APIC STLJDIFS IN DATA PROCESSING NO.

Q4

PROLOG FOR
PROGRAMMERS
Feliks Kluiniak
Stanislaw Szpakowicz
With a contribution by
Janusz S. Bieri
.-. .- _ _ . - -_ 1-
| 1 '
-,l I .l __ - 1
'I-.- - 1- -I | E Il-

___',_ I _.-

X .|_i._p.| iiitif ',_

l -1 _ ."l-'.'
._ , _.-
, .
- I -
! ' . ,- _-
;__ .
"' fr -"Y
_.= |" _ *- H. _ '.a "
'11 1.- .=- 1-,_.:-=-{ 11 3-. '
H
R

-. -|- -1 _
| 1

'\- I1

_'._"._ _ --\| '\._

.- ,___ _'_
X l-FIN _ _ i| ., y " "' "J

_ 1 _ ____ F | ll

. I
|' \ \

*11' +i '* 3"--| H - I-' H

* .___.. : I I [~;__ S

"-/ \ K :,-J)
xa .-1"
/
\\ _J______. ____,.,--"'
PROLOG FOR PROGRAMMERS
This rs volume 24 in A.P.|.C. Studies in Data Processing
General Editors: Fraser Duncan and M. J. R. Shave
A complete list of ri es in this series appears at the end of thrs volume
PROLOG FOR PROGRAMMERS

Feliks Kluiniak
Stanislaw Szpakowicz

Institute of Informatics
Warsaw University
Warsaw, Poland

With a contribution by Janusz S. Bieri

1985

ACADEMIC PRESS
(Harcourt Brace Jovanovich, Publishers)

London Orlando San Diego New York Austin

Toronto Boston Sydney Tokyo
Coryntotrr I985, ay Acantamc Pnsss Inc. (Lmtoom Lrn
ALL mot-rrs nassnven.
no PART or 11-us |>ueuc.a'r|ou my as neraooucsn on
TRANSMl'l"l'ED IN my FORM on ay my nuns. 1-11-zcraomc
on uscumttcn. tnctunmo Puottooory. neconnmo. on
my |u|=omxnou srotutota AND nertutavtu. sysrm. wmtour
PERMISSION m wnmno mom THE PUBLISHER.
ILEPRINTED WIT]-I cotuuscrtous I987

ACADEMIC PRESS INC. (LONDON) LTD.

24-2a Oval Rood
LONDON mvt vnx

United States Edition published by

ACADEMIC PRESS. INC.
Orlando. Florida 32887

British Library Cataloguing in Publication Data

Kluiniak. I-'eIiks

Prolog for programmers.

I. Programming languages (Electronic computers)
2. Electronic digital computers -Programming
I. Title II. Szpakowicz,Stanislaw
000.64'24 QA76.7

Library of Congress Cataloguing in Publication Data

Kluiniak, Feliks.
Prolog for programmers.

Includes index.
I. Prolog (Computer program language) I. Szpakowicz,
Stanislaw. II. Bieit, Janusz St. III. Title.
QA76.73.P76K58 I985 00I.64'24 84-I4520
ISBN 0I2-4I6520-6 (hardback)
0I2-4I6$2I-4 (paperback)
PRINTED IN GREAT BRITAIN
AT THE ALDEN PRESS. OXFORD

BSIGITIS 98765432!
 Agacie Sarze, i jej mamieF. K.

Pakowi, Mikofajowi, Blazejowi i GumceSzp

Zosi i BasiJ. S. B.
777 7 1 _1_i _

CONTENTS

PREFACE xi

INTRODUCTION TO PROLOG I
Data Structures I
Operations II
I-I-I") !"!"1"2 Control 24

2. PROLOG AND LOGIC 4|

2.1. Introduction 4|
2.2. Formulae and Their Interpretations 4|
2.3. Formal Reasoning 44
2.4. Resolution and Horn Clauses 45
2.5. Strategy 51

3. Ml-ITAMORPHOSIS GRAMMARS:
A POWERFUL EXTENSION 59
3.l. Prolog Representation of the Parsing Problem 59
3.2. The Simplest Form of Grammar Rules 67
3.3. Parameters of Non-terminal Symbols 69
3.4. Extensions 73
3.5. Programming Hints BI

4. SIMPLE PROGRAMMING TECHNIQUES 87

4. I. Introduction 87
4.2. Examples of Data Structures 88
4.3. Some Programming Hints I21
4.4. Examples of Program Design I30

vll
VIII Contents

5. SUMMARY OI-' SYNTAX ANI) BUILT-IN PROCI-IIJURI-IS I43

5. I. Prolog Syntax I43
5.2. Built-in Procedures: General Information I47
5.3. Convenience I49
5.4. Arithmetic I50
5.5. Comparing Integers and Names ISI
5.6. Testing Term Equality I52
5.7. Input/Output I53
5.8. Testing Characters I58
5.9. Testing Types I58
5.I0. Accessing the Structure of Terms I59
5.II. Accessing Procedures I60
5.12. Control I63
5.13. Debugging I64
5.I4. Grammar Processing I65
5.I5. Miscellaneous I65

6. PRINCIPLES OI-' PROl.OG IMPLEMI-INTATION I67

6. I. Introduction I67
6.2. Representation of Terms I68
6.3. Control I78
6.4. Tail Recursion Optimisation I79
6.5. Bibliographic Notes I83

7. TOY: AN I-IXI-IRCISI-I IN IMPLEMENTATION I85

7. I. Introduction I85
7.2. General Information I86
7.3. The Toy-Prolog Interpreter I87
7.4. Interpretation of Prolog-I0 in Toy-Prolog 201

8. TWO CASE STUDIES 2I5

8.l. Planning 215
8.2. Prolog and Relational Data Bases 226

9. PROLOG DIALECTS 249

9.I. Prolog I 249
9.2. Prolog ll 250
9.3. Micro-Prolog and MPROLOG 253

APPENDICES 255
A. I. Kernel File 256
A.2. Bootstrapper" 258
Contents

A.3. User Interface and Utilities

A.4. Three Useful Programs

REFERENCES

INDEX
 g

PREFACE

Prolog is a non-conventional programming language for a wide spectrum

of applications, including language processing, data base modelling and im-
plementation, symbolic computing, expert systems, computer-aided design,
simulation, software prototyping and planning. A version of Prolog has been
chosen as a systems programming language for so-called fth-generation com-
puters; experiments with systems programming and concurrent programming
are in progress.
Prolog, devised by Alain Colmerauer, is a logic programming language.
Logic programming is a new discipline which lends a unifying view to many
domains of computer science. Prolog can be classified as a descriptive pro-
gramming language, as opposed to prescriptive (or imperative) languages such
as Pascal, C and Ada. In-principle, the programmer is only supposed to
specify what is to be done by his or her program, without bothering with
how this should be achieved. Robert A. Kowalski has coined the equation
Algorithm = Logic + Control,
which emphasizes the distinction between the what (logic) and the how (con-
trol). The programmer need not always specify the control component. In
practice, however, Prolog can be treated as a procedural language.
Prolog is not standardized, and it comes in many different avours.
The most widespread dialect of Prolog is ProIog- I0, originally implemented
by David H. D. Warren for DEC-I0 computers. We describe a variant of
this dialect, based on an interpreter written in Pascal especially for this book.
The main part of the book is Chapters I-5. Chapters I and 3
are an introduction to Prolog, intended for those who use prescriptive
languages in their everyday practice. Both intuitions and the presentation
are practically biased, but we assume the reader has a certain amount of
programming experience and sophistication. Chapter 2 explains Prolog in
xl
xii Preface

terms of logic. It requires no deep knowledge of mathematics and is intended

as a counterpoint to Chapter I, but can be skipped on a rst reading. Chapter
4 contains some useful programming techniques and hints. Chapter 5 is a
reference manual for the version of Prolog described in this book. In addi-
tion, Chapter 8 is a discussion of two rather illuminating applications.
For those who wish to gain more insight into the language and its inner
workings, Chapter 6 introduces basic principles of Prolog implementation.
An implementation of the dialect described in this book is presented in
Chapter 7. We used this implementation to test our examples, including the
case studies of Chapter 8.
Chapter 9, written by Janusz S. Bien (who also did most of the bibli-
ography), brie y outlines the most characteristic features of several other
Prolog dialects.
The diskette enclosed with this book contains source text of all the
programs listed in Chapter 8 and in the appendices, and of the Toy-Prolog
interpreter discussed in Chapter 7. The interpreter is written in TURBO
Pascal. You can use the diskette on any IBM PC compatible computer
running MS-DOS 2.l0 or 3.l0.
The material in this book, supplemented by some additional reading and
a programming assignment, can be used for a two-semester course at the
level of third-year computer science majors. Re-implementation of or exten-
sions to the interpreter of Chapter 7 might make interesting assignments for
a translator-writing course.
While working on this book, we used the computing facilities of the
Institute of Infortnatics, Warsaw University. We would like to thank Pawel
Gburzynski and Krzysztof Kimbler, who helped us switch almost painlessly
to a di 'erent machine when the one we originally used broke down for a
protracted period of time. We thank David I-I. D. Warren for permitting us to
include the listings of WARPLAN. We are also grateful to all those who have
provided us with logic programming literature for the past I0 years.
1 AN INTRODUCTION TO PROLOG

Prolog is an unconventional language. In particular, its data structures are

quite different from those found in other programming languages. As it is
dif cult to talk about a computation without understanding the sort of
data that can be processed, we shall discuss data structures at some
length before coming to the question of how to do anything with them.
I-lave patience.

1.1. DATA STRUCTURES

1.1.1. Constants

Constants are the primitive building blocks of data structures. Con-

stants have no structure, so they are often called atoms. They repre-
sent only themselvesthey can be thought of as identical with their
names.
In Basic or Fortran, I951 is a constant. The integer variable J is not,
because it represents both a memory cell andin certain contextsa
value. The value is something quite different from the variable itself.
One is accustomed to treating I95] as a number greater than I948, but
this is because in programming languages constants usually belong to
certain types. The usual properties of integer constants (their ordering,
ability to be used in arithmetic operations, etc.) are taken for granted by
virtue of their belonging to the type Integer, just as in Pascal blue is a
successor of red when one writes
colour = ( red, blue, green )
I
2 I An Introduction to Prolog

Such type de nitions impose a certain structure on the otherwise undiffer-

entiated universe of individual symbolic constants, each of which has
only one attribute: its name.
The interpretation of a constant rests solely with the programmer.
I95l can be the price of a computer, the weight of a truck, the time of day
or a year of birth. One can always multiply it by 4, but this seldom makes
sense when it represents a ears registration number. The constant blue is
less burdened with inadequate interpretations, but one might wish not to
have the colours ordered. Constants are the primitives, and collecting
them into types should only be done when necessary.
In Prolog, as in other symbolic languages (such as Lisp) there is no
need to declare constants or group them into types. One can use them
freely, simply by writing down their names.
A legal constant name is one of the following:
A sequence of digits, possibly pre xed by a minus sign; by convention,
such constants are called Integers (e.g. 0, -7, I951);
--An identi er, which may contain letters, digits and underscores but
must begin with a lower case letter (e.g. q. aName. number_9);
A symbol which is a nonempty sequence of any of the following char-
acters:
+ 1- / < = > . ? $ & (rt) # \ -
Any one of the characters
, or ; or !
The symbol [] (pronounced nil");
A quoted name, written according to the Pascal convention for strings:
an arbitrary sequence of characters enclosed in apostrophes, an
apostrophe being represented by two consecutive apostrophes (e.g.
Can"t do this. consists of I4 characters).
All of these constants are purely symbolic and have no inherent inter-
pretation. However, some primitive operations in Prolog do treat them in
a special way:
Arithmetic operations interpret integers as representations of integer
values (they can also create new integers):
Comparison operations interpret integers as integer values. and all
other constants as representations of the sequences of characters form-
ing their names (these are lexicographically ordered by the underlying
collating sequence);
Input/output operations interpret all symbols as sequences of charac-
ters forming their names.
 I. I. Data Structures 3

Each occun'ence of a constant's description (name) is treated as re-

ferring to the same constant, but of course we are free to interpret each
separately.

1.1.2. Compound Objects

An important aspect of the expressive power of a programming lan-

guage is its ability to directly describe various data structures. Of the
popular and widely used languages, Pascal is the most powerful in this
respect, but it has several shortcomings. (This is not a criticism of Pascal:
our point of view does not take into account important design objectives
such as a safe type mechanism.)
Firstly, type de nitions in Pascal are overspeci ed. It is impossible to
program general algorithms which process stacks or trees regardless of
the type of their elements. (Records with variants are only a rough ap-
proximation to generic data types found in some more recent program-
ming languages.)
Secondly, those data structures which change their form dynamically
can only be built with pointers. One must therefore deal with the struc-
tures at a very low level: the level of representation rather than the con-
ceptual level at which many other things are done in Pascal. Programs
using pointers are en'or-prone and hard to understand, because opera-
tions on such data structures are encoded rather than directly expressed.
The third shortcoming has a similar effect. Ironically, Pascal types
are also underspeci ed. in that there is no way to directly express certain
quite natural constraints on the arguments of operations. One cannot say
that the function POP can only be applied to a non-empty stack; one can
only write a piece of code (hopefully con'ect) which checks the argument.
It is interesting that these shortcomings are not shared by Prolog data
types (or rather by their counterparts, since type is not really a Prolog
concept). And yet Prolog data structures are very simple. Let us look at
the details.

Fu NCTORS
To describe a compound object, it is not enough to list its compo-
nents. The ordered pair (I9, 24) can be an object of the type

rectangle = record
height, width : integer
end
4 I An Introduction to Prolog

as well as an object of the type

timcofday = record
hour, minute : integer
end
The complete description of a compound object must include a de ni-
tion of its structure. Structure is de ned principally by describing the way
in which the object and its components are intenelated. Describing these
intenelationships often consists in simply giving a type name to an aggre-
gate of components (as in the example above). It is the programmer's
responsibility to interpret this name in terms of real-world relations be-
tween entities being modelled by the program.
In conventional programming languages, the structure of a compound
object is usually described in a declaration associating the object with a
type de nition. The type de nition lists the name of the objectcompo-
nent relationship (type name) and, possibly, additional information about
the structure (types) of the components. The object is described by its
name (or the name of a pointer). The name's de nition is textually remote
from its occunences.
A different approach is taken in Prolog. Here, the type name is an
integral part of all the occurrences of the object's description. The nota-
tion is very simple: a description of a compound object is the type name
followed by a parenthesized sequence of descriptions of its components,
separated by commas. We write either
rectangle( I9, 24 )
or
timeofday( I9, 24 ).
The notation is similar to that used for writing functions in mathematics.
Terminology re ects this similarity. The type-name is called a functor, and the
components are called arguments. There is more to it than super cial
similarity of two simple syntactic conventions. One can certainly regard a type
such as rectangle as a function mapping components into compound objects.
From this point of view it is not surprising that we can have functors with no
arguments: these are simply constants. Sometimes it is also useful to have one-
argument functors. For example, the integer 2 can be represented by the object
successor(successor(zero)) (the fact that 2 > I > 0 is evident from its struc-
ture).
From the discussion above, it should be obvious that the important
attributes of a functor are both its name and its arity (i.e. the number of
arguments it takes). In Prolog, we can use both
timeofday( I7, I3 )
 I . I . Data Structures 5

and
timeofday( I7I3 )
in the same program. Even if the intended interpretation is the same,
these are two different objects: one has two components, and the other
has one. There are also two different functors, both named timcofday.
Whenever we speak of a functor in a context which gives no indication of its
arity, the arity must be given explicitly. The usual notation is to write it after a
slash: timcofday/2 or timcofday/I.
The lexical rules for forming functor names are the same for all ari-
ties. but integers can only be constants. Thus
I23( a, b )
is incorrect, but
l23( a, b )
is perfectly all right. Also, [] is only a constant.

Oruscr Dsscatrrtous
Descriptions of constants and compound objects are referred to as
tenns. Usually, the objects themselves are also called terms: this causes
no confusion in practice, but in this chapter we shall try to distinguish
between the two meanings.
The arguments of a temt are arbitrary terms. For example, one can
write a term describing a record:
customer( name( john, smith ),
address( strect( north_ave ), number( I73 ) ) ).
Of the various functors in this example, the outermost, customer/2, can
be said to de ne the general structure of the temi. It is called the main
functor, or principal functor. Similarly, name/2 is the main functor of the
rst argument.
Here is another example of a common data structure. A list can be
de ned as either the empty list, or a list constructed of any object (a head)
and a list (a tail). A list of the rst three letters in the alphabet could then
be descn'bed by the term
cons( a, cons( b, cons( c, emptylist ) ) ).
The following term would be a description of the two-element list con-
structed of the above list and a list containing the integer zero:
cons( cons( a,cons( b,cons( c,emptylist ) ) ),cons( 0,emptylist ) )
6 I An Introduction to Prolog

Even this small example demonstrates that nested parentheses can be

dif cult to read. Prolog therefore provides syntactic sugar to hide this
standard or canonic form of terms. Instead of writing
successor( successor( zero ) )
one can choose to use successor as a pre x functor and write
successor successor zero.
Alternatively, successor can be made a post x functor:
zero successor successor.
Functors with two arguments can be declared as in x functors, e.g.
&( a, b )
can be written as
a & b.
The term
a&b&c
would be ambiguous, so an in x functor is either left-associative or right-
associatlve (or non-associative. in which case the term is incorrect). If & is
right associative, the terms standard form is
&(a.&(b.));
if & is left-associative, then the term stands for
&(&(a,b),c).
We can use parentheses to stress or override associativity. If & is right-
associative, then
a&b&c
is equivalent to
a&(b&c)
but not to
( a & b ) & c,
which stands for
&( &( a, b ). C )
regardless of associativity.
 I. I . Data Structures 7

To make parentheses even less frequent, pre x, post x and in x

functors are given priorities. Functors with lower priority take prece-
dence over those with a higher priority (a Prolog-I0 convention, different
from that used in other programming languages and mathematics). For
example, if the priority of 1- is lower than that of +, then
3 1- 4 + 5 and 5 + 3 1' 4
denote
+(*(3.4).5) and +(5."'(3.4))
We can use parentheses to stress or override priorities, by writing
(31-4)+5 or 31-(4+5).
Pre x, post x and in x functors are usually referred to by the generic
name operators. Remember that these are not operators in any conven-
tional sense: they are only a syntactic convenience.
Operator names may not be quoted. If an operator is to be written in
standard form or with a different number of arguments, it must be quoted.
If + is an in x functor,
a+ b, '+(a,b) and +(a,b,c)
are correct terms, but
+ and +( a, b )
are not.
It is also possible to declare mixed operators, i.e. functors such as the
minus sign, which is both pre x and in x in ordinary arithmetic. Details
about declaring pre x, post x and in x functors can be found in Sections
5.I and 5.7.3.
For the time being, we shall only use in x functors to write terms
representing lists. However, instead of
a cons b cons c cons emptylist
we shall use a more concise notation, modelled after Lisp. The empty list
will be denoted by the constant l] (pronounced nil), and the construct-
ing functor by the right-associative in x functor ./2. Our two lists are
then written as
a.b.c.[]
and
( a.b.c.[] ).0.[]
3 I An Introduction to Prolog

The convention is arbitrary, in that any constant and two-argument func-

tor would do in place of [] and the dot. It is more convenient than others,
because these are the symbols expected by several built-in procedures.
(You can write such terms after feeding Prolog with
:- op( 800, xfy, ).
However, a minor technical dif culty makes it impossible to use the
period as a functor when it is immediately followed by a white space
character, such as blank, tab or new line. This is a nuisance, and Prolog
provides special syntactic sugar for lists: it is somewhat confusing, so we
will put it off until Chapter 4.)

Sraruos
Characters are constants whose names consist of single characters.
One can use quoted names for characters which are not correct identi ers
(e.g. , (', 3; and x is equivalent to x).
Strings are lists of characters. One can also write them in double
quotes. For example
"string" and
stand for
s.t.r.i.n.g.[] and ".[]
(Actually, the convention adopted in this book is different from that
of Prolog-I0. There, a string denotes a list of ASCII codes and not a list of
characters, so "string" stands for
ll5. I I6. I I4. I05. I l0.l03.[].
Similarly, in Prolog- I0 operations for reading and writing characters deal
directly with ASCII codes. We refuse to accept these conventions.)

1.1.3. Variables

Objects discussed so far are all, in a sense, constant. Their structure

is xed, we know everything about them and cannot learn anything new.
A programming language in which one could specify only such fully de-
ned objects would hardly be interesting. One must be able to use objects
whose complete form is de ned dynamically during a computation.
In Prolog, the simplest such as-yet-unknown objects are called vari-
ables (do not confuse them even for a moment with the variables of
conventional programming languagesl). The term denoting a variable is
 I . I . Data Structures 9

called a variable name (this is also usually called a variable: as with terms
and objects, we shall try to maintain the distinction throughout chapter I).
A variable name is written as an identi er starting with an upper case
letter or an underscore (e.g. Q, Number_9, _nnn).
A variable is an object whose structure is totally unknown. As a
computation progresses, the variable may become instantiated, i.e. a
more precise description of the object may be determined. The term
embodying this description is called the variables instantiation. An in-
stantiated variable is identical with the object described by its instantia-
tion, so it ceases to be a variable, although the object can still be referred
to through the variables name. (In general, a variable may be instantiated
also to another variabIewe shall soon see the meaning of this.)
There is also an altemative terminology. One says that a free (or
unbound) variable becomes bound to another term and is henceforth in-
distinguishable from that term (which is called its binding). The variable
becomes ground if its binding contains no variables. This terminology
brings to mind the process of binding formal parameters to actual parame-
ters. If the formal parameters were not allowed to change their value (as
in pure Lisp, say), the similarity would be very close indeed. except that a
binding need not be ground.
intuitively, Prolog variables are somewhat like the variables used in
mathematics. When we say that
fl!) = 6 + 3x
is a function of one variable, we mean that the equation allows us to
determine the functions value for any (one) given argument. The variable
denotes a single (albeit arbitrary) substitution and is not in itself an object
to which values can be assigned.
You can also regard a Prolog variable as an invisible pointer. When
not free, the pointer is automatically dereferenced in all contexts, so it is
impossible to distinguish it from the referenced object: in particular, it is
impossible to exchange the object for something else.

1.1.4. Tenns

If one thinks of a type as a set of objects, then a term is also a

de nition of a type. The term VariabIe3 describes the set of all objects,
because a variable can be instantiated to anything. On the other hand, one
can have a very precise type speci cation. For example, the term a.b.c.[]
describes a set containing only one object: the list of length 3, whose rst
element is a, whose second element is b and whose third element is c.
There is a wide range of choices between these extremes.
I0 I An Introduction to Prolog

We describe objects by de ning those of their properties which we

nd interesting in a given context. We do so by using variable names to
denote objects (in particular: components of other objects) whose exact
form is either unknown or unimportant. Our descriptions thus denote sets
of objects satisfying the explicitly formulated properties.
A few examples should make it clear:

I. painting( Painter, 'Saskia )

all Saskias of an unknown artist
2. painting( rembrandt, Picture )
-all pictures by Rembrandt
3. painting( rembrandt, picture(Title,l646) )
alI pictures painted by Rembrandt in I646
4. Head.Tail
alI non-empty lists
5. One.Two.Three.[]
alI lists of three elements
6. One.Two.l3.Tail
all lists containing at least three elements, such that the third ele-
ment is the number I3.

Actually, our comments in examples 4 and 6 are somewhat imprecise,

as they re ect an intended interpretation. Since a variable name denotes
an arbitrary object, the type Head.Tail contains more than true lists: the
object one.two also answers this description. Similarly, the term in exam-
ple I describes objects such as painting(59, Saskia). Term notation does
not allow us to express directly our wish to consider only paintings whose
rst arguments are the names of painters. This is in keeping with the
principle that the type of a compound object is de ned primarily by the
intenelationship between the object and its components, rather than by
the types of the components. The restriction is not necessarily a bad
thing: we shall see that a procedure popping an element off a stack is most
naturally written so that it can handle all stacks, whatever the types of
their elements. If one considers it important to restrict the types of com-
ponents, one can do it easily enough (we shall see how), but only con-
sciously and only when needed.
As a computation progresses, variables in various terms may become
instantiated. As a result, more is known about the objects described by
these terms. We extend our terminology so that we can talk about instan-
tiating terms and terms which are instantiations of other tenns. For exam-
ple, f(X).Tail is an instantiation of Head.Tail; and it may, in due course,
be further instantiated to a yet more precise description. As we shall see,
 l.2. Operations II

such a multi-step approximation to a desired description is very charac-

teristic of Prolog.
We have proposed to regard a term as a type de nition, i.e. a descrip-
tion of a class of objects, or altematively as a description of a single, as
yet unde ned object. These are two sides of the same coin. A single
object whose fomr is known only in general outline can be thought of as a
representative of the class of all objects having that form. A term denotes
a set by virtue of denoting any one of its possible instantiations.
A comment on the role of variable names. They are used as handles
on the objects they denote. Through a name we can, at any moment, look
at what we have actually leamed about the shape of the object. For
example, no matter what the instantiation of Head.Tail, the variable name
Head denotes the rst element of this list. The term X.X.Tail is also quite
legal in Prolog and denotes a list whose rst and second elements are the
same object.
Notice that we said the same object, not identical objects." It is
important to note that different compound objects can share components.
In general, Prolog temrs describe data structures which can be repre-
sented as directed acyclic graphs (DAGs). If we use an arrow to denote
the relation of being built of" (X > Y means that Y is a component of
X), then Fig. I.l illustrates the object denoted by

one( two( A.B ), three( A.B. C ), B ).

Sometimes one is not interested in certain objects and needs no name
to refer to them. Such terms can be denoted by anonymous variables, each
of which is written as an underscore. For example
a._._.b.[]
describes a list of length four whose rst and last elements are a and b.
The second and third elements can be any two different objects, or the
same object: we don't care.

1.2. OPERATIONS

The majority of operations in a Prolog program are calls to procedures

de ned by the user. Standard operationsaddition, comparisons, input!
output etc.are used relatively infrequently. For uniformity, every oper-
ation is written as if it were a procedure call, and the principal property of
all standard operations is only that they need not (and must not) be de-
I2 I An Introduction to Prolog

it

@")*:E" iE
mB
E\ FIG. l.I
kg;
Recurring components of an object.

ned by the user. Standard operations are accordingly called built-in pro-
cedures (or system procedures).
Procedure calls are written according to the usual practice. The pro-
cedure name is followed by an optional list of termsactual parameters,
enclosed in parentheses and separated by commas, e.g.
show( painting(rembrandt,X), etching(rembrandt,X) ).
The procedure name" is often called a predicate symbol (sometimes
shortened to predicate). Like functors, a predicate symbol has two attri-
butes: a name and an arity. Two distinct procedures can share the same
name, provided one has a different number of parameters than the other.
Procedure calls (and, as we shall see, procedure de nitions) have the
same syntax as temis. This notational uniformity is useful when programs
are dynamically modi ed (as is normally the case during an interactive
session), but it may be confusing for the uninitiated. We shall try to help
by reserving the word "argument" for components of terms: procedures
will be said to have parameters.
 l.2. Operations I3

Some versions of Prolog-including those described in this book-

carry this uniformity to the point of allowing the user to use pre x, post x
and in x notation for predicate symbols (such symbols are also called
operators"). This is achieved exactly as for functors, but a number of
frequently used symbols are usually predeclared to give the language a
more conventional avour. A case in point is the built-in procedure is,
whose name is written in in x notation. It expects its second parameter to
be an "expression": a term representing the abstract syntax tree of an
integer arithmetic expression; the tree is evaluated and the result is re-
tumed through the rst parameter. The two-argument functors + , -, 1-, I
and mod are predeclared as in x functors with conventional priorities and
associativity. so one can write
VisX7Y1=Zmod(2+X)
to instantiate V to an integer (provided the instantiations of X, Y and Z
are integers or expressions).
Sequences of procedure calls use commas for separators, for ex-
ample:
buy( picture(rembrandt,Title), Price ),
NewPrice is Price 1- I35/ I00,
seII( picture(rembrandt,Title), NewPrice ),
drink( beer )
(We shall enlarge on this in Sections 1.2.2 and I.3.I.)
We shall need two built-in procedures for our examples:
nl/0 terminates an output line;
-- write/I outputs a temr (variables are written as XI, X2, etc.); for
example, if A and B are uninstantiated, then
write( f('an id',g(A,B),7,A) ), nl
writes
f( an id, g( XI, X2 ), 7, XI ).
More precise descriptions of system procedures can be found in
Chapter 5. We shall now see how to de ne user procedures.

1.2.1. The Simplest Fonn of a Procedure

Try to think of a procedure which computes the head and the tail of a
list: we shall call it carcdr. What should its speci cation be like?
Let the list be the rst parameter and let the second and third parame-
ters retum its head and its tail. Head and tail are de ned only for non-
I4 I An Introduction to Prolog

empty lists, so the rst parameters type is described by the term

Head.Tail (Ql .Q2 would do just as well, but it is better to use meaningful
names). This type speci cation is most naturally written in the procedure
heading, thus
carcdr( Head.Tail, )
If a list is denoted by Head.Tail, then Head denotes its head and Tail
denotes its tail. We can therefore write
carcdr( Head.Tail, Head, Tail ).
The fullstop temrinates the speci cation. It is rather concise, but it con-
tains all the necessary infomration; the procedure is called carcdr and has
three parameters; the rst parameter must be a non-empty list, the second
parameter is to become the head, and the third is to become the tail of this
list.
It turns out that what we have written is also the complete de nition
of this procedure in Prolog. The call
carcdr( l.2.3.[], H, T)
instantiates H to I and T to 2.3.[]. (Recall that l.2.3.[] is really
-(I.-(Z.-(3.lI))) .)
Actually, our de nition is somewhat more general, becauseas we
have already pointed outHead.Tail need not be a list, as no conditions
are imposed on the form of its tail. But this does not matter: there is no
misunderstanding about the desired effect of, say
carcdr( timeofday( I2, 30 ).any( Object,at all" ). F, S ).
What we have here is a general procedure for getting at the rst and
second arguments of a term whose main functor is ./2.
We shall now specify the reverse of carcdr: a procedure which re-
tums, via its third parameter, a list constructed of its rst and second
parameters. We shall call it cons.
We do not really mind if the rst two parameters are not lists, so there
are no restrictions on their types:
cons( Object, Another, )
If Object describes an object and Another describes an object, then apply-
ing the list constructor to the two gives us a third object, whose descrip-
tion is Object.Another. This term is suf cient as a speci cation of the
third parameter, so we get
cons( Object, Another, Object.Another ).
 I .2. Operations I5

Here, again, we have a complete de nition of this procedure. But

notice that the order of parameters has no meaning in itself, so we might
as well have decided to pass the constructed list through the rst pa-
rameter:
cons( Object.Another, Object, Another ).
Variable names have no inherent meaning either, so cons is really the
same as carcdr. Indeed, when we read out the speci cation of carcdr, we
cheated a little: the parameter must be, or the parameter is to be-
come"-these distinctions were not present in the speci cation.
While both carcdr and cons could be so named to re ect their in-
tended use, they are both really a single procedure
conscarcdr( Head.Tail, Head, Tail ).
This is not very surprising, as cons is the reverse of the coin of which
carcdr is the face.
Now the call
conscarcdr( l.2.3.[I. H, T)
instantiates H to I and T to 2.3.[] and the call
conscarcdr( L, a, b.[] )
instantiates L to a.b.[ ]. But how is it done? We shall come to that, as soon
as we have cleared up a point of syntax.

1.2.2. Directives

In versions of Prolog deriving from Prolog-I0, the syntax of a simple

procedure de nition such as our conscarcdr example need not necessarily
differ from that of a procedure call. The meaning is de ned by context.
Such Prolog systems function in two modes: the command mode and
the de nition mode. Command mode is the default.
In command mode, the system reads and executes directives. The
directives are read in from the users terminal or from a le. Each direc-
tive is terminated by a fullstop (the character., immediately followed by a
white space character, including newline), and is either a query or a
command. A query is a procedure call or a sequence of procedure calls
separated by commas. Roughly, its execution consists of executing its call
and printing the resulting variable instantiations (see the end of Section
I.2.3 for a more precise description). For example, if conscarcdr has been
lb I An lntroduction to Prolog

de ned, then after reading the query

conscarcdr( l.2.3.[], H, T ).
the system writes
H =l
T=2.3.[]
(The actual printout might be in the special syntax used for lists; see
Section 4.2.1.)
A command has the form of a query pre xed by the symbol :-. Its
calls are executed but the variable instantiations are not written out auto-
matically. To get the same printout with a command, one would write
:- conscarcdr( l.2.3.[], H, T),
write(H = ), write( H ), nl,
write(T = '), write( T ), nl.
The terminology is somewhat uid: directives are often called goal
statements, while queries and commands are not always recognized under
those names. (Sometimes there are also slight syntactic differences. We
try to follow the original de nition of Prolog-l0, but in this book the
standard is set by the version of Prolog described in Chapter 7.)
De nition mode is entered upon executing the system procedure con-
sult/l or reconsult/l. (The argument is the name of the le from which
procedure de nitions are to be read; user is the name of the user's termi-
nal. The details are in Section 5.11.) ln this mode, the system accepts
procedure de nitions, which are also terminated by fullstops. Our de ni-
tion of conscarcdr is an example, but see Section l.3.l for the complete
syntax. Commands are allowed and properly executed in this mode, but
queries are not. De nition mode is exited when the system encounters the
de nition
end.
A note about comments in Prolog. A comment starts with a % charac-
ter (not contained in a string or quoted name) and extends till the end of
line. Be careful not to place a comment immediately after a dot that
terminates a clause: a fullstop is required.
As a point of interest, all directives and the basic building blocks of
procedures (called clauseswe will describe them in due time) are simply
single terms. Standard operator declarations include the in x functor ,
(comma) and the pre x functor :-. so the directive
:- p( 2, X ), write( X ), nl.
is really the term
I-( '.'( P( 2. X ). '.'( write( X ). nl ) ) )-
 l.2. Operations I7

This data structure is interpreted as a directive, so you need not worry

about these things unless you are an advanced Prolog hacker.
There is one important point, though probably you will nd it obvi-
ous. The actual parameters of procedure calls are the current instantia-
tions of terms directly written in the call. Thus
:- conscarcdr( a.b.[], H, T ), conscarcdr( L, H, T ), write( L ), nl.
will print out
a.b.[]

1.2.3. Uni cation

Since we succeeded in packing the whole de nition of conscarcdr into

its headingthe part specifying its name and formal parameterswe can
expect that its execution boils down to applying a suf ciently powerful
and general parameter-passing mechanism. This mechanism is imple-
mented by a term-matching operation called uni cation.
We will describe this operation by a pidginPascaI algorithm. The
function UNIFY is applied in tum to each formal and actual parameter pair.
If it retums true for all such pairs of terms, we say that uni cation is suocesful
(or succeeds); otherwise uni cation fails. Uni cation fails when the terms
describing the parameters do not match. In a very general sense this means
that the types of actual parameters are incompatible with those of the fomral
parameters.
function UNIFY ( var Actual, Formal : term ) : boolean;
var success : boolean;
be-gin success:= true;
if Formal is a variable then
Formal is instantiated to Actual
else
if Actual is a variable then
Actual is instantiated to Formal
else
if the main functors of Formal and Actual have
different names or arities then success:= false
else
while success and unmatched arguments remain do
success:= UNIFY( next argument of Actual,
next argument of Fomral );
UNlFY:= success
end;
I8 I An Introduction to Prolog

Notice that if we treat both the call and the procedure heading as
terms, then the process of matching successive pairs of parameters is
subsumed by the loop in UNIFY. We extend our terminology accord-
ingly, and say thatlike a pair of termsa call and a procedure heading
do or do not match. Altematively, we say that they do or do not unify (are
or are not uni able). The algorithm uni es matching terms. Uni ed terms
are indistinguishable, so they describe the same object.
If both Formal and Actual describe variables, then uni cation binds
them together. Variables which are bound together also represent the
same object: both their names refer to the same variable. (It is pointless to
ask whether the formal becomes an instantiation of the actual or the other
way round. Our algorithm implements the latter case, but this is not
observable from the outside. You can envisage a set of bound-together
variables as a chain of invisible pointers.)
Time for a very detailed analysis of a simple example: the procedure
p( A. b( c. A ) )
called with the query
p(X.b(X,Y)).
Figure l.2 shows the situation immediately before uni cation. The
horizontal line separates objects local to the directive and objects local to

caller
callee

FIG. l.2 Uni cation: before matching.

 I .2. Operations I9

the procedure. Note that objectsincluding variablesare accessible

both directly through their names and as components of other objects.
The rst pair of parameters is matched by binding X and A together.
The variables will behave as if they had merged into a simple object
(somewhat like two drops of water). This object is accessible under two
different names (Fig. l.3).
The second pair of parameters is uni ed in two phases. First, c be-
comes the instantiation of the amalgamated variables X and A. They
cease to exist as variables, but c is now also accessible through the name
A inside the procedure and through the name X outside (Fig. 1.4).
In the second phase Y is instantiated to the instantiation of A. The
object c is now accessible as c, A, X and Y (Fig. l.5).
The procedure p now terminates, but its local object c remains, being
accessible from the outside as X or Y. The instantiation of b( X, Y ) is
b( c, c ) (Fig. 1.6).
It is convenient to use a special notation for showing the effects of
uni cation. We shall write
A < B.[]
instead of A is instantiated to B.[]; and
A <-> B
instead of A and B are bound together.

/Q.if?
Q - z::.':.'

FIG. l.3 Uni cation: the rst pair of parameters is matched.

 /
G. l.4 Uni cation: the rst pair of bs arguments is matched

caller
I
g 4 callee

FIG. l.$ Uni cation: matching was successful.

El
l l
X -(
I'D

FIG. l.6 Uni cation: the callee is terminated.

 I .2. Operations 2|

Here are some example calls to our procedure

conscarcdr( Head.Tail, Head, Tail ).
l. conscarcdr( element.[], Car, Cdr )
Head < element, Tail < [],
Car < element, Cdr <- [].
2. conscarcdr( L, one.two.[], 3.4.5.[] )
L < Head.Tail, Head < one.two.[],
Tail <- 3.4.5.[].
Hence the instantiation of L is
(one.two.[] ).3.4.5.[]
3. conscarcdr( A.B, 2, 2.[] )
A <-> Head, B <-> Tail,
Head < 2 (and hence also A <- 2 ),
Tail < 2.[] (and hence also B < 2.[] ).
The instantiation of A.B is now
2.2.[].
4. conscarcdr( A.B.C, I0, [] )
A <> Head, Tail < B.C,
Head < I0, failure.
Uni cation fails. This is not surprising, as the rst actual parameter
describes a list of at least two elements, while the list constructed of
the second and third actual parameters would have only one element.
We shall wind this up with four general remarks.
First, we want to stress that the uni cation algorithm treats actual and
formal parameters absolutely symmetrically. This results in a very char-
acteristic property of Prolog: there is no difference between formal pa-
rameters used to bring information into a procedure and those used to
carry information out of a procedure. The direction of information ow
changes from call to call, as in examples l and 2 above. We can even
make a parameter serve both for input and for output. An example is the
call
conscarcdr( A.2.[], l, B) .
Here, Head < A and Tail <- 2.[];
then Head <- l ( and therefore A < l )
and B < 2.[] .
The result is that the rst formal parameter was used both for obtaining
information (that 2.[]) and for yielding information (that I).
This multi-way functioning of procedure parameters sometimes
makes it possible to use procedures in unexpected ways. Whenever we
Z2 I An Introduction to Prolog

shall say that a procedure does this and this, we shall not worry about
what it does after an unreasonable call. But you might nd thinking
about these things a useful exercise.
The second remark: effects of uni cation such as merging two varia-
bles are quite consistent with the interpretation of terms as descriptions of
types. We shall illustrate this with a very simple example. The following
two procedures accept only three- eld records whose neighbouring elds
are identical:
rst2( record( Fieldl2, Fieldl2, FieId3 ) ).
last2( record( Fieldl, Field23, Field23 ) ).
Each of these procedures can be thought of as imposing a constraint on a
description of the record type. The constraints are not mutually inconsis-
tent, so the directive
:- rst2( record( Fl,F2,F3 ) ),
last2( record( FI,F2,F3 ) ),
write( record( FI,F2,F3 ) ), nl.
writes out the description of a record whose three elds are identical:
record( XI, XI, XI ).
Similarly, the query
rst2( record( F l,F2, eld ) ), last2( record( F l,F2, eld ) ).
is answered with
Fl = eld
F2 = eld
Third, the convention that only the most interesting aspects of an
object are captured in a type description tumed out to be quite useful.
Example 3 would not have worked if the type of the rst formal parameter
had speci ed that the tail must be a proper list. The term A.B would not
have been accepted.
The fourth, and last, remark. If you follow the uni cation algorithm
carefully, you will notice that it can create cyclic data structures. For
example, if the procedure
same( X, X ).
is invoked with
same( f( V ), V ), write( V ), nl.
 I.2. Operations Z3

then we are in trouble. First, X < f(V); then V <X, that is to say
V < f(V). As a result, f becomes its own component and the printout
will be potentially in nite:
f(f(f(f(f(f(f(f(f(f(f(f( ....
Such cyclic structures can also cause trouble during uni cation. If we
write
same( f( V ), V), same( f( W), W), same( V, W ),...
then the uni cation algorithm will not terminate for the third procedure
call (this will probably manifest itself as recursion stack over ow): f
matches f, their rst arguments are both f, and their rst arguments are
both f, and so on.
All this could be avoided if a variable were not uni able with a term in
which that variable occurs. The uni cation algorithm is borrowed from
automatic theorem proving (see Chapter 2). The original algorithm contains
this occur cheek, but most versions of Prolog do not, as it considerably
increases the algorithms time complexity. Fortunately, cyclic stnictures
seldom occur in practice, and one learns to live with the knowledge that terms
are not always DAGs if one blunders badly. One version of Prolog (Prolog II;
see Section 9.2) is built to take advantage of cyclic data structures. They are
called in nite trees and are treated as bona de representations of graphs
arising in the real world. If one is careful, one can use such stnictures even in
more conventional Prolog systems: an example is the calltree program listed in
Appendix A.4.

1.2.4. Clauses

If we want a procedure which computes the fourth element of a list,

we can write
fourth( _._._.E4._, E4 ).
But this method is useless if we want the n-th (or even the hundredth)
element.
After parameters are passed, a procedure canjust as in other lan-
guagesexecute a sequence of operations. For example, a procedure
which prints the fourth element of a list would be:
fourth( _._....E4._ ) :- write( E4 ). nl.
Its body is a sequence of calls, separated by commas and pre xed by a :-.
As you see, a command is like a procedure without a heading.
24 I An Introduction to Prolog

A procedure heading, possibly followed by a body, is called a clause.

We shall now see how to use clauses for less trivial tasks.

1.3. CONTROL

1.3.1. The General Fonn of a Procedure

What happens when uni cation fails?

Part of the answer is that a procedure can consist of a number of
clauses. All these clauses must have headings with the same predicate
symbol, but the parameter speci cations may differ. When uni cation of a
call with the rst clause's heading is successful, the rst clause executes
its body (if any). When uni cation fails, its effects are undone: all varia-
bles which were instantiated by the attempt at uni cation are restored to
their original, unbound state. The call is then matched against the heading
of the second clause. If this is successful, the second clause is executed;
otherwise the third clause is attempted and so on. To execute a procedure
is thus to execute the rst of its clauses whose head matches the call (but
see the next section for a re nement of this statement). Roughly, the
matching clause contains code for that particular combination of parame-
ter types.
An elementary example is provided by an extended version of the
procedure carcdr of Section l.2.l.
carcdr( Head.Tail, Head, Tail ).
carcdr( l], _, _ ) :- write( cant crack empty list ), nl.
Here is a somewhat less trivial example, an immortal classic of intro-
ductory Prolog courses. It is a procedure which appends a list at the end
of another list:
append( Hd.Tl, List, Hd.TlAndList ) :-
append( Tl, List, TlAndList ).
append( []. List, List ).
All terms written in a clause are local to that clause. Both occurrences
of List in the rst clause refer to the same variable, which has nothing to
do with the variable named List in the second clause. The second clause
might as well have been
append( ll. QI4, Q14 ).
As in other programming languages with recursion, activation of a
clause is accompanied by creation of new instances of all its local objects.
 I .3. Control Z5

The tenns appearing in the clause describe these instances. Before an

attempt to unify a call with a clause heading can be made, a new instance
of the clause is created. When uni cation fails, the instance is destroyed.
Armed with this knowledge, we can now watch the effects of calling
append in the query
append( a.b.[], c.d.[], Result ).
For clarity, we shall use X , X etc., to denote different instances of a
variable named X.
The original call will successfully activate the rst clause, after the
following instantiations:
Hd <- a, Tl < b.[], List < c.d.[],
Result < a.TIAndList' (because Hd' is now a).
This clause will execute the call
append( b.[], c.d.[]. TlAndList ),
activating a second instance of the rst clause:
Hd" < b, TI < [], List" < c.d.[],
TlAndList < b.TlAndList .
In this instance, the body is
append( []. c.d.[], TlAndList ) .
This call does not match the rst clauses heading, so the second clause is
used:
List"' < c.d.[], TlAndList" <- c.d.[] .
The third instance of append has no calls to execute, so it returns to the
second instance. The second instance is done with its body, so it retums
to the rst instance, which also terminates. The variable Result was in-
stantiated to a.TlAndList, and TlAndList to b.TlAndList, and
TlAndList" to c.d.[]. Therefore, the query can be answered with
Result = a.b.c.d.[]
It is sometimes useful to represent the state of a computation by the
sequence of calls which must be executed. The sequence is often called
the current resolvent (see Section 2.4). If we use our procedure in the
directive
:- append( a.b.[], c.d.[], Result ), write( Result ), nl.
then the successive resolvents are as follows:
Z6 I An Introduction to Prolog

append(a.b.[],c.d.[],ResuIt), write(Result), nl.

append(b.[ ],c.d.[ ],TlAndList), write(a.TlAndList), nl.
append([],c.d.[],TlAndList"), write(a.b.TlAndList"), nl.
write(a.b.c.d.[]), nl.
?':'*1" !*:" E.
When no calls remain, the directive is terminated.
Here is the procedure to nd the n-th element of a list. Its rst param-
eter is n and the second a list. The third parameter retums the n-th
element of the list; if the element does not exist, the constant ? is retumed
and an error message is printed. It is assumed that the rst parameter is
not negative (we will leam how to check this in the next section). The
procedure is
nth( 0, _, ? ) :- write( nth( 0...... )??' ), nl.
nth( N, [], ? ) :- write( nth( ..,too short,.. )?? ), nl.
nth( I, El.-. El ).
nth( N, _.Tail, El ) :- M is N I, nth( M, Tail, El ).
Do trace its execution for a few examples.

1.3.2. Baektraclring

But what if a call matches none of the clause headings? An example is

the call
conscarcdr( notalist, something, other )
This suggests the answer. If none of the clauses ts the call, then
evidently the call is wrong: its set of actual parameters does not conform
to any of the type speci cations describing parameters acceptable to the
procedure.
As in other modem programming languages, such an erroneous call
does not abnormally terminate a program's execution but activates an
error-handling mechanism. In contrast to other languages, however, the
error is not necessarily handled by an active procedure present on the
activation stack. Prolog uses a more general method and takes into ac-
count even those procedures which retumed to their caller after success-
ful termination. Procedure instances are looked at, one by one, in reverse
order of their activation. The nearest such procedure instancecall it p
which contains as-yet-unactivated clauses matching its call is assumed to
be able to handle the situation. The computation is undone: its state is
made to appear as if the heading of ps most recently activated clause did
not match its call, and p is given a chance to execute other clauses.
 I .3. Control 27

This process is called backtracking, and a call which does not match
any clause heading is said to fail. Backtracking closely resembles our
behaviour in systematically searching for a solution to a problem. If we
end up in a blind alley, we get back to the nearest point at which we could
have applied another approach, and apply it. If no approach seems to be
working at that point, we retum to the previous point in which we appar-
ently made a wrong choice, and so on.
In implementation terms, each time a selected clause is not the last in
its procedure, a record is pushed onto a special stack of fail points (also
called choice points). The record contains all information necessary to
restore the state of the computation. When a procedure fails, the topmost
fail point is popped off the stack, the state described by it is restored and
the computation proceeds with the next clause.
It is important to note that not all effects of a computation are obliter-
ated on backtracking. Some system procedures do things which cannot be
undone, such as writing infonnation on a terminal screen. We say that
these procedures have side-effects.
Using our description of backtracking, try to follow the execution of
procedure p in the following example:
P =- q(X). writ(t1'yins(X)). nl.
female( X ), write( ok ), nl.
p :- write( Sorry! ), nl.
q( X ) :- writer( X ).
q( ? ) :- write( No more writers. ), nl.
writer( hesse ).
writer( mann ).
writer( grass ).
female( austen ).
female( sand ).
You should get the following printout:
trying( hesse )
trying( mann )
trying( grass )
No more writers.
trying( ?)
Sorry!
You may have noticed from this example that error handling is some-
what inadequate as a metaphor for backtracking. This is the subject of the
next section.
28 I An Introduction to Prolog

1.3.3. I-low to Use Backtracking

When a procedure instance is backtracked to, it behaves as if its most

recently activated clause did not match its call. We can therefore use
backtracking to implement extended type checking.
Recall from Section l.l.4 that we found it impossible to write terms
which could describe properties such as the object is a painter's name
or the tail is a properly constructed list. In other words, while rather
powerful in certain respects, this kind of type speci cation is weak in
others. This can be remedied by using procedures which do additional
type checking and either fail or successfully terminate, depending on the
outcome. If we want procedure q to accept only properly constructed
representations of paintings, we can write
q( painting( Painter, Name ) ):-
ispainter( Painter ), process( painting( Painter,Name ) ).
ispainter( rembrandt ).
ispainter( velasquez ).

Here, the role of ispainter is similar to the declaration of an enumeration

type in Pascal.
Prolog has several built-in procedures which can be used to check
properties of objects. For example, one can check whether the object
denoted by Something is an integer, by seeing whether the call
integer( Something )
succeeds or fails.
A number of built-in procedures implement comparison operations.
Like is, procedures for comparing integer values evaluate terms resem-
bling conventional arithmetic expressions. The procedures are <, =<,
=:= (equality), =\= (inequality), >= and >. Their names are prede-
clared as in x predicate symbols. For example the call
7*2+5=:=I+3*6
will be successful. There are also procedures comparing non-integer con-
stants according to their lexicographic ordering: @<, @= <, @> = , @>.
For example,
alpha @> beta
is a failing call.
Equality of constants can be determined by means of the procedure =
 l.3. Control 29

(= is predeclared as in x). The procedure is most easily expressed in

Prolog
X=X.
It can be used for any two terms, but of course it does more than checking
equality. It may cause its parameters to become equal, as it attempts to
unify them. For example,
a(b.X)=a(Y.)
will succeed after instantiating
X < c, Y < b .
Note that 7=7 succeeds, but 5+2=2+5 fails, as these are different terms.
Remember that all these procedures do not yield a Boolean result:
they only succeed or fail.
When we are interested in the structure of a compound object, we can
use a recursive procedure which does nothing but accepting the object.
Here is a version of carcdr which works only for true lists and fails for
objects such as a.b.c (but not for objects with variable tails, which match
ll)-
carcdr( Head.Tail, Head, Tail ) :- islist( Tail ).
islist( [] ).
islist( _.L ) :- islist( L ).
Such type checking can be quite general. For example, we can pro-
cess objects differently according to whether they are or are not members
of a set represented by a list:
process( Obj, Set ) :- member( Obj, Set ),
yes-action( Obj ).
process( Obj, _ ) :- no_action( Obj ).
member( El, El.Tail ).
member( El, _.Tail ) :- member( El, Tail ).
Try to trace the execution of process for a couple of simple calls, and
notice how member is called with successively shorter tails of the list,
until it nds a tail whose head is uni able with the rst parameter.
The fact that the rst clause of member expresses uni ability rather
than equality has very interesting consequences. The most obvious is that
the procedure can be used to retrieve information from a dictionary repre-
scnted by a list. The call
30 I An Introduction to Prolog

member( phone( krull, Number ),

phone( mann,II ).phone( hesse,5 ).phone( krull,II ).[] )
instantiates Number to ll.
When this information-retrieving effect is coupled with backtracking,
the result is rather striking. Consider the procedure
intersect( Ll, L2 ) :- member( E, Ll ), member( E, L2 ).
When given two sets represented by lists, the procedure terminates suc-
cessfully if the sets intersect and fails if they are disjoint. Here is a trace of
what happens when we call it with
intersect( a.b.c.d.[], c.d.[] ), write( ok ), nl.
I. intersect(a.b.c.d.[],c.d.[]), write(ok), nl.
(this activates the procedure:
Ll < a.b.c.d.[], L2 < c.d.[]
2. member(E,a.b.c.d.[]), member(E,c.d.[]), write(ok), nl.
(activates the rst clause of member:
E<->El,El<a)
3. member(a,c.d.{]), write(ok), nl.
(only the second clause matches the call)
4. member(a,d.[]), write(ok), nl.
(only the second clause matches the call)
5. member(a.[]). write(ok), nl.
(the call to member fails, nearest handler is in the procedure acti-
vated in step 2, so we backtrack to that situation)
6. member(E,a.b.c.d.[]), member(E,c.d.[]), write(ok), nl.
(the second clause now:
E <> El, Tail < b.c.d.[])
7. member(E,b.c.d.[]), member(E,c.d.[]), write(ok), nl.
(the rst clause:
E <> El, El <- b)
8. member(b,c.d.[]), write(ok), nl.
9. member(b,d.[]), write(ok), nl.
I0. member(b,[]), write(ok), nl.
(failure, backtracking to step 7)
ll. member(E,b.c.d.[]), member(E,c.d.[]), write(ok), nl.
(the second clause now:
E <> El, Tail' < c.d.[])
I2. member(E,c.d.[]),member(E,c.d.[]), write(ok), nl.
(the rst clause:
E <> El, El < c)
 l.3. Control 3]

I3. member(c,c.d.[]), write(ok), nl.

(the rst clause)
I4. write(ok), nl.
I5. nl.
Success.
Notice how the rst call to member in intersect is used as a back-
track driven" generator of successive elements on the list. A terminated
procedure can be reactivated if the effects of its execution prove unsatis-
factory. It can retum several resultsor behave in several waysand its
nal effect is determined not only by its actual parameters, but also by
what happens to the computation later on. It is this multiplicity of possible
behaviours that we have in mind when we say that, in general, a Prolog
procedure is nondeterministic. (This does not mean that its behaviour
cannot be predicted to the smallest detail.)
If one wants to see all the results produced by a nondeterministic
procedure, one can force Prolog to backtrack by calling an unde ned
procedure (the call will fail, because there is no matching clause). It is
customary to use the name fail, both for readability and because Prolog
makes it impossible to declare a procedure with this name. To print the
elements of a list, one can write
:- member( E, a.b.c.[] ), write( E ), nl, fail.
Altematively, one can use a query. After answering a query, the
system accepts a single printing character from the terminal. If the char-
acter is a semicolon, it backtracks; otherwise it terminates the query.
When all the possibilities are exhausted, the word no is printed and the
system reads another directive. For example,
user: female( W ).
system: W = austen
user: ;
system: W = sand
user: ;
system: no
If a successful query contains no non-anonymous variables (i.e. no instan-
tiations to show), the answer is yes.
Our intersect example does more than check for common elements. If
the elements are not ground, the sets are modi ed. For example,
intersect( one.X.three.l]. I.Y.[] )
32 I An Introduction to Prolog

succeeds after binding Y to one. This has a natural explanation. Since Y is

unknown, we cannot say that the sets do not intersect, but by binding Y
we ensure that the computation will fail if the supposition that Y is one
will turn out to be unacceptable. We shall then assume that X is I, that X
and Y are the same object, etc., etc.

1.3.4. Static Interpretation of Procedures

Detailed simulation of a program is not a very attractive way of leam-

ing its meaning. We insisted on doing it to help you understand what
happens inside the computer and to introduce techniques which can
sometimes be useful for debugging, when things are not happening the
way they should. But it is often quite clear what should happen, as many
Prolog procedures can be read without giving a thought to details of
execu on.
A clause which has no body is called a unit clause. It is a direct
de nition of a relation between its parameters. The clause
phone( hermann, 5 ).
says that hermann and 5 are in the relation phone. Other clauses can
extend the relation to other objects:
phone( mann, ll ).
phone( hesse, 5 ).
phone( krull, II ).
Unary relations can be thought of as expressing properties of objects:
red( herring ).
red( square ).
Nullary relations can denote general facts:
tired.
debugging.
A look at the clauses of phone tells that the call
phone( siddhartha, N )
will fail, and the call
phone( Who, 5 )
will nondeterministically produce hermann and (after a failure) hesse.
Note that the calls can be read as establish whether the actual parame-
 l.3. Control 33

ters are in the relation phone, i.e. succeed if they are in the relation, or
instantiate them so that they will be in the relation and succeed, or fail.
Somewhat less trivially, the unit clause
conscarcdr( Head.Tail, Head, Tail ).
can be used to establish whether the rst parameter is a list formed of the
second and third parameters. It is self-evident that
l. conscarcdr( a, b, c ) fails, because the objects are certainly not in the
relation;
2. conscarcdr( A.2.B, l, C.[] ) succeeds, because there does exist a list
of at least two elements whose second element is 2, such that its head
is I and its tail is a one-element listthe list is I.2.[] and the tail is 2.[ ];
3. conscarcdr( A, B, B.[] ) succeeds, because there do exist objects A
and B such that A is a list constructed of B and B.[]A is B.B.[] and
B can be any object.
Nonunit clauses are indirect de nitions of relations. Thus
append(
append( F; 1'11- ri- :'I',.""'-II" append( T, L, TL ).
can be read as
L is L appended to [], and
H.TL is L appended to H.T tfTL is L appended to T.
It is usually convenient to flavour this a little with the intended meaning,
as in
a list L appended to an empty list is L itself, and
a list L appended to a non-empty list H.T is formed of the head of
that list, H, and the result of appending L to its tail, T.
And. most spectacularly.
intersect( Ll, L2 ) :- member( E,Ll ), member( E,L2 ).
reads:
Ll and L2 intersect if an object E is a member of Ll and a member
of L2.
in other words
two lists intersect if they have a common member.
You will nd more about this in Chapter 2. But note here that this
interpretation does not fully explain procedures such as process of Sec-
tion l.3.3this is further discussed in Section 4.3.l.
34 I An Introduction to Prolog

1.3.5. The Order of Calls and Clauses

In practice, static interpretation is not always suf cient to explain a

program's behaviour. It cannot account for the order of calls in a clause
and the order of clauses in a procedure, because x and y means the
same as y and x. Yet this order is important, for three principal
reasons.
The rst reason is that some procedures, such as write and nl, have
side-effects, i.e. their results are not only variable instantiations. The
order in which several things are written has an obvious effect on the form
of the printout.
Another important reason is ef ciency. Here is a famous example
(Kowalski I974) of a naive naive sort:
sort( List, Sorted ) :- permute( List, Sorted ),
ordered( Sorted ).
The procedure generates successive permutations of a list until it nds
one that is ordered. If permute and ordered can be used both to check
their parameters and as generators, then this could also be expressed as
sort( List, Sorted ) :- ordered( Sorted ),
permute( List, Sorted ).
Here, successive ordered lists are generated until a permutation of the
rst parameter is found. Both procedures express the same de nition of a
sorted list, but while the rst is only very costly, the second is absolutely
useless.
A third reason is that all computations should be nite. We will illus-
trate this point with the procedure append, which can be written either as
append( H.T, L, H.TL ) :- append( T, L, TL ).
append( ll. L. L ).
or, apparently equivalently, as
append( ll. L. L ).
append( H.T, L, H.TL ) :- append( T, L, TL ).
Both versions are equivalent when append is used for appending. But
note that its precise reading from section 1.3.4 allows for other uses. For
example the second clause, H.TL is L appended to H.T if TL is L
appended to T, de nes H.TL in terms of H.T and L, but also H.T and L
in terms of H.TL. Indeed, append is often used for splitting a list. If one
executes
 1.3. Control 35

:- append( Front, End, a.b.c.[] ),

write( Front ), write( ' & ' ),
write( End ), nl, fail.
then the rst version of append will produce (after successive failures)
a.b.c.[] & []
a.b.[] & c.[]
a.[] & b.c.[]
[] & a.b.c.[]
and the second version
[] 8:. a.b.c.[]
a.[] & b.c.[]
a.b.[] & c.[]
a.b.c.[] & [].
This difference is not very important. But when we write
append( Ll, a.[], L3)
we expect that append will succeed, after instantiating the terms so that
L3 is a.[] appended to Ll. The second version does just this: Ll 1- [] and
L3 < a.[]; then, if we backtrack, LI < XI.[] and L3 1- XI.a.[]; then, if
we backtrack again, Ll < Xl.X2.[] and L3 < Xl.X2.a.[]; and so on-
there are in nitely many such solutions.
The rst procedure, however, rst looks for the last solution in this
in nite set, and this causes endless recursion.
Nevertheless, with careful programming, considerations of this sort
are needed only to obtain re nements of the general meaning of proce-
dures given by their static interpretation. Moreover, the order of calls and
clauses is usually a local thing, seldom requiring looking beyond a single
procedure.

1.3.6. The Cut

We shall now pass on to so-called extralogical features of Prolog.

These are simple and powerful mechanisms which play a large part in
making Prolog a practical programming language, but cannot be under-
stood in terms of static interpretation, as outlined in Section 1.3.4.
Since we have generators, we must be able to stop them. Suppose
that we have two methods for nding the solution of a problem described
in temis of two sets. Assume one of these methods is signi cantly cheaper
36 I An Introduction to Prolog

than the other, but a necessarythough not suf cient!condition for its
applicability is that the problem-de ning sets intersect. We might write
something like
try( Setl, Set2, Solution ) :-
intersect( Setl, Set2 ),
methodl( Setl, Set2, Solution ).
try( Setl, Set2, Solution ) :-
method2( Setl, Set2, Solution ).
Now if methodl fails, we want to try method2. But if the sets are large
and have many elements in common, we are effectively stopped by a
generator. Backtracking from methodl will cause intersect to nd another
way of showing that the sets do indeed intersect: this changes nothing, so
methodl will be attempted again and again until intersect enumerates all
the elements in the intersection of Setl and Set2. In terms of processing
time, this might be a disaster. And note that we are lucky: the generator is
not in nite.
To help in such cases, Prolog provides a commit operation, written as
I and called the cut procedure (old Prolog hands tend to call it the slash,
after the character /, which was its name in the original Marseilles Prolog).
When procedure p executes a cut, everything that was done by p up to
that momentincluding its choice of current clauseis taken as xed
and not to be reconsidered on backtracking. In implementation terms, !
cuts away the top section of the fail point stack, leaving only fail points
created before p was called.
Our problem can be solved by modifying intersect:
intersect( Ll, L2 ) :- member( E, Ll ),
member( E, L2), !.
The cut kills the generator of elements from LI.
A more involved example might be useful in clearing up doubts about
the effects of a cut. We will try to move a single cut around in our example
of section 1.3.2:
p :- q( X ), write( trying( X ) ), nl,
female( X ), write( ok ), nl.
p :- write( Sorry! ), nl.
q( X ) :- writer( X ).
q( ? ) :- write( No more writers. ), nl.
writer( hesse ).
writer( mann ).
writer( grass ).
 1.3. Control 37

female( austen ).
female( sand ).
If we insert a cut into the rst clause of writer:
writer( hesse ) :- !.
the printout will be
trying( hesse )
No more writers.
trying( ?)
Sorry!
If we insert it into q instead:
q( X ) :- !, writer( X ).
we will get
trying( hesse )
trying( mann )
trying( grass )
Sorry!
But if we choose to insert it at the end of this clause:
q( X ) :- writer( X ), I.
the program will write
trying( hesse )
Sorry!
By inserting the cut after the call to q in the rst clause of p, we would
obtain only
trying( hesse )
As evidenced by these examples, the cut is a powerful tool. A single
cut can drastically alter the behaviour of a program. It must be used very
carefully: Section 4.3.1 contains some useful hints.
An important property of the cut is that it can be used to implement a
sort of negation. When we want to list all male writers, we can write
:- writer( X ), male( X ), write( X ), nl, fail.
If, however, the program contains only descriptions of female persons (as
in our example), we must de ne male in terms offemale:
male( X ) :-female( X ). !, fail.
malet _ ).
33 I An Introduction to Prolog

When X is such that female succeeds, the second clause of male is cut off
and the whole procedure fails. When female fails, the second clause
takes over and the procedure succeeds. The trick is dirty, but very useful.
One must be careful, however: if the constant christie is not listed among
the females, male(christie) will succeed. (More on this in Section 4.3.2.)

1.3.7. Variable Calls

The negation schema shown in the previous section is of quite general

utility. For example, we could write a procedure for checking that two
sets (represented as lists) do not intersect:
disjoint( SI, S2 ) :- intersect( SI, S2 ), !, fail.
disjoint( _, _ ).
Prolog provides a very convenient extension which allows us to use
such schemas without going to the trouble of rewriting them again and
again. A variable call is a variable occupying the position of a call in a
clause or directive. When the tum comes to execute the call occupying
this position, the variable's current instantiation is taken as the call, by
treating its main functor as a predicate symbol and its arguments as pa-
rameters. If we de ne
do( X ) :- X
then the call
carcdr( el.[], A, B)
is exactly equivalent to
do( carcdr( el.[], A, B ))
as well as to
do( do( carcdr( el.[], A, B ) ) ) .
We can use this feature to de ne
not( X ) :- X, !, fail.
not( _ ).
and write
male( X ) :- not( female( X ) ).
disjoint( SI, S2 ) :- not( intersect( SI, S2 ) ).
The dirty trick is now nicely packaged.
 1.3. Control 39

In versions of Prolog described here, not is prede ned and the predi-
cate symbol is predeclared as a pre x symbol. Expanding male in-line, we
would write the directive of Section 1.3.6 as
:- writer( X ), not female( X ), write( X ), nl, fail.
Variable calls can be used to de ne many useful procedures. We shall
end by showing two companions of not: and and or. The rst is
written as a comma and the second as a semicolon; the rst succeeds
when both its pararneterstaken as callssucceed, and the second suc-
ceeds when either of its parameters succeeds (but establishes a fail point if
it is the rst one). Their de nitions are
',( A, B ) :- A, B.
:'( A._):- A-
':( _. B) 1- B-
Comma and semicolon are predeclared as in x symbols. After de n-
ing do, we could write the directive above as
:- do( ( writer( X ), not female( X ), write( X ), nl, fail ) ).
The extra parentheses are needed to avoid confusion with a call to do/5.
Priorities are chosen so that
artwork( X, Y ) :- painting( X, Y ), oil( Y );
etching( X, Y ), brass( Y ).
is equivalent to
artwork( X, Y ) :- ;'( ',( printing( X,Y ),oiI( Y ) ),
','( etching( X,Y ),brass( Y ) ) ).
To make the comma and semicolon appear a part of ProIogs syntax,
Prolog-I0 and some of its offsprings made the cut behave somewhat dif-
ferently for these procedures: they are transparent to it. Thus
artwork( X, Y ) :- painting( X, Y), oil( Y ), ! ;
etching( X, Y ), brass( Y ).
avoids checking the second altemative if the rst succeeds.
A similar exception applies to variable calls. If the procedure
a(X):-b,X.
a(_):-c.
40 I An Introduction to Prolog

is called with
a( ( d, !, fail ))
then the cut will commit all choices made by d and b and athe proce-
dure will fail without executing c.
One should avoid taking advantage of this peculiar property of the
cut. It is doubtful whether it is necessary.
2 PROLOG AND LOGIC

2.1. INTRODUCTION

Prolog stands for Programmation en logique (programming in

logic). Static interpretation of procedures (see Section 1.3.4) is possible
because Prolog can also be viewed as a system for proving theorems
expressed in logic. Adopting this viewpoint can provide the programmer
with new insights about the nature of his task.
In this chapter we attempt to introduce the fundamentals of this as-
pect of Prolog in an intuitive manner. Full appreciation of the subject is
possible only for people with a solid background in mathematical logic,
and we assume your knowledge of logic is very elementary. Consequently, the
presentation is often not suf ciently precise, and sometimes the tenninology is
a little unconventional: we are interested in Prolog rather than logic. The
chapter is a shortcut, so in some places you will nd it heavy going. A more
detailed, but still non-technical treatment can be found in Kowalski (1979b).
Another relevant book is Robinson (1979). See also van Emden and Kowalski
(1979).

