Programming Languages: Application and Interpretation

Programming Languages:
Application and Interpretation
Shriram Krishnamurthi
Brown University
Copyright
c 2003, Shriram Krishnamurthi
This work is licensed under the

Creative Commons Attribution-NonCommercial-ShareAlike 3.0 United States License.
If you create a derivative work, please include the version information below in your attribution.
This book is available free-of-cost from the author’s Web site.

This version was generated on 2007-04-26.
ii
Preface
The book is the textbook for the programming languages course at Brown University, which is taken pri-
marily by third and fourth year undergraduates and beginning graduate (both MS and PhD) students. It
seems very accessible to smart second year students too, and indeed those are some of my most successful
students. The book has been used at over a dozen other universities as a primary or secondary text. The
book’s material is worth one undergraduate course worth of credit.
This book is the fruit of a vision for teaching programming languages by integrating the “two cultures”
that have evolved in its pedagogy. One culture is based on interpreters, while the other emphasizes a survey
of languages. Each approach has significant advantages but also huge drawbacks. The interpreter method
writes programs to learn concepts, and has its heart the fundamental belief that by teaching the computer to
execute a concept we more thoroughly learn it ourselves.
While this reasoning is internally consistent, it fails to recognize that understanding definitions does
not imply we understand consequences of those definitions. For instance, the difference between strict
and lazy evaluation, or between static and dynamic scope, is only a few lines of interpreter code, but the
consequences of these choices is enormous. The survey of languages school is better suited to understand
these consequences.
The text therefore melds these two approaches. Concretely, students program with a new set of features
first, then try to distill those principles into an actual interpreter. This has the following benefits:
• By seeing the feature in the context of a real language, students can build something interesting with
it first, so they understand that it isn’t an entirely theoretical construct, and will actually care to build
an interpreter for it. (Relatively few students are excited in interpreters for their own sake, and we
have an obligation to appeal to the remainder too.)
• Students get at least fleeting exposure to multiple languages, which is an important educational at-
tribute that is being crushed by the wide adoption of industrially fashionable languages. (Better still,
by experimenting widely, they may come to appreciate that industrial fashions are just that, not the
last word in technological progress.)
• Because they have already programmed with the feature, the explanations and discussions are much
more interesting than when all students have seen is an abstract model.
• By first building a mental model for the feature through experience, students have a much better
chance of actually discovering how the interpreter is supposed to work.
iii
iv PREFACE
In short, many more humans learn by induction than by deduction, so a pedagogy that supports it is much
more likely to succeed than one that suppresses it. The book currently reflects this design, though the survey
parts are done better in lecture than in the book.
Separate from this vision is a goal. My goal is to not only teach students new material, but to also change
the way they solve problems. I want to show students where languages come from, why we should regard
languages as the ultimate form of abstraction, how to recognize such an evolving abstraction, and how to
turn what they recognize into a language. The last section of the book, on domain-specific languages, is a
growing step in this direction.
Design Principles
• Concepts like design, elegance and artistic sensibility are rarely manifest in computer science courses;
in the name of not being judgmental, we may be running the risk of depriving our students of judg-
ment itself. We should reverse this trend. Students must understand that artificial objects have their
own aesthetic; the student must learn to debate the tradeoffs that lead to an aesthetic. Programming
languages are some of the most thoroughly designed artifacts in computer science. Therefore, the
study of programming languages offers a microcosm to study design itself.
• The best means we have to lead students to knowledge is through questions, not answers. The best
education prepares them to assess new data by confronting it with questions, processing the responses,
and iterating until they have formed a mental model of it. This book is therefore structured more like
a discussion than a presentation. It leads the reader down wrong paths (so don’t blindly copy code
from it!). It allows readers to get comfortable with mistaken assumptions before breaking them down
systematically.
• The programming languages course is one of the few places in the curriculum where we can tease
out and correct our students’ misconceptions about this material. They are often misled on topics
such as efficiency and correctness. Therefore, material on compilation, type systems and memory
management should directly confront their biases. For instance, a presentation of garbage collection
that does not also discuss the trade-offs with manual memory management will fail to address the
prejudices students bear.
Background and Prerequisite

This book assumes that students are comfortable reasoning informally about loop invariants, have modest
mathematical maturity, and are familiar with the existence of the Halting Problem. At Brown, they have all
been exposed to Java but not necessarily to any other languages (such as Scheme).
Supplementary Material
There is some material I use in my course that isn’t (currently) in this book:
preparation in Scheme For the first week, I offer supplementary sessions that teach students Scheme. The
material from these sessions is available from my course Web pages. In addition, I recommend the
v
use of a simple introduction to Scheme, such as the early sections of The Little Schemer or of How to
Design Programs.
domain-specific languages I discuss instances of real-world domain-specific languages, such as the access-
control language XACML. Students find the concepts easy to grasp, and can see why the language is
significant. In addition, it is one they may themselves encounter (or even decide to use) in their
programming tasks.
garbage collection I have provided only limited notes on garbage collection because I feel no need to offer
my own alternative to Paul Wilson’s classic survey, Uniprocessor Garbage Collection Techniques. I
recommend choosing sections from this survey, depending on student maturity, as a supplement to
this text.
model checking I supplement the discussion of types with a presentation on model checking, to show
students that it is possible to go past the fixed set of theorems of traditional type systems to systems
that permit developers to state theorems of interest. I have a pre-prepared talk on this topic, and would
be happy to share those slides.
Web programming Before plunging into continuations, I discuss Web programming APIs and demonstrate
how they mask important control operators. I have a pre-prepared talk on this topic, and would
be happy to share those slides. I also wrap up the section on continuations with a presentation on
programming in the PLT Scheme Web server, which natively supports continuations.
articles on design I hand out a variety of articles on the topic of design. I’ve found Dan Ingalls’s dissection
of Smalltalk, Richard Gabriel’s on Lisp, and Paul Graham’s on both programming and design the
most useful. Graham has now collected his essays in the book Hackers and Painters.
logic programming The notes on logic programming are the least complete. Students are already familiar
with unification from type inference by the time I arrive at logic programming. Therefore, I focus on
the implementation of backtracking. I devote one lecture to the use of unification, the implications
of the occurs-check, depth-first versus breadth-first search, and tabling. In another lecture, I present
the implementation of backtracking through continuations. Concretely, I use the presentation in Dorai
Sitaram’s Teach Yourself Scheme in Fixnum Days. This presentation consolidates two prior topics,
continuations and macros.
Exercises
Numerous exercises are sprinkled throughout the book. Several more, in the form of homework assignments
and exams, are available from my course’s Web pages (where hyeari is one of 2000, 2001, 2002, 2003,
2004 and 2005):
http://www.cs.brown.edu/courses/cs173/hyeari/
In particular, in the book I do not implement garbage collectors and type checkers. These are instead
homework assignments, ones that students generally find extremely valuable (and very challenging!).
vi PREFACE
Programs
This book asks students to implement language features using a combination of interpreters and little com-
pilers. All the programming is done in Scheme, which has the added benefit of making students fairly
comfortable in a language and paradigm they may not have employed before. End-of-semester surveys re-
veal that students are far more likely to consider using Scheme for projects in other courses after taking this
course than they were before it (even when they had prior exposure to Scheme).
Though every line of code in this book has been tested and is executable, I purposely do not distribute
the code associated with this book. While executable code greatly enhances the study of programming
languages, it can also detract if students execute the code mindlessly. I therefore ask you, Dear Reader, to
please type in this code as if you were writing it, paying close attention to every line. You may be surprised
by how much many of them have to say.
Course Schedule
The course follows approximately the following schedule:
Weeks Topics
1 Introduction, Scheme tutorials, Modeling Languages
2–3 Substitution and Functions
3 Laziness
4 Recursion
4 Representation Choices
4–5 State
5–7 Continuations
7–8 Memory Management
8–10 Semantics and Types
11 Programming by Searching
11–12 Domain-Specific Languages and Metaprogramming
Miscellaneous “culture lecture” topics such as model checking, extensibility and future directions consume
another week.
An Invitation
I think the material in these pages is some of the most beautiful in all of human knowledge, and I hope any
poverty of presentation here doesn’t detract from it. Enjoy!
Acknowledgments
This book has a long and humbling provenance. The conceptual foundation for this interpreter-based ap-
proach traces back to seminal work by John McCarthy. My own introduction to it was through two texts I
read as an undergraduate, the first editions of The Structure and Interpretation of Computer Programs by
Abelson and Sussman with Sussman and Essentials of Programming Languages by Friedman, Wand and
Haynes. Please read those magnificent books even if you never read this one.
My graduate teaching assistants, Dave Tucker and Rob Hunter, wrote and helped edit lecture notes that
helped preserve continuity through iterations of my course. Greg Cooper has greatly influenced my thinking
on lazy evaluation. Six generations of students at Brown have endured drafts of this book.
Bruce Duba, Corky Cartwright, Andrew Wright, Cormac Flanagan, Matthew Flatt and Robby Findler
have all significantly improved my understanding of this material. Matthew and Robby’s work on DrScheme
has greatly enriched my course’s pedagogy. Christian Queinnec and Paul Graunke inspired the presentation
of continuations through Web programming, and Greg Cooper created the approach to garbage collection,
both of which are an infinite improvement over prior approaches.
Alan Zaring, John Lacey and Kathi Fisler recognized that I might like this material and introduced me to
it (over a decade ago) before it was distributed through regular channels. Dan Friedman generously gave of
his time as I navigated Essentials. Eli Barzilay, John Clements, Robby Findler, John Fiskio-Lasseter, Kathi
Fisler, Cormac Flanagan, Matthew Flatt, Suresh Jagannathan, Gregor Kiczales, Mira Mezini, Prabhakar
Ragde, Marc Smith, and Éric Tanter have provided valuable feedback after using prior versions of this text.
The book’s accompanying software has benefited from support by several generations of graduate as-
sistants, especially Greg Cooper and Guillaume Marceau. Eli Barzilay and Matthew Flatt have also made
excellent, creative contributions to it.
My chairs at Brown, Tom Dean and Eli Upfal, have permitted me to keep teaching my course so I
could develop this book. I can’t imagine how many course staffing nightmares they’ve endured, and ensuing
temptations they’ve suppressed, in the process.
My greatest debt is to Matthias Felleisen. An early version of this text grew out of my transcript of his
course at Rice University. That experience made me realize that even the perfection embodied in the books
I admired could be improved upon. This result is not more perfect, simply different—its outlook shaped by
standing on the shoulders of giants.
vii
viii ACKNOWLEDGMENTS
Thanks
Several more people have made suggestions, asked questions, and identified errors:
Ian Barland, Hrvoje Blazevic, Daniel Brown, Greg Buchholz, Lee Butterman, Richard Cobbe, Bruce Duba,
Stéphane Ducasse, Marco Ferrante, Dan Friedman, Mike Gennert, Arjun Guha, Roberto Ierusalimschy,
Steven Jenkins, Eric Koskinen, Neel Krishnaswami, Benjamin Landon, Usman Latif, Dan Licata, Alice Liu,
Paulo Matos, Grant Miner, Ravi Mohan, Jason Orendorff, Klaus Ostermann, Pupeno [sic], Manos Renieris,
Morten Rhiger, Bill Richter, Peter Rosenbeck, Amr Sabry, Francisco Solsona, Anton van Straaten, Andre
van Tonder, Michael Tschantz, Phil Wadler, Joel Weinberger, and Greg Woodhouse.
In addition, Michael Greenberg instructed me in the rudiments of classifying flora.

Contents
Preface iii
Acknowledgments vii
I Prelude 1
1 Modeling Languages 3
1.1 Modeling Meaning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Modeling Syntax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 A Primer on Parsers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Primus Inter Parsers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
II Rudimentary Interpreters 11
2 Interpreting Arithmetic 13
3 Substitution 15
3.1 Defining Substitution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2 Calculating with with . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.3 The Scope of with Expressions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.4 What Kind of Redundancy do Identifiers Eliminate? . . . . . . . . . . . . . . . . . . . . . . 23
3.5 Are Names Necessary? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4 An Introduction to Functions 27
4.1 Enriching the Language with Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.2 The Scope of Substitution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.3 The Scope of Function Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5 Deferring Substitution 33
5.1 The Substitution Repository . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
5.2 Deferring Substitution Correctly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
ix
x CONTENTS
5.3 Fixing the Interpreter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
6 First-Class Functions 41
6.1 A Taxonomy of Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
6.2 Enriching the Language with Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.3 Making with Redundant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
6.4 Implementing Functions using Deferred Substitutions . . . . . . . . . . . . . . . . . . . . . 45
6.5 Some Perspective on Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6.5.1 Filtering and Sorting Lists . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6.5.2 Differentiation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
6.5.3 Callbacks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
6.6 Eagerness and Laziness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
6.7 Standardizing Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
III Laziness 57
7 Programming with Laziness 59

7.1 Haskell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
7.1.1 Expressions and Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
7.1.2 Lists . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
7.1.3 Polymorphic Type Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
7.1.4 Laziness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
7.1.5 An Interpreter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
7.2 Shell Scripting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
8 Implementing Laziness 73
8.1 Implementing Laziness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
8.2 Caching Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
8.3 Caching Computations Safely . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
8.4 Scope and Evaluation Regimes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
IV Recursion 87
9 Understanding Recursion 89
9.1 A Recursion Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
9.2 Environments for Recursion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
9.3 An Environmental Hazard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
10 Implementing Recursion 97
CONTENTS xi
V Intermezzo 103
11 Representation Choices 105

11.1 Representing Environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
11.2 Representing Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
11.3 Representing Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
11.4 Types of Interpreters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
11.5 Procedural Representation of Recursive Environments . . . . . . . . . . . . . . . . . . . . . 109
VI State 115
12 Church and State 117
13 Mutable Data Structures 119

13.1 Implementation Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
13.2 Insight . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
13.3 An Interpreter for Mutable Boxes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
13.3.1 The Evaluation Pattern . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
13.3.2 The Interpreter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
13.4 Scope versus Extent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
14 Variables 133
14.1 Implementing Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
14.2 Interaction Between Variables and Function Application . . . . . . . . . . . . . . . . . . . 135
14.3 Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
VII Continuations 145
15 Some Problems with Web Programs 147
16 The Structure of Web Programs 149

16.1 Explicating the Pending Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
16.2 A Better Server Primitive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
16.3 Testing Web Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
16.4 Executing Programs on a Traditional Server . . . . . . . . . . . . . . . . . . . . . . . . . . 154
17 More Web Transformation 157

17.1 Transforming Library and Recursive Code . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
17.2 Transforming Multiple Procedures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
17.3 Transforming State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
17.4 The Essence of the Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
17.5 Transforming Higher-Order Procedures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
xii CONTENTS
17.6 Perspective on the Web Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
18 Conversion into Continuation-Passing Style 169

18.1 The Transformation, Informally . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
18.2 The Transformation, Formally . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
18.3 Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
19 Programming with Continuations 177

19.1 Capturing Continuations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
19.2 Escapers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
19.3 Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
19.4 Web Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
19.5 Producers and Consumers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
19.6 A Better Producer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
19.7 Why Continuations Matter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
20 Implementing Continuations 193

