Book
Uploaded by
jaysonceojaytechBook
Uploaded by
jaysonceojaytechPrinciples of Programming Languages
Version 1.0.2
Mike Grant
Zachary Palmer
Scott Smith
http://pl.cs.jhu.edu/pl/book
i
Copyright © c 2002-2016 Scott F. Smith.
This work is licensed under the Creative Commons Attribution-Share Alike 3.0 United
States License. To view a copy of this license, visit
http://creativecommons.org/licenses/by-sa/3.0/us/ or send a letter to Creative
Commons, 171 Second Street, Suite 300, San Francisco, California, 94105, USA.
This document was last compiled on February 10, 2018.
Contents
Preface v
1 Introduction 1
1.1 The Pre-History of Programming Languages . . . . . . . . . . . . . . . . . 1
1.2 A Brief Early History of Languages . . . . . . . . . . . . . . . . . . . . . . 2
1.3 This Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Operational Semantics 4
2.1 A First Look at Operational Semantics . . . . . . . . . . . . . . . . . . . . 4
2.2 BNF grammars and Syntax . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2.1 Operational Semantics for Logic Expressions . . . . . . . . . . . . 6
2.2.2 Abstract Syntax . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.3 Operational Semantics and Interpreters . . . . . . . . . . . . . . . 13
2.3 The F[ Programming Language . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3.1 F[ Syntax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3.2 Variable Substitution . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3.3 Operational Semantics for F[ . . . . . . . . . . . . . . . . . . . . . 21
2.3.4 The Expressiveness of F[ . . . . . . . . . . . . . . . . . . . . . . . 29
2.3.5 Russell’s Paradox and Encoding Recursion . . . . . . . . . . . . . 33
2.3.6 Call-By-Name Parameter Passing . . . . . . . . . . . . . . . . . . . 39
2.3.7 F[ Abstract Syntax . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.4 Operational Equivalence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.4.1 Dening Operational Equivalence . . . . . . . . . . . . . . . . . . . 44
2.4.2 Properties of Operational Equivalence . . . . . . . . . . . . . . . . 46
2.4.3 Examples of Operational Equivalence . . . . . . . . . . . . . . . . 47
2.4.4 The λ-Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3 Tuples, Records, and Variants 50
3.1 Tuples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.2 Records . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.2.1 Record Polymorphism . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.2.2 The F[R Language . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.3 Variants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.3.1 Variant Polymorphism . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.3.2 The F[V Language . . . . . . . . . . . . . . . . . . . . . . . . . . 55
ii
CONTENTS iii
4 Side Eects: State and Exceptions 58
4.1 State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.1.1 The F[S Language . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.1.2 Cyclical Stores . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.1.3 The “Normal” Kind of State . . . . . . . . . . . . . . . . . . . . . 66
4.1.4 Automatic Garbage Collection . . . . . . . . . . . . . . . . . . . . 66
4.2 Environment-Based Interpreters . . . . . . . . . . . . . . . . . . . . . . . . 67
4.3 The F[SR Language . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.3.1 Multiplication and Factorial . . . . . . . . . . . . . . . . . . . . . . 69
4.3.2 Merge Sort . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.4 Exceptions and Other Control Operations . . . . . . . . . . . . . . . . . . 73
4.4.1 Interpreting Return . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.4.2 The F[X Language . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.4.3 Implementing the F[X Interpreter . . . . . . . . . . . . . . . . . . 77
5 Object-Oriented Language Features 79
5.1 Encoding Objects in F[SR . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.1.1 Simple Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.1.2 Object Polymorphism . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.1.3 Information Hiding . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.1.4 Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.1.5 Inheritance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.1.6 Dynamic Dispatch . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.1.7 Static Fields and Methods . . . . . . . . . . . . . . . . . . . . . . . 91
5.2 The F[OB Language . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.2.1 Concrete Syntax . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.2.2 A Direct Interpreter . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.2.3 Translating F[OB to F[SR . . . . . . . . . . . . . . . . . . . . . . 96
6 Type Systems 99
6.1 An Overview of Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
6.2 TF[: A Typed F[ Variation . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.2.1 Design Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.2.2 The TF[ Language . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
6.3 Type Checking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.4 Types for an Advanced Language: TF[SRX . . . . . . . . . . . . . . . . 109
6.5 Subtyping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6.5.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6.5.2 The STF[R Type System: TF[ with Records and Subtyping . . . 114
6.5.3 Implementing an STF[R Type Checker . . . . . . . . . . . . . . . 115
6.5.4 Subtyping in Other Languages . . . . . . . . . . . . . . . . . . . . 115
6.6 Type Inference and Polymorphism . . . . . . . . . . . . . . . . . . . . . . 116
6.6.1 Type Inference and Polymorphism . . . . . . . . . . . . . . . . . . 116
6.6.2 An Equational Type System: EF[ . . . . . . . . . . . . . . . . . . 116
6.6.3 PEF[: EF[ with Let Polymorphism . . . . . . . . . . . . . . . . . 121
6.7 Constrained Type Inference . . . . . . . . . . . . . . . . . . . . . . . . . . 124
CONTENTS iv
7 Concurrency 127
7.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
7.1.1 The Java Concurrency Model . . . . . . . . . . . . . . . . . . . . . 128
7.2 The Actor Model and AF[V . . . . . . . . . . . . . . . . . . . . . . . . . 129
7.2.1 Syntax of AF[V . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
7.2.2 An Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
7.2.3 Operational Semantics of Actors . . . . . . . . . . . . . . . . . . . 131
7.2.4 The Local Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
7.2.5 The Global Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
7.2.6 The Atomicity of Actors . . . . . . . . . . . . . . . . . . . . . . . . 133
8 Compilation by Program Transformation 134
8.1 Closure Conversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
8.1.1 The Ocial Closure Conversion . . . . . . . . . . . . . . . . . . . . 138
8.2 A-Translation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
8.2.1 The Ocial A-Translation . . . . . . . . . . . . . . . . . . . . . . . 141
8.3 Function Hoisting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
8.4 Translation to C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
8.4.1 Memory Layout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
8.4.2 The toC translation . . . . . . . . . . . . . . . . . . . . . . . . . . 151
8.4.3 Compilation to Assembly code . . . . . . . . . . . . . . . . . . . . 154
8.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
8.6 Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
8.7 Garbage Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Bibliography 156
Index 158
Preface
This book is an introduction to the study of programming languages. The material has
evolved from lecture notes used in a programming languages course for juniors, seniors,
and graduate students at Johns Hopkins University [21].
The book treats programming language topics from a foundational. It is foundational
in that it focuses on core concepts in language design such as functions, records, objects,
and types and not directly on applied languages such as C, C++, or Java. We show how
the particular core concepts are realized in these modern languages, and so the reader
should emerge from this book with a stronger sense of how they are structured.
While the book uses formal mathematical techniques such as operational semantics
and type systems, it does not emphasize proofs of properties of these systems. We will
sketch the intuitions of some properties but not do any detailed proofs.
The OCaml Language
The OCaml programming language [15] is used throughout the book, and assignments
related to the book are best written in OCaml. OCaml is a modern dialect of ML which
has the advantages of being reliable, fast, free, and available on just about any platform
through http://caml.inria.fr.
The FbDK
Complementing the book is the F[ Development Kit, FbDK. It is a set of OCaml utilities
and interpreters for designing and experimenting with the toy F[ and F[SR languages
dened in the book. It is available from the book homepage at http://pl.cs.jhu.edu/
pl/book, and is documented in Appendix ??.
Background Needed
The book assumes familiarity with the basics of OCaml, including the module system
(but not the objects, the “O” in OCaml). Beyond that there is no absolute prerequisite,
but knowledge of C, C++, and Java is helpful because many of the topics in this book are
implemented in these languages. The compiler presented in chapter 8 produces C code
as its target, and so a basic knowledge of C will be needed to implement the compiler.
More nebulously, a certain “mathematical maturity” greatly helps in understanding the
concepts, some of which are deep. for this reason, previous study of mathematics, for-
mal logic and other foundational topics in Computer Science such as automata theory,
grammars, and algorithms will be a great help.
v
Chapter 1
Introduction
General-purpose computers have the amazing property that a single piece of hardware
can do any computation imaginable. Before general-purpose computers existed, there
were special-purpose computers for arithmetic calculations, which had to be manually
recongured to carry out dierent calculations. A general-purpose computer, on the other
hand, has the conguration information for the calculation in the computer memory itself,
in the form of a program. The designers realized that if they equipped the computer with
the program instructions to access an arbitrary memory location, instructions to branch
to a dierent part of the program based on a condition, and the ability to perform basic
arithmetic, then any computation they desired to perform was possible within the limits
of how much memory, and patience waiting for the result, they had.
These initial computer programs were in machine language, a sequence of bit pat-
terns. Humans understood this language as assembly language, a textual version of
the bit patterns. So, these machine languages were the rst programming languages,
and went hand-in-hand with general-purpose computers. So, programming languages
are a fundamental aspect of general-purpose computing, in contrast with e.g., networks,
operating systems, and databases.
1.1 The Pre-History of Programming Languages
The concept of general-purpose programming in fact predates the development of com-
puters. In the eld of mathematical logic in the early 20th century, logicians created
their own programming languages. Their motivation originally sprang from the concept
of a proof system, a set of rules in which logical truths could be derived, mechanically.
Since proof rules can be applied mechanically, all of the logically true facts can be me-
chanically enumerated by a person sitting there applying all of the rules in every order
possible. This means the set of provable truths are recursively enumerable. Logicians
including Frege, Church, and Curry wanted to create a more general theory of logic and
proof; this led Church to dene the λ-calculus in 1932, an abstract language of functions
which also dened a logic. The logic turned out to be inconsistent, but by then logicians
had discovered that the idea of a theory of functions and their (abstract) computations
was itself of interest. They found that some interesting logical properties (such as the
collection of all truths in certain logical systems) were in fact not recursively enumerable,
1
CHAPTER 1. INTRODUCTION 2
meaning no computer program could ever enumerate them all. So, the notion of general-
purpose computation was rst explored in the abstract by logicians, and only later by
computer designers. The λ-calculus is in fact a general-purpose programming language,
and the concept of higher-order functions, introduced in the Lisp programming language
in the 1960’s, was derived from the higher-order functions found in the λ-calculus.
1.2 A Brief Early History of Languages
There is a rich history of programming languages that is well worth reading about; here
we provide a terse overview.
The original computer programming languages, as mentioned above, were so-called
machine languages: the human and computer programmed in same language. Machine
language is great for computers but not so great for humans since the instructions are
each very simple and so many, many instructions are required. High-level languages were
introduced for ease of programmability by humans. FORTRAN was the rst high-level
language, developed in 1957 by a team led by Backus at IBM. FORTRAN programs
were translated (compiled ) into machine language to be executed. They didn’t run as
fast as hand-coded machine language programs, but FORTRAN nonetheless caught on
very quickly because FORTRAN programmers were much more productive. A swarm of
early languages followed: ALGOL in ’58, Lisp in the early 60’s, PL/1 in the late 60’s,
and BASIC in 1966.
Languages have developed on many fronts, but there is arguably a major thread of
evolution of languages in the following tiers:
1. Machine language: program directly in the language of the computer
2. FORTRAN, BASIC, C, Pascal, . . . : rst-order functions, nested control structures,
arrays.
3. Lisp, Scheme, ML: higher-order functions, automated garbage collection, memory
safety; strong typing in ML
Object-oriented language development paralleled this hierarchy.
1. (There was never an object-oriented machine language)
2. Simula67 was the original object-oriented language, created for simulation. It was
FORTAN-like otherwise. C++ is another rst-order object-oriented language.
3. Smalltalk in the late 70’s: Smalltalk is a higher-order object-oriented language
which also greatly advanced the concept of object-oriented programming by show-
ing its applicability to GUI programming. Java is partly higher order, has auto-
mated garbage collection, and is strongly typed.
Domain-specic programming languages (DSLs) are languages designed to solve a
more narrow domain of problems. All languages are at least domain-specialized: FOR-
TRAN is most highly suited to scientic programming, Smalltalk for GUI programming,
Java for Internet programming, C for UNIX system programming, Visual Basic for Mi-
crosoft Windows. Some languages are particularly narrow in applicability; these are
CHAPTER 1. INTRODUCTION 3
called Domain-specic languages. SNOBOL and Perl are text processing languages.
UNIX shells such as sh and csh are for simple scripting and le and text hacking. Prolog
is useful for implementing rule-based systems. ML is to some degree a DSL for language
processing. Also, some languages aren’t designed for general programming at all, in that
they don’t support full programmability via iteration and arbitrary storage allocation.
SQL is a database query language; XML is a data representation language.
1.3 This Book
In this book, our goal is to study the fundamental concepts in programming languages,
as opposed to learning a wide range of languages. Languages are easy to learn, it is the
concepts behind them that are dicult. The basic features we study in turn include
higher-order functions, data structures in the form of records and variants, mutable
state, exceptions, objects and classes, and types. We also study language implementa-
tions, both through language interpreters and language compilers. Throughout the book
we write small interpreters for toy languages, and in Chapter 8 we write a principled
compiler. We dene type checkers to dene which programs are well-typed and which
are not. We also take a more precise, mathematical view of interpreters and type check-
ers, via the concepts of operational semantics and type systems. These last two concepts
have historically evolved from the logician’s view of programming.
Now, make sure your seat belts are buckled, sit back, relax, and enjoy the ride. . .
Chapter 2
Operational Semantics
2.1 A First Look at Operational Semantics
The syntax of a programming language is the set of rules governing the formation of
expressions in the language. The semantics of a programming language is the meaning
of those expressions.
There are several forms of language semantics. Axiomatic semantics is a set of ax-
iomatic truths in a programming language. Denotational semantics involves modeling
programs as static mathematical objects, namely as set-theoretic functions with specic
properties. We, however, will focus on a form of semantics called operational semantics.
An operational semantics is a mathematical model of programming language execu-
tion. It is, in essence, an interpreter dened mathematically. However, an operational
semantics is more precise than an interpreter because it is dened mathematically, and
not based on the meaning of the programming language in which the interpreter is writ-
ten. This might sound sound like a pedantic distinction, but interpreters interpret e.g. a
language’s if statements with the if statement of the language the interpreter is written
in. This is in some sense a circular denition of if. Formally, we can dene operational
semantics as follows.
Denition 2.1 (Operational Semantics). An operational semantics for a program-
ming language is a mathematical denition of its computation relation, e ⇒ v, where e
is a program in the language.
e ⇒ v is mathematically a 2-place relation between expressions of the language, e, and
values of the language, v. Integers and booleans are values. Functions are also values
because they don’t compute to anything. e and v are metavariables, meaning they
denote an arbitrary expression or value, and should not be confused with the (regular)
variables that are part of programs.
An operational semantics for a programming language is a means for understanding in
precise detail the meaning of an expression in the language. It is the formal specication
of the language that is used when writing compilers and interpreters, and it allows us to
rigorously verify things about the language.
4
CHAPTER 2. OPERATIONAL SEMANTICS 5
2.2 BNF grammars and Syntax
Before getting into meaning we need to take a step back and rst precisely dene language
syntax. This is done with formal grammars. Backus-Naur Form (BNF) is a standard
grammar formalism for dening language syntax. You could well be familiar with BNF
since it is often taught in introductory courses, but if not we provide a brief overview.
All BNF grammars comprise terminals, nonterminals (aka syntactic categories), and
production rules. Terminals are traditionally identied using lower-case letters; non-
terminals are identied using upper-case letters. Production rules describe how non-
terminals are dened. The general form of production rules is:
〈nonterminal〉 ::= 〈form 1〉 | · · · | 〈form n〉
where each “form” above describes a particular language form – that is, a string of
terminals and non-terminals. A term in the language is a string of terminals which
matches the description of one of these rules (traditionally the rst).
For example, consider the language Sheep. Let {S} be the set of nonterminals, {a, b}
be the set of terminals, and the grammar denition be:
S ::= b | Sa
Note that this is a recursive denition. Examples of terms in Sheep are
b, ba, baa, baaa, baaaa, . . .
That is, any string starting with the character b and followed by zero or more a characters
is a term in Sheep. The following are examples that are not terms in SHEEP:
• a: Terms in Sheep must start with a b.
• bbaaa: Sheep does not allow multiple b characters in a term.
• baah: h is not a terminal in Sheep.
• Saaa: S is a non-terminal in Sheep. Terms may not contain non-terminals.
Another way of expressing a grammar is by the use of a syntax diagram. Syntax
diagrams describe the grammar visually rather than in a textual form. For example, the
following is a syntax diagram for the language Sheep:
S b
S a
The above syntax diagram describes all terms of the Sheep language. To generate
a form of S, one starts at the left side of the diagram and moves until one reaches the
right. The rectangular nodes represent non-terminals while the rounded nodes represent
terminals. Upon reaching a non-terminal node, one must construct a term using that
non-terminal to proceed.
CHAPTER 2. OPERATIONAL SEMANTICS 6
As another example, consider the language Frog. Let {F, G} be the set of nontermi-
nals, {r, i, b, t} be the set of terminals, and the grammar denition be:
F ::= rF | iG
G ::= bG | bF | t
Note that this is a mutually recursive denition. Note also that each production rule
denes a syntactic category. Terms in FROG include:
ibit, ribbit, ribibibbbit . . .
The following terms are not terms in Frog:
• rbt: When a term in Frog starts with r, the following non-terminal is F . The
non-terminal F may only be exapnded into rF or iG, neither of which start with
b. Thus, no string starting with rb is a term in Frog.
• rabbit: a is not a terminal in Frog.
• rrrrrrF : F is a non-terminal in Frog; terms may not contain non-terminals.
• bit: The only forms starting with b appear as part of the denition of G. As F is
the rst non-terminal dened, terms in Frog must match F (which does not have
any forms starting with b).
The following syntax diagram describes Frog:
F r F
i G
G b G
b F
2.2.1 Operational Semantics for Logic Expressions
In order to get a feel for what an operational semantics is and how it is dened, we will
now examine the operational semantics for a very simple language: propositional boolean
logic with no variables. The syntax of this language is as follows. An expression e is
recursively dened to consist of the values True and False, and the expressions e And e,
e Or e, e Implies e, and Not e.1 This syntax is known as the concrete syntax,
1
Throughout the book we use syntax very similar to OCaml in our toy languages, but with the
convention of capitalizing keywords to avoid potential conicts with the OCaml language.
CHAPTER 2. OPERATIONAL SEMANTICS 7
because it is the syntax that describes the textual representation of an expression in the
language. We can express it in a BNF grammar as follows:
e ::= v | Not e | e And e | e Or e | e Implies e | (e) expressions
v ::= True | False values
The following is an equivalent syntax diagram:
e v
Not e
e And e
e Or e
e Implies e
v True
False
Note that the syntax above breaks tradition somewhat by using lower-case letters
for non-terminals. Terminals are printed in xed-width font. The rationale for this is
consistency with the metavariables we will be using in operational semantics below and
will become clear shortly.
We can now discuss the operational semantics of the boolean language. Operational
semantics are written in the form of logic rules, which are written as a series of pre-
conditions above a horizontal line and the conclusion below it. For example, the logic
rule
Red(x) Shiny(x)
(Apple Rule)
Apple(x)
indicates that if a thing is red and shiny, then that thing is an apple. This is, of course, not
true; many red, shiny things exist which are not apples. Nonetheless, it is a valid logical
statement. In our work, we will be dening logical rules pertaining to a programming
language; as a result, we have control over the space in which the rules are constructed.
We need not necessarily concern ourselves with intuitive sense so long as the programming
language has a mathematical foundation.
Operational semantics rules discuss how pieces of code evaluate. For example, let us
consider the And rule. We may dene the following rule for And:
(And Rule (Try 1))
True And False ⇒ False
CHAPTER 2. OPERATIONAL SEMANTICS 8
This rule indicates that the boolean language code True And False evaluates to
False. The absence of any preconditions above the line means that no conditions must
be met; this operational semantics rule is always true. Rules with nothing above the line
are termed axioms since they have no preconditions and so the conclusion always holds.
As a rule, though, it isn’t very useful. It only evaluates a very specic program.
This rule does not describe how to evaluate the program True And True, for instance.
In order to generalize our rules to describe a full language and not just specic terms
within the language, we must make use of metavariables.
To maintain consistency with the above BNF grammar, we use metavariables starting
with e to represent expressions and metavariables starting with v to represent values.
We are now ready to make an attempt at describing every aspect of the And operator
using the following new rule:
(And Rule (Try 2))
v1 And v2 ⇒ the logical and of v1 and v2
Using this rule, we can successfully evaluate True And False, True and True, and
so on. Note that we have used a textual description to indicate the value of the expression
v1 And v2 ; this is permitted, although most rules in more complex languages will not
use such descriptions.
We very quickly encounter limitations in our approach, however. Consider the pro-
gram True And (False And True). If we tried to apply the above rule to that program,
we would have v1 = True and v2 = (False And True). These two values cannot be ap-
plied to logical and as (False and True) is not a boolean value; it is an expression.
Our boolean language rule does not allow for cases in which the operands to And are
expressions. We therefore make another attempt at the rule:
e1 ⇒ v1 e2 ⇒ v2
(And Rule (Try 3))
e1 And e2 ⇒ the logical and of v1 and v2
This rule is almost precisely what we want; in fact, the rule itself is complete. Intu-
itively, this rule says that e1 And e2 evaluates to the logical and of the values represented
by e1 and e2 . But consider again the program True And False, which we expect to
evaluate to False. We can see that e1 = True and that e2 = False, but our evaluation
relation does not relate v1 or v2 to any value. This is because, strictly speaking, we do
not know that True ⇒ True.
Of course, we would like that to be the case and, since we are in the process of
dening the language, we can make it so. We simply need to declare it in an operational
semantics rule.
(Value Rule)
v⇒v
CHAPTER 2. OPERATIONAL SEMANTICS 9
The value rule above is an axiom declaring that any value always evaluates to itself.
This satises our requirement and allows us to make use of the And rule. Using this
formal logic approach, we can now prove that True And (False And True) ⇒ False
as follows:
False ⇒ False True ⇒ True
True ⇒ True False And True ⇒ False
True And (False And True) ⇒ False
One may read the above proof tree as an explanation as to why True And (False
And True) evaluates to False. We can choose to read that proof as follows: “True And
(False And True) evaluates to False by the And rule because we know True evaluates
to True, that False And True evaluates to False, and that the logical and of true and
false is false. We know that False And True evaluates to False by the And rule because
True evaluates to True, False evaluates to False, and the logical and of true and false
is false.”
An equivalent and similarly informal format for the above is:
True And ( False And True ) ⇒ False, because by the And rule
True ⇒ True, and
( False And True ) ⇒ False, the latter because
True ⇒ True, and
False ⇒ False
The important thing to note about all three of these representations is that they are
describing a proof tree. The proof tree consists of nodes which represent the application
of logical rules with preconditions as their children. To complete our boolean language,
we dene the ⇒ relation using a complete set of operational semantics rules:
(Value Rule)
v⇒v
e⇒v
(Not Rule)
Not e ⇒ the negation of v
e1 ⇒ v1 e2 ⇒ v2
(And Rule)
e1 And e2 ⇒ the logical and of v1 and v2
The rules for Or and Implies are left as an exercise to the reader (see Exercise 2.4).
These rules form a proof system as is found in mathematical logic. Logical rules
express incontrovertible logical truths. A proof of e ⇒ v amounts to constructing a
sequence of rule applications such that, for any given application of a rule, the items
above the line appeared earlier in the sequence and such that the nal rule application
is e ⇒ v. A proof is structurally a tree, where each node is a rule, and the subtree rules
have conclusions which exactly match what the parent’s assumptions are. For a proof
CHAPTER 2. OPERATIONAL SEMANTICS 10
tree of e ⇒ v, the root rule has as its conclusion e ⇒ v. Note that all leaves of a proof
tree must be axioms. A tree with a non-axiom leaf is not a proof.
Notice how the above proof tree is expressing how this logic expression could be
computed. Proofs of e ⇒ v corresponds closely to how the execution of e produces the
value v as result. The only dierence is that “execution” starts with e and produces the
v, whereas a proof tree describes a relation between e and v, not a function from e to v.
Lemma 2.1. The boolean language is deterministic: if e ⇒ v and e ⇒ v ′ , then v = v ′ .
Proof. By induction on the height of the proof tree.
Lemma 2.2. The boolean language is normalizing: For all boolean expressions e, there
is some value v where e ⇒ v.
Proof. By induction on the size of e.
When a proof e ⇒ v can be constructed for some program e, we say that e converges.
When no such proof exists, e diverges. Because the boolean language is normalizing, all
programs in that language are said to converge. Some languages (such as OCaml) are not
normalizing; there are syntactically legal programs for which no evaluation proof exists.
An example of a OCaml program which is divergent is let rec f x = f x in f 0;;.
2.2.2 Abstract Syntax
Our operational semantics rules have expressed the evaluation relation in terms of con-
crete syntax using metavariables. Operators, such as the inx operator And, have ap-
peared in textual format. This is a good representation for humans to read because it
appeals to our intuition; it is not, however, an ideal computational representation. We
read True And False as “perform a logical and with operands True and False”. We
read True And (False And True) as “perform a logical and with operands False and
True and then perform a logical and with operands True and the result of the last oper-
ation.” If we are to write programs (such as interpreters) to work with our language, we
need a representation which more accurately describes how we think about the program.
The abstract syntax of a language is such a representation. A term in an abstract
syntax is represented as a syntax tree in which each operation to be performed is a
node and each operand to that operation is a child of that node. In order to represent
abstract syntax trees for the boolean language, we might use the following OCaml data
type:
type boolexp =
True | False |
Not of boolexp |
And of boolexp * boolexp |
Or of boolexp * boolexp |
Implies of boolexp * boolexp;;
To understand how the abstract and concrete syntax relate, consider the following
examples:
CHAPTER 2. OPERATIONAL SEMANTICS 11
Example 2.1.
Concrete:
True
True
Abstract:
True
Example 2.2.
Concrete:
True And False And
Abstract: True False
And(True, False)
Example 2.3.
Concrete: Implies
(True And False) Implies
((Not True) And False) And And
Abstract: True False Not False
Implies( And(True,False) ,
And(Not(True),False) ) True
There is a simple and direct relationship between the concrete syntax of a language
and the abstract syntax. As mentioned above, the abstract syntax is a form which more
directly represents the operations being performed whereas the concrete syntax is the
form in which the operations are actually expressed. Part of the process of compiling
or interpreting a program is to translate the concrete syntax term (source le) into an
abstract syntax term (AST) in order to manipulate it. We dene a relation JcK = a
to map concrete syntax form c to abstract syntax form a (in this case for the boolean
language):
JTrueK = True
JFalseK = False
JNot eK = Not(e)
Je1 And e2 K = And(Je1 K, Je2 K)
Je1 Or e2 K = Or(Je1 K, Je2 K)
Je1 Implies e2 K = Implies(Je1 K, Je2 K)
For example, this relation indicates the following:
CHAPTER 2. OPERATIONAL SEMANTICS 12
J(True And False) Implies ((Not True) And False)K
= Implies( JTrue And FalseK, J(Not True) And FalseK )
= Implies( And(JTrueK, JFalseK), And(JNot TrueK, JFalseK) )
= Implies( And(True, False), And(Not(JTrueK), False) )
= Implies( And(True, False), And(Not(True), False) )
The grammar we give is ambiguous in that there are multiple parse trees for some
concrete expressions, but we implicitly assume the usual operator precedence applies
with And binding tighter than Or binding tighter than Implies. Consider the following
examples:
Example 2.4.
Concrete: And
True Or True And False
Or False
Abstract:
And(Or(True,True),False) True True
True ⇒ True True ⇒ True
True Or True ⇒ True False ⇒ False
True Or True And False ⇒ False
Example 2.5.
Concrete: Or
True Or (True And False)
True And
Abstract:
Or(True,And(True,False)) True False
True ⇒ True False ⇒ False
True ⇒ True True And False ⇒ False
True Or (True And False) ⇒ True
The expression in example 2.4 will evaluate to False because one must evaluate the Or
operation rst and then evaluate the And operation using the result. Example 2.5, on the
other hand, performs the operations in the opposite order. Note that in both examples,
though, the parentheses themselves are no longer overtly present in the abstract syntax.
This is because they are implicitly represented in the structure of the AST; that is, the
AST in example 2.5 would not have the shape that it has if the parentheses were not
present in the concrete syntax of the form.
CHAPTER 2. OPERATIONAL SEMANTICS 13
In short, parentheses merely change how expressions are grouped. In example 2.5, the
only rule we can match to the entire expression is the Or rule; the And rule obviously can’t
match because the left parentheses would be part of e1 while the right parenthesis would
be part of e2 (and expressions with unmatched parentheses make no sense). Similarly but
less obviously, example 2.4 can only match the And rule; the associativity implicitly forces
the Or rule to happen rst, giving the And operator that entire expression to evaluate.
This distinction is clearly and correspondingly represented in the ASTs of the examples,
a fact which is key to the applicability of operational semantics.
2.2.3 Operational Semantics and Interpreters
As alluded above, there is a very close relationship between an operational semantics
and an actual interpreter written in OCaml. Given an operational semantics dened via
the relation ⇒, there is a corresponding (OCaml) evaluator function eval.
Denition 2.2 (Faithful Implementation). A (OCaml) interpreter function eval faith-
fully implements an operational semantics e ⇒ v if:
e ⇒ v if and only if eval(JeK) returns result JvK.
To demonstrate this relationship, we will demonstrate the creation of an eval function
in OCaml. Our rst draft of the function will, for sake of simplicity, only consist of the
And rule and the value rule:
let eval exp =
match exp with
| True -> True
| False -> False
| And ( exp0 , exp1 ) ->
begin
match ( exp0 , exp1 ) with
| ( True , True ) -> True
| (_ , False ) -> False
| ( False , _ ) -> False
end
At rst glance, this function appears to have the behavior we desire. True evaluates to
True, False to False, and True And False to False. This is not, however, a complete
implementation.
To nd out why, consider the concrete term True And True And True. As we have
seen before, this translates to the abstract term And(And(True,True),True). When the
eval function receives that value as its parameter, it matches the value to the case And
and denes exp0 and exp1 as And(True,True) and True, respectively. We then enter the
inner match, and this match fails! None of the terms can match the tuple (exp0,exp1)
because exp0 is an entire expression and not just a value, as the match expression is
expecting.
Looking back at our attempts to write the And rule above, we can see why this eval
function is awed: this version of the function does not consider the operands of And
to be expressions - it expects them to be values. We can see, then, that the And clause
CHAPTER 2. OPERATIONAL SEMANTICS 14
in our function is a faithful implementation of Try 2 of our And rule, a rule which we
rejected precisely because it could not handle nested expressions.
How can we correct this problem? We are trying to write a faithful implementation of
our nal And rule, which relies on the evaluation of the And rule’s operands. Thus, in our
implementation, we must evaluate those operands; we make this possible by declaring
our evaluation function to be recursive.
let rec eval exp =
match exp with
| True -> True
| False -> False
| And ( exp0 , exp1 ) ->
begin
match ( eval exp0 , eval exp1 ) with
| ( True , True ) -> True
| (_ , False ) -> False
| ( False , _ ) -> False
end
Observe that, in the above code, we have changed very little. We modied the eval
function to be recursive. We also added a call to eval for each of the operands to the
And operation. That call alone is sucient to x the problem; the process of evaluating
those arguments represents the e1 ⇒ v1 and e2 ⇒ v2 preconditions on the And rule, while
the use of the resultings values in the tuple causes the match to be against v1 and v2
rather than e1 and e2 . The above code is a faithful implementation of the value rule and
the And rule.
We can now complete the boolean language interpreter by continuing the eval fuction
in the same form:
let rec eval exp =
match exp with
True -> True
| False -> False
| Not ( exp0 ) -> ( match eval exp0 with
True -> False
| False -> True )
| And ( exp0 , exp1 ) -> ( match ( eval exp0 , eval exp1 ) with
( True , True ) -> True
| (_ , False ) -> False
| ( False , _ ) -> False )
| Or ( exp0 , exp1 ) -> ( match ( eval exp0 , eval exp1 ) with
( False , False ) -> False
| (_ , True ) -> True
| ( True , _ ) -> True )
| Implies ( exp0 , exp1 ) -> ( match ( eval exp0 , eval exp1 ) with
( False , _ ) -> True
CHAPTER 2. OPERATIONAL SEMANTICS 15
| ( True , True ) -> True
| ( True , False ) -> False )
The only dierence between the operational semantics and the interpreter is that
the interpreter is a function. We start with the bottom-left expression in a rule, use
the interpreter to recursively produce the value(s) above the line in the rule, and nally
compute and return the value below the line in the rule.
Note that the boolean language interpreter above faithfully implements its opera-
tional semantics: e ⇒ v if and only if eval(JeK) returns JvK as result. We will go back
and forth between these two forms throughout the book. The operational semantics
form is used because it is independent of any particular programming language. The
interpreter form is useful because we can interpret real programs for nontrivial numbers
of steps, something that is dicult to do “on paper” with an operational semantics.
Denition 2.3 (Metacircular Interpreter). A metacircular interpreter is an inter-
preter for (possibly a subset of ) a language x that is written in language x.
Metacircular interpreters give you some idea of how a language works, but suer
from the non-foundational problems implied in Exercise 2.5. A metacircular interpreter
for Lisp (that is, a Lisp interpreter written in Lisp) is a classic programming language
theory exercise.
2.3 The F[ Programming Language
Now that we have seen how to dene and understand operational semantics, we will
begin to study our rst programming language: F[. F[ is a shunk (attened) pure func-
tional programming language. 2 It has integers, booleans, and higher-order anonymous
functions. In most ways F[ is much weaker than OCaml: there are no reals, lists, types,
modules, state, or exceptions.
F[ is untyped, and in this way is it actually more powerful than OCaml. It is
possible to write some programs in F[ that produce no runtime errors, but which will
not typecheck in OCaml. For instance, our encoding of recursion in Section 2.3.5 is not
typeable in OCaml. Type systems are discussed in Chapter 6. Because there are no
types, runtime errors can occur in F[, such as the application (5 3).
Although very simplistic, F[ is still Turing-complete. The concept of Turing-
completeness has been dened in numerous equivalent ways. One such denition is as
follows:
Denition 2.4 (Turing Completeness). A computational model is Turing-complete if
every partial recursive function can be expressed within it.
This denition, of course, requires a denition of partial recursive functions (also
known as computable functions). Without going into an extensive discussion of founda-
tional material, the following somewhat informal denition will suce:
2
Also, any readers familiar with the programming language C] as well as basic music theory should
nd this at least a bit humorous.
CHAPTER 2. OPERATIONAL SEMANTICS 16
Denition 2.5 (Partial Recursive Function). A function is a partial recursive function
if an algorithm exists to calculate it which has the following properties:
• The algorithm must have as its input a nite number of arguments.
• The algorithm must consist of a nite number of steps.
• If the algorithm is given arguments for which the function is dened, it must produce
the correct answer within a nite amount of time.
• If the algorithm is given arguments for which the function is not dened, it must
either produce a clear error or otherwise not terminate. (That is, it must not appear
to have produced an incorrect value for the function if no such value is dened.)
The above denition of a partial recursive function is a mathematical one and thus
does not concern itself with execution-specic details such as storage space or practical
execution time. No constraints are placed against the amount of memory a computer
might need to evaluate the function, the range of the arguments, or that the function
terminate before the heat death of the universe (so long as it would eventually terminate
for all inputs for which the function is dened).
The practical signicance of Turing-completeness is this: there is no computation
that could be expressed in another deterministic programming language that cannot be
expressed in F[.3 In fact, F[ is even Turing-complete without numbers or booleans. This
language, one with only functions and application, is known as the pure lambda-calculus
and is discussed briey in Section 2.4.4. No deterministic programming language can
compute more than the partial recursive functions.
2.3.1 F[ Syntax
We will take the same approach in dening F[ as we did in dening the boolean language
above. We start by describing the grammar of the F[ language to dene its concrete
syntax; the abstract syntax is deferred until Section 2.3.7. We can dene the grammar
of F[ using the following BNF:
3
This does not guarantee that the F[ representation will be pleasant. Programs written in F[ to
perform even fairly simplistic computations such as determining if one number is less than another are
excruciating, as we will see shortly.