2.2. FORMULAE AND THEIR

[NTERPRETATIONS

Below is a pair of formulae written in the language of predicate logic

(also known as rst-order logic or predicate calculus):
(2-I) Vx D(Z. x. x)
(2.2) VxVyVzD(x,y,z)=>D(S(x).y.S(z)).
41
42 2 Prolog and Logic

The basic building blocks of such formulae are predicates. A predicate

consists of a predicate symbol (e.g. D), optionally followed by argu-
mentsa list of terms in parentheses, separated by commas. A term is a
variable (e.g. x, y, z), or a functor (e.g. Z, S) with an optional list of
arguments, which are terms. Terms denote objects in some universe
(more on this presently) and predicates stand for relations between these
objects.
A single predicate is a formula. A larger formula can be built from
simpler ones by means of logical connectives. The commonly used con-
nectives, listed in order of decreasing priority, are
the negation (not), written as H
the conjunction (and), written as /\
the disjunction (or), written as V
the implication, written as =>
Parentheses can be used to increase clarity or ovenide priority.
A formula (i.e. also a subformula) can be pre xed by a number of
quanti ers, whose priority is lower than that of the connectives. A quanti-
er can be
the existential quanti er, written as ix and read as there exists an x.
the universal quanti er, written as Vx and read as for all x, or for
any x.
The formula pre xed by a quanti er is called its scope, and the quanti ed
variable is local to this scope (an occurrence of its name outside the scope
does not denote the same object). In this chapter we shall deal only with
hilly quanti ed fomrulae; i.e. our formulae will not contain unquanti ed
variables.
Our example formulae can be read as
for any objectcall it xthe object Z is in relation D with x and x
and
for any three (not necessarily distinct) objectscall them x, y and
zif x, y and z are in relation D, then so are objects S( x ), y and
S( z ).
In practice, it is more convenient to use a slightly abbreviated reading, in
which the second formula is
for all x, y and z, D( x, y, z ) implies D( S( x ), y, S( z ) ).
Formulae of this kind are purely formal statements. One cannot dis-
cuss whether they are true or false, because no particular meaning is
attributed to the functors and predicate symbols. To talk about a formu-
las meaning, we must give it an interpretation. An interpretation is a
 2.2. Fonnulae and Their Interpretations 43

de nition of a universe (the set of objects which can be denoted by terms)

and a decision to let predicate symbols and functors denote particular
relations and functions de ned in this universe.
A concrete interpretation maps a (fully quanti ed) f0lTlll1IEl to a state-
ment which is true or false, depending on what it says about relations
between objects. Somewhat imprecisely, we shall say that a formula is
true (or false) in an interpretation. Of course, some formulae are true in all
interpretations (the formula true is a trivial example, and A V H A is
another); others are false in all interpretations (e.g. false, A /\ H A). The
rst kind of fomiulae are called tautologies; fomiulae of the second kind
are called inconsistent.
As an example, consider the following two interpretations of fom1u-
lae (2.1) and (2.2). The rst interpretation is the following:
the universe is the set of natural numbers (positive integers);
Z stands for the number 1 (one);
S stands for the function S(x) = 2x;
D(x, y, z) is true if and only if xy = z.
Our formulae now become the true statements
for any natural number x, Ix = x
and
for all natural numbers x, y and z, xy = z implies 2xy = 22.
Another interpretation is:
the universe is the set of non-negative integers;
Z stands for the number 0 (zero);
S stands for the successor function S(x) = x+1;
D(x,y,z) is true if and only if x+y = z.
The fomiulae are now
for any non-negative integer x, 0+x = x
and
for all integers x, y and z, x+y = z implies (x+1)+y = 2+1.
If an interpretation maps a formula into a true statement, then this
interpretation is called a model of this formula. We can also speak about a
model of a set of for-mu1aean interpretation in which all of them are true.
Our two interpretations are models of the example fomiulae. If Z
stood for I in the second interpretation, then it would not be a model.
When an interpretation interests us as a model, fomrulae which are true
(or false) in that interpretation will be referred to as true (or false) in the
model.
44 2 Prolog and Logic

All interpretations are models of tautologies. Inconsistent formulae

have no models.
When we want to talk about a particular model, we prefer to use
symbols which have some mnemonic value. The formula
(2.3) Vh Vt conscarcdr( .( h,t ), h, t )
can be interpreted as
for all integers h and t, the difference between h+t and h is t".
but this is better written as
Vh Vt difference( +( h,t ), h, t ).
The similarity is interesting, though: looking for other models of our
statement of a problem often provides illuminating insights into its nature.
Notice that the natural interpretation of formula (2.3) is very
down-to-earth. A list constructor can be thought of as a function mapping
two objects (a head and a tail) into a third object: the universe can be a set
of data structures.

2.3. FORMAL REASONING

The notion of logical consequence allows us to perform formal reason-

ing, i.e. reasoning which takes into account only the syntactic fonn of
fomrulae and disregards their interpretations. We say that formula a is a
logical consequence of a set of formulae B, B, B" . .. if all models of the set
[3, B, 5" are also models of a. It is a fundamental fact of logic that there
exist inference rules, which are correct recipes for deriving logical conse-
quences (conclusions) of other formulae (premises), provided the latter
have a certain form. The inference rules are usually quite simple, but we
can use them as elementary steps in long derivations. This is the back-
bone of mathematics: a set of formulae (axioms) de nes a theory, which is
the set of all formulae (called theorems) true in all models of the axioms; a
formal derivation of a new theorem is called its proof. (The name axioms
is often reserved for a minimal set of theorems specifying the theory of
interest. We nd it more convenient to use the name for any given set
of theorems accepted without proofs.)
Some inference niles are relatively trivial applications of the de ni-
tions of logical connectives. A well-known example is the modus ponens:
from any formula ct and from any formula of the form a => B, derive
the formula B.
 2.4. Resolution and Horn Clauses 45

Now the de nition of implication can be stated as follows: if at and B are

arbitrary formulae, then, in any interpretation, ct => B is false if and only if
a is true and B is false in that interpretation. Hence, any model of both a
and at => B must also be a model of B.
Two other simple rules are
a :> B is equivalent to H a V B,
(i.e. one can be derived from the other)
and one of the De Morgan laws
H ( a /\ B ) is equivalent to H a V H B.
Do convince yourself of their validitywe will need them presently!
Armed with a number of inference rules, we can attempt to derive a
formula directly or by means of a technique known as reductlo ad absur-
dum. To derive formula at from a set of axioms, assume that H at is a
theorem: if the resulting theory is inconsistent, then a is a theorem. A
theory is inconsistent if it contains an inconsistent fomiula. In this method
of proof we often show inconsistency by nding a formula B such that we
can derive
B/\" -
It is worth noting that all formulae are theorems of an inconsistent
theory. This is because, there being no models of the theory, no formula is
false in any of the models. (This might not have sounded too convincing,
but notice that if we can derive false, then we can derive any formula a
using modus ponens and false :> at. For any at, the fomiula false => at is a
tautology, because it is equivalent to H false V at, that is to say true V at.)
Consequently, if the set of formulae
Ha:
B
Bf

is inconsistent, then at is certainly a theorem, regardless of whether the set

B, B, is consistent or not.

2.4. RESOLUTION AND HORN CLAUSES

We shall be interested in an inference rule which we shall call the rule

of resolution (Robinson I965). It says
from H ct V B and from a V -y derive B V -y.
46 2 Prolog and Logic

Its validity is not hard to explain. In any model of H a V B and a V -y,

either H a is false or a is false. In the rst case B must be true (or else
H a V B would not be true), in the second y must be true. If a model of
H a V B and a V 7 must also be a model ofB or a model of -y, thenby
de nition of disjunctionit is a model of B V -y.
There are two interesting special cases of this rule. One is
from H a V B and from a derive B,
and the other is
from H a and from a derive E1.
Here |:| stands for the empty formula, which must be treated as equiva-
lent to false if this f0l'l'l'l of the rule is to be valid.
The rule of resolution is useful for reductio ad absurdum proofs when
our formulae are written in a restricted form called clausal form. A clause
is a disjunction of literals. A literal is either a predicate (called positive
literal) or a negated predicate (called negative literal). All clauses are
pre xed by universal quanti ers, one for each variable in the clause.
We shall limit our attention to Horn clauses, which have at most one
positive literal each. Here is a set of four Hom clauses (the predicates are
all nullary):
AVHBVHC
BVHD
C
D
Now, if we want to prove that A can be derived from these clauses, we
can use the rule of resolution to show that by adding the Hom clause
HA
to our set of formulae, we obtain an inconsistent set of clauses. The proof
can be canied out in the following four steps (we use parentheses to make
things more clear):
1. fromHAandfromAV(HBVHC)deriveHBVHC
2. fromHBVHCandfromBVHDderiveHCVHD
3. fromHCVH Dand fromCderiveH D
4. fromH Dand fromDderive|]
Notice that this type of reductio ad absurdum proof is successful
when we derive the empty clause [1 (i.e. false). The special cases of the
resolution rule are used to shorten formulae, while the general rule is used
 2.4. Resolution and Hom Clauses 47

to generate formulae which can be shortened. Now, if in an application of

the resolution rule both the premises have one positive literal each, then
the conclusion must also have one positive literal (do you see why?).
Hence, the proof cannot be successful unless at least one of the clauses
has no positive literals. However, if one of the premises has only negative
literals, then so has the conclusion. If only one of the initial clauses has this
fonn, then the proof can be made particularly simple (Kowalski and Kuehner
1971; Hill 1974). One of the premises in the rst step is the clause without
positive literals. If this step cannot derive the empty clause, then the second
step must use the only other clause without positive literals, i.e. the conclusion
of the preceding step, and so on. Ifas in the exampleall our axioms have
positive literals, then the negated theorem must have none and the nal proof
has the form of an orderly chain, in which each step provides a premise for the
immediately succeeding one. In each step, we shall call the premise without
positive literals the current resolvent.
Each step consists in cancelling a negative literal Hx in the current
resolvent, by replacing it with the negative literals of a clause having A as
its only positive literal. The resolvent shrinks when one of its literals is
cancelled with a unit clause, which has only a single positive literal.
Because a clause is a disjunction of literals, it can be written as an
implication. By the De Morgan law (see Section 2.3) A V ( H B V H C )
is equivalent to A V H (B /\ C). This, in tum, is equivalent to
B /\ C => A.
By analogy, we can write
B/\C=>
to denote "B V "C (i.e. a Hom clause with no positive literals). The
empty consequent represents false, since false (or its equivalent) is the
only formula at such that a V (H B V H C) is equivalent to H B V H C,
for all B and C.
Similarly, we shall denote A, a Hom clause with no negative literals,
by
=>A
Here, the empty premise represents true: A is equivalent to true => A. An
empty clause has no literals and can be written
=>
i.e. true => false, which is equivalent to false.
It is preferable to write the implication from right to left:
A(=B/\C
43 2 Prolog and Logic

to suggest the reading

to prove A, prove B and C.
Recalling that our resolvents have no positive literals, we can now
write the resolution rule as
from (I a /\ B and from at <2 -y derive <2 -y /\ B
where B, y or both may be empty (in that case we do not write a /\). This
is a rather mechanical prescription: to get rid of at, nd an implication
whose consequent is a and replace a with its premises. This is clearly
justi ed: since at can be proven by proving y, then at A B can be proven by
provingy A B.
A clause being always pre xed with universal quanti ers for each of
its variables, it is convenient not to write the quanti ers. Our fomrulae
from Section 2.2 are two Hom clauses, written as
D(Z,x,x)<:
D(S(x).y.$(z))<=D(x.y.z)
Let us see whether these clauses are consistent with
<=D(S(Z).x.S($(Z)))
The clause can be thought of as a query whether there exists an x which is
in relation D with S(Z) and S(S(Z)). It is equivalent to
VX'D(5(Z).X.5(5(Z))).
and since this is the negation of what we are trying to prove, our derived
fonnulaif we succeedwill be
3xD(S(Z),x,S(S(Z)))
(if a is not false for all x, then there must be at least one x for which a is
true).
The rule of resolution, in the form presented above, is useless for this
example. In fact, it could only be used for nullary predicates, because the
argument for its validity does not apply to premises such as
VxHA(x)VHB(x) and VyA(Z)VHC(y).
Fortunately, we can also employ a simple inference rule called the
substitution rule. It says
from Vx a( x ) derive a(r), where -r is an arbitrary temi.
Here, a(x) means that the fomiula a contains occurrences of variable x.
a(-r) stands for a formula which looks exactly like 0:, except that all occur-
 2.4. Resolution and Horn Clauses 49

rences of x have been replaced by occurrences of term 1-. An example of

the rule's application is
from Vx D( S( Z ), x, S( x ) ) derive
D($(Z).3($(Z)).$(S(S(Z)))"
The substitution rule is valid, of course. In every model of Vx a(x), a is
true for any object x (that is what the quanti er says!): hence, it is true for
any particular object.
It should now be clear that
from Vx a( x ) and Vy B( y ) derive Vz a( z ) /\ B( z )
is also a valid inference rule. It can be looked on as an application of the
substitution rule: in all models of Vx a(x) and Vy B(y), a(1-) and B(1-) are
true for any 1-, hence (by the de nition of conjunction) a(-r) /\ B(-r) is true
for any 1-, and therefore Vz a(z) /\ B(z) is true.
The substitution rule allows us to match different formulae by using
appropriate variable substitutions. For example, we can easily show that
H D(x, y, z) is inconsistent with D(Z, Z, Z), because we can derive
H D(Z, Z, Z) from the rst formula by substituting Z for x, y and z. We
shall denote such substitutions by
x1-Z,y1-Z,z1-Z.
Our ability to match formulae allows us to apply to the problem at
hand implications which express general rules. This is best illustrated
with our running example. (We use apostrophes to distinguish between
variables similarly named, but quanti ed in different scopes or used in
different applications of a formula.)
1. Match <= D(S(Z), x, S(S(Z)))
and D(S(x), y, S(z')) <= D(x, y, z)
by substituting x 1- Z, y 1- x, z 1- S(Z).
2. From <: D(S(Z), x, S(S(Z)))
and D(S(Z), x, S(S(Z)) <= D(Z, x, S(Z))
derive <= D(Z, x, S(Z)) by the rule of resolution.
3. Match <= D(Z, x, S(Z)) and D(Z, x, x) <=
by substituting x 1- x, x 1- S(Z).
4. From <= D(Z, S(Z), S(Z)) and D(Z, S(Z), S(Z)) <2
derive I] (the empty resolvent) by the rule of resolution.
A noteworthy feature of this example is that the term S(Z), nally
substituted for the x in
<=D(S(Z),x,S(S(Z)))
50 2 Prolog and Logic

can be thought of as a counterexample to the disproved hypothesis that

this formula is consistent with the others. We leamed in effect that, in any
model of the two clauses playing the role of axioms,
it is not true that Vx H D( S( Z ), x, S( S( Z ) ) ), because
H D( S( Z ), x, S( S( Z ) ) ) is false when x is S( Z ).
But this is the same as saying that our question
does there exist an x such that D( S( Z ), x, S( S( Z ) ) )
is answered with
yes, S( Z ) is such an x.
Recall our two interpretations from Section 2.2. In the rst we asked
whether there is an x such that 2x = 4, and the answer is that 2 is such an
x. In the second interpretation, the question whether there is an x such
that I + x = 2 was answered by 1.
The various substitutions were used to narrow the set of interesting
objects to those objects for which the formula being disproved is not true.
Indeed, it is evident that for all x other than S(Z), the fomiula
H D(S(Z), x, S(S(Z))) is true in both interpretations. It is so in all models of
the two original formulae. but we shall not attempt to justify it directly.
Our example would be an indirect justi cation if we could be certain
that the substitutions did not lose other objects satisfying the dis-
proved formula. Such certainty would be of practical value, because the
answer to our query (i.e. our counterexample) can be a term containing
variables. We want it to be as general as possible. in the sense that the set
of all terms obtainable by substituting something for its variables should
be the set of all answers.
It is a fundamental fact of resolution theory that the algorithm of
uni cation (as presented in Section 1.2.3, but extended with the occur
check) nds the most general set of substitutions needed to match two
literals. Most general" means that it is contained by all sets of substitu-
tions which make the literals match. When we match A(x) and H A(y),
both x 1- y and y 1- x are possiblewe treat them as indistinguishable.
In a sense, this is the minimal necessary set of substitutions.
We used uni cation in our example, so we did not lose any solutions.
Notice, however, that our discussion is concemed with the effects of the
substitution rule; as we shall see, different proofs can come up with
different solutions. Try the conscarcdr and append examples from Chap-
ter 1 to get a feeling for this kind of proof. You may use in x notation for
functorsit does not matter. The intersection example might be a little
more dif cult: read on.
 2.5. Strategy 51

2.5. STRATEGY

Disjunction being commutative, we can apply the rule of resolution to

any literal in the current resolvent: in our examples, we always chose the
leftmost one. The choice does not affect our ability to nish the proof, as
we must be able to cancel all the literals before obtaining the empty
resolvent. As it tums out, the desirable properties of uni cation men-
tioned in the previous section ensure that the order in which the cancel-
ling of literals is perfomied does not in uence the nal outcome of the
proof. The length of a proof, however, can be affected by the choice of
literals very strongly indeed. We shall discuss this matter at the end of this
section.
In our examples, at most one clause could be used to cancel a literal in
each step. In general, a number of clauses can be applicable (after suitable
matching) to a given literal. Choosing the right clause could be important,
because some of them can lead into blind alleys. After many steps, we
may tum up with a resolvent to which the rule of resolution cannot be
applied (because there are no matching clauses for its literals), even
though another choice of clause at an earlier step might have speedily led
to the empty resolvent.
The situation is illustrated by Fig. 2.1b, which shows part of the
sear-ch space for the problem listed in Fig. 2.1a. The space is tree-shaped:
each path from the root to a leaf represents a possible derivation se-
quence; its nodes are labelled with the successive resolvents. Some of the
paths are successful, some end in failure.
Notice that several subtrees occur more than once. This effect would
be less pronounced if the resolvents re ected the history of substitutions,
but we did not feel up to creating such a drawing for predicates with
arguments. Try it for the application of intersect traced in Section 1.3.3.
(Figure 2.5 will show you how to do it.)
Prolog always tries to use the leftmost literal, so its search space is
considerably smaller, as illustrated in Fig. 2.1c. Whenever it is presented
with a number of applicable clauses, the system always attempts the rst
one rst. When it encounters failure, it backtracks and tries another path.
In effect, it executes an orderly preorder search of the search space tree.
Figure 2.1c also illustrates the effect of a cut: a part of the search
space is shom off, but one must be aware that this part may contain
solutions! While this is not important in the example (if we expect a yes!
no answer), in general different solutions may represent essentially differ-
ent instantiations (i.e. substitutions) of variables in the root of the tree.
This is not in contradiction to our earlier statements about the desir-
able properties of uni cation. As evidenced by Fig. 2.2, by choosing
52 2 Prolog and Logic

different literals we change only the order in which things are proved, but
the general structure of the proof is not changed. It is convenient to
represent the structure of a proof by means of a proof tree (do not confuse
it with the search space), as illustrated in Fig. 2.3. The rst tree shows the
proofs of the preceding gure, which all used B and C to prove A, and F
to prove B. The other proof trees represent classes of proofs obtained
through a different choice of clauses: the proofs are carried out quite
differently.
Structurally different proofs use different subsets of the available
clauses for performing various subproofs, so the counterexamples of
Section 2.4--which are descriptions of sets of objects for which the
clauses used cannot all be truemay turn out to be di 'erent. Therefore, not
all solutions are the same.
Proof trees are interesting also because they re ect the invocation
tree when clauses are treated as procedures. As long as there are no
failures, the conventional procedure activation stack can be regarded as
an equally conventional stack used for preorder traversal of the proof
tree, orif failure is imminentof a quasi-proof tree which has a failure

101 A== BAC

A"-= D

B== Dt\E

B<:= F

C<=

C"== FAD

D-1::

F-= 1=A

The axioms The (negated) theorem

FIG. 2.1 (a) A search space: the source formulae. (b) A search space: an initial part of
the complete space. (Choice of literals denoted by an arc, choice of clauses by a dot.) (c) A
search space: Prolog search space has no choices for literals. (Solution one is found, the
others could be found on backtracking. ! marks the subspaces made unreachable by execut-
ing a cut at the end of A's rst clause.) (continued)
 { L_hr__m_r _r _ _h_____~___Dar1?OO
OUrrZu
TO
_
O_L__ _ r_u_rn6_?r9_m_; _I
tr
____r .<r
wt_ ___m_

U_n_mr
'_h _h_rm {U_ I"_ O_" _ _
___

or _ B
ur>mr

_
Ur
Wr
IUrWr3
W!
__Ur
II:
_m___
U__ m_m
r_rr_m_?O_"3 r_<
/<<
\
nq_:2
miO
__.|rI_m_ _r __nmu_r
/_\/\_v\
/ A
A ___V\
_V\
r_ _ _ _m_'mt, __
___I
_II
____
.
.

mr r
_____
/_\ di
__N
_ m_ _ _ _G
K
_
DU
Ur_DDUD
_6_ _,mr
54 2 Prolog and Logic

(cl "A
I

e\

/By 1C
--Dv--Ev-1C -FvHC
I Us

I
HEV -rc WC
I

I I

tail Bl HFVHD

HD

FIG. 2.1 (Continued)

in its rightmost leaf -(see Fig. 2.4). Backtracking to the nearest choice
point in the search space may reopen an attempt to prove a node in a
nished branch of the quasi-proof tree, i.e. to reactivate a terminated
procedure (this would happen if D had a second clause in our example).
Obviously, this cannot be done with a single simple stack: the matter is
discussed in Chapter 6.
The word strategy refers to the way in which a theorem-prover
(whether automatic or human) nds its way through a search space. In
logic programming literature the preferred term is control. It is not all a
question of choosing the order of clauses and literals. Some logic pro-
gramming systems employ no backtracking, choosing to cover the search
space in a breadth- rst manner. This helps avoid problems caused by
misapplied generators and left recursion (see Fig. 2.5 and Section 3.5.2).
There is, of course, the problem of potentially exponential memory re-
quirements. Prologs appetite for memory is at most roughly linear in
number of attempted derivation steps. This is paid for with the cost of
backtracking: the time complexity is still exponential.
 HA

HBVHC

HFvHC "B

B//////\\\\\\a l A*:'Bt\C

1 1 --
HC HF HF B<:: F

|:1 1:1 El F:
The clauses used
lhe tree ot possible proot paths in these proofs
FIG. 2.2 Choice of literals with xed choice of clauses from Fig. 2.1. (Prolog would
choose the left.most path, but only after some backtracking caused by choosing the rst
clause of B.)

A A A

B C B C D

F E] F F 0 El

I
El El El El
FIG. 2.3 Proof tr'ees for the example of Fig. 2.1. (The leftmost tree represents the
proofs of Fig. 2.2. Backtracking would cause Prolog to build each tree in tum, from left to
right.)
 A

/\
El as
FIG. 2.4 A quasi-proof tree, representing a failing attempt to prove B by using its rst
clause. (The right son of A was not generated.)

Mle,.ll..l2.xlll

e<-I '
l'(-.12.!)

E] Mte..l2.xl
e<-2 '
l(x

U Mte.xl
x<-.le,l'l '

x(.le,l-ll

U Mle,l')

"<"- ll ' Mle .te tn<=

lI(__(_[1; Mle,.ia,l)l %Mle,l)
if-ll .-IL-12. Ill

U e'l) e I

Th; gpagg The clauses

FIG. 2.5 Prolog search space for an in nite generator. (The example is member.)
 2.5 Strategy $7

There have been attempts to decrease this cost by means of a more

sophisticated backtracking strategy (Bruynooghe 1978, Pereira and Porto
1980b, 1982, Bruynooghe and Pereira 1981).
The more ambitious scheme, appropriately called intelligent back-
tracking, attempts to retain subproofs (which would otherwise have
been discarded) in order to avoid recomputing them again and again. In
other words, it attempts to take advantage of the multiplicity of identical
subtrees in a search space (compare Fig. 2.1b).
A simpler approach, called selective backtracking, consists in ana-
lysing which variable instantiations caused the failure. It is then possible
to backtrack directly to the nearest point where one of these instantiations
was made or where the computation would take an entirely different
course. In some cases this can save us a lot of thrashing about in a failure-
infested region of the tree's crown.
Unfortunately, these interesting ideas have not in uenced Prolog im-
plementations. They require a further complication of the already com-
plex runtime data structures and they do not mesh well with side-effects
of system procedures. The fact that Prolog can be used as a practical
language is still largely due to our dexterity in ghting exponential com-
plexity with the cut.
Attempts to modify Prologs strategy so that it would incorporate
parallelism or coroutining have been a little more successful. Parallelism
consists in growing various branches of a proof tree (or even several
trees) simultaneously. It is dif cult, not only because it raises tricky tech-
nical problems, but because we still lack suf cient understanding of its
effects on both time!space complexity and the number of solutions gained
or lost. Several very different approaches have been documented, but
none of them seems to answer all pertinent questions. Some of the refer-
ences are (Clark and Gregory 1983, Shapiro 1983b, Conery and Kibler
1983, Wise 1984, Eisinger et al. 1982).
Coroutining is just that: switching control between several active pro-
cedures. In terms of our drawings, coroutining is a non-trivial traversal of
a proof tree. Roughly, it is a matter of choosing a different order of literals
(a different path in Fig. 2.2). It is possible to demonstrate spectacular
improvements in the perfomiance of some programs when they are exe-
cuted in coroutining fashion. A simple example is the naive naive sort (see
Section 1.3.5):
sort( List, Sorted ) :- permute( List, Sorted ),
ordered( Sorted ).
When the execution of permute is interleaved with that of ordered so that
the latter can cause failure as soon as the rst out-of-order element is
58 2 Prolog and Logic

produced by the fomier, the program's behaviour compares very fa-

vourably with that shown when each pemiutation must be completed
before ordered is called. When the initial sequence of a permutation is
rejected, all other permutations starting with the same sequence can at
once be rejected as well, resulting in a signi cant reduction of the search
space.
Section 9.2 contains two short examples of coroutining Prolog pro-
grams. Coroutining comes in many avours: some of the references are
(Clark et al. 1979, Clark et al. 1982, Porto 1982, Colmerauer et al. 1983).
Unfortunately, most of these schemes are of restricted utility (Kluiniak
1981). The problem is that coroutines do not mesh well with backtracking.
We comment on this at greater length in (Kluiniak and Szpakowicz 1984).
3 METAMORPHOSIS GRAMMARS:
A POWERFUL EXTENSION

3.1. PROLOG REPRESENTATION

OF THE PARSING PROBLEM

We shall begin with a very simple fomtulation of the parsing problem:

given a sequence of items, nd out whether it has some presupposed
structure. The problem appears e.g. in programming languages when we
want to make sure that some text is a syntactically valid statement. Ad-
missible structures are usually described by a context-free grammar. As
an example we shall consider the following small grammar in Backus-
Naur-Fomt, which describes simple list expressions:
< list > ::= ()|(< items >)
< items > ::= < item > | < item >, < items >
(3.1) < item > ::= < atom > | < list >
< atom > ::= < letter > | < letter > < atom >
<letter>::=a|b|c|d|e|f|g|h|i|j|k|l|m|
lolplqlrlslllulvlwlxlylz
Terminal symbols of this grammar are small letters, round brackets,
and a comma. For example, the list

(3.(bl8X))
consists of 12 terminal symbols.
There are several commonly used methods of describing the structure
of a list (or, more generally, of a valid sequence of terminal symbols). The
59
60 3 Metamorphosis Grammars

method we adopt here leads to an elegant formulation of the parsing

problem in Prolog.
We shall depict a sequence of terminal symbols in a graph (3.2):
I I , I b 9 , o I I I
1 In In In an _ arr i I-r It 0-0 at-I an I-1 I

Every node in this graph corresponds to a boundary between two consec-

utive terminal symbols; every edge connecting two nodes corresponds to
the tenninal symbol it is labelled with. Two edges are contiguous if they
share a node; a sequence e, , e,,, of edges is contiguous if e,- and e,-.1 are
contiguous fori = 1, 2, ..., m 1. For example, the edges labelled b, i, g
are contiguous. The labels of contiguous edges are also contiguous.
A sequence of contiguous labels may constitute a whole which is
meaningful in that it corresponds to the right-hand side of a production.
For example, the (only) label of the one-element sequence of edges
0
Oi-i

constitutes a letter; the labels of the sequence

O X
Q-i->-@-in
constitute an atom.
We shall describe such meaningful combinations by connecting the
extreme nodes of a contiguous sequence by an edge. The edge will be
labelled with the name of an appropriate non-terrninal symbol, as for
example in Fig. 3.1.
To be able to represent graphs in a program, we must give each node a
unique name. For example, we can name nodes with numbers:
I I , I b I g I 0 x ) )
0 I-I1 Fl In e_ I-If IIt- -- B - -2- ill I-F it-I - III It
1 2 3 t. 5 5 7 B 9 10 ll 12 13

We can represent such a graph as a set of edges, every edge ex-

pressed by a unit clause that speci es the label of the edge and the names
of the nodes it connects. Perhaps the most compact way is to use the label
as the clause name, e.g.
atom( 9, ll ).
letter( 9, 10 ).
o( 9, 10 ).

' This manner of presentation is due to Colmerauer; it was also used by I-(owalski
( 1979b).
i For other ways of representing graphs in Prolog, see Sections 4.2.4 and 4.4.3.
 3.1. Representation of the Parsing Problem 61

atom

letter
.41, Q
FIG. 3.1 Meaningful combinations of edges.

We now observe that clauses which represent edges labelled with

non-terminal symbols might be derived from those corresponding to ter-
minal symbols, by virtue of general structural relationships inherent in the
grammar. The reasoning would be roughly as follows:

letter( 9, I0 ) because o( 9, 10 ): Q is a letter;

letter( 10, ll ) because x( I0, ll ): 5 is a letter;
atom( 10, ll ) because letter( 10, ll ): a letter makes an atom;
atom( 9, ll ) because letter( 9, 10 ) and atom( 10, ll ): a letter and
an atom make an atom.
Relationships of this kind can be generalized in a straightforward
manner, e.g.
letter( K, L ) :- o( K, L ).
(3 3) letter( K, L ) :- x( K, L ).
' atom( K, L ) :- letter( K, L ).
atom( K, M ) :- letter( K, L ), atom( L, M ).
Contiguity of edges is assured by using the same tenn (variable name) to
denote every intermediate node: once at the end of an edge and once at
the beginning of the next one.
Given the clauses that describe edges with temtinal symbols, e.g.
o( 9, 10 ).
0'4) X( 10, rt ).
we might now derive all the remaining relevant edges. Strictly speaking,
they would be present only implicitly. For example, to con rm the pres-
ence of the edge
atom( 9, 11 )
we would issue the command
:- atom( 9, ll ).
from which the following computation might ensue:
62 3 Metamorphosis Grammars

atom( 9, ll ).
letter( 9, L ), atom( L, ll ).
o( 9, L), atom( L, 11 ).
(3.5) L 1- I0
atom( 10, ll ).
letter( 10, ll ).
x( 10, 11 ).
success
The method of specifying the initial graph is rather awkward, even for
this small example. Moreover, it requires that terminal symbols be only
identi ers (nullary functors)-the restriction is unnatural but, fortunately,
unnecessary. We shall now describe a slightly different and much handier
notation.
Names of nodes need not be consecutive integers. On the contrary, it
is much better to derive (unique) names from the original sequence of
terminal symbols than to introduce another, completely independent no-
menclature. We shall exploit the one-one correspondence between a
node and the sequence of (contiguous) edges following it. As the name of
a node we shall take the list of terminal symbols labelling the correspond-
ing sequence. For example, the leftmost node of the graph (3.2) will be
named
(.a.,.(.b.i.g.,.o.x.).)'.[]
and the name of the rightmost onecorresponding to the empty sequence
of nodes_will be
ll
With this notation, the (implicit) clause describing the atom ox becomes
atom( o.x.').).[], ).').[] ).
Notice how the underlying sequence of terminal symbols can be seen
without resorting to separate clauses for Q and : it is simply the differ-
ence of the rst and the second node names, Q and _x in our case. For a
terminal symbol this difference is guaranteed to consist of the symbol
itself, as, say, in
Jlt
0-)-0

X,')'.'l' . [1 'l.)'. E]
In other words, if an edge connects the nodes X , Y and is labelled with the
terminal symbol T, then
X=T.Y
 3.1. Representation of the Parsing Problem 63

In order to allow arbitrary terms as terminal symbols, we can write,

e.g.
tem1inal( o, K, L )
instead of o(K, L). Moreover, rather than writing

temtinal( o, o.x.)'.).[], x.).)'.[] ).

terminal( x, x.).').[], ).).[] ).

we shall use the general-purpose, one-clause auxiliary procedure

tem1inaI( T, T.Y, Y ).
However, now we need some other way of specifying the initial sequence
of terminal symbols, which in the previous fomtulation could be read
from the assertions (3.4). Before we explain this, we shall rewrite (3.3):
letter( K tertninal( o, K, L ).
letter( K l"l" tertninal( x, K, L ).
atom( K, L letter( K, L ).
atom( K, M ) :- letter( K, L ), atom( L, M ).
terminal( T, T.Y, Y ).
The computation analogous to that shown in (3.5) would now look as
follows:
atom( o.x.).).[], ).').[] ).
letter( o.x.).).[], L), atom( L, ).')'.[] ).
tem1inal( o, o.x.).).[], L), atom( L, ).).[] ).
L 1- x.)'.).[]
(3.6) atom( x.').').[], ).')'.[] ).
letter( x.).).[], ).)'.[] ).
terminal( x, x.)'.).[], ).').[] ).
success
All the necessary infomtation about the initial graph was supplied by the
rst call. What is more, the graph itself is now implicit: we only getand
manipulatethe two sequences of terminal symbols.
We are now in a good position to restate each instance of the parsing
problem in terms of Prolog. A grammar is given in the form of Prolog
clauses, each clause corresponding to some structural relationship be-
tween a unit and its immediate components (in particular, to a BNF rule).
For example,
items( K, N ) :-
item( K, L), tem1inal( ,, L, M), items( M, N ).
64 3 Metamorphosis Grammars

A call on one of these clauses (or, to be more precise, on the procedure to

which it belongs) fully speci es two lists of terminal symbols, the second
being the tail of the rst. As a matter of convention, the clause name will
also be the name of a nonterminal symbol, i.e. it will tell us what structure
we want to attribute to the underlying sequence of terminal symbols. For
example, the call
:- items( b.i.g.','.o.x.')'.')'.[], ')'.')'.[] ).

can be interpreted as the question: In the graph detennined by the param-

eters, can an edge labelled with items be validly drawn between the ex-
treme nodes? Or brie y: Is items the valid structure of a given sequence
of terminal symbols, big,ox in our case?
The answer to this question is YES if the call succeeds, and NO
otherwise. Examples of unsuccessful attempts to parse are;
:- items( ,'.o.x.).').[], ).').[] ).
/items cannot begin with a comma!
:- atom( o.x.')'.).[], ').[] ).
/atom cannot end with a bracket!
Procedural interpretation can be expressed in terms of the successive
augmentation of the original graph. Every successful call implicitly adds
an edge. Parsing succeeds if we can connect the extreme nodes with a
single edge. This construction proceeds bottom-up: we can imagine an
edge being added only after the successful termination of a corresponding
call.
We shall illustrate this by a complete program for parsing lists.

list( K, M ) :- terminal( (, K, L ), terminal( ), L, M ).

list( K, N ) :-
tenninal( (', K, L ), items( L, M ), tertninal( '), M, N ).
items( K, L ) :- item( K, L ).
items( K, N ) :-
item( K, L ), terminal( ,', L, M ), items( M, N ).
(3 7) item( K, L) :- atom( K, L ).
' item( K, L ) :- list( K, L).
atom( K, L ) :- letter( K, L ).
atom( K, M ) :- letter( K, L), atom( L, M ).
letter( K, L ) :- terminal( a, K, L ).

letter( K, L ) :- terminal( z, K, L ).
terminal( T, T.Y, Y ). -
 3.1. Representation of the Parsing Problem 65

The call on list in the command

=- list( (.a.','.'(.b.i.g.','.o.x.)'.').[], [1 ).
results in the implicit construction of the graph shown in Fig. 3.2. Notice
the similarity of this graph to a conventional parse tree (Fig. 3.3).
The parameters of a call that initiates the parsing serve as an input
and an output parameter. The former contains a given list of terminal
symbols. Some initial segment of this list is supposed to constitute the
unit under consideration. For example, in
atom( o.x.).')'.[], ').').[])
we expect that some initial part of the list
o.x.)'.')'.[]
constitutes an atom. Should that be the case, the computation succeeds
provided the second parameter matches the tail of the list which remains
after chopping off the initial segment. For example, ').')'.[] remains
after chopping o and x off the list o.x.)'.).[]. In most cases the second
parameter is a variable, so that it actually behaves like an output parame-
ter. As an example, the call
(3.8) atom( o.x.)'.').[]. Tail )
instantiates Tail as )'.').[]compare this with (3.6).
I151

|lt'n$ _

itg_ms
item
list J?

.1?
items

item _ i
rl ms

l 1 |g_j
FIG. 3.2 The graph for the list (a,(big,ox)).
66 3 Metamorphosis Grammars

IISI

I /..\ ern I

item i items

atom item

letter list

/....\.
atom item

letter \|lom at O rn

letter atom letter atom

i letter o letter

g it
FIG. 3.3 The parse tree for the list (a,(big,ox)).

If the second parameter is a variable, only the entry node of some

subgraph of the whole graph is known. Parsing then may give ambiguous
results. For example, the call
items( b.i.g.,.o.x.).)'.[], Tail )
might succeed with Tail instantiated to ','.o.x.')'.').[] or to ).).[]. In
general, the results depend on how the clauses of a parsing program are
ordered. In the program above, the recursive clause for items would only
 3.2. The Simplest Form of Grammar Rules 67

be activated because of forced failure coming after a successful parsing

of as items.
Recall now that the parameter of a Prolog procedure can, in principle,
be bi-directional, the directioninput or outputdepending on the form
of the corresponding actual parameter. This also applies to calls that
initiate parsing. If the rst parameter is a variable, what we ask is whether
there exists a sequence of terminal symbols that has a particular struc-
ture. For example, the call
Iist( AList, [] )
should instantiate AList to any valid list of terminal symbols; in other
words, some list should be constructed, or synthesized. One example of
such a list is the empty list.
However, the situation is not fully symmetric. For any given se-
quence of terminal symbols, a call on list either succeeds or fails, i.e.
every sequence can be classi ed as a list or a non-listcan be syntacti-
cally analysed. Not so with synthesis. It is easy to see that the two calls
Iist( AList, [] ), fail
will act as a generator of one-element lists:
() (a) (b) (z) (aa) (ab) (az) (aaa) (aab)
Moreover, if we reorder the two clauses for item, the call on item with a
variable rst parameter would result in in nite recursion.

3.2. THE SIMPLEST FORM

OF GRAMMAR RULES

The input and output parameters of the clauses that constitute a pars-
ing program, such as (3.7), are the basis of yet another interpretation of
those clauses: in terms of operations on sequences of terminal symbols.
Take the clause
items( K, N) :- item( K, L), terminal( L, M ), items( M, N).
It can be read as follows: (an instance of) items can be chopped off
(recognized at the beginning of) K, leaving N, if (an instance of) item can
be chopped off K, leaving L, and then a comma can be chopped off L,
leaving M, and nally (another instance of) items can be chopped off M,
leaving N. Now the essence of all this is that items consist of an item, a
comma, and items. The other information can be routinely added to this
fundamental fact. All we need is four variables to stand for successive
remainders of the initial sequence of terminal symbols.
68 3 Metamorphosis Grammars

In the notation we shall use henceforth, this routine information is

suppressed. The notation resembles BNF productions. The lefthand side
of a Prolog grammar rule names the construction, and the righthand side
enumerates its constituents. For example:
atom > letter, atom.
The symbol > is rendered in Prolog as --> (it must be written without
intervening blanks). There is a simple convention to distinguish nontermi-
nal and terminal symbols: the latter are enclosed in square brackets, e.g.
items > item, [ , ], items.
Contiguous terminal symbols can be enclosed in a single pair of brackets.
For example, the rule for empty lists can be written as
list > [ '(', ') ].
If all terminal symbols are characters (one-character nullary functors), we
can use string notation:
list > "().
Such grammar rules are merely syntactic sugar for the underlying
clauses. The translation is fairly straightforward, the gain in clarity signi -
cant. However, some Prolog implementations, especially on small com-
puters, do not support grammar rule notation. Even then it seems worth-
while to write a preprocessor in Prolog (we shall describe such a
preprocessor in Section 7.4.4).
The counterpart of a parsing program, written down as a collection of
grammar rules, will be called a metamorphosis grammar, or grammar for
short. Here is the grammar of lists, corresponding to the program (3.7).
list -> [ (, ')' ].
list -> [ '(' ], items, [ ') ].
items > item.
items -> item, [ ,' ], items.
item > atom.
item > list.
atom > letter.
atom > letter, atom.
letter > [ a ].

letter > [ z ].
This is the name invented by Colmerauer (I975. I978). The name de nite clause
grammars was later introduced by Pereira and Warren (I930) for metamorphosis grammars
in nonnal fonn (as de ned by Colmerauer).
 3.3. Parameters of Non-Terminal Symbols 69

The procedure terminal need not be explicitly given (it ought to be pro-
vided by the implementation).
This grammar deserves its name. It is best understood independently
of the Prolog program it has been used to conceal. Every rule re ects the
consist of relationship between a whole and its constituents, exactly as
the original BNF grammar does. However, it should be remembered that
the grammar is also a program in disguise, and is executable immediately,
without any additional effort on the programmer's part!
Parsing can be initiated in two ways. First, we can simply call one of
the underlying procedures, e.g.

:- Iist( '('.a.','.(.b.i.g.','.o.x.)'.')'.[], [] ).

Second, we can use the built-in procedure phrase with two parameters:
the nonterminal symbol and the sequence of terminal symbols (which is
supposed to be an instance of the nonterminal). For example:
:- ph|ase( list, '(.a.,'.(.b.i.g.,.o.x.).)'.[] ).
It should be pointed out that the rst way brings out the routine informa-
tion we just managed to hide. On the other hand, the second way is less
exible, e.g. we cannot use phrase to perform calls such as (3.8).

3.3. PARAMETERS OF NON-TERMINAL

SYMBOLS

Grammars of the kind described so far are of little practical use. We

seldom parse anythingjust to accept or reject it. More often than not, we
need to compute the representation of its structure or to transform it
somehow, and we must do this while accepting the input. The representa-
tion of the structure will be built step by step, with the terminal symbols
taken into account in succession.
We shall give an example. Suppose we want to build a parse treea
Prolog termfor every valid sequence of terminal symbols that consti-
tute a list; see Fig. 3.3. To this end, we shall give each of the procedures in
(3.7) an additional parameter to hold the representation (of a structure) to
be constructed upon exit from the procedure. We must not meddle with
input and output parameters: their role remains the same as before. Here
is the program.
Iist( Iist( (', ') ), K, M ) :-
terminal( '(, K, L ), terminal( ), L, M ).
70 3 Metamorphosis Grammars

Iist( Iist( (, ITEMS, ') ), K, N ) :-

terminal( (', K, L ), items( ITEMS, L, M ),
terminal( )', M, N ).
items( items( ITEM ), K, L ) :- item( ITEM, K, L ).
items( items( ITEM, ',', ITEMS ), K, N ) :-
item( ITEM, K, L ), terminal( L, M ),
items( ITEMS, M, N ).
item( item( ATOM ), K, L ) :- atom( ATOM, K, L ).
item( item( LIST ), K, L ) :- Iist( LIST, K, L ).
atom( atom( LETTER ), K, L ) :- letter( LETTER, K, L ).
atom( atom( LETTER, ATOM ), K, M ) :-
letter( LETTER, K, L ), atom( ATOM, L, M ).
letter( letter( a ), K, L ) :- terminal( a, K, L ).

letter( letter( z ), K, L ) :- terminal( z, K, L ).

Again, we shall suppress the routine information, i.e. leave out the
input and output parameters. The resulting grammar will be as follows:

Iist( Iist( '(. )' ) )-> I '('. )' ]-

list( Iist( '(, ITEMS, ) ) ) ->
[ ( ], items( ITEMS ), [ ) ].
items( items( ITEM ) )> item( ITEM ).
items( items( ITEM, ',', ITEMS ) ) >
item( ITEM ), [ ], items( ITEMS ).
item( item( ATOM ) ) > atom( ATOM ).
item( item( LIST ) )> Iist( LIST ).
atom( atom( LETTER ) ) > letter( LETTER ).
atom( atom( LETTER, ATOM ) ) > letter( LETTER ),
atom( ATOM ).
letter( letter( a ) ) > [ a ].

letter( letter( z ) ) -> [ z ].

To compute the parse tree of Fig. 3.3, call:

:- phrase( Iist( T ), '('.a.,'.'(.b.i.g.','.o.x.')'.)'.[] ).

The conciseness and power of metamorphosis grammars can hardly

be appreciated in this tiny example. We shall show a grammar that de-
scribes (and parses) sequences of statements of a simple programming
language. The admissible statements are: assignment, if-then-else- ,
 3.3. Parameters of Non-Terminal Symbols 71

at

/\ /\ while

. . .../
\:"3/B
1 /\ ll i 1

./\. /\.
/\.
/\ /\
i 1 i
FIG. 3.4 An abstract syntax tree.
1

while-do-od, and skip. The sequencing operator is the semicolon. The

condition is either an arithmetic relation (= or <) or a relation negated.
The intended meaning of a sequence of statements is the term that
shows its structure. We shall not go into details; instead, we shall give an
example which ought to explain the idea. Given the (one-element) se-
quence of statements:
ifn < 0 then skip else
i:= 0;
while not n < (i + l)1-(i + I) do
i:= i +1
od

we should obtain the abstract syntax tree (a Prolog term):

if( lt( n, 0 ), skip, seq( assign( i, 0 ),
(3.9) while( not( lt( n, 1='( +'( i, 1 ), '+( i, 1 ) ) ) ),
35518111 i. '+'( 1. I )))))
The same tree is shown in Fig. 3.4.

Both parts of this example, here and in Section 3.4.1, are modelled on the illustration
in Colmerauer's original paper (I975).
72 3 Metamorphosis Grammars

Terminal symbols of our grammar are tokens (lexical units of the

language), e.g. if, n, + , (. Variables and expressions are intentionally left
unde ned: we want to avoid too many details. A grammar for expressions
will be discussed in Section 3.5.2. The following ten rules take care of the
rest of language constructions.

statements( S )> statement( S ).

statements( seq( S, OtherS ) ) >
statement( S ), [ ; ], statements( OtherS ).
statement( assign( V, E ) ) >
variable( V ), [ := ], expression( E ).
statement( if( C, S1, S2 ) )>
[ if], condition( C ), [ then ], statements( SI ),
[ else ], statements( S2 ), [ ].
statement( while( C, S ) )>
[ while ], condition( C ), [ do ], statements( S ), [ od ].
statement( skip )> [ skip ].
condition( R ) > relation( R ).
condition( not( R ) ) -> [ not ], relation( R ).
relation( eq( El, E2 ) )>
expression( El ), [ = ], expression( E2 ).
relation( lt( El, E2 ) )>
expression( El ), [ <' ], expression( E2 ).
This grammar would probably be activated by calls such as
read_a_list..of_tokens( LisT ),
phrase( statements( Structure ), LisT)

which analyse LisT and instantiate Structure appropriately, or fail if LisT

is not a valid sequence of statements. Another possibility (not always
practical, though) is to buildsynthesize, if you prefera list of Tokens
starting from a given structure:

take_a_ structure( S ), phrase( statements( S ), Tokens )

Here, Tokens will be instantiated if only S is a proper structure. The

grammar establishes one-one correspondence between structures and
lists of tokens, and provides transformation both ways.
A more realistic example of synthesis based on a metamorphosis
grammar will be given in the next section. Here we only observe that in
both cases (analysis and synthesis) similar computations ensue. They
differ because, on analysis, the sequence of terminal symbols controls
 3.4. Extensions 73

the computation (i.e. determines the choice of rules) whereas, on synthe-

sis, it is controlled by the initial non-terminal symbol's parameter.

3.4. EXTENSIONS

3.4.1. Conditions

Grammar rules described so far correspond to clauses in which every

call manipulates the sequence of terminal symbols, i.e. every call has an
input and an output parameter. Other calls could be inserted in between
without affecting the transfer of terminal symbols. The question is: Would
it be useful, and how could it be interpreted?
As a simple possibility, consider the cut in the rst clause of list:
Iist( Iist( '(, )' ), K, M ) :-
tenninal( (', K, L ), terminal( )', L, M ), !.
The cut tums the computation based on the list procedure into a deter-
ministic process: it handles either the empty list or non-empty lists. It
does not matter when we want to recognize a list. However, it is now
impossible to generate lists. The command
:- Iist( L, T, [] ), write( L ), write( T), nL, fail.
will only write one instance of L and T, namely
Iist( '(', ') ) and '('.)'.[]
The gain from the cut is small in this case, anyway. Cuts would be of
much greater use, say, in the program that parses statements (see the
previous section), where long and deep computations may occur.
Another example: suppose we want to change the program for pars-
ing lists so that for an atom it produces a Prolog atom instead of a parse
tree, e.g. retums
Iist( (', items( item( big ). items( item( ox ) ) ), ') )
for the list (big,ox). One way to do so is to make the procedure for atoms
retum a Prolog list of letters, and apply the built-in procedure pname (see
Section 5.10) to this list
item( item( ATOM ), K, L ) :-
atom( LETTERS, K, L ), pname( ATOM, LETTERS ).
item( item( LIST), K, L ) :- Iist( LIST, K, L ).
atom( LETTER.[], K, L ) :- letter( LETTER, K, L ).
74 3 Metamorphosis Grammars

atom( LETTER.LETTERS, K, M ) :-
letter( LETTER, K, L ), atom( LETTERS, L, M ).
letter( a, K, L ) :- terminal( a, K, L ).

letter( z, K, L ) :- terminal( z, K, L ).
One nal example: in the program above we shall replace the 26
clauses that de ne letters by a single clause:
letter( LETTER, K, L ) :-
terminal( LETTER, K, L ), isletter( LETTER ).
with isletter de ned, say, as
isletter( LETT ) :- a @=< LETT, LETT @=< z.
This new clause can be used as follows:
letter( Lett, x.').).[], Tail ).
terminal( Lett, x.).')'.[], Tail ), isletter( Lett ).
Lett 1- x, Tail 1- ').).[]
isletter( x ).
etc.
The variable in the call on terminal matches every terminal symbol. If
the terminal symbol is not a letter, a call on isletter will fail and a letter
will not be recognized. We call such terminal symbols variable terminals:
the rst (still unprocessed) symbol is selected and is then either accepted
or rejected, e.g. according to the result of a test such as isletter.
Extra calls that do not comprise input and output parameters have
been known as conditions, but the name is slightly misleading. Only in the
last example isletter(Lett) can be interpreted as a condition: the clause
will only be applied if isletter succeeds. The call on pname in the second
example is rather an action performed on the parameters of non-terminal
symbols. Finally, the cut can be reasonably interpreted exactly as in any
other clause, as pragmatic information on the future use of the clause.
Conditions in metamorphosis grammars are enclosed in curly brack-
ets, so that they will not be confused with terminal and non-terminal
symbols. Examples:
Iist( Iist( '('. ')' ) )> I '(. ) 1. 1!}-
item( item( ATOM ) )-> atom( LETTERS ),
{ pname( ATOM, LETTERS ) }.
letter( LETTER )-+ [ LETTER 1, { isletter( LETTER ) }.
As an exception, the cut need not be placed within curly brackets, e.g.
Iist( Iist( '('. ') ) )"* I '('. ')' I. 1-
 3.4. Extensions 75

Contiguous conditions can be combined in a single pair of brackets, and in

general a condition can also contain altematives conjoined by semicolons,
e.g.
alphanum( Char )-> [ Char ], { isletter( Char ) ; isdigit( Char ) }.
We shall now present a small fragment of a metamorphosis grammar,
meant primarily for synthesis (but applicable both ways, although not
without reservations). We want to take a structure computed by the gram-
mar for statements (see the previous section) and produce its translation
into a machine-oriented symbolic language. We shall only give a hint of
the target language by showing schematic translations of while(C, S) and
if(C, Sl,__S2). _ _
Let C and S be the translations of C and S. The evaluation of C sets a
ag used implicitly by a conditional jump instruction. Let 8 I , 2 be unique
labels. The translation of while(C, S) will be
label( fl )
not(C)
j_umpiftrue( 2 )
S
jump( 1 )
label( 2 )
The translation of if(C, S1, S2) will be
C
jumpiftrue( 1 )
S2
jump( 2 )
libel( 1 )
SI
label( 2 )
The code generator can be written as a grammar of the target
language. By way of explanation, we shall show three of the rules that
belong to the uppermost level of the de nition:
code( seq( S, OtherS ) ) -r code( S ), code( OtherS ).
code( while( C, S ) ) >
{ newlabel( Ll ) }, [ label( Ll ) ], codecond( not( C ) ),
{ newlabel( L2 ) }, [ jumpiftrue( L2 ) ], code( S ),
[jump( Ll ), label( L2 ) ].
code( skip ) -> [].
The action newlabel can generate a new, unique label. The de nition
of codecond will be given below. The third rule illustrates a new feature of
76 3 Metamorphosis Grammars

grammar rules. If the righthand side contains no terminal and non-termi-

nal symbols, nothing will be produced during synthesis and nothing will
be chopped off during analysis. The underlying clause is
code( skip, K, K ).
Try to trace the execution of
:- code( seq( skip, skip ). Translation, [] ).
Assuming that coderel de nes the grammar of codes for relations eq
and It, the de nition of codecond can be as follows:
codecond( not( not( C ) ) ) > codecond( C ).
codecond( not( Rel ) ) -> coderel( Rel ). I revert(._) ].
codecond( Rel )> coderel( Rel ).
where revert is an instruction of the target language that resets the
condition ag.
The example would be completed after specifying the translation of
expressions and of assignments, in particular the handling of variables.
The code generator together with the grammar of statements might
constitute the core of a simple compiler. Its overall structure might be:
compile :- read_tokens( Token_list ),
parse( Token_list, Syntax_tree ),
generate_code( Syntax_tree, Object_ code ),
write_code( Object_code ).
with parse and generate-code de ned as
parse( T, S ) :- phrase( statements( S ), T ).
generate_code( S, O ) :- phrase( code( S ), O ).
The procedure read_tokens, reading the source program in and perform-
ing lexical analysis, might also be (partly) written as a metamorphosis
grammarsee Colmerauer (1975, 1978).

3.4.2. Context

Another feature of grammar rules in Prolog is a mechanism for modi-

fying the sequence of terminal symbols during the computation. In gen-
eral, this would require explicit manipulations on input and output param-
eters, but such general mechanisms seem only necessary in natural
language processing (an important application of Prolog). A very re-
stricted mechanism, so-called context grammar rules, is quite suf cient,
though, in most of the other applications.
 3.4. Extensions 77

In a context grammar rule, the lefthand side is supplemented by a

so-called context-: terminal symbols, preceding the arrow >. For ex-
ample:
otherst( S, S ). I Delim ]-> [ Delim ], { stsdelim( Delim ) }.
do, [ not ]-> dont.
The output parameter in the head of an underlying clause is appended to
the context. As clauses, the above rules are:
otherst( S, S, K, Delim.L ) :-
terminal( Delim, K, L ), stsdelim( Delim ).
do( K, 'not.L ) :- dont( K, L).
The rst rule can be interpreted without resorting to the corresponding
clause; we shall give the interpretation below. The second rule, however,
can only be explained in terms of manipulations on sequences of terminal
symbols: a new terminal symbol appears after recognizing an instance of
dont, and only then is an instance of do recognized as well. We shall
elaborate on this example a little, too.
First we come back to the grammar for statements. In its present
shape it performs rather poorly on incorrect inputs. It fails without giving
any message or diagnostics. We shall try to improve the de nition of
statements, leaving the other rules as an exercise. We observe that a
statement (other than the last) may be delimited by a semicolon (it indi-
cates that there are other statements in this sequence), by else, , or od.
Other delimiters are erroneous. In case of errors, no meaningful structure
may be found for the whole sequence of statements, but we elect to
continue the analysis, after skipping a portion of input up to the nearest
semicolon. Here are some rules of a grammar that implements these
ideas.
statements( Sts ) -> statement( St ), otherst( St, Sts ).
otherst( Stl, seq( Stl, Sts ) )>
[ ';' ], statement( St2 ), otherst( St2, Sts ).
otherst( St, St ), [ Delim ] ->
[ Delim ], { stsdelim( Delim ) }.
otherst( _, _ ) > [ T ], erroneous( T ).
otherst( St, St )-> []. % this for the last statement
erroneous( T ) -> { write( bad( T ) ), nl }, skipped.
skipped. l ':' 1> I ;' 1-
Readers familiar with context-sensitive grammars will notice that neither rule is a
proper context-sensitive rule. Even if we disregard parameters and conditions, the rules will
only belong to Chomskian type 0.
78 3 Metamorphosis Grammars

skipped -> [_ ], skipped.

skipped -> []. % if we are skipping the last statement
stsdelim( else ). stsdelim( ). stsdelim( od ).
The context rule can be interpreted in the following manner: the
remainder of a sequence of statements is empty if we have encountered a
proper delimiter; this delimiter is retained. Notice that we have actually
effected one-item lookahead on a list of terminal symbols. In general,
we can have lookahead for any xed number of terminal symbols, for
example
p,[Tl,T2]->[TI,T2],{test(TI,T2)}.
This translates into
p( K, Tl.T2.M ) =-
terminal( rt, K, 1. ), terminal( T2, 1., M ), test(T1, T2).
We can use p to make the test; e.g. in
a -> p, b, c.
p consumes no input, so that the rule is structurally equivalent to
a > b, c.
but it will only be applied if two leftmost terminal symbols of the current
sequence pass the test.
The second example is a very simpli ed little grammar that recog-
nizes auxiliary do not, don't, does not, doesn't. This particular
problem can easily be solved differently; the way we have chosen is
intended as an illustration of context grammar rules:
aux > do, [ not ].
do, [ not ] -> dont.
do -> [ do ].
do -> [ does ].
dont > [ don"t ]. %i.e. don't
dont > [ 'doesnt ]. %i.e. doesn't
The following computation should explain how this grammar is used:
aux( doesnt.like.it.[], Tail ).
do( doesnt'.like.it.[], Tl ), terminal( not, Tl, Tail ).
TI 1- not'.L
dont( doesnt'.like.it.[], L ),
terminal( not, not.L, Tail ).
 3.4. Extensions 79

terminal( 'doesnt', doesnt.like.it.[], L ),

terminal( not, not.L, Tail ).
L 1- like.it.[]
terminal( not, 'not'.like.it.[], Tail ).
Tail 1- like.it.[]
success
Our last example is a small grammar that discards leading zeroes from
an integer represented as a list of digits:
zeroes, [ D ]-> [ 0 ], zeroes, [ D ], { digit( D ) }.
zeroes -> [].
You may wish to trace the execution of the directives
:- zeroes( 0.3.[], Tail ).
:- zeroes( 0.0.[], Tail ).

3.4.3. Alternatives

Two or more grammar rules with the same lefthand side (including
context and parameters of the non-terminal symbol) can be combined into
a single rule with the common lefthand side and with the righthand side
taking the form of alternativesa sequence of original righthand sides
separated by semicolons. For example:
list -> I '('. ')' I ; I '( l.it1=mS. I ') ]-
items > item ; item, [ , ], items.
item > atom ; list.
atom > letter ; letter, atom.
letter > [ L ], { isletter( L ) }.
Notice howat lastwe managed to come back rather closely to the
original BNF grammar (3.1).
The translation of a rule with an alternative into an underlying clause
is straightforward. One example should be su icient:
items( K, N ) :- item( K, N ) ;
item( K, L), terminal( ',', L, M ), items( M, N ).
The notation with alternatives is, strictly speaking, a convenience
rather than a real extension, andlike altematives in ordinary clauses
(see Section 1.3.7)it can sometimes adversely affect the grammars
readability.
SO 3 Metamorphosis Grammars

3.4.4. Syntax of Grammar Rules: Summary

We shall now give a metamorphosis grammar that describes full syn-

tax of grammar rules supported by Prolog-10. The principles of mapping
rules onto underlying clauses have been discussed at length in the pre-
vious sections, so we choose not to overburden the grammar with param-
eters that would take care of the translation. However, we encourage you to
try and augment the grammar along these lines. A hint: most of the non-
terminal symbols should be given three parameters, two variables (to
construct an input and output parameter) and a term (to hold thepartial
translation). For example:
grammar_rule( ( Tr_of_left :- Tr_of_right ), In_var, Out_var )
> lefthand_side( Tr_of_left, In_var, Out_var ), [ > ],
righthand_side( Tr_of_right, In_var, Out_var ), [ ].
rule_items( ( Tr_of_item, Tr_of_items ), Curr_in_var, Out_var )
> rule_item( Tr_of_item, Curr_in_var, Mid_var ), [ ],
rule_items( Tr_of_items, Mid_var, Out_var ).
In the actual translation we might eliminate the calls on the procedure
terminal. Since terminal( T, K, L) means that K = T.L, we can substitute
in advance T.L for K elsewhere in the clause. For example, in the clause
Iist( K, N ) :-
terminal( (, K, L ), items( L, M ), terminal( ), M, N ).
we have K = ('.L and M = )'.N, and after replacing K and M we obtain
Iist( (.L, N ) :- items( L, ').N ).
This is, in fact, what is done in many implementations (see, e.g., Section
7.4.9). As we have executed both calls on terminal beforehand, every
computation started by a call on list will be at least two steps shorter.
Here are some other examples of such an improved translation of gram-
mar rules:
letter( Lett, Lett.L, L ) :- isletter( Lett ).
p( Tl.T2.M, Tl.T2.M ) :- test( Tl, T2).
zeroes( 0.L, D.N ) :- zeroes( L, D.N ), digit( D ).
We shall now present the grammar without parameters (it is, really,
equivalent to a BNF de nition).
grammar_rule > lefthand_side, [ >' ],
righthand_side, [ ].
lefthand_side > nonterminal, context.
context > terminals ; ll.
 3.5. Programming Hints 31

righthand- side > altematives.

altematives > altemative ;
altemative, [ ; ], altematives.
altemative -1- 1 ll I : rule_items.
rule-items > rule_item ; rule_item, [ ], rule_items.
rule..item > nontenninal ; terminals ; condition ; [ ! ] ;
[ ( ] , altematives , [ ) ].
nonterminal > name ;
name, [ ( ], list_of_terms, [ ) ].
terminals > [ [ ], list_of_tenns, [ '1 ] ; string.
condition > [ { ], procedure_body, [ } ].
list_of.terms > term ; term, [ , ], list_of_tenns.
De nitions of name, term, string and procedure._body are left as an
exercise.
It should be noted that the original appearance of grammar rules in the
Marseilles interpreter of Prolog I (Roussel 1975) was slightly different. In
particular, no altematives were allowed, and terminal symbols and condi-
tions could not be combined. Just to give the avour of it, we shall rewrite
in Marseilles syntax some of the grammar rules for statements (Section
3.4.2).
:STATEMENTS( 1-STS ) = = :STATEMENT( 1-ST )
:OTHERST( -I=ST, -FSTS ).
:OTHERST( 1-STI, SEQ( 1-STI, 1-STS ) ) ==
#; :STATEMENT( 1-ST2 ) :OTHERST( -I=ST2, 1-STS ).
:OTHERST( 1-ST, 1-ST ) #-FDELIM ==
#1-DELIM -STSDELlM( 1-DELIM ).
:OTHERST( *DUMMYl, 1-DUMMY2 ) ==
#*T :ERRONEOUS( -FT ).
:OTHERST( -FST, -FST ) == .
1-THIS FOR THE LAST STATEMENT.

3.5. PROGRAMMING HINTS

3.5.1. Ef ciency Considerations

Metamorphosis grammars correspond to Prolog programs which im-

plement a very general parsing strategy: nondeterministic top-down pars-
ing with backtracking (Aho and Ullman 1977; Gries 1971). The potential
cost of this strategy is exponential. This is the disadvantage of the gener-
82 3 Metamorphosis Grammars

ality and ease of programming with metamorphosis grammars. Well-

known parsing algorithms for restricted classes of context-free grammars
can be quite conveniently programmed in Prolog without metamorphosis
grammars. See for example the operator precedence parser described in
Section 7.4.3 and Appendix A.3. However, this requires explicit handling
of the parsing stack, attributes etc., while metamorphosis grammars by
themselves are as powerful as attribute grammars (Knuth 1968) or two-
level grammars (van Wijngaarden 1976)see the discussion in (Pereira
and Warren 1980). Parameters and conditions/actions make it possible to
construct an intuitively appealing, concise and readable metamorphosis
grammar of any existing programming language (and of reasonable sub-
sets of natural languages), capturing semantics as well as syntaxsee e.g.
(Moss 1979). At the same time, such a grammar can usually be used as a
translator of this language, without additional e ort on the part of the
programmer, but there is often a certain price to be paid in ef ciency.
One source of inef ciency is repetition. Consider two rules from the
grammar for statements (Section 3.3):
relation( eq( El, E2 ) )>
expression( El ), [ =' ], expression( E2 ).
relation( lt( El, E2 ) ) >
expression( E1 ), [ < ], expression( E2 ).
If a given relation is not an equality, we recognize this state of affairs only
after parsing the rst expression and failing to nd an equals sign. We
abandon the rule and choose the next but then we must once more parse
the rst expression (which may be quite large). The problem remains if we
change the order of the rules.
To avoid this inef ciency, we may apply factorlzationthe technique
already used in Section 3.4.2:
relation( R ) > expression( El ), op_and_expr( E1, R ).
op_and_expr( E1, eq( E1, E2 ) )> [ = ], expression( E2 ).
op_and_expr( El, lt( E1, E2 ) )> [ < ], expression( E2 ).
Another solution is to combine the original rules into a single rule by
replacing the terminal symbols with a variable terminal, and adding a
suitable condition:
relation( R )> expression( El ), [ Op ],
{ makestruct( Op, E1, E2, R ) },
expression( E2 ).
makestruct( =, E1, E2, eq( E1, E2 ) ).
makestruct( <, E1, E2, lt( El. E2 ) ).
 3.5. Programming Hints 33

Notice the position of the condition: if we placed it at the end of the rule,
we would run the risk of discovering an improper instance of Op only
after parsing the whole input, say,
(A + b/2)1=c blah_blah 21-( n (x + y )/4)
In its present position the condition fails as soon as it sees an invalid
operator.
Both improvements of the original grammar eliminate possible repeti-
tions. Both, though, seem to decrease the readability and elegance of the
original solution, and we recommend that they be applied (if at all neces-
sary) only in the late stages of program debugging.

3.5.2. Elimination of Left Recursion

We shall now discuss a problem which frequently arises with inexpert

use of metamorphosis grammars. As an example, we shall consider the
task of writing a workable grammar of simple arithmetic expressions (see
Section 3.3). Here is the de nition in BNF (for simplicity, we limit our-
selves to two operators only):
< expression > ::= < add_expr > I
< expression > + < add_expr > |
< expression > - < add_expr >
< add_expr > ::= < constant >
We now give an obvious transcription of this de nition into a meta-
morphosis grammar. Parameters are used to build the structure of a given
expressi0nsee (3.9).
expression( E )-> add_expr( E ).
expression( E1 + E2 )>
expression( E1 ), [ + ], add_expr( E2 ).
expression( E1 E2 ) >
expression( El ), [ ' ], add_expr( E2 ).
The de nition of add_expr will be left out (it can be simply an integer
constant).
Unfortunately, this grammar_as a programis not only inef cient
but also incorrect. It goes into in nite (left) recursion whenever we give it
an expression that contains a minus. Try to analyse the expression 2 3
+ 5 (represented by 2.-'.3.+.5.[]).
B4 3 Metamorphosis Grammars

At rst sight, it seems we can improve the situation by applying one of

the techniques shown in the previous section. For example, the second
technique gives the following rules:
expression( E )> add_expr( E ).
expression( E ) > expression( E1 ), [ Op ],
{ makesum( Op, E1, E2, E ) },
add_expr( E2 ).
makesum( +, E1, E2, E1 + E2 ).
makesum( -, E1, E2, E1 E2 ).
Now con'ect expressions will be parsed successfully, although an expres-
sion composed of n add-expressions will require n I backtracks before
reaching the solution. But the grammar will still fall into in nite recursion
on any incorrect input (you may wish to check this on 2. + .[]). This means that
it is of no practical value. As in all top-down parsing methods, we must
eliminate left recursion to avoid trouble.
Suppose we reverse nonterminal symbols in the recursive rules in
(3.10):
expression( E1 + E2 )> add_expr( El ), [ + ], expression( E2 ).
expression( El E2 ) > add_expr( E1 ), [ ], expression( E2 ).
Now incorrect input causes the grammar to fail (without any error mes-
sage, but this can be xed). However, this grammar interprets operators
as right-associative. The instantiation of its parameter for the expression
2 3 + 5 will be (2, +(3, 5)) rather than +((2, 3), 5). Here is a possi-
ble solution to this new problem:
expression( E )> add_expr( El ), rest_of_expression( El, E ).
rest_of_expression( El, E ) >
[ + ], add_expr( E2 ), rest_of_expression( E1 + E2, E ).
rest_of_expression( El, E )>
[ '' ], add_expr( E2 ), rest_of_expression( El E2, E ).
rest_of_expression( E1, E1 )> [].
When we parse an expression, the parameter is initially uninstan-
tiated. It is passed unchanged and instantiated after reaching the end of
the expression. (In the terminology of attribute grammars this is a synthe-
sized attribute.) The nal structure is accumulated step by step. For ex-
ample, during the parsing of the expression 2 3 + 4 - 5, rest_of_ex-
pression will be activated four times, with 2, 2 3, (2 3) + 4 and ((2
3) + 4) - 5 as the rst parameter. (This parameter is an inherited
attribute.) Eventually the third rule will be chosen and E instantiated to
((23)+4)-5.
 3.5. Programming Hints 35

We shall now present a grammar for expressions, complete with error

handling, that ts the grammar for statements (see Sections 3.3 and 3.4.2).
The de nition of erroneous was given in Section 3.4.2.

expression( E )-> add_expr( El ), rest_of_expression( El, E ).

rest_of_expression( E1, E )>
[ + ], add_expr( E2 ), rest_of_expression( E1 + E2, E ).
rest_of_expression( El, E ) ->
[ '' ], add_expr( E2 ), rest_of_expression( El E2, E ).
rest_of_expression( E1, El ), [ Termin ]->
[ Termin ], { expr_termin( Termin ) }.
rest_of_expression( _, _ ) > [ T ], erroneous( T ).
rest_of_expression( E1, E1 )> [].
expr_termin( then ). expr_termin( else ).
expr_termin( do ). expr_termin( od ).
expr_termin( ; ). expr_termin( ).
add_expr( E )> mult_expr( E1 ), rest_of_add_expr( E1, E ).
rest_of_add_expr( El, E ) >
[ 1= ], mult_expr( E2 ), rest_of_add_expr( E1*E2, E ).
rest_of_add_expr( E1, E )>
[ I ], mult_expr( E2 ), rest_of_add_expr( E1/E2, E ).
rest_of_add_expr( E1, E1 ), [ Termin ] >
[ Termin ], { add_expr_termin( Termin ) }.
rest_of_add_expr( _ , _ ) -> [ T ], erroneous( T ).
rest_of_add_expr( El, E1 )> [].
add_expr_termin( Termin ) :- expr_termin( Termin ).
add_expr_termin( + ).
add_expr_termin( ' ).
mult-expr( E )> variable( E ).
mult_expr( E ) > constant( E ).
mult_expr( E )-> [ ( ], expression( E ), [ ) ].