20.1 Representing Continuations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
20.2 Adding Continuations to the Language . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
20.3 On Stacks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
20.4 Tail Calls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
20.5 Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
VIII Memory Management 207
21 Automatic Memory Management 209

21.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
21.2 Truth and Provability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
IX Semantics 215
22 Shrinking the Language 217

22.1 Encoding Lists . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
22.2 Encoding Boolean Constants and Operations . . . . . . . . . . . . . . . . . . . . . . . . . . 219
22.3 Encoding Numbers and Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
22.4 Eliminating Recursion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
23 Semantics 231
CONTENTS xiii
X Types 235
24 Introduction 237
24.1 What Are Types? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
24.2 Type System Design Forces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
24.3 Why Types? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
25 Type Judgments 243

25.1 What They Are . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
25.2 How Type Judgments Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
26 Typing Control 249

26.1 Conditionals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
26.2 Recursion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
26.3 Termination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
26.4 Typed Recursive Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
27 Typing Data 255

27.1 Recursive Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
27.1.1 Declaring Recursive Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
27.1.2 Judgments for Recursive Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
27.1.3 Space for Datatype Variant Tags . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
28 Type Soundness 261
29 Explicit Polymorphism 265

29.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
29.2 Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
29.3 The Type Language . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
29.4 Evaluation Semantics and Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
29.5 Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
30 Type Inference 273

30.1 Inferring Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
30.1.1 Example: Factorial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
30.1.2 Example: Numeric-List Length . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
30.2 Formalizing Constraint Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276
30.3 Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
30.4 Example: Using First-Class Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
30.5 Solving Type Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
30.5.1 The Unification Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
30.5.2 Example of Unification at Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
30.5.3 Parameterized Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
30.5.4 The “Occurs” Check . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282
xiv CONTENTS
30.6 Underconstrained Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282

30.7 Principal Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
31 Implicit Polymorphism 285

31.1 The Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
31.2 A Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286
31.3 A Better Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
31.4 Recursion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
31.5 A Significant Subtlety . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
31.6 Why Let and not Lambda? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
31.7 The Structure of ML Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
31.8 Interaction with Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
XI Programming by Searching 291
32 Introduction 293
33 Programming in Prolog 295

33.1 Example: Academic Family Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
33.2 Intermission . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
33.3 Example: Encoding Type Judgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
33.4 Final Credits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
34 Implementing Prolog 307

34.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
34.1.1 Searching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
34.1.2 Satisfaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
34.1.3 Matching with Logic Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
34.2 Subtleties and Compromises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
34.3 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
XII Domain-Specific Languages and Metaprogramming 313
35 Domain-Specific Languages 315

35.1 Language Design Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
35.2 Languages as Abstractions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
35.3 Domain-Specific Languages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
36 Macros as Compilers 319

36.1 Language Reuse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
36.1.1 Example: Measuring Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
36.1.2 Example: Local Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
CONTENTS xv
36.1.3 Example: Nested Local Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . 323

36.1.4 Example: Simple Conditional . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
36.1.5 Example: Disjunction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
36.2 Hygiene . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
36.3 More Macrology by Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
36.3.1 Loops with Named Iteration Identifiers . . . . . . . . . . . . . . . . . . . . . . . . 329
36.3.2 Overriding Hygiene: Loops with Implicit Iteration Identifiers . . . . . . . . . . . . . 330
36.3.3 Combining the Pieces: A Loop for All Seasons . . . . . . . . . . . . . . . . . . . . 333
36.4 Comparison to Macros in C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
36.5 Abuses of Macros . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
36.6 Uses of Macros . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
37 Macros and their Impact on Language Design 337

37.1 Language Design Philosophy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
37.2 Example: Pattern Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
37.3 Example: Automata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
37.3.1 Concision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
37.3.2 Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
37.4 Other Uses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
37.5 Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
XIII What’s Next? 349
38 Programming Interactive Systems 351
39 What Else is Next 355

xvi CONTENTS
Part I
Prelude
1
Chapter 1
Modeling Languages
A student of programming languages who tries to study a new language can be overwhelmed by details.
Virtually every language consists of
• a peculiar syntax,
• some behavior associated with each syntax,
• numerous useful libraries, and
• a collection of idioms that programmers of that language use.
All four of these attributes are important to a programmer who wants to adopt a language. To a scholar,
however, one of these is profoundly significant, while the other three are of lesser importance.
The first insignificant attribute is the syntax. Syntaxes are highly sensitive topics,1 but in the end, they
don’t tell us very much about a program’s behavior. For instance, consider the following four fragments:
1. a [25]
2. (vector-ref a 25)
3. a [25]
4. a [25]
Which two are most alike? The first and second, obviously! Why? Because the first is in Java and the
second is in Scheme, both of which signal an error if the vector associated with a has fewer than 25 entries;
the third, in C, blithely ignores the vector’s size, leading to unspecified behavior, even though its syntax is
exactly the same as that of the Java code. The fourth, in ML or Haskell, is an application of a to the list
containing just one element, 25: that is, it’s not an array dereference at all, it’s a function appliction!
That said, syntax does matter, at least inasmuch as its brevity can help programmers express and under-
stand more by saying less. For the purpose of our study, however, syntax will typically be a distraction, and
1 Matthias Felleisen: “Syntax is the Viet Nam of programming languages.”
3
4 CHAPTER 1. MODELING LANGUAGES
will often get in the way of our understanding deeper similarities (as in the Java-Scheme-C example above).
We will therefore use a uniform syntax for all the languages we implement.
The size of a language’s library, while perhaps the most important characteristic to a programmer who
wants to accomplish a task, is usually a distraction when studying a language. This is a slightly tricky
contention, because the line between the core of a language and its library is fairly porous. Indeed, what one
language considers an intrinsic primitive, another may regard as a potentially superfluous library operation.
With experience, we can learn to distinguish between what must belong in the core and what need not. It
is even possible to make this distinction quite rigorously using mathematics. Our supplementary materials
will include literature on this distinction.
Finally, the idioms of a language are useful as a sociological exercise (“How do the natives of this lin-
guistic terrain cook up a Web script?”), but it’s dangerous to glean too much from them. Idioms are funda-
mentally human, and therefore bear all the perils of faulty, incomplete and sometimes even outlandish human
understanding. If a community of Java programmers has never seen a particular programming technique—
for instance, the principled use of objects as callbacks—they are likely to invent an idiom to take its place,
but it will almost certainly be weaker, less robust, and less informative to use the idiom than to just use
callbacks. In this case, and indeed in general, the idiom sometimes tells us more about the programmers
than it does about the language. Therefore, we should be careful to not read too much into one.
In this course, therefore, we will focus on the behavior associated with syntax, namely the semantics
of programming languages. In popular culture, people like to say “It’s just semantics!”, which is a kind of
put-down: it implies that their correspondent is quibbling over minor details of meaning in a jesuitical way.
But communication is all about meaning: even if you and I use different words to mean the same thing, we
understand one another; but if we use the same word to mean different things, great confusion results. In
this study, therefore, we will wear the phrase “It’s just semantics!” as a badge of honor, because semantics
leads to discourse which (we hope) leads to civilization.
Just semantics. That’s all there is.
1.1 Modeling Meaning

So we want to study semantics. But how? To study meaning, we need a language for describing meaning.
Human language is, however, notoriously slippery, and as such is a poor means for communicating what are
very precise concepts. But what else can we use?
Computer scientists use a variety of techniques for capturing the meaning of a program, all of which rely
on the following premise: the most precise language we have is that of mathematics (and logic). Tradition-
ally, three mathematical techniques have been especially popular: denotational, operational and axiomatic
semantics. Each of these is a rich and fascinating field of study in its own right, but these techniques are
either too cumbersome or too advanced for our use. (We will only briefly gloss over these topics, in sec-
tion 23.) We will instead use a method that is a first cousin of operational semantics, which some people
call interpreter semantics.
The idea behind an interpreter semantics is simple: to explain a language, write an interpreter for it. The
act of writing an interpreter forces us to understand the language, just as the act of writing a mathematical
description of it does. But when we’re done writing, the mathematics only resides on paper, whereas we can
run the interpreter to study its effect on sample programs. We might incrementally modify the interpreter
1.2. MODELING SYNTAX 5
if it makes a mistake. When we finally have what we think is the correct representation of a language’s
meaning, we can then use the interpreter to explore what the language does on interesting programs. We can
even convert an interpreter into a compiler, thus leading to an efficient implementation that arises directly
from the language’s definition.
A careful reader should, however, be either confused or enraged (or both). We’re going to describe
the meaning of a language through an interpreter, which is a program. That program is written in some
language. How do we know what that language means? Without establishing that first, our interpreters
would appear to be mere scrawls in an undefined notation. What have we gained?
This is an important philosophical point, but it’s not one we’re going to worry about much in practice.
We won’t for the practical reason that the language in which we write the interpreter is one that we under-
stand quite well: it’s succint and simple, so it won’t be too hard to hold it all in our heads. (Observe that
dictionaries face this same quandary, and negotiate it successsfully in much the same manner.) The supe-
rior, theoretical, reason is this: others have already worked out the mathematical semantics of this simple
language. Therefore, we really are building on rock. With that, enough of these philosophical questions for
now. We’ll see a few other ones later in the course.
1.2 Modeling Syntax

I’ve argued briefly that it is both futile and dangerous to vest too much emotion in syntax. In a platonic
world, we might say
Irrespective of whether we write
• 3+4 (infix),
• 3 4 + (postfix), or
• (+ 3 4) (parenthesized prefix),
we always mean the idealized action of adding the idealized numbers (represented by) “3” and
“4”.
Indeed, each of these notations is in use by at least one programming language.

If we ignore syntactic details, the essence of the input is a tree with the addition operation at the root
and two leaves, the left leaf representing the number 3 and the right leaf the number 4. With the right data
definition, we can describe this in Scheme as the expression
(add (num 3) (num 4))
and similarly, the expression
• (3 − 4) + 7 (infix),
• 3 4 - 7 + (postfix), or
• (+ (- 3 4) 7) (parenthesized prefix)
would be represented as
(add (sub (num 3) (num 4))
(num 7))
One data definition that supports these representations is the following:
(define-type AE
[num (n number?)]
[add (lhs AE?)
(rhs AE?)]
[sub (lhs AE?)
(rhs AE?)])
where AE stands for “Arithmetic Expression”.
Exercise 1.2.1 Why are the lhs and rhs sub-expressions of type AE rather than of type num? Provide sample
expressions permitted by the former and rejected by the latter, and argue that our choice is reasonable.
1.3 A Primer on Parsers

Our interpreter should consume terms of type AE, thereby avoiding the syntactic details of the source lan-
guage. For the user, however, it becomes onerous to construct terms of this type. Ideally, there should be a
program that translates terms in concrete syntax into values of this type. We call such a program a parser.
In more formal terms, a parser is a program that converts concrete syntax (what a user might type) into
abstract syntax. The word abstract signifies that the output of the parser is idealized, thereby divorced from
physical, or syntactic, representation.
As we’ve seen, there are many concrete syntaxes that we could use for arithmetic expressions. We’re
going to pick one particular, slightly peculiar notation. We will use a prefix parenthetical syntax that, for
arithmetic, will look just like that of Scheme. With one twist: we’ll use {braces} instead of (parentheses), so
we can distinguish concrete syntax from Scheme just by looking at the delimiters. Here are three programs
employing this concrete syntax:
1. 3
2. {+ 3 4}
3. {+ {- 3 4} 7}
Our choice is, admittedly, fueled by the presence of a convenient primitive in Scheme—the primitive
that explains why so many languages built atop Lisp and Scheme look so much like Lisp and Scheme (i.e.,
they’re parenthetical), even if they have entirely different meanings. That primitive is called read.
Here’s how read works. It consumes an input port (or, given none, examines the standard input port).
If it sees a sequence of characters that obey the syntax of a number, it converts them into the corresponding
number in Scheme and returns that number. That is, the input stream
1 7 2 9 <eof>
1.3. A PRIMER ON PARSERS 7
(the spaces are merely for effect, not part of the stream) would result in the Scheme number 1729. If the
sequence of characters obeys the syntax of a symbol (sans the leading quote), read returns that symbol: so
c s 1 7 3 <eof>
(again, the spaces are only for effect) evaluates to the Scheme symbol ’cs173. Likewise for other primitive
types. Finally, if the input is wrapped in a matched pair of parenthetical delimiters—either (parentheses),
[brackets] or {braces}—read returns a list of Scheme values, each the result of invoking read recursively.
Thus, for instance, read applied to the stream
(1 a)
returns (list 1 ’a), to
{+ 3 4}
returns (list ’+ 3 4), and to
{+ {- 3 4} 7}
returns (list ’+ (list ’- 3 4) 7).

The read primitive is a crown jewel of Lisp and Scheme. It reduces what are conventionally two quite
elaborate phases, called tokenizing (or scanning) and parsing, into three different phases: tokenizing, reading
and parsing. Furthermore, it provides a single primitive that does the first and second, so all that’s left to do
is the third. read returns a value known as an s-expression.
The parser needs to identify what kind of program it’s examining, and convert it to the appropriate
abstract syntax. To do this, it needs a clear specification of the concrete syntax of the language. We’ll
use Backus-Naur Form (BNF), named for two early programming language pioneers. A BNF description of
rudimentary arithmetic looks like this:
<AE> ::= <num>

| {+ <AE> <AE>}
| {- <AE> <AE>}
The <AE> in the BNF is called a non-terminal, which means we can rewrite it as one of the things on the
right-hand side. Read ::= as “can be rewritten as”. Each | presents one more choice, called a produc-
tion. Everything in a production that isn’t enclosed in <· · ·> is literal syntax. (To keep the description
simple, we assume that there’s a corresponding definition for <num>, but leave its precise definition to your
imagination.) The <AE>s in the productions are references back to the <AE> non-terminal.
Notice the strong similarity between the BNF and the abstract syntax representation. In one stroke, the
BNF captures both the concrete syntax (the brackets, the operators representing addition and subtraction)
and a default abstract syntax. Indeed, the only thing that the actual abstract syntax data definition contains
that’s not in the BNF is names for the fields. Because BNF tells the story of concrete and abstract syntax so
succintly, it has been used in definitions of languages ever since Algol 60, where it first saw use.
Assuming all programs fed to the parser are syntactically valid, the result of reading must be either a
number, or a list whose first value is a symbol (specifically, either ’+ or ’-) and whose second and third
values are sub-expressions that need further parsing. Thus, the entire parser looks like this:2
;; parse : sexp −→ AE
;; to convert s-expressions into AEs
(define (parse sexp)

(cond
[(number? sexp) (num sexp)]
[(list? sexp)
(case (first sexp)
[(+) (add (parse (second sexp))
(parse (third sexp)))]
[(-) (sub (parse (second sexp))
(parse (third sexp)))])]))
Here’s the parser at work. The first line after each invocation of (parse (read)) is what the user types;
the second line after it is the result of parsing. This is followed by the next prompt.
Language: PLAI - Advanced Student.