CHAPTER 2. OPERATIONAL SEMANTICS 17
x ::= (a | b | . . . | z) lower-case letters
(A | B | . . . | Z capital letters
| a | b | ... | z lower-case letters
| 0 | 1 | ... | 9 digits
| _ | ’ | · · · )∗ other characters
v ::= x variable values
| True | False boolean values
| 0 | 1 | -1 | 2 | -2 | . . . integer values
| Function x → e function values
e ::= v value expressions
| (e) parenthesized expressions
| e And e | e Or e | Not e boolean expressions
|e + e|e - e|e = e| numerical expression
|ee application expression
| If e Then e Else e conditional expressions
| Let x = e In e let expression
| Let Rec f x = e In e recursive let expression
Note that in accordance with the above BNF, we will be using metavariables e, v,
and x to represent expressions, values, and variables respectively. Note the last point:
the metavariable x refers to an arbitrary F[ variable, not necessarily to the F[ variable
x.
Associativity in F[ works in a fashion very similar to OCaml. Function application,
for instance, is left associative, meaning that a b c has the same meaning as (a b) c.
As with any language, this associativity is signicant in that it aects how source code
is parsed into an AST.
2.3.2 Variable Substitution
The main feature of F[ is higher-order functions, which also introduces variables. Recall
that programs are computed by rewriting them:
( Function x -> x + 2)(3 + 2 + 5) ⇒ 12
because
3 + 2 + 5 ⇒ 10
because
3 + 2 ⇒ 5
and
5 + 5 ⇒ 10
and
10 + 2 ⇒ 12
Note how in this example, the argument is substituted for the variable in the body—this
gives us a rewriting interpreter. In other words, F[ functions compute by substituting
the actual argument for the for parameter; for example,
CHAPTER 2. OPERATIONAL SEMANTICS 18
( Function x -> x + 1) 2
will compute by substituting 2 for x in the function’s body x + 1, i.e. by computing 2 + 1.
This is not a very ecient method of computing, but it is a very simple and accurate
description method, and that is what operational semantics is all about – describing
clearly and unambiguously how programs are to compute.
Bound and Free Occurrences of Variables We need to be careful about how
variable substitution is dened. For instance,
(Function x -> Function x -> x) 3
should not evaluate to Function x -> 3 since the inner x is bound by the inner param-
eter. To correctly formalize this notion, we need to make the following denitions.
Denition 2.6 (Variable Occurrence). A variable use x occurs in e if x appears some-
where in e. Note we refer only to variable uses, not denitions.
Denition 2.7 (Bound Occurrence). Any occurrences of variable x in the expression
Function x -> e
are bound, that is, any free occurrences of x in e are bound occurrences in this expression.
Similarly, in the expression
Let Rec f x = e1 In e2
occurrences of f and x are bound in e1 and occurrences of f are bound in e2 . Note that
x is not bound in e2 , but only in e1 , the body of the function.
Denition 2.8 (Free Occurrence). A variable x occurs free in e if it has an occurrence
in e which is not a bound occurrence.
Let’s look at a few examples of bound versus free variable occurrences.
Example 2.6.
Function x -> x + 1
x is bound in the body of this function.
Example 2.7.
Function x -> Function y -> x + y + z
x and y are bound in the body of this function. z is free.
CHAPTER 2. OPERATIONAL SEMANTICS 19
Example 2.8.
Let z = 5 In Function x -> Function y -> x + y + z
x, y, and z are all bound in the body of this function. x and y are bound by
their respective function declarations, and z is bound by the Let statement. Note
that, while F[ contains Let as syntax, it can be dened as a macro (see Section 2.3.4
below). Binding rules work similarly for Functions and Let statements.
Example 2.9.
Function x -> Function x -> x + x
x is bound in the body of this function. Note that both x usages are bound
to the inner variable x.
Denition 2.9 (Closed Expression). An expression e is closed if it contains no free
variable occurrences. All programs we execute are closed (no link-time errors) – non-
closed programs don’t diverge, we can’t even contemplate executing them because they are
not in the domain of the evaluation relation.
Of the examples above, Examples 2.6, 2.8, and 2.9 are closed expressions. Example
2.7 is not a closed expression.
Now that we have an understanding of bound and free variables, we can give a formal
denition of variable substitution.
Denition 2.10 (Variable Substitution). The variable substitution of x for e′ in e,
denoted e[e′ /x], is the expression resulting from the operation of replacing all free occur-
rences of x in e with e′ . For now, we assume that e′ is a closed expression.
Here is an equivalent inductive denition of substitution:
x[v/x] = v
x′ [v/x] = x′ x 6= x′
(Function x → e)[v/x] = (Function x → e)
(Function x′ → e)[v/x] = (Function x′ → e[v/x]) x= 6 x′
(Let x = e1 In e2 )[v/x] = Let x = e1 [v/x] In e2
(Let x′ = e1 In e2 )[v/x] = Let x′ = e1 [v/x] In e2 [v/x] x 6= x′
n[v/x] = n for n ∈ Z
True[v/x] = True
False[v/x] = False
(e1 + e2 )[v/x] = e1 [v/x] + e2 [v/x]
(e1 And e2 )[v/x] = e1 [v/x] And e2 [v/x]
..
.
For example, let us consider a simple application of a function: (Function x -> x +
1) 2. We know that, to evaluate this function, we simply replace all instances of x in
the body with 2. This is written (x + 1)[2/x]. Given the above denition, we can
conclude that the result must be 3.
CHAPTER 2. OPERATIONAL SEMANTICS 20
While this may not seem like an illuminating realization, the fact that this is mathe-
matically discernable gives us a starting point for more complex subsitutions. Consider
the following example.
Example 2.10.
Expression:
(Function x -> Function y -> (x + x + y)) 5
Substitution:
(Function y -> (x + x + y))[5/x]
= (Function y -> (x + x + y)[5/x])
= Function y -> (x[5/x] + x[5/x] + y[5/x])
= Function y -> (5 + 5 + y)
α-conversion
In Example 2.9, we saw that it is possible for two variables to share the same name.
The variables themselves are, of course, distinct and follow the same rules of scope in
F[ as they do in OCaml. But reading expressions which make frequent use of the same
variable name for dierent variables can be very disorienting. For example, consider the
following expression.
Let Rec f x =
If x = 1 Then
( Function f -> f ( x - 1)) ( Function x -> x )
Else
f ( x - 1)
In f 100
How does this expression evaluate? It is a bit dicult to tell simply by looking at it
because of the tricky bindings. We can make it much easier to understand by using
dierent names. α-conversion is the process of replacing a variable denition and all
occurrences bound to it with a variable of a dierent name.
Example 2.11.
Function x -> x + 1
becomes
Function z -> z + 1
Example 2.11 shows a simple case in which x is substituted for z. For cases in which
the same variable name is used numerous times, we can use the same approach. Consider
Example 2.12 in which the inner variable x is α-converted to z.
Example 2.12.
Function x -> Function x -> x
becomes
Function x -> Function z -> z
Similarly, we could rename the outer variable to z as shown in Example 2.13. Note
that in this case, the occurrence of x is not changed, as it is bound by the inner variable
and not the outer one.
CHAPTER 2. OPERATIONAL SEMANTICS 21
Example 2.13.
Function x -> Function x -> x
becomes
Function z -> Function x -> x
Let’s gure out what variable occurrences are bound to which function in our previous
confusing function and rewrite the function in a clearer way by using α-conversion. One
possible result is as follows:
Let Rec f x =
If x = 1 Then
(Function z -> z (x - 1)) (Function y -> y)
Else
f (x - 1)
In f 100
Now it’s much easier to understand what is happening. If the function f is applied
to an integer which is not 1, it merely applies itself again to the argument which is one
less than the one it received. Since we are evaluating f 100, that results in f 99, which
in turn evaluates f 98, and so on. Eventually, f 1 is evaluted.
When f 1 is evaluated, we explore the other branch of the If expression. We know
that x is 1 at this point, so we can see that the evaluated expression is (Function z ->
z 0) (Function y -> y). Using substitution gives us (Function y -> y) 0, which in
turn gives us 0. So we can conclude that the expression above will evaluate to 0.
Observe, however, that we did not formally prove this; so far, we have been treating
substitution and other operations in a cavalier fashion. In order to create a formal proof,
we need a set of operational semantics which dictates how evaluation works in F[. Section
2.3.3 walks through the process of creating an operational semantics for the F[ language
and gives us the tools needed to prove what we concluded above.
2.3.3 Operational Semantics for F[
We are now ready to begin dening operational semantics for F[. For the same reasons
as in our boolean language, we will need a rule which relates values to values in ⇒:
(Value Rule)
v⇒v
We can also dene boolean operations for F[ in the same way as we did for the boolean
language above. Note, however, that not all values in F[ are booleans. Fortunately, our
denition of the rules addresses this for us, as there is (for example) no logical and of
the values 5 and 3. That is, we know that these rule only apply to F[ boolean values
because they use operations which are only dened for F[ boolean values.
CHAPTER 2. OPERATIONAL SEMANTICS 22
e⇒v
(Not Rule)
Not e ⇒ the negation of v
e1 ⇒ v1 e2 ⇒ v2
(And Rule)
e1 And e2 ⇒ the logical and of v1 and v2
..
.
We can also dene operations over integers in much the same way. For sake of clarity,
we will explicitly restrict these rules such that they operate only on expressions which
evaluate to integers.
e1 ⇒ v1 , e2 ⇒ v2 where v1 , v2 ∈ Z
(+ Rule)
e1 + e2 ⇒ the integer sum of v1 and v2
e1 ⇒ v1 , e2 ⇒ v2 where v1 , v2 ∈ Z
(- Rule)
e1 - e2 ⇒ the integer dierence of v1 and v2
As with the boolean rules, observe that these rules allow the ⇒ relation to be ap-
plied recursively: 5 + (4 - 3) can be evaluated using the + rule because 4 - 3 can be
evaluated using the - rule rst.
These rules allow us to write F[ programs containing boolean expressions or F[ pro-
grams containing integer expressions, but we currently have no way to combine the two.
There are two mechanisms we use to mix the two in a program: conditional expressions
and comparison operators. The only comparison operator in F[ is the = operator, which
compares the values of integers. We dene the = rule as follows.
e1 ⇒ v1 , e2 ⇒ v2 where v1 , v2 ∈ Z
(= Rule)
e1 = e2 ⇒ True if v1 and v2 are identical, else False
Note that the = rule is dened only where v1 and v2 are integers. Due to this
constraint, the expression True = True is not evaluatable in F[. This is, of course, a
matter of choice; as a language designer, one may choose to remove that constraint and
allow boolean values to be compared directly. To formalize this, however, a change to
the rules would be required. A faithful implementation of F[ using the above = rule is
required to reject the expression True = True.
An intuitive denition of a conditional expression is that it evalutes to the value of
one expression if the boolean is true and the value of the other expression if the boolean
is false. While this is true, the particulars of how this is expressed in the rule are vital.
Let us consider the following awed attempt at a conditional expression rule:
e1 ⇒ v1 e2 ⇒ v2 e3 ⇒ v3
(Flawed If Rule)
If e1 Then e2 Else e3 ⇒ v2 if v1 is True, v3 otherwise
CHAPTER 2. OPERATIONAL SEMANTICS 23
It seems that this rule allows us to evaluate many of the conditional expressions we
want to evaluate. But let us consider this expression:
If True Then 0 Else (True + True)
If we attempted to apply the above rule to the previous expression, we would nd that
the precondition e3 ⇒ v3 would not hold; there is no rule which can relate True + True ⇒
v for any v since the + rule only applies to integers. Nonetheless, we want the expression
to evaluate to 0. So how can we achieve this result?
In this case, we have no choice but to write two dierent rules with distinct pre-
conditions. We can capture all of the relationships in the previous rule and yet allow
expressions such as the previous to evaluate by using the following two rules:
e1 ⇒ True, e2 ⇒ v2
(If True Rule)
If e1 Then e2 Else e3 ⇒ v2
e1 ⇒ False, e3 ⇒ v3
(If False Rule)
If e1 Then e2 Else e3 ⇒ v3
Again, the key dierence between these two rules is that they have dierent sets of
preconditions. Note that the If True rule does not evaluate e3 , nor does the If False
rule evaluate e2 . This allows the untraveled logic paths to contain unevaluatable expres-
sions without necessarily preventing the expression containing them from evaluating.
Application
We are now ready to approach one of the most dicult F[ rules: application. How can
we formalize the evaluation of an expression like (Function x -> x + 1) (5 + 2)?
We saw in Section 2.3.2 that we can evaluate a function application by using variable
substitution. As we have a mathematical denition for the substitution operation, we
can base our function application rule around it.
Suppose we wish to evaluate the above expression. We can view application in two
parts: the function being applied and the argument to the function. We wish to know to
what the expression evaluates; thus, we are trying to establish that e1 e2 ⇒ v for some v.
?
(Application Rule (Part 1))
e1 e2 ⇒ v
In our boolean operations, we needed to evaluate the arguments before attempting
an operation over them (in order to allow recursive expressions). The same is true of our
application rule; in (Function x -> x + 1) (5 + 2), we must evaluate 5 + 2 before
it can be used as an argument.4 We must do likewise with the function we are applying.
4
Actually, some languages would perform substitution before evaluating the expression, but F[ and
most traditional languages do not. Discussion of this approach is handled in Section 2.3.6.
CHAPTER 2. OPERATIONAL SEMANTICS 24
e1 ⇒ v1 e2 ⇒ v2 ?
(Application Rule (Part 2))
e1 e2 ⇒ v
We obviously aren’t nished, though, as we still don’t have any preconditions which
allow us to relate v to something. Additionally, we know we will need to use variable
substitution, but we have no metavariables representing F[ variables in the above rule.
We can x this by reconsidering how we evaluate the rst argument; we know that the
application rule only works when applying functions. In restricting our rule to applying
functions, we can name some metavariables to describe the function’s contents.
e1 ⇒ Function x -> e e2 ⇒ v2 ?
(Application Rule (Part 3))
e1 e2 ⇒ v
In the above rule, x is the metavariable representing the function’s variable while
e represents the function’s body. Now that we have this information, we can dene
function application in terms of variable substitution. When we apply Function x ->
x + 1 to a value such as 7, we wish to replace all instances of x, the function’s variable,
in the function’s body with 7, the provided argument. Formally,
e1 ⇒ Function x -> e e2 ⇒ v2 e[v2 /x] ⇒ v
(Application Rule)
e1 e2 ⇒ v
F[ Recursion
We now have a very complete set of rules for the F[ language. We do not, however,
have a rule for Let Rec. As we will see, Let Rec is not actually necessary in basic F[;
it is possible to build recursion out of the rules we already have. Later, however, we will
create variants of F[ with type systems in which it will be impossible for that approach
to recursion to work. For that reason as well as our immediate convenience, we will
dene the Let Rec rule now.
Again, we start with an iterative approach. We know that we want to be able to eval-
uate the Let Rec expression, so we provide metavariables to represent the components
of the expression itself.
(Recursive Application Rule (Part 1))
?
Let Rec f x = e1 In e2 ⇒ v
Let us consider what we wish to accomplish. Consider for a moment a recursive
approach to the summation of the numbers between 1 and 5:
CHAPTER 2. OPERATIONAL SEMANTICS 25
Let Rec f x =
If x = 1 Then
1
Else
f (x - 1) + x
In f 5
If we focus on the last line (In f 5), we can see that we want the body of the
recursive function to be applied to the value 5. We can write our rule accordingly by
replacing f with the function’s body. We must make sure to use the same metavariable
to represent the function’s argument in order to allow the new function body’s variable
to be captured. We reach the following rule.
(Recursive Application Rule (Part 2))
e2 [(Function x -> e1 )/f ] ⇒ v
Let Rec f x = e1 In e2 ⇒ v
We can test our new rule by applying it to our recursive summation above.
???
f (5-1) ⇒ v ′ 5 ⇒ 5
5 = 1 ⇒ False f (5-1) + 5 ⇒ v
Function x -> · · · ⇒ Function x -> · · · 5 ⇒ 5 If 5 = 1 Then 1 Else f (5-1) + 5 ⇒ v
(Function x -> If x = 1 Then 1 Else f (x-1) + x) 5 ⇒ v
Let Rec f x = If x = 1 Then 1 Else f (x-1) + x In f 5 ⇒ v
As foreshadowed by its label above, our recursion rule is not yet complete. When we
reach the evaluation of f (5-1), we are at a loss; f is not bound. Without a binding for
f, we have no way of repeating the recursion.
In addition to replacing the invocation of f with function application in e2 , we need
to ensure that any occurrences of f in the body of the function itself are bound. To
what do we bind them? We could try to replace them with function applications as
well, but this leads us down a rabbit hole; each function application would require yet
another replacement. We can, however, replace applications of f in e1 with recursive
applications of f by reintroducing the Let Rec syntax. This leads us to the following
application rule:
(Recursive Application Rule (Part 3))
e2 [Function x -> e1 [(Let Rec f x = e1 In f )/f ]/f ] ⇒ v
Let Rec f x = e1 In e2 ⇒ v
While this makes a certain measure of sense, it isn’t quite correct. In Section 2.3.2,
we saw that substitution must replace a variable with a value, while the Let Rec term
CHAPTER 2. OPERATIONAL SEMANTICS 26
above is an expression. Fortunately, we have functions as values; thus, we can put the
expression inside of a function and ensure that we call it with the appropriate argument.
(Recursive Application Rule)
e2 [Function x -> e1 [(Function x -> Let Rec f x = e1 In f x)/f ]/f ] ⇒ v
Let Rec f x = e1 In e2 ⇒ v
Now, instead of encountering f (5-1) when we evaluate the summation example, we
encounter Let Rec f x = If x = 1 Then 1 Else f (x-1) + x In f (5-1). This al-
lows us to recurse back into the Let Rec rule. Eventually, we may reach a branch which
does not evaluate the Else side of the conditional expression, in which case that Let
Rec is not expanded (allowing us to terminate). Each application of the rule eectively
“unrolls” one level of recursion.
In summary, we have the following operational semantics for F[ (excluding some
repetetive rules such as the - rule):
(Value Rule)
v⇒v
e⇒v
(Not Rule)
Not e ⇒ the negation of v
e1 ⇒ v1 e2 ⇒ v2
(And Rule)
e1 And e2 ⇒ the logical and of v1 and v2
e1 ⇒ v1 , e2 ⇒ v2 where v1 , v2 ∈ Z
(+ Rule)
e1 + e2 ⇒ the integer sum of v1 and v2
e1 ⇒ v1 , e2 ⇒ v2 where v1 , v2 ∈ Z
(= Rule)
e1 = e2 ⇒ True if v1 and v2 are identical, else False
e1 ⇒ True, e2 ⇒ v2
(If True Rule)
If e1 Then e2 Else e3 ⇒ v2
e1 ⇒ False, e3 ⇒ v3
(If False Rule)
If e1 Then e2 Else e3 ⇒ v3
e1 ⇒ Function x -> e, e2 ⇒ v2 , e[v2 /x] ⇒ v
(Application Rule)
e1 e2 ⇒ v
e1 ⇒ v1 e2 [v1 /x] ⇒ v2
(Let Rule)
Let x = e1 In e2 ⇒ v2
(Let Rec)
e2 [Function x -> e1 [(Function x -> Let Rec f x = e1 In f x)/f ]/f ] ⇒ v
Let Rec f x = e1 In e2 ⇒ v
CHAPTER 2. OPERATIONAL SEMANTICS 27
Let us consider a few examples of proof trees using the F[ operational semantics.
Example 2.14.
Expression:
If 3 = 4 Then 5 Else 4 + 2
Proof:
3⇒3 4⇒4 4⇒4 2⇒2
3 = 4 ⇒ False 4 + 2⇒6
If 3 = 4 Then 5 Else 4 + 2 ⇒ 6
Example 2.15.
Expression:
(Function x -> If 3 = x Then 5 Else x + 2) 4
Proof:
by Example 2.14
Function x -> · · · ⇒ Function x -> · · · 4 ⇒ 4 If 3 = 4 Then 5 Else 4 + 2 ⇒ 6
(Function x -> If 3 = x Then 5 Else x + 2) 4 ⇒ 6
Example 2.16.
Expression:
(Function f -> Function x -> f(f x))(Function y -> y - 1) 4
Proof:
Due to the size of the proof, it is broken into multiple parts. We use v ⇒ F as an abbreviation
for v ⇒ v (when v is lengthy) for brevity.
Part 1:
Function f -> Function x -> f(f x) ⇒ F Function y -> y - 1 ⇒ F (Function x -> (Function y -> y - 1) ((Function y -> y - 1) x)) ⇒ F
(Function f -> Function x -> f(f x))(Function y -> y - 1) ⇒ (Function x -> (Function y -> y - 1) ((Function y -> y - 1) x))
Part 2:
4⇒4 1⇒1
Function y -> y - 1 ⇒ F 4 ⇒ 4 4 - 1⇒3 3⇒3 1⇒1
Function y -> y - 1 ⇒ F ((Function y -> y - 1) 4) ⇒ 3 3 - 1⇒2
(by part 1) 4⇒4 (Function y -> y - 1) ((Function y -> y - 1) 4) ⇒ 2
(Function f -> Function x -> f(f(x)))(Function y -> y - 1) 4 ⇒ 2
Interact with F[. Tracing through recursive evaluations is dicult, and there-
fore the reader should invest some time in exploring the semantics of Let Rec.
A good way to do this is by using the F[ interpreter. Try evaluating the expression we
looked at above:
# Let Rec f x =
If x = 1 Then 1 Else x + f (x - 1)
In f 3;;
==> 6
Another interesting experiment is to evaluate a recursive function without applying
it. Notice that the result is equivalent to a single application of the Let Rec rule. This
is a good way to see how the “unwrapping” actually takes place:
CHAPTER 2. OPERATIONAL SEMANTICS 28
# Let Rec f x =
If x = 1 Then 1 Else x + f (x - 1)
In f;;
==> Function x ->
If x = 1 Then
1
Else
x + (Let Rec f x =
If x = 1 Then
1
Else
x + (f) (x - 1)
In
f) (x - 1)
As we mentioned before, one of the main benets of dening an operational semantics
for a language is that we can rigorously verify claims about that language. Now that we
have dened the operational semantics for F[, we can prove a few things about it.
Lemma 2.3. F[ is deterministic.
Proof. By inspection of the rules, at most one rule can apply at any time.
Lemma 2.4. F[ is not normalizing.
Proof. To show that a language is not normalizing, we simply show that there is some e
such that there is no v with e ⇒ v. Let e be (Function x -> x x)(Function x -> x
x). e ; v for any v. Thus, F[ is not normalizing.
The expression in this proof is a very interesting one which we will examine in more
detail in Section 2.3.5. It does not evaluate to a value because each step in its evaluation
produces itself as a precondition. This is roughly analogous to trying to prove proposition
A by using A as a given.
In this case, the expression does not evaluate because it never runs out of work to do.
This is not the only kind of non-normalizing expression which appears in F[; another
kind consists of those expressions for which no evaluation rule applies. For example, (4
3) is a simpler expression that is non-normalizing. No rule exists to evaluate (e1 e2 )
when e1 is not a function expression.
Both of these cases look like divergence as far as the operational semantics are con-
cerned. In the case of an interpreter, however, these two kinds of expressions behave
dierently. The expression which never runs out of work to do typically causes an in-
terpreter to loop innitely (as most interpreters are not clever enough to realize that
they will never nish). The expressions which attempt application to non-functions, on
the other hand, cause an interpreter to provide a clear error in the form of an excep-
tion. This is because the error here is easily detectable; the interpreter, in attempting
to evaluate these expressions, can quickly discover that none of its available rules apply
to the expression and respond by raising an exception to that eect. Nonetheless, both
are theoretically equivalent.
CHAPTER 2. OPERATIONAL SEMANTICS 29
2.3.4 The Expressiveness of F[
F[ doesn’t have many features, but it is possible to do much more than it may seem.
As we said before, F[ is Turing complete, which means, among other things, that any
OCaml operation may be encoded in F[. We can informally represent encodings in our
interpreter as macros using OCaml let statements. A macro is equivalent to a statement
like “let F be Function x -> . . . ”
Logical Combinators First, there are the classic logical combinators, simple func-
tions for recombining data.
combI = Function x -> x
combK = Function x -> Function y -> x
combS = Function x -> Function y -> Function z -> (x z) (y z)
combD = Function x -> x x
Tuples Tuples and lists are encodable from just functions, and so they are not needed
as primitives in a language. Of course for an ecient implementation you would want
them to be primitives; thus doing this encoding is simply an exercise to better understand
the nature of functions and tuples. We will dene a 2-tuple (pairing) constructor; From a
pair you can get a n-tuple by building it from pairs. For example, (1, (2, (3, 4))) represents
the 4-tuple (1, 2, 3, 4).
First, we need to dene a pair constructor, pr. A rst (i.e., slightly buggy) approxi-
mation of the constructor is as follows.
def
pr (e1 , e2 ) = Function x -> x e1 e2
def
We use the notation a = b to indicate that a is an abbreviation for b. For example,
we might have a problem in which the incrementer function is commonly used; it would
def
make sense, then, to dene inc = Function x -> x + 1. Note that such abbreviations
do not change the underlying meaning of the expression; it is simply for convenience.
The same concept applies to macro denitions in programming languages. By creating
a new macro, one is not changing the math behind the programming language; one is
merely dening a more terse means of expressing a concept.
Based on the previous denition of pr, we can dene the following macros for pro-
jection:
def
left (e) = e (Function x -> Function y -> x)
def
right (e) = e (Function x -> Function y -> y)
Now let’s take a look at what’s happening. pr takes a left expression, e1 , and a right
expression, e2 , and packages them into a function that applies its argument x to e1 and
CHAPTER 2. OPERATIONAL SEMANTICS 30
e2 . Because functions are values, the result won’t be evaluated any further, and e1 and
e2 will be packed away in the body of the function until it is applied. Thus pr succeeds
in “storing” e1 and e2 .
All we need now is a way to get them out. For that, look at how the projection
operations left and right are dened. They’re both very similar, so let’s concentrate
only on the projection of the left element. left takes one of our pairs, which is encoded
as a function, and applies it to a curried function that returns its rst, or leftmost,
element. Recall that the pair itself is just a function that applies its argument to e1 and
e2 . So when the curried left function that was passed in is applied to e1 and e2 , the
result is e1 , which is exactly what we want. right is similar, except that the curried
function returns its second, or rightmost, argument.
Before we go any further, there is a technical problem involving our encoding of pr.
Suppose e1 or e2 contain a free occurrence of x when pr is applied. Because pr is dened
as Function x -> x e1 e2 , any free occurrence x contained in e1 or e2 will become
bound by x after pr is applied. This is known as variable capture. To deal with capture
here, we need to change our denition of pr to the following.
def
pr (e1 , e2 ) = (Function e1 -> Function e2 -> Function x -> x e1 e2) e1 e2
This way, instead of textually substituting for e1 and e2 directly, we pass them in as
functions. This allows the interpreter evaluate e1 and e2 to values before passing them
in, and also ensures that e1 and e2 are closed expressions. This eliminates the capture
problem, because any occurrence of x is either bound by a function declaration inside e1
or e2 , or was bound outside the entire pr expression, in which case it must have already
been replaced with a value at the time that the pr subexpression is evaluated. Variable
capture is an annoying problem that we will see again in Section 2.4.
Now that we have polished our denitions, let’s look at an example of how to use
these encodings. First, let’s create the pair p as (4, 5).
def
p = pr (4, 5) ⇒ Function x -> x 4 5
Now, let’s project the left element of p.
left p ≡ (Function x -> x 4 5) (Function x -> Function y -> x)
We use the notation a ≡ b to indicate the expansion of a specic macro instance. In this
case, left p expands to what we see above, which then becomes
(Function x -> Function y -> x) 4 5 ⇒ 4.
CHAPTER 2. OPERATIONAL SEMANTICS 31
right
head
left
0
False 1 False 2 False 3 True 0
Figure 2.1: Lists Implemented Via Pairs
This encoding works, and has all the expressiveness of real tuples. There are, nonethe-
less, a few problems with it. First of all, consider
left (Function x -> 0) ⇒ 0.
We really want the interpreter to produce a run-time error here, because a function is
not a pair.
Similarly, suppose we wrote the program right(pr(3, pr(4, 5))). One would
expect this expression to evaluate to pr(4, 5), but remember that pairs are not values in
our language, but simply encodings, or macros. So in fact, the result of the computation
is Function x -> x 4 5. We can only guess that this is intended to be a pair. In this
respect, the encoding is awed, and we will, in Chapter 3, introduce “real” n-tuples into
an extension of F[ to alleviate these kinds of problems.
Lists Lists can also be implemented via pairs. In fact, pairs of pairs are technically
needed because we need a ag to mark the end of list. The list [1; 2; 3] is rep-
resented by pr (pr(false,1), pr (pr(false,2), pr (pr(false,3), emptylist)))
def
where emptylist = pr(pr(true,0),0). The true/false ag is used to mark the end of
the list: only the empty list is agged true. The implementation is as follows.
def
cons (x, y) = pr(pr(Bool false, x), y)
def
emptylist = pr(pr(Bool true, Int 0),Int 0)
def
head x = right(left x)
def
tail x = right x
def
isempty l = (left (left l))
def
length = Let Rec len x =
If isempty(x) Then 0 Else 1 + len (tail x) In len
In addition to tuples and lists, there are several other concepts from OCaml that we
can encode in F[. We review a few of these encodings below. For brevity and readability,
we will switch back to the concrete syntax.
CHAPTER 2. OPERATIONAL SEMANTICS 32
Functions with Multiple Arguments Functions with multiple arguments are done
with currying just as in OCaml. For example
Function x -> Function y -> x + y
The Let Operation Although F[ includes syntax for Let, it is quite simple to encode:
def
(Let x = e In e′ ) = (Function x -> e′ ) e
For example,
(Let x = 3 + 2 In x + x) ≡ (Function x -> x + x) (3 + 2) ⇒ 10.
Sequencing Notice that F[ has no sequencing (;) operation. Because F[ is a pure
functional language, writing e; e′ is really just equivalent to writing e′ , since e will
never get used. Hence, sequencing really only has meaning in languages with side-eects.
Nonetheless, sequencing is denable in the following manner.
def
e; e′ = (Function n -> e′ ) e,
where n is chosen so as not to be free in e′ . This will rst execute e, throw away the
value, and then execute e′ , returning its result as the nal result of e; e′ .
Freezing and Thawing We can stop and re-start computation at will by freezing and
thawing.
def
Freeze e = Function n -> e
def
Thaw e = e 0
We need to make sure that n is a fresh variable so that it is not free in e. Note that
the 0 in the application could be any value. Freeze e freezes e, keeping it from being
computed. Thaw e starts up a frozen computation. As an example,
Let x = Freeze (2 + 3) In (Thaw x) + (Thaw x)
CHAPTER 2. OPERATIONAL SEMANTICS 33
This expression has same value as the equivalent expression without the freeze and thaw,
but the 2 + 3 is evaluated twice. Again, in a pure functional language the only dierence
is that freezing and thawing is less ecient. In a language with side-eects, if the frozen
expression causes a side-eect, then the freeze/thaw version of the function may produce
results dierent from those of the original function, since the frozen side-eects will be
applied as many times as they are thawed.
2.3.5 Russell’s Paradox and Encoding Recursion
F[ has a built-in Let Rec operation to aid in writing recursive functions, but its actually
not needed because recursion is denable in F[. The encoding is a non-obvious one,
and so before we introduce it, we present some background information. As we will see,
the encoding of recursion is closely related to a famous set-theoretical paradox due to
Russell.
Let us begin by posing the following question. How can programs compute forever
in F[ without recursion? The answer to this question comes in the form of a seemingly
simple expression:
(Function x -> x x)(Function x -> x x)
Recall from Lemma 2.2, that a corollary to the existence of this expression is that
F[ is not normalizing. This computation is odd in some sense. (x x) is a function
being applied to itself. There is a logical paradox at the heart of this non-normalizing
computation, namely Russell’s Paradox.
Russell’s Paradox
In Frege’s set theory (circa 1900), sets were written as predicates P (x). We can view
predicates as single-argument functions which return a boolean value: true if the argu-
ment is in the set represented by the predicate and false if it is not. Testing membership
in a set is done via application of the predicate. For example, consider the predicate
Function x -> (x = 2 Or x = 3 Or x = 5)
This predicate represents the integer set {2, 3, 5} since it will return True for any of the
elements in that set and False for all other arguments. If we were to extend F[ to
include a native integer less-than operator, the predicate
Function x -> x < 2
would represent an innitely-sized set containing all integer values less than 2 (as F[ still
has no notion of real numbers). In general, given a predicate P representing a set S,
CHAPTER 2. OPERATIONAL SEMANTICS 34
e ∈ S i P e ⇒ True
Russell discovered a paradox in Frege’s set theory, and it can be expressed in the
following way.
Denition 2.11 (Russell’s Paradox). Let P be the set of all sets that do not contain
themselves as members. Is P a member of P ?
Asking whether or not a set is a member of itself seems like strange question, but
in fact there are many sets that are members of themselves. The innitely receding
set {{{{. . .}}}} has itself as a member. The set of things that are not apples is also a
member of itself (clearly, a set of non-apples is not an apple). These kinds of sets arise
only in “non-well-founded” set theory.
To explore the nature of Russell’s Paradox, let us try to answer the question it poses:
Does P contain itself as a member? Suppose the answer is yes, and P does contain itself
as a member. If that were the case then P should not be in P , which is the set of all
sets that do not contain themselves as members. Suppose, then, that the answer is no,
and that P does not contain itself as a member. Then P should have been included in
P , since it doesn’t contain itself. In other words, P is a member of P if and only if it
isn’t. Hence Russell’s Paradox is indeed a paradox.
This can also be illustrated by using F[ functions as predicates. Specically, we will
write a predicate for P above. We must dene P to accept an argument (which we know
to be a set - a predicate in our model) and determine if it contains itself (pass it to
itself as an argument). Thus, our representation of P is Function x -> Not(x x). We
merely apply P to itself to get our answer.
Denition 2.12 (Computational Russell’s Paradox). Let
def
P = Function x -> Not(x x).
What is the result of P P ? Namely, what is
(Function x -> Not(x x)) (Function x -> Not(x x))?
If this F[ program were evaluated, it would run forever. We can informally detect
the pattern just by looking at a few passes of an evaluation proof:
..
.
Not (Not ((Function x -> Not (x x))(Function x -> Not (x x))))
Not ((Function x -> Not (x x)) (Function x -> Not (x x)))
(Function x -> Not (x x)) (Function x -> Not (x x))
CHAPTER 2. OPERATIONAL SEMANTICS 35
We know that Not (Not (e)) evaluates to the same value as e. 5 We can see that
we’re going in circles. Again, this statement tells us that P P ⇒ True if and only if
P P ⇒ False.
This is not how Russell viewed his paradox, but it has the same core structure; it
is simply rephrased in terms of computation, and not set theory. The computational
realization of the paradox is that the predicate doesn’t compute to true or false, so its
not a sensible logical statement. Russell’s discovery of this paradox in Frege’s set theory
shook the foundations of mathematics. To solve this problem, Russell developed his
ramied theory of types, which is the ancestor of types in programming languages. The
program
(function x -> not(x x)) (function x -> not(x x))
is not typeable in OCaml for the same reason the corresponding predicate is not typeable
in Russell’s ramied theory of types. Try typing the above code into the OCaml top-level
and see what happens.
More information on Russell’s Paradox may be found in [14].
Encoding Recursion by Passing Self
In the logical view, passing a function to itself as argument is a bad thing. From a
programming view, however, it can be an extremely powerful tool: passing a function to
itself allows recursive functions to be dened without the use of Let Rec. We now show
how we can modify the paradoxical combinator to do useful work. First, let’s start with
the original Mr. Bad:
(Function x -> Not (x x)) (Function x -> Not (x x))
The rst step to making Mr. Bad do some good is rather than making an unbounded
number of Nots, Not (Not ( Not( ...))), lets make an unbounded number of some
predened function F:
(Function x -> F (x x)) (Function x -> F (x x))
Well, that makes unbounded F (F ( F( ...))), but this sequence never terminates
and so it doesn’t do us any good. The next step is to freeze the computation at certain
points to directly stop it from continuing; recall from our freeze macro above all we need
to do is wrap some expression in a Function -> ... 6 to freeze it:
def
makeFroFs =
(Function x -> F (Function -> x x)) (Function x -> F (Function -> x x))
5
In this case, anyway, but not in general. General assertions about the equivalence of expressions
are hard to prove. In Section 2.4, we will explore a formal means of determining if two expressions are
equivalent.
6
We use “ ” to represent a wildcard pattern, just like in OCaml.
CHAPTER 2. OPERATIONAL SEMANTICS 36
Now, we have postponed the innite execution; supposing F were say
Function froF -> True
the above would compute to
F (Function -> makeFroFs)
and since the F we dened throws away its argument this computation would in turn
terminate with value True.