To make the grammar really complete, we should also de ne vari-

ables and constants. We choose not to do it, because variables require
symbol table handling-we shall discuss it in Section 4.2.2.
The techniques described above are only necessary if we want to
perform analysis with a metamorphosis grammar. Even more: the trans-
formed grammar is not good for synthesis, i.e. for constructing the se-
quence of terminal symbols given a (correct!) structure. Speci cally, for
synthesizing expressions, the only reasonable solution would be the origi-
nal grammar (3.10).
 7 I j

4 SIMPLE PROGRAIVIMING
TECHNIQUES
i 7 7171

4.1. INTRODUCTION

Programming in Prolog differs from programming in classical (Pascal-

style) languages primarily at the level of individual procedures. The larger
the program, the more suitable the general recommendations of program-
ming methodology. The advantages of systematic top-down design of
programs, modularity, clean interfaces, etc., are certainly independent of
the programming language used. Design and coding techniques speci c to
Prolog are due to its logical origin.
In Section 1.3.4 and Chapter 2 we discussed logicalstaticinter-
pretation of procedures. This interpretation makes it possible to design
programs without paying attentionat least initiallyto how the compu-
tation will proceed. One only needs to indicate what will be computed.
Kowalski (I974, 1979a) coined an equation,
Algorithm = Logic + Control
which helps clarify the distinctive feature of logic programming. It is
maintained that logic programming relieves the programmer of the burden
of specifying control information for her program. One would like to say:
completely relieves, but unfortunately (at least in Prolog) this is not the
case. Many useful built-in procedures, such as the cut, input/output and
program modi cation procedures (assert, etc.; see Section 5.11), cannot
be interpreted statically. As a result, a practical program may not usually
be designed without paying regard to control information.
' At least on a conceptual level: most existing Prolog implementations do not support it
explicitly.
87
88 4 Simple Programming Techniques

In Section 4.3 we shall brie y consider the advantages and disadvan-

tages of some side-effects in Prolog; we shall also present several simple
tricks that help increase the ef ciency of Prolog programs (especially their
space requirements) in many existing implementations. Earlier, in Section
4.2, we shall give a few examples of Prolog implementation of commonly
used data structures, in particular binary trees and linear lists. We shall
show basic operations on those structures and a few typical applications.
Section 4.4 contains small examples of program design.

4.2. EXAMPLES OF DATA STRUCTURES

We have chosen unbalanced binary search trees (BSTs) and one-way

linear lists as an illustration of methods of implementing recursive data
structures in Prolog. We assume you are familiar with basic de nitions
and algorithms; a detailed, though rather elementary presentation can be
found, for example, in Wirth (1976) or Sedgewick (I983). Here, we shall
refer only to common intuitions, and we shall concentrate on problems
speci c to Prolog. V
We shall also brie y discuss representation of data structures by
clausesin particular, Prolog counterparts of anays.

4.2.1. Simple Trees and Lists

Terms can usually be regarded as trees: the main functor labels the
root, subtrees correspond to arguments. This is slightly imprecise, be-
cause multiple occurrences of variables represent more general struc-
turesdirected acyclic graphs (DAGs). However, the term f(A, A)
which should be depicted as
t

can be thought of as

A A
 4.2. Examples of Data Structures 39

We must only remember that the two subtrees will remain identical, so
instantiating variables in one will affect the other. Another dif culty is
that it is possible to compute tenns which are not even DAGs, and which
should therefore be regarded as corresponding to in nite trees (see Sec-
tion l.2.3). All the same, an ordinary tree is a good intuition of the (gen-
eral) term.
Terms are a convenient and concise representation of trees with irreg-
ular structure, where the information in the nodes detennines both the
shape of the tree and the repertory of applicable operations. The abstract
syntax tree of Fig. 3.3, Section 3.3, is a typical example. However, pro-
grams that manipulate such irregular structures are usually problem-de-
pendent, in that every principal functor (i.e. every type of node may
require di 'erent computations).
There are other situations, typi ed by binary search trees, when we
need a more uniform representation, because we use trees for contents
rather than for structure. Suppose we represent the BST of Fig. 4.1 as the
term

few(people(many(languages), speak))

Even if we disregard the ambiguity (is languages the left or right de-
scendant of many?), main functors and their arguments must be iso-
lated, that is, we must use the built-in procedure = .. (univ; see Section
5.10). To modify the tree, e.g. by adding a node, we must rebuild it com-
pletely, also using univ. This is not only inelegant, but ine icient as well
(but see Section 4.2.6 for a discussion of such techniques).
We shall therefore represent empty binary trees by the atom

fevv\

people

../ \...
FIG. 4.1 A binary search tree.
90 4 Simple Programming Techniques

fl/...\.
./.\. .%l.\.
nil
\
languages nil
FIG. 4.2 A representation of the tree of Fig. 4.1.

and nonempty trees by three-argument terms

t( Left_subtree, Node_info, Right_subtree )
For example, the BST of Fig. 4.1 will be represented by the term
t( nil, few, t( t( t( nil, languages, nil ), many, nil ),
people, t( nil, speak, nil ) ) )
The term can be drawn as a tree (see Fig. 4.2). This method of represent-
ing binary trees can be readily adapted to trees of a different xed degree.
e.g. non-empty temary trees can be represented by four-argument terms
tt( Node_info, Left_subt, Middle_subt, Right_subt )
In Fig. 4.2 the contents of each node is only a key, but of course in
practical applications nodes contain other information as well. The tree
shown in Fig. 4.3 holds names and phone numbers of several persons-
names are keys in lexicographic order. We use a nonassociative in x
functor : to separate keys from other data.
An inorder traversal of a BST visits the nodes in increasing order,
according to the ordering relation in the set of keys. For example, the
 4.2. Examples of Data Structures 91

'1:

t thompson 2432

\7
nrl
\
rncbr|de:l7Bl
FIG.4.3 Another BST.
nil

following procedure can be used to write out name-phone pairs, sorted

alphabetically by names:

uritc_sorted( nil ).

ur1te_sorted( tt LeFt_subtreev Node_infov Risht_subtree ) ) =-

urite_sorted( Left_subtree)v

write( N0de_inF0)v nlr

write_sortud( Risht_subtree).

In this procedure, we need not test the actual ordering of nodes; this
would not be the case if we wanted, say, to locate a node in a tree. Let the
call
precedes( Nodel, Node2 )
succeed iff Nodel comes before Node2. For our name-number pairs the
procedure can be de ned simply as
precedes( Namel :_, Name2 :_ ) :- Namel Gr< Name2.
It is reasonable to expect that nodes are correctly built, e.g. that each
key is a name, and other information a number. A good place to check this
would be a procedure for inserting a node into a tree:
92 4 Simple Programming Techniques

insert( Node, Tree, Newtree ) :-

con'ect( Node ), !, ins( Node, Tree, Newtree ).
insert( Node, _ , _ ) :- signal_error( Node ).

However, such defensive programming is seldom necessary in practice.

The insertion procedure ins is rather straightforward. We must only
take care to preserve the ordering relation:

Z on emote tree will be replaced by o new leaf

ins( Nader nilv t( nil: Noder nil ) ).

ins( Node: tt Leftv Root: RiSht)r tt Neuleftr Rootr Risht ) ) :-

precedest Node: Root )1 ins( Nodev Leftr Neuleft ).

ins( Nnder t( LOFLr Root: Ri9ht)r t( Left! Root: Neurisht ) ) =-

DPGCGdG5( Root: Node )1 ins( Noder Rishtr Neurisht ).

The procedure fails when it tries to duplicate a key (both calls on precedes
fail). If the keys need not be unique, we must relax one of the tests, e.g.
by changing

precedes(Root, Node)
into
not precedes(Node, Root)

A BST can be built by successive insertions. We shall not discuss

balanced trees. They present problems of their own, which can be solved
by far in the same way as in classical programming languages (see e.g.
Sedgewick 1983) but which can cause memory problems with some Pro-
log implementations. One example is an AVL-tree insertion program (van
Emden 1981, Vasey 1982).
We need some thought to delete a node even from an unbalanced tree.
If either of the subtrees of the deleted node is empty, the other subtree
moves up and replaces the node. For example, deleting adams : _ from
the tree in Fig. 4.3 gives the tree in Fig. 4.4. Suppose now that both
subtrees are nonempty; we shall preserve the ordering if we replace the
deleted node by that with the largest key in the left subtree (or else that
with the smallest key in the right subtree). For example, deleting
thompson : _ in Fig. 4.3 gives the tree in Fig. 4.5.
 4.2. Examples of Data Structures 93

/\ /\
mt tndxme l ml t wmne1M32 mt
FIG. 4.4 The tree of Fig. 4.3 alter deleting adams : _ .

The following procedures implement this algorithm. The second

clause is for symmetry (and for ef ciency) but it is not really necessary.
delt Node1 Lt n111 Node1 Risht )1 Risht ).

de1( Noder t( Left1 Noder nil )1 Left).

delt Node1 tt Left1 Node1 Risht )1 tt Newleftv Leftnoxr Rilht ) ) I-

remove_mox( Left1 Leftmox1 Newleft ).

de1( Noder tt Left1 Rootr Risht )1 tt New1eft1 Root1 Risht ) ) 8-

orecedest Node1 Root )1 de1( Node1 Left1 Newleft ).

delt Node1 t( Left1 Rootr Risht )1 tt Left1 Root1 Newriaht ) ) 8-

precede5( Root1 Node )1 de1( Nodev Riahtr Newriaht ).

Z find and remove the node with the lorsest key

remove_mox( tt LeFt1 Hox1 nil )1 Hoxr Left ).

remove_mox( t( Left1 Rootr Risht )1 Hox1 t( Leftr Root1 Newrilht ) ) I-

remove_mox( Ri9ht1 Hax1 Newrisht )-

Nonnally we would call the procedure del with only the key given.
We might encapsulate such calls:
delete( Key, Oldtree, Newtree ) :-
del( Key : _ , Oldtree, Newtree ).
94 4 Simple Programming Techniques

-E

/\ A\
nil adams=54BB nil nil white:2t.32 nil
FIG. 4.5 The tree of Fig. 4.3 after deleting thompson : _.

The last basic operation on BSTs is the search itself:

search( Node1 tt "1 Node1 _ ) ).

search( Node1 tt Left1 Root1 _ ) ) =-

erecedest Node1 Root )1 search( Node1 Left ).

search( Node1 tt _1 Root1 Risht ) ) =-

nrocedest Root1 Node )1 search( Node1 Risht-).

Again, we can encapsulate typical calls nd information associated

with a given key:
nd( Tree, Key, Data ) :- search( Key : Data, Tree ).
A slightly different method of representing binary trees consists in
using
l( Node )
for leaves, instead of t(nil, Node, nil). However, with this representation
we would have to distinguish empty trees from leaves of non-empty
trees. For example, two more clauses would be necessary in the proce-
dure for tree insertion.
As a very special case, we can consider trees of degree I, that is, lists.
Recall that a widespread convention (introduced in Chapter 1) is to denote
empty lists by the atom
ll
 4.2. Examples of Data Structures 95

and non-empty lists by in x tenns

Head.Tail
The period is used to build trees of degree 2, which are a convenient
representation of lists. It plays the same role as t in our BST example. In
Prolog-10 a special notation has been invented as yet another application
of syntactic sugar. It is very commonly used, even though its advantages
over dot notation are debatable. Instead of Head.Tail we shall write
[ Head I Tail ]
the list a.b.c.Tail will be written as
[ a, b, c | Tail ]
and the list a.b.c.[] as
[ a, b, c ]
To make sure you have mastered this notation, check that [c I [d]] is the
same as [c, d].
We shall remind you of two list-manipulating procedures from Chapter 1.
Membership:
member( Item, [ Item I Tail ] ).
member( Item, [ _ | Tail ] ) :- member( Item, Tail ).
And list concatenation:
append( [], Second, Second ).
append( [ Head I First_tail ], Second, [ Head I Third_tail ] ) :-
append( First_tail, Second, Third_tail ).
Here is another small example of operations on lists. Consider the
following simple-minded sorting algorithm: given a list, put all its mem-
bers in a BST and then apply the procedure write_sorted, de ned above.

sort( List ) 2- buildtraat List1 ni11 Tree )1 write-sortedt Tree

Z 2nd and 3rd ariuoentt the tree built so far1 the final Lroo

buildtroet [J1 Fina1tree1 Finaltroo ).

buildtraat Elton I ILansJ1 Currenttree1 Finaltree ) 2-

insurtt Itenr Currenttree1 Nexttree )1

buildtroet Iteosr NextLroo1 Finaltrea ).

i Sometimes an equivalent notation is used: lHead... Tail]. with written without

blanks.
96 4 Simple Programming Techniques

In Section 4.2.3 we shall de ne a more useful sorting procedure based

on BSTs. It will construct the sorted permutation of a given list.
Just as in other programming languages, lists are used in Prolog pri-
marily to represent sequences and sets. They can also be used in a stand-
ard way to represent trees of unspeci ed degree. For example, the tree of
Fig. 4.6 might be represented by the list
la.lb.[e]].l].Id.lfl.ls]]l
Lists are best utilized when items are processed sequentially from left
to right, or when all processing takes place at the beginning of the list. In
the latter case the list is used as a stack. The basic stack operations, push
and pop, can be easily written in one procedure, e.g.
stack_op( Top, Rest_of_stack, [ Top I Rest_of_ stack ] ).
with the call
stack_op( Newtop, Stack, Newstack )
serving as push, and the call
stack_op( Top, Newstack, Stack )
to execute pop. However, in practice we would rather operate on the stack
implicitly, by using appropriate terms in clause heads. One example is the
procedure reduce (see Section 7.4.3) with old and new stacks as
parameters. The clause
reducet I br( r. '1) 1. t( X 1. br( I, '1) ). idt l ) I 5 1.
ll(lI'(l.X))|3])-
describes an action that consists of four pops followed by one push.
Nonsequential access to a list requires, as might be expected, time
proportional to the lists length. To build a list in linear time, we can

a\

1/ C <1

FIG. 4.6 A non-binary tree.

 4.2. Examples of Data Structures 97

successively push incoming items, but the original sequence will be re-
versed. Alternatively, we can use append to preserve the original order of
items, but this would square the nrnning time. Moreover, each call on
append entails not only a traversal of the entire list, but also creation of its
copy. Strictly speaking, a series of variables is produced and instantiated
to successive tails. When executing the call
append([ Itl, I12], [I13 ], X )
the following instantiations take place:
X 1- [ ltl I Third_tail' ]
Third_tail' 1- [ lt2 I Third_tail" ]
Third_tail 1- [ lt3 ]

As a result, only the top-level structure is copied. The situation is roughly

as in Fig. 4.7: the two lists share all items but the last.
We had a similar situation in the tree insertion procedure. Check
that Fig. 4.8 properly illustrates the picture after inserting tumer : 6481
into the tree of Fig. 4.3: we copy the top-level structure of the whole
branch.
Copying structures upon modi cation is necessary because of the
semantics of the operations: when we call append(Ll, L2, L3) to concat-
enate Ll and L2, we may wish to preserve an unmodi ed Ll. If we want
destructive modi cation operations, we must express this desire ex-
plicitly.

... . ... . . -sit]

- - @-

.. -__ -. _ _ ..___._;[]

FIG. 4.7 The result of appending two lists.

93 4 Simple Programming Techniques

i
old tree

'1. '5

,/\
I mcbride:17Bl nil

t/<L\.
FIG. 4.8 The result of insertion into a BST.

4.2.2. Open Lists and Trees

If we want to build lists ef ciently, we must avoid copying longer and

longer initial segments of the nal list. Recall how append extends the list
piece by piece. After the call
append( [ Itl, It2 ], [ It3 ], X )
 4.2. Examples of Data Structures 99

carom-: __

AFTER

Q<>---- --~::;'f f11i:1 @

@-<5 eds -5> r FIG. 4.9 Extending a list.

we get
x <-1 111 | Third_tail' 1
Third_tail' <-1 112 | Third_tail" 1
and nally bind Third_tail". The trick is to keep Third_tail" ready for a
subsequent instantiation:
Third_tail" 1- [ It3 I Third_tail' ]
The situation will be roughly as in Fig. 4.9. Figuratively speaking, we
shall be able to resume append in the next step of computation. We only
need to get hold of the variable Third_tail"', instantiate it:
Third_tail' 1- [ It4 I Third_tail'' ]
and so on. When we are through, we can instantiate, say,
Third_tail' 1- []
and come up with the nal instance of X,
[It], It2, It3, It4]
I00 4 Simple Programming Techniques

We shall illustrate this with a procedure that reads in a sequence of

letters (up to the rst non-letter) and puts them in a list:
read_letters( [ L I Tail ] ) :-
lastch( L ), letter( L ), !, rch, read_letters( Tail ).
read_letters( [] ).
(See Section 5.7.4 for the description of lastch and rch,)
The last tail variable can be left uninstantiated. Although the resulting
structure will not be a proper list, it will be equally good as a representa-
tion of sequences. We shall call such structures open lists, and to avoid
confusion we shall call proper lists, with [] at the end, closed lists. Empty
open lists will be uninstantiated variables.
We must exercise some care if we deal with open lists. Consider the
procedure that extends a given list by instantiating its tail variable:
extend( List, Ext ) :- var( List ), List = Ext.
extend( [ _ I Tail ], Ext ) :- extend( Tail, Ext ).
For example, after the call
extend([a,b|V],[c,d|W])
the rst parameter becomes la, b, c, d I W].
It is essential that the instantiation of the tail variable be delayed.
Consider what would happen if we changed the rst clause to (apparently
equivalent)
extend( Ext, Ext ).
The result of the call
extend( [ X, Y, Z I Endl ], [ a, b I End2 1)
(i.e. the rst parameters instantiation) would be [a, b, Z I Endl] instead of the
expected [X, Y, Z, a, b I End2].
Procedure extend can reasonably be used only in strictly detenninistic
fashion. Failure after a successful computation causes dummy elements to be
inserted after the rst list. For example, the calls
extend( [ a I El ], [ b I E2 I ). fail
instantiate El as [b | E2], [- , b I E2], [_ , _ , b I E2], etc. Therefore a more
reasonable version would be that with a cut at the end of the rst clause.
The reasoning that has led us to open lists can also be applied to trees.
Uninstantiated variables represent empty open trees. Non-empty open
trees will be represented as before. For example, the following term rep-
resents the tree of Fig. 4.1 (El, E6 are distinct variables):
 4.2. Examples of Data Structures [OI

t( El, few, t( t( t( E2, languages, E3 ), many, E4 ),

people, t( E5, speak, E6 ) ) )
Again, we shall refer to trees discussed before as closed trees.
We need not copy anything to insert a node into an open tree. We can
go down the appropriate branch, locate a suitable empty tree, i.e. a vari-
able, and instantiate it to a new leaf:

ins( Node: Empty ) =-

var( Emntu )r Emuts = t( E1: Node: E2 ).

ins( Node: t( Left: Root: H ) ) =~

precedes( Node: Root )7 ins( Node: Left ).

ins( Nodev tt xv Root Rjsht ) ) =-

nrecedest Root Node )r ins( Node: Risht ).

If we rewrite the rst clause as

ins( Node, t( El, Node, E2 ) ).
a subtle change in the procedure's behaviour will ensue. The procedure
will insert nothing if this Node was already present in the tree. Surpris-
ingly it will also be identical to the procedure search from the previous
section, and (as might be expected) will serve almost the same purpose.
The overall effect of this insertion/search procedure can be described as
follows. lt looks for a given Node and succeeds after nding it. However,
if there is no matching node in the tree. the procedure inserts Node and
then " nds" it as well.
There are some strikingly elegant applications of this. A well-known
example is maintenance of symbol tables for translators written in Prolog.
If the translated language is not. block structured, a symbol table usually
cannot contain duplicate entn'es, and it nonnally only grows, so that
keeping it in an open tree will require no copying at all.
The example we are going to present is, of necessity, rather involved.
Before we proceed, you might nd it helpful to retum to Sections 3.3 and
3.4.1, where we described a simple Algol-like language and sketched a
parser and a code generator.
We intend to produce object code for a single-address target machine.

-The only dilTerence is strictly technical: in some Prolog implementations dummy

variables cannot be used to pass information. so we must insert a leaf with fresh named
variables.
|02 4 Simple Programming Techniques

For simplicity, we assume the code will not contain extemal references
(we shall also not attempt any optimisations).
The code generators output should be a list of "symbolic" instruc-
tionsterms described schematically as
Opcode( Address )
Each Address is an uninstantiated variable. There should be a unique
Address for every addressable symbol of the source program (variable,
constant, label), and for every label created by the code generator. By
way of explanation, we give a possible translation of the assignment
x := x + y * y + 2
most opcodes have obvious meaning.
[ load( AI ),
store( A2 ),
load( A3 ),
mult( A3 ),
add( A4 ),
add( A2 ),
store( Al ),
$lP( - ).
label( Al ), data( _ ), % x
label( A2 ), data( _ ), % temporary
label( A3 ), data( _ ), % y
label( A4 ), data( 2 ) % constant 2
1
We want the same Prolog variable for all occurrences of a source variable;
for example, Al always stands for x.
To assemble this section of code, we should determine the base ad-
dress and go down the list, counting bytes (or other units of storage).
Each executable instruction would be assigned a nal address. The pseu-
doinstruction label would be treated differently. We would instantiate
Address as the current value of the location counter (without advancing
the counter); this would instantiate all occurrences of Address (or of
van'ables bound to it, if one wants such ne distinctions). Assuming each
instruction takes four bytes and the fragment of code starts at location
1000, we would obtain
1000 : load( 1032 ),
1004 : store( 1036),
1008 : load( 1040 ),
1012 : mult( 1040 ),
etc.
 4.2. Examples of Data Stmctures I03

Conveniently enough, all we need to achieve this remarkable beha-

viour is the procedure ins (it should have rather been christened
rable-lookup). Whenever the translator encounters a symbol, say x, in
the source program. it allocates a fresh variable V, to represent the sym-
bol in subsequent processing. lt also calls ins to locate or place the pair
P(X.V)
in the symbol table. On the rst occurrence of x the pair will actually be
inserted. A subsequent "insertion" of p(x, U) only binds V and U to-
gether, i.e. nds xs symbolic address".
For this scheme to work properly, each non-terminal symbol in the
grammar that implements our code generator (see Section 3.4.1) must be
fumished with one additional parameter to pass the symbol table. The
whole grammar should be called with an empty table:
generate_code( S, O ) :-
phrase( code( S, SymTab ), O ).
And here is a rule that might be used to generate code for assignments:
code( assign( Name, Expr ), SymTab )>
codeexpr( Expr, SymTab ).
% code for this arithmetic expression,
% the value will be left in the accumulator
[ store( Addr ) ],
{ ins( p( Name, Addr ), SymTab ) }.
Symbol tables can also be implemented in open lists. For short tables
lack of overhead due to key ordering tests can outweigh the loss due to
worse performance. The simplest lookup procedure for open lists can be
written as follows:
lookup( Entry, [ Entry I Tail ] ).
lookup( Entry, [ _ | Tail ] ) :- lookup( Entry, Tail ).
This procedure, and two other versions (a bit more sophisticated) have
been used in the Prolog part of ToyProlog implementation (see Section
7.4, Appendices A.2 and A.3), and in the program described in Sec-
tion 8.2.
Open lists were rst used in the bootstrapped Prolog interpreter from
Marseilles (Battani and Mloni 1973, Roussel 1975). The technique shown
in the code generator example was presented by Colmerauer (1975, 1978).
Open trees were introduced by Warren (l9'7'7b, l980b).

For simplicity, we omitted the symbol table while developing the parser. We can save
this particular program by doing symbol table management in the back-end, but of course the
more proper way is to install symbols in the table in the front-end.
I04 4 Simple Programming Techniques

4.2.3. Difference Lists

If the application does not require shortening a list, open lists can be
constructed with no copying whatsoever. Successive instances of the
originally empty lista variableare longer and longer open lists (as-
suming, of course, that we are careful to instantiate nal variables appro-
priately). However, each time we add an item, the list must be traversed
to nd the nal variable. To avoid this, we can keep this variable ready for
instantiation:
End = [ Newltem I NewEnd ]
and make NewEnd available for further processing.
The pair consisting of a list and its nal variable can be considered
another representation of the lista little redundant for the sake of ef -
ciency. It is reasonable to represent the term as a single term. We shall
write it as
OpenList -- ItsFinalVariable
with -- a nonassociative in x functor. For example:
[ a, b I X ] -- X
To add an item at the end of a list we use the procedure
additem (Item, List -- [Item I NewEnd], List -- NewEnd).
The call
additem( 4, [ l, 2, 3 I X ] -- X. NewList )
instantiates, as expected,
NewList < [ l, 2, 3, 4 I NewEnd ] -- NewEnd
because
X<[4INewEnd]
Consequently, the old list becomes
[1, 2, 3, 4 I NewEnd ] -- [4 I NewEnd ]
To get a new list, we had to destroy the old one.
Fortunately, the destruction is apparent. The pair can still be re-
garded as a representation of the sequence l, 2, 3. Notice that [4I

Difference lists (d-lists) were introduced by Clark and Tarnlund (I977).

 4.2. Examples of Data Structures I05

NewEnd] is a tail of [I , 2, 3, 4 I NewEnd]. The sequence consists of those

items we must pop off the rst list to get its tail, i.e. of items by which the
two lists differ-hence the name of this data structure: difference list
(d-list for short).
Actually, a pair consisting of an open list and its tail is only a special
case: a difference list is de ned as a pair X -- Y such that X = Y or X =
[AI , An I Y] for some n 2 I. In general, no restrictions need be placed
on the fomi of Y, although the most interesting applications of difference
lists are those where Y is an open list.
Difference lists can be used to advantage whenever activity is ex-
pected at both ends of the sequence, e.g. when it is used as a queue. The
procedure additem enqueues an item. To dequeue an item, we can use the
obvious
remitem( Item, [ Item I List ] -- End, List -- End ).
but the behaviour of this procedure is unsatisfactory for empty lists. The
call
remitem( ltem, E -- E, NewList )
instantiates NewList as List -- [Item I List], i.e. as a negative difference
list". A procedure which fails, given an empty list, may be written as
follows:
removeitem( Item, List -- End, NewList -- End ) :-
not List = End, List = [ Item I NewList ].
Another nice feature of difference lists is the way they can be concat-
enated. Suppose we have two lists:
[a,bIX]--X and [c,d,e|Y]--Y,
and we want to compute a list holding the sequence a, b, c, d, e. If we can
assure that I
X=[c,d,e|Y],
we shall have la, b I X] = la, b, c, d, e I Y], and
[ a, b I X ] -- Y
will be a solution. This is readily generalized as a procedure:
d_conc( Listl -- Tail], Tail] -- Tail2, Listl -- Tail2 ).

This structure can be very useful in its own rights; see Shapiro (l983b, Section 4.8).
I06 4 Simple Programming Techniques

Once again, it must be stressed that modi cation of such lists is destruc-
tive. For example, the second call below fails, because [c, d, e I Y] does
not match [p I Z]:
d_conc( [ a, b I X ] -- I Y ] -- Y, ABCDE ),
d_conc( [ a, b I X ] -- .?<.?< '. NP- -0 -- Z, ABP ).
We now return to the sorting algorithm based on BSTs (see Section
4.2.1). Instead of traversing the tree, built of a given list, and merely
writing out the nodes, we would rather traverse it in order to construct the
sorted permutation of the list:
tree- sort( List, SortedList ) :-
buildtree( List, nil, Tree ),
buildlist( Tree, SortedList ).
The procedure buildlist " attens" the tree (see Fig. 4.10 for an exam-
ple). The general outline of the algorithm is rather obvious: we atten the
subtrees (recursively) and concatenate the resulting lists together with the
root in between. Difference lists can be used to avoid numerous appends.
Let the results of recursive calls be denoted by
LFlat -- LFlatE and RFlat -- RFlatE
The algorithm is programmed as follows:
atten( nil, X -- X ).
atten( t( L, Root, R ), Flat ) :-
atten( L, LFlat -- LFlatE ),
atten( R, RFlat -- RFlatE ),

/\%
d_conc( LFlat -- LFlatE, [ Root I X ] -- X, A ).
d_conc( A, RFlat -- RFlatE, Flat ).

./\./\. FIG. 4.l0 (a) A tree. (b) The tree attened.

 4.2. Examples of Data Structures I07

This version is good for didactic purposes. Actually, we know that

the following instantiations take place:
I..FlatE < [ Root I X ], A < LFlat -- X,
X < RFlat, Flat < LFlat -- RFlatE
We can remove both calls on d_c0nc and end up with an equivalent form
of the second clause:
atten( t( L, Root, R ), LFlat -- RFlatE ) :-
atten( L, LFlat -- [ Root I RFlat ] ),
atten( R, RFlat -- RFlatE ).
We might similarly derive a short cut" clause for leaves. We begin with
atten( t( nil, Root, nil ), LFlat -- RFlatE ) :-
atten( nil, LFlat -- [ Root I RFlat ] ),
atten( nil, RFlat -- RFlatE ).
then make LFlat = [Root I RFlat] and RFlat = RFlatE, and remove the
recursive calls. The special case becomes:
atten( t( nil, Root, nil ), [ Root I RFlatE ] -- RFlatE ).
(as expectedl).
After the call
atten( Tree, List -- ListEnd )
we shall have List instantiated as
[Node|, Node" I ListEnd ],
and all we shall need to get SortedList is close List by binding ListEnd to
[]. This is easily achieved by de ning
build|ist( Tree, SortedList ) :-
atten( Tree, SortedList -- [] ).
(or replacing the buildlist call in rree_sorr. for that matter).
See Section 7.4.1 for a little more sophisticated application of differ-
ence lists.

4.2.4. Clausal Representation of Data Structures

A Prolog procedure built of unit clauses is a natural representation of

sets and sequences. Under the static interpretation of programs, such a
I03 4 Simple Programming Techniques

procedure models a relation. i.e. a set of tuples for which a certain rela-
tionship holds. For example:
name_phone( thompson, 2432 ).
name_phone( adams, 5488 ).
name_phone( white, 2432 ).
name_phone( mcbride, I781 ).
In practice, unit clause procedures are sequences rather than sets, in
that they are accessed sequentially. It is therefore possible to represent a
list by a procedure, e.g.
Iist( b ).
Iist( k ).
Iist( q ).
Iist( y ).
The call
Iist( X )
tests membership for instantiated X, and serves as a generator for unin-
stantiated X. The whole list can be processed thus:
process_list :- Iist( X ), process_item( X ), fail.
process_list.
In general, clauses may be used to represent multidimensional matrices-
we shall discuss this brie y in the next section.
The restriction to unit clauses is not essential. The clause
name_phone( X, 4396 ) :- of ce( X, rooml I9 ).
will generate tuples one at a time, exactly as the other four clauses do.
It is wonh emphasizing that explicit and generated data are functionally
indistinguishable. If ve people sit in room I19, we can get up to nine
name-phone pairs, without ever becoming aware of the "indirection" in
one of the clauses.
Any structure expressible in terms of relations can be naturally cast in
clauses. For example, a tree can be described as follows:

t( nodel , node2, thompson : 2432, node3 ).

t( node2, nil, adams : 5488, node4 ).
t( node3, nil, white : 2432, nil ).
t( node4, nil, mcbride : I78], nil ).
root( nodel ).
 4.2. Examples of Data Structures I09

In particular, we can represent a list in this way:

l( iteml, b, item4 ).
l( item2, y, nil ).
l( item3, q, item2 ).
l( item4, k, item3 ).
head( item] ).
In general, every graph can be expressed as a unit-clause procedure. By
way of explanation, here is a possible representation of the graph of Fig.
4.11 (see also Fig. 3.1):
edge( el, e2, o ).
edge( el, e2, letter ).
edge( el, e3, atom ).
edge( e2, e3, x ).
edge( e2, e3, letter ).
edge( e2, e3, atom ).
And a representation of the graph of Fig. 4. I5 (Section 4.4.3):
arc( a, b). arc( a, c ). arc( b, c ). arc( b, d ).
arc( b, e ). arc( c, d). arc( c, e ). arc( d, e ).
Clausal representation of trees, lists and the like is rather less conven-
ient than representations described in previous sections. It cannot be
passed as an actual parameter, so that its use can only be recommended
when the bulk of data remains unchanged (see Section 8. I for a non-trivial
example). Since variables are local in clauses, clever techniques shown in
Section 4.2.2 are hardly applicable here. To build and modify data dynami-
cally (e.g. add a node to a tree), we must apply extralogical built-in
procedures assert, retract, etc., to the detriment of static interpretation of
programs.
There are advantages, too. First of all, in Prolog implementations
which support clause indexing, direct access to components can be possi-
ble. Indexing consists in nding matching clauses by hashing rather than
by linear search, so that e.g. a node in a tree with n nodes can be
located in constant time rather than in log;n steps (on the average).

atom

,1):
e1 e2 e3
FIG. 4. || A graph.
 ]|() 4 Simple Programming Techniques

DEC-I0 Prolog was the rst to offer this possibility. If absent, it can be
mimicked by means of the built-in procedure =.. (see the next section).
Clausal representation sometimes helps reduce the problem at hand
to its bare essence. A case in point is an amazingly concise solution to a
map colouring problem; we quote it after Pereira and Porto (l980b). A
planar map is to be coloured with at most four colours so that contiguous
regions are coloured differently. First we de ne the contiguity relation for
colours:

nextt red: blue ). nextt red: sreen ). nextt red: sellou ).

nextt blue: red ). nextt blue: sreen ). next( blue: sellou ).

nextt sreen: red ). nextt sreen: blue ). next( sreen: sellou )

next( yellow: red ). nextt yellow: blue ). nextt sellou: sreen )

The original map of Pereira and Porto (l980b) is shown in Fig. 4.12. A
region is represented by its colourthis decision makes the solution
beautifully terse. To nd a colouring (if any) of the map, we must only call

nextt R1: R2 ): nextt R1: R3 ): nextt R1: R5 ): next( R1: R6 ):

next( R2: R3 ): nextt R2: R4 ): next( R2: R5 ): next( R2: R6 ):

nextt R3: R4 ): next( R3: R6 ): next( R5: R6 ):

UPitG( (R1: R2: R3: R4: R5: R6) )9 nl.

Structures represented by temts are usually traversed and manipu-

lated by recursive procedures. Clauses are traversed by backtracking,
either implicit (e.g. in the call above), or explicit (e.g. in the procedure

3 _.- -6

2
I
I

i I l

FIG. 4.l2 A map to be coloured.

 4.2. Examples of Data Stnrctures III

process_list). There is a fundamental discrepancy between these two

modes, because backtracking destroys variable instantiations which are
essential to recursive operations on data structures. Consider the task of
computing a list of arcs exiting vertex b of the graph in Fig. 4. I5. Arcs are
available one at a time to a routine that backtracks through" the proce-
dure arc. If we want them to survive backtracking, we must put aside".
i.e. assert, those which contain b:

Dut_oside =- arc( X: Y ): nut_oside_if_b( X: Y ): foil.

nut_o5ide_if_b( X: Y ) 8- has( b: X: Y ): o5sert( uith_b( X: Y ) )

hO5( X: _: X )-

= nuL_o5ide.

Z Now: o list con be created as follous=

collect_uith_b( ThisList: FinolList ) =-

retroct( uith_b( X: Y ) ): !:

collect_with_b( E (X: Y) I ThisLi5tJ: FinolList ).

collect_uith_b( Fino1List: FinolList ).

8- collect_with_b( E]: TheList): write( TheList ): nl.

Such operations are usually cast in terms of a general-purpose proce-

dure that nds a set of all items for which a given condition holds. In our
example, items would be (X, Y), and the condition
(arc( X, Y ), has( b, X, Y))
The set is represented by a list, possibly with repetitions, so that it is
called bag in the folklore. Here is our version of the procedure:
 I I2 4 Simple Programming Techniques

basoftltem: Condition: _) =-

ossert('BAG'('BAG')): Z a marker

Condition: Z senerates an instance of Item

ossert('BAG'(Item)): Z saves it

fail. Z this clause eventualls fails

basof(_: _: Baa) =-

retract('HAG'(Item)): !: Z set the last Item saved

collect(Item: E]: Bus).

collect('BAG': FinalBa9: Final as) =- !. Z this was the marker

collect(Item: This as: Finol as) :

retract('BAG'(NextItem)): I:

collect(NextItem: [Item I ThisBa9]: Final as).

The marker enables us to use the procedure bagof within Condition.

An example of such nested computation is the following pair of calls
(Graph is to be a list of bunches"lists of arcs entering or leaving a
given vertex; the condition in the rst call is an altemative, in the embed-
ded one a conjunction):

= basoF( X: ( arc(X: _)i arc(_: X) ): Uertices ):

basof( Bunch:

( member(U: Uertices):

basoft (Y: Z):

( arc(Y: Z): has(U: Y: Z) ):

Bunch

>
>:
Graph ).

Repetitions in a bag may be undesired. For example, the rst call

above should rather nd a set of all verticesas it stands, Graph will
contain numerous duplicates. The procedure setof would call bagof and
then lter the resulting bag. In Prolog- I0 and some other implementations
 4.2. Examples of Data Structures I I3

both bagof and setofare built-in procedures: setofeven retums its output
sorted. An implementation of setof in Prolog was presented by Pereira
and Porto (l98l).

4.2.5. Array Analogues in Prolog

There is no addressing mechanism in Prolog, no memory cells di-

rectly available to the programmerfor most applications this is simply
unnecessary. Consequently, there are no arrays interpreted as contigu-
ous, addressable areas. From a mathematical standpoint, arrays corre-
spond to nite matrices, i.e. to mappings from nite sets of subscripts to
sets of values. In theory, there are no restrictions on the fonn of sub-
scripts, although integers are most commonly used.
In Prolog we can represent such mappings as procedures consisting of
unit clauses, one clause for each sequence of subscripts and the corre-
sponding value. This is but a special case of relation in the relational model
of data (see Section 8.2).
Unit clauses are particularly convenient as a representation of sparse
matrices, provided that clause indexing is supported by the Prolog imple-
mentation.
Another possibility is to represent a mapping as a list of n-tuples
(subscripts, value), and to use list manipulation procedures. As a special
case, a sequence subscripted by consecutive integers may be represented
as a list of values. This approach may work for short lists, but in general it
is prohibitively inef cient.
We shall now present an alternative way of storing integer-sub-
scripted sequences, which is rather unlikely to outperform Prolog data
bases (with indexing), but may be reasonable for sequences of moderate
size. The method makes use of digital search trees (see e.g. Sedgewick
I983).
Branching in digital search trees is based on the values of successive
digits of the key being looked for. Keys cannot be negative. The order of
every node is equal to the base of the digital system, e.g. to I0 if keys are
expressed in decimal. In Fig. 4.13 we show two trees, each containing I5
items numbered 0 through I4. A,- denotes the i-th item, the root is empty
(i.e. contains a dummy value), and branches are labelled with digits. To
nd AI; in the binary tree, we take ll0l, the binary code of I3, and go
down selecting branches labelled with I, I, 0 and I. To nd AI; in the
temary tree, we use Ill, the ternary code of I3.
Digital search trees are best implemented in Prolog by open trees. We
shall demonstrate it in the case of temary trees (other cases are basically
 /C\

/\
/\ Q0 I

ltlx:kZ
/.L\
ID/
6)

tree.
/ .m\
FIG.4.l3 Digital search trees: la) Binary digital search tree lb)Tem rydtgtt lse rch
 4.2. Examples of Data Stmctures I I5

identical, although impractical if nodes have more than a few branches).

A non-empty tree will be represented as
t3( Value, Left, Middle, Right )
and an empty tree as a variable. Explicit labels are unnecessary, as we
may simply select Left or Middle, or Right upon encountering 0, I or 2,
respectively.
A procedure that nds a value, given a ternary subscript and a tree, is
quite straightforward. We assume that subscripts are represented as
closed lists of digits:

find_3( E]: t3( Value: _: _: _ ): Value ).

find_3( [0 I Sub]: t3( _: Left: 2: _ ): Value ) :-

find_3( Sub: Left: Value )-

find_3( [1 I Sub]: t3( _: _: Middle: _ ): Value ) :-

find_3( Sub: Middle: Value ).

find_3( [2 I Sub]: t3( _: _: _: Risht ): Value ) :-

find_3( Sub: Risht: Value ).

The procedure fails if the rst parameter is not a correct temary subscript,
or if the second parameter is not an open temary tree. However, it does
not fail when a nonexistent item is referred to. We shall discuss this
phenomenon presently.
Now for a procedure that replaces an item. Two tree parameters are
required, and the new tree is a copy of the old one, except for the replaced
item. The amount of copying is similar to that illustrated in Fig. 4.8.

chanse_3( E]: NewVal: t3( _: L: H: R ):

t3( NeuVa1: L: H: R ) ).

chanse_3( [0 I Sub]: NeuVal: t3( OldVal: L: H: R ):

t3( OldVal: NewL: H: R ) ) =-

chanse_3( Sub: NewVal: L: NeuL ).

chan9e_3( [1 I Sub]: N0uVa1: t3( 0ldVal: L: H: R ):

t3( 0ldVa1: L: New : R ) ) =-

chanse_3( Sub: NeuVal: H: New ).

 I16 4 Simple Programming Techniques

chanse_3( 2 : Sub]: NeuVal: tat 01dVal: L: a: R >.

e3< 0ldVal: L: H: NeuR > > :-
chanse_3( Sub: NeuVal: R: NewR )-

Both procedures behave in the same way when the subscript is too
large: they create a missing part of the tree, and then nd or change
the newly inserted item. For example, the call
nd_3( [2, l, 0, I ], Tree, A64)
applied to the tree of Fig. 4. I3b changes the node with item A7 into the
tree of Fig. 4.l4, or, in term notation, into
t3( A7, t3( Dummy2l, Empty_i.
t3( A64, Empty_ii, Empty_iii, Empty_iv ),
Empty_v ),
Empty_vi, Empty_vii )
The same effect will be achieved by the call
change_3( [ 2, I, 0, I ], A64, Tree, Tree )
The moral is that, rst, no special insertion procedure is needed, and,
second, the tree need not be full. It will contain only the inserted nodes
together with the branches required to reach these nodes, but intermedi-
ate nodes may contain no meaningful information.
To make the story complete, here is a little procedure that converts
nonnegative integers into lists of temary digits. Note that there are two
procedures here: conv_3/2 and (auxiliary) conv_3/3.
conv_3( O: [O] ).

conv_3( N: TerN ) = inteser( N ): O ( N: conv_3( N: E]: TerN )

conv_3< 0: ll isits: ll isits ) '.

conv_3( N: Z: ll isits ) =-

Disit is N mod 3: Nbs3 is N / 3:

conv_3( Nbs3: [Disit I Z]: All isits ).

4.2.6. Access to the Structure of Terms

In Section 4.2.1 we dismissed the possibility of representing tree

nodes with main functors: the term
few( nil, people( many( languages, nil ), speak ) )
 4.2. Examples of Data Structures II7

Q
FIG. 4.l4 Creation of the missing part of a tree.

would stand for the BST of Fig. 4.l. We shall now show an insertion
routine for such trees. The built-in procedure = .. (univ) is used to circum-
vent the problem raised by the potential diversity of the functors.

insert( Node: Tree: NewTree ) =-

Tree =.. ERoot: Left: Risht]:

insert( Node: Root: Left: Risht: NeuLeft: NeuRi9ht ):

NewTree =.. [Root: NewLeft: NeuRisht].

insert( Node: nil: Node ).

insert( Node: Leaf: NeuTree ) =-

insert( Node: Leaf: nil: nil: Left: Risht ):

NeuTree =.. [Leaf: Left: Risht].

insert( Node: Root: L: R: NeuL: R ) =-

vrecedest Node: Root ): insert( Node: L: NeuL ).

insert( Node: Root: L: R: L: NeuR ) =-

nrecedest Root: Node ): insert( Node: R: NeuR )-

This application of univ is far from typical. As a more realistic exam-

ple, consider the problem of translating an arithmetic expression into
reverse Polish form, e.g.
y*$qrl(sqr(x)+f(l.y))
into
I y. X. sqr. I. y. f. '+'. sqrl. * I
II3 4 Simple Programming Techniques

Here is a possible solution:

revwoll Exnr: RQVEXPP ) I-

Earr =.. [Fun I Arss]: ravarast hrss: I]: Revarus J:

arwendt Rav ris: [Fun]: RevExpr J.

revarast E]: Revall: Revall ).

revarlst [Ara I Aris]: RevTillNou: Revall ) I-

PEVPO1( Ari: Revhrs ):

append( RevTillNou: Revara: Rev ne ore ):

revarsst r bv Rev0neHorQ: Rev ll ).

We could use difference lists to decrease the cost of multiple append-

ings. but the procedures would become even less readable (but try itthis
would be an application of the " atten" schema, although a little un-
wieldy because of the unknown number of arguments). However, a read-
able version would not only be much longer, but also less exible:
revpol( A + B, RevExpr ) :-
revpol( A, RevA ), revpol( B, RevB ),
append( RevA. RevB, Aux ), append( Aux, [ + ], RevExpr ).

revpol( sin( A ), RevExpr ) :-

revpol( A, RevA ),
append( RevA, [ sin ], RevExpr ).

revpol( Atom, [ Atom ] ).

This is a closed schema: to be able to recognize a new function or opera-
tor, we must add a branch to this case statement".
Perhaps one of the most important applications of univ (and related
built-in procedures) is in bootstrapped implementations of Prolog. A basic
interpreter (see Chapter 6 and Section 7.3) may support Prolog (with a
very rudimentary syntax) furnished with built-in procedures analogous to
call and univ. Various user interfaces can then be written in this simpli ed
Prolog (see Section 7.4). provided we can convert texts to terms.
Assume we input the text
foo( e( X ), ok, X )
 4.2. Examples of Data Structures I I9

and produce its (intermediate) representation:

[[f.0.0].[[f.i.=l.V].[[0.l<]].V]
with uninstantiated V. (Try to write this reading program: a symbol table
such as those described in Section 4.2.2 must be used to handle variable
names properly.) Now we can glue the intermediate representation to-
gether:

sluet Inter: Inter ) = var( Inter ): '.

alue( CFunChars I Inter rss]: Term ) =-

not alldisitst FunChars ):

sluearss( Inter rss: Ares ):

pname( Fun: FunChors ): Term =.. [Fun I rss].

9lue( [Disits]: Number ) :-

alldisitst Disits ):

onamei( Number: Disits )-

sluearsst E]: L] ).

sluear9s( Elnter rs I lnter rss]: [ rs I Arss] ) =-

slue( Inter rs: rs ): sluearss( Inter rss: Arss ).

alldisitst CDi9it I Uisits] ) =-

d1sit( Di it ): I: a11di9it5( Uiyits ).

al1disits( [J J.

The procedure glue should be called with the second parameter uninstan-
tiated.
In implementations that do not support indexing (see Section 4.2.4),
unlv helps avoid linear search of matching clauses. Consider. for exam-
ple, a natural language application program which maintains a dictionary
whose entries can look as follows:
dict( program, noun( inanim ) or verb( intrans ) ).
dict( modular, adj ).
dict( an, article( indef ) ).
I20 4 Simple Programming Techniques

Next, assume that each word on input is ltered through this dictionary:
input_a_word( W, Features ) :-
read_a_word( W ),
( dict( W, Features ), I: signal_unknown( W ) ).
Without indexing, dictionary lookup requires time proportional to the
number of entries. Access to a procedure, i.e. to its rst clause, usually
requires approximately constant time (some form of hashing is used). We
can have our dictionary in the form
program( noun( inanim ) or verb( intrans ) ).
modular( adj ).
an( article( indef ) ).
and de ne dict as
dict( W, Features ) :-
Entry =.. [ W, Features ], Entry.
orequivalentlyas
dict( W, Features ) :-
functor( Entry, W, I ), Entry, arg( I, Entry, Features ).
A particularly simple dictionary is a table of keywords for a scanner
of, say, Pascal:
const. type. array. record.
function. var. begin. do.
etc. To create the representation of a source program name, we can use
this procedure:
key_or_id( Name, keyword( Name ) ) :- Name, !.
key_or_id( Name, ident( Name ) ).
As a nal example, here is the crucial part of a de nition of the
procedure phrase which initiates processing based on a metamorphosis
grammar:
phrase( InitialNonterminal, Terminals ) :-
InitialNonterminal =.. [ Name | Parameters ],
InitialCall =.. I Name, Terminals, [] | Parameters ].
InitialCall.
Note that input and output parameters are added at the beginning of the
parameter list (rather than at the end, as suggested in Section 3.3).
 4.3. Some Programming Hints IZI

4.3. SOME PROGRAMMING HINTS

We have collected here some down-to-earth suggestions which may

help improve your coding technique in Prolog. Although style is largely a
matter of taste, some of the things we have to say have long been present
in Prolog folklore, and we feel fairly con dent they are worth presenting.

4.3.1. Using the Cut Procedure

Essentially, the cut commits the currently executing procedure to

whatever it might have done since its activation. This is precisely what
makes the cut a controversial feature: that it can only be interpreted
dynamically. On the other hand, the variety of its uses and its power make
it an important factor in the emergence of Prolog as a practical program-
ming language.
In Chapters I and 2 we discuss the cutin a very general manner-
both as an extralogical mechanism and as a tool for improving ef ciency.
Here, we shall concentrate on its applications.
Despite Prologs inherent nondeterminism, the usual computation is
mostly deterministic: the majority of procedures are expected to produce
a single, well-de ned response to any particular set of input data. Most
procedures are strictly deterministic: at most one clause of a procedure
applies, regardless of the actual data.
With the procedural interpretation of Prolog in mind, clauses are
commonly written as
head :- tests, actions.
A clause is executed for its actions which can be performed if and only if
head matches the call and all tests succeed. This conforms to the funda-
mental notion of guarded commands (Dijkstra I975). Some Prolog dia-
lects, e.g. IC-Prolog (Clark et al. l982b), even provide special syntax for
guards.
If, during a deterministic computation, tests have succeeded, a cut
executed immediately after tests commits our choice of the clause. The
cut saves us furtherunnecessaryattempts to execute the procedure in
the case of a failure later on. As an example of this fairly typical situation,
consider the following:
% Retrieve ( fetch ) the grammatical description of a word,
% fail if there is no such word in the dictionary.
I22 4 Simple Programming Techniques

% The word may be given as a string:

nd( String, Description ) :-
isletterstring( String ), % yes, a string
pname( Word, String ), fetch( Word, Description ).
% or as a word, i.e. nullary functor:
nd( Word, Description ) :-
isword( Word ), % yes, a word
fetch( Word, Description ).
% Reject bad data:
nd( Bad, _ ) :-
not isletterstring( Bad ),
not isword( Bad ), % yes, bad data
signal( Bad ), fail.
In this procedure, cuts may be safely placed after tests. Notice, however,
that a cut inserted earlier changes the procedures behaviour, and a cut
afterfetch does not work if a word is absent from the dictionary. (lt would
also have ruinous effects if fetch were a nondeterministic generator of
synonyms.)
When we adhere to the guarded command style of programming.
the built-in procedure not is frequently used to invert tests (but see the
beginning of the next section for a brief discussion of nots peculiaritiesl).
However, we would not like to perform expensive tests twice, as in this
example:
addunique( Item. List ) :-
presentinalonglist( Item, List ), signal_dupl( Item ).
addunique( Item, List ) :-
not presentinalonglist( Item. List ), additem( Item, List ).
We can replace the inverted test with a cut after the original test:
addunique( Item, List ) :-
presentinalonglist( Item, List ), !,
signal_dupl( Item ).
addunique( Item, List ) :- % not present...( Item, List )
additem( Item, List ).
This procedure can be interpreted as
if present...( Item. List ) then signal_dupl( Item )
else additem( Item, List )
This is, perhaps, the most frequent application of the cut. It must be
remembered, though, that this use of the cut is extralogical: a clause with
 4.3. Some Programming Hints I23

a test removed means something else, and it cannot be understood in

separation from the rest of the procedure. Still, the procedure as a whole
is usually suf ciently readable, if we view it as a (possibly nested)
if then else if then etc.
Sometimes cuts inside a procedure are undesirable. One example is a
procedure that holds data, e.g.
father( jack, tom ).
father( bill, john ).
etc.
(with empty tests and actions). With a cut in each clause this would not
only look ugly, the procedure would be of no use as a generator! Instead,
we should commit the call on father, as in this procedure:
is_father( Person ) :- father( Person, _ ), !.
The cut serves as a rewall against unwanted backtracking.
Another example. Consider this group of grammar rules:
command( Cmd ) > stop( Cmd ).
command( Cmd ) > dump( Cmd ).
command( Cmd )> load( Cmd ).
command( Cmd )> create( Cmd ).
etc.
A switch such as command is best committed by the cut after a call.
e.g.
phrase( command( Cmd ), Tokens ), I
This technique, however, has a disadvantage. The committing cut
affects not only the call but also the calling procedure. If the call being
committed happens to be the last test in a clause, then the cut plays two
roles at once. Otherwise we should make it invisible to the surrounding
clause. To achieve this, we can use this general-purpose call-and-com-
mit procedure:
once( Call ) :- Call, !.
Other arguments against cutting high are implementation-depen-
dent. First of all, in many implementations memory requirements are
smaller when there are fewer fail-points, so it may be desirable to perfomi
cuts as soon as possible. Some implementations also optimise storage
utilisation of tail-recursive procedures (see Section 6.4). A procedure may
I24 4 Simple Programming Techniques

become tail-recursive dynamically, after having its remaining clauses cut

off. For example:
% Recognize a sequence of letters/digits.
Id( [Ch |Chs] )-> [ Ch ],{Ietter( Ch ) }, !, Id( Chs ).
Id( [ Ch I Chs ] )> [ Ch ], { digit( Ch ) }, !, Id( Chs ).
Id( ll )> [I
(Here, the cuts may protect us against deep recursion, effectively chang-
ing it into iteration.)
Sometimes the use of cuts should be recommended for clarity. We
shall present two versions of the procedure that translates the term (A, ,
..., A,,) into the list [A] , ..., A,,] and the term A (other than a comma-term)
into [A]. First the version with full guards:
c_list( AA, [ AA ] ) :- var( AA ).
c_list( AA, [ A | As ] ) :-
not var( AA ), AA = ( A, AATail ), c_list( AATail, As ).
c_list( AA, [ AA ] ) :-
not var( AA ), not AA = (_, _ ).
And the version with cuts (here the order of clauses is crucial):
c_list( AA, [ AA ] ) 2- var( AA ), !.
c_list( ( A, AATail ), I AI As] ) :-
!, c_list( AATail, As ).
c_list( AA, [ AA ] ).

In nondeterminisitic procedures cuts should be used cautiously, if we

do not want to inadvertently lose some solutions. In particular, proce-
dures that compute multiple answers (such as append) should not contain
cuts. A cut after a call on a generator makes it yield only its rst satisfac-
tory answer, as in this small example:
int( 0 ).
int( NextN ) :- int( N ), NextN is N + I.
:- int( X ), satisfactory( X ), !.
Cuts after tests in a procedure written according to the guarded com-
mand style implement Dijkstra's don't-care nondeterminism of if state-
ments: anyexactly oneof the branches with true guards is chosen (in
Prolog, the rst one).
Special care must be exercised when adding cuts to procedures in-
tended to be used in more than one way (such as grammar rules intended
both for analysis and synthesis).
 4.3. Some Programming Hints I25

4.3.2. Failure as a Programming Tool

The procedure not, used to invert tests, owes its power and concise-
ness to the combined effect of three extralogical mechanisms in Prolog:
variable calls, the cut, and forced failure. Recall the de nition:
not X :-- X, I, fail.
llOI _.
Observe that the second clause performs no instantiations, and any in-
stantiations in X must have been undone on failure. If not succeeds, its
parameter will remain intact. Therefore, not will not retum anything. For
example, the call
not student( X )
with uninstantiated X will not nd a nonstudent (as might have been
expected). Instead, it will fail if there is at least one student, e.g.
student( jim ). student( jill ).
Otherwise it will succeed with X still a variable. If we insist on nding
nonstudents, we can look for them among NewYorkers:
newyorker( tim ). newyorker( jim ).
newyorker( jill ). newyorker( amy ).
Now the command
:- newyorker( X ), not student( X ),
write( X ), nl, fail.
will print:
tim
amy
It must be emphasized that not called with a term containing variables
does not implement negation properly (see Clark I978). If the call not
student(X) succeeds, then we shall actually prove that

m 3x student( x )
which is equivalent to

Vx m student( x )
On the other hand, suppose not means 1 The command
:- not student( X ).
I26 4 Simple Programming Techniques

would then be interpreted (see Chapter 2) as

Vx mm student( x )

i.e. as Vx student( x ). Its negationto be proved by reductio ad absur-

dumis

3x m student( x )

This discrepancy was commented upon, for example. by Clark and McCabe
(l980a, l980b) and Dahl ( I980). In IC-Prolog (Clark et al. l982b) the problem
was solved by treating not calls with variables as erroneous. This is to say.
negation in their system is only applicable to ground predicates.
Except for not, forced failure is used primarily for ef ciency. Many
Prolog implementations have no garbage collection, but upon backtrack-
ing almost all of them very ef ciently recover some storage holding con-
trol information and term instances (see Chapter 6). We can take advan-
tage of this in a few rather unobvious but effective tricks. One of them is
double not.
On the face of it, the trick is pointless: the call
not not C
succeeds if and only if C does. We shall trace the execution of this call to
show its hidden effect. Assume rst that C succeeds; here are successive
snapshots:
not not C
not C, !, fail
C, !, fail, !, fail
!, fail, !, fail
% the cut will commit the intemal not
fail, I, fail
% RECOVER the storage used by C,
% and backtrack in the external not
SUCCESS
Now, let C fail:
not not C
not C, !, fail
C, !, fail, !, fail
% backtrack in the internal not.
% succeed via the second clause
 4.3. Some Programming Hints I27

!, fail
% the cut commits the external not
fail
FAILURE
Since double not does not instantiate anything. it can only be used
in two situations. Either we want to perform a complicated yes/no test
(with all interesting variables already instantiated), or we are only inter-
ested in some side-effects of C but we want to recover storage after its
execution. For readability, we usually de ne two procedures:
check( Cond ) :- not not Cond.
side_effects( Goals ) :- not not Goals.
One example should suf ce:
prettyprint( Term ) :- side_effects( doprettyprinting( Term ) ).
Suppose now that we need instantiations produced when executing a
call, and that space still matters. To preserve the results (i.e. the appropri-
ate tenns) over backtracking, we must put them aside. Only stored
clauses are immune to failure. The following general-purpose procedure
executes a call, and at the same time garbage collects the storage used
by the call:
with_gc( Call ) :-
once( Call ), assert( ASIDE( Call ) ), fail.
with_gc( Call ) :-
retract( 'ASIDE'( Call ) ), !. % commit retract
This method makes sense when assert requires less storage than Call. or
when the implementation has no general garbage collector but reclaims
storage left by retracted clauses.
with..gc can be employed in loop optimisation, which is an important
application of forced failure. Essentially, recursion is the most natural
Prolog counterpart of Pascal-like iteration. Consider a program that takes
large chunks of an even larger text. extracts some data from them. and
puts these data into an open tree. The storage for a step is worth recover-
ing. Let step assert basta. after having encountered the nal chunk. The
loop can be written as follows:
buildtree( _ ) :- retract( basta ), !. % remove the signal
buildtree( Tree ) :-
with_gc( step( Tree ) ). buildtree( Tree ).
(Find a similar solution for closed trees.)
I This technique was advocated by R. A. Kowalski at the Logic Programming Work-
shop in Debrecen. Hungary. I980.
I28 4 Simple Programming Techniques

Suppose now that steps of a loop have no common terms (which

would have to be passed down the loop). This means that a step is exe-
cuted only for its side-effects. For example, consider the problem of
reading in a Prolog program up to the clause end.. Let the procedure
clause_in perform one step: read a clause and assert it (unless it is end. or
incorrect). The following procedure repeatedly calls on clause_in, and
recovers storage after each step:
getprog :- clause_in( Clause ), Clause = end, !.
getprog :- getprog.
This loop can be made even more concise if we use a failure
screen. This is a procedure that always succeeds nondetenninistically,
i.e. leaves room for yet another success:
repeat.
repeat :- repeat.
(it is standard in some Prolog implementation). The loop can be expressed
as
getprog :- repeat, clause_in( C ), C = end, !.
After C = end succeeds, the cut will remove the pending choice in repeat,
and so terminate the loop.
For this technique to work, the core of the loop must be deterministic,
as otherwise a failure of C = end would evoke another attempt to execute
an already executed step. Usually it suf ces to enclose the call for a step
in once(_):
getprog :- repeat, once( clause_in( C ) ), C = end, !.
A special form of forced failure is caused by tagfail". This built-in
procedure is described in Section 5.12, together with other associated
procedures. They are all primarily used for error handling, as they allow
bypassing of large fragments of a computation. Here we shall present an
application of tag and tagfail for exiting loops.
An extremely simpli ed interactive executor of Prolog commands can
be programmed as follows:
ear :- tag( loop ).
ear.
loop :- repeat, read( C ), once( C ), fail.

' It is only available in Toy (see Section 5.12), but something similar is present or can
be programmed in several other implementations of Prolog.
 4.3. Some Programming Hints I29

The execution of
tagfail( loop )
terminates the loop: tag(loop) fails, and the second clause of ear promptly
succeeds. With a step de ned as
step :- read( C ), once( C ).
and loop rede ned as
loop :- repeat, tag( step ), fail.
we can also exit one step by calling
tagfail( step )

4.3.3. Clauses as Global Data

The program modi cation procedures-assert, retract and the like-

are rst of all used to maintain Prolog data bases (see Section 8.2). They
can also be used in automodifying procedures, those which assert or
retract their own clauses; this is an extremely dubious programming trick,
and is not recommended, especially since such programs tend to be rather
subtly implementation-dependent.
Modi cation procedures are also used to store so-called global data.
In Prolog implementations that do not support modularisation, the data
kept in program clauses (notably unit clauses) are accessible to all proce-
dures, i.e. global. Such data are signi cant in Prolog because they are not
affected by backtrackingsee with_gc in the previous section. Also, they
are sometimes more convenient to handle than information passed around
via parameters. One example is a switcha parameterless unit clause
whose presence or absence provides a simple yes/no test. For instance,
we can supply terse or wordy error messages:
message( Code ) :- terse, short_mes( Code ), nl, !.
message( Code ) :- long- mes( Code ), nl, !.
short_mes( sym( S ) ) :- display( ?sym ), display( S ).

long_mes( sym( S ) ) :-
display( Unexpected symbol on input: ' ),
display( S ), nl,
display( The remainder of the command will be ignored. ).
A switch can be easily tumed on:
tumon( Switch ) :- Switch, !. % already on
tumon( Switch ) :- asser1( Switch ).
|3O 4 Simple Programming Techniques

and off:
turnoff( Switch ) :- retract( Switch ), !.
% fails if Switch was off
turnoff( _ ). % already off
We can also revert the state of a switch (on > off, off > on):
ip( Switch ) :- retract( Switch ), !.
ip( Switch ) :- assert( Switch ).
Switches are really cumbersome to program without clausal data. It is
not dif cult to rewrite message:
message( Code, terse ) :- short_mes( Code ), nl, !.
message( Code, wordy ) :- long_mes( Code ), nl, !.
but the Terseness parameter ought to be carried everywhere throughout
the program: and dynamic reversal of a switch can be somewhat messy.
Our nal example demonstrates how assertions can be used to memo-
rize results of expensive computations for future use. Let the procedure
integrate perform symbolic integration of a given formula (and fail if it
cannot be done). If we are going to use this procedure frequently, we may
wish to avoid recomputing integrals. To this end, we should store every
integral, once computed, and always try to nd a ready answer before
launching actual integration. Here is a possible solution:
integral( Expr, IExpr) :- stored_integral( Expr, IExpr ), !.
integral( Expr, IExpr ) :-
integrate( Expr, IExpr ),
assert( stored_integral( Expr, IExpr ) ).
In fact. we have thus furnished our program with a primitive learning
capacity.

4.4. EXAMPLES OF PROGRAM DESIGN

In this section we look at several tiny programming problems and

their solutions which result from more or less formal analysis. This is not
a real exercise in derivation of programs from formal speci cations (see
Hogger I979; Gregory I980: Burstall and Darlington I977). This is. at
best. an illustration of such derivation. not very rigorous and with formu-
lae kept as simple as possible.
These particular problems present no dif culty to experienced pro-
 4.4. Examples of Program Design ]3|

grammers, who can readily solve them without resorting to sophisticated

techniques. Simple as they are, they help demonstrate how logic formu-
lae, which lend justi cation to a program designed in a traditional way,
can also be viewed as the same program (modulo some clean transfor-
mations). Implications of this observation for logic programming are far-
reaching and largely uninvestigated; see Shapiro (l983a) for fascinating
examples of Prolog programs which are but a by-product of theoretical
considerations.
Some of the procedures discussed below can be bi-directional, but we
intentionally neglect such possibilities. As an exercise, try to discover
some of their less obvious applications.
Formulae will be written according to the conventions of Prolog-I0:
variable names are capitalized, functor names begin with small letters.