> (parse (read))
3
(num 3)
> (parse (read))
{+ 3 4}
> (parse (read))
{+ {- 3 4} 7}
(add (sub (num 3) (num 4)) (num 7))
This, however, raises a practical problem: we must type programs in concrete syntax manually every
time we want to test our programs, or we must pre-convert the concrete syntax to abstract syntax. The
problem arises because read demands manual input each time it runs. We might be tempted to use an
intermediary such as a file, but fortunately, Scheme provides a handy notation that lets us avoid this problem
entirely: we can use the quote notation to simulate read. That is, we can write

> (parse ’3)
(num 3)
> (parse ’{+ 3 4})
2 Thisis a parser for the whole language, but it is not a complete parser, because it performs very little error reporting: if a user
provides the program {+ 1 2 3}, which is not syntactically legal according to our BNF specification, the parser silently ignores
the 3 instead of signaling an error. You must write more robust parsers than this one.
1.4. PRIMUS INTER PARSERS 9
> (parse ’{+ {- 3 4} 7})

(add (sub (num 3) (num 4)) (num 7))
This is the last parser we will write in this book. From now on, you will be responsible for creating
a parser from the concrete syntax to the abstract syntax. Extending the parser we have seen is generally
straightforward because of the nature of syntax we use, which means it would be worthwhile to understand
the syntax better.
1.4 Primus Inter Parsers

Most languages do not use this form of parenthesized syntax. Writing parsers for languages that don’t is
much more complex; to learn more about that, study a typical text from a compilers course. Before we drop
the matter of syntax entirely, however, let’s discuss our choice—parenthetical syntax—in a little more depth.
I said above that read is a crown jewel of Lisp and Scheme. In fact, I think it’s actually one of the great
ideas of computer science. It serves as the cog that helps decompose a fundamentally difficult process—
generalized parsing of the input stream—into two very simple processes: reading the input stream into an
intermediate representation, and parsing that intermediate representation. Writing a reader is relatively sim-
ple: when you see a opening bracket, read recursively until you hit a closing bracket, and return everything
you saw as a list. That’s it. Writing a parser using this list representation, as we’ve seen above, is also a
snap.
I call these kinds of syntaxes bicameral,3 which is a term usually used to describe legislatures such as
that of the USA. No issue becomes law without passing muster in both houses. The lower house establishes
a preliminary bar for entry, but allows some rabble to pass through knowing that the wisdom of the upper
house will prevent excesses. In turn, the upper house can focus on a smaller and more important set of
problems. In a bicameral syntax, the reader is, in American terms, the House of Representatives: it rejects
the input
{+ 1 2)
(mismatched delimiters) but permits both of
{+ 1 2}
{+ 1 2 3}
the first of which is legal, the second of which isn’t in our arithmetic language. It’s the parser’s (Senate’s)
job to eliminate the latter, more refined form of invalid input.
Exercise 1.4.1 Based on this discussion, examine XML. What do the terms well-formed and valid mean,
and how do they differ? How do these requirements relate to bicameral syntaxes such as that of Scheme?
3 Two houses.
Part II
Rudimentary Interpreters
11
Chapter 2
Interpreting Arithmetic
Having established a handle on parsing, which addresses syntax, we now begin to study semantics. We will
study a language with only numbers, addition and subtraction, and further assume both these operations are
binary. This is indeed a very rudimentary exercise, but that’s the point. By picking something you know
well, we can focus on the mechanics. Once you have a feel for the mechanics, we can use the same methods
to explore languages you have never seen before.
The interpreter has the following contract and purpose:
;; calc : AE −→ number
;; consumes an AE and computes the corresponding number
which leads to these test cases:
(test (calc (parse ’3)) 3)
(test (calc (parse ’{+ 3 4})) 7)
(test (calc (parse ’{+ {- 3 4} 7})) 6)
(notice that the tests must be consistent with the contract and purpose statement!) and this template:
(define (calc an-ae)
(type-case AE an-ae
[num (n) · · ·]
[add (l r) · · · (calc l) · · · (calc r) · · ·]
[sub (l r) · · · (calc l) · · · (calc r) · · ·]))
In this instance, we can convert the template into a function easily enough:
(define (calc an-ae)
(type-case AE an-ae
[num (n) n]
[add (l r) (+ (calc l) (calc r))]
[sub (l r) (- (calc l) (calc r))]))
Running the test suite helps validate our interpreter.
13
14 CHAPTER 2. INTERPRETING ARITHMETIC
What we have seen is actually quite remarkable, though its full power may not yet be apparent. We have
shown that a programming language with just the ability to represent structured data can represent one of
the most interesting forms of data, namely programs themselves. That is, we have just written a program
that consumes programs; perhaps we can even write programs that generate programs. The former is the
foundation for an interpreter semantics, while the latter is the foundation for a compiler. This same idea—
but with a much more primitive language, namely arithmetic, and a much poorer collection of data, namely
just numbers—is at the heart of the proof of Gödel’s Theorem.
Chapter 3
Substitution
Even in a simple arithmetic language, we sometimes encounter repeated expressions. For instance, the
Newtonian formula for the gravitational force between two objects has a squared term in the denominator.
We’d like to avoid redundant expressions: they are annoying to repeat, we might make a mistake while
repeating them, and evaluating them wastes computational cycles.
The normal way to avoid redundancy is to introduce an identifier.1 As its name suggests, an identifier
names, or identifies, (the value of) an expression. We can then use its name in place of the larger compu-
tation. Identifiers may sound exotic, but you’re used to them in every programming language you’ve used
so far: they’re called variables. We choose not to call them that because the term “variable” is semantically
charged: it implies that the value associated with the identifier can change (vary). Since our language ini-
tially won’t offer any way of changing the associated value, we use the more conservative term “identifier”.
For now, they are therefore just names for computed constants.
Let’s first write a few sample programs that use identifiers, inventing notation as we go along:
{with {x {+ 5 5}} {+ x x}}
We will want this to evaluate to 20. Here’s a more elaborate example:
{with {x {+ 5 5}}
{with {y {- x 3}}
{+ y y}}} [+ operation]
= {with {x 10} {with {y {- x 3}} {+ y y}}} [substitution]
= {with {x 10} {with {y {- 10 3}} {+ y y}}} [descent]
= {with {y {- 10 3}} {+ y y}} [- operation]
= {with {y 7} {+ y y}} [substitution]
= {with {y 7} {+ 7 7}} [descent]
= {+ 7 7} [+ operation]
= 14
1 As the authors of Concrete Mathematics say: “Name and conquer”.
15
16 CHAPTER 3. SUBSTITUTION
En passant, notice that the act of reducing an expression to a value requires more than just substitution;
indeed, it is an interleaving of substitution and calcuation steps. Furthermore, when we have completed
substition we “descend” into the inner expression to continue calculating.
Now let’s define the language more formally. To honor the addition of identifiers, we’ll give our language
a new name: WAE, short for “with with arithmetic expressions”. Its BNF is:
<WAE> ::= <num>

| {+ <WAE> <WAE>}
| {- <WAE> <WAE>}
| {with {<id> <WAE>} <WAE>}
| <id>
Notice that we’ve had to add two rules to the BNF: one for associating values with identifiers and another for
actually using the identifiers. The nonterminal <id> stands for some suitable syntax for identifiers (usually
a sequence of alphanumeric characters).
To write programs that process WAE terms, we need a data definition to represent those terms. Most
of WAE carries over unchanged from AE, but we must pick some concrete representation for identifiers.
Fortunately, Scheme has a primitive type called the symbol, which serves this role admirably.2 Therefore,
the data definition is
(define-type WAE
[num (n number?)]
[add (lhs WAE?) (rhs WAE?)]
[sub (lhs WAE?) (rhs WAE?)]
[with (name symbol?) (named-expr WAE?) (body WAE?)]
[id (name symbol?)])
We’ll call the expression in the named-expr field the named expression, since with lets the name in the id
field stand in place of that expression.
3.1 Defining Substitution

Without fanfare, we used substitution to explain how with functions. We were able to do this because
substitution is not unique to with: we’ve studied it for years in algebra courses, because that’s what happens
when we pass arguments to functions. For instance, let f (x, y) = x3 + y3 . Then
f (12, 1) = 123 + 13 = 1728 + 1 = 1729
f (10, 9) = 103 + 93 = 1000 + 729 = 1729 3
Nevertheless, it’s a good idea to pin down this operation precisely.

2 In
many languages, a string is a suitable representation for an identifier. Scheme does have strings, but symbols have the
salutary property that they can be compared for equality in constant time.
3 What’s the next smallest such number?
3.1. DEFINING SUBSTITUTION 17
Let’s make sure we understand what we’re trying to define. We want a crisp description of the process
of substitution, namely what happens when we replace an identifier (such as x or x) with a value (such as 12
or 5) in an expression (such as x3 + y3 or {+ x x}).
Recall from the sequence of reductions above that substitution is a part of, but not the same as, cal-
culating an answer for an expression that has identifiers. Looking back at the sequence of steps in the
evaluation example above, some of them invoke substitution while the rest are calcuation as defined for AE.
For now, we’re first going to pin down substitution. Once we’ve done that, we’ll revisit the related question
of calculation. But it’ll take us a few tries to get substitution right!
Definition 1 (Substitution) To substitute identifier i in e with expression v, replace all identifiers in e that
have the name i with the expression v.
Beginning with the program
{with {x 5} {+ x x}}
we will use substitution to replace the identifier x with the expression it is bound to, 5. The definition of
substitution above certainly does the trick: after substitution, we get
{with {x 5} {+ 5 5}}
as we would want. Likewise, it correctly substitutes when there are no instances of the identifier:
{with {x 5} {+ 10 4}}
to
{with {x 5} {+ 10 4}}
(since there are no instances of x in the expression, no substitutions occur). Now consider:
{with {x 5} {+ x {with {x 3} 10}}}
The rules reduce this to
{with {x 5} {+ 5 {with {5 3} 10}}}
Huh? Our substitution rule converted a perfectly reasonable program (whose value is 15) into one that isn’t
even syntactically legal, i.e., it would be rejected by a parser, because the program contains 5 where the BNF
tells us to expect an identifier. We definitely don’t want substitution to have such an effect! It’s obvious that
the substitution algorithm is too naı̈ve. To state the problem with the algorithm precisely, though, we need
to introduce a little terminology.
Definition 2 (Binding Instance) A binding instance of an identifier is the instance of the identifier that
gives it its value. In WAE, the <id> position of a with is the only binding instance.
Definition 3 (Scope) The scope of a binding instance is the region of program text in which instances of the
identifier refer to the value bound by the binding instance.
Definition 4 (Bound Instance) An identifier is bound if it is contained within the scope of a binding in-
stance of its name.
Definition 5 (Free Instance) An identifier not contained in the scope of any binding instance of its name is
said to be free.
With this terminology in hand, we can now state the problem with the first definition of substitution
more precisely: it failed to distinguish between bound instances (which should be substituted) and binding
instances (which should not). This leads to a refined notion of substitution.
Definition 6 (Substitution, take 2) To substitute identifier i in e with expression v, replace all identifiers in
e which are not binding instances that have the name i with the expression v.
A quick check reveals that this doesn’t affect the outcome of the examples that the previous definition
substituted correctly. In addition, this definition of substitution reduces
{with {x 5} {+ x {with {x 3} 10}}}
to
{with {x 5} {+ 5 {with {x 3} 10}}}
Let’s consider a closely related expression:
{with {x 5} {+ x {with {x 3} x}}}
Hopefully we can agree that the value of this program is 8 (the left x in the addition evaluates to 5, the right
x is given the value 3 by the inner with, so the sum is 8). The refined substitution algorithm, however,
converts this expression into
{with {x 5} {+ 5 {with {x 3} 5}}}
which, when evaluated, yields 10.

What went wrong here? Our substitution algorithm respected binding instances, but not their scope.
In the sample expression, the with introduces a new scope for the inner x. The scope of the outer x
is shadowed or masked by the inner binding. Because substitution doesn’t recognize this possibility, it
incorrectly substitutes the inner x.
Definition 7 (Substitution, take 3) To substitute identifier i in e with expression v, replace all non-binding
identifiers in e having the name i with the expression v, unless the identifier is in a scope different from that
introduced by i.
3.1. DEFINING SUBSTITUTION 19
While this rule avoids the faulty substitution we’ve discussed earlier, it has the following effect: after
substitution, the expression
{with {x 5} {+ x {with {y 3} x}}}
whose value should be that of {+ 5 5}, or 10, becomes
{with {x 5} {+ 5 {with {y 3} x}}}
The inner expression should result in an error, because x has no value. Once again, substitution has changed
a correct program into an incorrect one!
Let’s understand what went wrong. Why didn’t we substitute the inner x? Substitution halts at the with
because, by definition, every with introduces a new scope, which we said should delimit substitution. But
this with contains an instance of x, which we very much want substituted! So which is it—substitute
within nested scopes or not? Actually, the two examples above should reveal that our latest definition for
substitution, which may have seemed sensible at first blush, is too draconian: it rules out substitution within
any nested scopes.
Definition 8 (Substitution, take 4) To substitute identifier i in e with expression v, replace all non-binding
identifiers in e having the name i with the expression v, except within nested scopes of i.
Finally, we have a version of substitution that works. A different, more succint way of phrasing this
definition is
Definition 9 (Substitution, take 5) To substitute identifier i in e with expression v, replace all free instances
of i in e with v.
Recall that we’re still defining substitution, not evaluation. Substitution is just an algorithm defined over
expressions, independent of any use in an evaluator. It’s the calculator’s job to invoke substitution as many
times as necessary to reduce a program down to an answer. That is, substitution simply converts
{with {x 5} {+ x {with {y 3} x}}}
into
{with {x 5} {+ 5 {with {y 3} 5}}}
Reducing this to an actual value is the task of the rest of the calculator.
Phew! Just to be sure we understand this, let’s express it in the form of a function.
;; subst : WAE symbol WAE → WAE
;; substitutes second argument with third argument in first argument,
;; as per the rules of substitution; the resulting expression contains
;; no free instances of the second argument
(define (subst expr sub-id val)

(type-case WAE expr
[num (n) expr]

[add (l r) (add (subst l sub-id val)
(subst r sub-id val))]
[sub (l r) (sub (subst l sub-id val)
[with (bound-id named-expr bound-body)
(if (symbol=? bound-id sub-id)
expr
(with bound-id
named-expr
(subst bound-body sub-id val)))]
[id (v) (if (symbol=? v sub-id) val expr)]))
3.2 Calculating with with

We’ve finally defined substitution, but we still haven’t specified how we’ll use it to reduce expressions to
answers. To do this, we must modify our calculator. Specifically, we must add rules for our two new source
language syntactic constructs: with and identifiers.
• To evaluate with expressions, we calculate the named expression, then substitute its value in the
body.
• How about identifiers? Well, any identifier that is in the scope of a with is replaced with a value when
the calculator encounters that identifier’s binding instance. Consequently, the purpose statement of
subst said there would be no free instances of the identifier given as an argument left in the result. In
other words, subst replaces identifiers with values before the calculator ever “sees” them. As a result,
any as-yet-unsubstituted identifier must be free in the whole program. The calculator can’t assign a
value to a free identifier, so it halts with an error.
;; calc : WAE → number

;; evaluates WAE expressions by reducing them to numbers
(define (calc expr)