This particular example terminated too well, it threw away all the copies of F we
painstakingly made. The way we can get recursion is to make an F which sometimes uses
its argument Function -> makeFroFs to make recursive calls by thawing it. Con-
sider the following revised F, aiming to ultimately be a function that sums the integers
{0, 1, . . . , n} for argument n:
def
F = Function froFs -> Function n ->
If n = 0 Then 0 Else n + froFs 0 (n - 1)
If makeFroFs were to use this F, it would again compute to F (Function ->
makeFroFs), but the argument Function -> makeFroFs is not thrown out by this
new F. Consider the computation of makeFroFs 5, by the above it is equivalent to com-
puting
(F (Function -> makeFroFs)) 5
and so for the above denition of F we see parameter froFs will be instantiated with
Function -> makeFroFs, and parameter n with 5. Because 5 = 0 is False we then
compute the else clause which after the above instantiation is
5 + (Function -> makeFroFs) 0 (5-1)
And, to compute the right-hand side of the addition, rst the 0 dummy argument is
applied to un-freeze makeFroFs, so we are now setting to compute makeFroFs (5-1).
Look above – this is nearly the exact spot we started at except the argument is one
smaller! So, makeFroFs is now a recursive summation function and this example will
ultimately compute to 15. We have succeeded in repurposing the paradoxical combinator
to write recursive functions.
There are a few minor clean-up steps we can perform on the above. First, since the F
we care about are curried functions of two arguments (we need the additional argument
e.g. n for the recursive call at a dierent value), we can make a revised freezer Function
n -> F n which doesn’t need to use 0 as the thawer but can “pun” and use the argument
itself to do the thawing. So, we can redo the above denitions as
def
makeFs = (Function x -> F’ (Function n -> (x x) n))
(Function x -> F’ (Function n -> (x x) n))
CHAPTER 2. OPERATIONAL SEMANTICS 37
def
F’ = Function fs -> Function n ->
If n = 0 Then 0 Else n + fs (n - 1)
– we can remove the 0 argument from the recursive call here since the n-1 argument is
doing the unfreezing work via our pun.
One last refactoring we can do to clean this up is to make a generic recursive-function-
maker, by pulling out the F’ in makeFs above as an explicit parameter f. This gives
us
def
Y = Function f ->
(Function x -> f (Function n -> (x x) n))
(Function x -> f (Function n -> (x x) n))
and we can apply to some concrete F, e.g. Y F’, to create our recursive summing
function. We call the above expression Y because this recursive function creator was
discovered by logicians many years ago and given that name.
Another route to Y via direct self-passing
Since the construction of Y is complex we now present another path to its construction.
This approach is more based on how object-oriented languages such as C++ implement
messaging to self: every time an object method is invoked, (a pointer to) the object itself
is passed as an additional parameter.
We demonstrate this approach by again writing a summate function. This time we
follow the C++ idea and write the function to accept two arguments: arg, which is the
number to which we will sum, and this, which is the copy of the function we require to
be passed in so that recursion may occur. We will name our function summate0 to reect
the fact that it is not quite the summate we want. It can be dened as
def
summate0 = Function this -> Function arg ->
If arg = 0 Then 0 Else arg + this this (arg - 1)
Note the use of this this (arg - 1). The rst use of this names the function to be
applied; the second use of this is one of the arguments to that function. The argument
this allows the recursive call to invoke the function again, thus allowing us to recurse
as much as we need.
We can now sum the integers {0, 1, . . . , 7} with the expression
summate0 summate0 7
CHAPTER 2. OPERATIONAL SEMANTICS 38
summate0 always expects its rst argument this to be itself. It can then use one copy
for the recursive call (the rst this) and pass the other copy on for future duplication.
So summate0 summate0 “primes the pump”, so to speak, by giving the process an initial
extra copy of itself.
Better yet, recall that currying allows us to obtain the inner function without applying
it. In essence, a function with multiple arguments could be partially evaluated, with some
arguments xed and others waiting for input. We can use this to our advantage to dene
the summate function we want:
def
summate = summate0 summate0
This allows us to hide the self-passing from the individual using our summate function,
which cleans things up considerably. We can summarize the entire process as follows,
recalling that even Let itself could be a macro for a function call:
def
summate = Let summ = Function this -> Function arg ->
If arg = 0 Then 0 Else arg + this this (arg - 1)
In summ summ
We now have a model for dening recursive functions without the use of the Let
Rec operator. This means that untyped languages with no built-in recursion can still
be Turing-complete. While this is an accomplishment, we can do even better; we can
abstract the idea of self-passing almost entirely out of the body of the function itself.
The Y -Combinator The Y -combinator is a further abstraction of self-passing. The
idea is that the Y -combinator does the self-application with an abstract body of code
that is passed in as an argument. We rst dene a function called almostY, and then
revise that denition to arrive at the real Y -combinator.
def
almostY = Function body -> body body
using almostY, we can dene summate as follows.
def
summate = almostY (Function this -> Function arg ->
If arg = 0 Then 0 Else arg + this this (arg - 1))
The true Y -combinator actually goes one step further and allows us to write recursive
calls in the more natural style of just “this (arg - 1)”, avoiding the extra this pa-
rameter. To do this, we assume that the body argument is already in this simple form.
We then dene a new form, wrapper, which replaces this with (this this) in body:
CHAPTER 2. OPERATIONAL SEMANTICS 39
Denition 2.13 (Y -Combinator).
def
combY = Function body ->
Let wrapper = Function this -> Function arg -> body (this this) arg
In wrapper wrapper
This is the same Y as was dened in the previous subsection.
The transformation steps performed in these examples are also good examples of
the power of higher-order functional programming: “code surgery” is performed on
body to produce wrapper by simply using function abstraction and application. The
Y -combinator can then be used to dene summate as
def
summate = combY (Function this -> Function arg ->
If arg = 0 Then 0 Else arg + this (arg - 1))
2.3.6 Call-By-Name Parameter Passing
In call-by-name parameter passing, the argument to the function is not evaluated
at function call time, but rather is only evaluated if it is used. This style of parameter
passing is largely of historical interest now; Algol uses it but no modern languages are call-
by-name by default (The Digital Mars D language does allow parameters to be treated
as call-by-name via use of the lazy qualier). The reason is that it is much harder to
write ecient compilers if call-by-name parameter passing is used. Nonetheless, it is
worth taking a brief look at call-by-name parameter passing.
Let us dene the operational semantics for call-by-name.
e1 ⇒ Function x -> e, e[e2 /x] ⇒ v
(Call-By-Name Application)
e1 e2 ⇒ v
Freezing and thawing, dened in Section 2.3.4, is a way to get call-by-name behavior
in a call-by-value language. Consider, then, the computation of
(Function x -> Thaw x + Thaw x) (Freeze (3 - 2))
(3 - 2) is not evaluated until we are inside the body of the function where it is thawed,
and it is then evaluated two separate times. This is precisely the behavior of call-by-
name parameter passing, so Freeze and Thaw can encode it by this means. The fact that
(3 - 2) is executed twice shows the main weakness of call by name, namely repeated
evaluation of the function argument.
Lazy or call-by-need evaluation is a version of call-by-name that caches evaluated
function arguments the rst time they are evaluated so it doesn’t have to re-evaluate them
in subsequent uses. Haskell [13, 7] is a pure functional language with lazy evaluation.
CHAPTER 2. OPERATIONAL SEMANTICS 40
2.3.7 F[ Abstract Syntax
The previous sections thoroughly describe the operational semantics of F[ in terms of
its concrete syntax. Recall from our boolean language, however, that an interpreter
operates over a representation of a program which is more conducive to programmatic
manipulation; this representation is termed its abstract syntax. To dene the abstract
syntax of F[ for a OCaml interpreter, we need to dene a variant type that captures the
expressiveness of F[. The variant types we will use are as follows.
type ident = Ident of string
type expr =
Var of ident | Function of ident * expr | Appl of expr * expr |
Let of ident * expr * expr | LetRec of ident * ident * expr * expr |
Plus of expr * expr | Minus of expr * expr | Equal of expr * expr |
And of expr * expr| Or of expr * expr | Not of expr |
If of expr * expr * expr | Int of int | Bool of bool
type fbtype = TInt | TBool | TArrow of fbtype * fbtype | TVar of string;;
One important point here is the existence of the ident type. Notice where ident is
used in the expr type: as variable identiers, and as function parameters for Function
and Let Rec. The ident type attaches additional semantic information to a string,
indicating that the string specically represents an identier.
Note, though, that ident is used in two dierent ways: to signify the declaration of
a variable (such as in Function (Ident "x",...)) and to signify the use of a variable
(such as in Var(Ident "x")). Observe that the use of a variable is an expression and so
can appear in the AST anywhere that any expression can appear. The declaration of a
variable, on the other hand, is not an expression; variables are only declared in functions.
In this way, we are able to use the OCaml type system to help us keep the properties of
the AST nodes straight.
For example, consider the following AST:
Plus
Ident Int
"x" 5
At rst glance, it might appear that this AST represents the F[ expression x + 5.
However, the AST above cannot actually exist using the variants we dened above. The
Plus variation accepts two expressions upon construction and Ident is not a variant;
thus, the equivalent OCaml code Plus(Ident "x",Int 5) would not even typecheck.
The F[ expression x + 5 is represented instead by the AST
CHAPTER 2. OPERATIONAL SEMANTICS 41
Plus
Var Int
Ident 5
"x"
which is represented by the OCaml code Plus(Var(Ident "x"),Int 5).
Being able to convert from abstract to concrete syntax and vice versa is an important
skill for one to develop, but it takes some time to become procient at this conversion.
Let us look as some examples F[ in pursuit of rening this skill.
Example 2.17.
Concrete: Plus
1 + 2
Int Int
Abstract:
Plus(Int 1,Int 2) 1 2
Example 2.18.
Concrete: Or
True Or False
Bool Bool
Abstract:
Or(Bool true,Bool false) true false
Example 2.19.
If
Concrete:
Not Int Int
If Not(1 = 2) Then 3 Else 4
Equal 3 4
Abstract:
If(Not(Equal(Int 1, Int 2)),
Int 3, Int 4) Int Int
1 2
CHAPTER 2. OPERATIONAL SEMANTICS 42
Example 2.20.
Appl
Concrete:
(Function x -> x + 1) 5
Function Int
Abstract:
Ident Plus 5
Appl(
Function(
Ident "x", "x" Var Int
Plus(Var(Ident "x"),
Int 1)), Ident 1
Int 5)
"x"
Example 2.21.
Appl
Concrete:
(Function x -> Function y ->
x + y) 4 5 Appl Int
Abstract: Function Int 5
Appl(
Appl( Ident 4
Function(
Ident "x", "x" Function
Function(
Ident "y", Ident Plus
Plus(
Var(Ident "x"), "y" Var Var
Var(Ident "y")
))), Ident Ident
Int 4),
Int 5) "x" "y"
CHAPTER 2. OPERATIONAL SEMANTICS 43
Example 2.22.
Concrete:
Let Rec fib x =
If x = 1 Or x = 2 Then 1 Else fib (x - 1) + fib (x - 2)
In fib 6
Abstract:
Letrec(Ident "fib", Ident "x", If(Or(Equal(Var(Ident "x"), Int 1),
Equal(Var(Ident "x"), Int 2)), Int 1, Plus(Appl(Var(Ident "fib"),
Minus(Var(Ident "x"), Int 1)), Appl(Var(Ident "fib"), Minus(Var(Ident
"x"), Int 2)))), Appl(Var(Ident "fib"), Int 6))
Letrec
Ident Ident Appl
"fib" "x" Ident Int
If "fib" 6
Or Int
Equal Equal 1 Plus
Var Int Var Int
Ident 1 Ident 2
"x" "x" Appl Appl
Var Minus Var Minus
Ident Var Int Ident Var Int
"fib" Ident 1 "fib" Ident 2
"x" "x"
Notice how lengthy even simple expressions can become when represented in the
abstract syntax. Review the above examples carefully, and try some additional examples
of your own. It is important to be able to comfortably switch between abstract and
concrete syntax when writing compilers and interpreters.
CHAPTER 2. OPERATIONAL SEMANTICS 44
2.4 Operational Equivalence
One of the most basic operations dened over a space of mathematical objects is the
equivalence relation (∼=). We have, using operational semantics, dened programs as
mathematical entities. As a result, equivalence makes sense for programs too.
First, we are compelled to consider what we mean intuitively by “equivalent” pro-
grams. Two things which are equivalent are interchangable; thus, two equivalent pro-
grams must do the same thing. It is not necessary that they do so in exactly the same
way however. For example, we can easily see that the programs (1 + 1 + 1 - 1) and
(2) are equivalent, but the prior requires more computation to evaluate.
Because two equivalent programs can be substituted for each other but may have
dierent execution-time properties, we can use operational equivalence when optimizing
compilers and other such tools. Operational equivalence provides a rigorous and reliable
foundation for such work; we do not need to worry about an optimization changing the
behavior of an application because we can prove mathematically that such a change is
impossible. Operational equivalence also denes the process of refactoring, a process by
which developers can change the structure of a program for maintainability or enhance-
ment without changing the program’s behavior.
Dening an equivalence relation for programs is actually not as straightforward as
one might expect. The initial idea is to dene the relation such that two programs are
equivalent if they always lead to the same results when used. As we will see, however, this
denition is not sucient, and we will need to do some work to arrive at a satisfactory
denition.
Let us begin by looking at a few sample equivalences to get a feel for what they
are. η-conversion (or eta-conversion) is one example of an interesting equivalence. It is
dened as follows.
Function x -> e ∼
=
Function z -> (Function x -> e) z, for z not free in e
η-conversion is similar to the proxy pattern in object oriented programming[12]. A closely
related law for our freeze/thaw syntax is
Thaw (Freeze e) ∼
=e
In both examples, one of the expressions may be replaced by the other without ill eects
(besides perhaps changing execution time), so we say they are equivalent. To write
formal proofs, however, we will need to develop a more rigorous denition of equivalence.
2.4.1 Dening Operational Equivalence
Let’s begin by informally strengthening our denition of operational equivalence. We
dene equivalence in a manner dating all the way back to Leibniz[18]:
CHAPTER 2. OPERATIONAL SEMANTICS 45
Denition 2.14 (Operational Equivalence (Informal)). Two program expressions are
equivalent if and only if one can be replaced with the other at any place, and no external
change in behavior will be noticed.
We wish to study equivalence for possibly open programs, because there are good equiva-
lences such as x + 1 - 1 ∼ = x + 0. We dene “at any place” by the notion of a program
context, which is, informally, a F[ program with some holes (•) in it. Using this infor-
mal denition, testing if e1 ∼
= e2 would be roughly equivalent to performing the following
steps (for all possible programs and all possible holes, of course).
1. Select any program context (that is, program containing a hole).
2. Place e1 in the • position and run the program.
3. Do the same for e2 .
4. If the observable result is dierent, e1 is not equivalent to e2 .
5. Repeat steps 1-4 for every possible context. If none of these innitely many contexts
produces dierent results, then e1 is equivalent to e2 .
Now let us elaborate on the notion of a program context. Take an F[ program with
some “holes” (•) punched in it: replace some subterms of any expression with •. Then
“hole-lling” in this program context C, written C[e], means replacing • with e in C.
Hole lling is like substitution, but without the concerns of bound or free variables. It
is direct replacement with no conditions.
Let us look at an example of contexts and hole-lling using η-conversion as we dened
above. Let
def
C = (Function z -> Function x -> • ) z
Now, lling the hole with x + 2 is simply written
((Function z -> Function x -> • ) z)[x + 2] =
(Function z -> Function x -> x + 2) z
Finally, we are ready to rigorously dene operational equivalence.
Denition 2.15 (Operational Equivalence). e ∼ = e′ if and only if for all contexts C
′
such that C[e] and C[e ] are both closed expressions, C[e] ⇒ v for some v if and only if
C[e′ ] ⇒ v ′ for some v ′ .
Another way to phrase this denition is that two expressions are equivalent if in
any possible context, C, one terminates if the other does. We call this operational
equivalence because it is based on the interpreter for the language, or rather it is based
on the operational semantics. The most interesting, and perhaps nonintuitive, part of
CHAPTER 2. OPERATIONAL SEMANTICS 46
this denition is that nothing is said about the relationship between v and v ′ . In fact,
they may be dierent in theory. However, intuition tells us that v and v ′ must be very
similar, since equivalence holds for any possible context.
The only problem with this denition of equivalence is its “incestuous” nature—there
is no absolute standard of equivalence removed from the language. Domain theory is
a mathematical discipline which denes an algebra of programs in terms of existing
mathematical objects (complete and continuous partial orders). We are not going to
discuss domain theory here, mainly because it does not generalize well to programming
languages with side eects. [16] explores the relationship between operational semantics
and domain theory.
2.4.2 Properties of Operational Equivalence
In this section, we present some general equivalence principles for F[.
Denition 2.16 (Reexivity).
e∼
=e
Denition 2.17 (Symmetry).
e∼
= e′ if e′ ∼
=e
Denition 2.18 (Transitivity).
e∼ = e′ and e′ ∼
= e′′ if e ∼ = e′′
Denition 2.19 (Congruence).
C[e] ∼
= C[e′ ] if e ∼
= e′
Denition 2.20 (β-Equivalence).
((Function x -> e) v) ∼
= (e[v/x])
provided v is closed (if v had free variables they could be captured when v is placed deep
inside e).
Denition 2.21 (η-Equivalence).
(Function x -> e) ∼
= (Function z -> (Function x -> e) z)
Denition 2.22 (α-Equivalence).
(Function x -> e) ∼
= (Function y -> (e[y/x]))
Denition 2.23.
(n + n′ ) ∼
= the sum of n and n′
Similar rules hold for -, And, Or, Not, and =.
CHAPTER 2. OPERATIONAL SEMANTICS 47
Denition 2.24.
(If True Then e Else e′ ) ∼
=e
A similar rule holds for If False. . .
Denition 2.25.
If e ⇒ v then e ∼
=v
Equivalence transformations on programs can be used to justify results of computa-
tions instead of directly computing with the interpreter; it is often easier. An important
component of compiler optimization is applying transformations, such as the ones above,
that preserve equivalence.
2.4.3 Examples of Operational Equivalence
To solidify the concept of operational equivalence (one which reliably boggles newcomers
to programming language theory), we provide a number of examples of equivalent and
non-equivalent expressions. We start with a very simple example of two expressions
which are not equivalent.
Lemma 2.5. 2 3
def
Proof. By example. Let C = If •= 2 Then 0 Else (0 0). C[2] ⇒ 0 while C[3] ; v
for any v. Thus, by denition, 2 3.
Note that, in the above proof, we used the expression (0 0). This expression cannot
evaluate; the application rule applies only to functions. As a result, this expression makes
an excellent tool for intentionally making code get stuck when building a proof about
operational equivalence. Other similar get-stuck expressions exist, such as True + True,
Not 5, and the ever-popular (Function x -> x x)(Function x -> x x).
It should be clear that we can use the approach in Lemma 2.5 to prove that any
two integers which are not equal are not operationally equivalent. But can we make a
statement about a non-value expression?
Lemma 2.6. x x + 1 - 1.
At rst glance, this inequivalence may seem counterintuitive. But the proof is fairly
simple:
def
Proof. By example. Let C = (Function x -> •) True. Then C[x] ⇒ True. C[x + 1 - 1] ≡
(Function x -> x + 1 - 1) True, which cannot evaluate because no rule allows us to
evaluate True + 1. Thus, by denition, x x + 1 - 1.
Similar proofs could be used to prove inequivalences such as Not(Not(e)) e. The
key here is that some of the rules in F[ distinguish between integer values and boolean
values. As a result, equivalences cannot hold if one side makes assumptions about the
type of value to which it will evaluate and the other side does not.
CHAPTER 2. OPERATIONAL SEMANTICS 48
We have proven inequivalences; can we prove an equivalence? It turns out that some
equivalences can be proven but that this process is much more involved. Thus far, our
proofs of inequivalence have relied on the fact that they merely need to demonstrate an
example context in which the expressions produce dierent results. A proof of equiv-
alence, however, must demonstrate that no such example exists among innitely many
contexts. For example, consider the following:
Lemma 2.7. If e ; v for any v, e′ ; v ′ for any v ′ , and both e and e′ are closed
= e′ . For example, (0 0) ∼
expressions, then e ∼ = (Function x -> x x)(Function x ->
x x).
First, we need a denition:
Denition 2.26 (Touch). An expression is said to touch one of its subexpressions if
the evaluation of the expression causes the evaluation of that subexpression.
That is, If True Then 0 Else 1 touches the subexpression 0 because it is evaluated
when the whole expression is evaluated. 1 is not touched because the evaluation of the
expression never causes 1 to be evaluated.
We can now prove Lemma 2.7.
Proof. By contradiction. Without loss of generalization, suppose that there is some
context C for which C[e] ⇒ v while C[e′ ] ; v ′ for any v ′ .
Because e is a closed expression, C ′ [e] ; v for all v and all contexts C ′ which touch
e. Because C[e] ⇒ v, we know that C does not touch the hole.
Because C does not touch the hole, how the hole evaluates cannot aect the evaluation
of C; that is, C[e∗ ] ⇒ v ∗ for some v ∗ must be true for all e∗ or for no e∗ .
Because C[e] ⇒ v, C[e∗ ] ⇒ v ∗ for some v ∗ for all e∗ . But C[e′ ] ; v ′ for any v ′ . Thus,
by contradiction, C does not exist. Thus, by denition, e ∼ = e′ .
The above proof is much longer and more complex than any of the proofs of in-
equivalence that we have encountered and it isn’t even very robust. It could benet, for
example, from a proof that a closed expression cannot be changed by a replacement rule
(which is taken for granted above). Furthermore, the above proof doesn’t even prove a
very useful fact! Of far more interest would be proving the operational equivalence of
two expressions when one of them is executed in a more desirable manner than the other
(faster, fewer resources, etc.) in order to motivate compiler optimizations.
Lemma 2.7 is, however, eective in demonstrating the complexities involved in prov-
ing operational equivalence. It is surprisingly dicult to prove that an equivalence holds;
even proving 1 + 1 ∼ = 2 is quite challenging. See [16] for more information on this topic.
2.4.4 The λ-Calculus
We briey consider the λ-calculus. In Section 2.3, we saw how to encode tuples, lists,
Let statements, freezing and thawing, and even recursion in F[. The encoding approach
is very powerful, and also gives us a way to understand complex languages based on
our understanding of simpler ones. Even numbers, booleans, and if-then-else statements
are encodable, although we will skip these topics. Thus, all that is needed is functions
CHAPTER 2. OPERATIONAL SEMANTICS 49
and application to make a Turing-complete programming language. This language is
known as the pure λ calculus, because functions are usually written as λx.e instead
of Function x -> e.
Execution in λ calculus is extremely straightforward and concise. The main points
are as follows.
• Even programs with free variables can execute (or reduce in λ-calculus terminol-
ogy).
• Reduction can happen anywhere, e.g. inside a function body that hasn’t been
called yet.
• (λx.e)e′ ⇒ e[e′ /x] is the only reduction rule, called β-reduction. (It has a special
side-condition that it must be capture-free, i.e. no free variables in e′ become bound
in the result. Capture is one of the complications of allowing reduction anywhere.)
This form of computation is conceptually very interesting, but is more distant from
how actual computer languages execute and so we do not put a strong focus on it here.
Exercises
Exercise 2.1. How would you change the Sheep language to allow the terms bah, baah, · · ·
without excluding any terms which are already allowed?
Exercise 2.2. Is the term it in the Frog language? Why or why not?
Exercise 2.3. Is it possible to construct a term in Frog without using the terminal t?
If so, give an example. If not, why not?
Exercise 2.4. Complete the denition of the operational semantics for the boolean
language described in section 2.2.1 by writing the rules for Or and Implies.
Exercise 2.5. Why not just use interpreters and forget about the operational semantics
approach?
Chapter 3
Tuples, Records, and Variants
In Chapter 2 we saw that, using a language with only functions and application, we
could represent advanced programming constructs such as tuples and lists. However,
we pointed out that these encodings have fundamental problems, such as a low degree
of eciency, and the fact that they necessarily expose their details to the programmer,
making them dicult and dangerous to work with in practice. Recall how we could take
our encoding of a pair from Chapter 2 and apply it like a function; clearly the wrong
behavior. In this chapter we look at how we can build some of these advanced features
into the language, namely tuples and records, and we conclude the chapter by examining
variants.
3.1 Tuples
One of the most fundamental forms of data aggregation in programming is the notion of
pairing. With pairs, or 2-tuples, almost any data structure can be represented. Tripling
can be represented as (1, (2, 3)), and in general n-tuples can be represented with pairs
in a similar fashion. Records and C-style structs can be represented with sets (n-tuples)
of (label, value)-pairs. Even objects can be built up from pairs, but this is stretching it
(just as encoding pairs as functions was stretching it).
In Chapter 2, we showed an encoding of pairs based on functions. There were two
problems with this representation of pairs. First of all, the representation was inecient.
More importantly, the behavior of pairs was slightly wrong, because we could apply
them like functions. To really handle pairs correctly, we need to add them directly to
the language. We can add pairs to F[ in a fairly straightforward manner. We show how
to add pair functionality to the interpreter, and leave the operational semantics for pairs
as an exercise for the reader.
First, we extend the expr type in our interpreter to include the following.
type expr =
...
| Pr of expr * expr | Left of expr | Right of expr
Next, we add the following clauses to our eval function.
50
CHAPTER 3. TUPLES, RECORDS, AND VARIANTS 51
let rec eval e =
match e with
...
| Pr(e1, e2) -> Pr(eval e1, eval e2)
| Left(expr) -> (match eval expr with
Pr(e1,e2) -> e1
| _ -> raise TypeMismatch)
| Right(expr) -> (match eval expr with
Pr(e1, e2) -> e2
| _ -> raise TypeMismatch)
Notice that our pairs are eager, that is, the left and right components of the pair are
evaluated, and must be values for the pair itself to be considered a value. For example,
(2, 3+4) ⇒ (2, 7). OCaml tuples exhibit this same behavior. Also notice that our
space of values is now bigger. It includes:
• numbers 0, 1, -1, 2, -2, . . .
• booleans True, False
• functions Function x -> . . .
• pairs (v1 , v2 )
Exercise 3.1. How would we write our interpreter to handle pairs in a non-eager way?
In other words, what would need to be in the interpreter so that (e1 , e2 ) was considered
a value (as opposed to only (v1 , v2 ) being considered a value)?
Now that we have 2-tuples, encoding 3-tuples, 4-tuples, and n-tuples is easy. We
simply do them as (1, (2, (3, (. . . , n)))). As we saw before, lists can be encoded as n-
tuples.
3.2 Records
Records are a variation on tuples in which the elds have names. Records have several
advantages over tuples. The main advantage is the named eld. From a software engi-
neering perspective, a named eld “zipcode” is far superior to “the third element in the
tuple.” The order of the elds of a record is arbitrary, unlike with tuples.
Records are also far closer to objects than tuples are. We can encode object poly-
morphism via record polymorphism. Record polymorphism is discussed in Section 3.2.1.
The motivation for using records to encode objects is that a subclass is composed of a
superset of the elds of its parent class, and yet instances of both classes may be used in
the context of the superclass. Similarly, record polymorphism allows records to be used
in the context of a subset of their elds, and so the mapping is quite natural. We will
use records to model objects in Chapter 5.
Our F[ records will have the same syntax as OCaml records. That is, records are
written as {l1 =e1 ; l2 =e2 ; . . . ; ln =en }, and selection is written as e.lk , which selects
CHAPTER 3. TUPLES, RECORDS, AND VARIANTS 52
the value labeled lk from record e. We use l as a metavariable ranging over labels, just
as we use e as a metavariable indicating an expression; an actual record is for instance
{x=5; y=7; z=6}, so x here is an actual label.
If records are always statically known to be of xed size, that is, if they are known
to be of xed size at the time we write our code, then we may simply map the labels to
integers, and encode the record as a tuple. For instance,
{x=5; y=7; z=6} = (5, (7, 6))
e.x = Left e
e.y = Left (Right e)
e.z = Right (Right e)
Obviously, the makes for ugly, hard-to-read code, but for C-style structs, it works.
But in the case where records can shrink and grow, this encoding is fundamentally too
weak. C++ structs can be subtypes of one another, so elds that are not declared may,
in fact, be present at runtime.
On the other hand, pairs can be encoded as records quite nicely. The pair (3, 4)
can simply be encoded as the record {l=3; r=4}. More complex pairs, such as those
used to represent lists, can also be encoded as records. For example, the pair (3, (4,
(5, 6))), which represents the list [3; 4; 5; 6], can be encoded as the record {l=3;
r={l=4; r={l=5; r=6}}}.
A variation of this list encoding is used in the mergesort example in Section 4.3.2. This
variation encodes the above list as {l=3; r={l=4; r={l=5; r={l=6; r=emptylist}}}}.
This encoding has the nice property that the values are always contained in the l elds,
and the rest of the list is always contained in the r elds. This is much closer to the
way real languages such as OCaml, Scheme, and Lisp represent lists (recall how we write
statements like let (first::rest) = mylist in OCaml).
3.2.1 Record Polymorphism
Records do more that just add readability to programs. For instance, if you have
{size=10; weight=100} and {weight=10; name="Mike"}, either of these two records
can be passed to a function such as
Function x -> x.weight.
In scripting languages, this exibility is informally known as duck typing – if Function
d -> d.quack () does not produce a run-time error when invoked we will assume the
d passed to it is a duck since it quacked like a duck.
This form of exibility is called subtype polymorphism in a typed language. In
the function above, x can be any record with a weight eld. Subtype polymorphism on
records is known as record polymorphism. OCaml disallows record polymorphism,
so the OCaml version of the above code will not typecheck.
CHAPTER 3. TUPLES, RECORDS, AND VARIANTS 53
In typed object-oriented languages, subtype polymorphism is known as object poly-
morphism, or, more commonly, as simply polymorphism. The latter is, unfortunately,
confusing with respect to the parametric polymorphism of OCaml.
3.2.2 The F[R Language
We will now dene the F[R language: F[ with records. Again, we will concentrate on
the interpreter, and leave the operational semantics as an exercise to the reader.
The rst thing we need to consider is how to represent record labels. Record labels
are symbols, and so we could use our identiers (Ident "x") as labels, but it is better
to think of record labels as a dierent sort. For instance, labels are never bound or
substituted for. So we will dene a new type in our interpreter.
type label = Lab of string
Next, we need a way to represent the record itself. Records may be of arbitrary length,
so a list of (label, expression)-pairs is needed. In addition, we need a way to represent
selection. The F[R expr type now looks like the following.
type expr = ...
| Record of (label * expr) list | Select of expr * label
Let’s look at some concrete to abstract syntax examples for our new language.
Example 3.1.
{size=7; weight=255}
Record [(Lab "size", Int 7); (Lab "weight", Int 255)]
Example 3.2.
e.size
Select(Var(Ident "e"), Lab "size")
In addition, our denition of values must now be extended to include records. Specif-
ically, {l1 =v1 ; l2 =v2 ; . . . ; ln =vn } is a value, provided that v1 , v2 , . . . , vn are values.
Finally, we add the necessary rules to our interpreter. Because records can be of
arbitrary length, we will have to do a little more work when evaluating them. The
let-rec-and syntax in OCaml is used to declare mutually recursive functions, and we
use it below.
(* A function to project a given field *)
let rec lookupRecord body (Lab l) = match body with
[] -> raise FieldNotFound
| (Lab l’, v)::t -> if l = l’ then v else lookupRecord t (Lab l)
CHAPTER 3. TUPLES, RECORDS, AND VARIANTS 54
(* The eval function, with an evalRecord helper *)
let rec eval e = match e with
...
| Record(body) -> Record(evalRecord body)
| Select(e, l) -> match eval e with
Record(body) -> lookupRecord body l
| _ -> raise TypeMismatch
and evalRecord body = match body with
[] -> []
| (Lab l, e)::t -> (Lab l, eval e)::evalRecord t
Notice that our interpreter correctly handles {}, the empty record, by having it
compute to the itself since it is, by denition, a value.
Interact with F[SR. We can use our F[SR interpreter to explore records (F[SR
is F[ with records and state, and is introduced in Chapter 4). First, let’s try a
simple example to demonstrate the eager evaluation of records.
# {one = 1; two = 2;
three = 2 + 1; four = (Function x -> x + x) 2};;
==> {one=1; two=2; three=3; four=4}
Next, let’s try a more interesting example, where we use records to encode lists. Note
that we dene emptylist as -1. The function below sums all values in a list (assuming
it has a list of integers).
# Let emptylist = 0 - 1 In
Let Rec sumlist list =
If list = emptylist Then
0
Else
(list.l) + sumlist (list.r) In
sumlist {l=1; r={l=2; r={l=3; r={l=4; r=emptylist}}}};;
==> 10
3.3 Variants
We have been using variants in OCaml, as the types for expressions expr. Now we study
untyped variants more closely. OCaml actually has two (incompatible) forms of variant,
regular variants and polymorphic variants . In the untyped context we are working in,
the OCaml polymorphic variants are more appropriate and we will use that form.
We briey contrast the two forms of variant in OCaml for readers unfamiliar with
polymorphic variants. Recall that in OCaml, regular variants are rst declared as types
CHAPTER 3. TUPLES, RECORDS, AND VARIANTS 55
type feeling =
Vaguely of feeling | Mixed of feeling * feeling |
Love of string | Hate of string | Happy | Depressed
which allows Vaguely(Happy), Mixed(Vaguely(Happy),Hate("Fred")), etc. Polymor-
phic variants require no type declaration; thus, for the above we can directly write
‘Vaguely(‘Happy), ‘Mixed(‘Vaguely(‘Happy),‘Hate("Fred")), etc. The ‘ must be
prexed each variant name, indicating it is a polymorphic variant name.
3.3.1 Variant Polymorphism
Like records, variants are polymorphic. In records, many dierent forms of record could
get through a particular selection (any record with the selected eld). In variants, the
polymorphism is dual in that many dierent forms of match statement can process a
given variant.
3.3.2 The F[V Language
We will now dene the F[V language, F[ with . . . Variants.
The new syntax requires variant syntax and match syntax. Just as we restrict func-
tions to have one argument only, we also restrict variant constructors to take one argu-
ment only; multiple- or zero-argument variants must be encoded. In concrete syntax,
we construct variants by n(e) for n a named variant and e its parameter, for exam-
ple ‘Positive(3). Variants are then used via match: Match e With n1 (x1 ) -> e1 |
...| nm (xm ) -> em . We don’t dene a general pattern match as found in OCaml—our
Match will matching a single variant eld at a time, and won’t work on anything besides
variants.
The abstract syntax for F[V is as follows. First, each variant needs a name.
type name = Name of string
The F[V abstract syntax expr type now looks like
type expr = ...
| Variant of (name * expr)
| Match of expr * (name * ident * expr) list
Let’s look at some concrete to abstract syntax examples for our new language.
Example 3.3.
‘Positive(4)
Variant(Name "Positive", Int 4)
CHAPTER 3. TUPLES, RECORDS, AND VARIANTS 56
Example 3.4.
Match e With
‘Positive(x) -> 1 | ‘Negative(y) -> -1 | ‘Zero(p) -> 0
Match(Var(Ident("e")),[(Name "Positive",Ident "x",Int 1);
(Name "Negative",Ident "y",Int -1);
(Name "Zero",Ident "p",Int 0)])
Note in this example we can’t just have a variant Zero since 0-ary variants are not
allowed, and a dummy argument must be supplied. Multiple-argument variants may be
encoded by a single argument variant over a pair or record (since we have neither pair or
records in F[V, the only recourse is the encoding of pairs used in F[ in Section 2.3.4).
In addition, our denition of F[V values must also be extended from the F[ ones to
include variants: n(v) is a value, provided v is. To dene the meaning of F[V execution,
we extend the operational semantics of F[ with the following two rules:
e⇒v
(Variant Rule)
n(e) ⇒ n(v)
e ⇒ nj (vj ), ej [vj /xj ] ⇒ v
(Match Rule)
Match e With
n1 (x1 ) -> e1 | . . .
⇒ v
| nj (xj ) -> ej | . . .
| nm (xm ) -> em
The Variant rule constructs a new variant labeled n; its argument is eagerly evaluated
to a value, just as in OCaml: ‘Positive(3+2) ⇒ ‘Positive(5). The Match rule rst
computes the expression e being matched to a variant nj (vj ), and then looks up that
variant in the match, nding nj (xj ) -> ej , and then evaluating ej with its variable xj
given the value of the variant argument.
Example 3.5.