4.4.1. List Reversal

Let rev(X) denote the reverse of list X, let X with A denote the result
ofattaching A at the end oflist X (e.g.. lp. q] with r = lp. q. r]). Let X = Y
mean: X matches Y.
Assuming that X with A has already been de ned. a possible de nition
of rev is:
(4-I) rev(ll)=ll
(4.2) rev( [ A I Tail ] ) = rev( Tail ) with A
Now, recall that in Prolog we can comfortably express relations such as
the reversal of X is Y (which implicitly de nes Y as rev(X)) without
resorting to the notion of equality. To re-express (4.2) accordingly, we
begin with the introduction of a new variable to denote rev(Tail):
(4.3) T = rev( Tail ) I) rev( I A I Tail ] ) = T with A
This formula is equivalent to (4.2). Another new variable will denote T
with A:
(4.4) T = rev( Tail ) I) I TA = T with A =>
rev(lA|Tail])=TA)
which is equivalent to
(4.5) (T = rev( Tail ) /\ TA = T with A ) =>
rev(lA|Tail])=TA
We shall rewrite this implication. and the formula (4.l). using re-
ver.s'e(X. Y) instead of rev(X) = Y, and attut'h(X, Y. Z) instead 0fZ = X
tt't'lh Y:
I32 4 Simple Programming Techniques

(4.6) reverse( Tail, T ) /\ attach( T, A, TA ) =>

reverse( I A I Tail ], TA )
(4.7) reverse( l], [] )
These two formulae are exactly the logical interpretation of the following
procedure:
reverse( [ A I Tail ], TA ) :-
reverse( Tail, T ), attach( T, A, TA ).
reverse( I], [] ).
The procedure attach can be derived in a similar way:
[]withA=lA]
a-an-~. :5?~00:- /'- I [BITail]withA=[B|(TaiIwithA)]
From (4.9) we can obtain
(4.10) TA=TailwithA=>lBITaiI]withA=[BITA]
and this (together with (4.8)) is rewritten as
(4.ll) attach( Tail, A, TA ) :> attach( I B I Tail ], A, [ B I TA ] )
(4.I2) attach ( []. A. I A I )
These derivations are by no means unique. Here is another reasoning
that starts with (4.2). We rst introduce TA to denote rev([A I Tarll). and
get
(4. I3) TA = rev( [ A I Tail ] ) :> TA = rev( Tail ) with A
which is equivalent to (4.2). Now we introduce T:
(4.I4) (TA=rev([AITail])/\T=rev(Tail))=>
TA = T with A

This is easily translated into Prolog:

attach( T, A, TA ) :-
reverse( I A I Tail ], TA ), reverse( Tail, T ).
In short: we managed to de ne attach by reverse, but the de nition is only
useful if we can de ne reverse independently of attach.
Another method of reversing a list stems from its interpretation as a
stack (see Section 4.2.1). If we move the items of one stack onto another,
they will come up in reversed order. Let Stackl and Stack2 denote the
stacks before this reversal. The nal content of the second stack will be
rev( Stackl ) ++ Stack2
 4.4. Examples of Program Design I33

with X + + Y denoting the result of appending Y to X. The following two

equalities de ne rev recursively:
(4.15) rev( I] ) ++ Stack2 = Stack2
(4.16) rev( I A I Tail ] ) ++ Stack2 = ( rev( Tail ) ++ [A ] ) ++
Stack2
Now, ++ is associative, and [A] ++ Stack2 = [A I Stack2], so that we
can transfonn (4.16) into
(4.17) rev( I A I Tail ] ) ++ Stack2 = rev( Tail ) ++ [ A I Stack2 ]
Next, we introduce a new variable Final:
(4.I3) Final = rev( Tail ) + + I A I Stack2] I}
Final = rev( [ A I Tail ] ) ++ Stack2
Let reverse2(X, I, Z) denote the formula Z = rev(X) + + I. From
(4.15) and (4.18) we get
(4.19) reverse2( [], Stack2, Stack2 )
(4.20) reverse2( Tail, I A I Stack2 ], Final ) =>
reverse2( I A I Tail ], Stack2, Final )
(or, accordingly, the same in Prolog). For Y = [], reverse2(X, Y, Z) reads
Z = rev(X), so to get the reversal of L we must call
reverse2( L, [], LReversed )
indeed, Stack2 must be initially empty.
Notice that in going from (4.17) to (4.18) another direction of the
implication could have been chosen. This choice would lead to the Prolog
clause
reverse2a( Tail, [ A I Stack2 ], Final ) :-
reverse2a( [ A I Tail ], Stack2, Final ).
which de ned the shorter list in tenns of the longer. Even though it is
logically correct, operationally it is unrealistic: neither this nor (4.19)
would match the initial call with non-empty L.
We shall conclude this section with an even less formal derivation of
difference-list reversal. Let the list to be reversed be L -- Z, where L =
[A,, An I Z]. We can write
rev(L--Z)=[A,,IX]--Y
with X -- Y = rev([A,, A,,-, I W] -- W). Since W is an arbitrary term,
we can assume W = [An I Z], so that
X--Y=rev(L--[A,,IZ])
I34 4 Simple Programming Techniques

We now express the longer list by the shorter (see reverse2a above!):
rev(L--lA,,|Z])=X--Y:>
rev(L--Z)=[A,,|X]--Y
and rewrite it in Prolog, with reverse_d(X, Y) instead of rev(X) = Y:
reverse_d( L -- Z, I An I X ] -- Y) :-
reverse_d( L -- I An I Z ], X -- Y ).
The base clause,
reverse_d( Z -- Z, Y -- Y ).
must come (i.e. be tried) rst, because otherwise each call with a variable
second parameter will fall into in nite recursion (you may wish to check
this more thoroughly). This is where the peculiarities of Prolog come into
play, and obscure the so-far clean derivation. It is even worse: we have
missed one weakness of almost all Prolog implementations: the absence
of so-called occur check during uni cation (see Section I.2.3). Therefore,
the base clause matches calls with a non-empty list as the rst parameter,
if only the list ends with a variable. For example, the call
reverse_d( [ a, b I Z I -- Z, Rev )
instantiates Z < [a, b I Z] and Rev < Y -- Y. contrary to our expecta-
tions. One possible remedy is to instantiate the nal variable as [] before
going on, but to this end both clauses of reverse_d must be duplicated.
The complete procedure follows.

reverse_d( E] ~ L]: Y Y ).

reverse"d( L -" E]: [ n I X] " Y ) =-

reverse_d( L [ n]: X " Y )-

reverse_d( Z -- Z: Y ~~ Y ).

reverse_d( L ~ Z: [An I X] - Y ) =-

reverse_d( L [An I Z]: X ~ Y ).

Check that even this improved version loops for negative lists such as
[b] -- la, b]. As you see. difference lists are useful but can be rather tricky.

4.4.2. Sorting

We shall derive three procedures to sort a closed list of integers in

ascending order. The rst two implement insertion sort and a very simple
 4.4. Examples of Program Design I35

transposition sort, a variation of bubble sort. Both have running time

proportional to the square of list length (but both seem passable because
few Prolog applications require fast sorting procedures). The third proce-
dure is the simplest quicksort, which takes less time but uses more space
(the same justi cation applies).
Let X into I denote the list that results from inserting the integer X
into the ordered list Y. For example.
5into[4,7, 10] = l4,5,7, I0].
A possible de nition of insertion consists of three formulae:
(4.2I) A into I] = [A]
(4.22) A>B=>Ainto[BITail]=IBI(AintoTail)]
(4.23) A=<B:>Ainto[BITail]=[A,BITail]
The formula (4.22) can be rewritten as
(4.24) A>B/\AT=AintoTail:>Ainto[BITail]=[BIAT]
and then we can use insert(X. Y, Z) instead of X into Y = Z to get the
following procedure:
insert( A, [], I A ] ).
insert( A, [ B I Tail ], [ B I AT] ) :- A > B, insert( A, Tail, AT).
insert(A,[BITail],IA,BITail]):-A =< B.
Now, let sorted(X) denote the sorted permutation of the list X. We
can de ne sorted in the following way:
(4.25) sorted( I] ) = I]
(4.26) sorted( I A I Tail ] ) = A into sorted( Tail )
The latter formula can be replaced by
(4.27) ST = sorted( Tail ) /\ AST = A into ST I)
sorted( I A I Tail ] ) = AST
To express it in Prolog, we shall rewrite sorted(X) = Y as ins_sort(X, Y),
and get these two clauses:
ins_sort( I], I] ).
ins_sort( I A I Tail ], AST ) :-
ins_sort( Tail, ST ), insert( A, ST, AST ).
The order of calls in the second clause is not accidental: the tests in insert
require fully instantiated parameters. so the procedure would not work if
insert came rst!
Transposition sorting results from the observation that a sequence is
unordered iff it contains an unordered pair of contiguous items (e.g. A, B
I36 4. Simple Programming Techniques

such that A > B, if we consider the ascending order). Each step of a

sorting algorithm should increase the orderedness of the sequence, e.g.
by swapping A and B.
The following formula characterizes this sorting method:

(4.28) L=X++[A,BIY]/\A>B/\LT=X++[B,AIY]
=> sorted( L) = sorted( LT)
where X + + Y denotes Y appended to X. We should describe explicitly
the less ordered sequence by the more ordered, e.g. thus:

(4.29) L=X++[A,BIY]/\A>B/\LT=X++IB,AIY]
/\SL= sorted(LT):>SL= sorted(L)

Using trans_sort(X, Y) for sorted(X) = Y, and append(X, Y, Z) for

X + + Y = Z, we can rewrite (4.29) into Prolog:

trans_sort( L, SL ) :-
=1PPnd(X.lA Yl.L).A>B.
append( X, I re 2.1:: <-M ], LT), trans_sort( LT, SL ).

Suppose now that for no X, A, B, Y we have L = X ++ IA, B I Y]

/\ A > B, i.e. that L is either ordered or too short (and also orderedl).
More formally:

(4.30) (L=X++IA,BIY]/\A>B)I>L=sorted(L)

When we rewrite this in Prolog, we shall drop the premise and place the
resulting clause after the recursive one. The rst two calls in that clause
can be regarded as tests: does L contain a two-item subsequence, and is
this subsequence unordered? The clause fails if this is not the case, and
the premise of (4.30) becomes trivially true. We are left with the clause
trans_sort( L, L ).
which is exactly the required base clause: we proceed from less or-
dered sequences, so that eventually we must get an ordered permu-
tation.
The arrangement of calls in the rst clause is crucial. To begin with,
we repeatedly isolate any two contiguous items (this fails if the list is too
short), and we look at their ordering. The rst improperly ordered pair
terminates this process, and we recursively sort the improved se-
quence. The procedure is attributed to van Emden (Coelho et al. I980).
 4.4. Examples of Program Design I37

The last sorting algorithm we are going to program in Prolog is the

well-known quicksort (Hoare 1962). For a given sequence L and its ele-
ment A, let small(L, A) denote the subsequence consisting of all items
smaller than A, and large(L, A) those larger than A. Items equal to A will
fall, say, into small(L, A). The following formulae describe two possible
situations:
(4.31) sorted( I A I L ] ) = sorted( small( L, A )) ++ A ++
sorted( large( L, A ) )
(4.32) sorted( I] ) = I]
The usual transfonnations of (4.31) give, for example,
(4.33) small( L, A ) = LAs /\ large( L, A ) = LAI /\
sorted( LAs ) = SLAs /\ sorted( LAI ) = SLAI =>
sorted([AIL])=SLAs ++ [A]++ SLAI
When implementing quicksort, a standard practice is to compute
small(L, A) and large(L, A) simultaneously, i.e. to introduce
partition( L, A, LAs, LAI )
instead of the rst two equalities in (4.33). Here is the Prolog code for
partition (it can be derived in a straightforward way):
partition( I X I Tail ], A, I X I Small ], Large ) :-
X =< A, partition( Tail, A, Small, Large ).
partition( I X I Tail ], A, Small, I X I Large ] ) :-
X > A, partition( Tail, A, Small, Large ).
P iliont ll. -. ll. ll ).
The formula (4.33) should be transformed in the usual way:
(4.34) partition( L, A, LAs, LAI ) /\
sorted( LAs ) = SLAs /\ sorted( LAI ) = SLAI /\
SLAs ++ I A I SLAI ] = Sorted I)
sorted( I A I L ] ) = Sorted
This is directly expressible in Prolog, with quick_sort(X, Y) denoting the
equality sorted(X) = Y:
quick_sort( I A I L ], Sorted ) :-
partition( L, A, LAs, LAI ),
quick_sort( LAs, SLAs ), quick_sort( LAI, SLAI ),
append( SLAs, I A I SLAI ], Sorted ).
quick_sort( I], I] ).
I33 4. Simple Programming Techniques

ln the worst case. the cost of appending sorted fragments is propor-

tional to n2 for a list of length n. We can avoid appending altogether
exactly as we did in reverse2 in the previous section. We take the empty
stack, son large(L. A) and push sorted(large(L. A)) onto the stack. Next.
we stack A. and nally sorted(sma|l(L. A)).
The formulae corresponding to (4.3l). (4.32)and similar to (4.15).
(4. l6)are as follows:
(4.35) sorted( [ A I L ] ) ++ Stack2 =
sorted( sma|l( L. A )) ++ l A ] ++
sorted( large( L. A ) ) ++ Stack2
(4.36) sorted( [] ) ++ Stack2 = Stack2
We can now repeat the same reasoning and replace (4.35) with
(4.37) partition( L. A. LAs. LAI ) /\
sorted( LAs) + + [A] + + sorted( LAl ) + + Stack2=Sorted
=> sorted( I A I L ] ) + + Stack2 = Sorted
The lefthand side equality in (4.37) must be rewritten as
(4.38) sorted( LAI ) ++ Stack2 = LargeStacked /\
sorted( LAs ) ++ [ A ] ++ LargeStacked = Sorted
We introduce q_sort(X. Y. Z) for the equality sorted(X) ++ Y = Z. and
get the following procedure:
q_sort( [ A I L ]. Stack2. Sorted ) :-
partition( L. A. LAs. LAI ).
q_sort( LAI, Stack2. LargeStacked ),
q_sort( LAs. l A I LargeStacked ]. Sorted ).
q_sort( []. Stack2, Stack2 ).
And wrap it in
quick_sort_2( List. Sorted ) :- q_sort( List. ll. Sorted ).
This version of quicksort is also attributed to van Emden (Coelho er
al. I980).
Notice that the recursive calls on q_s0r! can be interchanged. A
partly uninstantiated stack will be appended to sorted(small(L. A)): the
other call will then fully instantiate the stack. Actually. the pairs Sorted.
[A I LargeStacked] and LargeStacked. Stack2 can be interpreted as dif-
ference lists. Try to derive more formally a version of quicksort with
difference lists.
 4.4. Examples of Program Design I39

4.4.3. Euler Paths

K FIG. 4.l$ A graph.

We shall try to solve in Prolog the problem of nding Euler paths in an

undirected graph. For the sake of completeness. here are the basic de ni-
tions. An undirected graph is the pair (V, 6), with V a nite set of vertice-s
and S a set of edges. A vertex is labelled with a unique name. An edge is
an unordered pair of different vertices, usually interpreted as a connec-
tion between them. A graph is often modelled by a drawing with a point
for each vertex and a line (connecting the two vertices) for each edge.
Figure 4. l5 shows a graph which consists of ve venices and eight edges.
A path from vertex X to vertex Y is a sequence of edges such that
contiguous edges share a vertex. X belongs to the rst edge. Y to the last.
For example,
(&.b).(b.).(.d).(d.)
is a path from a to e in the graph of Fig. 4.l5. The same path can be
unambiguously represented as a sequence of vertices:
abcde
An Euler path is a path passing through all vertices. in which every
edge occurs exactly once. An Euler graph is a graph that contains an
Euler path. For our graph.
dbacbecde

The problem (described as drawing a picture") was solved in Prolog by Szercdi

(I977).
'40 4. Simple Programming Techniques

is an example of Euler path. If we remove the edge be. the resulting graph
will not be an Euler graph (you may wish to check this).
We shall develop a very simple program. depending only on the most
intuitive properties, which looks somewhat blindly for Euler paths. A
more ef cient algorithm arises from a theorem that characterizes Euler
graphs. We shall quote the theorem at the end of this section.
At rst, we must choose a method of representing graphs. We can
assume that the graph contains no isolated venices (vertices which do not
belong to any edge); otherwise. it is certainly not an Euler graph. A graph
without isolated venices can be represented by its set of edges alone. An
edge is an unordered pair. i.e. a set of two vertices. Since we have no sets
in Prolog (as in most programming languages). we shall represent edges
with ordered pairs:
Vl<>V2
(<> is a non-associative in x functor). and we shall try to make the pro-
gram account for the cummutativity of pairs.
Euler graphs have the following two properties:
l. A graph with one edge is an Euler graph.
2. Suppose we take out an edge, and what remains is a Euler graph with
an Euler path starting with one of this edges vertices: then the whole
graph is an Euler graph (and we happened to have removed a tenninal
edge of an Euler path).
Let paths be represented with lists of vertex names. and let parh(E.P)
mean E is the set of edges of an Euler graph. and P is an Euler path in
this graph". The rst property above can be rewritten as two fonnulae:
(4.39) path({Vl <>V2 },[Vl,V2])
(4.40) path({V2<>Vl }, I Vl, V2])
In other words, both arrangements of vertices are equally satisfactory.
Here is how the second property can be formalized (\ denotes set
subtraction):
(4.41) path( E \ { Vl <> V2 }. [ V2 I RestofPath ] ) I}
path( E, [ VI. V2 I RestofPath ] )
(4.42) path( E \{ V2 <> Vl }. [ V2 I RestofPath I ) I)
path( E. [ Vl. V2 I RestofPath ] )
Before we rewrite (4.39-4.42) into Prolog, we must nally decide how
to represent sets. We can use any structure capable of holding uniform
data; to keep things simple we shall use lists. (Another possibility would
be to represent each edge as a separate clause, but then we would have no
 4.4. Examples of Program Design l4l

easy way of passing a set of edges as a parameter to a path- nding proce-

dure.)
The fonnulae (4.41) and (4.42) must be transformed to get rid of the
complicated expression inside path: for example. (4.41) becomes
(4.43) El = E \{ Vl <-> V2 } /\ path( El, I V2 I RestofPath I ) =}
path( E, [ VI, V2 I RestofPath ] )
With the set {VI <-> V2} represented as the list [VI <-> V2]. and
with takeout(X. Y. Z) denoting the equality X \ {Y} = Z. we can write
down the procedure path:
path( [ Vl <-> V2 ]. [ VI. V2 ] ).
path( I V2 <-> Vl ]. [ VI. V2 ] ).
path( E, [ VI, V2 I RestofPath ] ) :-
takeout( E. VI <-> V2, El ). path( El. I V2 I RestofPath ] ).
path( E. [ VI. V2 I RestofPath ] ) :-
takeout( E. V2 <-> Vl, El ). path( El. [ V2 I RestofPath ] ).
Notice that details of set representation are transparent to the recursive
clauses.
A list version of takeout can be de ned in a straightforward manner.
so we shall skip a detailed derivation:
takeout( I VI <-> V2I El ], VI <-> V2. El ).
takeout( [ Edge I Edges ], TheEdge, [ Edge I Remainder ] ) :-
takeout( Edges, TheEdge, Remainder ).
This program is crying out for optimisation: in the worst cases we can
traverse the list E twice before locating the edge to be taken out. One
solution. actually presented in Szeredi (I977). is to make takeout. rather
than path. sensitive to the order of vertices. This can be easily achieved
by adding another base clause to takeout:
takeout( [ V2 <-> Vl I El ]. Vl <-> V2. El ).
and deleting any one of the two recursive clauses of path.
The procedure path can be used non-deterministically. to produce all
Euler paths in a given graph, or with a cut, to check whether the graph is
an Euler graph (and nd an instance of Euler path). It can also be used the
other way round: given a path it computes a list that represents the Euler
graph with this path (or all such lists, but this would be overzealous).
We shall need a few more de nitions to formulate Eulers fundamen-
tal theorem on Euler graphs. A graph is connected if for each two vertices
Vl , V2 there is a path from Vl to V2. For example. the graph of Fig. 4.15
is connected. The degree of a vertex is the number of edges which contain
I42 4. Simple Programming Techniques

the vertex. For example. h in our graph is a vertex of degree 4. and e of

degree 3.
The theorem states that a graph is an Euler graph if and only if it
is connected and contains either no vertices of an odd degree. or exactly
two such vertices. In the latter case. the two odd-degree vertices are
terminal vertices of each Euler path. In the former case. each Euler path
is a cycle. i.e. a path that returns to the starting point. In our example. d
and e are the only vertices of odd degree.
If the graph is known to be an Euler graph. an Euler path can be found
in time proportional to the number of edges. Once removed. the edge can
be attached to the path for good. You may nd it amusing to modify the
above program in this direction.
 _ ittt

5 SUMMARY OF SYNTAX
AND BUILT-IN PROCEDURES

This chapter describes Prolog as de ned by Toy. the implementation

presented in Chapter 7. The supported dialect is very similar to Prolog-I0
(Pereira et al. I978. Bowen I981. Clocksin and Mellish I981); some. but
not all. differences are noted. Other standard versions will be similar:
use the appropriate reference manuals.
The user communicates with Toy through an interactive interface
(see Sections I.2.2 and 7.4.2).

5.1. PROLOG SYNTAX

A program can be regarded (roughly) as a sequence of clauses. De ni-

tions and grammar rules in the sequence are grouped in procedures. There
are quite a few principles that govem "consulting"/reconsulting". and
dynamically asserting/retracting clauses (with the rede nition switch on
or off). Therefore a formal de nition of procedures would be unnecessar-
ily involved: it should account for the fact that procedures change in time.
The notation used is extended BNF. Non-terminal symbols will be
boldfaced. and some of them subscripted (this is the rst extension). A
BNF rule takes the general form
lhsnontenn ::= rhsl I .... .. I rhs,
(lhsnontenn is either rhs, or or rhs,).
Each rhs is a sequence of non-terminals and terminals. The second
extension: zero or more occurrences of a sequence s are denoted as {s}.
To avoid confusion. terminal symbols I I} will be boldface. Throughout
the description. we assume standard operator declarations are in force.
Many special forms. such as integer expressions to be evaluated by is.
I43
I44 5 Summary of Syntax and Built-in Procedures

=:=. etc., or lists of single characters. will not be described. See Section
5.2 and on for applications of these forms in built-in procedures.
Some comments in plain English will be interspersed in the BNF
description. See also the notes at the end of this section.

clause :: = de nition I grammamlle I directive

de nition ::= nonunitclause I unitclause
nonunitclause ::= head :- body
unitclause ::= head
COMMENT the main functor of head is not a
binary :-
head ::= nonvarint
body ::= bodyalt { ; bodyalt }
bodyalt ::= call I . call }
call ::= nonvarint I variable I ( body )
nonvarint ::= term
COMMENT not a variable or an integer (a formal
de nition would be straightforward but
cumbersome)
grammarrule ::= lhside > rhside
COMMENT the arrow is written as -->
lhside ::= nonterminal context I nonterminal
nonterminal ::= nonvarint
context ::= terminals
rhside ::= altematives
altematives ::= altemative { ; altemative }
altemative ::= ruleitem I . ruleitem }
ruleitem ::= nonterminal I terminals I
condition I I I ( altematives )
terminals ::= list I string
COMMENT only closed lists are allowed
condition ::= curlyterm
directive ::= command I query
command ::= :- body
query ::= body
COMMENT body's main functor is not a unary :-
te|1n ::= tennmo
termN ::= op,._N termN_| I op,,_N IBTIIIN I
terms-I 0p.t.~ I terms ems I
WTIIIN-| oPxfx.N IIBITIIN-| I
tenns-I 0p.s.~ terms I
terms 0Pm.~ lemma I terms-I
 5. I. Prolog Syntax I45

COMMENT I =< N =< I200; op-r,,,_N is an operator of

type Type and priority N; termN can be
called term with priority N
termo ::= variable I integer I string I
IistInoop|
noop( term{ , term } ) I
( term ) I curlyterm
curlytenn ::= I term }
noop ::= functor
op-LN ::= functor
COMMENT T is one of fx, fy, xf. yf, xfx. xfy,
yfx, N is in the range l..l200; see
also note I
list=:= lllllermq-{.lrm9-}lI
Ill!l'llI999I .ll'lIl999}|lBl'III]
COMMENT terms with priority 999 can be safely
conjoined by commas which are in x
functors with priority I000
functor :: = word I qname I
symbol I soloehar
word :: = wordstart { alphanum }
wordstart ::= smalletter
alphanum ::= srnalletter I bigletter I
digit I -
qname ::= I qitem }'
qitem ::= " I nonquote
COMMENT nonquote is any character other than
symbol :: = symch I syrnch }
variable ::= varstart I alphanum }
varstart ::= bigletter I
integer ::= - digit I digit } I diglt I digit }
string ::= "I sitem }
COMMENT in Toy a string is equivalent to a list of
character names; in Prolog-I0. to a list
of their ASCII codes
tem::= nondquote
COMMENT nondquote is any character other than
smalletter::=aIbIcIdIeIfIgIhI|I
JIkIlImInI0IPIqIPI
SIIIUIVIWIXIYIZ
bigletter::=AIBIC I EIFIGI
=1:
22., *5 x|v I K0:
JIKIL OIP
SITIU
I46 5 Summary of Syntax and Built-in Procedures

digit::=0I I I2I3I4I5I6I7I8I9
symch::=.I:I-I<I=I>I+I/I
*I?I&I$I(wI#I-aI\
COMMENT a lone dot followed by white space is not a symch
but a fullstop
solochar ::= . I ; I !
token ::= functor I variable I integer I
string I hracketbar
COMMENT tokens are listed to explain note 6 below
bracketbar::=(I)I[I]I{I}II
comment ::= % I nonlineend } lineend
COMMENT lineend is an end-of-line ( linefeed )
character; nonlineend is any other
character. Toy converts line-ends to
single linefeeds
whitespace :: = I layoutchar I
COMMENT layoutchar is blank or tab or lineend
or any nonprintable character ( in
ASCII these are characters with codes
=< 3| )
fullstop ::= . layoutchar

Notes:
I. Mixed functors have not been described. but their inclusion is straight-
forward:
term ::= op|,;,_,|_N termN_|
and I I other combinations. In Toy. a mixed functor can only have one
binary and one unary type. both with the same priority.
2. There are numerous ambiguous combinations of contiguous operators.
This grammar does not account for them. See Section 7.4.3 (and Ap-
pendix A.3) for a rather detailed description in Prolog.
3. Not all functors can be declared as operators. Quoted names are al-
ways taken as normal functors.
4. In the de nition of body. commas and semicolons need not have been
actually singled out. because they are regular in x functors. The de ni-
tion
body ::= nonvarint

would not. however, emphasize the most common structure of body.

 5.2. Built-in Procedures: General Information I47

5. The syntax of directives conforms to the convention adopted in Toy.

See Section 7.4.2 for details.
6. Comments and whitespace can be freely inserted before and after a
token. and cannot be inserted in the middle of a token. Remember that
a comment extends till end-of-line.
7. Whitespace must be inserted between an unsigned integer and a minus
which is to be treated as a functor. A minus immediately preceding a
sequence of digits is taken as a part of the integer.
8. If curly brackets are not available. the usual practice is to use deco-
rated brackets: %( and %). This requires some care in the treatment
of comments.
9. A term on input must be terminated with a full stop not embedded in a
quoted name. string or comment.

5.2. BUILT-IN PROCEDURES:

GENERAL INFORMATION

For the purposes of this chapter. built-in procedures fall into two
groups. System procedures are implemented in the interpreter described
in Section 7.3. Prede ned procedures are written in Prolog; they belong to
the user interface described in Section 7.4. Together. these two groups
cover the basic set of Prolog-I0 procedures. Differences and extensions
are noted where appropriate but this is a description of Toy and is not
intended as a replacement for the Prolog-I0 manual. The procedures are
roughly classi ed according to their purpose.
A system procedure call may fail. succeed or raise an error. Failure or
success is equivalent to a failure or success of a normal procedure call.
The only difference is that success is usually accompanied by a side-
effect, such as writing a character. setting a switch, etc. A failing system
procedure does not usually cause any side-effects (input procedures are a
notable exception).
An error is raised when a system procedure detects an incorrect
parameter (or parameters). If the description of a procedure mentions the
form of expected parameters, parameters of unlisted forms will cause an
error to be raised. There is no guarantee that the error will be raised
before any actions are performed. though this is usually so.
Raising an error consists in invoking procedure error! I . with its single
parameter instantiated to the offending system procedure call. In general.
error behaves as if its call were present in the program instead of the
I43 5 Summary of Syntax and Built-in Procedures

erroneous system procedure call. An explicit call to error is also possible.

error is a Prolog procedure: the standard library contains a simple version
which outputs a message and fails. The user can augment this procedure
to his liking. possibly providing different clauses as error handlers" for
different system procedures. Rede nition of error requires removing it
from the standard library (see Section 7.4.5)in the present version it is
protected together with the whole library. Some prede ned procedures
invoke error. and so can the users programs.
error is not in Prolog-I0.
The following are conventions observed throughout this chapter. (Ad-
ditional conventions or explanations appear under some group headings.)
Whenever we say that a procedure tries to unify" we mean that it
fails or succeeds depending on the outcome. Success means that uni ca-
tion is performed.
When we say that a procedure "tests" something. we mean that it
fails or succeeds according to the result.
Acceptable parameters are indicated by conventional names listed
below:

TERMany term will do

INTEGERan integer
VARa variable
NONVARINTa non-variable, non-integer term
CALLsame as NONVARINT
ATOMa NONVARINT without arguments
NAMEsame as ATOM
CHARa NAME consisting of a single character
FILENAMEa NAME conforming to the implementation-dependent
conventions for specifying les
CALLISTa list (possibly empty) of CALLs
CHARLISTa list (possibly empty) of CHARs
DIGITLISTa CHARLIST built of digit characters

In descriptions, PARI , PAR2 etc. stand for actual parameters in the built-
in procedure call.
Note that I23 is a name, and I23 an integer. 9 is the integer nine. and
'9' is the digit (character). The output procedures do not always distin-
guish between the two (writeq does).
Toy introduces a number of prede ned operators. Some of them are
used as in x or pre x procedure names. Table 5.l is the list of prede ned
operators:
 5.3. Convenience I49

TABLE 5.1
Prede ned Operators

Name Type Priority

xfx I 200
fx I 2(1)
--> xfx I 2(1)
xfy I III)
xfy I (D0
l'|OI fv 9%
xfx 7(1)
is xfx 7(1)
xfx 7(1)
=\= xfx 7(1)
< xfx 7(1)
=< xfx 7(1)
> xfx 700
>= xfx 700
@< xfx 700
@=< xfx 700
@> xfx 7(1)
@> = xfx 7(1)
xfx 7(1)
\== xfx 7G3
xfx 7(1)
+ yfx SM
+ fx 500
yfx 500
fx 511]
yfx 4-00
I yfx 4-00
mod xfx 3(1)

5.3. CONVENIENCE

true
always succeeds
fail
always fails.
not CALL
the not procedure (but see Section 4.3.2!): succeeds only when
the parameter fails. De ned in Prolog:
not C :- C. !, fail.
I101 _.
I50 5 Summary of Syntax and Built-in Procedures

CALL . CALL
the and procedure: succeeds only when both arguments succeed.
De ned in Prolog:
A, B :- A. B.

See also the description of the cut.

CALL ; CALL
the or procedure: succeeds only if either of the parameters suc-
ceeds. De ned in Prolog:
A;_:- A.
_; B :- B.
See also the description of the cut.
check(CALL)
succeeds only when the parameter succeeds. but instantiates no vari-
ablesonly side-effects of CALL remain. De ned in Prolog:
check( Call ) :- not not Call.
Not in Prolog-I0.
side_effects(CALL)
exactly equivalent to check(CalI). but used when the parameter is to
be executed for its side-effects rather than to test something. Not in
Prolog-I0.
once(CALL)
executes CALL deterministically. De ned in Prolog:
once( Call ) :- Call. I.
Not in Prolog-I0.

5.4. ARITHMETIC

In the descriptions. div stands for integer division, and mod for taking
the remainder of integer division.
The following are correct invocation patterns for sum/3 (not in Pro-
log-I0).
sum(INTEGER. INTEGER. INTEGER)
succeeds only if PARI + PAR2 = PAR3
sum(INTEGER. INTEGER. VAR)
succeeds after unifying PAR3 with the value of PARI + PAR2
sum(INTEGER, VAR, INTEGER)
succeeds after unifying PAR2 with the value of PAR3 PARI
 5.5. Comparing Integers and Names I51

sum(VAR. INTEGER, INTEGER)

succeeds after unifying PARI with the value of PAR3 PAR2

The following are correct invocation patterns for prod/4 (not in Pro-
log-I0).
prod(INTEGER, INTEGER. INTEGER, INTEGER)
succeeds only if PARI * PAR2 + PAR3 = PAR4
prod(INTEGER. INTEGER. INTEGER. VAR)
succeeds after unifying PAR4 with the value of PARI * PAR2 +
PAR3
prod(INTEGER. INTEGER, VAR. INTEGER)
succeeds after unifying PAR3 with the value of PAR4 PARI *
PAR2
prod(INTEGER. VAR, VAR, INTEGER)
succeeds after unifying PAR2 with the value of PAR4 div PARI and
PAR3 with the value of PAR4 mod PARI
prod(VAR. INTEGER. VAR. INTEGER)
like the previous one. but with PARI and PAR2 exchanged
prod(INTEGER. VAR, INTEGER. INTEGER)
fails if (PAR4 - PAR3) mod PARI is not zero; otherwise succeeds
after unifying PAR2 with the value of (PAR4 - PAR3) div PARI
prod(VAR. INTEGER. INTEGER. INTEGER)
like the previous one. but with PARI and PAR2 exchanged

TERM is TERM
the procedure is assumes PAR2 is an integer expression. i.e. a tenn
composed of integers by means of standard arithmetic functors: +
(binary and unary). (binary and unary), 1-. /. mod. The procedure
fails if PAR2 is not an integer expression. Otherwise it evaluates the
expression and tries to unify the value with PARI. According to
Prolog-I0 conventions. is can also evaluate a list
[ INTEGER ]
as this INTEGER; e.g. 55 is I55] succeeds. (This is needed in Prolog-
I0 mainly for evaluating single character strings to ASCII codes.)
De ned in Prolog.

5.5. COMPARING INTEGERS AND NAMES

less(INTEGER. INTEGER)
succeeds only if PARI < PAR2. Not in Prolog-I0.
I52 5 Summary of Syntax and Built-in Procedures

TERM =:= TERM

PARI and PAR2 are treated as integer expressions and evaluated.
The procedure succeeds only if both parameters are proper integer
expressions (see is/2) and their values are equal. De ned in Prolog.
TERM =\= TERM
as above. but tests whether the values are nonequal
TERM < TERM
as above. but tests whether the value of PARI is less than that of
PAR2
TERM =< TERM
as above. but tests whether the value of PARI is not greater than that
of PAR2
TERM > TERM
as above. but tests whether the value of PARI is greater than that of
PAR2
TERM >= TERM
as above. but tests whether the value of PARI is not less than that of
PAR2
NAME @< NAME
succeeds only when PARI precedes PAR2 in the lexicographic order
(as de ned by the underlying ASCII collating sequence).
NAME @=< NAME
like @<, but tests whether PAR2 does not precede PARI. De ned in
Prolog.
NAME @> NAME
like @<, but tests whether PAR2 precedes PARI. De ned in Prolog.
NAME (w>= NAME
like @<, but tests whether PARI does not precede PAR2. De ned in
Prolog.

5.6. TESTING TERM EQUALITY

TERM = TERM
tries to unify PARI and PAR2. De ned in Prolog:
X=X.
eqvar(VAR, VAR)
succeeds only when the parameters are two occurrences of the same
nondummy variable. Not in Prolog-I0.
 5.7. Input/Output I53

TERM == TERM
succeeds only when the parameters are two occurrences of the same
term. For example, if A. B are uninstantiated.

P( A ) = = P( B )
fails. even though
P(A)=P(B)
succeeds. De ned in Prolog.
TERM \== TERM
succeeds only when the parameters are not two occurrences of the
same tenn. De ned in Prolog.

5.7. INPUT!OUTPUT

5.7.1. Switching Streams

This set of procedures can be used to dynamically change the les

read or written by the input/output procedures. The users terminal is
treated like any other le: its name is user (both for input and output); the
terminal is read from and written on by default.
Ideally. one should be able to open a le with tell or see. stop using it
with another tell or see. start using it from the current position after a
second tell or see. and close it with told or seen. There should be no limits
on the interleaving introduced by using a le in the middle of using a le in
the middle etc.
The procedures are described as if this situation were real. In prac-
tice. things are very implementation-dependent. The version of Toy pre-
sented in Chapter 7 has only two input and two output streams: one for the
terminal and one for a disk le in each direction. Also. Toy has no code for
dealing with incorrect le names. nonexistent les and the like. All this is too
dependent on the environment in which it is implemented.
sec(FILENAME)
the speci ed le becomes the current input le; the terminals name
is user
seeing(TERM)
tries to unify the parameter with the name of the current input le
I54 5 Summary of Syntax and Built-in Procedures

seen
closes the current input le; user becomes current. Has no effect if
the current le is user
teIl(FILENAME)
the speci ed le becomes the current output le; the terminals name
is user
telling(TERM)
tries to unify the parameter with the name of the current output le
told
closes the current output le; user becomes current. I-Ias no effect if
the current le is user

5.7.2. Listing Control

The Toy-Prolog interpreter contains a listing switch. If the switch is

on. each line read in from the current input is listed on the users terminal:
this is useful when one wants to see what is being read from a disk le.
echo
succeeds after turning the listing switch on; has no effect if the switch
is already on. Not in Prolog-I0.
noecho
succeeds after turning the listing switch off; has no effect if the
switch is already off. Not in Prolog-I0.

5.7.3. Terms

display/(TERM)
writes the term onto the current output. The term is written in stand-
ard notation (pre x with parentheses) and identi ers are not quoted
even if they normally should be. Variables are written as _n. where
n is an address. There is no guarantee that a variable will be printed
as the same address in different invocations of display. In Prolog-I0.
display is a little different: it always writes on the user's terminal.
write(TERM)
writes the term onto the current output. The term is written accord-
ing to operator declarations currently in force. No identi ers are
quoted. Variables are written as X I . X2 etc. Each invocation of write
begins numbering from I. so that e.g. the calls
write( X ). write( f( Y. X ) )
will produce
XIf( XI. X2 )
 5.7. Input/Output I55

De ned in Prolog. CAUTION: in Toy. write uses numbervars (see

Section 5.I5) which binds variables in the term to V(N) for N = I.
2. etc. Hence, write cannot output any term 'V(INTEGER) prop-
erly.
wr|teq(TERM)
same as write. but quotes identi ers that are not proper words or
symbols. and also those identi ers that coincide with operator
names; e.g. a 3-parameter is would be quoted. However. a quote
within a quoted name will not be doubled (this is a bug. actually).
Otherwise. a term written by writeq can be read back by read.
read(TERM)
reads from the current input a term. terminated with a full stop.
Succeeds only when PARI uni es with this term. Operator declara-
tions currently in force are taken into account. Recall that a quoted
name cannot be an operator. If the text on input is not a correct term.
read prints the message

+ ++ Bad term on input. Text skipped:

skips and reprints the input until the rst (still unprocessed) full stop.
and tries to unify PARI with e r r. (If the erroneous line does not
contain a full stop. you should input one before Prolog resumes.) See
the next section for behaviour on le end detecting. De ned in
Prolog in terms of single-character input (see the next section).
op(INTEGER. TERM. ATOM)
declares an operator with PAR3the name. PARIthe priority
(I =< PARI =< I200. and PAR2the type. PARI is usually less
than I000. to avoid con icts with clause-constructing operators (see
the table in Section 5.2); operators with lower priority take prece-
dence over those with a higher priority. PAR2 must be a proper word
or symbol. Admissible types of operators are fx. fy (unary. pre x);
xf. yf (unary. post x); xfx. xfy. yfx (binary. in x). The types fx. xf.
xfx are non-associative; fy, yf. associative; xfy. right-associative;
yfx. left-associative. Any other PAR2 causes an error.
If an operator declaration with this name but another priority is al-
ready in force. the procedure replaces the old declaration with the
new one. If a declaration with the same name and priority exists.
three possibilities arise:

both operators are binary or both unary; the old de nition is re-
placed;
the old operator is unary (binary). the newbinary (unary); a
mixed functor is declared;
I56 5 Summary of Syntax and Built-in Procedures

the old operator is mixed. the newbinary (unary); the binary

(unary) type in the mixed functor declaration is replaced with
PAR2.
De ned in Prolog.
delop(ATOM)
the operator declaration with the name given by PARI is deleted. The
name should be quoted to prevent it from being treated as an (errone-
ous) operator with missing arguments. De ned in Prolog. Not in
Prolog- I0.

5.7.4. Single Characters

The Toy interpreter contains a single-character input buffer called the

current character. Initially. it contains a blank and is then re lled by each
reading operation. In the presented version. each line end is treated as if it
were a linefeed character (ordinal number I0. see the procedure iseoln).
Behaviour upon detection of end-of- le depends on the current input. If
the input is user (i.e. the terminal). Prolog is terminated; otherwise an
automatic seen is performed and the reading operation is restarted.
The operations presented here (except nl) differ from those in Prolog-
l0. In Toy, the arguments of input/output operations are characters, and
the internal buffer can be used to rescan the current input character. In
Prolog- I0 there is no such buffer and the arguments of the operations are
integers. i.e. character codes. These operations could be de ned as
follows:
get0( Ord ) :- rch. lastch( Ch ). ordchr( Ord. Ch ).
get( Ord ) :- rch. skipbl. lastch( Ch ).
ordchr( Ord. Ch ).
skip( X ) :- repeat. get0( X ). !.
put( Ord ) :- ordchr( Ord. Ch ). wch( Ch ).
This would not assure complete compatibility. however. Erroneous calls
would be handled a little differently and so would line ends. See also the
description of strings.
rch
succeeds after lling current character with the next character from
current input (but see the introductory remarks for effects of line end
or end-of- le)
skipbl
succeeds after ensuring that current character is a printing character
 5.7. Input/Output I57

with ordinal number greater than 32. Does nothing if it already is such
a character; otherwise repeatedly invokes rch.
lastch(TERM)
tries to unify its parameter with current character
wch(CHAR)
writes the character on current output (the linefeed character is inter-
preted as line terminator)
nl
terminates the current output line. De ned in Prolog:
:- ordchr( I0. Ch ), assert( ( nl :- wch( Ch ) ) ).

rdch(TERM)
gets the next character from current input (by invoking rch). Makes a
copy of current character. treating a non-printing character (including
line end) as a blank; tries to unify the copy with its parameter. De-
ned in Prolog.
rdchsk(TERM)
same as above. but preceded by a call on skipbl

5.7.5. Others

These procedures are not really concemed with input/output. but the
only effect of status is to write something. and ordchr is most useful when
reading or writing non-printing characters. They all are not in Prolog-I0.

ordchr(INTEGER, CHAR)
succeeds only when PARI is the ordinal number (ASCII code) of PAR2
ordchr(VAR. CHAR)
succeeds after unifying the variable with the ordinal number of the
character
ordchr(INTEGER. VAR)
succeeds after unifying the variable with the character whose ordinal
number is the value of PARI mod I28
iseoln(TERM)
tries to unify PARI with the end-of-line character. De ned in Prolog:
:- ordchr( I0. Ch ). assert( iseoln( Ch ) ).

status
writes memory utilisation information on the current output
See also consult/I. reconsult/I. listing/0 and listing/I in Section 5.ll.
I58 5 Summary of Syntax and Built-in Procedures

5.8. TESTING CHARACTERS

Each of these procedures fails or succeeds depending on whether its

parameter is a character belonging to a particular class. They are designed
to help the user interface (see Section 7.4) in reading Prolog temts. but
some of them are of general utility. The procedures iseoln/I and ordchr/2
(see Section 5.7.5) are also used to test characters.
smalletter(TERM)
tests whether the parameter is a lower case letter
bigletter(TERM)
tests whether the parameter is an upper case letter
letter(TERM)
tests whether the parameter is an upper or lower case letter
digit(TERM)
tests whether the parameter is a decimal digit
alphanum(TERM)
tests whether the parameter is a letter. a digit or an underscore char-
acter
bracket(TERM)
tests whether the parameter is one of the following characters:
()ll{}
solochar(TERM)
tests whether the parameter is one of the following characters:
I .
~19

symch(TERM)
tests whether the parameter is one of the following characters:
+--/=@#s&=.'.><>\

5.9. TESTING TYPES

These procedures fail or succeed depending on the form of their

arguments.
var(TERM)
tests whether the parameter is an uninstantiated variable
integer(TERM)
tests whether the parameter is an integer
 5.I0. Accessing the Structure of Terms I59

nonvarint(TERM)
tests whether the parameter is a NONVARINT (neither a variable
nor an integer); not in Prolog-I0
atom(TERM)
tests whether the parameter is an atom (a NONVARINT without
arguments)

5.10. ACCESSING THE STRUCTURE OF TERMS

The procedures pname and pnamei are not in Prolog-I0. They replace
name/2. which is similar. but which uses lists of integers (ASCII codes) in
place of our lists of characters (see Section 5.7.4).

pname(NAME. TERM)
builds a list of characters fomting the name and tries to unify it with
PAR2
pname(VAR. CHARLIST)
succeeds after unifying the variable with a NAME formed of the
characters on the list. (Note that pname(X. II. 2. 3]) binds X to the
name I23. and not to the integer I23).
pnamei(INTEGER. TERM)
builds a list of decimal digit characters (constituting the written form
of the integer) and tries to unify it with the term; the integer must not
be negative.
pnamei(VAR, DIGITLIST)
succeeds after unifying the variable with an integer whose written
form is given by the digit characters on the list. Even when the
parameters are formally correct, an error may be raised if the speci-
ed integer is too large.

functor(VAR. INTEGER. 0)
PAR3 is the integer zero; succeeds after unifying the variable with
PAR2 (this version is allowed for completeness. see below for sensi-
ble uses of functor)
functor(VAR. NAME. INTEGER)
succeeds after unifying the variable with a term whose main functor
has the name and atity de ned by PAR2 and PAR3, and whose argu-
ments are different variables; PAR3 must not be negative
functor(INTEGER. TERM. TERM)
tries to unify PAR2 with PARI and PAR3 with the integer zero
I60 5 Summary of Syntax and Built-in Procedures

functor(NONVARINT. TERM. TERM)

tries to unify PAR2 and PAR3 with the name and arity of the main
functor in PARI

arg(lNTEGER. NONVARINT. TERM)

fails if the integer is smaller than I or greater than the arity of the
main functor in PAR2. Otherwise tries to unify PAR3 with that argu-
ment of PAR2 whose number is given by PARI.
The following are correct invocation patterns for the procedure =..
(pronounced univ). which is de ned in Prolog.
VAR =.. [INTEGER]
succeeds after unifying the variable with the integer
VAR =.. [NAME I TERM]
if the term is not a closed list. an error in the procedure length/2 is
raised (=.. uses length). Otherwise a term with NAME as its name
and TERM as its argument list is created and uni ed with VAR.
INTEGER =.. TERM
tries to unify the term with [PARI]
NONVARINT =.. TERM
constructs a list. with PARI s main functor as the head and the list of
PARls arguments as the tail. Tries to unify the list with PAR2.

5.11. ACCESSING PROCEDURES

The Toy-Prolog interpreter supports assert/3, retract/3 and clause/5.

These are low-level. but quite powerful procedures (see the editor in
Appendix A.4). Parameters representing clause bodies have the form of
lists of calls (an empty body is []).
The Prolog library uses these low-level routines to de ne Prolog-I0
procedures assert/I. assertal I. assertzl I. retract! I and clause/2. Unlike
most other built-in procedures. retract/I and clause/2 are non-determinis-
tic. Parameters representing clause bodies have the form of terms used in
the external representation. i.e. sequences built with commas (an empty
body is true).
An attempt to apply any of these procedures to system routines de-
ned by the interpreter is treated as an erroneous call. So is an attempt to
modify protected procedures. (There is a diagnostic printout which can-
not be suppressed by rede ning errorl I.)
Caution: Remember the standard operator declarations listed in Sec-
tion 5.2. To be safe. always enclose a clause-representing term in paren-
 5. I I. Accessing Procedures I6]

theses. For example.

assert((a :-b.c))
rs okay. but
assert( a :- b. c )
rs a call on assert/2. In some versions of Prolog. even
assert( a :- b )
rs incorrect.
assert(NONVARINT. CALLIST. INTEGER)
PARI is treated as a clauses head. PAR2 as its body. The clause is
asserted immediately after the n-th clause of this procedure (where n
is PAR3 if a clause with this number exists, and the last clauses
position if PAR3 is too large; if the procedure is empty or PAR3 < I.
the clause is asserted as rst). Not in Prolog-I0.
retract(NAME. INTEGER. INTEGER)
PAR2 must not be negative. PARI and PAR2 de ne the name and
arity of a predicate symbol. If the associated procedure does not
contain a clause whose number is given by PAR3 (the rst clause has
number I). retract fails. If the clause does exist. it is logically re-
moved from the procedure and retract succeeds. A removed clause
does not disappear from storage and its active instances can still run
to completion. Not in Prolog-I0.
clause(NAME. INTEGER, INTEGER. TERM. TERM)
PAR2 must not be negative. PARI and PAR2 are treated as the name
and arity of a predicate symbol. If the associated procedure has no
clause whose number is given by PAR3 (in particular. if it is a system
routine) then clause fails. Otherwise it tries to unify PAR4 with the
head of the clause and PARS with its body. Not in Prolog-I0.
asserta(NONVARINT)
treats the parameter as a clause (non-unit if its main functor is :-/2.
unit otherwise). An error is raised if the rst argument of a :- is not a
NONVARINT. Asserts the clause at the beginning of its procedure.
creating the procedure if it does not exist. De ned in Prolog.
asse rtz(N ON VARINT)
same as above. but the assertion is at the end of the procedure.
assert(NONVARINT)
equivalent to asserta(PAR I ).
retract(NONVARINT)
the parameter is treated as a clause (non-unit if its main functor is :-/2
and unit otherwise). An error is raised if the rst argument of :- is not
I62 S Summary of Syntax and Built-in Procedures

a NONVARINT. The rst matching clause is retracted. and a fail

point created (see Section 5.12). On failure. the next matching clause
will be retracted. Note: if PARI has the form nonvarint:-var then it
matches only clauses with a single call in their bodies. De ned in
Prolog.
clause(NONVARlNT. TERM)
tries to locate the rst procedure whose head matches PARI and whose
body matches PAR2; the body of a unit clause is the term true. After
successful uni cation. establishes a fail point and succeeds; the next
matching clause is sought on failure. De ned in Prolog. Not in
Prolog-I0.
rede ne
this procedure is needed to implement reconsult and should not be
used directly. It modi es the effects of assert: if the procedure to
which a clause is added is different from that affected by the last
assertion. an automatic abolish is invoked before the assert. The next
invocation of rede ne restores the original situation.
protect
succeeds after ensuring that all procedures already de ned. except
those whose heads are single characters with no arguments (this
restriction is imposed by a minor technical dif culty). are protected.
An attempt to modify a protected procedure (by means of assert,
retract. abolish. consult. reconsult) is treated as an erroneous invo-
cation of the system procedure in question. (The user interface in
Toy protects all its procedures.) Not in Prolog-I0.
abolish(NAME. INTEGER)
PAR2 must not be negative. PARI and PAR2 are treated as the name
and arity of a predicate symbol. All the clauses of this procedure are
logically removed (retracted) and abolish succeeds.
prede ned(NAME. INTEGER)
PAR2 must not be negative. PARI and PAR2 are treated as the name
and arity of a predicate symbol. If the procedure associated with this
symbol is a system procedure. prede ned succeeds; otherwise it
fails. Not in Prolog-I0.
consuIt(FILENAME)
sees the named le and enters program de nition mode: successive
terms are read-in and stored via assertz (see the convention for
assertas parameters) and asserted (but see protect). There are two
exceptions: the term end causes the le to be closed and de nition
mode to be exited: terms with the unary :- as a main functor are
treated as commands. and immediately executed. De ned in Prolog.
reconsult(FILENAME)
as above. but rede ne is called at the beginning and at the end of
 5.|2. Control I63

processing. Contiguous sequences of clauses with the same predicate

symbol in their heads are treated as complete de nitions of proce-
dures and supersede previous de nitions.
listing(NONVARINT)
PARI must be an ATOM. or a term of the form ATOM! INTEGER or
a list of such terms (possibly multi-level). Each atom is treated as a
procedure's name, each integer as a procedures arity. All relevant
procedures are listed on the current output. De ned in Prolog.
listing
as above, but for all de ned procedures (including procedures de-
ned in the monitor and library, but excluding built-in system proce-
dures).

5.12. CONTROL

Whenever a procedure call activates a clause which is not the last

clause in its procedure, we say that a fail point is associated with the call.
A fail point is something to backtrack to: it saves information necessary
for reestablishing the state of the computation and proceeding with the
next clause.
The immediate descendants of a call C are the calls in the procedure
which C activated. The immediate ancestor of a call C is the call which
activated the procedure containing C. An ancestor is the immediate an-
cestor or an ancestor of the immediate ancestor. A descendant is de ned
similarly.
!
the cut procedure; succeeds after nding the nearest ancestor which
is not a call/I, tag/2, ,/2 or ;/2 and removing all existing fail points
associated with this ancestor and all its descendants.
repeat
an endless generator of successes (see Section 4.3.2). De ned in
Prolog:
repeat.
repeat :- repeat.
calI(CALL)
behaves exactly as if its parameter were in its place, with the excep-
tion that an incorrect parameter (an integer or uninstantiated varia-
ble) is detected at run time rather than at clause-de nition time. In
top-level syntax, one can use a variable instead of a predicatethis
is converted to an invocation of call.
I64 S Summary of Syntax and Built-in Procedures

halt(ATOM)
stops the interpreter after writing the atom. Not in Prolog-I0.
stop
stops the interpreter. Not in Prolog-I0. De ned in Prolog.
The following procedures are not in Prolog-I0. They are useful for error
handling, but are dirty. and should be used sparingly.
tag(CALL)
this is a form of call/I which can be referred to by tagfail/I . tagexit/2,
tagcut/2 and ancestor/2. The parameter of tag is called a tagged
ancestor of its descendants; it is never removed from the stack as a
result of tail recursion optimisation (see Sections 6.4 and 7. I).
NOTE: a tag is recognized only when explicitly written in its clause.
In particular call(tag(C)) is equivalent to call(call(C)).
ancestor(TERM)
searches for the nearest tagged ancestor uni able with the parameter;
fails if no such ancestor is found. otherwise uni es and succeeds.
tagcut(TERM)
searches for the nearest tagged ancestor uni able with the parameter.
Fails if no such ancestor is found: otherwise uni es. removes all
existing fail points associated with the ancestor and its decendants
and succeeds.
tagfail(TERM)
equivalent to
tagcut( PARI ), fail
i.e. if the appropriate tagged ancestor is found. the ancestor fails
immediately; otherwise tagcut fails.
tagexit(TERM)
searches for the nearest tagged ancestor uni able with the parameter;
fails if no such ancestor is found. otherwise uni es and passes control
to the ancestor, which succeeds immediately.

5.13. DEBUGGING

The built-in debugging facilities of Toy are very primitive. There is

only a wall-paper trace which displays all calls with a plus or a minus to
indicate success or failure respectively (e.g. if a call fails to match two
clauses and activates the third, it is shown twice with a minus and once
with a plus).
 5.l5. Miscellaneous I65

A more usefulselectivetracer is listed in Appendix A.5.

There is also a switch which may cause the interpreter to output
warning messages upon encountering calls on non-existent procedures. It
is good practice to turn it on when debugging a program.
All these procedures are not in Prolog-I0, which has a more sophisti-
cated set of debugging aids.
debug
succeeds after turning tracing on (no effect if already on)
nodebug
succeeds after turning tracing off (no effect if already off)
nonexistent
succeeds after turning on warning about calls on nonexistent proce-
dures (no effect if already on)
nononexistent
succeeds after turning off warning about calls on nonexistent proce-
dures (no effect if already off)

5.14. GRAMMAR PROCESSING

phrase(CALL, TERM)
treats CALL as a nonterminal symbol of a grammar rule. schemati-
cally
nt( ARGI, ..., ARGn ).
and initiates grammar processing-with this initial symbolby
calling
nt( TERM, []. ARGI, ARGn)
De ned in Prolog (see Section 4.2.6).

5.15. MISCELLANEOUS

length(NONVARINT. TERM)
PARI must be a closed list. Computes the length of this list and tries
to unify the resulting integer with PAR2. De ned in Prolog.
isclosedlist(TERM)
succeeds only when the term is a closed list. De ned in Prolog. Not
in Prolog-I0.
I66 5 Summary of Syntax and Built-in Procedures

numbervars(TERM, INTEGER, TERM)

instantiates PAR I s variables as V(i), V(i + I), ..., V(j) where i, i+ I,
..., j are consecutive integers, and i is the value of PAR2. Variables bound
together are of course instantiated as the same V(k). As a result, PARI
becomes ground (obviously, this is undone upon backtracking). Then the
procedure tries to unify PAR3 with j+ I. For example, the call
numbervars( [ X, Y, X ], 6, Next )
where Next is uninstantiated, will instantiate
X<'V'(6), Y<'V'(7), Next 4-3
The call
numbervars( [ X, Y, X ], 6, not_a_number)
will fail.
member(TERM. TERM)
establishes the relationship: PARI is a member of the list PAR2.
De ned in Prolog:
member( X. I X I Y ] ).
member( X, [ _ I Y ] ) :- member( X, Y ).
bagof(TERM, CALL, TERM)
tries to unify PAR3 with the list of PARIs instantiations after all
possible computations of PAR2 (see Section 4.2.4 for details). Pro-
log-I0 has a more sophisticated version of this procedure. De ned in
Prolog.
6 PRINCIPLES OF PROLOG
IIVIPLEIVIENTATION

6.1. INTRODUCTION

This chapter is but a bird's-eye view on implementation techniques

speci c to Prolog. We assume you know how conventional block struc-
ture languages are implemented: a competent programmer could hardly
escape learning these things. The discussion is kept at a level free of
representation details. Chapter 7 provides a rather detailed and complete
case study of one of the many ways in which the basic principles can be
applied in practice.
Two topics are missing: compilation and garbage collection. To com-
pile Prolog programs is to apply the general principles in such a way that a
program is executed particularly ef ciently. This is done partly by taking
advantage of the underlying machine (e.g. by using machine code instead
of a more compact representation of programs. trading speed for memory)
and partly by performing special case analysis to detect operations which
can be simpli ed (e.g. uni cation with a variable which is known to be
uninstantiated). We decided that compilation is beyond the scope of this
book (which already discusses implementation issues more thoroughly
than the usual introduction to a programming language). The problem and
techniques of garbage collection are well known, and are best studied
independently of a particular programming language (though you will nd
that in Prolog one has to do with one of the harder variants of the
problem).
I67
I63 6 Principles of Prolog Implementation

6.2. REPRESENTATION OF TERMS

If we disregard the possibility of forming cyclic structures (see Sec-

tion l.2.3), we can see that all terms are directed acyclic graphs (DAGs).
They are not necessarily trees. because different branches can converge
to a common component: in linear notation we express this phenomenon
by repetition, as in t(p(X), q(p(X), Y)).
In a Prolog program. several identical occurrences of a term within a
single clause denote the same object. Properly speaking, this is not an
object but a descriptor, or template. At execution time, it corresponds to
different objects in different instances of the clause. In this and the next
chapter we shall reserve the unadorned word term for tenn instances.
Terms written in a program will be referred to as tenn descriptions. A
description can have several occurrences; similarly. an instance can have
several parents in a DAG.
There are many possible representations of a DAG. For our present
purposes they are all equivalent, provided that it is possible to distinguish
nodes corresponding to Prolog variables. On a more abstract level. how-
ever, two very different methods are used to implement term instances.
Accordingly. all existing implementations of Prolog can be classi ed as
either Structure Sharing or Non-Structure Sharing (NSS).
In principle. to form a new term instance in a Non-Structure Sharing
system. one must create a new DAG. We are talking about creating new
instances that correspond to term descriptions (present in the program
text, or in clauses asserted after having been constructed by a program);
creation of new terms as a result of uni cation is different. Variables are
bound by being associated with pointers directed at their instantations.
These pointers are invisible. i.e. automatically dereferenced, whenever
the DAG is traversed. Figure 6.l illustratesin a representation-indepen-
dent manner-two terms, before and after uni cation.
A Structure Sharing system takes advantage of the fact that different
instances of the same term differ only in their variable bindings. Whereas
two instances of
t(p(X).q(p(X).Y))
canbe
l(P(c).q(P().d)) and
t(P(r(a)).l(P(r(a)).I'(a)))
respectively, their general structure remains the same. The main functor
must be a t of two arguments; ts rst argument must also be the rst
 6.2. Representation of Terms I69

(o) terrnl= U term2= T

I . 7).
./\@ .
(bl IE: If 10??

I P B

'/\
I I
FIG. 6.l The Non-Structure Sharing representation of terms: (a) t(A. q(A. Y))
and t(p(X). B) before uni cation. (b) t(A. q(A. Y)) and t(p(X). B) instantiated to t(p(X).
q(p(X), Y)) after uni cation.

argument of the two-argument q which is ts second argument; and so on.

Consequently, all instances of the term may share this structural informa-
tion, if only care is taken to let them have different variables. This is
easily achieved by associating each instance with a different variable
frame: a chunk of storage holding variable instances. The intemal repre-
sentation of a term descriptionwe shall call it prototype-is a DAG in
which each variable node is represented by infomration about the offset of
the variables location in a variable frame. All tenns-including variable
bindingsare now represented not by single pointers, but by two-pointer
tenn handles
< prototype, variable frame >

' Another ten'nin0IOBy. introduced by Warren (I977a). is to call prototypes skeletons.

and handles molecules. We do not like the mixed metaphor.
 (0) l1r'.!Ji=< ' >

t _. , 0

'\ I
p Y: -
GIEII)

><= '\
prototype variable trarnes

(bl terrnt:< v >

0
t _.,
....
Q'l'I'I'l I ... I
d

0 <2 >

I
\/ Y.
\
'

I
I

X: 0 3

FIG. 6.2 The Structure Sharing representation of terms: (a) Two instances of t(p(X).
q(p(X), Y)). both sharing the same prototype. (b) tennl instantiated to t(p(c). q(p(c). d)) and
term2 instantiated to t(p(r(a)). q(p(r(a)). l'( )))-
 6.2. Representation of Terms I-/I

Figure 6.2 illustrates the principle of Structure Sharing. Figure 6.3

corresponds to Fig. 6.l. If we nd general DAGs less convenient than
trees, Structure Sharing makes it easy to employ trees by providing im-
plicit links to variables from all occurrence sites. This is shown in Fig. 6.4.
Inside a clause. different occun'ences of the same variable description
can appear within different term descriptions. There is the problem of

(O) mml-< r >\ , .>

t \

I
I
Q P WI
GED

as-\-.\.9/ =<=

(b)?.> 7:.-

. \ . \ I
Q ?
/\ 21
FIG. 6.3 The Structure Sharing representation of tenns: (a) t(A.q(A. Y)) and
t(p(X). B) before uni cation. (b) t(A. q(A. Y)) and t(p(X). B) instantiated to t(p(X).
q(p(X), Y)) after uni cation.
I72 6 Principles of Prolog Implementation

< 1 >

/\q ?
/
FIG. 6.4 Structure Sharing: the DAG t(p(X). q(p(X), Y)) represented by a tree.

ensuring that the same variable becomes a part of all the corresponding
terms associated with a clause instance. With Structure Sharing, this is
done by allocating a single frame for all the variables appearing in a
clause, at the moment of its activation. All occurrences of a variable
within this clause's prototypes are encoded as the same ojfset in the
common variable frame (this is just an application of the technique dem-
onstrated in Fig. 6.4).
In practice, most NSS implementations use a very similar approach
to solve the problem (despite its name. it is a hybrid method). Term
instances are also encoded as prototypes, with variables represented by
offsets into a clause's variable frame. One difference is that variable
frame locations hold only single pointers rather than term handles. The
othermore importantdifference is that terms formed in this way are
only virtual instances. This is to say that they may be used only as data
selectors, directing uni cation to instantiate variables in the variable
frame. Whenever one of these terms is to become a variables instantia-
tion. a real instance (a new DAG) must be built. If this new instance
contains a variable. its variable node becomes a copy of the appropriate
location in the variable frame, while the location is made to hold a pointer
to the node. This ensures that all future references to the variable will end
up in the node.
The process is shown in Fig. 6.5. Note that here, too. prototypes can
 I ... ....

,;.. EB

@._. @._.:. B@/P

\n or
E._.. t\/ .3..-/
8
E BE
/I\ .\9
./ @\/
FIG. 6.5 Virtual and real instances in Non-Structure Sharing: (a) proc(t(p(A).
p(A). q(p(A).B))) is called with pr0c(t(p(X). Y. q(Z. I'tY))))both terms are virtual be-
fore uni cation. (b) The rst occurrence of p(A) acts as a selectorA is bound to X. (c) The
second occunence of p(A) acts as a constructorY is bound to its copy (a instance).
(d) q(p(A). B) and q(Z, r(Y)) both act as selectors. but p(A) and r(Y) are constn|ctorsboth
tenns are now t(p(X). p(X). q(p(X), r(p(X)))). represented by a mixture of real and vir-
tual instances. (continued)
I74 6 Principles of Prolog Implementation

I I - *- - -

I 5- 1 m I
anrr
/I

I
El I
I

.
@..____.. .P
\ .
@/ @.___.. @\. a:s_/
t I

8 GED

FIG. 6.5 (Continued)

be trees, and more general DAGs are implemented by variable bindings.

In general. the process of term creation can give rise to several copies of a
single term instance (p(A) in the example), but these are indistinguishable.
The process of copying an uninstantiated variable might seem a little
roundabout: why dont all copies simply contain pointers to the variable
frame location? Indeed, why are there any copies at all: is not Structure
Sharing always better?
Recall from chapter I (see Fig. I.6) that a term's lifetime may have to
exceed that of the encompassing clause instance. Yet it is obvious that we
would like to regard a variable framewhich is created when a clause is
activatedas a part of the clauses activation record. If we are careful to
represent variable-to-variable bindings so that younger variables point at
older ones rather than the other way round, and if term copies contain no
pointers into variable frames. then there is no risk of leaving dangling
 6.2. Representation of Tenns I75

td)
A=0 x=0

8... ,._;. '8/3 \I

B
\ I

p II

FIG. 6.5 (Continued)

pointers as activation records are deallocated upon procedure completion

(according to the normal stack regime).
The situation is quite similar to that encountered in Pascal, say: an
object which is to live longer than the procedure which has created it is
allocated in the heap, i.e. a memory area distinct from the activation
stack. The NSS heap is called a copy staclr. It is a true stack, because term
copies can obviously be discarded when the program backtracks past
their point of creation. They can become inaccessible much earlier, too,
and a garbage collector could be very useful, but it is not essential to
Prolog as it is to Lisp. One must remember, however. that without gar-
bage collection the copy stack's size is roughly proportional to the
amount of time spent in forward execution (without backtracking). and

1 But this is not always possible (see the next section).

I76 6 Principles of Prolog Implementation

that one may need to alleviate that by introducing arti cial failures in a
few well-chosen places (see Section 4.3.2).
In the simplest form of Structure Sharing. a variable frame is an
integral part of the representation of a number of term instances. and
cannotin principlebe deallocated so long as any of these terms is
accessible. It must be allocated on a variable stack. which closely resem-
bles the copy stack of NSS (except that a garbage collector. if required. is
harder to implement). The activation stack is smaller, as it only holds
control information.
The most important advantage of NSS is that retention past the mo-
ment of procedure termination concems only those terms which become
variable instantiations. With simple Structure Sharing. on the other hand,
all terms are retained. As it tums out. terms are often used as selectors
rather than constructors. and clauses frequently propel a computation
along without creating many long-lived objects. The copy stack is there-
fore usually smaller than the variable stack. and the effects of memory
requirements being a function of time are much less pronounced with
NSS.
Starting with DECProlog-I0, many Structure Sharing implementa-
tions take advantage of the difference between terms which must live
longer than their clauses and those which need not. As a clause is read in,
it is analysed to detect variables which cannot. under any circumstances.
be used to form instantiations of variables outside the clause. These are
classi ed as local variables, whereas the others are called global. The
variable names are all local to the clause, of coursethe terminology is to
convey that global variables are long-lived, while local variables may be
allocated (and deallocated) with the clause's activation frame. The activa-
tion stack is accordingly referred to as the local staclr, and the global stack
holds global variables.
A simple. though not necessarily the most subtle, classi cation crite-
rion is whether a variable appears inside a term (i.e. is not only a proce-
dures parameter). For example. in
a(X.f(Y)):-b(X.g(Z)).
we nd that X is local. The rule about directing variable-to-variable
references towards the bottom of the stack suf ces to ensure that its de-
allocation will not leave dangling pointers. The variable Y is obviously
global, as the clause can export it after having been activated by
a( Something. Variable ).
The status of Z is uncertain. It can be bound to a variable in b. but we are
really interested only in those outside variables which outlive a. If the
 6.2. Representation of Terms I77

body of a were
b(s(Z))
then Z could conceivably be classi ed as local (according to our experi-
ence, though, allowing such cases could complicate the implementation).
But as the clause stands. we need to analyse b (assuming it will not
subsequently be modi ed!) to check whether g(Z) can be made an instan-
tiation of a variable to which X is bound. For example. with b de ned as
b( V, V )
the call
a(P.Q)
would instantiate P < g(Z) and Q < f(Y)both Y and Z would be ex-
ported. Variables that do not appear in terms can only be used to carry
information around the clause; it is safest to assume that all others will be
used to form structures.
This assumption does not yet allow Structure Sharing to be really
competitive with NSS. To achieve this, we must declare our intentions by
providing so-called mode declarations. In Prolog-I0 one writes
:- mode member(?, +).
to inform Prolog that the second parameter of member will never be a
variable, though its rst parameter might be one. This means that the
procedure
member( E, [
member( E, [ ><l'I'l l"l" member( E, L ).
will not be invoked as a generator of lists, so the compound tenns will only be
used as selectors and all the variableseven those global by the general
criterioncan be classi ed as local.
Providing mode declarations may seem a nuisance, but they are good
documentation (and are not compulsory). The declarations are static and
must necessarily be less informative than the dynamic special-case analy-
sis of NSS. In common cases. however, the difference is not detectable
and this form of Structure Sharing is. in fact, as good as NSS with regard
to memory utilisation. This does not mean that the two behave identi-
cally. Programs can be written which make any one method almost arbi-
trarily worse than the other (how would you go about devising such a
program?).

A compiler can also use this information to generate faster code.

I73 6 Principles of Prolog Implementation

Structure Sharing tends to be faster. but it is more complicated. There

is the problem of analysing clauses. utilising mode declarations and ma-
nipulating term handles instead of single pointers. Moreover. system rou-
tines such as clause are harder to write because a clause can contain
references to local variables and its instance is not therefore a correct
term. If you want to write a simple memory-ef cient interpreter. use
Non-Structure Sharing.