(type-case WAE expr
[num (n) n]
[sub (l r) (- (calc l) (calc r))]
(calc (subst bound-body
bound-id
(num (calc named-expr))))]
[id (v) (error ’calc ”free identifier”)]))
3.3. THE SCOPE OF WITH EXPRESSIONS 21
Observe that the step we earlier labeled “descend” is handled by the recursive invocation of calc. One
subtlety: In the rule for with, the value returned by calc is a number, but subst is expecting a WAE
expression. Therefore, we wrap the result in (num · · ·) so that substitution will work correctly.
Here are numerous test cases. Each one should pass:
(test (calc (parse ’5)) 5)
(test (calc (parse ’{+ 5 5})) 10)
(test (calc (parse ’{with {x {+ 5 5}} {+ x x}})) 20)
(test (calc (parse ’{with {x 5} {+ x x}})) 10)
(test (calc (parse ’{with {x {+ 5 5}} {with {y {- x 3}} {+ y y}}})) 14)
(test (calc (parse ’{with {x 5} {with {y {- x 3}} {+ y y}}})) 4)
(test (calc (parse ’{with {x 5} {+ x {with {x 3} 10}}})) 15)
(test (calc (parse ’{with {x 5} {+ x {with {x 3} x}}})) 8)
(test (calc (parse ’{with {x 5} {+ x {with {y 3} x}}})) 10)
(test (calc (parse ’{with {x 5} {with {y x} y}})) 5)
(test (calc (parse ’{with {x 5} {with {x x} x}})) 5)
3.3 The Scope of with Expressions

Just when we thought we were done, we find that several of the test cases above (can you determine which
ones?) generate a free-identifier error. What gives?
Consider the program
{with {x 5}
{with {y x}
y}}
Common sense would dictate that its value is 5. So why does the calculator halt with an error on this test
case?
As defined, subst fails to correctly substitute in this program, because we did not account for the named
sub-expressions in with expressions. To fix this problem, we simply need to make subst treat the named
expressions as ordinary expressions, ripe for substitution. To wit:
(type-case WAE expr
[num (n) expr]
expr
(with bound-id
(subst named-expr sub-id val)

The boxed expression shows what changed.
Actually, this isn’t quite right either: consider
{with {x 5}
{with {x x}
x}}
This program should evaluate to 5, but it too halts with an error. This is because we prematurely stopped
substituting for x. We should substitute in the named expression of a with even if the with in question
defines a new scope for the identifier being substituted, because its named expression is still in the scope of
the enclosing binding of the identifier.
We finally get a valid programmatic definition of substitution (relative to the language we have so far):
(type-case WAE expr
[num (n) expr]
(with bound-id
bound-body)
(with bound-id
Observe how the different versions of subst have helped us refine the scope of with expressions. By
focusing on the small handful of lines that change from one version to the next, and studying how they
change, we progressively arrive at a better understanding of scope. This would be much harder to do
through mere prose; indeed, our prose definition has not changed at all through these program changes, but
translating the definition into a program and running it has helped us refine our intutition.
Exercise 3.3.1 What’s the value of
{with {x x} x}
? What should it be, and what does your calculator say it is? (These can be two different things!)
3.4. WHAT KIND OF REDUNDANCY DO IDENTIFIERS ELIMINATE? 23
3.4 What Kind of Redundancy do Identifiers Eliminate?

We began this material motivating the introduction of with: as a means for eliminating redundancy. Let’s
revisit this sequence of substitutions (skipping a few intermediate steps):
{with {x {+ 5 5}} {with {y {- x 3}} {+ y y}}}

= {with {x 10} {with {y {- x 3}} {+ y y}}}
= {with {y {- 10 3}} {+ y y}}
= {with {y 7} {+ y y}}
= {+ 7 7}
= 14
Couldn’t we have also written it this way?
{with {x {+ 5 5}} {with {y {- x 3}} {+ y y}}}

= {with {y {- {+ 5 5} 3}} {+ y y}}
= {+ {- {+ 5 5} 3} {- {+ 5 5} 3}}
= {+ {- 10 3} {- {+ 5 5} 3}}
= {+ {- 10 3} {- 10 3}}
= {+ 7 {- 10 3}}
= {+ 7 7}
= 14
In the first sequence of reductions, we first reduce the named expression to a number, then substitute that
number. In the second sequence, we perform a “textual” substitution, and only when we have no substitu-
tions left to do do we begin to perform the arithmetic.
Notice that this shows there are really two interpretations of “redundancy” in force. One is a purely
static4 notion of redundancy: with exists solely to avoid writing an expression twice, even though it will be
evaluated twice. This is the interpretation taken by the latter sequence of reductions. In contrast, the former
sequence of reductions manifests both static and dynamic5 redundancy elimination: it not only abbreviates
the program, it also avoids re-computing the same value during execution.
Given these two sequences of reductions (which we will call reduction regimes, since each is governed
by a different set of rules), which does our calculator do? Again, it would be hard to reason about this
verbally, but because we’ve written a program, we have a concrete object to study. In particular, the lines
we should focus on are those for with. Here they are again:
...
bound-id
(num (calc named-expr) )))]
...
4 Meaning, referring only to program text.
5 Meaning, referring to program execution.
The boxed portion tells us the answer: we invoke calc before substitution (because the result of calc is what
we supply as an argument to subst). This model of substitution is called eager: we “eagerly” reduce the
named expression to a value before substituting it. This is in contrast to the second sequence of reductions
above, which we call lazy, wherein we reduce the named expression to a value only when we need to (such
as at the application of an arithmetic primitive).
At this point, it may seem like it doesn’t make much difference which reduction regime we employ: both
produce the same answer (though they may take a different number of steps). But do keep this distinction in
mind, for we will see a good deal more on this topic in the course of our study.
Exercise 3.4.1 Can you prove that the eager and lazy regimes will always produce the same answer for the
WAE language?
Exercise 3.4.2 In the example above, the eager regime generated an answer in fewer steps than the lazy
regime did. Either prove that that will always be the case, or provide a counterexample.
Exercise 3.4.3 At the beginning of this section, you’ll find the phrase
an identifier names, or identifies, (the value of) an expression
Why the parenthetical phrase?
3.5 Are Names Necessary?

A lot of the trouble we’ve had with defining substitution is the result of having the same name be bound
multiple times. To remedy this, a computer scientist named Nicolaas de Bruijn had a good idea.6 He asked
the following daring question: Who needs names at all? De Bruijn suggested that instead, we replace
identifiers with numbers. The number dictates how many enclosing scopes away the identifier is bound.
(Technically, we replace identifiers not with numbers but indices that indicate the binding depth. A number
is just a convenient representation for an index. A more pictorially pleasing representation for an index is
an arrow that leads from the bound to the binding instance, like the ones DrScheme’s Check Syntax tool
draws.)
The idea is easy to explain with an example: intead of writing
{with {x 5} {+ x x}}
we would write
{with 5 {+ <0> <0>}}
Notice that two things changed. First, we replaced the bound identifiers with indices (to keep indices sepa-
rate from numbers, we wrap each index in pointy brackets). We’ve adopted the convention that the current
scope is zero levels away. Thus, x becomes <0>. The second change is that, because we no longer care
about the names of identifiers, we no longer need keep track of the x as the bound identifier. The presence
of with indicates that we’ve entered a new scope; that’s enough. Similarly, we convert
6 De Bruijn had many great ideas, particularly in the area of using computers to solve math problems. The idea we present here
was a small offshoot of that much bigger project, but as it so happens, this is the one many people know him for.
3.5. ARE NAMES NECESSARY? 25
{with {x 5}
{with {y 3}
{+ x y}}}
into
{with 5
{with 3
{+ <1> <0>}}}
Let’s consider one last example. If this looks incorrect, that would suggest you may have misunderstood
the scope of a binding. Examining it carefully actually helps to clarify the scope of bindings. We convert
{with {x 5}
{with {y {+ x 3}}
{+ x y}}}
into
{with 5
{with {+ <0> 3}
{+ <1> <0>}}}
De Bruijn indices are useful in many contexts, and indeed the de Bruijn form of a program (that is, a
program where all identifiers have been replaced by their de Bruijn indices) is employed by just about every
compiler. You will sometimes find compiler texts refer to the indices as static distance coordinates. That
name makes sense: the coordinates tell us how far away statically—i.e., in the program text— an identifier
is bound. I prefer to use the less informative but more personal moniker as a form of tribute.
(define-type WAE
[num (n number?)]
[add (lhs WAE?) (rhs WAE?)]
[sub (lhs WAE?) (rhs WAE?)]
[with (name symbol?) (named-expr WAE?) (body WAE?)]
[id (name symbol?)])
;; subst : WAE symbol WAE → WAE

(type-case WAE expr
[num (n) expr]
(with bound-id
bound-body)
(with bound-id
;; calc : WAE → number

(define (calc expr)
(type-case WAE expr
[num (n) n]
[sub (l r) (- (calc l) (calc r))]
bound-id
(num (calc named-expr))))]
[id (v) (error ’calc ”free identifier”)]))
Figure 3.1: Calculator with with

Chapter 4
An Introduction to Functions
Through the agency of with, we have added identifiers and the ability to name expressions to the language.
Much of the time, though, simply being able to name an expression isn’t enough: the expression’s value is
going to depend on the context of its use. That means the expression needs to be parameterized; that is, it
must be a function.
Dissecting a with expression is a useful exercise in helping us design functions. Consider the program
{with {x 5} {+ x 3}}
In this program, the expression {+ x 3} is parameterized over the value of x. In that sense, it’s just like a
function definition: in mathematical notation, we might write
f (x) = x + 3
Having named and defined f , what do we do with it? The WAE program introduces x and then immediately
binds it to 5. The way we bind a function’s argument to a value is to apply it. Thus, it is as if we wrote
f (x) = x + 3; f (5)
In general, functions are useful entities to have in programming languages, and it would be instructive to
model them.
4.1 Enriching the Language with Functions

We will initially model the DrScheme programming environment, which has separate windows for Defini-
tions and Interactions. The Interactions window is DrScheme’s “calculator”, and the part we are trying to
model with our calculators. The contents of the Definitions window are “taught” to this calculator by click-
ing the Run button. Our calculator should therefore consume an argument that reflects these definitions.
To add functions to WAE, we must define their concrete and abstract syntax. In particular, we must both
describe a function definition, or declaration, and provide a means for its use, also known as an application
or invocation. To do the latter, we must add a new kind of expression, resulting in the language F1WAE.1
1 The reason for the “1” will become clear in Section 6.
27
28 CHAPTER 4. AN INTRODUCTION TO FUNCTIONS
We will presume, as a simplification, that functions consume only one argument. This expression language
has the following BNF:
<F1WAE> ::= <num>
| {+ <F1WAE> <F1WAE>}
| {with {<id> <F1WAE>} <F1WAE>}
| <id>
| {<id> <F1WAE>}
(The expression representing the argument supplied to the function is known as the actual parameter.)
We have dropped subtraction from the language on the principle that it is similar enough to addition for
us to determine its implementation from that of addition. To capture this new language, we employ terms of
the following type:
(define-type F1WAE
[num (n number?)]
[add (lhs F1WAE?) (rhs F1WAE?)]
[with (name symbol?) (named-expr F1WAE?) (body F1WAE?)]
[id (name symbol?)]
[app (fun-name symbol?) (arg F1WAE?)])
Convince yourself that this is an appropriate definition.
Now let’s study function definitions. A function definition has three parts: the name of the function,
the names of its arguments, known as the formal parameters, and the function’s body. (The function’s
parameters may have types, which we will discuss in Chapter X.) For now, we will presume that functions
consume only one argument. A simple data definition captures this:
(define-type FunDef
[fundef (fun-name symbol?)
(arg-name symbol?)
(body F1WAE?)])
Using this definition, we can represent a standard function for doubling its argument as
(fundef ’double
’n
(add (id ’n) (id ’n)))
Now we’re ready to write the calculator, which we’ll call interp—short for interpreter—rather than
calc to reflect the fact that our language has grown beyond arithmetic. The interpreter must consume two
arguments: the expression to evaluate, and the set of known function definitions. This corresponds to what
the Interactions window of DrScheme works with. The rules present in the WAE interpreter remain the
same, so we can focus on the one new rule.
;; interp : F1WAE listof(fundef) → number
;; evaluates F1WAE expressions by reducing them to their corresponding values
(define (interp expr fun-defs)

4.2. THE SCOPE OF SUBSTITUTION 29
(type-case F1WAE expr

[num (n) n]
[add (l r) (+ (interp l fun-defs) (interp r fun-defs))]
(interp (subst bound-body
bound-id
(num (interp named-expr fun-defs)))
fun-defs)]
[id (v) (error ’interp ”free identifier”)]
[app (fun-name arg-expr)
(local ([define the-fun-def (lookup-fundef fun-name fun-defs)])
(interp (subst (fundef-body the-fun-def )
(fundef-arg-name the-fun-def )
(num (interp arg-expr fun-defs)))
fun-defs))] ))
The rule for an application first looks up the named function. If this access succeeds, then interpretation
proceeds in the body of the function after first substituting its formal parameter with the (interpreted) value
of the actual parameter. We see the result in DrScheme:
> (interp (parse ’{double {double 5}})
(list (fundef ’double
’n
(add (id ’n) (id ’n)))))
20
To make this interpreter function correctly, we must make several changes. First, we must adapt the parser
to treat the relevant inputs (as identified by the BNF) as function applications. Second, we must modify the
interpreter itself, changing the recursive calls to take an extra argument, and adding the implementation of
app. Third, we must extend subst to handle the F1WAE language. Finally, we must write lookup-fundef ,
the helper routine that finds function definitions. The last two changes are shown in Figure 4.1.
Exercise 4.1.1 Why is the argument expression of an application of type F1WAE rather than of type WAE?
Provide sample programs permitted by the former and rejected by the latter, and argue that these programs
are reasonable.
Exercise 4.1.2 Why is the body expression of a function definition of type F1WAE rather than of type WAE?
Provide a sample definition permitted by using the former rather than the latter, and argue that it is reason-
able.
4.2 The Scope of Substitution

Suppose we ask our interpreter to evaluate the expression
(app ’f (num 10))
in the presence of the solitary function definition

(fundef ’f
’n
(app ’n (id ’n)))
What should happen? Should the interpreter try to substitute the n in the function position of the application
with the number 10, then complain that no such function can be found (or even that function lookup funda-
mentally fails because the names of functions must be identifiers, not numbers)? Or should the interpreter
decide that function names and function arguments live in two separate “spaces”, and context determines in
which space to look up a name? Languages like Scheme take the former approach: the name of a function
can be bound to a value in a local scope, thereby rendering the function inaccessible through that name. This
latter strategy is known as employing namespaces and languages such as Common Lisp adopt it.
4.3 The Scope of Function Definitions