Match ‘Grilled(3+1) With
‘Stewed(x) -> 4 + x |
‘Grilled(y) -> 2 + y ⇒ 6, because
‘Grilled(3+1) ⇒ ‘Grilled(4) and (2 + y)[4/y] ⇒ 6
Exercise 3.2. Extend the F[V syntax and operational semantics so the Match expression
always has a nal match of the form of the form“| -> e”. Is this Match strictly more
expressive than the old one, or not?
Variants and Records are Duals Here we see how the denition of a record is
modeled as a use of a variant, and a use of a record is the denition of a variant.
CHAPTER 3. TUPLES, RECORDS, AND VARIANTS 57
Variants are the dual of records: a record is this eld and that eld and that eld;
a variant is this eld or that eld or that eld. Since they are duals, dening a record
looks something like using a variant, and dening a variant looks like using a record.
Variants can directly encode records and vice-versa, in a programming analogy of
how DeMorgan’s Laws allows logical and to be encoded in terms of or, and vice-versa:
p Or q = Not(Not p And Not q); p And q = Not(Not p Or Not q).
Variants can be encoded using records as follows.
Match s With ‘n1 (x1 ) -> e1 | . . . | ‘nm (xm ) -> em =
s{n1 =Function x1 -> e1 ; . . . ; nm =Function xm -> em }
‘n(e) = (Function x -> (Function r -> r.n x )) e
The tricky part of the encoding is that denitions must be turned in to uses and
vice-versa. This is done with functions: an injection is modeled as a function which is
given a record and will select the specied eld.
Here is how records can be encoded using variants.
{l1 =e1 ; . . . ; ln =en } = (Function k1 -> . . . Function kn ->Function s ->
Match s With ‘l1 (x) -> k1 | . . . | ‘ln (x) -> kn )(e1 ) . . . (en )
e.lk = e ‘lk (0)
where x above is any fresh variable.
One other interesting aspect about the duality between records and variants is that
both records and variants can encode objects. A variant is a message, and an object
is a case on the message. In the variant encoding of objects, it is easy to pass around
messages as rst-class entities. Using variants to encode objects makes objects that are
hard to typecheck, however, and that is why we think of objects as more record-like.
Chapter 4
Side Eects: State and Exceptions
We will now leave the world of pure functional programming, and begin considering
languages with side-eects. For now we will focus solely on two particular side-eects,
state and exceptions. There are, however, many other types of side-eects, including the
following.
• Goto-statements or loop-breaks, which are similar to exceptions. Note that loop-
breaks require loops, which require state.
• Input and output.
• Distributed message passing.
4.1 State
Languages like F[, F[R, and F[V are pure functional languages. Once we add any
kind of side-eect to a language it is not pure functional anymore. Side-eects are non-
local, meaning they can aect other parts of the program. As an example, consider the
following OCaml code.
let x = ref 9 in
let f z = x := !x + z in
x := 5; f 5; !x
This expression evaluates to 10. Even though x was dened as a reference to 9 when
f was declared, the last line sets x to be a reference to 5, and during the application of f,
x is reassigned to be a reference to (5 + 5), making it 10 when it is nally dereferenced
in the last line. Clearly, the use of side eects makes a program much more dicult
to analyze, since it is not as declarative as a functional program. When looking at
programs with side-eects, one must examine the entire body of code to see which side-
eects inuence the outcome of which expressions. Therefore, it is a good programming
moral to use side-eects only when they are strongly needed.
Let us begin by informally discussing the semantics of references. In essence, when
a reference is created in a program, a cell is created that contains the specied value,
and the reference itself is a pointer to that cell. A good metaphor for these reference
58
CHAPTER 4. SIDE EFFECTS: STATE AND EXCEPTIONS 59
Figure 4.1: The “junction box” view of reference cells.
cells is to think of each one as a junction box. Then, each assignment is a path into
the junction, and each read, or dereference, is a path going out of the junction. The
reference cell, or junction, sits to the side of the program, and allows distant parts
of the program to communicate with one another. This “distant communication” can
be a useful programming paradigm, but again it should be used sparingly. Figure 4.1
illustrates this metaphor.
In the world of C++ and Java, non-const (or non-nal) global variables are the most
notorious form of reference. While globals make it easy to do certain tasks, they generally
make programs dicult to debug, since they can be altered anywhere in the code.
4.1.1 The F[S Language
Adding state to our language involves making signicant modications to the pure func-
tional languages that we’ve looked at so far. Therefore, we will spend a bit of time
developing a solid operational semantics for our state-based language before looking at
the interpreter. We will dene a language F[S: F[ with state.
The most signicant departure from the pure functional operational semantics we
have looked at so far is that the evaluation relation e ⇒ v is not sucient to capture the
semantics of a state-based language. The evaluation of an expression can now produce
side-eects, and our ⇒ rule needs to incorporate this somehow. Specically, we will need
CHAPTER 4. SIDE EFFECTS: STATE AND EXCEPTIONS 60
to have someplace to record all these side-eects: a store. In C, a store is a stack and a
heap, and memory locations are referenced by their addresses. When we are writing our
interpreter, we will only have access to the heap, so we will need to create an abstract
store in which to record side-eects.
Denition 4.1 (Store). A store is a nite map c 7→ v of cell names to values.
Cell names are an abstract representation of a memory location. A store is therefore
an abstraction of a runtime heap. The C heap is a low-level realization of a store where
the cell names are simply numerical memory addresses. A store is also known as a
dictionary from a data-structures perspective. We write Dom(S) to refer the domain of
this nite map, that is, the set of all cells that it maps.
Let us begin by extending the concrete syntax of F[ to include F[S expressions. The
additions are as follows.
• Referencing, Ref e.
• Assignment, e := e′ .
• Dereferencing, !e.
• Cell names, c.
We need cell names because Ref 5 needs to evaluate to a location, c, in the store. Because
cell names refer to locations in the heap, and the heap is initially empty, programs have
no cell names when the execution begins.
Although not part of the F[S syntax, we will need a notation to represent operations
on the store. We write S{c 7→ v} to indicate the store S extended or modied to contain
the mapping c 7→ v. We write S(c) to denote the value of cell c in store S.
Now that we have developed the notion of a store, we can dene a satisfactory
evaluation relation for F[S. Evaluation is written as follows.
〈e, S0 〉 ⇒ 〈v, S〉,
where at the start of computation, S0 is initially empty, and where S is the nal store
when computation terminates. In the process of evaluation, cells, c, will begin to appear
in the program syntax, as references to memory locations. Cells are values since they do
not need to be evaluated, so the value space of F[S also includes cells, c.
Evaluation Rules for F[S
Finally, we are ready to write the evaluation rules for F[S. We will need to modify all of
the F[ rules with the store in mind (recall that our evaluation rule is now 〈e, S0 〉 ⇒ 〈v, S〉,
not simply e ⇒ v). We do this by threading the store along the ow of control. Doing
this introduces a great deal more dependency between the rules, even the ones that do
not directly manipulate the store. We will rewrite the function application rule for F[S
to illustrate the types of changes that are needed in the other rules.
CHAPTER 4. SIDE EFFECTS: STATE AND EXCEPTIONS 61
(Function Application)
〈e1 , S1 〉 ⇒ 〈Function x -> e, S2 〉, 〈e2 , S2 〉 ⇒ 〈v2 , S3 〉, 〈e[v2 /x], S3 〉 ⇒ 〈v, S4 〉
〈e1 e2 , S1 〉 ⇒ 〈v, S4 〉
Note how the store here is threaded through the dierent evaluations, showing how
changes in the store in one place propagate to the store in other places, and in a xed order
that reects the indented evaluation order. The rules for our new memory operations
are as follows.
〈e, S1 〉 ⇒ 〈v, S2 〉
(Reference Creation)
〈Ref e, S1 〉 ⇒ 〈c, S2 {c 7→ v}〉, for c ∈
/ Dom(S2 )
〈e, S1 〉 ⇒ 〈c, S2 〉
(Dereference)
〈!e, S1 〉 ⇒ 〈v, S2 〉, where S2 (c) = v
〈e1 , S1 〉 ⇒ 〈c, S2 〉, 〈e2 , S2 〉 ⇒ 〈v, S3 〉
(Assignment)
〈e1 := e2 , S1 〉 ⇒ 〈v, S3 {c 7→ v}〉
These rules can be tricky to evaluate because the store needs to be kept up to date
at all points in the evaluation. Let us look at a few example expressions to get a feel for
how this works.
Example 4.1.
!(!(Ref Ref 5)) + 4
〈!(!(Ref Ref 5)) + 4〉, {}〉 ⇒ 〈9, {c1 7→ 5, c2 7→ c1 }〉, because
〈!(!(Ref Ref 5)), {}〉 ⇒ 〈5, {c1 7→ 5, c2 7→ c1 }〉, because
〈!(Ref Ref 5), {}〉 ⇒ 〈c1 , {c1 7→ 5, c2 7→ c1 }〉, because
〈Ref Ref 5, {}〉 ⇒ 〈c2 , {c1 7→ 5, c2 7→ c1 }〉, because
〈Ref 5, {}〉 ⇒ 〈c1 , {c1 7→ 5}〉
Example 4.2.
(Function y -> If !y = 0 Then y Else 0) Ref 7
〈(Function y -> . . . ) Ref 7, {}〉 ⇒ 〈0, {c1 7→ 7}〉, because
〈Ref 7, {}〉 ⇒ 〈c1 , {c1 7→ 7}〉, and
〈(If !y = 0 Then y Else 0)[c1 /y], {c1 7→ 7}〉 ⇒
〈0, {c1 7→ 7}〉, because
〈!c1 = 0, {c1 7→ 7}〉 ⇒ 〈False, {c1 7→ 7}〉, because
〈!c1 , {c1 7→ 7}〉 ⇒ 〈7, {c1 7→ 7}〉
CHAPTER 4. SIDE EFFECTS: STATE AND EXCEPTIONS 62
F[S Interpreters
Just as we had to modify the evaluation relation in our F[S operational semantics to
support state, writing a F[S interpreter will also require some additional work. There are
two obvious approaches to take, and we will treat them both below. The rst approach
involves mimicking the operational semantics and dening evaluation on an expression
and a store together. This approach yields a functional interpreter in which eval(e, S0 )
for expression e and initial state S0 returns the tuple (v, S), where v is the resulting
value and S is the nal state.
The second and more ecient design involves keeping track of state in a global, muta-
ble dictionary structure. This is generally how real implementations work. This approach
results in more familiar evaluation semantics, namely eval e returns v. Obviously such
an interpreter is no longer functional, but rather, imperative. We would like such an
interpreter to faithfully implement the operational semantics of F[S, and thus we would
ideally want a theorem that states that this approach is equivalent to the rst approach.
Proving such a theorem would be dicult, however, mainly because our proof would rely
on the operational semantics of OCaml, or whatever implementation language we chose.
We will therefore take it on good faith that the two approaches are indeed equivalent.
The Functional Interpreter The functional interpreter implements a stateful lan-
guage in a functional way. It threads the state through the evaluation, just as we did when
we dened our operational semantics. Imperative style programming can be “hacked”
into a functional style by threading the state along, and there are regular methods for
threading state through any functional programs, namely monads. The threading ap-
proach to modeling state functionally was rst employed by Strachey in the late 1960’s
[22]. Note that an operational semantics is always pure functional, since mathematics is
always purely functional.
To implement the functional interpreter, we model the store using a nite mapping
of integer keys to values. The skeleton of the implementation is presented below.
(* Declare all the expr, etc types globally for convenience. *)
(* The store functionality is a separate module. *)
module type STORE =
sig
(* ... *)
end
(* The Store structure implements a (functional) store. A simple
* implementation could be via a list of pairs such as
*
* [((Cell 2),(Int 4)); ((Cell 3),Plus((Int 5),(Int 4))); ... ]
*)
module Store : STORE =
CHAPTER 4. SIDE EFFECTS: STATE AND EXCEPTIONS 63
type store = (* ... *)
struct
let empty = (* initial empty store *)
let fresh = (* returns a fresh Cell name *)
let count = ref 0 in
function () -> ( count := !count + 1; Cell(!count) )
(* Note: this is not purely functional! It is quite
* difficult to make fresh purely functional.
*)
(* Look up value of cell c in store s *)
let lookup (s,c) = (* ... *)
(* Add or modify cell c to point to value v in the store s.
* Return the new store.
*)
let modify(s,c,v) = (* ... *)
end
(* The evaluator is then a functor taking a store module *)
module FbSEvalFunctor =
functor (Store : STORE) ->
struct
(* ... *)
let eval (e,s) = match e with
(Int n) -> ((Int n),s) (* values don’t modify store *)
| Plus(e,e’) ->
let (Int n,s’) = eval(e,s) in
let (Int n’,s’’) = eval(e’,s’) in
(Int (n+n’),s’’)
(* Other cases such as application use a similar store
* threading technique.
*)
| Ref(e) -> let (v,s’) = eval(e,s) in
let c = Store.fresh() in
(c,Store.modify(s’,c,v))
| Get(e) -> let (Cell(n),s’) = eval(e,s) in
(Store.lookup(Cell(n)),s’)
| Set(e,e’) -> (* exercise *)
CHAPTER 4. SIDE EFFECTS: STATE AND EXCEPTIONS 64
end
module FbSEval = FbSEvalFunctor(Store)
The Imperative Interpreter The stateful, imperative F[S interpreter is more e-
cient than its functional counterpart, because no threading is needed. In the imperative
interpreter, we represent the store as a dictionary structure (similar to Java’s HashMap
class or the C++ STL map template). The eval function needs no extra store parameter,
provided that it has a reference to the global dictionary. Non-store-related rules such as
Plus and Minus are completely ignorant of the store. Only the direct store evaluation
rules, Ref, Set, and Get actually extend, update, or query the store. A good evaluator
would also periodically garbage collect, or remove unneeded store elements. Garbage col-
lection is discussed briey in Section 4.1.4. The implementation of the stateful interpreter
is left as an exercise to the reader.
Side-Eecting Operators
Now that we have a mutable store, our code has properties beyond the value that is
returned, namely, side-eects. Operators such as sequencing (;), and While- and For
loops now become relevant. These syntactic concepts are easily dened as macros, so we
will not add them to the ocial F[S syntax. The macro denitions are as follows.
e1 ; e2 = (Function x -> e2 ) e1
While e Do e′ = Let Rec f x = If e Then f e′ Else 0 In f 0
Exercise 4.1. Why are sequencing and loop operations irrelevant in pure functional
languages like F[?
4.1.2 Cyclical Stores
An interesting phenomenon is possible with stateful languages. It is possible to make a
cyclical store, that is, a cell whose contents are a pointer to itself. Creating a cyclic
store is quite easy:
Let x = Ref 0 in x := x
This is the simplest possible store cycle, where a cell directly points to itself. This type
of store is illustrated in Figure 4.2.
Exercise 4.2. In the above example, what does !!!!!!!!x return? Can a store cycle
like the one above be written in OCaml? Why or why not? (Hint: What is the type of
such an expression?)
CHAPTER 4. SIDE EFFECTS: STATE AND EXCEPTIONS 65
Figure 4.2: A simple cycle.
A more subtle form of a store cycle is when a function is placed in the cell, and the
body of the function refers to that cell. Consider the following example.
Let c = Ref 0 In
c := (Function x -> If x = 0 Then 0 Else 1 + !c(x-1));
!c(10)
Cell c contains a function, which, in turn, refers to c. In other words, the function
has a reference to itself, and can apply itself, giving us recursion. This form of recursion
is known as tying the knot, and is the method used by most compilers to implement
recursion. A similar technique is used to make objects self-aware, although C++ explicitly
passes self, and is more like the Y -combinator.
Exercise 4.3. Tying the knot can be written in OCaml, but not directly as above. How
must the reference be declared for this to work? Why can we create this sort of cyclic
store, but not the sort described in Exercise 4.2?
Interact with F[SR. Let’s write a recursive multiplication function in F[SR,
rst using Let Rec, and then by tying the knot.
# Let Rec mult x = Function y ->
If x = 0 Then
0
Else
y + mult (x - 1) y In
mult 8 9;;
==> 72
Now we’ll use tying the knot. Because F[SR does not include a sequencing operation,
we use the encoding presented in Section 4.1.1.
# Let mult = Ref 0 In
(Function dummy -> (!mult) 9 8)
(mult := (Function x -> Function y ->
If x = 0 Then
0
Else
y + (!mult) (x - 1) y));;
==> 72
CHAPTER 4. SIDE EFFECTS: STATE AND EXCEPTIONS 66
4.1.3 The “Normal” Kind of State
Languages like C++, Java, and Scheme have a dierent form for expressing mutation.
There is no need for an explicit dereference (!) operator to get the value of a cell.
This form of mutation is more dicult to capture in an interpreter, because variables
mean something dierent depending on whether they are on the left or right side of an
assignment operator. An l-value occurs on the left side of the assignment, and represents
a memory location. An r-value to the right of an assignment, or elsewhere in the code,
and represents an actual value.
Consider the C/C++/Java assignment statement x = x + 1. The x on the left of the
assignment is an l-value, and the x in x + 1 is an r-value. In OCaml, we would write a
similar expression as x := !x + 1. Thus, OCaml is explicit in which is being referred to,
the cell or the value. For a Java l-value, the cell is needed to perform the store operation,
and for a Java r-value, the value of the cell is needed, which is why, in OCaml, we must
write !x.
l-values and r-values are distinct. Some expressions may be both l-values and r-
values, such as x and x[3]. Other values may only be r-values, such as 5, (0 == 1), and
sin(3.14159). Therefore, l-values and r-values are expressed dierently in the language
grammar, specically, l-values are a subset of r-values. There are some shortcomings
to such languages. For example, in OCaml we can say f(3) := 1, which assigns the
value 1 to the cell returned by the function f. Expressions of this sort are invalid in
Java, because f(3) is not an l-value. We could revise the F[S grammar to use this more
restrictive notion, where l-values must only be variables, in the following way:
type expr =
...
| Get of expr | Set of ident * expr | Ref of expr
For the variable on the left side of the assignment, we would need the address, and not
the contents, of the variable.
A nal issue with the standard notion of state in C and C++ is the problem of
uninitialized variables. Because variables are not required to be initialized, runtime
errors due to these uninitialized variables are possible. Note that the Ref syntax of F[S
and OCaml requires that the variable be explicitly initialized as part of its declaration.
4.1.4 Automatic Garbage Collection
Memory that is allocated may also, at some point, be freed. In C and C++, this is done
explicitly with free() and delete() respectively. However, Java, OCaml, Smalltalk,
Scheme, and Lisp, to name a few, support the automated freeing of unused memory
through a garbage collector.
Garbage collection is a large area of research (see [23] for a survey), but we will briey
sketch out what an implementation looks like.
First, something triggers the garbage collector. Generally, the trigger is that the store
is getting too full. In Java, the garbage collector may be explicitly invoked by the method
System.gc(). When the garbage collector is invoked, evaluation is suspended until the
garbage collector has completed. The garbage collector proceeds to scan through the
CHAPTER 4. SIDE EFFECTS: STATE AND EXCEPTIONS 67
current computation to nd cells that are directly used. The set of these cells is known
as the root set. Good places to look for roots are on the evaluation stack and in global
variable locations.
Once the root set is established, the garbage collector marks all cells not in the root
set as initially free. Then, it recursively traverses the memory graph starting at the root
set, and marks cells that are reachable from the root set as not free, thus at the end of
the traversal, all cells that are in a dierent connected component from any of the cells
in the root set are marked free, and these cells may be reused. There are many dierent
ways to reuse memory, but we will not go into detail here.
4.2 Environment-Based Interpreters
Now that we have discussed stateful languages, we are going to briey touch on some
eciency issues with the way we have been dening interpreters. We will focus on
eliminating explicit substitutions in function application.
A “low-level” interpreter would never duplicate the function argument for each vari-
able usage. The problem with doing so is that it can severely increase the size of the
data that must be maintained by the interpreter. Consider an expression in the following
form.
(Function x -> x x x) (whopping expr)
This expression is evaluated by computing
(whopping expr)(whopping expr)(whopping expr),
tripling the size of the data.
In order to solve this problem, we dene an explicit environment interpreter. In
an explicit environment interpreter, instead of doing direct variable substitution, we keep
track of each variable’s value in a runtime environment. The runtime environment is
simply a mapping of variables to values. We write environments as {x1 7→ v1 , x2 7→ v2 },
to mean that variable x1 maps to value v1 and variable x2 maps to value v2 .
Now, to compute
(Function x -> x x x)(whopping expr)
the interpreter computes (x x x) in the environment {x 7→ whopping expr}.
Technically, even the simple substituting interpreters we’ve looked at don’t really
make three copies of the data. Rather, they maintain a single copy of the data and three
pointers to it. This is because immutable data can always be passed by reference since
CHAPTER 4. SIDE EFFECTS: STATE AND EXCEPTIONS 68
it never changes. However, in a compiler the code can’t be copied around like this, so a
scheme like the one presented above is necessary.
There is a possibility for some anomalies with this approach, however. Specically,
there is a problem when a function is returned as the result of another function ap-
plication, and the returned function has local variables in it. Consider the following
example.
f = Function x -> If x = 0 Then
Function y -> y
Else Function y -> x + y
When f (3) is computed, the environment binds x to 3 while the body is computed, so
the result returned is
Function y -> x + y,
but it would be a mistake to simply return this as the correct value, because we would
lose the fact that x was bound to 3. The solution is that when a function value is
returned, the closure of that function is returned.
Denition 4.2 (Closure). A closure is a function along with an environment, such that
all free values in the function body are bound to values in the environment.
For the case above, the correct value to return is the closure
(Function y -> x + y, {x 7→ 3})
Theorem 4.1. A substitution-based interpreter and an explicit environment interpreter
for F[ are equivalent: all F[ programs either terminate on both interpreters or compute
forever on both interpreters.
This closure view of function values is critical when writing a compiler, since com-
piler’s should not be doing substitutions of code on the y. Compilers are discussed in
Chapter 8.
4.3 The F[SR Language
We can now dene the F[SR language. F[SR is the call-by-value language that included
the basic features of F[, and is extended to include records (F[R), and state (F[S).
In Section 4.4 and Chapters 5 and 6, we will study the language features missing from
F[SR, namely objects, exceptions, and types. In Chapter 8, we will discuss translations
for F[SR, but will not include these other language features. The abstract syntax type
for F[SR is as follows.
CHAPTER 4. SIDE EFFECTS: STATE AND EXCEPTIONS 69
type label = Lab of string
type ident = Ident of string
type expr =
Var of ident | Function of ident * expr | Appl of expr * expr |
LetRec of ident * ident * expr * expr | Plus of expr * expr |
Minus of expr * expr | Equal of expr * expr |
And of expr * expr | Or of expr * expr | Not of expr |
If of expr * expr * expr | Int of int | Bool of bool |
Ref of expr | Set of expr * expr | Get of expr | Cell of int |
Record of (label * expr) list | Select of label * expr |
Let of ident * expr * expr
type fbtype =
TInt | TBool | TArrow of fbtype * fbtype | TVar of string |
TRecord of (label * fbtype) list | TCell of fbtype;;
In the next two sections we will look at some nontrivial “real-world” F[SR programs
to illustrate the power we can get from this simple language. We begin by considering
a function to calculate the factorial function, and conclude the chapter by examining an
implementation of the merge sort algorithm.
4.3.1 Multiplication and Factorial
The factorial function is fairly easy to represent using F[SR. The main point of interest
here is the encoding of multiplication using a curried Let Rec denition. This example
assumes a positive integer input.
(*
* First we encode multiplication for positive nonnegative
* integers. Notice the currying in the Let Rec construct.
* Multiplication is encoded in the obvious way: repeated
* addition.
*)
Let Rec mult x = Function y ->
If y = 0 Then
0
Else
x + (mult x (y - 1)) In
(*
* Now that we have multiplication, factorial is easy to
* define.
*)
Let Rec fact x =
If x = 0 Then
CHAPTER 4. SIDE EFFECTS: STATE AND EXCEPTIONS 70
1
Else
mult x (fact (x - 1)) In
fact 7
4.3.2 Merge Sort
Writing a merge sort in F[SR is a fairly comprehensive example. One of the biggest
challenges is encoding integer comparisons, i.e. <, >, etc. Let’s discuss how this is
accomplished before looking at the code.
First of all, given that we already have an equality test, =, encoding the <= operation
basically gives us the other standard comparison operations for free. Assuming that we
have a lesseq function, the other operations can be trivially encoded as follows.
Let lesseq = (* real definition *) In
Let lessthan = (Function x -> Function y ->
(lesseq x y) And Not (x = y)) In
Let greaterthan = (Function x -> Function y ->
Not (lesseq x y)) In
Let greatereq = (Function x -> Function y ->
(greaterthan x y) Or (x = y)) In ...
Therefore it suces to encode lesseq. But how can we do this using only the regular
F[SR operators? The basic idea is as follows. To test if x is less than or equal to y we
compute a z such that x + z = y. If z is nonnegative, we know that x is less than or
equal to y. If z is negative, we know that x is greater than y.
At rst glance, we seem to have a “chicken and egg” problem here. How can we tell
if z is negative if we have no comparison operator. And how do we actually compute
z to begin with? One idea would be to start with z = 0 and loop, incrementing z by
1 at every step, and it testing if x + z = y. If we nd a z, we stop. We know that z
is positive, and we conclude that x is less than or equal to y. If we don’t nd a z, we
start at z = −1 and perform a similar loop, decrementing z by 1 every step. If we nd
a proper value of z this way, we conclude that x is greater then y.
The aw in the above scheme should be obvious: if x > y, the rst loop will never
terminate. Indeed we would need to run both loops in parallel for this idea to work!
Obviously F[SR doesn’t allow for parallel computation, but there is a solution along
these lines. We can interleave the values of z that are tested by the two loops. That is,
we can try the values {0, −1, 1, −2, 2, −3, . . .}.
The great thing about this approach is that every other value of z is negative, and we
can simply pass along a boolean value to represent the sign of z. If z is nonnegative, the
next iteration of the loop inverts it and subtracts 1. If z is negative, the next iteration
CHAPTER 4. SIDE EFFECTS: STATE AND EXCEPTIONS 71
simply inverts it. Note that we can invert the sign of an number x in F[SR by simply
writing 0 - x. Now, armed with our ideas about lesseq, let us start writing our merge
sort.
(* We need to represent the empty list somehow. Since we
* will need to test for it, and since equality is only
* defined on integers, emptylist will need to be an
* integer. We define it as -1.
*)
Let emptylist = (0 - 1) In
(* Next, let’s define some list operations, head, tail,
* cons, and length. There encodings are straightforward,
* and were covered in the text. Notice that we are
* encoding lists as records, using {l,r} records like
* pairs.
*)
Let head = Function seq -> seq.l In
Let tail = Function seq -> seq.r In
Let cons = Function elt -> Function seq -> {l=elt; r=seq} In
Let Rec length seq =
If seq = emptylist Then
0
Else
1 + length (seq.r) In
(* Now, we’re ready to define lesseq. Notice how lesseq is
* a wrapper function that uses le. le passes along the
* sign of v as an argument, as discussed in the text.
*)
Let lesseq = Function a -> Function b ->
Let Rec le x =
Function y -> Function v ->
Function v is non neg ->
If (x + v) = y Then
v is non neg
Else
If v is non neg Then
le x y (0 - v - 1) (Not v is non neg)
Else
le x y (0 - v) (Not v is non neg) In
CHAPTER 4. SIDE EFFECTS: STATE AND EXCEPTIONS 72
le a b 0 True In
(* This function takes a list and splits it into nearly
* equal halves. These halves are returned as a pair
* (encoded as an {left, right} record).
*)
Let split = Function seq ->
Let Rec splt seq1 = Function seq2 ->
If lesseq (length seq1) (length seq2) Then
{left=seq1; right=seq2}
Else
splt (tail seq1) (cons (head seq1) seq2) In
splt seq emptylist In
(* Here is where we merge two sorted lists. We scan through
* each list in parallel, chopping off the smaller of the
* two list heads and appending it to the result. This is
* where we make use of our lesseq function.
*)
Let Rec merge seq1 = Function seq2 ->
If seq1 = emptylist Then
seq2
Else If seq2 = emptylist Then
seq1
Else
If lesseq (head seq1) (head seq2) Then
cons (head seq1) (merge (tail seq1) seq2)
Else
cons (head seq2) (merge seq1 (tail seq2)) In
(* mergesort sorts a single list by breaking it into two
* smaller lists, recursively sorting those lists, and
* merging the two back into a single list. Recall that
* the base cases of the recursion are a single element list
* and an empty list, both of which are necessarily in
* sorted order by definition.
*)
Let Rec mergesort seq =
If lesseq (length seq) 1 Then
seq
Else
Let halves = split seq In
merge (mergesort (halves.left))
(mergesort (halves.right)) In
(* Finally we call mergesort on an actual list. Notice the
CHAPTER 4. SIDE EFFECTS: STATE AND EXCEPTIONS 73
* record encoding.
*)
mergesort {l=5; r=
{l=6; r=
{l=2; r=
{l=1; r=
{l=4; r=
{l=7; r=
{l=8; r=
{l=10; r=
{l=9; r=
{l=3; r=emptylist}}}}}}}}}}
4.4 Exceptions and Other Control Operations
Until now, expressions have been evaluated by marching through the evaluation rules one
at a time. To perform addition, the left and right operands were evaluated to values, after
which the addition expression itself was evaluated to a value. This addition expression
may have been part of a larger expression, which could then itself have been evaluated
to a value, etc.
In this section, we will discuss explicit control operations, concentrating on excep-
tions. Explicit control operations are operations that explicitly alter the control of the
evaluation process. Even simple languages have control operations. A common example
is the return statement in C++ and Java.
In F[, the value of the function is whatever its entire body evaluates to. If, in the
middle of some complex conditional loop expression, we have the nal result of the
computation, it is still necessary to complete the execution of the function. A return
statement gets around this problem by immediately returning from the function and
aborting the rest of the function computation.
Another common control operation is the loop exit operation, or break in C++ or
Java. break is similar to return, except that it exits the current loop instead of the
entire function.
These types of control operations are interesting, but in this section, we will concen-
trate more on two more powerful control operations, namely exceptions and exception
handlers. The reader should already be familiar with the basics of exceptions from
working with the OCaml exception mechanism.
There are some other control operations that we will not discuss, such as the call/cc,
shift/reset, and control/prompt operators.
We will avoid the goto operator in this section, except for the following brief discus-
sion. goto basically jumps to a labeled position in a function. The main problem with
goto is that it is too raw. The paradigm of jumping around between labels is not all that
useful in the context of functions. It is also inherently dangerous, as one may inadver-
tently jump into the middle of a function that is not even executing, or skip past variable
CHAPTER 4. SIDE EFFECTS: STATE AND EXCEPTIONS 74
initializations. In addition, with a rich enough set of other control operators, goto really
doesn’t provide any more expressiveness, at least not meaningful expressiveness.
The truth is that control operators are really not needed at all. Recall that F[, and
the lambda calculus for that matter, are already Turing-complete. Control operators are
therefore just conveniences that make programming easier. It is useful to think of control
operators as “meta-operators,” that is, operators that act on the evaluation process itself.
4.4.1 Interpreting Return
How are exceptions interpreted? Before we answer this question, we will consider adding
a Return operator to F[, since it is a simpler control operator than exceptions. The
trouble with Return, and with other control operators, is that it doesn’t t into the
normal evaluation scheme. For example, consider the expression
(Function x ->
(If x = 0 Then 5 Else Return (4 + x)) - 8) 4
Since x will not be 0 when the function is applied, the Return statement will get evalu-
ated, and execution should stop immediately, not evaluating the “- 8.” The problem is
that evaluating the above statement means evaluating
(If 4 = 0 Then 5 Else Return (4 + 4)) - 8,
which, in turn, means evaluating
(Return (4 + 4)) - 8.
But we know that the subtraction rule works by evaluating the left and right hand sides
of this expression to values, and performing integer subtraction on them. Clearly that
doesn’t work in this case, and so we need a special rules for subtraction with a Return
in one of the subexpressions.
First, we need to add Returns to the value space of F[ and provide an appropriate
evaluation rule:
e⇒v
(Return)
Return e ⇒ Return v
Next, we need the special subtraction rules, one for when the Return is on the left side,
and one for when the Return is on the right side.
CHAPTER 4. SIDE EFFECTS: STATE AND EXCEPTIONS 75
e ⇒ Return v
(- Return Left)
e - e′ ⇒ Return v
e ⇒ v, e′ ⇒ Return v ′
(- Return Right) v is not of the form Return v ′′
e - e′ ⇒ Return v ′
Notice that these subtraction rules evaluate to Return v and not simply v. This
means that the Return is “bubbled up” through the subtraction operator. We need to
dene similar return rules for every F[ operator. Using these new rules, it is clear that
Return (4 + 4) - 8 ⇒ Return 8.
Of course, we don’t want the Return to bubble up indenitely. When the Return
pops out of the function application, we only get the value associated with it. In other
words, our original expression,
(Function x ->
(If x = 0 Then 5 Else Return (4 + x)) - 8) 4
should evaluate to 8, not Return 8. To accomplish this, we need a special function
application rule.
e1 ⇒ Function x -> e, e2 ⇒ v2 , e[v2 /x] ⇒ Return v
(Appl. Return)
e1 e2 ⇒ v
A few other special application rules are needed for the cases when either the function
or the argument itself is a Return.
e1 ⇒ Return v
(Appl. Return Function)
e1 e2 ⇒ Return v
e1 ⇒ v1 , e2 ⇒ Return v
(Appl. Return Arg.)
e1 e2 ⇒ Return v
Of course, we still need the original function application rule (see Section 2.3.3) for the
case that function execution implicitly returns a value by dropping o the end of its
execution.
Let us conclude our discussion of Return by considering the eect of Return Return e.
There are two possible interpretations for such an expression. By the above rules, this
expression returns from two levels of function calls. Another interpretation would be to
add the following rule:
CHAPTER 4. SIDE EFFECTS: STATE AND EXCEPTIONS 76
e ⇒ Return v
(Double Return)
Return e ⇒ Return v
Of course we would need to restrict the original Return rule to the case where v was not
in the form Return v. With this rule, instead of returning from two levels of function
calls, the second Return actually interrupts and bubbles through the rst. Of course,
double returns are not a common construct, and these rules will not be used often in
practice.
4.4.2 The F[X Language
The translation of Return that was given above can easily be extended to deal with
general exceptions. We will dene a language F[X, which is F[ extended with a OCaml-
style exception mechanism. F[X does not have Return (nor does OCaml), but Return
is easily encodable with exceptions. For example, the “pseudo-OCaml” expression
(function x -> (if x = 0 then 5 else return (4 + x)) - 8) 4
can be encoded in the following manner.
exception Return of int;;
(function x ->
try
(if x = 0 then 5 else raise (Return (4 + x))) - 8
with
Return n -> n) 4;;
Return can be encoded in other functions in a similar manner.
Now, let’s dene our F[X language. The basic idea is very similar to the Return
semantics that we discussed above. We dene a new kind of value,
Raise #xn v,
which bubbles up the exception #xn. This is the generalization of the value class Return v
from above. #xn is a metavariable representing an exception name. An exception contains
two pieces of data: a name, and an argument. The argument can be an arbitrary expres-
sion. Although we only allow single-argument exceptions, zero-valued or multi-valued
versions of exceptions can be easily encoded by, for example, supplying the exception with
a record argument with zero, one, or several elds. We also add to F[X the expression
Try e With #xn x -> e′
CHAPTER 4. SIDE EFFECTS: STATE AND EXCEPTIONS 77
Note that Try binds free occurrences of x in e′ . Also notice that the F[X Try syntax
diers slightly from the OCaml syntax in that OCaml allows an arbitrary pattern-match
expression in the With clause. We allow only a single clause that matches all values of
#x.
F[X is untyped, so exceptions do not need to be declared. We use the “#” symbol to
designate exceptions in the concrete syntax, for example, #MyExn. Below is an example
of a F[X expression.
Function x -> Try
(If x = 0 Then 5 Else Raise #Return (4 + x)) - 8
With #Return n -> n) 4
Exceptions are side-eects, and can cause “action at a distance.” Therefore, like any
other side-eects, they should be used sparingly.
4.4.3 Implementing the F[X Interpreter
The abstract syntax type for F[X is as follows.
type exnlab = string
type expr =
...
| Raise of exnlab * expr
| Try of expr * exnlab * ident * expr
The rules for Raise and Try are derived from the return rule and the application
return rule respectively. Raise “bubbles up” an exception, just like Return bubbled
itself up. Try is the point at which the bubbling stops, just like function application was
the point at which a Return stopped bubbling. The operational semantics of exceptions
are as follows.
e ⇒ v, for v not of the form Raise . . .