6.3. CONTROL

One of the keys to the success of a Prolog implementation is the

ef ciency of backtracking. Whenever a fail point is established (see Sec-
tion I.3.2), the computations state must be saved. so that it can be
restored upon failure. Both the saving and the restoration of a state are
frequent events, which must take place as rapidly as possible.
The state of a computation can be reduced to the contents of the
control stack and the heap. Obviously, Prolog's special requirements
rule out checkpointing (i.e. dumping memory contents) as a means of
saving the state. Logging (i.e. recording changes made to the state) is a
more hopeful technique, as differences between successive states of inter-
est are usually minute in comparison to the amount of information con-
tained in a state. The technique is particularly suitableand universally
usedfor dealing with the evolution of variable instantiations. Only unin-
stantiated variables can be modi ed, so the old value need not be re-
membered and it is enough to record a modi ed variable's address.
While logging is also a viable method of handling activation record
traf c on the control stack, it would not be able to take advantage of the
disciplined manner in which procedure instances are created and de-
stroyed. A better method. well known since the appearance of (Bobrow
and Wegbreit I973). can roughly be described as using the log itself to
de ne a new state.
As a fail point is established, a fail point record is pushed onto a
special stack. (We are interested in a conceptual description. In practice.
this stack is often implemented by a chain of pointers threaded through
activation records.) The fail point record stores information about current
sizes of memory areas and a pointer to the list of untried clauses likely to
match the current call. In other words, it contains information essential to
Prologs ability to recommence computations from this fail point. To
The generic tenns are meant to emphasize that this discussion is valid for Structure
Sharing and NSS alike.
 6.4. Tail Recursion Optimisation 1'/9

make this information suf cient, stack and heap areas below the levels
indicated by a fail point record are treated as frozen, i.e. under special
protection.
Binding a frozen variable is allowed, but must be logged by pushing its
address onto a fourth stack, called the trail (its size is also remembered in
a fail point record). The control stack, however, is frozen quite literally.
Whenever a terminating procedure would cause control to be retumed to
an activation record (AR) within the frozen area, a copy of the AR is
created just above the protected part of the stack. The copy de nes the
current procedures environment: an ancestor link provides access to the
frozen AR of the procedures caller. To avoid copying that part of an AR
which contains variables, the variables ofa clause are associated with the A R of
its caller rather than with its own AR. An AR's copy will be used to perform a
new call: the original describes the previous call, so its variables are irrelevant.
All this is illustrated in Fig. 6.6.
With these precautions, backtracking consists in undoing bindings
made after creating the most recent fail point record FR (a simple matter
of resetting locations referenced in the top-most fragment of the trail),
popping all stacks to the levels indicated by FR, grabbing the untried
clause list and popping FR itself. This is rapid enough; the unescapable
penalty is that of maintaining (several copies of) frozen substacks which
would normally disappear with the shortening of call chains. One of the
reasons why judicious use of the cut is so important (see Section 4.3.1) is
that it allows Prolog to reclaim stack storage. To invoke the cut is to pop a
number of fail point records, thereby unfreezing areas of memory.

6.4. TAIL RECURSION OPTIMISATION

Many programming tasks are inherently iterative. For example, to

test whether an item is present in a list, we must look at successive
elements until either the list is exhausted or the item is found. But in
Prolog we can only de ne member as a recursive procedure. Recursion is
more expensive than iteration in that it requires not only time but also
stack storage which grows linearly with the number of tums. Storage is
often a scarce resource, and it would be very unsatisfactory if each deci-
sion to traverse a list had to be accompanied by speculations about the
potential length of the list. In Prolog, using recursion instead of iteration is
all the more serious because the stack may subsequently have to be frozen.

Local variables of Structure Sharing, "virtual" variable instances of NSS.

 lo) li) alXlI blXl.
liil blZl: clZ), d(Z).
liiil blll.
liv) ctel.
lvl dlVl.
2- alW), writetwl, fail.

lb
serum 3

.-Q 1

U i
['11
Wx wr|telW) mt
N_I --I,
ll *1
in
\
\-

1 tnlow $5 '
trail level
next clause Ill d-'""""Y
'_ 0
TRAIL FAIL FOTNT STACK comnot STACK COPY STACK

O
O i
7' """ "
Es
' a \\

.' - - - - - - - "1:-~

um 2
\ T

.- - - - - - -- In. -
\

a gw|,wm|'\Ti,Tait
Q

~ - -- Q
\
\ \\_..I H11
\

'-@=< =-- -- \ \
5"-I-
\Q- \\

E a....:..m
trail level
O
TRAIL FAIL POINT STACK CONTROL STACK COPY STAG!
FIG. 6.6 Control stack management: (a) The example program. (b) Clause (iv) is
invoked. (Solid lines in control stack are the ancestor links, dotted lines are variable bind-
ings. Active calls are underscored, remaining calls in each clause are also shown. The model
is NSS.) (c) One step later, d is ready to return. (d) After returning from d and b. (Frame 3 is
a copy of I, executing the next call. The variable Z was destroyed when the stack was
poppedit was just above the freezing level.) (e) After failure, before invocation of clause
(iii). (f) Clause (iii) and (i) terminated. directive in control.
 6.4. Tail Recursion Optimisation l3l

id) v:rite[W|,Iait I
3

X: F1111 \

\
ii W|,vvriteIW) hit \\
- . . . 1 I
I II e

t'reen'I'|5-hdon _ Q; ._____ T.-.--"I

trait level _
next diuserliill O-
0

TRAIL FAIL POINT STEK CONTROL STACK COPY STACI4

(ET blxlf I
_ 2

X; 0- - - - - -- -_
aIW],writeIW),tait ' \
I

. -- I
W: tree
dummy
0

TRAIL FAIL POINT STACK CONTROL STACK COPY STACK

If) v.m|w1,Fa1t |
' I
T
Y
1"
f

W1 0-*1- - - - - - "'

dummy
0

TRAIL FAIL POINT STACK CUITROL STACK COPY STACK

FIG. 6.6 (Continued)

Tail recursion optimisation (TRO) is the technique of replacing some

forms of recursion with iteration. Despite its name, it is also useful in
situations where there is no direct recursion, or even no recursion at all-
just a long chain of procedure calls.
The general idea is illustrated in Fig. 6.7. Assume that q is the last call
in p and that p is deterministic, i.e. there are no fail points between the
I82 6 Principles of Prolog Implementation

:- ..., p...
p:*" allulq.

q :-

FIG. 6.7 Tail recursion optimisation.

invocations of p and q (they were not established or were removed by a

cut). This means that both activation records are not frozen, and are not
separated by locations containing useful information. Now, if the acti-
vated clause of q is known to be the last clause matching its call, then
some of the information in the two activation records clearly becomes
redundant. The younger frame's control information is no longer needed,
because the only thing p can do after qs termination is retum immediately
to its caller: if q retumed to the caller of p, the effect would be the same.
Similarly, the older frames variables (local or virtual") will not be
needed by p. TRO is the technical term for replacing the twoeither
during or after qs invocationby one activation record, with qs varia-
bles and with control information needed to exit from p. If q is the same as
p (or contains a tail recursive call on p, or the like), many calls may be
executed without increasing the size of the control stack. (The heap may
grow, though, if the computation constructs some long-lived objects.)
Several methods of implementing TRO are described in the literature,
and we shall not discuss them here. An important feature of some of them is
that they allow delayed TRO, i.e. merging of our two activation records after q
performs a cut, even though its initial invocation is not deterministic. In
Section 7.3.4 you will nd one such method, which we favour for its
simplicity.
 6.5. Bibliographic Notes I33

6.5. BIBLIOGRAPHIC NOTES

The idea of structure sharing comes from Boyer and Moore (I972). It
was used in the original Marseilles interpreter (Battani and Mloni I973,
Roussel I975), which, actually, was preceded by an earlier, experimental
version (Colmerauer et al. I972). That interpreter did not have anything
like fail point records. Though variables were allocated on a separate
stack, control frames were alsoas a rulepopped only on backtrack-
ing. Classi cation of variables into local and global was introduced with
the DECProlog compiler. Warren (I977a) is the original reference, see
also Warren et al. (I977) and Warren (I980b). A preliminary report on the
rst NSS implementation is Bruynooghe (I976).
The idea of tail recursion optimisation is well known. Bruynooghe
was the rst to use TRO in Prolog, while Warren used a different method
as an afterthought; see Warren (I980a).
A good detailed explanation of the implementation principles is Bruy-
nooghe (l982b). It stresses both the similarity of structure sharing to
conventional handling of procedure instances and the similarity of Pro-
logs control structures to a proof tree. Van Emden (I982) contains a
disciplined derivation of the control algorithm, starting from search-tree
traversal.
Most implementations merge fail point records and control frames
into a single type of record. To our knowledge, they were rst separated
in Donz (I979), an early approach to global optimisation, where they were
talked of as the and-nodes and or-nodes of a search tree. We like the
separation because it brings to light the fact that backtracking is imple-
mented almost exactly as proposedin a more general settingin Bo-
brow and Wegbreit (I973), the classic paper on implementation of uncon-
ventional control structures.
A comparison of NSS and structure sharing can be found in Mellish
(I982), with some comments in Bruynooghe (l982b).
Mellish (I981) is an early approach to automatic production of mode
declarations by means of global ow analysis. Other papers concemed
with global analysis, though not for the sake of ef ciency, are Bruy-
nooghe (I982a) and Mycroft and O'Keefe (I983).
At the time of this writing we know of two new compilers being
developed. The references are Bowen et al. (I983) and Ballieu (I983).
See also Section 2.5 for references on Prolog implementations with
coroutining and parallelism.
As a point of interest, we shall mention two papers describing imple-
mentations of Prolog done by embedding it in another programming lan-
guage: Lisp (Komorowski I982) or POP-ll (Mellish and Hardy I983).
7 TOY: AN EXERCISE IN
IMPLEMENTATION

7.1. INTRODUCTION

This chapter is a case study of Toya simple but fairly complete

implementation of Prolog. Only the most important (or least obvious)
information is presented here, and it should be read together with the
source texts available on the diskette enclosed with this book (some of these
are listed in the appendices).
While designing Toy, we attempted to strike a compromise between
several con icting goals. We wanted to write:
A clean, readable interpreter which you could nd useful for getting a
feel" of what is involved in implementing a life-size Prolog system;
A usable interpreter, which we could use to test all the programming
examples in this book (our extant implementations were quite incom-
patible with Prolog-I0) and which you might use to experiment with
Prolog if you have a lot of time but no access to a machine running one
of the commercially available Prolog systems;
A large fragment of the implementation in Prolog itself, to provide a
sizable example of using the language for solving well-known but not
completely trivial programming tasks at a relatively low level;
An interpreter which, though useful, would have little commercial
value.
We decided to use Pascal, because it is easy to read, well known
and generally available. The program is not written to be very ef cient:
concem for readability and conciseness almost always prevailed. It is
not particularly short and elegant either, as we wanted it to support a
l36 7 Toy: An Exercise in Implementation

fairly complete version of Prolog modelled after the Prolog-I0 dialect.

There are two principal reasons why we call it Toy:
The user interface is written in Prolog, and this makes it rather slow;
There is no garbage collector, and moreover, partitioning storage
into several disjoint xed-length areas makes it easier to encounter a
memory over ow condition.
If you decide to use Toy, you will quickly nd that the time taken to
read and write terms requires some patience. We had to rewrite read,
write and op in Pascal for our purposes, and it is but a moderately dif cult
task. A rather straightforward implementation resulted in another I000
lines of code, but a lot of it is dedicated to handling mixed functors (see
Section 7.4.3).
We used Toy on two minicomputers: a PDP II/40 look-alike running
RSXI IM, and a Polish computer called Mera 400. The PDP has an ad-
dress space of 64KB; we used it to bring the system up, but it was a tight
squeeze. You might do better with a P-code system rather than with a
native-code compiler of Pascal, such as the one we had to use. The Mera
had a I28KB address space and a fairly good native-code compiler (but
with no attempt at global optimisation): we could easily load and execute
both the whole Prolog interface and programs such as WARPLAN or
Toy-Sequel (see Chapter 8). We tested all our programs and had quite a
bit of memory to spare, running in a 104KB space.
The original implementation was subsequently ported (almost painlessly!)
into Berkeley Pascal (on the VAX/780 running 4.2 BSD UNIX) and into
TURBO Pascal (on the IBM PC running MS-DOS 2.l0). You can nd the
TURBO version on the diskette enclosed with this book. Files READ.ME,
CONTENTS, INSTALL and TURBO.PAT contain general information
about the diskette and the implementation. The latter le summarizes changes
introduced into the original Mera Pascal which was listed in the hardcover
edition of this book.
Feel free to run Toy and play with it, but remember it is copyrighted. No
version of this implementation may be used or distributed for gain, all listings
must contain our copyright notice, and the heading produced by status (see
Section 5.7.5) must contain the texts Toy-Prolog and IIUW Warszawa".
Other than that, you are welcome to modify it, give it to friends, etc. If you
have any comment to make, we shall be happy to hear from you.

7.2. GENERAL INFORMATION

Toy is a Non-Structure Sharing interpreter (see Chapter 6). The pro-

gram written in Pascal supports a limited syntax, which we shall call Toy-
 7.3. The Toy-Prolog Interpreter I87

Prolog, and only a subset of the usual system (built-in) procedures. The
full user interface and library is implemented in Prolog (see Section 7.4)-
this approach was taken in the original Marseilles implementation, and in
a number of implementations since. A short program called the boot-
strapper (see Section 7.4.l), written in Toy-Prolog, is used to translate
into Toy-Prolog other parts of the user interface, which are written in a
slightly restricted fonn of the usual syntax. Next, various interface pro-
grams can be loaded during initialization (see Section 7.3.6).
A Prolog program called the monitor supports an interactive pro-
gramming regime (see Section 7.4.2). Full Prolog-I0 syntax can be used
(see sections 7.4.3-7.4.5). A program called the translator can be used
to convert Prolog-I0 programs into Toy-Prolog (see Section 7.4.6). The
translator shares most of the monitors routines. lt can be used for large
(interactively debugged) programs which are to be loaded quickly, with-
out repeated syntactic analysis by the rather slow parser in the monitor.
See Appendix A.4 for a few examples of such programs.
We shall nish this section with an example of Toy-Prolog syntax.
There is no point in providing a precise description of this language, as it
is very simple and the recursive-descent parser (a fragment called the
READER, see le READER.PAS on the diskette) is so straightforward that
it can easily be used to resolve all doubts. Our example is

P(la.lb.Ld|X1.Y)=-q(Y.X).r(s(Y).-).
:-p(Z,(t:-u,v)).

To make it directly acceptable to the READER, we write

p( a.( b.c.[] ).d.:0, :l ) : q( :l, :0) . r( s( :l ),_ ) . []
:p( :0. =-( t. '.( u. v ) ) ) - ll#
See Appendix A.2 for further examples. The syntax is not nice, but is
very close to the intemal representation of clauses.

7.3. THE TOY-PROLOG INTERPRETER

7.3.1. The Principal Storage Areas

Toy uses several disjoint areas of memory for its data structures (see
Fig. 7.I). They are listed below.
CT (character table), used to store strings: print names of Prolog func-
tors and predicate symbols;
AT (atom table), used to store atoms. In this chapter atom does not
denote a functor with no arguments. It is the generic name of a record
 I88 7 Toy: An Exercise in Implementation

HTPBH

CTHIGH ATl' H "''

f li
| 7'" op prototypes
--

=1""9I- .-_"9bt
prototypes
cttow mow "N
c1 AT
Ii F - t-| BTHIGH I'll-IIGH ground
pt0IOI)1>IS
| FROTLOW
tree I ree cop STACK HG-I
I
.____.__._...l 4- PT
hadnraclt
l.2 %- stildt

M,
__csbot
stacks
"an reep='ds I Two
|__i. FTLOW BTIDN TTLOW variable
ack
Ft B1 It st MTLDW
MT
FIG. 7.l The main data areas.

containing useful information about a symbol (a functor or predicate

symbol in our case);
MT (main table), used to store term instances and prototypes. There
are two subareas here:
Prototype storage, which is further divided into disjoint storage
areas for ground (variable-free) prototypes and for those that con-
tain variables. The classi cation is important because a ground
prototype can be used to represent all its instances, and need not
be copied onto the copy stack;
Stack storage, which is further divided into disjoint areas for the
copy stack and the variable stack (the variable stack holds varia-
bles from activation records: Pascal's type mechanism made it
more convenient to keep control information from activation re-
cords in a separate table FT);
FT (frame table), used as the activation record stack (but variables are
stacked in MT);
BT (backtrack table), used as the fail-point stack (here called back-
track-point stackwe just needed a different letter to label the table);
'l'I (trail table), used as the trail stack;
Pascal's heap, used to store procedure descriptors;
Pascal's stack, used for recursion in uni cation and tenn-copying oper-
ations.
In what follows, we shall use the word pointer, or address, to denote
both Pascal pointers and indices into the tables.
 7.3. The Toy-Prolog Interpreter I39

7.3.2. The Dictionary: Atoms and Procedure Descriptions

The character and atom tables fonn the dictionary: a data structure
used primarily as an aid in translating between the extemal and intemal
forms of Prolog terms and clauses. lt also supports access to procedures,
making it easier to implement variable calls and clause manipulation.
An atom is a record containing infonnation about a functor and/or a
predicate symbol. The difference between a tenn and a procedure de-
pends only on context and is not always recognized. Predicate symbols
are denoted by functors when clauses are treated as terms (e.g. in assert);
conversely, a functor may be used to invoke a procedure (as in call).
The attributes of an atom are

Its print name (a pointer to a string in CT);

Its arity;
The procedure of this name and arity (a pointer to a procedure descrip-
tor, or nil).

Atoms are accessed through direct pointers or through a hashing proce-

dure. Direct pointers are present in the representation of terms (including
clauses; see the next section). The pointers are used for

Printing a functor,
Determining arity,
Finding a procedure.

In particular, the representation of a call contains a pointer to an atom as

the only handle on its procedure. Addition and deletion of clauses in the
procedure does not therefore require modi cation of its calls.
Hashing is used to locate appropriate atoms during conversion from
extemal representation. Such conversion takes place when tenns are read
in or when they are created by functor and pname. For simplicity, linear
rehash is used in the current version: you might wish to improve it.
Print names are represented in CT by contiguous sequences of char-
acters terminated with EOS characters (zero bytes). As a name is created
by pname or the READER, its characters are pushed on top of the string
area in CT (procedure buildname). On termination of the string, wrap-
name is invoked to locate an atom with the same printname. If such an
atom is found, the string is obliterated; otherwise a new atom is created
and retumed. Since this atoms arity is unknown, the arity eld is set to
the special value of noarity (procedure ndname).
I90 7 Toy: An Exercise in Implementation

Atoms are located by the READER in a two-phase process. First,

buildname and wrapname are used to nd the rst atom with this name;
then a single scan through the (virtual) hash chain nds an atom with the
correct arity, or detects its absence and creates it (procedure ndatom).
Conversion between atoms of different priorities, needed to implement
functor, requires invocation of the hash algorithm to locate the beginning
of the appropriate hash chain (procedure samename).
Procedure descriptors are allocated in the Pascal heap. Descriptors
are formed of lists of records, each of them with:
a pointer to the next element in the list;
the number of variables in an activation record;
either the number of a system procedure, or pointers to the prototypes
of a clause's head and body.
A system procedure descriptor is formed of a single such record. The
descriptor of a Prolog procedure is a list of records, one for each clause.
The head predicate's atom always points at the rst element of this list.
A clause body is represented by the prototype of a Prolog list contain-
ing its calls. Figure 7.2a illustrates the layout (recall that the binary dot is
the Prolog list constructor).

7.3.3. Prototypes and Term Instances

The main table, MT, holds a variety of objects which are distin-
guished partly by their addresses and partly by their contents. Addresses
are used to distinguish between prototypes and term instances ( elds
denoting variables contain variable offsets in prototypes, and variable
bindings in term instances). Prototypes of ground terms, which contain no
variables, are also used as instances: this helps keep down the size of the
copy stack.
Instances of non-ground terms are kept in the stack area. It is divided
into the copy stack and the variable stack. The variable stack holds acti-
vation-record variables and is separated from the copy stack because it
can shrink on procedure return and not only upon backtracking (see
Chapter 6).
Object contents are used to distinguish between integers, variables,
and normal terms with functors.
Integers are two-word objects. The second word holds the integer and
the rsta special marker INT, which prevents the interpreter from
treating integers as pointers.
 7.3. The Toy-Prolog Interpreter I9]

Variables hold values less or equal to VARLIM (both INT and pointers
to MT or AT objects have values greater than VARLIM). VARLIM is
kept only inside the dummy variable (-) Prototype, whose address is
DUMVAR)(this prototype is treated as ground. The value
FREEVAR (equal to VARLIM I) lls free variable instances. Values
below FREEVAR are negative: in prototypes their absolute values
denote offsets in variable frames, and in instances their absolute values

Io)

member\

[1 '-IIEID

GIEID _

M}
.../\.. 6
NW
FIG. 7.2 The intemal representation of member:
member( :0, :0._ ) : I]
member( :0, _.:l ):member( :0, :l ). []
(a) The abstract form. (b) The data structures (variable o sets adjusted by offoff; []/0,
J2, J0, memberl2 denote addresses). (continued)
I92 7 Toy: An Exercise in Implementation

'

(bl I

' e

merrtler/2
N
I punot -uou

I
E
: 5 ./0 I

unuij nu:anru
1
IIEi il l
aan heap
I
E

I
Iii!
CT
HI
Hill
v_i._..|

AT
5
H

PUITO
NILPROTX
w DUMVARX
MT
I prototypes)
FIG. 7.2 (Continued)

are pointers to variable bindings. (Actually, the situation is slightly

different: INT = I, VARLIM = 0 and FREEVAR = I. All negative
entries denote non-dummy variables. MT's lower index is 2, but van
able frame offsets start from 0 and are therefore adjusted by the con
stant OFFOFF = 2; -2 stands for offset 0, -3 for offset I, etc.)
 7.3. The Toy-Prolog Interpreter I93

IIIIII
an mtegr; -5 a variable instance: a vanatle |:lototy|:e;
hotnd totheobject ottset3 Iadjtstedbydlo l
at address 5

qLLr..1..1.'
atom ot pl3 an instance ot atom ot qlt
|=lqtX),_.'H.
X is tree
FIG. 7.3 intemal representation of terms.

NormaI tenns are contiguous sequences of words. The rst word

holds a pointer to the main functors atom (its arity eld de nes the
length of the sequence). Other words represent arguments. For variable
arguments see above; other arguments are represented by pointers to
appropriate objects (see procedures getarity and getarg in the listing).
Figure 7.3 illustrates these conventions. Figure 7.2b shows the com-
plete intemal representation of a procedure.
As explained in Chapter 6, tenn instances are pushed onto the copy
stack only when absolutely necessary (when they become variable bind-
ings) and are otherwise represented as in Structure Sharing. It is therefore
convenient to represent all instances by a pair of pointers. If the rst
pointer addresses a prototype, the second (which we shall call the proto-
types environment) is a pointer to an area in the variable stack. If the rst
pointer addresses a term instance, the second is disregarded. Note that-
unlike in Structure Sharing implementationsthe environment need
never change as term arguments are accessed: variable bindings are never
nonground prototypes and require no environment.
The normal mechanisms of object recognition and creation are cir-
cumvented in two major cases (see procedure loadsyslcernel).
To avoid creation in the copy stack of too many integer objects repre-
senting intermediate results, a range of the most frequently used inte-
gers ( I.. I0 in this version) is maintained in the form of unique ground
prototypes.
To avoid the overhead of locating character atoms, checking whether
functors represent characters, and duplicating character prototypes or
instances, ground prototypes of ASCII characters are kept in a contigu-
I94 7 Toy: An Exercise in Implementation

ous area of MT. Accessing a character prototype requires only the

addition of its ordinal number to this area's address.
Certain other objects also have representations at addresses known to
the interpreter. Apart from "popular" integers, characters and the
dummy variable, there is also a prototype of the atom I] (see the begin-
ning of the global variable declarations for a listing of all these addresses).
Addresses of atoms requiring special treatment are kept in the table STD
(see the de nition of type stdatomld). There is also the prototype of a
dummy clause, whose body consists of a single call to errorl I , located at
address errcallseq.
Procedures for handling term representations are quite straightfor-
ward. Only prototype creation might not be immediately obvious. The
method is quite similar to that used for creating entries in the dictionary.
A prototype is allocated by invoking initprot with information about the
main functor. Arguments are then lled in by newparg and newpvararg,
and the process is tenninated by wrapprot. This procedure checks if all
the arguments are groundwhen this is the case, the prototype is moved
to the ground prototype area. Note that the process is inherently recur-
sive, as argument prototypes may be created before their parent terms
prototype is wrapped up. This is why initprot must be used: piecemeal
allocation would not preserve contiguity.
A short comment about terminst, the procedure usually used to create
terms on the copy stack. Non-variable arguments are represented not by
direct pointers, but by negative values, as if they were all fonned by
instantiating pre-existing variables. This is necessary because the proce-
dure argument (which follows chains of variable bindings to locate the
nal instantiation) expects variable arguments directly inside the repre-
sentation of their parent tenns. A recursive call on terminst can retum a
variable and treating the variable as a nonnal argumentby inserting a
positive pointer to itwould break the chain of references. (Such things
are not easily seen, and the erroneous situations are rather infrequent:
this bug was the hardest to locate!)

7.3.4. Control

In Toy, clause bodies are represented as prototypes of lists. The list

elements are prototypes of calls, and none of them is an integer or a
variable. While not directly related to the extemal form of clauses in
Prolog-I0, this representation is very regular and easy to handle.
The method of representing control state is almost exactly like that
described in Chapter 6. The principal difference is that the variable part of
 7.3. The Toy-Prolog Interpreter I95

?- . [1 ?_..._._._[]
1 ,,,l,_ til |
c oI ....
N
M--==m
utm_g
-

@ ,,,,.,

= |
8-43..
8""? 2

an
III! =-='='
H
__..-.,
=. %sonu ifIII
'95: 0 IE!
III l"I= _ 0
main
E 2
T$'IIr=
imI|va=t.

I'r:env=2
IT BI Ft m
Ieteetul
FIG. 7.4 A more detailed form of Fig. 6.6d.

an activation record is kept on a separate stack. Figure 7.4 is a detailed

version of Fig. 6.6d. We shall comment only on the variables used as
control registers".
The crucial variables are:
topf, a pointer to the current control frame (i.e. activation record),
which is always on top of the stack in FT;
topb, a pointer to the current backtrack point (i.e. fail point) record,
which is always on top of the stack in BT;
-csbot, a pointer to the rst free location below the copy stack in MT
(this stack grows downwards);
vtop, a pointer to the rst free location above the variable stack in MT;
ttop, a pointer to the rst free location above the trail stack in TI.
Five auxiliary variables contain copies of infonnation available else-
where. They are used for ef ciency:
ancf is a pointer to the current control frames parent frame;
ropenv is a pointer to the current variable frame (associated with topf);
ancenv is a pointer to the variable frame associated with ancf;
[96 7 Toy: An Exercise in Implementation

frozenheap is a pointer to the rst free location below the frozen part of
the copy stack;
frozenvars is a pointer to the rst free location above the frozen part of
the variable stack.
Execution of a Prolog program is driven by the procedure resolve.
Each tum of its loop is an attempt to match a call against a clause head, or
to execute a system procedure. At the beginning of this step the situation
is as shown in Fig. 7.4: a control frame for the current call is on top of the
stack, but the clause is not yet invoked and the associated variable frame
is empty. If the call was an erroneous system procedure call, the error
handler is activated (see below).
If the step is unsuccessful (the head did not match the call, or the
system procedure failed), the interpreter backtracks. Otherwise it enters
the procedure orif it was a system procedure or a unit clauseexits it.
Entering a procedure consists in setting up the control state so that the
next call to be executed will be the rst call in the freshly activated clause.
Exiting is the process of nding the next pending call: either the one
immediately following the successful current call, or (if this was the last in
its clause) a call following the nearest ancestor which is not the last call in
its clause.
To stop the execution, the ag stop must be set. This is done either by
the system procedure halt, or by backtrack when there are no fail points
left (i.e. when the directive failed) or by exitt when it cannot nd a pend-
ing call (i.e. when the directive succeeded).
Two auxiliary variables play the role of a program counter:
ccall contains a pointer to the prototype of the current call (it is the
prototype of the rst element in the list indicated by the current control
frames calls eld, unless that element is an invocation of call or tag:
ccall is then the outermost argument which is neither of these);
cproc contains a pointer to the descriptor of the procedure invoked by
ccall (for Prolog procedures, this is the rst clauses descriptor when in
forward execution, and a pointer recovered from a backtrack point
record's resume eld when immediately after a failure).
Notice that a fail points resume eld points at the predecessor of the
clause which is to be retried. This is so to make retract correct.
The algorithm used for tail recursion optimisation (procedure
trooverlay) merits some explanation. We employ the naive method sug-
gested by Fig. 6.7. After uni cation is over, procedure candotro checks
whether the current call is an untagged tail call and whether the ancestor
frame is not frozen. If so, neither the call nor the variable frame associ-
ated with the ancestor frame will ever be needed again. The current
 7.3. The Toy-Prolog Interpreter I97

variable frame is shifted to replace the ancestor variable frame, and the
control stack is popped so that the ancestor control frame becomes top-
most (the most recently activated clause is still accessible through cproc).
The algorithm is made a little complicated by the fact that the shifted
variables may be instantiated to one another or to the destroyed (overlaid)
variables. Both cases are illustrated in Fig. 7.5.
The cut procedure simply removes as many backtrack point records
as necessary (possibly none) to ensure that the call invoking the procedure
containing the cutand all subsequent callswill not be retried. (There
are exceptions to this rule: notice that ,/2, ;/2 and call/I are transparent to
the cut.) After popping off backtrack points, the interpreter must purge
the topmost section of the trail to remove references to variables which
are no longer frozen. This is necessary, because such variables can be
popped off, or shifted, during TRO. Notice that the method of TRO applied

Io)
- vtcp=80

I
8 IIIII
3$!-'5E

Q
---~ F -1

HT
(variable stacltl
FIG. 7.5 Tail recursion optimisation: merging two frames. (a) The initial situation.
Both frames are not frozen, the call is tail recursive. The variables at 57 and $8 are instan-
tiated to the same free variable, the variable at 59 is instantiated to the variable at 44. (b)
Adjustment pass. (i) The rst variable (at 57) points at an overlaid free variable (at 56). The
direction of the pointer is reversed. (ii) The second variable (at 58) is dereferenced to that at
$7 through that at $6. The reference is remapp-ed: the second variable points at 5$the
future location of the variable now at S7. (iii) The third variable (at $9) is dereferenced to that
at 44 through that at $5. (c) Shifting pass overlays the parent's variable frame with the
current variable frame; the parent's control frame becomes current. (continued)
 I93 7 Toy: An Exercise in Implementation

I bl

\'TP=5O vtop= 50 vtQ=G0

-~ 5 1 -
5 -
51 iB 51 I 51 msavaa
56
ss - -
ss ss

iE 1
3
ti
8 a 3
S2
it

Ii) Iiil Iiii)

Icl
vtop:5l

ill I

- =41
1|
I-I he P

HT
(variable stack)
FIG. 7.5 (Continued)

here makes it fairly easy to perform delayed frame merging after things
are made detenninistic by the cut. We shall not enter into the details of
this and of tagcut: this is a simple exercise.
The last thing worth mentioning is the handling of erroneous calls to
 7.3. The Toy-Prolog Interpreter I99

system procedures. This involves pushing a dummy variable frame, with

a single variable instantiated to the erroneous call. The current control
frame (in which the call was invoked) is associated with this variable
frame and becomes the ancestor of errorl I. As a result, the parameter of
error/I is the rigl1t instance of the erroneous call. The process is illus-
trated in Fig. 7.6.
The program maintains several important invariants, such as there
are no outside references to non-frozen variable frames except from vari-
ables higher in the variable stack". We decided to let you have the fun of
discovering them for yourself (after all, these are the real trade secrets).

7.3.5. System Procedures

We shall not give a detailed description of the system routines. There

are too many of them, and the listing is more or less self-explanatory. The
general principles are as follows:
All system procedures are invoked through procedure sysroutcall;
sysroutcall sets up pointers to their parameters in table SPAR (the
values of integer parameters are also passed through table SPARV);
System procedures that can fail or succeed indicate the result by setting
a Boolean parameter (success) passed by sysroutcall;
Whenever a system procedure detects an error, it sets the global ag
syserror, which forces the interpreter to invoke errorl I (see the end of
the previous section).
There are no tricks, except in the procedure concemed with creating
new clauses. It is important that several occurrences of a variable be
represented by occurrences of the same offset when a tenn is translated
into a prototype. To achieve this, addresses of variables appearing in an
asserted clause are stacked in the free area above the topmost variable
frame. With each variable occurrence, this temporary variable dictionary
is searched linearly and, possibly, augmented. The position of a variable
in this dictionary is treated as its offset.
To add a new system procedure, one must:
Write its code;
Insert its identi er in type sysroutid (its place there de nes its position);
Insert its call in procedure sysroutcall (in the same position);
Insert its name and arity in the kemel le (in the same position)see
the next section.
 i
'

| I
I I

to)

Gill

st- Irr
8
*.
_.. :._._-______.. _.l- _. -. 3.
.-

MT
Ietacke)

I'_"-_'_":
I

(bl

e-II] '1

topt
ElHi) csho

lifl til
MT
htldol
FIG. 7.6 Handling erroneous calls to system routines. (a) wch detects an incorrect
argument: a(V). (b) a call to error/I is set up.
 7.4. Interpretation of Prolog-I0 in Toy-Prolog ZUI

7.3.6. Inltlalisation

initialisation is done in three phases. First, most of the variables are

set by procedure lrritvars. Then two portions of data are read from the so-
called kemel le. One portion de nes the names and arities of standard
atoms whose addresses must be known to the interpreter. They are cre-
ated and their addresses are stored in table STD. The other portion de-
nes the names and arities of system procedures: as the atoms are cre-
ated, they are associated with system procedure descriptors. The number
and order of all these atoms is known to the interpreter. Arities are impor-
tant, but printnames are arbitrary and can be changed at will.
The last phase of initialisation consists in creating a number of stand-
ard objects. Their addresses are known to the interpreter but they cannot
be created before the addresses of standard atoms are xed. The objects
are:
The prototype of [ ];
The prototypes of characters;
The prototypes of the integers I, 0, I0;
The dummy clause body used to invoke error (it is the prototype of
[err0r(X)]);
The prototype of user, needed by the stream switching procedures (see
Section 5.7.I).
After initialisation, the interpreter begins normal execution, reading
the current le. This is normally the kemel le, containing some useful
library procedures. One can also append the bootstrapper or the trans-
lated monitor (see below).

7.4. INTERPRETATION OF PROLOG-10

IN TOY-PROLOG

7.4.1. Intennediate Language

Even a modest program in Toy-Prolog can be unmanageable. To write

the monitor, we use a subset of full Prolog, without operators and gram-
mar rule notation. Commas and the symbol :- are treated as separators.
List notation is allowed, with one restriction: an X in I .... .. | X] must be a
variable. This subset is translated into Toy-Prolog by a "bootstrapper"
written in Toy-Prolog. Debugging and testing the monitor required fre-
quent retranslations of its small pieces, but the gain in readability was worth
ZO2 7 Toy: An Exercise in Implementation

this extra effort. Of course, once the monitor works, the bootstrapper is
no longer needed.
The bootstrapper is listed in Appendix A.2. Comments starting with
%% associate mnemonics with variable numbers. The main procedure is
translate (lines 2-I3), with two parametersthe names of the source and
output les. The unit processed with each tum of the failure-driven loop is
a single clause or a comment. The loop stops upon encountering a @ in
place of the rst non-blank character of a unit. The translation of a clause
is a string which is built on the y on a difference list of characters; the
list is represented by the two parameters christened termrepr and rest.
of_termrepr. Here is how the clause in lines 54-55 would look after
rewriting it into full Prolog and combining those parameters:

ctailaux( Fterm_ rstch, Termrepr -- Rest_of_termrepr,

Sym_tab ) :-
fterm( Fterm_ rstch, Fterms_ rstch,
Termrepr -- [' ", | Middletermrepr ], Sym_tab ),
fterms( Fterms_ rstch, Middletermrepr -- ResLof_termrepr,
Sym_tab ).

Fterms_ rstch is the rst non-blank following a functor-term; in a correct

clause, it can only be a dot, or a comma (see lines 58-66).
Comments embedded in a clause are copied at once (lines 50-53).
Moreover, the string contains end-of-line and blank characters which
improve the appearance of the translation.
Error in a clause causes a message to be printed and the input up to
the nearest dot to be reprinted and skipped (see lines I5-21). The program
assumes the data are correct, and protests upon encountering the rst
unexpected character.
Output for each clause with variables is followed by a comment that
associates variable numbers with source names taken from a symbol table
for this clause (lines 2I9-226). The table is an open list of names. Their
positions are used as variable numbers in the translation. Up to 99 varia-
bles can occur in a clause. The number-name pairs are written six in a
line (line 224).
There are some other minor points worth noticing. For example, the
output string gets closed eventually due to the [] in the initial call on
clause (line II); translations of lists within lists are parethesized, see the
fth parameter of term (lines I31, I36, I37); identi ers are enclosed in
quotes by fterm (lines 69-70); etc. etc. However, the rest of the program
should be self-explanatory. A hint: it can be viewed as a metamorphosis
grammar used for synthesis, driven by input data, with the two compo-
 7.4. Interpretation of Prolog-I0 in Toy-Prolog 203

nents of a difference list serving as an input and output parameter (see

Section 3.I).

7.4.2. Overview of the Monitor

The core of the monitor is an implementation of the built-in procedure

read that is used in user programs (see Section 7.4.3). The user
communicaes with Prolog via an interactive "driver" which operates in a
loop terminated by executing the procedure stop. In each cycle the driver
prompts the user with
?-

and then reads and executes a directive. The symbol table (retumed by
the two-parameter read; see the end of the next section) pairs source
names of variables with variables proper. After successful execution,
the symbol table is used to display nal instances of these variables, and
the driver awaits a pI'Il'lIlbIC character. If it is a semicolon, execution
resumes with forced failure, else processing of this directive terminates.
A directive can be pre xed with :- (we call such a directive a com-
mand, and that without the pre x a query). It will then be executed deter-
ministically, and variable instances will not be printed. However, neither
a non-unit clause nor a grammar rule make sense when read directly by
the driver: a two-parameter procedure :- or --> (presumably unde ned)
would be called. User procedures can be de ned by calling the built-in
procedure consult or reconsult; both are implemented in the driver. In
consult mode", term L --> R is treated as a grammar rule and translated
by the procedure transl_rule (see Section 7.4.4). A one-parameter term
:-C is treated as a command and executed. Other terms are treated as
program clauses.
The monitor is listed in Appendix A.3.

7.4.3. Reader

The syntax of Prolog-I0 is only deceptively simple, so the reader is

rather involved. One wonders whether a simpler syntax would necessar-
ily be less user-friendly.
The main component of the reader is a parser which produces intemal
representations of tenns on input (Appendix A.3, lines 90-332). Transla-
tion of an internal representation into a term proper is quite straightfor-
ward (look at the listing of the procedure malceterm, lines 334-357, after
reaching the end of this section).
204 7 Toy: An Exercise in Implementation

The parser is a classic operator precedence parser; those parsers

belong to the shift-reduce" classthey are bottom-up and determinis-
tic (Gries I97I, Aho and Ullman I977).
Recall that, roughly, an operator precedence grammar has no produc-
tion with two consecutive non-terminals, and all its productions are such
that a shift-reduce parser can determine the handle by comparing
neighbouring terminals in a sentential form. This is possible when each
pair of terminals is in at most one of the three relations denoted by <, =,
>. The relations are de ned as follows (p, q are terminals, U, V, W non-
terminals):
p = q if there exists a production of the fonn
[H eee ooo

or Uaoelpvqool

-p < q if there exists a production of the form

U _,, p V
where q or W q--- can be derived from V
_p > q if there exists a production of the form
U ,__) ...... v q ---

where p or p W can be derived from V

A parser shifts (i.e. scans a sentential form from left to right) until it
detects a pair of terminals related by >. It then scans backwards until the
nearest pair of terminals related by <. The < and > are assumed to be
brackets delimiting the handle in a canonic parse: the handle is reduced
and the process continues.
Note that <, = and > have nothing to do with the common number-
ordering relations. However, if terminals are operators as in arithmetic
expressions, these relations re ect operator priority: the grammar is
structured so that higher priority operators (with operands) are reduced
rst. The situation is similar in the case of Prolog "operators" (even
though in Prolog-I0 weaker operators are given the higher priority). We
shall sayvery informallythat f is weaker than g if f < g or g > f. But
note that, for example, + < (, ( = ), and + > ).
We shall now return to our program. We assume that the input is
delimited by two additional operators. The rightmost delimiter is weaker
than any operator to its left; the leftmost is weaker than any operator to its
right (except the other delimiter). Notice that an empty input is erro-
neous.
The parser maintains a stack of symbols. Initially the stack contains
 7.4. Interpretation of Prolog-I0 in Toy-Prolog 20$

only the leftmost delimiter. The rst true terminal becomes the current
input tenninal. In each step, the current input tenninal is compared to the
topmost terminal on the stack. Three situations are possible:

I. The input is erroneousthe parser stops with error";

2. The topmost terminal is strongerthere must be a production with the
righthand side consisting of a number of topmost symbols on the stack;
we reduce the stack by replacing all these symbols with a correspond-
ing lefthand side;
3. The topmost terminal is not stronger, i.e. no righthand side has been
completedwe shift the current input terminal onto the stack and
make the next tenninal current.

Our operator grammar of Prolog-I0 terms assumes seven classes of

terminals and one class of nonterrninals, t (for terms). Parameters of
symbols are used to build the intemal representation of a given term.
Terminal symbols are read by a scanner (see Appendix A.3, lines
36l-480). The procedure absorbtolren (lines 379-409) reads and con-
structs a raw token:
id(NameString) from words, symbols, and solo-characters;
qid(NameString) from quoted names;
-var(NameString) from variables;
num(NumberString) from integers;
str(String) from strings;
br(LeftRight, Type) from brackets (LeftRight is I or r,
Type is '0. ll. or {}"):
-bar from |;
dot from a full stop.
Next, the procedure maketolren (lines 457-480) constructs a terminal
symbol:
-vns(Variable) from var(NameString);
vns(Number) from num(NumberString);
vns(String) from str(String);
ff(Name, Types, Priority) from id(NarneString) (when this functor
is an operator);
id(Name) from id(NameString) (when this functor
is not an operator) and from qid(Name-
String) (i.e. a quoted name never de-
notes an operator);
br(LR, T) from br(LR, T);
Z06 7 Toy: An Exercise in Implementation

bar from bar;

dot from dot.
The terminal symbol dot is used as the rightmost delimiter of the input.
The leftmost delimiter (and the seventh terminal) is bottom. It is never
retumed by the scanner: the parser's main procedure, gettr, pushes it
onto the initially empty stack. Both delimiters never appear in produc-
tions.
The Types argument of ff is a list of functor types: [Binary], or
[Unary], or [Binary, Unary] (see the de nition of the built-in procedure
op lines 656-718)).
A symbol table in an open list is used to relate a variables name to a
Prolog variable.
The grammar underlying the parser is given in the listing (lines 99-
I07). The de nition of intemal representation can be read off the reduce
procedure (lines I58-I79). Incidentally, the procedure can be para-
phrased as a metamorphosis grammar. For example, the fth and sixth clause
would be rewritten as
t(tr(TyP.X))->lbr(l.Tyr>)l.t(X).
lbr(r.TyP<=)1-
t(bar(X.Y))>lbr(l.ll)l.t(X).lbarl.
t(Y).Ibr(r.ll)1.
Notice, however, that top-down analysis based on such a grammar would
not be deterministic.
There are ve types of intemal representations:
arg0(X) for X a variable, name, string, or nullary functor;
trI(Narne, X) for a pre x or post x term (X is the representa-
tion of the argument);
tr2(Name, X, Y) for an in x term-(X, Y are the representations of
the arguments; in particular, the comma is an in-
x functor, so comma-lists" of terms are repre-
sented with tr2for example, the representation
of
a, b, c
is
1r2( ars0( a ).
tr2( ars0( b ). ars0( c ) ) );
bar(X, Y) for a list with front X and tail Y; X is often the
representation of a comma-list;
 7.4. Interpretation of Prolog-I0 in Toy-Prolog 207

tr(Name, X) for all other valid situations:

tr(()', X) is equivalent to X;
tr([], X) represents a list (of de nite length), X
usually represents a comma-term
tr('{}, X) represents the term {(Cond)} where
Cond is the term represented by X (this is
used in grammar rules);
tr(Narne, X) with Name other than a bracket
type (and Xusually the representation of a
comma-term) represents a nonnal term; for
example, the term
foo( rapes,

is represented by
tr(f00.tr2(.. rs0(lPl).ars0(5)))
The parser's entry point is the procedure gettr (lines I25-I27), and
the main loop is implemented as the procedure parse (lines I29-I38). The
loop tenninates successfully when the original input (bottom and dot
included) reduces to the sequence
bottom t( IntemaIRepresentation ) dot
The parser fails in two situations:
when the procedure establish_precedence fails, i.e. when the topmost
tenninal on the stack and the current input terminal do not compare;
-when the procedure reduce fails, i.e. the top segment of the stack does
not match any production.
The procedure topterminal (lines I40-I43) retums Top, the topmost
stack terminal, and its position: I means Top is the top item, 2 means it is
covered by a t(_).
The precedence relations are summarized in Table 7.I. We treat all
operators jointly with respect to other terminals. Empty slots signify erro-
neous combinations of contiguous tenninals.
A functor-functor relationship is the only potentially con icting
one: to establish the precedence relation for a given Top and Input, we
must consider their priorities and types (sometimes even some broader
context should be considered but this might require changes in the other-
wise deterministic algorithm). If the priorities differ, the functor with
lower priority is taken as stronger, according to the conventions of Pro-
log-I0. (Notice, however, that when Top is stronger, Input cannot be a
203 7 Toy: An Exercise in Implementation

TABLE 7.1

Precedence Relations for the Operator Grammar of Terms

Qvns id I bottom dot

mrmll
~ IIH
< nu
>-liar
I -=< <IIHl
]
-< <IIHl >* >
{ =<
}
rVIunn-V vsVI I- ._.__|_.
vu-
VV 1
I >3

I VvI-IV
<e
-'3! A
AAA A
A
AW I
>3

bottom <<<<<<<<<< >

dot . I

' Top can be any pre x or in x functor, i.e. Types = [xf] and Types = Iyfl are ex-
cluded.
Input can be any in x or post x functor, i.e. Types = [fx] and Types = [fy] are
excluded.

pre x functor, and when Input is stronger, Top cannot be post x.) If Top
and Input have equal priorities, their types must be examined (see below).
Mixed functors require special treatment. In most contexts, their
inherent ambiguity is apparent: only one of a functors types can be
properly attributed to it. For example, let Input be &, an [xfy, fy] functor,
and Top a left parenthesis not covered by a non-terminal:
.... .. ( &
Surely, & can only be a pre x variation of this mixed functoran in x
variation is excluded. Likewise, if Top is $, an [xfx, xfl functor, covered
' Recall that our version of Prolog allows a mixed functor to have only one binary and
one unary type, both with the same priority.
 7.4. Interpretation of Prolog-I0 in Toy-Prolog 209

by a non-terminal, and Input a right bracket:

.... .. $ Term ]

then $ certainly cannot be a post x functor. In such situations, we can

disambiguate the mixed functor by removing the incompatible type
from its representation. For example, we replace ff(&', [xfy, fy], Prior-
ity) with ff(&', [fy], Priority).
The relation in Table 7.I is implemented by the procedure es-
tablish_precedence (lines I95-204), which takes the two terminals and the
position of Top. It fails given an incorrect combination, otherwise it suc-
ceeds with the fourth parameter instantiated as gt (Top is stronger) or lseq
(Top is not stronger). When both tenninals are mixed functors, the proce-
dure tries to disambiguate their types. The last two parameters are instan-
tiated as the new top and new input tenninal, to be used in the next step
(usually they remain unchanged).
The real job is done by the procedure p which retums gt or lseq, or-
when functors are involvedgt(NewTop. Newlnput) or lseq(NewTop,
Newlnput). It fails given an erroneous pair of tenninals.
Table 7.l has 80-odd nonempty entries, but it can be easily simpli ed.
First of all, we can treat bottom and dot separately; see the last two
clauses of p (lines 240-241). Next, we consider slots with ="the rst
four clauses (lines 206-209) take care of this, and the remainder of p can
operate with the six slots cleared. Now we are left with a I0 x I0 table
with three different rows and three columns. Table 7.2 depicts the situa-
tion after combining identical rows and columns.
TABLE 7.2
Slrnpllletl Precedence Relations

vnsid

I.

I:t|-I

' Top and Input cannot be separated by a non-

terminal.
Input cannot be a pre x functor.
Top cannot be a post x functor.
ZIO 7 Toy: An Exercise in Implementation

The next six clauses of p (lines 2I I-222) take care of the six noncon-
icting slots in Table 7.2. The procedure restrict (lines 265-271) is used to
test and possibly disambiguate the type of a functor. The procedure per-
fonns set subtraction for sets given as lists; it will fail if the difference is
an empty set.
Now we must try to resolve a con ict in the remaining slot. A closer
look at the grammar allows a re nement of this slot (see Table 7.3). The
l2th and l3th clauses of p (lines 229-238) are responsible for situations
when the priorities differ. Again, we also attempt a disambiguation of
types.
The Ilth clause (lines 225-227) applies to functors with equal priori-
ties. Table 7.4 shows the precedence relation in this case. We allow all
combinations that can be disarnbiguated without analysing broader con-
text to the left or to the right of the two functors. For example, an xfy
functor f is weaker than an xfx functor g because the tenn
AfBgC
cannot be interpreted as
( AfB )gC
gs left argument would have, incorrectly, the same priority as g.
The relation of Table 7.4 is implemented by the procedure ip (lines
3I9-332), which returns lseq, gt or err. Con ict resolution is performed by
the procedure res_con (lines 273-291), which also returns lseq, gt or err
(err is later rejected by do_rels called in p). It also retums disambi-
guatedsometimes unchangedfunctors.
If only one of the terminals is a mixed functor, we choose a non-
con icting interpretation by comparing slots in Table 7.4. This is done by

TABLE 7.3
A Re nement for Two Operators

Hpre x III x 3 st x

pre x

in x

post x vvnVA
 7.4. Interpretation of Prolog-I0 in Toy-Prolog ZI I

TABLE 7.4
Precedence Relations for Operators with Eqnal Priorities

lnputs .
lYP
Top's I
type xfy xfx xf 5- *5. ::= re
>- II
~ II
I I AI A
>~ >- II
A

yf >* >*

H I. -

> > --
......

' Top and Input must be separated by a non-tenninal.

I

Top and Input must not be separated by a non-tenninal.

the procedure match_rels (lines 297-300). For two mixed functors we

extract a subtable of relations for each possible pair of interpretations; see
Table 7.5 (and lines 286-289). The situation is clear if all four slots are the
same. Otherwise there are only four pattems which can be correct: when
one of the rows or one of the columns contains two err slots. Detailsin
the procedure res_mixed (lines 302-3l7).
The procedure read(Term, SymbolTable) perfonns the two phases of
the readersee lines 65-69. It retums the symbol table with variables
from this tenn. The table is used by the interactive driver (see Section

TABLE 7.5
The Snbtable Template for Two Mixed Fnnctore:
tbeBInaryandUnary'I"ypesAreCo|||paredIrIth
EachOther

- -...,... -...,...
TTopBin RelBB

"M @ R-"=8
ZI2 7 Toy: An Exercise in Implementation

7.4.2). If data are incorrect, the parser will stop on the rst bad symbol
and read/2 will skip characters up to the nearest full stop after this symbol
(which may also be a full stop). The built-in procedure read/I simply
encapsulates read/2.

7.4.4. Grammar. Preproeessor

The grammar rule preprocessor (lines 482-583 in Appendix A.3) operates

according to the principles presented in Chapter 3. The list of lefthand side
terminals (usually empty) is connected to the output variable of the lefthand
side non-terminal. Calls on the procedure terminal (Section 3.1) are pre-
executed" for e iciency. By way of explanation, here are two examples. The
rule
a>lPl.b.lq.rl.-
is translated into
a(lPlXl.Z)=-b(X.lq.r|Yl).(Y.Z).
Therule
a+b,[q,r],c,[s].
is translated into
a(X.Z)=-b(X.lq.flYl).(Y.l$|Zl)-
(The translation of a list of terminals is true, absorbed by the next item's
translation; see combine, lines 540-542).
Conditions/actions (other than a single cut) are passed to the prepro-
cessor as {}(C); see the procedure malreterm in the reader, lines 345-
346). The functor {} is stripped off by the procedure transl_item, line
550.
Righthand sides separated by semicolons are preceded by a non-
terminal de ned as
dummy > [].
This is necessary when altematives start with different terminals. For
example, the rules
a>[p],b. and a>[q],c.
would be translated with
(lP|Y].Z) and a(lq|Yl.Z)
 7.4. Interpretation of Prolog-I0 in Toy-Prolog ZI3

as a lefthand side. Consequently the rule

a>[Pl.b:lql..
must be translated as
3(X.Z-)2-'dummy(X,[p|Y]),b(Y,Z);
dummy'(X,lllV]),c(V,Z).
For simplicity, this has been applied to all rules with altematives.

1.4.5. Library
The library (Appendix A.3, lines 585-1002) contains de nitions of about
20 built-in procedures (note that several simple procedures are also de ned in
the kernel le, appendix A.l). Their de nitions in Chapter 5 can
be treated as design documentation. Their implementation is largely
straightforward. We shall comment on a few not quite obvious passages.
The procedures clause(Head, Body) and retract(Clause) are back-
trackable, i.e. can be used in failure-driven loops that generate or re-
move all matching clauses. Here is a description of the generator (the
other procedure is programmed similarly). We are going to visit all
clauses of a procedure and suspend execution each time we get to a
matching one. This is achieved by setting up a recursive loop with its step
distributed between two clauses (see the procedure remcls/7, lines 814-
822). The rst clause does the matching. Upon mismatch, we immediately
proceed with the second clause, i.e. conclude the step. If the matching
succeeds, the generator succeeds, too, but with a pending altemative. A
failure later on resumes the second half of the step.
The procedures write and writeq both encapsulate the procedure out-
term(Tenn, With_or_withouLquotes) which rst uses numbervars (lines
623-632) to bind all variables in Term, and next calls
outt( TernLafter_numbervars, Context, With_or_withouLquotes ).
Context speci es the essential features of a functor whose argument is
Term. If it is not an operator, or there is no extemal functor, then Context
is fd(_, _). Otherwise, Context is fd(ff(Priority, Associativity), Dir). Term
may be to the left (Dir = I) or to the right (Dir = r) of the functor. Associ-
ativity may be a(l) or a(r) for left- and right-associative functors, and na(l),
na(r), or na(_) for non-associative functors. Context is tested by the pro-
cedure out)fl5 (lines 933-935) to decide whether Term should be paren-
thesized to avoid ambiguity in the case of equal priorities. Actually, the
ZI4 7 Toy: An Exercise in Implementation

testperformed by agree (lines 939-943)is rather crude (see the pre-

vious section!): sometimes we overparenthesize. The parameter of na has
only been added for homogeneity, but it could be used in a more subtle
detection of non-ambiguous cases.

7.4.6. Translator

The translator of Prolog-I0 into Toy-Prolog (Appendix A.3, lines I004-

l088) is invoked by the call
translate( SourceFileName, OutputFileName ).
Commands are translated and also executed (detenninistically), so that,
for example, a declaration of an in x functor affects subsequent parts of
the input program. The translator terminates (and succeeds) after reading
in the unary clause
end.
The program is quite easy to understand. Only the procedure lookup
may require an explanation. The table pairs variables of the clause with
consecutive integers, starting from 0. A variable is a key, so we must use
the built-in procedure eqvar to locate variables already present in the
table. The third parameter of the procedure loolcup indicates the last
number encountered (initially, -1), so that only a new variable requires
one addition. A more simple-minded solution would be to keep only varia-
bles in the table, and count them during lookup. This would require at
least (n I) * nl2 additions for a clause with n variables. ( In Toy, integers
are implemented in a particularly simple way, so this might ll the copy
stack with many dead integers). Another possibility is to apply the procedure
numbervarsinside putto Head and Body jointly.
The translator outputs bare translations. It would be helpful to have
source comments transferred to the translation, and to get source variable
names paired with numbers (see Section 7.4.1). Try this exercise for
yourself.
8 TWO CASE STUDIES

8.1. PLANNING

We shall consider planning with respect to a nite. usually small. set

of objects to which simple actions from a nite. and also small. set are
applicable. Objects constitute a closed world . The state of the "world"
is. by de nition. the set of all relationships that hold between its objects;
we also call these relationships facts about objects. As a result of an
action. some relationships cease or begin to hold; we say that an action
deletes or adds facts. A fact established by an action is also called a goal
achieved by this action. Every action transforms one state into another.
Planning consists in nding a sequence of actions that lead from a given
initial state to a given nal state.
As an example, we shall describe one of the so-called cube worlds.
There are three cubes. a, b. c, and oor. All we can do with them is stack
cubes on cubes or on the oor. There are two types of facts concerning a
cube U and an object W: U is sitting on W. and U is clear (this means that
nothing is sitting on U). The set of possible states is determined by naming
all meaningless (i.e. impossible or forbidden) combinations of facts:
-A cube X sitting on a clear cube Y;
-A cube sitting on two different objects;
Two different cubes sitting on the same cube;
An object sitting on itself.
There is one kind of action: move a single clear block, either from
another block onto the oor, or from an object onto another clear block
(the object must differ from both blocks). As a result of moving X from Y
onto Z, X is sitting on Z instead of Y, Y is clear (unless it is the oor), Z is
not clear (unless it is the oor).
215
216 8 Two Case Studies

(o)_ _ (bl
FIG. B.l (a) An initial state of the cubes world. (b) A nal state of the cubes world.

Even in this microscopic world, planning may require some sophisti-

cation. It is reasonable to postulate that a desirable fact, once added, will
never be deleted (otherwise we risk an in nite loop). However, let the
initial state be that of Fig. 8. la, described by a conjunction of ve facts:
a on oor, b on oor, c on a, clear(b), clear(c).

Let the nal state be that of Fig. 8.lb, described by a conjunction of two
goals: c on a, a on b. The rst goal is trivially achieved. To put a on b,
though, we must remove c from a, i.e. destroy an already achieved goal.
The simple strategy of achieving goals one by one (and freezing all rele-
vant facts) would not work in this case.
ln a more crowded world, a state might comprise so many facts
that its direct representation (as a list, say) would be impractical. More-
over, even a small change might require copying large data structures.
Clausal representation is free from this disadvantage but it is unwieldy
when a change must be undone, and of course planning is a trial-and-
error process. What we need is a method of incrementally describing
incremental changes, and making them easily undoable.
A state and an action determine the next state, if we assume that the
action does not affect facts not mentioned explicitly in the description of
the actions effects as added or deleted. Given an initial state and a plan,
i.e. a sequence of actions, we can check whether a fact holds in the
resulting nal state. To undo an action, we remove it from the plan (in
practice, this may be slightly more complicated).
For any particular planning problem, the initial state can be consid-
ered xed. The nal state should be given implicitly, as a conjunction of
facts to be established by a plan we are going to nd. This approach was
taken by D. H. D. Warren in his remarkable planning program, WAR-
PLAN.
 8. I. Planning 217

ln WARPLAN, a world description is separated from the planning

procedure (see Listing 8. l, pp. 221-223, lines I-26, for the description of our
cube world). Objects are given implicitly, in descriptions of actions and facts.
Actions are de ned by three procedures. The two-parameter procedure
can( Action, Precondition )
serves as a catalogueone clause per action; Precondition is a con_iunc-
tion of facts that must hold for Action to be applicable. A conjunction is
either a fact, or a pair of conjunctions constructed by the in x functor &,
e.g.cona&aonb.
Two other procedures,
add( Fact, Action )
del( Fact, Action )
give facts added and deleted by available actions (and, conversely,
actions which can add or delete a fact). Impossible combinations of facts
are listed in the procedure
imposs( Conjunction )
In these four procedures, we can use variables instead of world objects to
express general laws, e.g. a clear cube U is sitting on a cube V":

U on V & notequal( V, oor ) 8: clear( U )

For ef ciency, facts that hold in the initial state, and are unaffected by
any action, are listed in the procedure
always( Fact )
Other facts that hold in the initial state are supplied by the procedure
given( InitialStateName, Fact )
The initial state is denoted by its name, e.g. srarr. A state derived
from it by actions Al, An is denoted by the term
lnitia]StateName : Al : : An,
e.g.
start : move( c,a, oor) : move( a, oor, b ) : move(c, oor, a)
The planning program (Listing 8.2, pp. 224-226) operates indepen-
dently of speci c world descriptions. It assumes the presence of an appro-
priate data base whose coherence is the responsibility of the user.
Zl8 8 Two Case Studies

The program begins with a conjunction of facts (i.e. the description of

a desired nal state) and the empty plan. In each step, the conjunction
shrinks and/or the plan grows; successive intermediate states approxi-
mate the nal state. Roughly speaking, the plan is constructed backwards:
we look for preconditions of actions that achieve the nal state, then for
preconditions of actions that achieve those preconditions, etc. Unless a
fact holds in an intermediate state, the program chooses an action that
adds this fact, inserts the action into the current partial plan, removes the
fact from the current conjunction and adds to it the actions precondi-
tions.
A partial plan usually contains variables. For example, to achieve a
on b, we use the action move(a, V, b), whose precondition includes the
fact a on V (for an unknown V). Such variables require some care: the fact
U on c may, in general, differ from a on V, even though the two terms are
uni able. We can either use the built-in procedure = = to compare facts,
or temporarily instantiate their variables (by the built-in procedure num-
bervars) prior to the comparison.
In addition to the current conjunction and plan, the program main-
tains a conjunction of desirable facts already planned for. No newly in-
serted action can destroy any of these preserved facts.
The program is amazingly concise. In Warren's original paper it was
accompanied by many pages of detailed considerations. Hence, the ab-
sence of proper comments in the program text. Below we shall present, in
our own words, some indispensable technical explanations.
The main planning routine, plan, is called only if the nal state de-
scription is not inconsistent (lines I0-l3), i.e. if it does not imply one of
the impossible combinations of facts. plan has three input parameters-
facts to be achieved, facts already achieved (initially true; see line l3) and
the current planand one output parameter, the nal plan. The proce-
dure solve is called for each fact of the initial goal list (see lines 30-32). lt
has ve parameters: a fact to be established, preserved facts. the current
plan, preserved facts after solve has succeeded and the new plan.
Every clause of solve accounts for a different status of the fact (lines
35-39). It may be always true; it may be true by virtue of general laws
extemal to worlds (e.g. equality or inequality of objects will be
checked by this clause); it may hold in the state described by the current
plan (to preserve it, we add it to the facts planned for; see lines 83-84);
otherwise (the last clause) we choose an action and call achieve.
The procedure achieve (lines 4|-49) tries to apply a given action. i.e.
to insert it into the current plan (as the last action, or as the last but one,
etc.). The action U is applicable if it deletes none of the preserved facts,
and if its precondition is consistent with these facts and if a plan for
 8. l. Planning ZI9

achieving this precondition can be constructed. Notice that possible addi-

tions to P (preserved facts) made by the recursive call on plan are invisible
to achieve: they are only needed "locally" during the construction of the
intermediate plan Tl. The additional call on preserves (line 45) is neces-
sary because of variables in the plan. For example, the action move(b, a,
W) need not delete the fact clear(c), so preserves lets it through; however,
plan may instantiate W as c, and this ought to cause a failure.
If, for any of these reasons, the action U cannot be added at the end of
the plan, achieve will try to undo the last action V and insert U earlier into
the plan. This is only possible if V does not delete the fact to be added by
U. The procedure retrace (lines 65-73) removes from the set of preserved
facts all facts that may be established by V but are different from Vs
preconditions. Speci cally, it removes the facts added by V (lines 68-69)
and the facts that constitute the precondition of V (lines 70-7l)the
latter facts will be re-inserted by append (see lines 66, 86-87).
A few comments on the remaining procedures. A fact holds after
executing a given plan (lines 52-55), if it is given or added by one of the
actions, and preserved by all subsequent actions (if any). Two conjunc-
tions, C and P, are inconsistent (lines 76-78, 93-97) if C&P contains all
facts of an impossible combination S. except those whichlike not-
eqaalare tested "metaphysically" (see line 95). For disjoint C, S this
cannot be the casehence the call on intersect which is relatively cheap.
Two object descriptions X and Y, with variables instantiated by number-
vars in mkground (line l0l), may refer to the same object if X = Y or
X = V(_) or Y = V(_)see line 99. The procedure elem (lines 89-9|)
extracts single facts from a nested conjunction; it can be used both to test
membership, and to generate facts.
Now that you have acquainted yourself with the planning program,
try it on a richer world. Here is the world of a robot that walks around
several rooms, moves some boxes, etc. (see Listing 8.], lines 33-l0l).
Figure 8.2 depicts an initial state of this world. There are six points, ve
rooms, four doors, three boxes, a light switch, and the robot. Nine types
of facts are considered: at(Object, Point). on(Object, Box), nextto(Ob-
jectl, Object2), pushable(Object), inroom(Object, Room), locinroom-
(Point, Room), connects(Door, Rooml, Room2), status(Lightswitch,
OnO '), on oorthe latter characterizes the robot. Only the robot per-
forms actionsthere are seven of them (see lines 64-77).
' The special treatment of V's preconditions is necessary for actions which add facts
listed among their own preconditions. lf retrace simply deleted V's effects. such precondi-
tions could be lost from the list of facts which must be presented by U, and those parts of the
plan which achieve locally desirable" goals could inadvertently be destroyed in the inser-
tion process.
Z20 8 Two Case Studies

roomltl roornl2l roeml3l re-orntl.)

lightswiteh ll l
borlll
O
point (L)

oponntttl boll2)

Iq ,rr

' w ttlt U
.

pomll l
opomtllll
O
pe|ntl5l

doorltl dootl2l dootl3l dcorll

room I5)

FIG. 8.2 "STRIPS" world.

The procedure del merits a comment. It is supposed to delete more

than it shouldwe count on add to straighten the situation out. For
example, the action tumon(S) removes whatever status of S may be re-
corded (line 54); a moment later" it adds the appropriate fact (line 39).
The clauses in lines 49-50 say that a moved object X is no longer next
to" anything. However, this does not apply to the robot manipulating a
box (lines 4648)del fails, i.e. the fact is not deleted.
For sample results, see Listing 8.3, p. 227.
Although WARPLAN is a feat of ingenuity, there is much more to
planning than it does account for. For one thing, the plans it generates
need not be optimal, i.e. contain the least possible number of actions. For
example, action U in achieve (lines 47-49) is executed when it preserves
V's precondition P; if we checked that U establishes P, we might delete
actions which had been planned to establish it. A much more profound
problem: in general, it is likely that conditional or iterative plans will be
required, rather than sequential (the robot explores the world).
Even with these (and other) limitations, and despite exponential time
complexity, WARPLAN is an excellent tool for experiments with rigor-
ous world descriptions. One example is the world of a robot that assem-
bles cars. Warren has also demonstrated how his program can be used to
compile arithmetic expressions into machine code (the code is treated as a
plan for placing some values in some registers).
LISTING 8.1 WARPLANExa|:|plea of worlds.

%%%%%% WARPLAN-eubewonds

op(50, xlx, on).

del( UonW, rmva(U, V, W) ).
elear(V) meva(U V W) ).
el( UonZ, move( F
el( elear(W), move( .<F2 ii
0.0.DD

E ( rmva( U, V, oor),
UonV 8. notequal(V, oor) 8. elear(U) ).
(move(U, V, W),
elear(W) 8. UonV 8. notequal(U, W) 8. etear(U))
-I l -L.-L-L-L-L lr|'poas(XonY8.elear(Y)).
Qm hu~_,v.0en~|o>m:~uro- l|rpoas()tonY&XonZ8-netequal(Y, 2)).
18 lrrpoas(XonZ&YonZ&netequal(Z, oor)&notaqud(X, Y))
19 lrrpoea( Xonx ).
20
21 %1'hatl1reeblocltsp|'oblem.
22 glven( start, aon oor ).
23 glvant start, bontloor ).
24 glvent start, eona ).
25 olvent start, eteartb) )-
26 glven( stall, elear(e) ).
27
28 :- plans(eona 8. a onb. start ).
29 :- plans(aonb 8. b one. Start ).
2- d0b 'Ol')'), rede ne.
f
I! %%%%%WARPt.AN-t:l1oSTFllPSproblem

EEEEEEEE ;%?iahtoud1(D, Fl1,Fl2) ).

8'F
5a .3. .75N

as
c
5E E .75c
.2). U)
:- l, lall.
$ci $ 38 3$li58lt3-' 8
4-B
49 *
no
l( nextto
22 I, tall.
I, tall.

50
51
52
del( nextto
del( Ont33X. 53? 3.'.,=:;.~: t==.:. gages?
del( orilcor, eirrbon l Bl) .
gs
::- -. . xH"cc
?". . "

53 del( lnmom(robot,Z), gothroud1(D,Fl1,H2) ).

54 del( atatus(S,Z), tumon(S) ).
55
56 rmved( robot, goto1(
57 movedt robot, goto2(
58 movedt robot, pushto FF?
59 rnovedt X, pus)-tto( X. Y. "<33 1-vm"'.""1"
60 rmved( robot, ellmbon(B) ).
211
LISTING 8.1 (Continued)

ed( robot, ellmbotl(B) ).