Suppose our definition list contains multiple function definitions. How do these interact with one another?
For instance, suppose we evaluate the following input:
(interp (parse ’{f 5})
(list (fundef ’f ’n (app ’g (add (id ’n) (num 5))))
(fundef ’g ’m (sub (id ’m) (num 1)))))
What does this program do? The main expression applies f to 5. The definition of f, in turn, invokes
function g. Should f be able to invoke g? Should the invocation fail because g is defined after f in the list
of definitions? What if there are multiple bindings for a given function’s name?
We will expect this program to evaluate to 9. We employ the natural interpretation that each function
can “see” every function’s definition, and the natural assumption that each name is bound at most once so
we needn’t disambiguate between definitions. Is is, however, possible to define more sophisticated scopes.
Exercise 4.3.1 If a function can invoke every defined function, that means it can also invoke itself. This is
currently of limited value because the language F1WAE lacks a harmonious way of terminating recursion.
Consider adding a simple conditional construct (such as if0, which succeeds if the term in the first position
evaluates to 0) and writing interesting programs in this language.
4.3. THE SCOPE OF FUNCTION DEFINITIONS 31
(define-type F1WAE
[num (n number?)]
[id (name symbol?)]
(define-type FunDef
(arg-name symbol?)
(body F1WAE?)])
;; lookup-fundef : symbol listof(FunDef) −→ FunDef

(define (lookup-fundef fun-name fundefs)
(cond
[(empty? fundefs) (error fun-name ”function not found”)]
[else (if (symbol=? fun-name (fundef-fun-name (first fundefs)))
(first fundefs)
(lookup-fundef fun-name (rest fundefs)))]))
;; subst : F1WAE symbol F1WAE → F1WAE

[num (n) expr]
(with bound-id
bound-body)
(with bound-id
[id (v) (if (symbol=? v sub-id) val expr)]
(app fun-name (subst arg-expr sub-id val))]))
Figure 4.1: Implementation of Functions: Support Code

;; interp : F1WAE listof(fundef) → number

(define (interp expr fun-defs)
[num (n) n]
[add (l r) (+ (interp l fun-defs) (interp r fun-defs))]
bound-id
(num (interp named-expr fun-defs)))
fun-defs)]
fun-defs))]))
Figure 4.2: Implementation of Functions: Interpreter

Chapter 5
Deferring Substitution
Let’s examine the process of interpreting the following small program. Consider the following sequence of
evaluation steps:
{with {x 3}
{with {y 4}
{with {z 5}
{+ x {+ y z}}}}}
= {with {y 4}
{with {z 5}
{+ 3 {+ y z}}}}
= {with {z 5}
{+ 3 {+ 4 z}}}
= {+ 3 {+ 4 5}}
at which point it reduces to an arithmetic problem. To reduce it, though, the interpreter had to apply substi-
tution three times: once for each with. This is slow! How slow? Well, if the program has size n (measured
in abstract syntax tree nodes), then each substitution sweeps the rest of the program once, making the com-
plexity of this interpreter at least O(n2 ). That seems rather wasteful; surely we can do better.
How do we avoid this computational redundancy? We should create and use a repository of deferred
substitutions. Concretely, here’s the idea. Initially, we have no substitutions to perform, so the repository
is empty. Every time we encounter a substitution (in the form of a with or application), we augment the
repository with one more entry, recording the identifier’s name and the value (if eager) or expression (if
lazy) it should eventually be substituted with. We continue to evaluate without actually performing the
substitution.
This strategy breaks a key invariant we had established earlier, which is that any identifier the interpreter
encounters is of necessity free, for had it been bound, it would have been replaced by substitution. Because
we’re no longer using substitution, we will encounter bound identifiers during interpretation. How do we
handle them? We must substitute them with by consulting the repository.
Exercise 5.0.2 Can the complexity of substitution be worse than O(n2 )?
33
34 CHAPTER 5. DEFERRING SUBSTITUTION
5.1 The Substitution Repository

Let’s provide a data definition for the repository:
(define-type DefrdSub
[mtSub]
[aSub (name symbol?) (value number?) (ds DefrdSub?)])
where DefrdSub stands for “deferred substitutions”. A DefrdSub has two forms: it’s either empty (mtSub1 )
or non-empty (represented by an aSub structure). The latter contains a reference to the rest of the repository
in its third field.
The interpreter obviously needs to consume a repository in addition to the expression to interpret. There-
fore, its contract becomes
;; interp : F1WAE listof(fundef) DefrdSub → number
It will need a helper function that looks up the value of identifiers in the repository. Its code is:
;; lookup : symbol DefrdSub → F1WAE
(define (lookup name ds)

(type-case DefrdSub ds
[mtSub () (error ’lookup ”no binding for identifier”)]
[aSub (bound-name bound-value rest-ds)
(if (symbol=? bound-name name)
bound-value
(lookup name rest-ds))]))
With that introduction, we can now present the interpreter:
(define (interp expr fun-defs ds)
[num (n) n]
[add (l r) (+ (interp l fun-defs ds) (interp r fun-defs ds))]
(interp bound-body
fun-defs
(aSub bound-id
(interp named-expr
fun-defs
ds)
ds))]
[id (v) (lookup v ds)]
1 “Empty sub”—get it?
5.2. DEFERRING SUBSTITUTION CORRECTLY 35
(interp (fundef-body the-fun-def )

fun-defs
(aSub (fundef-arg-name the-fun-def )
(interp arg-expr fun-defs ds)
ds)))]))
Three clauses have changed: those for with, identifiers and applications. Applications must look up the
value of an identifier in the repository. The rule for with evaluates the body in a repository that extends
the current one (ds) with a binding for the with-bound identifier to its interpreted value. The rule for
an application similarly evaluates the body of the function with the repository extended with the formal
argument bound to the value of the actual argument.
To make sure this is correct, we recommend that you first study its behavior on programs that have no
identifiers—i.e., verify that the arithmetic rules do the right thing—and only then proceed to the rules that
involve identifiers.
5.2 Deferring Substitution Correctly

Consider the evaluation of the expression
{with {n 5}
{f 10}}
in the following list of function definitions:

(list (fundef ’f ’p (id ’n)))
That is, f consumes an argument p and returns the value bound to n. This corresponds to the Scheme
definition
(define (f p) n)
followed by the application
(local ([define n 5])
(f 10))
What result does Scheme produce?
Our interpreter produces the value 5. Is this the correct answer? Well, it’s certainly possible that this is
correct—after all, it’s what the interpreter returns, and this could well be the interpreter for some language.
But we do have a better way of answering this question.
Recall that the interpreter was using the repository to conduct substitution more efficiently. We hope that
that’s all it does—that is, it must not also change the meaning of a program! Our “reference implementation”
is the one that performs explicit substitution. If we want to know what the value of the program really “is”,
we need to return to that implementation.
What does the substitution-based interpreter return for this program? It says the program has an unbound
identifier (specifically, n). So we have to regard our interpreter with deferred substitutions as buggy.
While this new interpreter is clearly buggy relative to substitution, which it was supposed to represent,
let’s think for a moment about what we, as the human programmer, would want this program to evaluate to.
It produces the value 5 because the identifier n gets the value it was bound to by the with expression, that
is, from the scope in which the function f is used. Is this a reasonable way for the language to behave? A
priori, is one interpretation better than the other? Before we tackle that, let’s introduce some terminology to
make it easier to refer to these two behaviors.
Definition 10 (Static Scope) In a language with static scope, the scope of an identifier’s binding is a syn-
tactically delimited region.
A typical region would be the body of a function or other binding construct. In contrast:
Definition 11 (Dynamic Scope) In a language with dynamic scope, the scope of an identifier’s binding is
the entire remainder of the execution during which that binding is in effect.
That is, in a language with dynamic scope, if a function g binds identifier n and then invokes f, then f can
refer to n—and so can every other function invoked by g until it completes its execution—even though f
has no locally visible binding for n.
Armed with this terminology, we claim that dynamic scope is entirely unreasonable. The problem is
that we simply cannot determine what the value of a program will be without knowing everything about its
execution history. If the function f were invoked by some other sequence of functions that did not bind a
value for n, then that particular application of f would result in an error, even though a previous application
of f in the very same program’s execution completed successfully! In other words, simply by looking at the
source text of f, it would be impossible to determine one of the most rudimentary properties of a program:
whether or not a given identifier was bound. You can only imagine the mayhem this would cause in a large
software system, especially with multiple developers and complex flows of control. We will therefore regard
dynamic scope as an error and reject its use in the remainder of this text.
5.3 Fixing the Interpreter

Let’s return to our interpreter. Our choice of static over dynamic scope has the benefit of confirming that
the substituting interpreter did the right thing, so all we need do is make the new interpeter be a correct
reimplementation of it. We only need to focus our attention on one rule, that for function application. This
currently reads:
fun-defs
ds)))]
5.3. FIXING THE INTERPRETER 37
When the interpreter evaluates the body of the function definition, what deferred substitutions does it rec-
ognize? It recognizes all those already are in ds, with one more for the function’s formal parameter to
be replaced with the value of the actual parameter. But how many substitutions should be in effect in the
function’s body? In our substitution-based interpreter, we implemented application as follows:
fun-defs))]
This performs only one substitution on the function’s body: the formal parameter for its value. The code
demonstrates that none of the substitutions applied to the calling function are in effect in the body of the
called function. Therefore, at the point of invoking a function, our new interpeter must “forget” about the
current substitutions. Put differently, at the beginning of every function’s body, there is only one bound
identifier—the function’s formal parameter—independent of the invoking context.
How do we fix our implementation? We clearly need to create a substitution for the formal parameter
(which we obtain using the expression (fundef-arg-name the-fun-def )). But the remaining substitutions must
be empty, so as to not pick up the bindings of the calling context. Thus,
fun-defs
(mtSub) )))]
That is, we use the empty repository to initiate evaluation of a function’s body, extending it immediately with
the formal parameter but no more. The difference between using ds and (mtSub) in the position of the box
succintly captures the implementation distinction between dynamic and static scope, respectively—though
the consequences of that distinction are far more profound that this small code change might suggest.
Exercise 5.3.1 How come we never seem to “undo” additions to the repository? Doesn’t this run the risk
that one substitution might override another in a way that destroys static scoping?
Exercise 5.3.2 Why is the last ds in the interpretation of with also not replaced with (mtSub)? What
would happen if we were to effect this replacement? Write a program that illustrates the difference, and
argue whether the replacement is sound or not.
Exercise 5.3.3 Our implementation of lookup can take time linear in the size of the program to find some
identifiers. Therefore, it’s not clear we have really solved the time-complexity problem that motivated the
use of a substitution repository. We could address this by using a better data structure and algorithm for
lookup: a hash table, say. What changes do we need to make if we use a hash table?
Hint: This is tied closely to Exercise 5.3.1!
(define-type F1WAE
[num (n number?)]
[id (name symbol?)]
(define-type FunDef
(arg-name symbol?)
(body F1WAE?)])
[mtSub]
[aSub (name symbol?) (value number?) (ds DefrdSub?)])
;; lookup-fundef : symbol listof(FunDef) −→ FunDef

(define (lookup-fundef fun-name fundefs)
(cond
[(empty? fundefs) (error fun-name ”function not found”)]
[else (if (symbol=? fun-name (fundef-fun-name (first fundefs)))
(first fundefs)
(lookup-fundef fun-name (rest fundefs)))]))
;; lookup : symbol DefrdSub → F1WAE

bound-value
Figure 5.1: Functions with Deferred Substitutions: Support Code

5.3. FIXING THE INTERPRETER 39
;; interp : F1WAE listof(fundef) DefrdSub → number

(define (interp expr fun-defs ds)
[num (n) n]
[add (l r) (+ (interp l fun-defs ds) (interp r fun-defs ds))]
(interp bound-body
fun-defs
(aSub bound-id
(interp named-expr
fun-defs
ds)
ds))]
fun-defs
(mtSub))))]))
Figure 5.2: Functions with Deferred Substitutions: Interpreter

Chapter 6
First-Class Functions
We began Section 4 by observing the similarity between a with expression and a function definition applied
immediately to a value. Specifically, we observed that
{with {x 5} {+ x 3}}
is essentially the same as

f (x) = x + 3; f (5)
Actually, that’s not quite right: in the math equation above, we give the function a name, f , whereas there is
no identifier named f anywhere in the WAE program. We can, however, rewrite the mathematical formula-
tion as
f = λ (x).x + 3; f (5)
which we can then rewrite as
(λ (x)x + 3)(5)
to get rid of the unnecessary name ( f ).
Notice, however, that our langugae F1WAE does not permit anonymous functions of the style we’ve
used above. Because such functions are useful in their own right, we now extend our study of functions.
6.1 A Taxonomy of Functions

The translation of with into mathematical notation exploits two features of functions: the ability to create
anonymous functions, and the ability to define functions anywhere in the program (in this case, in the func-
tion position of an application). Not every programming language offers one or both of these capabilities.
There is, therefore, a taxonomy that governs these different features, which we can use when discussing
what kind of functions a language provides:
first-order Functions are not values in the language. They can only be defined in a designated portion of
the program, where they must be given names for use in the remainder of the program. The functions
in F1WAE are of this nature, which explains the 1 in the name of the language.
41
42 CHAPTER 6. FIRST-CLASS FUNCTIONS
higher-order Functions can return other functions as values.
first-class Functions are values with all the rights of other values. In particular, they can be supplied as the
value of arguments to functions, returned by functions as answers, and stored in data structures.
We would like to extend F1WAE to have the full power of functions, to reflect the capability of Scheme. In
fact, it will be easier to return to WAE and extend it with first-class functions.
6.2 Enriching the Language with Functions

To add functions to WAE, we must define their concrete and abstract syntax. First let’s examine some
concrete programs:
{{fun {x} {+ x 4}}
5}
This program defines a function that adds 4 to its argument and immediately applies this function to 5,
resulting in the value 9. This one
{with {double {fun {x} {+ x x}}}
{+ {double 10}
{double 5}}}
evaluates to 30. The program defines a function, binds it to double, then uses that name twice in slightly
different contexts (i.e., it instantiates the formal parameter with different actual parameters).
From these examples, it should be clear that we must introduce two new kinds of expressions: function
applications (as before), as well as (anonymous) function definitions. Here’s the revised BNF corresponding
to these examples:
<FWAE> ::= <num>
| {+ <FWAE> <FWAE>}
| {with {<id> <FWAE>} <FWAE>}
| <id>
| {fun {<id>} <FWAE>}
| {<FWAE> <FWAE>}
Note that F1WAE did not have function definitions as part of the expression language, since the definitions
were assumed to reside outside the expression being evaluated. In this language, functions can be defined
anywhere, including in the function position of an application (the last BNF production): instead of just
the name of a function, programmers can write an arbitrary expression that must be evaluated to obtain the
function to apply. The corresponding abstract syntax is:
(define-type FWAE
[num (n number?)]
[add (lhs FWAE?) (rhs FWAE?)]
[with (name symbol?) (named-expr FWAE?) (body FWAE?)]
6.2. ENRICHING THE LANGUAGE WITH FUNCTIONS 43
[id (name symbol?)]