(Raise)
Raise #xn e ⇒ Raise #xn v
e ⇒ v for v not of the form Raise #xn v
(Try)
Try e With #xn x -> e′ ⇒ v
e ⇒ Raise #xn v, e′ [v/x] ⇒ v ′
(Try Catch)
Try e With #xn x -> e′ ⇒ v ′
In addition, we need to add special versions of all of the other F[ rules so that the
Raise bubbles up through the computation just as the Return did. For example
e ⇒ Raise #xn v
(- Raise Left)
e - e′ ⇒ Raise #xn v
CHAPTER 4. SIDE EFFECTS: STATE AND EXCEPTIONS 78
Note that we must handle the unusual case of when a Raise bubbles through a Raise,
something that will not happen very often in practice. The rule is very much like the “-
Raise Left” rule above.
e ⇒ Raise #xn v
(Raise Raise)
Raise e ⇒ Raise #xn v
Now, let’s trace through the execution of
Function x -> Try
(If x = 0 Then 5 Else Raise #Return (4 + x)) - 8
With #Return n -> n) 4
After the function application and the evaluation of the If statement, we are left with
Try (Raise #Return (4 + 4)) - 8
With #Return n -> n
which is
Try (Raise #Return 8) - 8
With #Return n -> n
which by bubbling in the subtraction computes to
Try Raise #Return 8
With #Return n -> n
which by the Try Catch rule returns the value 8, as expected.
Chapter 5
Object-Oriented Language
Features
Object-oriented programming has become an industry standard. Yet, object-oriented
features do not fundamentally add much to a language. They certainly do not lead to
shorter programs. The success of object-oriented programming is due mainly to the fact
that it is a style that is appropriate for the human psychology. Humans, as part of their
basic function, are highly adept at recognizing and interacting with everyday objects.
And, objects in programming are enough like objects in the everyday world that our rich
intuitions can apply to them.
Before we discuss the properties of objects in object-oriented programming, let us
briey review some of the more important properties of everyday objects that would
make them useful for programming.
• Everyday objects are active, that is, they are not fully controlled by us. Objects
have internal and evolving state. For instance, a running car is an active object,
and has a complex internal state including the amount of gas remaining, the engine
coolant and transmission uid levels, the amount of battery power, the oil level,
the temperature, the degree of wear and tear on the engine components, etc.
• Everyday objects are communicative. We can send messages to them, and we can
get feedback from objects as a result of sending them messages. For example,
starting a car and checking the gas gauge can both be viewed as sending messages
to the car.1
• Everyday objects are encapsulated. They have internal properties that we can not
see, although we can learn some of them by sending messages. To continue the car
example, checking the gas gauge tells us how much gas is remaining, even though
we can’t see into the gas tank directly, and we don’t know if it contains regular or
premium gas.
1
It may, at rst, seem unnatural to view checking the gas gauge as sending a message. After all, it is
really the car sending a message to us. We view ourselves as the “sender,” however, because we initiated
the check, that is, in a sense, we asked the car how much gas was remaining.
79
CHAPTER 5. OBJECT-ORIENTED LANGUAGE FEATURES 80
• Everyday objects may be nested, that is, objects may be made up of several smaller
object components, and those components may themselves have smaller compo-
nents. Once again, a car is a perfect example of a nested object, being made of
of smaller objects including the engine and the transmission. The transmission is
also a nested object, including an input shaft, an output shaft, a torque converter,
and a set of gears.
• Everyday objects are, for the most part, uniquely named. Cars are uniquely named
by license plates or vehicle registration numbers.
• Everyday objects may be self-aware, in the sense that they may intentionally in-
teract with themselves, for instance a dog licking his paw.
• Everyday object interactions may be polymorphic in that a diverse set of objects
may share a common messaging protocol, for example the accelerator/brake/steer-
ing wheel messaging system common to all models of cars and trucks.
The objects in object-oriented programming also have these properties. For this
reason, object-oriented programming has a natural and familiar feel to most people. Let
us now consider objects of the programming variety.
• Objects have an internal state in the form of instance variables or elds. Objects
are generally active, and their state is not fully controlled by their callers.
• Objects support messages, which are named pieces of code that are tied to a particu-
lar object. In this way objects are communicative, and groups of objects accomplish
tasks by sending messages to each other.
• Objects consist of encapsulated code (methods or operations) along with a mutable
state (instance variables or elds). Right away it should be clear that objects are
inherently non-functional. This mirrors the statefulness of everyday objects.
• Objects are typically organized in a nested fashion. For example, consider an
object that represents a graphical web browser. The object itself is frame, but that
frame consists of a toolbar and a viewing area. The toolbar is made up of buttons,
a location bar, and a menu, while the viewing area is made up of a rendered
panel and a scroll bar. This concept is illustrated in Figure 5.1. Object nesting
is generally accomplished by the outer object storing its inner objects as elds or
instance variables.
• Objects have unique names, or object references, to refer to them. This is analogous
to the naming of real-world objects, with the advantage that object references are
always unique whereas real-world object names may be ambiguous.
• Objects are self aware, that is, objects contain references to themselves. this in
Java and self in Smalltalk are self-references.
• Objects are polymorphic, meaning that a “fatter” object can always be passed to
a method that takes a “thinner” one, that is, one with fewer methods and public
elds.
CHAPTER 5. OBJECT-ORIENTED LANGUAGE FEATURES 81
Figure 5.1: A nested object representing a web browser.
CHAPTER 5. OBJECT-ORIENTED LANGUAGE FEATURES 82
There are several additional features objects commonly have. Classes are nearly
always present in languages with objects. Classes are not required: it is perfectly valid
to have objects without classes. The language Self [19, 20, 2] has no classes, instead it has
prototype objects which are copied that do the duty of classes. JavaScript copied this idea
from Self and so now many programmers are using prototype-based objects. Important
concepts of classes include creation, inheritance, method overriding, superclass access,
and dynamic dispatch. We will address these concepts later.
Information hiding for elds and methods is another feature that most object-oriented
languages have, generally in the form of public, private, and protected keywords.
Other aspects of objects include object types and modules. Types are discussed
briey, but modules are beyond the scope of this book. We also ignore method over-
loading, as it is simply syntactic sugar and adds only to readability, not functionality.
Recall that overloading a method means there are two methods with the same name,
but dierent type signatures. Overriding means redening the behavior of a superclass’s
method in a subclass. Overriding is discussed in Section 5.1.5.
5.1 Encoding Objects in F[SR
One of the most important aspects of objects is how close they are to existing concepts
that we have already discussed. In this section we demonstrate how objects may easily be
encoded in F[SR. We will model objects as records of functions and references. Record
labels and values can be thought of as slots for either methods or elds. A slot with a
function is a method, and a slot with a reference is a eld.
5.1.1 Simple Objects
Consider the object in Figure 5.22 that represents a point in two-dimensional space. The
object has elds x and y, along with two methods: magnitude and iszero, with obvious
functionality. To encode this object as a record, we are going to need a record that has
the following structure.
Let point = {
x = 4;
y = 3;
magnitude = Function -> . . . ;
iszero = Function -> . . .
} In . . .
We can’t write the magnitude method yet, because we need to dene it in terms of
x and y, but there is no way to refer to them in the function body. The solution we
will use is the same one that C++ uses behind the scenes: we pass the object itself as an
argument to the function. Let’s revise our rst encoding to the following one (assuming
we’ve dened a sqrt function).
2
We use UML to diagram objects. UML is fully described in [11]. Although the diagram in Figure
5.2 is technically a class diagram, for our purposes we will view it simply as the induced object.
CHAPTER 5. OBJECT-ORIENTED LANGUAGE FEATURES 83
Figure 5.2: The “point” object.
Let point = {
x = 4;
y = 3;
magnitude = Function this -> Function ->
sqrt(sqr this.x + sqr this.y);
iszero = Function this -> Function ->
((this.magnitude) this {}) = 0
} In . . .
There are a few points of interest about the above example. First of all, the object
itself needs to be explicitly passed as an argument when we send a message. For example,
to send the magnitude message to point, we would write
point.magnitude point {}
For convenience, we can use the abbreviation obj <- method to represent the message
(obj.method obj).
Even inside the object’s methods, we need to pass this along when we call another
method. iszero illustrates this point in the way it calls magnitude. This encoding of
self-reference is called the self-application encoding. There are a number of other
encodings, and we will treat some of them below.
There is a problem with this encoding, though; our object is still immutable. An
object with immutable elds can be a good thing in many cases. For example, windows
that can’t be resized and sets with xed members can be implemented with immutable
elds. In general, though, we need our objects to support mutable elds. Luckily this
problem has an easy solution. Consider the following revision of our point encoding,
which uses Refs to represent elds.
CHAPTER 5. OBJECT-ORIENTED LANGUAGE FEATURES 84
Let point = {
x = Ref 4;
y = Ref 3;
magnitude = Function this -> Function ->
sqrt(sqr !(this.x) + sqr !(this.y));
iszero = Function this -> Function ->
(this.magnitude this {}) = 0;
setx = Function this -> Function newx -> this.x := newx;
sety = Function this -> Function newy -> this.y := newy
} In . . .
To set x to 12, we can write either point <- setx 12, or we can directly change the
eld with (point.x) := 12. To access the eld, we now write !(point.x), or we could
dene a getx method to abstract the dereferencing. This strategy gives us a faithful
encoding of simple objects. In the next few sections, we will discuss how to encode some
of the more advanced features of objects.
5.1.2 Object Polymorphism
Suppose we dene the following function.
Let tallerThan = Function person1 -> Function person2 ->
greatereq (person1 <- height) (person2 <- height)
If we were to dene person objects that supported the height message, we could pass
them as arguments to this function. However, we could also create specialized person
objects, such as mother, father, and child. As long as these objects still support
the height message, they are all valid candidates for the tallerThan function. Even
objects like dinosaurs and buildings can be arguments. The only requirement is that
any objects passed to tallerThan support the message height. This is known as object
polymorphism.
We already encountered a similar concept when we discussed record polymorphism.
Recall that a function
Let getheight = Function r -> r.height . . .
can take any record with a height eld:
. . . In getheight {radius = 4; height = 4; weight = 44}
CHAPTER 5. OBJECT-ORIENTED LANGUAGE FEATURES 85
Object polymorphism is really the same as record polymorphism, evidenced by how
we view objects as records. So simply by using the record encoding of objects, we can
easily get object polymorphism.
Let eqpoint = {
(* all the code from point above: x, y, magnitude . . . *)
equal = Function this -> Function apoint ->
!(this.x) = !(apoint.x) And !(this.y) = !(apoint.y)
} In eqpoint <- equal({
x = Ref 3;
y = Ref 7;
(* . . . *)
})
The object passed to equal needs only to dene x and y. Object polymorphism in
our embedded language is thus more powerful than in C++ or Java: C++ and Java look
at the type of the arguments when deciding what is allowed to be passed to a method.
Subclasses can always be passed to methods that take the superclass, but nothing else is
allowed. In our encoding, which is closer to Smalltalk, any object is allowed to be passed
to a function or method as long as it supports the messages needed by the function.
One potential diculty with object polymorphism is that of dispatch: since we
don’t know the form of the object until runtime, we do not know exactly where the
correct methods will be laid out in memory. Thus hashing may be required to look up
methods in a compiled object-oriented language. This is exactly the same problem that
arises when writing the record-handling code in our F[SR compiler (see Chapter 8),
again illustrating the similarities between objects and records.
5.1.3 Information Hiding
Most object-oriented languages allow information hiding which is used to protect
methods and instances variables from direct outside use. Information hiding is one of
the key tools for the encapsulation of object data. In C++ and Java, the public, private,
and protected qualiers are used to control the degree of information hiding for both
elds and methods.
For now we will simply encode hiding in F[SR. Note that only public and private
data makes sense in F[SR (protected data only makes sense in the context of classes
and inheritance, which we have not dened yet). In our encoding, we accomplish hiding
by simply making data inaccessible. In real, typed languages, it is the type system itself
(and the bytecode verier in the case of Java) that enforces the privacy of data.
Let’s begin with a partial encoding of information hiding.
CHAPTER 5. OBJECT-ORIENTED LANGUAGE FEATURES 86
Let pointImpl = (* point from before *) In
Let pointInterface = {
magnitude = pointImpl.magnitude pointImpl;
setx = pointImpl.setx pointImpl;
sety = pointImpl.sety pointImpl
} In . . .
In this encoding, each method is “preapplied” to pointImpl, the full point object,
while pointInterface contains only public methods and instances. Methods are now
invoked simply as
pointInterface.setx 5
This solution has a aw, though. Methods that return this re-expose the hidden elds
and methods of the full object. Consider the following example.
Let pointImpl = {
(* . . . *)
sneaky = Function this -> Function -> this
} In Let pointInterface = {
magnitude = pointImpl.magnitude pointImpl;
setx = pointImpl.setx pointImpl;
sety = pointImpl.sety pointImpl;
sneaky = pointImpl.sneaky pointImpl
} In pointInterface.sneaky {}
The sneaky method returns the full pointImpl object instead of the pointInterface
version. To remedy this problem, we need to change the encoding so that we don’t have
to pass this every time we invoke a method. Instead, we will give the object a pointer
to itself from the start.
Let prePoint = Let privateThis = {
x = Ref 4;
y = Ref 3
} In Function this -> {
magnitude = Function ->
sqrt !(privateThis.x) + !(privateThis.y);
setx = Function newx -> privateThis.x := newx;
sety = Function newy -> privateThis.y := newy;
getThis = Function -> this
} In Let point = prePoint prePoint In . . .
CHAPTER 5. OBJECT-ORIENTED LANGUAGE FEATURES 87
Now the message send operation is just
point.magnitude {}
The method getThis still returns this, but this contains the public parts only. Note
that for this encoding to work, privateThis can be used only as a target for messages,
and can not be returned.
The disadvantage of this encoding is that this will need to be applied to itself every
time it’s used inside the object body. For example, instead of writing
(point.getThis {}).magnitude {}
we would, instead, have to write
((point.getThis {}) (point.getThis {})).magnitude {}
These encodings are relatively simple. Classes and inheritance are more dicult
to encode, and are discussed below. Object typing is also particularly dicult, and is
covered in Chapter 6.
5.1.4 Classes
Classes are foremost templates for creating objects. A class is, in essence, an object
factory. Each object that is created from a class must have its own unique set of instance
variables (with the exception of static elds, which we will treat later).
It is relatively easy to produce a simple encoding of classes. Simply freeze the code
that creates the object, and thaw to create new object. Ignoring information hiding, the
encoding looks as follows.
Let pointClass = Function -> {
x = Ref 4;
y = Ref 3;
magnitude = Function this -> Function ->
sqrt !(this.x) + !(this.y);
setx = Function this -> Function newx -> this.x := newx
sety = Function this -> Function newy -> this.y := newy
} In . . .
We can dene new pointClass to be pointClass {}. Some typical code which creates
and uses instances might look as follows.
CHAPTER 5. OBJECT-ORIENTED LANGUAGE FEATURES 88
Let point1 = pointClass {} In
Let point2 = pointClass {} In
point1 <- setx 5 . . .
point1 and point2 will have their own x and y values, since Ref creates new store cells
each time it is thawed. The same freeze and thaw trick can be applied to our encoding
of information hiding to get hiding in classes. The dicult part of encoding classes is
encoding inheritance, which we discuss in the next section.
A more useful notion of class is to not just freeze an object, but make a function
returning the object; the parameters to that function are analogous to the constructor
values of a class.
Let pointClass’ = Function ix -> Function iy -> {
x = Ref ix;
y = Ref iy;
magnitude = Function this -> Function ->
sqrt !(this.x) + !(this.y);
setx = Function this -> Function newx -> this.x := newx
sety = Function this -> Function newy -> this.y := newy
} In . . .
Then pointClass’ 0 0 creates a point object initialized to (0, 0).
5.1.5 Inheritance
As a rule, about 80 percent of the utility of objects is realized in the concepts we have
talked about above, namely
• Objects that encapsulate data and code under a single name to achieve certain
functionality.
• Polymorphic objects.
• Templates for generating objects (the main function of classes)
The other 20 percent of the utility comes from from inheritance. Inheritance allows
related objects to be dened that share some common code. Inheritance can be encoded
in F[SR by using the following scheme. In the subclass, we create an instance of the
superclass object, and keep the instance around as a “slave.” We use the slave to access
methods and elds of the superclass object. Thus, the subclass object only contains
new and overridden methods. Real object oriented languages tend not to implement
inheritance this way for reasons of eciency (imagine a long inheritance chain in which
a method call has to be delegated all the way back to the top of the hierarchy before
it can be invoked). Still, the encoding is good enough to illustrate the main points of
inheritance. For example, consider the following encoding of ColorPoint, a subclass
CHAPTER 5. OBJECT-ORIENTED LANGUAGE FEATURES 89
Figure 5.3: The point, colorPoint inheritance hierarchy.
of Point, illustrated in Figure 5.3. (In this encoding we will assume there are no class
constructor parameters, i.e. we will work from pointClass and not pointClass’ above.)
Let pointClass = . . . In
Let colorPointClass = Function ->
Let super = pointClass {} In {
x = super.x; y = super.y;
color = Ref {red = 45; green = 20; blue = 20};
magnitude = Function this -> Function ->
mult(super.magnitude this {})(this.brightness this {});
brightness = Function this -> Function ->
(* compute brightness. . . *)
setx = super.setx; sety = super.sety
} In . . .
There are several points of interest in this encoding. First of all, notice that to inherit
methods and elds from the superclass, we explicitly link them together (i.e. x, y, setx,
and sety). To override a method in the superclass, we simply redene it in the subclass
instead of linking to the superclass; magnitude is an example of this. Also notice that we
can still invoke superclass methods from the subclass. For instance, magnitude invokes
super.magnitude in its body. Notice how super.magnitude is passed this instead of
super as its argument. This has to do with dynamic dispatch, which we will address
now.
5.1.6 Dynamic Dispatch
We say that a method is dynamically dispatched if, looking at a message send v<-m, we
are not sure precisely what method m will be executed at runtime. Dynamic dispatch
is related to object polymorphism: a variable v could contain many dierent kinds of
objects, so v<-m could be sending m to any one of those dierent kinds of objects. Unless
we know what kind of objects v is at runtime, we cannot know which method m will be
invoked.
The term dynamic dispatch refers to how the method invoked by a message send
is not xed at compile-time. If we had a method isNull declared in pointClass and
inherited by colorPointClass with code
CHAPTER 5. OBJECT-ORIENTED LANGUAGE FEATURES 90
Function this -> Function -> (this.magnitude this {}) = 0,
the magnitude method here is not xed at compile-time. Concretely, assuming pointClass
has this isNull and we execute
Let p = pointClass {} In
Let cp = colorPointClass {} In
p <- isNull {}; cp <- isNull {};
In p’s call to isNull, this is a point, and so isNull will internally invoke point’s
magnitude method. On the other hand for cp, this is a colorPoint, and so colorPoint’s
magnitude method will be invoked. In conclusion, the magnitude method call from
within isNull is dynamically dispatched.
The superclass dispatching above also now can be explained. If isNull in colorPointClass
had additionally been overridden to be
Function this -> Function -> (super.isNull this {}) = 0 And (* etc *),
then tracing through cp’s call to isNull it would have correctly called the colorPointClass
magnitude, whereas if we had instead written
Function this -> Function -> (super.isNull super {}) = 0 And (* etc *),
(note change in bold), the pointClass magnitude would have been invoked from within
isNull instead, and the brightness would mistakenly not have been taken into account
in the magnitude calculation for a colorPoint.
Another example For additional clarication on this issue let us do one more exam-
ple. Consider the following classes, rectClass and its subclass squareClass.
Let rectClass = Function w -> Function l -> {
getWidth = Function this -> Function -> w;
getLength = Function this -> Function -> l;
area = Function this -> Function ->
mult ((this.getLength) this {}) ((this.getWidth) this {})
} In
Let squareClass = Function e ->
Let super = rectClass 1 e In {
getLength = (super.getLength);
CHAPTER 5. OBJECT-ORIENTED LANGUAGE FEATURES 91
(* We override width to be the same as length *)
getWidth = (super.getLength);
areaStatic = Function this -> Function ->
(super.area) super {};
areaDynamic = Function this -> Function ->
(super.area) this {}
} In
Let mySquare = squareClass 10 In
mySquare <-areaDynamic {} (* returns 100 *);
mySquare <-areaStatic {} (* returns 10 *)
Notice that in the squareClass, getLength has been overridden to behave the same
as getWidth. There are two ways to calculate the area in squareClass. areaStatic calls
(super.area) super {}. This means that when rectClass’s area method is invoked,
it is invoked with rectClass’s getLength and getWidth as well. The result is 1×10 = 10
since the rectangle was initialized to have width 1.
On the contrary, the dynamically dispatched area method, areaDynamic is written
(super.area) this {}. This time, rectClass’s area method is invoked, but this
is an instance of squareClass, rather than rectClass, and squareClass’s overridden
getLength and getWidth are used. The result is 1, the correct area for a square. The
behavior of dynamic dispatch is almost always the behavior we want. The key to dy-
namic dispatch is that inherited methods get a revised notion of this when they are
inherited. Our encoding promotes dynamic dispatch, because we explicitly pass this
into our methods.
Java (and almost every other object-oriented language) uses dynamic dispatch to
invoke methods by default. In C++, however, unless a method is declared as virtual, it
will not be dynamically dispatched.
5.1.7 Static Fields and Methods
Static elds and methods are found in almost all object-oriented languages, and are
simply elds and methods that are not tied to a particular instance of an object, but
rather to the class itself.
We can trivially encode static elds and methods by simply making our class def-
initions records instead of functions. The creation of the object can itself be a static
method of the class. This is, in fact, how Smalltalk implements constructors. Consider
the following reimplementation of pointClass in which we make use of static elds.
Let pointClass = {
newWithXY = Function class ->
Function newx ->
Function newy -> {
x = Ref newx;
y = Ref newy;
CHAPTER 5. OBJECT-ORIENTED LANGUAGE FEATURES 92
magnitude = Function this -> Function ->
sqrt ((!(this.x)) + (!(this.y)))
};
new = Function class -> Function ->
(class.newWithXY) class (class.xdefault)
(class.ydefault);
xdefault = 4;
ydefault = 3
} In
Let point = (pointClass.new) pointClass {} In
(point.magnitude) point {}
Notice how the class method newWithXY is actually responsible for building the point
object. new simply invokes newWithXY with some default values that are stored as class
elds. This is a very clean way to encode multiple constructors.
Perhaps the most interesting thing the encoding is how classes with static elds start
to look like our original encoding of simple objects. Look closely—notice that class
methods take an argument class as their rst parameter, for the exact same reason that
regular methods take this as a parameter. So in fact, pointClass is really just another
primitive object that happens to be able to create objects.
Viewing classes as objects is the dominant paradigm in Smalltalk. Java has some
support for looking at classes as objects through the reection API as well. Even in
languages like C++ that don’t view classes as objects, design patterns such as the Factory
patterns[12] capture this notion of objects creating objects.
This encoding is truly in the spirit of object-oriented programming, and it is a clean,
and particularly satisfying way to think about classes and static members.
5.2 The F[OB Language
Now that we have looked at encodings of objects in terms of the known syntax of F[SR,
we may now study how to add these features directly to a language. We will call this
language F[OB, F[ with objects. Our encodings of objects were operationally correct,
but not adding syntactic support for objects makes them too dicult to work with
in practice. The colorPoint encoding, for example, is quite dicult to read. This
readability problem only gets worse when types are introduced into the language.
F[OB includes most of the features we discussed above: classes, message send, meth-
ods, elds, super, and this. We support information hiding in the same manner in which
Smalltalk does: all instance variables are hidden (protected) and all methods are exposed
(public).
CHAPTER 5. OBJECT-ORIENTED LANGUAGE FEATURES 93
F[OB also supports primitive objects. Primitive objects are objects that are
dened “inline,” that is, objects that are not created from a class. They are a more
lightweight form of object, and are similar to Smalltalk’s blocks and Java’s anonymous
classes. Primitive objects aren’t very common in practice, but we include them in F[OB
because it requires very little work. The value returned by the expression new aClass is
an object, which means that an object is a rst class expression. As long as our concrete
syntax allows us to directly dene primitive objects, no additional work is needed in the
interpreter.
An interesting consequence of having primitive objects is that we could get rid of
functions entirely (not methods). Functions could simply be encoded as methods of
primitive objects. For this reason, object-oriented languages that support primitive
objects have many of the advantages of higher-order functions.
5.2.1 Concrete Syntax
Let’s introduce the F[OB concrete syntax with an example. The following code is an
implementation of the pointClass / colorPointClass hierarchy from before (Figure
5.3). For simplicity this encoding does not include class constructor arguments.
Let pointClass =
Class Extends EmptyClass
Inst
x = 0;
y = 0
Meth
magnitude = Function -> sqrt(x + y);
setx = Function newx -> x := newx;
sety = Function newy -> y := newy
In Let colorPointClass =
Class Extends pointClass
Inst
x = 0;
y = 0;
(* A use of a primitive object: *)
color = Object
Inst
Meth red = 0; green = 0; blue = 0
Meth
magnitude =
Function ->
mult(Super <- magnitude {})(This <- brightness {})
(* An unnormalized brightness metric *)
brightness = Function ->
color <- red + color <- green + color <- blue;
setx = Super <- setx (* explicitly inherit *)
CHAPTER 5. OBJECT-ORIENTED LANGUAGE FEATURES 94
sety = ...; setcolor = ...
In Let point = New pointClass
In Let colorPoint = New colorPointClass In
(* Some sample expressions *)
point <- setx 4; point <- sety 7;
point <- magnitude{};
colorpoint <- magnitude {}
There is a lot going on with this syntax, so let’s take some time to point out some
of the major elements. First of all, notice that This and Super are special “reserved
variables.” In our F[ encodings, we had to write “Function this -> ” and pass this as
an explicit parameter. Now, self-awareness happens implicitly, and is This is a reference
to self.
Note the use of a primitive object to dene the color eld of the colorPointClass.
The red, green, and blue values are implemented as methods, since elds are always
“private.”
In our previous encodings we used one style when dening the base class, and another
dierent style when dening the subclass. In F[OB we use the same syntax for both by
always specifying the superclass. To allow for base classes that do not inherit from any
other class, we allow a class to extend EmptyClass, which is simply a special class that
does not dene anything.
F[OB instance variables use the l/r-value form of state that was discussed in Section
4.1.3. There is no need to explicitly use the ! operator to get the value of an instance
variable. F[OB instance variables are therefore mutable, not following the Caml conven-
tion that all variables are immutable. Note that method arguments are still immutable.
There are thus two varieties of variable: immutable method parameters, and mutable
instances. Since it is clear which variables are instances and which are not, there is no
great potential for confusion. It is the job of the parser to distinguish between instance
variables and regular variables. An advantage of this approach is that it keeps instance
variables from being directly manipulated by outsiders.
Method bodies are generally functions, but need not be; they can be any immutable,
publicly available value. For example, immutable instances can be considered methods
(see the color primitive object in the example above).
Note that we still have to explicitly inherit methods. This is not the cleanest syntax,
but it simplies the interpreter, and facilitates the translation to F[SR discussed in
Section 5.2.3.
Also, there is no constructor function. new is used to create new instances, and,
following Smalltalk, initialization is done explicitly by writing an initialize method.
For simplicity, F[OB does not support static elds and methods, nor does it take
the “classes as objects” view discussed in the previous section.
5.2.2 A Direct Interpreter
We rst consider a direct interpreter for F[OB. The abstract syntax can be expressed
by the following Caml type.
CHAPTER 5. OBJECT-ORIENTED LANGUAGE FEATURES 95
type ide = Ide of string | This | Super
type label = Lab of string
type expr = (* the F[ expressions, including Let *)
(* Object holds the instance list and the method list *)
| Object of ((label * expr) list) * ((label * expr) list)
| Class of expr * ((label * expr) list) * ((label * expr) list)
| EmptyClass
| New of expr
| Send of expr * label
(* parser has to decide if a var. is InstVar or just a Var *)
| InstVar of label
| InstSet of label * expr
Here is a rough sketch of the interpreter. This interpreter is not complete, but it gives a
general idea of what one should look like.
(* Substitute "sinst" for Super in all method bodies *)
let rec subst_super sinst meth =
match meth with
[] -> []
| (l, body)::rest ->
(l, subst(body, InstVar(sinst), Super))::
(subst_super sinst rest)
let rec eval e =
match e with
...
| Object(inst, meth) -> Object(eval_insts inst, meth)
| Send(term1, label) ->
(match (eval term1) with
Object(inst, meth) ->
subst(selectMeth(meth, label),
Object(inst, meth), This)
_ -> raise TypeMismatch)
| Class(super, inst, meth) -> Class(super, inst, meth)
| New(Class(super, inst, meth)) ->
(match super with
EmptyClass -> eval (Object(inst, meth))
| s -> let sobj = eval(New super) in
let sinst = (* A fresh instance variable label *) in
let newinst = (sinst, sobj)::inst in
let newmeth = subst_super sinst meth in
eval (Object(newinst, newmeth))
...
CHAPTER 5. OBJECT-ORIENTED LANGUAGE FEATURES 96
and eval_insts inst =
match inst with
[] -> []
| (l, body)::rest -> (l, eval(body))::(eval_insts rest)
This code sketch for the interpreter does a nice job of illustrating the roles of This
and Super. We only substitute for this when we send a message. That’s because This is
a dynamic construct, and we don’t know what it will be until runtime (see the discussion
of dynamic dispatching in Section 5.1.6). On the other hand, Super is a static construct,
and is known at the time we write the code, which allows us to substitute for Super as
soon as the Class expression is evaluated.
5.2.3 Translating F[OB to F[SR
Another way to give meaning to F[OB programs is by dening a translation mapping
F[OB programs into F[SR programs. This is a complete characterization of how objects
can be encoded in F[SR, the topic of Section 5.1.
In Chapter 8 below, we devleop a compiler for F[SR. To obtain a compiler for F[OB,
we can simply add a translation step that translates F[OB to F[SR, and then simply
compile the F[SR using this compiler. Thus, we will have a F[OB compiler “for free”
when we are all done. This section only discusses the F[OB to F[SR translation.
The translation is simply a formalization of the encodings given in Section 5.1. Al-
though real object-oriented compilers are much more sophisticated, this section should
at least provide an understandable view of what an object-oriented compiler does. The
concrete syntax translation is inductively dened below in a piecewise manner.
toFbSR(Object Inst x1 =e1 ; . . . ; xn =en Meth m1 =e′1 ; . . . ; mk =e′k ) =
{inst = {x1 =Ref(toFbSR(e1 )); . . . ; xn =Ref(toFbSR(en ))};
meth = {m1 = Function this -> toFbSR(e′1 ); . . . ;
mk = Function this -> toFbSR(e′k )} }
toFbSR(Class Extends e Inst x1 =e1 ; . . . ; xn =en
Meth m1 =e′1 ; . . . ; mk =e′k ) =
Function -> Let super = (toFbSR(e)) {} In
{inst = {x1 =Ref(toFbSR(e1 )); . . . ; xn =Ref(toFbSR(en ))};
meth = {m1 = Function this -> toFbSR(e′1 ); . . . ;
mk = Function this -> toFbSR(e′k )} }
toFbSR(New e) = (toFbSR(e)) {}
toFbSR(EmptyClass) = Function -> {}
toFbSR(Super <- m args) = super.meth.m this args
toFbSR(e <- m args) =
Let ob = toFbSR(e) In ob.meth.m ob args, for e not Super
toFbSR(x := e) = this.inst.x := toFbSR(e)
toFbSR(x) = !(this.inst.x), for x an instance variable
toFbSR(y) = y, for y a function variable
toFbSR(anything else) = homomorphic
CHAPTER 5. OBJECT-ORIENTED LANGUAGE FEATURES 97
The translation is fairly clean, except that messages to Super have to be handled a bit
dierently than other messages in order to properly implement dynamic dispatch. Notice
again that instance variables are handled dierently than function variables, because they
are mutable. Empty function application, i.e. f (), may be written as empty record
application: f {}. The “ ” variable in Function -> e is any variable not occurring in
e. The character “ ” itself is a valid variable identier in the FbDK implementation of
F[SR, however (see Appendix ??).
As an example of how this translation works, let us perform it on the F[OB version
of the point / colorPoint classes from Section 5.2.1. The result is the following:
Let pointClass =
Function -> Let super = (Function -> {}) {} In {
inst = {
x = Ref 3;
y = Ref 4
};
meth = {
magnitude = Function this -> Function ->
sqrt ((!(this.inst.x)) + (!(this.inst.y)));
setx = Function this -> Function newx ->
(this.inst.x) := newx;
sety = Function this -> Function newy ->
(this.inst.y) := newy
}
}
In Let colorPointClass =
Function -> Let super = pointClass {} In {
inst = {
x = Ref 3;
y = Ref 4;
color = Ref ({inst = {}; meth = {
red = Function this -> 45;
green = Function this -> 20;
blue = Function this -> 20
}})
};
meth = {
magnitude = Function this -> Function ->
mult ((super.meth.magnitude) this {})
((this.meth.brightness) this {});
brightness = Function this -> Function ->
(((!(this.inst.color)).meth.red) this) +
(((!(this.inst.color)).meth.green) this) +
(((!(this.inst.color)).meth.blue) this);
setx = Function this -> Function newy ->
CHAPTER 5. OBJECT-ORIENTED LANGUAGE FEATURES 98
(super.meth.setx) this newy;
sety = Function this -> Function newy ->
(super.meth.setx) this newy;
setcolor = Function this -> Function c ->
(this.inst.color) := c
}
} In
(* Let colorPoint = New colorPointClass In
* colorPoint <- magnitude {}
*)
Let colorPoint = colorPointClass {} In
(colorPoint.meth.magnitude) colorPoint {};;
Interact with F[SR. The translated code above should run ne in F[SR, pro-
vided you dene the mult and sqrt functions rst. mult is easily dened as
Let Rec mult x = Function y ->
If y = 0 Then
0
Else
x + (mult x (y - 1)) In ...
sqrt is not so easily dened. Since we’re more concerned with the behavior of the
objects, rather than numerical accuracy, just write a dummy sqrt function that returns
its argument:
Let sqrt = Function x -> x In ...
Now, try running it with the F[SR le-based interpreter. Our dummy sqrt function
returns 7 for the point version of magnitude, and the colorPoint magnitude multiplies
that result by the sum of the brightness (85 in this case). The result is
$ FbSR fbobFBsr.fbsr
==> 595
After writing a Caml version of toFbSR for the abstract syntax, a F[OB compiler is
trivially obtained by combining toFbSR with the functions dened in Chapter 8:
let FbOBcompile e = toC(hoist(atrans(clconv(toFbSR e))))
Finally, there are several other ways to handle these kinds of encodings. More infor-
mation about encoding objects can be found in [10].
Chapter 6
Type Systems
In F[, if we evaluate the expression
3 + (If False Then 3 Else False),
we will get some kind of interpreter-specic error at runtime. If F[ had a type system,
such an expression would not have been allowed to evaluate.
In Lisp, if we dene a function
(defun f (x) (+ x 1)),
and then call (f "abc"), the result is a runtime type error. Similarly, the Smalltalk
expression
String new myMessage
results in a “message not supported” exception when run. Both of these runtime errors
could have been detected before runtime if the languages supported static type systems.
The C++ code
int a[10]; a[123] = 5;
executes unknown and potentially harmful eects, but the equivalent Java code will throw
an ArrayIndexOutOfBoundsException at runtime, because array access is checked by a
dynamic type system.
These are just a few examples of the kinds of problems that type systems are designed
to address. In this chapter we discuss such type systems, as well as algorithms to infer
and to check types.
99
CHAPTER 6. TYPE SYSTEMS 100
6.1 An Overview of Types
A type is simply a property with which a program is implicitly or explicitly annotated
before runtime. Type declarations are invariants that hold for all executions of a pro-
gram, and can be expressed as statements such as “this variable always holds a String
object,” or “this function always returns a tree expression.”
Types have many other advantages besides simply cutting down on runtime errors.
Since types (and module signatures) specify invariant properties of the program, they
serve as precise and descriptive comments on the functionality of the code. In this
manner, types aid in large software development.
With a typed language, more information is known at compile-time, and this helps
the compiler produce much faster code. For example, the record implementation we will
use in our F[SR compiler (via hashing, see Chapter 8) is not needed in C++ due to its
static type system. We know the size of all records at compile-time, and therefore know
where they should be laid out in memory. In contrast, Smalltalk is very slow, and the
lack of a static type system counts for much of the slowness. The Strongtalk system [8]
attempts to bring a static type system into Smalltalk.
