a problem with this definition: they would prefer definitions to have to form x = e
in which x is the name of the thing being defined and e is the defining expression. We can write
definition in this way in Haskell:
inc2 = `n -> n + 1
We can read this definition as “inc2 is the function that maps n to n + 1”. The backslash
indicates that n is a parameter and the “arrow” -> links the parameter to the body of the function.
Backslash looks a bit like λ and the Haskell notation suggests the way in which function is written
in λ-calculus: λn. n + 1.
The expression \n -> n + 1 can be used alone: it is an anonymous function.
Lec1> (`n -> n + 1) 7
8 :: Integer
There is another way of writing this anonymous function. Its form is ? = 1 in which ? stands for
the unknown parameter. Putting parentheses around +1 makes this expression into a function:
Lec1> (+1) 6
7 :: Integer
In general, given a binary operator, we can create functions from it by giving it either a left or a
right operand. For example, (/3) is a function that divides its argument by 3 and (3/) is a function
that divides its argument into 3:
Lec1> (/3) 9
3.0 :: Double
Lec1> (12/) 5
2.4 :: Double
This provides us with another definition of inc:
inc3 = (+1)
Operators with one argument missing are called sections. The general rules are:
(op expr ) is equivalent to \x -> (x op expr)
(expr op ) is equivalent to \x -> (expr op x)
There is an important exception to this rule: (-1) is not a section: it is just −1. However,
(subtract expr) is a section equivalent to \x -> x - expr.
So far, all of our functions have had exactly one argument. In fact, all Haskell functions take
exactly one argument. Nevertheless, we can write definitions such as
add x y = x + y
and use the functions defined:
Lec1> add 347589 23874
371463 :: Integer
How do we account for this? The function add is actually a second order function that expects
one argument and returns a function that expects another argument. To see this, define
5 HASKELL 42
inc4 = add 1
and try it:
Lec1> inc4 5
6 :: Integer
Thus add 1 is “the function that adds 1 to its argument” and add 99 is “the function that adds
99 to its argument”, and so on. Haskell interprets the expression add 2 3 as “apply the function
add 2 to the argument 3”.
5.1.4 Syntax
Syntax conventions for functional programming are different from “traditional” programming
styles. You will have noticed, for example, that we write inc n rather than inc(n). Here are
some of the rules.
The “blank” operator denotes function application. Thus f x means “the result of applying the
function f to the argument x”. The technical name for the blank operator is juxtaposition.
Juxtaposition binds higher than any other operation.
Because application binds strongly, we must put parentheses around arguments that are not
simple. For example, the Haskell expression f n+1 denotes f(n) + 1; if we want the effect of
f(n + 1), we must write f(n+1).
Haskell interprets f x y as ((f x) y). First, the function f is applied to the argument x;
then the result, which must be a function, is applied to the argument y.
The application f(g(x)) applies g to x and then applies f to the result. We can write this in
Haskell either as f(g(x)) or as f (g x).
Compared to programs in most other languages, Haskell programs are remarkably free from “syn-
tactic clutter” — braces, semicolons, and such. The trade-off is that Haskell programmers must
respect layout rules. Unlike C
++
, for example, in Haskell indentation and line breaks make a
difference. The rules are simple and intuitive and, if you use existing source code as a model, you
may never have to learn them explicitly.
It is important to know that the rules exist, however, because otherwise Haskell’s error messages
can be quite confusing. For example, inserting this definition into lec1.hs
fc6 n = if
n == 0 then 1 else n * fc6(n - 1)
gives this error:
Reading file "e:/courses/comp348/src/lec1.hs":
Parsing
ERROR "e:/courses/comp348/src/lec1.hs":27 -
Syntax error in expression (unexpected ‘;’, possibly due to bad layout)
Haskell allows semicolons but does not require them. However, the parser inserts a semicolon at
the end of every line. Consequently, if you break a line in the wrong place, you get the mysterious
diagnostic “unexpected ‘;’ ”.
There’s nothing wrong with line breaks, provided that they occur in sensible places. The following
definition is fine:
fc6 n =
5 HASKELL 43
if n == 0
then 1
else n * fc6(n - 1)
We have already seen the Haskell case expression. Cases must appear on separate lines:
case e of
p1 -> e1
p2 -> e2
p3 -> e3
or separated by semicolons:
case e of p1 -> e1 ; p2 -> e2 ; p3 -> e3
There are a few other Haskell constructs for which layout is important, and we will state the rules
when we introduce the constructs.
5.1.5 Tracing
Debugging functional programs is difficult. One problem is that we cannot interleave print state-
ments with other statements because functions programs don’t have statements. Why? Because
statements have side-effects and functional programs are not supposed to have side-effects.
The Hugs Trace module provides a way around the problem of no side-effects. It defines a function
trace with two arguments: a string and an expression. When trace is called, it prints the message
as a side-effect and returns the expression.
To use the Trace module, write
import Trace
at the beginning of your program. To use it, replace any expression e by
(trace (s) (e))
in which s is a string. The parentheses are not always necessary (depending on the context) but
they do no harm.
Since trace returns the value of e, the meaning of your program should not be affected. However,
you will be able to monitor its progress by looking at what trace prints.
Here is a demonstration of Trace in use. We import the Trace module and define our old friend,
the factorial function. We keep the original version for later use and define another version with a
call to trace.
module Demo where
import Trace
¦-
fac 0 = 1
fac n = n * fac (n - 1)

fac 0 = 1
fac n = trace ("fac " ++ show n ++ "`n") (n * fac (n - 1))
Calling the traced version of fac gives:
5 HASKELL 44
Demo> fac 6
fac 6
fac 5
fac 4
fac 3
fac 2
fac 1
720 :: Integer
5.2 Details
5.2.1 Lists
We have already seen (in Section 5) that Haskell supports lists. Here are the basic built-in opera-
tions for lists:
Lists are enclosed in square brackets: [...].
The empty list is [].
Items within lists are separated by commas: [1,2,3].
A list may contain items of any type, but all of the items in a given list must have the same
type (note the use of -- to separate comments from code):
Lec1> [1,2,3] -- integer list - OK
[1,2,3] :: [Integer]
Lec1> ["abc","def"] -- string list - OK
["abc","def"] :: [[Char]]
Lec1> [’a’,’b’,’c’] -- a character list is a string
"abc" :: [Char]
Lec1> [1, 1.5] -- integers are promoted if necessary
[1.0,1.5] :: [Double]
Lec1> [1,’a’] -- mixed types - BAD!
ERROR - Illegal Haskell 98 class constraint in inferred type
*** Expression : [1,’a’]
*** Type : Num Char => [Char]
The infix operator : builds lists: it takes an item on the left and a list on the right. The
operator : is right-associative, so parentheses are not required in the second example below:
Lec1> 1:[]
[1] :: [Integer]
Lec1> 1:2:[]
[1,2] :: [Integer]
A very common idiom in Haskell progamming is x:xs (read “x followed by xes”). This is a
list consisting of an item x and a list xs. The initial letter may vary: for example, a:as,
y:ys, and so on.
5 HASKELL 45
The expression [x..y] constructs a list, provided that x and y are members of the same
enumerated type:
Lec1> [2..4]
[2,3,4] :: [Integer]
Lec1> [3,6,..20]
[3,6,9,12,15,18] :: [Integer]
Lec1> [’a’..’z’]
"abcdefghijklmnopqrstuvwxyz" :: [Char]
Lec1> [5..2]
[] :: [Integer]
The infix operator ++ concatenates lists.
There are functions for extracting the first item of a list (head) and the remaining items of the list
(tail) but we don’t often need them. The reason is that we can define list functions by using the
constructors as patterns. Most function definitions need only two cases: what is the result for an
empty list, []? and what is the result for a non-empty list x:xs?
For example, to find the length of a list, we define:
len [] = 0
len (x:xs) = 1 + len xs
The parentheses are required: since juxtaposition has higher precedence than colon, Haskell would
parse len x:xs as (len x):xs and report a type error.
If the append operator ++ did not exist, we could append lists using our own function app defined
by
app [] ys = ys
app (x:xs) ys = x:app xs ys
Any Haskell function with two arguments
1
can be used as an infix operator by putting back quotes
(grave accents) around it. The second example below demonstrates this. The third example shows
what happens if we try to build a list with mixed types.
Lec1> app [1,2,3] [4,5,6]
[1,2,3,4,5,6] :: [Integer]
Lec1> [1,2,3] ‘app‘ [4,5,6]
[1,2,3,4,5,6] :: [Integer]
Lec1> app [’a’,’b’] [1,2]
ERROR - Illegal Haskell 98 class constraint in inferred type
*** Expression : app [’a’,’b’] [1,2]
*** Type : Num Char => [Char]
A common operation on lists is to select items that satisfy a predicate. For example, to obtain the
even items in a list of integers, we could use the Prelude function even:
evens [] = []
evens (x:xs) = if even x then x:evens xs else evens xs
1
Of course, it hasn’t really got two arguments — see Section 5.1.3.
5 HASKELL 46
In functional programming, however, we usually want to generalize, using functions to do so.
The function filt takes a predicate p and a list, and returns the items in the list that satisfy
the predicate. (A better name would be filter but a function with this name is defined in the
Prelude: it does the same as filt.)
filt f [] = []
filt p (x:xs) = if p x then x:filt p xs else filt p xs
Now we can write:
Lec1> evens [1..20]
[2,4,6,8,10,12,14,16,18,20] :: [Integer]
Lec1> filt even [1..20]
[2,4,6,8,10,12,14,16,18,20] :: [Integer]
Using filt, we can write a better definition for evens:
evens = filt even
5.2.2 Lazy Evaluation
Haskell evaluates expressions lazily: this means that it never evaluates an expression until the
value is actually needed. It is perfectly alright to introduce a non-terminating definition into a
Haskell program, but things go wrong when we try to use it. Here is a non-terminating definition
loop = loop
and here is an attempt to use it (the message ¦Interrupted!¦ appears when the user interrupts
the program, for example, by entering ctrl-C):
Lec1> loop + 3
¦Interrupted!¦
Lec1>
There are other things that fail:
Lec1> 1/0
Program error: primDivDouble 1.0 0.0
But now define a constant function that doesn’t use its argument:
one n = 1
Since Haskell never evaluates the argument of this function, we can give it any argument at all:
Lec1> one 77
1 :: Integer
Lec1> one (1/0)
1 :: Integer
Lec1> one loop
1 :: Integer
Lazy evaluation allows us to work with infinite lists. The lists do not have an infinite number of
items: what “infinite” means in this context is “unbounded” — the list appears to have as many
items as we need. For example, if we define
from n = n:from (n+1)
nats = from 1
5 HASKELL 47
then from n yields an unbounded list starting at n and nats yields as many natural numbers as
we might need: [1,2,3,...]. We can use the built-in function take to bite off finite chunks from
the beginning of an infinite list:
Lec1> take 6 nats
[1,2,3,4,5,6] :: [Integer]
Lec1> take 6 (evens nats)
[2,4,6,8,10,12] :: [Integer]
Here is an alternative definition of nats:
nats = [1..]
As an application of infinite lists, the function nomul returns the items in list l that are not
multiples of n. The function sieve takes a list, and returns its first item n followed by the list
obtained by applying itself to a list with multiples of n.
nomul n l = filter (`m -> mod m n /= 0) l
sieve (n:ns) = n:sieve (nomul n ns)
This is what happens when we apply sieve to the list [2,3,4,...].
Lec1> sieve (from 2)
[2,3,5,7,11,13,17,19,23,29,31,37,41,43,47,53,59,61,67,71,73,79,83,89
Interrupted!
5.2.3 Pattern Matching
It is time to look more closely at some of the simple definitions we have already seen:
fac 0 = 1
fac n = n * fac(n - 1)
len [] = 0
len (x:xs) = 1 + len xs
The object following the function name on the left of = is a pattern. When the function is called,
the argument of the function is matched against the available patterns. If a match succeeds,
the expression on the right side of = is evaluated. Pattern matching may include binding, which
assigns values to variables in the expression.
Examples:
In the call fac 0, the argument 0 matches the pattern 0. The expression 1 is evaluated,
yielding 1.
In the call fac 3, the argument 3 matches the pattern n, yielding the binding ¦ n = 3 ¦. The
expression n * fac(n - 1) is evaluated with this binding, yielding 3 * fac(2).
In the call len [], the argument [] matches the pattern []. The expression 0 is evaluated,
yielding 0.
In the call len [27..30], the argument is equivalent to 27:28:29:30. This matches the
pattern x:xs, yielding the bindings ¦ x = 27, xs = [28..30]¦. The expression 1 + len xs is
evaluated with these bindings, yielding 1 + len [28..30].
Pattern matching occurs in a number of contexts in Haskell. All of these contexts can be expressed
in terms of the case expression:
5 HASKELL 48
Kind Pattern Expression Condition for match Binding
Wild card e none none
Constant k e k==e none
Variable v e none v = e
Constructor C p
1
p
1
C

e
1
e
2
C = C

, p
1
∼ e
1
, p
2
∼ e
2
Figure 6: Pattern matching
case e of p
1
-> e
1
; p
2
-> e
2
; ...; p
n
-> e
n
Here is how other kinds of expression map to case expressions:
` p
1
p
2
. . . p
n
→ e = `x
1
x
2
. . . x
n
→ case (x
1
, x
2
, . . . , x
n
) of (p
1
, p
2
, . . . , p
n
) → e
where the x
i
are new variables
if e
1
then e
2
else e
3
= case e
1
of True → e
2
; False → e
3
let p
1
= e
1
; p
2
= e
2
; . . . p
n
= e
n
; in e = case (e
1
, e
2
, . . . , e
n
) of (p
1
, p
2
, . . . , p
n
) → e
2
The case expression
case e of p
1
-> e
1
; p
2
-> e
2
; ...; p
n
-> e
n
is evaluated as follows:
The expression e is matched against p
1
, p
2
, . . . , p
n
in turn.
If e matches p
i
, then the value of the expression is obtained by evaluating e
i
with any bindings
obtained by matching.
If none of the patterns match, the expression fails.
Note that the order of patterns is significant. For example, the following definition of the factorial
function does not work, because the first pattern always succeeds and evaluation never terminates:
fac n = n * fc3(n - 1)
fac 0 = 1
Figure 6 summarizes the important kinds of pattern and the way in which they match. The
expression p ∼ e stands for “pattern p matches expression e”.
The underscore character, , is called a “wild card” because it matches anything but produces
no binding. Use the wild card when you don’t need the value of the argument in the function.
For example, the head and tail selectors for lists can be defined as
head (x: ) = x
tail ( :xs) = xs
A constant matches a constant with the same value but produces no binding. For example,
[] matches the empty list but nothing else.
A pattern consisting of a variable name matches any expression and binds that expression to
the variable.
If the pattern is a constructor with patterns as arguments, the expression must be an invocation
of the same constructor to expressions. The corresponding expressions are matched and the
bindings from the matches are combined. Since “:” is the list constructor, the pattern (3:xs)
matches any list whose first element is 3.
5 HASKELL 49
Haskell patterns are linear. This means that each variable in a pattern must be distinct. Suppose
that we are defining a function merge that has two lists and arguments and that we need to consider
the case in which the first item in both lists is the same. We are not allowed to write
merge (x:xs) (x:ys) = ... -- BAD!
Instead, we must write
merge (x:xs) (y:ys) = if x == y then ... else ...
Here are some examples of the pattern matching rules. First, the most common two cases for a
list function are the empty list pattern, [], and the non-empty list pattern, (x:xs). This pattern
matches the list with one item, in which case xs is bound to [].
Sometimes a third pattern is necessary. For example, if we define
swap [] = []
swap (x:y:ys) = y:x:swap ys
the expression “swap [3]” gives an error because the list with one element does not match any
pattern. The solution is to include this case in the definition:
swap [] = []
swap [x] = [x]
swap (x:y:ys) = y:x:swap ys
The tuple is another kind of constructor in Haskell. (A tuple corresponds roughly to a C
++
struct.) Tuples are enclosed in parentheses and the items inside them are separated by commas.
The items need not be of the same type. Unlike lists, tuples of different lengths have different
types.
A tuple with two items is usually called a pair. We can define selectors for pairs like this:
first (x, ) = x
second ( ,y) = y
However, these definitions are not really necessary, because the Prelude defines functions fst and
snd that do the same thing.
We can also define
plus (x,y) = x + y
This is not the same function as the Prelude function
add x y = x + y
The argument of plus must be a pair; there is no equivalent to “add 1”.
Functions can return tuples. The following function breaks a list into two roughly equal parts and
returns the two parts as a pair:
brk xs = (take (len xs ‘div‘ 2) xs, drop (len xs ‘div‘ 2) xs)
Obviously, we would like to avoid the repeated evaluation of “len xs ‘div‘ 2”. In a procedural
language, we would introduce a local variable. The Haskell equivalent is:
brk xs = (take m xs, drop m xs)
where m = len xs ‘div‘ 2
This may not look “functional” but in fact it is. It is “syntactic sugar” for
brk xs = (`m -> (take m xs, drop m xs)) (len xs ‘div‘ 2)
5 HASKELL 50
Description Syntax Associativity
simple term t
parenthesized term (t)
as-patterns a@p right
application f x left
do, if, left, `, case (left) right
case (leftwards) right
infix operators — see Figure 8
function types -> right
contexts =>
type constraints ::
do, if, left, `, case (rightwards) right
sequences ..
generators <-
grouping , n-ary
guards |
cases ->
definitions =
separation ;
Figure 7: Parsing precedence, from highest (at the top) to lowest (at the bottom).
Associativity
Precedence Left None Right
9 !! .
8 ^ ^^ **
7 * / ‘div‘ ‘mod‘
‘rem‘ ‘quot‘
6 + -
5 : ++
4 == /= < <= > >=
‘elem’ ‘notElem’
3 &&
2 ||
1 >> >>=
0 $ $! ‘seq’
Figure 8: Operator precedence, from highest (at the top) to lowest (at the bottom).
5.2.4 Operators
Figures 7 and 8 show the precedence rules that Haskell uses for parsing expressions and operators.
We have not encountered all of them yet.
Some constructs (do, if, let, \, and case) apparently have two precedences: one leftwards
and one rightwards. This is because of a metarule which says that each of these constructs
5 HASKELL 51
extends as far to the right as possible. In the first line in the table below, let seems to have
lower precedence than +, because the whole expression “a + a” is part of it. In the second
line, let seems to have higher precedence than +, because its components group together. The
second line is an application of the metarule: the let expression extends as far as possible.
Note parentheses override the metarule, as shown in the third lines of the table. Note also
that all three parsings are what we might intuitively expect.
The expression parses as
let a = 3 in a + a let a = 3 in (a + a)
k + let a = 3 in a + a k + (let a = 3 in (a + a))
k + (let a = 3 in a) + a (k + (let a = 3 in a)) + a
5.2.5 Building Haskell Programs
In procedural programming, we have a collection of control structures, or abstraction mechanisms
(assignment, sequence, conditional, loop, procedure call, etc.), and we use them to build programs.
In Functional Programming (FP), we have only one abstraction mechanism: function definition.
This makes the task of programming both simpler and harder: simpler because there are fewer
things to worry about, harder because we have a more limited toolkit.
The tendency in FP is to define a lot of small functions and then to build programs using these
functions. This can be confusing for the beginner, because the number of small functions is large.
But, with experience, this approach to programming becomes quite manageable.
The Prelude is a collection of functions that many programmers have found useful. It is often easier
to define a new function using Prelude functions than it is to write the function “from scratch”.
Here are some of the Prelude functions and their definitions.
Since Haskell does not have an explicit loop construct, recursion is used frequently. The key step
in writing a function is often to choose the appropriate parameter for recursion.
Recall that “take n xs” returns the first n items of the list x. If n is zero, the list is unchanged.
Otherwise, we use recursion on the value of n. To be safe, we define the appropriate response when
the list is empty.
take 0 xs = []
take [] = []
take n (x:xs) = x:take (n-1) xs
Recall also that “drop n xs” throws away the first n items of the list xs and returns the rest. The
definition is similar.
drop 0 xs = xs
drop [] = []
drop n ( :xs) = drop (n-1) xs
The functions takeWhile and dropWhile are only a little more complicated.
takeWhile p [] = []
takeWhile p (x:xs) = if p x then x : takeWhile p xs else []
dropWhile p [] = []
dropWhile p (x:xs) = if p x then dropWhile p xs else x:xs
Haskell provides another way of writing the last line. The infix operator @, read “as”, allows us to
write a pattern in two equivalent ways. In this case, we want xs and x:xs’ to stand for the same
5 HASKELL 52
list. Note that one or more primes may be added to a variable name to give a new name, as in
mathematics.
dropWhile p xs@(x:xs’) = if p x then dropWhile p xs’ else xs
The functions break and span are complementary. They split a list into two parts depending on
the value of a predicate.
Prelude> span (<5) [1..20]
([1,2,3,4],[5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20])
Prelude> break (>=5) [1..20]
([1,2,3,4],[5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20])
In particular, note:
Prelude> break isSpace "Here is a sentence."
("Here"," is a sentence.") :: ([Char],[Char])
The definitions of span and break are as follows:
span p [] = ([],[])
span p xs@(x:xs’) = if p x
then (x:ys, zs)
else ([],xs)
where (ys,zs) = span p xs’
break p = span (not . p)
With these functions working, we can write words, which splits a string into words separated by
blanks, and unwords, which puts the words together again.
words s = case dropWhile isSpace s of
"" -> []
s’ -> w : words s’’
where (w,s’’) = break isSpace s’
unwords [] = ""
unwords [w] = w
unwords (w:ws) = w ++ ’ ’ : unwords ws
Now we are ready for something more elaborate. Suppose we have a very long string of words,
perhaps an entire file, and we want to format so that each line is no longer than 80 characters.
The first step is to break into up into a list of words using the function words.
Here is the first version of a function that does the rest. The idea is to put words into a line until
the next word won’t fit — that is, it will make the line more than 80 characters long. When this
happens, we put the line into the output and carry on.
The problem is that we have no local variables and no state. Where do we store the information in
this “line”? The answer is that we need an accumulating parameter and recursion. We define
brk1 as
brk1 [] line = [line]
brk1 (w:ws) line = if length w + 1 + length line <= 80
then brk1 ws (line ++ " " ++ w)
else line : brk1 (w:ws) []
5 HASKELL 53
If there is something in our word string w:ws, this function calculates how long the line will become
if we append a blank and the word w to it. If the length is less than 80 characters, the word is
moved into the line. Otherwise, the result is a list consisting of the line and the result of processing
the rest of the word list. We start it off with an empty line:
brk1 sentence []
This function has one error: the first character of every line that it outputs is a blank. The next
version corrects that error.
brk2 [] line = [line]
brk2 (w:ws) line = if length w + 1 + length line <= 80
then
if line == ""
then brk2 ws w
else brk2 ws (line ++ " " ++ w)
else line : brk2 (w:ws) []
Finally, to make our function presentable to users, we wrap its parts up into a single definition:
format sen = brk (words sen) [] where
brk [] line = [line]
brk ws@(w:ws’) line = if length w + 1 + length line <= 80
then
if line == ""
then brk ws’ w
else brk ws’ (line ++ " " ++ w)
else line : brk ws []
5.2.6 Guards
The pattern-matching mechanism is handy but rather inflexible because it only allows for exact
matching. For example, if we want a factorial function that does not blow up for negative numbers,
we have to write
fac n = if n <= 0 then 1 else n * fac(n-1)
We can do better with guards. A guard is a condition that qualifies a pattern. As we know, the
pattern
n
matches any number. The pattern
n | n < 0
matches any negative number. Once we have provided a pattern, we can specify several guards,
each with their appropriate definitions. This version of factorial works for all integers:
fac n
| n <= 0 = 1
| n > 0 = n * fc(n-1)
The guard otherwise includes any conditions not covered by previous guards. For example, the
actual definitions of take and takeWhile in the Prelude are:
take n | n <= 0 = []
take [] = []
take n (x:xs) = x : take (n-1) xs
5 HASKELL 54
takeWhile p [] = []
takeWhile p (x:xs)
| p x = x : takeWhile p xs
| otherwise = []
5.3 Types
Haskell is a strongly typed language. This means that, if a program compiles without error, it
will either give an answer of the expected type or it will fail. Semantically, every type includes a
value ⊥ (pronounced “bottom” because it is the minimal element of a lattice) that denotes failure.
For example, if the type of a function is Int, the function will return either an integer or ⊥. Failure
is discussed in Section 5.3.2. For the rest of this section, we will discuss programs that do not fail.
Haskell uses the Hindley-Milner type system. An expression is written without type declara-
tions but, if it is type correct, the compiler can find a principle type or most general type for
it. There are a small number of situations in which the type checker needs a type declaration to
help it.
The following examples illustrate principle types. First, declare i to be the identity function. In
all of these examples, we follow the convention of declaring the type before the function. The type
declaration is accepted in this form by the Haskell compiler. It can also be obtained using the
:type command in the Hugs interpreter.
i :: a -> a
i x = x
The Haskell type is a -> a which we can read as “the type of a function that takes an argument
of any type and returns a result of the same type”. The letter a is called a type variable. In the
formal type system, this type is written ∀ a . a → a and is read “for all types a, the type a to a”.
The next example is a constant function: it takes two arguments and throws away the second one.
k :: a -> b -> a
k x y = x
There are two points to note. First, the type is a → b → a although you might have expected
ab → a. This is because Haskell functions take their arguments “one at a time”. The right arrow
associates to the right and the type a → b → a is parsed as a → (b → a). In other words, k is a
function that takes one argument of type a and returns a function with type b → a. The following
tests verifies this:
Types> k ’x’
k ’x’ :: a -> Char
Types> k ’x’ 5
’x’ :: Char
The second point is simply that the type contains two type variables, a and b. Since there is
nothing in the definition that connects the types, the compiler assumes that they are different
(although they don’t have to be).
In the next example, the two arguments are used to build a list. Consequently, they must have the
same type. The type [a] is the type of a list whose elements have type a.
two :: a -> a -> [a]
two x y = [x,y]
5 HASKELL 55
When we apply a function to an actual argument, the types become specialized. Providing a Char
as the first argument of two forces the second argument to be Char as well.
Types> two ’a’
two ’a’ :: Char -> [Char]
The function two ’a’ accepts an argument of type Char but an argument of another kind gives a
type error.
Types> two ’a’ ’b’
"ab" :: [Char]
Types> two ’a’ 3
ERROR - Illegal Haskell 98 class constraint in inferred type
*** Expression : two ’a’ 3
*** Type : Num Char => [Char]
The next function introduces a new feature: it uses the equality predicate ==. Haskell allows this
even without further type information, but adds a requirement to the type: it must be an equality
type. The expression Eq => is called a context and it says that the type a must be an instance
of the class Eq.
cmp :: Eq a => a -> a -> [a]
cmp x y = if x == y then [x,y] else []
If we change == to >, we get a different requirement: the type a must be an instance of the class
Ord, also called an ordinal type. The class Ord is a subclass of the class Eq: this means that any
type that provides > must also provide ==.
gt :: Ord a => a -> a -> [a]
gt x y = if x > y then [x,y] else []
Something else happens if we change one argument of > to the digit “1”. The context becomes
(Ord a, Num a), asserting that a must be both an ordinal type and a numeric type.
n1 :: (Ord a, Num a) => a -> a -> a
n1 x y = if x > 1 then x + y else 0
If we write 1.5 instead of 1, we obtain yet another context: a must be an ordinal type and a
fractional type.
n2 :: (Ord a, Fractional a) => a -> a -> a
n2 x y = if x > 1.5 then x + y else 0
As stated above, Haskell finds the most general type. We can specialize types in two ways. First,
we can add a type annotation to any variable or expression in the definition. Note that we are not
allowed to annotate a pattern — only an expression. If we specify that x is a Double, the compiler
infers that y must be a Double too, because x and y both occur in the expression x + y.
n3 :: Double -> Double -> Double
n3 x y = if (x::Double) > 1.5 then x + y else 0
We can also specify the type of the entire function. Here is the same function again, but with the
type Int specified by the programmer. Asking Haskell for the type of this function shows that it
has used the type we provided rather than the more general type.
n4 :: Int -> Int -> Int
n4 x y = if x > 1 then x + y else 0
Types> :t n4
5 HASKELL 56
n4 :: Int -> Int -> Int
Of course, if we do provide a type, we’d better get it right. In the next example, we provide an
incorrect type for the same function:
n5 :: Int -> Char -> Int
n5 x y = if x > 1 then x + y else 0
This attempt does not fool the compiler:
Reading file "e:/courses/comp348/src/types.hs":
Type checking
ERROR "e:/courses/comp348/src/types.hs":31 - Type error in application
*** Expression : x + y
*** Term : x
*** Type : Int
*** Does not match : Char
It is often easier to write functions than to figure out their types, which can get quite complicated.
In this example, we apply a function to a list. Haskell infers: that the most general type of the
function is a → b; the second argument of apply must be a list of as; and the result is a list of bs.
apply :: (a -> b) -> [a] -> [b]
apply f [] = []
apply f (x:xs) = f x : apply f xs
Types can get even longer, as the following example shows. The individual types are f :: a → b,
g :: c → a, h :: d → c; the type of x is d, and the type of the result is b.
compose :: (a -> b) -> (c -> a) -> (d -> c) -> d -> b
compose f g h x = f (g (h x))
The type of a tuple is written in the same way as values of the type. In this example, the type
(a, b) is “the type of pairs whose first element has type a and whose second element has type b”.
Similar notation is used for longer tuples.
cp :: a -> b -> (a,b)
cp x y = (x, y)
Summarizing what we have seen so far:
All valid Haskell functions and expressions have a type.
The type inferred by the compiler is the most general type.
The most general type may be specialized by type annotations provided by the user.
A context indicates that a type is an instance of a class.
5.3.1 Using Types
Since Haskell infers types, it is possible to virtually ignore types while programming. This is
unwise, for two reasons:
It is often useful to have a clear idea of the types of your variables when writing functions.
Haskell errors are very often caused by type mismatches. If you don’t understand the type
system, you will find these errors very hard to understand. If you do understand the type
system, you will find them quite helpful.
5 HASKELL 57
For example, one version of the function brk function in Section 5.2.5 looked like this:
brk1 [] line = [line]
brk1 (w:ws) line = if length w + 1 + length line <= 80
then brk1 ws (line ++ ’ ’ ++ w)
else line : brk1 (w:ws) []
Haskell did not accept this version, reporting:
Reading file "e:/courses/comp348/src/lec1.hs":
Type checking
ERROR "e:/courses/comp348/src/lec1.hs":9 - Type error in application
*** Expression : ’ ’ ++ w
*** Term : ’ ’
*** Type : Char
*** Does not match : [a]
To make sense of this error:
In the line beginning ERROR, the :9 after the path name is the line number. It identifies then
brk1 ws (line ++ ’ ’ ++ w) as the line with the error.
The expression which cannot be typed is ’ ’ ++ w.
The term (that is, part of the expression) which is causing the problem is ’ ’. It has type
Char, but the type checker is expecting something of type [a] — that is, a list.
The correction is to replace ’ ’ by something which is a list; in this case, [’ ’] or simply "
".
When you are reading someone else’s function definitions, knowing the types of the functions often
makes it easier to understand them. For this reason, it is a common convention in Haskell (and other
functional languages) to write the type declaration, as inferred by the compiler, above the function
declaration. Haskell accepts these type declarations, provided they are correctly formatted, and
reports an error if there is an inconsistency between the declaration type and the inferred type.
However, you are allowed to declare a more specialized type than the type that Haskell infers. For
example, if your module includes the declaration
app :: [Int] -> [Int] -> [Int]
app [] ys = ys
app (x:xs) ys = x:app xs ys
Haskell will accept the declaration, record the type you have specified, and give an error if you try
to use it incorrectly:
Lec1> :info app
app :: [Int] -> [Int] -> [Int]
Lec1> app "abc" "def"
ERROR - Type error in application
*** Expression : app "abc" "def"
*** Term : "def"
*** Type : String
*** Does not match : [Int]
There are some occasions when the type system doesn’t work as you would expect. You may have
already discovered this problem when representing sets as lists:
Lec1> union [] []
5 HASKELL 58
ERROR - Unresolved overloading
*** Type : Eq a => [a]
*** Expression : union [] []
In fact, even [] gives a type error:
Prelude> []
ERROR - Cannot find "show" function for:
*** Expression : []
*** Of type : [a]
This is a deficiency of the type system, but it is not an important one. If the expression union [] []
arises in a context, the empty lists will be the result of some computation and their types will be
known. Here is a contrived example: we define
comb n = union [1,3..n] [1,5..n]
and evaluate comb n with n = 0. This calls union [] [], but Haskell does not complain:
Lec1> comb 0
[] :: [Integer]
5.3.2 Failure
There are only two main causes of failure:
A function is given an argument that is of the correct type but outside the domain of the
function. E.g., head [] or 1/0.
A function fails to terminate.
Although the semantics says that the value of a failed program is always ⊥, the symptoms of
failure can vary quite a lot.
Domain errors are usually trapped by the runtime system:
Prelude> 1/0
Program error: primDivDouble 1.0 0.0
Prelude> head []
Program error: head []
A non-terminating program may run until it is interrupted, whether or not it is producing
output:
Prelude> [1..]
[1,2,3,4,5,6,Interrupted!
Prelude> f where f = f
Interrupted!
Some implementations of Hugs, including WinHugs, do not detect stack overflow, with the
result that programs may crash:
Prelude> g 1 where g n = g(n+1)
5.3.3 Reporting Errors
There is a built-in function
5 HASKELL 59
error :: String -> a
The result returned by this function is the only value that is shared by all types: ⊥ (pronounced
“bottom”). As a side-effect, it displays the string.
Here is an example of its use. The definition of head in the Prelude is:
head (x:xs) = x
head [] = error "headPreuldeList: head[]"
5.3.4 Standard Types
There are a number of types that are either built-in, defined in the Prelude, or defined in libraries.
The first letter of a type is always upper-case. This is not a convention: the compiler uses the case
of the first letter to distinguish type names from type variables, ordinary variables, and function
names.
The type Bool has values True and False.
The type Char has the ASCII characters as its values. The type String is equivalent to
[Char].
There are a number of numeric types. These include Int (fixed size integers); Integer (integers
of any size); Float (single-precision floats); Double (double precision floats); and Rational
(ratios of Integers).
The type () has only two values: (), which is the empty tuple, and ⊥ (bottom). This type
corresponds roughly to void in C
++
.
5.3.5 Standard Classes
The Standard Classes of Haskell are defined in the Prelude. Here are a couple of examples. The
equality class specifies that a type instance must provide the functions == and /=. However, if
the user supplies one of these functions, the compiler will provide the other one automatically, as
shown by the mutually recursive definitions.
class Eq a where
(==), (/=) :: a -> a -> Bool
-- Minimal complete definition: (==) or (/=)
x == y = not (x/=y)
x /= y = not (x==y)
The ordinal class inherits from the equality class and provides additional functions for comparison.
An instance type must provide at least <=; all of the other functions can be inferred by the compiler.
class (Eq a) => Ord a where
compare :: a -> a -> Ordering
(<), (<=), (>=), (>) :: a -> a -> Bool
max, min :: a -> a -> a
-- Minimal complete definition: (<=) or compare
-- using compare can be more efficient for complex types
compare x y | x==y = EQ
| x<=y = LT
| otherwise = GT
x <= y = compare x y /= GT
5 HASKELL 60

Eq

©
d
d
d

Show

©

Ord
d
d
d

Num

©
d
d
d

Enum
d
d
d

Real

©
d
d
d

Fractional

©
d
d
d

Integral

RealFrac
d
d
d

Floating

©

RealFloat
Figure 9: Standard Class Inheritance Graph
x < y = compare x y == LT
x >= y = compare x y /= LT
x > y = compare x y == GT
max x y | x <= y = y
| otherwise = x
min x y | x <= y = x
| otherwise = y
Figure 9 shows the class inheritance graph for the standard classes. There is no need to memorize
this graph: you will need it only if you are designing your own numeric type and, even then,
Haskell’s error messages will guide you. What is more important is to know the functions that a
class is supposed to implement; you can get this information from the Prelude, as shown above.
The class Show must provide a function that translates a value of the type into a string.
5.3.6 Defining New Types
As an illustration of how programmers can define new Haskell types, we discuss the type Frac,
shown in Figure 10. This type is actually unnecessary, because there is already a type Rational,
but it is not defined in the Prelude.
The first step is to declare the type and constructors for it. The name of the type is Frac and
the name of the only constructor is F. These names do not have to be different: we could have
used Frac for both. To construct a fraction, we must provide two integers:
Frac> F 9 12
9/12 :: Frac
5 HASKELL 61
module Frac where
data Frac = F Integer Integer
instance Show Frac where
show (F n d) = if d == 1
then show n
else show n ++ "/" ++ show d
instance Eq Frac where
F n d == F n’ d’ = n * d’ == n’ * d
instance Ord Frac where
F n d <= F n’ d’ = n * d’ <= n’ * d
instance Num Frac where
F n d + F n’ d’ = normalize( F (n * d’ + n’ * d) (d * d’) )
F n d - F n’ d’ = normalize( F (n * d’ - n’ * d) (d * d’) )
F n d * F n’ d’ = normalize( F (n * n’) (d * d’) )
negate (F n d) = normalize( F (negate n) d )
fromInteger n = F n 1
instance Fractional Frac where
recip (F n d) = normalize( F d n )
normalize :: Frac -> Frac
normalize (F n d) =
if d == 0
then error ("Bad fraction: " ++ show (F n d))
else if n == 0
then F 0 1
else if d < 0
then normalize (F (negate n) (negate d))
else F (n ‘div‘ g) (d ‘div‘ g)
where g = gcd n d
-- Harmonic series: 1/1 + 1/2 + 1/3 + ...
harm :: Integer -> [Frac]
harm n
| n <= 0 = []
| otherwise = harm (n - 1) ++ [F 1 n]
Figure 10: The Class Frac
5 HASKELL 62
We could introduce new function names for operations on fractions. For example, addFrac
might add two fractions. There are many advantages, however, to making the type Frac an
instance of numeric classes. If we do this, we can overload function names and, for example,
add fractions using +.
From Figure 9, we can see that any numeric class must be an instance of class Show. We
provide a version of the required function show that writes a fraction in the form "2/3". As
a nice touch, it shows fractions with denominator equal to 1 as integers.
The next step is to make Frac an equality type. We define ==, and Haskell automatically
provides /=.
We can provide an ordering for fractions. The function that we choose is not arbitrary: if
we provide <=, as here, Haskell will provide a number of other functions, including <, >, >=,
compare, min, and max.
Caution: providing a different comparison, such as >=, may lead to looping!
The next step is to make Frac an instance of Num, which requires overloaded definitions of +,
-, and *. It is useful also to define negate, because Haskell treats -x as (negate x), and a
function that converts integers to Fracs. Note that the Num class does not contain division
operations. This enables us to define types like Vector, which can be added, subtracted, and
multiplied, but not divided.
Since we do intend Fracs to be divided, we make Frac an instance of Fractional. All
that is needed is a function recip defined so that recip x = 1/x. Haskell will then define
x/y = x recip y.
We want to work with fractions in “normal form”. For example, 4/6 should be converted to 2/3
and 5/(−16) should be converted to −5/16. We also want to check for zero denominators and
put the fraction zero into the normal form 0/1. All of these tasks are performed by normalize,
which uses the standard function gcd to cancel common factors. We use normalize for every
operation that constructs a new fraction.
Finally and for fun, we introduce the function harm to construct a list of the fractions
1/1, 1/2, 1/3, . . . , 1/n.
Figure 11 shows a few simple examples of the new type. The functions work as expected. The
last example is of interest, because we did not define sum. However, sum is defined for all instances
of Num, so it works correctly for Frac. The result is exact because the components of a Frac are
Integer (“large integers”) rather than Int (32-bit integers).
5.3.7 Defining Data Types
A type may have more than one constructor. Enumerations have several constructions but no
data, as in these declarations:
data Bool = False | True
data Colour = Red | Green | Blue
Haskell provides special syntax (if/then/else) for conditional expressions because programmers
are familiar with the English words. The special syntax is actually unnecessary, because we could
define
if True x y = x
if False x y = y
relying on lazy evaluation to ensure that only one of x or y is actually evaluated.
5 HASKELL 63
Frac> F 3 4 > F 5 6
False :: Bool
Frac> F 3 4 + F 7 12
4/3 :: Frac
Frac> $$ > F 6 5
True :: Bool
Frac> F 7 4 / F 2 5
35/8 :: Frac
Frac> harm 5
[1,1/2,1/3,1/4,1/5] :: [Frac]
Frac> sum (harm 50)
13943237577224054960759/3099044504245996706400 :: Frac
Figure 11: Trying out type Frac
Binary Search Trees Constructors may also define data components. In Figure 12, we define
a type for binary trees with integers at the nodes. A BTree is either Empty, or consists of an Int
and two subtrees of type BTree. Here are some things to note about this definition:
Function show uses a “helper” function with an additional argument to display trees. Using
the tree t defined at the bottom of Figure 12:
BTree> t
1
2
3
4
5
6
8 :: BTree
The functions size, height, and treeToList are trivial to define by cases on the constructors.
The function insert assumes that the trees are binary search trees. A new entry is inserted
in the left or right subtree unless it is equal to an entry already in the tree, in which case it is
ignored (an alternative would be to report an error).
Comparing trees for equality, using == is straightforward. The important thing to note is that
the apparently unnecessary case Empty == Empty = True must be included — otherwise
the function loops!
Constructors may define data components. As an example, we will develop a type for binary trees
with two constructors: Empty constructs an empty tree, and Tree n l r constructs a tree with
integer n at the root, a left subtree l, and a right subtree r. Note that this is a recursive type
definition because BTree occurs on both sides of =. Just as with recursion in functions, we need a
base case (in this example the base case is Empty) and an inductive case.
data BTree = Empty | Tree Int BTree BTree
5 HASKELL 64
module BTree where
import Trace
data BTree = Empty | Tree Int BTree BTree
instance Show BTree where
show t = sh t ""
where
sh Empty sp = ""
sh (Tree n l r) sp = let sps = sp ++ " "
in sh r sps ++ sp ++ show n ++ "\n" ++ sh l sps
instance Eq BTree where
Empty == Empty = True
(Tree n l r) == (Tree n’ l’ r’) = n == n’ && l == l’ && r == r’
size Empty = 0
size (Tree n l r) = 1 + size l + size r
height Empty = 0
height (Tree n l r) = 1 + max (height l) (height r)
treeToList Empty = []
treeToList (Tree n l r) = treeToList l ++ treeToList r ++ [n]
listToTree [] = Empty
listToTree (n:ns) = insert n (listToTree ns)
insert m Empty = Tree m Empty Empty
insert m t@(Tree n l r) =
if m < n then Tree n (insert m l) r
else if m > n then Tree n l (insert m r)
else t
find Empty m = False
find (Tree n l r) m =
if m == n then True
else if m < n then find l m
else find r m
merge x y = listToTree(treeToList x ++ treeToList y)
Figure 12: Class BTree
5 HASKELL 65
We can build trees using the constructors. This is a bit tedious, and we will discuss easier ways
later.
t = Tree 1
( Tree 2
( Tree 3
Empty
( Tree 4
Empty
( Tree 5
Empty
Empty ) ) )
Empty )
( Tree 6 Empty Empty )
The simple functions below illustrate how we use the constructors as patterns in functions that
operate on trees. The size of a tree is the number of nodes that it contains and the height of a tree
is the length of the longest path from the root to a leaf node.
size Empty = 0
size (Tree n l r) = 1 + size l + size r
height Empty = 0
height (Tree n l r) = 1 + max (height l) (height r)
The function treeToList converts a tree to a list. The ordering of the list is significant: it would be
natural to use inorder (left/root/right), because this gives an ordered list of integers. However, we
want to be able to use the list to recreate the tree, and for this purpose postorder (left/right/root)
is better.
Here are these functions in use:
BTree> size t
6 :: Integer
BTree> height t
5 :: Integer
BTree> treeToList t
[5,4,3,2,6,1] :: [Int]
As with previous types, we make the type Btree an instance of class Show so that we can convert
trees to strings. The implementation of show below uses a “helper function” sh with a second
argument giving the amount of indentation to use when printing a node. In recursive calls, the
indentation is controlled by the depth of the node in the tree. For example:
BTree> t
6
1
2
5
4
3
:: BTree
5 HASKELL 66
We can easily define an equality relation on trees: trees are equal if the integers at their root nodes
are equal and their subtrees are (recursively) equal. Surprisingly, we must declare the Empty is
equal to Empty — without this component of the definition, the function loops.
instance Eq BTree where
Empty == Empty = True
(Tree n l r) == (Tree n’ l’ r’) = n == n’ && l == l’ && r == r’
Let us assume that our binary trees are going to be used as binary search trees (BST). That
is, we must ensure that all entries in every left subtree are less than the root and that all entries
in every right subtree are greater than the root. The following function inserts an integer into a
BST, maintaining the BST property. We have to decide what to do with duplicate entries, which
are not allowed in a binary search tree. This version of function insert ignores them.
It is important to “rebuild” the parts of the tree that we have taken apart. The following code for
the case m < n returns only part of the tree:
if m < n
then insert m l
The function insert provides another way of building trees. Note that the last node in the
expression is the first node to be inserted in the tree — and it becomes the root.
t1 = insert 1
(insert 8
(insert 3
(insert 2
(insert 5
(insert 2
(insert 6
(insert 4 Empty)))))))
BTree> t1
8
6
5
4
3
2
1
Although this is a slight improvement over calling constructors, it is still rather clumsy. Here is a
function that takes integers from a list and builds a BST from them.
listToTree [] = Empty
listToTree (n:ns) = insert n (listToTree ns)
t2 = listToTree [5, 2, 7, 0, 8, 3, 6, 1, 4]
BTree> t2
8
7
6
5
4
5 HASKELL 67
3
2
1
0
We have written treeToList and listToTree so that they are inverses.
BTree> t2 == listToTree (treeToList t2)
True :: Bool
It is easy to merge trees using treeToList and listToTree.
merge x y = listToTree (treeToList x ++ treeToList y)
Then, with
t3 = listToTree [14, 9, 13, 16, 11, 17, 19, 18, 10, 20, 12]
we have
BTree> merge t2 t3
20
19
18
17
16
14
13
12
11
10
9
8
7
6
5
4
3
2
1
0
:: BTree
The function find uses the familiar algorithm to decide whether a given integer is stored in the
tree. To make trees useful, we should store a key/value pair at each node. Then find would match
the key and return the corresponding value.
The fringe of a tree is a list of the leaf nodes as they would be visited in a left to right scan.
Assuming that we can recognize a leaf, it is straightforward to define fringe, as in Figure 13.
We can test these functions with the trees that we already have.
BTree> fringe t1
[1,3,5,8] :: [Int]
BTree> fringe t2
[0,2,5,7] :: [Int]
5 HASKELL 68
isLeaf (Tree _ Empty Empty) = True
isLeaf _ = False
fringe Empty = []
fringe t@(Tree n l r) =
if isLeaf t
then trace ("Leaf " ++ show n ++ "\n") (n : (fringe l ++ fringe r))
else fringe l ++ fringe r
sameFringe x y = fringe x == fringe y
Figure 13: The fringe of a tree
There is a famous problem in Computer Science called the “same fringe problem”. Given two trees,
how do we decide whether they have the same fringe? Since list comparison is provided for any
equality type, and Integer is an equality type, we have immediately:
sameFringe x y = fringe x == fringe y
BTree> sameFringe t1 t2
False :: Bool
Let us investigate this more closely. Here is a modified version of fringe that traces each discovery
of a leaf node:
fringe Empty = []
fringe t@(Tree n l r) =
if isLeaf t
then trace ("Leaf " ++ show n ++ "`n") (n : (fringe l ++ fringe r))
else fringe l ++ fringe r
This is what happens when we call sameFringe again with the traced version of fringe:
BTree> sameFringe t1 t2
Leaf 1
Leaf 0
False :: Bool
5.3.8 Types with Parameters
In this section, we revisit trees, but with a new twist. The trees in Section 5.3.7 were defined with
integers at the nodes. In the declaration below, we define trees with an arbitrary type, indicated
by the type variable a, at the nodes. On the left side of the declaration, Tree a indicates that the
type variable a is a parameter of the type. Whenever we refer to the type on the right side, we
must specify it as Tree a again.
module Tree where
data Tree a = Empty | T a (Tree a) (Tree a)
In practice, trees with integer nodes are not very useful. We need to store key/value pairs at
the nodes so that we can store useful information in the tree, as shown in Figure 14. The type
5 HASKELL 69
module Tree where
data (Ord k, Show k, Show v) => Tree k v =
Empty
| T k v (Tree k v) (Tree k v)
instance (Ord k, Show k, Show v) => Show (Tree k v) where
show Empty = ""
show (T key value left right) =
show left ++
leftjust (show key) 12 ++ show value ++ "\n" ++
show right
leftjust :: [Char] -> Int -> [Char]
leftjust s n
| length s >= n = s
| otherwise = leftjust (s ++ " ") n
out t = putStr (show t)
insert :: (Ord a, Show a, Show b) => a -> b -> Tree a b -> Tree a b
insert newKey newValue Empty = T newKey newValue Empty Empty
insert newKey newValue t@(T key value left right) =
if newKey == key then T key newValue left right
else if newKey < key then T key value (insert newKey newValue left) right
else T key value left (insert newKey newValue right)
build :: (Ord a, Show a, Show b) => [(a,b)] -> Tree a b
build [] = Empty
build ((key,value):ks) = insert key value (build ks)
t = build [
("Charles", 4163548695),
("Anne", 5144831733),
("George", 8147586977),
("Edward", 4504657464),
("Diana", 2124756843),
("Feridoon", 6063647586),
("Bill", 4508982212) ]
Figure 14: Key-value tree
5 HASKELL 70
variables k and v stand for “key” and “value” respectively. We require that k is an ordered type
so that we can compare keys.
When we construct a non-empty tree, we must now provide four arguments: the key, the value,
the left subtree, and the right subtree. The insertion function insert can use the operators ==, <,
and > with keys because the type of the keys is an instance of class Ord. When we supply a new
pair with a key that already exists in the tree, the new value replaces the old value.
As before, we can reduce the amount of writing required by providing a function, build, that
builds a tree from a list of pairs.
Here is a test:
t = build [ (6,"Anne"), (3,"Bill"), (8,"Charles"), (1,"Diana") ]
There is a catch with parameterized types: when we use them as instances of a class, we must
repeat all of the constraints. For example, to make Tree an instance of Show, we must write:
instance (Ord k, Show k, Show v) => Show (Tree k v) where
....
In the source file:
t = build [
("Charles", 4163548695),
("Anne", 5144831733),
("George", 8147586977),
("Edward", 4504657464),
("Diana", 2124756843),
("Feridoon", 6063647586),
("Bill", 4508982212) ]
A Haskell session:
Tree> t
"Anne" 5144831733
"Bill" 4508982212
"Charles" 4163548695
"Diana" 2124756843
"Edward" 4504657464
"Feridoon" 6063647586
"George" 8147586977
:: Tree [Char] Integer
Formulas It is often useful to define types with many constructors. The following declaration
introduces a type suitable for formulas in propositional calculus, with constants true and false,
propositional symbols (or names), conjunctions, disjunctions, negations, and implications.
module Logic where
data Formula =
TT
| FF
| Var String
| Con [Formula]
| Dis [Formula]
| Neg Formula
5 HASKELL 71
| Imp Formula Formula
When we define a function for a type like this one, we must usually include cases for all of the
constructors. The following definition of show is incomplete because it does not indicate operator
precedence.
instance Show Formula where
show TT = "true"
show FF = "false"
show (Var name) = name
show (Con []) = ""
show (Con [f]) = show f
show (Con (f:fs)) = show f ++ " and " ++ show (Con fs)
show (Dis []) = ""
show (Dis [f]) = show f
show (Dis (f:fs)) = show f ++ " or " ++ show (Dis fs)
show (Neg f) = "not " ++ show f
show (Imp p q) = show p ++ " => " ++ show q
As before, we can build formulas using the constructors, although this is rather tedious.
f = (Imp p (Imp p (Imp p p))) where p = Var "p"
g = Imp (Var "p") (Con [Var "q", Neg (Var "r")])
Logic> f
p => p => p => p :: Formula
Logic> g
p => q and not r :: Formula
5.4 Input and Output
Haskell handles input and output using monads. Moands are pure functions that provide the
effect of managing state changes. In this course, we will not discuss monads in detail; instead we
will just provide some simple examples of their use.
The first program reads a character and echoes it. The keyword do is followed by a sequence of
actions. Each action must have the same indentation. The function getChar performs an action
and returns a value, in this case a Char. The function putChar takes a Char as its argument and
performs an action — writing the character.
test1 = do
c <- getChar
putChar c
The next example shows how to write functions that perform actions. The function ready performs
the following sequence of actions: output a prompt; read a character; return a Bool value that
depends on the character read.
ready = do
putStr "Ready? "
char <- getChar
return (char == ’y’)
5 HASKELL 72
test2 = do
t <- ready
if t
then putStr "\nReady"
else putStr "\nNot ready"
The function getline in the next example is recursive. The else part requires an action sequence
that is introduced by a second do. The program test3 reads a line, applies a function (switch)
to it, and then writes the line.
getline = do
char <- getChar
if char == ’\n’
then return ""
else do
line <- getLine
return (char:line)
switch s = map sw s where
sw char =
if isLower char
then toUpper char
else if isUpper char
then toLower char
else char
test3 = do
line <- getline
putStr (switch line)
The next example uses files instead of console functions.
process s = filter (not . isSpace) s
test4 = do
s <- readFile "e:/temp/test.dat"
writeFile "e:/temp/test.res" (process s)
Since it is not good practice to build file names into programs, the next example reads the input
and output file names from the console.
main = do
putStr "Input file: "
iFile <- getLine
putStr "Output file: "
oFile <- getLine
s <- readFile iFile
writeFile oFile s
putStr (iFile ++ " copied to " ++ oFile ++ "\n")
5 HASKELL 73
5.4.1 String Formatting
By default, Hugs does not interpret formatting characters from strings. The following exchange
with the interpreter illustrates this (the standard function unlines converts a list of strings into a
single string with the list items separated by newline characters):
Prelude> unlines ["first line", "second line", "third line"]
"first line\nsecond line\nthird line\n"
The IO function putStr interprets the formatting characters correctly:
Prelude> putStr(unlines ["first line", "second line", "third line"])
first line
second line
third line
Since putStr :: String -> IO () is an IO function, you can use it only at the “top level” —
that is, as a command that you enter or as part of a do sequence. This does not matter in practice,
because there is no need to format strings until you actually want to see them at the end of a
calculation.
5.4.2 Interactive Programs
The standard IO functions include interact, a function that enables you to write interactive
programs easily. The argument for interact is a function of type String -> String. A simple
example of such a function is map toUpper. If you write
interact (map toUpper)
the interpreter will echo each character as you type it, converting lower case letters to upper case
letters.
The effect of lazy evaluation is important here. The effect is not to read a string of characters and
then output the same string with lower case letters converted. Instead, it reads each character
and then immediately echoes the converted character.
The following example provides a more elaborate version of the same idea. The function machine
:: String -> String simulates a vending machine. The argument string is a sequence of “com-
mands” interpreted as follows:
Command Effect
’n’ Insert a nickel (5 cents)
’d’ Insert a dime (10 cents)
’q’ Insert a quarter (25 cents)
’r’ Demand a refund
The output from the machine is a string defined as follows:
String Condition
"" the amount entered is less than 25 cents
"candy" the amount entered was exactly 25 cents
"candy and n cents change" the amount entered was 25 +n cents
"Refund: n" response to ’r’ with n cents stored
5 HASKELL 74
Here is the definition of the function machine:
machine commands = mac commands 0
where
mac "" stored = "Amount left: " ++ show stored
mac (’q’:cs) stored = respond (stored+25) cs
mac (’d’:cs) stored = respond (stored+10) cs
mac (’n’:cs) stored = respond (stored+5) cs
mac (’r’:cs) stored = "Refund: " ++ show stored ++ "\n" ++ mac cs 0
mac (_:cs) stored = mac cs stored
respond stored cs =
if stored == 25
then "candy\n" ++ mac cs 0
else if stored > 25
then "candy and " ++ show (stored-25) ++ " cents change\n" ++ mac cs 0
else mac cs stored
The following dialog with Hugs shows the result when the user enters the string "qdddnnr". To
end the input string, and hence the dialogue, the user enters ctrl-z. The second example shows
the response to "dddd" followed by ctrl-z, showing that the machine reports the amount it has
left before terminating.
VM> interact machine
candy
candy and 5 cents change
Refund: 10
Amount left: 0
VM> interact machine
candy and 5 cents change
Amount left: 10
VM> ....
5.5 Reasoning about programs
Haskell function definitions not only look like equations but, in an important sense, they are
equations. We will illustrate this with a simple example of program manipulation. The following
definitions of append (append two lists) and length (length of a list) should be familiar by now.
The only difference is that we have numbered the equations for future reference.
append [] ys = ys (1)
append (x:xs) ys = x : append xs ys (2)
length [] = 0 (3)
length (x:xs) = 1 + length xs (4)
Suppose that we want to define a new function len2 such that len2 xs ys returns the sum of the
lengths of the lists xs and ys. Here is a simple definition:
len2 xs ys = length (append xs ys) (5)
5 HASKELL 75
The obvious problem with this definition is that it is inefficient: it constructs a new list just for the
purpose of finding its length. How can we avoid this inefficiency? As a first step, we instantiate
(5) with xs = [] and then use (1) to simplify the resulting expression:
len2 [] ys = length (append [] ys) (6)
= length ys (7)
Next, we instantiate (5) with xs = (x:xs) and then use (2), (4), and (5) to simplify the resulting
expression:
len2 (x:xs) = length (append (x:xs) ys (8)
= length (x : append xs ys) (9)
= 1 + length (append xs ys) (10)
= 1 + len2 xs ys (11)
We can now combine (7) and (11) to obtain a definition of len2 that does not use append:
len2 [] ys = length ys
len2 (x:xs) ys = 1 + len2 xs ys
We could, of course, have obtained this definition of len2 with a little thought and no formal
deduction. Here is a somewhat more complicated example in which the final answer is less obvious.
Suppose we would like a function isSubList such that isSubList xs ys returns True if and only if
xs is a sublist of ys. The following definition is clearly correct but not, as it stands, implementable
(think of bs as the “before list” and as as the “after list”):
isSubList xs ys = ∃ bs,as • bs ++ xs ++ as == ys
To make progress, we assume that ys is not empty (and so can be written y:ys’) and that there
are two cases: either bs = [] or bs = y:bs’ (clearly, if bs is not empty, its first element must be
y). Then we have:
isSubList xs (y:ys’) = ∃ as • [] ++ xs ++ as == ys’
∨ ∃ bs’,as • y:bs’ ++ xs ++ as == y:ys’
= ∃ as • xs ++ as == ys’
∨ ∃ bs’,as • bs’ ++ xs ++ as == ys’
= ∃ as • xs ++ as == ys’
∨ isSubList xs ys’
The second disjunct is a straightforward recursion. The first requires another function which, on
the basis of its required behaviour, we will call isPrefix. The definition of isPrefix is
isPrefix xs ys = ∃ as • xs ++ as == ys
Assuming both lists are not empty, we can write this as
isPrefix (x:xs’) (y:ys’) = ∃ as • (x:xs’) ++ as == (y:ys’)
= x == y ∧ ∃ as’ • xs’ ++ as’ == ys’
= x == y ∧ isPrefix xs’ ys’
It is straightforward to add the base cases to obtain the following definition:
5 HASKELL 76
isSubList [] _ = True
isSubList _ [] = False
isSubList xs (y:ys) = isPrefix xs (y:ys) || isSubList xs ys
where
isPrefix [] _ = True
isPrefix _ [] = False
isPrefix (x:xs) (y:ys) = x == y && isPrefix xs ys
6 The Object Oriented Paradigm
A fundamental feature of our understanding of the world is that we organize our ex-
perience as a number of distinct object (tables and chairs, bank loans and algebraic
expressions, . . .) . . . . When we wish to solve a problem on a computer, we often
need to construct within the computer a model of that aspect of the real or conceptual
world to which the solution of the problem will be applied. (Hoare 1968)
Object Oriented Programming (OOP) is the currently dominant programming paradigm. For
many people today, “programming” means “object oriented programming”. Nevertheless, there
is a substantial portion of the software industry that has not adopted OOP, and there remains
widespread misunderstanding about what OOP actually is and what, if any, benefits it has to offer.
6.1 Simula
Many programs are computer simulations of the real world or a conceptual world. Writing such
programs is easier if there is a correspondence between objects in the world and components of the
program.
Simula originated in the Norwegian Computing Centre in 1962. Kristen Nygaard proposed a
language for simulation to be developed by himself and Ole-Johan Dahl (1978). Key insights were
developed in 1965 following experience with Simula I:
the purpose of the language was to model systems;
a system is a collection of interacting processes;
a process can be represented during program execution by multiple procedures each with its
own Algol-style stacks.
The main lessons of Simula I were:
the distinction between a program text and its execution;
the fact that data and operations belong together and that most useful programming constructs
contain both.
Simula 67 was a general purpose programming language that incorporated the ideas of Simula I
but put them into a more general context. The basic concept of Simula 67 was to be “classes of
objects”. The major innovation was “block prefixing” (Nygaard and Dahl 1978).
Prefixing emerged from the study of queues, which are central to discrete-event simulations. The
queue mechanism (a list linked by pointers) can be split off from the elements of the queue (ob-
jects such as trucks, buses, people, and so on, depending on the system being simulated). Once
recognized, the prefix concept could be used in any context where common features of a collection
of classes could be abstracted in a prefixed block or, as we would say today, a superclass.
Dahl (1978, page 489) has provided his own analysis of the role of blocks in Simula.
Deleting the procedure definitions and final statement from an Algol block gives a pure data
record.
Deleting the final statement from an Algol block, leaving procedure definitions and data dec-
larations, gives an abstract data object.
Adding coroutine constructs to Algol blocks provides quasi-parallel programming capabilities.
77
6 THE OBJECT ORIENTED PARADIGM 78
Listing 19: A Simula Class
class Account (real balance);
begin
procedure Deposit (real amount)
balance := balance + amount;
procedure Withdraw (real amount)
balance := balance - amount;
end;
Adding a prefix mechanism to Algol blocks provides an abstraction mechanism (the class
hierarchy).
All of these concepts are subsumed in the Simula class.
Listing 19 shows a simple Simula-style class, with modernized syntax. Before we use this class, we
must declare a reference to it and then instantiate it on the heap:
ref (Account) myAccount;
MyAccount := new Account(1000);
Note that the class is syntactically more like a procedure than a data object: it has an (optional)
formal parameter and it must be called with an actual parameter. It differs from an ordinary Algol
procedure in that its AR remains on the heap after the invocation.
To inherit from Account we write something like:
Account class ChequeAccount (real amount);
. . . .
This explains the term “prefixing” — the declaration of the child class is prefixed by the name of
its parent. Inheritance and subclasses introduced a new programming methodology, although this
did not become evident for several years.
Simula provides:
coroutines that permit the simulation of concurrent processes;
multiple stacks, needed to support coroutines;
classes that combine data and a collection of functions that operate on the data;
prefixing (now known as inheritance) that allows a specialized class to be derived from a
general class without unnecessary code duplication;
a garbage collector that frees the programmer from the responsibility of deallocating storage.
Simula itself had an uneven history. It was used more in Europe than in North America, but it
never achieved the recognition that it deserved. This was partly because there were few Simula
compilers and the good compilers were expensive.
On the other hand, the legacy that Simula left is considerable: a new paradigm of programming.
Exercise 23. Coroutines are typically implemented by adding two new statements to the language:
suspend and resume. A suspend statement, executed within a function, switches control to a
scheduler which stores the value of the program counter somewhere and continues the execution of
another function. A resume statement has the effect of restarting a previously suspended function.
At any one time, exactly one function is executing. What features might a concurrent language
provide that are not provided by coroutines? What additional complications are introduced by
concurrency? 2
6 THE OBJECT ORIENTED PARADIGM 79
6.2 Smalltalk
Smalltalk originated with Alan Kay’s reflections on the future of computers and programming in
the late 60s. Kay (1996) was influenced by LISP, especially by its one-page metacircular interpreter,
and by Simula. It is interesting to note that Kay gave a talk about his ideas at MIT in November
1971; the talk inspired Carl Hewitt’s work on his Actor model, an early attempt at formalizing
objects. In turn, Sussman and Steele wrote an interpreter for Actors that eventually became
Scheme (see Section 4.2). The cross-fertilization between OOPL and FPL that occurred during
these early days has, sadly, not continued.
The first version of Smalltalk was implemented by Dan Ingalls in 1972, using BASIC (!) as the
implementation language (Kay 1996, page 533). Smalltalk was inspired by Simula and LISP; it
was based on six principles (Kay 1996, page 534).
1. Everything is an object.
2. Objects communicate by sending and receiving messages (in terms of objects).
3. Objects have their own memory (in terms of objects).
4. Every object is an instance of a class (which must be an object).
5. The class holds the shared behaviour for its instances (in the form of objects in a program
list).
6. To [evaluate] a program list, control is passed to the first object and the remainder is treated
as its message.
Principles 1–3 provide an “external” view and remained stable as Smalltalk evolved. Principles
4–6 provide an “internal” view and were revised following implementation experience. Principle
6 reveals Kay’s use of LISP as a model for Smalltalk — McCarthy had described LISP with a
one-page meta-circular interpreter; one of Kay’s goals was to do the same for Smalltalk.
Smalltalk was also strongly influenced by Simula. However, it differs from Simula in several ways:
Simula distinguishes primitive types, such as integer and real, from class types. In Smalltalk,
“everything is an object”.
In particular, classes are objects in Smalltalk. To create a new instance of a class, you send a
message to it. Since a class object must belong to a class, Smalltalk requires metaclasses.
Smalltalk effectively eliminates passive data. Since objects are “active” in the sense that they
have methods, and everything is an object, there are no data primitives.
Smalltalk is a complete environment, not just a compiler. You can edit, compile, execute, and
debug Smalltalk programs without ever leaving the Smalltalk environment.
The “block” is an interesting innovation of Smalltalk. A block is a sequence of statements that
can be passed as a parameter to various control structures and behaves rather like an object. In
the Smalltalk statement
10 timesRepeat: [Transcript nextPutAll: ’ Hi!’]
the receiver is 10, the message is timesRepeat:, and [Transcript nextPutAll: ’ Hi!’] is
the parameter of the message.
Blocks may have parameters. One technique for iteration in Smalltalk is to use a do: message
which executes a block (its parameter) for each component of a collection. The component is
6 THE OBJECT ORIENTED PARADIGM 80
passed to the block as a parameter. The following loop determines whether "Mika" occurs in the
current object (assumed to be a collection):
self do: [:item | item = "Mika" ifTrue: [^true]].
^false
The first occurrence of item introduces it as the name of the formal parameter for the block.
The second occurrence is its use. Blocks are closely related to closures (Section 4.2). The block
[:x | E] corresponds to λx . E.
The first practical version of Smalltalk was developed in 1976 at Xerox Palo Alto Research Center
(PARC). The important features of Smalltalk are:
everything is an object;
an object has private data and public functions;
objects collaborate by exchanging “messages”;
every object is a member of a class;
there is an inheritance hierarchy (actually a tree) with the class Object as its root;
all classes inherit directly or indirectly from the class Object;
blocks;
coroutines;
garbage collection.
Exercise 24. Discuss the following statements (Gelernter and Jagannathan 1990, page 245).
1. [Data objects are replaced by program structures.] Under this view of things, particularly
under Smalltalk’s approach, there are no passive objects to be acted upon. Every entity is an
active agent, a little program in itself. This can lead to counter-intuitive interpretations of simple
operations, and can be destructive to a natural sense of hierarchy.
2. Algol 60 gives us a representation for procedures, but no direct representation for a procedure
activation. Simula and Smalltalk inherit this same limitation: they give us representations for
classes, but no direct representations for the objects that are instantiated from classes. Our
ability to create single objects and to specify (constant) object “values” is compromised. 2
Exercise 25. In Smalltalk, an object is an instance of a class. The class is itself an object, and
therefore is a member of a class. A class whose instances are classes is called a metaclass. Do
these definitions imply an infinite regression? If so, how could the infinite regression by avoided?
Is there a class that is a member of itself? 2
6.3 CLU
CLU is not usually considered an OOPL because it does not provide any mechanisms for incre-
mental modification. However:
The resulting programming methodology is object-oriented: programs are developed
by thinking about the objects they manipulate and then inventing a modular structure
based on these objects. Each type of object is implemented by its own program module.
(Liskov 1996, page 475)
Parnas (1972) introduced the idea of “information hiding”: a program module should have a public
interface and a private implementation. Liskov and Zilles (1974) introduced CLU specifically to
support this style of programming.
6 THE OBJECT ORIENTED PARADIGM 81
An abstract data type (ADT) specifies a type, a set of operations that may be performed on
instances of the type, and the effects of these operations. The implementation of an ADT provides
the actual operations that have the specified effects and also prevents programmers from doing
anything else. The word “abstract” in CLU means “user defined”, or not built in to the language:
I referred to the types as “abstract” because they are not provided directly by a pro-
gramming languages but instead must be implemented by the user. An abstract type
is abstract in the same way that a procedure is an abstract operation. (Liskov 1996,
page 473)
This is not a universally accepted definition. Many people use the term ADT in the following
sense:
An abstract data type is a system consisting of three constituents:
1. some sets of objects;
2. a set of syntactic descriptions of the primitive functions;
3. a semantic description — that is, a sufficiently complete set of relationships that
specify how the functions interact with each other.
(Martin 1986)
The difference is that an ADT in CLU provides a particular implementation of the type (and, in
fact, the implementation must be unique) whereas Martin’s definition requires only a specification.
A stack ADT in CLU would have functions called push and pop to modify the stack, but a
stack ADT respecting Martin’s definition would define the key property of a stack (the sequence
push(); pop() has no effect) but would not provide an implementation of these functions.
Although Simula influenced the design of CLU, it was seen as deficient in various ways. Simula:
does not provide encapsulation: clients could directly access the variables, as well as the
functions, of a class;
does not provide “type generators” or, as we would say now, “generic types” (“templates” in
C++);
associates operations with objects rather than types;
handles built-in and user-defined types differently — for example, instances of user-defined
classes are always heap-allocated.
CLU’s computational model is similar to that of LISP: names are references to heap-allocated
objects. The stack-only model of Algol 60 was seen as too restrictive for a language supporting
data abstractions. The reasons given for using a heap include:
Declarations are easy to implement because, for all types, space is allocated for a pointer.
Stack allocation breaks encapsulation because the size of the object (which should be an
implementation “secret”) must be known at the point of declaration.
Variable declaration is separated from object creation. This simplifies safe initialization.
Variable and object lifetimes are not tied. When the variable goes out of scope, the object
continues to exist (there may be other references to it).
The meaning of assignment is independent of type. The statement x := E means simply
that x becomes a reference (implemented as a pointer, of course) to E. There is no need to
worry about the semantics of copying an object. Copying, if needed, should be provided by a
member function of the class.
6 THE OBJECT ORIENTED PARADIGM 82
Listing 20: A CLU Cluster
intset = cluster is create, insert, delete, member, size, choose
rep = array[int]
create = proc () returns (cvt)
return (rep$new())
end create
insert = proc (s: intset, x: int)
if ∼member(s, x) then rep$addh(down(s), x) end
end insert
member = proc (s: cvt, x: int) returns (bool)
return (getind(s, x) ≤ rep$high(s))
end member
. . . .
end intset
Reference semantics requires garbage collection. Although garbage collection has a run-time
overhead, efficiency was not a primary goal of the CLU project. Furthermore, garbage collec-
tion eliminates memory leaks and dangling pointers, notorious sources of obscure errors.
CLU is designed for programming with ADTs. Associated with the interface is an implementation
that defines a representation for the type and provides bodies for the functions.
Listing 20 shows an extract from a cluster that implements a set of integers (Liskov and Guttag
1986, page 63). The components of a cluster, such as intset, are:
1. A header that gives the name of the type being implemented (intset) and the names of
the operations provided (create, insert, . . . .).
2. A definition of the representation type, introduced by the keyword rep. An intset is repre-
sented by an array of integers.
3. Implementations of the functions mentioned in the header and, if necessary, auxiliary func-
tions that are not accessible to clients.
The abstract type intset and the representation type array[int], referred to as rep, are distinct.
CLU provides special functions up, which converts the representation type to the abstract type,
and down, which converts the abstract type to the representation type. For example, the function
insert has a parameter s of type intset which is converted to the representation type so that it
can be treated as an array.
It often happens that the client provides an abstract argument that must immediately be converted
to the representation type; conversely, many functions create a value of the representation type
but must return a value of the abstract type. For both of these cases, CLU provides the keyword
cvt as an abbreviation. In the function create above, cvt converts the new instance of rep to
intset before returning, and in the function member, the formal parameter s:cvt indicates that
the client will provide an intset which will be treated in the function body as a rep.
All operations on abstract data in CLU contain the type of the operator explicitly, using the syntax
type$operation.
CLU draws a distinction between mutable and immutable objects. An object is mutable if its value
can change during execution, otherwise it is immutable. For example, integers are immutable but
arrays are mutable (Liskov and Guttag 1986, pages 19–20).
6 THE OBJECT ORIENTED PARADIGM 83
The mutability of an object depends on its type and, in particular, on the operations provided by
the type. Clearly, an object is mutable if its type has operations that change the object. After
executing
x: T = intset$create()
y: T
y := x
the variables x and y refer to the same object. If the object is immutable, the fact that it is shared
is undetectable by the program. If it is mutable, changes made via the reference x will be visible
via the reference y. Whether this is desirable depends on your point of view.
The important contributions of CLU include:
modules (“clusters”);
safe encapsulation;
distinction between the abstract type and the representation type;
names are references, not values;
the distinction between mutable and immutable objects;
analysis of copying and comparison of ADTs;
generic clusters (a cluster may have parameters, notably type parameters);
exception handling.
Whereas Simula provides tools for the programmer and supports a methodology, CLU provides
the same tools and enforces the methodology. The cost is an increase in complexity. Several
keywords (rep, cvt, down, and up) are needed simply to maintain the distinction between abstract
and representation types. In Simula, three concepts — procedures, classes, and records — are
effectively made equivalent, resulting in significant simplification. In CLU, procedures, records,
and clusters are three different things.
The argument in favour of CLU is that the compiler will detect encapsulation and other errors.
But is it the task of a compiler to prevent the programmer doing things that might be completely
safe? Or should this role be delegated to a style checker, just as C programmers use both cc (the
C compiler) and lint (the C style checker)?
The designers of CLU advocate defensive programming:
Of course, checking whether inputs [i.e., parameters of functions] are in the permit-
ted subset of the domain takes time, and it is tempting not to bother with the checks,
or to use them only while debugging, and suppress them during production. This is
generally an unwise practice. It is better to develop the habit of defensive programming,
that it writing each procedure to defend itself against errors. . . .
Defensive programming makes it easier to debug programs. . . . (Liskov and Guttag
1986, page 100)
Exercise 26. Do you agree that defensive programming is a better habit than the alternative
suggested by Liskov and Guttag? Can you propose an alternative strategy? 2
6.4 C++
C++ was developed at Bell Labs by Bjarne Stroustrup (1994). The task assigned to Stroustrup was
to develop a new systems PL that would replace C. C++ is the result of two major design decisions:
6 THE OBJECT ORIENTED PARADIGM 84
first, Stroustrup decided that the new language would be adopted only if it was compatible with
C; second, Stroustrup’s experience of completing a Ph.D. at Cambridge University using Simula
convinced him that object orientation was the correct approach but that efficiency was essential for
acceptance. C++ is widely used but has been sharply criticized (Sakkinen 1988; Sakkinen 1992;
Joyner 1992).
C++:
is almost a superset of C (that is, there are only a few C constructs that are not accepted by
a C++ compiler);
is a hybrid language (i.e., a language that supports both imperative and OO programming),
not a “pure” OO language;
emphasizes the stack rather than the heap, although both stack and heap allocation is provided;
provides multiple inheritance, genericity (in the form of “templates”), and exception handling;
does not provide garbage collection.
Exercise 27. Explain why C++ is the most widely-used OOPL despite its technical flaws. 2
6.5 Eiffel
Software Engineering was a key objective in the design of Eiffel.
Eiffel embodies a “certain idea” of software construction: the belief that it is possible to
treat this task as a serious engineering enterprise, whose goal is to yield quality software
through the careful production and continuous development of parameterizable, scien-
tifically specified, reusable components, communicating on the basis of clearly defined
contracts and organized in systematic multi-criteria classifications. (Meyer 1992)
Eiffel borrows from Simula 67, Algol W, Alphard, CLU, Ada, and the specification language Z.
Like some other OOPLs, it provides multiple inheritance. It is notable for its strong typing, an
unusual exception handling mechanism, and assertions. By default, object names are references;
however, Eiffel also provides “expanded types” whose instances are named directly.
Eiffel does not have: global variables; enumerations; subranges; goto, break, or continue state-
ments; procedure variables; casts; or pointer arithmetic. Input/output is defined by libraries rather
than being built into the language.
The main features of Eiffel are as follows. Eiffel:
is strictly OO: all functions are defined as methods in classes;
provides multiple inheritance and repeated inheritance;
supports “programming by contract” with assertions;
has an unusual exception handling mechanism;
provides generic classes;
may provide garbage collection.
6.5.1 Programming by Contract
The problem with writing a function to compute square roots is that it is not clear what to do if
the argument provided by the caller is negative. The solution adopted in Listing 21 is to allow for
the possibility that the argument might be negative.
6 THE OBJECT ORIENTED PARADIGM 85
Listing 21: Finding a square root
double sqrt (double x)
¦
if (x ≥ 0.0)
return . . . . //