[fun (param symbol?) (body FWAE?)]
[app (fun-expr FWAE?) (arg-expr FWAE?)])
To define our interpreter, we must think a little about what kinds of values it consumes and produces.
Naturally, the interpreter consumes values of type FWAE. What does it produce? Clearly, a program that
meets the WAE description must yield numbers. As we have seen above, some program that use functions
and applications also evaluate to numbers. How about a program that consists solely of a function? That is,
what is the value of the program
{fun {x} x}
? It clearly doesn’t represent a number. It may be a function that, when applied to a numeric argument,
produces a number, but it’s not itself a number (if you think differently, you need to indicate which number
it will be: 0? 1? 1729?). We instead realize from this that functions are also values that may be the result of
a computation.
We could design an elaborate representation for function values, but for now, we’ll remain modest. We’ll
let the function evaluate to its abstract syntax representation (i.e., a fun structure). (We will soon get more
sophisticated than this.) For consistency, we’ll also let numbers evaluate to num structures. Thus, the result
of evaluating the program above would be the value
(fun ’x (id ’x))
Now we’re ready to write the interpreter. We must pick a type for the value that interp returns. Since
we’ve decided to represent function and number answers using the abstract syntax, it makes sense to use
FWAE, with the caveat that only two kinds of FWAE terms can appear in the output: numbers and functions.
Our first interpreter will use explicit substitution, to offer a direct comparison with the corresponding WAE
and F1WAE interpreters.
;; interp : FWAE → FWAE

;; evaluates FWAE expressions by reducing them to their corresponding values
;; return values are either num or fun
(define (interp expr)

(type-case FWAE expr
[num (n) expr]
[add (l r) (add-numbers (interp l) (interp r))]
bound-id
(interp named-expr)))]
[fun (bound-id bound-body)

expr]
[app (fun-expr arg-expr)
(local ([define fun-val (interp fun-expr)])
(interp (subst (fun-body fun-val)
(fun-param fun-val)
(interp arg-expr))))] ))
(We’ve made a small change to the rule for add: it uses a helper named add-numbers because interp now
returns an FWAE. As an exercise, define this helper function for yourself.)
The rule for a function says, simply, to return the function itself. (Notice the similarity to the rule for
numbers, the other kind of constant in our language!) That leaves only the rule for applications to study.
This rule first evaluates the function position of an application. This is because that position may itself
contain a complex expression that needs to be reduced to an actual function. For instance, in the expression
{{{fun {x} x}
{fun {x} {+ x 5}}}
3}
the outer function position consists of the application of the identity function to a function that adds five to
its argument.
When evaluated, the function position had better reduce to a function value, not a number (or anything
else). For now, we implicitly assume that the programs fed to the interpreter have no errors. (Chapter X will
study how we can detect erroneous programs.) Given a function, we need to evaluate its body after having
substituted the formal argument with its actual value. That’s what the rest of the program does: evaluate
the expression that will become the bound value, bind it using substitution, and then interpet the resulting
expression. The last few lines are very similar to the code for with.
To understand this interpreter better, consider what it produces in response to evaluating the following
term:
{with {x 3}
{fun {y}
{+ x y}}}
DrScheme prints the following:
(fun ’y (add (num 3) (id ’y)))
Notice that the x inside the function body has been replaced by 3 as a result of substitution, so the function
has no references to x left in it.
Exercise 6.2.1 What induced the small change in the rule for add? Explain, with an example, what would
go wrong if we did not make this change.
Exercise 6.2.2 Did you notice the small change in the interpretation of with?
Exercise 6.2.3 What goes wrong if the interpreter fails to evaluate the function position (by invoking the
interpreter on it)? Write a program and present the expected and actual results.
6.3. MAKING WITH REDUNDANT 45
6.3 Making with Redundant

Now that we have functions and function invocation as two distinct primitives, we can combine them to
recover the behavior of with as a special case. Every time we encounter an expression of the form
{with {var named} body}
we can replace it with
{{fun {var} body}

named}
and obtain the same effect. The result of this translation does not use with, so it can be evaluated by a more
primitive interpreter: one for AE enriched with functions.
Exercise 6.3.1 Implement a pre-processor that performs this translation.
We will assume the existence of such a pre-processor, and use the language FAE as our basis for subse-
quent exploration. Section 36 discusses such pre-processors in great detail.
6.4 Implementing Functions using Deferred Substitutions

As Section 5 described, our implementation will be more sprightly if we deferred substitutions instead of
performing them greedily. Let’s study how to adapt our interpreter to use this more efficient strategy.
First, we must provide a definition of the substitution repository. The repository associates identifiers
with their values. Previously, the value had always been a number, but now our set of values is richer. We
therefore use the following type, with the understanding that the value will always be a num or fun:
[mtSub]
[aSub (name symbol?) (value FAE?) (ds DefrdSub?)])
Relative to this, the definition of lookup remains the same (only it now returns values of type FAE).
Our first attempt at adapting the interpreter is
(define (interp expr ds)
(type-case FAE expr
[num (n) expr]
[add (l r) (add-numbers (interp l ds) (interp r ds))]
expr]
(local ([define fun-val (interp fun-expr ds)])
(interp (fun-body fun-val)
(aSub (fun-param fun-val)
(interp arg-expr ds)

ds)))]))
When we run a battery of tests on this interpreter, we find that the expression
{with {x 3}
{with {f {fun {y} {+ x y}}}
{with {x 5}
{f 4}}}}
evaluates to 9. This should be surprising, because we seem to again have introduced dynamic scope! (Notice
that the value of x depends on the context of the application of f, not its definition.)
To understand the problem better, let’s return to this example, which we examined in the context of the
substitution interpreter: the result of interpreting
{with {x 3}
{fun {y}
{+ x y}}}
in the substitution interpreter is

That is, it had substituted the x with 3 in the procedure. Now we are deferring substitutions, but we haven’t
modified how the interpreter evaluates a function definition. As a result, the above expression evaluates to
(fun ’y (add (id ’x) (id ’y)))
But what happened to the substitutions deferred in its body?
The moral is, to properly defer substitution, the value of a function should be not only its text, but also
the substitutions that were due to be performed on it. We therefore define a new datatype for the interpreter’s
return value that attaches the definition-time repository to every function value:
(define-type FAE-Value
[numV (n number?)]
[closureV (param symbol?)
(body FAE?)
(ds DefrdSub?)])
Accordingly, we change the rule for fun in the interpreter to
(closureV bound-id bound-body ds)]
We call this constructed value a closure because it “closes” the function body over the substitutions that are
waiting to occur.
When the interpreter encounters a function application, it must ensure that the function’s pending substi-
tutions aren’t forgotten. It must, however, ignore the substitutions pending at the location of the invocation,
for that is precisely what led us to dynamic instead of static scope. It must instead use the substitutions of the
6.4. IMPLEMENTING FUNCTIONS USING DEFERRED SUBSTITUTIONS 47
invocation location to convert the function and argument into values, hope that the function expression eval-
uated to a closure, then proceed with evaluating the body employing the repository of deferred substitutions
stored in the closure.
(interp (closureV-body fun-val)
(aSub (closureV-param fun-val)
(closureV-ds fun-val) )))]
That is, having evaluated fun-expr to yield fun-val, we obtain not only the actual function body from fun-
val’s closure record but also the repository stored within it. Crucially, while we evaluate arg-expr in ds, the
repository active at the invocation location, we evaluate the function’s body in its “remembered” repository.
Once again, the content of this boxed expression determines the difference between static and dynamic
scope. Figure 6.2 presents the complete interpreter.
It is worth looking back at the interpreter in Figure 5.2. Observe that in the application rule, the reposi-
tory used (in place of the boxed expression above) is (mtSub). This is just the special case when functions
cannot be nested within one another (so there are no deferred substitutions to record). The C programming
language implements a middle-ground by allowing functions (technically, pointers to functions) to serve as
values but allowing them to be defined only at the top-level. Because they are defined at the top-level, they
have no deferred substitutions at the point of declaration; this means the second field of a closure record
is always empty, and can therefore be elided, making the function pointer a complete description of the
function. This purchases some small implementation efficiency, at the cost of potentially great expressive
power, as the examples in Section 6.5 illustrate.
Returning to the interpreter above, the input
{with {x 3}
{fun {y}
{+ x y}}}
results in the following output:
(closureV ’y
(add (id ’x) (id ’y))
(aSub ’x (numV 3) (mtSub)))
That is, the closure has the same effect as the output under substitution, namely
except that the deferred substitution of 3 for x has been recorded in the closure’s repository.
Exercise 6.4.1 This interpreter does not check whether the function position evaluates to a closure value.
Modify the interpreter to check and, if the expression fails to yield a closure, report an error.
Exercise 6.4.2 Suppose we explicitly implemented with in the interpreter. Given that with is just a
shorthand for creating and applying a closure, would the changes we made to closure creation and function
application have an effect on with too?
6.5 Some Perspective on Scope

The example above that demonstrated the problem with our caching interpreter might not be a very convinc-
ing demonstration of the value of static scope. Indeed, you might be tempted to say, “If you know x is 3,
why not just use 3 instead of x inside the procedure declaration? That would avoid this problem entirely!”
That’s a legitimate response for that particular example, which was however meant only to demonstrate the
problem, not to motivate the need for the solution. Lets’s now consider three examples that do the latter.
6.5.1 Filtering and Sorting Lists

In the process of defining a sort routine like quicksort, we would want to define a helper procedure that
filters all the elements less than the pivot:
;; filter-<-pivot: number list(number) → list(number)
(define (filter-<-pivot pivot l)
(cond
[(empty? l) empty]
[(cons? l)
(if (< (first l) pivot)
(cons (first l) (filter-<-pivot pivot (rest l)))
(filter-<-pivot pivot (rest l)))]))
Unfortunately, we need another procedure to filter all the elements greater than the pivot as well. It is poor
programming practice to copy this code when we should, instead, be abstracting over it; furthermore, this
code assumes that we will only want to sort lists of numbers (because it uses <, which compares only
numbers), whereas we will want to sort lots of other types of lists, too. Therefore, a more sensible filtering
routine would be this:
;; filter-pivot: X (X X → boolean) list(X) → list(X)
(define (filter-pivot pivot comp l)
(cond
[(empty? l) empty]
[(cons? l)
(if (comp (first l) pivot)
(cons (first l) (filter-pivot pivot comp (rest l)))
(filter-pivot pivot comp (rest l)))]))
We can then pass different procedures, such as <, >=, string<=?, and so on as the second argument,
thereby using this procedure on lists of any comparable type.
While this is superior to the previous filter procedure, it is somehow unsatisfying because it does not
properly reflect the essence of filtering. The essence, we would think, is a procedure with the following
type:
;; filter-any: (X → boolean) list(X) → list(X)
6.5. SOME PERSPECTIVE ON SCOPE 49
That is, given a predicate that determines whether or not to keep an element in a list, and given a list, the
filter procedure returns a list of just those elements that were chosen to be kept. The implementation of this
is straightforward:
(define (filter-any comp l)
(cond
[(empty? l) empty]
[(cons? l)
(if (comp (first l))
(cons (first l) (filter-any comp (rest l)))
(filter-any comp (rest l)))]))
But now, consider the use of this procedure in a context such as quicksort: the comparator argument must
close over the pivot. That is, the sorting routine has this form:
(define (qs comp l)
(cond
[(empty? l) empty]
[(cons? l)
(local ([define pivot (first l)])
(append
(qs comp (filter-any (lambda (x) (comp x pivot)) (rest l)))
(list pivot)
(qs comp (filter-any (lambda (x) (not (comp x pivot))) (rest l)))))]))
In a language with closures, it is entirely natural to write this program. If, however, the language did not
have static scope, how would we be able to define the procedure given as an argument to filter-any? Examine
the definition of sorting procedures in libraries for languages like C.
6.5.2 Differentiation
Let’s implement (a simple version of) numeric differentiation in Scheme. The program is
(define H 0.0001)
(define (d/dx f )
(lambda (x)
(/ (- (f (+ x H)) (f x))
H)))
In this example, in the algebraic expression, the identifier f is free relative to the inner function. However,
we cannot do what we proposed earlier, namely to substitute the free variable with its value; this is because
we don’t know what values f will hold during execution, and in particular f will likely be bound to several
different values over the course of the program’s lifetime. If we run the inner procedure under dynamic
scope, one of two things will happen: either the identifier f will not be bound to any value in the context of
use, resulting in an unbound identifier error, or the procedure will use whatever f is bound to, which almost
certainly will not correspond to the value supplied to d/dx. That is, in a hypothetical dynamically-scoped
Scheme, you would get
> (define diff-of-square (d/dx (lambda (x) (∗ x x))))
> (diff-of-square 10)
reference to undefined identifier: f
> (define f ’greg)
procedure application: expected procedure, given: greg; arguments were: 10.0001
> (define f sqrt)
0.15811348772487577
That is, f assumes whatever value it has at the point of use, ignoring the value given to it at the inner
procedure’s definition. In contrast, what we really get from Scheme is
20.000099999890608 ;; approximately 10 × 2 = 20
6.5.3 Callbacks
Let’s consider another example, this one from Java. This program implements a callback, which is a com-
mon programming pattern employed in programming GUIs. In this instance, the callback is an object in-
stalled in a button; when the user presses the button, the GUI system invokes the callback, which brings up
a message box displaying the number of times the user has pressed the button. This powerful paradigm lets
the designer of the GUI system provide a generic library of graphical objects independent of the behavior
each client wants to associate with that object.
// GUI library code

public class JButton {
public void whenPressed(ActionEvent e) {
for (int i = 0; i < listeners.length; ++i)
listeners[i].actionPerformed(e);
}
}
// User customization
public class GUIApp {
private int count = 0;
public class ButtonCallback implements ActionListener {

public void actionPerformed(ActionEvent e) {
count = count + 1;
JOptionPane.showMessageDialog(null,
6.5. SOME PERSPECTIVE ON SCOPE 51
"Callback was invoked " +

count + " times!");
}
}
public Component createComponents() {

JButton button = new JButton("Click me!");
button.addActionListener(new ButtonCallback());
return button;
}
}
Stripped to its essence, the callback code is really no different from

;; GUI library code
(define (button callback)
(local [(define (sleep-loop)
(when button-pressed
(begin
(callback)
(sleep-loop))))]
(sleep-loop)))
;; User customization
(local [(define count 0)
(define (my-callback)
(begin
(set! count (add1 count)) ;; increment counter
(message-box
(string-append ”Callback was invoked ”
(number→string count)
” times!”))))]
(button my-callback))
That is, a callback is just a function passed to the GUI toolbox, which the toolbox invokes when it has an
argument. But note that in the definition of my-callback (or ButtonCallback), the identifier count is not
bound within the function (or object) itself. That is, it is free in the function. Therefore, whether it is scoped
statically or dynamically makes a huge difference!
How do we want our callback to behave? Naturally, as the users of the GUI toolbox, we would be very
upset if, the first time the user clicked on the button, the system halted with the message
error: identifier ‘count’ not bound
The bigger picture is this. As programmers, we hope that other people will use our functions, perhaps
even in fantastic contexts that we cannot even imagine. Unfortunately, that means we can’t possibly know
what the values of identifiers will be at the location of use, or whether they will even be bound. If we must
rely on the locus of use, we will produce highly fragile programs: they will be useful only in very limited
contexts, and their behavior will be unpredictable everywhere else.
Static scoping avoids this fear. In a language with static scope, the programmer has full power over
choosing from the definition and use scopes. By default, free identifiers get their values from the definition
scope. If the programmer wants to rely on a value from the use scope, they simply make the corresponding
identifier a parameter. This has the added advantage of making very explicit in the function’s interface which
values from the use scope it relies on.
Dynamic scoping is primarily interesting as a historical mistake: it was in the earliest versions of Lisp,
and persisted for well over a decade. Scheme was created as an experimental language in part to experiment
with static scope. This was such a good idea that eventually, even Common Lisp adopted static scope.
Most modern languages are statically scoped, but sometimes they make the mistake of recapitulating this
phylogeny. So-called “scripting” languages, in particular, often make the mistake of implementing dynamic
scope (or the lesser mistake of just failing to create closures), and must go through multiple iterations before
they eventually implement static scope correctly.
6.6 Eagerness and Laziness