Finally, in an untyped language its easier to do very ugly “hacking”. For instance,
a list [1;true;2;false;3;true] can be written in an untyped language, but this is dan-
gerous and it would be much preferred to represent this as [(1,true);(2,false);(3,true)],
which, in OCaml, has the type (int * bool) list.
However, untyped languages have one distinct advantage over typed ones: they are
more expressive. For example, consider the F[SR encoding of F[OB from Chapter
5. This encoding would not work with OCaml as the target language, because the
OCaml type system disallows record polymorphism. The Y -combinator is another ex-
ample of where an untyped language really shines. F[ could support recursion via the
Y -combinator, but a simple typed version of F[ can’t since it cannot be typed. Recall
from Section 2.3.5 that OCaml can not type the Y -combinator, either.
All of us have seen types used in many ways. The following is a list of some of the
more common dimensions of types.
• Atomic types: int, float, . . .
• Type constructors, which produce types from types: ’a -> ’b, ’a * ’b
• OCaml-style type constructor denitions via type
• C-style type denitions via struct and typedef
• Object-oriented types: class types, object types
• Module types, or signatures
• Java-style interfaces
• Exception types: 〈method〉 throws 〈exception〉
• Parametric polymorphism: ’a -> ’a
CHAPTER 6. TYPE SYSTEMS 101
• Record/Object polymorphism: pass a ColorPoint to a function which expects a
Point.
There are also several newer dimensions of types that are currently active research areas.
• Eect types: the type “int -x,y-> int” indicates that variables x and y will
assigned to in this function. Java’s throws clauses for methods are a form of eect
types.
• Concrete class analysis: for variable x:Point, a concrete class analysis produces
a set such as {Point, ColorPoint, DataPoint}. This means at runtime x could
either be a Point, a ColorPoint, or a DataPoint (and, nothing else). This is useful
in optimization.
• Typed Assembly Language [3, 17]: put types on assembly-level code and have a
type system that guarantees no unsafe pointer operations.
• Logical assertions in types: int -> { x:int | odd(x) } for a function returning
odd numbers.
There is an important distinction that needs to be made between static and dynamic
type systems. Static type systems are what we usually mean when we talk about type
systems. A static type system is the standard notion of type found in C, C++, Java and
OCaml. Types are checked by the compiler, and type-unsafe programs fail to compile.
Dynamic type systems, on the other hand, check type information at runtime.
Lisp, Scheme, and Smalltalk are, in fact, dynamically typed. In fact, F[ and F[SR are
technically dynamically typed as well, since they will raise a typeMismatch when the
type of an expression is not what it expects. Any time you use a function, the runtime
environment makes sure that its a function. If you use an integer, it makes sure it’s an
integer, etc. These runtime type checks add a lot of overhead to the runtime environment,
and thus cause programs to run slowly.
There is some dynamic typechecking that occurs in statically typed languages too.
For instance, in Java, downcasts are veried at run-time and can raise exceptions. Out-
of-bounds array accesses are also checked at run-time in Java and OCaml, and are thus
dynamically typed. Array accesses are not typed in C or C++, since no check is performed
at all. Note that the type of an array (i.e. int, float) is statically checked, but the size
is dynamically checked.
Finally, languages can be untyped. The FbSR compiler produces untyped code,
since runtime errors cause core dumps. It is important to understand the distinction
between an untyped language and a dynamically typed one. In an untyped language
there is no check at all and anomalous behavior can result at runtime. In Chapter 8,
we compile F[SR to untyped C code, using casts to “disable” type system. Machine
language is another example of an untyped language.
To really see the dierence between untyped and dynamically typed languages, con-
sider the following two program fragments. The rst is C++ code with the type system
“disabled” via casts.
CHAPTER 6. TYPE SYSTEMS 102
#include <iostream>
class Calculation {
public: virtual int f(int x) { return x; }
};
class Person {
public: virtual char *getName() { return "Mike"; }
};
int main(int argc, char **argv) {
void *o = new Calculation();
cout << ((Person *)o)->getName() << endl;
return 0;
}
The code compiles with no errors, but when we run it the output is “ã¿.” But if
we compile it with optimization, the result is “Ãà.” Run it on a dierent computer and
it may result in a segmentation fault. Use a dierent compiler, and the results may be
completely exotic and unpredictable. The point is that because we’re working with an
untyped language, there are no dynamic checks in place, and meaningless code like this
results in undened behavior.
Contrast this behavior with that of the equivalent piece of Java code. Recall that
Java’s dynamic type system checks the type of the object when a cast is made. We will
use approximately the same code:
class Calculation {
public int f(int x) { return x; }
}
class Person {
public String getName() { return "Mike"; }
}
class Main {
public static void main(String[] args) {
Object o = new Calculation();
System.out.println(((Person) o).getName());
}
}
When we run this Java code, the behavior is quite predictable.
Exception in thread "main" java.lang.ClassCastException: Calculation
at Main.main(example.java:12)
CHAPTER 6. TYPE SYSTEMS 103
The ClassCastException is an exception raised by Java’s dynamic type system. The
unsafe code is never executed in Java, and so the behavior of the program is consistent
and well-dened. With the C++ version, there is no dynamic check, and the unsafe code
is executed, resulting in wild and unexpected behavior.
6.2 TF[: A Typed F[ Variation
We will use the prex “T” to represent a typed version of a language which we have pre-
viously studied. Thus we have several possible languages: TF[, TF[R, TF[S, TF[SR,
TF[OB, TF[X, TF[SRX, etc. There are too many languages to consider individually,
so we will rst look at TF[ as a warm-up, and then consider full-blown TF[SRX.
6.2.1 Design Issues
Before we begin to investigate TF[ or any typed language, there are a few general
design issues to address. We will take a moment to discuss these issues before giving the
specication for TF[.
The rst question to ask is how much explicit type information must our language
contain? How much type information must the program be decorated with, and how
much can be inferred by the compiler? A spectrum of possibilities exists.
On one end of the spectrum, we can use no decoration at all. We simply stick to our
untyped language syntax, and let the compiler infer all the type information.
Alternatively, we can use limited decoration, and have the compiler do partial infer-
ence. There is a wide range of possibilities. C for instance requires function argument
and return types to be specied, and declared variables to be given types. However,
within the body of a function, individual expressions do not need to be typed as those
types may be inferred. Some languages only require function argument types be declared,
and the return values of functions is then inferred.
At the other end of the spectrum, every subexpression and its identier must be
decorated with its type. This is too extreme, however, and makes the language unusable.
Instead of writing
Function x -> x + 1,
we would need some gross syntax like
(Function x -> (x:int + 1:int):int):(int -> int).
For TF[ and TF[SRX, we will concentrate on the C and Pascal view of explicit
type information. We specify function argument and return types, and declared variable
types, and allow the rest to be inferred.
We should also illustrate the dierence between type checking and type inference.
In any typed language, the compiler should typecheck the program before generating
CHAPTER 6. TYPE SYSTEMS 104
code. Type inference algorithms infer types and check that the program body has
no type errors. OCaml is an example of this.
A type checker generally just checks the body is well-typed given the types listed
on declarations. This is how C, C++, and Java work, although technically they are also
inferring some types, such as the type of 3+4.
In any case it must be possible to run the type inference or type checking algorithm
quickly. OCaml in theory can take exponential time to infer types, but in practice it is
linear.
6.2.2 The TF[ Language
Finally, we are ready to look at TF[, a typed F[ language. To simplify things, we will
not include the Let Rec syntax of F[, but only non-recursive, anonymous functions. We
will discuss typing recursion in Section 6.4.
In analogue with our development of operational semantics and interpreters, we will
dene two things when discussing types. First we dene type systems, which are
language-independent notations for assigning types to programs. Type systems are anal-
ogous to operational semantics. Secondly, we dene type checkers, which are OCaml
implementations of type systems analogous to interpreters. We begin with type systems.
Type Systems
Type systems are rule-based formal systems that are similar to operational semantics.
Type systems rigorously and formally specify what program have what types, and have
a strong and deep parallel with formal logic (recall our discussion of Russell’s Paradox
in Section 2.3.5). Type systems are generally a set of rules about type assertions.
Denition 6.1. (Type Environment) A type environment, Γ, is a set {x1 : τ1 , . . . , xn :
τn } of bindings of free variables types. If a variable x is listed twice in Γ, the rightmost
(innermost) binding is the proper type. We write Γ(x) = τ to indicate that τ is the
innermost type for x in Γ.
Denition 6.2. (Type Assertion) A type assertion, Γ ` e : τ , indicates that in type
environment Γ, e is of type τ .
The TF[ types in the concrete syntax are
τ ::= Int | Bool | τ -> τ.
We can represent this in a OCaml abstract syntax as
type fbtype = Int | Bool | Arrow of fbtype * fbtype
The expressions of TF[ are almost identical to those of F[, except that we must
explicitly decorate functions with type information about the argument. For example,
in the concrete syntax we write
CHAPTER 6. TYPE SYSTEMS 105
Function x:τ -> e
where the abstract syntax representation is
Function of ide * fbtype * expr
The TF[ Type Rules
We are now ready to dene the TF[ types rules. These rules have the same structure
as the operational semantics rules we have looked at before; the horizontal line reads
“implies.” The following three rules are the axioms for out type system (recall that an
axiom is a rule that is always true, that is, a rule with nothing above the line).
(Hypothesis)
Γ ` x : τ for Γ(x) = τ
(Int)
Γ ` n : Int for n an integer
(Bool)
Γ ` b : Bool for b True or False
The Hypothesis rule simply says that if a variable x contained in a type environment
Γ has type τ then the assertion Γ ` x : τ is true. The Int and Bool rules simply give
types to literal expressions such as 7 and False. These rules make up the base cases of
our type system.
Next, we have rules for simple expressions.
Γ ` e : Int, Γ ` e′ : Int
(+)
Γ ` e + e′ : Int
Γ ` e : Int, Γ ` e′ : Int
(-)
Γ ` e - e′ : Int
Γ ` e : Int, Γ ` e′ : Int
(=)
Γ ` e = e′ : Bool
These rules are fairly straightforward. For addition and subtraction, the operands
must typecheck to Ints, and the result of the expression is an Int. Equality is similar,
but typechecks to a Bool. Note that equality will only typecheck with integer operands,
not boolean ones. The And, Or, and Not rules are similar, and their denition should be
obvious.
CHAPTER 6. TYPE SYSTEMS 106
The If rule is a bit more complicated. Clearly, the conditional part of the expression
must typecheck to a Bool. But what about the Then and Else clauses? Consider the
following expression.
If e Then 3 Else False
Should this expression typecheck? If e evaluates to True, then the result is 3, an Int.
If e is False, the result is False, a Bool. Clearly, then this expression should not
typecheck, because it does not always evaluate to the same type. This tells us that for
an If statement to typecheck, both clauses must typecheck to the same type, and the
type of the entire expression is then the same as the two clauses. The rule is as follows.
Γ ` e : Bool, Γ ` e′ : τ, Γ ` e′′ : τ,
(If)
Γ ` If e Then e′ Else e′′ : τ
We have now covered all the important rules except for functions and application.
The Function rule is a bit dierent from other rules, because functions introduce new
variables. The type of the function body depends on the type of the variable itself, and
will not typecheck unless that variable is in Γ, the type environment. To represent this
in our rule, we need to perform the type assertion with the function’s variable appended
to the type environment. We do this in the following way.
Γ, x : τ ` e : τ ′
(Function)
Γ ` (Function x:τ -> e) : τ -> τ ′
Notice the use of the type constructor -> to represent the type of the entire function
expression. This type constructor should be familiar, as the OCaml type system uses the
same notation. In addition, the Function rule includes the addition of an assumption
to Γ. We assume the function argument, x, is of type τ , and add this assumption to the
environment Γ to derive the type of e. The Application rule follows from the Function
rule:
Γ ` e : τ -> τ ′ , Γ ` e′ : τ
(Application)
Γ ` e e′ : τ ′
Just as in operational semantics, a derivation of Γ ` e : τ is a tree of rule applications
where the leaves are axioms (Hypothesis, Int or Bool rules) and the root is Γ ` e : τ .
Let’s try an example derivation.
CHAPTER 6. TYPE SYSTEMS 107
` (Function x:Int -> (Function y:Bool ->
If y Then x Else x+1)) :
Int -> Bool -> Int
Because by the function rule, it suces to prove
x : Int ` (Function y: Bool -> (If y Then x Else x+1)) :
Bool->Int
Because by the function rule again, it suces to prove
x : Int, y : Bool ` If y Then x Else x+1 : Int
Because by the If rule, it suces to prove
x : Int, y : Bool ` y : Bool
x : Int, y : Bool ` x : Int
x : Int, y : Bool ` x+1 : Int
all of which either follow by the Hypothesis rule or + and
Hypothesis.
Given the above and letting
f = (Function x:Int -> (Function y:Bool ->
If y Then x Else x+1))
we then have
` f 5 True : Int
Because by the application rule,
` f : Int -> Bool -> Int
(which we derived above)
` 5 : Int by the Int rule
And thus
` f 5 : Bool -> Int by the Application rule.
Given this and
` True : Bool by the Bool rule
we can get
` f 5 True : Int by the Application rule.
As we mentioned before, TF[ is a very weak language. No recursive functions can be
dened. In fact, all programs are guaranteed to halt. TF[ is thus normalizing. Without
recursion, TF[ is not very useful. Later we will add recursion to TF[SRX and show
how to type it.
Exercise 6.1. Try to type the Y -combinator in TF[.
Now that we have a type system for the TF[ language, we can detect whether or not
our programs are well-typed. But what does this mean? The answer comes in the form
of the following type soundness theorem.
CHAPTER 6. TYPE SYSTEMS 108
Theorem 6.1. If ` e : τ , then in the process of evaluating e, a “stuck state” is never
reached.
We will not precisely dene the concept of a stuck state. It is basically a point at
which evaluation can not continue, such as 0 (Function x -> x) or (Function x ->
x) + 4. In terms of a F[ interpreter, stuck stated are the cases that raise exceptions.
This theorem asserts that a type system prevents runtime errors from occurring. Similar
theorems are the goal of most type systems.
6.3 Type Checking
To write an interpreter for TF[, we simply modify the F[ interpreter to ignore type
information at runtime. What we are really interested in is writing a type checker.
This is the type system equivalent of an interpreter; given the language-independent type
rules, dene a type checking algorithm in a particular language, namely OCaml.
A type-checking algorithm typeCheck takes as input a type environment Γ and an
expression e. It either returns the type τ of e or it raises an exception indicating that e
is not well-typed in the environment Γ.
Some type systems do not have an easy corresponding type-checking algorithm. In
TF[ we are fortunate in that the type checker mirrors the type rules in a nearly direct
fashion. As was the case with the interpreters, the outermost structure of the expression
dictates the rule that applies. The ow of the recursion is that we pass the environment
Γ and expression e down, and return the result type τ back up.
Here is a rst pass at the TF[ typechecker, typecheck : envt -> expr -> fbtype.
Γ can be implemented as a (ident * fbtype) list, with the most recent item at the front
of the list.
let rec typecheck gamma e =
match e with
(* lookup returns the first mapping of x in gamma *)
Var x -> lookup gamma x
| Function(Ide x,t,e1) ->
let t’ = typecheck (((Ide x),t)::gamma) e1 in
Arrow(t,t’)
| Appl(e1,e2) ->
let Arrow(t1,t2) = typecheck gamma e1 in
if typecheck gamma e2 = t1 then
t2
else
raise TypeError
| Plus(e1,e2) ->
if typecheck gamma e1 = Int and
typecheck gamma e2 = Int then
Int
else
raise TypeError
CHAPTER 6. TYPE SYSTEMS 109
| (* ... *)
Lemma 6.1. typecheck faithfully implements the TF[ type system. That is,
• ` e : τ if and only if typecheck [] e returns τ , and
• typecheck [] e raises a typeError exception if and only if ` e : τ is not provable
for any τ .
Proof. Omitted (by case analysis).
This Lemma implies the typecheck function is a sound implementation of the type
system for TF[.
6.4 Types for an Advanced Language: TF[SRX
Now that we’ve looked at a simple type system and type checker, let us move on to a
type system for a more complicated language: TF[SRX. We include just about every
piece of syntax we have used up to now, except for F[OB’s classes and objects and F[V’s
variants. Here is the abstract syntax dened in terms of a OCaml type.
type exnid = string
and expr =
Var of ident
| Function of ident * fbtype * expr
| Letrec of ident * ident * fbtype * expr * fbtype * expr
| Appl of expr * expr
| Plus of expr * expr | Minus of expr * expr
| Equal of expr * expr | And of expr * expr
| Or of expr * expr | Not of expr
| If of expr * expr * expr | Int of int | Bool of bool
| Ref of expr | Set of expr * expr | Get of expr
| Cell of int | Record of (label * expr) list
| Select of label * expr | Raise of expr * fbtype |
| Try of expr * exnid * ident * expr
| Exn of exnid * expr
and fbtype =
Int | Bool | Arrow of fbtype * fbtype
| Rec of label * fbtype list | Rf of fbtype
Next, we will dene the type rules for TF[SRX. All of the TF[ type rules apply, and
so we can move directly to the more interesting rules. Let’s begin by tackling recursion.
What we’re really typing is the In clause, but we need to ensure that the rest of the
expression is also well-typed.
Γ, f : τ -> τ ′ , x : τ ` e : τ ′ , Γ, f : τ -> τ ′ ` e′ : τ ′′
(Let Rec)
Γ ` (Let Rec f x:τ = e:τ ′ In e′ ) : τ ′′
CHAPTER 6. TYPE SYSTEMS 110
Next, we move on to records and projection. The type of a record is simply a map of
the eld names to the types of the values associated with each eld. Projection is typed
as the type of the value of the eld projected. The rules are
Γ ` e 1 : τ1 , . . . , Γ ` e n : τn
(Record)
Γ ` {l1 = e1 ; . . . ; ln = en } : {l1 : τ1 ; . . . ; ln : τn }
Γ ` e : {l1 : τ1 ; . . . ; ln : τn }
(Projection)
Γ ` e.li : τi for 1 ≤ i ≤ n
We’ll also need to be able to type side eects. We can type Ref expressions with the
special type τ Ref. Set and Get expressions easily follow.
Γ`e:τ
(Ref)
Γ ` Ref e : τ Ref
Γ ` e : τ Ref, Γ ` e′ : τ
(Set)
Γ ` e := e′ : τ
Γ ` e : τ Ref
(Get)
Γ ` !e : τ
Finally, the other kind of side eects we need to type are exceptions. To type an
exception itself, we will simply use the type Exn. This allows all exceptions to be typed
in the same way. For example, the code
If b Then (#IntExn 1) Else (#BoolExn False)
will typecheck, and has type Exn. We do this to allow maximum exibility.
Because they alter the ow of evaluation, Raise expressions should always typecheck,
provided the argument typechecks to an Exn type. It is dicult to know what type to
give a raise expression, though. Consider the following example.
If b Then Raise (#Exn True) Else 4
This expression should typecheck to type Int. From our If rule, however, we know that
the Then and Else clause must have the same type. We infer, therefore, that the Raise
expression must have type Int for the If to typecheck. In Section 6.6.2 we see how to
handle this inference automatically. For now, we will simply type Raise expressions with
the arbitrary type τ . Note that this is a perfectly valid thing for a type rule to do, but
it is dicult to implement in an actual typechecker.
CHAPTER 6. TYPE SYSTEMS 111
Next, notice that the With clause of the Try expression is very much like a function.
Just as we did with functions, we will need to decorate the identier with type information
as well. However, as we see below, this decoration can be combined with the nal kind
of type decoration, which we will discuss now.
Consider the expression
Try Raise (#Ex 5)
With #Ex x:Int -> x + 1
The type of this expression is clearly Int. But suppose the example were modied a bit.
Try Raise (#Ex False)
With #Ex x:Int -> x + 1
This expression will also type to Int. But suppose we were to evaluate the expression
using our operational semantics for exceptions. When the exception is raised, False will
be substituted for x, which could cause a runtime type error. The problem is that our
operational semantics is ignorant of the type of the exception argument.
We can solve this problem without changing the operational semantics, however.
Suppose, instead of writing #Ex False, we wrote #Ex@Bool False. The @Bool would be
used by the type rules to verify that the argument is indeed a Bool, and the interpreter
will simply see the string “#Ex@Bool”, which will be used to match with the With clause.
This also eliminates the need for type decoration on the With clause identier, since it
serves the same purpose. In a sense, this is very much like overloading a method in Java
or C++. When a method is overloaded, the type and the method name are needed to
uniquely identify the correct method. Our nal exception syntax looks like this:
Try Raise (#Ex@Bool False)
With #Ex@Int x -> x + 1
This expression typechecks and has type Int. When evaluated, the result is Raise
#Ex@Bool False, i.e. the exception is not caught by the With clause. This is the behavior
we want.
Now that we’ve got a type-friendly syntax worked out, let’s move on to the actual
type rules. They are fairly straightforward.
Γ ` e : τ′
(Raise)
Γ ` (Raise #xn@τ ′ e) : τ for arbitrary τ
Γ ` e : τ, Γ, x : τ ′ ` e′ : τ
(Try)
Γ ` (Try e With #xn@τ ′ x -> e′ ) : τ
Using these rules, let’s type the expression from above:
CHAPTER 6. TYPE SYSTEMS 112
` (Try Raise (#Ex@Bool False) With #Ex@Int x -> x + 1) : Int
Because by the Raise rule,
` Raise (#Ex@Bool False) : τ for arbitrary τ
Because by the Bool rule, ` False : Bool
And, by the + rule
x : Int ` x + 1 : Int
By application of the Int and Hypothesis rules
Therefore, by the Try rule, we deduce the type Int for the original expression.
Exercise 6.2. Why are there no type rules for cells?
Exercise 6.3. How else could we support recursive functions in TF[SRX without using
Let Rec, but still requiring that recursive functions properly typecheck? Prove that your
solution typechecks.
Exercise 6.4. Attempt to type some of the untyped programs we have studied up to
now, for example, the Y -combinator, Let, sequencing abbreviations, a recursive factorial
function, and the encoding of lists. Are there any that can not typecheck at all?
Exercise 6.5. Give an example of a non-recursive F[SR expression that evaluates prop-
erly to a value, but does not typecheck when written in TF[SRX.
6.5 Subtyping
The type systems that we covered above are reasonably adequate, but there are still
many types of programs that have no runtime errors that will nonetheless not typecheck.
The rst extension to our standard type systems that we would like to consider is what
is known as subtyping. The main strength of subtyping is that it allows record and
object polymorphism to typecheck.
Subtypes should already be a familiar concept from Java and C++. Subclasses are
subtypes, and extending or implementing an interface gives a subtype.
6.5.1 Motivation
Let us motivate subtypes with an example. Consider a function
Function x:{l:Int} -> (x.l + 1):Int.
This function takes as an argument a record with eld l of type Int. In the untyped
F[R language the record passed into the function could also include other elds besides
l, and the call
(Function x -> x.l + 1) {l = 4; m = 6}
CHAPTER 6. TYPE SYSTEMS 113
would generate no run-time errors. However, this would not type-check by our TF[SRX
rules: the function argument type is dierent from the type of the value passed in.
The solution is to re-consider record types such as {m:Int; n:Int} to mean a record
with at least the m and n elds of type Int, but possibly other elds as well, of unknown
type. Think about the previous record operations and their types: under this interpre-
tation of record typing, the Record and Projection rules both still make sense. The old
rules are still sound, but we need a new rule to reect this new understanding of record
types:
Γ ` e : {l1 : τ1 ; . . . ; ln : τn }
(Sub-Record0 )
Γ ` e : {l1 : τ1 ; . . . ; lm : τm } for m < n
This rule is valid, but it’s not as good as we could do. To see why, consider another
example,
F = Function f -> f ({x=5; y=6; z=3}) + f({x=6; y=4}).
Here the function f should, informally, take a record with at least x and y elds, but
also should accept records where additional elds are present. Let us try to type the
function F .
F : ({x:Int; y:Int} -> Int) -> Int
Consider the application F G for
G = Function r -> r.x + r.x.
If we were to typecheck G, we would end up with G : {x:Int} -> Int, which does
not exactly match F ’s argument, {x:Int; y:Int} -> Int, and so typechecking F G
will fail even though it does not cause a runtime error.
In fact we could have given G a type {x:Int; y:Int} -> Int, but its too late to
know that was the type we should have used back when we typed G . The Sub-Rec0 rule
is of no help here either. What we need is a rule that says that a function with a record
type argument may have elds added to its record argument type, as those elds will be
ignored:
Γ ` e : {l1 : τ1 ; . . . ; ln : τn } -> τ
(Sub-Function0 )
Γ ` e : {l1 : τ1 ; . . . ; ln : τn ; . . . ; lm : τm } -> τ
CHAPTER 6. TYPE SYSTEMS 114
Using this rule, F G will indeed typecheck. The problem is that we still need other rules.
Consider records inside of records:
{pt = {x=4; y=5}; clr = 0} : {pt:{x:Int}; clr:Int}
should still be a valid typing since the y eld will be ignored. However, there is no type
rule allowing this typing either.
6.5.2 The STF[R Type System: TF[ with Records and Subtyping
By now it should be clear that the strategy we were trying to use above can never
work. We would need a dierent type rule for every possible combination of records and
functions!
The solution is to have a separate set of subtyping rules just to determine when
one type can be used in the place of another. τ <: τ ′ is read “τ is a subtype of τ ′ ,” and
means that an object of type τ may also be considered an object of type τ ′ . The rule
added to the TF[ type system (along with the record rules of TF[SRX) is
Γ ` e : τ, ` τ <: τ ′
(Sub)
Γ ` e : τ′
We also need to make our subtyping operator reexive and transitive. This can be
accomplished with the following two rules.
(Sub-Re)
` τ <: τ
` τ <: τ ′ , ` τ ′ <: τ ′′
(Sub-Trans)
` τ <: τ ′′
Our rule for subtyping records needs to do two things. It needs to ensure that if a
record B is the same as record A with some additional elds, then B is a subtype of A.
It also needs to handle the case of records within records. If B’s elds are all subtypes
of A’s elds, then B should also be a subtype of A. We can reect this concisely in a
single rule as follows.
(Sub-Record)
` τ1 <: τ1′ , . . . , τn <: τn′
` {l1 : τ1 ; . . . ; ln : τn ; . . . ; lm : τm } <: {l1 : τ1′ ; . . . ; ln : τn′ }
The function rule must also do two things. If functions A and B are equivalent except
that B returns a subtype of what A returns, then B is a subtype of A. However, if A
CHAPTER 6. TYPE SYSTEMS 115
and B are the same except that B’s argument is a subtype of A’s argument, then A is a
subtype of B. Simply put, for a function to be a subtype of another function, it has to
take less and give more. The rule should make this clear:
τ0′ <: τ0 , τ1 <: τ1′
(Sub-Function)
τ0 -> τ1 <: τ0′ -> τ1′
From our discussions and examples in the previous section, it should be clear that
this more general set of rules will work.
6.5.3 Implementing an STF[R Type Checker
Automated typechecking of STF[R is actually quite dicult. There are two ways to
make the task easier. The rst is to add more explicit type decoration to help the
typechecker. The second it to completely infer the types in a constraint form, a topic
covered in Section 6.7.
Here, we briey sketch how the typecheck function for STF[R may be written. The
TF[ typechecker requires certain types to be identical, for example, the function domain
type must be identical to the type of the function argument in an application e e′ .
Γ ` e : τ -> τ ′ , Γ ` e′ : τ
(Application)
Γ ` e e′ : τ ′
In STF[R at this point, we need to see if subtyping is possible. typecheck(e′ ) returns
τ ′′ and then τ ′′ <: τ is checked via a function areSubtypes(τ ′′ , τ ). This produces a valid
proof by the Sub rule. Other rules where the TF[ rules require a type match similarly
are generalized to allow the Sub rule to be used.
Exercise 6.6. Implement the areSubtypes function.
6.5.4 Subtyping in Other Languages
It is interesting to see how subtyping is used in other languages. Consider, for example,
Java and C++. In these languages, subclassing is the main form of subtyping. A subclass
is a subtype of the class from which it extends. In Java, a class is also a subtype of any
interfaces it implements.
Java and C++ are thus more restrictive. Suppose there were classes with structure
{x:Int; y:Int; color:Int} and {x:Int; y:Int} that weren’t one of the two cases
above (that is, the prior is not a subclass of the latter, and the two share no common
interface). The two classes, then, are not subtypes in Java or C++, but they are in
STF[R. Declared subtyping has the advantage, however, that subtype relationships do
not have to be completely inferred.
OCaml objects are more exible in that there is no restriction to a hierarchy, but it’s
also less exible in that there is really no object polymorphism— an explicit coercion of
a ColorPoint to a Point is required. There are several research languages with type
inference and subtyping, but the types are often complex or hard to read [9].
CHAPTER 6. TYPE SYSTEMS 116
6.6 Type Inference and Polymorphism
Type inference was originally discovered by Robin Milner, the original creator of ML,
and independently by the logician J. Roger Hindley. The key idea of Milner’s “Algorithm
W” is to initially give all variables arbitrary types, ’a, and then to unify or equate the
types, if indicated by the program. For example, if the application f x has type ’a ->
’b and x has type ’c, we may equate ’a and ’c.
We will look at full type inference, that is, type inference on programs with no explicit
type information.
6.6.1 Type Inference and Polymorphism
Type inference goes hand-in-hand with parametric polymorphism (often called “generic
types”). Consider the function Function x -> x. Without polymorphism, what type
can be inferred for this function? Int -> Int is a awed answer, because the function
could be used in a context where it is passed a boolean. With type inference we need to
achieve something called a principal type.
Denition 6.3. (Principal Type) A principal type τ for expression e (where ` e : τ )
has the following property. For any other type τ ′ such that ` e : τ ′ , for any context C
for which ` C[e : τ ′ ] : τ ′′ for any τ ′′ , then ` C[e : τ ] : τ ′′′ as well, for some τ ′′′ .
What principality means is no other typing will let more uses of the program type-
check, so the principal typing will always be best. The desired property of our type
inference algorithm is that it will always infer principal types. An important corollary is
that with a principal-type algorithm, we know inference will never “get in the way” of
the programmer by, for example, inferring Bool -> Bool for the identity function when
the programmer wants to use it on Int -> Int.
OCaml infers the type ’a -> ’a for the identity function, which can be shown to be
a principal type for it. In fact, OCaml type inference always infers principal types.
6.6.2 An Equational Type System: EF[
We are going to present type inference in a nonstandard way. Milner’s “Algorithm W”
eagerly unies ’a and ’c: it replaces one with the other everywhere. We will present
equational inference, in which we lazily accumulate equations like ’a = ’c, and then
solve the system of equations at the end of the algorithm.
We will study EF[, a simple, equationally typed version of F[. EF[ uses the same
grammar for expressions as the F[ language did, since EF[ does not have any type
decorations. The EF[ types are
τ ::= Int | Bool | τ -> τ | α types
α ::= ’a | ’b | . . . type variables
EF[ types during inference are going to include an extra set of constraining equations,
E, which constrain the behavior of the type variables. Type judgments for EF[ are thus
of the form Γ ` e : τ \E, the same as before but tacking a set of equations on the side.
Each member of E is an equation like ’a = ’c. Equational types will be used to aid
inference. Here is an outline of the overall approach.
CHAPTER 6. TYPE SYSTEMS 117
1. Infer equational types for the whole programs.
2. If the equations are inconsistent, pronounce that there is a type error.
3. If the equations are consistent, simplify them to give an inferred type.
It is a fact that if there are no inconsistencies in the equations, they can always be
simplied to give an equation-free type.
Denition 6.4. (Equational Type) An equational type is a type of the form
τ \{τ1 = τ1′ , . . . , τn = τn′ }
Each τ = τ ′ is an equation on types, meaning τ and τ ′ have the same meaning as
types. We will let E mean some arbitrary set of type equations. For instance,
Int -> ’a\{’a = Int -> ’a1, ’a1 = Bool}
is an equational type. If you think about it, this is really the same as the type
Int -> Int -> Bool.
This is known as equation simplication and is a step we will perform in our type inference
algorithm. It is also possible to write meaningless types such as
Int -> ’a\{’a = Int -> ’b, ’a = Bool}
which cannot be a type since it implies that functions and booleans are the same type.
Such equation sets are deemed inconsistent, and will be equated with failure of the type
inference process. There are also possibilities for circular (self-referential) types that
don’t quite look inconsistent:
Int -> ’a\{’a = Int -> ’a }
OCaml disallows such types, and we will also disallow them initially. These types can’t
be simplied away, and that is the main reason why OCaml disallows them: users of the
language would have to see some type equations.
CHAPTER 6. TYPE SYSTEMS 118
The EF[ Type Rules
The EF[ system is the following set of rules. Note that Γ serves the same role as it did in
the TF[ rules. It is a type environment that binds variables to simple (non-equational)
types. Our axiomatic rules look much like the TF[ rules.
(Hypothesis)
Γ ` x : τ \∅ for Γ(x) = τ
(Int)
Γ ` n : Int\∅ for n an integer
(Bool)
Γ ` b : Bool\∅ for b a boolean
The rules for +, -, and = also look similar to their TF[ counterparts, but now we must
take the union of the equations of each of the operands to be the set of equations for the
type of the whole expression, and add an equation to reect the type of the operands.
Γ ` e : τ \E, Γ ` e′ : τ ′ \E ′
(+)
Γ ` e + e′ : Int\E ∪ E ′ ∪ {τ = Int, τ ′ = Int}
Γ ` e : τ \E, Γ ` e′ : τ ′ \E ′
(-)
Γ ` e - e′ : Int\E ∪ E ′ ∪ {τ = Int, τ ′ = Int}
Γ ` e : τ \E, Γ ` e′ : τ ′ \E ′
(=)
Γ ` e = e′ : Bool\E ∪ E ′ ∪ {τ = Int, τ ′ = Int}
The And, Or, and Not rules are dened in a similar way. The rule for If is also similar
to the TF[ If rule. Notice, though, that we do not immediately infer a certain type like
we did in the previous rules. Instead, we infer a type α, and equate α to the types of the
Then and Else clauses.
(If)
Γ ` e : τ \E, Γ ` e′ : τ ′ \E ′ , Γ ` e′′ : τ ′′ \E ′′
Γ ` (If e Then e′ Else e′′ ) : α\E ∪ E ′ ∪ E ′′ ∪ {τ = Bool, τ ′ = τ ′′ = α}
Finally, we are ready for the function and application rules. Functions no longer
have explicit type information, but we may simply choose a fresh type variable ’a as the
function argument, and include it in the equations later. The application rule also picks
a fresh type variable ’a type, and adds an equation with ’a as the right hand side of a
function type. The rules should make this clear.
Γ, x : α ` e : τ \E
(Function)
Γ ` (Function x -> e) : α -> τ \E
Γ ` e : τ \E, Γ ` e′ : τ ′ \E ′
(Application)
Γ ` e e′ : α\E ∪ E ′ ∪ {τ = τ ′ -> α}
CHAPTER 6. TYPE SYSTEMS 119
These rules almost directly dene the equational type inference procedure: the proof
can pretty much be built from the bottom (leaves) on up. Each equation added denotes
two types that should be equal.
Solving the Equations
One thing that should be immediately clear from the EF[ type rules is that any syn-
tactically correct program may be typed by these rules. Whether or not a program
is well-typed is not determined until we actually solve the system of equations in the
equational type of the entire program. The EF[ type rules are really just accumulating
constraints.
Solving the equations involves two steps. First we compute the closure of the equa-
tions, producing new equations that hold by transitivity, etc. Next, we check for any
inconsistent equations, such as Int = Bool that denote type errors.
The following algorithm computes the equational closure of set E.
• For each equation of the form τ0 -> τ0′ = τ1 -> τ1′ in E, add τ0 = τ1 and τ0′ = τ1′
to E.
• For each set of equations τ0 = τ1 and τ1 = τ2 in E, add the equation τ0 = τ2 to E
(by transitivity).
• Repeat (1) and (2) until no more equations can be added to E.
Note that we will implicitly use the symmetric property on these equations, and so there
is no need to add τ1 = τ0 for every equation τ0 = τ1 .
The closure serves to uncover inconsistencies. For instance,
Closure({’a = Int -> ’b, ’a = Int -> Bool, ’b = Int}) =
{’a = Int -> ’b, ’a = Int -> Bool,
’b = Int, Int -> ’b = Int -> Bool,
Int = Int, ’b = Bool, Int = Bool},
directly uncovering the inconsistency Int = Bool.