x
else
??
¦
Listing 22: A contract for square roots
sqrt (x: real): real is
require x ≥ 0.0
Result := . . . . --

x
ensure abs(Result * Result / x - 1.0) ≤ 1e-6
end
This approach, applied throughout a software system, is called defensive programming
(Section 6.3).
Testing arguments adds overhead to the program.
Most of the time spent testing is wasted, because programmers will rarely call sqrt with a
negative argument.
The failure action, indicated by “??” in Listing 21, is hard to define. In particular, it is
undesirable to return a funny value such as −999.999, or to send a message to the console
(there may not be one), or to add another parameter to indicate success/failure. The only
reasonable possibility is some kind of exception.
Meyer’s solution is to write sqrt as shown in Listing 22. The require/ensure clauses constitute
a contract between the function and the caller. In words:
If the actual parameter supplied, x, is positive, the result returned will be approximately

x.
If the argument supplied is negative, the contract does not constrain the function in any way at all:
it can return -999.999 or any other number, terminate execution, or whatever. The caller therefore
has a responsibility to ensure that the argument is valid.
The advantages of this approach include:
Many tests become unnecessary.
Programs are not cluttered with tests for conditions that are unlikely to arise.
The require conditions can be checked dynamically during testing and (perhaps) disabled
during production runs. (Hoare considers this policy to be analogous to wearing your life
jacket during on-shore training but leaving it behind when you go to sea.)
Require and ensure clauses are a useful form of documentation; they appear in Eiffel inter-
faces.
Functions in a child class must provide a contract at least as strong as that of the corresponding
function in the parent class, as shown in Listing 23. We must have R
p
⇒ R
c
and E
c
⇒ E
p
. Note
6 THE OBJECT ORIENTED PARADIGM 86
Listing 23: Eiffel contracts
class Parent
feature
f() is
require R
p
ensure E
p
end
. . . .
class Child inherit Parent
feature
f() is
require R
c
ensure E
c
. . . .
the directions of the implications: the ensure clauses are “covariant” but the require clauses are
“contravariant”. The following analogy explains this reversal.
A pet supplier offers the following contract: “if you send at least $200, I will send you an animal”.
The pet supplier has a subcontractor who offers the following contract: “if you send me at least
$150, I will send you a dog”. These contracts respect the Eiffel requirements: the contract offered
by the subcontractor has a weaker requirement ($150 is less than $200) but promises more (a dog
is more specific than an animal). Thus the contractor can use the subcontractor when a dog is
required and make a profit of $50.
Eiffel achieves the subcontract specifications by requiring the programmer to define
R
c
≡ R
p

E
c
≡ E
p

The ∨ ensures that the child’s requirement is weaker than the parent’s requirement, and the ∧
ensures that the child’s commitment is stronger than the parent’s commitment.
In addition to require and ensure clauses, Eiffel has provision for writing class invariants. A
class invariant is a predicate over the variables of an instance of the class. It must be true when
the object is created and after each function of the class has executed.
Eiffel programmers are not obliged to write assertions, but it is considered good Eiffel “style” to
include them. In practice, it is sometimes difficult to write assertions that make non-trivial claims
about the objects. Typical contracts say things like: “after you have added an element to this
collection, it contains one more element”. Since assertions may contain function calls, it is possible
in principle to say things like: “this structure is a heap (in the sense of heap sort)”. However:
this does not seem to be done much in practice;
complex assertions increase run-time overhead, increasing the incentive to disable checking;
a complex assertion might itself contain errors, leading to a false sense of confidence in the
correctness of the program.
6.5.2 Repeated Inheritance
Listing 24 shows a collection of classes (Meyer 1992, page169). The features of HomeBusiness
include:
6 THE OBJECT ORIENTED PARADIGM 87
Listing 24: Repeated inheritance in Eiffel
class House
feature
address: String
value: Money
end
class Residence inherit House
rename value as residenceValue
end
class Business inherit House
rename value as businessValue
end
class HomeBusiness inherit Residence Business
. . . .
end
address, inherited once but by two paths;
residenceValue, inherited from value in class House;
businessValue, inherited from value in class House.
Repeated inheritance can be used to solve various problems that appear in OOP. For example, it is
difficult in some languages to provide collections with more than one iteration sequence. In Eiffel,
more than one sequence can be obtained by repeated inheritance of the successor function.
C++ has a similar mechanism, but it applies to entire classes (via the virtual mechanism) rather
than to individual attributes of classes.
6.5.3 Exception Handling
An Eiffel function must either fulfill its contract or report failure. If a function contains a rescue
clause, this clause is invoked if an operation within the function reports failure. A return from
the rescue clause indicate that the function has failed. However, a rescue clause may perform
some cleaning-up actions and then invoke retry to attempt the calculation again. The function in
Listing 25 illustrates the idea. The mechanism seems harder to use than other, more conventional,
exception handling mechanisms. It is not obvious that there are many circumstances in which it
makes sense to “retry” a function.
Exercise 28. Early versions of Eiffel had a reference semantics (variable names denoted references
to objects). After Eiffel had been used for a while, Meyer introduced expanded types. An instance
of an expanded type is an object that has a name (value semantics). Expanded types introduce
various irregularities into the language, because they do not have the same properties as ordinary
types. Why do you think expanded types were introduced? Was the decision to introduce them
wise? 2
6.6 Java
Java (Arnold and Gosling 1998) is an OOPL introduced by Sun Microsystems. Its syntax bears
some relationship to that of C++, but Java is simpler in many ways than C++. Key features of
Java include the following.
6 THE OBJECT ORIENTED PARADIGM 88
Listing 25: Exceptions in Eiffel
tryIt is
local
triedFirst: Boolean -- initially false
do
if not triedFirst
then firstmethod
else secondmethod
end
rescue
if not triedFirst
then
triedFirst := true
retry
else
restoreInvariant
-- report failure
end
end
Java is compiled to byte codes that are interpreted. Since any computer that has a Java byte
code interpreter can execute Java programs, Java is highly portable.
The portability of Java is exploited in network programming: Java bytes can be transmitted
across a network and executed by any processor with an interpreter.
Java offers security. The byte codes are checked by the interpreter and have limited function-
ality. Consequently, Java byte codes do not have the potential to penetrate system security
in the way that a binary executable (or even a MS-Word macro) can.
Java has a class hierarchy with class Object at the root and provides single inheritance of
classes.
In addition to classes, Java provides interfaces with multiple inheritance.
Java has an exception handling mechanism.
Java provides concurrency in the form of threads.
Primitive values, such as int, are not objects in Java. However, Java provides wrapper
classes, such as Integer, for each primitive type.
A variable name in Java is a reference to an object.
Java provides garbage collection.
6.6.1 Portability
Compiling to byte codes is an implementation mechanism and, as such, is not strongly relevant to
this course. The technique is quite old — Pascal P achieved portability in the same way during
the mid-seventies — and has the disadvantage of inefficiency: Java programs are typically an order
of magnitude slower than C++ programs with the same functionality. For some applications, such
as simple Web “applets”, the inefficiency is not important. More efficient implementations are
promised. One interesting technique is “just in time” compilation, in which program statements
are compiled immediately before execution; parts of the program that are not needed during a
particular run (this may be a significant fraction of a complex program) are not compiled.
6 THE OBJECT ORIENTED PARADIGM 89
Listing 26: Interfaces
interface Comparable
¦
Boolean equalTo (Comparable other);
// Return True if ‘this’ and ‘other’ objects are equal.
¦
interface Ordered extends Comparable
¦
Boolean lessThan (Ordered other);
// Return True if ‘this’ object precedes ‘other’ object in the ordering.
¦
The portability of Java has been a significant factor in the rapid spread of its popularity, particularly
for Web programming.
6.6.2 Interfaces
A Java class, as usual in OOP, is a compile-time entity whose run-time instances are objects. An
interface declaration is similar to a class declaration, but introduces only abstract methods and
constants. Thus an interface has no instances.
A class may implement an interface by providing definitions for each of the methods declared in
the interface. (The class may also make use of the values of the constants declared in the interface.)
A Java class can inherit from at most one parent class but it may inherit from several interfaces
provided, of course, that it implements all of them. Consequently, the class hierarchy is a rooted
tree but the interface hierarchy is a directed acyclic graph. Interfaces provide a way of describing
and factoring the behaviour of classes.
Listing 26 shows a couple of simple examples of interfaces. The first interface ensures that instances
of a class can be compared. (In fact, all Java classes inherit the function equals from class Object
that can be redefined for comparison in a particular class.) If we want to store instances of a class
in an ordered set, we need a way of sorting them as well: this requirement is specified by the second
interface.
Class Widget is required to implement equalTO and lessThan:
class Widget implements Ordered
¦
. . . .
¦
6.6.3 Exception Handling
The principal components of exception handling in Java are as follows:
There is a class hierarchy rooted at class Exception. Instances of a subclass of class Exception
are used to convey information about what has gone wrong. We will call such instances
“exceptions”.
The throw statement is used to signal the fact that an exception has occurred. The argument
of throw is an exception.
6 THE OBJECT ORIENTED PARADIGM 90
Listing 27: Declaring a Java exception class
public class DivideByZero extends Exception
¦
DivideByZero()
¦
super("Division by zero");
¦
¦
Listing 28: Another Java exception class
public class NegativeSquareRoot extends Exception
¦
public float value;
NegativeSquareRoot(float badValue)
¦
super("Square root with negative argument");
value = badValue;
¦
¦
The try statement is used to attempt an operation and handle any exceptions that it raises.
A try statement may use one or more catch statements to handle exceptions of various types.
The following simple example indicates the way in which exceptions might be used. Listing 27
shows the declaration of an exception for division by zero. The next step is to declare another
exception for negative square roots, as shown in Listing 28
Next, we suppose that either of these problems may occur in a particular computation, as shown
in Listing 29. Finally, we write some code that tries to perform the computation, as shown in
Listing 30.
6.6.4 Concurrency
Java provides concurrency in the form of threads. Threads are defined by a class in that standard
library. Java threads have been criticized on (at least) two grounds.
Listing 29: Throwing exceptions
public void bigCalculation ()
throws DivideByZero, NegativeSquareRoot
¦
. . . .
if (. . . .) throw new DivideByZero();
. . . .
if (. . . .) throw new NegativeSquareRoot(x);
. . . .
¦
6 THE OBJECT ORIENTED PARADIGM 91
Listing 30: Using the exceptions
¦
try
¦
bigCalculation()
¦
catch (DivideByZero)
¦
System.out.println("Oops! divided by zero.");
¦
catch (NegativeSquareRoot n)
¦
System.out.println("Argument for square root was " + n);
¦
¦
¦
The synchronization methods provided by Java (based on the early concepts of semaphores
and monitors) are out of date and do not benefit from more recent research in this area.
The concurrency mechanism is insecure. Brinch Hansen (1999) argues that Java’s concurrency
mechanisms are unsafe and that such mechanisms must be defined in the base language and
checked by the compiler, not provided separately in a library.
Exercise 29. Byte codes provide portability. Can you suggest any other advantages of using byte
codes? 2
6.7 Kevo
All of the OOPLs previously described in this section are class-based languages. In a class-based
language, the programmer defines one or more classes and, during execution, instances of these
classes are created and become the objects of the system. (In Smalltalk, classes are run-time
objects, but Smalltalk is nevertheless a class-based language.) There is another, smaller, family of
OOPLs that use prototypes rather than classes.
Although there are several prototype-based OOPLs, we discuss only one of them here. Antero
Taivalsaari (1993) has given a thorough account of the motivation and design of his language,
Kevo. Taivalsaari (1993, page 172) points out that class-based OOPLs, starting with Simula, are
based on an Aristotelian view in which
the world is seen as a collection of objects with well-defined properties arranged [ac-
cording] to a hierarchical taxonomy of concepts.
The problem with this approach is that there are many classes which people understand intuitively
but which are not easily defined in taxonomic terms. Common examples include “book”, “chair”,
“game”, etc.
Prototypes provide an alternative to classes. In a prototype-based OOPL, each new type of object
is introduced by defining a prototype or exemplar, which is itself an object. Identical copies
of a prototype — as many as needed — are created by cloning. Alternatively, we can clone a
6 THE OBJECT ORIENTED PARADIGM 92
Listing 31: Object Window
REF Window
Object.new → Window
Window ADDS
VAR rect
Prototypes.Rectangle.new → rect
VAR contents
METHOD drawFrame . . . .;
METHOD drawContents . . . .;
METHOD refresh
self.drawFrame
self.drawContents;
ENDADDS;
Listing 32: Object Title Window
REF TitleWindow
Window.new → TitleWindow
TitleWindow ADDS
VAR title
"Anonymous" → title
METHOD drawFrame . . . .;
ENDADDS;
prototype, modify it in some way, and use the resulting object as a prototype for another collection
of identical objects.
The modified object need not be self-contained. Some of its methods will probably be the same
as those of the original object, and these methods can be delegated to the original object rather
than replicated in the new object. Delegation thus plays the role of inheritance in a prototype-
based language and, for this reason, prototype-based languages are sometimes called delegation
languages. This is misleading, however, because delegation is not the only possible mechanism for
inheritance with prototypes.
Example 16: Objects in Kevo. This example illustrates how new objects are introduced in
Kevo; it is based on (Taivalsaari 1993, pages 187–188). Listing 31 shows the introduction of a
window object. The first line, REF Window, declares a new object with no properties. The second
line defines Window as a clone of Object. In Kevo terminology, Window receives the basic printing
and creation operations defined for Object. In the next block of code, properties are added to the
window using the module operation ADDS.
The window has two components: rect and contents. The properties of rect are obtained
from the prototype Rectangle. The properties of contents are left unspecified in this example.
Finally, three new methods are added to Window. The first two are left undefined here, but the
third, refresh, draws the window frame first and then the contents.
Listing 32 shows how to add a title to a window. It declares a new object, TitleWindow with
Window as a prototype. Then it adds a new variable, title, with initial value "Anonymous", and
provides a new method drawFrame that overrides Window.drawFrame and draws a frame with a
title. 2
6 THE OBJECT ORIENTED PARADIGM 93
6.8 Other OOPLs
The OOPLs listed below are interesting, but lack of time precludes further discussion of them.
Beta. Beta is a descendant of Simula, developed in Denmark. It differs in many respects from
“mainstream” OOPLs.
Blue. Blue (K¨olling 1999) is a language and environment designed for teaching OOP. The language
was influenced to some extent by Dee (Grogono 1991b). The environment is interesting in
that users work directly with “objects” rather than program text and generated output.
CLOS. The Common LISP Object System is an extension to LISP that provides OOP. It has a
number of interesting features, including “multi-methods” — the choice of a function is based
on the class of all of its arguments rather than its first (possibly implicit) argument. CLOS
also provides “before” and “after” methods — code defined in a superclass that is executed
before or after the invocation of the function in a class.
Self. Self is the perhaps the best-known prototype language. Considerable effort has been put into
efficient implementation, to the extent that Self programs run at up to half of the speed of
equivalent C++ programs. This is an impressive achievement, given that Self is considerably
more flexible than C++.
6.9 Issues in OOP
OOP is the programming paradigm that currently dominates industrial software development. In
this section, we discuss various issues that arise in OOP.
6.9.1 Algorithms + Data Structures = Programs
The distinction between code and data was made at an early stage. It is visible in FORTRAN and
reinforced in Algol 60. Algol provided the recursive block structure for code but almost no facilities
for data structuring. Algol 68 and Pascal provided recursive data structures but maintained the
separation between code and data.
LISP brought algorithms (in the form of functions) and data structures together, but:
the mutability of data was downplayed;
subsequent FPLs reverted to the Algol model, separating code and data.
Simula started the move back towards the von Neumann Architecture at a higher level of abstrac-
tion than machine code: the computer was recursively divided into smaller computers, each with its
own code and stack. This was the feature of Simula that attracted Alan Kay and led to Smalltalk
(Kay 1996).
Many OOPLs since Smalltalk have withdrawn from its radical ideas. C++ was originally called
“C with classes” (Stroustrup 1994) and modern C++ remains an Algol-like systems programming
language that supports OOP. Ada has shown even more reluctance to move in the direction of
OOP, although Ada 9X has a number of OO features.
6 THE OBJECT ORIENTED PARADIGM 94
6.9.2 Values and Objects
Some data behave like mathematical objects and other data behave like representations of physical
objects (MacLennan 1983).
Consider the sets ¦ 1, 2, 3 ¦ and ¦ 2, 3, 1 ¦. Although these sets have different representations (the
ordering of the members is different), they have the same value because in set theory the only
important property of a set is the members that it contains: “two sets are equal if and only if they
contain the same members”.
Consider two objects representing people:
[name="Peter", age=55]
[name="Peter", age=55]
Although these objects have the same value, they might refer to different people who have the same
name and age. The consequences of assuming that these two objects denote the same person could
be serious: in fact, innocent people have been arrested and jailed because they had the misfortune
to have the same name and social security number as a criminal.
In the following comparison of values and objects, statements about values on the left side of the
page are balanced by statements about objects on the right side of the page.
Values
Objects
Integers have values. For example, the integer 6 is a value.
Objects possess values as attributes. For example, a counter might
have 6 as its current value.
Values are abstract. The value 7 is the common property shared by
all sets with cardinality 7.
Objects are concrete. An object is a particular collection of sub-
objects and values. A counter has one attribute: the value of the
count.
Values are immutable (i.e., do not change).
Objects are usually, but not necessarily, mutable (i.e., may change).
It does not make sense to change a value. If we change the value 3,
it is no longer 3.
It is meaningful, and often useful, to change the value of an object.
For example, we can increment a counter.
Values need representations. We need representations in order to
compute with values. The representations 7, VII, sept, sju, and
a particular combination of bits in the memory of a computer are
different representations of the same abstract value.
Objects need representations. An object has subobjects and values
as attributes.
It does not make sense to talk about the identity of a value. Suppose
we evaluate 2 +3, obtaining 5, and then 7 −2, again obtaining 5. Is
it the same 5 that we obtain from the second calculation?
Identity is an important attribute of objects. Even objects with
the same attributes may have their own, unique identities: consider
identical twins.
We can copy a representation of a value, but copying a value is
meaningless.
6 THE OBJECT ORIENTED PARADIGM 95
Copying an object certainly makes sense. Sometimes, it may even
be useful, as when we copy objects from memory to disk, for exam-
ple. However, multiple copies of an object with a supposedly unique
identity can cause problems, known as “integrity violations” in the
database community.
We consider different representations of the same value to be equal.
An object is equal to itself. Whether two distinct objects with equal
attributes are equal depends on the application. A person who is
arrested because his name and social insurance number is the same
as those of a wanted criminal may not feel “equal to” the criminal.
Aliasing is not a problem for values. It is immaterial to the pro-
grammer, and should not even be detectable, whether two instances
of a value are stored in the same place or in different places.
All references to a particular object should refer to the same object.
This is sometimes called “aliasing”, and is viewed as harmful, but it
is in fact desirable for objects with identity.
Functional and logical PLs usually provide values.
Object oriented PLs should provide objects.
Programs that use values only have the property of referential trans-
parency.
Programs that provide objects do not provide referential trans-
parency, but may have other desirable properties such as encap-
sulation.
Small values can be held in machine words. Large values are best
represented by references (pointers to the data structure representing
the value.
Very small objects may be held in machine words. Most objects
should be represented by references (pointers to the data structure
holding the object.
The assignment operator for values should be a reference assignment.
There is no need to copy values, because they do not change.
The assignment operator for objects should be a reference assign-
ment. Objects should not normally be copied, because copying en-
dangers their integrity.
The distinction between values and objects is not as clear as these comparisons suggest. When we
look at programming problems in more deatil, doubts begin to arise.
Exercise 30. Should sets be values? What about sets with mutable components? Does the model
handle objects with many to one abstraction functions? What should a PL do about values and
objects? 2
Conclusions:
The CM must provide values, may provide objects.
What is “small” and what is “large”? This is an implementation issue: the compiler must
decide.
What is “primitive” and what is “user defined”? The PL should minimize the distinction as
much as possible.
Strings should behave like values. Since they have unbounded size, many PLs treat them as
objects.
6 THE OBJECT ORIENTED PARADIGM 96
Copying and comparing are semantic issues. Compilers can provide default operations and
“building bricks”, but they cannot provide fully appropriate implementations.
6.9.3 Classes versus Prototypes
The most widely-used OOPLs — C++, Smalltalk, and Eiffel — are class-based. Classes provide a
visible program structure that seems to be reassuring to designers of complex software systems.
Prototype languages are a distinct but active minority. The relation between classes and prototypes
is somewhat analogous to the relation between procedural (FORTRAN and Algol) and functional
(LISP) languages in the 60s.
Prototypes provide more flexibility and are better suited to interactive environments than classes.
Although large systems have been built using prototype languages, these languages seem to be
more often used for exploratory programming.
6.9.4 Types versus Objects
In most class-based PLs, functions are associated with objects. This has the effect of making binary
operations asymmetric. For example, x = y is interpreted as x.equal(y) or, approximately,
“send the message equal with argument y to the object x”. It is not clear whether, with this
interpretation, y = x means the same as x = y. The problem is complicated further by inheritance:
if class C is a subclass of class P, can we compare an instance of C to an instance of P, and is this
comparison commutative?
Some of the problems can be avoided by associating functions with types rather than with objects.
This is the approach taken by CLU and some of the “functional” OOPLs. An expression such as
x = y is interpreted in the usual mathematical sense as equal(x,y). However, even this approach
has problems if x and y do not have the same type.
6.9.5 Pure versus Hybrid Languages
Smalltalk, Eiffel, and Java are pure OOPLs: they can describe only systems of objects and pro-
grammers are forced to express everything in terms of objects.
C++ is a hybrid PL in that it supports both procedural and object styles. Once the decision was
made to extend C rather than starting afresh (Stroustrup 1994, page 43), it was inevitable that
C++ could be a hybrid language. In fact, its precursor, “C with classes”, did not even claim to be
object oriented.
A hybrid language can be dangerous for inexperienced programmers because it encourages a hybrid
programming style. OO design is performed perfunctorily and design problems are solved by
reverting to the more general, but lower level, procedural style. (This is analogous to the early
days of “high level” programming when programmers would revert to assembler language to express
constructs that were not easy or possible to express at a high level.)
For experienced programmers, and especially those trained in OO techniques, a hybrid language
is perhaps better than a pure OOPL. These programmers will use OO techniques most of the
time and will only resort to procedural techniques when these are clearly superior to the best OO
solution. As OOPLs evolve, these occasions should become rarer.
Perhaps “pure or hybrid” is the wrong question. A better question might be: suppose a PL
provides OOP capabilities. Does it need to provide anything else? If so, what? One answer is: FP
6 THE OBJECT ORIENTED PARADIGM 97
capabilities. The multi-paradigm language LEDA (Budd 1995) demonstrates that it is possible to
provide functional, logic, and object programming in a single language. However, LEDA provides
these in a rather ad hoc way. It more tightly-integrated design would be preferable, but is harder
to accomplish.
6.9.6 Closures versus Classes
The most powerful programming paradigms developed so far are those that do not separate code
and data but instead provide “code and data” units of arbitrary size. This can be accomplished
by high order functions and mutable state, as in Scheme (Section 4.2) or by classes. Although
each method has advantages for certain applications, the class approach seems to be better in most
practical situations.
6.9.7 Inheritance
Inheritance is one of the key features of OOP and at the same time one of the most troublesome.
Conceptually, inheritance can be defined in a number of ways and has a number of different uses.
In practice, most (but not all!) OOPLs provide a single mechanism for inheritance and leave it to
the programmer to decide how to use the mechanism. Moreover, at the implementation level,
the varieties of inheritance all look much the same — a fact that discourages PL designers from
providing multiple mechanisms.
One of the clearest examples of the division is Meyer’s (1988) “marriage of convenience” in which
the class ArrayStack inherits from Array and Stack. The class ArrayStack behaves like a stack
— it has functions such as push, pop, and so on — and is represented as an array. There are other
ways of obtaining this effect — notably, the class ArrayStack could have an array object as an
attribute — but Meyer maintains that inheritance is the most appropriate technique, at least in
Eiffel.
Whether or not we agree with Meyer, it is clear that there two different relations are involved.
The relation ArrayStack ∼ Stack is not the same as the relation ArrayStack ∼ Array. Hence the
question: should an OOPL distinguish between these relations or not?
Based on the “marriage of convenience” the answer would appear to be “yes” — two distinct ends
are achieved and the means by which they are achieved should be separated. A closer examination,
however, shows that the answer is by no means so clear-cut. There are many practical programming
situations in which code in a parent class can be reused in a child class even when interface
consistency is the primary purpose of inheritance.
6.10 A Critique of C++
C++ is both the most widely-used OOPL and the most heavily criticized. Further criticism might
appear to be unnecessary. The purpose of this section is to show how C++ fails to fulfill its role
as the preferred OOPL.
The first part of the following discussion is based on a paper by LaLonde and Pugh (1995) in which
they discuss C++ from the point of view of a programmer familiar with the principles of OOP,
exemplified by a language such as Smalltalk, but who is not familiar with the details of C++.
We begin by writing a specification for a simple BankAccount class with an owner name, a balance,
and functions for making deposits and withdrawals: see Listing 33.
6 THE OBJECT ORIENTED PARADIGM 98
Data members are prefixed with “p” to distinguish them from parameters with similar names.
The instance variable pTrans is a pointer to an array containing the transactions that have
taken place for this account. Functions that operate on this array have been omitted.
We can create instances of BankAccount on either the stack or the heap:
BankAccount account1("John"); // on the stack
BankAccount *account2 = new BankAccount("Jane"); // on the heap
The first account, account1, will be deleted automatically at the end of the scope containing the
declaration. The second account, account2, will be deleted only when the program executes a
delete operation. In both cases, the destructor will be invoked.
The notation for accessing functions depends on whether the variable refers to the object directly
or is a pointer to the object.
account1.balance()
account2→balance()
A member function can access the private parts of another instance of the same class, as shown in
Listing 34. This comes as a surprise to Smalltalk programmers because, in Smalltalk, an object
can access only its own components.
The next step is to define the member functions of the class BankAccount: see Listing 35.
Listing 36 shows a simple test of the class BankAccount. This test introduces two bugs.
The array account1.pTrans is lost because it is over-written by account2.pTrans.
At the end of the test, account2.pTrans will be deleted twice, once when the destructor for
account1 is called, and again when the destructor for account2 is called.
Although most introductions to C++ focus on stack objects, it is more interesting to consider heap
objects: Listing 37.
The code in Listing 37 — creating three special accounts — is rather specialized. Listing 38
generalizes it so that we can enter the account owner’s names interactively.
Listing 33: Declaring a Bank Account
class BankAccount
¦
public:
BankAccount(char * newOwner = "");
∼BankAcocunt();
void owner(char *newOwner);
char *owner();
double balance();
void credit(double amount);
void debit(double amount);
void transfer(BankAccount *other, double amount);
private:
char *pOwner;
double pBalance;
Transaction *pTrans[MAXTRANS];
¦;
6 THE OBJECT ORIENTED PARADIGM 99
Listing 34: Accessing Private Parts
void BankAccount::transfer(BankAccount *other, double amount)
¦
pBalance += amount;
other→pBalance -= amount;
¦
Listing 35: Member Functions for the Bank Account
BankAccount::BankAccount(char *newOwner)
¦
pOwner = newOwner;
pBalance = 0.0;
¦
BankAccount::∼BankAccount()
¦
// Delete individual transactions
delete [] pTrans; // Delete the array of transactions
¦
void BankAccount owner(char *newOwner)
¦
pOwner = newOwner;
¦
char *BankAccount owner()
¦
return pOwner;
¦
. . . .
Listing 36: Testing the Bank Account
void test
¦
BankAccount account1("Anne");
BankAccount account2("Bill");
account1 = acount2;
¦
Listing 37: Using the Heap
BankAccount *accounts[] = new BankAccount * [3];
accounts[0] = new BankAccount("Anne");
accounts[1] = new BankAccount("Bill");
accounts[2] = new BankAccount("Chris");
// Account operations omitted.
// Delete everything.
for (int i = 0; i < 3; i++)
delete accounts[i];
delete accounts;
6 THE OBJECT ORIENTED PARADIGM 100
Listing 38: Entering a Name
BankAccount *accounts[] = new BankAccount * [3];
char * buffer = new char[100];
for (int i = 0; i < 3; i++)
¦
cout << "Please enter your name: ";
cin.getline(buffer, 100);
accounts[i] = new BankAccount(buffer);
¦
// Account operations omitted.
// Delete everything.
for (int i = 0; i < 3; i++)
delete accounts[i];
delete accounts;
delete buffer;
Listing 39: Adding buffers for names
// char * buffer = new char[100]; (deleted)
for (int i = 0; i < 3; i++)
¦
cout << "Please enter your name: ";
char * buffer = new char[100];
cin.getline(buffer, 100);
accounts[i] = new BankAccount(buffer);
¦
// Account operations omitted.
// Delete everything.
for (int i = 0; i < 3; i++)
¦
delete accounts[i]→owner();
delete accounts[i];
¦
delete accounts;
delete buffer;
The problem with this code is that all of the bank accounts have the same owner: that of the last
name entered. This is because each account contains a pointer to the unique buffer. We can fix
this by allocating a new buffer for each account, as in Listing 39.
This version works, but is poorly designed. Deleting the owner string should be the responsibility
of the object, not the client. The destructor for BankAccount should delete the owner string. After
making this change, we can remove the destructor for owner in the code above: Listing 40.
Unfortunately, this change introduces another problem. Suppose we use the loop to create two
accounts and then create the third account with this statement:
accounts[2] = new BankAccount("Tina");
The account destructor fails because "Tina" was not allocated using new. To correct this problem,
we must modify the constructor for BankAccount. The rule we must follow is that data must be
allocated with the constructor and deallocated with the destructor, or controlled entirely outside
6 THE OBJECT ORIENTED PARADIGM 101
Listing 40: Correcting the Destructor
BankAccount::∼BankAccount()
¦
delete pOwner;
// Delete individual transactions
delete [] pTrans; // Delete the array of transactions
¦
Listing 41: Correcting the Constructor
BankAccount::BankAccount(char *newOwner)
¦
char *pOwner = new char (strlen(newOwner) + 1);
strcpy(pOwner, newOwner);
¦
the object. Listing 41 shows the new constructor.
The next problem occurs when we try to use the overloaded function owner() to change the name
of an account owner.
accounts[2]→owner("Fred");
Once again, we have introduced two bugs:
The original owner of accounts[2] will be lost.
When the destructor is invoked for accounts[2], it will attempt to delete "Fred", which was
not allocated by new. To correct this problem, we must rewrite the member function owner(),
as shown in Listing 42.
We now have bank accounts that work quite well, provided that they are allocated on the heap. If
we try to mix heap objects and stack objects, however, we again encounter difficulties: Listing 43.
The error occurs because transfer expects the address of an account rather than an actual account.
We correct it by introducing the reference operator, &.
account1→transfer(&account3, 10.0);
We could make things easier for the programmer who is using accounts by declaring an overloaded
version of transfer() as shown below. Note that, in the version of Listing 44, we must use the
operator “.” rather than “→”. This is because other behaves like an object, although in fact an
address has been passed.
The underlying problem here is that references (&) work best with stack objects and pointers (*)
work best with heap objects. It is awkward to mix stack and heap allocation in C++ because two
Listing 42: Correcting the Owner Function
void BankAccount owner(char *newOwner)
¦
delete pOwner;
pOwner = new char(strlen(newOwner) + 1);
strcpy(pOwner, newOwner);
¦
6 THE OBJECT ORIENTED PARADIGM 102
Listing 43: Stack and Heap Allocation
BankAccount *account1 = new BankAccount("Anne");
BankAccount *account2 = new BankAccount("Bill");
BankAccount account3("Chris");
account1→transfer(account2, 10.0); // OK
account1→transfer(account3, 10.0); // Error!
Listing 44: Revising the Transfer Function
void BankAccount::transfer(BankAccount &other, double amount)
¦
pBalance += amount;
other.pBalance -= amount;
¦
sets of functions are needed.
If stack objects are used without references, the compiler is forced to make copies. To implement
the code shown in Listing 45, C++ uses the (default) copy constructor for BankAccount twice:
first to make a copy of account2 to pass to test() and, second, to make another copy of account2
that will become myAccount. At the end of the scope, these two copies must be deleted.
The destructors will fail because the default copy constructor makes a shallow copy — it does not
copy the pTrans fields of the accounts. The solution is to write our own copy constructor and
assignment overload for class BankAccount, as in Listing 46. The function pDuplicate() makes
a deep copy of an account the function; it must make copies of all of the components of a back
account object and destroy all components of the old object. The function pDelete has the same
effect as the destructor. It is recommended practice in C++ to write these functions for any class
that contains pointers.
LaLonde and Pugh continue with a discussion of infix operators for user-defined classes. If the
objects are small — complex numbers, for example — it is reasonable to store the objects on
the stack and to use call by value. C++ is well-suited to this approach: storage management is
automatic and the only overhead is the copying of values during function call and return.
If the objects are large, however, there are problems. Suppose that we want to provide the standard
algebraic operations (+, −, ∗) for 4 4 matrices. (This would be useful for graphics programming,
for example.) Since a matrix occupies 448 = 128 bytes, copying a matrix is relatively expensive
and the stack approach will be inefficient. But if we write functions that work with pointers to
matrices, it is difficult to manage allocation and deallocation. Even a simple function call such as
project(A + B);
will create a heap object corresponding to A + B but provides no opportunity to delete this object.
Listing 45: Copying
BankAccount test(BankAccount account)
¦
return account;
¦
BankAccount myAccount;
myAccount = test(account2);
6 THE OBJECT ORIENTED PARADIGM 103
Listing 46: Copy Constructors
BankAccount::BankAccount(BankAccount &account)
¦
pDuplicate(account);
¦
BankAccount & BankAccount operator = (BankAccount &rhs)
¦
if (this == &rhs)
return this;
pDelete();
pDuplicate(rhs);
return *this;
¦
Listing 47: Using extended assignment operators
A = C;
A *= D;
A += B;
(The function cannot delete it, because it cannot tell whether its argument is a temporary.) The
same problem arises with expressions such as
A = B + C * D;
It is possible to implement the extended assignment operators, such as +=, without losing objects.
But this approach reduces us to a kind of assembly language style of programming, as shown in
Listing 47.
These observations by LaLonde and Pugh seem rather hostile towards C++. They are making a
comparison between C++ and Smalltalk based only on ease of programming. The comparison is
one-sided because they do not discuss factors that might make C++ a better choice than Smalltalk
in some circumstances, such as when efficiency is important. Nevertheless, the discussion makes an
important point: the “stack model” (with value semantics) is inferior to the “heap model” (with
reference semantics) for many OOP applications. It is significant that most language designers
have chosen the heap model for OOP (Simula, Smalltalk, CLU, Eiffel, Sather, Dee, Java, Blue,
amongst others) and that C++ is in the minority on this issue.
Another basis for criticizing C++ is that is has become, perhaps not intentionally, a hacker’s
language. A sign of good language design is that the features of the language are used most of
the time to solve the problems that they were intended to solve. A sign of a hacker’s language is
that language “gurus” introduce “hacks” that solve problems by using features in bizarre ways for
which they were not originally intended.
Input and output in C++ is based on “streams”. A typical output statement looks something like
this:
cout << "The answer is " << answer << ’.’ << endl;
The operator << requires a stream reference as its left operand and an object as its right operand.
Since it returns a stream reference (usually its left operand), it can be used as an infix operator
in compound expressions such as the one shown. The notation seems clear and concise and, at
first glance, appears to be an elegant way of introducing output and input (with the operator >>)
capabilities.
6 THE OBJECT ORIENTED PARADIGM 104
Listing 48: A Manipulator Class
class ManipulatorForInts
¦
friend ostream & operator << (ostream & const ManipulatorForInts &);
public:
ManipulatorForInts (int (ios::*)(int), int);
. . . .
private:
int (ios::*memberfunc)(int);
int value;
¦
Obviously, << is heavily overloaded. There must be a separate function for each class for which
instances can be used as the right operand. (<< is also overloaded with respect to its left operand,
because there are different kinds of output streams.) In the example above, the first three right
operands are a string, a double, and a character.
The first problem with streams is that, although they are superficially object oriented (a stream
is an object), they do not work well with inheritance. Overloading provides static polymorphism
which is incompatible with dynamic polymorphism (recall Section 9.6). It is possible to write
code that uses dynamic binding but only by providing auxiliary functions.
The final operand, endl, writes end-of-line to the stream and flushes it. It is a global function with
signature
ios & endl (ios & s)
A programmer who needs additional “manipulators” of this kind has no alternative but to add
them to the global name space.
Things get more complicated when formatting is added to streams. For example, we could change
the example above to:
cout << "The answer is " << setprecision(8) << answer << ’.’ << endl;
The first problem with this is that it is verbose. (The verbosity is not really evident in such a small
example, but it is very obvious when streams are used, for example, to format a table of heteroge-
neous data.) The second, and more serious problem, is the implementation of setprecision. The
way this is done is (roughly) as follows (Teale 1993, page 171). First, a class for manipulators with
integer arguments is declared, as in Listing 48, and the inserter function is defined:
ostream & operator << (ostream & os, const ManipulatorForInts & m)
¦
(os.*memberfunc)(m.value);
return os;
¦
Second, the function that we need can be defined:
ManipulatorForInts setprecision (int n)
¦
return ManipulatorForInts (& ios::precision, n);
¦
It is apparent that a programmer must be quite proficient in C++ to understand these definitions,
let alone design similar functions. To make matters worse, the actual definitions are not as shown
6 THE OBJECT ORIENTED PARADIGM 105
Listing 49: File handles as Booleans
ifstream infile ("stats.dat");
if (infile) . . . .
if (!infile) . . . .
here, because that would require class declarations for many different types, but are implemented
using templates.
There are many other tricks in the implementation of streams. For example, we can use file handles
as if they had boolean values, as in Listing 49. It seems that there is a boolean value, infile, that
we can negate, obtaining !infile. In fact, infile can be used as a conditional because there is
a function that converts infile to (void *), with the value NULL indicating that the file is not
open, and !infile can be used as a conditional because the operator ! is overloaded for class
ifstream to return an int which is 0 if the file is open.
These are symptoms of an even deeper malaise in C++ than the problems pointed out by authors of
otherwise excellent critiques such as (Joyner 1992; Sakkinen 1992). Most experienced programmers
can get used to the well-known idiosyncrasies of C++, but few realize that frequently-used and
apparently harmless constructs are implemented by subtle and devious tricks. Programs that look
simple may actually conceal subtle traps.
6.11 Evaluation of OOP
OOP provides a way of building programs by incremental modification.
Programs can often be extended by adding new code rather than altering existing code.
The mechanism for incremental modification without altering existing code is inheritance.
However, there are several different ways in which inheritance can be defined and used, and
some of them conflict.
The same identifier can be used in various contexts to name related functions. For example,
all objects can respond to the message “print yourself”.
Class-based OOPLs can be designed so as to support information hiding (encapsulation) —
the separation of interface and implementation. However, there may be conflicts between
inheritance and encapsulation.
An interesting, but rarely noted, feature of OOP is that the call graph is created and modified
during execution. This is in contrast to procedural languages, in which the call graph can be
statically inferred from the program text.
Section 6.9 contains a more detailed review of issues in OOP.
Exercise 31. Simula is 33 years old. Can you account for the slow acceptance of OOPLs by
professional programmers? 2
7 Backtracking Languages
The PLs that we discuss in this section differ in significant ways, but they have an important
feature in common: they are designed to solve problems with multiple answers, and to find as
many answers as required.
In mathematics, a many-valued expression can be represented as a relation or as a function re-
turning a set. Consider the relation “divides”, denoted by [ in number theory. Any integer greater
than 1 has at least two divisors (itself and 1) and possible others. For example, 1 [ 6, 2 [ 6, 3 [ 6,
and 6 [ 6. The functional analog of the divides relation is the set-returning function divisors:
divisors(6) = ¦ 1, 2, 3, 6 ¦
divisors(7) = ¦ 1, 7 ¦
divisors(24) = ¦ 1, 2, 3, 4, 6, 8, 12, 24 ¦
We could define functions that return sets in a FPL or we could design a language that works di-
rectly with relations. The approach that has been most successful, however, is to produce solutions
one at a time. The program runs until it finds the first solution, reports that solution or uses it
in some way, then backtracks to find another solution, reports that, and so on until all the choices
have been explored or the user has had enough.
7.1 Prolog
The computational model (see Section 10.2) of a logic PL is some form of mathematical logic.
Propositional calculus is too weak because it does not have variables. Predicate calculus is too
strong because it is undecidable. (This means that there is no effective procedure that determines
whether a given predicate is true.) Practical logic PLs are based on a restricted form of predicate
calculus that has a decision procedure.
The first logic PL was Prolog, introduced by Colmerauer (1973) and Kowalski (1974). The discus-
sion in this section is based on (Bratko 1990).
We can express facts in Prolog as predicates with constant arguments. All Prolog textbooks start
with facts about someone’s family: see Listing 50. We read parent(pam, bob) as “Pam is the
parent of Bob”. This Bob’s parents are Pam and Tom; similarly for the other facts. Note that,
since constants in Prolog start with a lower case letter, we cannot write Pam, Bob, etc.
Prolog is interactive; it prompts with ?- to indicate that it is ready to accept a query. Prolog
responds to simple queries about facts that it has been told with “yes” and “no”, as in Listing 51.
If the query contains a logical variable — an identifier that starts with an upper case letter —
Prolog attempts to find a value of the variable that makes the query true.
If it succeeds, it replies with the binding.
Listing 50: A Family
parent(pam, bob).
parent(tom, bob).
parent(tom, liz).
parent(bob, ann).
parent(bob, pat).
parent(pat, jim).
106
7 BACKTRACKING LANGUAGES 107
Listing 51: Prolog queries
?- parent(tom, liz).
yes
?- parent(tom, jim).
no
?- parent(pam, X).
X = bob
If there is more than one binding that satisfies the query, Prolog will find them all. (The exact
mechanism depends on the implementation. Some implementations require the user to ask for more
results. Here we assume that the implementation returns all values without further prompting.)
?- parent(bob, C).
C = ann
C = pat
Consider the problem of finding Jim’s grandparents. In logic, we can express “G is a grandparent
of Jim” in the following form: “there is a person P such that P is a parent of Jim and there is a
person G such that G is a parent of P”. In Prolog, this becomes:
?- parent(P, jim), parent(G, P).
P = pat
G = bob
The comma in this query corresponds to conjunction (“and”): the responses to the query are
bindings that make all of the terms valid simultaneously. Disjunction (“or”) is provided by multiple
facts or rules: in the example above, Bob is the parent of C is true if C is Ann or C is Pat.
We can expression the “grandparent” relation by writing a rule and adding it to the collection
of facts (which becomes a collection of facts and rules). The rule for “grandparent” is “G is a
grandparent of C if G is a parent of P and P is a parent of C”. In Prolog:
grandparent(G, C) :- parent(G, P), parent(P, C).
Siblings are people who have the same parents. We could write this rule for siblings:
sibling(X, Y) :- parent(P, X), parent(P, Y).
Testing this rule reveals a problem.
?- sibling(pat, X).
X = ann
X = pat
We may not expect the second response, because we do not consider Pat to be a sibling of her-
self. But Prolog simply notes that the predicate parent(P, X), parent(P, Y) is satisfied by the
bindings P = bob, X = pat, and Y = pat. To correct this rule, we would have to require that X
and Y have different values:
sibling(X, Y) :- parent(P, X), parent(P, Y), different(X, Y).
It is clear that different(X, Y) must be built into Prolog somehow, because it would not be
feasible to list every true instantiation of it as a fact.
Prolog allows rules to be recursive: a predicate name that appears on the left side of a rule may
also appear on the right side. For example, a person A is an ancestor of a person X if either A is
a parent of X or A is an ancestor of a parent P of X:
ancestor(A, X) :- parent(A, X).
7 BACKTRACKING LANGUAGES 108
parent(pam, bob)
ancestor(pam, bob) parent(bob, pat)
ancestor(pam, pat) parent(pat, jim)
ancestor(pam, jim)
Figure 15: Proof of ancestor(pam, jim)
ancestor(A, X) :- ancestor(A, P), parent(P, X).
We confirm the definition by testing:
?- ancestor(pam, jim).
yes
We can read a Prolog program in two ways. Read declaratively, ancestor(pam, jim) is a
statement and Prolog’s task is to prove that it is true. The proof is shown in Figure 15, which uses
the convention that
P
Q
means “from P infer Q”. Read procedurally, ancestor(pam, jim) defines
a sequence of fact-searches and rule-applications that establish (or fail to establish) the goal. The
procedural steps for this example are as follows.
1. Using the first rule for ancestor, look in the data base for the fact parent(pam, jim).
2. Since this search fails, try the second rule with A = pam and X = jim. This gives two new
subgoals: ancestor(pam, P) and parent(P, jim).
3. The second of these subgoals is satisfied by P = pat from the database. The goal is now
ancestor(pam, pat).
4. The goal expands to the subgoals ancestor(pam, P) and parent(P, pat). The second of
these subgoals is satisfied by P = bob fromthe database. The goal is now ancestor(pam, bob).
5. The goal ancestor(pam, bob) is satisfied by applying the first rule with the known fact
parent(pam, bob).
In the declarative reading, Prolog finds multiple results simply because they are there. In the
procedure reading, multiple results are obtained by backtracking. Every point in the program at
which there is more than one choice of a variable binding is called a choice point. The choice
points define a tree: the root of the tree is the starting point of the program (the main goal) and
each leaf of the tree represents either success (the goal is satisfied with the chosen bindings) or
failure (the goal cannot be satisfied with these bindings). Prolog performs a depth-first search of
this tree.
Sometimes, we can tell from the program that backtracking will be useless. Declaratively, the
predicate minimum(X,Y,Min) is satisfied if Min is the minimum of X and Y. We define rules for the
cases X ≤ Y and X > Y .
minimum(X, Y, X) :- X ≤ Y.
minimum(X, Y, Y) :- X > Y.
It is easy to see that, if one of these rules succeeds, we do not want to backtrack to the other,
because it is certain to fail. We can tell Prolog not to backtrack by writing “!”, pronounced “cut”.
7 BACKTRACKING LANGUAGES 109
Program Meaning
p :- a,b. p :- c. p ⇐⇒ (a ∧ b) ∨ c
p :- a,!,b. p:- c. p ⇐⇒ (a ∧ b) ∨ (∼a ∧ c)
p :- c. p :- a,!,b. p ⇐⇒ c ∨ (a ∧ b)
Figure 16: Effect of cuts
minimum(X, Y, X) :- X ≤ Y, !.
minimum(X, Y, Y) :- X > Y, !.
The introduction of the cut operation has an interesting consequence. If a Prolog program has no
cuts, we can freely change the order of goals and clauses. These changes may affect the efficiency
of the program, but they do not affect its declarative meaning. If the program has cuts, this is
no longer true. Figure 16 shows an example of the use of cut and its effect on the meaning of a
program.
Recall the predicate different introduced above to distinguish siblings, and consider the following
declarations and queries.
different(X, X) :- !, fail.
different(X, Y).
?- different(Tom, Tom).
no
?- different(Ann, Tom).
yes
The query different(Tom, Tom) matches different(X, X) with the binding X = Tom and suc-
ceeds. The Prolog constant fail then reports failure. Since the cut has been passed, no backtrack-
ing occurs, and the query fails. The query different(Ann, Tom) does not match different(X, X)
but it does match the second clause, different(X, Y) and therefore succeeds.
This idea can be used to define negation in Prolog:
not (P) :- P, !, fail.
true.
However, it is important to understand what negation in Prolog means. Consider Listing 52,
assuming that the entire database is included. The program has stated correctly that crows fly and
penguins do not fly. But it has made this inference only because the database has no information
about penguins. The same program can produce an unlimited number of (apparently) correct
responses, and also some incorrect responses. We account for this by saying that Prolog assumes
a closed world. Prolog can make inferences from the “closed world” consisting of the facts and
rules in its database.
Now consider negation. In Prolog, proving not(P) means “P cannot be proved” which means “P
cannot be inferred from the facts and rules in the database”. Consequently, we should be sceptical
about positive results obtained from Prolog programs that use not. The response
?- not human(peter).
yes
means no more than “the fact that Peter is human cannot be inferred from the database”.
The Prolog system proceeds top-down. It attempts to match the entire goal to one of the rules
and, having found a match, applies the same idea to the parts of the matched formulas. The
“matching” process is called unification.
7 BACKTRACKING LANGUAGES 110
Listing 52: Who can fly?
flies(albatross).
flies(blackbird);
flies(crow).
?- flies(crow).
yes
?- flies(penguin).
no
?- flies(elephant).
no
?- flies(brick).
no
?- flies(goose).
no
Suppose that we have two terms A and B built up from constants and variables. A unifier of A
and B is a substitution that makes the terms identical. A substitution is a binding of variable
names to values. In the following example we follow the Prolog convention, using upper case letters
X, Y, . . .to denote variables and lower case letters a, b, . . .to denote constants. Let A and B be the
terms
A ≡ left(tree(X, tree(a, Y ))
B ≡ left(tree(b, Z))
and let
σ = (X = b, Z = tree(a, Y ))
be a substitution. Then σA = σB = left(tree(b, tree(a, Y )) and σ is a unifier of A and B. Note
that the unifier is not necessarily unique. However, we can define a most general unifier which
is unique.
The unification algorithm contains a test called the occurs check that is used to reject a substi-
tution of the form X = f(X) in which a variable is bound to a formula that contains the variable.
If we changed A and B above to
A ≡ left(tree(X, tree(a, Y ))
B ≡ left(tree(b, Y )
they would not unify.
You can find out whether a particular version of Prolog implements the occurs check by trying the
following query:
?- X = f(X).
If the reply is “no”, the occurs check is implemented. If, as is more likely, the occurs check is not
performed, the response will be
X = f(f(f(f( . . . .
Prolog syntax corresponds to a restricted form of first-order predicate calculus called clausal
form logic. It is not practical to use the full predicate calculus as a basis for a PL because it
is undecidable. Clausal form logic is semi-decidable: there is an algorithm that will find a proof
7 BACKTRACKING LANGUAGES 111
Listing 53: Restaurants
good(les halles).
expensive(les halles).
good(joes diner).
reasonable(Restaurant) : not expensive(Restaurant).
?- good(X), reasonable(X).
X = joes diner.
?- reasonable(X), good(X).
no
for any formula that is true in the logic. If a formula is false, the algorithm may fail after a finite
time or may loop forever. The proof technique, called SLD resolution, was introduced by Robinson
(1965). SLD stands for “Selecting a literal, using a Linear strategy, restricted to Definite clauses”.
The proof of validity of SLD resolution assumes that unification is implemented with the occurs
check. Unfortunately, the occurs check is expensive to implement and most Prolog systems omit
it. Consequently, most Prolog systems are technically unsound, although problems are rare in
practice.
A Prolog program apparently has a straightforward interpretation as a statement in logic, but the
interpretation is slightly misleading. For example, since Prolog works through the rules in the
order in which they are written, the order is significant. Since the logical interpretation of a rule
sequence is disjunction, it follows that disjunction in Prolog does not commute.
Example 17: Failure of commutativity. In the context of the rules
wife(jane).
female(X) :- wife(X).
female(X) :- female(X).
we obtain a valid answer to a query:
?- female(Y).
jane
But if we change the order of the rules, as in
wife(jane).
female(X) :- female(X).
female(X) :- wife(X).
the same query leads to endless applications of the second rule. 2
Exercise 32. Listing 53 shows the results of two queries about restaurants. In logic, conjunction
is commutative (P ∧ Q ≡ Q∧ P). Explain why commutativity fails in this program. 2
The important features of Prolog include:
Prolog is based on a mathematical model (clausal form logic with SLD resolution).
“Pure” Prolog programs, which fully respect the logical model, can be written but are often
inefficient.
In order to write efficient programs, a Prolog programmer must be able to read programs both
declaratively (to check the logic) and procedurally (to ensure efficiency).
7 BACKTRACKING LANGUAGES 112
Prolog implementations introduce various optimizations (the cut, omitting the occurs check)
that improve the performance of programs but compromise the mathematical model.
Prolog has garbage collection.
7.2 Alma-0
One of the themes of this course is the investigation of ways in which different paradigms of
programming can be combined. The PL Alma-0 takes as its starting point a simple imperative
language, Modula–2 (Wirth 1982), and adds a number of features derived from logic programming
to it. The result is a simple but highly expressive language. Some features of Modula–2 are omitted
(the CARDINAL type; sets; variant records; open array parameters; procedure types; pointer types;
the CASE, WITH, LOOP, and EXIT statements; nested procedures; modules). They are replaced by
nine new features, described below. This section is based on Apt et al. (1998).
BES. Boolean expressions may be used as statements. The expression is evaluated. If its value
is TRUE, the computation succeeds and the program continues as usual. If the value of the
expression is FALSE, the computation fails. However, this does not mean that the program
stops: instead, control returns to the nearest preceding choice point and resumes with a
new choice.
Example 18: Ordering 1. If the array a is ordered, the following statement succeeds.
Otherwise, it fails; the loop exits, and i has the smallest value for which the condition is
FALSE.
FOR i := 1 to n - 1 DO a[i] <= a[i+1] END
2
SBE. A statement sequence may be used as a boolean expression. If the computation of a sequence
of statements succeeds, it is considered equivalent to a boolean expression with value TRUE.
If the computation fails, it is considered equivalent to a boolean expression with value FALSE.
Example 19: Ordering 2. The following sequence decides whether a precedes b in the
lexicographic ordering. The FOR statement finds the first position at which the arrays differ.
If they are equal, the entire sequence fails. If there is a difference at position i, the sequence
succeeds iff a[i] < b[i].
NOT FOR i :=1 TO n DO a[i] = b[i] END;
a[i] < b[i]
2
EITHER–ORELSE. The construct EITHER-ORELSE is added to the language to permit back-
tracking. An EITHER statement has the form
EITHER <seq> ORELSE <seq> . . . .ORELSE <seq> END
Computation of an EITHER statement starts by evaluating the first sequence. If it succeeds,
computation continues at the statement following END. If this computation subsequently
fails, control returns to the second sequence in the EITHER statement, with the program in
the same state as it was when the EITHER statement was entered. Each sequence is tried in
turn in this way, until one succeeds or the last one fails.
7 BACKTRACKING LANGUAGES 113
Listing 54: Creating choice points
EITHER y := x
ORELSE x > 0; y := -x
END;
x := x + b;
y < 0
Listing 55: Pattern matching
SOME i := 0 TO n - m DO
FOR j := 0 to m - 1 DO
s[i+j] = p[j]
END
END
Example 20: Choice points. Assume that the sequence shown in Listing 54 is entered
with x = a > 0. The sequence of events is as follows:
The assignment y := x assigns a to y. This sequence succeeds and transfers control to
x := x + b.
The assignment x := x + b assigns a + b to x.
The test y < 0 fails because y = a > 0.
The original value of x is restored. The test x > 0 succeeds because x = a > 0, and y
receives the value −a.
The final sequence is performed and succeeds, leaving x = a +b and y = −a.
2
SOME. The SOME statement permits the construction of statements that work like EITHER state-
ments but have a variable number of ORELSE clauses. The statement
SOME i := 1 TO n DO <seq> END
is equivalent to
EITHER i := 1; <seq>
ORELSE SOME i := 2 TO n DO <seq> END
Example 21: Pattern matching. Suppose that we have two arrays of characters: a
pattern p[0..m-1] and a string s[0..n-1]. The sequence shown in Listing 55 sets i to the
first occurrence of p in s, or fails if there is no match. 2
COMMIT. The COMMIT statement provides a way of restricting backtracking. (It is similar to
the cut in Prolog.) The statement
COMMIT <seq> END
evaluates the statement sequence <seq>; if seq succeeds (possibly by backtracking), all of
the choice points created during its evaluation are removed.
Example 22: Ordering 3. Listing 56 shows another way of checking lexicographic or-
dering. With the COMMIT, this sequence yields that value of a[i] < b[i] for the smallest i
7 BACKTRACKING LANGUAGES 114
Listing 56: Checking the order
COMMIT
SOME i := 1 TO n DO
a[i] <> b[i]
END
END;
a[i] < b[i]
such that a[i] <> b[i]. Without the COMMIT, it succeeds if a[i] < b[i] is TRUE for some
value of i such that a[i] <> b[i]. 2
FORALL. The FORALL statement provides for the exploration of all choice points generated by a
statement sequence. The statement
FORALL <seq1> DO <seq2> END
evaluates <seq1> and then <seq2>. If <seq1> generates choice points, control backtracks
through these, even if the original sequence succeeded. Whenever <seq2> succeeds, its choice
points are removed (that is, there is an implicit COMMIT around <seq2>).
Example 23: Matching. Suppose we enclose the pattern matching code above in a func-
tion Match. The following sequence finds the number of times that p occurs in s.
count := 0;
FORALL k := Match(p, s) DO count := count + 1 END;
2
EQ. Variables in Alma-0 start life as uninitialized values. After a value has been assigned to
a variable, it ceases to be uninitialized and becomes initialized. Expressions containing
uninitialized values are allowed, but they do not have values. For example, the comparison
s = t is evaluated as follows:
if both s and t are initialized, their values are compared as usual.
if one side is an uninitialized variable and the other is an expression with a value, this
value is assigned to the variable;
if both sides are uninitialized variables, an error is reported.
Example 24: Remarkable sequences. A sequence with 27 integer elements is remark-
able if it contains three 1’s, three 2’s, . . . ., three 9’s arranged in such a way that for
i = 1, 2, . . . , 9 there are exactly i numbers between successive occurrences of i. Here is a
remarkable sequence:
(1, 9, 1, 2, 1, 8, 2, 4, 6, 2, 7, 9, 4, 5, 8, 6, 3, 4, 7, 5, 3, 9, 6, 8, 3, 5, 7).
Problem 1. Write a program that tests whether a given array is a remarkable sequence.
Problem 2. Find a remarkable sequence.
Due to the properties of “=”, the program shown in Listing 57 is a solution for both of these
problems! If the given sequence is initialized, the program tests whether it is remarkable; if
it is not initialized, the program assigns a remarkable sequence to it. 2
7 BACKTRACKING LANGUAGES 115
Listing 57: Remarkable sequences
TYPE Sequence = ARRAY [1..27] OF INTEGER;
PROCEDURE Remarkable (VAR a: Sequence);
VAR i, j: INTEGER;
BEGIN
For i := 1 TO 9 DO
SOME j := 1 TO 25 - 2 * i DO
a[j] = i;
a[j+i+1] = i;
a[j+2*i+1] = i
END
END
END;
Listing 58: Compiling a MIX parameter
VAR v: integer;
BEGIN
v := E;
P(v)
END
MIX. Modula–2 provides call by value (the default) and call by reference (indicated by writing
VAR before the formal parameter. Alma-0 adds a third parameter passing mechanism: call
by mixed form. The parameter is passed either by value or by reference, depending on the
form of the actual parameter.
We can think of the implementation of mixed form in the following way. Suppose that
procedure P has been declared as
PROCEDURE P (MIX n: INTEGER); BEGIN . . . .END;
The call P(v), in which v is a variable, is compiled in the usual way, interpreting MIX as VAR.
The call P(E), in which E is an expression, is compiled as if the programmer had written the
code shown in Listing 58.
Example 25: Linear Search. The function Find in Listing 59 uses a mixed form param-
eter.
The call Find(7,a) tests if 7 occurs in a.
The call Find(x, a), where x is an integer variable, assigns each component of a to x
by backtracking.
Listing 59: Linear search
Type Vector = ARRAY [1..N] of INTEGER;
PROCEDURE Find (MIX e: INTEGER; a: Vector);
VAR i: INTEGER;
BEGIN
SOME i := 1 TO N DO e = a[i] END
END;
7 BACKTRACKING LANGUAGES 116
Listing 60: Generalized Addition
PROCEDURE Plus (MIX x, y, z: INTEGER);
BEGIN
IF KNOWN(x); KNOWN(y) THEN z = x + y
ELSIF KNOWN(y); KNOWN(z) THEN x = z - y
ELSIF KNOWN(x); KNOWN(z) THEN y = z - x
END
END;
The sequence
Find(x, a); Find(x, b)
tests whether the arrays a and b have an element in common; if so, x receives the value
of the common element.
The statement
FORALL Find(x, a) DO Find(x, b) END
tests whether all elements of a are also elements of b.
The statement
FORALL Find(x, a); Find(x, b) DO WRITELN(x) END
prints all of the elements that a and b have in common.
2
KNOWN. The test KNOWN(x) succeeds iff the variable x is initialized.
Example 26: Generalized Addition. Listing 60 shows the declaration of a function
Plus. The call Plus(x,y,z) will succeed if x+y = z provided that at least two of the actual
parameters have values. For example, Plus(2,3,5) succeeds and Plus(5,n,12) assigns 7 to
n. 2
The language Alma-0 has a declarative semantics in which the constructs described above
correspond to logical formulas. It also has a procedural semantics that describes how the
constructs can be efficiently executed, using backtracking. Alma-0 demonstrates that, if appropriate
care is taken, it is possible to design PLs that smoothly combine paradigms without losing the
advantages of either paradigm.
The interesting feature of Alma-0 is that it demonstrates that it is possible to start with a simple,
well-defined language (Modula-2), remove features that are not required for the current purpose,
and add a small number of simple, new features that provide expressive power when used together.
Exercise 33. The new features of Alma-0 are orthogonal, both to the existing features of Modula-
2 and to each other. How does this use of orthogonality compare with the orthogonal design of
Algol 68? 2
7.3 Other Backtracking Languages
Prolog and Alma-0 are by no means the only PLs that incorporate backtracking. Others include
Icon (Griswold and Griswold 1983), SETL (Schwartz, Dewar, Dubinsky, and Schonberg 1986), and
2LP (McAloon and Tretkoff 1995).
7 BACKTRACKING LANGUAGES 117
Constraint PLs are attracting increasing attention. A program is a set of constraints, expressed
as equalities and inequalities, that must be satisfied. Prolog is strong in logical deduction but weak
in numerical calculation; constraint PLs attempt to correct this imbalance.
8 Structure
The earliest programs had little structure. Although a FORTRAN program has a structure — it
consists of a main program and subroutines — the main program and the subroutines themselves
are simply lists of statements.
8.1 Block Structure
Algol 60 was the first major PL to provide recursive, hierarchical structure. It is probably significant
that Algol 60 was also the first language to be defined, syntactically at least, with a context-
free grammar. It is straightforward to define recursive structures using context-free grammars —
specifically, a grammar written in Backus-Naur Form (BNF) for Algol 60. The basis of Algol 60
syntax can be written in extended BNF as follows:
PROG → BLOCK
BLOCK → "begin" ¦ DECL¦ ¦ STMT¦ "end"
STMT → BLOCK [
DECL → HEADER BLOCK [
With blocks come nested scopes. Earlier languages had local scope (for example, FORTRAN
subroutines have local variables) but scopes were continuous (no “holes”) and not nested. It was
not possible for one declaration to override another declaration of the same name. In Algol 60,
however, an inner declaration overrides an outer declaration. The outer block in Listing 61 declares
an integer k, which is first set to 6 and then incremented. The inner block declares another integer
k which is set to 4 and then disappears without being used at the end of the block.
See Section 9.8.1 for further discussion of nesting.
8.2 Modules
By the early 70s, people realized that the “monolithic” structure of Algol-like programs was too
restricted for large and complex programs. Furthermore, the “separate compilation” model of
FORTRAN — and later C — was seen to be unsafe.
The new model that was introduced by languages such as Mesa and Modula divided a large program
into modules.
Listing 61: Nested Declarations
begin
integer k;
k := 6;
begin
integer k;
k := 1;
end;
k := k + 1;
end;
118
8 STRUCTURE 119
A module is a meaningful unit (rather than a collection of possibly unrelated declarations)
that may introduce constants, types, variables, and functions (called collectively features).
A module may import some or all of its features.
In typical modular languages, modules can export types, functions, and variables.
Example 27: Stack Module. One module of a program might declare a stack module:
module Stack
¦
export initialize, push, pop, . . . .
. . . .
¦
Elsewhere in the program, we could use the stack module:
import Stack;
initialize();
push(67);
. . . .
2
This form of module introduces a single instance: there is only one stack, and it is shared by all
parts of the program that import it. In contrast, a stack template introduces a type from which
we can create as many instances as we need.
Example 28: Stack Template. The main difference between the stack module and the stack
template is that the template exports a type.
module Stack
¦
export type STACK, initialize, push, pop, . . . .
. . . .
¦
In other parts of the program, we can use the type STACK to create stacks.
import Stack;
STACK myStack;
initialize(myStack);
push(myStack, 67);
. . . .
STACK yourStack;
. . . .
2
The fact that we can declare many stacks is somewhat offset by the increased complexity of the
notation: the name of the stack instance must appear in every function call. Some modular
languages provide an alternative syntax in which the name of the instance appears before the
function name, as if the module was a record. The last few lines of the example above would be
written as:
import Stack;
STACK myStack;
8 STRUCTURE 120
myStack.initialize();
myStack.push(67);
. . . .
8.2.1 Encapsulation
Parnas (1972) wrote a paper that made a number of important points.
A large program will be implemented as a set of modules.
Each programmer, or programming team, will be the owners of some modules and users of
other modules.
It is desirable that a programmer should be able to change the implementation of a module
without affecting (or even informing) other programmers.
This implies that a module should have an interface that is relatively stable and an imple-
mentation that changes as often as necessary. Only the owner of a module has access to its
implementation; users do not have such access.
Parnas introduced these ideas as the “principle of information hiding”. Today, it is more usual to
refer to the goals as encapsulation.
In order to use the module correctly without access to its implementation, users require a speci-
fication written in an appropriate language. The specification answer the question “What does
this module do?” and the implementation answers the question “How does this module work?”
Most people today accept Parnas’s principle, even if they do not apply it well. It is therefore
important to appreciate how radical it was at the time of its introduction.
[Parnas] has proposed a still more radical solution. His thesis is that the programmer
is most effective if shielded from, rather than exposed to, the details of construction of
system parts other than his own. This presupposes that all interfaces are completely and
precisely defined. While that is definitely sound design, dependence upon its perfect
accomplishment is a recipe for disaster. A good information system both exposes
interface errors and stimulates their correction. (Brooks 1975, page 78)
Brooks wrote this in the first edition of The Mythical Man Month in 1975. Twenty years later, he
had changed his mind.
Parnas was right, and I was wrong. I am now convinced that information hiding, today
often embodied in object-oriented programming, is the only was of raising the level of
software design. (Brooks 1995, page 272)
Parnas’s paper was published in 1972, only shortly after Wirth’s (1971) paper on program de-
velopment by stepwise refinement. Yet Parnas’s approach is significantly different from Wirth’s.
Whereas Wirth assumes that a program can be constructed by filling in details, perhaps with occa-
sional backtracking to correct an error, Parnas sees programming as an iterative process in which
revisions and alterations are part of the normal course of events. This is a more “modern” view
and certainly a view that reflects the actual construction of large and complex software systems.
8 STRUCTURE 121
8.3 Control Structures
Early PLs, such as FORTRAN, depended on three control structures:
sequence;
transfer (goto);
call/return.
These structures are essentially those provided by assembly language. The first contribution of
PLs was to provide names for variables and infix operators for building expressions.
The “modern” control structures first appeared in Algol 60. The most significant contribution
of Algol 60 was the block: a program unit that could contain data, functions, and statements.
Algol 60 also contributed the familiar if-then-else statement and a rather complex loop statement.
However, Algol 60 also provided goto statements and most programmers working in the sixties
assumed that goto was essential.
The storm of controversy raised by Dijkstra’s (1968) letter condemning the goto statement indicates
the extent to which programmers were wedded to the goto statement. Even Knuth (1974) entered
the fray with arguments supporting the goto statement in certain circumstances, although his
paper is well-balanced overall. The basic ideas that Dijkstra proposed were as follows.
There should be one flow into, and one flow out of, each control structure.
All programming needs can be met by three structures satisfying the above property: the
sequence; the loop with preceding test (while-do); and the conditional (if-then-else).
Dijkstra’s proposals had previously been justified in a formal sense by Bohm and Jacopini (1966).
This paper establishes that any program written with goto statements can be rewritten as an
equivalent program that uses sequence, conditional, and loop structures only; it may be necessary
to introduce Boolean variables. Although the result is interesting, Dijkstra’s main point was not
to transform old, goto-ridden programs but to write new programs using simple control structures
only.
Despite the controversy, Dijkstra’s arguments had the desired effect. In languages designed after
1968, the role of goto was down-played. Both C and Pascal have goto statements but they are
rarely used in well-written programs. C, with break, continue, and return, has even less need
for goto than Pascal. Ada has a goto statement because it was part of the US DoD requirements
but, again, it is rarely needed. C
++
inherited goto statements from C. Java does not have a goto
statement, although some uses of continue look rather like jumps.
The use of control structures rather than goto statements has several advantages.
It is easier to read and maintain programs. Maintenance errors are less common.
Precise reasoning (for example, the use of formal semantics) is simplified when goto statements
are absent.
Certain optimizations become feasible because the compiler can obtain more precise informa-
tion in the absence of goto statements.
8.3.1 Loop Structures.
The control structure that has elicited the most discussion is the loop. The basic loop has a single
test that precedes the body:
8 STRUCTURE 122
Listing 62: Reading in Pascal
var n, sum : int;
sum := 0;
read(n);
while n ≥ 0 do
begin
sum := sum + n;
read(n);
end
Listing 63: Reading in C
int sum = 0;
while (true)
¦
int n;
cin >> n;
if (n < 0)
break;
sum += n;
¦
while E do S
It is occasionally useful to perform the test after the body: Pascal provides repeat/until and C
provides do/while.
This is not quite enough, however, because there are situations in which it is useful to exit from
a loop from within the body. Consider, for example, the problem of reading and adding numbers,
terminating when a negative number is read without including the negative number in the sum.
Listing 62 shows the code in Pascal. In C and C++, we can use break to avoid the repeated code,
as shown in Listing 63.
Some recent languages, recognizing this problem, provide only one form of loop statement consisting
of an unconditional repeat and a conditional exit that can be used anywhere in the body of the
loop.
8.3.2 Procedures and Functions
Some languages, such as Ada and Pascal, distinguish the terms procedure and function. Both
names refer to “subroutines” — program components that perform a specific task when invoked
from another component of the program. A procedure is a subroutine that has some effect on the
program data but does not return a result. A function is a subroutine that returns a result; a
function may have some effect on program data, called a “side effect”, but this is often considered
undesirable.
Languages that make a distinction between procedures and functions usually also make a distinction
between “actions” and “data”. For example, at the beginning of the first Pascal text (Jensen and
Wirth 1976), we find the following statement:
An algorithm or computer program consists of two essential parts, a description of
actions which are to be performed, and a description of the data, which are manip-
8 STRUCTURE 123
ulated by the actions. Actions are described be so-called statements, and data are
described by so-called declarations and definitions.
In other languages, such as Algol and C, the distinction between “actions” and “data” is less
emphasized. In these languages, every “statement” has a value as well as an effect. Consistently,
the distinction between “procedure” and “function” also receives less emphasis.
C, for example, does not have keywords corresponding to procedure and function. There is a
single syntax for declaring and defining all functions and a convention that, if a function does not
return a result, it is given the return type void.
We can view the progression here as a separation of concerns. The “concerns” are: whether a
subroutine returns a value; and whether it has side-effects. Ada and Pascal combine the concerns
but, for practical reasons, allow “functions with side-effects”. In C, there are two concerns: “having
a value” and “having an effect” and they are essentially independent.
Example 29: Returning multiple results. Blue (K¨olling 1999) has a uniform notation for
returning either one or more results. For example, we can define a routine with this heading:
findElem (index: Integer) -> (found: Boolean, elem: Element) is . . . .
The parameter index is passed to the routine findElem, and the routine returns two values: found
to indicate whether something was found, and elem as the object found. (If the object is not found,
elem is set to the special value nil.)
Adding multiple results introduces the problem of how to use the results. Blue also provides
multi-assignments: we can write
success, thing := findElem("myKey")
Multi-assignments, provided they are implemented properly, provide other advantages. For exam-
ple we can write the standard “swap sequence”
t := a;
a := b;
b := t;
in the more concise and readable form
a, b := b, a
2
Exercise 34. Describe a “proper” implementation of the multi-assignment statement. Can you
think of any other applications of multi-assignment? 2
Exercise 35. Using Example 29 and other examples, discuss the introduction of new features into
PLs. 2
8.3.3 Exceptions.
The basic control structures are fine for most applications provided that nothing goes wrong.
Although in principle we can do everything with conditions and loops, there are situations in
which these constructs are inadequate.
8 STRUCTURE 124
Example 30: Matrix Inversion. Suppose that we are writing a function that inverts a matrix.
In almost all cases, the given matrix is non-singular and the inversion proceeds normally. Occa-
sionally, however, a singular matrix is passed to the function and, at some point in the calculation,
a division by zero will occur. What choices do we have?
We could test the matrix for singularity before starting to invert it. Unfortunately, the test
involves computing the determinant of the matrix, which is almost as much work as inverting
it. This is wasteful, because we know that most of the matrices that we will receive are
non-singular.
We could test every divisor before performing a division. This is wasteful if the matrix is
unlikely to be singular. Moreover, the divisions probably occur inside nested loops: each level
of loop will require Boolean variables and exit conditions that are triggered by a zero divisor,
adding overhead.
Rely on the hardware to signal an exception when division by zero occurs. The exception
can be caught by a handler that can be installed at any convenient point in the calling
environment.
2
PL/I (Section 3.5) was the first major language to provide exception handling. C provided prim-
itives (setjmp/longjmp) that made it possible to implement exceptin handling. For further de-
scriptions of exception handling in particular languages, see Section 3.
9 Names and Binding
Names are a central feature of all programming languages. In the earliest programs, numbers were
used for all purposes, including machine addresses. Replacing numbers by symbolic names was one
of the first major improvements to program notation (Mutch and Gill 1954).
9.1 Free and Bound Names
The use of names in programming is very similar to the use of names in mathematics. We say that
“x occurs free” in an expression such as x
2
+2x +5. We do not know anything about the value of
this expression because we do not know the value of x. On the other hand, we say that “n occurs
bound” in the expression
5