E3edt robot, gothroul(D,Fl1,Fl2) ).
goto1( P, H ),
35.nroom(P,Ft) 8. hroom(robot,Fl) 8. orllloor ).
can( goto2( X, Fl ),
38 1283 9 lnroorn(X,Fl) 8 lnreom(robot,H) 8 orlloor ).
68 can( tumon( lights\vltch(S) ).
69 on(robot.box(1)) 8 na1rtto(box(1), llghtswlteh(S)) ).
70 can( pushtot X, Y, Fl
71 pusl1able(X) 8 I 8 lnmorn(X,Fl) 8
72 nexttot robot, X "'5
73
74
can( gothroud1( D, Fl N-< 5%.-1.?_m
eennaets(D,H1,Fl2) 8- lnmom(robot,Rt) 8
75 naxtto(robet, D) 8- orllloor ).
76 can( cllmbol1(box(B) ), on(robot,box(B)) ).
77 can( elirnbont box(B)), nexttot robot. box(B)) 8- onlloor ).
78
79 alwayst lnreom(D,Fl1) ) :- alwayst connoets(D, Fl1,Fl2)
80 alwayst eonnoctst D, H2, H1 i
eonnaets1( D, H1, H2 )
-0 I

81 alwayst eonnacts(D,Fl1,Fl2) :- eonnaetstt D,Fl1,Fl2 )

ii

82 alwayst pushablet box(N)) ).

83 alwayst loelnroom(polr|t(N),roorn(1)) ) :- ranget N, 1,5 )
84 alwayst loelnroom(polnt(6),room(4)) ).
85 alwaya( lnroom( lightswlteh(1),roorn(1)) ).
86 iltritsl ll5o|1l5Wll|1(1)-PlI'll4)) )-
87
8-8 cenneetstt deor(N),room(N),roorn(5) ) :- ranget N, 1,4 )
89
90 Ii
91
92
ra
L A-3-2 _zFF. . C:-_: If...L+1, ranget M, L1,N ).
93
94 |n'poss( at(X,Y) 8- at(X,Z) 8. notaquaI(Y,Z) ).
95
96 glV9r)( strbst, 8I( box(N),polr|t(N)) ) 2- range( N, 1,3 ).
97 glvent WW1, 8I( robot, po|r|t(5)) ).
98 glven( strbst, lnroom(box(N),roorn(1)) ) :- rangat N,1,3 )
99 glven( strbst, ontlcor ).
100 givert( strlpst, statua( Igt1tawlteh(1),olt) ).
101 glvant strbst, lnroom(robot,roorn(1)) ).
102
103 % A tow tests.
104 :- plane( at( robot, polnt(5) ), etrbsl ).
105 :- plane( at( robot, polnt(1) ) 8| at( robot, polnt(2) ), strbst ).
106 1- Pll t Bit robot.P0|rI1t4))- W951 )-
107 :- plans( statua( I|gl'ltcwltch(1), on ), strbst ).
108 :- plans( at(robot. polr|t(6) ), etrbst ).
109 :- pla.na( nextto( box(1), box(2) ) 8|
110 nextto( box(3), box(2) ), strbst ).

Z22
LISTING 8.2 WARPLANTl|e general planner

1 %%%%%% WAFIPLAN-ASystemtorGenerathgPlam
2 %%%%%% (pnlbhltadwllttrnklrdpemlbsbnolthehnlnr
3 %%%%%% DavklH.D.Warren)
4
5 %%% Ihegenerdplanner.
6 9;, ..__.._._._.__...
7 :- op(200,Ily,&), op(100,yhr,:).
8
9 % Generate and output a plan.
10 Plarwt 0. _) =-
11 lnconslstamt C, true ), I, write( 'lrrpossble.' ), nl.
12 Plarwt 0. T) =-
13 plan( C, true, T, T1 ), outputt T1 ), I.
14 an _I _) :'
15 write( Cannot do thls.' ), nl.

it#2
16
17 output()ts:X) :-
18
19
20
nurrbarvars(Xs:X,1,_), output1(
output(_) :- wrIo('Nothlngneed rs '.' ).

21 output1()ta:X) :- I, outputttlts), output2(X.':').

22 output1(X) :- out|:|ut2(X,':').
23
24 output2(ltem.Punet) :- wrttetltem), wrIe(Punet), nl
25
26
27
28
29
30
% Maln planning routhe.
% De nllom ol
% see epeeltb world deserbt
plant X80. P. fl
si umuumumuw-

31
32
33
I, eelvet X, P, ."1,5
plan(X,P,T,T1) :-
ill:-I

5"is
-1.. U ." -I ."T2).
.17 _-|.'-4 all

eolvlngagoal.
always(X).
aw? holcls(X,T), and(X,P,P1).
55525? FxxxxsSw mmi gaaaa552::xv??? U,P,T,T1).
j
in-P ...X

%Metl\odsolael\lev||-|gaqoal-
% byextenslon:
levo(_, U, P,T,T1:U) :-
U,P), ean(U,O), notheonsbterl(C,P),
T,T1), preserves(U,P).
rtlon:
U, P, T:V, T1:V) :-
)t,V), retraee(P,V,P1),
Pt,T,T1), preserved(X,V).
Ilataetholdslnaglvonstate.
_:V) :- acld(X,V).

* holds(X,T), preservecKX,V).
E5525
ggisgg
$ iE83.'S$-5$ci8$ $3-138i % Prove that an action preserves a tact.
58 preservest U, X80 ) :- proservod( X, U ), praservest U, C)
59 preserveat _, tme ).

223
LISTING 8.2 (Continued)

v) =- check(pres(X,V)).
8828
ii i -x
...rrltground( xav 1. not del( x,v).
64 % Fletraclng a goal already achieved.
65 retraeet P, V, P2 ) :-
66 can( V, C ), retraee( P, V. C, P1 ), append( C. P1, P2 ).
67
68 retraee( X&P, V, C, P1 ) :-
69 add( Y, V ), X --. Y, I, reIrace( P, V, C, P1 ).
70 retraeet X&P, V, C, P1 ) :-
71 eIem( Y, C ), X - Y, I, retracet P, V, C, P1 ).
72 retraeet XGP, V, C, X&P1 ) :- retraee( P, V, C, P1 ).
73 retraee( true. _, _, tme ).
74
75 t. lneonsbteney with a goal already achieved.
76 lnconsbletit C, P) :-
77 Wlkawuodt CGP I. lmtm t 8 I.
78 check( lnterseett C. S ) ), ln'pIled( S, CGP ), I.
79
80 % % % Utilities.
81 % .......-
82
8I8l'l')( Y, P], XII Y, I.

"J";sf:
3-.555.,l'._.1)3% l. GPIZIOIIK C, P, P1 ).

8918 128
iiii
E3
89 eIern( X,Y&_) :- eIern( X,Y).
90 elern( X,_8|C) :- I. elem( X,C).
91 elem( X,X).
92
93 ln'olled(S1&S2,C) :- I, h'plIed(S1,C), In'pIled(S2,C)
94 lrrplIed(X,C) :- eIem(X,C).
95 ln'plled( X,_) :- X.
96
97 interseeI(S1,S2) :- elernt X,S1), elem( X,S2).
98
99 notequal( X, Y) :- not X-Y, notlt-'V'(_), notY-'V'(_).
100
101 rrI<ground(X) :- run'bervars(X,O,_).

Z24
LISTING 8.3 WARPLANSam||Ie reldtn

Toy-Prolog Btenlng:
?-)t.IiIeswItl'|outIesteasesbut\vIth end. atthe end
:- eonsuttt planner ). consult cubes ).
?-
:-planstcona 8. aonb, start).
start :
move(
move(
move(
7.
3?;
:-planstaonb 8. bone, start).
start :
rmve(
move(
move(
7.
3?;
;delop( on ), reconailt strbs ).
:- planst at( robot, polri(5) ), strbst ).
Nothlng need be done.
7.
1- 010rBt01t robot. 0051(1) ) 8- it r0001. 0010112) ). 8111001 )-
lnposslble.
7-
:- planst aI( robot, poIrn(4) ), strbst ).
stripst :
goto1(poIr|I(4),room(1)).
:- plans( etaIus( ll9|'|tswltch(1), on ), etrbst ).
strbst :
goto2(box(1),roorn(1)):
pt.|sl1to(box(1),ll9htswltch(1),roorrI(1)):
elimbont box( 1 ) ) :
tumon( llghtswtteht 1 ) ).
7-
:- planet at( robot, poIrd(6) ). strbst ).
slripst :
2
iii
90102( d0'0r( 1 ).
00Ihr00oh(d00rt ).l'00l'l1l5))
0010210004 4).
a0Ihr00oh(d00rt4 ).l'00l'l1l4ll
0010110010116).
?-
r00mt J5 -u-n

:- plans( nert1o( box(1), box(2) ) 8| next'to( box(3). box(2) ). elrbst )

stripst :
goto2( box( 1 ), roorn(
pusl'Io( box( 1 ). box( mt1))=
coto2( box( 3 ), room(
pusl1to( box( 3 ), box( N-L~_l 1" -"1" '-P -a I)-
?- stop.
Toy-Prolog. end ol eesslon.
Z26 8 Two Case Studies

BIBLIOGRAPHIC Nores
WARPLAN is described in Warren (I974). Our presentation has been
greatly in uenced by this excellent paper. The program we publish here is
a slightly cleaned-up version of the text given in Coelho et al. (I980),
where all the mentioned examples of worlds can also be found. The ro-
bot's world was introduced by Fikes and Nilsson (I971) as a test case for
their system STRIPS; Warren (I974) used it to compare the performance
of the two systems. An extension of WARPLAN, intended for generating
conditional plans, was described in Warren (I976).

8.2. PROLOG AND RELATIONAL DATA BASES

ln this section, we shall be primarily concemed with data bases in the

limited sense: a data base is a purposefully structured collection of stored
data, often pertaining to an organisation (e.g. a bank, factory, university,
warehouse). ln a relational data base all data are conceptually grouped
into relations, which are usually depicted as rectangular tables as in Fig.
8.3. A column in the table is called an attribute and referred to by a name,
e.g. dno. All values of an attribute belong to a common domain, e.g. each
salary belongs to integers. A relation is a set of tuples (table rows) which

empno name dno salary rngrno

13 Miller 1500 ' 19 -1

. 21 Jones 1000 13 I
35 Brown 1000 21 I
- 38 White 800 1 35
1.3 Smith 1200 1 13
61 Thomas 85-0 21
89 Morgan 1050 35
i 42 Miller _a-0-AhJ_a-Q-lo 850 35
-- . . _ I

dno name rngrno

-It Pubttclilela t ions 21

2 '
Security 1.3
FIG. 8.3 Contents of a relational data base.
 8.2. Prolog and Relational Data Bases 227

consist of attribute values, e.g.

< 38, White, I, 800, 35 >.
Tuples belong to the set described by a relation schema which speci es
names, domains and order of attributes, e.g.
EMP < integer empno, string name, integer dno,
integer salary, integer mgrno >
DEPT < integer dno, string name, integer mgrno >
Two tuples may share the value of an attribute, and thus implicitly fall
into one group; for example, Brown and Thomas are both subordinates of
a manager whose number is 21.
A relation can be changed by inserting, deleting or updating some of
its tuples. These operations are referred to as data manipulation. ,_
A query to the data base is answered by enumerating tuples of the l I
resulting relation (or by computing an aggregate function, such as ""aver- 1
age" or "total", over these tuples). Most queries are expressible in tenns
of the following primitive operations on relations. /'""
Selection chooses tuples for which a given condition holds; for exam-
ple, we can select from EMP those employees of department I who
cam over 900 (there are three such tuples).
Projection neglects some attributes and (possibly) reorders the remain-
ing ones; for example, we can project EMP over name, empno, and
salary, to get
< Miller, I3, I500 >
and seven other triples.
Joln of two relations A, B fomis a new relation. It consists of those
concatenations of tuples from A with tuples from B, for which a given
condition holds. For example, the join of EMP and DEPT, such that
department numbers coincide, consists of the tuple
< 2|, Jones, l, I000, I3, I, PubIicRelations, 2l >
and six other 8-tuples.
Unconditional join is a product; for EMP and DEPT the product con-
sists of sixteen 8-tuples.
Finally, set operations, namely union, intersection and difference, can
be applied to two relations whose corresponding attributes belong to
the same domain, i.e. whose schemata differ only in names.
Much of this conceptual framework is naturally translated into Pro-
log. A relation is modelled as a procedure made of unit clauses which
Z28 8 Two Case Studies

correspond to tuples, for example:

'EMP( I3, Miller, 0, I500, I9 ).
'EMP'( 2], Jones, I, I000, I3 ).
etc.
(we use quotes to prevent capitalized names from being treated as varia-
bles). To change a relation, we use the built-in procedures assert and
NZIFGCI.
Primitive operations on relations are expressed in terms of procedure
calls. For example, the procedure
s( Empno, Name, Dno, Salary, Mgmo) :-
EMP( Empno, Name, Dno, Salary, Mgmo ),
Dno = I, Salary > 900.
can be used to generate all tuples for employees of department I who earn
over 900 (i.e. to implement selection):
:- s( E, N, D, S, M), write( ( E, N, D, S, M ) ), nl, fail.
Better still, we can substitute I for Dno and remove the test:
s( E, N, I, S, M) :- EMP( E, N, l, S, M), S > 900.
The procedure p can be used to implement projection:
p( Name, Empno, Salary ) :-
EMP( Empno, Name, _, Salary, _ ).
The composition of these two operations can be expressed in Prolog quite
succintly:
s_then_p( Name, Empno, Salary ) :-
EMP'( Empno, Name, I, Salary. _ ). Salary > 900.
Or we can put this directly into a query:
:- EMP'( E _
write( ( N $0.2 <0:-' jIIs'= -'1-.. Qua ="'v
Finally, here is the join of EMP and DEPT over coinciding department
numbers:
j( Empno, NameE, DnoE, Salary, MgmoE, DnoE, NameD,
MgmoD ) :-
EMP'( Empno, NameE, DnoE, Salary, MgmoE ),
DEPT'( DnoE, NameD, MgrnoD ).
 8.2. Prolog and Relational Data Bases Z29

All these operations are neatly explained in tenns of static interpreta-

tion of procedures (try for yourselfl). Set operations are even more
straightforward. Let a(X| , X,,) and b(X| , X,,) denote generators of
tuples, such as DEPT( D, N, M ) or p( N, E, S ). We have
aUNIONb( X1, ..., X) :-
a( X), ..., X" ) 1 X|
aINTERSECTIONb( X|,
a( X1, ..., X ) , b( X|, .2
3*-~.?<
aDIFFERENCEb( X|, ..., X,,) :-
a( X|, ..., X,, ) , not b( X1, X,, ).
Queries which involve only primitive operations can be answered
without actually creating the resulting relation. Its tuples can be generated
by a failure-driven loop and displayed immediately. To compute an aggre-
gate function, however, we need the whole attribute (column) at once. We
can construct it by means of the procedure bagof (see Section 4.2.4); for
example:
:- bagof( Salary, EMP( _, _ , _, Salary, _ ), Salaries ),
max_of( Salaries, MaxSaI ), write( MaxSaI ), nl.
Sometimes we can also use bagof for ef ciency. For example, we look
for employees of department I who earn no more than a I000 and who are
FTU members since at least I980:
:- EMP( E, N, I, S, _ ), S =< I000,
FTU( E, _, _, Dateloined ), Dateloined =< I980,
write( ( E, N ) ), nl, fail.
With those implementations of Prolog which do not support clause index-
ing, the entire relation FTU would be scanned many times. Instead, we
can precompute the necessary set:
:- bagof( Empno, ( FTU( Empno, _, _, Dateloined ),
Dateloined =< I980 ), FTUMembers ),
EMP( E, N, I, S, _ ). S =< I000.
member( E, FTUMembers ), write( ( E, N ) ), nl, fail.
There are queries which cannot be expressed as a composition of
selections, projections, joins and aggregate functions. A classical exam-
ple: nd every employee who eams more than at least one of herlhis
superiors. The relation is a superior of" is inherently transitive, but we
can only express the relations is an immediate manager of is an
immediate manager of an immediate manager of " etc. In Prolog, how-
230 8 Two Case Studies

ever, the problem is easily solved. For example, we can de ne a proce-

dure to generate managers salaries:
mgr_sal( Mgmo, Salary ) :- EMP( Mgmo, _ , _, Salary, _ ).
mgr_ sal( Mgmo, Salary ) :-
EMP'( Mgmo, _, _, _ , MgrMgrno ),
mgr_ sal( MgrMgmo, Salary ).
(try to rewrite it so as to avoid the repeated pass through EMP). From the
standpoint of the caller, this generator is indistinguishable from those made of
unit clauses. The query can be written as follows:
:- EMP'( Empno, _, _ , Salary, Mgrno ),
once( ( mgr_sal( Mgmo, MgrSal ), MgrSal < Salary ) ),
write( Empno ), nl, fail.
A relation which is computed rather than stored (e.g. s, p, s_then_p, j
above) is called a view in the relational data base terminology. A view
results from primitive operations on stored relationsalso indirectly, via
other views-and it changes as those relations change (and conversely, a
change in a view might in uence those relationsbut this poses quite
nontrivial problems). The relation mgr_sal, however, can only be ob-
tained by embedding primitive operations in a host programming lan-
guage, fumished with recursion or iteration. An important advantage of
Prolog is its ability to express tuples, views, and special programs in the
same language. In particular, it offers a possibility of enforcing integrity
constralntsapplication-speci c conditions of the coherence of data.
Constraints should be tested prior to any change to a relation. For exam-
ple, we can use this procedure to insert only correct tuples:
insert( Tuple ) :-
correct_insert( Tuple ), !, assertz( Tuple ).
insert( Tuple ) :- signal_violation( Tuple ).
correct_insert( EMP( E, _, D, _, M ) ) :-
!, E=\= M, DEPT( D, _, _ ). % there is such a dept
correct_insert( _ ). % others are OK
On the whole, Prolog is a powerful tool for data base applications.
Admittedly, there is more to data base systems than our presentation
suggests. For one thing, the size of a real data base may far exceed the
capacity of any existing Prolog implementation. The model described in
Chapter 6 ought to be augmented: clauses would be stored on disk and
handled by standard or specialized access methods. Second, every practi-
cal data base implementation should address problems such as concurrent
 8.2. Prolog and Relational Data Bases 231

execution of users commands, recovery after hardware failures, etc.,

etc. There are no ready solutions in Prolog but presumably they can be
programmed into it.
Surprisingly, Prolog is, in a sense, too strong, too unrestricted. For
example, to ensure the conformance of a tuple with the relation schema,
some form of type checking is required, presumably as explicit tests.
Unrestrained use of assert/retract may also ruin the integrity of a data
base in other ways. Consequently, Prolog should rather be considered a
tool for implementing more restricted user interfaces: queries and com-
mands in a user language are analysed (types checked, integrity ensured,
etc.), and only then translated into Prolog.
A particularly attractive option would be to query the data base in a
natural language. Several encouraging small experiments have been car-
ried out. At the moment, though, this is much more a research problem in
its own right than a generally available programming technique.
Relational data languages, notably Sequel and Query-by-Example,
provide syntactic sugar for relation schema de nitions, data manipulation
and queries (involving Boolean expressions and compositions of primitive
relational operations). In contrast with natural language interfaces, a Pro-
log implementation of a relational data language is a programming task of
moderate complexity.
We shall now present Toy-Sequel, a relational data language pat-
temed after Sequel and implemented in Prolog (see Listing 8.4). With the
exception of aggregate functions, expressions in tuple speci cations and
some exotic features, it supports all that is essential in Sequel. Extensions
are relatively easy to introduce (we left them out to make the program
shorter). To give the avour of the language, here is an annotated conver-
sation with our program, initiated by the call

:- toysequel.
To begin with, we specify a few relation schemas:
create EMP < string name, integer salary, integer dno >.
create DEPT < integer dno, string manager >.
create BoardMembers < string name, string position,
integer seniority >.
Now we insert some tuples:
into EMP insert < "Brown", I000, I >, < "White". 800, l >.
< "Miller", 850, I >, < "Barry", 900, 2 >.
< "Thomas", 850, I >, < "Morgan". I050, I >.
into DEPT insert < I, "Jones" >, < 2, "Smith" >.
232 8 Two Case Studies

We can ask what relations the data base contains; Toy-Sequel dis-
plays their names (its responses are italicized):
relations.
B0ardMembers
DEPT
EMP
What is the schema of EMP?
relation EMP.
string name
integer salary
integer dno
A select expression determines a set of tuples. They may be dis-
played. For example, who in departments other than 2 earns at least I000?
select from EMP tuples < name, salary >
where dno < > 2 and salary >= I000.
Brown l0OO
Morgan I050
Or they may be inserted elsewhere:
into EMP insert
select from DEPT tuples < manager, I000, dno >.
In the absence of where the condition is taken as true.
Both managers have the salary I000. We can given Smith a raise:
update EMP so that salary = I200 where name = "Smith".
Fire Barry:
from EMP delete tuples where name = "Barry".
If several relations are involved. e.g. in a join, attribute names may be
ambiguous. To disambiguate, qualify them with relation names. For ex-
ample:
select from EMP, DEPT tuples < name, EMP_dno, manager >-
where EMP_dno = DEPT_dno.
(Actually, EM P_dno may be replaced with dno: an unquali ed attribute
name is quali ed with the leftmost appropriate relation name.)
A relation may be accessed in several places at once. For example, to
compare salaries of different employees we need the product of EMP by
EM P. We must give one of the occurrences an alias name and so allow
 8.2. Prolog and Relational Data Bases 233

unambiguous references to attribute values. The following query joins the

relations EMP, DEPT and EMP alias Mgr, to nd employees who eam
more than their (immediate) manager:
select from EMP, DEPT, Mgr = EMP tuples < EMP_name >-
where EMP_dno = DEPT_dno
and DEPT_manager = Mgr_name
and Mgr_salary < EMP_salary.
Morgan
Again, the quali cation with EMP is super uous, as well as the quali ca-
tion of manager.
A similar condition can be used to give a raise of half the difference in
salaries to those who eam over I00 less than their manager:
update EMP using DEPT, Mgr = EMP
so that salary = EMP_ salary+ ( Mgr_saIary - EMP_salary)/2

where EMP_dno = DEPT_dno

and manager = Mgr_name
and Mgr_salary > EMP_ salary + I00.
Two miscellaneous queries. Find employees whose names do not
begin with M:
select from EMP tuples < name > J_
where name < "M" or name > = l
( ve of them). And nd EMP tuples with a nonexistent department num-
ber-this is a kind of (manual...) integrity checking:
select from EMP tuples < name. salary, dno > where
not < dno > in select from DEPT tuples < dno >.
in denotes set membership. The name dno in the nested select expression
pertains to DEPT.
Time to nish. The relation BoardMembers will not be necessary.
after all:
cancel BoardMembers.
Store the data base in a le:
dump to AAA.
(next time we shall begin with
load from AAA.
234 8 Two Case Studies

and resume at this point). Finally, retum control to Prolog:

stop.
We shall not go into details of the Toy-Sequel interpreter. The ration-
ale for its design was given above; the program is (almost) self-document-
ing. The following remarks account for a few central technical decisions.
The main procedure, toysequel (lines 4-7 in the listing), repeatedly
reads and executes commands. The procedure getcommand (lines 9-I I)
retums a Prolog goal, which is the translation of a command, and a ag.
The ag remains uninstantiated if the command is correct, otherwise it is
instantiated as error. The procedure docommand (lines I3-I4) executes a
correct command's translation, and does nothing in the case of errors.
A command is processed in three phases. The text, terminated with a
dot, is read in (lines 32-43) and then passed through a scanner, imple-
mented as a metamorphosis grammar (lines 45-84). It classi es tokens as
names, strings, integers and single non-alphanumeric characters. A list of
tokens goes to the command compiler-a metamorphosis grammar which
is the core of the interpreter. The grammar consists of ll parts, one for
each Toy-Sequel command (see lines I13-I23).
All commands, except load and stop, manipulate the relation cata-
logue. The catalogue is implemented as a three-parameter procedure
r e I, with a unit clause for each relation schema. A schema stores the
name of a relation, a generator of this relation's tuples and a "frame" of
symbol table entries linking attribute names and types to variables in the
generator (see lines l3I-I41). For example, the command
create EMP< string name, integer salary, integer dno >.
adds the clause
r e l( EMP, ' EMP( Name, Salary, Dno ),
[ attr( name, string, Name ), attr( salary, integer, Salary ),
attr( dno, integer, Dno ) ] ).
Blanks are added to relation names in generators to make con icts with
other procedures less plausible.
The command processors for select, insert, delete and update main-
tain a symbol tablea stack of frames taken from the catalogue. For
example, attribute names in the command
select from EMP, Mgr = EMP, DEPT
tuples < name, Mgr_dno, manager >
where dno = DEPT_dno and manager = Mgr_name
and salary > Mgr_saIary.
 8.2. Prolog and Relational Data Bases 235

will be looked for in the following symbol table (see lines 208-223, and
96-I02):
[ EMP : [ attr( name, string, NameEMP ),
attr( salary, integer, SalaryEMP ),
attr( dno, integer, DnoEMP ) ],
Mgr : [ attr( name, string, NarneMgr ),
attr( salary, integer, SalaryMgr ),
attr( dno, integer, DnoMgr ) ],
DEPT : [ attr( dno, integer, DnoDEPT ),
attr( manager, string, ManagerDEPT ) ] ]
(Nested select expressions would push their own frames onto this stack-
see line 277.)
The product of these three relations will be generated by the following
calls retrieved from the catalogue:
EMP( NameEMP, SalaryEMP, DnoEMP ),
EMP'( NarneMgr, SalaryMgr, DnoMgr ),
DEPT'( DnoDEPT, ManagerDEPT )
The condition will be translated into a Prolog goal (see lines 236-237, 239-
362). The goal will be executed immediately after the generators (lines
I66-I68). Attribute names in the condition will be translated into varia-
bles from the symbol table. Thus,
salary > Mgr_salary
will become
SalaryEMP > SalaryMg:r
The "equalities"
DnoEMP = DnoDEPT, ManagerDEPT = NameMg:r
will be processed at compile time, by binding variables together (line 318),
so that actually only six different variables will occur in the generators.
The tuple pattem (lines 225-234) will also contain variables from the
symbol table:
[ NameEMP, DnoMgr, ManagerDEPT ]
One such tuple will be displayed in every step of the failure-driven loop
(lines I66-I68).
A construction that would certainly bene t from a more detailed ex-
planation is update. We shall comment on the example shown in the
236 8 Two Case Studies

listing (lines 396-411):

update EMP using DEPT, Mgr = EMP
so that salary = salary + ( Mgr_salary salary ) / 5
where salary < Mgr_salary I000 and Mgr_name = manager
and DEPT_dno = dno and not< Mgr_name > in
select from BoardMembers tuples < name >.
First, two copies of the stack frame are created, and two call pattems
(OldTup and NewTup):
' EMP( Name, Salary, Dno )
' EMP( NewName, NewSalary, NewDno )
Now. makemodlist creates a raw modi cation list:

[ modif( attr( name, string, Name ), NewName, ModName ),

modif( attr( salary, integer, Salary ). NewSalary, ModSalary ),
modif( attr( dno, integer, Dno ). NewDno, ModDno ) ]
A symbol table is constructed, rst a frame for EMP (note that old attrib-
ute values will be retrieved and used), next for DEPT and EMP (with the
alias name Mgr). During the construction of Modi cations (lines 420-421,
410), the raw list is changed by ndmname (lines 454-465, 434): ModSal-
ary is instantiated as true (line 462) to note that the salary will be modi-
ed. Finally, closemodlist (lines 447-452, 422) binds together variables
that stand for unmodi ed attributes, i.e. Name with NewName and Dno
with NewDno. Also, equalities in lines 398-399 cause two other pairs of
variables to be bound together (see line 318).
A comment on error treatment. Incorrect data do not terminate pro-
cessing. Instead, the procedure ancestor instantiates the variable Err ag
(lines 5, 484, 488)this prevents the command from being executed (lines
13-14), but the analysis continues. The grammar rules synerrc (lines 486-
498) display the troublesome token and the others to its right, and then
succeed leaving the token list intact.
In actual use, Toy-Sequel would probably be found too simple. How-
ever, many extensions are quite straightforward. As an exercise, try to
augment Toy-Sequel with de ned views, e.g.
view EMP1 < name. salary > as
select from EMP tuples < name, salary > where dno = 1.
Another extension: wild card" tuple speci cations, e.g.
select from EMP tuples *.
( i.e. tuples < name, salary, dno > )
 8.2. Prolog and Relational Data Bases 237

select from EMP, DEPT tuples < EMP_*, manager >-

where EMP..dno = DEPT_dno.
(i.e. tuples < name, salary, EMP_dno, manager > ).
And aggregate functions, e.g.
select from EMP average of < salary >.
An example of less straightforward modi cations is query optimisa-
tion. Consider the command:
select from EMP, DEPT tuples < name, salary >
where dno = DEPT_dno and manager = "Jones".
The answer will be generated by the calls
EMP'( Name, Salary, Dno ), ' DEPT( Dno, "Jones" )
which access every EMP tuple, even though only one department is in-
volved. The same set of tuples would be generated by the calls
' DEPT'( Dno, "Jones" ), EMP( Name, Salary, Dno )
but now other departments tuples would never be retrieved. This optimi-
sation could speed things up considerably for Prolog implementations
with clause indexing.

Bratrooaxm-11c NOTES
The bibliography on data bases is enormous. We shall only name a
few positions relevant to our presentation which (of necessity) has only
touched on basic facts. Two widely accepted introductory textbooks on
data bases in general are Ullman (1982) and Date (1982). The relational
model of data was introduced by Codd (1970) and further elaborated by
many, including Codd himself (1979). The most popular relational data
languages are probably Quel, used in the data base system INGRES
(Stonebraker et al. I976), Sequel, created for the system R (Astrahan
I976; Chamberlin et al. 1976) and Query-by-Example (Zloof I977).
In the proceedings of conferences on logic in data bases (Gallaire and
Minker I978, Gallaire et al. I981) there are, in particular, papers on the
role of logic programming in data base theory and applications. The ad-
vantages of Prolog (and logic programming at large) for data bases have
been advocated by quite a few authors, e.g. Kowalski (1978), Gallaire
(I983) and Lloyd (I982). A practical demonstration of Prolog's power is
Chat80 (Warren and Pereira 1982; Warren 1981), a system with a natural
language interface. Queries in English are translated into Prolog calls;
they are similar to those produced by Toy-Sequel, but Chat80 performs
238 8 Two Case Studies

some query optimisation. Several other data base applications with natu-
ral language interface are described in Dahl (1977), Coelho (I982), and
Filgueiras and Pereira (1983).
Another example of data base application of Prolog is an implementa-
tion of Query-by-Example (Neves and Williams 1983; Neves et al. I983).
Chomicki and Grudziriski (1983) describe a system, based on extendible
hashing, that manipulates tuples stored on disk. The system has been
designed to support data-base-oriented implementations of Prolog.
The Toy-Sequel interpreter was rewritten as a sized-down version of
SPOQUEL, a program which we had written with Wlodek Grudziriski in
early I982. It helped us through a dif cult winter.
 LISTING 8.4 ToySeqneI lIIer|I'eter

Toy-Sequel herpreter ------

c) COPYRIGHT 1933 - Feirs I(lrz1ii1.StanlslawSzpaIrowi:z
Instlute ol Irlorrnatlee, Warsaw Unlverely
giliiiel :- wrle( '- Toy-Sequel, IIUW Warszawa 1983 -' ), nl,
109001. 11111011). 1001 0010001001111 C1110. 5111100) l.
tag( docorrmar:I( Cmd, Err ag ) ),
Cmd - eequehtop, 1.
geteomrnartrl Cmd, Err ag) :-
readcmd( CmdStrlno ).
aean( Cmdstrhg, TLBI ), oor1ple( TLlst, Cmd ).
do-eomrnar|:l(Cmd,Errt1ag) :- var( Err ag), I, Cmd.
; :3@@~@maw- do-eomrnan:l(_,_).
15
16 sean( CntlStri1g,TLlst) :-
17 ph1ase( tokenet TLlst ), CrndStrlng ), traeeaean( TLIst ).
18
19 o:>rnp1le(TL1st,Cmd) :-
20 phrase(eo1'r|rnar|1:l(Crr\d),TLIst),l,traeeco|'rpb(C>rr1d).
21 corrp1le(_,error) :- synerr(badeommu1d).
22
23 traeesean(Cmd):- traeesean, I, wrle('--aear'rned'(Cmd)), nl.
24 traeesean(_).
25
26 traoeconpl write( 'con1:>led'( CI'rI:l)), ri.
27 traceeorrpIe(_
28
29 traces-can. traoecorrplle.
30
31 % - - - - --readerandamnner - - - - --
32 %Fleaderstopsonthe rudoto1.rtsldestr1ngs.
3-3 readernr:I(S1rhg) :- rdehsk( Ch), readcmd( Ch, String).
34
35 readernd( []) :- I, rch.
36 readerndt "", [" I Flestl) :-
37 I, rdcht Ch ), readstrt Ch, Fleet, Flest llter ),
38 rdcht Nexteh ), readcmdt Nerrteh, Rest tlter ).
39 readem:I( Ch, [Ch | Ftestl) :- rdeht Nexteh ), readcmd( Nertch, Rest )
40
41 rea&tr( "", [" | Fleet], Rest) :- I.
42 rea&tr( Ch, [Ch | Fleet], FlestAlter) :-
43 rdcht Nextoh ), readstrt Nerrteh, Rest, Rest tlter ).
44
45 % This scanner recognizes names, strings, lriegere, and slngle
48 %eI-raraetere.Strlr1gsareretumedasI1steoIeI-raraeterrr.
47 It. The tokens are: n(Name), e(Strhg), ltlrieger), Aslngleoharaeter.
48 tol-tenet [T | Ts]) -> tokent T), I, sp, tokenst Ts ).
49 Iokemt I] I -> 11-
50
51 Iolten(n(Narne)) ->
52 le ertl-). nam00hntNN). IPMI00tName.l1-IN1~l1)}-
53 I0l100t0t$Irh0)) > 1"'l. 01rh00h0I0t$I1'lI'o)-
54 I0l100tltl"l000r)) >
55 0101118). 0101110). =00lstDD).
58 trmm0lt1.10lDD1). 0l0n0dt$.l.lrIle00r)}-
57 tol-can( Ch) -> [Ch].
58
59 letter( Ch) -> [Ch], { letter( Ch) }.

Z39
LISTING 8.4 (Continued)

i
namechss([Ch|Chs]) -> letter( Ch ), I, nameeha1s(Chs).
namechars(hCh|Cha]) --> dblt( Ch ), I, nameehatst Chs).
)->

%'"'1nastrlngstandsIcraslngIe"'
.

% (reademd treats "..."...' as two adjacent strlngs).

str1ngeha1s( 1" | Chs] ) --> [", '1, I, etr1ngehars( Chs ).
strIngehars( I] ) -> ["'], I.
strlngcharst [Ch | Chs] ) --> [Ch], strIngehars( Chs ).
38 3 -28328 2 8
71 d1glt(Ch) -> [en], {db1t(Ch) 1.
72
73 d1011=110l0D1) --> 01111110). 1.0101101001-
74 dleltstlll --> 11-
75
76
77 01001 '+')
am 1-
-I

78 010111-I-') 1'r1'.rV m-:"+"*-.L-.

II

79
80 slgnedt '+', I, I ).
slgnedt I, Integer) :- Integer Is - I.
-> [' '1, 1, sp. %opt1onalspaeee
-8-8 -> [].

323E882 % These are used Ior attrbutee.

87 qnarne( Qual-Name) -> [n( Oual ), '_', n( Name )], 1.
88 qnarne( Varlmle-Name ) -> [n( Name )]. It. Ie Ills all qualllers
89
90 constant( Int, Integer) -> [1(l1'ill. 1.
ccnstar|t(Str, strhg 1 -> [s(Str)].
all syn-bot table operations - - - - - -
:- op(100,xlx , ':').
It. Glven a relatlon name and an atlas (cl select expression, procedure
% relname), use the reletlon's schema to push a new set ct Items onto
83882833 % sy11'l:1o1 table stadt and Io retum the generdor. The Icrrnat
99 It. ct a schema ls described wlth create (prceechrre newrel).
100 newrelname( Re1Nm,A11as, Generator, O1dST, [Al1as:Fle1ST I OIdST) ) :-
101 'r e I'( Fle1Nm, Generator, He1ST ), I.
102 newre1name( Fte1Nm, _, Ialt, OldST, OIdST) :- synerr(nore1name(FlelNn-1))
103
104 It. Glven a quali ed name, return Its associated varlable and type.
105 tlndattrt Q-Nm, Var, Type, [Q : FleIST | S1) ) :-
106 rnerrb-er( attr( Nm, Type, Var ), Rel-ST ), I.
107 f1ndattr( QNm, Var, Type, [_|S1')) :- I, ndattr(QNm, Var, Type, ST).
108 tlndattrt QNm, __, _, |]) :- synerrt ncattri:1ute( 0Nm ) ).
109
110 % - - - - - - command corrpller - - - - - -
111 %Seethevar1ouscom1nar1dsIorexarrpIesoluse.Cornmar1d
112 % Interpretation routines Bted alonwlde corrrnand gramnar processors.
113 cornrnand( Cmd ) --> create( Cmd ).
114 ccrnrnand( Cmd) -> cancel( Cmd).
115 command( C -> se1eet( Cmd ).
116 command( -> relat1orn( Cmd ).
117 corn-narr.I( relation( Cmd ).
118 corn-narr.I( -> hsert( Cmd ).
119 corn-narr.I( OOOO E5353
deIete( Cmd ).
240
LISTING 8.4 (Continued)

120 command(C -> updaIe( Cmd).

121 ccmmand(C stop(Cmd).
122 ccnI'nan:l(C -> cirrrp( Cmd).
123
124
ccrnrnand(C E333 -> load(Cmd).
125 %---createanewrelatlcn-~
126 % Eg: create EMP < strhg name, lrneger salary, Irieger dno >.
127 )1. Eg: create DEPT<Integer dnc,str1ng manager>.
128 It. Eg: create Boardlvlerrberscstrlng narne, strhg pos, Integer senlority>
120 % Note that lcwertupper case matters. Keywords trust be
130 % In lower ease,othe1wIse use any ccnventlon you lire.
131 create( newreI( RelNarne, [V I Vs], [attr(Nm, Type, V) | As] ) ) -->
132 [n( create ), n( Fte1Name )],
133 ['<], typnam( Type, Nm ), typnarnst Vs, As ), [>'].
134
135 typnarrIstlVlV0l-l00rtN"I.Tv00.V)l00l) ->
136 [','], I, typnam(Type, Nm), typnan's(Vs,As).
:3; Iv1=>narmtl1.Il) --> 1]-
139 typnam(strhg, Nm) -> [n( strlng), n( Nm )], I.
140 typnarn( Integer, Nm) --> [n( heger ), n( Nm )], 1.
141 typnam( nctype, Nm) -> synerrct Iypeexpected ).
142
143 It. A schema stores a pattem lor Involthg the relatIon's tuples (the
144 )t.generdor).ar1dallstolsyn'l:oltd:leent11es1i'lrhgattrbute
145 %na1'r|esanrltypeswlthvar1d:leshthegeneruor.
146 newreI( ReINarne, Vars, Rel-ST) :-
147 not 'r e I'( FlelName, _, _ ), 1.
143 rnlrgen( Fle1Name, Vars. Generator ),
149 assen( 'r e I'( RelName, Generator, Flel-ST ) ).
150 newreI( FlelName, _, _) :- narnerr( ci.preharne( FIe1Narne ) ).
151
152 %Addab1ankIcr(rudI111e1'lary) security.
153 11-Itgen( RelName, Vars, Generator) :-
154 pname( ReIName , Chars ), pname( Ftel m, [' ' | Chars] ).
155 Generator -.. [HelNm|Vars].
156
157 %---canceIarelat1cn---
156 It. Eg: cancel EMP.
159 caneel( canceI( Relllame) ) --> [n( cancel), n( ReIName )].
160
161 caneeI( Fte1Name ) :- retract( r e I'( Fte1Name, Generator, _ ) ), I,
162 retract( Generator), lal.
163 caneel( ReIName) :- namerrt urltncwnt FtelName) ).
164
165 % - - - queries - - -
166 )t.Ll1sttheeetgenerdedbyaseIeeIexpreaslon.
167 se1ect( (Generators, Flter, wrleIuple( Tup ), tall ->
166 se1ectexp( set( Generators, Filer, Tup, _ ), =11-an .,_'\u-I.
169
170 w1lletuple(|]) :- 1, nl.
171 wr110100IetIVIV0I)=-wrltV).w1tt0t' ').wrII0t0p101Vs)-
172 wr1t1XlY1) =- I. wr11e1exttIXIY1)-
17:3 wrI(X) :- write()t).
174
175 % Llst al relatlorn. Eg: relations.
1 76relatlore( (re I'( FtelNm, _, _), wrle( FlelNm), nI,Ia1l)) ->
177 [n( relatlorn )].
176 % Llst the attrtbutesol a relatlon. Eg: relation EMP.
179 relation( relation( Name )) -> [n( relation ), n( Name )].
Z41
LISTING 8.4 (Continued)

160
161 relation( Ftelll) :- re I'( FtelN,_, Altrs ), I, l1stattrs( Altrs ).
162 relatlont RelN) :- wrIe( Relll ), wrIe(' B not a relatlonr ), nl.
163
184 ll6lall1s( I] ) :- 1.
165 l1slatt1s(|attr( Narne,Type,_)|Attrs]) :-
166 wrl1e( Type ), wrlte(' '), wrle( Name ), nl,
167 1lstat:trs( Attrs ).
166
169 ll.---selectexpresslon-H
190 % Eg: select Irom EMP, Mgr-EMP, DEPT tuples < name, dno >
191 % where salary > l1lgr_aalary65!100
192 % and Mgr_name - manager and DEPT_dno - EMP_dno
193 )6 and < manager > h (<'SmlIh'>, <'Jones">, <'Brown'>) .
194 %(legelname=sanddepartrnerl nrrntlersollltosesrrbordhatesclsmith.
195 %JonesorBrownwhoeamnnrethan65%oltl1elrmar1.ager'ssaIary).
196
197 ll. Generators plclr up tuples Irom named relatlom, Flters pass only
196 lbtuplesllthgtltewhere-darse,Tq:>leBhstar|tlatedlothepassed
199 % tuples (one by one), Types B the tuple's pattem wlh types Instead
200 %ol attrbutes (usedlortypechecltlng).
201 % The exarrple oonplles to :
202 % set((' EMP'(Name. Salary, Dno), ' EMP'(MgrName, MgrSala1y, MgrDno)
203 ll. ' DEPT(Dr1c, l1lgrName) ),
204 ll. (Salary > MgrSalary65I100, tme, true.
205 % (rneni:ler(MgrNarne, <'Srnith', '..lones', 'Brown">), true)),
206 % [Name,Dr1o], [strlng,lnteger] )
207
206 selectexpt set( Generators, Fllter, Tuple, Types ), InitST ) -->
209 [n( select ), n(l1'oI1'|)], relnames( Generators, lnltST, ST),
210 [n( tuples )]. tuplepdtern( Tuple, Types, ST ),
211 whereelarset Fllter, ST ).
212
213 ll. One or more relation names, possbly 'allased'. Syntlol table trag-
214 It. meris are stadted In reverse order, so attrbute search orderwlll
215 % be that or the Irom-Bl (uslng-H lot update) relatlom.
216 reInarnes( ( Gen, Gens ), OldST, NewST ) -->
217 relname( Name, Alas ), |','], I, relnames( Gene, OldST, TempST ),
216 { newrelnarnet Name, Alas, Gen, TernpST, NewST ) }.
219 relnamest Gen, OldST, NewST ) -> relname( Name, Alas ).
220 { newrelnamet Name, Alias, Gen, OldST, NewST) }.
221
222 relnarnet Name,Allas)--> [n( Alias). '-'. I'll Name)], 1.
223 relnamel Name, Name) --> [n( Name )].
224
225 ll. luplepdtem B also Invoked by lnexp.
226 Ivnlemlteml IA I A01. 1'11 Tel. ST I -->
227 [<'], atlrpatt( A, T, ST ), attrpaIts( As, Ts, ST), ['>'].
220
229 e rpattst 141401.111 T -->
zao [','], I, attrpatt(A, .-1.3rpm -I-I atlrpatts( As, Ts, ST ).
201 01110011011]-1]-_) > 1]-
232
233 atIrpatt(Attrl:ute,Type,_) --> corBtar1t(Altrbute,Type), 1.
234 altrpatt(A,T,ST) -> qname(ON), {llnda1tr(ON,A,T,ST)).
235
236 whereelarset Fller, ST) -> [n(where)], I, boolexpt Fllter, ST).
237 whereelarse(true,_) --> [].
236
239 ll.---Booleanexpresslons-~
242
LISTING 8.4 (Continued)

240 )0 Eg: salary > Mgr_aalary ' 851100

241 % or <name> h select from Boarttlllllerrbers tuples <narne>
242 % Note that embedded select expressions do not rnodlly the symbol table
243 %whose errtenslone arevlsbleonlyhtheneaed constructs.
244 boolexpt E, ST) -> bterm( T, ST), rboole:p(T, E, ST).
245
246 rboolexpl -> [n(or)], I, boolexpt , ST).
247
246
rboolexp( I11! 1'11? -2F V3 Iv

249 bterm( T, ST) -> bIactor( F, ST ). Ibterrnl F, T, ST).

250
251 rbterm( -> [n( and )], I, bterm( Fl, ST).
252 Ibterrnl fl!- ."l'l= .
253
254 blaclort not F, - T).
255 blaclor( E, S - 2. -tn
256
257
258
blaclor(
blar:tor(
E, S
E, S -I-I-I-
--
-:rs. :." '
gssv
:-!-|-Cg
=" xi";-R-"--;In!"333%;
259 %---setmeni:lershb---
260 %Eg: <dno,name>ln<1,'Jones' >,<2,'SrnIl'l'>
261 )1. Eg: < name > In (select Irom Boardlrlenbers tuples <name>)
262 lnexp( ( Generator, Fller ), ST) ->
263 tuplepattem( Patt, Type, ST), [n( In )],
264 setexpl eet( Generator, Fler, Tuple, Types ), ST ),
265 matcl'patterns( Palt, Type, Tuple, Types ).
266
267 %matclpatten'BBarule,aothatsynenccanshowconten.
266 rnatchpaltemsl Patt, Types, Patt, Types) -> I.
269 rnatchpatlernsl P1, T1, P2, T2 ) -->
270 synenc( badlnexppatternl T1, P1, T2, P2 ) ).
271
272 %---setexpresslom-~
273 ll. A sequence ol tuples or a select expression, posslbly In parerlheses.
274 ')t.ThegeneraIorlorasequeneeoIluplesB acallon merrberwiththe
275 ltseoonrlparameterhnarltlatedtoallstolthesettrples.
276 setexpl S61, ST ) -> [('], I, eet6lt|:)( Set, ST ), [)'].
277 setexpl Set, ST) -> seleetexpl Set, ST), I.
276 setexp( set( menber( Patt, lTup|Tups] ). tme, Patt, Types ), ST) -->
279 tvplet Tun. T111700 ). Iunlesl Tum. T7004 1.
260 { rrirpattem( Types, Part) ), I.
261 setexpl sell lal, lal, l], I] ), _) -> synerrct badsetexpr ).
262
263 w1>Ie=tlTu1>|T01=0l.Tr1m) --> I'.'l. I. 10171011017. 1017171100).
264 14000700117044. 1017171700) 1. 101000110170. Tynes)-
265 100100111-_) -> 11-
266
207 100101 141401-lTlT01) -->
266 ['<'], constant( A, T), consta11ts(As,Ts), [>'], I.
209 I01>let11.l]) -> l'<'l. 0700110100111-1010). Ital)-
290
291 00rBlnrl0tlAIA0l-l1'lT01)->
292 [','], 1, conetari(A,T), oonstarls( As,Ts).
293 00retamstIl-l1)->l1-
294
295 01100107001 Type. T1100) =- I-
296 checlrtypel T1,T2) :- syr1err(lnror1sBter1t(T1,T2)).
297
296 %Pattlneet(_,_, Patl,_)lsal1stolnlreshvarlmles
299 %(nBthelengthoIlupleshthIsset).
243
LISTING 8.4 (Continued)

300 n'Irpattem( l],|]) :- I.

301 rrIrpattem([_|Types],[V|Vs]) :- rriq:1attem( Types, Vs ).
302
303 %---relatlonalexpresslons-~
304 relexp( E, ST ) ->
305 sln'plexp( Lel1E, LeltType, ST), relop( Op ), I,
306 6ln'plexp( Fllgl1tE, Ftlgl-1tType, ST ),
307 { consrel( LeltE, LeltType, Op, HbhtE, FlbhtType, E ) ).
308
309 I'6|Op( 'I<' ) > ['u', '<']. I'6bp( '02-) V 1-: I a
310 relop( '--') -> [<', >1. relopl '<') - I

311 reIop('>-') --> ['>','-'1. relop( '>') - VV

-7* ,_!'.-.-1
312
313 consrell L, Type, Op, R, Type, E ) :- consrell L, Op, Fl, Type, E ), I.
314 corsrell L, LType, Op, Fl, RType, tall):-
315 E -.. [Op, L, R], synerrcl typeconlllctl LType, RType, E ) ).
316
317 ll. The llrst clause does corrpile-tlme equally.
318 consrel( Arg, '-:-', Arg, _. true ).
319 consrel( L, '-:-', Ft, strlng, tall ).
320 corBrel( L, '--', Ft, strlng, not L - Fl ).
321 consrel( L, Op, R, heger, E) :- E -.. IOP. L, Fl].
322 cornrel( L, '<', Fl, strlng, lstr( ).
323 consrel( L, '-<', Ft, strlng, ( ; L - Fl ) ).

326
5-=~e
324 cornrel( L, '>', Ft, string, let .113!
325 consrell L, '>-', R, strlng.( l -' ; Fl - L) ).
.mr-rir-:1)

327 ll. Compare strlngs lexlcogrq:>hleally.

326 W11]-Ll_1) 1- 1-
329 lstr([Ch1|_],[Ch2|_]) :- Ch1@<Ch2, I.
330 lstr( [Ch1 |Chs1],[Ch1 |Chs2]) :- lstr( Chs1,Chs2).
331
332 %---sln'pleexpresslons---
333 s1nplexp( E. strlng, ST) -> strln@xp( E, ST), 1.
334 sll1plexp( E, Integer, ST ) --> ar1tl'|exp( E, ST).
335
336 stringexp(Slr, _) -->[s( Slr)], I.
337 % Type checklng delayed to avold error messages - rnlght be Integer.
338 slrlrlgeIp( Var, ST) -->
339 qname( ON ), { flndattrl ON, Var, Type, ST), Type - strhg }.
340
341 arithexp( E, ST ) --> 9t6I'I1'l( T, ST ), rarlthexp( T, E, ST ).
342
343 rarlthexpt L, E, ST) ->
344 ['+'], I, atennt T, ST), rarltl1exp( L+T, E, ST).
345 rarlthexp( L, E, ST) >
346 |'-'), I, atom( T, ST ), rarltl'lexp( L-T, E, ST).
347 rarlthexp(E, E,__) --> [].
346
349 atem1( T. ST) --> alaclor( F, ST), ratem1( F, T, ST ).
350
351 raterm( L, T, ST) -->
352 ["'], I, aiactort F, ST ), raterm( LF, T, ST).
353 raterm( L, T, ST) ->
354 ['1'], I, alactorl F, ST), ralem'|( UF, T, ST).
355 ratennl T, T, _ ) -> [].
356
357 alactort E, ST) ifxp( E, ST ), [')].
358 alactort Int, _ ) it 5; "5
359 alactor( Var, ST) -v=V
24-4
LISTING 8.4 (Continued)

360 qname( ON ), { l1ndaltr( ON, Var, Type, ST ), Typg - Integer }, I.

361 alac-tor(0,_) -> qname(ON), I,synerrc(noIlr|teger(ON)).
362 a1actor( 0, _ ) -> synerrc( nohegertactor ).
363
364 % - - - Insert - - -
365 % Eg: lrio EMP hsert e'Jones',1000,1>, <'Smlth',1200,2>.
366 ll. Eg: Into EMP lnsen select Irom DEPT tuples <1-nanager, 1050, dno>.
367 hsenl (Generators, Fllter, assertz( NewTuple ), lal ) ) ->
368 [n( Into ), n( Relttlame )].
389 { r e I'( RelName, _, FtelST) ), I, [n( Insert )],
370 setexpl set( Generators, Flter, Ttple, Types ), I] ),
371 { checlr.types( Types, RelST ),
372 n'ltgen( Fteltllame, Ttple, NewTuple) ).
373 lnsert( lall) --> [n( Into ), n( Ftelllm )],
374 synerrct norelnamel RelNm ) ).
375
376 01100107170011]-1]) 1- |-
377 checlttypes( [T|Ts], [attr( _, T, _ )|As]) :- 1, cheeltlypes( Ts, As ).
378 checlrtypes( Types, Altrs) :- synerr( badsettype( Types, Attrs ) ).
379
380 % - - - delete - - -
381 % Eg: Irom EMP delete all tuples.
382 % Eg: Irom EMP delete tuples where salary < 1000 and
383 ll. <dno>hselectlromDEPT triples <dno>
384 )1. where manager - 'SmIh' .
385 %(le llrealsuborchatesol Smlhwhoeamlessthana 1000)
386 delete( ( RelGen, FtelFler, retract( FtelGen ), lal ) ) -->
387 [n( Irom ), n( ReINrn )],
388 { newrelnarne( Flelllm, FlelNm, FtelGen, |], ST) },
389 [n( delete )], deIIllter( FlelFler, ST ).
390
391 delIllter( true, _) --> [n( all), n( 1uples)), I.
392 deI|'ller( RelFIler, ST) ->
393 [n( tuples ), n( where )], boolexpl FtelFilter, ST ).
394
395 % - - - update - - -
ll. Eg: update EMP uslng DEPT, Mgr-EMP
so that salary - salary + (Mgr_salary - sala|y)l5
where salary < Mgr_sala1y -1000 anrl Mgr_name - manager
and DEPT_dno - dno
and not <Mgr_name> In
select Irom Board errbete tuples <name>.
le to all errployees who earn over a 1000 less than thelr manager
glve a ralse equal to 20% ol the dillerence, provlded the manager
does not at on the board)
l'hls B oormlled to :
' EMP'( Name, Sal, Dno ). % Olt:ITup
DEPT'(Dno,Manager), ' EMP'(Manager,ltIgrSal,MgrDno)), ll. UseGens
$ $ $ $ s0s0s! Sal < I.lgrSal - 1000. true, true.
% not (' Boardlllen'be1s'(Manager,_,_),1rue) ). % Fllter
% NewSalBSal+(MgrSal-Sal)l5, %
% retract( EMP'(Name,Sal,Dno)).assett(' EMP'(Name,NewSal,Dno)),lail
412 update( ( Olt:ITup, UseGerB, Fller, Moc lleatlore,
413 retract( Olt:lTup ), assen( NewTup ), tall ) ) ->
414 [n( update ), n( RelNm )].
415 { 'r e I'( FlelNm, OldT1.1p. OldST ),
416 r e I'( Flelllm, NewTup, NewST ), I.
417 rnalrerncrIst( OldST, NewST, MLBI ) }.
418 uslngelmset UseGerB, UseST ), { ST - [F1eINm : OldST | UseSl') }.
419 I01 00 ). I'll 11101 )].
245
LISTING 8.4 (Continued)

420 moc ller( Modltlmtlon, l1lLBt, ST),

421 rnociflersl Mocilleatlon, l1lodlllcat1ons,MLlst,ST),
422 {closemocl1st( Must ) ), whereclmsel Fller, ST).
4:3 update(lall) -> [n( update )], synerrc( noupdaledrelmlon).
425 uslngela.rse( Gerts, ST) --> [n(uslng)], reInames( Gerts, []. ST).
426 uslngclt|.rse(true, ST) -> [].
427
426 rnodIlers( M, (I.I,Ms), IvlLBt. ST) -->
429 ['.'], l, modlllerl MM,MLBt,S1'),
430 rnocitlersl MM,l1ls, I.ILlst, ST).
431 rnodllletsl l1l,M,_,_) -> [].
432
433 rnodlller( AttrVar ls Expr, Must, ST ) -->
434 [n( Nm )], {llndmnarne( Nm, Altrvar, Type, MLlst) ),
435 ['-'), sln'plexp( Expr, EType, ST),
435 1 "TYP01 T700. ETYP0. N111) )-
437
436 % A 'rnorIsl' lists updated relatlon's attrbutes together wlth new
439 It. varlables lonnlng new tq:>le's altrbutes and Mod variables which are
440 %usedIoIlaganattrI:1ute's|'not:llllcatlonwhenltlsdetectedonthe
441 % lelt hanrl slde ol an equalty In so that-list.
442 malt_emodist( [Old | Olds], [attr( _, __, Newv) | NewVs].
443 [mocil(Old,NewV,lvlod)|lllods]) :-
444 I, maltemodist( Olds, NewVs, Mods ).
g maltemodistt [], []. [] ).
447 ll. Blnd old and new variables In "modllst' entrles wlth clear flag.
446 closemodlBt([l1llod| Modal) :-
449 closemodt Mod ), I, closemodlst( Mods ).
450 closemo tl I] ).
451 closen'od(modil(atlr(_,_,OldV),OldV,lvlod)) :- var(Mod).
452 closemod( _ ).
453
454 % Flag an updated attrbute In '|'nodlBt'.
455 flndn'name( Nm, NewV, T, MLBI) :-
456 meni:ler( rnod1l(attr( Nm, T, _ ), NewV, Mod ), MLlst ), l,
457 mmod( Mod, Nm ).
456 llndn'narne( Nm, _, _, _) :- eynerrl rlotlrupdatedrell Nm ) ).
459
460 % ll no errors, the Ilrst clause ol mmod lals, the second blrnls. _
461 rnmoo(Mod,Nm) :- notvar(ltlod), l, synerr(updaIeclwlee(Nm)).
462 mmod(tn.1e,_).
463
464 rrIype(Typ6.Type,_) :- I.
465 rrIype(T1,T2,Nm) :- synerr( typeconlllct(T1,Nm,T2)).
466
467 ')t.---controlcommands---
466 stop( sequelstop) --> [n(s1op)].
469
470 sequeBtop. )0 do nothlng (cl the maln procechre)
471
472 load( oorBult( FlleName)) ->[n(load).n(lrom),rt(FlleName)].
473
cIrrrp( t:lrrrp( FlleName)) --> [n( durrp ), n( to ), n( FlleName)].
476 cirmp( FlleName) :- tell FlleName).
477 'rel'( Nm, Gen, ST), welause('rel'( Nm, Gen,ST)),
476 Gen, wclause( Gen ), tall.
479 t:irrrp(_) :- wrlte('enrl.'), nl, told.
246
us'm~1c 11.4 (C0110-I111)

ct) ;- wrtteqt ct 1, write( ).0|-

iii ---errorharItlr\r|1'1i)0101-"""'"1'"'1 ' ' ' ' H

lrlol :- synrrlesl 1110 l. '1'51'19'1mmd1 -' worn

3 1101lrIo) -> 10711110011010). WIl0('Context:')).

coraext. It. wll lal everlualyl
i .7. ..> {nl, a1\cest0r10010""'"d1-'"'111'
:_ N w|iQ('--- QITDTI '), TII.

95 6595 6053
492 context -> lT01101'l- 1'11""1Tk 1 1' ends
493
494 wtoken(T) =- wt1T.F1001T). wrltet FlealT), write("). |-
495 v1rt(n(N8rI16).1131"01-
498 wq I( Integer), |rI090r I-
497 wll 01 $11100 I. $11100 )-
498 wt( Char, Char).
499
500 narnerr(lrIo) ;- nl, wr1tet'E"0'"-'1- "'-
501 M, "1, ,_ .1, tegte11toeeornneror_._))-

241
9 PROLOG DlALECIS'

9.1. PROLOG I

The idea of logic programming emerged in Marseilles in the rst half

of 1972 while Robert Kowalski was visiting the arti cial intelligence team
founded by Alain Colmerauer at the University of Marseilles. Colmerauer
with his team prepared the design speci cation of the programming lan-
guage Prolog (Colmerauer et al. 1972). The language resembled a theorem
prover rather closely, but it already possessed the essential properties of
contemporary Prolog, and even some features reintroduced quite re-
cently, e.g. delaying calls till appropriate instantiation of their arguments.
Almost at the same time Kowalski advocated predicate calculus as a
formalism for expressing algorithms without commitment to a speci c
strategy of their execution; the short note (1972) was later expanded to a
larger paper (1974). Hence the two pioneers of logic programming took
from the outset different approaches to the problem of changing the idea
into reality. Although developed in close interaction, these different atti-
tudes still manifest themselves in logic programming research.
The language described in Colmerauer et al. (1972) was implemented
in Algol W on IBM 360/67 by Philippe Roussel and used at once in several
applications (Colmerauer et al. 1973, Pasero I973, Kanoui I973, Joubert
I974; Bergman and Kanoui I973, 1975. Battani and Mloni 1975. Guizol
1975). It was quickly replaced by an improved version. coded partly in
Fortran by Battani and Mloni (I973) and partly in Prolog by Colmerauer
and Roussel. This was the rst version of Prolog used outside Marseilles.
Although not christened so by its authors, it deserves the name of Prolog
I. especially as its commonly used name Marseille Prolog" has become
ambiguous.
' This chapter was contributed by Janusz S. Bien, Institute of Informatics, Warsaw
University, Warsaw. Poland.
249
250 9 Prolog Dialects

The original reference to Prolog I is Roussel (1975); some historical

infonnation can be found in Battani and Mloni (1973) and l(Iu2niak
(1984). The syntax of the language is illustrated below by the sample
clauses:
+APPEND( NIL, -I-X, -I-X ).
+APPEND( -1-X.-1-Y, -I-Z, -1-X.-1-V ) APPEND( -1-Y, -1-Z, -1-V ).
This should be preceded by the decimation of the in x dot:
AJOP( 1, "X'( XX ) )!
And a sample directive (SORT means write):
APPEND( A.B.NIL, C.NIL, -I-X ) SORT( -I-X ) LIGNE!
Positive literals (see Chapter 2) are preceded with + , negative with ; the
notation allows representation of non-Hom clauses. This was a natural
requirement, because the early versions of Prolog were intended to imple-
ment a general theorem-proving method known as linear resolution with
selection function (Kowalski and Kuehner 1971). Program clauses were
distinguished from directives by a different tenninator. The syntax sur-
vived its original motivation and is still used in some versions of the
language (Kluiniak and Szpakowicz 1983).

9.2. PROLOG II

After Prolog I was released, Colmerauers team experimented with

various mutations of the language; some of them have been described in
Guizol and Mloni (I976), Colmerauer er al. (1979) and Kanoui and Van
Caneghem (1980). Finally it was announced that the goal of creating the
ultimate Prolog was achieved (Colmerauer et al. 1981). The new lan-
guage was called Prolog II by its authors (Colmerauer 1982, Van
Caneghem I982, Kanoui I982).
The most important innovation of Prolog II is the treatment of cyclic
data structures (see Section 1.2.3). They are simply valid representations
of in nite trees, which can be manipulated in a similar way to other terms
(Colmerauer 1979, 1982). However, the same in nite tree can be repre-
sented by different data stnrctures; to let them be matched correctly it
appeared necessary to treat functors in the sameway as arguments (Fil-
gueiras I982). As a result, a functor can be a variable or a compound term.
In Prolog II, the standard form of terms is considered just a shorthand
notation for a more general form called tuple. For example, ff(x) stands
 9.2. Prolog ll 251

for <ff, x>, while <x, y> and <<ff(x)>, y> are also legal terms (single-
letter names denote variables, ff is a constant). Instead of unifying two
tuples, Prolog II constructs a system of equations. For example, matching
<x> with <ff(x)> corresponds to solving in x the equation
x=ff(x).
The solution of this equation is the in nite tree ff(ff(ff(...))), which is
represented by an appropriate cyclic data structure.
The behavior of Prolog programs is described as incremental solving
of the system of equations introduced by program clauses, which can also
be seen as rewriting rules. The execution of a call is viewed as the opera-
tion of erasing it by applying the rules and solving the appropriate equa-
tions. This viewpoint manifests itself in the syntax of clauses by an arrow
leading from the head to the (possibly empty) body, e.g.

0000001 nil. x. X )>;

0000001 X-Y. z. x.v )> 0000001 Y. z. v ) ;
This approach makes it possible to describe the principles of Prolog II in a
compact and self-contained way, relieved from the references to theorem-
proving techniques and relying only on the most fundamental and intui-
tive notions of logic (Colmerauer I983).
Prolog II offers a simple yet powerful coroutining mechanism (see
Chapter 2). A call may require its parameter to be instantiated. Given a
variable, it waits for it to become bound. This is achieved by the b_uilt-in
procedure geler (freeze), whose rst argument is a trigger (usually a
variable to be bound) and the second a call (to be delayed). If the trig-
ger is already bound, geler simply executes the call.
Another coroutining primitive is drf, which succeeds if its parameters
are not perfectly equal (see the built-in procedure = =, Section 5.6).
If during the execution of dif (down the tenns structure") a variable is
encountered, dif waits until it becomes bound and only then resumes the
comparison.
The coroutining mechanism will be illustrated by two examples
adapted from Colmerauer et al. (I983). The rst example is a procedure
that takes two trees represented as deeply nested dotted list structures. It
succeeds when both structures can be attened to the same linear list. A
built-in procedure ident is used to check whether the argument is a
constant.
sameleaves( a, b ) > leaves( a, u ) leaves( b, u ) Iist( u );
leaves( a, u )> geler( u, leavesl( a, u ) );
252 9 Prolog Dialects

leavesl( a, a.nil ) > ident( a );

leavesl( a.l, a.u )> ident( a) leaves( I, u );
leavesl( ( a.b ).l. u )> leavesl( a.b.l, u );
Iist( nil )>;
Iist( a.u )> Iist( u );
The second example is a procedure that generates a list of digits 1, 2, 3
such that all three elements are different.
perm( x.y.z.nil )> alldifferent( x.y.z.nil )
alldigits( x.y.z.nil );
alldigits( nil ) >;
alldigits( x.l )> digit( x ) alldigits( I );
digit( I ) >; digit( 2 )>; digit( 3 )>;
alldifferent( nil ) >;
alldifferent( x.l )> outside( x, I ) alldifferent( I );
outside( x, nil );
outside( x, y.l )> dif( x. y ) outside( x, I );
Prolog II supports a kind of modularisation implemented by so-called
worlds, each with a unique name. Worlds are organised into a tree
structure. The root is the world origine, which has two subworlds
ordinaire and ?????. Subworlds of ordinaire (which is the de-
fault) can be created by the user who can walk up and down the tree, and
also create and discard worlds. ????? contains the Prolog II supervisor
and cannot be used as the current world. Every procedure name is associ-
ated with the world in which it was rst mentioned (declared). It is
accessible in this world and its descendants but not in its siblings. More-
over, a name N and the same name N declared in a superworld later on
refer to different procedures.
Clauses are available for all manipulations (including initial de nition)
only in the world where the procedure name has been declared and in its
direct subworlds. For example, standard procedure names are introduced
in originewith clauses de ned in ?????and used in ordi-
naire. So. the user cannot change a standard procedure de nition.
Clause indexing is provided. or rather tuple indexing. The leftmost name
in the tuple is used as a key.
The purpose of Toys tag, tagexit etc. (see Section 5.12) is served in
Prolog ll by a pair of built-in procedures bloc. n-bloc. In the call bloc-(l,
t), t is the call to be executed in the block. When. during the execution of
t, a call of the form n-bloc-(ll) is encountered, the most recent call on
 9.3. Micro-Prolog and MPROLOG 253

bloc(l, t) with l uni able with ll is sought. If none is found. an error

condition is raised, else this call on bloc succeeds deterministically. This
feature is used for error handling and for exiting loops.
We have presented here only some of the Prolog ll features, and the
interested reader is referred to Colmerauer er al. (1983). It is interesting
that the pilot implementation of Prolog ll which is described here was
done on an Apple ll microcomputer using software paging on oppy
disks.

9.3. MICRO-PROLOG AND MPROLOG

micro-Prolog is the dialect used by l(owalskis team at Imperial Col-

lege of Science and Technology in London. micro-Prolog was developed
and implemented by McCabe (I981); his main goal was to install Prolog on
a cheap 8-bit microcomputer. micro-Prolog uses lists to represent terms.
calls and clauses. e.g.
((APP00d()xx))
((APP00d(X|X)Y(X|Z))(APP001XYZ))
The list notation has several advantages: predicates and functors need not
have a xed number of arguments (i.e. the whole argument list can be
bound to a single variable); their names can be arbitrary list structures (for
technical reasons this is not allowed for predicates). micro-Prolog also
supports a simple form of modularity: modules are created dynamically
and the accessibility of names is detennined by export/import lists.
A typical user is not expected to interact directly with micro-Prolog,
because a special front-end called Simple is provided to conceal the low-
level language features. Here is the append example in the Simple syntax:
A000nd(()xx)
Append((x|X)Y(x|Z))ifAppend(XYZ)
Simple was used as a computer language for children; other interesting
applications are expert systems. The language is subject to various exper-
iments and extensions. such as an explanation facility or an original form
of input/output operations called query-the-user (Sergot I982); Simple
and other extensions are all written in micro-Prolog. Both micro-Prolog
and some of its applications are extensively documented in Ennals (I983)
and Clark and McCabe (I984).
MPROLOG was developed at the Institute for Co-ordination of Com-
puter Techniques. in Budapest (Bendl et al. 1980, SzKl 1982), using the
programming language CDL2 (Koster 1974). MPROLOG is an upward-
254 9 Prolog Dialects

compatible extension of Prolog-10, intended for creating production soft-

ware for mainframe computers. The crucial extension consists in intro-
ducing a form of modularity, based on the ideas of Szeredi (1982) and
similar in spirit to that found in many other languages. The modules are
syntactic units and contain explicit export/import lists detennining the
visibility (i.e. the accessibility) of names; a visible name can serve as a
functor or as a predicate name. When a program is entered, only the main
module is loaded and executed; other modules must be loaded explicitly
by calling appropriate built-in procedures.
MPROLOG is a large system which includes several components.
The pretranslator produces an intemal form of a program module; the
consolidator links the modules into a program; the interpreter executes it.
The program development support system (PDSS) provides a dedicated
editor and debugging aids. A compiler and an optimizer are under devel-
opment. The language offers a multitude of built-in procedures (probably
more than any other Prolog system) and interfaces to user-supplied proce-
dures written in CDL2 and Fortran.
APPENDICES

Appendix A.l Itemel File ...................................................... ..