The closure of E can be computed in polynomial time. After computing the closure,
the constraints are consistent if
1. No immediate inconsistencies are uncovered, such as Int = Bool, Bool = τ -> τ ′ ,
or Int = τ -> τ ′ .
2. No self-referential equations exits (we will deal with this issue shortly).
If the equations are consistent, the next step is to solve the equational constraints.
We do this by substituting type variables with actual types. The algorithm is as follows.
Given τ \E,
CHAPTER 6. TYPE SYSTEMS 120
1. Replace some type variable α in τ with τ ′ , provided α = τ ′ or τ ′ = α occurs in E
and either
• τ ′ is not a type variable, or
• τ ′ is a type variable α′ which lexographically succeeds α.
2. Repeat (1) until no more such replacements are possible.
Notice that step (1) considers the symmetric equivalent of each equation, which is why
we didn’t include them in the closure. The algorithm has a aw though: the replacements
may continue forever. This happens when E contains a circular type. Recall the example
of a self-referential type
Int -> ’a\{’a = Int -> ’a}.
Trying to solve these constraints results in the nonterminating chain
Int -> Int -> ’a\{’a = Int -> ’a}
Int -> Int -> Int -> ’a\{’a = Int -> ’a}
Int -> Int -> Int -> Int -> ’a\{’a = Int -> ’a}
...
The solution is to check for such cycles before trying to solve the equations. The
best way to do this it to phrase the problem in graph-theoretical context. Specically,
we dene a directed graph G in which the nodes are the the type variables in E. There
is a directed edge from ’a to ’b if ’a = τ is an equation in E and ’b occurs in τ .
We raise a typeError if there is a cycle in G for which there is at least one edge
representing a constraint that isn’t just between type variables (’a = ’b).
In summary, our entire EF[ type inference algorithm is as follows. For expression e,
1. Produce a proof of ` e : τ \E by applying the EF[ type rules. Such a proof always
exists.
2. Extend E by computing its closure.
3. Check if E is immediately inconsistent. If so, raise a typeError.
4. Check for cycles in E using the algorithm described above. If there is a cycle, raise
a typeError.
5. Solve E by the above equation solution algorithm. This algorithm will always
terminate if there are no cycles in E.
6. Output: the solution type τ ′ for e produced by the solution algorithm.
Theorem 6.2. The typings produced by the above algorithm are always principal.
CHAPTER 6. TYPE SYSTEMS 121
The proof of this theorem is beyond the scope of this book.
Let’s conclude with an example of the type inference algorithm in action. Suppose
we want to infer the type of the expression
(Function x -> If x Then 3 Else 4) False
First we produce the following proof.
By the Application rule,
` ((Function x -> If x Then 3 Else 4) False) :
’c\{’a = Bool, Int = ’b, ’a -> ’b = Bool -> ’c}
Because, by the Function rule,
(Function x -> If x Then 3 Else 4) :
’a -> ’b\{’a = Bool, Int = ’b}
Because, by the If rule,
x : ’a ` If x Then 3 Else 4 : ’b\{’a = Bool, Int = ’b}
Because by the Int and Hypothesis rules,
x : ’a ` x : ’a\∅,
x : ’a ` 3 : Int\∅, and
x : ’a ` 4 : Int\∅
And, by the Bool rule,
` False : Bool\∅
Given the proof of
` ((Function x -> If x Then 3 Else 4) False) :
’c\{’a = Bool, Int = ’b, ’a -> ’b = Bool -> ’c}
we compute the closure of the set of equations to be
{’a = Bool, Int = ’b, ’a -> ’b = Bool -> ’c, ’b = ’c, Int = ’c}
The set is not immediately inconsistent, and does not contain any cycles. Therefore, we
solve the equations. In this case, τ contains only ’c, and we can replace ’c with Int.
We output Int as the type of the expression, which is clearly correct.
6.6.3 PEF[: EF[ with Let Polymorphism
After all our work on EF[, we still don’t have polymorphism, we only have type variables.
To illustrate this, consider the function
Let x = Function y -> y In (x True); (x 0)
CHAPTER 6. TYPE SYSTEMS 122
Recalling our encoding of Let as a function, this expression is equivalent to
(Function x -> (x True); (x 0)) (Function y -> y)
In OCaml, such programs typecheck ne. Dierent uses of Function y -> y can
have dierent types. Consider what EF[ would do when typing this expression, though.
` (Function x -> (x True); (x 0)) :
’a -> ’c\{’a = Bool -> ’b, ’a = Int -> ’c, . . .}
But when we compute the closure of this equational type, we get the equation Int =
Bool! What went wrong? The problem in this case is that each use of x in the body
used the same type variable ’a. In fact, when we type Function y -> y, we know that
’a can be anything, so for dierent uses, ’a can be dierent things. We need to build
this intuition into our type system to correctly handle cases like this. We dene such a
type system in PEF[, which is EF[ with Let and Let-polymorphism.
PEF[ has a special Let typing rule, in which we allow a new kind of type in Γ:
∀α1 . . . αn . τ . This is called a type schema, and may only appear in Γ. An example of
a type schema is ∀α. α -> α. Note that the type variables α1 . . . αn are considered to
be bound by this type expression.
The new rule for Let is
Γ ` e : τ \E, Γ, x : ∀α1 . . . αn . τ ′ ` e′ : τ ′′ \E ′
(Let)
Γ ` (Let x = e In e′ ) : τ ′′ \E ′
where τ ′ is a solution of ` e : τ \E using the above algorithm, and τ ′ has free type
variables α1 . . . αn that do not occur in Γ.
Notice that since we are invoking the simplication algorithm in this rule, it means
the full algorithm is not the clean 3-pass infer-closure-simplify form give above: the rules
need to call close-simplify on some sub-derivations.
We also need to add an axiom to ensure that a fresh type variable is given to each
Let usage. The rule is
(Let Inst.)
Γ, x : ∀α1 . . . αn . τ ′ ` x : R(τ ′ )\∅
where R(τ ′ ) is a renaming of the variables α1 . . . αn to fresh names. Since these names
are fresh each time x is used, the dierent uses won’t conict like above.
It will help to see an example of this type system in action. Let’s type the example
program from above:
CHAPTER 6. TYPE SYSTEMS 123
Let x = Function y -> y In (x True); (x 0)
We have
` Function y -> y : ’a -> ’a\∅
This constraint set trivially has the solution type ’a -> ’a. Thus, we then typecheck
the Let body under the assumption that x has type ∀’a.’a -> ’a.
x : ∀’a.’a -> ’a ` x : ’b -> ’b\∅
by the Let-Inst rule. Then
x : ∀’a.’a -> ’a ` x True : ’c\{’b -> ’b = Bool -> ’c}
Similarly,
x : ∀’a.’a -> ’a ` x 0 : ’e\{’d -> ’d = Int -> ’e}
The important point here is that this use of x gets a dierent type variable, ’d, by
the Let-Inst rule. Putting the two together, the type is something like
x : ∀’a.’a -> ’a ` x True; x 0 :
’e\{’b -> ’b = Bool -> ’c, ’d -> ’d = Int -> ’e}
which by the Let rule then produces
` (Let x = Function y -> y In (Function x -> x True; x 0)) :
’e\{’b -> ’b = Bool -> ’c, ’d -> ’d = Int -> ’e}
Since ’b and ’d are dierent variables, we don’t get the conict we got previously.
CHAPTER 6. TYPE SYSTEMS 124
6.7 Constrained Type Inference
There was a reason why we presented Hindley-Milner type inference in the form above: if
we replace equality constraints by subtyping constraints, <:, we can perform constrained
type inference. To understand why it is useful to perform this generalization, it is easiest
to just look at the rules.
F[ is not the best system to show o the power of replacing equality with subtyping.
Since the language does not have records, there is not any interesting subtyping that
could happen. To show the usefulness of subtyping, we thus dene the constraints in
an environment where we have records, F[R. F[R plus constraints is CF[R. We can
contrast CF[R with the EF[R language which we did not study but is simply EF[ with
support for records. Instead of types τ \E for a set of equations E, CF[R has types
τ \{τ1 <: τ1′ , . . . , τn <: τn′ }
CF[R has the following set of type rules. These are direct generalizations of the EF[
rules, replacing = by <:. The <: is always in the direction of information ow. We let
C represent a set of subtyping constraints.
(Hypothesis)
Γ ` x : τ \∅ for Γ(x) = τ
(Int)
Γ ` n : Int\∅ for n an integer
(Bool)
Γ ` b : Bool\∅ for b a boolean
Γ ` e : τ \C, Γ ` e′ : τ ′ \C ′
(+)
Γ ` e + e′ : Int\C ∪ C ′ ∪ {τ <: Int, τ ′ <: Int}
Γ ` e : τ \C, Γ ` e′ : τ ′ \C ′
(-)
Γ ` e - e′ : Int\C ∪ C ′ ∪ {τ <: Int, τ ′ <: Int}
Γ ` e : τ \C, Γ ` e′ : τ ′ \C ′
(=)
Γ ` e = e′ : Bool\C ∪ C ′ ∪ {τ <: Int, τ ′ <: Int}
(If)
Γ ` e : τ \C, Γ ` e′ : τ ′ \C ′ , Γ ` e′′ : τ ′′ \C ′′ ,
Γ ` (If e Then e′ Else e′′ ) : α\C ∪ C ′ ∪ C ′′ ∪ {τ <: Bool, τ ′ <: α, τ ′′ <: α}
Γ, x : α ` e : τ \C
(Function)
Γ ` (Function x -> e) : α -> τ \C
Γ ` e : τ \C, Γ ` e′ : τ ′ \C ′
(Application)
Γ ` e e′ : α\C ∪ C ′ ∪ {τ <: τ ′ -> α}
CHAPTER 6. TYPE SYSTEMS 125
The two rules we have not seen in EF[ are the Record and Projection rules. There is
nothing particularly special about these rules, however.
Γ ` e1 : τ1 \C1 , . . . , Γ ` en : τn \Cn
(Record)
Γ ` {l1 =e1 ; . . . ; ln =en } : {l1 : τ1 ; . . . ; ln : τn }\C1 ∪ . . . ∪ Cn
Γ ` e : τ \C
(Projection)
Γ ` e.l : α\{τ <: {l : α}} ∪ C
As with EF[, these rules almost directly dene the type inference procedure and the
proof can pretty much be built from the bottom up.
The complete type inference algorithm is as follows. Given an expression e,
1. Produce a proof of ` e : τ \C using the above type rules. Such a proof always
exists.
2. Extend C by computing the closure as described below.
3. If C is immediately inconsistent, raise a typeError.
4. Check C for cycles as described below. If C contains a cycle, raise a typeError.
5. The inferred type is e : τ \C.
The algorithms for computing the closure of C and doing cycle detection are fairly
obvious generalizations of the EF[ algorithms. Closure is computed as follows.
1. For each constraint {l1 : τ1 , . . . , ln : τn , . . . , lm : τm } <: {l1 : τ1′ , . . . , ln : τn′ } in C,
add τ1 <: τ1′ , . . . , τn <: τn′ to C.
2. For each constraint τ0 -> τ0′ <: τ1 -> τ1′ in C, add τ1 <: τ0 and τ0′ <: τ1′ to C.
3. For constraints τ0 <: τ1 and τ1 <: τ2 , add τ0 <: τ2 to C (by transitivity).
4. Repeat until no more constraints can be added.
A constraint set is immediately inconsistent if τ <: τ ′ and τ and τ ′ are dierent kinds
of type (function and record, Int and function, etc), or two records are ordered by <:
and the right record has a eld the left record does not.
To perform cycle detection in C, we use the following algorithm. Dene a directed
graph G where nodes are type variables in C. There is an edge from a ’a node to a ’b
node if there is an equation ’a <: τ ′ in C, and ’b occurs in τ ′ . Additionally, there is an
edge from ’b to ’a if τ ′ <: ’a occurs in C and ’b occurs in τ ′ . C has a cycle if and only
if G has a cycle.
It seems that there is a major omission in our constrained type inference algorithm: we
never solve the constraints! The algorithm is correct, however. The reason we don’t want
to solve the constraints is that any substitution proceeds with possible loss of generality.
CHAPTER 6. TYPE SYSTEMS 126
Consider, for example, a constraint ’a <: τ , and the possibility of substituting ’a with τ .
This precludes the possibility that the ’a position be a subtype of τ , as the substitution in
eect asserts the equality of ’a and τ . In simpler terms, we need to keep the constraints
around as part of the type. This is the main weakness of constrained type systems; the
types include the constraints, and are therefore dicult to read and understand.
We have the same shortcomings as in the equational case at this point: there is as
of yet no polymorphism. The solution used in the equational case won’t work here, as it
required the constraints to be solved.
The solution is to create constrained polymorphic types
∀α1 , . . . , αn . τ \C
in the assumptions Γ, in place of the polymorphic types (type schema) we had in the
equational version. The details of this process are quite involved, and we will not go into
them. Constrained polymorphic types make very good object types, since polymorphism
is needed to type inheritance.
Chapter 7
Concurrency
A concurrent computation is working on more than one thing at once. We assume some
familiarity with concurrency but provide a brief overview to get us going. The Wikipedia
article on parallel computing provides a more detailed overview.
7.1 Overview
Concurrent execution can be loosely grouped into three implementation categories, in
order of loosest to tightest coupling. Distributed computation is computation on mul-
tiple computers which share no memory and are sending messages between each other
to communicate data. There is still a wide range of how tightly these computers can
be decoupled. Grid computing is distributed computing where the Internet is the com-
munication medium. Cluster computing is over a fast LAN network. Massive Parallel
Processing (MPP) involves specialized communication hardware for very high commu-
nication bandwidth. Distributed shared memory is the case where dierent processes
are running on dierent processors but sharing some special memory via a memory
bus. Lastly, multithreaded computation is the case where multiple threads of execution
share a single memory which is local. Multithreaded computations may run on a single
(core) CPU, meaning the concurrent execution is an illusion achieved by interleaving
the execution steps of two threads, or if the computer has multiple cores there can be
true concurrent execution of a multithreaded program. The Threads Wikipedia article
claries how threads and processes dier.
All of the above models still support independent foci of control. That is the primary
focus of our study in this chapter – it captures a wide range of models and they are
the most elegant forms of concurrent architecture. There are several other models that
we are not addressing. Vector processors are computers that can do an operation on a
whole array in one step. These architectures used to be called SIMD (Single Instruction
Multiple Data). Stream processors are the modern version of vector processors which
can work on more than just arrays in parallel, for example sparse arrays. The GPGPU
(General Purpose Graphics Processing Units) is a recent version of a stream processor
which arose as a generalization of specialized graphics processors. Lastly, FPGA’s are
Field Programmable Circuits: you can create your own (parallel) logic circuitry on the
y.
127
CHAPTER 7. CONCURRENCY 128
Historically, concurrent programming arose as library extensions to existing lan-
guages. Early models included UNIX sockets for IPC, and process forking in C. While
some concurrency is easily programmable via library extensions to an underlying sequen-
tial language, some aspects of concurrency are fundamental enough that it is important
to integreate concurrency into the programming language.
The standard model these days for Java, C, etc concurrency is via multithreading.
Multithreaded programming is coming on in a big way – newer CPUs have multiple cores
and can run multiple threads simultaneously. Mulithreaded programming unfortunately
is also a disaster waiting to happen – the programs are just too hard to debug. The
root of the problem is there are too many possible cases of interleaving of the dierent
threads that can occur when accessing the shared memory. Some of them may show up
in testing, but many of them will only show up after deployment.
The primary “Bad Things” that can occur in multithreaded programming include the
following. A race condition is when two operations are interleaved and leave the data in
an inconsistent state. For example, a simultaneous variable update where one thread set
the high word of a double and another thread set the low word due to near- simultaneous
access – the double’s value is not a sensible value from either thread. Deadlock is when
threads may need to wait for resources to free up (they were locked to begin with to
prevent race conditions), and can get in a cycle of waiting: A waits for B waits for C
waits for A.
Race conditions may be avoided via the use of locks of various kinds. Monitors
are regions of mutual exclusion in the source program, only one thread can execute a
monitor block at any one point. They are similar to the synchronized keyword of
Java. A Semaphore is a low-level locking mechanism: grab a lock before a critical
operation; block if someone else has lock; once you have the lock you know you are the
only one accessing; free it when you are done. General semaphores allow n threads to
simultaneously access a critical region, not just one.
Atomicity is another key design concept of concurrent programming languages. An
atomic region is a region of code that may have been interleaved with other thread
executions, but you can’t tell – it always appears to have run atomically , in one step.
The more atomicity you can get in your language design the fewer interleavings need
to be considered in debugging and the fewer bugs. The Sun Java Concurrency Tutorial
provides more details on the model.
7.1.1 The Java Concurrency Model
Let us briey review the Java approach to concurrency. Java implementats the standard
notion of thread: memory is shared between concurrent threads of control (i.e., each
thread has its own runtime stack but only one heap). The synchronized keyword
is used to declare zones of mutual exclusion; they are a form of monitor. There are
several variants: synchronized methods or blocks of code can both be dened. When
a synchronized method/block is running, no other synchronized method/block for that
object can start from another thread - any such threads would have to wait until the
currently running method/block nishes. An advantage of Java’s synchronized is how
it is object-based : locks are per-object and this is arguably a good level of abstraction
for locking. The disadvanatge of Java however is monitors give only mutual exclusion,
CHAPTER 7. CONCURRENCY 129
not atomicity.
The Java concurrency model has some other nice features which we briey review
here. Much of this is found in the java.util.concurrent package.
• Atomic integers, oats, etc - there will never be any race conditions on setting or
getting the value from AtomicInteger etc.
• Locks - java.util.concurrent.locks.Lock
• Concurrent collections - e.g. ConcurrentHashMap which supports concurrent ad-
d/lookup atomically.
7.2 The Actor Model and AF[V
We will study one model of concurrent programming in more detail here, the Actor
Model. Actors are a simple, elegant model of concurrent programming. It is to some
degree the “functional programming” analogy in the concurrent programming world, and
Actors in fact t very well with functional programming style. The Erlang programming
language [1] is a real-world language with exactly that structure: a functional basis with
an actor concurrency layar on top.
The actor model was originally developed by Karl Hewitt at MIT in the 1970’s. The
Actor model Wikipedia entry has a good historical overview and description, and we
now briey summarize the key points. Each actor is an autonomous, distributed agent,
and has a name that is not forgeable. All messaging is asynchronous – sending actors
never wait for a reply from the receiving actor. The arrival order of the messages is
nondeterministic, they may arrive in a dierent order they were sent in. Actors can be
thought of as a non-shared-memory model with respect to our earlier classication: all
communication between actors is via explicit messaging, not by implicit communication
through a shared variable in a common memory. Along with their local computation
actions and sending of messages, actors can create other actors. If an actor is busy when
a message arrives is it put in a message queue – only one message is processed at a time
by an actor. Concurrency is achieved by the fact that many actors can be processing
their messages in parallel, not by concurrent behavior of an individual actor. In our
idealized model we assume there are no faults: all messages eventually arrive (but, they
may take arbitrarily long to do so). Each actor is a reactive system: normally the actor
is idle, is wakes up and process a single message in a nite amount of time, and goes
back to sleep until another message comes in. This is the life of the actor, always and
forever reacting to events from the outside.
One of the biggest advantages of the actor model is the built-in Atomicity: multiple
actors can be running in parallel, but any interleaved run is equivalent to a run where
each actor runs all of its steps in one big step. That is because actors can never accept
messages in the middle of processing their original message.
In this section we dene AF[V, an actor layer on top of the F[V language. We need
the “V” because we will use variants to dene messages. Recall F[Vvariants are like
OCaml’s inferred variants - ‘foo(4) is the variant foo with argument 4. Notice how
we can also view this as the message foo with argument 4. F[Vvariants always have
exactly one argument only, for simplicity.
CHAPTER 7. CONCURRENCY 130
7.2.1 Syntax of AF[V
The AF[V expressions are the following.
v ::= . . . the F[V values . . . | a values
e ::= . . . the F[V expressions . . . | e e | Create(e, e) | a expressions
where a are taken from an unbounded set of actor names. They are like the cells c of
F[S in that they cannnot appear in source programs but can show up at runtime, and
there are innitely many unique ones; they are just names (nonces).
AF[V syntax e e′ indicates a message send; it expects e to evaluate to an actor
name, and sends that actor the message which is the value of e′ . Create(e, e′ ) creates
an actor with behavior e, and with initial local data e′ . e should evaluate to a function
and that function is the (whole) code for the actor. The Create returns the (new) name
of this new actor as its result.
Some features of AF[V include the following. The code of an actor is nothing but one
function – there is no state within an actor in the form of elds. At the end of processing
each message, the actor goes idle; the behavior it is going to have upon receiving the next
message is the value at the end of the previous message send. The latter point is the key
to how actors can mutate over time: each message processing is purely functional, but
at the end the actor gets to pick what state it wants to mutate to before processing the
next message. This is an interesting method for mixing a bit of imperative programming
into a pure functional model.
In AF[V you must write your own explicit message dispatch code using the Match
syntax of F[V. We are here taking the dual approach to encoding to objects compared
to the objects-as-records approach we used in F[OB, here taking objects as functions
and messages as variants. Lastly, in order to allow actors to know their own name, at
creation time it is passed to them.
7.2.2 An Example
Before getting into the operational semantics lets do a simple example. Here is an actor
that gets a start message and then counts down from its initial value to 0. Here, we
use the term Y to represent the Y-combinator and the term to represent a value of no
importance (as with the empty record {} before).
Function myaddr ->
Y (Function this -> Function localdata -> Function msg ->
Match msg With
‘main(n) -> myaddr <- ‘count(n); this(_)
| ‘count(n) -> If n = 0
Then this(_)
Else
myaddr <- ‘count(n-1);
this(_) /* set to respond to next message */
)
CHAPTER 7. CONCURRENCY 131
Here is a code fragment that another actor could use to re up a new actor with the
above behavior and get it started. Suppose the above code we abbreviated CountTenBeh.
Let x = Create(CountTenBeh,_) /* _ is the localdata - unused */
In x <- ‘main(10)
Here is an alternative way to count down, where the localdata eld holds the value,
and its not in the message.
Function myaddr ->
Y (Function this -> Function localdata -> Function msg ->
Match msg With
‘count(_) ->
If localdata = 0
Then _
Else
myaddr <- ‘count(_);
this(localdata - 1) /* set to respond to next msg */
)
Suppose the above code was abbreviated CountTenBeh2; using it is then
Let x = create(CountTenBeh2,10) /* 10 is the localdata */
In x <- ‘count(_)
The latter example is the correct way to give actors local data – in the former example
the counter value had to be forwarded along every message.
Here is another usage fragment for the rst example:
Let x = create(CountTenBeh,_)
In x <- ‘main(10); x <- ‘main(5)
In this case the actor x will in parallel and independently counting down from 10 .. 0
and 5 .. 0 - these counts may also interleave in random ways. For the second example
an analogue might be:
Let x = create(CountTenBeh2,10)
In x <- ‘count(_); x <- ‘count(_)
This does nothing but get one more count message queued up; since the actor sends a
new one out each time it gets one until 0, the eect will be to have a leftover message at
the end.
7.2.3 Operational Semantics of Actors
The operational semantics for actors has two layers: the local computation within an
actor, which is not too dierent than F[V, and the concurrent global stepping of all the
actors. Lets start with the latter.
CHAPTER 7. CONCURRENCY 132
A global state G is a “soup” of the active actors and sent messages:
{〈a, v〉 | a is an actor name, v is its behavior}
G ::= ∪
{[a v] | a is an actor name, v is the message sent to a}
Actor systems can run forever, there is no notion of a nal value. So, the system does
small steps of computation:
G1 → G2 → G3 → . . .
– this indicates one step of computation; in each single step one actor in the soup
competely processed one message in the soup. We will dene this → relation below. →∗
is the reexive, transitive closure of → – many steps of actor computation. In general
this continues innitely since actor systems may not terminate. The nal meaning of a
run of an actor system is this innite stream of states.
7.2.4 The Local Rules
Lets start with the local rules. They are dened with a similar relation ⇒ in F[V
operational semantics, but the local executions additionally have side eects of the actors
they create and messages they send. We will make any such side eects be labels on this
S
arrow relation. So we make the local compuration relation be =⇒ – S here is the soup
of eects. S is in fact the same syntactic form as a G: it contains two kinds of elements,
[a v] indicating a send to a of message v that this local actor performed over its
execution, and 〈a, v〉 indicating a new actor it created, named a, with behavior (body)
v.
S
Now let us look at the operational semantics rules for =⇒. Most of the rules are very
minor changes to the corresponding F[V rule; we just give the + rule to show how the
existing F[V rules must be tweaked to add the potential for side eects S:
S S′
e1 =⇒ v1 , e2 =⇒ v2 where v1 , v2 ∈ Z
(+ Rule)
S∪S ′
e1 + e2 =⇒ the integer sum of v1 and v2
Since e or e′ above could in theory have each created actors or sent messages, we
need to append their eects to the nal result. These eects are like state, they are on
the side. A major dierence with F[S is the eects here are ”write only” – they don’t
change the direction of local computation in any way, they are only spit out. In that
sense local actor computation stays functional.
Here is the send rule:
S S′
e1 =⇒ a, e2 =⇒ v
(Send Rule)
S∪S ′ ∪{[av]}
e1 e2 =⇒ v
CHAPTER 7. CONCURRENCY 133
The main consequence is the message [a v] is added to the soup. (The return result
v here is largely irrelevant, the goal of a message send is to add the side eect.)
Lastly, here is the create rule:
S S′ S ′′
e1 =⇒ v1 , e2 =⇒ v2 , v1 a v2 =⇒ v3
(Create Rule)
S∪S ′ ∪S ′′ ∪{〈a,v3 〉}
Create(e1 , e2 ) =⇒ a, for a a fresh actor name
This time the return result matters - it is the name of the new actor. The running
of v1 a v2 passes the actor its own name and its initial values to initialize it, and so v1
will need to be a curried function of two arguments to accept these parameters a and v2
(in fact it needs to be a curried function of three arguments, because later the message
will also be passed as a parameter; more on that very soon).
7.2.5 The Global Rule
Last but not least, here is the global single-step rule for one actor in the soup processing
in its entirety one message:
S
(G ∪ {[a v ′ ]} ∪ {〈a, v〉}) → (G ∪ {〈a, v ′′ 〉} ∪ S) if (v v ′ =⇒ v ′′ )
This is the only global rule. It matches an actor with a message in the global soup that
is destined for it, uses the local semantics to run that actor (in isolation), and throws
back into the soup all of the S, which contains all the messages sent by this one actor run
as well as any new actors created by this one actor run. A global actor run is just the
repeated application of this rule. Notice how the actor behavior which was v is changed
to v ′′ , the result of this run.
To test these rules you can run the example programs above.
7.2.6 The Atomicity of Actors
The above semantics has actors executing atomically: each actor runs independently to
completion. However, it would be possible to make an alternative semantics in which
multiple actors run in parallel Since the actors are completely local in their executions,
the two semantics should be provably equivalent.
Chapter 8
Compilation by Program
Transformation
The goal of this chapter is to understand the core concepts behind compilation by writing
a F[SR compiler. Compilers are an important technology because code produced by a
compiler is faster than interpreted code by several orders of magnitude. At least 95% of
the production software running is compiled code. Compilation today is a very complex
process: compilers make multiple passes on a program to get source code to target code,
and perform many complex optimizing transformations. Our goals in this chapter are to
understand the most basic concepts behind compilation: how a high-level program can
be mapped to machine code.
We will outline a compiler of F[SR to a very limited subset of C (“pseudo-assembly”).
The reader should be able to implement this compiler in Caml by lling in the holes we
have left out. The compiler uses a series of program transformations to express the
compilation process. These program transformations map F[SR programs to equivalent
F[SR programs, removing high-level features one at a time. In particular the following
transformations are performed in turn on a F[SR program by our compiler:
1. Closure conversion
2. A-translation
3. Function hoisting
After a program has gone through these transformations, we have a F[SR program that
is getting close to the structure of machine language. The last step is then the translate
of this primitive F[SR program to C.
Real production compilers such as gcc and Sun’s javac do not use a transformation
process, primarily because the speed of the compilation itself is too slow. It is in fact
possible to produce very good code by transformation. The SML/NJ ML compiler uses a
transformational approach [5]. Also, most production compilers transform the program
to an intermediate form which is neither source nor target language (“intermediate lan-
guage”) and do numerous optimizing transformations on this intermediate code. Several
textbooks cover compiler technology in detail [6, 4].
134
CHAPTER 8. COMPILATION BY PROGRAM TRANSFORMATION 135
Our main goal, unlike a production compiler, is understanding: to appreciate the gap
between high- and low-level code, and how the gaps may be bridged. Each transformation
that we dene bridges one gap. Program transformations are interesting in their own
right, as they give insights into the F[SR language. Optimization, although a central
topic in compilation, is beyond the scope of this book. Our focus is on the compilation
of higher-order languages, not C/C++; some of the issues are the same but others are
dierent. Also, our executables will not try to catch run-time type errors or garbage
collect unused memory.
The desired soundness property for each F[SR program translation is: programs
before and after translation have the same execution behavior (in our case, termination
and same numerical output, but in general the same I/O behavior). Note that the
programs that are output by the translation are not necessarily operationally equivalent
to the originals.
The F[SR transformations are now covered in the order they are applied to the
source program.
8.1 Closure Conversion
Closure conversion is a transformation which eliminates nonlocal variables in functions.
For example, x in Function y -> x * y is a nonlocal variable: it is not the param-
eter and is used in the body. Via closure conversion, all such nonlocal variables can
be removed, obtaining an equivalent program where all variables used in functions are
parameters of the function. C and C++ have global variables (as does Java via static
elds), but global variables are not problematic nonlocal variables. The problematic
ones which must be removed are those that are parameters to other functions, where
function denitions have been nested. C and C++ have no such problematic nonlocal
variables since function denitions cannot be nested. In Java, inner classes are nested
class denitions, and there is also an issue of nonlocal variables which must be addressed
in Java compilation.
Consider for example the following curried addition function.
add = Function x -> Function y -> x + y
In the body x + y of the inner Function y, x is a nonlocal and y is a local variable for
that function.
Now, we ask the question, what should add 3 return? Let us consider some obvious
choices:
• Function y -> x + y wouldn’t make sense because the variable x would be un-
dened, we don’t know its value is 3.
• Function y -> 3 + y seems like the right thing, but it amounts to code substi-
tution, something a compiler can’t do since compiled code must be immutable.
Since neither of these ideas work, something new is needed. The solution is to return
a closure, a pair consisting of the function and an environment which remembers the
values of any nonlocal variables for later use:
CHAPTER 8. COMPILATION BY PROGRAM TRANSFORMATION 136
(Function y -> x + y, { x |-> 3 })
Function denitions are now closure denitions; to invoke such a function a new process
is needed. Closure conversion is a global program transformation that explicitly per-
forms this operation in the language itself. Function values are dened to be closures,
i.e. tuples of the function and an environment remembering the values of nonlocal vari-
ables. When invoking a function which is dened as a closure, we must explicitly pass
it the nonlocals environment which is in the closure so it can be used nd values of the
nonlocals.
The translation is introduced by way of example. Consider the inner Function y ->
x + y in add above translates to the closure
{ fn = Function yy -> (yy.envt.x) + (yy.arg);
envt = { x = xx.arg } };
Let us look at the details of the translation. Closures are dened as tuples in the form
of records
{ fn = Function ...; envt = {x = ...; ...}}
consisting of the original function (the fn eld) and the nonlocals environment (the envt
eld), which is itself a record. In the nonlocals environment { x = xx.arg }, x was a
nonlocal variable in the original function, and its value is remembered in this record using
a label of the same name, x. All such nonlocal variables are placed in the environment;
in this example x is the only nonlocal variable.
Functions that used to take an argument y are modied to take an argument named
yy (the original variable name, doubled up). We don’t really have to change the name
but it helps in understanding because the role of the variable has changed: the new
argument yy is expected to be a record of the form { envt = ..; arg = ..}, passing
both the environment and the original argument to the function.
If yy is indeed such a record at function invocation, then within the body we can use
yy.envt.x to access what was a nonlocal variable x in the original function body, and
yy.arg to access what was the argument y to the function.
The whole add function is closure-converted by converting both functions:
add’ = {
fn = Function xx -> {
fn = Function yy -> (yy.envt.x) + (yy.arg);
envt = { x = xx.arg }
};
envt = {}
}
CHAPTER 8. COMPILATION BY PROGRAM TRANSFORMATION 137
The outer Function x -> ... arguably didn’t need to be closure-converted since it had
no nonlocals, but for uniformity it is best to closure convert all functions.
Translation of function application In the above example, there were no applica-
tions, and so we didn’t dene how closure-converted functions are to be applied. Appli-
cation must change, because functions are now in fact represented as records. Function
call add 3 after closure conversion then must pass in the environment since the caller
needs to know it:
(add’.fn)({ envt = add’.envt; arg = 3})
So, we rst pull out the function part of the closure, (add’.fn), and then pass it a
record consisting of the environment add’.envt also pulled from the closure, and the
argument, 3. Translation of add 3 4 takes the result of the above, which should evaluate
to a function closure { fn = ...; envt = ...}, and does the same trick to apply 4 to
it:
Let add3’ = (add’.fn){ envt = add’.envt; arg = 3 } In
(add3’.fn){ envt = add3’.envt; arg = 4}
and the result would be 12, the same as the original result, conrming the soundness of
the translation in this case. In general applications are converted as follows. At function
call time, the remembered environment in the closure is passed to the function in the
closure. Thus, for the add’ 3 closure above, add3’, when it is applied later to e.g. 7,
the envt will know it is 3 that is to be added to 7.
One more level of nesting Closure conversion is even slightly more complicated if
we consider one more level of nesting of function denitions, for example
triadd = Function x -> Function y -> Function z -> x + y + z
The Function z needs to get x, and since that Function z is dened inside Function
y, Function y has to be an intermediary to pass from the outermost function x. Here
is the translation.
CHAPTER 8. COMPILATION BY PROGRAM TRANSFORMATION 138
triadd’ = {
fn = Function xx -> {
fn = Function yy -> {
fn = Function zz ->
(zz.envt.x) + (zz.envt.x) + (zz.arg);
envt = { x = yy.envt.x; y = yy.arg }
};
envt = { x = xx.arg }
};
envt = {}
}
Some observations can be made. The inner z function has nonlocals x and y so both
of them need to be in its environment; The y function doesn’t directly use nonlocals, but
it has nonlocal x because the function inside it, Function z, needs x. So its nonlocals
envt has x in it. Function z can get x into its environment from y’s environment, as
yy.envt.x. Thus, Function y serves as middleman to get x to Function z.
8.1.1 The Ocial Closure Conversion
With the previous example in mind, we can write out the ocial closure conversion trans-
lation. We will use the notation clconv(e) to express the closure conversion function,
dened inductively as follows (this code is informal; it uses concrete F[SR syntax which
in the case of e.g. records looks like Caml syntax).
Denition 8.1 (Closure Conversion).
1. clconv(x) = x (* variables *)
2. clconv(n) = n (* numbers *)
3. clconv(b) = b (* booleans *)
4. clconv(Function x -> e) = letting x, x1, ..., xn be precisely the free vari-
ables in e, the result is the F[SR expression
{ fn = Function xx -> SUB[clconv(e)];
envt = { x1 = x1; ...; xn = xn } }
where SUB[clconv(e)] is clconv(e) with substitutions (xx.envt.x1)/x1, . . . ,
(xx.envt.xn)/xn and (xx.arg)/x performed on it, but not substituting in Function’s
inside clconv(e) (stop substituting when you hit a Function).
5. clconv(e e’) = Let f = clconv(e) In (f.fn){ envt = f.envt; arg = clconv(e’)}
6. clconv(e op e’) = clconv(e) op clconv(e’) for all other operators in the lan-
guage (the translation is homomorphic in all of the other operators). This is pretty
clear in every case except maybe records which we will give just to be sure. . .
CHAPTER 8. COMPILATION BY PROGRAM TRANSFORMATION 139
7. clconv({ l1 = e1; ...; ln = en }) = { l1 = clconv(e1);...; ln = clconv(en)
}
For the above example, clconv(add) is add’. The desired soundness result is
Theorem 8.1. Expression e computes to a value if and only if clconv(e) computes
to a value. Additionally, if one returns numerical value n, the other returns the same
numerical value n.