n=0
(2n + 1). (12)
More precisely, the binding occurrence of n is in n = 0. The n in the expression (2n + 1) is a
reference to, or use of , to this binding.
There are two things to note about (12). First, it has a definite value, 36, because n is bound to a
value (actually, the set of values ¦ 0, 1, . . . , 5 ¦). Second, the particular name that we use does not
change the value of the expression. If we systematically change n to k, we obtain the expression

5
k=0
(2k + 1), which has the same value as (12).
Similarly, the expression


0
e
−x
dx,
which contains a binding occurence of x (in dx) and a use of x (in e
−x
), has a value that does not
depend on the particular name x.
The expression

e
−x
dx,
is interesting because it binds x but does not have a numerical value; the convention is that it
denotes a function, −e
−x
. The conventional notation for functions, however, is ambiguous. It is
clear that −e
−x
is a function of x because, by convention, e denotes a constant. But what about
−e
−ax
: is it a function of a or a function of x? There are several ways of resolving this ambiguity.
We can write
x → −e
−ax
,
read “x maps to −e
−ax
”. Or we can use the notation of the λ-calculus:
λx . − e
−ax
.
In predicate calculus, predicates may contain free names, as in the following expression, whose
truth value depends on the value of n:
n mod 2 = 0 ∧ n mod 3 = 1. (13)
Names are bound by the quantifiers ∀ (for all) and ∃ (there exists). We say that the formula
∀ n. n mod 2 = 0 ⇒ n
2
mod 2 = 0
is closed because it contains no free variables. In this case, the formula is also true because (13) is
true for all values of n. Strictly, we should specify the range of values that n is allowed to assume.
125
9 NAMES AND BINDING 126
We could do this implicitly: for example, it is very likely that (13) refers to integers. Alternatively,
we could define the range of values explicitly, as in
∀ n ∈ Z . n mod 2 = 0 ⇒ n
2
mod 2 = 0.
Precisely analogous situations occur in programming. In the function
int f(int n)
¦
return k + n;
¦
k occurs free and n occurs bound. Note that we cannot tell the value of n from the definition of
the function but we know that n will be given a value when the function is called.
In most PLs, a program with free variables will not compile. A C compiler, for example, will
accept the function f defined above only if it is compiled in a scope that contains a declaration of
k. Some early PLs, such as FORTRAN and PL/I, provided implicit declarations for variables that
the programmer did not declare; this is now understood to be error-prone and is not a feature of
recent PLs.
9.2 Attributes
In mathematics, a variable normally has only one attribute: its value. In (12), n is bound to the
values 0, 1, . . . , 5. Sometimes, we specify the domain from which these values are chosen, as in
n ∈ Z.
In programming, a name may have several attributes, and they may be bound at different times.
For example, in the sequence
int n;
n = 6;
the first line binds the type int to n and the second line binds the value 6 to n. The first binding
occurs when the program is compiled. (The compiler records the fact that the type of n is int;
neither this fact nor the name n appears explicitly in the compiled code.) The second binding
occurs when the program is executed.
This example shows that there are two aspects of binding that we must consider in PLs: the
attribute that is bound, and the time at which the binding occurs. The attributes that may be
bound to a name include: type, address, and value. The times at which the bindings occur include:
compile time, link time, load time, block entry time, and statement execution time.
Definition. A binding is static if it occurs during before the program is executed: during
compilation or linking. A binding is dynamic if it occurs while the program is running: during
loading, block entry, or statement execution.
Example 31: Binding. Consider the program shown in Listing 64. The address of k is bound
when the program starts; its value is bound by the scanf call. The address and value of n are
bound only when the function f is called; if f is not called (because k ≤ 0), then n is never bound.
2
9 NAMES AND BINDING 127
Listing 64: Binding
void f()
¦
int n=7;
printf("%d", n);
¦
void main ()
¦
int k;
scanf("%d", &k);
if (k>0)
f();
¦
9.3 Early and Late Binding
An anonymous contributor to the Encyclopedia of Computer Science wrote:
Broadly speaking, the history of software development is the history of ever-later bind-
ing time.
As a rough rule of thumb:
Early binding ⇒ efficiency;
Late binding ⇒ flexibility.
Example 32: Variable Addressing. In FORTRAN, addresses are bound to variable names at
compile time. The result is that, in the compiled code, variables are addressed directly, without
any indexing or other address calculations. (In reality, the process is somewhat more complicated.
The compiler assigns an address relative to a compilation unit. When the program is linked, the
address of the unit within the program is added to this address. When the program is loaded, the
address of the program is added to the address. The important point is that, by the time execution
begins, the “absolute” address of the variable is known.)
FORTRAN is efficient, because absolute addressing is used. It is inflexible, because all addresses
are assigned at load time. This leads to wasted space, because all local variables occupy space
whether or not they are being used, and also prevents the use of direct or indirect recursion.
In languages of the Algol family (Algol 60, Algol 68, C, Pascal, etc.), local variables are allocated on
the run-time stack. When a function is called, a stack frame (called, in this context, an activation
record or AR) is created for it and space for its parameters and local variables is allocated in the
AR. The variable must be addressed by adding an offset (computed by the compiler) to the address
of the AR, which is not known until the function is called.
Algol-style is slightly less efficient than FORTRAN because addresses must be allocated at block-
entry time and indexed addressing is required. It is more flexible than FORTRAN because inactive
functions do not use up space and recursion works.
Languages that provide dynamic allocation, such as Pascal and C and most of their successors, have
yet another way of binding an address to a variable. In these languages, a statement such as new(p)
allocates space on the “heap” and stores the address of that space in the pointer p. Accessing the
variable requires indirect addressing, which is slightly slower then indexed addressing, but greater
9 NAMES AND BINDING 128
flexibility is obtained because dynamic variables do not have to obey stack discipline (last in, first
out). 2
Example 33: Function Addressing. When the compiler encounters a function definition, it
binds an address to the function name. (As above, several steps must be completed before the
absolute address of the function is determined, but this is not relevant to the present discussion.)
When the compiler encounters a function invocation, it generates a call to the address that it has
assigned to the function.
This statement is true of all PLs developed before OOP was introduced. OOPLs provide “virtual”
functions. A virtual function name may refer to several functions. The actual function referred to
in a particular invocation is not in general known until the call is executed.
Thus in non-OO PLs, functions are statically bound, with the result that calls are efficiently
executed but no decisions can be made at run-time. In OOPLs, functions are dynamically bound.
Calls take longer to execute, but there is greater flexibility because the decision as to which function
to execute is made at run-time.
The overhead of a dynamically-bound function call depends on the language. In Smalltalk, the
overhead can be significant, because the CM requires a search. In practice, the search can be
avoided in up to 95% of calls by using a cache. In C++, the compiler generates “virtual function
tables” and dynamic binding requires only indirect addressing. Note, however, that this provides
yet another example of the principle: Smalltalk binds later than C++, runs more slowly, but
provides greater flexibility. 2
9.4 What Can Be Named?
The question “what can be named?” and its converse “what can be denoted?” are important
questions to ask about a PL.
Definition. An entity can be named if the PL provides a definitional mechanism that associates
a name with an instance of the entity. An entity can be denoted if the PL provides ways of
expressing instances of the entity as expressions.
Example 34: Functions. In C, we can name functions but we cannot denote them. The
definition
int f(int x)
¦
B
¦
introduces a function with name f, parameter x, and body B. Thus functions in C can be named.
There is no way that we can write a function without a name in C. Thus functions in C cannot be
denoted.
The corresponding definition in Scheme would be:
(defun f (lambda (x)
E))
9 NAMES AND BINDING 129
This definition can be split into two parts: the name of the function being defined (f) and the
function itself ((lambda (x) E)). The notation is analogous to that of the λ-calculus:
f = λx . E
FPLs allow expressions analogous to λx . E to be used as functions. LISP, as mentioned in Sec-
tion 4.1, is not one of these languages. However, Scheme is, and so are SML, Haskell, and other
FPLs. 2
Example 35: Types. In early versions of C, types could be denoted but not named. For
example, we could write
int *p[10];
but we could not give a name to the type of p (“array of 10 pointers to int”). Later versions of C
introduced the typedef statement, making types nameable:
typedef int *API[10];
API p;
2
9.5 What is a Variable Name?
We have seen that most PLs choose between two interpretations of a variable name. In a PL
with value semantics, variable names denote memory locations. Assigning to a variable changes
the contents of the location named by the variable. In a PL with reference semantics, variable
names denote objects in memory. Assigning to a variable, if permitted by the language, changes
the denotation of the name to a different object.
Reference semantics is sometimes called “pointer semantics”. This is reasonable in the sense that
the implementation of reference semantics requires the storage of addresses — that is, pointers.
It is misleading in that providing pointers is not the same as providing reference semantics. The
distinction is particularly clear in Hoare’s work on PLs. Hoare (1974) has this to say about the
introduction of pointers into PLs.
Many language designers have preferred to extend [minor, localized faults in Algol 60
and other PLs] throughout the whole language by introducing the concept of reference,
pointer, or indirect address into the language as an assignable item of data. This
immediately gives rise in a high-level language to one of the most notorious confusions
of machine code, namely that between an address and its contents. Some languages
attempt to solve this by even more confusing automatic coercion rules. Worse still, an
indirect assignment through a pointer, just as in machine code, can update any store
location whatsoever, and the damage is no longer confined to the variable explicitly
names as the target of assignment. For example, in Algol 68, the assignment
x := y
always changes x, but the assignment
p := y + 1
may, if p is a reference variable, change any other variable (of appropriate type) in the
whole machine. One variable it can never change is p! . . . . References are like jumps,
leading wildly from one part of a data structure to another. Their introduction into
high-level languages has been a step backward from which we may never recover.
9 NAMES AND BINDING 130
One year later, Hoare (1975) provided his solution to the problem of references in high-level PLs:
In this paper, we will consider a class of data structures for which the amount of
storage can actually vary during the lifetime of the data, and we will show that it can
be satisfactorily accommodated in a high-level language using solely high-level problem-
oriented concepts, and without the introduction of references.
The implementation that Hoare proposes in this paper is a reference semantics with types. Explicit
types make it possible to achieve
. . . . a significant improvement on the efficiency of compiled LISP, perhaps even a
factor of two in space-time cost for suitable applications. (Hoare 1975)
All FPLs and most OOPLs (the notable exception, of course, being C++) use reference semantics.
There are good reasons for this choice, but the reasons are not the same for each paradigm.
In a FPL, all (or at least most) values are immutable. If X and Y are have the same value,
the program cannot tell whether X and Y are distinct objects that happen to be equal, or
pointers to the same object. It follows that value semantics, which requires copying, would be
wasteful because there is no point in making copies of immutable objects.
(LISP provides two tests for equality. (eq x y) is true if x and y are pointers to the same
object. (equal x y) is true if the objects x and y have the same extensional value. These
two functions are provided partly for efficiency and partly to cover up semantic deficiencies of
the implementation. Some other languages provide similar choices for comparison.)
One of the important aspects of OOP is object identity. If a program object X corresponds
to a unique entity in the world, such as a person, it should be unique in the program too. This
is most naturally achieved with a reference semantics.
The use of reference semantics in OOP is discussed at greater length in (Grogono and Chalin 1994).
9.6 Polymorphism
The word “polymorphism” is derived from Greek and means, literally, “many shapes”. In PLs,
“polymorphism” is used to describe a situation in which one name can refer to several different en-
tities. The most common application is to functions. There are several kinds of polymorphism; the
terms “ad hoc polymorphism” and “parametric polymorphism” are due to Christopher Strachey.
9.6.1 Ad Hoc Polymorphism
In the code
int m, n;
float x, y;
printf("%d %f", m+ n, x +y);
the symbol “+” is used in two different ways. In the expression m + n, it stands for the function
that adds two integers. In the expression x +y, it stands for the function that adds two floats.
In general, ad hoc polymorphism refers to the use of a single function name to refer to two or more
distinct functions. Typically the compiler uses the types of the arguments of the function to decide
which function to call. Ad hoc polymorphism is also called “overloading”.
9 NAMES AND BINDING 131
Almost all PLs provide ad hoc polymorphism for built-in operators such as “+”, “−”, “*”, etc.
Ada, C++, and other recent languages also allow programmers to overload functions. (Strictly,
we should say “overload function names” but the usage “overloaded functions” is common.) In
general, all that the programmer has to do is write several definitions, using the same function
name, but ensuring that the either the number or the type of the arguments are different.
Example 36: Overloaded Functions. The following code declares three distinct functions in
C++.
int max (int x, int y);
int max (int x, int y, int z);
float max (float x, float y);
2
9.6.2 Parametric Polymorphism
Suppose that a language provides the type list as a parameterized type. That is, we can make
declarations such as these:
list(int)
list(float)
Suppose also that we have two functions: sum computes the sum of the components of a given
list, and len computes the number of components in a given list. There is an important difference
between these two functions. In order to compute the sum of a list, we must be able to choose an
appropriate “add” function, and this implies that we must know the type of the components. On
the other hand, there seems to be no need to know the type of the components if all we need to
do is count them.
A function such as len, which counts the components of a list but does not care about their type,
has parametric polymorphism.
Example 37: Lists in SML. In SML, functions can be defined by cases, “[]” denotes the empty
list, and “::” denotes list construction. We write definitions without type declarations and SML
infers the types. Here are functions that sum the components of a list and count the components
of a list, respectively.:
sum [] = 0
| sum (x::xs) = x + sum(xs)
len [] = 0
| len(x::xs) = 1 + len(xs)
SML infers the type of sum to be list(int) → int: since the sum of the empty list is (integer)
0, the type of the operator “+” must be int int → int, and all of the components must be
integers.
For the function len, SML can infer nothing about the type of the components from the function
definition, and it assigns the type list(α) → int to the function. α is a type variable, acting as
a type parameter. When len is applied to an actual list, α will implicitly assume the type of the
list components. 2
9 NAMES AND BINDING 132
9.6.3 Object Polymorphism
In OOPLs, there may be many different objects that provide a function called, say, f. However,
the effect of invoking f may depend on the object. The details of this kind of polymorphism,
which are discussed in Section 6, depend on the particular OOPL. This kind of polymorphism is a
fundamental, and very important, aspect of OOP.
9.7 Assignment
Consider the assignment x := E. Whatever the PL, the semantics of this statement will be
something like:
evaluate the expression E;
store the value obtained in x.
The assignment is unproblematic if x and E have the same type. But what happens if they have
different types? There are several possibilities.
The statement is considered to be an error. This will occur in a PL that provides static type
checking but does not provide coercion. For example, Pascal provides only a few coercions
(subrange to integer, integer to real, etc.) and rejects other type differences.
The compiler will generate code to convert the value of expression E to the type of x. This
approach was taken to extremes in COBOL and PL/I. It exists in C and C++, but C++ is
stricter than C.
The value of E will be assigned to x anyway. This is the approach taken by PLs that use
dynamic type checking. Types are associated with objects rather than with names.
9.8 Scope and Extent
The two most important properties of a variable name are scope and extent.
9.8.1 Scope
The scope of a name is the region of the program text in which a name may be used. In C++, for
example, the scope of a local variable starts at the declaration of the variable and ends at the end
of the block in which the declaration appears, as shown in Listing 65. Scope is a static property
of a name that is determined by the semantics of the PL and the text of the program. Examples
of other scope rules follow.
FORTRAN. Local variables are declared at the start of subroutines and at the start of the main
program. Their scope starts at the declarations and end at the end of the program unit (that
is, subroutine or main program) within which the declaration appears. Subroutine names
have global scope. It is also possible to make variables accessible in selected subroutines
by means of COMMON statements, as shown in Listing 66. This method of “scope control” is
insecure because the compiler does not attempt to check consistency between program units.
Algol 60. Local variables are declared at the beginning of a block. Scopes are nested: a decla-
ration of a name in an inner block hides a declaration of the same name in an outer block,
as shown in Listing 67.
9 NAMES AND BINDING 133
Listing 65: Local scope in C++
¦
// x not accessible.
t x;
// x accessible
. . . .
¦
// x still accessible in inner block
. . . .
¦
// x still accessible
¦
// x not accessible
Listing 66: FORTRAN COMMON blocks
subroutine first
integer i
real a(100)
integer b(250)
common /myvars/ a, b
. . . .
return
end
subroutine second
real x(100), y(100)
integer k(250)
common /myvars/ x, k, y
. . . .
return
end
Listing 67: Nested scopes in Algol 60
begin
real x; integer k;
begin
real x;
x := 3.0
end;
x := 6.0;
end
9 NAMES AND BINDING 134
The fine control of scope provided by Algol 60 was retained in Algol 68 but simplified in
other PLs of the same generation.
Pascal. Control structures and nested functions are nested but declarations are allowed only at
the start of the program or the start of a function declaration. In early Pascal, the order of
declarations was fixed: constants, types, variables, and functions were declared in that order.
Later dialects of Pascal sensibly relaxed this ordering.
Pascal also has a rather curious scoping feature: within the statement S of the statement
with r do S, a field f of the record r may be written as f rather than as r.f.
C. Function declarations cannot be nested and all local variables must appear before executable
statements in a block. The complete rules of C scoping, however, are complicated by the fact
that C has five overloading classes (Harbison and Steele 1987):
preprocessor macro names;
statement labels;
structure, union, and enumeration tags;
component names;
other names.
Consequently, the following declarations do not conflict:
¦
char str[200];
struct str ¦int m, n; ¦ x;
. . . .
¦
Inheritance. In OOPLs, inheritance is, in part, a scoping mechanism. When a child class C
inherits a parent class P, some or all of the attributes of P are made visible in C.
C++. There are a number of additional scope control mechanisms in C++. Class attributes may
be declared private (visible only with the class), protected (visible within the class and
derived classes), and public (visible outside the class). Similarly, a class may be derived as
private, protected, or public. A function or a class may be declared as a friend of a class,
giving it access to the private attributes of the class. Finally, C++ follows C in providing
directives such as static and extern to modify scope.
Global Scope. The name is visible throughout the program. Global scope is useful for pervasive
entities, such as library functions and fundamental constants (π = 3.1415926) but is best
avoided for application variables.
FORTRAN does not have global variables, although programmers simulate them by over
using COMMON declarations. Subroutine names in FORTRAN are global.
Names declared at the beginning of the outermost block of Algol 60 and Pascal programs
have global scope. (There may be “holes” in these global scopes if the program contains local
declarations of the same names.)
The largest default scope in C is “file scope”. Global scope can be obtained using the storage
class extern.
9 NAMES AND BINDING 135
Listing 68: Declaration before use
void main ()
¦
const int i = 6;
¦
int j = i;
const int i = 7;
cout << j << endl;
¦
¦
Listing 69: Qualified names
class Widget
¦
public:
void f();
¦;
Widget w;
f(); // Error: f is not in scope
w.f(); // OK
Local Scope. In block-structured languages, names declared in a block are local to the block.
There is a question as to whether the scope starts at the point of definition or is the entire
block. (In other words, does the PL require “declaration before use”?) C++ uses the “declare
before use” rule and the program in Listing 68 prints 6.
In Algol 60 and C++, local scopes can be as small as the programmer needs. Any statement
context can be instantiated by a block that contains local variable declarations. In other
languages, such as Pascal, local declarations can be used only in certain places. Although a
few people do not like the fine control offered by Algol 60 and C++, it seems best to provide
the programmer with as much freedom as possible and to keep scopes as small as possible.
Qualified Scope. The components of a structure, such as a Pascal record or a C struct, have
names. These names are usually hidden, but can be made visible by the name of the object.
Listing 69 provides an example in C++. The construct w. opens a scope in which the public
attributes of the object w are visible.
The dot notation also appears in many OOPLs, where is it used to select attributes and
functions of an object. For an attribute, the syntax is the same: o.a denotes attribute a
of object o. If the attribute is a function, a parameter list follows: o.f(x) denotes the
invocation of function f with argument x in object o.
Import and Export. In modular languages, such as Modula-2, a name or group of names can
be brought into a scope with an import clause. The module that provides the definitions
must have a corresponding export clause.
Typically, if module A exports name X and module B imports name X from A, then X may be
used in module B as if it was declared there.
Namespaces. The mechanisms above are inadequate for very large programs developed by large
teams. Problems arise when a project uses several libraries that may have conflicting names.
9 NAMES AND BINDING 136
Namespaces, provided by PLs such as C++ and Common LISP, provide a higher level of
name management than the regular language features.
In summary:
Nested scopes are convenient in small regions, such as functions. The advantage of nested
scopes for large structures, such as classes or modules, is doubtful.
Nesting is a form of scope management. For large structures, explicit control by name quali-
fication may be better than nesting.
Qualified names work well at the level of classes and modules, when the source of names is
obvious.
The import mechanism has the problem that the source of a name is not obvious where it
appears: users must scan possibly remote import declarations.
Qualified names are inadequate for very large programs.
Library designers can reduce the potential for name conflicts by using distinctive prefixes. For
example, all names supplied by the commercial graphics library FastGraph are have fg_ as a
prefix.
Namespaces provide a better solution than prefixes.
Scope management is important because the programmer has no “work arounds” if the scoping
mechanisms provided by the PL are inadequate. This is because PLs typically hide the scoping
mechanisms.
If a program is compiled, names will normally not be accessible at run-time unless they are included
for a specific purpose such as debugging. In fact, one way to view compiling is as a process that
converts names to numbers.
Interpreters generally have a repository that contains the value of each name currently accessible
to the program, but programmers may have limited access to this repository.
9.8.2 Are nested scopes useful?
As scope control mechanisms became more complicated, some people began to question to need
for nested scopes (Clarke, Wilden, and Wolf 1980; Hanson 1981).
On a small scale, for example in statement and functions, nesting works fairly well. It is,
however, the cause of occasional errors that could be eliminated by requiring all names in a
statement or function to be unique.
On a large scale, for example in classes and modules, nesting is less effective because there is
more text and a larger number of names. It is probably better to provide special mechanisms
to provide the scope control that is needed rather than to rely on simple nesting.
C++ allows classes to be nested (this seems inconsistent, given that functions cannot be nested).
The effect is that the nested class can be accessed only by the enclosing class. An alternative
mechanism would be a declaration that states explicitly that class A can be accessed only by class
B. This mechanism would have the advantage that it would also be able to describe constraints
such as “class A can be accessed only by classes B, C, and D”.
9 NAMES AND BINDING 137
9.8.3 Extent
The extent, also called lifetime, of a name is the period of time during program execution during
which the object corresponding to the name exists. Understanding the relation between scope and
extent is an important part of understanding a PL.
Global names usually exist for the entire lifetime of the execution: they have global extent.
In FORTRAN, local variables also exist for the lifetime of the execution. Programmers assume
that, on entry to a subroutine, its local variables will not have changed since the last invocation of
the subroutine.
In Algol 60 and subsequent stack-based languages, local variables are instantiated whenever control
enters a block and they are destroyed (at least in principle) when control leaves the block. It is an
error to create an object on the stack and to pass a reference to that object to an enclosing scope.
Some PLs attempt to detect this error (e.g., Algol 68), some try to prevent it from occurring (e.g.,
Pascal), and others leave it as a problem for the programmer (e.g., C and C++).
In PLs that use a reference model, objects usually have unlimited extent, whether or not their
original names are accessible. Examples include Simula, Smalltalk, Eiffel, CLU, and all FPLs. The
advantage of the reference model is that the problems of disappearing objects (dangling pointers)
and inaccessible objects (memory leaks) do not occur. The disadvantage is that GC and the
accompanying overhead are more or less indispensable.
The separation of extent from scope was a key step in the evolution of post-Algol PLs.
10 Abstraction
Abstraction is one of the most important mental tools in Computer Science. The biggest problems
that we face in Computer Science are concerned with complexity; abstraction helps us to manage
complexity. Abstraction is particularly important in the study of PLs because a PL is a two-fold
abstraction.
A PL must provide an abstract view of the underlying machine. The values that we use in pro-
gramming — integers, floats, booleans, arrays, and so on — are abstracted from the raw bits of
computer memory. The control structures that we use — assignments, conditionals, loops, and so
on — are abstracted from the basic machine instructions. The first task of a PL is to hide the
low-level details of the machine and present a manageable interface, preferably an interface that
does not depend on the particular machine that is being used.
A PL must also provide ways for the programmer to abstract information about the world. Most
programs are connected to the world in one way or another, and some programs are little more
than simulations of some aspect of the world. For example, an accounting program is, at some
level, a simulation of tasks that used to be performed by clerks and accountants. The second
task of a PL is to provide as much help as possible to the programmer who is creating a simplified
model of the complex world in a computer program.
In the evolution of PLs, the first need came before the second. Early PLs emphasize abstraction
from the machine: they provide assignments, loops, and arrays. Programs are seen as lists of
instructions for the computer to execute. Later PLs emphasize abstraction from the world: they
provide objects and processes. Programs are seen as descriptions of the world that happen to be
executable.
The idea of abstracting from the world is not new, however. The design of Algol 60 began in
the late fifties and, even at that time, the explicit goal of the designers was a “problem oriented
language”.
Example 38: Sets and Bits. Here is an early example that contrasts the problem-oriented and
machine-oriented approaches to PL design.
Pascal provides the problem-oriented abstraction set. We can declare types whose values are sets;
we can define literal constants and variables of these types; and we can write expressions using set
operations, as shown in Listing 70. In contrast, C provides bit-wise operations on integer data, as
shown in Listing 71.
These features of Pascal and C have almost identical implementations. In each case, the elements
of a set are encoded as positions within a machine word; the presence or absence of an element
Listing 70: Sets in Pascal
type SmallNum = set of [1..10];
var s, t: SmallNum;
var n: integer;
begin
s := [1, 3, 5, 7, 9];
t := s + [2, 4, 6, 8];
if (s <= t) . . . .
if (n in s) . . . .
end
138
10 ABSTRACTION 139
Listing 71: Bit operations in C
int n = 0x10101010101;
int mask = 0x10000;
if (n & mask) . . . .
is represented by a 1-bit or a 0-bit, respectively; and the operations for and-ing and or-ing bits
included in most machine languages are used to manipulate the words.
The appearance of the program, however, is quite different. The Pascal programmer can think
at the level of sets, whereas the C programmer is forced to think at the level of words and bits.
Neither approach is necessarily better than the other: Pascal was designed for teaching introductory
programming and C was designed for systems programmers. The trend, however, is away from
machine abstractions such as bit strings and towards problem domain abstractions such as sets. 2
10.1 Abstraction as a Technical Term
“Abstraction” is a word with quite general meaning. In Computer Science in general, and in PL
in particular, it is often given a special sense. An abstraction mechanism provides a way of
naming and parameterizing a program entity.
10.1.1 Procedures.
The earliest abstraction mechanism was the procedure, which exists in rudimentary form even in
assembly language. Procedural abstraction enables us to give a name to a statement, or group of
statements, and to use the name elsewhere in the program to trigger the execution of the statements.
Parameters enable us to customize the execution according to the particular requirements of the
caller.
10.1.2 Functions.
The next abstraction mechanism, which also appeared at an early stage was the function. Func-
tional abstraction enables us to give a name to an expression, and to use the name to trigger
evaluation of the expression.
In many PLs, procedures and functions are closely related. Typically, they have similar syntax and
similar rules of construction. The reason for this is that an abstraction mechanism that works only
for simple expressions is not very useful. In practice, we need to abstract form arbitrarily complex
programs whose main purpose is to yield a single value.
The difference between procedures and functions is most evident at the call site, where a procedure
call appears in a statement context and a function call appears in an expression context.
FORTRAN provides procedures (called “subroutines”) and functions. Interestingly, COBOL does
not provide procedure or function abstraction, although the PERFORM verb provides a rather re-
stricted and unsafe way of executing statements remotely.
LISP provides function abstraction. The effect of procedures is obtained by writing functions with
side-effects.
Algol 60, Algol 68, and C take the view that everything is an expression. (The idea seems to be
that, since any computation leaves something in the accumulator, the programmer might as well
be allowed to use it.) Thus statements have a value in these languages. It is not surprising that
10 ABSTRACTION 140
this viewpoint minimizes the difference between procedures and functions. In Algol 68 and C, a
procedure is simple a function that returns a value of type void. Since this value is unique, it
requires log 1 = 0 bits of storage and can be optimized away by the compiler.
10.1.3 Data Types.
Although Algol 68 provided a kind of data type abstraction, Pascal was more systematic and
more influential. Pascal provides primitive types, such as integer, real, and boolean, and type
expressions (C denotes an integer constant and T denotes a type):
C. . C
(C, C, . . . ., C)
array [T] of T
record v: T, . . . .end
set of T
file of T
Type expressions can be named and the names can be used in variable declarations or as compo-
nents of other types. Pascal, however, does not provide a parameter mechanism for types.
10.1.4 Classes.
Simula introduced an abstraction mechanism for classes. Although this was an important and
profound step, with hindsight we can see that two simple ideas were needed: giving a name to an
Algol block, and freeing the data component of the block from the tyranny of stack discipline.
10.2 Computational Models
There are various ways of describing a PL. We can write an informal description in natural language
(e.g., the Pascal User Manual and Report (Jensen and Wirth 1976)), a formal description in natural
language (e.g., the Algol report (Naur et al. 1960)), or a document in a formal notation (e.g., the
Revised Report on Algol 68 (van Wijngaarden et al. 1975)). Alternatively, we can provide a formal
semantics for the language, using the axiomatic, denotational, or operational approaches.
There is, however, an abstraction that is useful for describing a PL in general terms, without
the details of every construct, and that is the computational model (also sometimes called
the semantic model ). The computational model (CM) is an abstraction of the operational
semantics of the PL and it describes the effects of various operations without describing the actual
implementation.
Programmers who use a PL acquire intuitive knowledge of its CM. When a program behaves in an
unexpected way, the implementation has failed to respect the CM.
Example 39: Passing parameters. Suppose that a PL specifies that arguments are passed by
value. This means that the programs in Listing 72 and Listing 73 are equivalent. The statement
T a = E means “the variable a has type T and initial value E” and the statement x ← a means
“the value of a is assigned (by copying) to x”.
A compiler could implement this program by copying the value of the argument, as if executing
the assignment x ← a. Alternatively, it could pass the address of the argument a to the function
but not allow the statement S to change the value of a. Both implementations respect the CM. 2
10 ABSTRACTION 141
Listing 72: Passing a parameter
void f (T x);
¦
S;
¦
T a = E;
f(a);
Listing 73: Revealing the parameter passing mechanism
T a = E;
¦
T x;
x ← a;
S;
¦
Example 40: Parameters in FORTRAN. If a FORTRAN subroutine changes the value
of one of its formal parameters (called “dummy variables” in FORTRAN parlance) the value of
the corresponding actual argument at the call site also changes. For example, if we define the
subroutine
subroutine p (count)
integer count
. . . .
count = count + 1
return
end
then, after executing the sequence
integer total
total = 0
call p(total)
the value of total is 1. The most common way of achieving this result is to pass the address of the
argument to the subroutine. This enables the subroutine to both access the value of the argument
and to alter it.
Unfortunately, most FORTRAN compilers use the same mechanism for passing constant arguments.
The effect of executing
call p(0)
is to change all instances of 0 (the constant “zero”) in the program to 1 (the constant “one”)! The
resulting bugs can be quite hard to detect for a programmer who is not familiar with FORTRAN’s
idiosyncrasies.
The programmer has a mental CM for FORTRAN which predicts that, although variable arguments
may be changed by a subroutine, constant arguments will not be. The problem arises because
typical FORTRAN compilers do not respect this CM. 2
A simple CM does not imply easy implementation. Often, the opposite is true. For example, C is
easy to compile, but has a rather complex CM.
10 ABSTRACTION 142
Example 41: Array Indexing. In the CM of C, a[i] ≡ ∗(a + i). To understand this, we must
know that an array can be represented by a pointer to its first element and that we can access
elements of the array by adding the index (multiplied by an appropriate but implicit factor) to this
pointer.
In Pascal’s CM, an array is a mapping. The declaration
var A : array [1..N] of real;
introduces A as a mapping:
A : ¦ 1, . . . , N ¦ → {.
A compiler can implement arrays in any suitable way that respects this mapping. 2
Example 42: Primitives in OOPLs. There are two ways in which an OOPL can treat prim-
itive entities such as booleans, characters, and integers.
Primitive values can have a special status that distinguishes them from objects. A language
that uses this policy is likely to be efficient, because the compiler can process primitives without
the overhead of treating them as full-fledged objects. On the other hand, it may be confusing
for the user to deal with two different kinds of variable — primitive values and objects.
Primitive values can be given the same status as objects. This simplifies the programmer’s
task, because all variables behave like objects, but a naive implementation will probably be
inefficient.
A solution for this dilemma is to define a CM for the language in which all variables are objects.
An implementor is then free to implement primitive values efficiently provided that the behaviour
predicted by the CM is obtained at all times.
In Smalltalk, every variable is an object. Early implementations of Smalltalk were inefficient,
because all variables were actually implemented as objects. Java provides primitives with efficient
implementations but, to support various OO features, also has classes for the primitive types. 2
Example 43: λ-Calculus as a Computational Model. FPLs are based on the theory of
partial recursive functions. They attempt to use this theory as a CM. The advantage of this point
of view is that program text resembles mathematics and, to a certain extent, programs can be
manipulated as mathematical objects.
The disadvantage is that mathematics and computation are not the same. Some operations that
are intuitively simple, such as maintaining the balance of a bank account subject to deposits and
withdrawals, are not easily described in classical mathematics. The underlying difficulty is that
many computations depend on a concept of mutable state that does not exist in mathematics.
FPLs have an important property called referential transparency. An expression is referentially
transparent if its only attribute is its value. The practical consequence of referential transparency
is that we can freely substitute equal expressions.
For example, if we are dealing with numbers, we can replace E + E by 2 E; the replacement
would save time if the evaluation of E requires a significant amount of work. However, if E was not
referentially transparent — perhaps because its evaluation requires reading from an input stream or
the generation of random numbers — the replacement would change the meaning of the program.
2
10 ABSTRACTION 143
Example 44: Literals in OOP. How should an OOPL treat literal constants, such as 7? The
following discussion is based on (K¨olling 1999, pages 71–74).
There are two possible ways of interpreting the symbol “7”: we could interpret it as an object
constructor, constructing the object 7; or as a constant reference to an object that already exists.
Dee (Grogono 1991a) uses a variant of the first alternative: literals are constructors. This introduces
several problems. For example, what does the comparison 7 = 7 mean? Are there two 7-objects
or just one? The CM of Dee says that the first occurrence of the literal “7” constructs a new
object and subsequent occurrences return a reference to this (unique) object.
Smalltalk uses the second approach. Literals are constant references to objects. These objects,
however, “just magically exist in the Smalltalk universe”. Smalltalk does not explain where they
come from, or when and how they are created.
Blue avoids these problems by introducing the new concept of a manifest class. Manifest classes
are defined by enumeration. The enumeration may be finite, as for class Boolean with members
true and false. Infinite enumerations, as for class Integer, are allowed, although we cannot see
the complete declaration.
The advantage of the Blue approach, according to K¨olling (1999, page 72), is that:
. . . . the distinction is made at the logical level. The object model explicitly recognizes
these two different types of classes, and differences between numbers and complex
objects can be understood independently of implementation concerns. They
cease to be anomalies or special cases at a technical level. This reduces the dangers of
misunderstandings on the side of the programmer. [Emphasis added.]
2
Exercise 36. Do you think that K¨olling’s arguments for introducing a new concept (manifest
classes) are justified? 2
Summary Early CMs:
modelled programs as code acting on data;
structured programs by recursive decomposition of code.
Later CMs:
modelled programs as packages of code and data;
structured programs by recursive decomposition of packages of code and data.
The CM should:
help programmers to reason about programs;
help programmers to read and write programs;
constrain but not determine the implementation.
11 Implementation
This course is concerned mainly with the nature of PLs. Implementation techniques are a sepa-
rate, and very rich, topic. Nevertheless, there are few interesting things that we can say about
implementation, and that is the purpose of this section.
11.1 Compiling and Interpreting
11.1.1 Compiling
As we have seen, the need for programming languages at a higher level than assembly language
was recognized early. People also realized that the computer itself could perform the translation
from a high level notation to machine code. In the following quote, “conversion” is used in roughly
the sense that we would say “compilation” today.
On the other hand, large storage capacities will lead in the course of time to very
elaborate conversion routines. Programmes in the future may be written in a very
different language from that in which the machines execute them. It is to be hoped (and
it is indeed one of the objects of using a conversion routine) that many of the tiresome
blunders that occur in present-day programmes will be avoided when programmes can
be written in a language in which the programmer feels more at home. However, any
mistake that the programmer does commit may prove more difficult to find because it
will not be detected until the programme has been converted into machine language.
Before he can find the cause of the trouble the programmer will either have to investigate
the conversion process or enlist the aid of checking routines which will ‘reconvert’ and
provide him with diagnostic information in his own language. (Gill 1954)
A compiler translates source language to machine language. Since practical compilation systems
allow programs to be split into several components for compiling, there are usually several phases.
1. The components of the source program is compiled into object code.
2. The object code units and library functions required by the program are linked to form an
executable file.
3. The executable file is loaded into memory and executed.
Since machine code is executed directly, compiled programs ideally run at the maximum possible
speed of the target machine. In practice, the code generated by a compiler is usually not optimal.
For very small programs, a programmer skilled in assembly language may be able to write more
efficient code than a compiler can generate. For large programs, a compiler can perform global
optimizations that are infeasible for a human programmer and, in any case, hand-coding such large
programs would require excessive time.
All compilers perform some checks to ensure that the source code that they are translating is
correct. In early languages, such as FORTRAN, COBOL, PL/I, and C, the checking was quite
superficial and it was easy to write programs that compiled without errors but failed when they
were executed. The value of compile-time checking began to be recognized in the late 60s, and
languages such as Pascal, Ada, and C++ provided more thorough checking than their predecessors.
The goal of checking is to ensure that, if a program compiles without errors, it will not fail when
144
11 IMPLEMENTATION 145
it is executed. This goal is hard to achieve and, even today, there are few languages that achieve
it totally.
Compiled PLs can be classified according to whether the compiler can accept program components
or only entire programs. An Algol 60 program is a block and must be compiled as a block. Standard
Pascal programs cannot be split up, but recent dialects of Pascal provide program components.
The need to compile programs in parts was recognized at an early stage: even FORTRAN was
implemented in this way. The danger of separate compilation is that there may be inconsistencies
between the components that the compiler cannot detect. We can identify various levels of checking
that have evolved.
A FORTRAN compiler provides no checking between components. When the compiled com-
ponents are linked, the linker will report subroutines that have not been defined but it will
not detect discrepancies in the number of type of arguments. The result of such a discrepancy
is, at best, run-time failure.
C and C++ rely on the programmer to provide header files containing declarations and imple-
mentation files containing definitions. The compiler checks that declarations and definitions
are compatible.
Overloaded functions in C++ present a problem because one name may denote several func-
tions. A key design decision in C++, however, was to use the standard linker (Stroustrup
1994, page 34). The solution is “name mangling” — the C++ compiler alters the names of
functions by encoding their argument types in such a way that each function has a unique
name.
Modula-2 programs consist of interface modules and implementation modules. When the com-
piler compiles an implementation module, it reads all of the relevant interfaces and performs
full checking. In practice, the interface modules are themselves compiled for efficiency, but
this is not logically necessary.
Eiffel programs consist of classes that are compiled separately. The compiler automatically
derives an interface from the class declaration and uses the generated interfaces of other classes
to perform full checking. Other OOPLs that use this method include Blue, Dee, and Java.
11.1.2 Interpreting
An interpreter accepts source language and run-time input data and executes the program di-
rectly. It is possible to interpret without performing any translation at all, but this approach
tends to be inefficient. Consequently, most interpreters do a certain amount of translation before
executing the code. For example, the source program might be converted into an efficient internal
representation, such as an abstract syntax tree, which is then “executed”.
Interpreting is slower than executing compiled code, typically by a factor of 10 or more. Since no
time is spent in compiling, however, interpreted languages can provide immediate response. For
this reason, prototyping environments often provide an interpreted language. Early BASIC and
LISP systems relied on interpretation, but more recent interactive systems provide an option to
compile and may even provide compilation only.
Most interpreters check correctness of programs during execution, although some interpreters may
report obvious (e.g., syntactic) errors as the program is read. Language such as LISP and Smalltalk
are often said to be untyped, because programs do not contain type declarations, but this is
incorrect. The type checking performed by both of these languages is actually more thorough than
11 IMPLEMENTATION 146
that performed by C and Pascal compilers; the difference is that errors are not detected until the
offending statements are actually executed. For safety-critical applications, this may be too late.
11.1.3 Mixed Implementation Strategies
Some programming environments, particularly those for LISP, provide both interpretation and
compilation. In fact, some LISP environments allow source code and compiled code to be freely
mixed.
Byte Codes. Another mixed strategy consists of using a compiler to generate byte codes and an
interpreter to execute the byte codes. The “byte codes”, as the name implies, are simply streams
of bytes that represent the program: a machine language for an “abstract machine”. One of the
first programming systems to be based on byte codes was the Pascal “P” compiler.
The advantage of byte code programs is that they can, in principle, be run on any machine. Thus
byte codes provide a convenient way of implementing “portable” languages.
The Pascal “P” compiler provided a simple method for porting Pascal to a new kind of machine.
The “Pascal P package” consisted of the P-code compiler written in P-code. A Pascal implementor
had to perform the following steps:
1. Write a P-code interpreter. Since P-code is quite simple, this is not a very arduous job.
2. Use the P-code interpreter to interpret the P-code compiler. This step yields a P-code pro-
gram that can be used to translate Pascal programs to P-code. At this stage, the implementor
has a working Pascal system, but it is inefficient because it depends on the P-code interpreter.
3. Modify the compiler source so that the compiler generates machine code for the new machine.
This is a somewhat tedious task, but it is simplified by the fact that the compiler is written
in Pascal.
4. Use the compiler created in step 2 to compile the modified compiler. The result of this step
is a P-code program that translates Pascal to machine code.
5. Use the result of step 4 to compile the modified compiler. The result is a machine code
program that translates Pascal to machine code — in other words, a Pascal compiler for the
new machine.
Needless to say, all of these steps must be carried out with great care. It is important to ensure
that each step is completed correctly before beginning the next.
With the exercise of due caution, the process described above (called bootstrapping) can be
simplified. For example, the first version of the compiler, created in step 2 does not have to
compile the entire Pascal language, but only those parts that are needed by the compiler itself.
The complete sequence of steps can be used to generate a Pascal compiler that can compile itself,
but may be incomplete in other respects. When that compiler is available, it can be extended to
compile the full language.
Java uses byte codes to obtain portability in a different way. Java was designed to be used as an
internet language. Any program that is to be distributed over the internet must be capable of
running on all the servers and clients of the network. If the program is written in a conventional
language, such as C, there must be a compiled version for every platform (a platform consists of
a machine and operating system). If the program is written in Java, it is necessary only to provide
a Java byte code interpreter for each platform.
11 IMPLEMENTATION 147
Just In Time. “Just in time” (JIT) compilation is a recent strategy that is used for Java and a
few research languages. The idea is that the program is interpreted, or perhaps compiled rapidly
without optimization. As each statement is executed, efficient code is generated for it. This
approach is based on the following assumptions:
in a typical run, much of the code of a program will not be executed;
if a statement has been executed once, it is likely to be executed again.
JIT provides fast response, because the program can be run immediately, and efficient execution,
because the second and subsequent cycles of a loop are executed from efficiently compiled code.
Thus JIT combines some of the advantages of interpretation and compilation.
Interactive Systems A PL can be designed for writing large programs that run autonomously
or for interacting with the user. The categories overlap: PLs that can be used interactively, such
as BASIC, LISP, SML, and Smalltalk, can also be used to develop large programs. The converse
is not true: an interactive environment for C++ would not be very useful.
An interactive PL is usually interpreted, but interpretation is not the only possible means of
implementation. PLs that start out with interpreters often acquire compilers at a later stage —
BASIC, LISP, and a number of other interactive PLs followed this pattern of development.
A PL that is suitable for interactive use if programs can be described in small chunks. FPLs are
often supported interactively because a programmer can build up a program as a collection of func-
tion definitions, where most functions are fairly small (e.g., they fit on a screen). Block structured
languages, and modular languages, are less suitable for this kind of program development.
11.1.4 Consequences for Language Design
Any PL can be compiled or interpreted. A PL intended for interpretation, however, is usually
implemented interactively. This implies that the user must be able to enter short sequences of
code that can be executed immediately in the current context. Many FPLs are implemented in
this way (often with a compiler); the user can define new functions and invoke any function that
has previously been defined.
11.2 Garbage Collection
There is a sharp distinction: either a PL is implemented with garbage collection (GC) or it is
not. The implementor does not really have a choice. An implementation of LISP, Smalltalk, or
Java without GC would be useless — programs would exhaust memory in a few minutes at most.
Conversely, it is almost impossible to implement C or C++ with GC, because programmers can
perform arithmetic on pointers.
The imperative PL community has been consistently reluctant to incorporate GC: there is no
widely-used imperative PL that provides GC. The functional PL community has been equally
consistent, but in the opposite direction. The first FPL, LISP, provides GC, and so do all of its
successors.
The OOP community is divided on the issue of GC, but most OOPLs provide it: Simula, Smalltalk,
CLU, Eiffel, and Java. The exception, of course, is C++, for reasons that have been clearly
explained (Stroustrup 1994, page 219).
11 IMPLEMENTATION 148
I deliberately designed C++ not to rely on [garbage collection]. I feared the very
significant space and time overheads I had experienced with garbage collectors. I feared
the added complexity of implementation and porting that a garbage collector would
impose. Also, garbage collection would make C++ unsuitable for many of the low-level
tasks for which it was intended. I like the idea of garbage collection as a mechanism
that simplifies design and eliminates a source of errors. However, I am fully convinced
that had garbage collection been an integral part of C++ originally, C++ would have
been stillborn.
It is interesting to study Stroustrup’s arguments in detail. In his view, the “fundamental reasons”
that garbage collection is desirable are (Stroustrup 1994, page 220):
1. Garbage collection is easiest for the user. In particular, it simplifies the building and use of
some libraries.
2. Garbage collection is more reliable than user-supplied memory management schemes for
some applications.
The emphasis is added: Stroustrup makes it clear that he would have viewed GC in a different
light if it provided advantages for all applications.
Stroustrup provides more arguments against GC than in favour of it (Stroustrup 1994, page 220):
1. Garbage collection causes run-time and space overheads that are not affordable for many
current C++ applications running on current hardware.
2. Many garbage collection techniques imply service interruptions that are not acceptable for
important classes of applications, such as hard real-time applications, device drivers, control
applications, human interface code on slow hardware, and operating system kernels.
3. Some applications do not have the hardware resources of a traditional general-purpose com-
puter.
4. Some garbage collection schemes require banning several basic C facilities such as pointer
arithmetic, unchecked arrays, and unchecked function arguments as used by printf().
5. Some garbage collection schemes impose constraints on object layout or object creation that
complicates interfacing with other languages.
It is hard to argue with Stroustrup’s conclusion (Stroustrup 1994, page 220):
I know that there are more reasons for and against, but no further reasons are needed.
These are sufficient arguments against the view that every application would be better
done with garbage collection. Similarly, these are sufficient arguments against the view
that no application would be better done with garbage collection. (Stroustrup 1994,
page 220) [Emphasis in the original.]
Stroustrup’s arguments do not preclude GC as an option for C++ and, in fact, Stroustrup advo-
cates this. But it remains true that it is difficult to add GC to C++ and the best that can be done
is probably “conservative” GC. In the meantime, most C++ programmers spend large amounts of
time figuring out when to call destructors and debugging the consequences of their wrong decisions.
Conservative GC depends on the fact that it is possible to guess with a reasonable degree of
accuracy whether a particular machine word is a pointer (Boehm and Weiser 1988). In typical
11 IMPLEMENTATION 149
modern processors, a pointer is a 4-byte quantity, aligned on a word boundary, and its value is a
legal address. By scanning memory and noting words with these properties, the GC can identify
regions of memory that might be in use. The problem is that there will be occasionally be words
that look like pointers but actually are not: this results in about 10% of garbage being retained.
The converse, and more serious problem, of identifying live data as garbage cannot occur: hence
the term “conservative”.
The implementation of Eiffel provided by Meyer’s own company, ISE, has a garbage collector that
can be turned on and off. Eiffel, however, is a public-domain language that can be implemented
by anyone who chooses to do so — there are in fact a small number of companies other than ISE
that provide Eiffel. Implementors are encouraged, but not required to provide GC. Meyer (1992,
page 335) recommends that an Eiffel garbage collector should have the following properties.
It should be efficient, with a low overhead on execution time.
It should be incremental, working in small, frequent bursts of activity rather than waiting for
memory to fill up.
It should be tunable. For example, programmers should be able to switch off GC in critical
sections and to request a full clean-up when it is safe to do so.
The concern about the efficiency of GC is the result of the poor performance of early, mark-sweep
collectors. Modern garbage collectors have much lower overheads and are “tunable” in Meyer’s
sense. Larose and Feeley (1999), for example, describe a garbage collector for a Scheme compiler.
Programs written in Erlang, translated into Scheme, and compiled with this compiler are used in
ATM switches. (Erlang is a proprietary functional language developed by Ericsson.)
GC has divided the programming community into two camps: those who use PLS with GC and
consider it indispensable, and those who use PLs without GC and consider it unnecessary. Expe-
rience with C++ is beginning to affect the polarization, as the difficulties and dangers of explicit
memory management in complex OO systems becomes apparent. Several commercial products
that offer “conservative” GC are now available and their existence suggests that developers will
increasingly respect the contribution of GC.
12 Conclusion
Values and Objects. Mathematically oriented views of programming favour values. Simulation
oriented views favour objects. Both views are useful for particular applications. A PL that
provides both in a consistent and concise way might be simple, expressive, and powerful.
The “logical variable” (a variable that is first unbound, then bound, and later may be unbound
during backtracking) is a third kind of entity, after values and objects. It might be fruitful to
design a PL with objects, logical variables, unification, and backtracking. Such a language
would combine the advantages of OOP and logic programming.
Making Progress. It is not hard to design a large and complex PL by throwing in numerous
features. The hard part of design is to find simple, powerful abstractions — to achieve more
with less. In this respect, PL design resembles mathematics. The significant advances in
mathematics are often simplifications that occur when structures that once seemed distinct
are united in a common abstraction. Similar simplifications have occurred in the evolution
of PLs: for example, Simula (Section 6.1) and Scheme (Section 4.2). But programs are not
mathematical objects. A rigorously mathematical approach can lead all too rapidly to the
“Turing tar pit”.
Practice and Theory. The evolution of PLs shows that, most of the time, practice leads theory.
Designers and implementors introduce new ideas, then theoreticians attempt to what they
did and how they could have done it better. There are a number of PLs that have been based
on purely theoretical principles but few, if any, are in widespread use.
Theory-based PLs are important because they provide useful insights. Ideally, they serve
as testing environments for ideas that may eventually be incorporated into new mainstream
PLs. Unfortunately, it is often the case that they show retroactively how something should
have been done when it is already too late to improve it.
Multiparadigm Programming. PLs have evolved into several strands: procedural program-
ming, now more or less subsumed by object oriented programming; functional programming;
logic programming and its successor, constraint programming. Each paradigm performs well
on a particular class of applications. Since people are always searching for “universal” solu-
tions, it is inevitable that some will try to build a “universal” PL by combining paradigms.
Such attempts will be successful to the extent that they achieve overall simplification. It is
not clear that any future universal language will be accepted by the programming community;
it seems more likely that specialized languages will dominate in the future.
150
References
Abelson, H. and G. Sussman (1985). Structure and Interpretation of Computer Programs. MIT
Press.
˚
Ake Wikstr¨om (1987). Functional Programming Using Standard ML. Prentice Hall.
Apt, K. R., J. Brunekreef, V. Partington, and A. Schaerf (1998, September). Alma-0: an impera-
tive language that supports declarative programming. ACM Trans. Programming Languages
and Systems 20(5), 1014–1066.
Arnold, K. and J. Gosling (1998). The Java Programming Language (Second ed.). The Java
Series. Addison Wesley.
Ashcroft, E. and W. Wadge (1982). R
x
for semantics. ACM Trans. Programming Languages and
Systems 4(2), 283–294.
Backus, J. (1957, February). The FORTRAN automatic coding system. In Proceedings of the
Western Joint Computer Conference.
Backus, J. (1978, August). Can programming be liberated from the von Neumann style? A
functional style and its algebra of programs. Communications of the ACM 21(8), 613–64.
Bergin, Jr., T. J. and R. G. Gibson, Jr. (Eds.) (1996). History of Programming Languages–II.
ACM Press (Addison Wesley).
Boehm, H.-J. and M. Weiser (1988). Garbage collection in an uncooperative environment. Soft-
ware: Practice & Experience 18(9), 807–820.
Bohm, C. and G. Jacopini (1966, May). Flow diagrams, Turing machines, and languages with
only two formation rules. Communications of the ACM 9(5), 366–371.
Bratko, I. (1990). PROLOG: Programming for Artificial Intelligence (Second ed.). Addison Wes-
ley.
Brinch Hansen, P. (1999, April). Java’s insecure parallelism. ACM SIGPLAN 34(4), 38–45.
Brooks, F. P. (1975). The Mythical Man-Month. Addison Wesley.
Brooks, F. P. (1987, April). No silver bullet: Essence and accidents for software engineering.
IEEE Computer 20(4), 10–18.
Brooks, F. P. (1995). The Mythical Man-Month (Anniversary ed.). Addison Wesley.
Budd, T. A. (1995). Multiparadigm Programming in Leda. Addison Wesley.
Church, A. (1941). The calculi of lambda-conversion. In Annals of Mathematics Studies, Vol-
ume 6. Princeton Univesity Press.
Clarke, L. A., J. C. Wilden, and A. L. Wolf (1980). Nesting in Ada programs is for the birds. In
Proc. ACM SIGPLAN Symposium on the Ada programming Language. Reprinted as ACM
SIGPLAN Notices, 15(11):139–145, November 1980.
Colmerauer, A., H. Kanoui, P. Roussel, and R. Pasero (1973). Un systeme de communication
homme-machine en francais. Technical report, Groupe de Recherche en Intelligence Artifi-
cielle, Universit´e d’Aix-Marseille.
Dahl, O.-J., E. W. Dijkstra, and C. A. R. Hoare (1972). Structured Programming. Academic
Press.
Dijkstra, E. W. (1968, August). Go to statement considered harmful. Communications of the
ACM 11(8), 538.
Gelernter, D. and S. Jagannathan (1990). Programming Linguistics. MIT Press.
151
REFERENCES 152
Gill, S. (1954). Getting programmes right. In International Symposium on Automatic Digi-
tal Computing. National Physical Laboratory. Reprinted as pp. 289–292 of (Williams and
Campbell-Kelly 1989).
Griswold, R. E. and M. T. Griswold (1983). The Icon Programming Language. Prentice Hall.
Grogono, P. (1991a, January). The Dee report. Technical Report
OOP–91–2, Department of Computer Science, Concordia University.
ttp://www.cs.concordia.ca/∼faculty/grogono/dee.html.
Grogono, P. (1991b, January). Issues in the design of an object oriented
programming language. Structured Programming 12(1), 1–15. Available as
www.cs.concordia/~faculty/grogono/oopissue.ps.
Grogono, P. and P. Chalin (1994, May). Copying, sharing, and aliasing. In Colloquium on Object
Orientation in Databases and Software Engineering (ACFAS’94), Montreal, Quebec. Avail-
able as www.cs.concordia/~faculty/grogono/copying.ps.
Hanson, D. R. (1981, August). Is block structure necessary? Software: Practice and Experi-
ence 11(8), 853–866.
Harbison, S. P. and G. L. Steele, Jr. (1987). C: A Reference Manual (second ed.). Prentice Hall.
Hoare, C. A. R. (1968). Record handling. In Programming Languages, pp. 291–347. Academic
Press.
Hoare, C. A. R. (1974). Hints on programming language design. In C. Bunyan (Ed.), Computer
Systems Reliability, Volume 20 of State of the Art Report, pp. 505–34. Pergamon/Infotech.
Reprinted as pages 193–216 of (Hoare 1989).
Hoare, C. A. R. (1975, June). Recursive data structures. Int. J. Computer and Information
Science 4(2), 105–32. Reprinted as pages 217–243 of (Hoare 1989).
Hoare, C. A. R. (1989). Essays in Computing Science. Prentice Hall. Edited by C. B. Jones.
Jensen, K. and N. Wirth (1976). Pascal: User Manual and Report. Springer-Verlag.
Joyner, I. (1992). C++? a critique of C++. Available from ian@syacus.acus.oz.au. Second
Edition.
Kay, A. C. (1996). The early history of Smalltalk. In History of Programming Languages–II, pp.
511–578. ACM Press (Addison Wesley).
Kleene, S. C. (1936). General recursive functinos of natural numbesr. Mathematische An-
nalen 112, 727–742.
Knuth, D. E. (1974). Structured programming with GOTO statements. ACM Computing Sur-
veys 6(4), 261–301.
K¨olling, M. (1999). The Design of an Object-Oriented Environment and Language for Teaching.
Ph. D. thesis, Basser Department of Computer Science, University of Sydney.
Kowalski, R. A. (1974). Predicate logic as a programming language. In Information processing
74, pp. 569–574. North Holland.
LaLonde, W. and J. Pugh (1995, March/April). Complexity in C++: a Smalltalk perspective.
Journal of Object-Oriented Programming 8(1), 49–56.
Landin, P. J. (1965, August). A correspondence between Algol 60 and Church’s lambda-notation.
Communications of the ACM 8, 89–101 and 158–165.
Larose, M. and M. Feeley (1999). A compacting incrmental collector and its performce in a pro-
duction quality compiler. In ACM SIGPLAN International Symposiuum on Memory Man-
agement (ISMM’98). Reprinted as ACM SIGPLAN Notices 343:1–9, March 1999.
REFERENCES 153
Lientz, B. P. and E. B. Swanson (1980). Software Maintenance Management. Addison Wesley.
Lindsey, C. H. (1996). A history of Algol 68. In History of Programming Languages, pp. 27–83.
ACM Press (Addison Wesley).
Liskov, B. (1996). A history of CLU. In History of Programming Languages–II, pp. 471–496.
ACM Press (Addison Wesley).
Liskov, B. and J. Guttag (1986). Abstraction and Specification in Program Development. MIT
Press.
Liskov, B. and S. Zilles (1974, April). Programming with abstract data types. ACM SIGPLAN
Notices 9(4), 50–9.
MacLennan, B. J. (1983, December). Values and objects in programming languages. ACM SIG-
PLAN Notices 17(12), 70–79.
Martin, J. J. (1986). Data Types and Data Structures. Prentice Hall International.
McAloon, K. and C. Tretkoff (1995). 2LP: Linear programming and logic programming. In P. V.
Hentenryck and V. Saraswat (Eds.), Principles and Preactice of Constraint Programming,
pp. 101–116. MIT Press.
McCarthy, J. (1960, April). Recursive functions of symbolic expressions and their computation
by machine, Part I. Communications of the ACM 3(4), 184–195.
McCarthy, J. (1978). History of LISP. In R. L. Wexelblat (Ed.), History of Programming Lan-
guages, pp. 173–197. Academic Press.
McKee, J. R. (1984). Maintenance as a function of design. In Proc. AFIPS National Computer
Conference, pp. 187–93.
Meyer, B. (1988). Object-oriented Software Construction. Prentice Hall International. Second
Edition, 1997.
Meyer, B. (1992). Eiffel: the Language. Prentice Hall International.
Milne, R. E. and C. Strachey (1976). A Theory of Programming Language Semantics. John
Wiley.
Milner, B. and M. Tofte (1991). Commentary on Standard ML. MIT Press.
Milner, B., M. Tofte, and R. Harper (1990). The Definition of Standard ML. MIT Press.
Mutch, E. N. and S. Gill (1954). Conversion routines. In International Symposium on Automatic
Digital Computing. National Physical Laboratory. Reprinted as pp. 283–289 of (Williams and
Campbell-Kelly 1989).
Naur, P. (1978). The European side of the last phase of the development of Algol. In History of
Programming Languages. ACM Press. Reprinted as pages 92–137 of (Wexelblat 1981).
Naur, P. et al. (1960, May). Report on the algorithmic language Algol 60. Communications of
the ACM 3(5), 299–314.
Nygaard, K. and O.-J. Dahl (1978). The development of the SIMULA languages. In R. L.
Wexelblat (Ed.), History of Programming Languages, pp. 439–478. Academic Press.
Parnas, D. L. (1972, December). On the criteria to be used in decomposing systems into modules.
Communications of the ACM 15(12), 1053–1058.
Perlis, A. J. (1978). The American side of the development of Algol. In History of Programming
Languages. ACM Press. Reprinted as pages 75–91 of (Wexelblat 1981).
Radin, G. (1978). The early history and characteristics of PL/I. In History of Programming
Languages. ACM Press. Reprinted as pages 551–574 of (Wexelblat 1981).
REFERENCES 154
Ritchie, D. M. (1996). The development of the C programming language. In History of Program-
ming Languages, pp. 671–686. ACM Press (Addison Wesley).
Robinson, A. J. (1965). A machine-oriented logic based on the resolution principle. Journal of
the ACM 12, 23–41.
Sakkinen, M. (1988). On the darker side of C++. In S. Gjessing and K. Nygaard (Eds.), Pro-
ceedings of the 1988 European Conference of Object Oriented Programming, pp. 162–176.
Springer LNCS 322.
Sakkinen, M. (1992). The darker side of C++ revisited. Structured Programming 13.
Sammet, J. (1969). Programming Languages: History and Fundamentals. Prentice Hall.
Sammett, J. E. (1978). The early history of COBOL. In History of Programming Languages.
ACM Press. Reprinted as pages 199–241 of (Wexelblat 1981).
Schwartz, J. T., R. B. K. Dewar, E. Dubinsky, and E. Schonberg (1986). Programming with sets
— an introduction to SETL. Springer Verlag.
Steele, Jr., G. L. (1996). The evolution of LISP. In History of Programming Languages, pp.
233–308. ACM Press (Addison Wesley).
Steele, Jr., G. L. et al. (1990). Common LISP: the Language (second ed.). Digital Press.
Steele, Jr., G. L. and G. J. Sussman (1975). Scheme: an interpreter for the extended lambda
calculus. Technical Report Memo 349, MIT Artificial Intelligence Laboratory.
Strachey, C. (1966). Towards a formal semantics. In T. B. Steel (Ed.), Formal Language De-
scription Languages for Computer Programming. North-Holland.
Stroustrup, B. (1994). The Design and Evolution of C++. Addison Wesley.
Taivalsaari, A. (1993). A Critical View of Inheritance and Reusability in Object-oriented Pro-
gramming. Ph. D. thesis, University of Jyv¨askyl¨a.
Teale, S. (1993). C++ IOStreams Handbook. Addison Wesley. Reprinted with corrections, Febru-
ary 1994.
Turing, A. M. (1936). On computable numbers with an application to the Entscheidungsproblem.
Proc. London Math. Soc. 2(42), 230–265.
Turner, D. (1976). The SASL Langauge Manual. Technical report, University of St. Andrews.
Turner, D. (1979). A new implementation technique for applicative languages. Software: Practice
& Experience 9, 31–49.
Turner, D. (1981). KRC language manual. Technical report, University of Kent.
Turner, D. (1985, September). Miranda: a non-strict functional language with polymorphic
types. In Proceedings of the Second International Conference on Functional Programming
Languages and Computer Architecture. Springer LNCS 301.
van Wijngaarden, A. et al. (1975). Revised report on the algorithmic language Algol 68. Acta
Informatica 5, 1–236. (In three parts.).
Wexelblat, R. L. (Ed.) (1981). History of Programming Languages. Academic Press. From the
ACM SIGPLAN History of programming Languages Conference, 1–3 June 1978.
Wheeler, D. J. (1951). Planning the use of a paper library. In Report of a Conference on High
Speed Automatic Calculating-Engines. University Mathematical Laboratory, Cambridge.
Reprinted as pp. 42–44 of (Williams and Campbell-Kelly 1989).
Whitaker, W. A. (1996). Ada — the project. In History of Programming Languages, pp. 173–228.
ACM Press (Addison Wesley).
REFERENCES 155
Williams, M. R. and M. Campbell-Kelly (Eds.) (1989). The Early British Computer Conferences,
Volume 14 of The Charles Babbage Institute Reprint Series for the History of Computing.
The MIT Press and Tomash Publishers.
Wirth, N. (1971, April). Program development by stepwise refinement. Communications of the
ACM 14(4), 221–227.
Wirth, N. (1976). Algorithms + Data Structures = Programs. Prentice Hall.
Wirth, N. (1982). Programming in Modula-2. Springer-Verlag.
Wirth, N. (1996). Recollections about the development of Pascal. In History of Programming
Languages, pp. 97–110. ACM Press (Addison Wesley).
A Abbreviations and Glossary
Actual parameter. A value passed to a function.
ADT. Abstract data type. A data type described in terms of the operations that can be performed
on the data.
AR. Activation record. The run-time data structure corresponding to an Algol 60 block. Sec-
tion 3.3.
Argument. A value passed to a function; synonym of actual parameter.
Bind. Associate a property with a name.
Block. In Algol 60, a syntactic unit containing declarations of variables and functions, followed
by code. Successors of Algol 60 define blocks in similar ways. In Smalltalk, a block is a
closure consisting of code, an environment with bindings for local variables, and possibly
parameters.
BNF. Backus-Naur Form (originally Backus Normal Form). A notation for describing a context-
free grammar, introduced by John Backus for describing Algol 60. Section B.4.2.
Call by need. The argument of a function is not evaluated until its value is needed, and then
only once. See lazy evaluation.
CBN. Call by name. A parameter passing method in which actual parameters are evaluated when
they are needed by the called function. Section 3.3
CBV. Call by value. A parameter passing method in which the actual parameter is evaluated
before the function is called.
Closure. A data structure consisting of one (or more, but usually one) function and its lexical
environment. Section 4.2.
CM. Computational Model. Section 10.2
Compile. Translate source code (program text) into code for a physical (or possibly abstract)
machine.
CS. Computer Science.
Delegation. An object that accepts a request and then forwards the request to another object is
said to “delegate” the request.
Dummy parameter. Formal parameter (FORTRAN usage).
Dynamic. Used to describe something that happens during execution, as opposed to static.
Dynamic Binding. A binding that occurs during execution.
Dynamic Scoping. The value of a non-local variable is the most recent value that has been
assigned to it during the execution of the program. Cf. Lexical Scoping. Section 4.1.
Dynamic typing. Performing type checking “on the fly”, while the program is running.
Eager evaluation. The arguments of a function are evaluated at the time that it is called, whether
or not their values are needed.
156
A ABBREVIATIONS AND GLOSSARY 157
Early binding. Binding a property to a name at an early stage, usually during compilation.
Exception handling. A PL feature that permits a computation to be abandoned; control is
transferred to a handler in a possibly non-local scope.
Extent. The time during execution during which a variable is active. It is a dynamic property,
as opposed to scope.
Formal parameter. A variable name that appears in a function heading; the variable will receive
a value when the function is called. “Parameter” without qualification usually means “formal
parameter”.
FP(L). Functional Programming (Language). Section 4.
Garbage collection. Recycling of unused memory when performed automatically by the imple-
mentatin rather than coded by the programmer. (The term “garbage collection” is widely
used but inaccurate. The “garbage” is thrown away: it is the useful data that is collected!)
GC. Garbage collection.
Global scope. A name that is accessible throughout the (static) program text has global scope.
Heap. A region for storing variables in which no particular order of allocation or deallocation is
defined.
High order function. A function that either accepts a function as an argument, or returns a
function as a value, or both.
Higher order function. Same as high order function.
Inheritance. In OOPLs, when an object (or class) obtains some of its features from another
object (or class), it is said to inherit those features.
Interpret. Execute a program one statement at a time, without translating it into machine code.
Lambda calculus. A notation for describing functions, using the Greek letter λ, introduced by
Church.
Late binding. A binding (that is, attachment of a property to a name) that is performed at a later
time than would normally be expected. For example, procedural languages bind functions to
names during compilation, but OOPLs “late bind” function names during execution.
Lazy evaluation. If the implementation uses lazy evaluation, or call by need, a function does not
evaluate an argument until the value of the argument is actually required for the computation.
When the argument has been evaluated, its value is stored and it is not evaluated again during
the current invocation of the function.
Lexical Scoping. The value of a non-local variables is determined in the static context as deter-
mined by the text of the program. Cf. Dynamic Scoping. Section 4.1.
Local scope. A name that is accessible in only a part of the program text has local scope.
OOP(L). Object Oriented Programming (Language). Section 6.
Orthogonal. Two language features are orthogonal if they can be combined without losing any
of their important properties. Section 3.6.
A ABBREVIATIONS AND GLOSSARY 158
Overloading. Using the same name to denote more than one entity (the entities are usually
functions).
Own variable. A local variable in an Algol 60 program that has local scope (it can be accessed
only within the procedure in which it is declared) but global extent (it is created when the
procedure is first called and persists until the program termiantes). Section 3.3.
Parameter. Literally “unknown” or “unmeasurable” (from Greek). Used in PLs for the objects
passed to functions.
Partial function. A function that is defined over only a part of its domain.