Recall that a lazy evaluator was one that did not reduce the named-expression of a with to a value at the
time of binding it to an identifier. What is the corresponding notion of laziness in the presence of functions?
Let’s look at an example: in a lazy evaluator,
{with {x {+ 3 3}}
{+ x x}}
would first reduce to
{+ {+ 3 3} {+ 3 3}}
But based on what we’ve just said in section 6.3 about reducing with to procedure application, the treatment
of procedure arguments should match that of the named expression in a with. Therefore, a lazy language
with procedures is one that does not reduce its argument to a value until necessary in the body. The following
sequence of reduction rules illustrates this:
{{fun {x} {+ x x}}

{+ 3 3}}
= {+ {+ 3 3} {+ 3 3}}
= {+ 6 {+ 3 3}}
= {+ 6 6}
= 12
which is just an example of the with translation described above; a slightly more complex example is

{double {double 5}}}
6.7. STANDARDIZING TERMINOLOGY 53
= {{fun {x} {+ x x}}

{{fun {x} {+ x x}}
5}}
= {{fun {x} {+ x x}}
{+ 5 5}}
= {+ {+ 5 5} {+ 5 5}}
= {+ 10 {+ 5 5}}
= {+ 10 10}
= 20
What do the corresponding reductions look like in an eager regime? Are there significant differences be-
tween the two?
Exercise 6.6.1 Modify the interpreter with deferred substitution to handle a lazy language with first-class
functions.
6.7 Standardizing Terminology

We have thus far been using the terms “deferred substitution” and “repository”. There is actually a standard
term used for these, and we’ll adopt this from now on.
Definition 12 (environment) An environment1 is a repository of deferred substitutions.
1 Cormac Flanagan: “Save the environment! Create a closure today!”

(define-type FAE
[num (n number?)]
[add (lhs FAE?) (rhs FAE?)]
[id (name symbol?)]
[fun (param symbol?) (body FAE?)]
[app (fun-expr FAE?) (arg-expr FAE?)])
[numV (n number?)]
(body FAE?)
(ds DefrdSub?)])
[mtSub]
[aSub (name symbol?) (value FAE-Value?) (ds DefrdSub?)])
;; lookup : symbol DefrdSub → FAE-Value

bound-value
;; num+ : numV numV −→ numV

(define (num+ n1 n2)
(numV (+ (numV-n n1) (numV-n n2))))
Figure 6.1: First-Class Functions with Deferred Substitutions: Support Code

6.7. STANDARDIZING TERMINOLOGY 55
;; interp : FAE DefrdSub → FAE-Value

(define (interp expr ds)
(type-case FAE expr
[num (n) (numV n)]
[add (l r) (num+ (interp l ds) (interp r ds))]
(closureV bound-id bound-body ds)]
(closureV-ds fun-val))))]))
Figure 6.2: First-Class Functions with Deferred Substitutions: Interpreter

Part III
Laziness
57
Chapter 7
Programming with Laziness
We have seen several teasers about the difference between eager and lazy evaluation. As you may have
guessed, once the language becomes sufficiently rich, this seemingly small difference does have a large
impact on the language’s behavior. We will now study these differences in two different contexts.
7.1 Haskell
The paradigmatic modern lazy programming language is called Haskell, in honor of Haskell Curry, who laid
the foundation for a great deal of modern programming language theory. We will study the experience of
programming in Haskell (using its own syntax) to get a feel for the benefits of laziness.
What follows is a partisan sample of Haskell’s many wonders. It is colored by the fact that this text uses
Haskell primarily to illustrate specific linguistic features, as opposed to providing a general introduction to
Haskell. The Haskell language Web site1 has references to several texts and tutorials that describe far more
of the language and do so from several perspectives.
7.1.1 Expressions and Definitions

Like Scheme, simple Haskell programs do not need to be wreathed in scaffolding; and like most Scheme
implementations, most Haskell implementations provide an interactive environment. These notes use one
called Helium; others have a similar interface.
Prelude> 3
3
Prelude> True
True
(Prelude> is a Haskell prompt whose significance will soon become clear.) Haskell employs a traditional
algebraic syntax for operations (with the corresponding order of precedence), with parentheses representing
only grouping:
1 http://www.haskell.org/
59
60 CHAPTER 7. PROGRAMMING WITH LAZINESS
Prelude> 2*3+5
11
Prelude> 2+3*5
17
Prelude> mod 7 4
3
Prelude> mod (mod 7 4) 2
1
As in most programming languages other than Scheme, some built-in operators are written in infix notation
while most others, including user-defined ones, are written in prefix. A prefix operator can always be used
in an infix position, however, using a special syntactic convention (note that these are backquotes):
Prelude> 7 ‘mod‘ 4
3
Prelude> (7 ‘mod‘ 4) ‘mod‘ 2
1
and infix operators can, similarly, be treated as a procedural value, even used in a prefix position:
Prelude> ((<) 4 ((+) 2 3))
True
The latter is syntactically unwieldy; why would one need it?2
We have seen integers (Int) and booleans (Bool). Haskell also has characters (of type Char) that are
written inside single-quotes: ’c’ (the character ‘c’), ’3’ (the character ‘3’), ’\n’ (the newline character),
and so on.
Based on what we’ve seen so far, we can begin to write Haskell functions. This one proposes a grading
scale for a course:
scoreToLetter :: Int -> Char
scoreToLetter n
| n > 90 = ’A’
| n > 80 = ’B’
| n > 70 = ’C’
| otherwise = ’F’
The first line of this excerpt tells Haskell the type to expect for the corresponding definition (read :: as
“has the type”). The rest of the excerpt defines the actual function using a series of rules, akin to a Scheme
conditional. Loading this definition into the Haskell evaluator makes it available to execute.3 To test the
function, we use it in the evaluator:
2 Answer: Because we may want to use a traditionally infix operator, such as +, as an argument to another function. Think about
what would happen without such a notation.

3 In Helium, this definition must be saved in a file whose name begins with a capital letter. Helium’s file functions can be
accessed from the menus or, as in most other Haskell implementations, from the Haskell command-line: :l followed by a filename
loads the definitions in the named file, :r reloads the file loaded most recently, and so on. The implementation manual will describe
other short-cuts.
7.1. HASKELL 61
CS173> scoreToLetter 83
’B’
’F’
’A’
Note that in typical Haskell implementations, upon loading a file, the prompt changes from Prelude to
the name of the file, indicating in which context the expression will be evaluated. The Prelude is the set of
definitions built into Haskell.
7.1.2 Lists
Haskell naturally has more sophisticated types as well. As in Scheme, lists are inductively (or recursively)
defined data-structures; the empty list is written [] and non-empty list constructor is written :. Haskell also
offers a convenient abbreviation for lists. Thus:
Prelude> []
[]
Prelude> 1:[]
[1]
Prelude> 1:2:[]
[1,2]
Prelude> 1:[2,3,2+2]
[1,2,3,4]
Note, however, that lists must be homogenous: that is, all values in a list must be of the same type.
Prelude> [1,’a’]
Type error in element of list
expression : [1, ’a’]
term : ’a’
type : Char
does not match : Int
(The exact syntax of the type error will depend on the specific Haskell implementation, but the gist should
be the same. Here, Helium tells us that the second element of the list has type Char, whereas Helium was
expecting a value of type Int based on the first list element.)
Haskell’s Prelude has many useful functions already built in, including standard list manipulatives:
CS173> filter odd [1, 2, 3, 4, 5]
[1,3,5]
CS173> sum [1, 2, 3, 4, 5]
15
CS173> product [1, 2, 3, 4, 5]
120
We can, of course, use these in the context of our new definition:

CS173> map scoreToLetter [83, 51, 99]
"BFA"
CS173> length (map scoreToLetter [83, 51, 99])
3
It takes a little practice to know when one can safely leave out the parentheses around an expression. Eliding
them in the last interaction above leads to this error:
CS173> length map scoreToLetter [83, 51, 99]
Type error in application
expression : length map scoreToLetter [83, 51, 99]
term : length
type : [a] -> Int
does not match : ((b -> c) -> [b] -> [c]) -> (Int -> Char) -> [Int] -> d
probable fix : remove first and second argument
What?!? With practice (and patience), we realize that Haskell is effectively saying that length takes
only one argument, while the use has three: map, scoreToLetter and [83, 51, 99]. In this case,
Helium’s suggestion is misleading: the fix is not to remove the arguments but rather to inform Haskell of
our intent (first map the function, the determine the length of the result) with parentheses.
Suppose Haskell didn’t have length built in. We could build it easily, using Haskell’s pattern-matching
notation:
len [] = 0
len (x:s) = 1 + len s
Here, the argument (x:s) automatically deconstructs the list, though Haskell also provides the operators
head and tail for explicit manipulation. Notice, however, that we haven’t written a type declaration for
length. This brings us to two interesting aspects of Haskell.
7.1.3 Polymorphic Type Inference

We can ask Haskell for the type of any expression using the :type or :t directive. For instance:
CS173> :t 3
3 :: Int
CS173> :t True
True :: Bool
CS173> :t 3 + 4
3 + 4 :: Int
What should we expect when we ask Haskell for the type of len? Haskell responds with
CS173> :t len
len :: [a] -> Int
7.1. HASKELL 63
What does this type mean? It says that len consumes a list and returns an Int, but it says a little more.
Specifically, it says that the list consumed by len must have elements (recall that lists are homogenous) of
type. . . a. But a is not a concrete type like Int or Bool; rather, it is a type variable. Mathematically, we
would write this as
∀α . len : [α] → Int
That is, α is bound to a concrete type, and remains bound to that type for a particular use of len; but
different uses of len can bind α to different concrete types. We call such types polymorphic, and will study
them in great detail in Section 29.
We can see the type parameter at work more clearly using the following (trivial) function:
listCopy [] = []
listCopy (x:s) = x : listCopy s
Haskell reports this type as
CS173> :t listCopy
listCopy :: [a] -> [a]
which is Haskell’s notation for
∀α . listCopy : [α] → [α]
When we apply listCopy to different argument list types, we see that it produces lists of the same type
as the input each time:
CS173> :t listCopy [1,2,3]
listCopy [1,2,3] :: [Int]
CS173> :t listCopy [’a’,’b’,’c’]
listCopy [’a’,’b’,’c’] :: [Char]
CS173> :t listCopy [[1], [1, 2], []]
listCopy [[1], [1, 2], []] :: [[Int]]
In the last instance, notice that we are applying listCopy to—and obtaining as a result—a list of type list
of Int (i.e., a nested list of integers).
Why does Haskell assign the type parameter a name (a)? When there is only one parameter the name
isn’t necessary, but some functions are parameterized over multiple types. For instance, map is of this form:
CS173> :t map
map :: (a -> b) -> [a] -> [b]
which we might write with explicit quantifiers as
∀α, β . map : (α → β ) → [α] → [β ]
Just from reading the type we can guess map’s behavior: it consumes a function that transforms each α into
a corresponding β , so given a list of α’s it generates the corresponding list of β ’s.
In the process of studying polymorphism, we may have overlooked something quite remarkable: that
Haskell was able to generate types without our ever specifying them! This process is known as type in-
ference. Indeed, not only is Haskell able to infer a type, it infers the most general type it can for each
expression. We will study the machinery behind this remarkable power, too, in Section 29.
Exercise 7.1.1 What would you expect is the type of the empty list? Check your guess using a Haskell
implementation.
Exercise 7.1.2 Why does Haskell print the type of multi-argument functions with arrows between each
adjacent pair of arguments? Experiment with Haskell by providing what appears to be a two-argument
function (such as map) with only one argument.
7.1.4 Laziness
Is Haskell eager or lazy? We can test this using a simple interaction:
CS173> head []
exception: Prelude.head: empty list.
This tells us that attempting to ask for the first element of the empty list will result in a run-time exception.
Therefore, if Haskell used eager evaluation, the following expression should also result in an error:
CS173> (\ x -> 3) (head [])
3
The expression (\ x -> 3) uses Haskell’s notation for defining an anonymous procedure: it is the syn-
tactic analog of Scheme’s (lambda (x) 3). Thus, the whole expression is equivalent to writing
((lambda (x) 3) (first empty))
which in Scheme would indeed result in an error. Instead, Haskell evaluates it to 3. From this, we can posit
that Haskell does not evaluate the argument until it is used, and therefore follows a lazy evaluation regime.
Why is laziness useful? Clearly, we rarely write a function that entirely ignores its argument. On
the other hand, functions do frequently use different subsets of their arguments in different circumstances,
based on some dynamic condition. Most programming languages offer a form of short-circuited evaluation
for the branches of conditional (based on the value of the test expression, only one or the other branch
evaluates) and for Boolean connectives (if the first branch of a disjunction yields true the second branch
need not evaluate, and dually for conjunction). Haskell simply asks why this capability should not be lifted
to function arguments also, and demonstrates what we get when we do.
In particular, since Haskell treats all function applications lazily, this also encompasses the use of most
built-in constructors, such as the list constructor. As a result, when confronted with a definition such as
ones = 1 : ones
Haskell does not evaluate the second argument to : until necessary. When it does evaluate it, there is a
definition available for ones: namely, a 1 followed by . . . . The result is therefore an infinite list, but only
the act of examining the list actually constructs any prefix of it.
How do we examine an infinite list? Consider a function such as this:
front :: Int -> [a] -> [a]
front _ [] = []
front 0 (x:s) = []
front n (x:s) = x : front (n-1) s
7.1. HASKELL 65
When used, front causes as many list constructions of ones as necessary until the recursion terminates—
CS173> front 5 ones

[1,1,1,1,1]
CS173> front 10 ones
[1,1,1,1,1,1,1,1,1,1]
—but no more. Because the language does not force front to evaluate its arguments until necessary,
Haskell does not construct any more of ones than is needed for front to terminate. That is, it is the act
of pattern-matching that forces ones to grow, since the pattern-matcher must determine the form of the list
to determine which branch of the function to evaluate.
Obtaining the prefix of a list of ones may not seem especially impressive, but there are many good uses
for front. Suppose, for instance, we have a function that generates the eigenvalues of a matrix. Natural
algorithms for this problem generate the values in decreasing order of magnitude, and in most applications,
only the first few are meaningful. In a lazy language, we can pretend we have the entire sequence of
eigenvalues, and use front to obtain just as many as the actual application needs; this in turn causes only
that many to be computed. Indeed, any application can freely generate an infinite list of values, safe in the
knowledge that a consumer can use operators such as front to inspect the prefix it cares about.
The function front is so useful when programming in Haskell that it is actually built into the Pre-
lude, under the name take. Performing the same computation in an eager language is considerably more
complex, because the computation that generates values and the one that consumes them must explicitly
coordinate with each other: in particular, the generator must be programmed to explicitly expect requests
from the consumer. This complicates the construction of the generator, which may already have complex
domain-specific code; worse, if the generator was not written with such a use in mind, it is not easy to adapt
it to behave accordingly.
Where else are infinite lists useful? Consider the process of generating a table of data whose rows cycle
between a fixed set of colors. Haskell provides a function cycle that consumes a list and generates the
corresponding cyclic list:
CS173> take 5 (cycle ["blue", "rondo"])