Closure conversion produces programs where functions have no nonlocal variables,
and all functions thus could have been dened at the “top level” like in C. In fact, in
Section 8.3 below we will explicitly hoist all inner function denitions out to the top.
8.2 A-Translation
Machine language programs are linear sequences of atomic instructions; at most one
arithmetic operation is possible per instruction, so many instructions are needed to evalu-
ate complex arithmetic (and other) expressions. The A-translation closes the gap between
expression-based programs and linear, atomic instructions, by rephrasing expression-
based programs as a sequence of atomic operations. We represent this as a sequence of
Let statements, each of which performs one atomic operation.
The idea should be self-evident from the case of arithmetic expressions. Consider for
instance
4 + (2 * (3 + 2))
Our F[SR interpreter dened a tree-notion of evaluation order on such expressions. The
order in which evaluation happens on this program can be made explicitly linear by using
Let to factor out the parts in the order that the interpreter evaluates the program
Let v1 = 3 + 2 In
Let v2 = 2 * v1 In
Let v3 = 4 + v2 In
v3
This program should give the same result as the original since all we did was to make
the computation sequence more self-evident. Notice how similar this is to 3-address
machine code: it is a linear sequence of atomic operations directly applied to variables
or constants. The v1 etc variables are temporaries; in machine code they generally
end up being assigned to registers. These temporaries are not re-used (re-assigned to)
above. Register-like programming is not possible in F[SR but it is how real 3-address
intermediate language works. In the nal machine code generation temporaries are re-
used (via a register allocation strategy).
We are in fact going to use a more naive (but uniform) translation, that also rst
assigns constants and variables to other variables:
CHAPTER 8. COMPILATION BY PROGRAM TRANSFORMATION 140
Let v1 = 4 In
Let v2 = 2 In
Let v3 = 3 In
Let v4 = 2 In
Let v5 = v3 + v4 In
Let v6 = v2 * v5 In
Let v7 = v1 + v6 In
v7
This simple translation closely corresponds to the operational semantics—every node in
a derivation of e ⇒ v is a Let in the above.
Exercise 8.1. Write out the operational semantics derivation and compare its structure
to the above.
This translation has the advantage that every operation will be between variables.
In the previous example above, 4+v2 may not be low-level enough for some machine
languages since there may be no add immediate instruction. One simple optimization
would be to avoid making fresh variables for constants. Our emphasis at this point is
on correctness as opposed to eciency, however. Let is a primitive in F[SR—this is
not strictly necessary, but if Let were dened in terms of application, the A-translation
results would be harder to manipulate.
Next consider F[SR code that uses higher-order functions.
((Function x -> Function y -> y)(4))(2)
The function to which 2 is being applied rst needs to be computed. We can make this
explicit via Let as well:
Let v1 = (Function x -> Function y -> y)(4) In
Let v2 = v1(2) In
v2
The full A-translation will, as with the arithmetic example, do a full linearization of all
operations:
Let v1 =
(Function x ->
Let v1’ = (Function y -> Let v1’’ = y in v1’’) In v1’)
In
Let v2 = 4 In
Let v3 = v1 v2 In
Let v4 = 2 In
Let v5 = v3 v4 In
v5
CHAPTER 8. COMPILATION BY PROGRAM TRANSFORMATION 141
All forms of F[SR expression can be linearized in similar fashion, except If:
If (3 = x + 2) Then 3 Else 2 * x
can be transformed into something like
Let v1 = x + 2 In
Let v2 = (3 = v1) In
If v2 Then 3 Else Let v1 = 2 * x In v1
but the If still has a branch in it which cannot be linearized. Branches in machine code
can be linearized via labels and jumps, a form of expression lacking in F[SR. The above
transformed example is still “close enough” to machine code: we can implement it as
v1 := x + 2
v2 := 3 = v1
BRANCH v2, L2
L1: v3 := 3
GOTO L3
L2: v4 := 4
L3:
8.2.1 The Ocial A-Translation
We dene the A-translation as a Caml function, atrans(e) : term -> term. We will
always apply A-translation to the result of closure conversion, but that fact is irrelevant
for now.
The intermediate result of A-translation is a list of tuples
[(v1,e1); ...; (vn,en)] : (ide * term) list
which is intended to represent
Let v1 = e1 In . . . In Let vn = en In vn ...
but is a form easier to manipulate in Caml since lists of declarations will be appended
together at translation time. When writing a compiler, the programmer may or may not
want to use this intermediate form. It is not much harder to write the functions to work
directly on the Let representation.
We now sketch the translation for the core primitives. Assume the following auxiliary
functions have been dened:
CHAPTER 8. COMPILATION BY PROGRAM TRANSFORMATION 142
• newid() which returns a fresh F[SR variable every time called,
• The function letize which converts from the list-of-tuples form to the actual Let
form, and
• resultId, which for list [(v1,e1); ...; (vn,en)] returns result identier vn.
Denition 8.2 (A Translation).
let atrans e = letize (atrans0 e)
and atrans0(e) = match e with
(Var x) -> [(newid(),Var x)]
| (Int n) -> [(newid(),Int n)]
| (Bool b) -> [(newid(),Bool b)]
| Function(x,e) -> [(newid(),Function(x,atrans e)]
| Appl(e,e’) -> let a = atrans0 e in let a’ = atrans0 e’ in
a @ a’ @ [(newid(),Appl(resultId a,resultId a’)]
(* all other D binary operators + - = AND etc. of form
* identical to Appl
*)
| If(e1,e2,e3) -> let a1 = atrans0 e1 in
a1 @ [(newid();If(resultId a1,atrans e2,atrans e3)]
(* ... *)
At the end of the A-translation, the code is all “linear” in the way it runs in the
interpreter, not as a tree. Machine code is also linearly ordered; we are getting much
closer to machine code.
Theorem 8.2. A-translation is sound, i.e. e and atrans(e) both either compute to
values or both diverge.
Although we have only partially dened A-translation, the extra syntax of F[SR
(records, reference cells) does not provide any major complication.
8.3 Function Hoisting
So far, we have dened a front end of a compiler which performs closure conversion and
A-translation in turn:
let atrans clconv e = atrans(clconv(e))
After these two phases, functions will have no nonlocal variables. Thus, we can hoist
all functions in the program body to the start of the program. This brings the program
CHAPTER 8. COMPILATION BY PROGRAM TRANSFORMATION 143
structure more in line with C (and machine) code. Since our nal target is C, the leftover
code from which all functions were hoisted then is made the main function. A function
hoist carries out this transformation. Informally, the operation is quite simple: take
e.g.
4 + (Function x -> x + 1)(4)
and replace it by
Let f1 = Function x -> x + 1 In 4 + f1(4)
In general, we hoist all functions to the front of the code and give them a name via
Let. The transformation is always sound if there are no free variables in the function
body, a property guaranteed by closure conversion. We will dene this process in a simple
iterative (but inecient) manner:
Denition 8.3 (Function Hoisting).
let hoist e =
if e = e1[(Function ea -> e’)/f] for some e1 with f free,
and e’ itself contains no functions
(i.e. Function ea -> e’ is an innermost function)
then
Let f = (Function ea -> e’) In hoist(e1)
else e
This function hoists out innermost functions rst. If functions are not hoisted out
innermost-rst, there will still be some nested functions in the hoisted denitions. So,
the order of hoisting is important.
The denition of hoisting given above is concise, but it is too inecient. A one-pass
implementation can be used that recursively replaces functions with variables and accu-
mulates them in a list. This implementation is left as an exercise. Resulting programs
will be of the form
Let f1 = Function x1 -> e1 In
...
Let fn = Function xn -> en In
e
where each e, e1 , . . . en contain no function constants.
CHAPTER 8. COMPILATION BY PROGRAM TRANSFORMATION 144
Theorem 8.3. If all functions occurring in expression e contain no nonlocal variables,
then e ∼
= hoist(e).
This Theorem may be proved by iterative application of the following Lemma:
Lemma 8.1.
e1 [(Function x -> e′ )/f ] ∼
=
(Let f = (Function x -> e′ ) In e1
provided e′ contains at most the variable x free.
We lastly transform the program to
Let f1 = Function x1 -> e1 In
...
fn = Function -> Function xn -> en In
main = Function dummy -> e In
main(0)
So, the program is almost nothing but a collection of functions, with a body that just
invokes main. This brings the program closer to C programs, which are nothing but a
collection of functions and main is implicitly invoked at program start.
Let Rec denitions also need to be hoisted to the top level; their treatment is similar
and will be left as an exercise.
8.4 Translation to C
We are now ready to translate into C. To summarize up to now, we have
let hoist atrans clconv e = hoist(atrans(clconv(e)))
We have done about all the translation that is possible within F[SR. Programs are indeed
looking a lot more like machine code: all functions are declared at the top level, and each
function body consists of a linear sequence of atomic instructions (with exception of If
which is a branch). There still are a few things that are more complex than machine
code: records are still implicitly allocated, and function call is atomic, no pushing of
parameters is needed. Since C has function call built in, records are the only signicant
gap that needs to be closed.
The translation involves two main operations.
1. Map each function to a C function
2. For each function body, map each atomic tuple to a primitive C statement.
CHAPTER 8. COMPILATION BY PROGRAM TRANSFORMATION 145
Atomic Tuples Before giving the translation, we enumerate all possible right-hand
sides of Let variable assignments that come out of the A-translation (in the following
vi, vj, vk, and f are variables).These are called the atomic tuples.
Fact 8.1 (Atomic Tuples). F[SR programs that have passed through the rst three phases
have function bodies consisting of tuple lists where each tuple is of one of the following
forms only:
1. x for variable x
2. n for number n
3. b for boolean b
4. vi vj (application)
5. vj + vk
6. vj - vk
7. vj And vk
8. vj Or vk
9. Not vj
10. vj = vk
11. Ref vj
12. vj := vk
13. !vj
14. { l1 = v1; ...; ln = vn }
15. vi.l
16. If vi Then tuples1 Else tuples2 where tuples1 and tuples2 are the lists of
variable assignments for the Then and Else bodies.
Functions should have all been hoisted to the top so there will be none of those in
the tuples. Observe that some of the records usages are from the original program, and
others were added by the closure conversion process. We can view all of them as regular
records. All we need to do now is generate code for each of the above tuples.
CHAPTER 8. COMPILATION BY PROGRAM TRANSFORMATION 146
8.4.1 Memory Layout
Before writing any compiler, a xed memory layout scheme for objects at run-time is
needed. Since objects can be read and written from many dierent program points, if
read and write protocols are not uniform for a given object, the code simply will not work!
So, it is very important to carefully design the strategy beforehand. Simple compilers
such as ours will use simple schemes, but for eciency it is better to use a more complex
scheme.
Lets consider briey how memory is laid out in C. Values can be stored in several
dierent ways:
• In registers: These must be temporary, as registers are generally local to each
function/method. Also, registers are only one word in size (or a couple words, for
oats) so can’t directly hold arrays or structs.
• On the run-time stack in the function’s activation record. The value is then refer-
enced as the memory location at some xed oset from the stack pointer (which is
itself in a register).
• In xed memory locations (globals).
• In dynamically allocated (malloc’ed) memory locations (on the heap).
Figure 8.1 illustrates the overall model of memory we are dealing with.
To elaborate a bit more on stack storage of variables, here is some C pseudo-code to
give you the idea of how the stack pointer sp is used.
register int sp; /* compiler assigns sp to a register */
*(sp - 5) = 33;
printf("%d", *(sp - 5));
Stack-stored entities are also temporary in that they will be junk when the function/method
returns.
Another important issue is whether to box or unbox values.
Denition 8.4. A register or memory location vi ’s value is stored boxed if vi holds a
pointer to a block of memory containing the actual value. A variable’s value is unboxed
if it is directly in the register or memory location v1 .
Figure 8.2 illustrates the dierence between boxed and unboxed values.
For multi-word entities such as arrays, storing them unboxed means variables directly
hold a pointer to the rst word of the sequence of space. To clarify the above concepts
we review C’s memory layout convention. Variables may be declared either as globals,
register (the register directive is a request to put in a register only), or on the call stack;
all variables declared inside a function are kept on the stack. Variables directly holding
ints, floats, structs, and arrays are all unboxed. (Examples: int x; float x; int
arr[10]; snork x for snork a struct.) There is no such thing as a variable directly
holding a function; variables in C may only hold pointers to functions. It is possible to
write “v = f” in C where f is a previously declared function and not “v = &f”, but that
CHAPTER 8. COMPILATION BY PROGRAM TRANSFORMATION 147
Figure 8.1: Our model of memory
CHAPTER 8. COMPILATION BY PROGRAM TRANSFORMATION 148
Figure 8.2: Boxed vs. unboxed values. The integer value 123 is stored as an unboxed
value, while the record {x=5; y=10} is stored as a boxed value.
CHAPTER 8. COMPILATION BY PROGRAM TRANSFORMATION 149
is because the former is really syntactic sugar for the latter. A pointer to a function is in
fact a pointer to the start of the code of the function. Boxed variables in C are declared
explicitly, as pointer variables. (Examples: int *x; float *x; int *arr[10]; snork
*x for snork a struct.) All malloc’ed structures must be stored in a pointer variable
because they are boxed: variables can’t directly be heap entities. Variables are static
and the heap is dynamic.
Here is an example of a simple C program and the Sun SPARC assembly output
which gives some impressionistic idea of these concepts:
int glob;
main()
{
int x;
register int reg;
int* mall;
int arr[10];
x = glob + 1;
reg = x;
mall = (int *) malloc(1);
x = *mall;
arr[2] = 4;
/* arr = arr2; --illegal: arrays are not boxed */
}
In the assembly language, %o1 is a register, [%o0] means dereference, [%fp-24] means
subtract 24 from frame pointer register %fp and dereference. The assembly representation
of the above C code is as follows.
main:
sethi %hi(glob), %o1
or %o1, %lo(glob), %o0 /* load global address glob into %o0 */
ld [%o0], %o1 /* dereference */
add %o1, 1, %o0 /* increment */
st %o0, [%fp-20] /* store in [%fp-20], 20 back from fp -- x */
/* x directly contains a number,
/* not a pointer */
ld [%fp-20], %l0 /* %l0 IS reg (its in a register directly) */
mov 1, %o0
call malloc, 0 /* call malloc. resulting address to %o0 */
nop
st %o0, [%fp-24] /* put newspace location in mall [%fp-24] */
ld [%fp-24], %o0 /* load mall into %o0 */
ld [%o0], %o1 /* this is a malloced structure -- unbox. */
st %o1, [%fp-20] /* store into x */
mov 4, %o0
st %o0, [%fp-56] /* array is a sequence of memory on stack */
.LL2:
CHAPTER 8. COMPILATION BY PROGRAM TRANSFORMATION 150
ret
restore
Our memory layout strategy is more like a higher-level language such as Java or ML.
The Java JVM uses a particular, xed, memory layout scheme: all object references are
boxed pointers to heap locations; the primitive bool, byte, char, short, int, long,
float, and double types are kept unboxed. Since Java arrays are objects they are
also kept boxed. There is no reference (&) or dereference (*) operator in Java. These
operations occur implicitly.
Memory layout for our F[SR compiler F[SR is (mostly) a Caml subset and so its
memory is also managed more implicitly than C memory. We will use a simple, uniform
scheme in our compilers which is close in spirit to Java’s: Box Refs and records and
function values, but keep boolean values and integers unboxed. Also, as in C (and Java),
all local function variables will be allocated on the stack and accessed as osets from
the stack pointer. We will achieve the latter by implementing F[SR local variables as C
local variables, which will be stack allocated by the C compiler.
Since a Ref is nothing but a mutable location, there may not seem to be any reason
to box it. However, if a function returns a Ref as result, and it were not boxed, it would
have been allocated on the stack and thus would be deallocated. Here is an example that
reects this problem:
Let f = (Function x -> Ref 5) In !f( ) + 1
If Ref 5 were stored on the stack, after the return it could be wiped out. All of the
Let-dened entities in our tuples (the vi variables) can be either in registers or on the
call stack: none of those variables are directly used outside the function due to lexical
scoping, and they don’t directly contain values that should stay alive after the function
returns. For eciency, they can all be declared as register Word variables:
register Word v1, v2, v3, v4, ...;
One other advantage of this simple scheme is every variable holds one word of data,
and thus we don’t need to keep track of how much data a variable is holding. This
scheme is not very ecient, and real compilers optimize signicantly. One example is
Ref’s which are known to not escape a function can unboxed and stack allocated.
All that remains is to come up with a scheme to compile each of the above atomic
tuples and we are done. Records are the most dicult so we will consider them before
writing out the full translation.
Compiling untyped records Recall from when we covered records that the elds
present in a record cannot be known in advance if there is no type system. So, we won’t
know where the eld that we need is exactly. Consider, for example,
(Function x -> x.l)(If y = 0 Then {l = 3} Else {a = 4; l = 3})
CHAPTER 8. COMPILATION BY PROGRAM TRANSFORMATION 151
Field l will be in two dierent positions in these records so the selection will not have
a sole place it can nd the eld in. Thus we will need to use a hashtable for record
lookup. In a typed language such as Caml this problem is avoided: the above code is
not well-typed in Caml because the if-then can’t be typed. Note that the problems with
records are closely related to problems with objects, since objects are simply records with
Refs.
This memory layout diculty with records illustrates an important relationship be-
tween typing and compilation. Type systems impose constraints on program structure
that can make compilers easier to implement. Additionally, typecheckers will obviate
the need to deal with certain run-time errors. Our simple F[SR compilers are going to
core dump on e.g. 4 (5); in Lisp, Smalltalk, or Scheme these errors would be caught
at run-time but would slow down execution. In a typed language, the compiler would
reject the program since it will not typecheck. Thus for typed languages they will both
be faster and safer.
Our method for compilation of records proceeds as follows. We must give records
a heavy implementation, as hash tables (i.e., a set of key-value pairs, where the keys
are label names). In order to make the implementation simple, records are boxed so
they take one word of memory, as mentioned above when we covered boxing. A record
selection operation vk .l is implemented by hashing on key l in the hash table pointed to
by vk at runtime. This is more or less how Smalltalk message sends are implemented,
since records are similar to objects (and Smalltalk is untyped).
The above is less than optimal because space will be needed for the hashtable, and
record eld accessing will be much slower than, for example, struct access in C. Since
closures are records, this will also signicantly slow down function call. A simple opti-
mization would be to treat closure records specially since the eld positions will always
be xed, and use a struct implementation of closure (create a dierent struct type for
each function).
For instance, consider
(Function x -> x.l)(If y = 0 Then {l = 3} Else {a = 4; l = 3})
The code x.l will invoke a call of approximate form hashlookup(x,"l"). {a = 4; l =
3} will create a new hash table and add mappings of "a" to 4 and "l" to 3.
8.4.2 The toC translation
We are now ready to write the nal translation to C, via functions
• toCTuple mapping an atomic tuple to a C statement string,
• toCTuples mapping a list of tuples to C statements,
• toCFunction mapping a primitive F[SR function to a string dening a C function,
• toC mapping a list of primitive F[SR functions to a string of C functions.
CHAPTER 8. COMPILATION BY PROGRAM TRANSFORMATION 152
The translation as informally written below takes a few liberties for simplicity. Strings
"..." below are written in shorthand. For instance "vi = x" is shorthand for tostring(vi)
^" = " ^tostring(x). The tuples Let x1 = e1 In Let ...In Let xn = en In xn
of function and then/else bodies are assumed to have been converted to lists of tu-
ples [(x1,e1),...,(xn,en)], and similarly for the list of top-level function denitions.
When writing a compiler, it probably will be easier just to simply keep them in Let form,
although either strategy will work.
toCTuple(vi = x) = "vi = x;" (* x is a FbSR variable *)
toCTuple(vi = n) = "vi = n;"
toCTuple(vi = b) = "vi = b;"
toCTuple(vi = vj + vk) = "vi = vj + vk;"
toCTuple(vi = vj - vk) = "vi = vj - vk;"
toCTuple(vi = vj And vk ) = "vi = vj && vk;"
toCTuple(vi = vj Or vk ) = "vi = vj || vk;"
toCTuple(vi = Not vj ) = "vi = !vj;"
toCTuple(vi = vj = vk) = "vi = (vj == vk);"
toCTuple(vi = (vj vk) = "vi = *vj(vk);"
toCTuple(vi = Ref vj) = "vi = malloc(WORDSIZE); *vi = vj;"
toCTuple(vi = vj := vk) = "vi = *vj = vk;"
toCTuple(vi = !vj) = "vi = *vj;"
toCTuple(vi = { l1 = v1; ... ; ln = vn }) =
/* 1. malloc a new hashtable at vi
2. add mappings l1 -> v1 , ... , ln -> vn */
toCTuple(vi = vj.l) = "vi = hashlookup(vj,"l");"
toCTuple(vi = If vj Then tuples1 Else tuples2) =
"if (vj) { toCTuples(tuples1) } else { toCTuples(tuples2) };"
toCtuples([]) = ""
toCtuples(tuple::tuples) = toCtuple(tuple) ^ toCtuples(tuples)
toCFunction(f = Function xx -> tuples) =
"Word f(Word xx) {" ^ ... declare temporaries ...
toCtuples(tuples) ^
"return(resultId tuples); };"
toCFunctions([]) = ""
toCFunctions(Functiontuple::Functiontuples) =
toCFunction(Functiontuple) ^ toCFunctions(Functiontuples)
(* toC then invokes toCFunctions on its list of functions. *)
The reader may wonder why a fresh memory location is allocated for a Ref, as opposed
to simply storing the existing address of the object being referenced. This is a subtle
issue, but the code vi = &vj, for example, would denitely not work for the Ref case
(vj may go out of scope).
This translation sketch above leaves out many details. Here is some elaboration.
Typing issues We designed out memory layout so that every entity takes up one
word. So, every variable is of some type that is one word in size. Type all variables
CHAPTER 8. COMPILATION BY PROGRAM TRANSFORMATION 153
as Word’s, where Word is a 1-word type (dened as e.g. typedef void *Word;). Many
type casts need to be inserted; we are basically turning o the type-checking of C, but
there is no ”switch” that can be icked. So, for instance vi = vj + vk will really be
vi = (Word (int vj) + (int vk)) – cast the words to ints, do the addition, and
cast back to a word. To cast to a function pointer is a tongue-twister: in C you can
use (*((Word (*)()) f))(arg). The simplest way to avoid confusion when actually
writing a compiler is to include the following typedefs to the resulting C code:
/*
* Define the type ‘Word’ to be a generic one-word value.
*/
typedef void *Word;
/*
* Define the type ‘FPtr’ to be a pointer to a function that
* consumes Word and returns a Word. All translated
* functions should be of this form.
*/
typedef Word (*FPtr)(Word);
/*
* Here is an example of how to use these typedefs in code.
*/
Word my function(Word w) {
return w;
}
int main(int argc, char **argv) {
Word f1 = (Word) my function;
Word w1 = (Word) 123;
Word w2 = ( *((FPtr) f1) )(w1); /* Computes f1(123) */
printf("%d\n", (int) w2); /* output is "123\n". */
return 0;
}
Global Issues Some global issues you will need to deal with include the following.
You will need to print out the result returned by the main function (so, you probably
want the FbSR main function to be called something like FbSRmain and then write your
own main() by hand which will call FbSRmain); The C functions need to declare all the
temporary variables they use. One solution is to declare in the function header a C array
Word v[22]
where 22 is the number of temporaries needed in this particular function, and use names
v[0], v[1], etc for the temporaries. Note, this works well only if the newid() function
CHAPTER 8. COMPILATION BY PROGRAM TRANSFORMATION 154
is instructed to start numbering temporaries at zero again upon compiling each new
function. Every compiled program is going to have to come with a standard block of
C code in the header, including the record hash implementation, main() as alluded to
above, etc.
Other issues that present problems include the following. Record creation is only
sketched; but there are many C hash set libraries that could be used for this purpose.
The nal result (resultId) of the Then and Else tuples needs to be in the same variable
vi, which is also the variable where the result of the tuple is put, for the If code to be
correct. This is best handled in the A-translation phase.
There are several other issues that will arise when writing the compiler. The full
implementation is left as an exercise.
8.4.3 Compilation to Assembly code
Now that we have covered the compilation to C, we informally consider how a compiler
would compile all the way to machine code. The C code we produce above is very close
to assembly code. It would be conceptually easy to translate into assembly, but we skip
the topic due to the large number of cases that arise in the process (saving registers,
allocating space on the stack for temporaries).
8.5 Summary
let frontend e = hoist(atrans(clconv(e)));;
let translator e = toC(frontend(e));;
We can assert the correctness of our translator.
Theorem 8.4. FbSR program e terminates in the F[SR operational semantics (or eval-
uator) just when the C program translator(e) terminates, provided the C program
does not run out of memory. Core dump or other run-time errors are equated with
nontermination. Furthermore, if F[SR’s eval(e) returns a number n, the compiled
translator(e) will also produce numerical output n.
8.6 Optimization
Optimization can be done at all phases of the translation process. The above translation
is simple, but inecient. There is always a tradeo between simplicity and eciency in
compiler designs, both the eciency of compilation itself, and eciency of code produced.
In the phases before C code is produced, optimizations consist of replacing chunks of the
program with operationally equivalent chunks.
Some simple optimizations include the following. The special closure records {fn =
.., envt = .. } could be implemented as a pointer to a C struct with fn and envt
elds, instead of using the very slow hash method,1 will signicantly speed up the code
1
Although hash lookups are O(1), there is still a large amount of constant overhead, whereas struct
access can be done in a single load operation.
CHAPTER 8. COMPILATION BY PROGRAM TRANSFORMATION 155
produced. Records which do not not have eld names overlapping with other records
can also be implemented in this manner (there can be two dierent records with the
same elds, but not two dierent records with some elds the same and some dierent).
Another optimization is to modify the A-translation to avoid making tuples for variables
and constants. Constant expressions such as 3 + 4 can be folded to 7.
More fancy optimizations require a global ow analysis be performed. Simply put,
a ow analysis nds all possible uses of a particular denition, and all possible denitions
corresponding to a particular use.
A denition is a record, a Function, or a number or boolean, and a use is a record
eld selection, function application, or numerical or boolean operator.
8.7 Garbage Collection
Our compiled code mallocs but never frees. We will eventually run out of memory. A
garbage collector is needed.
Denition: In a run-time image, memory location n is garbage if it never will be
read or written to again.
There are many notions of garbage detection. The most common is to be somewhat
more conservative and take garbage to be memory locations which are not pointed to by
any known (”root”) object.
Bibliography
[1] Open source erlang. http://www.erlang.org/.
[2] The Self programming language. http://research.sun.com/self/language.
html.
[3] Typed assembly language. http://www.cs.cornell.edu/talc/.
[4] A.V. Aho, R. Sethi, and J.D. Ullman. Compilers: Principles, Techniques and Tools.
Addison-Wesley, 1986.
[5] A. Appel. Compiling with Continuations. Cambridge University Press, 1992.
[6] Andrew Appel. Modern Compiler Implementation in ML. Cambridge University
Press, 1998.
[7] Richard Bird. Introduction to Functional Programming using Haskell. Prentice Hall,
2nd edition, 1998.
[8] Gibblad Bracha and David Griswold. Strongtalk: Typechecking Smalltalk in a
Production Environment. In Proceedings of the OOPSLA ’93 Conference on Object-
oriented Programming Systems, Languages and Applications, pages 215–230, 1993.
[9] Kim Bruce. Foundations of Object-Oriented Languages: Types and Semantics. MIT
Press, 2002.
[10] Jonathan Eifrig, Scott Smith, Valery Trifonov, and Amy Zwarico. Application of
OOP type theory: State, decidability, integration. In OOPSLA ’94, pages 16–30,
1994.
[11] Martin Fowler. UML Distilled. Addison Wesley, 2nd edition, 2000.
[12] Erich Gamma, Richard Helm, Ralph Johnson, and John Vlissides. Design Patterns:
Elements of Reusable Object-Oriented Software. Addison-Wesley Professional Com-
puting Series. Addison-Wesley, 1994.
[13] Paul Hudak, John Peterson, and Joseph Fasel. A gentle introduction to Haskell,
version 98, June 2000. http://www.haskell.org/tutorial/.
[14] Andrew D. Irvine. Russell’s paradox. Stanford Encyclopedia of Philosophy, June
2001. http://plato.stanford.edu/entries/russell-paradox/.
156
BIBLIOGRAPHY 157
[15] Xavier Leroy. The Objective Caml system release 3.11, documentation and
user’s manual, November 2008. http://caml.inria.fr/pub/docs/manual-ocaml/
index.html.
[16] Ian A. Mason, Scott F. Smith, and Carolyn L. Talcott. From operational semantics
to domain theory. Information and Computation, 128(1):26–47, 1996.
[17] Greg Morrisett, Karl Crary, Neal Glew, Dan Grossman, Richard Samuels, Frederick
Smith, David Walker, Stephanie Weirich, and Steve Zdancewic. Talx86: A realistic
typed assembly language. In 1999 ACM SIGPLAN Workshop on Compiler Support
for System Software, pages 25–35, Atlanta, GA, USA, May 1999.
[18] J J O’Connor and E F Robertson. Gottfried Wilhelm von Leibniz. The MacTu-
tor History of Mathematics Archive, October 1998. http://www-history.mcs.
st-andrews.ac.uk/history/Mathematicians/Leibniz.html.
[19] Randall B. Smith and David Ungar. Self: The power of simplicity. In Conference
proceedings on Object-oriented programming systems, languages and applications,
pages 227–242. ACM Press, 1987.
[20] Randall B. Smith and David Ungar. Programming as an experience: The inspiration
for Self. Lecture Notes in Computer Science, 952:303–??, 1995.
[21] Scott Smith. Programming languages course. http://pl.cs.jhu.edu/pl.
[22] J. Stoy. Denotational Semantics: The Scott-Strachey Approach to Programming
Language Theory. MIT Press, 1977.
[23] Paul R. Wilson. Uniprocessor garbage collection techniques. ACM Computing Sur-
veys, 2002. ftp://ftp.cs.utexas.edu/pub/garbage/bigsurv.ps.
Index
A-translation, 139–142 environment-based interpreters, see inter-
abstract syntax, 10, 16, 40 preters, environment-based
Actor Model, 129 equational inference, 116
Actors, 129 equational types, see types, equational
α-equivalence, 46 equivalence, see operational equivalence
atomic region, 128 η-conversion, 44
atomic tuples, 145 η-equivalence, 46
Atomicity, 128 exceptions, 73–78
axiomatic semantics, 4 explicit environment interpreter, 67
axioms, 8
faithful implementation, 13, 62
β-equivalence, see also β-reduction, 46 ow analysis, 155
β-reduction, 49 freezing, see also thawing, 32
bound occurrence, see occurrence, bound function hoisting, 142–144
boxed values, 146
garbage collection, 64, 66–67
call-by-name, 39 generic types, 116
call-by-need, 39
capture, 30 Haskell, 39
capture-avoiding substitution, 49
information hiding, 85–87
classes, 87–88
inheritance, 88–89
closed expression, 19
interpreters
closure, 68
environment-based, 67–68
closure conversion, 135–139
compilation, 134–155 l-value, see also r-value, 66
concrete syntax, 6, 16 lambda-calculus, 48–49
concurrency, 127–133 lazy evaluation, 39
congruence, 46 logical combinators, 29
control operations, 73 loops, 64
curried function, 32
cycle, 128 message, 129
cyclical store, 64–66 metacircular interpreter, 15
metavariables, 4
Deadlock, 128 Monitors, 128
denotational semantics, 4
deterministic languages, 10, 28 nonlocal variable, 135
dispatch, 85 normalizing languages, 10, 28, 107
domain theory, 46
dynamic dispatch, 89–91 object polymorphism, 53, 84–85
158
INDEX 159
relation to record polymorphism, 84 symmetry, 46
objects, 79–98 syntax, 4
encoding as records, 51 syntax diagram, 5
occurrence, 18 syntax tree, 10
bound, 18
free, 18 thawing, see also freezing, 32
operational equivalence, 44–49 threading, 60
operational semantics, 4–49 touch, 48
denition, 4 transitivity, 46
tuples, 50–51
pairs, see tuples turing completeness
parametric polymorphism, 116 denition, 15
partial recursive function, 15 Turing-complete, 15, 49, 74
denition, 16 tying the knot, 65
polymorphism, 52–53 type assertion, 104
on objects, see object polymorphism type checking, 104, 108–109
on records, see record polymorphism type environment, 104
primitive objects, 93 type inference, 103, 104, 116–126
principal type, 116 and polymorphism, 116
program context, 45 constrained, 124–126
proof, 9 type schema, 122
proof system, 9 type soundness, 107
proof tree, 9 type systems, 104–105
dynamic, 101
r-value, see also l-value, 66 equational, 116–123
race condition, 128 overview of, 100–103
record polymorphism, 52–53 static, 101
records, 51–54 types, 15, 99–126
reexivity, 46 checking, see type checking
renaming substitution, see capture-avoiding inference, see type inference
substitution
Return, 74–76 untyped languages, 101
runtime environment, 67
Russell’s Paradox, 33–35 variable capture, see capture
variable substitution, 17–21
self-application encoding, 83 denition, 19
semantics, 4 variants, 54–57
Semaphore, 128 polymorphism, 55
sequencing, 64
side-eects, 58–78 Y -combinator, 38–39, 100, 107
state, 58–73
static elds, 91–92
static methods, 91–92
store, 60
stuck state, 108
subtype polymorphism, 52
subtyping, 112–115
You might also like
- Data Structure Optimization For Functional Programs: Maciej GodekNo ratings yetData Structure Optimization For Functional Programs: Maciej Godek176 pages
- ACSL: ANSI/ISO C Specification LanguageNo ratings yetACSL: ANSI/ISO C Specification Language95 pages
- Programming in Standard ML: Robert Harper Carnegie Mellon University Spring Semester, 2005No ratings yetProgramming in Standard ML: Robert Harper Carnegie Mellon University Spring Semester, 2005293 pages
- Ada For Software Engineers (Ben-Ari, 2005)No ratings yetAda For Software Engineers (Ben-Ari, 2005)357 pages
- Write You A Haskell: Building A Modern Functional Compiler From First Principles100% (3)Write You A Haskell: Building A Modern Functional Compiler From First Principles193 pages
- Visual Based Product Identification For Blind: Project Report OnNo ratings yetVisual Based Product Identification For Blind: Project Report On23 pages
- Chapter 2: Building Abstractions With ObjectsNo ratings yetChapter 2: Building Abstractions With Objects55 pages
- Peter Grogono - The Evolution of Programming LanguagesNo ratings yetPeter Grogono - The Evolution of Programming Languages125 pages
- ACSL: ANSI/ISO C Specification Language: Preliminary Design (Version 1.4, November 5, 2008)No ratings yetACSL: ANSI/ISO C Specification Language: Preliminary Design (Version 1.4, November 5, 2008)87 pages
- Mena As Functional Programming Ideas For The Curious KotlineNo ratings yetMena As Functional Programming Ideas For The Curious Kotline186 pages
- Programming Language Concepts: Lecture NotesNo ratings yetProgramming Language Concepts: Lecture Notes182 pages
- School of Computer Science: August 2, 2012No ratings yetSchool of Computer Science: August 2, 2012224 pages
- Shed Type Structures - Steel Vs Bamboo: Mini Project Report On100% (2)Shed Type Structures - Steel Vs Bamboo: Mini Project Report On33 pages
- 1999 - Grogono Evolution Programming LanguagesNo ratings yet1999 - Grogono Evolution Programming Languages122 pages
- Comparing Performance of L7x Vs L6x Using Logix5000 Task Monitor ToolNo ratings yetComparing Performance of L7x Vs L6x Using Logix5000 Task Monitor Tool13 pages
- Operating System Interview Questions With AnswersNo ratings yetOperating System Interview Questions With Answers12 pages
- Manual MRS TOOLS V 2 .1 English LanguageNo ratings yetManual MRS TOOLS V 2 .1 English Language19 pages
- Assignment 1, 2 and Punishment Assignment QuestionsNo ratings yetAssignment 1, 2 and Punishment Assignment Questions2 pages
- Time-dependent Behaviour and Design of Composite Steel-concrete StructuresFrom EverandTime-dependent Behaviour and Design of Composite Steel-concrete StructuresNo ratings yet
- Advanced Multiplayer Game Development with Ureal Engine 5: A Comprehensive Guide to C++ ScriptingFrom EverandAdvanced Multiplayer Game Development with Ureal Engine 5: A Comprehensive Guide to C++ ScriptingNo ratings yet
- The Linux Terminal for Advanced Users - The Command Line Made Easy: First EditionFrom EverandThe Linux Terminal for Advanced Users - The Command Line Made Easy: First EditionNo ratings yet