x is partial over
the domain of reals because it has no real value if x < 0.
PL. Programming Language.
Polymorphism. Using one name to denote several distinct objects. (Literally “many shapes”).
Recursion. A process or object that is defined partly in terms of itself without circularity.
Reference semantics. Variable names denote objects. Used by all functional and logic languages,
and most OOPLs.
RE. Regular Expression. Section B.4.
Scope. The region of text in which a variable is accessible. It is a static property, as opposed to
extent.
Semantic Model. An alternative term for Computational Model (CM). Section 10.2.
Semantics. A system that assigns meaning to programs.
Stack. A region for storing variables in which the most recently allocated unit is the first one to
be deallocated (“last-in, first out”, LIFO). The name suggests a stack of plates.
Static. Used to describe something that happens during compilation (or possible during linking
but, anyway, before the program is executed), as opposed to dynamic.
Static Binding. A value that is bound before the program is executed. Cf. Dynamic Binding.
Static typing. Type checking performed during compilation.
Strong typing. Detect and report all type errors, as opposed to weak typing.
Template. A syntactic construct from which many distinct objects can be generated. A template
is different from a declaration that constructs only a single object. (The word “template”
has a special sense in C++, but the usage is more or less compatible with our definition.)
Thunk. A function with no parameters used to implement the call by name mechanism of Algol
60 and a few other languages.
Total function. A function that is defined for all values of its arguments in its domain. e
x
is
total over the domain of reals.
Value semantics. Variable names denote locations. Most imperative languages use value seman-
tics.
Weak typing. Detect and report some type errors, but allow some possible type errors to remain
in the program.
B Theoretical Issues
As stated in Section 1, theory is not a major concern of the course. In this section, however, we
briefly review the contributions that theory has made to the development of PLs and discuss the
importance and relevance of these contributions.
B.1 Syntactic and Lexical Issues
There is a large body of knowledge concerning formal languages. There is a hierarchy of languages
(the “Chomsky hierarchy”) and, associated with each level of the hierarchy, formal machines that
can “recognize” the corresponding languages and perform other tasks. The first two levels of the
hierarchy are the most familiar to programmers: regular languages and finite state automata;
and context-free languages and push-down automata. Typical PLs have context-free (or almost
context-free) grammars, and their tokens, or lexemes, can be described by a regular language.
Consequently, programs can be split into lexemes by a finite state automaton and parsed by a
push-down automaton.
It might appear, then, that the problem of PL design at this level is “solved”. Unfortunately,
there are exceptions. The grammar of C++, for example, is not context-free. This makes the
construction of a C++ parser unnecessarily difficult and accounts for the slow development of
C++ programming tools. The preprocessor is another obstacle that affects program development
environments in both C and C++.
Exercise 37. Give examples to demonstrate the difficulties that the C++ preprocessor causes in
a program development environment. 2
B.2 Semantics
There are many kinds of semantics, including: algebraic, axiomatic, denotational, operational,
and others. They are not competitive but rather complementary. There are many applications of
semantics, and a particular semantic system may be more or less appropriate for a particular task.
For example, a denotational semantics is a mapping from program constructs to abstract math-
ematical entities that represent the “meaning” of the constructs. Denotational semantics is an
effective language for communication between the designer and implementors of a PL but is not
usually of great interest to a programmer who uses the PL. An axiomatic semantics, on the other
hand, is not very useful to an implementor but may be very useful to a programmer who wishes
to prove the correctness of a program.
The basic ideas of denotational semantics are due to Landin (1965) and Strachey (1966). The most
complete description was completed by Milne (1976) after Strachey’s death. Denotational semantics
is a powerful descriptive tool: it provides techniques for mapping almost any PL feature into a
high order mathematical function. Denotational semantics developed as a descriptive technique
and was extended to handle increasingly arcane features, such as jumps into blocks.
Ashcroft and Wadge published an interesting paper (1982), noting that PL features that were
easy to implement were often hard to describe and vice versa. They suggested that denotational
semantics should be used prescriptively rather then descriptively. In other words, a PL designer
should start with a simple denotational semantics and then figure out how to implement the
language — or perhaps leave that task to implementors altogether. Their own language, Lucid,
was designed according to these principles.
159
B THEORETICAL ISSUES 160
To some extent, the suggestions of Ashcroft and Wadge have been followed (although not neces-
sarily as a result of their paper). A description of a new PL in the academic literature is usually
accompanied by a denotational semantics for the main features of the PL. It is not clear, however,
that semantic techniques have had a significant impact on mainstream language design.
B.3 Type Theory
There is a large body of knowledge on type theory. We can summarize it concisely as follows. Here
is a function declaration.
T f(S x) ¦ B; return y; ¦
The declaration introduces a function called f that takes an argument, x of type S, performs
calculations indicated as B, and returns a result, y of type T. If there is a type theory associated
with the language, we should be able to prove a theorem of the following form:
If x has type S then the evaluation of f(x) yields a value of type T.
The reasoning used in the proof is based on the syntactic form of the function body, B.
If we can do this for all legal programs in a language L, we say that L is statically typed. If L is
indeed statically typed:
A compiler for L can check the type correctness of all programs.
A program that is type-correct will not fail because of a type error when it is executed.
A program that is type-correct may nevertheless fail in various ways. Type checking does not
usually detect errors such as division by zero. (In general, most type checking systems assume that
functions are total : a function with type S → T, given a value of type S, will return a value of
type T. Some commonly used functions are partial : for example x/y ≡ divide(x,y) is undefined
for y = 0.) A program can be type-correct and yet give completely incorrect answers. In general:
Most PLs are not statically typed in the sense defined above, although the number of loopholes
in the type system may be small.
A type-correct program may give incorrect answers.
Type theories for modern PLs, with objects, classes, inheritance, overloading, genericity, dy-
namic binding, and other features, are very complex.
Type theory leads to attractive and interesting mathematics and, consequently, tends to be over-
valued by the theoretical Computer Science community. The theory of program correctness is more
difficult and has attracted less attention, although it is arguably more important.
Exercise 38. Give an example of an expression or statement in Pascal or C that contains a type
error that the compiler cannot detect. 2
B.4 Regular Languages
Regular languages are based on a simple set of operations: sequence, choice, and repetition. Since
these operations occur in many contexts in the study of PLs, it is interesting to look briefly at
regular languages and to see how the concepts can be applied.
B THEORETICAL ISSUES 161
In the formal study of regular languages, we begin with an alphabet, Σ, which is a finite set of
symbols. For example, we might have Σ = ¦ 0, 1 ¦. The set of all finite strings, including the empty
string, constructed from the symbols of Σ is written Σ

.
We then introduce a variety of regular expressions, which we will refer to as RE here. (The
abbreviation RE is also used to stand for “recursively enumerable”, but in this section it stands
for “regular expression”.) Each form of RE is defined by the set of strings in Σ

that it generates.
This set is called a “regular language”.
1.
g

is RE and denotes the empty set.
2. (the string containing no symbols) is RE and denotes the set ¦ ¦.
3. For each symbol x ∈ Σ, x is RE and denotes the set ¦ x ¦.
4. If r is RE with language R, then r

is RE and denotes the set R

(where
R

= ¦ x
1
x
2
. . . x
n
[ n ≥ 0 ∧ x
i
∈ R¦ ).
5. If r and s are RE with languages R and S respectively, then (r + s) is RE and denotes the
set R∪ S.
6. If r and s are RE with languages R and S respectively, then (rs) is RE and denotes the set
RS (where
RS = ¦ rs [ r ∈ R∧ s ∈ S ¦ ).
We add two further notations that serve solely as abbreviations.
1. The expression a
n
represents the string aa . . . a containing n occurrences of the symbol a.
2. The expression a
+
is an abbreviation for the RE aa

that denotes the set ¦ a, aa, aaa, . . . ¦.
B.4.1 Tokens
We can use regular languages to describe the tokens (or lexemes) of most PLs. For example:
UC = A+ B + +Z
LC = a + b + +z
LETTER = UC +LC
DIGIT = 0 + 1 + + 9
IDENTIFIER = LETTER( LETTER + DIGIT)

INTCONST = DIGIT
+
FLOATCONST = DIGIT
+
. DIGIT

. . . .
We can use a program such as lex to construct a lexical analyzer (scanner) from a regular expression
that defines the tokens of a PL.
Exercise 39. Some compilers allow “nested comments”. For example, ¦. . . ¦. . .¦ . . .¦ in Pascal or
/ ∗ . . . / ∗ . . .∗ / . . . ∗ / in C. What consequences does this have for the lexical analyzer (scanner)? 2
B THEORETICAL ISSUES 162
RE Control Structure
x statement
rs statement; statement
r

while expression do statement
r +s if condition then statement else statement
Figure 17: REs and Control Structures
B.4.2 Context Free Grammars
Context free grammars for PLs are usually written in BNF (Backus-Naur Form). A grammar rule,
or production, consists of a non-terminal symbol (the symbol being defined), a connector (usually
::= or →), and a sequence of terminal and non-terminal symbols. The syntax for the assignment,
for example, might be defined:
ASSIGNMENT → VARIABLE “ := ” EXPRESSION
Basic BNF provides no mechanism for choice: we must provide one rule for each possibility. Also,
BNF provides no mechanism for repetition; we must use recursion instead. The following example
illustrates both of these limitations.
SEQUENCE → EMPTY
SEQUENCE → STATEMENT SEQUENCE
BNF has been extended in a variety of ways. With a few exceptions, most of the extended forms
can be described simply: the sequence of symbols on the right of the connector is replaced by a
regular expression. The extension enables us to express choice and repetition within a single rule.
In grammars, the symbol [ rather than + is used to denote choice.
STATEMENT = ASSIGNMENT [ CONDITIONAL [
SEQUENCE = ( STATEMENT)

Extended BNF (EBNF) provides a more concise way of describing grammars than BNF. Just as
parsers can be constructed from BNF grammars, they can be constructed from EBNF grammars.
Exercise 40. Discuss the difficulty of adding attributes to an EBNF grammar. 2
B.4.3 Control Structures and Data Structures
Figure 17 shows a correspondence between REs and the control structures of Algol-like languages.
Similarly, Figure 18 shows a correspondence between REs and data structures..
Exercise 41. Why might Figures 17 and 18 be of interest to a PL designer? 2
B.4.4 Discussion
These analogies suggest that the mechanisms for constructing REs — concatenation, alternation,
and repetition — are somehow fundamental. In particular, the relation between the standard
control structures and the standard data structures can be helpful in programming. In Jackson
B THEORETICAL ISSUES 163
RE Data Structure
x int n;
rs struct ¦ int n; float y; ¦
r
n
int a[n];
r

int a[]
r +s union ¦ int n; float y; ¦
Figure 18: REs and Data Structures
Structured Design (JSD), for example, the data structures appropriate for the application are
selected first and then the program is constructed with the corresponding control structures.
The limitations of REs are also interesting. When we move to the next level of the Chomsky
hierarchy, Context Free Languages (CFLs), we obtain the benefits of recursion. The corresponding
control structure is the procedure and the corresponding data structures are recursive structures
such as lists and trees (Wirth 1976, page 163).
C Haskell Implementation
Several modern functional programming languages, including Haskell, are implemented using com-
binators and graph reduction. The technique described below was introduced into mathematical
logic by M. Schoenfinkel in 1924 and was rediscovered as an implementation method by D.A. Turner
in 1979.
We start with three observations.
Operators are really just functions. For example, we can write x+y as add x y. Thus we only
have to consider prefix functions.
Functions of more than one variable can be reduced to functions of one variable. We understand
add x y as (add x) y: we evaluate (add x), obtaining a function of one variable, which we then
apply to y.
In a definition such as
succ x = add 1 x
we can distinguish between bound variables, or simply variables, such as x in this examples,
and constants, such as 1 and +. Although add might be defined elsewhere, it is a constant
as far as this definition is concerned.
We would like to write the definition of succ in the form
succ = ???
so that the distinction between the name being defined and the definition is clear. Schoenfinkel
showed how to do this using the following abstraction algorithm. If e is an expression possibly
containing x, the abstraction of e with respect to x is written [x] e. It is an expression that does
not contain x. The rules for abstraction follow. In the second rule, we assume x = y, because
otherwise the first rule would apply.
[x] x = I
[x] y = K y
[x] (f g) = S ([x] f) ([x] g)
Apply these rules to the function succ, defined above, gives
succ = [x] (add 1 x)
= S ([x] (add 1)) ([x] x)
= S (S (K add) (K 1)) I
These expressions quickly become quite long. If we abstract n from
fac n = if (n == 0) 1 (n fac(n −1))
(where we treat if as a function of three arguments), we obtain
fac = S(S(S(K I if)(S(S(K ==)(K0))I))(K1))(S(S(K)I)(S(Kfac)(S(S(K −)I)( 1))))
In general, the size of the SKI expression increases exponentially with the size of the original
expression. Turner introduced new functions B and C and optimizations to overcome the size
problem. Using these, the definition of fac becomes:
fac = S(C(B if (== 0))1)(S (B fac (C − 1)))
164
C HASKELL IMPLEMENTATION 165
S

©
d

S

©
d

I
K
c
K
c
+ 1
Figure 19: The graph for succ
To make use of these expressions, we apply an argument and then make use of the following
reduction rules:
I x = x
K x y = x
S f g x = (f x) (g x)
B f g x = f (g x)
C f g x = f x g
Using the simpler example, succ, we have:
succ 4 = S (S (K add) (K 1)) I 4
= (S (K add) (K 1) 4) (I 4)
= ((K + 4) (K 1 4)) 4
= + 1 4
= 5
For practical implementation, these expressions are represented as graphs: Figure 19 shows the
graph corresponding to succ. To evaluate an expression such as succ 4, we add a third subtree
to the top node, giving S three arguments, and transform the graph, always using the rule for
whatever function is at the root of the tree.
The basic technique of graph reduction is quite slow. When Turner presented one his functional
programming languages at a conference in 1982, he said that it was about 500 times slower than
conventional languages such as C and FORTRAN. Twenty years of research, however, have led
to steady improvements, and a modern implementation of a functional language has performance
comparable to that of a compiled procedural language.

CONTENTS

ii

Contents
1 Introduction 1.1 1.2 How important are programming languages? . . . . . . . . . . . . . . . . . . . . . Features of the Course . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1 2 3 7 7 9 10 14 14 16 16 17 18 18 20 20 24 24 26 26 29 30 32 34 34 36 38 39 42 43 44 44 46 Hugs Commands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Literate Haskell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Defining Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Syntax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tracing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Lists . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Lazy Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2 Anomalies 3 The Procedural Paradigm 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 Early Days . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . FORTRAN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Algol 60 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . COBOL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . PL/I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Algol 68 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Pascal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Modula–2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.10 Ada . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 The Functional Paradigm 4.1 LISP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.1 4.1.2 4.1.3 4.2 4.3 4.4 4.5 Basic LISP Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A LISP Interpreter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dynamic Binding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . SASL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . SML . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Other Functional Languages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

5 Haskell 5.1 Getting Started . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1.1 5.1.2 5.1.3 5.1.4 5.1.5 5.2 5.2.1 5.2.2

Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

CONTENTS 5.2.3 5.2.4 5.2.5 5.2.6 5.3 5.3.1 5.3.2 5.3.3 5.3.4 5.3.5 5.3.6 5.3.7 5.3.8 5.4 5.4.1 5.4.2 5.5 Pattern Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Building Haskell Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . Guards . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Using Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

iii 47 50 51 53 54 56 58 58 59 59 60 62 68 71 73 73 74 77 77 79 80 83 84 84 86 87 87 88 89 89 90 91 93 93 93 94 96

Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Failure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Reporting Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Standard Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Standard Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Defining New Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Defining Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Types with Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . String Formatting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Interactive Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Input and Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Reasoning about programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

6 The Object Oriented Paradigm 6.1 6.2 6.3 6.4 6.5 Simula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Smalltalk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . CLU . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C++ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Eiffel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.5.1 6.5.2 6.5.3 6.6 6.6.1 6.6.2 6.6.3 6.6.4 6.7 6.8 6.9 Programming by Contract . . . . . . . . . . . . . . . . . . . . . . . . . . . . Repeated Inheritance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Exception Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Portability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Interfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Exception Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Concurrency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Java . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Kevo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Other OOPLs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Issues in OOP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.9.1 6.9.2 6.9.3 Algorithms + Data Structures = Programs . . . . . . . . . . . . . . . . . . Values and Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Classes versus Prototypes . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Scope and Extent . . . . . Polymorphism . . . . . . . . . .11 Evaluation of OOP . . . . . . . . . . . . . . . . . . . . . . . Early and Late Binding . . . . . iv 96 96 97 97 97 105 106 106 112 116 118 118 118 120 121 121 122 123 125 125 126 127 128 129 130 130 131 132 132 132 132 136 137 6. . . . . . . . . . Other Backtracking Languages .6. . . . . . . . . . . . . . . . . . .2 7. . Control Structures .9. . . . . . . . . . . . . What is a Variable Name? . . . . . . . . .1 8. . . . . . . . . . . . .6 6. . .1 8. . . .3 Encapsulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 Structure 8. . . . . .6 Free and Bound Names . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Loop Structures. . . . . . .4 6. .2 9. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8. . . . . . . . . .2. . . . . .8. . . . . . . . . . . . . .3 9. . Pure versus Hybrid Languages . . . .10 A Critique of C++ . . .5 9. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 8. Procedures and Functions . . . . . . . . . . . . . . . . .3 Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8. . . . . . . . . . . . . . . . . . . . . . . . . .5 6. . . . . . . Object Polymorphism . . . . . . . . . . . . . . . . . . .8 Ad Hoc Polymorphism . . . . . . . . . . . . . . . . . . . . . . . .3. . . . . . . . . . . . .1 8. . .7 Types versus Objects . . . . . . . . . . . . . Alma-0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3. Parametric Polymorphism . . .4 9. . . . . . . . . Inheritance . . . . . . . . . . . . . . . . . .3. . . . . . . . . . . . . . . . 6. . . . . . . . . . . . . 9. . . . . . . . . . . . . . . . . .9. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Closures versus Classes . . . . . . . . . . . .6. . . . . . 9. . . . . . . . .1 9. . . . . . . . . . . . . . . . . . . . . . . . Modules . . . . . . . . .1 9. . Extent . . . . . . . . . .3 9. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Are nested scopes useful? . . . . . . . . . . . . . . . . . . . . .8. . . . . . . . . . . . . . . . .9. . . . .3 Prolog . . . . . . .3 Block Structure . . . . . . . . . . . . . . . . . . . . . . . . . . .2 9. . . . . . . . . . . . . . . . . . . . . . Attributes . . . . . . . . . . . . . . . . . . . . . . . .1 9. . . . What Can Be Named? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 7. . . . . . . . . . . . . . . . . . . . . . . . .7 9.CONTENTS 6. . . . . . . Assignment .9. . . . . Exceptions. . . . . . . 7 Backtracking Languages 7. . . . . . . . . . . . . . . . . . . . . . . 9 Names and Binding 9. . . . . . .2 9. . . . . . .

1. . .2 Context Free Grammars . . . . . . . . . . . . . A Literate Haskell program . . . .1 Tokens . . . . . . . . . . . . . . . 10. . . . . . . . . . .4 Consequences for Language Design . . 11. . . . . . . . . . . 11. . . . . . . . . . . . . . . . . .1 Syntactic and Lexical Issues . . . . . . . . .1 Procedures. . . . . . . . . . . B. . . . . . . . . . . . . . . . . . . . . . .1. . . . . . . . . . . . . . . . . . . . . . B. . . . . . . . . . . . . . . . . . . . . . . . . . . . .1. . . . . . . . . . . . . . .2 Functions. . . . . . . . . . . . . . . . . . . . . . . . . . .4 Discussion . . . . . . . . 21 35 36 37 40 . . . .3 Type Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . .1. . .2 Interpreting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4. . . . . . . . . . . .4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1.3 Control Structures and Data Structures . . . . . . . .1 Compiling and Interpreting . . . . . . . . . . . . . . . . . . B. . . . . 10. . . . . . . . . . . . . . . . . . 10. . . . . . . . . . . . . 10. . . . . . . 11. .4. . . . . . . . . . . 12 Conclusion References A Abbreviations and Glossary B Theoretical Issues B. . . . . . . .4 Regular Languages . . . . . . . . . . . . . .3 Mixed Implementation Strategies . . . . . . . . 10. . . . . . . . . . . . . . . . . . . . . . . . . Hugs commands . . . . . . . . . . . . . . . . . . . Parameters for :set . . . . . . B. . . . . . . . . . . . . . . . . . . . 11. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A short session with Hugs . . . . . . . .2 Semantics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Computational Models . . . .1.1. . . . . . . . . . . C Haskell Implementation v 138 139 139 139 140 140 140 144 144 144 145 146 147 147 150 151 156 159 159 159 160 160 161 162 162 162 164 List of Figures 1 2 3 4 5 The list structure (A (B C)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Compiling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4. . . . . . . . . . . . . . . . .LIST OF FIGURES 10 Abstraction 10. . . . . . . . . . . .1. . . . B. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Data Types. . . . .1 Abstraction as a Technical Term . . . . . . . . 11. . . . .4 Classes. . . . B. . . . . . . . . . . . 11 Implementation 11. . B. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Garbage Collection . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .LIST OF FIGURES 6 7 8 9 10 11 12 13 14 15 16 17 18 19 Pattern matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . jim) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Parsing precedence . . . . . . . . . . . . . . . . . . . . . . . . . REs and Control Structures . . . . . . The graph for succ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Effect of cuts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . REs and Data Structures . . . . . . . . . . Class BTree . . . . . . . . . . . . . . . . . . The fringe of a tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Operator precedence . . . . . . . . vi 48 50 50 60 61 63 64 68 69 108 109 162 163 165 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Standard Class Inheritance Graph . . . . . . . . . . . . . . . . . . . . . Trying out type Frac . Proof of ancestor(pam. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . The Class Frac . . . . . . . . . . . . . . . . . . . . . . Key-value tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Tcl/Tk. Another factor that we should consider in studying PLs is their ubiquity: PLs are everywhere. C++. if coding requires 15% of total project time. The following scenario has occurred often in the history of programming. Estimates of maintenance time range from “at last 50%” (Lientz and Swanson 1980) to “between 65% and 75%” (McKee 1984) of the total time. Javascript. even if they had principles that they could use to help them. input for complex applications. Consequently. . however. . long-since rejected programming languages. it should help us to find useful answers to questions such as these. we could deduce that reducing coding time to zero would reduce the total project time by only 15%. and to predict how they might develop in the future. COBOL. . . . Dennis Ritchie. For example. Perl. Python. the notation develops into a programming language with many of the bad features of old. and we try to avoid unproductive “language wars”. PLs no longer seem so central to Computer Science (CS). . We can summarize our goals as follows: if the course is successful. these figures suggest that the PL might have a significant effect on total development time. . . . . . By the time that the developers realize this.1 How important are programming languages? As the focus of practitioners and researchers has moved from programming to large-scale software development. The assumption behind this reasoning is that the choice of PL affects only the coding phase of the project. but the main focus of the course is on contemporary and evolving PLs. The “obvious” PLs are the major implementation languages: Ada. we need to know something about how they evolved. because coding occupies only a small fraction of the total time required for project development. Since the choice of PL can influence the ease of maintenance. We not discuss details of syntax. is widely acknowledged to be a major component of software development. macro languages. PLs for the World Wide Web: Java. We consider early languages. . What is a good programming language? How should we choose an appropriate language for a particular task? How should we approach the problem of designing a programming language? 1. The format increases in complexity until it becomes a miniature programming language. Here is a useful epigram for the course: C and Pascal are the same language. . Brooks (1987) and others have argued that. it is too late for them to redesign the format. Maintenance. Scripting PLs: Javascript. “Hidden” languages: spreadsheets.COMP 348 1 Introduction Principles of Programming Languages In order to understand why programming languages (PLs) are as they are today. The treatment is fairly abstract. FORTRAN. . 1 . . Developers realize that an application requires a format for expressing input data. developments in programming languages cannot lead to dramatic increases in software productivity. .

. For example. we consider Haskell in some depth. . C and C++ are both widely-used languages but. a rounded work of art . 1. . . we can recognize the mistakes and omissions of language designers and criticize their PLs accordingly.1 INTRODUCTION 2 There is an unfortunate tendency in Computer Science to re-invent language features without carefully studying previous work. but the coverage is very uneven. On the contrary. Algol deserves our affection and appreciation. Given the state of the art at the time when they designed the PLs. However. Rarely has a construction so useful and elegant emerged as the output of a committee of 13 meeting for about 8 days . We conclude that PLs are important. pages 88–91) in saying: [Algol 60] was more racehorse than camel . we attempt to focus on the deep and influential ideas that have developed with PLs. however. Exercises are provided mainly to provoke thought. semantic details. These criticisms do not imply that the designers were stupid or incompetent. although safe and provably correct methods of programming concurrent processes were developed by himself and Hoare in the mid-seventies. Many references are provided. . . Instead. To provide insight into functional programming. and yet [Algol 60] was a noble begin but never intended to be a satisfactory end. or legalistic issues. . The time devoted to a PL depends on its relative importance in the evolution of PLs rather than the extent to which it was used or is used. if you are interested in a particular aspect of the course. With hindsight. we are better prepared to understand and appreciate its solution later. We do not dwell on fine lexical or syntactic distinctions. they are discussed briefly. If we spend time to think about a problem. In many cases. The intention is not that you read them all but rather that. We can echo Perlis (1978. the designers could hardly be expected to foresee the understanding that would come only with another twenty or thirty years of experience. sometimes with unfortunate consequences. the solution to an exercise can be found (implicitly!) elsewhere in these notes. their importance is often under-estimated. Some of the exercises may appear in revised form in the examination at the end of the course. Java does not make use of these methods and is insecure. since they did not contribute profound ideas. we have chosen to study PLs whose designers had deep insights and made profound contributions. The approach of the course is conceptual rather than technical. . you should be able to find out more about it by reading the indicated material.2 Features of the Course We will review a number of PLs during the course. . . Brinch Hansen (1999) points out that.

In Pascal. In C. In PLs such as C and Pascal. we can use a function call in place of E but not in place of v. One way is to look at some anomalies: programs that look as if they ought to work but don’t. or programs that look as if they shouldn’t work but do. for example. but a function cannot return an array. In other PLs. to replace the top element of a stack. Yet PLs treat data structures and files in completely different ways. What other changes would you have to make to the language? 2 Exercise 3. Suppose you added types as parameters to a language. In any case. We access a component of a by writing a[i. Example 1: Assignments.f. Explain why most PLs treat data structures and files differently. The variable v is not necessarily a simple name: it may be an expression such as a[i] or r.j] (except in C. t a[]) {. it yields an address into which the value of E is stored. PLs are fussy about the kinds of objects that can be passed as parameters to a function or returned as results by a function. For example. such as Algol 60. and a function cannot return an array. The important difference is the more or less accidental feature that data structures usually live in memory and files usually live in some kind of external storage medium. such as a disk. arrays can be passed by value or by reference to a function. 2 Example 4: Arrays.} and then call sort(float. there is not a great deal of difference between a data structure and a file. In some PLs. .j]? 3 . where we have to write a[i][j]). . even ones that are older than C and Pascal. It would be useful. Explain why the asymmetry mentioned in Example 1 exists. Conceptually. but only by reference. 2 Exercise 2. Can you see any problems that might be encountered in implementing a function like top? 2 Example 2: Parameters and Results. Is this equivalent to having types as parameters? 2 Example 3: Data Structures and Files. scores). Most languages do not allow types to be passed as parameters. Most PLs provide an assignment statement of the form v := E in which v is a variable and E is an expression. What difference does this make to the implementor of the PL and the programmers who use it? Suppose that we have a two-dimensional array a.2 Anomalies There are many ways of motivating the study of PLs. 2 Exercise 4. C++ provides templates. the size of an array is fixed at compile-time. 2 Exercise 1. We can access a row of the array by writing a[i]. to write a procedure such as void sort (type t. The expression E yields a value. we cannot write a statement such as top(stack) = 6. the size of an array is not fixed until run-time. . an array can be passed to a function. Why can’t we access a column by writing a[.

although experienced C programmers have ways of avoiding the problem. Some PLs. Why is there usually no provision for triangular. symmetric. Suggest some applications for records with both data and function components. such as C. but it cannot contain function declarations. . Both languages also provide a data aggregation — record in Pascal and struct in C. char *money (int amount) { char buffer[16]. such as Pascal. or sparse arrays? 2 Exercise 5. amount % 100). whereas others. 2 Example 5: Data Structures and Functions. For example. do not. Suggest ways of declaring arrays of the kinds mentioned in Example 4. and we defined the function shown in Listing 1. we are accustomed to factoring expressions. In algebra. Pascal and C programs contain declarations for functions and data. as z = (A + B) * x. 2 Example 6: Memory Management. Most PLs allow us to write something like this: if (x > PI) y = sin(x) else y = cos(x) Only a few PLs allow the shorter form: y = if (x > PI) then sin(x) else cos(x). allow nested functions. Sometimes this is possible in PLs: we can certainly rewrite z = A * x + B * x. A record looks rather like a small program: it contains data declarations.2 ANOMALIES 4 It is easy to declare rectangular arrays in most PLs. amount / 100. 2 Exercise 7. it remains a defect of both C and C++. Very few PLs recognize the even shorter form: y = (if (x > PI) then sin else cos)(x). char *)? 2 Example 7: Factoring. banded. What is wrong with the function money in Example 6? Why is it difficult for a C function to return a string (i. } Writing C functions that return strings is awkward and.e. sprintf(buffer.. Most C/C++ programmers learn quite quickly that it is not a good idea to write functions like this one. 2 Exercise 6. return buffer. when we see the expression Ax + Bx we are strongly tempted to write it in the form (A + B)x.%d". 2 Example 8: Returning Functions. Suppose that we were allowed to nest functions in C. "%d.

3. add6(4)). i++) a[i-a] < a[i] and ∃ i • 0 ≤ i < N ∧ a[i] = k . which should print “10”: int add6 (int) = addk(6).n) assigns 5 to n. Example 8 is just one of several oddities about the way programs handle functions. intint addk (int k) { int f (int n) { return n + k. f(x). } return f. Add(i. } 5 The idea is that addk returns a function that “adds k to its argument”.5) assigns 2 to i.3. For instance: Add(2.2 ANOMALIES Listing 1: Returning a function typedef int intint(int).z) attempts to satisfy the constraint x + y = z. i < N. printf("%d". “It would be fairly easy to change C so that a function could return a function. Some PLs do not permit this sequence: if (x > PI) f = sin else f = cos. Why are statements like this permitted in C (with appropriate syntax) but not in Pascal? 2 Example 10: Functions that work in more than one way. 2 Example 11: Logic computation. 2 Exercise 9. The function Add(x. 2 Exercise 8. It would be convenient if we could encode logical assertions such as ∀ i .y. We would use it as in the following program.) 2 Example 9: Functions as Values.” True or false? (Note that you can represent the function itself by a pointer to its first instruction. 0 < i < N ⇒ a[i − 1] < a[i] in the form all (int i = 1.

When type-checking was improved. functions like this actually worked. Garbage collection provides an example. . or they may believe (because they have been told) that garbage collection has an unacceptably high overhead. } printf("%d". 6 in the form exists (int i = 0. fac(3. i < N. Working with low-level PLs leads to the belief that “this is all that computers can do”. Many programmers have struggled for years writing programs that require explicit deallocation of memory. It is less well-known that we can eliminate the recursion as well: see Listing 2. they would no longer compile. TYPE fac (int n. however.) These programmers may be unaware of the existence of PLs that provide garbage collection. Memory “leaks” are one of the most common forms of error in large programs and they can do serious damage. else return n * f(n-1. when the rocket was already in lunar orbit. fac) = 3 × 2 × fac(1. . Can you complete the definition of TYPE in Example 2? 2 The point of these examples is to show that PLs affect the way we think. It would not occur to most programmers to write programs in the style of the preceding sections because most PLs do not permit these constructions. fac) = 3×2×1 = 6 In early versions of Pascal. fac) = 3 × fac(2. (One of the Apollo missions almost failed because of a memory leak that was corrected shortly before memory was exhausted. fac)).2 ANOMALIES Listing 2: The Factorial Function typedef TYPE . . This program computes factorials using neither loops nor recursion: fac(3. It is a well-known fact that we can rewrite programs that use loops as programs that use recursive functions. TYPE f) { if (n <= 1) return 1. . f). i++) a[i] == k 2 Example 12: Implicit Looping. 2 Exercise 10.

ideas that enter the mainstream and ideas that end up in a backwater. like the history of any technology. Sources for this section include (Wexelblat 1981. Williams and Campbell-Kelly 1989. in this course. under program control. There are advances and setbacks. Briefly. instead of by its address in the store. it enables the programmer to refer to any word in a programme by means of a label or tag attached to it arbitrarily by the programmer.3 The Procedural Paradigm The introduction of the von Neumann architecture was a crucial step in the development of electronic computers. is complex. In fact. (Mutch and Gill 1954) The key point in this quotation is the phrase “instead of by its address in the store”. 3. The second. for example. The first programs used numbers to refer to machine addresses. Thus the stored-program concept led to the development of programming “languages”. even ideas that are submerged for a while and later surface in an unexpected place. One of the first additions to programming notation was the use of symbolic names rather than numbers to represent addresses. Instead of writing Location 100 101 102 103 104 105 the programmer would write Order A 104 A 2 T 104 H 24 C 50 T 104 7 . we can identify several strands in the evolution of PLs. consequence is that computers can process programs themselves. The history of PLs. The programmer could then write simply ‘A a3’ to denote the operation of adding this number into the accumulator. The first consequence of this idea is that changing and modifying the stored program is simple and efficient. The basic idea is that instructions can be encoded as data and stored in the memory of the computer. without having to specify just where the number is located in the machine. a very different situation from earlier computers that were programmed by plugging wires into panels. These strands are commonly called “paradigms” and. Thus.1 Early Days The first PLs evolved from machine code. changes can take place at electronic speeds. a number appearing in the calculation might be labelled ‘a3’. With the benefit of hindsight. and ultimately more far-reaching. Bergin and Gibson 1996). we survey the paradigms separately although their development was interleaved. a computer can translate a program from one notation to another. In particular.

Another difficulty arises from the fact that. . parameters which vary during the solution of the problem. Find in it the origins of concern for: correctness. This quotation introduces a theme that has continued with variations to the present day. Subroutines may be divided into two types. EXTERNAL parameters. . space/time trade-offs. . maintenance. parameterization. The importance and subroutines and subroutine libraries was recognized before high-level programming languages had been developed. 3.3 THE PROCEDURAL PARADIGM A A T H C T a3 2 a3 24 50 a3 8 a3) systematically replacing the address 104 by the symbol a3 and omitting the explicit addresses. The following advantages arise from the use of such a library: 1. although it is desirable to have subroutines available to cover all possible requirements. it is also undesirable to allow the size of the resulting library to increase unduly. It enables routines to be more readily understood by other users. . . Library subroutines may be used directly. and making it much easier to locate errors. which we have called OPEN and CLOSED. However. This establishes the principle that a variable name stands for a memory location. as the following quotation shows. . 4. . It simplifies the task of preparing problems for the machine. thus reducing the total size of the library. parameters which are fixed throughout the solution of a problem and arise solely from the use of the library. genericity. thus greatly reducing the overall chance of error in a complete routine. being subroutines instead of individual orders. reuse. a principle that influenced the subsequent development of PLs and is now known — perhaps inappropriately — as value semantics.e. i. INTERNAL parameters. . We may divide the parameters associated with subroutines into two classes. Library subroutines are known to be correct. as conventions are standardized and the units of a routine are much larger. encapsulation. (Wheeler 1951) Exercise 11. without detailed coding and punching.e. a subroutine can be made more versatile by the use of parameters associated with it. . 2. i. An open subroutine is one which is included in the routine as it stands whereas a closed subroutine is placed in an arbitrary position in the store and can be called into use by any part of the main routine.

it was clear that large benefits could be achieved. The major achievements of FORTRAN are: efficient compilation. the individual steps of procedural PLs. FORTRAN. which transferred control to another part of the program. are often called “statements”. will comprise a large set of programs to enable the IBM 704 to accept a concise formulation of a problem in terms of a mathematical notation and to produce automatically a high-speed 704 program for the solution of the problem. In these notes. (Backus 1957) It is interesting to note that. The data structures of these early languages were usually rather simple: typically primitive values (integers and floats) were provided. was very similar to the solution he advocated in 1957 — programming must become more like mathematics: “we should be focusing on the form and content of the overall result”. Although FORTRAN did not eliminate programming. the commands of a procedural program are executed sequentially.2 FORTRAN FORTRAN was introduced in 1957 at IBM by a team led by John Backus. The designers focused on efficient implementation rather than elegant language design. FORTRAN has value semantics. He saw the assignment statement as a source of inefficiency: “the von Neumann bottleneck”. By default.) This suggests that the IBM team’s goal was to eliminate programming! The following quotation seems to confirm this: If it were possible for the 704 to code problems for itself and produce as good programs as human coders (but without the errors). Confusingly. The solution. The style of programming that this viewpoint developed became known as the “imperative” or “procedural” programming paradigm. Backus (1978) criticized FORTRAN and similar languages as “lacking useful mathematical properties”. The “Preliminary Report” describes the goal of the FORTRAN project: The IBM Mathematical Formula Translation System or briefly. which transferred control but also stored a “link” to which control would be returned after executing a subroutine. non-technical sense) or recipes rather than “commands”. we use the term “procedural” rather than “imperative” because programs resemble “procedures” (in the English. however. although in logic a “statement” is a sentence that is either true or false. such as Pascal and C. 20 years later. and the “jump and store link” command. it was a major step towards the elimination of assembly language coding. (Quoted in (Sammet 1969). knowing that acceptance depended on the high performance of compiled programs. . Procedural PLs provide various ways of escaping from the sequence.and multi-dimensioned arrays of primitive values. Variable names stand for memory addresses that are determined when the program is loaded. along with single.3 THE PROCEDURAL PARADIGM 2 9 Machine code is a sequence of “orders” or “instructions” that the computer is expected to execute. The earliest mechanisms were the “jump” command. 3.

Limited control structures. is feasible. This enables different parts of a program to share regions of memory. The goal was a “universal programming language”.3 THE PROCEDURAL PARADIGM 10 separate compilation (programs can be presented to the compiler as separate subroutines. all designed for the IBM 704. The principal limitations of FORTRAN are: Flat. In another sense. Algol was a failure: few complete. but the compiler does not check for consistent usage of these regions. FORTRAN also provides the EQUIVALENCE statement. FORTRAN. including the parameters and local variables of subroutines. A program consists of a sequence of subroutines and a main program. In other words. Naur 1978. labels provide the only indication that a sequence of statements form a group. DO. IBM dominated. The control structures of FORTRAN are IF. high-quality compilers were written and the language was not widely used (although it was used more in Europe than in North America). Variables are either global or local to subroutines. with COBOL. Exercise 12. One program component might use a region of memory to store an array of integers. which allows variables with different names and types to share a region of memory. 1960. partly to provide a PL that was independent of any particular company and its computers. FORTRAN borrows the concept of COMMON storage from assembly language program. and GOTO. uniform structure. How would you interpret this statement if you were: writing a FORTRAN program? writing a FORTRAN compiler? 2 3. FORTRAN programs are rather similar to assembly language programs: the main difference is that a typical line of FORTRAN describes evaluating an expression and storing its value in memory whereas a typical line of assembly language specifies a machine instruction (or a small group of instructions in the case of a macro). demonstration that high-level programming. and another might assume that the same region contains reals. most of the development of PLs was coming from industry. and FLPL (FORTRAN List Processing Language). Algol was a huge success: it became the . with automatic translation to machine code. FORTRAN allocates all data. Unsafe memory allocation.3 Algol 60 During the late fifties. Algol 60 (Naur et al. Recursion is forbidden because only one instance of a subroutine can be active at one time. but the compiler does not check for consistency between components). Since there are no compound statements. No recursion. In one sense. statically. To conserve precious memory. The committee included both John Backus (chief designer of FORTRAN) and John McCarthy (designer of LISP). The FORTRAN 1966 Standard stated that a FORTRAN implementation may allow recursion but is not required to do so. Perlis 1978) was designed by an international committee. There is no concept of nesting in FORTRAN.

end.. begin function f(x) begin . In the Algol block shown in Listing 3. one kind of statement is a block. Despite the simplicity of the implementation. The following code works fine in Algol 60. the ACM required submissions to the algorithm collection to be written in Algol. successor PLs such as C and Pascal dropped this useful feature. standard language for describing algorithms.14159. The AR is created on entry to the block and destroyed after the statements of the block have been executed. The designers of Algol realized that it was relatively simple to allow the size of an array to be determined at run-time. real y. Algol programs are recursively structured. in particular. The compiler statically allocates space for a pointer and an integer (collectively called a “dope vector”) on the stack. The essence of the copy rule is that the program behaves as if the text of the formal parameter in the function is replaced by the text of the actual parameter. The recursive structure of programs means that large programs can be constructed from small programs. the two assignments to x refer to two different variables.. . . procedure average (n). x := 2. The complications arise because it may be necessary to rename some of the variables during the textual substitution. this in turn means that ARs can be allocated on a stack . begin real array a[1:n]. when the size of the array is known. Dynamic Arrays. the appropriate amount of space is allocated on the stack and the components of the “dope vector” are initialized. The run-time entity corresponding to a block is called an activation record (AR). A program is a block . The default method of passing parameters in Algol was “call by name” and it was described by a rather complicated “copy rule”. integer x.. A block consists of declarations and statements. integer n... The syntax of Algol ensures that blocks are fully nested. x := 1.3 THE PROCEDURAL PARADIGM Listing 3: An Algol Block begin integer x. The major innovations of Algol are discussed below. Block structure and stacked ARs have been incorporated into almost every language since Algol. y := 3. end. Block Structure. At run-time. The usual implementation strategy was to translate the actual parameter into a procedure with no arguments (called a “thunk”). A variable or function name declared in a block can be accessed only within the block: thus Algol introduced nested scopes. There are various kinds of statement. Call By Name. For the better part of 30 years. end 11 end.

Call by name provides control structure abstraction.0/x).) For example. the Algol committee had several valid reasons for introducing it. i. if we declare the procedure try as in Listing 6. integer max. The first parameter specifies the number of iterations. i. sum := s end 12 each occurrence of the formal parameter inside the function was translated into a call to this function. (This is in contrast with call by value. and the third is the loop body. i. Call by name evaluates the actual parameter exactly as often as it is accessed. the expression 1. for i := 1 until n do s := s + val. integer n. call by value. The procedure in Listing 5 provides a form of abstraction of a for loop. if x ≤ 0. begin n := n + 1 end Listing 5: A General Sum Function integer procedure sum (max. does not allow a procedure to alter the value of its actual parameters in the calling environment: the parameter behaves like an initialized local variable.0/x will not be evaluated. s := 0. Own Variables. If the procedure count is defined as in Listing 4. begin integer s. The statement sum(3. the statement count(widgets) has the same effect as the statement begin widgets := widgets + 1 end The other parameter passing mechanism provided by Algol. A variable in an Algol procedure can be declared own. on entry to the procedure. Call by name enables procedures to alter their actual parameters. The effect is that the variable has local scope (it can be accessed only by the statements within the procedure) but global extent (its lifetime is the execution of the entire program). However.3 THE PROCEDURAL PARADIGM Listing 4: Call by name procedure count (n). Exercise 13. the second is the loop index. Why does i appear in the parameter list of Sum? 2 . The mechanism seems strange today because few modern languages use it. 1. because. a[i]) computes a[1]+a[2]+a[3]. val. val). where the parameter is usually evaluated exactly once. it is safe to call try(x > 0.

Before discussing these. . Applying this idea consistently throughout the language would have led to high order functions and paved the way to functional programming. Discuss the initialization of own variables. a record. however. begin try := if b then x else 0.3 THE PROCEDURAL PARADIGM Listing 6: Using call by name real procedure try (b. x). anything that jeopardized it was not acceptable. But the concept was important: it is the separation of scope and extent that ultimately leads to objects. but not quite powerful enough. Own variables were in fact rather problematic in Algol. With hindsight. Algol would be a language with modules. Yet Algol 60 does not provide records. A variable name stands for a memory addresses that is determined when the block containing the variable declaration is entered at run-time. An Algol block consists of declarations followed by statements. is actually a parameterless procedure. In particular. Suppose that declarations and statements could be interleaved in a block. like FORTRAN. The actual parameter in an Algol call. The local data of an AR is destroyed after the statements of the AR have been executed. for various reasons including the difficulty of reliably initializing them (see Example 14).0 end 13 Exercise 14. as mentioned above. If the data was retained rather than destroyed. since they believed that the stack discipline obtained with nested blocks was crucial for efficiency. in effect. 2 Algol 60 and most of its successors. D denotes a sequence of declarations and S denotes a sequence of statements. In the following block. They knew that incorporating the “missed opportunities” described above would have led to significant implementation problems. The Algol committee knew what they were doing. assuming the default calling mechanism. begin D1 S1 D2 S2 end A natural interpretation would be that S1 and S2 are executed concurrently . An Algol block without statements is. Algol 60 was simple and powerful. The call by name mechanism was a first step towards the important idea that functions can be treated as values. we can see that Algol made important contributions but also missed some very interesting opportunities. has a value semantics. The dominant trend after Algol was towards languages of increasing complexity. such as PL/I and Algol 68. boolean b. we take a brief look at COBOL. real x.

$$9. but not necessarily. The problem of type conversion had not arisen previously because only a small number of types were provided by the PL. SALREP PICTURE $$$. The assignment statement in COBOL has several forms. to PL/I) that would combine the best features of these three languages.99 The first indicates that SALARY is to be stored in a form suitable for computation (probably. (Only one dollar symbol will be printed.4 COBOL COBOL (Sammett 1978) introduced structured data and implicit type conversion. The choice made by the designers of COBOL was radical: type conversion should be automatic. Can you explain why? 2 3. the dominant languages were Algol. including MOVE X to Y. When COBOL was introduced. COBOL. Insiders at the time referred to the new language as “CobAlgoltran”. binary form) and the second provides a format for reporting salaries as amounts in dollars. If X and Y have different types. The design principles of PL/I (Radin 1978) included: the language should contain the features necessary for all kinds of programming.3 THE PROCEDURAL PARADIGM 14 3. a single statement could generate a large amount of machine code. it remains true that FORTRAN is widely used for “number crunching” and COBOL is widely used for data processing. the COBOL compiler will attempt to find a conversion from one type to the other. a programmer could learn a subset of the language. immediately before the first significant digit). The data division of a COBOL program contained descriptions of the data to be processed. might contain these declarations: 77 77 The Data Division of a COBOL program SALARY PICTURE 99999. suitable for a particular application. USAGE IS COMPUTATIONAL. with appropriate formatting. COBOL introduced many new types. 2 Exercise 15. COBOL introduced “data processing”. “programming” was more or less synonymous with “numerical computation”. Example 13: Automatic conversion in COBOL. The Procedure Division would probably contain a statement like MOVE SALARY TO SALREP. In COBOL.5 PL/I During the early 60s. which implicitly requires the conversion from binary to character form. without having to learn the entire language. Despite significant advances in the design and implementation of PLs. The continuing desire for a “universal language” that would be applicable to a wide variety of problem domains led IBM to propose a new programming language (originally called NPL but changed. In most PLs of the time. after objections from the UK’s National Physical Laboratory. FORTRAN. in the sense that data could have various degrees of precision. and different representations as text. . where data meant large numbers of characters. Another important innovation of COBOL was a new approach to data types. a single statement translated into a small number of machine instructions.

but their existence encouraged others to produce better designs. Some of these were later incorporated into C. Exercise 16. however. if the handler contains a GOTO statement. Add 17 to 578. control is transferred to whichever ON statement for that condition was most recently executed. the compiler will detect the error and provide a diagnostic message that is incomprehensible to the programmer because it refers to a part of the language outside the learned subset. as far as possible. form of exception handling. If the condition (which might be OVERFLOW. 2. like all programmers. might be paraphrased as: “Do everything possible to compile this statement. again because it is outside the learned subset. control returns to the statement that raised the exception or. Convert the integer 8 to the string ’8’. to the target of that statement. on encountering an assignment x = E. Since we can execute the statement allocate x as often as necessary. After the statements between BEGIN and END (the handler) have been executed. PL/I did introduce some important new features into PLs. to make a mistake. 4. the expression (Gelernter and Jagannathan 1990) (’57’ || 8) + 17 is evaluated as follows: 1. Concatenate the strings ’57’ and ’8’.) becomes TRUE. For example.. They were not all well-designed. Discuss potential problems of the PL/I exception handling mechanism. The compiler’s policy. . the compiler will not detect the error and the program will behave in a way that is inexplicable to the programmer. could not be named. Every variable has a storage class: static. Convert the integer 595 to the string ’595’. Statements of the following form are allowed anywhere in the program: ON condition BEGIN. avoid issuing any diagnostic message that would tell the programmer what is happening”. END. PL/I provided a simple. An object associated with a based variable x requires explicit allocation and is placed on the heap rather than the stack. obtaining ’578’. PL/I extends the automatic type conversion facilities of COBOL to an extreme degree. etc. automatic.3 THE PROCEDURAL PARADIGM 15 An important lesson of PL/I is that these design goals are doomed to failure.. and not very safe. based variables provide a form of template.. Convert the string ’578’ to the integer 578. based. 5. or controlled. Types. obtaining 595. 2 . PRINTER OUT OF PAPER. 3. PL/I provides a wide range of programmer-defined types. A programmer who has learned a “subset” of PL/I is likely. With luck. More probably.