Appendix A.2 Bootstrapper ................................................ ..
Appendix A.3 User Interface and Utilitim ................................. ..
Appendix A.4 Three Useful Programs
A Simple Editor ............................................. ..
A Primitive Tracing Tool .................................... ..
A Program Structure Analyser with Analyser Analysed

255
APPENDIX A.l
Kernel Flle

1 % KEFINEL llle
2 % standard atoms
3 ';'/2 ','/2 call/1 tag/1
U/0 '.'/2 'error'l1 'user'I0
% atoms ldentilylng system routlnes (keep tall rst and tn.re' last)
'lal|'/0 tag/1 call/1 I70
'tagcut'l1 'Iaglall'/1 tagexlt/1 ancestor/1
halt/1 'status'l0
310 :)-ro>r.nt:> display/1 rch'I0 'lastch'/1 'sk'pbl'I0 'wch'/1
11 'echo'/0 'noecho'l0
12 see/1 'seelng'/1 'seen'I0 tell/1 telling/1 'Iold'I0
13 'orclchr'/2 'sum'l3 'prod'l4 less/2 '@<'/2
14 'smalletter'/1 'blgIeIter'l1 'letter'l1 'dIgll'l1 'alphanum'l1
15 bracket/1 'solochar'/1 symch/1
16 'eqvar'/2 var/1
17 atom/1 Integer/1 nonvarint/1
18 'lunctor'I3 arg'I3 pname/2 pnamei/2
19 '$procl1 '$proclknlt'l0 $proclnltl0
20 clause/5 retract'l3 abolish/2 'assert'l3 'redellne'l0
21 'predellned'l2 protectl0
22 'nonexlstent'I0 'nononexlstent'/0
23 'debugl0 nodebug/0
24 'tn.re'I0
25
26 % kemel lbrary
27 error(:0) : nl . display(-1-r-+ System call error: ') . dlsplay(:0) .
28 nl . lall . U
29 :ordchr(10, :0) . assert(lseoh(:0). U, 0) .
30 assert(nl, wch(:0).U, 0) . U I
31 '-'(:0, :0) :U
32 ','(:0, :1) :call(:0) . calI(:1) . U
33 ';'(:0, _) :call(:0) . U
34 ';'(_, :0) : call(:0) . U
35 not(:0) :call(:0) . '1' . lall . U
36 not(_) :U
37 check(:0) : not(not(:0)) . U
38 side_ettects(:0) : not(not(:0)) . U
39
40 or1ce(:0) :call(:0) . '1' . U
41
42 '@-<'(:0, :1) : '@<'(:1, :0) . '1 . lall . U
43 @-<'(_, _) :U
4-4 '@>'(:0, :1) : '@<'(:1, :0) . U
45 '@>-'(:0, :1) : '@-<'(:1, :0) . U
46
47 % - - - - - - baslc lnp1.rl procedures - - - - - -
48 rdchsk(:0) : rch . skbbl . lastchm) . U
49 r1:lch(:0) : rch . laslch(:1) . sch(:1, :0) . U
50 % convert nonprhtable characters to blanks
51 sch(:0, :0) : '@<'(' ', :0) . '1' . U
52 sch(:0, ' ') : U
53
756
APPENDIX A.l (Continued)

54 repeat : U
55 repeat : repeat . U
56 merr|ber(:0, :0.:1) : U
57 merr|ber(:0, _.:1) : merrber(:0, :1) . U
58
59 proo(:0) : '$proclnIt' . $pr(:0) . U
60 '$pr(:0) : '$procii'nit . I . tall . U
61 '$pr(:0) : '$proc'(:0) . U
62 '$pr(:0) : $pr(:0) . U
63
64 % b a g o I (preserves order of solutions)
65 bagot(:0, :1, _) : asserta('BAG'('BAG')) . call(:1) .
66 asserta('BAG'(1))) . tall . U
67 %% 0 Item, 1 Condition.
68 bago1(_, _, :0) : 'BAG'(:1) . I . lntobagtzt, U, :0) . U
69 %% 0 Bag, 1 Item,
70 imobag('BAG', :0, :0) : I . retract('BAG, 1, 1) . U
71 %% 0 Flnal_bag,
72 intobag(:0, :1, :2) : retract('BAG, 1, 1) . 'BAG'(:3) . I .
73 |ntobag(:3, :0.:1, :2) . |]
74 %% 0 Item, 1 Thls_bag, 2 FInaI_bag, 3 Next_ltem,
75
76 % end oi le - toyprolog will now read Irom the terminal
77 :dlsplay('Keme| le Ioaded.') .nl . see(user) . U I

257
APPENDIX A.2
Bootstnpper

1 % % % translator ot Prolog-10(mlnl) Into kemel-prolog" % % %

2 transIate(:0, :1) : see(:0) . telI(:1) . program . seen . told .
seeluser) . te|l(ueer) . display(transIated(:0)) . nl . U
%% 0 trom_llle, 1 to_ le
U'l-I -(D %-----------------
6 % maln loop
7 program : rch . skpb(:0) . tag(trar|sl(:0)) . lsendsym(:0) . I . U
8 PYlJ| "'l I P"tI'al'" - I]
9 lransl('@') : I . rch . U
10 transl(%') : oomment(%, :0, U) . I . puttrm) . U
11 transI(:0) :oIauee(:0, :1, U, :2) . putlr(:1) . putvamames(:2, 0).U
12 %% 0 stench, 1 tennrepr, 2 sym_tab
13 Isendsyrn(@') : U % otherwise tall, Ie loop
14 %-----------------
15 % error handing: skip to the nearest dot
16 err(:0, :1) : display( error tn) . dlsplay(:0) .
17 display(: unexpected "') . dIspIay(:1) . lastch(:2) .
18 display( . text skbped: ') . sItb(:2) . nt . taglall(transl(_)).U
19 /0% 0 proo_name, 1 bad_item, 2 rst_sklpped_char
20 sl<Ip('.') :wch(.') . U
21 sl-tb(:0) :wch(:0) . rch . laslch(:1) . skb(:1) . U
22 %----~------------
23 % a oomment extends till end_oI_lhe
24 oomment(:0, 11:1, :1) : IseoIn(:0) . U
25 %% 0 eoln, 1 rest_ot_termrepr
26 oomment(:0, :0.:1, :2) : rch . lastoh(:3) . oomment(:3, :1, 2) . U
27 %% 0 char, 1 termrepr, 2 rest_oI_termrepr, 3 nextchar
28 %---------~-------
29 % read a goal
30 olause(':, :'.:0, :1, :2) : I . ctaIl(':', :0, .@.:1, :2) . U
31 %% 0 termrepr, 1 rest_ot_termrepr, 2 sym_tab
32 % read an asserllon/n.|le
33 olause(:0, :1, :2, :3) :lten'n(:0, :4, :1, .':'.:5, :3) .
34 I . ctall(:4, 5, :2, :3) . U
35 %% 0 ltem1_ rstch, 1 tem1repr, 2 rest_ot_tennrepr,
36 %% 3 sym_tab, 4 otal|_ rstoh, 5 mlddletennrepr
37 oIause(:0, _, _, _) : err(clause, :0) . |]
38 %-----------------
39 % clause tall
40 ctall('.', '.[.'].:0, :0, _) : I . [1
41 %% 0 rest_ol_termrepr
42 % rlghthaml slde ol a non-unit clause. or a goal
43 % eoln and blari-ts Inserted to make the output look tidy
44 ctall(':', :4. .20, :1, :2) : rdch(-') . I . Iseoln(:4) .
45 rdchsl<(:3) . ctallauxm, :0, :1, :2) . U
46 %% 0 termrepr, 1 rest_ot_terrrrepr, 2 sym_tab,3 calls_tIrstch,
47 %% 4 eoln
48 ctall(:0. _, _, _) : err(ctall, 1)) . U
49 % get the righthand slde ol a clause (errbedded oomments not displaced)
50 ctaIlaux('%, :0, :1, :2) :oomment(%, :0, '.:5) . I .
51 rdchskl ) . ctallauxl , :5, :1, -2) . [1
52 %% 0 lerrnrepr, 1 rest_ot_termrepr, 2 sym_tab, 3 rest_lirstch,

258
APPENDIX A.2 (Continued)

53 /6% 5 mlddletermrepr
54 ctal|aux(:0, :1, :2, :3) :fterm(:0, :4, :1, :3) .
55 Iterl'ns(:4, :5, :2, :3) . U
56 %% 0 tterm_Ilrstoh, 1 termrepr, 2 rest_ol_termrepr,
57 %% 3 sym_tab, 4 item\s_Ilrstoh, 5 rriddleterrrtepr
56 % a IIst oI functor-temte (Ie calls)
59 items(), '.'[.']'.:0, :0, _) : I . U
60 %% 0 rest_oI_termrepr
61 % eoInandbIarits-dotalI/2/
62 Iterms(,, :4. .:0, :1, :2) : I . Ise-oln(:4) .
63 rdchslt(i3) . ctaIIaux(3, :0, :1, :2) . U
64 %% 0 termrepr, 1 rest_oI_termrepr, 2 sym_tab,3 ctaIl_lIrstoh,
65 I/0/o 4 BOII1
66 Item's(:0, _, _, _) : err(ftemB, :0) . U
67 %-----------------
66 %atunctor-term
69 Iterrn(:0, :1, ".:2, :3, :4):
70 Iderlt(:0, :5, :2, ".:6) . I . args(:5, :1, :6, :3, :4) . U
71 %% 0 Id_Ilrstch, 1 lerrrinator, 2 terrnrepr,3 rest_oI_terrnrepr,
72 %% 4 sym_lab, 5 Id_lermInator, 6 mlddleterrrtepr
73 % Identlltersz words, I, quoted names, synbols
74 Ident(:0, :1, 11:2, :3) :
75 wordstart(:0) . rdch(:4) . abhanurns(:4, :1, :2, :3) . U
76 %% 0 Id_llrstoh, 1 terrrlnator, 2tem1repr,
77 %% 3 rest_ot_termrepr, 4 nextch
76 Ident('I', 1), 'I'.:1, :1) : rch . skpb(:0) . U
79 %% 0 terrnlnator, 1 terrrrepr
60 ident(", .1), :1, :2) : rdoht ) .qIderl(3, :0, :1, :2) . U
61 /6% 0 lem1lnator, 1 temrepr, 2 rest_oI_terrnrepr, 3 nextoh
82 iderrt(:0, :1, :0.:2, :3):
83 syrnch(:0) . rdoh(:4) . sy|'riJol(:4, :1, :2, 3) . U
64 %% 0 syrnb_Iirstoh, 1 lerrnlnator, 2 tennrepr,
65 %% 3 resl_oI_termrepr, 4 nextch
66 % quoted Identi ers
87 qldent("", :0, :1, :2) :
68 rdch(:3) . qtdentalll , :0, :1, :2) . I . U
69 %% 0 termlnator, 1 termrepr, 2 rest_ot_termropr, 3 nextoh
90 qIdent(:0, :1, :0.:2, :3) : rdch(:4) .qIdent(:4, :1, :2, :3) . U
91 ' %% 0 char, 1 lerrnlnator, 2 termrepr,
92 %% 3 rest_oI_termrepr, 4 nextch
93 qIderlaIl(", :0, :2) :
94 rdch(:3) . qIdent(:3, :0, :1, :2) . |]
95 %% 0 terrnlnator, 1 termrepr, 2 rest_oI_termrepr, 3 nextoh
96 qiderlail(_, :0, :1, :1) :skpb(:0) . U
97 %% 0 terrnlnator, 1 rest_ot_lerrnrepr
96 % words and symbols
99 3bII3ll.IlTB(IO, :1, :0.:2, :3):
100 alphanum(:0) . I . rdch(:4) . abhanurns(:4, :1, :2, :3) . U
101 /6% 0 an_alphanum, 1 terminator, 2 lennrepr,
102 /6% 3 rest_oI_termrepr, 4 nextch
103 aIphanums(_, :0, :1, :1) :skpb(:0). U
104 %% 0 terrrlnator, 1 rest_ot_lerrnrepr

2.59
APPENDIX A.2 (Continued)

105 symboI(:0, :1, :0.:2, :3)

106 syrnch(:0) . I . rdch(:4) . symboI(:4, :1, :2, :3) . U
107 %% 0 a_synt>oIchar, 1 terrnlnator, 2 termrepr,
108 %% 3 rest_oi_tennrepr, 4 nextch
109 syrnboI(_, :0, :1, :1) :skpb(:0). U
110 %% 0 terrrlnator, 1 rest_ot_termrepr
111 % get argumert list: nothlng or a sequence oi temts in brackets
112 args((, :0, (.:1, :2, :3)
113 I . rdchsk(:4) .terms(:4, :1, :2, :3) . rdchslt(:0) . U
114 %% 0 nextch, 1 tem1repr, 2 rest_oI_terrnrepr,
115 %% 3 sym_tab, 4 ten'ns_Iirstch
116 args(:0, :0, :1, :1, _) : U
117 /6% 0 nextoh. 1 rest_oI_tennrepr
118 % get a sequence oi temts
119 tem1s(:0, :1, :2, 3) :terrn(:0, :4, :1, :5, inargs, :3) .
120 termstaII(:4, .5, :2, :3) . U
121 %% 0 terrn_IIrstch, 1 temtrepr, 2 rest_ot_termrepr, 3 sym_tab
122 /6% 4 teminator. 5 mlddletermrepr
123 terrnstaiI(), ).:0, :0, _) : I . U
124 %% 0 rest_ot_termrepr
125 tem1staII(,, .:0, :1 . 2)
126 I . rdchsk(:3) .tem':s(:3, :0, :1, :2) . U
127 %% 0 middletermrepr, 1 rest_oI_tem\repr, 2 sym_tab, 3 nextch
128 tem1staII(:0. _, _, _) : errltennstail, :0) . U
129 %---
130 % get a term (context used to Iorce brackets around Ilsts within lists)
131 terrn(:0, :1, :2, :3, :4, 5) : t(:0, :1, 2, :3, :4, :5) . I . U
132 /6% 0 tlrstch, 1 terminator, 2 termrepr,
133 %% 3 rest_oI_termrepr, 4 context. 5 sym_tab
134 terrn(:0, _, _ _, _, _) : err(term, :0) . U
135 t(:0, :1 '4) :varIable(:0, :1, 0
136 t(:0, :1 nargs, :4) : Iist(:0, :1, 6
137 t(:0, :1 -"'.NI :3, Iniist, :4) : Iist(:0, :1 -NPPP? 5 :4).U
136 % a dirty pa -5.|"5'-5ST:-'5' tor negative numbers
139 ti-, :0, :1, :2, _, :3) :
140 rdch(:4) . numberorItem1(:4, :0, :1, :
141 %% 0 terrrlnator, 1 termrepr, "N 3..0:8 9.1::
i 6" rmrepr,
142 %% 3 sym_tab, 4 nextch
143 :3, _, _) : number(:0, :1, :2, :3) .U
144 9 _, :4) : Iterm(:0, :1, :2, :3, :4) . U
145 a9<=-T-.=':~ ~99I'.1'_:
146 nurnberorItem1(:0, :1,
147 digit(:0) . I . nurrbe :3) . U
146 %'yo 0 TIBXICII r, 2 termrepr, 3 rest_ot_lermrepr
149 numberortterm(:0, :1, ;_& $355 @iL
150 symboI(:0, :5, :2, .:6) . args(:5, :1, :6, :3, :4) . U
151 %% 0 nextoh, 1 terminator, 2 termrepr, 3 rest_oI_terrnrepr
152 %% 4 sym_tab, 5 syrnboI_termInator, 6 middleterrnrepr
153 %~-
154 % get a variable
155 variabIe(:0, :1, : : : :v I. ar;tart(:0) . aIphanums(:0, :1, :5,U).
156 tindv(:5, PP a.u$.6- i.

260
APPENDIX A.2 (Conrinned)

157 %% 0 Iirstoh, 1 terminator, 2 lennrepr,

156 /0% 3 rest_oi_tem\repr, 4 sym_tab, 5 name
159 Iindv(_.U, _.:0, :0, _) : U % no search: an anonymous variable
160 %% 0 rest_oI_terrnrepr
161 Ilndv(:0, ':.:1, :2, :3) : look(:0, 0, :4, :3) . setn(:4, :1, :2).U
162 /0% 0 name, 1 termrepr, 2 rest_oI_tem1repr, 3 sym_tab, 4 num
163 % look counts Irom 0 and Ilnds the position oi a name In the syn1tab
164 IooI<(:0, :1, :1, :0.:2) : U
165 %% 0 name, 1 num, 2 syrntabtail
166 looi<(:0, '2, :1, _.."3) :sum(:2, 1, :4) . looi<(:0, :4, :1, :3) . U
167 %% 0 name, 1 num, 2 ourrnum, 3 symtabtail, 4 GUlItlUmpl.|$1
166 % set a number: no more than two digits (should be enough)
169 setn(:0, :1.:2, :2) : lee-s(1), 10) .
170 ordohr(:3, '0') . surn(:3, :0, :4) .ordchr(:4, :1) . U
171 %% 0 num, 1 char, 2 rest_oi_termrepr, 3 k, 4 kpiusnum
172 setn(:0, :1, :2) : iess'(:0, 100) . prod(10, :3, :4, :0) .
173 setn(:3, :1, :5) . setn(:4, ".5, :2) . [1
174 /6% 0 num, 1 termrepr, 2 rest_oi_termrepr,
175 %% 3 nurnby10, 4 nummod10, 5 mlddletermrepr
176 setn(:0, _, _) : err(setn, .13) . U
177 % - - - - - - - - - - - - - - - - -
176 % get a list In square brackets
179 Iist([, :0, :1, :2, ii) : rdchsk(:4) . endIist(:4, :1, :2, :3) .
160 rdchsk(:0) . U
161 %% 0 terrrlnator, 1 termrepr, 2 rest_oi_termrepr,
162 %% 3 sym_lab, 4 nextch
163 endIIst('1, [.'1'.1J, :0, _) : U
164 %% 0 rest_oI_terrnrepr
165 endIist(:0, :1, '2, :3) :
166 term(:0, :4, :1, ..:5, lnlist, :3) . ItaiI(:4, :5, :2, :3) . U
167 %% 0 lirstoh, 1 lennrepr, 2 rest_ot_tem\repr,
166 %% 3 sym_tab, 4 nextch, 5 middlelermrepr
169 ItaII(']', '['.'1'.:0, :0, _) :I . [1
190 %% 0 rest_oi_termrepr
191 Itali(|, :0, :1, :2) :I . rdchsl-r(:3) . variable(:3.']'.1J,:1,:2).[1
192 /0% 0 termrepr, 1 rest_oi_termrepr, 2 sym_tab, 3 nextch
193 italIi'.'. :0, :1, :2) : I . rdchsiq ) .
194 terrn(:3, :4, :0, ..:5, iniist, :2) . ltaIi(:4, :5, :1, :2) . U
195 %% 0 termrepr, 1 rest_oi_termrepr, 2 sym_tab,
196 %% 3 tem1_iirstoh, 4 nettch, 5 middleterrnrepr
197 ItaiI(:0, _, _, _) : err(Itail, :0) . U
196 % - - - - - - - - - - - - - - - - -
199 % numbers: only natural ones
200 nurrber(:0, :1, :2, :3) :digIt(:0) . digIts(:0, :1, :2, 3) . U
201 %% 0 rstch, 1 non_digit, 2 tennrepr, 3 rest_oi_termrepr
202 dlgIts(:0, :1, :0.:2, .13) : dIgit(:0) .
203 I . rdch(:4) . dlgits(:4, :1, :2, :3) . U
204 %% 0 iirstch, 1 non_digit, 2 lennrepr, 3 resl_ot_terrnrepr.
/070 4 DBXICII
206 dIgits(_, :0, :1, :1) :skpb(:0) . U
207 %% 0 non_digit, 1 rest_oi_termrepr
20s % - - - - - - - - - - - - - - - - -

261
APPENDIX A.2 (Conlinned)

209 % auxiliary tests

210 vvordstart(:0) :smaIIetter(:0) . U
211 varstart(:0) : blgietter(:0) . U
212 varstart(_) : U
213 %-----------------
214 skpb(:0) : skbbl . Iastch(:0) . U
215 %-----------------
216 /oOUIDI.i1lI'IOIl8I'IS|81bI1
217 puttr(U) : I . U
216 puttr(:0.:1) :vvch(:0) . puttr(:1) . U
219 putvamames(:0, _) :var(:0) . I . nl . U
220 /0/o 0 sym_tab_end
221 putvamal'ne=s(:0.:1, :2) : nextIine(:2) .wch( ) . dlspiay(:2) .
222 puttr( .:0) .woh(,) . surn(:2, 1, :3) . putvamames(:1, :3) . U
223 %% 0 ourrname, 1 sym_tab_taiI, 2 currnum, 3 nextnum
224 nextllne(:0) : prod(6, _, 0, :0) . I . ni . display( %%') . U
225 %% 0 a_rnultIpie_ot_Ilne_sIze
226 nextI|ne(_) : U
227 /0/o%IhO B| /o%%
226 :dlspIay("'BO0TSTRAPPEFI" Ioaded.') . nI . see(user) . U I

262
APPENDIX A.3
Um Interface and Utilities

Interpreter oi Toy-Prolog - the Prolog part.

) COPYRIGHT 1963 - Fellrs Khznlak. Stanislaw Szpakowicz
institute oi lniomtatlce, Warsaw University
I I 0 0 I I O I Q I I I I O 0 I I 0 I I I I I I 0 0 I 0 Q I I n O O Q I Q Q Q Q I I I I I 0 0 0 I 0 I 0 0 I I I O O O O I I on
I Q O O Q Q I 0 O 0 I I I I 0 I I I I I I I I 0 n I Q I I O 0 I 0 O I 0 O I I I I I 0 I I 0 0 0 Q I O O O O I n n I O 0 O 0 II

interactive driver - top level

O O I I I I I O I I I I I I O I I O O O I I O I O O O I I O O I I I I O I I O O O O I I I I I I O I I O I O Q I O I I I O I II
I I I I I I I 0 I I I I I I I I O I O O I O I I I I I I I I 0 I 0 I I 0 I I I I 0 O O I Q I Q I I n I O I I I 0 I I I I I 0 II

:- nl, dispiay('Toy-Prolog Ilstenlng:'), nl, tag(Ioop).

Ea?69E8?a9%9$ :- hait('Toy-Prolog, end oi sesslon.).
loop :- repeat.
jacoda-udum-bun:- dlsplay(?- '), read(Term, Sym_tab), exeo(Tem1, Sym_tab), Iall
12
13 S1013 2- I3QI&|l(I0Op).
14
15 exec(err', _) :- I. %this oovers variables, too
16 exec(:-(Goals), _) :- I, once(GoaIs).
17 exec(N,_) :- Integer(N), I, num_clar.se.
18 exec(GoaIs. Sym_tab) :-
19 call(Goals), nurrbervars(Goals, O,_),
20 printvars(Sym_tab), enoudt, I.
21 exeo(_, _) :- dlsplay(no), nl. %l1caII(Goals) tails
22
23 enough :- rch, skipbl, lastch(Ch), rch, not(-(Ch, :'l).
24
25 prIntvars(Sym_tab) :- var(Sym_tab), dlspIay(yes), nl, I.
26 prlntvars(Sym_tab) :- prvars(Sym_tab).
27
28 prvars(Sym_tab) :- var(Sym_tab), I.
29 prvars([var(NarneStrlng, Instance) | Sym_tab_tail]) :-
30 writetext(NameStn'ng), display( - '),
31 slde_eiIects(outt(Instance, ld(_, _), q)),
32 % this is equivalent to writeq(lrstance) but we avoid
33 % suporlluous calls on nunbervars - cl WRITE
34 nl, prvars(Sym_tab_taiI).
35
36 num_clause :- dispIay(+++ A nurrber cant be a cIause.'), nl.
37
38 % read a program upto end. (the only way to de ne user procedures)
39 % oonsultlreoonsult must be Issued Irom the terrnlnal, and it retums
40 % there ( oonsult(user) is correct, too 1
41 oonsult(FIle) :- seelng(0IdF), reacbrog(Flie), see(OldF).
42 reconsul(FIle) :-
43 redeline, seelng(OidF), reaq:>rog(Flie), see(OidF), redefine.
44 readprog(user) :- I, getprog.
45 readprog(FlIe) :- see(FIle), echo, getprog, noecho, seen.
46
47 % the actual lob is done by this procedure
48 getprog :- repeat, read(T). asslrnIIate(l'), -(T.end), I.
49
50 asstmIlate(e r r) :- l. % a variable is erroneous. too

263
APPENDIX A.3 (Continued)

51 assimilate( -->(Leit, Right) ) :-

52 I, tag(transI_ruIe(LeIt, Right, Clause)), assertz(Clause).
53 asslmiIate( :-(Goal)) :- I, once(GoaI).
54 asslmilate(end) :- I.
55 assimilate(N) :- integer(N), I, num_clause.
56 % otherwise - store the clause
57 assimiIate(Clause) :- assertz(Clause).
58
59
60

reading a term
$a\$
read(1') :- read(T, Sym_tab).
read(T, Sym_tab) :-
gettr(T_Intemal, Sym_tab), I, maketerrn(T_lntemal,1').
% ii gettr iails, then...
read('err,_) :-
nl, dispIay(+-I-+ Bad term on input. Text skipped: '), skip, nl.
B'8 3$ 28 9
71 % skip to the nearest lull stop not In quotes or in oomment
72 skip :- lastch(Ch), wch(Ch), sklp(Ch).
73
74 skip(.) :- rch, lastch(Ch), e_skb(Ch), I.
75 sklp(%') :- skip_commerl, I, rch, skb.
76 skip(Q) :- isquote(Q), skip_s(0), I, rch. skip.
77 skip(_) :- rch, skip.
78
79 % stop on a layout character
80 e_skip(Ch) :- @-<(Ch, '1.
81 e_skIp(Ch) :- wch(Ch), rch, skip.
82
83 skip_commerlt :- repeat. rch, lastch(Ch), wch(Ch), lseoln(Ch), I.
84
85 isquote(""). lsquote("").
86
87 /0 HSIIIHQ
88 skip_s(Q) :- repeat, rch, lastch(Ch), wch(Ch), -(Ch, Q), I.
89
90 % IIIIIIIIIIZZIIIHIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIII
91 % parse r
92 /o IIIIIIIIIIIZIZIIIIIIIZIIZIIIIIIIIIIIIIIIIIIIIIIIIIIIZIIIIIIIIIIIIII
93 % This is an operator precedence parser Ior Prolog-10. g e t t r
94 % constructs the intemal representation oi a term. Next, m a k e-
95 % t e r m constructs the temt proper - see r e a d. Here is an In-
96 % Iorrnal description oi the underlying operator precedence grammar
97 % (each rule corresponds to one clause oi r e d u c e). Sides are
98 % separated by --> and multiple righthand sides by OR.
99 % t --> variable OFI integer OR string
100 /o I II) IdBl'III1I8l
101 /o I II) Id6l'III1IOI' ( I )
102 % I --> U OFI {}

264
APPENDIX A.3 (c liillltd)

% t-->(t) OFI [t] OFI {t}

% t --> [ I I t ]
% t --> t postiix_Iunctor
0/0 I II) I Il'lIIX_IU|'BIOI'I
% t --> preiix_Iunctort
% Sequerces oi temts separated by commas - in rules 3, 5. 6 - will be
% recognised as comma-temts (commas are iniix functors, covered by
% rule 6). There are ve types oi operators, vns(_), ld(_).
% I1(_, _, _), br(_, _), bar: see the scanner. The temtinal symbol dot
% never gets onto the stack. The temtinal symbol bottom is never re-
% tumed by the scanner; it is only used to initiate and terminate the
% main loop (p a r s e). The only nontemtinal syrnboi is t(_).
% There are live types oi intemal represeriations (Args denotes the
% represerlatlon oi argumerls - usually a comrna-term):
% tr(Name, Args) - tor iunctor-terms.
% argO(X) - ior X a variable, atom, number, or stn'ng,
% bar(X, Y) - tor a list with Irom X and tail Y,
% tr1(Name, X) - Iorpreiix and post x Iunclors.
% tr2(Name, X, Y) - Ior lniix functors.
% A Name in tr mey be a bracket type. See r e d u c e (clauses 5, 6)
% and makete rm iordetails.
% - - - get the intemal representation oi a term
gettr(X, Sym_tab) :-
gettoken(T, Sym_tab). P rse([bottom], T, X, Sym_tab).
% p a r s e takes 4 parameters: the current stack, the current token
% Irom input, the variable used to bring the intemal representation
%tothe suriace, andthe symboltabie (usedby getto ke n)
parse([t(X), bottom], dot, X, _) :- I.
parse(Stack, Irput, X, Sym_tab) :-
topterminaI(Stack, Top, Pos),
establish_precedence(Top, Irput, Pos, Rel, FITop, Filnput),
exch_top(Top, FITop, Stack, FIStack).
step(FIel, Fiinput, FIStack, NewStack, Newlnput, Sym_tab).
parse(NewStack, Newlrput, X, Sym_tab).
% the topmost terrninai will be covered by at most one nonterrninai
% (the third parameter gives Tops position: 1 on the top, 2 covered)
tootemtlnaltilu. Top I _l. Ton. 2) 1- I-
topterminaI([Top | _], Top, 1).
% change the topmost terminal (applies only to mixed iunctors)
exch_top(Top. Top, Stack, Stack) :- I.
exch_top(_, FITop. [t()()_ _|S]. [t(X), FITop |S1) :- I.
exch_top(_, FITop, L I S1, [FITop | S1).
% - - - periorrn one step: shift (stack the current token) or reduce
step(lseq, Fiirput. Stadt, [Rlrput I Stack], Newlnput, Sym_tab) :-
I, gettoken(Newlnput, Sym_tab).
step(gt, Filnput, Stack, NewStad<, Fiinput, _) :-
reduce(Stack, NewStack), I.

265
APPENDIX A.3 (Coniinned)

155 % iail ii reduction Irrpossibie (parse and gettr will Iail, too -
156 % this iallure will be intercepted by gettre caller)
157
156 % reduce top segment oi the stack according to the underlying grammar
159 reduce([vns(X) | S], [t(arg0(X)) 1 S1).
160 redvceil Idtll I SI. ltlaro illl I SD-
161 redvs llbrir. 0'l- lIX).brtl. '0'). ldlll I $1.
162 (t(tr(l. Xll I Sil-
163 redus ilbrir. TYP9). br(l.TvPBl I $1.
164 IIIHYOOITYPBII I $1) 1- "OIHTYPB. (I'll-
165 %'U or {}, see p, 2nd clause
166 wdvcelibrtr. Tyne). IIXI. bI'(|-TYPBI I Si.
167 ltltr(TrP@.Xll I SD-
16B reduceilbrir. 1]). IIYI. bar. IIXI. br(l.[i'l I $1.
169 [t(bar(X,Y)) | S1).
170 reduce([I1(l,Type,_), t(X) 1 S],
171 [t(tr1(I, X1) | S1) :-
172 isrrpost1(Type).
173 reduce([t(Y), I1(I, Type,_), t(X) | S],
174 [t(tr2(l, X,Y)) | S1) :-
175 isminI(Type).
176 reduce([t(X), il(l,Type,_) | S],
177 [t(tr1(I, X)) | S1) :-
176 Isl1'preI(Type).
179 % otherwlseIalI(cI step)
160
161 %- - -auxiliary tests ior the parser
162 ispreI(iy). lspreI(Ix).
163
184 ispostI(yI). ISDO-SI'I(X1).
185
186 isrnpreI(['i'Un]) :- Isprei(TUn).
188
187 Ismprei(L, TUn]) :- IspreI(TUn).
169 isminI([i'Bln]) :- mer1'ber('i'Bln,[rriy,ybr,:rix]).
190 isminI(L, _1).
191
192 ismposti(1'TUn1) :- lspostl(TUn).
193 ismpost1(L, TUn]) :- Ispost1(TUn).
194
195 % - - - establish precedence relation between the topmost
196 % temtinal on the stack and the cunent lrput terrnlnal
197 estabiish_precedence(Top, Input, Pos, Rel, RTop, Rlrput) :-
196 p(Top,I npu t,Pos,Rei0),
199 Iinalize(Rel0. Top, Input, Rel, RTop, Rlnput), I.
200
201 Iinalize(lseq, Top, Input, Iseq, Top, Input).
202 linalize(gt, Top, input, gt, Top, Input).
203 Iinalize(lseq(RTop, Rlrput), _, _, Iseq, RTop, Rlnput).
204 linalIze(gt(RTop, Rlnput). _, _, gt, RTop, Rlnput).
205
206 p(id(_), mu, '0'), 1, iseq).
ms
APPENDIX A.3 (Continued)

207 p(br(I, Type), br(r, Type), _, lseq).

206 p(br(I, U), bar, 2, iseq).
209 p(bar, br(r, U), 2, lseq).
210
211 p(Top,lrput,1,gt) :-
212 vns_Id_br(Top, r), br_bar(lnput, r).
21a P(TP.I1(N.TlIPB3. Pl. 1. at(T=>r>. "IN. FITYP$- Pill =-
214 vns_ld_br(Top, r), restrIct(Types, [fx. IY]. RTypes).
215 p(Top,lnput,1,lseq) :-
216 br_bar(Top, I), vns_id_br(lmut, I).
217 ntTen. IN.Tvr>e=.Pl.Pos.lseqlT0p. tN.HTvrm.Plll =-
216 br_bar('l'op, I), pre_lnpost(Pos, Types, RTypes).
219 p(I1(N, Types, P), Input, Pos, gt(ll(N, RTypes, P), Imut)) :-
22O br_bar(Input, r), post_lrpre(Pos, Types, RTypes).
221 PIIIIN. TYP68. P). Input. 1. |99q(II(N.FITYP03.P). Irwin I-
222 vns_id_br(lnput, I), restrict(Types,[x'i,yI].RTypes).
223
224 %Iunctors with equal priorities
225 p(I1(NTop, TeTop, P),I1(Nlrp,TsIrrp, P), Pos, Rel) :-
226 ree_confi(TsTop, Tslrp, Pos, RTsTop, RTsIrp, ReI0),
227 I, do_rel(RelO, Ii(NTop, RTsTop, P), Ii(Nlnp, RTslnp, P), Rel).
226 % different priorities
229 r>l tNTer>. TsTop. PTor>l. tNInr>. TellP- Plwl. Poe.
230 gt(Ii(NTop, RTsTop, PTop), i1(Nlrp, RTslnp, P|rrp))) :-
231 slronger(PTop, PIFP). I,
232 restrlct(Tslrp, (ix. Ivl. FITeIrp),
233 pO6I_lt|prg(PO6. TeTop, RTsTop).
234 p(li(NTop, TsTop, PTop), li(NInp, Tslrp, PIIIJ). Pos,
235 iseq(l1(NTop, RTsTop, PTop), l1(NIrp, RTelrp, Plrp))) :-
236 stronger(Plnp,PTop), I.
237 restrict(TsTop, (xi, yl], RTeTop),
238 pre_lrpost(Pos, Tslnp, RTslnp).
239
240 p(_, dot, _, gt).
241 p(bottom, _, _, Iseq).
242 % otherwise fail (p a r e e fails, too)
243
244 vrls_Id_br(vns(_), _).
245 vns_id_br(id(_), _).
246 vns_id_br(br(LeitRIght, _), LeitRIght).
247
246 br_bar(br(LeitRIght, _), LeItFIIghl).
249 br_bar(ba, _).
250
251 stronger(Prior1, Prlor2) :- less(Prlor1, Prlor2).
252
253 pre_lnpost(1, Types, RTypes) :- % the Iunctor trust be prefix
254 reslrlct(Types, (xi, yi1, A),
gig restricts. lriv. vir. ml. FlTvr>e=l-
pre_irpost(2, Types, RTypes) :- % the functor must not be prefix
:2; F6811?!-'I(TV1568. III. Ill]. FIT!!!)FBI-

267
APPENDIX A.3 (Continued)

259 post_inpre(1, Types, RTypes) :- % the functor must be postiix

260 restricl(Types, (ix, fy], A),
261 restrict(A, [xfy, yix, xix], RTypes).
262 post_Inpre(2, Types, RTypes) :- % the functor rnust not bepostiix
263 restrict(Types, [xI, yl], RTypes).
264
265 % leave only those types that do not belong to RSet,
266 % tail ii this would leave no types at all (RSet
267 % contains only binary types, or only unary types)
266 restrict([T1, RSet, (T1) :- I, not(merrber(T, RSet)).
269 restrict([TBin, TUn], RSet, [TBin]) :- mel'r|ber(TUn, RSet), I.
270 restrict([TBin, TUn], RSet, [TUn]) :- men'ber(TBin, RSet), I.
271 restrlct(Types, _, Types).
272
273 % compute relation for two functors with equal priorities; four cases:
274 % both nom1aI, Top rnlxed, Input mixed, both rnlxed
275 res_confl([Ti'op1, [Tlnp], Pos, [Tl'op], [Tim], ReI0) :-
276 I, I1_p(TTop,TInp, Pos, Rei0).
277 res_coniI(1Ti'opBIn,I"I'opUn],[Tlnp], Pos, RTsTop,1TIrrp], ReI0) :-
276 I, I1_p(iTopBIn, Tirp, Pos, RelB),
279 f1_p(TTopUn, Tlrp, Pos, RelU),
260 match_reIs(RelB, ReIU, ReI0,iTopBin, 'iTopUn, RTsTop).
261 res_confi([TI'op], 1TlnpBin, TInpUn], Pos, [TI'op], RTelrp, Rei0) :-
262 I, I1_p(iTop, TlnpBin, Pos, RelB),
263 I1_p(TTop, TlnpUn, Pos, ReIU),
264 match_reIs(RelB, RelU, Rel0, TIrpBin, TIrpUn, RTsInp).
265 res_conil(1Ti'opBin,Tl'opUn],1'l'InpBin, TlnpUn], Pos, RTsTop, RTsinp
266 Rel0) :- I1_p(1l'opBin, TlnpBin, Pos, ReIBB),
267 I1_p(iTopBIn,TIrpUn, Pos, RelBU),
266 l1_p(iTopUn,TlnpBin, Pos, ReIUB),
269 I1_p(iTopUn, TinpUn, Pos, RelUU),
290 res_mixed(RelBB, RelBU, RelUB, ReIUU, Rel0,
291 TTopBin, lTopUn, TlnpBin, TInpUn, RTsTop, RTslnp), I.
292
293 do_rel(lseq, TopF, InpF, Iseq(TopF, InpF)).
294 do_rel(gt, TopF, InpF, gt(TopF, InpF)).
295 % fail II ReI0 - err
296
297 match_reIs(Rel, Rel, Rel, TBin, TUn, [TBin, TUn]) :- I. % err Included
296 match_reis(en, Rel, Rel,_,TUn,[TUn1) :- I.
299 match_reis(Rei, err, Rel, TBin, _, [TBln]) :- I.
300 match_reIs(_. _, en, TBin, TUn, [TBin, TUn]).
301
302 res_mlxed(Rel0, Rel0. Rel0, Rel0, FleI0,
303 TTopBin, iTopUn, TlnpBIn, TIrpUn,
304 ['l'l'opBin, TTopUn], [TInpBin, TlnpUn]).
305 res_mixed(err, err, RelUB, ReIUU, Rel0.
306 _, iTopUn, TInpBIn,TInpUn,1'lTopUn], RTsInp) :-
307 match_reis(ReIUB, ReIUU, Rel0, TInpBIn, TlrrpUn, RTsInp).
306 res_mixed(ReIBB, RelBU, err, err, Rel0,
309 TTopBin, _, TlnpBin, TlnpUn, [Ti'opBin], RTslnp) :-
310 match reis(ReiBB, ReIBU, Rel0, TInpBIn, TlnpUn, RTslnp).

168
APPENDIX A.3 (Continued)

311 res_mixed(en, RelBU, err, ReIUU, Rel0,

312 TTopBin, TTopUn, _, TInpUn, RTsTop, |TlnpUn1) :-
313 match_rels(ReiBU, ReIUU, Rel0, TTopBin, TTopUn, RTsTop)
314 res_mixed(ReIBB, err, ReiUB, err, Rel0,
315 iTopBln,TTopUn,TlnpBln,_,RTsTop,['i'InpBin]) :-
316 match_reis(RelBB, ReIUB, Rel0, TTopBin, TTopUn, RTsTop)
317 res_mlxed(_, _, _, _, err, _, _, _. _, _, _).
316
319 % establish precedence relation for two (basic) types
320 l1_p(i'i'op,TInp, Pos,lseq) :-
321 merrber(Ti'op, [xiy, iy]), % right_associative
322 il_p_aux1(Pos, Tlnp), !.
323 I1_p(1l'op,Tlrp,Pos.gt) :-
324 merrber(Tlnp, [yfx, yI1), % Ielt_associative
325 I1_p_aux2(Pos, iTop), I.
326 i1_p(_,_,_, en).
327
326 i1_p_aux1(1, Tlrp) :- lspreI(TInp).
3323 l1_p_aux1(2,Tlrp) :- merrber(Tlrp, [xiy, xi, xix]).
331 I1_p_aux2(1, lTop) :- Ispos'li('lTop).
332 I1_p_aux2(2, iTop) :- mernber(iTop,[yix,fx, x1x1).
333
334 /o IIIIIIIIIIZIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIII
335 % intemal representation --->tem1
336 % IIIIIIII!IIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIZIZIIIIIIIZIIIIIIIIZ
337 maketerm(arg0(X), X) :- I. %variabIe, atom, nurnber,strlng
336 maketenn(tr((), RawTerm), T) :-
339 I, maketenn(RawTerm, T).
340 maketem1(bar(RawLIst, RawTaII),T) :-
341 I, maketen'n(RawTail, Tail),
342 makeIist(RawUst, Tail, T).
343 maketerm(tr('U, RawLIst), T) :-
344 I, makelIst(RawList, 'U,1').
345 maketenn(tr((), RawArg), {}(Arg)) :-
346 I, maketenn(RawArg, Arg).
347 maketerrn(tr(Name, RawArgs), T) :-
346 I, makeIlst(RawArgs, U, Args),
349 -..(T, (Name | Args]).
350 maketerm(tr2(Name, RawArg1, RawArg2), T) :-
351 I, maketem1(RawArg1,Arg1), maketerm(RawArp2.Arg2),
352 -..(T, (Name, Argt, Arg2]).
353 maketerm(tr1(Name, RawArg), T) :-
354 maketerm(RawArg.Arg), -..(T,[Name,Arg]).
355
356 % comma-term to dot-list-with-Tai
357 makeIist(tr2('.'. RawArg, RawArgs), Tall, [Arg | Args]) :-
356 I, maketenn(RawArg, Arg). makeIist(RawArgs,TaII, Args).
359 makellst(RawArg, Tail, [Atq | Tall1) :- rn-aketenn(RawArp, Atq).
360

169
APPENDIX A.3 (Continued)

361 %IIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIII
362 % ecanner
363 %:::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
364 % this scanner retums six kinds ol tokens:
365 % vns(_) variables, nurrbers, strings
366 % ld(Name) atons
367 % iI(Name, Types, Prior) fix Iunctors
366 % br(Which, Type) brackets (lelt/right. ll/I1/1})
369 % bar | (in lists)
370 % dot . followed by a layout character
371
372 % - - - read a token and construct its intemal form
373 %the lrx:tt.ttissupposedtobeposItioned
374 % over the first character oi a token (or preceding white space)
375 gettoken(Token, Sym_tab) :-
376 skipbl, Iastch(Startch), absotbtoken(Startch, Rawtoken), I.
377 maketoken(RawIoken, Token, Sym_tab), I.
376
379 % - - - read in a suitable sequence oi characters
360 % a word, le a regular alphanumeric identifier
361 absotbtoken(Ch, ld([Ch | WordtaIl1)) :-
362 won:Istatt(Ch), getword(WorclaII).
383 yo 8 VHTIQDIB
364 absorbtoken(Ch, var([Ch | Tall1)) :-
365 vatstatt(Ch), getword(Tai).
366 % a solo character is a comma, a semicolon or an exclamation mark
367 absorbtoken(Ch, ld([Ch])) :- solochar(Ch), rch.
366 %abracket,le()[]{}
369 absor1:ttoken(Ch, br(Wh, Type)) :-
390 bracket(Ch), bracket(Ch, Wh, Type), rch.
391 absorbtoken(|, bar) :- rch.
392 % a string In quotes or In double quotes
393 absotbtoken(, qId(Qname)) :-
394 rdch(Nextch), getstrlng("", Nextch, Qname).
395 absorbtoken("", str(Strlng)) :-
396 rdch(Nextch), getstrIng(", Nextch, String).
397 % a positive nunber
396 absorbtoken(Ch, nt.tm([Ch | Digits])) :-
399 dlglt(Ch), getdlgls(DIgIts).
400 % a negative nurrber or a dash (possbly starting a syrrbol, see below)
401 absotbtoken(-, Rawtoken) :- rdch(Ch), nt.tm_or_sym(Ch, Rawtoken).
402 absotbtoken(., Rawtoken) :- rdch(Ch), dot_or_eym(Ch, Rawtoken).
403 %asymboI,buIitoi.:-<->+/'?&$@l_"
404 absorbtoken(Ch, ld([Ch | Syn-bs])) :- symch(Ch), getsym(Symbs).
405 % an embedded comment
406 absotbtoken(%, Rawtoken) :-
407 skbcommert, lastch(Ch), absotbtoken(Ch, Rawtoken).
406 % this shouldn't happen:
409 absorbtoken(Ch, _) :- dlspIay(errInscan(Ch)), nl, fail.
410
41 1 num_or_sym(Ch, num([-, Ch | Dlgits])) :-
412 dlglt(Ch), getdlgls(Dlgits).

270
APPENDIX A.3 (Continued)

413 num_or_sym(Ch, id([-, Ch 1 Syn'l:s])) :- symch(Ch), getsym(Syn'bs)

num_or_sym(_, id([-1)).
416 % layout characters precede in ASCII
417 dot_or_sytn(Ch, dot) :- @-<(Ch, '). I6 no advance
416 dot_or_eym(Ch, id([., Ch | Symbs])) :- symch(Ch), getsyrn(Syrrbs).
3 dot_or_sym(_, Id([.1)).
421 skbcommert :- Iastch(Ch). BeoIn(Ch). skipbl. I.
422 skipcommert :- rch, skipcomrnerl.
423
424 % - - - auxiliary Input procedtxes
425 % read an abhanumerlc identi er
426 getword([Ch | Word1) :-
427 rdch(Ch), aIphartum(Ch), I, getwordtword).
428 getword(U).
429
430 % read a sequence ol digits
431 getdIgls([Ch | Digltsl) :-
432 rdch(Ch), dlglt(Ch), I, getdlgls(DlgIis).
433 getdIgIs(U).
434
435 % read a symbol
436 ttsislrrttili?-I1 I Svmbsll =-
437 rdch(Ch), sytnch(Ch), I, getsym(Syn'l:s).
g c@tBvmiI]l-
440 % read a quoted id orstrlng (Dellrn B either or '1
441 getstrlng(Deilm, Dellrn, Str) :-
442 I, rdch(Nextch), twodeIIms(DeIIm, Nextch, Str).
443 getstrlng(DelIm, Ch, [Ch | Strl) :-
444 rdch(Nextch). 99Istrlng(DeIlm, Nextch, Str).
445 twodellms(DeIIm, Delim, [Delim 1 Str]) :-
446 I, rdch(Nextch), getstrlng(Dellm, Nextch, Str).
447 twodelims(_, _, U). %cIose the list
446
449 % - - - auxiliary tests
450 wordstart(Ch) :- smalIetter(Ch).
451 varstart(Ch) :- blgletter(Ch).
452 vars1art(_).
45s btacket((, t, '()'). breoket('), r, '()').
454 bracket('[, t, 11'). bracket(1, r, 1]).
4552 bracket((, I, '{)'). bracket(), r, '{)').
457 % - - - translonn a raw token Irtto its final form
456 maketoken(var(Namestrlng), vns(Ptr), Sym_tab) :-
459 makeptr(Namestring, Ptr, Sym_tab).
460 maketoken(k:I(Namestring), Token, _) :-
461 pname(Name, Namestrlng), make_iI_or_k:i(Name, Token).
462 maketoken(qld(Nanlestrl e). id(Name), _) :-
463 pname(Name, Namestrlng).
464 maketoken(rum([- 1 DigIts]), vns(N). _) :-

2'7]
APPENDIX A.3 (Continued)

465 pnamei(N1, Digits), sum(N, N1, 0).

466 maketoken(mm(Digits), vns(N), _) :- pnameI(N, Digits).
467 maketoken(str(Chars), vns(Chats), _).
466 maketoken(Token, Token, _). % br(_,_) and bar and dot
469
470 % variables are kept In a syrnboi table (an open IBI)
471 makeptr([_], _, _). %no search - an anonymous variable
472 makeptr(Nmstr, Ptr, Sym_tab) :- iook_var(var(Nmstr, Ptr), Sym_tab)
473
474 % look-up
475 look_var(ltem, [Item | Sym_tab]).
476 look_var(ltem, L I Sym_tab]) :- look_var(ltem, Sym_tab).
477
476 make_I1_or_ld(Name, Ii(Name, Types, Prlor)) :-
479 FF(Name, Types, Prior), I.
460 make_I1_or_k:I(Name, Id(Name)).
461
462 /o IIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIZIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIII
463 % grammar rule preprocessor
464 % :::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
465 transI_t'ule(Lelt. Right, Clause) :-
466 two_ok(Leit, Right),
467 lsolate_Ihs_t(Left, Nont, Lhs_t),
466 oonnect(Lhs_t, Ot.ttpar, Finaivar),
469 expand(Nont, Inltvar, Outpar, Head),
490 makebody(RIght, Inltvar, Finalvar, Body, AIt_ ag).
491 do_clause(Body, Head, Clause).
492
493 do_cIause(true, Head, Head) :- I.
494 do_clause(Body, Head, :-(Head, Body)).
495
496 % Lhs_t is a list (possibly enpty) oi leithand side tem'tinaB
497 isoIate_Ihs_t(,(Nont, Lhs_t), Nont, Lhs_t) :-
496 ;(nonvarint(Nont), ruienor(varint)),
499 ;(isclosedlist(Lhs_t), ruIerror(ter)), I.
500 isoIate_lhs_t(Nont, Noni, U).
501
502 %fail II notaclosed list
503 IscIosedilst(L) :- check(BcII(L)).
504 lscII(L) :- var(L), I, lail.
505 lscIl(U).
506 Iscli(L | L1) :- BclI(L).
507
506 % connect terminals to the nearest nontem'tInals input parameter
509 % (actually, open a closed list)
510 connect(U, Nextvar, Nextvar) :- I.
511 connect(lTsym | Tsyms]. ['l'sym | Outpar], Nextvar) :-
512 oonnect(Tsyt'ns, Outpar, Nextvar).
513
514 % - - - translate the righthand slde (loop over altematives)
515 % In altematives. each righthand slde B preceded by a dummy
516 % nontenninal, as defined by ch.tmmy' --> U. (sirtce temtinals

272
APPENDIX A.3 (Continued)

517 % are appended to lrput parameters, the lrput parameter of a common

516 % leithand side must be a variable)
519 makebody(;(Alt, Alts), Inltvar, Finalvar.
520 ;(,( dummy(InItvar, Nextvar), AIt_b), AIt_bs), _) :-
521 I, two_ok(AIt, Alts),
522 makerigix(AIt, Nextvar, Finalvar, AIt_b),
523 makebody(Aits, Inltvar, Finalvar, Alt_bs, alt).
524 makebody(Right, Inltvar, Finalvar, Body, AIt_fIag) :-
525 var(Alt_Ilag), I, % only one altemative
526 makerigix(Right, Inltvar, Finalvar, Body).
527 makebody(Right, lnitvar, Finalvar,
526 ,(dummy(Initvar, Nextvar), Body), alt) :-
529 makerlgix(Right, Nextvar, Finalvar, Body).
530
531 % - - - trarslate one altemative
532 makerigix(,(ltem, Items), Thispar, Finalvar, T_item_items) :-
533 I, two_ok(ltem, items).
534 transI_ltem(item, Thispar, Nextvar, T_item),
535 makerlgix(ltems, Nextvar, Finalvar, T_itemsl.
536 conbIne(T_ltem, T_itenB, T_item_ltems).
537 makerlgl'I(ltem, Thlspar, Finalvar, T_item) :-
536 transl_Iiem(ltem, Thispar, Finalvar, T_item).
539
540 conbIne(true, T_ltems, T_item) :- I.
541 cornbIne(T_ltem, true, T_item) :- I.
542 cot'nbine(T_item, T_IIenB, ,(T_ltem, T_ltems)).
543
544 %---translate one item (sure tobeafunclor-temt)
545 transl_ltem(Ten'ninals, 'I'hlspar, Nextvar, true) :-
546 Bciosediist(Tenninals),
547 I, connect(TennlnaIs, 'I'hBpar, Nextvar).
546 % conditions (the out and others)
549 transI_item(I, ThBpar, Thlspar, I) :- I.
550 transI_ltem({}(Corxl), 'i'hispar, Thlspar, call(Corxl)) :- I.
551 /abadlist of tem1inaB (missed the rst clause)
552 transl_item(L|_1, _, _,_) :- t'ulerror(ter).
553 %a nested altemative
554 transl_ltem(:(X, Y), ThBpar, Nextvar, Trarsl) :-
555 I, makebody(:(X,Y)_ThBpar,Nextvar,Transl,_).
556 %finaliy,aregularnontem1lrtal
557 Il'3iL5|_II6l'l1(NOI'II, Thispar, Nextvar, Transl) :-
556 expand(Nont, Thlspar, Nextvar, Trartsl).
559
560 % add input parameter and output parameter
561 expand(Nont, In_par, Out_par, Call) :-
562 -..(Nont, [Fun | Args]),
g -..(CalI, [Fun, In_par, Out_par | Args]).
565 % - - - error handling
566 two_ok(X,Y) :- nonvarht(X), nonvarht(Y), I.
:3; two_oI<(_, _) :- mlerror(varint).

273
APPENDIX A.3 (Continued)

see ruterrorwtsssaoel :-
570 nl, dBpIay(+++ Error In this rule: '), mes(Messa09). nl,
571 tagiaIl(trarBI_rule(_, _, _)).
572 % diagnostim are only very brief (and not too informative ...)
573 mes(varlnt) :- display(variable or Integer ltem.).
574 mes(ter) :- dIsplay('terrnlnaIs not on a closed iBt.).
575
576 % - - - Initiate grammar processing
577 phrase(Nortt, TennlnaB) :-
576 nonvarht(Nont), I,
579 expand(Nont, TermlnaB, U, Init_caII),
560 calI(InII_caIl).
561 phrase(N,T) :- error(phrase(N,T)).
562
563 ch.tmmy(X, X).
564
$5 % IIOIIOI-IIIOIOIOOIQOIIIIOIQOIIIOOI

586 % IIOIIOIIIIIOIIIIIIIIOOIOIOOIOIOIOIO

567 % I I b r a r y
566 % IOIIIOIIOIOOIOIIIIIOIIIOIOOIIIOIO

569 % IIOOIIOIQOIIIOIIIQOIOI-QOIOOOI-QOIOO

590 % :::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
591 % -.. (readasunIv')
592 % I u I u u I u I I u I I I I I u u O O O u O I O u u I O u u O u u u u u I I u I I u I u u I u u I I I I I u u I I u u I u u u nu
I u I I u u I u I u u I u u u I I u u I u I O u I O u O u I Q u Q u O O u I u I u u I u u u u u I u I u I u u I u I u I u u I II

593 -..(X,Y) :- var(X), var(Y), I, error(-..(X,Y)).

594 -..(Num,[Num1) :- Integer(Num), I.
595 -..(Tenn, [Fun1Args1) :-
596 setarlty(Tenn, Args, N),
597 iunctor(Tenn, Fun, N), % this works both ways
596 not(integer(Fun)), %we don't wanteg 17(X)
599 setargs(Tenn, Args. 0, N). %this works both ways,too
600
601 setarity(Term, Args, N) :- var(Tenn), I, length(Args, N).
602 %notlcethatbadArgsgiveanerrorIn length
603 setarity(_,_,_). %AritywIIIbesetby functor in-..
604
605 % both numeric parameters are given,
606 % the loop stops when the third reaches the fourth
607 %(worIt.sbothwaysbecause a rg does)
606 setargs(_, U, N, N) :- I.
609 setargs(Ten'n, [Arg 1 Args1, K, N) :-
610 sum(l<, 1, K1), arg(I(1, Tert'n, Arg),
611 setargs(Tenn, Args, K1, N).
612
613 % find the length of a closed list; error ll not closed
614 length(LBt, N) :- length(LBt, 0. N).
615
616 % this is a tall-recursive lomtulatlon oi length
617 length(L,_, _) :- var(L), t, error(length(L, _)).
616 Iength(U, N, N) :- I.
619 Iength(L|Llst],l(,N) :-
620 I, sum(K,1,l<1), length(LIst,l<1,N).

274
APPENDIX A.3 (Continued)

621 Iength(Blzarre,_,_) :- error(Ierqth(Bizane,_)).

622
623 %blnd every variable to adBtlnct 'V(N)
624 nunbervats('V(N), N, NextN) :- I, sum(N,1,NextN).
625 nunbervats(V'(_), N, N) :- I.
626 nun'betvars(X, N, N) :- lnteger(X), I.
627 nun'bervats(X, N, NextN) :- nurnbervars(X, 1, N, NextN)
626
629 nurrbervars(X, K, N, NextN) :-
630 arg(K, X, A), I, nurrbervars(A, N, MIdN),
631 sum(K,1,K1), nut'nbetvars(X,K1,MIdN,NextN).
632 nurrbervars(_, _, N, N).
633
634 % ::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
635 % predeiIned'ilx" iunctorsand op
636 % ::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
637 % (ordered according to probable frequency)
636 FF(,,[xiy], 1000).
639 FF(:- , 1xix, ix], 1200).
640 'FF'(';',[xly], 1100).
641 FF(not, [fy], 900).
642 FF'(- ,[xix1, 700).
643 FF(B,[xix1, 700).
644 FF(->,[xix],1200).
645 FF'(+ ,1yIx,ix], 500). FF'(- ,1yix,ix], 500).
646 FF'( ,[yix], 400). 'FF(l ,[yix1, 400).
647 'FF(mod.[xix], 300).
646 FF(< ,[xix], 700). FF(-< , [xfx], 700).
649 FF(:- ,[x1x], 700). FF(:- ,[x1x], 700).
650 'FF(-:-,[xfx], 700). FF'(--,[x1x], 700).
651 FF(@<,[xix], 700). FF(@-<, [xix], 700).
652 FF(@> , [xix], 700). 'FF'(@>-, [xix], 700).
653 FF'(-.., [xix], 700).
654 FF'(--,[xix], 700). 'FF'(--, [xix], 700).
655
656 % thB Inplemerlatlon of op takes care oi redeflnlions
657 % and oi mixed Iunclots
656 op(Prlor, Type, Name) :-
559 8I0"\(N&t'ne). pname(Name, String), noq(String),
660 % noq - see WRITE
661 Integer(Prior), less(0, Prior), less(Prlor, 1201),
662 set_klnd(Type, Kirxi), I,
663 do_op(Prior, Type, Name, Kind).
664 % Ii not all parameters are OK -
665 op(P, T, N) :- error(op( P, T, N )1.
666
667 % set Kind to bin or un
668 set_klnd(Type, bin) :- blnary(Type, _), I.
669 set_klnd(Type, un) :- unary(Type, _, _), I.
670
671 % test lor binary and instantiate Assoc
672 binarylxiy, a(r)). % right associative

2'75
APPENDIX A.3 (Continued)

673 binary(yfx, a(I)). % lelt associative

674 binarylxix, na(_)). % non-associative
675 % test for unary, instantiate Kind and Assoc
676 unary(fy, pre, a(r)). % right associative
677 unary(fx, pre,na(r)). %rlght non-associative
676 unarytyf. post, a(l)). % left associative
679 unary(xI, post, na(I)). % left non-associative
680
681 do_op(P, T, N, Kind) :-
662 FF(N, Oidtypes, Oidprior), I,
663 addii(Oidtypes, Oldprior, P, T, N, Kind).
664 do_op(P, T, N, _) :- assertz(FF'(N, [T1, P1).
665
666 % add or redeline a functor
667 % for rnlxed functors, keep the binary type before the unary
666
669 % the same priority: redefine or make mixed
690 addlf([Oidtype], P, P, T, N, Kind) :-
691 I, set_kind(0ldtype, Oldkind),
692 addff1(0Iclkind, Kind, Oidtype, T, N, P).
693 addlf([OIdtype1, 0ldtype2], P, P, T, N, Kind) :-
694 I, addii2(Kind, Oldtypel, OIdtype2, T, P, N).
695 % otherwise the priorities were diiierent: redefine
696 addil(_, _, P, T, N, _) ;- redel1(N,1T1, P).
697
696 % make a mixed iunctor or change type
699 addii1(bin. un, Oidtype, T, N, P) :- mk_mixed(N, [Oidtype, T1, P).
700 addlf1(un, bln, Oidtype, T, N, P) :- rnk mixed(N, [T, Oldtype], P).
701 addil1(KInd, Kind, _, T, N, P) 1- redef1(N,1T1, P).
702
703 % adjust a mixed functor by changing one ol its types
704 addlf2(bin, _, OIdtype2, T, P, N) 2- t'r|k_mIxed(N, [T, Oldtype2], P)
705 addii2(un, Oldlypet, _, T, P, N) :- rnk_mIxed(N, [Oldlype1, T], P).
706
707 mk_rnixed(N, Types, P) :-
708 retract(FF(N, _, _)), I, assertz(FF'(N, Types. P1).
709
710 % redeline and issue a waming
711 redeI1(N, T, P) :-
712 nl, dispIay(1unctor"), display(N),
713 display( redefined), nl,
714 retract(FF(N, _, _)), I, assetta(FF(N, T, P)).
715
716 % remove adeclaratlon
717 delop(Name) :- atom(Name), retract(FF(Name,_, _)), I.
716 delop(Name) :- enor(delop(Name)).
719
720
721 %IIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIZIIIIIIIIIIIIIIIIII
722 % evaluate an arithmetic expression
723 YoIIIIIIIIIIIIIIIIIIIIIIIIZIIIIIIIIIIIIIIIIIIIIIIII1211211IIIIIIII
724 is(N,N) :- integer(N), I.

276
APPENDIX A.3 (Continued)

725 is(Val, +(A, B)) :-

726 I, Is(Av, A), Is(Bv, B), surn(Av, Bv, Val).
727 is(Val, -(A, B)) :-
72B I, is(Av, A), i5(Bv, B), surn(Bv, Val, Av).
729 is(VaI, '(A, B)) :-
730 I, Is(Av, A), i5(Bv, B), prod(Av, Bv, O, Val).
731 Is(VaI,l(A, B)) :-
732 I, Is(Av, A), Is(Bv, B), prod(Bv, Val, _, Av).
733 is(Val, rnod(A, B) ) :-
734 I, Is(Av, A), ls(Bv, B), prod(Bv,_, Val, Av).
735 is(VaI, +(A)) :- I, Is(Val, A).
736 is(Val, -(A)) :- I, Is(Av, A), sum(Val, Av, O).
737 Is(N, [N]) :- InIeger(N).
738 %otherwise I all
739
740 % - - - - - - EVALUATE AN ARITHMETIC RELATION - - - -
-:-(X, I X), Is(XV,
742 <(X,Y) :- Is(XV, X), Is(YV,Y), Iess(XV,YV).
743 -<(X, Y) :- Is(XV,X), is(YV,Y), noI(Iess(YV,XV)).
744 >(X,Y) :- Is(XV, X), is(YV,Y), Iess(YV, XV).
745 >-(X,Y) :- is(XV,X), i5(YV,Y), not(I0ss(XV,YV)).
746 --(X, Y) 2- not(-:-(X,
747
748 % :::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
749 % perlect equaity oItem1s
750 % :::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
751 --(T1, T2) :- var(T1), var(T2), I, eqvar(T1, T2).
II(T1, I dm(-I?(T1,
753
754 III(T'|,T2) 2- l'lOI(III?(T1,T2)).
755
756 II?(T1,T2) I-
757 Ir|toger(T1), Intoger(T2), I, -(T1, T2).
756 --?(T1,T2) :-
759 nonvarint(T1), nor|varhl(T2),
760 Iunctor(T1, Fun, Arity), Iunclor(T2, Fun, Arity),
761 equaIargs(T1, T2, 1).
762
763 equaIargs(T1, T2, Argnumber) :-
764 arg(Argnurrber, T1, Arg1), arg(Argnurnber, T2, Arg2)
765 % arg IaIIs given too largo a number
766 I, II(A.I'g1, N52). Q-||'|xA|'gl'I.l|'|'&l', 1, NBXI|'I.||'|'bB'l')|
767 equaIargs(T1, T2, Nextnumber).
723 equalargs(_, _, _).
770 % :::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
771 % assert, asserla, assortz, retract, clause
772 %:::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
773 %- - -addaclause(usIng buiit-In assert(_, _, _))
774 asserl(CI) :- assona(Cl).
775 assena(CI) :-
776 nonvarht(CI), c:onverl(Cl, Head, Body), I.

277
APPENDIX A.3 (Continued)

777 sssert(Head, Body, 0).

776 asse|1a(Cl) :- error(asse|1a(Cl)).
779
760 assertz(C|) :-
761 nonvarint(Cl), convert(C|,Head,BOdy). I,
762 sssert(Head, Body, 32767). %Ie 2 to 15th mlnus1
763 assertz(Cl) :- error(assertz(CI)).
764
765 % convert the extemal1ormoIaBodylnto adotted llst
766 convert(:-(Head, B), Head, Body) :- conv_body(B,Body).
767 convert(Unl_cI, UnIt_cl, |]).
766
769 % this procedure works both ways
790 conv_body(B, [caI|(B)]) :- var(B), I.
791 cor|v_body(trus, []).
792 conv_body(B, Body) :- conv_b(B, Body).
793
794 conv_b(B, [Body1) :- var(B), I, conv_calI(B, Body).
795 conv_b(','(C, B), [Call | Bodyl) :-
796 I, conv_cal|(C, Call), conv_b(B, Body).
797 conv_b(CaII, [CaIl]). % not avarisble
796
799 % Interpreter can process varlable calls only within c a I I
600 conv_cal|(C, call(C)) :- var(C), I.
601 conv_caII(C, C).
602
603 % - - - remove a clause (thls procedure Is backtrackable)
604 retract(C|) :-
605 nonvar'nt(Cl), convert(C|, Head, Body), I,
606 lunctor(Head, Fun, Arity), rsmcls(Fun, Arity, 1, Head, Body)
B07 retract(CI) :- error(retract(CI)).
B08
B09 % ultimate Iallure It N too big (retractl3 tails)
810 remcks(Fun, Arlty, N, Head, Body) :-
611 clause(Fun, Arlty, N, N_head, N_body),
612 remcIs(Fun, Arity, N, N_head. Head, N_body, Body).
613
614 remcks(Fun, Arity, N, Head, Head, Body. Body) :-
615 retract(Fun, Anty, N).
616 % user's backtracking resumes r e t r a ct here
617 % (after removing the Nth clause the next becomes Nth)
616 rerncls(Fun, Arlty, N, N_head, Head, N_body, Body) :-
619 checl-t(-(N_hsad, Head)). check(-(N_body, Body)),
620 I, remcls(Fun, Arlty, N, Head, Body).
621 remcIs(Fun, Arlty, N, _, Head, _, Body) :-
622 surn(N, 1, N1), remcIs(Fun, Arity, N1, Head, Body).
623
624 % - - - generate nondeterrnlnlsttcally all clauses whose head
625 % andbodymatchthe parametersot clause
626 c|ause(Head, Body) :-
627 nonvarIr|t(Head), I, Iunctor(Head. Fun, Arity),
626 gencIs(Fun, Arity, 1, Head. Body).

278
APPENDIX AJ (Con ned)

629 clause(Head, Body) :- error(clause(Head, Body)).