). Like Algol 60. . dynamic data structures. Even the priority of these operators can be altered. and other features to the language. . the argument list is a collateral clause. In a function call f (x. constants. perhaps. .. the expressions x. Algol 68 was not widely used. Pascal demonstrated that a PL could be simple yet powerful. Wirth made extensive use of the ideas of Dijkstra and Hoare (later published as (Dahl. Dijkstra.7 Pascal Pascal was designed by Wirth (1996) as a reaction to the complexity of Algol 68. 1975. z). it does not necessarily follow that orthogonality is always a good thing. “use the address rather than the value”. and z can be evaluated in any order. for all these entities. z). variables. The important contributions of Pascal included the following. for an assignment x := E. to optimize. The main design principle of Algol 68 was orthogonality : the language was to be defined using a number of basic concepts that could be combined in arbitrary ways. weak context. To what extent can it be checked by the compiler? 2 3. . PL/I.) and mechanisms for building structured types (array. file. some of which have become part of the culture (cast. and other languages that were becoming popular in the late 60s.). . Although it is true that lack of orthogonality can be a nuisance in PLs. Operator overloading: programmers can provide new definitions for standard operators such as “+”. The ideas that Algol 68 introduced. bool. The language was described in a formal notation that specified the complete syntax and semantics of the language (van Wijngaarden et al. y. 1975). y. record. its successor Algol 68 (van Wijngaarden et al. . however. set. have been widely imitated. The type system of Pascal was based on primitives (integer. Lindsey 1996) is much more complex. .) and some of which have not (mode. narrowing. . Collateral clauses provide a good. Exercise 17. This single keyword introduces call by reference. It appears in C in the form of the operators “*” and “&”. and functions. pointers. y. the Algol 68 report does not specify the order of evaluation of the expressions in a collateral clause. Thus data types in Pascal form a recursive hierarchy just as blocks do in Algol 60. coercion. the lifetime of the variable x must be less than or equal to the lifetime of the object obtained by evaluating E. example of the idea that a PL specification should intentionally leave some implementation details undefined. For example. . In a collateral clause of the form (x. . real. . Algol 68 has a very uniform notation for declarations and other entities. . . A large vocabulary of PL terms. . Explain the motivation for this rule. and Hoare 1972)). Algol 68 uses the same syntax (mode name = expression) for types. The fact that the Report was very hard to understand may have contributed to the slow acceptance of the language. there must be forms of expression that yield appropriate values. The important features introduced by Algol 68 include the following. roughly. This gives the implementor freedom to use any order of evaluation and hence. or concurrently. although it was popular for a while in various parts of Europe. voiding. The operator ref stands for “reference” and means. This implies that. . In this example. and early. . especially Hoare’s ideas of data structuring. Algol 68 has a rule that requires.6 Algol 68 Whereas Algol 60 is a simple and expressive language.3 THE PROCEDURAL PARADIGM 16 3.

supersets were developed and. Because of the perceived missing features. How serious do you think this problem was? 2 3. removes Pascal’s weaknesses. Pascal is a kind of “fill in the blanks” language in which all programs have a similar structure. This implies that the size must be deducible from the interface which implies. The first version of “Standard Pascal” was almost useless as a practical programming language and the Revised Standard described a usable language but appeared only after most people had lost interest in Pascal. The record type was a useful innovation (although very similar to the Algol 68 struct) but allowed data only. the introduction of modules. Allowing functions in a record declaration would have paved the way to modular and even object oriented programming. Pascal had a strong influence on many later languages. each of which makes certain decisions and adds new details. All other type conversions are explicit (even when no action is required) and the compiler checks type correctness. Like Algol 60. in turn. to some extent. It is well-known that the biggest loop-hole in Pascal’s type structure was the variant record. For example.8 Modula–2 Wirth (1982) followed Pascal with Modula–2. which inherits Pascal’s strengths and. of course. Pascal missed important opportunities. was never completed. Nevertheless. Modula–2 was the product of a sabbatical year in California. The implementation contains the “secret” information about the module. where Wirth worked with the designers of Mesa. List some of the “missing features” of Pascal. The monolithic structure that this idea imposes on programs is a drawback of Pascal because it prevents independent compilation of components.) Modula–2 provides a limited escape from this dilemma: a programmer can define an “opaque” type with a hidden representation. Coroutines. The important contribution of Modula–2 was. The important features of Modula–2 are: Modules with separated interface and implementation descriptions (based on Mesa). Modula. In this case.3 THE PROCEDURAL PARADIGM 17 Pascal provides no implicit type conversions other than subrange to integer and integer to real. (Wirth’s first design. This design has the unfortunate consequence that some information that should be secret must be put into the interface. that the interface must contain the representation of the object. these became incompatible. inevitably. the interface contains only a pointer to the instance and the representation can be placed in the implementation module. data type declarations. The interface provides information about the use of the module to both the programmer and the compiler. Exercise 18.) A module in Modula–2 has an interface and an implementation. determined by the relatively strict syntax. 2 Exercise 19. another early modular language. the compiler must know the size of the object in order to declare an instance of it. (The same problem appears again in C++. Its most important innovations were probably the combination of simplicity. and static type checking. . Pascal was designed to match Wirth’s (1971) ideas of program development by stepwise refinement. Programmers are expected to start with a complete but skeletal “program” and flesh it out in a series of refinement steps. Pascal was a failure because it was too simple.

record types. arrays are defined in terms of pointers and pointer arithmetic.. It is not clear why four distinct mechanisms are required (Gelernter and Jagannathan 1990).10 Ada Ada (Whitaker 1996) represents the last major effort in procedural language design. but this aspect of the language proved hard to implement. For example. and packages and tasks (not representable in the language). type element is private.. the template generic max: integer. C is a low-level language by comparison with the other PLs discussed in this section. A procedure definition looks like this: procedure procname ( parameters ) is body A record type looks like this: type recordtype ( parameters ) is body The parameters of a record type are optional. Since UNIX depended heavily on C.3 THE PROCEDURAL PARADIGM 18 3. package Stack is . The corresponding objects are: blocks and records (representable in the language). 3. strings. might be instantiated by a declaration such as package intStack is new Stack(20. The syntactic differences suggest that the designers did not look for similarities between these constructs. The enormous success of C is partly accidental. It was the first widely-used language to provide full support for concurrency.. This is both the strength and weakness of C. For example. but C does not provide real support for arrays. += and other assignment operators) and others are unique to C and C++ (for example. If present. or boolean operations. integer) . It is a large and complex language that combines then-known programming features with little attempt at consolidation. with good reason. and task types. The number of concepts is small. The compiler is assumed to be rather unsophisticated (a reasonable assumption for a compiler running on a PDP/11 in the late sixties) and in need of hints such as register. after Bell released it to universities. with interactions checked by the compiler. C is based on a small number of primitive concepts. UNIX. A generic package looks like this: generic ( parameters ) package packagename is package description The parameters can be types or values. Ritchie (Ritchie 1996) designed it for a particular task — systems programming — for which it has been widely used. C is notable for its concise syntax. the spread of UNIX inevitably led to the spread of C. It is designed to be easy to compile and to produce efficient object code. postfix and prefix ++ and --). generic packages. Ada provides templates for procedures.9 C C is a very pragmatic PL. Some syntactic features are inherited from Algol 68 (for example. became popular. they have a different form than the parameters of procedures.

3 THE PROCEDURAL PARADIGM Finally. 2 . Exercise 20. Changing the syntax would be a minor improvement. a task template looks like this (no parameters are allowed): task type templatename is task description 19 Of course. just as the identity declaration of Algol 68 suggested new and interesting possibilities. But it is disturbing that the language designers apparently did not consider passible relationships between these four kinds of declaration. Propose a uniform style for Ada declarations. programmers hardly notice syntactic differences of this kind: they learn the correct incantation and recite it without thinking. but uncovering deep semantic similarities might have a significant impact on the language as a whole.

4

The Functional Paradigm

Procedural programming is based on instructions (“do something”) but, inevitably, procedural PLs also provide expressions (“calculate something”). The key insight of functional programming (FP) is that everything can be done with expressions: the commands are unnecessary. This point of view has a solid foundation in theory. Turing (1936) introduced an abstract model of “programming”, now known as the Turing machine. Kleene (1936) and Church (1941) introduced the theory of recursive functions. The two theories were later shown (by Kleene) to be equivalent: each had the same computational power. Other theories, such as Post production systems, were shown to have the same power. This important theoretical result shows that FP is not a complete waste of time but it does not tell us whether FP is useful or practical. To decide that, we must look at the functional programming languages (FPLs) that have actually been implemented. Most functional language support high order functions. Roughly, a high order function is a function that takes another function as a parameter or returns a function. More precisely: A zeroth order expression contains only variables and constants. A first order expression may also contain function invocations, but the results and parameters of functions are variables and constants (that is, zeroth order expressions). In general, in an n-th order expression, the results and parameters of functions are (n − 1)-th order expressions. A high order expression is an n-th order expression with n ≥ 2. The same conventions apply in logic with “function” replaced by “function or predicate”. In first-order logic, quantifiers can bind variables only; in a high order logic, quantifiers can bind predicates.

4.1

LISP
Anyone could learn LISP in one day, except that if they already knew FORTRAN, it would take three days. Marvin Minsky

Functional programming was introduced in 1958 in the form of LISP by John McCarthy. The following account of the development of LISP is based on McCarthy’s (1978) history. The important early decisions in the design of LISP were: to provide list processing (which already existed in languages such as Information Processing Language (IPL) and FORTRAN List Processing Language (FLPL)); to use a prefix notation (emphasizing the operator rather than the operands of an expression); to use the concept of “function” as widely as possible (cons for list construction; car and cdr for extracting list components; cond for conditional, etc.); to provide higher order functions and hence a notation for functions (based on Church’s (1941) λ-notation); to avoid the need for explicit erasure of unused list structures. McCarthy (1960) wanted a language with a solid mathematical foundation and decided that recursive function theory was more appropriate for this purpose than the then-popular Turing machine model. He considered it important that LISP expressions should obey the usual mathematical laws allowing replacement of expressions and: 20

4 THE FUNCTIONAL PARADIGM
E s c s E s s E NIL

21

A

c s s c

E

s c

s

E NIL

B

C

Figure 1: The list structure (A (B C)) Another way to show that LISP was neater than Turing machines was to write a universal LISP function and show that it is briefer and more comprehensible than the description of a universal Turing machine. This was the LISP function eval [e, a], which computes the value of a LISP expression e, the second argument a being a list of assignments of values to variables. . . . Writing eval required inventing a notation for representing LISP functions as LISP data, and such a notation was devised for the purpose of the paper with no thought that it would be used to express LISP programs in practice. (McCarthy 1978) After the paper was written, McCarthy’s graduate student S. R. Russel noticed that eval could be used as an interpreter for LISP and hand-coded it, thereby producing the first LISP interpreter. Soon afterwards, Timothy Hart and Michael Levin wrote a LISP compiler in LISP; this is probably the first instance of a compiler written in the language that it compiled. The function application f (x, y) is written in LISP as (f x y). The function name always comes first: a + b is written in LISP as (+ a b). All expressions are enclosed in parentheses and can be nested to arbitrary depth. There is a simple relationship between the text of an expression and its representation in memory. An atom is a simple object such as a name or a number. A list is a data structure composed of cons-cells (so called because they are constructed by the function cons); each cons-cell has two pointers and each pointer points either to another cons-cell or to an atom. Figure 1 shows the list structure corresponding to the expression (A (B C)). Each box represents a cons-cell. There are two lists, each with two elements, and each terminated with NIL. The diagram is simplified in that the atoms A, B, C, and NIL would themselves be list structures in an actual LISP system. The function cons constructs a list from its head and tail: (cons head tail). The value of (car list) is the head of the list and the value of (cdr list) is the tail of the list. Thus: (car (cons head tail)) → head (cdr (cons head tail)) → tail The names car and cdr originated in IBM 704 hardware; they are abbreviations for “contents of address register” (the top 18 bits of a 36-bit word) and “contents of decrement register” (the bottom 18 bits). It is easy to translate between list expressions and the corresponding data structures. There is a function eval (mentioned in the quotation above) that evaluates a stored list expression. Consequently, it is straightforward to build languages and systems “on top of” LISP and LISP is often used in this way.

4 THE FUNCTIONAL PARADIGM

22

It is interesting to note that the close relationship between code and data in LISP mimics the von Neumann architecture at a higher level of abstraction. LISP was the first in a long line of functional programming (FP) languages. Its principal contributions are listed below. Names. In procedural PLs, a name denotes a storage location (value semantics). In LISP, a name is a reference to an object, not a location (reference semantics). In the Algol sequence int n; n := 2; n := 3; the declaration int n; assigns a name to a location, or “box”, that can contain an integer. The next two statements put different values, first 2 then 3, into that box. In the LISP sequence (progn (setq x (car structure)) (setq x (cdr structure))) x becomes a reference first to (car structure) and then to (cdr structure). The two objects have different memory addresses. A consequence of the use of names as references to objects is that eventually there will be objects for which there are no references: these objects are “garbage” and must be automatically reclaimed if the interpreter is not to run out of memory. The alternative — requiring the programmer to explicitly deallocate old cells — would add considerable complexity to the task of writing LISP programs. Nevertheless, the decision to include automatic garbage collection (in 1958!) was courageous and influential. A PL in which variable names are references to objects in memory is said to have reference semantics. All FPLs and most OOPLs have reference semantics. Note that reference semantics is not the same as “pointers” in languages such as Pascal and C. A pointer variable stands for a location in memory and therefore has value semantics; it just so happens that the location is used to store the address of another object. Lambda. LISP uses “lambda expressions”, based on Church’s λ-calculus, to denote functions. For example, the function that squares its argument is written (lambda (x) (* x x)) by analogy to Church’s f = λx . x2 . We can apply a lambda expression to an argument to obtain the value of a function application. For example, the expression ((lambda (x) (* x x)) 4) yields the value 16. However, the lambda expression itself cannot be evaluated. Consequently, LISP had to resort to programming tricks to make higher order functions work. For example, if we want to pass the squaring function as an argument to another function, we must wrap it up in a “special form” called function: (f (function (lambda (x) (* x x))) . . . .) Similar complexities arise when a function returns another function as a result. Dynamic Scoping. Dynamic scoping was an “accidental” feature of LISP: it arose as a side-effect of the implementation of the look-up table for variable values used by the interpreter. The C-like program in Listing 7 illustrates the difference between static and dynamic scoping. In

It was not long before a form for defining functions was introduced (originally called define. void f () { printf("%d". f (). the value of the expression is obtained by taking the car of the cadr (that is. } void main () { int x = 7. contains expressions such as the one shown in Listing 8. written in LISP. but its internal state did not change. a program is the same list of functions. The initial environment is empty. On paper. How much can you learn about a language by reading an interpreter written in the language? What can you not learn? 2 . The LISP interpreter. A LISP program has no real structure. the LISP interpreter behaved as a calculator: it evaluated expressions entered by the user. Exercise 22. Some programs gave one answer when interpreted and another answer when compiled! Exercise 21. the environment contains the global binding for x: x = 4 . We might paraphrase this as: “if the car of the expression that we are currently evaluating is car. however.4 THE FUNCTIONAL PARADIGM Listing 7: Static and Dynamic Binding int x = 4. x = 4 . and the overhead of searching a linear list for a variable value just makes them slightly slower still. The interpreter then evaluates the call f (). At run-time. a program is a list of function definitions. later changed to defun) to enable users to add their own functions to the list of built-in functions. and forcing it to maintain a linear list would be unacceptably inefficient. Since the value of this variable is 4. } // 1 23 // 2 C. Describe a situation in which dynamic scoping is useful. The environment behaves like a stack (last in. 1990). LISP was the first major language to be interpreted. When the interpreter evaluates the function main. Although dynamic scoping is natural for an interpreter. It is a much larger and more complex language than the original LISP and includes many features of Scheme (described below). Interpreters are slow anyway. Common LISP provides static scoping with dynamic scoping as an option. The current dialect of LISP is called Common LISP (Steele et al. early LISP systems had an unfortunate discrepancy: the interpreters used dynamic scoping and the compilers used static scoping. (Do not confuse dynamic scoping with dynamic binding!) A LISP interpreter constructs its environment as it interprets. x). it uses the first value of x in the environment and prints 7. Originally. the program prints 4. has more efficient ways of accessing variables. the variable x in the body of the function f is a use of the global variable x defined in the first line of the program. which we denote by . Consequently. A compiler. translated into internal form. After interpreting the LISP equivalent of the line commented with “1”. the functions may invoke one another with either direct or indirect recursion. it inserts the local x into the environment. first out). it is inefficient for a compiler. obtaining x = 7. the second term) of the expression”. when it encounters x in the body of f . added to the interpreter. 2 Interpretation.

. cond is the conditional construct has the form (cond (p1 e1 ) (p2 e2 ) . 24 4.1 Basic LISP Functions The principal — and only — data structure in classical LISP is the list. Lists are written (a b c).. return e1. (t e)) end works like this: if p1 is true.. null is a predicate that returns true if its argument is the empty list.1. first (originally called car) returns the head (first component) of a list. . def is used to define functions: (def name (lambda parameters body )) 4. The following functions are provided for lists.. .1. The function pairs builds an environment from a list of names and a list of expressions. .2 A LISP Interpreter The following interpreter is based on McCarthy’s LISP 1. else return e... second (originally called cadr. The table is represented as a list of pairs. cons builds a list. atom is a predicate that return true if its argument is not a list — and must therefore be an “atom”. The important forms are: quote takes a single argument and does not evaluate it. if p2 is true.5 Programmer’s Manual by McCarthy et al. given its head and tail. that is. MIT Press 1962. . (def pairs (lambda (names exps env) (cond ((null names) env) (t (cons (cons (first name) (first exps)) (pairs (rest names) (rest exps) env) )) ) )) The function lookup finds the value of a name in a table. . An environment is a list of name/value pairs.4 THE FUNCTIONAL PARADIGM Listing 8: Defining car (cond ((eq (car expr) ’car) (car (cadr expr)) ) . Lists also represent programs: the list (f x y) stands for the function f applied to the arguments x and y. rest (originally called cdr) returns the tail of a list. A form looks like a function but does not evaluate its argument. return e2 .5 Evaluator from Lisp 1. short for car cdr) returns the second element of a list. a variable name or other primitive object.

and other built-in functions not considered here. it takes a list of pairs of the form (cnd exp) and returns the value of the expression corresponding to the first true cnd.4 THE FUNCTIONAL PARADIGM (def lookup (lambda (name table) (cond ((eq name (first (first table))) (rest (first table))) (t (lookup name (rest table))) ) )) 25 The heart of the interpreter is the function eval which evaluates an expression exp in an environment env. The cases it considers are: Built-in functions: first. (def apply (lambda (fun args env) (cond ((eq fun (quote first)) (first (first args))) ((eq fun (quote second)) (second (first args))) ((eq fun (quote cons)) (cons (first args) (second args))) ((atom fun) (apply (eval fun env) args env)) ((eq (first fun) (quote lambda)) (eval (third fun) (pairs (second fun) args env)) ) ) )) . second. It is a straightforward recursion. Any other function with an atomic name: this is assumed to be a user-defined function. A function which is a lambda form. and lookup is used to find the lambda form corresponding to the name. (def evcon (lambda (tests env) (cond ((null tests) nil) (t (cond ((eval (first (first tests)) env) (eval (second (first tests)) env)) (t (evcon (rest tests) env)) ) ) ) )) The function evlist evaluates a list of expressions in an environment. (def eval (lambda (exp env) (cond ((null exp) nil) ((atom exp) (lookup exp env)) ((eq (first exp) (quote quote)) (second exp)) ((eq (first exp) (quote cond)) (evcon (second exp) env) ) (t (apply (first exp) (evlist (rest exp) env) env)) ) )) The function evcon is used to evaluate cond forms. cons. (def evlist (lambda (exps env) (cond ((null exps) nil) (t (cons (eval (first exps) env) (evlist (rest exps) env) )) ) )) The function apply applies a function fun to arguments args in the environment env.

in 1978. Steele and Sussman implemented the actor model using a small LISP . Steele Jr. It contains an important defect that was not discovered for some time. and C++.2). Haskell. we effectively build a list of activation records). the dynamic binding problem was not discovered until LISP was well-entrenched: it would be almost 20 years before Guy Steele introduced Scheme. (The interpreter above does not incorporate error detection and would simply fail with this function. We use a simpler example to explain the problem. and so that is the value that show returns.) However. including Scheme. It is very similar to LISP in both syntax and semantics. Suppose we define: (def show (lambda () x )) Calling this function generates an error because x is not defined. we can wrap show inside another function: (def try (lambda (x) (show) )) We then find that (try t) evaluates to t. and Gerald Jay Sussman (1975).t) is in the environment. use static binding . In other words. The starting point of Scheme was an attempt by Steele and Sussman to understand Carl Hewitt’s theory of actors as a model of computation. the binding (x. However. When show is called. with static binding. The model was object oriented and influenced by Smalltalk (see Section 6.3 Dynamic Binding 26 The classical LISP interpreter was implemented while McCarthy was still designing the language. James Slagle defined a function like this (def testr (lambda (x p f u) (cond ((p x) (f x)) ((atom x) (u)) (t (testr (rest x) p f (lambda () (testr (first x) p f u)))) ) )) and it did not give the correct result. People wrote LISP compilers. Consequently.4 THE FUNCTIONAL PARADIGM 4.1. We say that LISP uses dynamic binding whereas most other languages. Correcting the LISP interpreter to provide static binding is not difficult: it requires a slightly more complicated data structure for environments (instead of having a simple list of name/value pairs. there were many LISP systems that provided dynamic binding during interpretation and static binding for compiled programs! 4.2 Scheme Scheme was designed by Guy L. but it is hard to write a compiler with dynamic binding. the value of a variable in LISP depends on the dynamic behaviour of the program (which functions have been called) rather than on the static text of the program. but it corrects some of the errors of LISP and is both simpler and more consistent.

as in Listing 10. The interpreter provided lexical scoping. but the form shown is accepted by the Scheme compiler and we use it in these notes. to spare programmers the trouble of writing lambda all the time. For example. The answer to the question “what is a function?” is “a function is an expression (the body of the function) and an environment containing the values of all variables accessible at the point of definition”. Implementing the interpreter brought an odd fact to light: the interpreter’s code for handling alpha was identical to the code for handling lambda! This indicated that closures — the objects created by evaluating lambda — were useful for both high order functional programming and object oriented programming (Steele 1996). Scheme. as in LISP. but awkward because they require special forms. uses static scoping . as in Listing 9. the factorial function could be represented either as a function. LISP ducks the question “what is a function?” It provides lambda notation for functions. Scheme provides an answer to this question: the value of a function is a closure. which is the value of n in the environment when (f) is evaluated. Thus in Scheme we can write both (define num 6) which binds the value 6 to the name num and (define square (lambda (x) (* x x))) which binds the squaring function to the name square. a Scheme interpreter yields 4 as the value of (f). (Both are possible in LISP.) Since the Scheme interpreter accepts a series of definitions. and an alpha operation for creating actors.n 1) (alpha (f) (c (* f n))))))) 27 interpreter. The closure created by evaluating the definition of f includes all name bindings in effect at the time of definition. not evaluated itself. 4 or 7? If this sequence could be written in LISP. but which value. Closures in Scheme are ordinary values. a lambda operation for creating functions. it is important to understand the effect of the following sequence: (define n 4) (define f (lambda () n)) (define n 7) (f) The final expression calls the function f that has just been defined. They can be passed as arguments and returned by functions.) .n 1)))))) Listing 10: Factorial with actors (define actorial (alpha (n c) (if (= n 0) (c 1) (actorial (. Consequently. the result would be 7.4 THE FUNCTIONAL PARADIGM Listing 9: Factorial with functions (define factorial (lambda (n) (if (= n 0) 1 (* n (factorial (. or as an actor. The value returned is the value of n. (Scheme actually provides an abbreviated form of this definition. however. but a lambda expression can only be applied to arguments.

001 = 2x + · · · We can apply Dsq like this: (->” is the Scheme prompt): -> (Dsq 3) 6. Approximately (Abelson and Sussman 1985): D f (x) = f (x + dx) − f (x) dx We define the Scheme functions shown in Listing 11. Scheme introduced continuations.001)2 − x2 0. set! changes the value of a variable.) . effectively.001)) Listing 12: Banking in Scheme (define (make-account balance) (define (withdraw amount) (if (>= balance amount) (sequence (set! balance (.001 2 Scheme avoids the problem of incompatibility between interpretation and compilation by being statically scoped.(f (+ x dx)) (f x)) dx)))) (define square (lambda (x) (* x x))) (define Dsq (derive sq 0. Differentiation is a function that maps functions to functions. side-effects are allowed. After these definitions have been evaluated. In particular.balance amount)) balance) ("Insufficient funds")) (define (deposit amount) (set! balance (+ balance amount)) balance) (define (dispatch m) (cond ((eq? m ’withdraw) withdraw) ((eq? m ’deposit) deposit) (else (error "Unrecognized transaction" m)))) dispatch) 28 Example 14: Differentiating. In addition to full support for high order functions. whether it is interpreted or compiled. The interpreter uses a more elaborate data structure for storing values of local variables to obtain the effect of static scoping.4 THE FUNCTIONAL PARADIGM Listing 11: Differentiating in Scheme (define derive (lambda (f dx) (lambda (x) (/ (. (The “!” is a reminder that set! has side-effects. the function f (x) = (x + 0. Dsq is. Although Scheme is primarily a functional language.

functions. . The following examples use SASL notation. takes an amount as argument. -> (define acc (make-account 100)) The value of acc is a closure consisting of the function dispatch together with an environment in which balance = 100. evaluating an expression more than once will always give the same result. The definition nums(n) = n::nums(n+1) apparently defines an infinite list: nums(0) = 0::nums(1) = 0::1::nums(2) = . The function dispatch takes a single argument which must be withdraw or deposit. as in Algol 60. it returns one of the functions withdraw or deposit which. Thus combinator reduction is (in this sense) the most efficient way to pass parameters to functions. when compared to OOPLs such as Simula and Smalltalk (described in Section 6) is that we can define only one function at a time. Combinator reduction implements call by name (the default method for passing parameters in Algol 60) but with an optimization. In SASL. in turn. The limitation of this approach. The function shown in Listing 12 shows how side-effects can be used to define an object with changing state (Abelson and Sussman 1985. and never more than once. page 173). it is evaluated exactly once. 4. we can find its second element in SASL: . all of which are implemented with combinator reduction. Andrew’s Symbolic Language) was introduced by David Turner (1976). The expression x::xs denotes a list with first element (head) x and remaining elements (tail) xs. 2 This example demonstrates that with higher order functions and control of state (by side-effects) we can obtain a form of OOP. Turner subsequently designed KRC (Kent Recursive Calculator) (1981) and Miranda (1985). The value returned is the new balance. The following dialog shows how banking works in Scheme. Evaluating an expression only when it is needed.3 SASL SASL (St.4 THE FUNCTIONAL PARADIGM Listing 13: Using the account -> ((acc ’withdraw) 50) 50 -> ((acc ’withdraw) 100) Insufficient funds -> ((acc ’deposit) 100) 150 29 Example 15: Banking in Scheme. . is called call by need or lazy evaluation. The first step is to create an account. The quote sign (’) is required to prevent the evaluation of withdraw or deposit. we can define it like this: second (x::y::xs) = y Although nums(0) is an “infinite” list. Listing 13 shows some simple applications of acc. it is not evaluated. It has an Algollike syntax and is of interest because the compiler translates the source code into a combinator expression which is then processed by graph reduction (Turner 1979). The function second returns the second element of a list. If it is needed one or more times. Since SASL expressions do not have side-effects. local. . This function must be used to dispatch messages to other. If the parameter is not needed in the function.

The second.3 and that most types can be inferred by the compiler. as in the following alternative declaration of the factorial function: .X. .fun fac 0 = 1 = | fac n = n * fac(n-1). val it = 720 : int SML also allows function declaration by cases. the required result is known. there are two patterns. In the declaration of fac. SML is run interactively. as soon as the argument of second is in the form 0::1::.4 SML SML (Milner. If we consider if as a function. it changes the prompt to “=” on the second line. The vertical bar “|” indicates that we are declaring another “case” of the declaration. In this example.fac 6. . SML assigns the result to the variable it. and prompts with “-”..4 THE FUNCTIONAL PARADIGM second(nums(0)) = second(0::nums(1)) = second(0::1::nums(2)) = 1 30 This works because SASL evaluates a parameter only when its value is needed for the calculation to proceed. Call by need is the only method of passing arguments in SASL but it occurs as a special case in other languages. . so that the expression if P then X else Y is a fancy way of writing if(P. or SML. The programmer then tests the factorial function with argument 6. we would not be able to write expressions such as if x = 0 then 1 else 1/x In C. and Harper 1990. the programmer defines the factorial function and SML responds with its type. val it = 720 : int Since SML recognizes that the first line of this declaration is incomplete. . 0. The distinguishing feature of SML is that it is statically typed in the sense of Section B. which can be used in the next interaction if desired. . val fac = fn : int -> int . . .fac 6.Y). \tt n. Milner and Tofte 1991) was designed as a “metalanguage” (ML) for reasoning about programs as part of the Edinburgh Logic for Computable Functions (LCF) project. 4. Each case of a declaration by cases includes a pattern. the functions && (AND) and || (OR) are defined as follows: X && Y X || Y ≡ if X then Y else false ≡ if X then true else Y These definitions provide the effect of lazy evaluation and allow us to write expressions such as if (p != NULL && p->f > 0) . Tofte. The first. . If it did not. val fac = fn : int -> int . The language survived after the rest of the project was abandoned and became “standard” ML. is a constant pattern. and matches only itself. In the following example.fun fac n = if n = 0 then 1 else n * fac(n-1). then we see that if must use call by need for its second and third arguments.

The declaration of hasfactor introduces two functions. val it = 81. as in hasfactor 2. as shown in Listing 16. val quad = fn : real -> real .4 THE FUNCTIONAL PARADIGM Listing 14: Function composition .0.even 6. val it = true : bool . Listing 16: A function with one argument . adapted from (Ake Wikstr¨m 1987). defined as an infix operator. The function o (intended to resemble the small circle that mathematicians use to denote functional composition) is built-in. All functions in SML have exactly one argument.0 : real Listing 15: Finding factors . Note that the definition fun sq x = x * x.val even = hasfactor 2. Functions like hasfactor take their arguments one at a time. val it = 289. The symbols ’a. but this is not the case. shows how functions can be o used to build a programmer’s toolkit.fun hasfactor f n = n mod f = 0. they indicate that SML has recognized o as a polymorphic function.0 : real We can pass functions as arguments to other functions.quad 3. as in Listing 14. It may be helpful to consider the types involved: hasfactor : int -> int -> bool hasfactor 2 : int -> bool hasfactor 2 6 : bool ˚ The following brief discussion. The functions here are for list manipulation. Applying the first argument. val it = true : bool 31 is a variable pattern.fun sq x:real = x * x.infix o.sq 17. val even = fn : int -> bool. . which is a widely used example but not the only way in which a FPL can be used. The function hasfactor defined in Listing 15 returns true if its first argument is a factor of its second argument.hasfactor 3 9. We start with a list generator.fun (f o g) x = g (f x).val quad = sq o sq. after the American logician Haskell Curry. . and ’c are type names. val sq = fn : real -> real . val hasfactor fn : int -> int -> bool . yields a new function. It might appear that hasfactor has two arguments. ’b. but even if it wasn’t. would fail because SML cannot decide whether the type of x is int or real.0. we could easily declare it and use it to build the fourth power function. and matches any value of the appropriate type. val o = fn : (’a -> ’b) * (’b -> ’c) -> ’a -> ’c . The trick of applying one argument at a time is called “currying”. .

5). We can use reduce to obtain one-line definitions of sum and prod. We note that sum and prod have a similar form. following LISP and Scheme.= fun : int * int -> int list .4. val it = 120 : int Listing 18: Using reduce .5). and those that provide mainly “lazy evaluation”.prod (1 -. the notation introduced by Backus (1978). val -.5] : int list The functions sum and prod in Listing 17 compute the sum and product. Although Backus’s speech (and subsequent publication) turned many people in the direction of functional programming. and a list.fun sum xs = reduce add 0 xs.n) = if m < n then m :: (m+1 -.4 THE FUNCTIONAL PARADIGM Listing 17: Sums and products .5. Since 1978.5 Other Functional Languages We have not discussed FP. val sum = fn : int list -> int .fun reduce f u [] = u = | reduce f u (x::xs) = f x (reduce f u xs).3. call by value). of a list of integers.fun insert x:int [] = x::[] | insert x (y::ys) = if x <= y then x::y::ys else y::insert x ys. val prod = fn : int list -> int .fun (m -. respectively. his ideas for the design of FPLs were not widely adopted.2. val it = 15 : int . This suggests that we can abstract the common features into a function reduce that takes a binary function. preserving the ordering: . . val reduce = fn : (’a -> ’b -> ’b) -> ’b -> ’a list -> ’b We can also define a sorting function as . a value for the empty list.val sort = reduce insert nil.n) else []. val it = [1.infix --. . val insert = fn : int -> int list -> int list 4.sum (1 -. val prod = fn : int list -> list 32 .1 -.fun prod [] = 1 = | prod (x::xs) = x * prod xs.fun sum [] = 0 = | sum (x::xs) = x + sum xs. .fun prod xs = reduce mul 1 xs. as in Listing 18. val sort = fn : int list -> int list where insert is the function that inserts a number into an ordered list. FPLs have split into two streams: those that provide mainly “eager evaluation” (that is. val sum = fn : int list -> list . The idea of processing a list by recursion has been captured in the definition of reduce.

4 THE FUNCTIONAL PARADIGM 33 following SASL. The most important language in the first category is still SML. while Haskell appears to be on the way to becoming the dominant FPL for teaching and other applications. .

5. if you enter :browse (or :b for short). the environment recognizes various commands with the prefix ”:”. and prints the value.org/bio. its foundation is the mathematical concept of a function. ’\n’.html). You can find information about Haskell at www. and Basic. Prelude> 1234567890987654321 / 1234567 1. The concatenation operator for lists is ++. download it from http://www. Haskell strings are not terminated by ’\0’. evaluates it. including Haskell. This is the same as the concatenation operator for strings — Haskell considers a string to be simply a list of characters. Figure 2 shows Hugs working out 2 + 2. you need to know more than one programming language. we will be using a system derived from Gofer and called HUGS (Haskell Users’ Gofer System).).org. Prelude> [1. is the list. For instance.3] ++ [4.5 Haskell COMP 348 is about the principles of programming languages. Integers have unlimited size and conversion are performed as required. It is based on a “read/eval/print” loop: the system reads an expression.org/hugs. Haskell uses the same escape sequences C++(e. Haskell gets its name from the American logician Haskell Curry (1900–82 — see http://www. Although it has many of the features that we associate with other languages. and data structures.3. 5. Java. Hugs has been installed on both Solaris8 machines and linux hosts in pkg/hugs98. In order to understand the principles. as in C++. Even this is not enough: you really need to know more than one style of programming.0e+012 :: Double Prelude> 1234567890987654321 ‘div‘ 1234567 1000000721700 :: Integer Strings are built into Haskell..5. Prelude> "Hallo" ++ " " ++ "everyone!" "Hallo everyone!" :: [Char] The basic data structure of almost all functional programming languages. One of the first Haskell implementations was called “Gofer”.2.4. 34 . As well as expressions. The infix operator ++ concatenates strings. Hugs will display the functions and structures defined in the current environment.6] :: [Integer] The prelude provides a number of basic operations on lists. Executables are in /site/bin and can be run without any special preparation.g.haskell. installation is straightforward. In Haskell. If you want to run Hugs on your own PC. The prompt Prelude> indicates that the only module that has been loaded is the “prelude”. Accordingly. such as variables. the course begins with an introduction to a language that is quite different to languages that you may already be familiar with such as C++. Hugs provides an interactive programming environment. However. At Concordia. expressions.haskell.6] [1.2. lists are enclosed in square brackets with items separated by commas. Haskell computes with integers and floating-point numbers. A string literal is enclosed in double quotes. Haskell is a functional programming language in the sense of Section 4. etc.1 Getting Started There are several Haskell compilers and all of them are free.haskell. Documentation is in pkg/hugs98/docs.

you have to write your own programs — which.3.hs" Reading file "e:/courses/comp348/src/lec1. but it quickly gets boring.8.19.18.19.15..6.5.20] :: Prelude> sum [1. The command is :load and it is followed by a path name..Syntax error in input (unexpected ‘=’) Fortunately.hs Type :? for help Prelude> 2+2 4 :: Integer Prelude> Figure 2: A short session with Hugs Prelude> [1.20] [2.2.4..5 HASKELL __ __ || || ||___|| ||---|| || || || || __ __ ____ ___ || || || || ||__ ||__|| ||__|| __|| ___|| Version: December 2001 _____________________________________________ Hugs 98: Based on the Haskell 98 standard Copyright (c) 1994-2001 World Wide Web: http://haskell.5] :: [Integer] Prelude> drop 10 [1.. means mainly writing functions (although later we will also introduce new data types).14.12.8.20] [1. The environment does not allow you to introduce any new definitions: Prelude> n = 99 ERROR .13.3.4.org/hugs Report bugs to: hugs-bugs@haskell.7.2.hs": .11.17.20] [1.20] 2432902008176640000 :: Integer Prelude> take 5 [1.16.14.10. in Haskell. it does allow you read definitions from a file.10.6.14.13.9.20] 210 :: Integer Prelude> product [1.20] :: [Integer] Prelude> take 5 "Peter Piper pickled a pepper" "Peter" :: [Char] [Integer] Evaluating expressions is fun for a few minutes. To do anything interesting..18.20] :: [Integer] Prelude> filter even [1.17.16.20] [11.hs": Hugs session for: C:\Program Files\Hugs98\lib\Prelude.4.16.18.15.12. Here is how the definitions for this lecture are loaded.org _____________________________________________ 35 Haskell 98 mode: Restart with command line option -98 to enable extensions Reading file "C:\Program Files\Hugs98\lib\Prelude.12.. Note that the prompt changes to reflect the newly loaded module: Prelude> :load "e:/courses/comp348/src/lec1.

:set options :cd directory :edit file :find name :gc :info name . :also filename .. .. .hs The file lec1.hs looks like this (the “. :module module :names pattern :project project file :type expression :version Brief description List all Hugs commands Execute a shell command Quit the current Hugs session Load Haskell source from a file Load additional files Repeat last load command Display functions exported by module(s) Set flags and environment parameters (see Figure 4) Change working directory Start the editor Start the editor with the cursor on the object named Force garbage collection Display information about files.hs e:/courses/comp348/src/lec1. we will include the prompt. classes. .1. In the examples that follow. 5. When we need to show the effect of evaluating an expression. you can enter :b instead of :browse.1) . . as in the examples above. . The first definition introduces a function called fc1... This code introduces a Haskell module called Lec1 that contains a number of definitions. Figure 4 shows options for the :set command. The WinHugs environment has icons for most of these commands. etc. . we digress briefly to explain the commands provided by the Hugs environment. Figure 3 provides a brief summary of each command. Notes for Figure 3 . . . we will show Haskell code as it appears in the source file. these options can also be specified on the command line.. Note that any command can be abbreviated: for example. :reload :browse module .” indicates material that is discussed below): module Lec1 where fc1 n = if n == 0 then 1 else n * fc1(n ..1 Hugs Commands Before describing the Haskell language.5 HASKELL 36 Command :? :! command :quit :load filename . Change module Find names that match the pattern Load a project (obsolete) Print the type of expression without evaluating it Display Hugs version Figure 3: Hugs commands Note (1) (1) (1) (2) (2) (2) (3) Hugs session for: C:\Program Files\Hugs98\lib\Prelude.

±q ±w ±k ±u ±i p string r string P string E string F string h num c num Print # of reductions and cells after evaluation Print type after evaluation Terminate evaluation after first error Print cells recovered by garbage collector Use literate modules as default Warn about errors in literate modules Print dots to show progress (not recommended!) Print nothing to show progress Always show which modules are loaded Show “kind” errors in full Use show to display results Chase imports when loading modules Set prompt to string Set repeat last expression string to string Set search path for modules to string Use editor setting given by string Set preprocessor filter to string Set heap size to num bytes (has no effect on current session) Set constraint cutoff limit to num Figure 4: Parameters for :set. In this case. The editor commands assume that Hugs has been linked to an editor. that make the :project command unnecessary in most cases. :edit suspends Hugs and starts the editor. The default editor for Hugs in Windows is notepad. You may find it more convenient to ignore these commands and simply run Hugs and your favourite editor as independent processes. The other commands take a string or a number as argument. Recent versions of Hugs use a smarter method of loading projects. The project command reads a project file containing information about the files that Hugs should load.5 HASKELL 37 ±s ±t ±f ±g ±l ±e ±. for a Windows session. but it positions the editor cursor at the chosen object. :find is similar. One easy way to develop a Hugs program is to: Write the first version in a file fun. 3. :reload repeats the previous :load command. Hugs resumes.hs Make corrections in the editor window Reload the file in the Hugs window: :r 2. :also is similar but it does not remove previous definitions. The symbol ± indicates a toggle: +x turns x on and -x turns it off. You can obtain the information in Figure 4 by entering :set without any arguments. The output also includes the current settings which. :load removes all definitions (except those belonging to the Prelude) and loads new definitions from the named file(s). you may like to use emacs with Hugs. 1. when you close the editor. as shown. called “import chasing”. For example.hs Leaving the editor running. will look something like this: . all you have to do is save the corrected file in the editor window and enter :r in Hugs to reload the modified source code. load this file into Hugs and try the functions: :l fun.

lhs": Parsing ERROR "e:/courses/comp348/src/vector.5 HASKELL Current settings: Search path : Project Path : Editor setting : Preprocessor : Compatibility : 5. One is to turn off the e switch (its default setting is +e): Prelude> :set -e and then reload the source file — it will load without errors.{Hugs}\lib\exts. A program written in this way is called a literate program and the art of writing such programs is called literate programming . Here are the rules: • Literate Haskell source files use the extension .{Hugs}\lib\hugs. • Any other lines are ignored by the compiler.qQkI -h250000 -p"%s> " -r$$ -c40 -P{Hugs}\lib.lhs instead of . it will report: Prelude> :l e:/courses/comp348/src/vector.lhs Reading file "e:/courses/comp348/src/vector.exe -F Haskell 98 (+98) 38 Literate Haskell Donald Knuth proposed the idea that programs should be readable documents with detailed explanations and commentary built in. the e switch in the compiler is on (as if we had entered :set +e).1. The check for blank lines can be suppressed by turning off the e switch in Hugs. • A line beginning with > is interpreted as Haskell code. " ++ show y ++ ". The other way is to put blank lines in the source file: . The first step towards literate programming is to make “comments” very easy to write and read. When Hugs loads this file. This enables easy conversion of vectors to strings. Here’s an example to illustrate how this works. Suppose that the source file vector.lhs":14 . " ++ show z ++ ")" By default.hs. • The lines before and after lines containing code must be either blank or lines containing code (but see the e switch below). > instance Show Vector where > show (V x y z) = "(" ++ show x ++ ".Program line next to comment There are two ways to fix this problem.2 +tfewuiAR -sgl. rather than just code listings with occasional — and often useless — comments.{Hugs}\lib\win32 -EC:\WINNT\notepad. The overhead for code in literate Haskell (at least the Hugs version) is relatively light.lhs contains this text: Vector is an instance of the type class Show. A consequence of this step is that the code itself becomes slightly harder to write and read.

We can compute n ! by constructing the list [1. Our first definition of the factorial function uses recursion: fc1 n = if n == 0 then 1 else n * fc1(n .1. . to make the syntax legal. are defined by cases. . " ++ show z ++ ")" In general. since everything is a value in Haskell. The preferred form of definition is a set of equations: fc3 0 = 1 fc3 n = n * fc3(n . it is a good idea to put blank lines in the source file anyway.. It can be written on one line but. " ++ show y ++ ".3. 39 > instance Show Vector where > show (V x y z) = "(" ++ show x ++ ".. Figure 5 shows a small but complete example of a literate Haskell program.4): fc2 n = case n of 0 -> 1 .. the factorial function: 0! = 1 n ! = 1 × 2 × 3 × .n] and then using the Prelude function product to multiply all of the list items: fc4 n = product [1. there must be a semicolon between the two case clauses (see Section 5. We will start by considering the favourite function of functional programmers. several ways to define functions. n -> n * fc2(n .1) This definition works fine. there is no need for a return statement: you write just the value that you want returned. There are also “closed form” expressions for some functions. but fc3 4 would cause an error. × n for n ≥ 1 The first point to note is that.n] This works correctly when n = 0 because the “product” of the empty list is defined to be 1: Lec1> product [] .5 HASKELL Vector is an instance of the type class Show... but it is not the preferred style for Haskell. 5. Many functions.3 Defining Functions Since Haskell is a functional language.2. is not idiomatic Haskell. in fact. because they make the code stand out more clearly. however. they are in fact independent assertions for Haskell. If we wrote the first but not the second. The cases for factorial are n = 0 and n > 0.1) Even this. fc3 0 would evaluate correctly to 1.1.1) Although we have written these equations together. defining functions is fundamental and there are.1) This is the first definition that is written on more than one line. including this one. This enables easy conversion of vectors to strings. Haskell has a case expression (analogous to switch in C++ and Java) and we use it here for the second definition of factorial: fc2 n = case n of 0 -> 1 n -> n * fc2(n .

> instance Show Vector where > show (V x y z) = "(" ++ show x ++ ". The expression ’norm v’ yields a unit vector with the same direction as v. " ++ show y ++ ". The following definitions provide sums. We define V to be the constructor for vectors.x’ * y) (negate z) A vector can be scaled by a real number. > dot (V x y z) (V x’ y’ z’) = x * x’ + y * y’ + z * z’ > len v = sqrt(dot v v) > norm v = if l == 0 > then error "Norm of zero vector" > else scale v (1/l) > where l = len v Figure 5: A Literate Haskell program .y’ * z) (z * x’ .z’) * y’ . differences. This enables easy conversion of vectors to strings.z’ * x) (x > negate (V x y z) = V (negate x) (negate y) (z + z’) (z . Vectors are equal if all of their components are equal. It might be an improvement to consider vectors to be equal if their components are approximately equal.x’) (y .5 HASKELL Author: Peter Grogono Date: 5 October 2003 Description: Example of a Literate Haskell source file > module Vector where This is a module for vectors in 3D Euclidean space. and outer (cross) products of vectors. > instance Num Vector where > V x y z + V x’ y’ z’ = V (x + x’) (y + y’) > V x y z . " ++ show z ++ ")" Vector is an equality type. > data Vector = V Double Double Double 40 Vector is an instance of the type class Show.y’) > V x y z * V x’ y’ z’ = > V (y * z’ . > instance Eq Vector where > V x y z == V x’ y’ z’ = x == x’ && y == y’ && z == z’ Vector is a numeric class. > scale (V x y z) d = V (x * d) (y * d) (z * d) These auxiliary functions define the dot product and length of a vector. Haskell does not allow us to overload ’*’ and ’/’ for vector/scalar operations.V x’ y’ z’ = V (x .

The backslash indicates that n is a parameter and the “arrow” -> links the parameter to the body of the function.5 HASKELL 1 :: Integer 41 To illustrate some other styles of function definition. However. Putting parentheses around +1 makes this expression into a function: Lec1> (+1) 6 7 :: Integer In general. we can write definitions such as add x y = x + y and use the functions defined: Lec1> add 347589 23874 371463 :: Integer How do we account for this? The function add is actually a second order function that expects one argument and returns a function that expects another argument. we can create functions from it by giving it either a left or a right operand. We can write definition in this way in Haskell: inc2 = \n -> n + 1 We can read this definition as “inc2 is the function that maps n to n + 1”. The general rules are: (op expr ) is equivalent to \x -> (x op expr ) (expr op ) is equivalent to \x -> (expr op x) There is an important exception to this rule: (-1) is not a section: it is just −1. define . In fact. Its form is ? = 1 in which ? stands for the unknown parameter. all of our functions have had exactly one argument. n + 1. To see this. The expression \n -> n + 1 can be used alone: it is an anonymous function. Nevertheless. Lec1> (\n -> n + 1) 7 8 :: Integer There is another way of writing this anonymous function.0 :: Double Lec1> (12/) 5 2.4 :: Double This provides us with another definition of inc: inc3 = (+1) Operators with one argument missing are called sections. (/3) is a function that divides its argument by 3 and (3/) is a function that divides its argument into 3: Lec1> (/3) 9 3. The “obvious” definition is a single equation: inc1 n = n + 1 Purists have a problem with this definition: they would prefer definitions to have to form x = e in which x is the name of the thing being defined and e is the defining expression. Backslash looks a bit like λ and the Haskell notation suggests the way in which function is written in λ-calculus: λ n . For example. we will use an even simpler function: inc adds 1 to its argument. all Haskell functions take exactly one argument.expr . given a binary operator. So far. (subtract expr ) is a section equivalent to \x -> x .

’ ”. The “blank” operator denotes function application. and such. The following definition is fine: fc6 n = . It is important to know that the rules exist. You will have noticed.hs fc6 n = if n == 0 then 1 else n * fc6(n . 5. possibly due to bad layout) Haskell allows semicolons but does not require them. the parser inserts a semicolon at the end of every line. For example.4 Syntax Syntax conventions for functional programming are different from “traditional” programming styles. Juxtaposition binds higher than any other operation. however. for example. The application f (g(x)) applies g to x and then applies f to the result. There’s nothing wrong with line breaks. Compared to programs in most other languages. First. Here are some of the rules.5 HASKELL inc4 = add 1 and try it: Lec1> inc4 5 6 :: Integer 42 Thus add 1 is “the function that adds 1 to its argument” and add 99 is “the function that adds 99 to its argument”. that we write inc n rather than inc(n). if we want the effect of f (n + 1).hs": Parsing ERROR "e:/courses/comp348/src/lec1. which must be a function. For example. if you break a line in the wrong place. in Haskell indentation and line breaks make a difference. and so on. the Haskell expression f n+1 denotes f (n) + 1. provided that they occur in sensible places.1. you get the mysterious diagnostic “unexpected ‘.hs":27 Syntax error in expression (unexpected ‘. Consequently. Thus f x means “the result of applying the function f to the argument x”. The rules are simple and intuitive and. is applied to the argument y. we must write f(n+1). Because application binds strongly. We can write this in Haskell either as f(g(x)) or as f (g x). Haskell interprets the expression add 2 3 as “apply the function add 2 to the argument 3”. because otherwise Haskell’s error messages can be quite confusing. The trade-off is that Haskell programmers must respect layout rules.’. Haskell programs are remarkably free from “syntactic clutter” — braces. you may never have to learn them explicitly. then the result. for example. Unlike C++. the function f is applied to the argument x. The technical name for the blank operator is juxtaposition.1) gives this error: Reading file "e:/courses/comp348/src/lec1. Haskell interprets f x y as ((f x) y). inserting this definition into lec1. semicolons. we must put parentheses around arguments that are not simple. if you use existing source code as a model. However.

To use the Trace module. replace any expression e by (trace (s) (e)) in which s is a string. module Demo where import Trace {fac 0 = 1 fac n = n * fac (n . Why? Because statements have side-effects and functional programs are not supposed to have side-effects.5 HASKELL if n == 0 then 1 else n * fc6(n . To use it. it prints the message as a side-effect and returns the expression. Since trace returns the value of e.1) -} fac 0 = 1 fac n = trace ("fac " ++ show n ++ "\n") (n * fac (n . We keep the original version for later use and define another version with a call to trace. The parentheses are not always necessary (depending on the context) but they do no harm.1) We have already seen the Haskell case expression. 5. p2 -> e2 . Here is a demonstration of Trace in use. you will be able to monitor its progress by looking at what trace prints. and we will state the rules when we introduce the constructs. write import Trace at the beginning of your program. When trace is called. One problem is that we cannot interleave print statements with other statements because functions programs don’t have statements.5 Tracing Debugging functional programs is difficult. The Hugs Trace module provides a way around the problem of no side-effects. p3 -> e3 There are a few other Haskell constructs for which layout is important. However. We import the Trace module and define our old friend. the meaning of your program should not be affected. Cases must appear on separate lines: case e p1 p2 p3 of -> e1 -> e2 -> e3 43 or separated by semicolons: case e of p1 -> e1 . It defines a function trace with two arguments: a string and an expression.1)) Calling the traced version of fac gives: .1. the factorial function.

.0.3].1.2. .a character list is a string "abc" :: [Char] Lec1> [1.5 HASKELL Demo> fac 6 fac 6 fac 5 fac 4 fac 3 fac 2 fac 1 720 :: Integer 44 5.2. This is a list consisting of an item x and a list xs. Here are the basic built-in operations for lists: Lists are enclosed in square brackets: [."def"] -.3] :: [Integer] Lec1> ["abc". y:ys. The operator : is right-associative.integer list .. a:as. A list may contain items of any type. 1.Illegal Haskell 98 class constraint in inferred type *** Expression : [1.to separate comments from code): Lec1> [1. and so on.integers are promoted if necessary [1. Items within lists are separated by commas: [1.OK [1.OK ["abc".mixed types .string list .2 5. The initial letter may vary: for example.3] -."def"] :: [[Char]] Lec1> [’a’.BAD! ERROR .2.1 Details Lists We have already seen (in Section 5) that Haskell supports lists.’a’] -.2. but all of the items in a given list must have the same type (note the use of -.’a’] *** Type : Num Char => [Char] The infix operator : builds lists: it takes an item on the left and a list on the right.5] :: [Double] Lec1> [1.’b’.]. The empty list is [].5] -. so parentheses are not required in the second example below: Lec1> 1:[] [1] :: [Integer] Lec1> 1:2:[] [1.’c’] -.2] :: [Integer] A very common idiom in Haskell progamming is x:xs (read “x followed by xes”).

6] :: [Integer] Lec1> [1.2] ERROR .6] [1.6.3.Illegal Haskell 98 class constraint in inferred type *** Expression : app [’a’. []? and what is the result for a non-empty list x:xs? For example. If the append operator ++ did not exist..4.5 HASKELL 45 The expression [x..4] :: [Integer] Lec1> [3.4. [Char] There are functions for extracting the first item of a list (head) and the remaining items of the list (tail) but we don’t often need them.15. provided that x and y are members of the same enumerated type: Lec1> [2. we could append lists using our own function app defined by app [] ys = ys app (x:xs) ys = x:app xs ys Any Haskell function with two arguments1 can be used as an infix operator by putting back quotes (grave accents) around it.12.2] [] :: [Integer] The infix operator ++ concatenates lists. Haskell would parse len x:xs as (len x):xs and report a type error. we define: len [] = 0 len (x:xs) = 1 + len xs The parentheses are required: since juxtaposition has higher precedence than colon.3. it hasn’t really got two arguments — see Section 5. to find the length of a list. The reason is that we can define list functions by using the constructors as patterns.2.’z’] "abcdefghijklmnopqrstuvwxyz" :: Lec1> [5.’b’] [1.2] *** Type : Num Char => [Char] A common operation on lists is to select items that satisfy a predicate.3] ‘app‘ [4.9. The third example shows what happens if we try to build a list with mixed types. .2.’b’] [1.5.2.2.3.18] :: [Integer] Lec1> [’a’. For example.20] [3.4] [2..3.6] [1.6] :: [Integer] Lec1> app [’a’. Lec1> app [1.3] [4.5..6. The second example below demonstrates this.5.. to obtain the even items in a list of integers. we could use the Prelude function even: evens [] = [] evens (x:xs) = if even x then x:evens xs else evens xs 1 Of course..1.y] constructs a list. Most function definitions need only two cases: what is the result for an empty list.5.

5 HASKELL

46

In functional programming, however, we usually want to generalize, using functions to do so. The function filt takes a predicate p and a list, and returns the items in the list that satisfy the predicate. (A better name would be filter but a function with this name is defined in the Prelude: it does the same as filt.) filt f [] = [] filt p (x:xs) = if p x then x:filt p xs else filt p xs Now we can write: Lec1> evens [1..20] [2,4,6,8,10,12,14,16,18,20] :: Lec1> filt even [1..20] [2,4,6,8,10,12,14,16,18,20] :: evens = filt even 5.2.2 Lazy Evaluation [Integer] [Integer]

Using filt, we can write a better definition for evens:

Haskell evaluates expressions lazily : this means that it never evaluates an expression until the value is actually needed. It is perfectly alright to introduce a non-terminating definition into a Haskell program, but things go wrong when we try to use it. Here is a non-terminating definition loop = loop and here is an attempt to use it (the message {Interrupted!} appears when the user interrupts the program, for example, by entering ctrl-C): Lec1> loop + 3 {Interrupted!} Lec1> There are other things that fail: Lec1> 1/0 Program error: one n = 1 Since Haskell never evaluates the argument of this function, we can give it any argument at all: Lec1> 1 :: Lec1> 1 :: Lec1> 1 :: one 77 Integer one (1/0) Integer one loop Integer primDivDouble 1.0 0.0

But now define a constant function that doesn’t use its argument:

Lazy evaluation allows us to work with infinite lists. The lists do not have an infinite number of items: what “infinite” means in this context is “unbounded” — the list appears to have as many items as we need. For example, if we define from n = n:from (n+1) nats = from 1

5 HASKELL

47

then from n yields an unbounded list starting at n and nats yields as many natural numbers as we might need: [1,2,3,...]. We can use the built-in function take to bite off finite chunks from the beginning of an infinite list: Lec1> take 6 nats [1,2,3,4,5,6] :: [Integer] Lec1> take 6 (evens nats) [2,4,6,8,10,12] :: [Integer] Here is an alternative definition of nats: nats = [1..] As an application of infinite lists, the function nomul returns the items in list l that are not multiples of n. The function sieve takes a list, and returns its first item n followed by the list obtained by applying itself to a list with multiples of n. nomul n l = filter (\m -> mod m n /= 0) l sieve (n:ns) = n:sieve (nomul n ns) This is what happens when we apply sieve to the list [2,3,4,...]. Lec1> sieve (from 2) [2,3,5,7,11,13,17,19,23,29,31,37,41,43,47,53,59,61,67,71,73,79,83,89 Interrupted! 5.2.3 Pattern Matching

It is time to look more closely at some of the simple definitions we have already seen: fac 0 = 1 fac n = n * fac(n - 1) len [] = 0 len (x:xs) = 1 + len xs The object following the function name on the left of = is a pattern. When the function is called, the argument of the function is matched against the available patterns. If a match succeeds, the expression on the right side of = is evaluated. Pattern matching may include binding , which assigns values to variables in the expression. Examples: In the call fac 0, the argument 0 matches the pattern 0. The expression 1 is evaluated, yielding 1. In the call fac 3, the argument 3 matches the pattern n, yielding the binding { n = 3 }. The expression n * fac(n - 1) is evaluated with this binding, yielding 3 * fac(2). In the call len [], the argument [] matches the pattern []. The expression 0 is evaluated, yielding 0. In the call len [27..30], the argument is equivalent to 27:28:29:30. This matches the pattern x:xs, yielding the bindings { x = 27, xs = [28..30] }. The expression 1 + len xs is evaluated with these bindings, yielding 1 + len [28..30]. Pattern matching occurs in a number of contexts in Haskell. All of these contexts can be expressed in terms of the case expression:

5 HASKELL

48

Kind Wild card Constant Variable Constructor

Pattern k v C p 1 p1

Expression e e e C e1 e2

Condition for match none k==e none C = C , p1 ∼ e 1 , p2 ∼ e 2

Binding none none v=e

Figure 6: Pattern matching

case e of p1 -> e1 ; p2 -> e2 ; ...; pn -> en Here is how other kinds of expression map to case expressions: \ p1 p2 . . . pn → e = \x1x2 . . . xn → case (x1 , x2, . . . , xn ) of (p1, p2, . . ., pn ) → e where the xi are new variables if e1 then e2 else e3 = case e1 of True → e2 ; False → e3 let p1 = e1 ; p2 = e2 ; . . . pn = en ; in e = case (e1, e2, . . . , en ) of (p1, p2, . . ., pn ) → e2 The case expression case e of p1 -> e1 ; p2 -> e2 ; ...; pn -> en is evaluated as follows: The expression e is matched against p1 , p2, . . . , pn in turn. If e matches pi , then the value of the expression is obtained by evaluating ei with any bindings obtained by matching. If none of the patterns match, the expression fails. Note that the order of patterns is significant. For example, the following definition of the factorial function does not work, because the first pattern always succeeds and evaluation never terminates: fac n = n * fc3(n - 1) fac 0 = 1 Figure 6 summarizes the important kinds of pattern and the way in which they match. The expression p ∼ e stands for “pattern p matches expression e”. The underscore character, , is called a “wild card” because it matches anything but produces no binding. Use the wild card when you don’t need the value of the argument in the function. For example, the head and tail selectors for lists can be defined as head (x: ) = x tail ( :xs) = xs A constant matches a constant with the same value but produces no binding. For example, [] matches the empty list but nothing else. A pattern consisting of a variable name matches any expression and binds that expression to the variable. If the pattern is a constructor with patterns as arguments, the expression must be an invocation of the same constructor to expressions. The corresponding expressions are matched and the bindings from the matches are combined. Since “:” is the list constructor, the pattern (3:xs) matches any list whose first element is 3.

Instead.y) = x + y This is not the same function as the Prelude function add x y = x + y The argument of plus must be a pair. It is “syntactic sugar” for brk xs = (\m -> (take m xs. The Haskell equivalent is: brk xs = (take m xs. we would introduce a local variable. drop m xs)) (len xs ‘div‘ 2) -..5 HASKELL 49 Haskell patterns are linear . We are not allowed to write merge (x:xs) (x:ys) = . else . The solution is to include this case in the definition: swap [] = [] swap [x] = [x] swap (x:y:ys) = y:x:swap ys The tuple is another kind of constructor in Haskell. Here are some examples of the pattern matching rules.) Tuples are enclosed in parentheses and the items inside them are separated by commas.. We can also define plus (x.. The items need not be of the same type. there is no equivalent to “add 1”. in which case xs is bound to []. we must write merge (x:xs) (y:ys) = if x == y then . tuples of different lengths have different types. ) = x second ( . (A tuple corresponds roughly to a C++ struct. First. these definitions are not really necessary. Unlike lists.y) = y However. Suppose that we are defining a function merge that has two lists and arguments and that we need to consider the case in which the first item in both lists is the same. and the non-empty list pattern. The following function breaks a list into two roughly equal parts and returns the two parts as a pair: brk xs = (take (len xs ‘div‘ 2) xs. This pattern matches the list with one item. the most common two cases for a list function are the empty list pattern. []. if we define swap [] = [] swap (x:y:ys) = y:x:swap ys the expression “swap [3]” gives an error because the list with one element does not match any pattern.BAD! . Functions can return tuples. because the Prelude defines functions fst and snd that do the same thing.. we would like to avoid the repeated evaluation of “len xs ‘div‘ 2”. A tuple with two items is usually called a pair .. In a procedural language. This means that each variable in a pattern must be distinct. For example. (x:xs). We can define selectors for pairs like this: first (x. drop m xs) where m = len xs ‘div‘ 2 This may not look “functional” but in fact it is. drop (len xs ‘div‘ 2) xs) Obviously. Sometimes a third pattern is necessary..

We have not encountered all of them yet. \. if. Figure 7: Parsing precedence.5 HASKELL 50 Description simple term parenthesized term as-patterns application do. \. 5. if.2. <. left. case (rightwards) sequences generators grouping guards cases definitions separation Syntax t (t) a@p fx Associativity right left right right Figure 8 -> right => :: right . and case) apparently have two precedences: one leftwards and one rightwards. Some constructs (do. case (left) case (leftwards) infix operators — see function types contexts type constraints do. from highest (at the top) to lowest (at the bottom). left. Precedence 9 8 7 6 5 4 3 2 1 0 Left !! * / ‘div‘ ‘mod‘ ‘rem‘ ‘quot‘ + - Associativity None Right .. ^ ^^ ** : == /= < <= > >= ‘elem’ ‘notElem’ && || >> >>= $ ++ $! ‘seq’ Figure 8: Operator precedence. \. if.4 Operators Figures 7 and 8 show the precedence rules that Haskell uses for parsing expressions and operators. This is because of a metarule which says that each of these constructs . from highest (at the top) to lowest (at the bottom). n-ary | -> = . let.

the list is unchanged. Otherwise. conditional. let seems to have higher precedence than +. The key step in writing a function is often to choose the appropriate parameter for recursion. we want xs and x:xs’ to stand for the same .5 expression a = 3 in a + a let a = 3 in a + a (let a = 3 in a) + a parses as let a = 3 in (a + a) k + (let a = 3 in (a + a)) (k + (let a = 3 in a)) + a Building Haskell Programs In procedural programming.). recursion is used frequently. Since Haskell does not have an explicit loop construct. It is often easier to define a new function using Prelude functions than it is to write the function “from scratch”. takeWhile p [] = [] takeWhile p (x:xs) = if p x then x : takeWhile p xs else [] dropWhile p [] = [] dropWhile p (x:xs) = if p x then dropWhile p xs else x:xs Haskell provides another way of writing the last line. sequence. with experience. drop 0 xs = xs drop [] = [] drop n ( :xs) = drop (n-1) xs The functions takeWhile and dropWhile are only a little more complicated. because the number of small functions is large. In this case. In Functional Programming (FP). The infix operator @. But. allows us to write a pattern in two equivalent ways. let seems to have lower precedence than +. we have only one abstraction mechanism: function definition. This can be confusing for the beginner. Recall that “take n xs” returns the first n items of the list x. as shown in the third lines of the table.2. Note parentheses override the metarule. To be safe. In the second line. The Prelude is a collection of functions that many programmers have found useful. and we use them to build programs. we define the appropriate response when the list is empty. because its components group together. because the whole expression “a + a” is part of it. In the first line in the table below. Note also that all three parsings are what we might intuitively expect. read “as”.5 HASKELL 51 extends as far to the right as possible. The definition is similar. we use recursion on the value of n. we have a collection of control structures. Here are some of the Prelude functions and their definitions. loop. The second line is an application of the metarule: the let expression extends as far as possible. harder because we have a more limited toolkit. The tendency in FP is to define a lot of small functions and then to build programs using these functions. or abstraction mechanisms (assignment. take 0 xs = [] take [] = [] take n (x:xs) = x:take (n-1) xs Recall also that “drop n xs” throws away the first n items of the list xs and returns the rest. If n is zero. The let k + k + 5. This makes the task of programming both simpler and harder: simpler because there are fewer things to worry about. procedure call. etc. this approach to programming becomes quite manageable.

Prelude> span (<5) [1.15.xs) where (ys.20] ([1.6.[5. note: Prelude> break isSpace "Here is a sentence. and we want to format so that each line is no longer than 80 characters." ("Here".[Char]) The definitions of span and break are as follows: span p [] = ([].2.3. which puts the words together again. The idea is to put words into a line until the next word won’t fit — that is.18.8.17.19. Where do we store the information in this “line”? The answer is that we need an accumulating parameter and recursion..13. The first step is to break into up into a list of words using the function words. The problem is that we have no local variables and no state.12.4].10.14.") :: ([Char].11.20]) In particular.[]) span p xs@(x:xs’) = if p x then (x:ys. We define brk1 as brk1 [] line = [line] brk1 (w:ws) line = if length w + 1 + length line <= 80 then brk1 ws (line ++ " " ++ w) else line : brk1 (w:ws) [] .20] ([1. which splits a string into words separated by blanks.13.7. words s = case dropWhile isSpace s of "" -> [] s’ -> w : words s’’ where (w. perhaps an entire file.7.11. zs) else ([]." is a sentence. we can write words.14.2.9.18. Suppose we have a very long string of words.6. and unwords.10. it will make the line more than 80 characters long.[5.5 HASKELL 52 list.s’’) = break isSpace s’ unwords [] = "" unwords [w] = w unwords (w:ws) = w ++ ’ ’ : unwords ws Now we are ready for something more elaborate.15. They split a list into two parts depending on the value of a predicate. p) With these functions working. Here is the first version of a function that does the rest.. dropWhile p xs@(x:xs’) = if p x then dropWhile p xs’ else xs The functions break and span are complementary.16. When this happens.12.16.20]) Prelude> break (>=5) [1. as in mathematics.9.17.8. Note that one or more primes may be added to a variable name to give a new name.zs) = span p xs’ break p = span (not .19.3. we put the line into the output and carry on.4].

the actual definitions of take and takeWhile in the Prelude are: take n | n <= 0 = [] take [] = [] take n (x:xs) = x : take (n-1) xs . The next version corrects that error. we have to write fac n = if n <= 0 then 1 else n * fac(n-1) We can do better with guards. this function calculates how long the line will become if we append a blank and the word w to it. If the length is less than 80 characters. to make our function presentable to users. We start it off with an empty line: brk1 sentence [] This function has one error: the first character of every line that it outputs is a blank.5 HASKELL 53 If there is something in our word string w:ws. For example.6 Guards The pattern-matching mechanism is handy but rather inflexible because it only allows for exact matching. the word is moved into the line. A guard is a condition that qualifies a pattern. each with their appropriate definitions. Otherwise. As we know. Once we have provided a pattern. if we want a factorial function that does not blow up for negative numbers. brk2 [] line = [line] brk2 (w:ws) line = if then if line == "" then brk2 else brk2 else line : brk2 length w + 1 + length line <= 80 ws w ws (line ++ " " ++ w) (w:ws) [] Finally. For example. The pattern n | n < 0 matches any negative number. we wrap its parts up into a single definition: format sen = brk (words sen) [] where brk [] line = [line] brk ws@(w:ws’) line = if length w + 1 + length line <= 80 then if line == "" then brk ws’ w else brk ws’ (line ++ " " ++ w) else line : brk ws [] 5. This version of factorial works for all integers: fac n | n <= 0 = 1 | n > 0 = n * fc(n-1) The guard otherwise includes any conditions not covered by previous guards. the pattern n matches any number. the result is a list consisting of the line and the result of processing the rest of the word list.2. we can specify several guards.

The letter a is called a type variable. In the next example. the function will return either an integer or ⊥. This is because Haskell functions take their arguments “one at a time”. the compiler assumes that they are different (although they don’t have to be). declare i to be the identity function. In the formal type system. This means that. if the type of a function is Int. The type declaration is accepted in this form by the Haskell compiler. a → a and is read “for all types a. Semantically. For example. a and b. The next example is a constant function: it takes two arguments and throws away the second one. The type [a] is the type of a list whose elements have type a. First. the two arguments are used to build a list. Since there is nothing in the definition that connects the types. Haskell uses the Hindley-Milner type system. An expression is written without type declarations but. we follow the convention of declaring the type before the function.5 HASKELL 54 takeWhile p [] = [] takeWhile p (x:xs) | p x = x : takeWhile p xs | otherwise = [] 5. the type is a → b → a although you might have expected a × b → a. There are a small number of situations in which the type checker needs a type declaration to help it. For the rest of this section. we will discuss programs that do not fail. two :: a -> a -> [a] two x y = [x. they must have the same type. the type a to a”.3. Failure is discussed in Section 5. every type includes a value ⊥ (pronounced “bottom” because it is the minimal element of a lattice) that denotes failure. It can also be obtained using the :type command in the Hugs interpreter.2. In other words. i :: a -> a i x = x The Haskell type is a -> a which we can read as “the type of a function that takes an argument of any type and returns a result of the same type”. it will either give an answer of the expected type or it will fail. In all of these examples. if a program compiles without error. First. if it is type correct. k is a function that takes one argument of type a and returns a function with type b → a. The right arrow associates to the right and the type a → b → a is parsed as a → (b → a).3 Types Haskell is a strongly typed language.y] . the compiler can find a principle type or most general type for it. The following examples illustrate principle types. Consequently. k :: a -> b -> a k x y = x There are two points to note. this type is written ∀ a . The following tests verifies this: Types> k ’x’ k ’x’ :: a -> Char Types> k ’x’ 5 ’x’ :: Char The second point is simply that the type contains two type variables.

5 HASKELL 55 When we apply a function to an actual argument. Num a) => a -> a -> a n1 x y = if x > 1 then x + y else 0 If we write 1.5 instead of 1.y] else [] If we change == to >. Fractional a) => a -> a -> a n2 x y = if x > 1. but with the type Int specified by the programmer. Here is the same function again. because x and y both occur in the expression x + y. Asking Haskell for the type of this function shows that it has used the type we provided rather than the more general type. First.y] else [] Something else happens if we change one argument of > to the digit “1”.5 then x + y else 0 We can also specify the type of the entire function. but adds a requirement to the type: it must be an equality type. Providing a Char as the first argument of two forces the second argument to be Char as well. Note that we are not allowed to annotate a pattern — only an expression. asserting that a must be both an ordinal type and a numeric type. n2 :: (Ord a. The expression Eq => is called a context and it says that the type a must be an instance of the class Eq. we can add a type annotation to any variable or expression in the definition. We can specialize types in two ways. cmp :: Eq a => a -> a -> [a] cmp x y = if x == y then [x. Haskell allows this even without further type information. Types> two ’a’ ’b’ "ab" :: [Char] Types> two ’a’ 3 ERROR . Num a). If we specify that x is a Double. n1 :: (Ord a. also called an ordinal type.5 then x + y else 0 As stated above.Illegal Haskell 98 class constraint in inferred type *** Expression : two ’a’ 3 *** Type : Num Char => [Char] The next function introduces a new feature: it uses the equality predicate ==. The context becomes (Ord a. Haskell finds the most general type. The class Ord is a subclass of the class Eq: this means that any type that provides > must also provide ==. Types> two ’a’ two ’a’ :: Char -> [Char] The function two ’a’ accepts an argument of type Char but an argument of another kind gives a type error. gt :: Ord a => a -> a -> [a] gt x y = if x > y then [x. n4 :: Int -> Int -> Int n4 x y = if x > 1 then x + y else 0 Types> :t n4 . the compiler infers that y must be a Double too. we obtain yet another context: a must be an ordinal type and a fractional type. the types become specialized. we get a different requirement: the type a must be an instance of the class Ord. n3 :: Double -> Double -> Double n3 x y = if (x::Double) > 1.

In this example. If you don’t understand the type system. the type of x is d. the type (a. as the following example shows.5 HASKELL n4 :: Int -> Int -> Int 56 Of course. The most general type may be specialized by type annotations provided by the user. . compose :: (a -> b) -> (c -> a) -> (d -> c) -> d -> b compose f g h x = f (g (h x)) The type of a tuple is written in the same way as values of the type. h :: d → c.Type error in application *** Expression : x + y *** Term : x *** Type : Int *** Does not match : Char It is often easier to write functions than to figure out their types. The individual types are f :: a → b.hs":31 . we’d better get it right. The type inferred by the compiler is the most general type. which can get quite complicated. In this example. cp :: a -> b -> (a. the second argument of apply must be a list of as. If you do understand the type system. b) is “the type of pairs whose first element has type a and whose second element has type b”. we apply a function to a list. apply :: (a -> b) -> [a] -> [b] apply f [] = [] apply f (x:xs) = f x : apply f xs Types can get even longer. g :: c → a. Haskell errors are very often caused by type mismatches. Similar notation is used for longer tuples.1 Using Types Since Haskell infers types. In the next example. and the type of the result is b. you will find these errors very hard to understand.b) cp x y = (x.3.hs": Type checking ERROR "e:/courses/comp348/src/types. Haskell infers: that the most general type of the function is a → b. for two reasons: It is often useful to have a clear idea of the types of your variables when writing functions. 5. and the result is a list of bs. you will find them quite helpful. y) Summarizing what we have seen so far: All valid Haskell functions and expressions have a type. A context indicates that a type is an instance of a class. it is possible to virtually ignore types while programming. we provide an incorrect type for the same function: n5 :: Int -> Char -> Int n5 x y = if x > 1 then x + y else 0 This attempt does not fool the compiler: Reading file "e:/courses/comp348/src/types. This is unwise. if we do provide a type.

hs":9 . it is a common convention in Haskell (and other functional languages) to write the type declaration. if your module includes the declaration app :: [Int] -> [Int] -> [Int] app [] ys = ys app (x:xs) ys = x:app xs ys Haskell will accept the declaration.Type error in application *** Expression : app "abc" "def" *** Term : "def" *** Type : String *** Does not match : [Int] There are some occasions when the type system doesn’t work as you would expect. The correction is to replace ’ ’ by something which is a list. a list. It has type Char. part of the expression) which is causing the problem is ’ ’.Type error in application *** Expression : ’ ’ ++ w *** Term : ’ ’ *** Type : Char *** Does not match : [a] To make sense of this error: 57 In the line beginning ERROR. you are allowed to declare a more specialized type than the type that Haskell infers. the :9 after the path name is the line number. one version of the function brk function in Section 5. but the type checker is expecting something of type [a] — that is. It identifies then brk1 ws (line ++ ’ ’ ++ w) as the line with the error. The expression which cannot be typed is ’ ’ ++ w. You may have already discovered this problem when representing sets as lists: Lec1> union [] [] . The term (that is. in this case. However.5 looked like this: brk1 [] line = [line] brk1 (w:ws) line = if length w + 1 + length line <= 80 then brk1 ws (line ++ ’ ’ ++ w) else line : brk1 (w:ws) [] Haskell did not accept this version.2. [’ ’] or simply " ".5 HASKELL For example. When you are reading someone else’s function definitions. Haskell accepts these type declarations. For example. reporting: Reading file "e:/courses/comp348/src/lec1. knowing the types of the functions often makes it easier to understand them. and give an error if you try to use it incorrectly: Lec1> :info app app :: [Int] -> [Int] -> [Int] Lec1> app "abc" "def" ERROR .hs": Type checking ERROR "e:/courses/comp348/src/lec1. record the type you have specified. as inferred by the compiler. provided they are correctly formatted. above the function declaration. For this reason. and reports an error if there is an inconsistency between the declaration type and the inferred type.

4. whether or not it is producing output: Prelude> [1. do not detect stack overflow.n] and evaluate comb n with n = 0.3. the symptoms of failure can vary quite a lot. Domain errors are usually trapped by the runtime system: Prelude> 1/0 Program error: primDivDouble 1.Interrupted! Prelude> f where f = f Interrupted! Some implementations of Hugs.n] [1. A function fails to terminate.5 HASKELL ERROR .3..g.. If the expression union [] [] arises in a context.Cannot find "show" function for: *** Expression : [] *** Of type : [a] 58 This is a deficiency of the type system.5.5. This calls union [] [].0 0.Unresolved overloading *** Type : Eq a => [a] *** Expression : union [] [] In fact. the empty lists will be the result of some computation and their types will be known. but it is not an important one.6.2 Failure There are only two main causes of failure: A function is given an argument that is of the correct type but outside the domain of the function.2. with the result that programs may crash: Prelude> g 1 where g n = g(n+1) 5.3... but Haskell does not complain: Lec1> comb 0 [] :: [Integer] 5. Although the semantics says that the value of a failed program is always ⊥. E. Here is a contrived example: we define comb n = union [1.0 Prelude> head [] Program error: head [] A non-terminating program may run until it is interrupted. including WinHugs.] [1.3. even [] gives a type error: Prelude> [] ERROR .3 Reporting Errors There is a built-in function . head [] or 1/0.

and ⊥ (bottom). and Rational (ratios of Integers).5 HASKELL error :: String -> a 59 The result returned by this function is the only value that is shared by all types: ⊥ (pronounced “bottom”). As a side-effect.3. Here is an example of its use. Integer (integers of any size). ordinary variables. 5. These include Int (fixed size integers). There are a number of numeric types. The type Bool has values True and False. and function names.3. (/=) :: a -> a -> Bool -. or defined in libraries. (>=). (<=). min :: a -> a -> a -. if the user supplies one of these functions. An instance type must provide at least <=.4 Standard Types head[]" There are a number of types that are either built-in.5 Standard Classes The Standard Classes of Haskell are defined in the Prelude.Minimal complete definition: (<=) or compare -. However. Double (double precision floats). The type Char has the ASCII characters as its values. as shown by the mutually recursive definitions.Minimal complete definition: x == y = not (x/=y) x /= y = not (x==y) (==) or (/=) The ordinal class inherits from the equality class and provides additional functions for comparison. The type () has only two values: (). The equality class specifies that a type instance must provide the functions == and /=. class Eq a where (==). class (Eq a) => Ord a where compare :: a -> a -> Ordering (<). defined in the Prelude. which is the empty tuple. all of the other functions can be inferred by the compiler. The first letter of a type is always upper-case.using compare can be more efficient for complex types compare x y | x==y = EQ | x<=y = LT | otherwise = GT x <= y = compare x y /= GT . Float (single-precision floats). Here are a couple of examples. This is not a convention: the compiler uses the case of the first letter to distinguish type names from type variables. The definition of head in the Prelude is: head (x:xs) = x head [] = error "headPreuldeList: 5. This type corresponds roughly to void in C++. (>) :: a -> a -> Bool max. The type String is equivalent to [Char]. it displays the string. the compiler will provide the other one automatically.

5 HASKELL        60 Eq    Ord   Enum d d d ‚ d   d d d ‚ d        ©      d d d ‚ d       ©      Show Num Real Integral       ©     d d d ‚ d   d d d ‚ d        ©    Fractional RealFrac  d d d ‚ d        ©     d d d ‚ d     Floating RealFloat       ©     Figure 9: Standard Class Inheritance Graph x < y = compare x y == LT x >= y = compare x y /= LT x > y = compare x y == GT max x y | x <= y = y | otherwise = x min x y | x <= y = x | otherwise = y Figure 9 shows the class inheritance graph for the standard classes. To construct a fraction. even then. The name of the type is Frac and the name of the only constructor is F. These names do not have to be different: we could have used Frac for both. The class Show must provide a function that translates a value of the type into a string. we must provide two integers: Frac> F 9 12 9/12 :: Frac . because there is already a type Rational. but it is not defined in the Prelude. There is no need to memorize this graph: you will need it only if you are designing your own numeric type and. we discuss the type Frac. What is more important is to know the functions that a class is supposed to implement. as shown above.6 Defining New Types As an illustration of how programmers can define new Haskell types. you can get this information from the Prelude. 5. Haskell’s error messages will guide you. This type is actually unnecessary. shown in Figure 10.3. The first step is to declare the type and constructors for it.

1) ++ [F 1 n] Figure 10: The Class Frac ..Harmonic series: 1/1 + 1/2 + 1/3 + .5 HASKELL 61 module Frac where data Frac = F Integer Integer instance Show Frac where show (F n d) = if d == 1 then show n else show n ++ "/" ++ show d instance Eq Frac where F n d == F n’ d’ = n * d’ == n’ * d instance Ord Frac where F n d <= F n’ d’ = n * d’ <= n’ * d instance Num Frac where F n d + F n’ d’ = normalize( F (n * d’ + n’ * d) (d * d’) ) F n d . harm :: Integer -> [Frac] harm n | n <= 0 = [] | otherwise = harm (n .n’ * d) (d * d’) ) F n d * F n’ d’ = normalize( F (n * n’) (d * d’) ) negate (F n d) = normalize( F (negate n) d ) fromInteger n = F n 1 instance Fractional Frac where recip (F n d) = normalize( F d n ) normalize :: Frac -> Frac normalize (F n d) = if d == 0 then error ("Bad fraction: " ++ show (F n d)) else if n == 0 then F 0 1 else if d < 0 then normalize (F (negate n) (negate d)) else F (n ‘div‘ g) (d ‘div‘ g) where g = gcd n d -.F n’ d’ = normalize( F (n * d’ ..

we can see that any numeric class must be an instance of class Show. This enables us to define types like Vector. 1/3. as in these declarations: data Bool = False | True data Colour = Red | Green | Blue Haskell provides special syntax (if/then/else) for conditional expressions because programmers are familiar with the English words. If we do this. Finally and for fun. -. for example. The function that we choose is not arbitrary: if we provide <=. sum is defined for all instances of Num. but not divided. so it works correctly for Frac. we make Frac an instance of Fractional. because we could define if True x y = x if False x y = y relying on lazy evaluation to ensure that only one of x or y is actually evaluated. such as >=. The next step is to make Frac an equality type. From Figure 9. .5 HASKELL 62 We could introduce new function names for operations on fractions. . as here. All of these tasks are performed by normalize. and max. Caution: providing a different comparison. Figure 11 shows a few simple examples of the new type. compare. There are many advantages. addFrac might add two fractions. All that is needed is a function recip defined so that recip x = 1/x. Enumerations have several constructions but no data. The special syntax is actually unnecessary. and *. . For example. We want to work with fractions in “normal form”. We can provide an ordering for fractions. However. . . >. we can overload function names and. As a nice touch. Haskell will then define x/y = x × recip y. because Haskell treats -x as (negate x). add fractions using +. 5. >=. because we did not define sum. We use normalize for every operation that constructs a new fraction. We also want to check for zero denominators and put the fraction zero into the normal form 0/1. we introduce the function harm to construct a list of the fractions 1/1.3. We provide a version of the required function show that writes a fraction in the form "2/3". and Haskell automatically provides /=. The result is exact because the components of a Frac are Integer (“large integers”) rather than Int (32-bit integers). Note that the Num class does not contain division operations. 4/6 should be converted to 2/3 and 5/(−16) should be converted to −5/16. to making the type Frac an instance of numeric classes. subtracted. Since we do intend Fracs to be divided.7 Defining Data Types A type may have more than one constructor. however. which requires overloaded definitions of +. 1/2. which can be added. including <. The functions work as expected. min. We define ==. which uses the standard function gcd to cancel common factors. and a function that converts integers to Fracs. may lead to looping! The next step is to make Frac an instance of Num. 1/n. For example. The last example is of interest. Haskell will provide a number of other functions. it shows fractions with denominator equal to 1 as integers. and multiplied. It is useful also to define negate.

and a right subtree r. Note that this is a recursive type definition because BTree occurs on both sides of =.1/5] :: 63 [Frac] Frac> sum (harm 50) 13943237577224054960759/3099044504245996706400 :: Figure 11: Trying out type Frac Frac Binary Search Trees Constructors may also define data components.1/2. we need a base case (in this example the base case is Empty) and an inductive case. As an example. Here are some things to note about this definition: Function show uses a “helper” function with an additional argument to display trees. Just as with recursion in functions. A BTree is either Empty.1/4. data BTree = Empty | Tree Int BTree BTree . in which case it is ignored (an alternative would be to report an error). In Figure 12. height. using == is straightforward. or consists of an Int and two subtrees of type BTree.1/3. we define a type for binary trees with integers at the nodes. we will develop a type for binary trees with two constructors: Empty constructs an empty tree. The function insert assumes that the trees are binary search trees. A new entry is inserted in the left or right subtree unless it is equal to an entry already in the tree. Comparing trees for equality. Using the tree t defined at the bottom of Figure 12: BTree> t 1 2 3 4 5 6 8 :: BTree The functions size. The important thing to note is that the apparently unnecessary case Empty == Empty = True must be included — otherwise the function loops! Constructors may define data components.5 HASKELL Frac> F 3 4 > F 5 6 False :: Bool Frac> F 3 4 + F 7 12 4/3 :: Frac Frac> $$ > F 6 5 True :: Bool Frac> F 7 4 / F 2 5 35/8 :: Frac Frac> harm 5 [1. and treeToList are trivial to define by cases on the constructors. and Tree n l r constructs a tree with integer n at the root. a left subtree l.

5 HASKELL 64 module BTree where import Trace data BTree = Empty | Tree Int BTree BTree instance Show BTree where show t = sh t "" where sh Empty sp = "" sh (Tree n l r) sp = let sps = sp ++ " " in sh r sps ++ sp ++ show n ++ "\n" ++ sh l sps instance Eq BTree where Empty == Empty = True (Tree n l r) == (Tree n’ l’ r’) = n == n’ && l == l’ && r == r’ size Empty = 0 size (Tree n l r) = 1 + size l + size r height Empty = 0 height (Tree n l r) = 1 + max (height l) (height r) treeToList Empty = [] treeToList (Tree n l r) = treeToList l ++ treeToList r ++ [n] listToTree [] = Empty listToTree (n:ns) = insert n (listToTree ns) insert insert if m else else m Empty = Tree m Empty Empty m t@(Tree n l r) = < n then Tree n (insert m l) r if m > n then Tree n l (insert m r) t find Empty m = False find (Tree n l r) m = if m == n then True else if m < n then find l m else find r m merge x y = listToTree(treeToList x ++ treeToList y) Figure 12: Class BTree .

because this gives an ordered list of integers. The ordering of the list is significant: it would be natural to use inorder (left/root/right).3. In recursive calls. The size of a tree is the number of nodes that it contains and the height of a tree is the length of the longest path from the root to a leaf node. we make the type Btree an instance of class Show so that we can convert trees to strings.1] :: [Int] As with previous types. For example: BTree> t 6 1 2 5 4 :: 3 BTree .5 HASKELL 65 We can build trees using the constructors. This is a bit tedious.2. we want to be able to use the list to recreate the tree. and for this purpose postorder (left/right/root) is better.6. size Empty = 0 size (Tree n l r) = 1 + size l + size r height Empty = 0 height (Tree n l r) = 1 + max (height l) (height r) The function treeToList converts a tree to a list.4. and we will discuss easier ways later. Here are these functions in use: BTree> size t 6 :: Integer BTree> height t 5 :: Integer BTree> treeToList t [5. t = Tree 1 ( Tree 2 ( Tree 3 Empty ( Tree 4 Empty ( Tree 5 Empty Empty ) ) ) Empty ) ( Tree 6 Empty Empty ) The simple functions below illustrate how we use the constructors as patterns in functions that operate on trees. The implementation of show below uses a “helper function” sh with a second argument giving the amount of indentation to use when printing a node. However. the indentation is controlled by the depth of the node in the tree.

we must ensure that all entries in every left subtree are less than the root and that all entries in every right subtree are greater than the root. we must declare the Empty is equal to Empty — without this component of the definition. instance Eq BTree where Empty == Empty = True (Tree n l r) == (Tree n’ l’ r’) = n == n’ && l == l’ && r == r’ Let us assume that our binary trees are going to be used as binary search trees (BST). which are not allowed in a binary search tree. Surprisingly. Here is a function that takes integers from a list and builds a BST from them. The following code for the case m < n returns only part of the tree: if m < n then insert m l The function insert provides another way of building trees. Note that the last node in the expression is the first node to be inserted in the tree — and it becomes the root. 2. 1. maintaining the BST property. 0.5 HASKELL 66 We can easily define an equality relation on trees: trees are equal if the integers at their root nodes are equal and their subtrees are (recursively) equal. 8. the function loops. That is. 3. 6. 4] BTree> t2 8 7 6 5 4 . This version of function insert ignores them. 7. It is important to “rebuild” the parts of the tree that we have taken apart. The following function inserts an integer into a BST. it is still rather clumsy. listToTree [] = Empty listToTree (n:ns) = insert n (listToTree ns) t2 = listToTree [5. We have to decide what to do with duplicate entries. t1 = insert 1 (insert 8 (insert 3 (insert 2 (insert 5 (insert 2 (insert 6 (insert 4 Empty))))))) BTree> t1 8 6 5 4 3 2 1 Although this is a slight improvement over calling constructors.

2. as in Figure 13. 16. 12] we have BTree> merge t2 t3 20 19 18 17 16 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0 :: BTree 67 The function find uses the familiar algorithm to decide whether a given integer is stored in the tree. 20.3. We can test these functions with the trees that we already have.5 HASKELL 3 2 1 0 We have written treeToList and listToTree so that they are inverses. 11. Then find would match the key and return the corresponding value. 9.5. BTree> fringe t1 [1. 18. 10. we should store a key/value pair at each node.5. 17. 19. it is straightforward to define fringe. The fringe of a tree is a list of the leaf nodes as they would be visited in a left to right scan. 13. Assuming that we can recognize a leaf. with t3 = listToTree [14. To make trees useful. merge x y = listToTree (treeToList x ++ treeToList y) Then.7] :: [Int] . BTree> t2 == listToTree (treeToList t2) True :: Bool It is easy to merge trees using treeToList and listToTree.8] :: [Int] BTree> fringe t2 [0.

3. and Integer is an equality type. we revisit trees. we define trees with an arbitrary type.5 HASKELL isLeaf (Tree _ Empty Empty) = True isLeaf _ = False fringe Empty = [] fringe t@(Tree n l r) = if isLeaf t then trace ("Leaf " ++ show n ++ "\n") (n : (fringe l ++ fringe r)) else fringe l ++ fringe r sameFringe x y = fringe x == fringe y Figure 13: The fringe of a tree 68 There is a famous problem in Computer Science called the “same fringe problem”. Tree a indicates that the type variable a is a parameter of the type. The type . On the left side of the declaration. as shown in Figure 14. Here is a modified version of fringe that traces each discovery of a leaf node: fringe Empty = [] fringe t@(Tree n l r) = if isLeaf t then trace ("Leaf " ++ show n ++ "\n") (n : else fringe l ++ fringe r BTree> sameFringe t1 t2 Leaf 1 Leaf 0 False :: Bool 5. we must specify it as Tree a again. Given two trees.3. In the declaration below. but with a new twist. how do we decide whether they have the same fringe? Since list comparison is provided for any equality type. trees with integer nodes are not very useful.7 were defined with integers at the nodes. we have immediately: sameFringe x y = fringe x == fringe y BTree> sameFringe t1 t2 False :: Bool Let us investigate this more closely. Whenever we refer to the type on the right side. at the nodes. module Tree where data Tree a = Empty | T a (Tree a) (Tree a) In practice. indicated by the type variable a. The trees in Section 5. We need to store key/value pairs at the nodes so that we can store useful information in the tree.8 Types with Parameters (fringe l ++ fringe r)) This is what happens when we call sameFringe again with the traced version of fringe: In this section.

5144831733).value):ks) = insert key value (build ks) t = build [ ("Charles". 4508982212) ] Figure 14: Key-value tree .b)] -> Tree a b build [] = Empty build ((key. Show v) => Tree k v = Empty | T k v (Tree k v) (Tree k v) instance (Ord k. 8147586977).5 HASKELL 69 module Tree where data (Ord k. ("Bill". Show b) => [(a. ("Diana". ("George". Show a. 4163548695). 6063647586). ("Anne". Show b) => a -> b -> Tree a b -> Tree a b insert newKey newValue Empty = T newKey newValue Empty Empty insert newKey newValue t@(T key value left right) = if newKey == key then T key newValue left right else if newKey < key then T key value (insert newKey newValue left) right else T key value left (insert newKey newValue right) build :: (Ord a. ("Edward". Show v) => Show (Tree k v) where show Empty = "" show (T key value left right) = show left ++ leftjust (show key) 12 ++ show value ++ "\n" ++ show right leftjust :: [Char] -> Int -> [Char] leftjust s n | length s >= n = s | otherwise = leftjust (s ++ " ") n out t = putStr (show t) insert :: (Ord a. Show k. ("Feridoon". Show k. 2124756843). 4504657464). Show a.

"Anne"). we must now provide four arguments: the key. When we construct a non-empty tree."Bill"). and implications. Show v) => Show (Tree k v) where . that builds a tree from a list of pairs. disjunctions. <. ("Edward". 5144831733). 6063647586). we must repeat all of the constraints. For example.."Diana") ] There is a catch with parameterized types: when we use them as instances of a class. The following declaration introduces a type suitable for formulas in propositional calculus. Here is a test: t = build [ (6. ("Anne". with constants true and false. we can reduce the amount of writing required by providing a function. 8147586977). In the source file: t = build [ ("Charles". conjunctions. the value. and the right subtree. As before. The insertion function insert can use the operators ==. negations. 2124756843). ("George"."Charles"). (3.. module Logic where data Formula = TT | FF | Var String | Con [Formula] | Dis [Formula] | Neg Formula . ("Bill". the left subtree. build. (8. Show k. 4508982212) ] A Haskell session: Tree> t "Anne" 5144831733 "Bill" 4508982212 "Charles" 4163548695 "Diana" 2124756843 "Edward" 4504657464 "Feridoon" 6063647586 "George" 8147586977 :: Tree [Char] Integer Formulas It is often useful to define types with many constructors.5 HASKELL 70 variables k and v stand for “key” and “value” respectively. the new value replaces the old value.. ("Diana". We require that k is an ordered type so that we can compare keys. When we supply a new pair with a key that already exists in the tree. to make Tree an instance of Show. propositional symbols (or names). and > with keys because the type of the keys is an instance of class Ord. 4163548695). (1. we must write: instance (Ord k. 4504657464). ("Feridoon".

The function putChar takes a Char as its argument and performs an action — writing the character. return a Bool value that depends on the character read. although this is rather tedious.5 HASKELL | Imp Formula Formula 71 When we define a function for a type like this one. instance show show show show show show show show show show show Show TT = FF = (Var (Con (Con (Con (Dis (Dis (Dis (Neg (Imp Formula where "true" "false" name) = name []) = "" [f]) = show f (f:fs)) = show f ++ " and " ++ show (Con fs) []) = "" [f]) = show f (f:fs)) = show f ++ " or " ++ show (Dis fs) f) = "not " ++ show f p q) = show p ++ " => " ++ show q As before. read a character. The keyword do is followed by a sequence of actions. The function getChar performs an action and returns a value. Each action must have the same indentation. we must usually include cases for all of the constructors. in this case a Char.4 Input and Output Haskell handles input and output using monads. we will not discuss monads in detail. The following definition of show is incomplete because it does not indicate operator precedence.getChar putChar c The next example shows how to write functions that perform actions. f = (Imp p (Imp p (Imp p p))) where p = Var "p" g = Imp (Var "p") (Con [Var "q". instead we will just provide some simple examples of their use. The first program reads a character and echoes it. Neg (Var "r")]) Logic> f p => p => p => p :: Logic> g p => q and not r :: Formula Formula 5. we can build formulas using the constructors. Moands are pure functions that provide the effect of managing state changes.getChar return (char == ’y’) . test1 = do c <. ready = do putStr "Ready? " char <. In this course. The function ready performs the following sequence of actions: output a prompt.

getChar if char == ’\n’ then return "" else do line <. applies a function (switch) to it. The program test3 reads a line.dat" writeFile "e:/temp/test. process s = filter (not .getLine s <.getLine putStr "Output file: " oFile <. and then writes the line. the next example reads the input and output file names from the console.readFile iFile writeFile oFile s putStr (iFile ++ " copied to " ++ oFile ++ "\n") .ready if t then putStr "\nReady" else putStr "\nNot ready" The function getline in the next example is recursive.5 HASKELL 72 test2 = do t <. main = do putStr "Input file: " iFile <.getLine return (char:line) switch s = map sw s where sw char = if isLower char then toUpper char else if isUpper char then toLower char else char test3 = do line <. getline = do char <. The else part requires an action sequence that is introduced by a second do. isSpace) s test4 = do s <.res" (process s) Since it is not good practice to build file names into programs.getline putStr (switch line) The next example uses files instead of console functions.readFile "e:/temp/test.

This does not matter in practice.2 Interactive Programs The standard IO functions include interact. The effect of lazy evaluation is important here. The following exchange with the interpreter illustrates this (the standard function unlines converts a list of strings into a single string with the list items separated by newline characters): Prelude> unlines ["first line". because there is no need to format strings until you actually want to see them at the end of a calculation. The following example provides a more elaborate version of the same idea. The argument for interact is a function of type String -> String. The argument string is a sequence of “commands” interpreted as follows: Command ’n’ ’d’ ’q’ ’r’ Effect Insert a nickel (5 cents) Insert a dime (10 cents) Insert a quarter (25 cents) Demand a refund The output from the machine is a string defined as follows: String "" "candy" "candy and n cents change" "Refund: n" Condition the amount entered is less than 25 cents the amount entered was exactly 25 cents the amount entered was 25 + n cents response to ’r’ with n cents stored .5 HASKELL 5. a function that enables you to write interactive programs easily. converting lower case letters to upper case letters. you can use it only at the “top level” — that is. as a command that you enter or as part of a do sequence.4. "third line"]) first line second line third line Since putStr :: String -> IO () is an IO function. "third line"] "first line\nsecond line\nthird line\n" The IO function putStr interprets the formatting characters correctly: Prelude> putStr(unlines ["first line". Instead. The function machine :: String -> String simulates a vending machine. 5. it reads each character and then immediately echoes the converted character.4. "second line".1 String Formatting 73 By default. "second line". The effect is not to read a string of characters and then output the same string with lower case letters converted. Hugs does not interpret formatting characters from strings. A simple example of such a function is map toUpper. If you write interact (map toUpper) the interpreter will echo each character as you type it.

append [] ys = ys append (x:xs) ys = x : length [] = 0 length (x:xs) = 1 + length xs append xs ys (1) (2) (3) (4) Suppose that we want to define a new function len2 such that len2 xs ys returns the sum of the lengths of the lists xs and ys. Here is a simple definition: len2 xs ys = length (append xs ys) (5) . and hence the dialogue. The second example shows the response to "dddd" followed by ctrl-z. they are equations.. in an important sense. VM> interact machine candy candy and 5 cents change Refund: 10 Amount left: 0 VM> interact machine candy and 5 cents change Amount left: 10 VM> . We will illustrate this with a simple example of program manipulation. showing that the machine reports the amount it has left before terminating. The only difference is that we have numbered the equations for future reference.5 HASKELL Here is the definition of the function machine: 74 machine commands = mac commands 0 where mac "" stored = "Amount left: " ++ show stored mac (’q’:cs) stored = respond (stored+25) cs mac (’d’:cs) stored = respond (stored+10) cs mac (’n’:cs) stored = respond (stored+5) cs mac (’r’:cs) stored = "Refund: " ++ show stored ++ "\n" ++ mac cs 0 mac (_:cs) stored = mac cs stored respond stored cs = if stored == 25 then "candy\n" ++ mac cs 0 else if stored > 25 then "candy and " ++ show (stored-25) ++ " cents change\n" ++ mac cs 0 else mac cs stored The following dialog with Hugs shows the result when the user enters the string "qdddnnr". the user enters ctrl-z. The following definitions of append (append two lists) and length (length of a list) should be familiar by now.. To end the input string. 5.5 Reasoning about programs Haskell function definitions not only look like equations but..

5 HASKELL 75 The obvious problem with this definition is that it is inefficient: it constructs a new list just for the purpose of finding its length. we instantiate (5) with xs = (x:xs) and then use (2). The definition of isPrefix is isPrefix xs ys = ∃ as • xs ++ as == ys Assuming both lists are not empty. we can write this as isPrefix (x:xs’) (y:ys’) = ∃ as • (x:xs’) ++ as == (y:ys’) = x == y ∧ ∃ as’ • xs’ ++ as’ == ys’ = x == y ∧ isPrefix xs’ ys’ It is straightforward to add the base cases to obtain the following definition: . we assume that ys is not empty (and so can be written y:ys’) and that there are two cases: either bs = [] or bs = y:bs’ (clearly. on the basis of its required behaviour. of course. have obtained this definition of len2 with a little thought and no formal deduction. Then we have: isSubList xs (y:ys’) = ∃ as • [] ++ xs ++ as == ys’ ∨ ∃ bs’. if bs is not empty.as • bs’ ++ xs ++ as == ys’ = ∃ as • xs ++ as == ys’ ∨ isSubList xs ys’ (8) (9) (10) (11) The second disjunct is a straightforward recursion. (4). How can we avoid this inefficiency? As a first step.as • bs ++ xs ++ as == ys To make progress. The following definition is clearly correct but not. implementable (think of bs as the “before list” and as as the “after list”): isSubList xs ys = ∃ bs. Suppose we would like a function isSubList such that isSubList xs ys returns True if and only if xs is a sublist of ys. Here is a somewhat more complicated example in which the final answer is less obvious. we instantiate (5) with xs = [] and then use (1) to simplify the resulting expression: len2 [] ys = length (append [] ys) = length ys (6) (7) Next. as it stands. The first requires another function which. its first element must be y).as • y:bs’ ++ xs ++ as == y:ys’ = ∃ as • xs ++ as == ys’ ∨ ∃ bs’. and (5) to simplify the resulting expression: len2 (x:xs) = length (append (x:xs) ys = length (x : append xs ys) = 1 + length (append xs ys) = 1 + len2 xs ys We can now combine (7) and (11) to obtain a definition of len2 that does not use append: len2 [] ys = length ys len2 (x:xs) ys = 1 + len2 xs ys We could. we will call isPrefix.

5 HASKELL isSubList [] _ = True isSubList _ [] = False isSubList xs (y:ys) = isPrefix xs (y:ys) || isSubList xs ys where isPrefix [] _ = True isPrefix _ [] = False isPrefix (x:xs) (y:ys) = x == y && isPrefix xs ys 76 .

. if any. When we wish to solve a problem on a computer. which are central to discrete-event simulations. “programming” means “object oriented programming”. we often need to construct within the computer a model of that aspect of the real or conceptual world to which the solution of the problem will be applied. Key insights were developed in 1965 following experience with Simula I: the purpose of the language was to model systems. Simula 67 was a general purpose programming language that incorporated the ideas of Simula I but put them into a more general context.) . The basic concept of Simula 67 was to be “classes of objects”. . . buses. people. . page 489) has provided his own analysis of the role of blocks in Simula. depending on the system being simulated). as we would say today. Nevertheless. a superclass.6 The Object Oriented Paradigm A fundamental feature of our understanding of the world is that we organize our experience as a number of distinct object (tables and chairs. Once recognized. Simula originated in the Norwegian Computing Centre in 1962. Writing such programs is easier if there is a correspondence between objects in the world and components of the program. bank loans and algebraic expressions. (Hoare 1968) Object Oriented Programming (OOP) is the currently dominant programming paradigm. Deleting the procedure definitions and final statement from an Algol block gives a pure data record. a system is a collection of interacting processes. leaving procedure definitions and data declarations. The main lessons of Simula I were: the distinction between a program text and its execution. the fact that data and operations belong together and that most useful programming constructs contain both. a process can be represented during program execution by multiple procedures each with its own Algol-style stacks. the prefix concept could be used in any context where common features of a collection of classes could be abstracted in a prefixed block or. The major innovation was “block prefixing” (Nygaard and Dahl 1978). The queue mechanism (a list linked by pointers) can be split off from the elements of the queue (objects such as trucks. benefits it has to offer. and so on. there is a substantial portion of the software industry that has not adopted OOP. Deleting the final statement from an Algol block. Adding coroutine constructs to Algol blocks provides quasi-parallel programming capabilities. 6.1 Simula Many programs are computer simulations of the real world or a conceptual world. Prefixing emerged from the study of queues. Dahl (1978. For many people today. . and there remains widespread misunderstanding about what OOP actually is and what. gives an abstract data object. Kristen Nygaard proposed a language for simulation to be developed by himself and Ole-Johan Dahl (1978). . 77 .

the legacy that Simula left is considerable: a new paradigm of programming. classes that combine data and a collection of functions that operate on the data. Simula itself had an uneven history. Exercise 23. begin procedure Deposit (real amount) balance := balance + amount. 78 Adding a prefix mechanism to Algol blocks provides an abstraction mechanism (the class hierarchy). At any one time. It was used more in Europe than in North America. switches control to a scheduler which stores the value of the program counter somewhere and continues the execution of another function. executed within a function. a garbage collector that frees the programmer from the responsibility of deallocating storage. but it never achieved the recognition that it deserved.amount. To inherit from Account we write something like: Account class ChequeAccount (real amount). Simula provides: coroutines that permit the simulation of concurrent processes. A suspend statement. procedure Withdraw (real amount) balance := balance . with modernized syntax. end. What features might a concurrent language provide that are not provided by coroutines? What additional complications are introduced by concurrency? 2 . Before we use this class. This was partly because there were few Simula compilers and the good compilers were expensive. we must declare a reference to it and then instantiate it on the heap: ref (Account) myAccount. This explains the term “prefixing” — the declaration of the child class is prefixed by the name of its parent. exactly one function is executing. multiple stacks. Coroutines are typically implemented by adding two new statements to the language: suspend and resume. Listing 19 shows a simple Simula-style class. Note that the class is syntactically more like a procedure than a data object: it has an (optional) formal parameter and it must be called with an actual parameter. . needed to support coroutines.6 THE OBJECT ORIENTED PARADIGM Listing 19: A Simula Class class Account (real balance).. All of these concepts are subsumed in the Simula class. It differs from an ordinary Algol procedure in that its AR remains on the heap after the invocation.. On the other hand. Inheritance and subclasses introduced a new programming methodology. MyAccount := new Account(1000). prefixing (now known as inheritance) that allows a specialized class to be derived from a general class without unnecessary code duplication.. although this did not become evident for several years. A resume statement has the effect of restarting a previously suspended function.

Kay (1996) was influenced by LISP. In turn.2 Smalltalk Smalltalk originated with Alan Kay’s reflections on the future of computers and programming in the late 60s. Every object is an instance of a class (which must be an object). Smalltalk was inspired by Simula and LISP. and [Transcript nextPutAll: the parameter of the message. the talk inspired Carl Hewitt’s work on his Actor model. Since a class object must belong to a class. Smalltalk was also strongly influenced by Simula. 1. and everything is an object. execute. from class types. Since objects are “active” in the sense that they have methods. not continued. and by Simula. Sussman and Steele wrote an interpreter for Actors that eventually became Scheme (see Section 4. However. 6. one of Kay’s goals was to do the same for Smalltalk. you send a message to it. the message is timesRepeat:. Smalltalk requires metaclasses. “everything is an object”. The first version of Smalltalk was implemented by Dan Ingalls in 1972. Everything is an object. using BASIC (!) as the implementation language (Kay 1996. page 533). In particular. classes are objects in Smalltalk. Blocks may have parameters. Principle 6 reveals Kay’s use of LISP as a model for Smalltalk — McCarthy had described LISP with a one-page meta-circular interpreter. page 534). such as integer and real. and debug Smalltalk programs without ever leaving the Smalltalk environment. It is interesting to note that Kay gave a talk about his ideas at MIT in November 1971. To create a new instance of a class. Objects communicate by sending and receiving messages (in terms of objects).2). especially by its one-page metacircular interpreter. In the Smalltalk statement 10 timesRepeat: [Transcript nextPutAll: ’ Hi!’] ’ Hi!’] is the receiver is 10. 4. 5. Principles 4–6 provide an “internal” view and were revised following implementation experience.6 THE OBJECT ORIENTED PARADIGM 79 6. The component is . In Smalltalk. To [evaluate] a program list. not just a compiler. control is passed to the first object and the remainder is treated as its message. compile. The “block” is an interesting innovation of Smalltalk. A block is a sequence of statements that can be passed as a parameter to various control structures and behaves rather like an object. it was based on six principles (Kay 1996. One technique for iteration in Smalltalk is to use a do: message which executes a block (its parameter) for each component of a collection. Smalltalk effectively eliminates passive data. 2. The class holds the shared behaviour for its instances (in the form of objects in a program list). Smalltalk is a complete environment. 3. You can edit. Principles 1–3 provide an “external” view and remained stable as Smalltalk evolved. The cross-fertilization between OOPL and FPL that occurred during these early days has. it differs from Simula in several ways: Simula distinguishes primitive types. there are no data primitives. Objects have their own memory (in terms of objects). an early attempt at formalizing objects. sadly.

Exercise 24. particularly under Smalltalk’s approach. The second occurrence is its use. However: The resulting programming methodology is object-oriented: programs are developed by thinking about the objects they manipulate and then inventing a modular structure based on these objects. 2 Exercise 25. Simula and Smalltalk inherit this same limitation: they give us representations for classes. objects collaborate by exchanging “messages”. Algol 60 gives us a representation for procedures. every object is a member of a class. Each type of object is implemented by its own program module. there is an inheritance hierarchy (actually a tree) with the class Object as its root. The class is itself an object. how could the infinite regression by avoided? Is there a class that is a member of itself? 2 6. [Data objects are replaced by program structures. A class whose instances are classes is called a metaclass. all classes inherit directly or indirectly from the class Object. Discuss the following statements (Gelernter and Jagannathan 1990. there are no passive objects to be acted upon. The following loop determines whether "Mika" occurs in the current object (assumed to be a collection): self do: ^false [:item | item = "Mika" ifTrue: [^true]]. . page 245). but no direct representation for a procedure activation. garbage collection. In Smalltalk. E. (Liskov 1996. The first practical version of Smalltalk was developed in 1976 at Xerox Palo Alto Research Center (PARC). blocks. The important features of Smalltalk are: everything is an object. and can be destructive to a natural sense of hierarchy. Blocks are closely related to closures (Section 4. an object is an instance of a class.3 CLU CLU is not usually considered an OOPL because it does not provide any mechanisms for incremental modification.] Under this view of things. and therefore is a member of a class. The first occurrence of item introduces it as the name of the formal parameter for the block. The block [:x | E] corresponds to λx . an object has private data and public functions. Do these definitions imply an infinite regression? If so. a little program in itself. 2.6 THE OBJECT ORIENTED PARADIGM 80 passed to the block as a parameter. This can lead to counter-intuitive interpretations of simple operations. Our ability to create single objects and to specify (constant) object “values” is compromised. but no direct representations for the objects that are instantiated from classes. 1. coroutines. Every entity is an active agent.2). page 475) Parnas (1972) introduced the idea of “information hiding”: a program module should have a public interface and a private implementation. Liskov and Zilles (1974) introduced CLU specifically to support this style of programming.

a set of operations that may be performed on instances of the type. This simplifies safe initialization. (Martin 1986) The difference is that an ADT in CLU provides a particular implementation of the type (and. 3. some sets of objects. page 473) This is not a universally accepted definition. pop() has no effect) but would not provide an implementation of these functions. as well as the functions. . for all types. the object continues to exist (there may be other references to it).6 THE OBJECT ORIENTED PARADIGM 81 An abstract data type (ADT) specifies a type. but a stack ADT respecting Martin’s definition would define the key property of a stack (the sequence push(). 2. does not provide “type generators” or. a set of syntactic descriptions of the primitive functions. a sufficiently complete set of relationships that specify how the functions interact with each other. if needed. The stack-only model of Algol 60 was seen as too restrictive for a language supporting data abstractions. and the effects of these operations. Variable declaration is separated from object creation. the implementation must be unique) whereas Martin’s definition requires only a specification. Many people use the term ADT in the following sense: An abstract data type is a system consisting of three constituents: 1. The meaning of assignment is independent of type. it was seen as deficient in various ways. of a class. Simula: does not provide encapsulation: clients could directly access the variables. Although Simula influenced the design of CLU. When the variable goes out of scope. A stack ADT in CLU would have functions called push and pop to modify the stack. There is no need to worry about the semantics of copying an object. of course) to E. The statement x := E means simply that x becomes a reference (implemented as a pointer. “generic types” (“templates” in C++). The reasons given for using a heap include: Declarations are easy to implement because. instances of user-defined classes are always heap-allocated. associates operations with objects rather than types. (Liskov 1996. as we would say now. or not built in to the language: I referred to the types as “abstract” because they are not provided directly by a programming languages but instead must be implemented by the user. handles built-in and user-defined types differently — for example. An abstract type is abstract in the same way that a procedure is an abstract operation. The word “abstract” in CLU means “user defined”. The implementation of an ADT provides the actual operations that have the specified effects and also prevents programmers from doing anything else. Variable and object lifetimes are not tied. Stack allocation breaks encapsulation because the size of the object (which should be an implementation “secret”) must be known at the point of declaration. Copying. CLU’s computational model is similar to that of LISP: names are references to heap-allocated objects. should be provided by a member function of the class. a semantic description — that is. in fact. space is allocated for a pointer.

conversely. and down. which converts the representation type to the abstract type. page 63). auxiliary functions that are not accessible to clients. . . .. x: int) returns (bool) return (getind(s. CLU is designed for programming with ADTs. and in the function member. An object is mutable if its value can change during execution. such as intset. end intset 82 Reference semantics requires garbage collection. The components of a cluster. Associated with the interface is an implementation that defines a representation for the type and provides bodies for the functions. integers are immutable but arrays are mutable (Liskov and Guttag 1986. introduced by the keyword rep. the formal parameter s:cvt indicates that the client will provide an intset which will be treated in the function body as a rep. CLU provides the keyword cvt as an abbreviation. which converts the abstract type to the representation type.. . Furthermore. the function insert has a parameter s of type intset which is converted to the representation type so that it can be treated as an array. For both of these cases.. choose rep = array[int] create = proc () returns (cvt) return (rep$new()) end create insert = proc (s: intset. referred to as rep. efficiency was not a primary goal of the CLU project. insert. pages 19–20). x: int) if ∼member(s. x) ≤ rep$high(s)) end member . notorious sources of obscure errors. x) end end insert member = proc (s: cvt. using the syntax type$operation.6 THE OBJECT ORIENTED PARADIGM Listing 20: A CLU Cluster intset = cluster is create. member. For example. . For example. Implementations of the functions mentioned in the header and. A header that gives the name of the type being implemented (intset) and the names of the operations provided (create. CLU draws a distinction between mutable and immutable objects. The abstract type intset and the representation type array[int]. if necessary. CLU provides special functions up. delete. It often happens that the client provides an abstract argument that must immediately be converted to the representation type. All operations on abstract data in CLU contain the type of the operator explicitly. are: 1. 3. garbage collection eliminates memory leaks and dangling pointers. Listing 20 shows an extract from a cluster that implements a set of integers (Liskov and Guttag 1986. otherwise it is immutable. An intset is represented by an array of integers. insert. cvt converts the new instance of rep to intset before returning.). Although garbage collection has a run-time overhead. In the function create above. x) then rep$addh(down(s). 2. A definition of the representation type. many functions create a value of the representation type but must return a value of the abstract type. size. are distinct.

6 THE OBJECT ORIENTED PARADIGM 83 The mutability of an object depends on its type and. the fact that it is shared is undetectable by the program. C++ is the result of two major design decisions: . In Simula. page 100) Exercise 26. . The cost is an increase in complexity. records. and clusters are three different things. just as C programmers use both cc (the C compiler) and lint (the C style checker)? The designers of CLU advocate defensive programming : Of course. Several keywords (rep. It is better to develop the habit of defensive programming. parameters of functions] are in the permitted subset of the domain takes time. cvt. . changes made via the reference x will be visible via the reference y. Do you agree that defensive programming is a better habit than the alternative suggested by Liskov and Guttag? Can you propose an alternative strategy? 2 6. that it writing each procedure to defend itself against errors. and records — are effectively made equivalent. . classes.. But is it the task of a compiler to prevent the programmer doing things that might be completely safe? Or should this role be delegated to a style checker. and up) are needed simply to maintain the distinction between abstract and representation types. This is generally an unwise practice.e. After executing x: T = intset$create() y: T y := x the variables x and y refer to the same object. The task assigned to Stroustrup was to develop a new systems PL that would replace C. In CLU. names are references. distinction between the abstract type and the representation type. safe encapsulation. Clearly. the distinction between mutable and immutable objects. and suppress them during production. and it is tempting not to bother with the checks. not values. (Liskov and Guttag 1986. down. . CLU provides the same tools and enforces the methodology. notably type parameters). an object is mutable if its type has operations that change the object. procedures. checking whether inputs [i. If the object is immutable.4 C++ C++ was developed at Bell Labs by Bjarne Stroustrup (1994). . Defensive programming makes it easier to debug programs. in particular. three concepts — procedures. The important contributions of CLU include: modules (“clusters”). . Whether this is desirable depends on your point of view. resulting in significant simplification. generic clusters (a cluster may have parameters. on the operations provided by the type. analysis of copying and comparison of ADTs. or to use them only while debugging. The argument in favour of CLU is that the compiler will detect encapsulation and other errors. If it is mutable. Whereas Simula provides tools for the programmer and supports a methodology. exception handling.

Alphard. provides multiple inheritance. procedure variables. Eiffel: is strictly OO: all functions are defined as methods in classes.6 THE OBJECT ORIENTED PARADIGM 84 first.5. Explain why C++ is the most widely-used OOPL despite its technical flaws. Sakkinen 1992. it provides multiple inheritance. however. break. may provide garbage collection. Stroustrup decided that the new language would be adopted only if it was compatible with C. provides multiple inheritance and repeated inheritance. object names are references. scientifically specified. 6.. and exception handling. Like some other OOPLs. CLU. although both stack and heap allocation is provided. goto. C++ is widely used but has been sharply criticized (Sakkinen 1988. or pointer arithmetic. is a hybrid language (i. and the specification language Z. or continue statements. whose goal is to yield quality software through the careful production and continuous development of parameterizable. . By default. Eiffel embodies a “certain idea” of software construction: the belief that it is possible to treat this task as a serious engineering enterprise.1 Programming by Contract The problem with writing a function to compute square roots is that it is not clear what to do if the argument provided by the caller is negative. Ada. communicating on the basis of clearly defined contracts and organized in systematic multi-criteria classifications. at Cambridge University using Simula convinced him that object orientation was the correct approach but that efficiency was essential for acceptance. Eiffel does not have: global variables. The solution adopted in Listing 21 is to allow for the possibility that the argument might be negative. Joyner 1992). and assertions. emphasizes the stack rather than the heap. Algol W.5 Eiffel Software Engineering was a key objective in the design of Eiffel. there are only a few C constructs that are not accepted by a C++ compiler). (Meyer 1992) Eiffel borrows from Simula 67. second. Stroustrup’s experience of completing a Ph. casts. 2 6. supports “programming by contract” with assertions. It is notable for its strong typing. reusable components. C++: is almost a superset of C (that is. enumerations. Eiffel also provides “expanded types” whose instances are named directly. provides generic classes.e. The main features of Eiffel are as follows. has an unusual exception handling mechanism.D. not a “pure” OO language. genericity (in the form of “templates”). subranges. does not provide garbage collection. an unusual exception handling mechanism. Exercise 27. a language that supports both imperative and OO programming). Input/output is defined by libraries rather than being built into the language.

it is undesirable to return a funny value such as −999. because programmers will rarely call sqrt with a negative argument. or to send a message to the console (there may not be one). .999.3). In particular. . applied throughout a software system.999 or any other number. x. // x else ?? } Listing 22: A contract for square roots sqrt (x: real): real is require x ≥ 0.6 THE OBJECT ORIENTED PARADIGM Listing 21: Finding a square root double sqrt (double x) { if (x ≥ 0. In words: If the actual parameter supplied. (Hoare considers this policy to be analogous to wearing your life jacket during on-shore training but leaving it behind when you go to sea.1. . The caller therefore has a responsibility to ensure that the argument is valid. is hard to define. . terminate execution. the contract does not constrain the function in any way at all: it can return -999.) Require and ensure clauses are a useful form of documentation. Programs are not cluttered with tests for conditions that are unlikely to arise. The require/ensure clauses constitute a contract between the function and the caller. The only reasonable possibility is some kind of exception. Meyer’s solution is to write sqrt as shown in Listing 22. The failure action. or to add another parameter to indicate success/failure. indicated by “??” in Listing 21. The require conditions can be checked dynamically during testing and (perhaps) disabled during production runs. Note . The advantages of this approach include: Many tests become unnecessary. as shown in Listing 23.x ensure abs(Result * Result / x . Testing arguments adds overhead to the program. they appear in Eiffel interfaces. Most of the time spent testing is wasted. or whatever. . is positive. -. is called defensive programming (Section 6. .0) √ return . If the argument supplied is negative.0 √ Result := .0) ≤ 1e-6 end 85 This approach. the result returned will be approximately √ x. We must have Rp ⇒ Rc and Ec ⇒ Ep. Functions in a child class must provide a contract at least as strong as that of the corresponding function in the parent class.

It must be true when the object is created and after each function of the class has executed.. However: this does not seem to be done much in practice. In addition to require and ensure clauses. The following analogy explains this reversal. 86 the directions of the implications: the ensure clauses are “covariant” but the require clauses are “contravariant”.. it contains one more element”. Eiffel programmers are not obliged to write assertions. it is possible in principle to say things like: “this structure is a heap (in the sense of heap sort)”. complex assertions increase run-time overhead.5. a complex assertion might itself contain errors... A class invariant is a predicate over the variables of an instance of the class. Eiffel has provision for writing class invariants. but it is considered good Eiffel “style” to include them. The features of HomeBusiness include: . Eiffel achieves the subcontract specifications by requiring the programmer to define Rc ≡ Rp ∨ · · · E c ≡ Ep ∧ · · · The ∨ ensures that the child’s requirement is weaker than the parent’s requirement. and the ∧ ensures that the child’s commitment is stronger than the parent’s commitment. 6.2 Repeated Inheritance Listing 24 shows a collection of classes (Meyer 1992. I will send you an animal”. Typical contracts say things like: “after you have added an element to this collection. I will send you a dog”. The pet supplier has a subcontractor who offers the following contract: “if you send me at least $150. Thus the contractor can use the subcontractor when a dog is required and make a profit of $50. A pet supplier offers the following contract: “if you send at least $200. In practice. class Child inherit Parent feature f() is require Rc ensure Ec . These contracts respect the Eiffel requirements: the contract offered by the subcontractor has a weaker requirement ($150 is less than $200) but promises more (a dog is more specific than an animal).. leading to a false sense of confidence in the correctness of the program.6 THE OBJECT ORIENTED PARADIGM Listing 23: Eiffel contracts class Parent feature f() is require Rp ensure Ep end . page169). increasing the incentive to disable checking. it is sometimes difficult to write assertions that make non-trivial claims about the objects.. Since assertions may contain function calls.

6. inherited once but by two paths. end 87 address. businessValue.. However. but Java is simpler in many ways than C++. more than one sequence can be obtained by repeated inheritance of the successor function. a rescue clause may perform some cleaning-up actions and then invoke retry to attempt the calculation again. Early versions of Eiffel had a reference semantics (variable names denoted references to objects). inherited from value in class House. inherited from value in class House. more conventional. it is difficult in some languages to provide collections with more than one iteration sequence. A return from the rescue clause indicate that the function has failed. It is not obvious that there are many circumstances in which it makes sense to “retry” a function.3 Exception Handling An Eiffel function must either fulfill its contract or report failure.6 THE OBJECT ORIENTED PARADIGM Listing 24: Repeated inheritance in Eiffel class House feature address: String value: Money end class Residence inherit House rename value as residenceValue end class Business inherit House rename value as businessValue end class HomeBusiness inherit Residence Business . Meyer introduced expanded types. If a function contains a rescue clause. Why do you think expanded types were introduced? Was the decision to introduce them wise? 2 6. exception handling mechanisms. but it applies to entire classes (via the virtual mechanism) rather than to individual attributes of classes. After Eiffel had been used for a while. this clause is invoked if an operation within the function reports failure. Expanded types introduce various irregularities into the language... Key features of Java include the following.6 Java Java (Arnold and Gosling 1998) is an OOPL introduced by Sun Microsystems. Repeated inheritance can be used to solve various problems that appear in OOP. In Eiffel. residenceValue.5. The mechanism seems harder to use than other. C++ has a similar mechanism. Exercise 28. Its syntax bears some relationship to that of C++. because they do not have the same properties as ordinary types. . An instance of an expanded type is an object that has a name (value semantics). The function in Listing 25 illustrates the idea. For example.

For some applications.initially false do if not triedFirst then firstmethod else secondmethod end rescue if not triedFirst then triedFirst := true retry else restoreInvariant -. in which program statements are compiled immediately before execution. parts of the program that are not needed during a particular run (this may be a significant fraction of a complex program) are not compiled. is not strongly relevant to this course. More efficient implementations are promised.6 THE OBJECT ORIENTED PARADIGM Listing 25: Exceptions in Eiffel tryIt is local triedFirst: Boolean -. The technique is quite old — Pascal P achieved portability in the same way during the mid-seventies — and has the disadvantage of inefficiency: Java programs are typically an order of magnitude slower than C++ programs with the same functionality. Java offers security . Java byte codes do not have the potential to penetrate system security in the way that a binary executable (or even a MS-Word macro) can. Java has an exception handling mechanism. such as simple Web “applets”. the inefficiency is not important. as such.report failure end end 88 Java is compiled to byte codes that are interpreted.1 Portability Compiling to byte codes is an implementation mechanism and. A variable name in Java is a reference to an object. The byte codes are checked by the interpreter and have limited functionality. Java provides concurrency in the form of threads. However. are not objects in Java. Java has a class hierarchy with class Object at the root and provides single inheritance of classes. Java provides interfaces with multiple inheritance. for each primitive type. Java is highly portable. Since any computer that has a Java byte code interpreter can execute Java programs. Primitive values. In addition to classes. such as int. The portability of Java is exploited in network programming: Java bytes can be transmitted across a network and executed by any processor with an interpreter. Java provides wrapper classes. Consequently. . such as Integer. Java provides garbage collection. 6. One interesting technique is “just in time” compilation.6.

We will call such instances “exceptions”. The throw statement is used to signal the fact that an exception has occurred. Interfaces provide a way of describing and factoring the behaviour of classes. A class may implement an interface by providing definitions for each of the methods declared in the interface. is a compile-time entity whose run-time instances are objects. particularly for Web programming. the class hierarchy is a rooted tree but the interface hierarchy is a directed acyclic graph. (In fact.) A Java class can inherit from at most one parent class but it may inherit from several interfaces provided..2 Interfaces A Java class.6. we need a way of sorting them as well: this requirement is specified by the second interface. .6 THE OBJECT ORIENTED PARADIGM Listing 26: Interfaces 89 interface Comparable { Boolean equalTo (Comparable other). Class Widget is required to implement equalTO and lessThan: class Widget implements Ordered { .. The first interface ensures that instances of a class can be compared. Consequently. 6.) If we want to store instances of a class in an ordered set. Instances of a subclass of class Exception are used to convey information about what has gone wrong. as usual in OOP. Listing 26 shows a couple of simple examples of interfaces. // Return True if ‘this’ object precedes ‘other’ object in the ordering. Thus an interface has no instances. } interface Ordered extends Comparable { Boolean lessThan (Ordered other). all Java classes inherit the function equals from class Object that can be redefined for comparison in a particular class.. (The class may also make use of the values of the constants declared in the interface.3 Exception Handling The principal components of exception handling in Java are as follows: There is a class hierarchy rooted at class Exception. An interface declaration is similar to a class declaration. } The portability of Java has been a significant factor in the rapid spread of its popularity. // Return True if ‘this’ and ‘other’ objects are equal. The argument of throw is an exception. } 6. of course. that it implements all of them.6. but introduces only abstract methods and constants.

The following simple example indicates the way in which exceptions might be used. as shown in Listing 30. Listing 27 shows the declaration of an exception for division by zero. we suppose that either of these problems may occur in a particular computation. Listing 29: Throwing exceptions public void bigCalculation () throws DivideByZero. . 6. Finally. we write some code that tries to perform the computation. . if (. .... Threads are defined by a class in that standard library. A try statement may use one or more catch statements to handle exceptions of various types... as shown in Listing 28 Next.. if (. } } 90 The try statement is used to attempt an operation and handle any exceptions that it raises. value = badValue. NegativeSquareRoot { .) throw new DivideByZero().. as shown in Listing 29..4 Concurrency Java provides concurrency in the form of threads. NegativeSquareRoot(float badValue) { super("Square root with negative argument"). Java threads have been criticized on (at least) two grounds.6. . } } Listing 28: Another Java exception class public class NegativeSquareRoot extends Exception { public float value. The next step is to declare another exception for negative square roots.6 THE OBJECT ORIENTED PARADIGM Listing 27: Declaring a Java exception class public class DivideByZero extends Exception { DivideByZero() { super("Division by zero"). . } . .) throw new NegativeSquareRoot(x). .. .

during execution. starting with Simula. The problem with this approach is that there are many classes which people understand intuitively but which are not easily defined in taxonomic terms.6 THE OBJECT ORIENTED PARADIGM Listing 30: Using the exceptions { try { bigCalculation() } catch (DivideByZero) { System. Antero Taivalsaari (1993) has given a thorough account of the motivation and design of his language. } } } 91 The synchronization methods provided by Java (based on the early concepts of semaphores and monitors) are out of date and do not benefit from more recent research in this area. Common examples include “book”. “chair”. In a prototype-based OOPL. In a class-based language.println("Argument for square root was " + n). Exercise 29. } catch (NegativeSquareRoot n) { System."). (In Smalltalk. instances of these classes are created and become the objects of the system. family of OOPLs that use prototypes rather than classes. “game”. Prototypes provide an alternative to classes.out. the programmer defines one or more classes and.7 Kevo All of the OOPLs previously described in this section are class-based languages. we can clone a . page 172) points out that class-based OOPLs. Although there are several prototype-based OOPLs. Brinch Hansen (1999) argues that Java’s concurrency mechanisms are unsafe and that such mechanisms must be defined in the base language and checked by the compiler. smaller. Alternatively.out. each new type of object is introduced by defining a prototype or exemplar . are based on an Aristotelian view in which the world is seen as a collection of objects with well-defined properties arranged [according] to a hierarchical taxonomy of concepts. Can you suggest any other advantages of using byte codes? 2 6.) There is another. The concurrency mechanism is insecure. etc. Byte codes provide portability. Kevo. Taivalsaari (1993. classes are run-time objects. but Smalltalk is nevertheless a class-based language. which is itself an object. we discuss only one of them here. not provided separately in a library. Identical copies of a prototype — as many as needed — are created by cloning .println("Oops! divided by zero.

drawContents. modify it in some way. .drawFrame and draws a frame with a title. . with initial value "Anonymous". 2 . Example 16: Objects in Kevo. This is misleading. . but the third. refresh. ENDADDS.drawFrame self. Listing 31 shows the introduction of a window object. TitleWindow with Window as a prototype. The window has two components: rect and contents. The properties of contents are left unspecified in this example. Listing 32 shows how to add a title to a window. Some of its methods will probably be the same as those of the original object. It declares a new object. 92 prototype. Delegation thus plays the role of inheritance in a prototypebased language and. This example illustrates how new objects are introduced in Kevo. .new → TitleWindow TitleWindow ADDS VAR title "Anonymous" → title METHOD drawFrame . . prototype-based languages are sometimes called delegation languages. . properties are added to the window using the module operation ADDS. The first line. however. The first two are left undefined here.new → rect VAR contents METHOD drawFrame . pages 187–188).. In the next block of code. . REF Window. it is based on (Taivalsaari 1993.new → Window Window ADDS VAR rect Prototypes. title. . declares a new object with no properties. because delegation is not the only possible mechanism for inheritance with prototypes.. for this reason. Window receives the basic printing and creation operations defined for Object. and these methods can be delegated to the original object rather than replicated in the new object. ENDADDS. The modified object need not be self-contained. draws the window frame first and then the contents.6 THE OBJECT ORIENTED PARADIGM Listing 31: Object Window REF Window Object. In Kevo terminology.. The properties of rect are obtained from the prototype Rectangle. Listing 32: Object Title Window REF TitleWindow Window. Finally. and use the resulting object as a prototype for another collection of identical objects. and provides a new method drawFrame that overrides Window. METHOD drawContents .Rectangle. three new methods are added to Window. METHOD refresh self. . Then it adds a new variable. The second line defines Window as a clone of Object.