630
B31 % generate; ultlmate laibre It N too blg (clausel5 talks)
B32 gencls(Fun, Arity, N, Head. Body) :-
833 clause(Fun, Arlty, N, N_head, N_body),
B34 gencls(Fun. Arlty, N, N_head, Head, N_body, Body).
835
836 % tall It N_head does not match Head.
837 % or ll N_body converted does not match Body
B3-B gencls(_, _, _, N_head, N_head, N_body, Body) :-
B39 conv_body(Body, N_body).
B40 % user's bacldracklng resumes c I a u s e here
641 gencls(Fun, Arlty, N, _, Head, _, Body) :-
B42 surn(N, 1, N1), gencls(Fun, Anty, N1, Head, Body).
B43
644 % :::::::::::::::::::::::::::::::::::::::::::::::::::::::
845 /e Iistlng
646 % I I Q Q Q Q I Q I n I I I G 0 Q Q 0 Q I I O 0 0 I I I I 0 I I O I I I Q I 0 I 0 I O O I I 0 I I I I 0 I I OI
I n a I Q G Q u I Q 0 O I O O I I I I I I I O O O I I I I I I I I I I I I 0 I I I I I I I I I I I I I I I II

B47 % list procedures detemtlned by the parameter ( Ilstlng(_) )

646 % or all user's pnooechtres ( Ilstlng )
B49 listlng :-
650 proc(Head), llstproc(Head), nl, tall.
651 Ilstlng. % catch the llnal tall Irom p r o c
652
653 IIstlng(Fun) :- atom(Fun), I, lIstbyname(Fun).
854 llstIng(I(Fun, ArIty)) :-
sss atom(Fun), Integsr(ArIy), -<(o, Aritv). |,
656 lunctor(Head, Fun, Arlty), lIstproc(Head).
657 IIstIng(L) :-
656 lsclosedIlst(L), llstseveral(L), I.
659 lIstlng(X) :- srror(listIng(X)).
660 % lsclosedlst - cl grammar rule preprocessor
661
662 IisIssveraI([]).
663 llstseveralq Item | ltems]) :-
664 llstlng(ltem), llstsevsral(ltems).
665
666 % all procedures wlththlsname
667 llstbynarne(Fun) :-
666 proc(Head). Iunctor(Hsad, Fun, _),
669 llstproc(Hea:l), nl, tall.
670 llstbyname(_). %succeed
671
672 % one procedure
673 listproc(Head) :-
674 dause(Head, Body),
675 wrlteclause(Head. Body). wch(.), nl, tall.
676 listproc(_). %succeed
677
676 wrIteclause(Head, Body) :-
679 not(var(Body)), -(Body,true), I, wrlteq(Head).
660 writecIause(Head. Body) :- wrlteq(:-(Heal, Body)).

279
APPENDIX AJ (Continued)

661
662 /o IIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIZZIIIIIZIIZIZIIIIIIIIIIII
663 % wrIte
664 /e IIIZIIIIIIIIZIIIIIIIIIIIIIZIZIIZIIIIIZiIIIIiIZIII1211112I
665 write(Tem1) :- side_eI1ects(outterm(Tem1, noq)).
666
667 % writeq encloses In quotes all Identiliers except words,
666 % syrrbols and solochars (not colnclding with 'llx" functors)
669 writeq(Term) :- sIde_el1ects(outterm(Term, q)).
690
691 writetext([Ch | Chsl) :- l, wch(Ch), wrltetext(Chs).
s92 writetext([]).
693
694 outtem1(T, Q) :- nurrbervars(T, 1, _), outt(T, Id(_,_), Q).
695
696 % the real iob ls done here
697 outt('V'(N), _, _) :- lntsger(N), I, wch('X'). dispIay(N).
696 % C A U T I 0 N : cult Is unable to write 'V'(Integer)
699 outt(Tem1, _, _) :- lnteger(Tem1), display(Tem\), I.
900 % the second parameter specltles a context lor 'llx' functors:
901 % the nearest extemal functor and Tem1's position
902 % (to the left or to the right ol the extemal functor)
903 outt(Term, Context, Q) :-
904 -..(Tem1, [Name | Argsl),
905 outlun(Name, Args, Context, Q).
906
907 % - - - output a lunctor-terrn
906 % - asa"lix"tenn
909 out1un(Name, Args, Context, Q) :-
910 kslix(Name, Args,This_l1, Kind), l,
911 outI1(Kind,Thls_I1, [Name | Args], Context, Q).
912 % - asalist
913 out1un(., [Larg, Rarg], _. Q) :-
914 I, outlist([Larg|Flarg],Q).
915 % - asanormaltunctor-tenn
916 out1un(Name, Args, _, Q) :-
917 outname(Name, Q). outargs(Args, Q).
916
919 % isllx constructs a palr l1(Prlor, Assoclatlvlty) , and
920 % In or pro or 'post' (tails ll not a " x" Iunctor)
921 isIix(Narns, L, _], (PrIor, Assoc), In) :-
922 FF(Name, Types, Prior), rnl<_bIn(Types, Assoc).
923 istix(Name, L], (Prior, Assoc), Kind) :-
924 FF(Name, Types, Prior), rnl-t_un('l'ypes,Kind,Assoc).
925
926 % Bintype (it any) ls belore Untype (it any)
927 mk_bin([Bintype I _], Assoc) :- blnary(Birlype, Assoc).
926 ml-t_un([Untype], Kind, Assoc) :- unary(Ur1ype,Klnd,Assoc).
929 mk_un(L,Untype],KInd,Assoc) :- unary(Untype,KInd,Assoc)
930 % tests - see o p
931
932 % - - - output a "lix' tem1 (thls out has 5 parameters)

260
APPENDIX A.3 (Colltilled)

933 out1I(Klnd, This_I1, NameArgs, Context, O) :-

934 agree(ThIs_tI, Context), I,
935 outI1(l-(Ind, Thls_t1, NameArgs, Q).
936 out (Klnd, Thls_t1, NameArgs, _, Q) :-
937 wch('('), out11(KInd, Thls_t1, NameArgs, Q), wch(')').
938
939 % agree helps avoid (some) unnecessary brackets around the tem-I
940 agree(_, Id(Ext_lI, _)) :- var(Ext_I1).
941 agree(t1(Prior1, _), ld(t1(Prior2, _), _)) :-
942 stronger(Prior1, Prlor2). % cl the parser
943 agres(t1(PrIor, a(Dir)),1d(11(PrIor, a(Dir)), DIr)).
944
945 % output the functor and the arguments (this outlt has 4 parameters)
946 out11(in, This_l1, [Name, Larg, Rarg], Q) :-
947 outt(Larg, Id(This_I1, I), Q),
946 outln(Name, ' '). outt(FIarg, Id(This_I1, r), Q).
949 outl1(pre,Thls_t1, [Name, Arg], Q) :-
950 outIn(Name, ' '). outt(Arg, ld(ThIs_I1, n, Q).
951 outl1(post, TI1Is_t1,[Nams.N9]. Q) :-
ssz outt(Arg,Id(ThIs_I1, I), 0). outIn(Name, ' ').
953
954 %outputIunctor's name encIosedInEncl
955 outtn(Name, Encl) :- wch(EncI), dIsplay(Name), wch(EncI).
956
957 % - - - print a name (In quotes, ll necessary)
956 outname(Name, noq) :- I, display(Name).
959 outname(Name, q) :-
960 'FF'(Name,_,_), I, out1n(Name,
961 outname(Name, q) :-
962 pname(Name, Namestrlng).
963 check(noq(NamestrIng)), I, display(Name).
964 outname(Name,q) :- outln(Name,"").
965
966 noq([Ch |StrIng1) :- wordstart(Ch), lsword(StrIng).
967 noq([Cl'l]) :- solochar(Ch).
969 00qI['I'- I'D-
969 noq([Ch |String1) :- symch(Ch), Issym(String).
970
971 iavord(|]).
972 isword([Ch | Strlng]) :- abhanum(Ch), Isword(Strlng).
973 issyn1([]).
974 lssym([Ch | Strlngl) :- symch(Ch), lssym(Strlng).
975
976 % - - - output a list ol argumerls (cl outtun)
977 _) I" I-

976 outargs(Args, Q) :-
979 lake(Context), wch('('), outargs(Args, Context, Q), wch(')').
960
961 outargs([Last]. Context. Q) :- I. outt(Last. Context. Q).
962 outargs([Arg|Args], Context. Q) :-
963 outt(Arg, Context, O), dlsplay(', '), outargs(Args, Context, Q).
964
APPENDIX A.3 (Continned)

985 % commas are used to delimit list Items, so we must bracket commas
986 % w I t h I n ltens (it's a trick: we deperd on having
987 % the priority 1000 and being associative)
988 Iake(Id(I1(1000, na(_)), _)).
989
990 % - - - output a list In square brackets (ci outlun - the main
991 % Iunctor is the dot, and the list cannot be empty)
992 outlist([F|rst | Tail], Q) :-
993 Iake(Context), wch('['), outt(Flrst, Context, Q),
994 outIist(Tail, Context, Q), wch(')').
995
996 outIist([], _, _) :- I.
997 outIist([Item | Items], Context, Q) :-
998 I, display(, '), outt(item, Context, Q),
999 outlist(Iten's, Context, O).
1000 % the bar and the closing Item (still bracketed Ii It contains commas)
1001 outIist(CIos'ng, Context, Q) :-
1002 display(' I '). outt(CIosing, Context, 0).
1003
%IOIIOQIIIIOIQOOIQIQQOQIIOQOIIQOOIOQ
1004
% OOOIIIQ-iiQIIQQOIQIIIIIQOIIOQQIOIIOI
1005
1006 % tr a n s I at o r
% IIIOQIOIIIQQOIIOIIOIOIII-IIOIIII-Oil!-I
1007
% IIIOQIOI-liliiliiliIIIIOOOQOOOOIIOOOQO
1008
1009 % read aprogram upto end. and translate it Into "kemel"lom1
1010 translate(Inlile, Out le) :-
1011 see(lnlile), telI(Outlile),
1012 nl, repeat,
1013 read(Cla:se), pul(Clause), nl, -(Clause, end), I.
1014 seen, told, see(user), telI(user).
1015
1016 % - - - produce and output the translation oi one clause
1017 put(:-(Head, Body)) :-
1018 I, puthead(Head, Sym_tab), putbody(Body, Sym_tab).
1019 put(-->(Lelt, FIight)) :-
1020 I, tag(transI_ruIs(Lelt, Right, :-(Head, Body).
1021 puthead(Head, Sym_tab), putbody(Body, Sym_tab).
1022 put(:-(Goal)) :-
1023 I, putbody(GoaI, Sym_tab), wch(l), nl.
1024 once(Goal). % a iailure here wouldn't matter (cl translate)
1025 put(snd) :- I.
1026 put('err) :- I.
1027 put(UnitcIause) :- puthead(Unltclause, Sym_tab), putbody(true, _).
1028
1029 % - - - put a head call (it nus! be a Iunctor-temt)
1030 puthead(Head, Sym_tab) :-
1031 nonvar'nt(Head), I, puttem1(Head, Sym_tab).
1032 puthead(Head, _) :- transl_err(Head).
1033
1034 %---putalistoIcallsand|]attheend
1035 putbodg/(Body, Sym_tab) :-
1036 punct(:), conv_body(Body, B), I, putbody_c(B, Sym_tab).

282
APPENDIX A.3 (Continued)

1037 %seeassertetcIor conv_body

1038
1039 pu1b0dv_l[l- J 1- L di=-Pl=wlI]l-
1040 putbody_c([Tem1 | Terms], Sym_tab) :-
1041 not(integer(Tenn)), I, puttem1(Term, Sym_tab),
1042 punct(.), putbody_c(Tem1s, Sym_tab).
1043 putbody_c([Tem1|_], _) :- transl_err(Tem1).
1044
1045 pUIlCi(Ch) 2- wch("), wch(Ch), nl, display( ').
1046
1047 % - - - put a tenn (with in x dots, and canonical otherwise)
1048 puttem1(Term,Sym_tab) :-
1049 var(Tem1), I, Iookup(Tem1, Sym_tab, -1, N),
1050 wch(:), dlsplay(N).
1051 putterm(Term, _) :- lnteger(Tem1), I. display(Terrn).
1052 puttem1([Head | Tail], Sym_tab) :-
1053 I, puttenn_InIist(Head, Sym_tab).
1054 display( . '), puttem1(TaiI, Sym_tab).
1055 puttem1(Term, Sym_tab) :-
1056 -..(Tem1,[Name|Args]), outln(Name.""), %clWFIITE
1057 putargs(Args, Sym_tab).
1058
1059 % Sym_tab Is an open Ibt cl palrs vn(VarIabIe, Number)
1060 % (this Iomtulation helps avoid too many additions)
1061 lookup(V, S_t_end, PrevlousN, N) :-
1062 var(S_t_end), I, sum(PrevIousN, 1, N),
1063 -(S_t_end, [vn(V, N) | New_s_t_end]).
1064 lookup(V, [vn(CurrV, CurrN) | _], _, CurrN) :-
1065 eqvar(V, CurrV), I.
1066 lookup(V, [vn(_, CurrN) | S_t_tail], _, N) :-
1067 lookup(V, S_t_tail, CurrN, N).
1068
1069 % arguments - nothing, or a list ol terms in parentheses
1070 putargs([|, _) :- I.
1071 putargs(Args, Sym_tab) :-
1072 wcl1('('), putarglist(Args, Sym_tab), wch(')').
1073
1074 putarglist([Arg]. Sym_tab) :- I, puttem1(Arg, Sym_tab).
1075 putarglist([Arg | Args], Sym_tab) :-
1076 puttem1(Arg, Sym_tab), display(, '),
1077 putarglist(Args, Sym_tab).
1078
1079 % - - - a list within a list nust be enclosed In parentheses
1080 puttenn_lnlist(Tenn, Sym_tab) :-
1081 nonvarlnt(Term), -(Term, L | _)), I,
1082 wch('('), putterm(Tem1, Sym_tab), wch(')').
1083 puttern1_inIist(Term, Sym_tab) :- putterm(Term, Sym_tab).
1084
1085 % - - - error handing (only one error ks discovered by translate)
1086 lransl_err(X) :-
1087 nl, display(-+++ Bad head or call: '), display/(X), nl, Iail.
1088
1089 :- see( user ), ear.
263
APPENDIX A.4
Three Useful Programs

A simple ctlltor

% A slrrpie Interactive clause editor.

% Watch tor name con icts with its procedures I
% Note that this version has no saieguards agaimt Proiog's crash
% (eg. due to stack overllow).
% Cali edit( name/arity ) to edit the procedure oi this name and arity.
% Each invocation oi edit is associated with a cursor, which is the number
% oi a clause. Initially the cursor is at clause 0, i.e. belore the rst
% clause In this procedure. The cursors value and its associated clause
% is usually displayed between commands.
% Commands are listed below. Temtlnate the line immediately alter typing
% last character. Don't use blanks where not shown and only one where shown
%
% Commands :
%
%
% e Name/Arity - Invoke a nested instance to edit another procedure.
% The current cursor stays in place unless you happen
% to modlly this procedure within a nested Instance.
%x - exit Irom the current editor instance.
%+ - move the cursor to the next clause, no action it none.
% <cr> - an empty line is an altemative lorrn oi +.
%- - move the cursor to the previous clause, no action ii at 0.
%t - top : move the cursor to 0.
%b - bottom : move the cursor to the bottom clause
% (0 lor empty procedures).
%I - list the whole procedure.
%d - delete the current clause and move the cursor to
% the next (or to the new bottom ll bottom is deleted).
%i - lnsen alter the current clause. In the following
% lines write clauses as you would alter consuit(user)
% (tenninate the sequence with end.).The cursor is
% positioned at the last inserted clause.
% I Fiiename - like I, but read the clauses from a iile.
% Take care I lename correctness is not checked.
/9 p - invoke a nested instance oi Prolog. II there is
% no memory overliow, invoking stop will retum
% control to the editor.
%

edit( Nams!Arity) :- not ( atom( Name ), integer( Arity) ), I,

write( Bad parameters : ' ),
write( edit( NameIAnty) ), nl, iail.
edit( Name/Arity) :- predelined( Name, Arity), I,
write( 'Can"t edit system routine : ' ),
write( NamelArlty), nl, Iaii.
edit( NameArity ):- tag( ed( NameArity, 0 ) ).

284
APPENDIX A.4 (Continned)

ed( NameArity, Cursor) :- show( NameArity, Cursor), I.

docmd( NameArity, Cursor, NewCursor ).
ed( NameArity, NewCursor ).
ed( NameArity, Cursor) :- display( 'Cu|sorout oi range : ' ).
display( Cursor ), nl, ed( NameArity, 0 ).
docmd( NameArity, Cursor, NewCursor) :-
repeat, % repeat over incorrect commands
getline( Une ), cmd( Line, NameArity, Cursor, NewCursor ),
I.
getIine([]) :- rch, iastch(C), kseoln(C), I.
getIine([C | L]) :- iastch( C ), getline( L).
% cmd tails tor incorrect commands.
cmd( [1, NmAr, Cur, NCur) :- next_cursor( NmAr, Cur, NCur ).
cmd( ['+'], NmAr, Cur, NCur) :- next_cursor( NmAr, Cur, NCur ).
cmd( ['-'], _, Cur, NCur) :- prev_cursor( Cur, NCur ).
cmd( [I], NmAr, _, 0 ).
cmd( lb], NmAr, Cur, NCur) :- bottom_cursor( NmAr, Cur, NCur ).
cmd( [I], NmAr, Cur, Cur ) :- iistlng( NmAr).
cmd( [d], NmAr, Cur, NCur) :- deIete( NmAr, Cur, NCur) .
cmd( [I], NmAr, Cur, NCur) :- insert( NmAr, Cur, NCur ).
cmd( [i,' ' | NameStrlng], NmAr, Cur, NCur) :-
le_lnse|1( NameString, NmAr, Cur, NCur ).
cmd( ls, | Args], NmAr, Cur, Cur) :-
append( NameString, ['1' | ArityStrlng], Args ),
caIl_edlt( NameString, ArityStrlng ).
Cindi III. _. _. _i 1- it-IOBXIII Bdi _. _ I )-
cmd( [p], NmAr, Cur, Cur) :- invoke_Proiog.
cmd(Str1|-|g_ __ __ _) :- display( '--incorrect command : ' ),
wrItetext( String), nl, lail.

% check is provided with the standard library ( check(C) :- not not C )

next_cursor( Name/Arity, Cursor, Next ) :-
Next is Cursor + 1, check( ciause( Name, Arity, Next, _, _) ), I.
next_cursor( _, Cursor, Cursor). % cursor at last clause
prev_cursor( 0, 0 ).
prev_cursor( Cursor, Prev ) :- Cursor > 0, Prev is Cursor - 1.
bottom_cursor( NamelArlty, Cursor, Bottom) :-
Next is Cursor + 1, check( ciause( Name, Arity, Next, _, _ )),
I, bottom_cursor( NamelArity, Next, Bottom).
bottom_cursor( _, Cursor, Cursor ).

deIete( _, 0, 0) :- I, display( 'Can"t delete clause 0' ), nl.

delete( NameIArlty, Cursor, NewCursor) :-
retract( Name, Anty, Cursor).
cursor_ln_range( Name, Arity, Cursor, NewCursor ).
cursor_ln_range( Nm, Ar, Cur, Cur ) :-
check( clause( Nm, Ar, Cur, _, _ )), I.
APPENDIX A.4 (Continned)

cursor_ln_range( _. _. Cur. Prev) :- Prev is Cur - 1.

% conven is dellned in the standard library

insert( NameArity, Cursor, NewCursor) :-
repeat, % get end. or a clause oi NamelArlty, skb others
read( Clause), convert( Clause, Head, Body ),
accept( Head, NameArity, Clause ),
I.
end_or_proceed( Head, Body, NameArity, Cursor, NewCursor ).
end_or_proceed( end, []. _, Cursor, Cursor) :- I.
end_or_proceed( Head, Body, NameArity, Cursor, NewCursor) :-
Next is Cursor + 1, assert( Heal, Body, Cursor),
insert( NameArity, Next, NewCursor ).
accept( _. _. end ).
accept( Head, NamelArity, _) :- lunctor( Head, Name, Arity ).
accept( _, _, Clause) :-
dlspiay( '---clause not in edited procedure - ignored ),
nl, write( Clause), iail.

Iiie_insert( FNameStrlng, NameArity, Cursor, NewCursor) :-

pname( FiIeName, FNa|'neStrlng ),
see( FiIeName ), insert( NameArity, Cursor, NewCursor ),
seen, see( user).

cail_edit( NameString, ArityStrlng ) :-

pname( Name, NameString ), pnamei( Arity, ArityStrlng ),
edit( Name/Arlty ).

invoke_Prolog :- tag( loop ). % this works only tor the Toy-Prolog monitor
invoke_Proiog. %( loop temtinated by tagiall )

% conv_body is de ned in the standard library (asserta etc.),

% so is writeclause.
show(NameArity,0) :- I, write( '[0]('), write( NameArity),
rit BI I I. nl.
show( Name/Arity:~Cursor) :-
side_ei1ects( ( clause( Name, Anty, Cursor, Head, Body ),
conv_body( NIceBody, Body ),
display( I I. display( Cursor I.
display( '1 ' I.
writeclause( Head, NIceBody ),
display( ), nl) ).

rl:l([],L,L).
:3I:MiEILi-L2.IEI2II=-aPPBI1dil-.L2.l-l-2I-
286
APPENDIX A.4 (Continned)

Aprhnltiveu-ndngtool

% A primitive tracing package.

% Watch tor name con icts with its procedures I
% Use spy( Pattem ) to trace calls matching Pattem,
% nospy( Pattem ) to stop tracing.
% To trace, execute trace( Goal ) Instead oi Goal.
% Successful calls are displayed with a plus, tailing calls with a rnlnus.
% Note: tagcut, tagexlt, tagiall and ancestor will not be executed properly
% tracelsslow: ilyouwlshtohevethelnsldesolacorrectand
% costly procedure executed at nomtal speed, add
% a prede ned(...) assertion tor its call.
spy(Ail) :- var( Ail), I, assert(spied(All)).
spy( Pattem) :- spied( Pattem ), I. % spied already
spy( Pattem) :- asse|1(spled( Pattem ) ).
nospy( Pattem) :- retract(spied( Pattem)), tall.
"$PY( _ I-
trace( Goal) :- tag( runbody( Goal ) ).

runbody((A,B)):- I, nunbody(A), nunbody(B).

runbdvlIA:BII=- I. irvnbOdviAI:runb0dyiBII-
runbody( caIl( Call ) ) :-
var( Call), I, showiaiiure(call( Call) ), tail.
runbody(calI(CalI)) :- I, nunbody(CalI).
wrvbvdvltaslwlll :- I . runbody( callI Call II-
runbody( Cali) :- predelIned( Call), I, runsystem( Call).
runbody( Cali) :- tag( runuser( Call ) ).

runsystem( I ) :- runcut.
runsystem( Forbidden ) :-
lslo|bidden(Forbldden), I, nl,
display( FORBIDDEN CALL ' ), write( Forbidden ),
display( ' FAILS I), nl, tall.
runsystern( Call) :- not spIed( Call ), I, Cali.
runsystem( Call) :- Call, I, showsuccess(CaIl).
runsystern( Call) :- showlalhure( Call ), tall.

runuser( Call) :- not spied( Call), I,

clause(Cali,Body), runbody(Body).
nunuser( Call) :- clause( Call, Body ), showsuccess( Cali ),
runbody( Body I-
nunuser( Call) :- showialltxe( Call ), tall.
runcut :- spied(i), simuiatecut, showsuocess(l).
runcut :- simuiatecut.
sinluiatecut :- tagcut( nunuser( _ ) ).
sinluiatecut :- tagcut( nurbody( _ ) ). % cut In inltlaigoai

287
APPENDIX A.4 (Continned)

showsuccess(CaiI) :- dispiay('+'), write(CalI), nl.

showiaIIure( Call) :- display( ' -'), write(CaII), nl.
isIorbidden( tagexit( II-
isiolbIdden( taglaiI( _' i

lslorblddenl tagcut( i ii
n

isiorbidden( ancesto?( _ ) )
predeiined( Call ) :-
Cli0Cl<( ( 1t.lI'tCt0t'( Call F N ) prede md( F N ) ) )
APPENDIX A.4 (Continued)

Aprognmsu'ucturennnlyserwIthnnnlyse|-analysed

% Given a procedure name and arity, print its call tree.

% The main data structure is a queue oi procedures whose tall contains
% calls which were not yet seen. Each elemert oi the queue contains a list
% oi calls (references to main queue elemeris) and a variable to hold its
% ordinal nurrber in the listed tree.
% Queues are searched linearly : the algorithm is costly Ior large trees.
% CAUTION : don't atterrpt to list a trace oi this program - cyclic structures
% are formed as a rule.

calitree( NamelArity) :- add( proc( Name, Arlty, Ord, Calls), Queue),

lill( Queue, Queue ),
prlnt_caiis([proc(Name.Arity.0rd,Cails)].3,1 ,_).

% add iinds (inserts) an element in ( to ) an open list

add( El, [Ei|Tall]) :- I.
add( El, L | Tail] ) :- add( El, Tail ).
% iill walks the queue and expands procedures, lnsenlng their calls Into
% the queue ll not yet seen. Queue begimlng ls passed along to allow search.
iiII( []. _) :- I. % eviderlly reached the tennlnatlng variable
IiII( [proc(Name,Arlty,_,|])|QTalI]. Q) :- prede ned( NameIArlty ), I,
Iiil( QTaii, Q ).
liII( [proc(Name,Arlty,_,undeIIned)|QTail], Q ) :-
not clause( Name. Arity, 1, _, _ ), I, ( QTaiI, Q ).
IiiI( [proc(Name,ArIty,_,Calls)iQTail], Q) :-
add_cals( Name. Arity, 1, Calls. Q ), liII( QTail, Q ).
% system procedures and procedures deiined In the monitor should not be shown
predelined( Name I Arity) :- predeIined( Name, Arity ).
% only the more commonly used procedures (but the list is easily extended)
predeiined( not I 1 ). prede ned( nil 0 ).
predellned( read I 1 ). prede ned( '-..' I 2 ).
predeiined( op I 3 ). predeilmd( is I 2 ).
predeIinsd( assen I 1 ). prede ned( assert: I 1 ).
prede ned( retract I 1 ). prede ned( clause I 2 ).
prede ned( write I 1 ).prede ned( writeq I 1).
% add_cals processes the clauses oi a procedure. adding calls to its list
% oi calls and to the queue (only nding In the queue ii already there)
add_caiIs( Name. Arity, N, Calls, Q ) :-
clause( Name, Arlty, N, _, Body ), I,
body_caIis( Body, Calls, Q ), N1 is N + 1.
add_cals( Name, Arity, N1, Calls, Q ).
add_caiIs( _. _. _, |], _) :- I. % close the list it empty
% ( only unit clauses )
add_calls( _. _, _, _, _ ). % non-empty list lelt open
body_calls( [], _, _ ) :- I.

289
APPENDIX A.4 (Continued)

body_cals( [Call | BodyTaIl], Calls, Q ) :-

lunctor( Call, Name, Arity ),
add( proc(Name,Arlty,Ord,CaIlees), Calls I.
add( proc(Name,Arlty,Ord,Callees), Q I.
add_insides( Call, Calls, Q I.
body_cals( BodyTall, Calls, Q ).
% add_insides unpacks melaioglcal calls: ll their arguments are not variable
% or integer, they are added to the queues.
add_lrsides( Call. Q1, Q2 ) :- meta_call_1( Cali, Arg ), I,
add_inside( Arg, Q1, Q2 ).
add_insldes( Call, Q1, Q2 ) :- meta_call_2( Call, Arg1, Arg2 ), I,
add_inside( Arg1, Q1, Q2 ),
add_inside( Arg2. Q1, Q2 ).
add_irsides( _, _, _ ).
add_inside( V, _, _) :- (var( V ) ;lnteger( V ) ), I.
add_inside( Cali, Q1, Q2) :- lunctor( Cali, Name, Arity ),
add( proc(Name,Arity,Ord.Callees), Q1 ),
add( proc(Name,Anty,Ord.Caiiees), Q2 I.
add_lnsides( Call, Q1, Q2).

meta_call_1( caIl( Call ), Cali ).

meta_call_1( tag( Call ), Ca ).
meta_call_1( not Call, Call ).
meta_call_1( check( Cali ), Call ).
meta_call_1( slde_ellects( Cali ), Cali ).
meta_call_1( once( Ca ), Call ).
meta_calI_2(
meta_caii_2( P) UJID }>_> I IOJ
% Print calls, staning at given tab setting and ordinal, retuming next ordinal
% number. Third clause tails ii ordinal numbers don't match, i.e. proc
% was already printed In another line.
prinI_0a|ls( I]. _, Ord. Ord) :- I. % this matches the temtinating var
% oi a ca list.
print_caIks( [proc(Name,Arity,0rd,undellned)|Calks], Tab, Ord, NOrd ) :-
I, start_undeiined( Ord, Tab),
writeq( NamelArlty ), display( unde ned ), nl,
TOrd is Ord + 1, print_calks( Calls, Tab, TOrd, NOrd ).
print_caIis( [proc(Name,Arlty,Ord,Callees)|Cdls], Tab. Ord, NOrd) :-
I, slart_llne( Ord, Tab), writeq( NamelArlty ), nl,
InnerTab is Tab + 3, InnerOrd is Ord + 1,
prInt_calls( Cailees, InnerTab, InnerOrd, TOrd ),
prlnt_calis( Calhs, Tab. TOrd, NOrd ).
print_caiks( [proc(Name,Arity,AnotherOrd,_)|Cails], Tab, Ord, NOrd) :-
start_unnumbered__Ilne( Tab ), writeq( NameIArity ),
repetitioni Name, Arlty, AnotherOrd ), nl,
print_caIls( Calls, Tab, Ord, NOrd ).
APPENDIX A.4 (Continued)

repetitlon( Name. Arlty, _) :- prede ned( Name Arity) I

repetitlon( _, _, Ord) :- display( ' (see ' ), display( Ord)
display( I I-
% Ord numbers are printed In 4 columns, right lustilled
start_llne( Ord, Tab) :- nun'ber_lne( Ord) I tab( Tab )
number_lne( N ) :- N < 10, display( ' ), display( N)
number_llne( N ) :- N < 100, display( '). display( N)
number_lne( N ) :- N < 1000, display( ' ' ), display( N)
number_Iine( N ) :- display( N ).

start_unnuni:ered_Iine( Tab) :- display( ) tab( Tab ' )

start_unde ned( Ord, Tab) :- nun-ber_Ine( Ord) tab( Tab I
tab( 0,_): -I .
tab(N,Ch):- wch(Ch),N1isN-1,tab(N1Ch)
7 %
% a sanple call and results
:- calltree( calltree I 1 ).
1 calltree/1
2 add/2
3 I/0
addI2 (see 2)
4 IiIlI2
II0
5 prede ned I 1
6 prede ned I 2
liill 2 (see 4)
7 'not' I 1
8 clause I 5
9 add_cals I 5
clause I 5
II0
10 body_cals/3
no
11 Iunctor! 3
add I 2 (see 2)
12 add_lnsldes/ 3
13 meta_caII_1 I 2
II0
14 add_irslde I 3
15 ';' I 2
16 call! 1
17 var! 1
18 integer! 1
II0
Iunctor I 3
add! 2 (see 2)
add_lnsides/ 3 (see 12)
APPENDIX A.4 (Continued)

19 meta_cail_2 I 3
body_caiis I 3 (see 10)
20 is I 2
add_cals I 5 (see 9)
21 pn'nt_cails I 4
II0
22 sta|1_undeiIned I 2
23 nurrber_iine/ 1
24 '<' I 2
is I 2 (see 20)
25 less I 2
26 display! 1
27 tab I 2
II0
26 wch I 1
'is'/ 2 (see 20)
tab I 2 (see 27)
29 writeq I 1
display I 1
30 nl I 0
'is' I 2 (see 20)
print_caiB I 4 (see 21)
31 start_line I 2
nurrber_|ineI 1 (see 23)
II0
tab I 2 (see 27)
32 start_unnumbered_iine/ 1
display! 1
tab I 2 (see 27)
33 repetition I 3
predeiinsd I 2
II0
display! 1
REFERENCES

Aho. A. V. and Ullman. J. D. (I977). Principles of Compiler Design. Addison-Wesley.

Reading. Massachusetts.
Astrahan. A. M. (I976). System R: Relational Approach to Database Management. ACM
Trans. Data Base Systems l(2). pp. 97-I37.
Ballieu, G. (I983). A Virtual Machine to Implement Prolog. In Pereira er al. I983. pp. 40-52.
Battani. G. and Meloni, H. (I973). Interpreteur du langage de programmation PROLOG.
Groupe d'Intelligence Arti cielle. Universite dAix-Marseille.
Battani. G. and Mloni. H. (I975). Mise en oeuvre des constraintes phonologiques, syntax-
iques et semantiques dans un systeme de comprehension automatique de la parole.
Ph.D. thesis. Universite dAix-Marseille.
Bendl. J .. Koves. P. and Szeredi, P. (I980). The MPROLOG System. In T mlund I980. pp.
20l-209.
Bergman, M. and Kanoui, H. (I973). Application of Mechanical Theorem-Proving to Sym-
bolic Calculus. In Proceedings of the 3rd Colloquium on Advanced Computing Meth-
ods in Theoretical Physics." Marseille.
Bergman, M. and Kanoui, H. (I975). Sycophante: Systeme de calcul formel et dintegration
symbolique sur ordinateur. Groupe dIntelligence Arti cielle. Universit dAix-Mar-
seille.
Bobrow, D. G. and Wegbreit. B. (I973). A Model and Stack Implementation of Multiple
Environments. Commun. ACM l6(l0). 59l-603.
Bowen. D. L. (I98I). DECSystem-I0 Prolog User's Manual. Department of Arti cial Intelli-
gence. University of Edinburgh.
Bowen. D.. Byrd. L. and Clocksin. W. (I983). A Ponable Prolog Compiler. In Pereira er al.
I983. pp. 74-83.
Boyer, R. S. and Moore. J. S. (I972). The Sharing of Structure in Theorem Proving Pro-
grams. In Machine Intelligence 7" (B. Meltzer and D. Michie. eds.). pp. l0l-II6.
Edinburgh University Press.
Bruynooghe. M. (I976). An Interpreter for Predicate Programs: Part I. Report CWI6.
Katholieke Universiteit Leuven. g
Bruynooghe, M. (I978). Intelligent Backtracking for an Interpreter of Hom Clause Logic
Programs. Report CWl6. Katholieke Universiteit Leuven. Also in "Mathematical
Logic in Computer Science" (B. Domolki and T. Gergely. eds.). pp. 2I5-258. North-
Holland Publ.. Amsterdam.
Bruynooghe. M. (l982a). Adding Redundancy to Obtain More Reliable and More Readable
Prolog Programs. In Van Caneghem l982a. pp. 52-55.
293
294 References

Bruynooghe. M. (l982b). The Memory Management of PROLOG Implementations. In Clark

and Tamlund I982. pp. 83-98.
Bmynooghe. M. and Pereira. L. M. (l98l). Revison of Top-Down Logical Reasoning
through Intelligent Backtracking. Report CIUNL-8/8|. Universidade Nova de Lis-
boa.
Bundy. A.. ed. (I983). Proceedings of the 8th Intemational Joint Conference on Arti cial
Intelligence". 8-I2 August I983. Karslruhe. William Kaufmann. Inc.. Los Altos. Cali-
fomia.
Burstall. R. M. and Darlington. J. (I977). Transformation for Developing Recursive Pro-
grams. J. ACM 24(I). 44-67.
Campbell. J. A.. ed. (I984). Implementations of PROLOG. Ellis Horwood Ltd.. Chich-
ester.
Chamberlin. D. D.. Astrahan. M. M.. Eswaran. K. P.. Grif th. P. P.. Lorie. R. A.. Mehl.
J. W.. Reisner. P. and Wade. B. W. (I976). SEQUEL2: A Uni ed Approach to Data
De nition. Manipulation and Control. IBM. Res. Dev. 20(6). pp. 560-575.
Chomicki. J. and Gmdzinski. W. (I983). A Database Support System for Prolog. In Pereira
et al. I983. pp. 290-303.
Clark. K. L. (I978). Negation as Failure. In Gallaire and Minker I978. PP. 293-322.
Clark. K. L. and Gregory. S. (I983). PARLOG: A Parallel Logic Programming Language.
Research Report DOC 83/5. Imperial College. London.
Clark, K. L. and McCabe. F. G. (l980a). IC-PROLOGAspects of Its Implementation. In
Tamlund I980. pp. I90-I97.
Clark, K. L. and McCabe. F. G. (I980b). IC-PROLOGLa.nguage Features. In Tamlund
I980. pp. 45-52.
Clark. K. L. and McCabe. F. G. (I984). micro-PROLOG. Programming in Logic". Pren-
tice-Hall. Englevvood Cliffs. New Jersey.
Clark. K. L. and Tamlund. S.-A. (I977). A First Order Theory of Data and Programs. In
Information Processing 77" (B. Gilchrist. ed.). pp. 939-944. North-Holland. Am-
sterdam.
cum. K. L. and Tamlund. s.-A.. eds. (I982). Logic Programming". Academic Press.
New York and London.
Clark. K. L.. McCabe. F. G. and Gregory. S. (I979). The Control Facilities of IC-PROLOG.
In Expert Systems in Micro-Electronic Age" (D. Michie. ed.). pp. I29-I49. Edin-
burgh University Press.
Clark. K. L.. Ennals. J. R. and McCabe. F. G. (I982a). A micro-PROLOG Primer". Logic
Programming Associates Ltd.. London.
Clark. K. L.. McCabe. F. G. and Gregory. S. (l982b). IC-PROLOGLanguage Features.
In Clark and T rnlund I982. pp. 253-266.
Clocksin. W. F. and Mellish. C. S. (l98l). Programming in Prolog. Springer-Verlag.
Berlin and Heidelberg.
Codd. E. F. (I970). A Relational Model of Data for Large Shared Data Banks. Commun.
ACM 13(6). 377-387.
Codd. E. F. (I979). Extending the Relational Data Base Model to Capture More Meaning.
ACM Trans. Data Base Systems 4(4). pp. 397-434.
Coelho. H. (I982). Man-Machine Communication in PortugueseA Friendly Library Ser-
vice System. Information Systems 7(2). I63-l8l.
Coelho. H.. Cotta. J. C. and Pereira. L. M. (I980). How to Solve It in Prolog. Laboratorio
Nacional dc Engenharia Civil. Lisboa.
Colmerauer. A. (I975). Le grammaires de metamorphose. Groupe dIntelligence Arti cielle.
Universite d'Aix-Marseille.
 References 295

Colmerauer. A. (I978). Metamorphosis Grammars. In "Natural Language Communica-

tion with Computer" (L. Bolc. ed.). pp. I33-I89. Springer-Verlag. Berlin and Heidel-
berg.
Colmerauer. A. (I979). Prolog and In nite Trees. Groupe d'InteIligence Arti cielle. Univer-
site dAix-Marseille. Also in Clark and Tamlund I982. pp. 45-66.
Colmerauer. A. (I982). PROLOG II. Manuel de reference et modele theorique. Groupe
d'Intelligence Arti cielle. Universite Marseille II.
Colmerauer. A. (I983). Prolog in Ten Figures. In Bundy I983. pp. 487-499.
Colmerauer. A.. Kanoui. H.. Roussel. P. and Pasero. R. (I972). Un systeme de conununica-
tion homme-machine en franqais. Rapport preliminaire. Groupe d'lntelligence A|1i -
cielle. Universite dAix-Marseille.
Colmerauer. A.. Kanoui. H.. Roussel. P. and Pasero. R. (I973). Un systeme de communica-
tion homme-machine en francais. Rapport de rechereche sur le contrat CR1 no 72-I8 de
fevrier 72 a juin 73. Groupe d'Intelligence Arti cielle. Universite dAix-Marseille.
Colmerauer. A.. Kanoui. H. and Van Caneghem. M. (I979). Etude et realisation d'un
systeme Prolog. Groupe dIntelligence Arti cielle. Universite dAix-Marseille.
Colmerauer. A.. Kanoui. H. and Van Caneghem. M. (l98I). Last Steps toward an Ultimate
Prolog. In Proceedings of the 7th Intemational Joint Conference on Arti cial Intelli-
gence" (R. Schank. ed.). Vancouver. Canada. pp. 947-948.
Colmerauer. A.. Kanoui. H. and Van Caneghem. M. (I983). Prolog. bases theoriques et
developpements actuels. Techniques et Science Inforrnatiques 2(4). 27l-3| l. English
translation in Technology and Science of Informatics 2(4).
Conery. J. S. and Kibler. D. F. (I983). AND Parallelism in Logic Programs. In Bundy I983.
pp. 539-543.
Dahl. V. (I977). Un systeme deductif dinterrogation de banques de donnees en espagnol.
Ph.D. thesis. Universite dAix-Marseille.
Dahl. V. (I980). Two Solutions for the Negation Problem. In Tamlund I980. pp. 6|-72.
Date. C. J. (I982). An Introduction to Database Systems. 3rd. ed.. Vol. I. Addison-
Wesley. Reading. Massachusetts.
Dijkstra. E. W. (I975). Guarded Commands. Nondeterrninacy and Fonnal Derivation of
Programs. Commun. ACM I8. 453-457.
Donz. P. (I979). Une methode de transformation et doptimisation de programmes Prolog:
de nition et implementation. Ph.D. thesis. Universite dAix-Marseille.
Eisinger. N.. Kasif. B. and Minker. J . (I982). Logic Programminga Parallel Approach. In
Van Caneghem l982a. pp. 7|-77.
Emden. M. H. van (l98l). AVL-Tree Insertion: A Benchmark Program Biased towards Prolog.
Logic Programming Newsiett. 2. 4.
Emden. M. H. van (I982). An Interpreting Algoritlun for Prolog Programs. In Van Caneghem
l982a. pp. 56-64. Also in Campbell I984. pp. 93-I I0.
Emden. M. H. van and Kowalski. R. A. (I979). The Semantics of Predicate Logic as a
Programming Language. J. ACM 13(4). 733-744.
Ennals. J . R. (I983). Beginning Micro-Prolog. Ellis Horwood. Chichester. U.K.
Fikes. R. E. and Nilsson. N. J. (I971). STRIPS: A New Approach to the Application of
Theorem-Proving to Problem Solving. Artif. Inteil. 2. 235-246.
Filgueiras. M. (I982). A Prolog Interpreter Working with In nite Terms. Repon FCIlUNL-
20/82, Universidade Nova de Lisboa. Also in Campbell I984. pp. 250-258.
Filgueiras. M. and Pereira. L. M. (I983). Relational Databases a La Carte. In Pereira et al.
I983. pp. 389-407.
Gallaire. H. (I983). Logic Databases versus Deductive Databases. In Pereira ct al. I983. pp.
608-622.
296 References

Gallaire. H. and Minker. J .. eds. ( I978). Logic and Databases". Plenum Press. New York.
Gallaire. H.. Minker. J . and Nicolas. J.-H.. eds. (l98l). Advances in Database Theory".
vol. I. Plenum Press. New York.
Gregory. S. (I980). Towards the Compilation of Annotated Logic Programs. Research Re-
port DOC 80lI6. Imperial College. London.
Gries. D. (I97I). Compiler Construction for Digital Computers. Wiley and Sons. New
York.
Guizol. J . (I975). Synthese du francais a partir dune representation en logique du premier
ordre. Ph.D. thesis. Universite d'Aix-Marseille.
Guizol. J . and Meloni. H. (I976). Prolog modulaire. Groupe d'lntelIigence Arti cielle. Uni-
versite d'Aix-Marseillc.
Hill. R. (I974). LUSH Resolution and Its Completeness. DCL Memo 78. University of
Edinburgh.
Hoare. C. A. D. (I962). Quicksor1. Comput. J. 5(I). I0-I5.
Hogger. C. J. (I979). Derivation of Logic Programs. Ph.D. thesis. Imperial College.
London.
Joubert. M. (I974). Un systeme de resolution de problemes a tendance naturelle. Ph.D.
thesis. Universite d'Aix-Marseille.
Kanoui. H. (I973). Application de la demonstration automatique aux manipulations algebri-
que et a l'integration formelle sur ordinateur. Ph.D. thesis. Universite d'Aix-Marseille.
Kanoui. H. (I982). PROLOG II. Manuel d'exemples. Groupe dIntelligence Arti cielle.
Universite d'Aix-Marseille ll.
Kanoui. H. and Van Caneghem. M. (I980). Implementing a Very High Level Language on a
Very Low Cost Computer. In Information Processing 80" (S. Lavington. ed.). pp.
349-354. North-Holland. Amsterdam.
Klutniak. F. (l98l). Remarks on Coroutines in Prolog. In Szpakowicz I981. pp. I9-29.
Kluiniak. F. (I984). The Marseille lnterpretera Personal Perspective. In Campbell
I984. pp. 65-70.
Kluiniak. F. and Szpakowicz. S. (I983). "Prolog" WNT (Wydawnictwa Naukowo-Tech-
niczne). Warsaw.
Kluiniak. F. and Szpakowicz. S. (I984). Prologa Panacea? In Campbell I984. pp. 7|-84.
Knuth. D. E. (I968). Semantics of Context-Free Languages. Math. Sysl. Theory 2. I27-I45.
Komorowski. H. J. (I982). QLOGThe Programming Environment for Prolog in LISP. In
Clark and T rnlund I982. pp. 3l5-322.
Koster. C. H. A. (I974). Using the CDL Compiler-Compiler. In Compiler Construction.
An Advanced Course" (F. J . Bauer and J . Eickel. eds.). pp. 366-426. Lecture Notes in
Computer Science 2|. Springer-Verlag. Berlin and Heidelberg.
Kowalski. R. A. (I972). The Predicate Calculus as a Programming Language. In Proceed-
ings of the Intemational Symposium and Summer School on Mathematical Foundations
of Computer Science". Jablonna near Warsaw. Poland.
Kowalski. R. A. (I974). Predicate Logic as Programming Language. In "Proceedings of the
IFIP Congress". pp. 569-574. North-Holland. Amsterdam.
Kowalski. R. A. (I978). Logic for Data Description. In Gallaire and Minker I978. pp. 77-
I03.
Kowalski. R. A. (l979a). Algorithm = Logic + Control. Cornmun. ACM 22. 42443I.
Kowalski. R. A. (l979b). Logic for Problem SoIving". North-Holland. Amsterdam.
Kowalski. R. A. and Kuehner. D. (I97l). Linear Resolution with Selection Function. Arti -
cial Intelligence 2(3l4). 227-260. Also in The Automation of Reasoning II" (J. H.
Siekmann and G. Wrightson. eds.) Springer-Verlag. Berlin and Heidelberg. I983.
 References 297

Lloyd. J . W. (I982). An Introduction to Deductive Database Systems. Department of Com-

puter Science Report TR 8ll3. University of Melboume.
McCabe. F. G. (l98l). Micro PROLOG Programmer's Reference Manual. Logic Pro-
gramming Associates Ltd.. London.
Mellish. C. S. (l98l). Automatic Generation of Mode Declarations in Prolog Programs.
Paper presented at Workshop on Logic Programming. Long Beach. Los Angeles. Cali-
fomia.
Mellish. C. S. (I982). An Alternative to Structure Sharing in the Implementation of a Prolog
Interpreter. In Clark and Tamlund I982, pp. 99-I06.
Mellish. C. and Hardy. S. (I983). Integrating Prolog into the Poplog Environment. In Bundy
I983. pp. 533-535. Also in Campbell I984. pp. I47-I62.
Moss. C. D. S. (I979). A New Grammar for Algol 68. Department of Computing repor1 79/6.
Imperial College. L.ondon.
Mycroft. A. and O'Keefe. R. (I983). A Polymorphic Type System for Prolog. In Pereira et
al. I983. pp. I07-I22.
Neves. J . and Williams. M. (I983). Towards a Co-operative Data Base Management System.
In Pereira ct al. I983. pp. 34I-370.
Neves. J .. Anderson. S. and Williams. M. (I983). Security and Integrity in Logic Data Bases
using QBE. In Pereira et al. I983. pp. 304-340.
Pasero. R. (I973). Representation du francais en logique du Ier ordre en vue de dialoguer
avec un ordinateur. Ph.D. thesis. Universite d'Aix-Marseille.
Pereira. L. M. and Porto. A. (I980a). An Interpreter for Logic Programs using Selective
Backtracking. Report 3/80. Centro de Inforrnatica da Universidade Nova de Lisboa.
Pereira. L. M. and Porto. A. (l980b). Selective Backtracking for Logic Programs. In Pro-
ceedings of the 5th Conference on Automated Deduction" (W. Bibel and R. Kowalski.
eds.). pp. 306-3l7. Springer-Verlag. Berlin and Heidelberg.
Pereira. L. M. and Porto. A. (l98l). All Solutions. Logic Programming Newslett. 2. 9-I0.
Pereira. L. M. and Porto. A. (I982). Selective Backtracking. In Clark and Tamlund I982.
pp. I07-II4.
Pereira. F. C. N. and Warren. D. H. D. (I980). De nite Clause Grammars for Language
Analysisa Survey of the Fonnalism and a Comparison with Augmented Transition
Networks. Artif. Intell. 13(3). 23l-278.
Pereira. L. M.. Pereira. F. C. N. and Warren. D. H. D. (I978). User's Guide to DECSys-
tem-l0 Prolog (Provisional Version). Department of Arti cial Intelligence. University
of Edinburgh.
Pereira. L. M.. Porto. A.. Monteiro. L. and Filgueiras. M.. eds. (I983). "Logic Program-
ming Workshop '83 Proceedings. Praia da Falesia. Algarve. Portugal. 26 June to I
July. I983. Universidade Novade Lisboa.
Porto. A. (I982). EPILOG: A Language for Extended Programming in Logic. In Van
Caneghem l982a. pp. 3|-37.
Robinson. J. A. (I965). A Machine-oriented Logic Based on the Resolution Principle. J.
ACM l2(I). pp. 23-4|.
Robinson. J. A. (I979). "Logic: Form and Functionthe Mechanization of Deductive
Reasoning. North-Holland. Amsterdam.
Roussel. P. (I975). PROLOG. manuel de reference et dutilisation. Groupe d'Intelligence
Arti cielle. Universite dAix-Marseille.
Sedgewick. R. (I983). "Algorithms". Addison-Wesley. Reading. Massachusetts.
Sergot. M. (I982). A Query-the-User Facility for Logic Programming. Research report DOC
82lI8. Imperial College. London.
293 References

Shapiro. E. Y. (I983a). Algorithmic Program Debugging". MIT Press. Cambridge. Massa-

chusetts.
Shapiro. E. Y. (l983b). A Subset of Concurrent Prolog and Its Interpreter. ICOT (Institute
for New Generation Computer Technology) Technical Report TR-003. Tokyo.
Stonebraker. M.. Wong. E.. Kreps. P. and Held. G. (I976). The Design and Implementation
of INGRES. ACM Trans. Data Base Systems l(3). pp. I89-222.
Szeredi. P. (I977). PROLOGa Very High Level Language Based on Predicate Logic.
Preprints of 2nd Hungarian Computer Science Conference. Budapest.
Szeredi. P. (I982). Module Concept for Prolog. Paper presented at the Prolog Programming
Environments Workshop. Linkoping.
SzKl (I982). MPROLOG User's Manual. Szamistastechnikai Koordinacios Intezet. Buda-
pest.
Szpakowicz. S.. ed. (l98l). Papers in Logic Programming I. IInfUW Report No. I04.
Warsaw University.
Tamlund. S.-A.. ed. (I980). Preprints of: Logic Programming Workshop. l4I6 July I980.
Debrecen. Hungary.
Ullman. J . D. (I982). Principles of Database Systems". 2nd ed. Computer Science Press.
Rockville. Maryland.
Van Caneghem. M.. ed. (l982a). Proceedings of the Ist Intemational Logic Programming
Conference". I4-I7 September I982. Faculte de Sciences de Luminy. Marseille.
France.
Van Caneghem. M. (l982b). PROLOG II. Manuel dutilisation. Groupe d'Intelligence Arti -
cielle. Universite Marseille II.
Vasey. P. (I982). AVL-Tree Insertion Revisited. Logic Programing Newslett. 3. II.
Warren. D. H. D. (I974). WARPLANa System for Generating Plans. DGL Memo 76.
University of Edinburgh.
Warren. D. H. D. (I976). Generating Conditional Plans and Programs. In Proceedings of
the AISB Summer Conference". Edinburgh. pp. 344-354.
Warren. D. H. D. (I977a). Implementing ProIogCompiling Predicate Logic Programs.
DAI Repor1 Nos. 39 and 40. University of Edinburgh.
Warren. D. H. D. (I977b). Logic Programming and Compiler Writing. DAI Report No. 44.
University of Edinburgh.
Warren. D. H. D. (I980a). An Improved Prolog Implementation Which Optimises Tail
Recursion. In Tarnlund I980. pp. I-I I.
Warren. D. H. D. (I980b). Logic Programming and Compiler Writing. SoftwarePractice
and Experience l0(2). 97-I25.
Warren. D. H. D. (l98l). Ef cient Processing of Interactive Relational Database Queries
Expressed in Logic. In Proceedings of the 7th Intemational Conference on Very Large
Data Bases". Cannes. pp. 272-281.
Warren. D. H. D. and Pereira. F. C. N. (I982). An Ef cient Easily Adaptable System for
Interpreting Natural Language Queries. Am. J. Computational Linguistics 8(3-4). lI0-
II9.
Warren. D. H. D.. Pereira. L. M. and Pereira. F. C. N. (I977). Prolog--the Language and
Its Implementation Compared with Lisp. Presented at the ACM Symposium on Arti -
cial Intelligence and Programming Languages. Rochester. New York. SIGART News-
letter No. 64. SIGPLAN Notices 12(8). I977. pp. I09-II5.
Wijngaarden. van A.. ed. (I976). "Revised Report on the Algorithmic Language Algol 68".
Springer-Verlag. Berlin and Heidelberg.
 References 299

Wirth. N. (I976). Algorithms + Data Structures = Programs". Prentice-Hall. Englewood

Cliffs. New Jersey.
Wise. M. J. (I984). EPILOG: Re-interpreting and Extending Prolog for a Multiprocessor
Environment. In Campbell I984. pp. 34I-35I.
Zloof. M. M. (I977). Query-by-Example: A Data Base Language. IBM Syst. J. l6(4). 324-
342
The subject index was originally compiled from the rst edition of this volume
Therefore the page numbers for the following entries should now read

Chat80. 237-238 Planning. 2I 5-226

Coroutining. 57 Prolog I. 249-250
in Prolog. 25I-252 Prolog II. 250-253

Databases. relational. 226-237 Query-by-Example. 23l

Dialects
micro-Prolog and MPROLOG. 253-254 Sequel. 23l
Prolog I. 249-250 Simple. 253
Prolog II. 250-253
Toy-sequel. 23l-247
In nite trees. 23. 250 Tree(s)
in nite. 23. 250
micro-Prolog. 253
Modularisation. in Prolog. 252. 254 WARPLAN. 2I6226
MPROLOG. 253-254
INDEX

Page numbers in italics indicate material referring to implementation issues. P381: numbers
followed by n indicate material in footnotes.

A =:=! 2. 28. I52

=</ 2. 28. I52
Altematives. in grammar rules. 79-8| ==/ 2. I53
Anonymous variables. ll =\= I2. 28. I52
Arguments. 4 > / 2. 28. I52
Arity. 4-5 >=/ 2. 28. I52
An'ay analogues. in Prolog. II3-I I6 @</ 2. 28. I52
Associative functor. I55 @=<! 2. 28.152
Atoms. in Toy-Prolog interpreter. I89-I90 @>/ 2. 28. I52
Attribute grammars. 82. 84 @>=/ 2. 28. I52
Axioms. 44 abolish I 2. I62
alphanuml I. I58
B ancestorl I. I64
arg I 3. I60
Backtracking. 26-27 assert! I. I09. I29. I6I
how to use. 28-32 assen I 3. l6I
intelligent. 57 assertal I. I6l
selective. 57 assertzl I. I6I
Binary trees. representation in Prolog. 89- atoml I. I59
94 bagofI3. II2. I66
Bound variables. 9. I8 bigletter! I. I58
Built-in procedures. see also Procedure(s). bracket I I. I58
built-in calll I. II8. I63
H0. 36. l2l-I24. I63 checkI I. I50
, I 2. 39. I50 clause I 2. I62
;I2. 39. I50 clause I 5. I6l
\== I2. I53 consult! I. I6. I62
<I2. 28. I52 debugI0. I65
=.J 2. 39. II7. I60 delopl I. I56
= I 2. 28-29. I52 digit! I. I58
=\= I2. I52 display! I. I54
302 Index

echoI0. I54 statusI0. I57

eqvar I 2. I52 stopI0. I64
errorl I. I47-I48 sum I 3. I50
failI0. 3I. I49 symch I I. I58
functor I 3. I59 tag! I. I64
halt! I. I64 tagcutl I. I64
integer! I. 28. I58 tagexit I I. I64
is I 2. I3. I5l tagfaill I. I64
isclosedlistl I. I65 tellI I. I54
iseolnl I. I57 tellingl I. I54
Iastchl I. I57 toldI0. I54
length I 2. I65 true I 0. I49
IessI2. I5I varl I. I58
letter! I. I58 wchI I. I57
listing I 0. I63 writel I. I3. I54-I55
listing! I. I63 writeq! t. tss
memberI2. I66
nlI0. I3. I57
nodebug! 0. I65
noechoI0. I54 CaIl(s)
nonexistent I 0. I65 order of. 34-35
nononexistent I 0. I65 variable. 38-40
nonvarint I I. I59 Canonic form. 6
notI I. 39. I22. lb. I49 Characters. 8
numbervarsl I. I66 Chat80. B2-253
once] l. I23. I50 Choice points. 27
op I 3. 8. I55 Clausal representauon. of data structures.
ordchr I 2. I57 I07-ll3
phrase I 2. 69. I20. I65 Clause(s). 23-24. 46 I44
pname I 2. 73. I59 as global data. I29-I30
pnameiI2. I59 Hom. 46-50
prede ned I 2. I62 order of. 34-35
prodI4. I5I unit. 32. 47. I44
protect I0. I62 Closed lists. I00
rch I0. I56 Command. I5. I6. I44
rdchl I. I57 Command mode. I5
rdchskl I. I57 Comment. I6. I46
readI I. I55 Compound objects. 3
reconsultl I. I6. I62-163 descriptions. 5-8
rede ne I 0. I62 functors. 3-5
repeat I 0. I63 suings. 8
reuact I I. I09. I29, I6]-I62 Conditions. in grammar rules. 73-76 I44
retract I 3. l6l Connectives. logical 42
see! I. I53 Consequence. logical I44
seeing! I. I53 Constants. I-3
seen I 0. I54 Context. in grammar rules. 76-79
side_effectsI I. I50 Control. 54. I63-I64 I78-I79
sltipbll 0. I56 backtracking. 26-32
smalletterl I. I58 cut. 35-38
solocharl I. I58 general forr'n of a procedure 24-26
 Index 303

order of calls and clauses. 34-35 Fail point record. I78-I79

static interpretation of procedure. 32-33 Failure. 27
in Toy-Prolog interpreter. I94-I99 forced. I25
variable calls. 38-40 as programming tool. I25-I29
Convenience. built-in procedures. I49-I50 Fonnal reasoning. 44-45
Copy stack. I75-I76 Formulae. 4|-42
Coroutining. 57 inconsistent. 43
in Prolog. 257-258 interpretation of. 42-44
Current resolvent. 47 tautology, 43
Cut procedure. 35-38. see also Built-in true (false) in a model. 43
procedures. U0 true (false) in an interpretation.
43
D Free variable. 9
Frozen variable. I79
DAGs. see Directed acyclic graphs Fullstop. I5. I46
Data bases. relational. 228-253 Functor. 3-5. 42. I45
Data stnrctures. 88 associative. I55
access to stnrcture of terms. II6-I20 in x. 6. I55
array analogues. II3-II6 left-associative. 6. I55
clausal representation of. I07-I I3 main. 5
compound objects. 3-8 non-associative. I55
constants. I-3 post x. 6. I55
difference lists. I04-I07 pre x. 6. I55
open trees and lists. 98-I03 priority. 7
simple trees and lists. 88-98 right-associative. 6. I55
tenns. 9-ll G
variables. 8-9
Debugging. I64-I65 Generator. 3I
De nite clause grammars. 68n Global stack. I76
De nition mode. I5 Global variables. I 76
Dialects Goal statements. I6
micro-Prolog and MPROLOG. 259-260 Grammar(s)
Prolog I. 255-256 attribute. 82. 84
Prolog II. 256-259 de nite clause. 68n
Dictionary. I89-I90 metamorphosis. 68
Difference lists. I04-I07 two-level. 82
Directed acyclic graphs (DAGs). ll. I68 Grammar preprocessor. 2I2-213
Directives. I5-I7. I44 Grammar processing. I65
Grammar nrles. I44
extensions
alternatives. 79
Ef ciency. in grammar rules. 8l-83 conditions. 73-76
Euler paths. I39-I42 context. 76-79
Extralogical features. of Prolog. 35-38 in Prolog. 68
simplest form of. 67-69
Ground variable. 9
H
Factorization. 82
Fail point(s). 27. I 78-I79 Hom clauses. 46-50
304 Index

Metamorphosis grammars. 68
micro-Prolog. 259
Identi er. 2 Mixed operators. 7. I55. I56
Inconsistency. 43 Mode declarations. I77
Indexing. I09-IIO Model. 43
Inference rules. 44 true (false) formulae in. 43
In nite trees. 23. 256 Modularisation. in Prolog. 258. 260
In x functor. 6. I55 Molecules. l69n
Initialisationl 201 Monitor. 203
Input I output MPROLOG. 259-260
listing control. I54
single characters. I56-I57 N
switching streams. I53-I54
terms. I54-I56 Namets)
lnstantiation comparing. I5I-I52
of a term. I0-II quoted. 2
of a variable. 9 variable. 9
lnteger(s). 2. I45 Negation. in Prolog. 38
comparing. I5I-I52 Negative literals. 46
Interpretation Nondeterminism. in Prolog. 3I
of formulae. 42-44 Non-associative functor. I55
static. of procedures. 32-33 Non-structure sharing (NSS). I68. I72-I76
Nonterrninal (symbol). 68. I44
I.. Nonterrninal symbols. parameters of. 69-
73
Language. interrnediate. 20!-203 NSS. see Non-structure sharing
Left-associative functor. 6. I55
Library. 2I3-214 O
Linear lists. representation in Prolog. 95
List(s). I45 Occur check. 23
closed. I00 Open lists. 98-I03
difference. I04-I07 Operations. II-I3. see also Clause(s)
linear. 95 directives. I5-I7. I44
open. 98-I03 simplest forr'n of procedure. I3-I5
representation in Prolog. 7-8 uni cation. I7-23
reversal of. I3I-I34 Operator(s). 7
simple. 88-98 mixed. 7
List notation. 95 prede ned. I48-I49
Literals. 46 type of. I55
negative. 46
positive. 46 P
Local stack. I76
Local variables. I 76 Parallelism 57
Logical connectives. 42 Parsing problem. representation of. 59-67
Logical consequence. 44 Planning. 2l5-228
Positive literals. 46
Post x functor. 6. I55
Prede ned operators. I48-I49
Main functor. 5 Predicate(s). 42
Matching. see Uni cation Predicate symbol. I2. 42
 Index 305

Prg g functor, 6, |55 Right-associative functor. 6. I55

Priority. 7. I55 Rule of resolution. 45-46
Procedure(s). II-I3. see also Built-in
procedures S
accessing. I60-I63
built-in. I2. I47-I49 Scope. of quanti ers, 42
prede ned, I47 Search space. 5I
side-effects of. 27 Sequel. 233
system. I47 Side-effect. of built-in procedure. 27
general form of. 24-26 Simple. 259
nondeterministic. 3I Simple lists. 88-98
simplest forr'n of. I3-I5 Skeletons. I69n
static interpretation of. 32-33 Sorting. l34l33
system. I2. I47. I99 Stack
in Toy-Prolog interpreter. I89. I90. I99 copy. I75-I76
Procedure body. 23. I44 global. I 76
Procedure call. I44 local. I 76
Procedure de nition. I44 Storage areas. I87-I88
Procedure heading. r4. I44 Stratesv. 5|-$8
Prolog I. 255-256 Strings. 8. I45
Prolog ll. 256-259 Structure sharing. I68-I72. I76-I78
Prolog-I0. I43 Switching streams. I53-I54
Prolog-I0 character input Symbol. 2. I45
get. I56 Syntax of Prolog. I43-I47
get0. I56 System procedures. I2
put. I56 built-in. I47
skip. I56 prede ned. I47
Prolog grammar rule. 68 in Toy-Prolog interpreter. I99
Proof. 44
Proof tree. 52 T
PrototyP(s). I69. I90I94
Tail recursion optimisation (TRO). I 79-I82
Q delayed. I82
Tautology(ies). 43
Quanti ers. 42 Terrn(s). 5. 9-ll, 42. I44
scope of. 42 accessing stnrcture of. I59-I60
Query. I5-I6. I44 instantiation of. I0-I I
Query-by-Example. 233 Term descriptions. I68
Quoted name. 2. I45 Term handle. I69
Terminal (symbol). 68. 74. I44
Term instances. I68. I90-I94
Theorems. 44
Reader. for interpretation of Prolog-I0 in Theory. 44
Toy-Prolog. 203-2I2 Toy-Sequel. 233-253
Reasoning. formal. 44-45 Trail. I 79
Recursion. elimination of. in metamorpho- Translator. 214
sis grammars. 83-85 Tree(s)
Reductio ad absurdum. 45 binary. 89-94
Resolution. rule of. 45-46 closed. l0I
Resolvent. current. 25-26 in nite. 23. 256
 306 Index

open. 98-I03 bound. 9

proof. 52 bound together. I8
simple. 88-98 free. 9
TRO. see Tail recursion optimisation frozen. I 79
Two-level grammars. 82 global. I76
ground. 9
U
instantiation of. 9
Uni cation. I7-23 local. I 76
success and failure. I7 Variable binding. 9
Unit clause. 32. 47. I44 Variable calls. 38-40
Univ. see Built-in procedures. =.J 2 Variable frame. I69
Universe. 43 Variable name. 9
user ( lename of the user's terminal). I6. Variable temiinal (symbol). 74
I53

V W
VariabIes(s). 8-9. 42. I45
anonymous. II WARPLAN. 2I6-228
 A.P.l.C. Studies in Data Processing
General Editors: Fraser Duncan and M. J . R. Shave

Some Commercial Autocodes. A Comparative Study

E. L. Willey. A. d'Agapeyeff. Marion Tribe. B. J. Gibbens
and Michelle Clarke
A Primer of ALGOL 60 Programming
E. W. Dijkstra
lmput Language for Automatic Programming
A. P. Yershov. G. I. Kozhukhin and U. Voloshin
Introduction to System Programming
Edited by Peter Wegner
ALGOL 60 Implementation. The Translation and Use of
ALGOL 60 Programs on a Computer
B. Randell and L. J. Russell
Dictionary for Computer Languages
Hans Breuer
The Alpha Automatic Programming System
Edited by A. P. Yershov
Structured Programming
O.-J. Dahl. E. W. Dijkstra and C. A. R. Hoare
Operating Systems Techniques
Edited by C. A. R. Hoare and R. H. Perrott
ALGOL 60. Compilation and Assessment
B. A. Wichmann
Definition of Programming Languages by Interpreting Automata
Alexander Ollongren
Principles of Program Design
M. A. Jackson
Studies in Operating Systems
R. M. McKeag and R. Wilson
Software Engineering
R. J. Perrott
(continued)

Out of print.
Computer Architecture: A Structured Approach
R. W. Doran
Logic Programming
Edited by K. L. Clark and S.-A. Tarnlund
Fortran Optimization
Michael Metcalf
Multi-microprocessor Systems
Y. Paker
Introduction to the Graphical Kernel SystemGKS
F. R. A. Hopgood. D. A. Duce. J. R. Gallop and
D. C. Sutcliffe
Distributed Computing
Edited by Fred B. Chambers. David A. Duce and
Gillian P. Jones
Introduction to Logic Programming
Christopher John Hogger
Lucid. the Data ow Programming Language
William W. Wadge and Edward A. Ashcroft
Foundations of Programming
Jacques Arsac
Prolog for Programmers
Feliks Kluiniak and Stanislaw Szpakowicz
This is a self-contained handbook of advanced logic programming. It
assumes the reader is not a novice to computer science. has perhaps read
an introductory book on Prolog and would now like to do some non-trivial
work using the language.
A thorough discussion of Prolog treatment of data structures. an overview
of Prolog's logical foundations. a concise but comprehensive presentation
of logic grammars. simple and advanced programming techniques. style
and efficiency guidelines. e reel exercise in Prolog implementation. and
two non-trivial case studies make this an invaluable handbook for both
professional and student. Experienced Prolog programmers may find the
systematic exposition helpful. and the discussion of implementation issues
instructive.
One of the unique features of the book is its detailed coverage of Meta-
morphosis (or Definite Clause) Grammars: programmers applying Prolog
to natural or formal language processing will definitely appreciate this.
A small but complete interpreter is provided on the disk in source form
(Pascal and Prolog) so that readers can use it to run their programs or
take it apart to study in detail to achieve an in-depth understanding of
Prolog. '

Academic ' ress

Harcourt Brace Jovanovich. Publishers
LONDON OFILANDO SAN ottseo
NEW YOFIK AUSTIN BOSTON SYDNEY
TOKYO TOFIONTO
had-|:rnIeFlIIlrue.(LrrrlnI|)LIl.
zatanovu -ann.Lor\eonrrw1rox.Engune
arrerrSrnre_nrrnn.trrr..orterrao.t=rorraea2ur
_ _ __ r__'_."-.-I, . -
56BerberGrelneFlond.DcI1llIlt.0r-|tnrroll2At.Car\e:ln
Aedernleirenltltr n
POBt:I1II1HorthFlyde. MSW. 2'tt3.At.rriin
' A;rrernretIrun.tnpn. rrrr.
ltdeheehr I-leltnltu stag- 3-I1-is hdahasln. Chlyodl-lut. Ttiyo toe. Japan