Many OOPLs since Smalltalk have withdrawn from its radical ideas. In this section. CLOS also provides “before” and “after” methods — code defined in a superclass that is executed before or after the invocation of the function in a class. Beta is a descendant of Simula. This was the feature of Simula that attracted Alan Kay and led to Smalltalk (Kay 1996). separating code and data. developed in Denmark. each with its own code and stack. although Ada 9X has a number of OO features. Algol 68 and Pascal provided recursive data structures but maintained the separation between code and data. This is an impressive achievement.6 THE OBJECT ORIENTED PARADIGM 93 6. 6. Considerable effort has been put into efficient implementation. we discuss various issues that arise in OOP. C++ was originally called “C with classes” (Stroustrup 1994) and modern C++ remains an Algol-like systems programming language that supports OOP.1 Algorithms + Data Structures = Programs The distinction between code and data was made at an early stage. to the extent that Self programs run at up to half of the speed of equivalent C++ programs.9 Issues in OOP OOP is the programming paradigm that currently dominates industrial software development. Blue (K¨lling 1999) is a language and environment designed for teaching OOP. Blue. LISP brought algorithms (in the form of functions) and data structures together. but: the mutability of data was downplayed. including “multi-methods” — the choice of a function is based on the class of all of its arguments rather than its first (possibly implicit) argument. Self is the perhaps the best-known prototype language. given that Self is considerably more flexible than C++. The language o was influenced to some extent by Dee (Grogono 1991b). but lack of time precludes further discussion of them. The environment is interesting in that users work directly with “objects” rather than program text and generated output. CLOS. 6. Self.8 Other OOPLs The OOPLs listed below are interesting. It differs in many respects from “mainstream” OOPLs. It is visible in FORTRAN and reinforced in Algol 60. It has a number of interesting features.9. Ada has shown even more reluctance to move in the direction of OOP. Beta. Simula started the move back towards the von Neumann Architecture at a higher level of abstraction than machine code: the computer was recursively divided into smaller computers. subsequent FPLs reverted to the Algol model. . Algol provided the recursive block structure for code but almost no facilities for data structuring. The Common LISP Object System is an extension to LISP that provides OOP.

e.e. The representations 7. . Objects need representations. 1 }. For example. We can copy a representation of a value. It does not make sense to talk about the identity of a value. it is no longer 3. a counter might have 6 as its current value. unique identities: consider identical twins. statements about values on the left side of the page are balanced by statements about objects on the right side of the page... An object is a particular collection of subobjects and values. we can increment a counter. Objects possess values as attributes. but copying a value is meaningless. For example. Although these sets have different representations (the ordering of the members is different). innocent people have been arrested and jailed because they had the misfortune to have the same name and social security number as a criminal. age=55] [name="Peter". and a particular combination of bits in the memory of a computer are different representations of the same abstract value. We need representations in order to compute with values. If we change the value 3. obtaining 5. A counter has one attribute: the value of the count. they have the same value because in set theory the only important property of a set is the members that it contains: “two sets are equal if and only if they contain the same members”. For example. VII. may change). Is it the same 5 that we obtain from the second calculation? Identity is an important attribute of objects.9. Values Objects Integers have values. Consider the sets { 1. Objects are concrete. It does not make sense to change a value. and then 7 − 2. again obtaining 5.6 THE OBJECT ORIENTED PARADIGM 6. they might refer to different people who have the same name and age. Values are abstract. An object has subobjects and values as attributes. The value 7 is the common property shared by all sets with cardinality 7. age=55] Although these objects have the same value. mutable (i. 2. 3. Suppose we evaluate 2 + 3. sju. to change the value of an object. Consider two objects representing people: [name="Peter". do not change). Objects are usually. sept.2 Values and Objects 94 Some data behave like mathematical objects and other data behave like representations of physical objects (MacLennan 1983). In the following comparison of values and objects. but not necessarily. Even objects with the same attributes may have their own. Values are immutable (i. the integer 6 is a value. 3 } and { 2. and often useful. It is meaningful. Values need representations. The consequences of assuming that these two objects denote the same person could be serious: in fact.

but may have other desirable properties such as encapsulation. it may even be useful. and is viewed as harmful. may provide objects. Large values are best represented by references (pointers to the data structure representing the value. Most objects should be represented by references (pointers to the data structure holding the object. whether two instances of a value are stored in the same place or in different places. Aliasing is not a problem for values. Sometimes. Programs that use values only have the property of referential transparency. because they do not change. This is sometimes called “aliasing”. . but it is in fact desirable for objects with identity. for example. Since they have unbounded size. An object is equal to itself. Very small objects may be held in machine words. A person who is arrested because his name and social insurance number is the same as those of a wanted criminal may not feel “equal to” the criminal. known as “integrity violations” in the database community. The assignment operator for objects should be a reference assignment.6 THE OBJECT ORIENTED PARADIGM 95 Copying an object certainly makes sense. Exercise 30. We consider different representations of the same value to be equal. There is no need to copy values. Programs that provide objects do not provide referential transparency. The assignment operator for values should be a reference assignment. Objects should not normally be copied. many PLs treat them as objects. What is “primitive” and what is “user defined”? The PL should minimize the distinction as much as possible. All references to a particular object should refer to the same object. multiple copies of an object with a supposedly unique identity can cause problems. Strings should behave like values. as when we copy objects from memory to disk. Small values can be held in machine words. The distinction between values and objects is not as clear as these comparisons suggest. doubts begin to arise. When we look at programming problems in more deatil. Object oriented PLs should provide objects. However. It is immaterial to the programmer. Functional and logical PLs usually provide values. What is “small” and what is “large”? This is an implementation issue: the compiler must decide. because copying endangers their integrity. Should sets be values? What about sets with mutable components? Does the model handle objects with many to one abstraction functions? What should a PL do about values and objects? 2 Conclusions: The CM must provide values. and should not even be detectable. Whether two distinct objects with equal attributes are equal depends on the application.

6 THE OBJECT ORIENTED PARADIGM 96 Copying and comparing are semantic issues. these occasions should become rarer. but lower level.5 Pure versus Hybrid Languages Smalltalk. x = y is interpreted as x. For example. Prototypes provide more flexibility and are better suited to interactive environments than classes. it was inevitable that C++ could be a hybrid language. This has the effect of making binary operations asymmetric. and Java are pure OOPLs: they can describe only systems of objects and programmers are forced to express everything in terms of objects. Once the decision was made to extend C rather than starting afresh (Stroustrup 1994.3 Classes versus Prototypes The most widely-used OOPLs — C++.9. “send the message equal with argument y to the object x”.9. and especially those trained in OO techniques. A better question might be: suppose a PL provides OOP capabilities.equal(y) or. Prototype languages are a distinct but active minority. and is this comparison commutative? Some of the problems can be avoided by associating functions with types rather than with objects. Eiffel.9. “C with classes”. In fact. even this approach has problems if x and y do not have the same type. These programmers will use OO techniques most of the time and will only resort to procedural techniques when these are clearly superior to the best OO solution. with this interpretation. functions are associated with objects. An expression such as x = y is interpreted in the usual mathematical sense as equal(x. what? One answer is: FP . OO design is performed perfunctorily and design problems are solved by reverting to the more general. can we compare an instance of C to an instance of P . its precursor.) For experienced programmers. However. and Eiffel — are class-based. Smalltalk. Perhaps “pure or hybrid” is the wrong question. a hybrid language is perhaps better than a pure OOPL. page 43). these languages seem to be more often used for exploratory programming. procedural style. approximately. It is not clear whether. y = x means the same as x = y.4 Types versus Objects In most class-based PLs. The problem is complicated further by inheritance: if class C is a subclass of class P . Does it need to provide anything else? If so. but they cannot provide fully appropriate implementations.y). Compilers can provide default operations and “building bricks”. (This is analogous to the early days of “high level” programming when programmers would revert to assembler language to express constructs that were not easy or possible to express at a high level. The relation between classes and prototypes is somewhat analogous to the relation between procedural (FORTRAN and Algol) and functional (LISP) languages in the 60s. 6. This is the approach taken by CLU and some of the “functional” OOPLs. did not even claim to be object oriented. C++ is a hybrid PL in that it supports both procedural and object styles. 6. As OOPLs evolve. Although large systems have been built using prototype languages. 6. Classes provide a visible program structure that seems to be reassuring to designers of complex software systems. A hybrid language can be dangerous for inexperienced programmers because it encourages a hybrid programming style.

Conceptually.9. inheritance can be defined in a number of ways and has a number of different uses. We begin by writing a specification for a simple BankAccount class with an owner name. 6.7 Inheritance Inheritance is one of the key features of OOP and at the same time one of the most troublesome. The relation ArrayStack ∼ Stack is not the same as the relation ArrayStack ∼ Array. The multi-paradigm language LEDA (Budd 1995) demonstrates that it is possible to provide functional. but who is not familiar with the details of C++. as in Scheme (Section 4.2) or by classes. The purpose of this section is to show how C++ fails to fulfill its role as the preferred OOPL. There are other ways of obtaining this effect — notably. Whether or not we agree with Meyer. the class approach seems to be better in most practical situations. at the implementation level. the class ArrayStack could have an array object as an attribute — but Meyer maintains that inheritance is the most appropriate technique. however. and functions for making deposits and withdrawals: see Listing 33. There are many practical programming situations in which code in a parent class can be reused in a child class even when interface consistency is the primary purpose of inheritance. A closer examination. but is harder to accomplish. most (but not all!) OOPLs provide a single mechanism for inheritance and leave it to the programmer to decide how to use the mechanism. One of the clearest examples of the division is Meyer’s (1988) “marriage of convenience” in which the class ArrayStack inherits from Array and Stack.9. The first part of the following discussion is based on a paper by LaLonde and Pugh (1995) in which they discuss C++ from the point of view of a programmer familiar with the principles of OOP. . 6.10 A Critique of C++ C++ is both the most widely-used OOPL and the most heavily criticized. 6. In practice. The class ArrayStack behaves like a stack — it has functions such as push. and object programming in a single language. and so on — and is represented as an array. exemplified by a language such as Smalltalk. at least in Eiffel. Hence the question: should an OOPL distinguish between these relations or not? Based on the “marriage of convenience” the answer would appear to be “yes” — two distinct ends are achieved and the means by which they are achieved should be separated. This can be accomplished by high order functions and mutable state. Further criticism might appear to be unnecessary. Moreover. pop. it is clear that there two different relations are involved. Although each method has advantages for certain applications. It more tightly-integrated design would be preferable. shows that the answer is by no means so clear-cut. the varieties of inheritance all look much the same — a fact that discourages PL designers from providing multiple mechanisms.6 THE OBJECT ORIENTED PARADIGM 97 capabilities. However.6 Closures versus Classes The most powerful programming paradigms developed so far are those that do not separate code and data but instead provide “code and data” units of arbitrary size. logic. LEDA provides these in a rather ad hoc way. a balance.

void transfer(BankAccount *other. the destructor will be invoked.balance() account2→balance() A member function can access the private parts of another instance of the same class. an object can access only its own components. char *owner().6 THE OBJECT ORIENTED PARADIGM 98 Data members are prefixed with “p” to distinguish them from parameters with similar names. ∼BankAcocunt(). In both cases. We can create instances of BankAccount on either the stack or the heap: BankAccount account1("John"). // on the heap The first account. account1. At the end of the test. The code in Listing 37 — creating three special accounts — is rather specialized. as shown in Listing 34. This comes as a surprise to Smalltalk programmers because. once when the destructor for account1 is called. account1. in Smalltalk.pTrans is lost because it is over-written by account2. double amount). private: char *pOwner. }. The second account. Transaction *pTrans[MAXTRANS]. account2. void owner(char *newOwner). account2. .pTrans will be deleted twice. void credit(double amount). Functions that operate on this array have been omitted. double pBalance. will be deleted only when the program executes a delete operation. Although most introductions to C++ focus on stack objects. will be deleted automatically at the end of the scope containing the declaration. it is more interesting to consider heap objects: Listing 37. Listing 33: Declaring a Bank Account class BankAccount { public: BankAccount(char * newOwner = ""). This test introduces two bugs. void debit(double amount). The array account1. The notation for accessing functions depends on whether the variable refers to the object directly or is a pointer to the object. double balance(). Listing 36 shows a simple test of the class BankAccount. The instance variable pTrans is a pointer to an array containing the transactions that have taken place for this account.pTrans. // on the stack BankAccount *account2 = new BankAccount("Jane"). Listing 38 generalizes it so that we can enter the account owner’s names interactively. The next step is to define the member functions of the class BankAccount: see Listing 35. and again when the destructor for account2 is called.

i < 3.. for (int i = 0. delete accounts. // Account operations omitted. pBalance = 0. accounts[0] = new BankAccount("Anne"). account1 = acount2. accounts[2] = new BankAccount("Chris"). } Listing 37: Using the Heap BankAccount *accounts[] = new BankAccount * [3]. } char *BankAccount owner() { return pOwner. 99 .6 THE OBJECT ORIENTED PARADIGM Listing 34: Accessing Private Parts void BankAccount::transfer(BankAccount *other.. Listing 36: Testing the Bank Account void test { BankAccount account1("Anne"). } Listing 35: Member Functions for the Bank Account BankAccount::BankAccount(char *newOwner) { pOwner = newOwner.0. double amount) { pBalance += amount. i++) delete accounts[i]. } . // Delete everything. other→pBalance -= amount. // Delete the array of transactions } void BankAccount owner(char *newOwner) { pOwner = newOwner. BankAccount account2("Bill"). accounts[1] = new BankAccount("Bill").. } BankAccount::∼BankAccount() { // Delete individual transactions delete [] pTrans.

getline(buffer. delete buffer. but is poorly designed.getline(buffer. 100 The problem with this code is that all of the bank accounts have the same owner: that of the last name entered. for (int i = 0. i++) { cout << "Please enter your name: ". // Delete everything. as in Listing 39. delete accounts. i++) delete accounts[i]. cin. 100). // Delete everything.6 THE OBJECT ORIENTED PARADIGM Listing 38: Entering a Name BankAccount *accounts[] = new BankAccount * [3]. The rule we must follow is that data must be allocated with the constructor and deallocated with the destructor. char * buffer = new char[100]. Unfortunately. 100). char * buffer = new char[100]. } // Account operations omitted. for (int i = 0. We can fix this by allocating a new buffer for each account. not the client. cin. we can remove the destructor for owner in the code above: Listing 40. i++) { delete accounts[i]→owner(). i++) { cout << "Please enter your name: ". To correct this problem. i < 3. i < 3. Listing 39: Adding buffers for names // char * buffer = new char[100]. delete buffer. } // Account operations omitted. The destructor for BankAccount should delete the owner string. for (int i = 0. or controlled entirely outside . The account destructor fails because "Tina" was not allocated using new. i < 3. This version works. } delete accounts. (deleted) for (int i = 0. Suppose we use the loop to create two accounts and then create the third account with this statement: accounts[2] = new BankAccount("Tina"). Deleting the owner string should be the responsibility of the object. we must modify the constructor for BankAccount. This is because each account contains a pointer to the unique buffer. accounts[i] = new BankAccount(buffer). After making this change. i < 3. this change introduces another problem. delete accounts[i]. accounts[i] = new BankAccount(buffer).

6 THE OBJECT ORIENTED PARADIGM Listing 40: Correcting the Destructor BankAccount::∼BankAccount() { delete pOwner. provided that they are allocated on the heap. The underlying problem here is that references (&) work best with stack objects and pointers (*) work best with heap objects. although in fact an address has been passed. We could make things easier for the programmer who is using accounts by declaring an overloaded version of transfer() as shown below. We now have bank accounts that work quite well. which was not allocated by new.” rather than “→”. Once again. however. strcpy(pOwner. The next problem occurs when we try to use the overloaded function owner() to change the name of an account owner. // Delete the array of transactions } Listing 41: Correcting the Constructor BankAccount::BankAccount(char *newOwner) { char *pOwner = new char (strlen(newOwner) + 1). we must rewrite the member function owner(). We correct it by introducing the reference operator. } 101 the object. } . &. strcpy(pOwner. If we try to mix heap objects and stack objects. // Delete individual transactions delete [] pTrans. Note that. It is awkward to mix stack and heap allocation in C++ because two Listing 42: Correcting the Owner Function void BankAccount owner(char *newOwner) { delete pOwner. This is because other behaves like an object. account1→transfer(&account3. 10. Listing 41 shows the new constructor. To correct this problem. we again encounter difficulties: Listing 43. it will attempt to delete "Fred". we must use the operator “. When the destructor is invoked for accounts[2]. we have introduced two bugs: The original owner of accounts[2] will be lost. accounts[2]→owner("Fred").0). pOwner = new char(strlen(newOwner) + 1). newOwner). newOwner). in the version of Listing 44. as shown in Listing 42. The error occurs because transfer expects the address of an account rather than an actual account.

6 THE OBJECT ORIENTED PARADIGM Listing 43: Stack and Heap Allocation BankAccount *account1 = new BankAccount("Anne"). will create a heap object corresponding to A + B but provides no opportunity to delete this object. Suppose that we want to provide the standard algebraic operations (+. there are problems.pBalance -= amount. If the objects are small — complex numbers. for example — it is reasonable to store the objects on the stack and to use call by value. myAccount = test(account2). double amount) { pBalance += amount. The destructors will fail because the default copy constructor makes a shallow copy — it does not copy the pTrans fields of the accounts.) Since a matrix occupies 4×4×8 = 128 bytes. Listing 45: Copying BankAccount test(BankAccount account) { return account. C++ is well-suited to this approach: storage management is automatic and the only overhead is the copying of values during function call and return. . Even a simple function call such as project(A + B). But if we write functions that work with pointers to matrices. for example. To implement the code shown in Listing 45. } BankAccount myAccount. these two copies must be deleted. however. to make another copy of account2 that will become myAccount. If the objects are large. C++ uses the (default) copy constructor for BankAccount twice: first to make a copy of account2 to pass to test() and. 10. −. If stack objects are used without references. 10. ∗) for 4 × 4 matrices. (This would be useful for graphics programming. // Error! Listing 44: Revising the Transfer Function void BankAccount::transfer(BankAccount &other. LaLonde and Pugh continue with a discussion of infix operators for user-defined classes. copying a matrix is relatively expensive and the stack approach will be inefficient. It is recommended practice in C++ to write these functions for any class that contains pointers. The function pDelete has the same effect as the destructor. // OK account1→transfer(account3. the compiler is forced to make copies.0). as in Listing 46. } 102 sets of functions are needed. The solution is to write our own copy constructor and assignment overload for class BankAccount. second. BankAccount account3("Chris"). it must make copies of all of the components of a back account object and destroy all components of the old object. it is difficult to manage allocation and deallocation. BankAccount *account2 = new BankAccount("Bill").0). other. At the end of the scope. The function pDuplicate() makes a deep copy of an account the function. account1→transfer(account2.

6 THE OBJECT ORIENTED PARADIGM Listing 46: Copy Constructors BankAccount::BankAccount(BankAccount &account) { pDuplicate(account). without losing objects. Since it returns a stream reference (usually its left operand). pDuplicate(rhs). } BankAccount & BankAccount operator = (BankAccount &rhs) { if (this == &rhs) return this. Smalltalk. Another basis for criticizing C++ is that is has become. as shown in Listing 47. It is possible to implement the extended assignment operators. A sign of good language design is that the features of the language are used most of the time to solve the problems that they were intended to solve. Sather.) The same problem arises with expressions such as A = B + C * D. Nevertheless. It is significant that most language designers have chosen the heap model for OOP (Simula. Blue.’ << endl. } Listing 47: Using extended assignment operators A = C. such as when efficiency is important. it can be used as an infix operator in compound expressions such as the one shown. A typical output statement looks something like this: cout << "The answer is " << answer << ’. A sign of a hacker’s language is that language “gurus” introduce “hacks” that solve problems by using features in bizarre ways for which they were not originally intended. at first glance. return *this. perhaps not intentionally. These observations by LaLonde and Pugh seem rather hostile towards C++. because it cannot tell whether its argument is a temporary. The notation seems clear and concise and. A *= D. Java. the discussion makes an important point: the “stack model” (with value semantics) is inferior to the “heap model” (with reference semantics) for many OOP applications. appears to be an elegant way of introducing output and input (with the operator >>) capabilities. 103 (The function cannot delete it. The comparison is one-sided because they do not discuss factors that might make C++ a better choice than Smalltalk in some circumstances. pDelete(). such as +=. A += B. They are making a comparison between C++ and Smalltalk based only on ease of programming. amongst others) and that C++ is in the minority on this issue. Input and output in C++ is based on “streams”. The operator << requires a stream reference as its left operand and an object as its right operand. Dee. a hacker’s language. But this approach reduces us to a kind of assembly language style of programming. Eiffel. CLU. .

public: ManipulatorForInts (int (ios::*)(int). First. as in Listing 48. the actual definitions are not as shown . (<< is also overloaded with respect to its left operand. to format a table of heterogeneous data. return os. we could change the example above to: cout << "The answer is " << setprecision(8) << answer << ’. for example.’ << endl. } 104 Obviously. the function that we need can be defined: ManipulatorForInts setprecision (int n) { return ManipulatorForInts (& ios::precision. The first problem with this is that it is verbose. const ManipulatorForInts & m) { (os. There must be a separate function for each class for which instances can be used as the right operand.6). The first problem with streams is that.) The second. let alone design similar functions. .*memberfunc)(m. It is possible to write code that uses dynamic binding but only by providing auxiliary functions.) In the example above.. It is a global function with signature ios & endl (ios & s) A programmer who needs additional “manipulators” of this kind has no alternative but to add them to the global name space. and more serious problem. a double. they do not work well with inheritance. n). private: int (ios::*memberfunc)(int). Overloading provides static polymorphism which is incompatible with dynamic polymorphism (recall Section 9. For example. } It is apparent that a programmer must be quite proficient in C++ to understand these definitions. and the inserter function is defined: ostream & operator << (ostream & os. (The verbosity is not really evident in such a small example. The final operand. writes end-of-line to the stream and flushes it.. but it is very obvious when streams are used. int).. because there are different kinds of output streams. endl. Things get more complicated when formatting is added to streams. a class for manipulators with integer arguments is declared. the first three right operands are a string. The way this is done is (roughly) as follows (Teale 1993. } Second.value). To make matters worse. int value. and a character.6 THE OBJECT ORIENTED PARADIGM Listing 48: A Manipulator Class class ManipulatorForInts { friend ostream & operator << (ostream & const ManipulatorForInts &). although they are superficially object oriented (a stream is an object). page 171). is the implementation of setprecision. << is heavily overloaded.

Class-based OOPLs can be designed so as to support information hiding (encapsulation) — the separation of interface and implementation. For example. Programs can often be extended by adding new code rather than altering existing code. The mechanism for incremental modification without altering existing code is inheritance. Sakkinen 1992). there may be conflicts between inheritance and encapsulation. However. all objects can respond to the message “print yourself”. if (!infile) .11 Evaluation of OOP OOP provides a way of building programs by incremental modification. feature of OOP is that the call graph is created and modified during execution. that we can negate. Section 6. 6. in which the call graph can be statically inferred from the program text. but few realize that frequently-used and apparently harmless constructs are implemented by subtle and devious tricks. there are several different ways in which inheritance can be defined and used. if (infile) . . but are implemented using templates. There are many other tricks in the implementation of streams. In fact. with the value NULL indicating that the file is not open. Most experienced programmers can get used to the well-known idiosyncrasies of C++. . obtaining !infile. This is in contrast to procedural languages. Programs that look simple may actually conceal subtle traps. because that would require class declarations for many different types. The same identifier can be used in various contexts to name related functions. However.9 contains a more detailed review of issues in OOP. . . but rarely noted. we can use file handles as if they had boolean values.dat").6 THE OBJECT ORIENTED PARADIGM Listing 49: File handles as Booleans ifstream infile ("stats. It seems that there is a boolean value. Exercise 31. For example. infile can be used as a conditional because there is a function that converts infile to (void *). and some of them conflict. Can you account for the slow acceptance of OOPLs by professional programmers? 2 . and !infile can be used as a conditional because the operator ! is overloaded for class ifstream to return an int which is 0 if the file is open. . These are symptoms of an even deeper malaise in C++ than the problems pointed out by authors of otherwise excellent critiques such as (Joyner 1992. as in Listing 49. infile. . Simula is 33 years old. 105 here. An interesting.

2. bob). ann). 1 | 6.2) of a logic PL is some form of mathematical logic. parent(tom. Bob. is to produce solutions one at a time. Note that.7 Backtracking Languages The PLs that we discuss in this section differ in significant ways. 24 } We could define functions that return sets in a FPL or we could design a language that works directly with relations. In mathematics.to indicate that it is ready to accept a query. If the query contains a logical variable — an identifier that starts with an upper case letter — Prolog attempts to find a value of the variable that makes the query true. 12. The approach that has been most successful. The program runs until it finds the first solution. Any integer greater than 1 has at least two divisors (itself and 1) and possible others. bob) as “Pam is the parent of Bob”. The first logic PL was Prolog. bob). For example. introduced by Colmerauer (1973) and Kowalski (1974). 7. 3 | 6. 4. parent(pat. 8. 3. since constants in Prolog start with a lower case letter. We can express facts in Prolog as predicates with constant arguments. reports that. This Bob’s parents are Pam and Tom. Prolog responds to simple queries about facts that it has been told with “yes” and “no”.) Practical logic PLs are based on a restricted form of predicate calculus that has a decision procedure. liz). it replies with the binding. 7 } divisors(24) = { 1. We read parent(pam. it prompts with ?. 2. 3. The functional analog of the divides relation is the set-returning function divisors: divisors(6) = { 1. we cannot write Pam. Predicate calculus is too strong because it is undecidable. Consider the relation “divides”. and to find as many answers as required. parent(bob. denoted by | in number theory. but they have an important feature in common: they are designed to solve problems with multiple answers. and 6 | 6. 106 . 6. a many-valued expression can be represented as a relation or as a function returning a set. 6 } divisors(7) = { 1. The discussion in this section is based on (Bratko 1990). (This means that there is no effective procedure that determines whether a given predicate is true. parent(bob. jim). reports that solution or uses it in some way.1 Prolog The computational model (see Section 10. etc. 2 | 6. Listing 50: A Family parent(pam. pat). then backtracks to find another solution. All Prolog textbooks start with facts about someone’s family: see Listing 50. parent(tom. as in Listing 51. If it succeeds. Prolog is interactive. and so on until all the choices have been explored or the user has had enough. Propositional calculus is too weak because it does not have variables. however. similarly for the other facts.

parent(tom. Y) must be built into Prolog somehow.sibling(pat. (The exact mechanism depends on the implementation. C).7 BACKTRACKING LANGUAGES Listing 51: Prolog queries ?. X). Y) is satisfied by the bindings P = bob. parent(P. P = pat G = bob The comma in this query corresponds to conjunction (“and”): the responses to the query are bindings that make all of the terms valid simultaneously.parent(bob. In logic. Testing this rule reveals a problem. because we do not consider Pat to be a sibling of herself. X).parent(P. For example. parent(G. Siblings are people who have the same parents. jim). Disjunction (“or”) is provided by multiple facts or rules: in the example above. parent(P. a person A is an ancestor of a person X if either A is a parent of X or A is an ancestor of a parent P of X: ancestor(A. P).parent(A. liz).parent(P. It is clear that different(X. The rule for “grandparent” is “G is a grandparent of C if G is a parent of P and P is a parent of C”. and Y = pat. different(X. Bob is the parent of C is true if C is Ann or C is Pat. no 107 ?.parent(tom. We could write this rule for siblings: sibling(X. Some implementations require the user to ask for more results. C = ann C = pat Consider the problem of finding Jim’s grandparents. X) :. Y). We can expression the “grandparent” relation by writing a rule and adding it to the collection of facts (which becomes a collection of facts and rules). this becomes: ?. X). parent(P. X).parent(P. jim). In Prolog: grandparent(G. Here we assume that the implementation returns all values without further prompting. X = ann X = pat We may not expect the second response. C). parent(P. yes ?. X). . we can express “G is a grandparent of Jim” in the following form: “there is a person P such that P is a parent of Jim and there is a person G such that G is a parent of P”. X = pat. X = bob If there is more than one binding that satisfies the query. because it would not be feasible to list every true instantiation of it as a fact.parent(pam. Y) :. Prolog will find them all. Y). Y) :.) ?. we would have to require that X and Y have different values: sibling(X.parent(G. In Prolog. Y). P). C) :. ?. To correct this rule. Prolog allows rules to be recursive: a predicate name that appears on the left side of a rule may also appear on the right side. But Prolog simply notes that the predicate parent(P. X).

7 BACKTRACKING LANGUAGES

108

parent(pam, bob) ancestor(pam, bob) ancestor(pam, pat) ancestor(pam, jim) Figure 15: Proof of ancestor(pam, jim) ancestor(A, X) :- ancestor(A, P), parent(P, X). We confirm the definition by testing: ?- ancestor(pam, jim). yes We can read a Prolog program in two ways. Read declaratively , ancestor(pam, jim) is a statement and Prolog’s task is to prove that it is true. The proof is shown in Figure 15, which uses the convention that P means “from P infer Q”. Read procedurally , ancestor(pam, jim) defines Q a sequence of fact-searches and rule-applications that establish (or fail to establish) the goal. The procedural steps for this example are as follows. 1. Using the first rule for ancestor, look in the data base for the fact parent(pam, jim). 2. Since this search fails, try the second rule with A = pam and X = jim. This gives two new subgoals: ancestor(pam, P) and parent(P, jim). 3. The second of these subgoals is satisfied by P = pat from the database. The goal is now ancestor(pam, pat). 4. The goal expands to the subgoals ancestor(pam, P) and parent(P, pat). The second of these subgoals is satisfied by P = bob from the database. The goal is now ancestor(pam, bob). 5. The goal ancestor(pam, bob) is satisfied by applying the first rule with the known fact parent(pam, bob). In the declarative reading, Prolog finds multiple results simply because they are there. In the procedure reading, multiple results are obtained by backtracking . Every point in the program at which there is more than one choice of a variable binding is called a choice point. The choice points define a tree: the root of the tree is the starting point of the program (the main goal) and each leaf of the tree represents either success (the goal is satisfied with the chosen bindings) or failure (the goal cannot be satisfied with these bindings). Prolog performs a depth-first search of this tree. Sometimes, we can tell from the program that backtracking will be useless. Declaratively, the predicate minimum(X,Y,Min) is satisfied if Min is the minimum of X and Y. We define rules for the cases X ≤ Y and X > Y . minimum(X, Y, X) :- X ≤ Y. minimum(X, Y, Y) :- X > Y. It is easy to see that, if one of these rules succeeds, we do not want to backtrack to the other, because it is certain to fail. We can tell Prolog not to backtrack by writing “!”, pronounced “cut”. parent(bob, pat) parent(pat, jim)

7 BACKTRACKING LANGUAGES Program p :- a,b. p :- c. p :- a,!,b. p:- c. p :- c. p :- a,!,b. Meaning p ⇐⇒ (a ∧ b) ∨ c p ⇐⇒ (a ∧ b) ∨ (∼a ∧ c) p ⇐⇒ c ∨ (a ∧ b)

109

Figure 16: Effect of cuts minimum(X, Y, X) :- X ≤ Y, !. minimum(X, Y, Y) :- X > Y, !. The introduction of the cut operation has an interesting consequence. If a Prolog program has no cuts, we can freely change the order of goals and clauses. These changes may affect the efficiency of the program, but they do not affect its declarative meaning. If the program has cuts, this is no longer true. Figure 16 shows an example of the use of cut and its effect on the meaning of a program. Recall the predicate different introduced above to distinguish siblings, and consider the following declarations and queries. different(X, X) :- !, fail. different(X, Y). ?- different(Tom, Tom). no ?- different(Ann, Tom). yes The query different(Tom, Tom) matches different(X, X) with the binding X = Tom and succeeds. The Prolog constant fail then reports failure. Since the cut has been passed, no backtracking occurs, and the query fails. The query different(Ann, Tom) does not match different(X, X) but it does match the second clause, different(X, Y) and therefore succeeds. This idea can be used to define negation in Prolog: not (P) :- P, !, fail. true. However, it is important to understand what negation in Prolog means. Consider Listing 52, assuming that the entire database is included. The program has stated correctly that crows fly and penguins do not fly. But it has made this inference only because the database has no information about penguins. The same program can produce an unlimited number of (apparently) correct responses, and also some incorrect responses. We account for this by saying that Prolog assumes a closed world . Prolog can make inferences from the “closed world” consisting of the facts and rules in its database. Now consider negation. In Prolog, proving not(P) means “P cannot be proved” which means “P cannot be inferred from the facts and rules in the database”. Consequently, we should be sceptical about positive results obtained from Prolog programs that use not. The response ?- not human(peter). yes means no more than “the fact that Peter is human cannot be inferred from the database”. The Prolog system proceeds top-down. It attempts to match the entire goal to one of the rules and, having found a match, applies the same idea to the parts of the matched formulas. The “matching” process is called unification.

7 BACKTRACKING LANGUAGES Listing 52: Who can fly? flies(albatross). flies(blackbird); flies(crow). ?- flies(crow). yes ?- flies(penguin). no ?- flies(elephant). no ?- flies(brick). no ?- flies(goose). no

110

Suppose that we have two terms A and B built up from constants and variables. A unifier of A and B is a substitution that makes the terms identical. A substitution is a binding of variable names to values. In the following example we follow the Prolog convention, using upper case letters X, Y, . . .to denote variables and lower case letters a, b, . . .to denote constants. Let A and B be the terms A ≡ left(tree(X, tree(a, Y )) B ≡ left(tree(b, Z)) and let σ = (X = b, Z = tree(a, Y )) be a substitution. Then σA = σB = left(tree(b, tree(a, Y )) and σ is a unifier of A and B. Note that the unifier is not necessarily unique. However, we can define a most general unifier which is unique. The unification algorithm contains a test called the occurs check that is used to reject a substitution of the form X = f (X) in which a variable is bound to a formula that contains the variable. If we changed A and B above to A ≡ left(tree(X, tree(a, Y )) B ≡ left(tree(b, Y ) they would not unify. You can find out whether a particular version of Prolog implements the occurs check by trying the following query: ?- X = f(X). If the reply is “no”, the occurs check is implemented. If, as is more likely, the occurs check is not performed, the response will be X = f(f(f(f( . . . . Prolog syntax corresponds to a restricted form of first-order predicate calculus called clausal form logic. It is not practical to use the full predicate calculus as a basis for a PL because it is undecidable. Clausal form logic is semi-decidable: there is an algorithm that will find a proof

most Prolog systems are technically unsound. If a formula is false. which fully respect the logical model. the order is significant. ?. Consequently. Listing 53 shows the results of two queries about restaurants. In the context of the rules wife(jane). as in wife(jane). the algorithm may fail after a finite time or may loop forever.female(X). “Pure” Prolog programs.reasonable(X). no 111 for any formula that is true in the logic. since Prolog works through the rules in the order in which they are written. female(X) :.good(X).female(X). X = joes diner. the occurs check is expensive to implement and most Prolog systems omit it. the same query leads to endless applications of the second rule. Unfortunately. In logic. good(X). good(joes diner). jane But if we change the order of the rules.female(Y). can be written but are often inefficient.wife(X). Explain why commutativity fails in this program. For example. female(X) :. we obtain a valid answer to a query: ?. 2 Exercise 32. restricted to Definite clauses”. using a Linear strategy. it follows that disjunction in Prolog does not commute. was introduced by Robinson (1965). reasonable(Restaurant) : not expensive(Restaurant). . called SLD resolution. SLD stands for “S electing a literal.7 BACKTRACKING LANGUAGES Listing 53: Restaurants good(les halles). female(X) :. A Prolog program apparently has a straightforward interpretation as a statement in logic. ?. The proof of validity of SLD resolution assumes that unification is implemented with the occurs check.wife(X). The proof technique. female(X) :. although problems are rare in practice. Example 17: Failure of commutativity. reasonable(X). 2 The important features of Prolog include: Prolog is based on a mathematical model (clausal form logic with SLD resolution). conjunction is commutative (P ∧ Q ≡ Q ∧ P ). but the interpretation is slightly misleading. expensive(les halles). Since the logical interpretation of a rule sequence is disjunction. In order to write efficient programs. a Prolog programmer must be able to read programs both declaratively (to check the logic) and procedurally (to ensure efficiency).

If its value is TRUE. and EXIT statements. If the array a is ordered. . variant records. The FOR statement finds the first position at which the arrays differ. Each sequence is tried in turn in this way. with the program in the same state as it was when the EITHER statement was entered. the computation succeeds and the program continues as usual. the CASE. and i has the smallest value for which the condition is FALSE. sets. the sequence succeeds iff a[i] < b[i]. If the computation fails. 7. The PL Alma-0 takes as its starting point a simple imperative language. . pointer types. This section is based on Apt et al. Boolean expressions may be used as statements. the following statement succeeds. modules). it is considered equivalent to a boolean expression with value TRUE. Prolog has garbage collection. LOOP. FOR i := 1 to n . A statement sequence may be used as a boolean expression. WITH. If the computation of a sequence of statements succeeds. the loop exits. open array parameters. If there is a difference at position i. If the value of the expression is FALSE.2 Alma-0 One of the themes of this course is the investigation of ways in which different paradigms of programming can be combined. An EITHER statement has the form EITHER <seq> ORELSE <seq> . Some features of Modula–2 are omitted (the CARDINAL type. control returns to the second sequence in the EITHER statement. this does not mean that the program stops: instead. However. The expression is evaluated. Example 19: Ordering 2. it is considered equivalent to a boolean expression with value FALSE.7 BACKTRACKING LANGUAGES 112 Prolog implementations introduce various optimizations (the cut. The following sequence decides whether a precedes b in the lexicographic ordering. the entire sequence fails. BES. If it succeeds. the computation fails. it fails. Modula–2 (Wirth 1982). a[i] < b[i] 2 EITHER–ORELSE. nested procedures. Otherwise. described below.ORELSE <seq> END Computation of an EITHER statement starts by evaluating the first sequence. The result is a simple but highly expressive language. The construct EITHER-ORELSE is added to the language to permit backtracking. (1998). NOT FOR i :=1 TO n DO a[i] = b[i] END. Example 18: Ordering 1. They are replaced by nine new features.1 DO a[i] <= a[i+1] END 2 SBE. and adds a number of features derived from logic programming to it. until one succeeds or the last one fails. . computation continues at the statement following END. control returns to the nearest preceding choice point and resumes with a new choice. If they are equal. If this computation subsequently fails. omitting the occurs check) that improve the performance of programs but compromise the mathematical model. . procedure types.

The sequence shown in Listing 55 sets i to the first occurrence of p in s.n-1]. The final sequence is performed and succeeds. x := x + b. With the COMMIT. The test x > 0 succeeds because x = a > 0.m DO FOR j := 0 to m .. The sequence of events is as follows: The assignment y := x assigns a to y.) The statement COMMIT <seq> END evaluates the statement sequence <seq>. The test y < 0 fails because y = a > 0. leaving x = a + b and y = −a. Assume that the sequence shown in Listing 54 is entered with x = a > 0. This sequence succeeds and transfers control to x := x + b. The COMMIT statement provides a way of restricting backtracking. Suppose that we have two arrays of characters: a pattern p[0.1 DO s[i+j] = p[j] END END 113 Example 20: Choice points. y := -x END. (It is similar to the cut in Prolog. Listing 56 shows another way of checking lexicographic ordering.7 BACKTRACKING LANGUAGES Listing 54: Creating choice points EITHER y := x ORELSE x > 0. The statement SOME i := 1 TO n DO <seq> END is equivalent to EITHER i := 1. 2 COMMIT. if seq succeeds (possibly by backtracking). or fails if there is no match.. this sequence yields that value of a[i] < b[i] for the smallest i . Example 22: Ordering 3. The assignment x := x + b assigns a + b to x. all of the choice points created during its evaluation are removed. 2 SOME.m-1] and a string s[0. and y receives the value −a. The SOME statement permits the construction of statements that work like EITHER statements but have a variable number of ORELSE clauses. The original value of x is restored. <seq> ORELSE SOME i := 2 TO n DO <seq> END Example 21: Pattern matching. y < 0 Listing 55: Pattern matching SOME i := 0 TO n .

The following sequence finds the number of times that p occurs in s. Suppose we enclose the pattern matching code above in a function Match. Variables in Alma-0 start life as uninitialized values. 8. 7). . 4. three 9’s arranged in such a way that for i = 1. 2 FORALL. count := 0. . FORALL k := Match(p.. an error is reported. 1. the program shown in Listing 57 is a solution for both of these problems! If the given sequence is initialized. 2 .7 BACKTRACKING LANGUAGES Listing 56: Checking the order COMMIT SOME i := 1 TO n DO a[i] <> b[i] END END. The statement FORALL <seq1> DO <seq2> END evaluates <seq1> and then <seq2>. the program assigns a remarkable sequence to it. The FORALL statement provides for the exploration of all choice points generated by a statement sequence. 9. Write a program that tests whether a given array is a remarkable sequence. 7. Whenever <seq2> succeeds. if one side is an uninitialized variable and the other is an expression with a value. 6. 9. 5. it ceases to be uninitialized and becomes initialized . 6. 2. three 2’s. this value is assigned to the variable. control backtracks through these. 5. Example 24: Remarkable sequences. . Example 23: Matching. but they do not have values. 9 there are exactly i numbers between successive occurrences of i. Expressions containing uninitialized values are allowed. 2. 2 EQ. 2. if it is not initialized. Here is a remarkable sequence: (1. 4. Due to the properties of “=”. s) DO count := count + 1 END. it succeeds if a[i] < b[i] is TRUE for some value of i such that a[i] <> b[i]. Problem 1. 5. the program tests whether it is remarkable. 9. 6. . . there is an implicit COMMIT around <seq2>). 8. the comparison s = t is evaluated as follows: if both s and t are initialized. 2. its choice points are removed (that is. Problem 2. A sequence with 27 integer elements is remarkable if it contains three 1’s. 7. 3. 3. 8. if both sides are uninitialized variables. Find a remarkable sequence. For example. 3. 1.. . a[i] < b[i] 114 such that a[i] <> b[i]. even if the original sequence succeeded. their values are compared as usual. . 4. If <seq1> generates choice points. Without the COMMIT. After a value has been assigned to a variable.

a[j+2*i+1] = i END END END. The call P(E).. assigns each component of a to x by backtracking. depending on the form of the actual parameter. where x is an integer variable. Listing 58: Compiling a MIX parameter VAR v: integer. .END. Listing 59: Linear search Type Vector = ARRAY [1. P(v) END 115 MIX. VAR i: INTEGER. . Alma-0 adds a third parameter passing mechanism: call by mixed form.27] OF INTEGER.7 BACKTRACKING LANGUAGES Listing 57: Remarkable sequences TYPE Sequence = ARRAY [1. PROCEDURE Find (MIX e: INTEGER. Suppose that procedure P has been declared as PROCEDURE P (MIX n: INTEGER). BEGIN v := E. a[j+i+1] = i. Example 25: Linear Search. j: INTEGER.a) tests if 7 occurs in a. is compiled in the usual way. interpreting MIX as VAR. in which E is an expression. in which v is a variable. The parameter is passed either by value or by reference.. The call Find(x. . a: Vector). . BEGIN SOME i := 1 TO N DO e = a[i] END END. Modula–2 provides call by value (the default) and call by reference (indicated by writing VAR before the formal parameter. is compiled as if the programmer had written the code shown in Listing 58. The function Find in Listing 59 uses a mixed form parameter. BEGIN . We can think of the implementation of mixed form in the following way.N] of INTEGER. VAR i. PROCEDURE Remarkable (VAR a: Sequence). a). The call P(v). BEGIN For i := 1 TO 9 DO SOME j := 1 TO 25 . The call Find(7.2 * i DO a[j] = i.

both to the existing features of Modula2 and to each other. It also has a procedural semantics that describes how the constructs can be efficiently executed. KNOWN(z) THEN x = z . new features that provide expressive power when used together. a). and Schonberg 1986). Alma-0 demonstrates that. 2 The language Alma-0 has a declarative semantics in which the constructs described above correspond to logical formulas. 116 The sequence Find(x. b) tests whether the arrays a and b have an element in common. and 2LP (McAloon and Tretkoff 1995). Exercise 33.y. Find(x. x receives the value of the common element.5) succeeds and Plus(5.3 Other Backtracking Languages Prolog and Alma-0 are by no means the only PLs that incorporate backtracking. The interesting feature of Alma-0 is that it demonstrates that it is possible to start with a simple.12) assigns 7 to n. a) DO Find(x.y ELSIF KNOWN(x). KNOWN(y) THEN z = x + y ELSIF KNOWN(y). The call Plus(x. The statement FORALL Find(x.7 BACKTRACKING LANGUAGES Listing 60: Generalized Addition PROCEDURE Plus (MIX x. b) END tests whether all elements of a are also elements of b.z) will succeed if x + y = z provided that at least two of the actual parameters have values. y. KNOWN(z) THEN y = z . if appropriate care is taken. Listing 60 shows the declaration of a function Plus. if so. a). The new features of Alma-0 are orthogonal. b) DO WRITELN(x) END prints all of the elements that a and b have in common. remove features that are not required for the current purpose. Others include Icon (Griswold and Griswold 1983). z: INTEGER). SETL (Schwartz. Dubinsky. The statement FORALL Find(x. Find(x. and add a small number of simple. How does this use of orthogonality compare with the orthogonal design of Algol 68? 2 7. Dewar. Example 26: Generalized Addition. well-defined language (Modula-2).3. BEGIN IF KNOWN(x). For example. Plus(2.n.x END END. 2 KNOWN. it is possible to design PLs that smoothly combine paradigms without losing the advantages of either paradigm. using backtracking. . The test KNOWN(x) succeeds iff the variable x is initialized.

7 BACKTRACKING LANGUAGES 117 Constraint PLs are attracting increasing attention. Prolog is strong in logical deduction but weak in numerical calculation. . expressed as equalities and inequalities. A program is a set of constraints. that must be satisfied. constraint PLs attempt to correct this imbalance.

a grammar written in Backus-Naur Form (BNF) for Algol 60. 8. In Algol 60.2 Modules By the early 70s. The inner block declares another integer k which is set to 4 and then disappears without being used at the end of the block. It was not possible for one declaration to override another declaration of the same name. Furthermore. people realized that the “monolithic” structure of Algol-like programs was too restricted for large and complex programs. which is first set to 6 and then incremented. syntactically at least. It is probably significant that Algol 60 was also the first language to be defined. k := 6. however.1 for further discussion of nesting. Listing 61: Nested Declarations begin integer k. 118 .8 Structure The earliest programs had little structure. The outer block in Listing 61 declares an integer k. It is straightforward to define recursive structures using context-free grammars — specifically. FORTRAN subroutines have local variables) but scopes were continuous (no “holes”) and not nested. The new model that was introduced by languages such as Mesa and Modula divided a large program into modules. k := 1. The basis of Algol 60 syntax can be written in extended BNF as follows: PROG → BLOCK BLOCK → "begin" { DECL } { STMT } "end" STMT → BLOCK | · · · DECL → HEADER BLOCK | · · · With blocks come nested scopes. an inner declaration overrides an outer declaration. end.8. begin integer k. k := k + 1. Earlier languages had local scope (for example. 8. with a contextfree grammar. the “separate compilation” model of FORTRAN — and later C — was seen to be unsafe. end.1 Block Structure Algol 60 was the first major PL to provide recursive. See Section 9. Although a FORTRAN program has a structure — it consists of a main program and subroutines — the main program and the subroutines themselves are simply lists of statements. hierarchical structure.

. pop.. . ... STACK myStack. and functions (called collectively features). initialize. . In typical modular languages. . Example 27: Stack Module. import Stack.. as if the module was a record. push(67). functions. One module of a program might declare a stack module: module Stack { export initialize. . types. and it is shared by all parts of the program that import it.. . push. initialize(). Example 28: Stack Template. A module may import some or all of its features. we can use the type STACK to create stacks. variables. STACK yourStack. module Stack { export type STACK.. push(myStack.. . STACK myStack.... The last few lines of the example above would be written as: import Stack. . 2 This form of module introduces a single instance: there is only one stack. Some modular languages provide an alternative syntax in which the name of the instance appears before the function name.. initialize(myStack).. In contrast.. pop. The main difference between the stack module and the stack template is that the template exports a type. . . . a stack template introduces a type from which we can create as many instances as we need. } Elsewhere in the program. 2 The fact that we can declare many stacks is somewhat offset by the increased complexity of the notation: the name of the stack instance must appear in every function call.. } In other parts of the program. push..8 STRUCTURE 119 A module is a meaningful unit (rather than a collection of possibly unrelated declarations) that may introduce constants. and variables. modules can export types. 67). . . we could use the stack module: import Stack.

A large program will be implemented as a set of modules.8 STRUCTURE myStack. This presupposes that all interfaces are completely and precisely defined. Parnas introduced these ideas as the “principle of information hiding”. . It is desirable that a programmer should be able to change the implementation of a module without affecting (or even informing) other programmers. His thesis is that the programmer is most effective if shielded from. I am now convinced that information hiding. The specification answer the question “What does this module do?” and the implementation answers the question “How does this module work?” Most people today accept Parnas’s principle. users do not have such access.push(67). myStack.initialize(). even if they do not apply it well. Parnas was right. . and I was wrong. today often embodied in object-oriented programming. it is more usual to refer to the goals as encapsulation. 8. will be the owners of some modules and users of other modules. Today. While that is definitely sound design. or programming team. perhaps with occasional backtracking to correct an error. This implies that a module should have an interface that is relatively stable and an implementation that changes as often as necessary. Whereas Wirth assumes that a program can be constructed by filling in details. the details of construction of system parts other than his own. [Parnas] has proposed a still more radical solution. rather than exposed to. only shortly after Wirth’s (1971) paper on program development by stepwise refinement.. (Brooks 1975. Only the owner of a module has access to its implementation. A good information system both exposes interface errors and stimulates their correction. page 272) Parnas’s paper was published in 1972. Each programmer.. (Brooks 1995.1 Encapsulation 120 Parnas (1972) wrote a paper that made a number of important points. This is a more “modern” view and certainly a view that reflects the actual construction of large and complex software systems. users require a specification written in an appropriate language. Parnas sees programming as an iterative process in which revisions and alterations are part of the normal course of events. he had changed his mind..2. dependence upon its perfect accomplishment is a recipe for disaster. Twenty years later. is the only was of raising the level of software design. It is therefore important to appreciate how radical it was at the time of its introduction. In order to use the module correctly without access to its implementation. page 78) Brooks wrote this in the first edition of The Mythical Man Month in 1975. Yet Parnas’s approach is significantly different from Wirth’s.

Precise reasoning (for example. Both C and Pascal have goto statements but they are rarely used in well-written programs. although his paper is well-balanced overall. the role of goto was down-played.1 Loop Structures. Maintenance errors are less common. All programming needs can be met by three structures satisfying the above property: the sequence. goto-ridden programs but to write new programs using simple control structures only. The use of control structures rather than goto statements has several advantages. and one flow out of. The basic ideas that Dijkstra proposed were as follows. Dijkstra’s arguments had the desired effect. the use of formal semantics) is simplified when goto statements are absent.8 STRUCTURE 121 8. C.3. the loop with preceding test (while-do). The first contribution of PLs was to provide names for variables and infix operators for building expressions. and statements. However. although some uses of continue look rather like jumps. call/return. conditional. and return. Ada has a goto statement because it was part of the US DoD requirements but. has even less need for goto than Pascal. There should be one flow into. again. and loop structures only. and the conditional (if-then-else). Algol 60 also provided goto statements and most programmers working in the sixties assumed that goto was essential. This paper establishes that any program written with goto statements can be rewritten as an equivalent program that uses sequence. continue. Dijkstra’s main point was not to transform old. Dijkstra’s proposals had previously been justified in a formal sense by Bohm and Jacopini (1966). The most significant contribution of Algol 60 was the block : a program unit that could contain data. transfer (goto).3 Control Structures Early PLs. The storm of controversy raised by Dijkstra’s (1968) letter condemning the goto statement indicates the extent to which programmers were wedded to the goto statement. Java does not have a goto statement. each control structure. The basic loop has a single test that precedes the body: . depended on three control structures: sequence. It is easier to read and maintain programs. Algol 60 also contributed the familiar if-then-else statement and a rather complex loop statement. Certain optimizations become feasible because the compiler can obtain more precise information in the absence of goto statements. with break. These structures are essentially those provided by assembly language. it is rarely needed. Despite the controversy. Although the result is interesting. In languages designed after 1968. The control structure that has elicited the most discussion is the loop. it may be necessary to introduce Boolean variables. such as FORTRAN. C++ inherited goto statements from C. Even Knuth (1974) entered the fray with arguments supporting the goto statement in certain circumstances. The “modern” control structures first appeared in Algol 60. 8. functions.

read(n).3. Listing 62 shows the code in Pascal. for example. while (true) { int n. but this is often considered undesirable. because there are situations in which it is useful to exit from a loop from within the body. recognizing this problem. provide only one form of loop statement consisting of an unconditional repeat and a conditional exit that can be used anywhere in the body of the loop. however. end Listing 63: Reading in C int sum = 0. a description of actions which are to be performed. the problem of reading and adding numbers.8 STRUCTURE Listing 62: Reading in Pascal var n. terminating when a negative number is read without including the negative number in the sum. we can use break to avoid the repeated code. For example. Consider. we find the following statement: An algorithm or computer program consists of two essential parts. sum := 0. 8. and a description of the data. cin >> n. A function is a subroutine that returns a result. } 122 while E do S It is occasionally useful to perform the test after the body: Pascal provides repeat/until and C provides do/while.2 Procedures and Functions Some languages. which are manip- . called a “side effect”. a function may have some effect on program data. while n ≥ 0 do begin sum := sum + n. In C and C++. This is not quite enough. Both names refer to “subroutines” — program components that perform a specific task when invoked from another component of the program. as shown in Listing 63. if (n < 0) break. Languages that make a distinction between procedures and functions usually also make a distinction between “actions” and “data”. sum : int. sum += n. at the beginning of the first Pascal text (Jensen and Wirth 1976). distinguish the terms procedure and function. read(n). A procedure is a subroutine that has some effect on the program data but does not return a result. such as Ada and Pascal. Some recent languages.

. elem is set to the special value nil. . For example we can write the standard “swap sequence” t := a. b := t. (If the object is not found. For example. elem: Element) is . 123 In other languages. in the more concise and readable form a. In C. Ada and Pascal combine the concerns but. In these languages.3. there are situations in which these constructs are inadequate. such as Algol and C. Consistently. 2 8.8 STRUCTURE ulated by the actions. if a function does not return a result. we can define a routine with this heading: findElem (index: Integer) -> (found: Boolean. We can view the progression here as a separation of concerns. the distinction between “procedure” and “function” also receives less emphasis. provided they are implemented properly. There is a single syntax for declaring and defining all functions and a convention that. b := b. a 2 Exercise 34. for example. it is given the return type void. and the routine returns two values: found to indicate whether something was found. Describe a “proper” implementation of the multi-assignment statement. the distinction between “actions” and “data” is less emphasized. The “concerns” are: whether a subroutine returns a value.3 Exceptions. Using Example 29 and other examples. C. Although in principle we can do everything with conditions and loops. for practical reasons. thing := findElem("myKey") Multi-assignments. . Actions are described be so-called statements.) Adding multiple results introduces the problem of how to use the results. and elem as the object found. discuss the introduction of new features into PLs. a := b. The basic control structures are fine for most applications provided that nothing goes wrong. Example 29: Returning multiple results. every “statement” has a value as well as an effect. The parameter index is passed to the routine findElem. Can you think of any other applications of multi-assignment? 2 Exercise 35. Blue also provides multi-assignments: we can write success. and data are described by so-called declarations and definitions. and whether it has side-effects. does not have keywords corresponding to procedure and function. . Blue (K¨lling 1999) has a uniform notation for o returning either one or more results. allow “functions with side-effects”. provide other advantages. there are two concerns: “having a value” and “having an effect” and they are essentially independent.

C provided primitives (setjmp/longjmp) that made it possible to implement exceptin handling. however. which is almost as much work as inverting it. adding overhead. The exception can be caught by a handler that can be installed at any convenient point in the calling environment. We could test every divisor before performing a division. because we know that most of the matrices that we will receive are non-singular. This is wasteful. the divisions probably occur inside nested loops: each level of loop will require Boolean variables and exit conditions that are triggered by a zero divisor. the test involves computing the determinant of the matrix. . see Section 3. Moreover. This is wasteful if the matrix is unlikely to be singular. a singular matrix is passed to the function and. the given matrix is non-singular and the inversion proceeds normally. Suppose that we are writing a function that inverts a matrix. What choices do we have? We could test the matrix for singularity before starting to invert it. Unfortunately. 2 PL/I (Section 3.5) was the first major language to provide exception handling. For further descriptions of exception handling in particular languages. Rely on the hardware to signal an exception when division by zero occurs. at some point in the calculation.8 STRUCTURE 124 Example 30: Matrix Inversion. In almost all cases. a division by zero will occur. Occasionally.

If we systematically change n to k. 9. we say that “n occurs bound” in the expression 5 (2n + 1). the formula is also true because (13) is true for all values of n. We say that the formula ∀ n . or use of . because n is bound to a value (actually. Second. however. the expression 0 ∞ e−x dx. the convention is that it denotes a function. We say that “x occurs free” in an expression such as x2 + 2x + 5. n mod 2 = 0 ⇒ n2 mod 2 = 0 is closed because it contains no free variables. n=0 (12) More precisely.. we should specify the range of values that n is allowed to assume. which contains a binding occurence of x (in dx) and a use of x (in e−x ). Similarly. is interesting because it binds x but does not have a numerical value. In this case. First. is ambiguous. 1. Strictly. Or we can use the notation of the λ-calculus: λx . But what about −e−ax : is it a function of a or a function of x? There are several ways of resolving this ambiguity.9 Names and Binding Names are a central feature of all programming languages. In the earliest programs. whose truth value depends on the value of n: n mod 2 = 0 ∧ n mod 3 = 1. 125 (13) .1 Free and Bound Names The use of names in programming is very similar to the use of names in mathematics. the binding occurrence of n is in n = 0. − e−ax . 5 }). including machine addresses. We do not know anything about the value of this expression because we do not know the value of x. the set of values { 0. Replacing numbers by symbolic names was one of the first major improvements to program notation (Mutch and Gill 1954). . the particular name that we use does not change the value of the expression. by convention. we obtain the expression 5 k=0 (2k + 1). as in the following expression. In predicate calculus. There are two things to note about (12). The conventional notation for functions. which has the same value as (12). numbers were used for all purposes. 36. We can write x → −e−ax . The expression e−x dx. The n in the expression (2n + 1) is a reference to. read “x maps to −e−ax ”. On the other hand. to this binding. has a value that does not depend on the particular name x. it has a definite value. e denotes a constant. . It is clear that −e−x is a function of x because. predicates may contain free names. −e−x . Names are bound by the quantifiers ∀ (for all) and ∃ (there exists). .

will accept the function f defined above only if it is compiled in a scope that contains a declaration of k. In (12). then n is never bound. in the sequence int n. The times at which the bindings occur include: compile time. 1. block entry time. or statement execution. address. In the function int f (int n) { return k + n. as in ∀ n ∈ Z . such as FORTRAN and PL/I. 9.) The second binding occurs when the program is executed. for example. In programming. this is now understood to be error-prone and is not a feature of recent PLs. . neither this fact nor the name n appears explicitly in the compiled code.9 NAMES AND BINDING 126 We could do this implicitly: for example. The address and value of n are bound only when the function f is called. . Note that we cannot tell the value of n from the definition of the function but we know that n will be given a value when the function is called. } k occurs free and n occurs bound. Consider the program shown in Listing 64. In most PLs. if f is not called (because k ≤ 0). a variable normally has only one attribute: its value. n mod 2 = 0 ⇒ n2 mod 2 = 0. a name may have several attributes. Alternatively. and value. A C compiler. we specify the domain from which these values are chosen. Example 31: Binding. load time. Sometimes. (The compiler records the fact that the type of n is int. its value is bound by the scanf call. and they may be bound at different times. A binding is static if it occurs during before the program is executed: during compilation or linking. Some early PLs. n = 6. . and the time at which the binding occurs.. the first line binds the type int to n and the second line binds the value 6 to n. link time. A binding is dynamic if it occurs while the program is running: during loading. The address of k is bound when the program starts. Precisely analogous situations occur in programming. Definition. and statement execution time. as in n ∈ Z. The attributes that may be bound to a name include: type. block entry. we could define the range of values explicitly. n is bound to the values 0. it is very likely that (13) refers to integers. For example. The first binding occurs when the program is compiled. 2 . provided implicit declarations for variables that the programmer did not declare.2 Attributes In mathematics. a program with free variables will not compile. 5. This example shows that there are two aspects of binding that we must consider in PLs: the attribute that is bound.

without any indexing or other address calculations. In these languages. n). When a function is called. by the time execution begins. local variables are allocated on the run-time stack. if (k>0) f(). The compiler assigns an address relative to a compilation unit. When the program is loaded. In FORTRAN. } 127 9. such as Pascal and C and most of their successors. have yet another way of binding an address to a variable. } void main () { int k. and also prevents the use of direct or indirect recursion. In languages of the Algol family (Algol 60. the history of software development is the history of ever-later binding time.9 NAMES AND BINDING Listing 64: Binding void f () { int n=7. the address of the unit within the program is added to this address. As a rough rule of thumb: Early binding ⇒ efficiency. etc. which is slightly slower then indexed addressing. Example 32: Variable Addressing. It is more flexible than FORTRAN because inactive functions do not use up space and recursion works. because absolute addressing is used.) FORTRAN is efficient. The variable must be addressed by adding an offset (computed by the compiler) to the address of the AR. variables are addressed directly. addresses are bound to variable names at compile time. Pascal. Algol-style is slightly less efficient than FORTRAN because addresses must be allocated at blockentry time and indexed addressing is required. Algol 68. because all addresses are assigned at load time. a stack frame (called. Languages that provide dynamic allocation. (In reality. because all local variables occupy space whether or not they are being used. C. the address of the program is added to the address. The result is that.). which is not known until the function is called.3 Early and Late Binding An anonymous contributor to the Encyclopedia of Computer Science wrote: Broadly speaking. When the program is linked. It is inflexible. The important point is that. in the compiled code. scanf("%d". This leads to wasted space. the “absolute” address of the variable is known. &k). Late binding ⇒ flexibility. printf("%d". the process is somewhat more complicated. but greater . Accessing the variable requires indirect addressing. in this context. an activation record or AR) is created for it and space for its parameters and local variables is allocated in the AR. a statement such as new(p) allocates space on the “heap” and stores the address of that space in the pointer p.

the search can be avoided in up to 95% of calls by using a cache. and body B. Definition. Note. 2 9. This statement is true of all PLs developed before OOP was introduced. In Smalltalk. but provides greater flexibility. The corresponding definition in Scheme would be: (defun f (lambda (x) E)) . Example 34: Functions. we can name functions but we cannot denote them. A virtual function name may refer to several functions. OOPLs provide “virtual” functions. The actual function referred to in a particular invocation is not in general known until the call is executed. In OOPLs. When the compiler encounters a function definition. several steps must be completed before the absolute address of the function is determined. however. because the CM requires a search. it binds an address to the function name. In practice. 2 Example 33: Function Addressing. with the result that calls are efficiently executed but no decisions can be made at run-time. (As above. the overhead can be significant. but this is not relevant to the present discussion. Thus functions in C can be named. Calls take longer to execute. An entity can be denoted if the PL provides ways of expressing instances of the entity as expressions. functions are statically bound.) When the compiler encounters a function invocation. In C++. first out). the compiler generates “virtual function tables” and dynamic binding requires only indirect addressing. functions are dynamically bound.4 What Can Be Named? The question “what can be named?” and its converse “what can be denoted?” are important questions to ask about a PL. but there is greater flexibility because the decision as to which function to execute is made at run-time. parameter x. The overhead of a dynamically-bound function call depends on the language. that this provides yet another example of the principle: Smalltalk binds later than C++. runs more slowly. Thus functions in C cannot be denoted. There is no way that we can write a function without a name in C. In C. it generates a call to the address that it has assigned to the function.9 NAMES AND BINDING 128 flexibility is obtained because dynamic variables do not have to obey stack discipline (last in. An entity can be named if the PL provides a definitional mechanism that associates a name with an instance of the entity. The definition int f (int x) { B } introduces a function with name f . Thus in non-OO PLs.

Many language designers have preferred to extend [minor. API p. but we could not give a name to the type of p (“array of 10 pointers to int”). For example.5 What is a Variable Name? We have seen that most PLs choose between two interpretations of a variable name. leading wildly from one part of a data structure to another. variable names denote objects in memory. but the assignment p := y + 1 may. References are like jumps. Assigning to a variable. The notation is analogous to that of the λ-calculus: f = λx . pointer. .1. in Algol 68. E to be used as functions. . pointers. localized faults in Algol 60 and other PLs] throughout the whole language by introducing the concept of reference. One variable it can never change is p! . Worse still. if permitted by the language. It is misleading in that providing pointers is not the same as providing reference semantics. 2 9. the assignment x := y always changes x. E FPLs allow expressions analogous to λx . This immediately gives rise in a high-level language to one of the most notorious confusions of machine code. making types nameable: typedef int *API[10]. changes the denotation of the name to a different object. LISP. This is reasonable in the sense that the implementation of reference semantics requires the storage of addresses — that is. or indirect address into the language as an assignable item of data. Haskell. as mentioned in Section 4.9 NAMES AND BINDING 129 This definition can be split into two parts: the name of the function being defined (f) and the function itself ((lambda (x) E)). The distinction is particularly clear in Hoare’s work on PLs. and other FPLs. if p is a reference variable. 2 Example 35: Types. Later versions of C introduced the typedef statement. is not one of these languages. an indirect assignment through a pointer. For example. . and the damage is no longer confined to the variable explicitly names as the target of assignment. Scheme is. change any other variable (of appropriate type) in the whole machine. In early versions of C. just as in machine code. Assigning to a variable changes the contents of the location named by the variable. In a PL with value semantics. Some languages attempt to solve this by even more confusing automatic coercion rules. Reference semantics is sometimes called “pointer semantics”. can update any store location whatsoever. In a PL with reference semantics. . Their introduction into high-level languages has been a step backward from which we may never recover. namely that between an address and its contents. and so are SML. However. we could write int *p[10]. types could be denoted but not named. Hoare (1974) has this to say about the introduction of pointers into PLs. variable names denote memory locations.

the program cannot tell whether X and Y are distinct objects that happen to be equal. “many shapes”. would be wasteful because there is no point in making copies of immutable objects. This is most naturally achieved with a reference semantics. y. a significant improvement on the efficiency of compiled LISP. of course. These two functions are provided partly for efficiency and partly to cover up semantic deficiencies of the implementation. Some other languages provide similar choices for comparison. and without the introduction of references. . printf("%d %f". It follows that value semantics. the terms “ad hoc polymorphism” and “parametric polymorphism” are due to Christopher Strachey. If a program object X corresponds to a unique entity in the world. and we will show that it can be satisfactorily accommodated in a high-level language using solely high-level problemoriented concepts. n. .) One of the important aspects of OOP is object identity . “polymorphism” is used to describe a situation in which one name can refer to several different entities. Ad hoc polymorphism is also called “overloading”.6 Polymorphism The word “polymorphism” is derived from Greek and means. x + y). Explicit types make it possible to achieve . such as a person. Typically the compiler uses the types of the arguments of the function to decide which function to call. 9. In general. In the expression m + n. ad hoc polymorphism refers to the use of a single function name to refer to two or more distinct functions. (Hoare 1975) All FPLs and most OOPLs (the notable exception. float x. (eq x y) is true if x and y are pointers to the same object. (LISP provides two tests for equality.9 NAMES AND BINDING 130 One year later. There are several kinds of polymorphism. it stands for the function that adds two floats. m + n. literally. it stands for the function that adds two integers. 9. The use of reference semantics in OOP is discussed at greater length in (Grogono and Chalin 1994). If X and Y are have the same value. Hoare (1975) provided his solution to the problem of references in high-level PLs: In this paper. There are good reasons for this choice. . or pointers to the same object. which requires copying. we will consider a class of data structures for which the amount of storage can actually vary during the lifetime of the data. but the reasons are not the same for each paradigm. perhaps even a factor of two in space-time cost for suitable applications. In a FPL. In the expression x + y. (equal x y) is true if the objects x and y have the same extensional value.6. all (or at least most) values are immutable. . The most common application is to functions. The implementation that Hoare proposes in this paper is a reference semantics with types.1 Ad Hoc Polymorphism In the code int m. the symbol “+” is used in two different ways. it should be unique in the program too. In PLs. being C++) use reference semantics.

and “::” denotes list construction. The following code declares three distinct functions in C++. (Strictly. Example 37: Lists in SML. SML can infer nothing about the type of the components from the function definition. int z). int y). functions can be defined by cases.) In general. “−”. all that the programmer has to do is write several definitions. has parametric polymorphism.6. Ada. we can make declarations such as these: list(int) list(float) Suppose also that we have two functions: sum computes the sum of the components of a given list. int y. the type of the operator “+” must be int × int → int. float max (float x. we should say “overload function names” but the usage “overloaded functions” is common. α is a type variable.2 Parametric Polymorphism Suppose that a language provides the type list as a parameterized type. float y).9 NAMES AND BINDING 131 Almost all PLs provide ad hoc polymorphism for built-in operators such as “+”. and this implies that we must know the type of the components. A function such as len. In order to compute the sum of a list. Example 36: Overloaded Functions. acting as a type parameter.: sum [] = 0 | sum (x::xs) = x + sum(xs) len [] = 0 | len(x::xs) = 1 + len(xs) SML infers the type of sum to be list(int) → int: since the sum of the empty list is (integer) 0. There is an important difference between these two functions. C++. and it assigns the type list(α) → int to the function. int max (int x. When len is applied to an actual list. and other recent languages also allow programmers to overload functions. but ensuring that the either the number or the type of the arguments are different. In SML. and all of the components must be integers. Here are functions that sum the components of a list and count the components of a list. On the other hand. 2 . we must be able to choose an appropriate “add” function. which counts the components of a list but does not care about their type. using the same function name. “[]” denotes the empty list. respectively. int max (int x. α will implicitly assume the type of the list components. “*”. We write definitions without type declarations and SML infers the types. That is. there seems to be no need to know the type of the components if all we need to do is count them. and len computes the number of components in a given list. 2 9. For the function len. etc.

Their scope starts at the declarations and end at the end of the program unit (that is.6. The value of E will be assigned to x anyway. The details of this kind of polymorphism. However. the semantics of this statement will be something like: evaluate the expression E. 9. This method of “scope control” is insecure because the compiler does not attempt to check consistency between program units. The compiler will generate code to convert the value of expression E to the type of x. as shown in Listing 67. the effect of invoking f may depend on the object. This is the approach taken by PLs that use dynamic type checking. The assignment is unproblematic if x and E have the same type. Pascal provides only a few coercions (subrange to integer. For example. 9. say. integer to real. which are discussed in Section 6.) and rejects other type differences. The statement is considered to be an error. aspect of OOP. . 9.7 Assignment Consider the assignment x := E. In C++. the scope of a local variable starts at the declaration of the variable and ends at the end of the block in which the declaration appears. This approach was taken to extremes in COBOL and PL/I. It is also possible to make variables accessible in selected subroutines by means of COMMON statements. as shown in Listing 65. It exists in C and C++. Scope is a static property of a name that is determined by the semantics of the PL and the text of the program.9 NAMES AND BINDING 9. as shown in Listing 66. f . This kind of polymorphism is a fundamental. Subroutine names have global scope.8 Scope and Extent The two most important properties of a variable name are scope and extent. for example. subroutine or main program) within which the declaration appears. Scopes are nested : a declaration of a name in an inner block hides a declaration of the same name in an outer block.3 Object Polymorphism 132 In OOPLs. there may be many different objects that provide a function called. depend on the particular OOPL. store the value obtained in x. and very important.8. etc.1 Scope The scope of a name is the region of the program text in which a name may be used. but C++ is stricter than C. But what happens if they have different types? There are several possibilities. FORTRAN. Whatever the PL. Algol 60. Examples of other scope rules follow. Local variables are declared at the beginning of a block. This will occur in a PL that provides static type checking but does not provide coercion. Types are associated with objects rather than with names. Local variables are declared at the start of subroutines and at the start of the main program.

.. x := 3. y(100) integer k(250) common /myvars/ x.0 end. begin real x. y .. b .9 NAMES AND BINDING 133 Listing 65: Local scope in C++ { // x not accessible. t x.. end . } // x still accessible } // x not accessible Listing 66: FORTRAN COMMON blocks subroutine first integer i real a(100) integer b(250) common /myvars/ a. x := 6..... integer k.. k. { // x still accessible in inner block . return end subroutine second real x(100). // x accessible ..0... return end Listing 67: Nested scopes in Algol 60 begin real x.

n. C. Pascal. or public. structure. the order of declarations was fixed: constants. . a class may be derived as private. When a child class C inherits a parent class P. (There may be “holes” in these global scopes if the program contains local declarations of the same names. other names. Finally. Global scope is useful for pervasive entities. such as library functions and fundamental constants (π = 3. in part. Global Scope. Class attributes may be declared private (visible only with the class). protected (visible within the class and derived classes). statement labels. In early Pascal. component names. A function or a class may be declared as a friend of a class. Names declared at the beginning of the outermost block of Algol 60 and Pascal programs have global scope. C++.. and public (visible outside the class). some or all of the attributes of P are made visible in C.9 NAMES AND BINDING 134 The fine control of scope provided by Algol 60 was retained in Algol 68 but simplified in other PLs of the same generation. Later dialects of Pascal sensibly relaxed this ordering.f. Subroutine names in FORTRAN are global. however.1415926) but is best avoided for application variables. inheritance is. protected. a scoping mechanism. union. and enumeration tags. a field f of the record r may be written as f rather than as r. The name is visible throughout the program. Similarly. giving it access to the private attributes of the class.) The largest default scope in C is “file scope”. C++ follows C in providing directives such as static and extern to modify scope. Consequently. . Function declarations cannot be nested and all local variables must appear before executable statements in a block. types.. the following declarations do not conflict: { char str[200].. Pascal also has a rather curious scoping feature: within the statement S of the statement with r do S. are complicated by the fact that C has five overloading classes (Harbison and Steele 1987): preprocessor macro names. Control structures and nested functions are nested but declarations are allowed only at the start of the program or the start of a function declaration. struct str {int m. and functions were declared in that order. There are a number of additional scope control mechanisms in C++. In OOPLs. FORTRAN does not have global variables. variables. although programmers simulate them by over using COMMON declarations. Global scope can be obtained using the storage class extern. The complete rules of C scoping. } x. } Inheritance.

where is it used to select attributes and functions of an object. These names are usually hidden. const int i = 7.f(). There is a question as to whether the scope starts at the point of definition or is the entire block. // OK 135 f is not in scope Local Scope. In Algol 60 and C++. a parameter list follows: o. opens a scope in which the public attributes of the object w are visible. it seems best to provide the programmer with as much freedom as possible and to keep scopes as small as possible. Listing 69 provides an example in C++. The construct w. Import and Export. // Error: w. Qualified Scope. does the PL require “declaration before use”?) C++ uses the “declare before use” rule and the program in Listing 68 prints 6. such as a Pascal record or a C struct. Any statement context can be instantiated by a block that contains local variable declarations. In modular languages. The dot notation also appears in many OOPLs. }. but can be made visible by the name of the object. The mechanisms above are inadequate for very large programs developed by large teams. local scopes can be as small as the programmer needs. local declarations can be used only in certain places.9 NAMES AND BINDING Listing 68: Declaration before use void main () { const int i = 6. if module A exports name X and module B imports name X from A. Although a few people do not like the fine control offered by Algol 60 and C++. the syntax is the same: o. a name or group of names can be brought into a scope with an import clause. . For an attribute. If the attribute is a function. Widget w. then X may be used in module B as if it was declared there. In block-structured languages. such as Modula-2.a denotes attribute a of object o. } } Listing 69: Qualified names class Widget { public: void f(). Namespaces. The components of a structure. Problems arise when a project uses several libraries that may have conflicting names. (In other words. have names.f(x) denotes the invocation of function f with argument x in object o. f(). such as Pascal. Typically. cout << j << endl. { int j = i. names declared in a block are local to the block. In other languages. The module that provides the definitions must have a corresponding export clause.

when the source of names is obvious. and D”. the cause of occasional errors that could be eliminated by requiring all names in a statement or function to be unique. C. It is. for example in classes and modules. provided by PLs such as C++ and Common LISP. all names supplied by the commercial graphics library FastGraph are have fg_ as a prefix. for example in statement and functions. The effect is that the nested class can be accessed only by the enclosing class. Wilden. provide a higher level of name management than the regular language features. such as functions. In fact. one way to view compiling is as a process that converts names to numbers. such as classes or modules. Library designers can reduce the potential for name conflicts by using distinctive prefixes. . and Wolf 1980.8. C++ allows classes to be nested (this seems inconsistent. Hanson 1981). The advantage of nested scopes for large structures. however. given that functions cannot be nested). Qualified names are inadequate for very large programs. It is probably better to provide special mechanisms to provide the scope control that is needed rather than to rely on simple nesting. Nesting is a form of scope management. nesting is less effective because there is more text and a larger number of names. 9. For example. An alternative mechanism would be a declaration that states explicitly that class A can be accessed only by class B.2 Are nested scopes useful? As scope control mechanisms became more complicated. some people began to question to need for nested scopes (Clarke. This is because PLs typically hide the scoping mechanisms. Namespaces provide a better solution than prefixes. On a small scale. Interpreters generally have a repository that contains the value of each name currently accessible to the program. On a large scale. This mechanism would have the advantage that it would also be able to describe constraints such as “class A can be accessed only by classes B. If a program is compiled. explicit control by name qualification may be better than nesting. nesting works fairly well. but programmers may have limited access to this repository. In summary: Nested scopes are convenient in small regions. The import mechanism has the problem that the source of a name is not obvious where it appears: users must scan possibly remote import declarations. Scope management is important because the programmer has no “work arounds” if the scoping mechanisms provided by the PL are inadequate. is doubtful. For large structures.9 NAMES AND BINDING 136 Namespaces. names will normally not be accessible at run-time unless they are included for a specific purpose such as debugging. Qualified names work well at the level of classes and modules.

some try to prevent it from occurring (e. and all FPLs. Pascal). In PLs that use a reference model. Understanding the relation between scope and extent is an important part of understanding a PL. It is an error to create an object on the stack and to pass a reference to that object to an enclosing scope. Global names usually exist for the entire lifetime of the execution: they have global extent..8.g.g. Algol 68).3 Extent 137 The extent. on entry to a subroutine.. Smalltalk. The disadvantage is that GC and the accompanying overhead are more or less indispensable. Eiffel.. In FORTRAN. The separation of extent from scope was a key step in the evolution of post-Algol PLs.9 NAMES AND BINDING 9. whether or not their original names are accessible. Programmers assume that. and others leave it as a problem for the programmer (e. local variables are instantiated whenever control enters a block and they are destroyed (at least in principle) when control leaves the block. CLU. In Algol 60 and subsequent stack-based languages. of a name is the period of time during program execution during which the object corresponding to the name exists.g. Examples include Simula. local variables also exist for the lifetime of the execution. . C and C++). also called lifetime. Some PLs attempt to detect this error (e. The advantage of the reference model is that the problems of disappearing objects (dangling pointers) and inaccessible objects (memory leaks) do not occur. objects usually have unlimited extent. its local variables will not have changed since the last invocation of the subroutine.

5. loops. In the evolution of PLs. the explicit goal of the designers was a “problem oriented language”. .10 Abstraction Abstraction is one of the most important mental tools in Computer Science. and so on — are abstracted from the basic machine instructions. 4. . preferably an interface that does not depend on the particular machine that is being used. Most programs are connected to the world in one way or another. and so on — are abstracted from the raw bits of computer memory. begin s := [1. The control structures that we use — assignments. A PL must also provide ways for the programmer to abstract information about the world. an accounting program is. . if (n in s) . even at that time. the elements of a set are encoded as positions within a machine word. . . floats. conditionals. Here is an early example that contrasts the problem-oriented and machine-oriented approaches to PL design. 3. Abstraction is particularly important in the study of PLs because a PL is a two-fold abstraction. 9]. In contrast. 8]. end 138 . a simulation of tasks that used to be performed by clerks and accountants. Programs are seen as descriptions of the world that happen to be executable. the presence or absence of an element Listing 70: Sets in Pascal type SmallNum = set of [1. 6. as shown in Listing 71. The values that we use in programming — integers. A PL must provide an abstract view of the underlying machine. var n: integer. Programs are seen as lists of instructions for the computer to execute. and we can write expressions using set operations. loops. however. abstraction helps us to manage complexity. C provides bit-wise operations on integer data. var s. if (s <= t) . the first need came before the second. we can define literal constants and variables of these types. as shown in Listing 70. These features of Pascal and C have almost identical implementations..10]. The idea of abstracting from the world is not new. t := s + [2. The second task of a PL is to provide as much help as possible to the programmer who is creating a simplified model of the complex world in a computer program. 7. The first task of a PL is to hide the low-level details of the machine and present a manageable interface. For example. . and some programs are little more than simulations of some aspect of the world. Example 38: Sets and Bits. t: SmallNum. and arrays. booleans. In each case. Pascal provides the problem-oriented abstraction set. at some level. The biggest problems that we face in Computer Science are concerned with complexity. arrays. We can declare types whose values are sets. The design of Algol 60 began in the late fifties and. Early PLs emphasize abstraction from the machine: they provide assignments. Later PLs emphasize abstraction from the world: they provide objects and processes.

. 10.1. they have similar syntax and similar rules of construction. and the operations for and-ing and or-ing bits included in most machine languages are used to manipulate the words. and C take the view that everything is an expression. Procedural abstraction enables us to give a name to a statement. Interestingly. which exists in rudimentary form even in assembly language.) Thus statements have a value in these languages. where a procedure call appears in a statement context and a function call appears in an expression context.10 ABSTRACTION Listing 71: Bit operations in C int n = 0x10101010101. The trend. procedures and functions are closely related. and to use the name elsewhere in the program to trigger the execution of the statements. Algol 68. which also appeared at an early stage was the function. . respectively. Neither approach is necessarily better than the other: Pascal was designed for teaching introductory programming and C was designed for systems programmers. LISP provides function abstraction. Parameters enable us to customize the execution according to the particular requirements of the caller. is quite different.1. whereas the C programmer is forced to think at the level of words and bits. however. it is often given a special sense. An abstraction mechanism provides a way of naming and parameterizing a program entity. The difference between procedures and functions is most evident at the call site. The earliest abstraction mechanism was the procedure. int mask = 0x10000. or group of statements. since any computation leaves something in the accumulator. (The idea seems to be that. Typically. . 139 is represented by a 1-bit or a 0-bit. In Computer Science in general. The effect of procedures is obtained by writing functions with side-effects. Algol 60. the programmer might as well be allowed to use it.1 Procedures. COBOL does not provide procedure or function abstraction. we need to abstract form arbitrarily complex programs whose main purpose is to yield a single value. FORTRAN provides procedures (called “subroutines”) and functions. is away from machine abstractions such as bit strings and towards problem domain abstractions such as sets.2 Functions. however.1 Abstraction as a Technical Term “Abstraction” is a word with quite general meaning. In practice. and to use the name to trigger evaluation of the expression. if (n & mask) . The next abstraction mechanism. In many PLs. The Pascal programmer can think at the level of sets. 2 10. and in PL in particular. It is not surprising that . although the PERFORM verb provides a rather restricted and unsafe way of executing statements remotely. 10. Functional abstraction enables us to give a name to an expression. The appearance of the program. The reason for this is that an abstraction mechanism that works only for simple expressions is not very useful.

and boolean. Programmers who use a PL acquire intuitive knowledge of its CM.10 ABSTRACTION 140 this viewpoint minimizes the difference between procedures and functions. . Since this value is unique. the Algol report (Naur et al. 1960)). and that is the computational model (also sometimes called the semantic model ). 10.end set of T file of T Type expressions can be named and the names can be used in variable declarations or as components of other types.. the Pascal User Manual and Report (Jensen and Wirth 1976)). denotational. we can provide a formal semantics for the language. and type expressions (C denotes an integer constant and T denotes a type): C.C (C.. 10. Example 39: Passing parameters. . .g. without the details of every construct. . Pascal. Pascal was more systematic and more influential. 2 . a procedure is simple a function that returns a value of type void.3 Data Types. This means that the programs in Listing 72 and Listing 73 are equivalent. Alternatively.1. Although this was an important and profound step. We can write an informal description in natural language (e. such as integer.g. however. with hindsight we can see that two simple ideas were needed: giving a name to an Algol block. Pascal provides primitive types. The computational model (CM) is an abstraction of the operational semantics of the PL and it describes the effects of various operations without describing the actual implementation. or operational approaches.. . Simula introduced an abstraction mechanism for classes. it could pass the address of the argument a to the function but not allow the statement S to change the value of a. using the axiomatic. C) array [T] of T record v: T. . There is.. the implementation has failed to respect the CM. . a formal description in natural language (e. an abstraction that is useful for describing a PL in general terms. When a program behaves in an unexpected way. or a document in a formal notation (e.2 Computational Models There are various ways of describing a PL. Although Algol 68 provided a kind of data type abstraction. In Algol 68 and C..4 Classes. 1975)). 10. Both implementations respect the CM. The statement T a = E means “the variable a has type T and initial value E” and the statement x ← a means “the value of a is assigned (by copying) to x”. however.g. real.1. does not provide a parameter mechanism for types. Alternatively. as if executing the assignment x ← a. and freeing the data component of the block from the tyranny of stack discipline. the Revised Report on Algol 68 (van Wijngaarden et al. C. it requires log 1 = 0 bits of storage and can be optimized away by the compiler. A compiler could implement this program by copying the value of the argument. . Suppose that a PL specifies that arguments are passed by value.

The effect of executing call p(0) is to change all instances of 0 (the constant “zero”) in the program to 1 (the constant “one”)! The resulting bugs can be quite hard to detect for a programmer who is not familiar with FORTRAN’s idiosyncrasies. most FORTRAN compilers use the same mechanism for passing constant arguments. count = count + 1 return end then. but has a rather complex CM. S. C is easy to compile. although variable arguments may be changed by a subroutine. If a FORTRAN subroutine changes the value of one of its formal parameters (called “dummy variables” in FORTRAN parlance) the value of the corresponding actual argument at the call site also changes. Unfortunately.. This enables the subroutine to both access the value of the argument and to alter it. if we define the subroutine subroutine p (count) integer count . } T a = E.. Listing 73: Revealing the parameter passing mechanism T a = E. . after executing the sequence integer total total = 0 call p(total) the value of total is 1. Often.. The programmer has a mental CM for FORTRAN which predicts that. the opposite is true. { T x. The problem arises because typical FORTRAN compilers do not respect this CM. } 141 Example 40: Parameters in FORTRAN. constant arguments will not be. For example.10 ABSTRACTION Listing 72: Passing a parameter void f (T x). 2 A simple CM does not imply easy implementation. The most common way of achieving this result is to pass the address of the argument to the subroutine. For example. { S. x ← a. f (a).

10 ABSTRACTION 142 Example 41: Array Indexing. The underlying difficulty is that many computations depend on a concept of mutable state that does not exist in mathematics. In Smalltalk. because the compiler can process primitives without the overhead of treating them as full-fledged objects. 2 Example 42: Primitives in OOPLs. The practical consequence of referential transparency is that we can freely substitute equal expressions. but a naive implementation will probably be inefficient. Primitive values can have a special status that distinguishes them from objects. an array is a mapping. and integers. 2 introduces A as a mapping: . FPLs have an important property called referential transparency . To understand this. the replacement would save time if the evaluation of E requires a significant amount of work. The advantage of this point of view is that program text resembles mathematics and. FPLs are based on the theory of partial recursive functions. characters. we must know that an array can be represented by a pointer to its first element and that we can access elements of the array by adding the index (multiplied by an appropriate but implicit factor) to this pointer. They attempt to use this theory as a CM. . The disadvantage is that mathematics and computation are not the same. Early implementations of Smalltalk were inefficient. . An implementor is then free to implement primitive values efficiently provided that the behaviour predicted by the CM is obtained at all times. For example. programs can be manipulated as mathematical objects. because all variables behave like objects. However. There are two ways in which an OOPL can treat primitive entities such as booleans. On the other hand. An expression is referentially transparent if its only attribute is its value. to support various OO features. if E was not referentially transparent — perhaps because its evaluation requires reading from an input stream or the generation of random numbers — the replacement would change the meaning of the program. because all variables were actually implemented as objects. In the CM of C. Primitive values can be given the same status as objects. are not easily described in classical mathematics.. This simplifies the programmer’s task. In Pascal’s CM. also has classes for the primitive types. A language that uses this policy is likely to be efficient. A compiler can implement arrays in any suitable way that respects this mapping.. every variable is an object. it may be confusing for the user to deal with two different kinds of variable — primitive values and objects. Some operations that are intuitively simple. if we are dealing with numbers. The declaration var A : array [1. such as maintaining the balance of a bank account subject to deposits and withdrawals. A solution for this dilemma is to define a CM for the language in which all variables are objects. . 2 Example 43: λ-Calculus as a Computational Model. A : { 1. to a certain extent. a[i] ≡ ∗(a + i). N } → R. we can replace E + E by 2 × E. Java provides primitives with efficient implementations but.N ] of real.

structured programs by recursive decomposition of code. pages 71–74). page 72).10 ABSTRACTION 143 Example 44: Literals in OOP. “just magically exist in the Smalltalk universe”. Do you think that K¨lling’s arguments for introducing a new concept (manifest o classes) are justified? 2 Summary Early CMs: modelled programs as code acting on data. These objects. The advantage of the Blue approach. Dee (Grogono 1991a) uses a variant of the first alternative: literals are constructors. How should an OOPL treat literal constants. The enumeration may be finite. . o There are two possible ways of interpreting the symbol “7”: we could interpret it as an object constructor . Blue avoids these problems by introducing the new concept of a manifest class. Later CMs: modelled programs as packages of code and data. [Emphasis added. such as 7? The following discussion is based on (K¨lling 1999. and differences between numbers and complex objects can be understood independently of implementation concerns. This introduces several problems. is that: o . constructing the object 7. constrain but not determine the implementation. . . Literals are constant references to objects. For example. The object model explicitly recognizes these two different types of classes. Manifest classes are defined by enumeration. as for class Integer. however. structured programs by recursive decomposition of packages of code and data. The CM should: help programmers to reason about programs. They cease to be anomalies or special cases at a technical level. as for class Boolean with members true and false. according to K¨lling (1999. the distinction is made at the logical level. help programmers to read and write programs. . or as a constant reference to an object that already exists. Smalltalk uses the second approach. This reduces the dangers of misunderstandings on the side of the programmer. are allowed.] 2 Exercise 36. Smalltalk does not explain where they come from. what does the comparison 7 = 7 mean? Are there two 7-objects or just one? The CM of Dee says that the first occurrence of the literal “7” constructs a new object and subsequent occurrences return a reference to this (unique) object. Infinite enumerations. although we cannot see the complete declaration. or when and how they are created.

there are few interesting things that we can say about implementation. Programmes in the future may be written in a very different language from that in which the machines execute them. (Gill 1954) A compiler translates source language to machine language. the code generated by a compiler is usually not optimal. COBOL. The executable file is loaded into memory and executed. Implementation techniques are a separate. 3. Since practical compilation systems allow programs to be split into several components for compiling. People also realized that the computer itself could perform the translation from a high level notation to machine code. “conversion” is used in roughly the sense that we would say “compilation” today. the checking was quite superficial and it was easy to write programs that compiled without errors but failed when they were executed. such as FORTRAN. 11. For very small programs. The value of compile-time checking began to be recognized in the late 60s. Before he can find the cause of the trouble the programmer will either have to investigate the conversion process or enlist the aid of checking routines which will ‘reconvert’ and provide him with diagnostic information in his own language.1 Compiling and Interpreting Compiling As we have seen. In the following quote. large storage capacities will lead in the course of time to very elaborate conversion routines.1. and very rich. It is to be hoped (and it is indeed one of the objects of using a conversion routine) that many of the tiresome blunders that occur in present-day programmes will be avoided when programmes can be written in a language in which the programmer feels more at home.11 Implementation This course is concerned mainly with the nature of PLs. any mistake that the programmer does commit may prove more difficult to find because it will not be detected until the programme has been converted into machine language. hand-coding such large programs would require excessive time. and C. Since machine code is executed directly. a compiler can perform global optimizations that are infeasible for a human programmer and.1 11. in any case. the need for programming languages at a higher level than assembly language was recognized early. In practice. 1. The components of the source program is compiled into object code. compiled programs ideally run at the maximum possible speed of the target machine. and languages such as Pascal. For large programs. a programmer skilled in assembly language may be able to write more efficient code than a compiler can generate. Ada. The object code units and library functions required by the program are linked to form an executable file. All compilers perform some checks to ensure that the source code that they are translating is correct. However. it will not fail when 144 . there are usually several phases. and C++ provided more thorough checking than their predecessors. Nevertheless. and that is the purpose of this section. The goal of checking is to ensure that. 2. On the other hand. topic. In early languages. PL/I. if a program compiles without errors.

it reads all of the relevant interfaces and performs full checking.2 Interpreting An interpreter accepts source language and run-time input data and executes the program directly. C and C++ rely on the programmer to provide header files containing declarations and implementation files containing definitions. The solution is “name mangling” — the C++ compiler alters the names of functions by encoding their argument types in such a way that each function has a unique name. It is possible to interpret without performing any translation at all. even today. prototyping environments often provide an interpreted language. The compiler checks that declarations and definitions are compatible. run-time failure. Since no time is spent in compiling. at best. Consequently. most interpreters do a certain amount of translation before executing the code. page 34). such as an abstract syntax tree. however. because programs do not contain type declarations. Dee. the linker will report subroutines that have not been defined but it will not detect discrepancies in the number of type of arguments. The compiler automatically derives an interface from the class declaration and uses the generated interfaces of other classes to perform full checking. but more recent interactive systems provide an option to compile and may even provide compilation only. In practice. An Algol 60 program is a block and must be compiled as a block. Eiffel programs consist of classes that are compiled separately. Interpreting is slower than executing compiled code. syntactic) errors as the program is read. We can identify various levels of checking that have evolved. Most interpreters check correctness of programs during execution. When the compiler compiles an implementation module.11 IMPLEMENTATION 145 it is executed. Standard Pascal programs cannot be split up. 11. Language such as LISP and Smalltalk are often said to be untyped. A key design decision in C++. The result of such a discrepancy is. however. Other OOPLs that use this method include Blue. there are few languages that achieve it totally. but this is incorrect. Early BASIC and LISP systems relied on interpretation. but this is not logically necessary. This goal is hard to achieve and. although some interpreters may report obvious (e.1. A FORTRAN compiler provides no checking between components. and Java. interpreted languages can provide immediate response. was to use the standard linker (Stroustrup 1994. the interface modules are themselves compiled for efficiency. Modula-2 programs consist of interface modules and implementation modules. typically by a factor of 10 or more. The type checking performed by both of these languages is actually more thorough than . which is then “executed”.g. The need to compile programs in parts was recognized at an early stage: even FORTRAN was implemented in this way. For example. The danger of separate compilation is that there may be inconsistencies between the components that the compiler cannot detect. but this approach tends to be inefficient. For this reason. the source program might be converted into an efficient internal representation.. Overloaded functions in C++ present a problem because one name may denote several functions. but recent dialects of Pascal provide program components. Compiled PLs can be classified according to whether the compiler can accept program components or only entire programs. When the compiled components are linked.

When that compiler is available. it can be extended to compile the full language. it is necessary only to provide a Java byte code interpreter for each platform. all of these steps must be carried out with great care. Any program that is to be distributed over the internet must be capable of running on all the servers and clients of the network. be run on any machine. but it is simplified by the fact that the compiler is written in Pascal. 4. but only those parts that are needed by the compiler itself. 11. Since P-code is quite simple. Needless to say. created in step 2 does not have to compile the entire Pascal language. A Pascal implementor had to perform the following steps: 1. Use the P-code interpreter to interpret the P-code compiler.11 IMPLEMENTATION 146 that performed by C and Pascal compilers. The Pascal “P” compiler provided a simple method for porting Pascal to a new kind of machine. this is not a very arduous job. this may be too late. It is important to ensure that each step is completed correctly before beginning the next. 3. in principle. are simply streams of bytes that represent the program: a machine language for an “abstract machine”. but it is inefficient because it depends on the P-code interpreter. The “byte codes”. the first version of the compiler.3 Mixed Implementation Strategies Some programming environments. Java uses byte codes to obtain portability in a different way. For safety-critical applications. Byte Codes. . This step yields a P-code program that can be used to translate Pascal programs to P-code. At this stage. particularly those for LISP. The advantage of byte code programs is that they can. but may be incomplete in other respects. If the program is written in Java. such as C. 2. This is a somewhat tedious task. With the exercise of due caution. Use the compiler created in step 2 to compile the modified compiler. Java was designed to be used as an internet language. the process described above (called bootstrapping ) can be simplified. The result of this step is a P-code program that translates Pascal to machine code. Use the result of step 4 to compile the modified compiler. Write a P-code interpreter. In fact. the implementor has a working Pascal system. provide both interpretation and compilation. If the program is written in a conventional language. Thus byte codes provide a convenient way of implementing “portable” languages. Another mixed strategy consists of using a compiler to generate byte codes and an interpreter to execute the byte codes. some LISP environments allow source code and compiled code to be freely mixed. as the name implies. One of the first programming systems to be based on byte codes was the Pascal “P” compiler. a Pascal compiler for the new machine. For example. The complete sequence of steps can be used to generate a Pascal compiler that can compile itself. The result is a machine code program that translates Pascal to machine code — in other words. 5.1. The “Pascal P package” consisted of the P-code compiler written in P-code. the difference is that errors are not detected until the offending statements are actually executed. Modify the compiler source so that the compiler generates machine code for the new machine. there must be a compiled version for every platform (a platform consists of a machine and operating system).

The implementor does not really have a choice. if a statement has been executed once. The idea is that the program is interpreted. A PL intended for interpretation. but most OOPLs provide it: Simula. The converse is not true: an interactive environment for C++ would not be very useful. such as BASIC. This implies that the user must be able to enter short sequences of code that can be executed immediately in the current context. page 219). is usually implemented interactively. Many FPLs are implemented in this way (often with a compiler). or perhaps compiled rapidly without optimization. because the program can be run immediately. and Java. and so do all of its successors. This approach is based on the following assumptions: in a typical run. but in the opposite direction. provides GC.11 IMPLEMENTATION 147 Just In Time. but interpretation is not the only possible means of implementation. are less suitable for this kind of program development. and efficient execution. is C++.. because the second and subsequent cycles of a loop are executed from efficiently compiled code. FPLs are often supported interactively because a programmer can build up a program as a collection of function definitions.1. . efficient code is generated for it. of course. An implementation of LISP. The exception. 11. 11.2 Garbage Collection There is a sharp distinction: either a PL is implemented with garbage collection (GC) or it is not. and Smalltalk. The categories overlap: PLs that can be used interactively. The OOP community is divided on the issue of GC. CLU. LISP. The first FPL. or Java without GC would be useless — programs would exhaust memory in a few minutes at most. and a number of other interactive PLs followed this pattern of development. Smalltalk. Eiffel. the user can define new functions and invoke any function that has previously been defined. it is almost impossible to implement C or C++ with GC. Conversely. can also be used to develop large programs. As each statement is executed. LISP. An interactive PL is usually interpreted. The functional PL community has been equally consistent. Block structured languages. because programmers can perform arithmetic on pointers. it is likely to be executed again. Thus JIT combines some of the advantages of interpretation and compilation. where most functions are fairly small (e. “Just in time” (JIT) compilation is a recent strategy that is used for Java and a few research languages. PLs that start out with interpreters often acquire compilers at a later stage — BASIC.g. JIT provides fast response. A PL that is suitable for interactive use if programs can be described in small chunks. they fit on a screen).4 Consequences for Language Design Any PL can be compiled or interpreted. Smalltalk. Interactive Systems A PL can be designed for writing large programs that run autonomously or for interacting with the user. The imperative PL community has been consistently reluctant to incorporate GC: there is no widely-used imperative PL that provides GC. SML. however. LISP. and modular languages. much of the code of a program will not be executed. for reasons that have been clearly explained (Stroustrup 1994.

It is hard to argue with Stroustrup’s conclusion (Stroustrup 1994. I am fully convinced that had garbage collection been an integral part of C++ originally. these are sufficient arguments against the view that no application would be better done with garbage collection. In particular. (Stroustrup 1994. Many garbage collection techniques imply service interruptions that are not acceptable for important classes of applications. In typical . it simplifies the building and use of some libraries.] Stroustrup’s arguments do not preclude GC as an option for C++ and. page 220): I know that there are more reasons for and against. 148 It is interesting to study Stroustrup’s arguments in detail. 4. Some garbage collection schemes require banning several basic C facilities such as pointer arithmetic. and unchecked function arguments as used by printf(). Conservative GC depends on the fact that it is possible to guess with a reasonable degree of accuracy whether a particular machine word is a pointer (Boehm and Weiser 1988). The emphasis is added: Stroustrup makes it clear that he would have viewed GC in a different light if it provided advantages for all applications.11 IMPLEMENTATION I deliberately designed C++ not to rely on [garbage collection]. the “fundamental reasons” that garbage collection is desirable are (Stroustrup 1994. page 220): 1. I feared the added complexity of implementation and porting that a garbage collector would impose. 2. In the meantime. Some applications do not have the hardware resources of a traditional general-purpose computer. I like the idea of garbage collection as a mechanism that simplifies design and eliminates a source of errors. and operating system kernels. Stroustrup advocates this. In his view. most C++ programmers spend large amounts of time figuring out when to call destructors and debugging the consequences of their wrong decisions. 5. Similarly. Garbage collection is easiest for the user. These are sufficient arguments against the view that every application would be better done with garbage collection. Garbage collection causes run-time and space overheads that are not affordable for many current C++ applications running on current hardware. I feared the very significant space and time overheads I had experienced with garbage collectors. page 220) [Emphasis in the original. Stroustrup provides more arguments against GC than in favour of it (Stroustrup 1994. garbage collection would make C++ unsuitable for many of the low-level tasks for which it was intended. 3. 2. But it remains true that it is difficult to add GC to C++ and the best that can be done is probably “conservative” GC. control applications. Garbage collection is more reliable than user-supplied memory management schemes for some applications. However. Some garbage collection schemes impose constraints on object layout or object creation that complicates interfacing with other languages. human interface code on slow hardware. C++ would have been stillborn. such as hard real-time applications. but no further reasons are needed. Also. device drivers. in fact. unchecked arrays. page 220): 1.

Meyer (1992. working in small. however. and compiled with this compiler are used in ATM switches. and its value is a legal address. describe a garbage collector for a Scheme compiler. for example. It should be incremental. the GC can identify regions of memory that might be in use. Experience with C++ is beginning to affect the polarization. Implementors are encouraged. Modern garbage collectors have much lower overheads and are “tunable” in Meyer’s sense. with a low overhead on execution time. By scanning memory and noting words with these properties. For example. The concern about the efficiency of GC is the result of the poor performance of early. Several commercial products that offer “conservative” GC are now available and their existence suggests that developers will increasingly respect the contribution of GC. programmers should be able to switch off GC in critical sections and to request a full clean-up when it is safe to do so. (Erlang is a proprietary functional language developed by Ericsson. and those who use PLs without GC and consider it unnecessary. Larose and Feeley (1999). Programs written in Erlang. a pointer is a 4-byte quantity.11 IMPLEMENTATION 149 modern processors. . aligned on a word boundary. mark-sweep collectors. of identifying live data as garbage cannot occur: hence the term “conservative”. It should be tunable. but not required to provide GC. translated into Scheme. The implementation of Eiffel provided by Meyer’s own company. has a garbage collector that can be turned on and off. page 335) recommends that an Eiffel garbage collector should have the following properties.) GC has divided the programming community into two camps: those who use PLS with GC and consider it indispensable. frequent bursts of activity rather than waiting for memory to fill up. The converse. Eiffel. It should be efficient. is a public-domain language that can be implemented by anyone who chooses to do so — there are in fact a small number of companies other than ISE that provide Eiffel. and more serious problem. ISE. as the difficulties and dangers of explicit memory management in complex OO systems becomes apparent. The problem is that there will be occasionally be words that look like pointers but actually are not: this results in about 10% of garbage being retained.

A PL that provides both in a consistent and concise way might be simple. powerful abstractions — to achieve more with less. It might be fruitful to design a PL with objects. Ideally. Theory-based PLs are important because they provide useful insights. after values and objects. then bound. Such a language would combine the advantages of OOP and logic programming. functional programming. expressive. It is not clear that any future universal language will be accepted by the programming community. are in widespread use. Simula (Section 6. A rigorously mathematical approach can lead all too rapidly to the “Turing tar pit”. The evolution of PLs shows that. Both views are useful for particular applications. Multiparadigm Programming. most of the time. it seems more likely that specialized languages will dominate in the future. and powerful.12 Conclusion Values and Objects.2). In this respect. Similar simplifications have occurred in the evolution of PLs: for example. and later may be unbound during backtracking) is a third kind of entity. logical variables. now more or less subsumed by object oriented programming. Mathematically oriented views of programming favour values. PLs have evolved into several strands: procedural programming. Practice and Theory. Each paradigm performs well on a particular class of applications. logic programming and its successor. it is inevitable that some will try to build a “universal” PL by combining paradigms. Unfortunately. The “logical variable” (a variable that is first unbound. unification. Such attempts will be successful to the extent that they achieve overall simplification. they serve as testing environments for ideas that may eventually be incorporated into new mainstream PLs. PL design resembles mathematics. Simulation oriented views favour objects. The significant advances in mathematics are often simplifications that occur when structures that once seemed distinct are united in a common abstraction. The hard part of design is to find simple. Making Progress. Since people are always searching for “universal” solutions. it is often the case that they show retroactively how something should have been done when it is already too late to improve it. It is not hard to design a large and complex PL by throwing in numerous features. practice leads theory. if any.1) and Scheme (Section 4. then theoreticians attempt to what they did and how they could have done it better. constraint programming. Designers and implementors introduce new ideas. 150 . and backtracking. But programs are not mathematical objects. There are a number of PLs that have been based on purely theoretical principles but few.

W. T. April). F. Rx for semantics. Go to statement considered harmful. Java’s insecure parallelism. F. C. and languages with only two formation rules. Communications of the ACM 9 (5). J.-J. Kanoui. 10–18. and G. Academic Press. P. E. April). Partington. (1957. J. Programming Languages and Systems 20 (5). D. (1999. September). The Mythical Man-Month (Anniversary ed. and A. 1014–1066. Programming Linguistics. Groupe de Recherche en Intelligence Artificielle. The Java Series. and G. Technical report. Princeton Univesity Press. ACM SIGPLAN Symposium on the Ada programming Language.). Jacopini (1966. Un systeme de communication homme-machine en francais. and R. Backus. MIT Press. R. 366–371. ACM Trans. and A. I. P. R. Roussel. Gosling (1998). Wilden. Functional Programming Using Standard ML. E. Jr. J. (1978. Dijkstra. In Annals of Mathematics Studies. P. The calculi of lambda-conversion. May). Wadge (1982). Communications of the ACM 21 (8). Clarke. 283–294. The FORTRAN automatic coding system. H. ACM SIGPLAN 34 (4). Backus. L. Jagannathan (1990). e Dahl. Church. (1968. Sussman (1985). The Mythical Man-Month. A. Software: Practice & Experience 18 (9). Jr. Addison Wesley. G. Brunekreef. and S. Brooks.) (1996). Nesting in Ada programs is for the birds. C. Addison Wesley. Programming Languages and Systems 4 (2). IEEE Computer 20 (4). (Eds. and W. E. ACM Trans.References Abelson. In Proc. Boehm. Brooks.-J. 538. J. Turing machines. H. 807–820. The Java Programming Language (Second ed. A o Apt. Brinch Hansen. Universit´ d’Aix-Marseille.). Schaerf (1998. and M. Structured Programming. Wolf (1980). PROLOG: Programming for Artificial Intelligence (Second ed. In Proceedings of the Western Joint Computer Conference.. Addison Wesley. Budd. Gibson. Alma-0: an imperative language that supports declarative programming. and C. (1990). Pasero (1973). 151 . A. ACM Press (Addison Wesley). Bratko. P. K. History of Programming Languages–II.. November 1980. 613–64. P. Weiser (1988).. August). and J. Gelernter.). L. 38–45. V. Volume 6. Brooks. H. (1995).. (1941). K. Bergin. Flow diagrams. Addison Wesley. Garbage collection in an uncooperative environment. ˚ke Wikstr¨m (1987). (1987. and R. A. J. Structure and Interpretation of Computer Programs. Communications of the ACM 11 (8). February). A. Hoare (1972).. Prentice Hall. Ashcroft. No silver bullet: Essence and accidents for software engineering. Can programming be liberated from the von Neumann style? A functional style and its algebra of programs. T. MIT Press. W. O. Addison Wesley. Reprinted as ACM SIGPLAN Notices. Dijkstra. Colmerauer. F. (1995). Multiparadigm Programming in Leda. 15(11):139–145. A. Bohm. Arnold. August). (1975).

89–101 and 158–165. Hints on programming language design. (1992). E. S. Volume 20 of State of the Art Report. C. ttp://www. (1987). Copying.concordia. 291–347. C. In Programming Languages. Jones. Complexity in C++: a Smalltalk perspective. Available from ian@syacus. C. Springer-Verlag. A. D. Jensen. Grogono. In International Symposium on Automatic Digital Computing. D. ACM Computing Surveys 6 (4). Mathematische Annalen 112. Landin. Hoare.REFERENCES 152 Gill. In C. P. C: A Reference Manual (second ed. (1981. Communications of the ACM 8. Harbison. Is block structure necessary? Software: Practice and Experience 11 (8). R. A. K¨lling. pp.cs.ps. and G. P. R. K. Issues in the design of an object oriented programming language. (1974). 1–15. Bunyan (Ed. Griswold (1983). T. National Physical Laboratory.cs. (1991b. General recursive functinos of natural numbesr. Computer and Information Science 4 (2). and M. Academic Press.cs. In Information processing 74. 505–34. A. June). and aliasing. pp. Record handling. (1989). (1975. Steele. Larose. E. Int. Prentice Hall. Prentice Hall. C. Available as www. W. C. Hoare. Chalin (1994. Joyner. M. C.au. Hoare. LaLonde.concordia/~faculty/grogono/oopissue.ps. Hoare. Technical Report OOP–91–2. Jr. sharing. A compacting incrmental collector and its performce in a production quality compiler. Kay. (1965. (1974). o Ph.). and N. 289–292 of (Williams and Campbell-Kelly 1989). Kleene. 727–742. P. Recursive data structures. R. Feeley (1999). 511–578. Pergamon/Infotech. Essays in Computing Science. The Dee report. Wirth (1976). Structured Programming 12 (1). S. L. 569–574. Computer Systems Reliability. In Colloquium on Object Orientation in Databases and Software Engineering (ACFAS’94). M. R. Montreal. C++? a critique of C++. S. (1954). In History of Programming Languages–II.acus. Structured programming with GOTO statements. (1974). R. (1991a. Griswold. P. 49–56. Concordia University. The Icon Programming Language. Quebec. North Holland. Second Edition. Reprinted as pages 217–243 of (Hoare 1989). and P. March/April). . I. In ACM SIGPLAN International Symposiuum on Memory Management (ISMM’98). thesis. Pascal: User Manual and Report. The early history of Smalltalk. A correspondence between Algol 60 and Church’s lambda-notation. 853–866. May). (1999). January). R. and M. Available as www. J. January). R. A. B. (1968). ACM Press (Addison Wesley). J. (1936). Pugh (1995. Getting programmes right. Reprinted as ACM SIGPLAN Notices 343:1–9.ca/∼faculty/grogono/dee. Knuth. Journal of Object-Oriented Programming 8 (1). A. pp.html. Kowalski.concordia/~faculty/grogono/copying. P. Grogono.oz. (1996). Reprinted as pages 193–216 of (Hoare 1989). Edited by C. August). 105–32. University of Sydney. D. March 1999. 261–301. Prentice Hall.). Predicate logic as a programming language. and J. pp. A. Basser Department of Computer Science. Hanson. The Design of an Object-Oriented Environment and Language for Teaching. August). Grogono. Department of Computer Science. Reprinted as pp.

Meyer. Wexelblat (Ed.). B. Radin. B. Lindsey. K. A history of CLU. Communications of the ACM 3 (5). Swanson (1980). Zilles (1974. and R. M. and S. Milner. ACM SIGPLAN Notices 17 (12). G. (1988). (1978). ACM Press. Tofte. 471–496. (1983. pp. April). AFIPS National Computer Conference. 27–83. December). Strachey (1976). The European side of the last phase of the development of Algol. 439–478. (1996). B. ACM Press (Addison Wesley). J. Abstraction and Specification in Program Development. 299–314. Recursive functions of symbolic expressions and their computation by machine. (1986). Saraswat (Eds. Naur. E. P. Guttag (1986). pp. 2LP: Linear programming and logic programming. May). Addison Wesley. (1978). (1960. MIT Press. Academic Press. A history of Algol 68. Software Maintenance Management. Communications of the ACM 3 (4). In R.. Gill (1954). pp. H. In History of Programming Languages. History of LISP. et al. In History of Programming Languages–II. Prentice Hall International. Prentice Hall International. MIT Press. Part I. and C. . In History of Programming Languages.-J. and O. 1053–1058. December). J. Reprinted as pages 92–137 of (Wexelblat 1981). The early history and characteristics of PL/I. B. Maintenance as a function of design. V. ACM SIGPLAN Notices 9 (4). Reprinted as pp. J. 101–116. J. McKee. 50–9. 187–93. Harper (1990). (1984). Second Edition. D. R. L. Communications of the ACM 15 (12). K. In International Symposium on Automatic Digital Computing. E. Prentice Hall International. MIT Press. (1978). The development of the SIMULA languages. A. Commentary on Standard ML. Tofte (1991). Report on the algorithmic language Algol 60. On the criteria to be used in decomposing systems into modules. McCarthy. McAloon. History of Programming Languages. Perlis. A Theory of Programming Language Semantics. J. and E. Tretkoff (1995). Values and objects in programming languages. and S. 283–289 of (Williams and Campbell-Kelly 1989). 70–79. B. Wexelblat (Ed. N. Conversion routines. 173–197. Programming with abstract data types. (1996). McCarthy. 184–195. (1992). Milner. Meyer. Naur. Academic Press.). (1978). John Wiley. 1997. and M. B. B. Hentenryck and V. pp. and C. Mutch. J. In History of Programming Languages. L. Eiffel: the Language. The American side of the development of Algol. ACM Press. J. R. Martin. Liskov. L. Liskov.). B. C. pp. P. P. Milne. (1972. MIT Press. ACM Press (Addison Wesley). ACM Press. Liskov. Object-oriented Software Construction. MacLennan. and J. B. (1960. The Definition of Standard ML. Nygaard. Reprinted as pages 551–574 of (Wexelblat 1981). In R. Reprinted as pages 75–91 of (Wexelblat 1981).REFERENCES 153 Lientz. pp. April). National Physical Laboratory. History of Programming Languages. B. In Proc. Dahl (1978). Parnas. In P. In History of Programming Languages. Principles and Preactice of Constraint Programming. Data Types and Data Structures.

Turner.. S. A. Stroustrup. In History of Programming Languages. From the ACM SIGPLAN History of programming Languages Conference.). (1981). pp. Common LISP: the Language (second ed. Sammet. M. M. Dubinsky. Cambridge. D. D. (1994). MIT Artificial Intelligence Laboratory. et al. G. pp. A. On computable numbers with an application to the Entscheidungsproblem.. thesis. Soc. Prentice Hall. (1990). pp. 230–265. On the darker side of C++. T. A. and G. A machine-oriented logic based on the resolution principle. University of Kent. A. In History of Programming Languages. Springer Verlag. (1985. J. R. In S. Reprinted as pp. Academic Press. Robinson. Reprinted with corrections. Nygaard (Eds. Wheeler. In T. Turner. February 1994. 1–236. 42–44 of (Williams and Campbell-Kelly 1989). Technical report. ACM Press (Addison Wesley). Wexelblat. In Proceedings of the Second International Conference on Functional Programming Languages and Computer Architecture. 671–686. (In three parts. a a Teale. pp. Ph. Steele. Whitaker.REFERENCES 154 Ritchie. A Critical View of Inheritance and Reusability in Object-oriented Programming.). 173–228. University of Jyv¨skyl¨. KRC language manual. Planning the use of a paper library. Turner. The SASL Langauge Manual. Structured Programming 13. (1988). (1996). Towards a formal semantics. Schwartz. J. Turing.). Addison Wesley. Addison Wesley. (1996). M. K. C. ACM Press. (1965). D. D. Formal Language Description Languages for Computer Programming. London Math. B. 162–176. E. L. (1936). L. Sakkinen. 2 (42). B. and E. . In History of Programming Languages. Springer LNCS 301. Proceedings of the 1988 European Conference of Object Oriented Programming. Programming Languages: History and Fundamentals. Steel (Ed. (1992). September). B. (1951). J. Technical Report Memo 349. L. University Mathematical Laboratory. In History of Programming Languages. Turner. Scheme: an interpreter for the extended lambda calculus. (1996). Jr. ACM Press (Addison Wesley). 233–308. J. Software: Practice & Experience 9. J. (1975).. A. D. Taivalsaari. History of Programming Languages. D. 31–49. (1978). 1–3 June 1978. Sussman (1975). (1966). W. The development of the C programming language.). Gjessing and K. Acta Informatica 5. M. E. D. Miranda: a non-strict functional language with polymorphic types. The Design and Evolution of C++. Revised report on the algorithmic language Algol 68. Technical report. In Report of a Conference on High Speed Automatic Calculating-Engines. Sakkinen. Jr. G. L. (1969). (1993). J. G. Jr. Digital Press. R. Strachey. Sammett. University of St. A new implementation technique for applicative languages. Dewar. Journal of the ACM 12. Programming with sets — an introduction to SETL. (1976). Andrews. Reprinted as pages 199–241 of (Wexelblat 1981). Steele. Steele. 23–41. ACM Press (Addison Wesley). Ada — the project. et al. Springer LNCS 322. North-Holland. The darker side of C++ revisited. C++ IOStreams Handbook. Proc.. The early history of COBOL. (Ed.) (1981). Schonberg (1986). van Wijngaarden. (1979). The evolution of LISP. (1993).

REFERENCES

155

Williams, M. R. and M. Campbell-Kelly (Eds.) (1989). The Early British Computer Conferences, Volume 14 of The Charles Babbage Institute Reprint Series for the History of Computing. The MIT Press and Tomash Publishers. Wirth, N. (1971, April). Program development by stepwise refinement. Communications of the ACM 14 (4), 221–227. Wirth, N. (1976). Algorithms + Data Structures = Programs. Prentice Hall. Wirth, N. (1982). Programming in Modula-2. Springer-Verlag. Wirth, N. (1996). Recollections about the development of Pascal. In History of Programming Languages, pp. 97–110. ACM Press (Addison Wesley).

A

Abbreviations and Glossary

Actual parameter. A value passed to a function. ADT. Abstract data type. A data type described in terms of the operations that can be performed on the data. AR. Activation record. The run-time data structure corresponding to an Algol 60 block. Section 3.3. Argument. A value passed to a function; synonym of actual parameter . Bind. Associate a property with a name. Block. In Algol 60, a syntactic unit containing declarations of variables and functions, followed by code. Successors of Algol 60 define blocks in similar ways. In Smalltalk, a block is a closure consisting of code, an environment with bindings for local variables, and possibly parameters. BNF. Backus-Naur Form (originally Backus Normal Form). A notation for describing a contextfree grammar, introduced by John Backus for describing Algol 60. Section B.4.2. Call by need. The argument of a function is not evaluated until its value is needed, and then only once. See lazy evaluation. CBN. Call by name. A parameter passing method in which actual parameters are evaluated when they are needed by the called function. Section 3.3 CBV. Call by value. A parameter passing method in which the actual parameter is evaluated before the function is called. Closure. A data structure consisting of one (or more, but usually one) function and its lexical environment. Section 4.2. CM. Computational Model. Section 10.2 Compile. Translate source code (program text) into code for a physical (or possibly abstract) machine. CS. Computer Science. Delegation. An object that accepts a request and then forwards the request to another object is said to “delegate” the request. Dummy parameter. Formal parameter (FORTRAN usage). Dynamic. Used to describe something that happens during execution, as opposed to static. Dynamic Binding. A binding that occurs during execution. Dynamic Scoping. The value of a non-local variable is the most recent value that has been assigned to it during the execution of the program. Cf. Lexical Scoping. Section 4.1. Dynamic typing. Performing type checking “on the fly”, while the program is running. Eager evaluation. The arguments of a function are evaluated at the time that it is called, whether or not their values are needed. 156

A

ABBREVIATIONS AND GLOSSARY

157

Early binding. Binding a property to a name at an early stage, usually during compilation. Exception handling. A PL feature that permits a computation to be abandoned; control is transferred to a handler in a possibly non-local scope. Extent. The time during execution during which a variable is active. It is a dynamic property, as opposed to scope. Formal parameter. A variable name that appears in a function heading; the variable will receive a value when the function is called. “Parameter” without qualification usually means “formal parameter”. FP(L). Functional Programming (Language). Section 4. Garbage collection. Recycling of unused memory when performed automatically by the implementatin rather than coded by the programmer. (The term “garbage collection” is widely used but inaccurate. The “garbage” is thrown away: it is the useful data that is collected!) GC. Garbage collection. Global scope. A name that is accessible throughout the (static) program text has global scope. Heap. A region for storing variables in which no particular order of allocation or deallocation is defined. High order function. A function that either accepts a function as an argument, or returns a function as a value, or both. Higher order function. Same as high order function. Inheritance. In OOPLs, when an object (or class) obtains some of its features from another object (or class), it is said to inherit those features. Interpret. Execute a program one statement at a time, without translating it into machine code. Lambda calculus. A notation for describing functions, using the Greek letter λ, introduced by Church. Late binding. A binding (that is, attachment of a property to a name) that is performed at a later time than would normally be expected. For example, procedural languages bind functions to names during compilation, but OOPLs “late bind” function names during execution. Lazy evaluation. If the implementation uses lazy evaluation, or call by need , a function does not evaluate an argument until the value of the argument is actually required for the computation. When the argument has been evaluated, its value is stored and it is not evaluated again during the current invocation of the function. Lexical Scoping. The value of a non-local variables is determined in the static context as determined by the text of the program. Cf. Dynamic Scoping. Section 4.1. Local scope. A name that is accessible in only a part of the program text has local scope. OOP(L). Object Oriented Programming (Language). Section 6. Orthogonal. Two language features are orthogonal if they can be combined without losing any of their important properties. Section 3.6.

Regular Expression. as opposed to weak typing . Weak typing. Recursion. Literally “unknown” or “unmeasurable” (from Greek).3. √ Partial function. RE. Static. A process or object that is defined partly in terms of itself without circularity. Section 3. before the program is executed). A function that is defined over only a part of its domain. Value semantics. Using one name to denote several distinct objects.2. Reference semantics. Variable names denote objects. Semantic Model.4. A system that assigns meaning to programs. It is a static property. A syntactic construct from which many distinct objects can be generated. Section 10. A region for storing variables in which the most recently allocated unit is the first one to be deallocated (“last-in. but the usage is more or less compatible with our definition. The name suggests a stack of plates. Dynamic Binding. Stack. Semantics. Section B. Parameter. (Literally “many shapes”). Most imperative languages use value semantics. PL. Total function. Cf. A template is different from a declaration that constructs only a single object. ex is total over the domain of reals. as opposed to dynamic. Scope.) Thunk. Used in PLs for the objects passed to functions. . and most OOPLs. Programming Language. Detect and report all type errors. Detect and report some type errors. A function that is defined for all values of its arguments in its domain. A local variable in an Algol 60 program that has local scope (it can be accessed only within the procedure in which it is declared) but global extent (it is created when the procedure is first called and persists until the program termiantes). A value that is bound before the program is executed. but allow some possible type errors to remain in the program. Own variable. LIFO). An alternative term for Computational Model (CM). anyway. A function with no parameters used to implement the call by name mechanism of Algol 60 and a few other languages. (The word “template” has a special sense in C++. The region of text in which a variable is accessible. Static typing. Template. Static Binding. Used to describe something that happens during compilation (or possible during linking but. x is partial over the domain of reals because it has no real value if x < 0.A ABBREVIATIONS AND GLOSSARY 158 Overloading. Variable names denote locations. Using the same name to denote more than one entity (the entities are usually functions). Type checking performed during compilation. Strong typing. Polymorphism. first out”. Used by all functional and logic languages. as opposed to extent.

associated with each level of the hierarchy. Typical PLs have context-free (or almost context-free) grammars. B. is not context-free. There is a hierarchy of languages (the “Chomsky hierarchy”) and. axiomatic. programs can be split into lexemes by a finite state automaton and parsed by a push-down automaton. such as jumps into blocks. The basic ideas of denotational semantics are due to Landin (1965) and Strachey (1966). This makes the construction of a C++ parser unnecessarily difficult and accounts for the slow development of C++ programming tools. Their own language. They suggested that denotational semantics should be used prescriptively rather then descriptively . we briefly review the contributions that theory has made to the development of PLs and discuss the importance and relevance of these contributions. Denotational semantics is an effective language for communication between the designer and implementors of a PL but is not usually of great interest to a programmer who uses the PL. for example. and their tokens. 2 B. operational. In other words. including: algebraic. can be described by a regular language. then. Ashcroft and Wadge published an interesting paper (1982). was designed according to these principles. and others. An axiomatic semantics. They are not competitive but rather complementary. noting that PL features that were easy to implement were often hard to describe and vice versa.2 Semantics There are many kinds of semantics.B Theoretical Issues As stated in Section 1. or lexemes.1 Syntactic and Lexical Issues There is a large body of knowledge concerning formal languages. Give examples to demonstrate the difficulties that the C++ preprocessor causes in a program development environment. denotational. a denotational semantics is a mapping from program constructs to abstract mathematical entities that represent the “meaning” of the constructs. Consequently. and a particular semantic system may be more or less appropriate for a particular task. and context-free languages and push-down automata. is not very useful to an implementor but may be very useful to a programmer who wishes to prove the correctness of a program. For example. on the other hand. formal machines that can “recognize” the corresponding languages and perform other tasks. that the problem of PL design at this level is “solved”. It might appear. there are exceptions. a PL designer should start with a simple denotational semantics and then figure out how to implement the language — or perhaps leave that task to implementors altogether. Unfortunately. however. Denotational semantics developed as a descriptive technique and was extended to handle increasingly arcane features. 159 . The preprocessor is another obstacle that affects program development environments in both C and C++. In this section. The first two levels of the hierarchy are the most familiar to programmers: regular languages and finite state automata. Exercise 37. Denotational semantics is a powerful descriptive tool: it provides techniques for mapping almost any PL feature into a high order mathematical function. The grammar of C++. Lucid. theory is not a major concern of the course. There are many applications of semantics. The most complete description was completed by Milne (1976) after Strachey’s death.

It is not clear. it is interesting to look briefly at regular languages and to see how the concepts can be applied. A program that is type-correct may nevertheless fail in various ways. y of type T . Exercise 38. inheritance. If L is indeed statically typed: A compiler for L can check the type correctness of all programs. If we can do this for all legal programs in a language L. Since these operations occur in many contexts in the study of PLs. although the number of loopholes in the type system may be small. B. return y. T f (S x) { B. The theory of program correctness is more difficult and has attracted less attention. the suggestions of Ashcroft and Wadge have been followed (although not necessarily as a result of their paper). performs calculations indicated as B. Type checking does not usually detect errors such as division by zero. B.y) is undefined for y = 0. The reasoning used in the proof is based on the syntactic form of the function body. tends to be overvalued by the theoretical Computer Science community. consequently. In general: Most PLs are not statically typed in the sense defined above. that semantic techniques have had a significant impact on mainstream language design. A type-correct program may give incorrect answers. and returns a result. and repetition. 2 B. choice. given a value of type S. dynamic binding.4 Regular Languages Regular languages are based on a simple set of operations: sequence. however. } The declaration introduces a function called f that takes an argument. We can summarize it concisely as follows. . most type checking systems assume that functions are total : a function with type S → T .3 Type Theory There is a large body of knowledge on type theory. If there is a type theory associated with the language. Type theories for modern PLs. are very complex. (In general. we say that L is statically typed . will return a value of type T . and other features. Give an example of an expression or statement in Pascal or C that contains a type error that the compiler cannot detect. genericity. A description of a new PL in the academic literature is usually accompanied by a denotational semantics for the main features of the PL. classes.B THEORETICAL ISSUES 160 To some extent. Here is a function declaration.) A program can be type-correct and yet give completely incorrect answers. although it is arguably more important. we should be able to prove a theorem of the following form: If x has type S then the evaluation of f (x) yields a value of type T . Type theory leads to attractive and interesting mathematics and. Some commonly used functions are partial : for example x/y ≡ divide(x. x of type S. A program that is type-correct will not fail because of a type error when it is executed. overloading. with objects.

B THEORETICAL ISSUES 161 In the formal study of regular languages. .} in Pascal or / ∗ .. For example: UC LC LETTER DIGIT IDENTIFIER INTCONST FLOATCONST = = = = = = = . . a containing n occurrences of the symbol a. . We add two further notations that serve solely as abbreviations. For each symbol x ∈ Σ. We can use a program such as lex to construct a lexical analyzer (scanner) from a regular expression that defines the tokens of a PL. {.∗ / .}. we might have Σ = { 0. 1. If r and s are RE with languages R and S respectively. If r and s are RE with languages R and S respectively. . For example.∗ / in C.. 2.) Each form of RE is defined by the set of strings in Σ∗ that it generates. which is a finite set of symbols. What consequences does this have for the lexical analyzer (scanner)? 2 A+ B + ···+Z a+ b+ ···+z U C + LC 0+1+···+9 LETTER ( LETTER + DIGIT )∗ DIGIT+ DIGIT+ . 4. . . For example. . xn | n ≥ 0 ∧ xi ∈ R } ). then (r + s) is RE and denotes the set R ∪ S. . The expression a+ is an abbreviation for the RE aa∗ that denotes the set { a. If r is RE with language R. The set of all finite strings. .4. {. (the string containing no symbols) is RE and denotes the set { }./ ∗ . constructed from the symbols of Σ is written Σ∗ . Some compilers allow “nested comments”. This set is called a “regular language”. g 1. 6. . . including the empty string. Exercise 39.1 Tokens We can use regular languages to describe the tokens (or lexemes) of most PLs. aaa.} .. Σ. . we begin with an alphabet. x is RE and denotes the set { x }.   is RE and denotes the empty set. The expression an represents the string aa . DIGIT∗ . . B. but in this section it stands for “regular expression”. aa. 3. . 2. 1 }. We then introduce a variety of regular expressions. (The abbreviation RE is also used to stand for “recursively enumerable”. then (rs) is RE and denotes the set RS (where RS = { rs | r ∈ R ∧ s ∈ S } ). 5. . then r∗ is RE and denotes the set R∗ (where R∗ = { x1 x2 . . . which we will refer to as RE here. . .

Why might Figures 17 and 18 be of interest to a PL designer? 2 B.4 Discussion These analogies suggest that the mechanisms for constructing REs — concatenation. they can be constructed from EBNF grammars. alternation. Figure 18 shows a correspondence between REs and data structures.. the relation between the standard control structures and the standard data structures can be helpful in programming. The following example illustrates both of these limitations. statement while expression do statement if condition then statement else statement Figure 17: REs and Control Structures 162 B. and repetition — are somehow fundamental. The syntax for the assignment. a connector (usually ::= or →). A grammar rule. Just as parsers can be constructed from BNF grammars. In Jackson . SEQUENCE → EMPTY SEQUENCE → STATEMENT SEQUENCE BNF has been extended in a variety of ways. and a sequence of terminal and non-terminal symbols. or production.4. BNF provides no mechanism for repetition. most of the extended forms can be described simply: the sequence of symbols on the right of the connector is replaced by a regular expression. Similarly. we must use recursion instead.B THEORETICAL ISSUES RE x rs r∗ r+s Control Structure statement statement.4. Also. In grammars. might be defined: ASSIGNMENT → VARIABLE “ := ” EXPRESSION Basic BNF provides no mechanism for choice: we must provide one rule for each possibility. 2 B. Discuss the difficulty of adding attributes to an EBNF grammar. In particular.4. Exercise 41. Exercise 40. the symbol | rather than + is used to denote choice. STATEMENT = ASSIGNMENT | CONDITIONAL | · · · SEQUENCE = ( STATEMENT )∗ Extended BNF (EBNF) provides a more concise way of describing grammars than BNF.3 Control Structures and Data Structures Figure 17 shows a correspondence between REs and the control structures of Algol-like languages. With a few exceptions.2 Context Free Grammars Context free grammars for PLs are usually written in BNF (Backus-Naur Form). for example. The extension enables us to express choice and repetition within a single rule. consists of a non-terminal symbol (the symbol being defined).

When we move to the next level of the Chomsky hierarchy. } int a[n]. struct { int n. float y. int a[] union { int n. float y. the data structures appropriate for the application are selected first and then the program is constructed with the corresponding control structures. . Context Free Languages (CFLs).B THEORETICAL ISSUES RE x rs rn r∗ r+s Data Structure int n. for example. we obtain the benefits of recursion. page 163). The corresponding control structure is the procedure and the corresponding data structures are recursive structures such as lists and trees (Wirth 1976. } 163 Figure 18: REs and Data Structures Structured Design (JSD). The limitations of REs are also interesting.

Functions of more than one variable can be reduced to functions of one variable. the definition of fac becomes: fac = S(C(B if (== 0))1)(S × (B fac (C − 1))) 164 . gives succ = [x] (add 1 x) = S ([x] (add 1)) ([x] x) = S (S (K add) (K 1)) I These expressions quickly become quite long. We start with three observations. If e is an expression possibly containing x. such as 1 and +. the size of the SKI expression increases exponentially with the size of the original expression. The technique described below was introduced into mathematical logic by M. Schoenfinkel showed how to do this using the following abstraction algorithm. Using these. we obtain fac = S(S(S(K I if)(S(S(K ==)(K0))I))(K 1))(S(S(K ×)I)(S(K fac)(S(S(K −)I)( 1)))) In general. In the second rule. The rules for abstraction follow. It is an expression that does not contain x. Turner introduced new functions B and C and optimizations to overcome the size problem. Schoenfinkel in 1924 and was rediscovered as an implementation method by D. obtaining a function of one variable. We would like to write the definition of succ in the form succ = ??? so that the distinction between the name being defined and the definition is clear. because otherwise the first rule would apply. are implemented using combinators and graph reduction. and constants. it is a constant as far as this definition is concerned. which we then apply to y. If we abstract n from fac n = if (n == 0) 1 (n × fac(n − 1)) (where we treat if as a function of three arguments).A. the abstraction of e with respect to x is written [x] e. We understand add x y as (add x) y: we evaluate (add x). such as x in this examples. or simply variables. In a definition such as succ x = add 1 x we can distinguish between bound variables. including Haskell. Operators are really just functions. Turner in 1979. Although add might be defined elsewhere. we assume x = y.C Haskell Implementation Several modern functional programming languages. For example. we can write x + y as add x y. Thus we only have to consider prefix functions. defined above. [x] x = I [x] y = K y [x] (f g) = S ([x] f ) ([x] g) Apply these rules to the function succ.

we have: succ 4 = S (S (K add) (K 1)) I 4 = (S (K add) (K 1) 4) (I 4) = ((K + 4) (K 1 4)) 4 = +14 = 5 For practical implementation. and transform the graph. have led to steady improvements. these expressions are represented as graphs: Figure 19 shows the graph corresponding to succ.C HASKELL IMPLEMENTATION 165 S   ©   d ‚ d S   ©   d ‚ d I K c K c + 1 Figure 19: The graph for succ To make use of these expressions. When Turner presented one his functional programming languages at a conference in 1982. and a modern implementation of a functional language has performance comparable to that of a compiled procedural language. The basic technique of graph reduction is quite slow. . we apply an argument and then make use of the following reduction rules: Ix = x Kxy = x S f g x = (f x) (g x) B f g x = f (g x) Cf gx = f xg Using the simpler example. always using the rule for whatever function is at the root of the tree. succ. we add a third subtree to the top node. To evaluate an expression such as succ 4. however. giving S three arguments. Twenty years of research. he said that it was about 500 times slower than conventional languages such as C and FORTRAN.

Sign up to vote on this title
UsefulNot useful