["blue","rondo","blue","rondo","blue"]
The procedure for displaying the data can consume the cyclic list and simply extract as many elements from
it as necessary. The generator of the cyclic list doesn’t need to know how many rows there will be in the
table; laziness ensures that the entire infinite list does not get generated unless necessary. In other words,
programmers often find it convenient to create cyclic data structure not so much to build a truly infinite data
structure, but rather to produce one that is large enough for all possible consumers (none of which will ever
examine more than a finite prefix, but each of which may want a different number of prefix elements).
Consider one more example. At the end of some stages of the Tour de France, the top finishers receive a
“time bonus”, which we can think of as a certain number of bonus points. Let us suppose that the top three
finishers receive 20-, 12- and 8-second bonuses, respectively, while the others receive none. Given a list
reflecting the order in which contestants completed a stage, we would like a list that pairs each name with
the number of points that person received. That is, we would like a function timeBonuses such that
CS173> timeBonuses ["Lance", "Jan", "Tyler", "Roberto", "Iban"]

[("Lance",20),("Jan",12),("Tyler",8),("Roberto",0),("Iban",0)]
where ("Lance",20) is an anonymous tuple of two elements, the first projection a string and the second
a number. Note that the result is therefore a list of two-tuples (or pairs), where the heterogeneity of lists
forces each tuple to be of the same type (a string in the first projection and a number in the second).
We can write timeBonuses by employing the following strategy. Observe that every position gets a
fixed bonus (20, 12, and 8, followed by zero for everyone else), but we don’t know how many finishers there
will be. In fact, it isn’t even clear there will be three finishers if the organizers run a particularly brutal stage!
First let’s create a list of all the bonuses:
[20, 12, 8] ++ cycle [0]
where ++ appends lists. We can check that this list’s content matches our intuition:
Prelude> take 10 ([20, 12, 8] ++ cycle [0])

[20,12,8,0,0,0,0,0,0,0]
Now we need a helper function that will match up the list of finishers with the list of scores. Let’s define
this function in parts:
tB :: [String] -> [Int] -> [(String,Int)]

tB [] _ = []
Clearly, if there are no more finishers, the result must also be the empty list; we can ignore the second
argument. In contrast, if there is a finisher, we want to assign him the next available time bonus:
tB (f:fs) (b:bs) = (f,b) : tB fs bs
The right-hand side of this definition says that we create an anonymous pair out of the first elements of each
list ((f,b)), and construct a list (:) out of this pair and the natural recursion (tB fs bs).
At this point our helper function definition is complete. A Haskell implementation ought to complain
that we haven’t specified what should happen if the second argument is empty but the first is not:
(26,1): Warning: Missing pattern in function bindings:

tBb (_ : _) [] = ...
This message says that the case where the first list is not empty (indicated by (_ : _)) and the second one
is ([]) hasn’t been covered. Since we know the second list is infinitely long, we can ignore this warning.
Given this definition of tB, it is now straightforward to define timeBonuses:
timeBonuses finishers =
tB finishers ([20, 12, 8] ++ cycle [0])
This definition matches the test case above. We should also be sure to test it with fewer than three finishers:
CS173> timeBonuses ["Lance", "Jan"]

[("Lance",20),("Jan",12)]
7.1. HASKELL 67
The helper function tB is so helpful, it too (in a slightly different form) is built into the Haskell Prelude.
This more general function, which terminates the recursion when the second list is empty, too, is called zip:
zip [] _ = []
zip _ [] = []
zip (a:as) (b:bs) = (a,b) : zip as bs
Notice that the type of zip is entirely polymorphic:
CS173> :type zip

zip :: [a] -> [b] -> [(a, b)]
Its name is suggestive of its behavior: think of the two lists as the two rows of teeth, and the function as the
zipper that pairs them.
Haskell can equally comfortably accommodate non-cyclic infinite lists. To demonstrate this, let’s first
define the function zipOp. It generalizes zip by consuming an operator to apply to the pair of first
elements:
zipOp :: (a -> b -> c) -> [a] -> [b] -> [c]

zipOp f [] _ = []
zipOp f _ [] = []
zipOp f (a:as) (b:bs) = (f a b) : zipOp f as bs
We can recover the zip operation from zipOp easily:4
myZip = zipOp (\ a b -> (a,b))
But we can also pass zipOp other operators, such as (+):5
CS173> zipOp (+) [1, 1, 2, 3, 5] [1, 2, 3, 5, 8]

[2,3,5,8,13]
In fact, zipOp is also built into the Haskell Prelude, under the name zipWith.
In the sample interaction above, we are clearly beginning to build up the sequence of Fibonacci numbers.
But there is an infinite number of these and, indeed, there is no reason the arguments to zipOp must be
finite lists. Let us therefore generate the entire sequence. The code above is suggestive: clearly the first and
second argument are the same list (the list of all Fibonacci numbers), but the second is the first list “shifted”
by one, i.e., the tail of that list. We might therefore try to seed the process with the initial values, then use
that seed to construct the remainder of the list:
seed = [1, 1]
output = zipOp (+) seed (tail seed)
4 Recall that (\ · · ·) is Haskell’s equivalent of (lambda · · ·).
5 We have to enclose + to avoid parsing errors, since + is an infix operator. Without the parentheses, Haskell would try to add
the value of zipOp to the list passed as the first argument.
But this produces only one more Fiboanacci number before running out of input values, i.e., output is
bound to [2]. So we have made progress, but need to find a way to keep seed from exhausing itself. It
appears that we want a way to make seed and output be the same, so that each new value computed
triggers one more computation! Indeed,
fibs = 1 : 1 : zipOp (+) fibs (tail fibs)
We can test this in Haskell:
CS173> take 12 fibs
[1,1,2,3,5,8,13,21,34,55,89,144]
Sure enough fibs represents the entire infinite list of Fibonacci numbers, ready for further use.
Exercise 7.1.3 Earlier, we saw the following interaction:

Prelude> take 10 ([20, 12, 8] ++ cycle [0])
[20,12,8,0,0,0,0,0,0,0]
What happens if you instead write take 10 [20, 12, 8] ++ cycle [0]? Does it result in a type
error? If not, do you get the expected answer? If so, is it for the right reasons? Try this by hand before
entering it into Haskell.
Exercise 7.1.4 The definition of the Fibonacci sequence raises the question of which “algorithm” Haskell is
employing. Is it computing the nth Fibonacci number in time linear in n (assuming constant-time arithmetic)
or exponential in n?
1. First, try to determine this experimentally by asking for the nth term for large values of n (though you
may have trouble with arithmetic overflow).
2. Of course, even if you observe linear behavior, this is not proof; it may simply be that you didn’t use
a large enough value of n to observe the exponential. Therefore, try to reason about this deductively.
What about Haskell will determine the computation time of the nth Fibonacci?
7.1.5 An Interpreter
Finally, we demonstrate an interpreter for WAE written in Haskell. First we define some type aliases,
type Identifier = String
type Value = Int
followed by the two important type definitions:
type Env = [(Identifier, Value)]
data WAE = Num Int
| Add WAE WAE
| Id Identifier
| With Identifier WAE WAE
7.2. SHELL SCRIPTING 69
The core interpreter is defined by cases:

interp :: WAE -> Env -> Value
interp (Num n) env = n
interp (Add lhs rhs) env = interp lhs env + interp rhs env
interp (Id i) env = lookup i env
interp (With bound_id named_expr bound_body) env =
interp bound_body
(extend env bound_id (interp named_expr env))
The helper functions are equally straightforward:
lookup :: Identifier -> Env -> Value
lookup var ((i,v):r)
| (eqString var i) = v
| otherwise = lookup var r
extend :: Env -> Identifier -> Value -> Env

extend env i v = (i,v):env
This definition of lookup uses Haskell’s pattern-matching notation as an alternative to writing an explicit
conditional. Finally, testing these yields the expected results:
CS173> interp (Add (Num 3) (Num 5)) []
8
CS173> interp (With "x" (Add (Num 3) (Num 5)) (Add (Id "x") (Id "x"))) []
16
We can comment out the type declaration for interp (a line beginning with two dashes (--) is treated as
a comment), reload the file, and ask Haskell for the type of interp:
interp :: WAE -> Env -> Int
Remarkably, Haskell infers the same type as the one we ascribed, differing only in the use of the type alias.
Exercise 7.1.5 Extend the Haskell interpreter to implement functions using Haskell functions to represent
functions in the interpreted language. Does the resulting interpreted language have eager or lazy applica-
tion? How would you make it take on the other semantics?
7.2 Shell Scripting

While most programmers have never programmed in Haskell before, many have programmed in a lazy
language: the language of most Unix shells. In this text we’ll use the language of bash (the Bourne Again
Shell), though most of these programs work identically or have very close counterparts in other popular shell
languages.
The classical shell model assumes that all programs can potentially generate an infinite stream of output.
The simplest such example is the program yes, which generates an infinite stream of y’s:
> yes
y
y
y
and so on, forever. (Don’t try this at home without your fingers poised over the interrupt key!) To make
it easier to browse the output of (potentially infinite) stream generators, Unix provides helpful applications
such as more to page through the stream. In Haskell, function composition makes the output of one function
(the equivalent of a stream-generating application) the input to another. In a shell, the | operator does the
same. That is,
> yes | more
generates the same stream, but lets us view finite prefixes of its content in segments followed by a prompt.
Quitting from more terminates yes.
What good is yes? Suppose you run the following command:
> rm -r Programs/Sources/Java
Say some of these files are write-protected. For each such file, the shell will generate the query
rm: remove write-protected file ‘Programs/Sources/Java/frob.java’?
If you know for sure you want to delete all the files in the directory, you probably don’t want to manually
type y in response to each such question. How many can there be? Unfortunately, it’s impossible to predict
how many write-protected files will be in the directory. This is exactly where yes comes in:
> yes | rm -r Programs/Sources/Java
generates as many y inputs as necessary, satisfying all the queries, thereby deleting all the files.
We’ve seen that more is a useful way of examining part of a stream. But more is not directly analogous
to Haskell’s take. In fact, there is a Unix application that is: it’s called head. head prints the first n entries
in a stream, where n is given as an argument (defaulting to 10):
> yes | head -5

y
y
y
y
y
These examples already demonstrate the value of thinking of Unix programs as generators and con-
sumers of potentially infinite streams, composed with |. Here are some more examples. The application wc
counts the number of characters (-c), words (-w) and lines (-l) in its input stream. Thus, for instance,
> yes | head -5 | wc -l

5
7.2. SHELL SCRIPTING 71
(not surprisingly). We can similarly count the number of files with suffix .scm in a directory:
> ls *.scm | wc -l
2
We can compose these into longer chains. Say we have a file containing a list of grades, one on each
line; say the grades (in any order in the file) are two 10s, one 15, one 17, one 21, three 5s, one 2, and
ten 3s. Suppose we want to determine which grades occur most frequently (and how often), in descending
order.
The first thing we might do is sort the grades, using sort. This arranges all the grades in order. While
sorting is not strictly necessary to solve this problem, it does enable us to use a very useful Unix application
called uniq. This application eliminates adjacent lines that are identical. Furthermore, if supplied the -c
(“count”) flag, it prepends each line in the output with a count of how many adjacent lines there were. Thus,
> sort grades | uniq -c
2 10
1 15
1 17
1 2
1 21
10 3
3 5
This almost accomplishes the task, except we don’t get the frequencies in order. We need to sort one more
time. Simply sorting doesn’t do the right thing in two ways:
> sort grades | uniq -c | sort
1 15
1 17
1 2
1 21
2 10
3 5
10 3
We want sorting to be numeric, not textual, and we want the sorting done in reverse (decreasing) order.
Therefore:
> sort grades | uniq -c | sort -nr
10 3
3 5
2 10
1 21
1 2
1 17
1 15
There is something fundamentally beautiful—and very powerful!—about the structure of the Unix shell.
Virtually all Unix commands respect the stream convention, and so do even some programming languages
built atop it: for instance, by default, Awk processes its input one-line-at-a-time, so the Awk program
{print $1} prints the first field of each line, continuing until the input runs out of lines (if ever), at
which point the output stream terminates. This great uniformity makes composing programs easy, thereby
encouraging programmers to do it.
Alan Perlis recognized the wisdom of such a design in this epigram: “It is better to have 100 functions
operate on one data structure than 10 functions on 10 data structures” (the data structure here being the
stream). The greatest shortcoming of the Unix shell is that is is so lacking in data-sub-structure, relying
purely on strings, that every program has to repeatly parse, often doing so incorrectly. For example, if a
directory holds a filename containing a newline, that newline will appear in the output of ls; a program
like wc will then count the two lines as two different files. Unix shell scripts are notoriously fragile in these
regards. Perlis recognized this too: “The string is a stark data structure and everywhere it is passed there is
much duplication of process.”
The heart of the problem is that the output of Unix shell commands have to do double duty: they must
be readable by humans but also ready for processing by other programs. By choosing human readability as
the default, the output is sub-optimal, even dangerous, for processing by programs: it’s as if the addition
procedure in a normal programming language always returned strings because you might eventually want
to print an answer, instead of returning numbers (which are necessary to perform further arithmethic) and
leaving conversion of numbers to strings to the appropriate input/output routine.6
In short, Unix shell languages are both a zenith and a nadir of programming language design. Please
study their design very carefully, but also be sure to learn the right lessons from them!
6 Wemight fantasize the following way of making shell scripts more robust: all Unix utilities are forced to support a -xmlout
flag that forces them to generate output in a standard XML language that did no more than wrap tags around each record (usually,
but not always, line) and each field, and a -xmlin flag that informs them to expect data in the same format. This would eliminate
the ambiguity inherent in parsing text.
Chapter 8
Implementing Laziness
Now that we’ve seen Haskell and shell scripts at work, we’re ready to study the implementation of laziness.
That is, we will keep the syntax of our language unchanged, but alter the semantics of function application
to be lazy.
8.1 Implementing Laziness

Consider the following expression:
{with {x {+ 4 5}}
{with {y {+ x x}}
{with {z y}
{with {x 4}
z}}}}
Recall that in a lazy language, the argument to a function—which includes the named expression of a
with—does not get evaluated until use. Therefore, we can naively think of the expression above reducing
as follows:
{with {x {+ 4 5}}
{with {y {+ x x}}
{with {z y}
{with {x 4}
z}}}}
= {with {y {+ x x}}
{with {z y}
{with {x 4}
z}}} [x -> {+ 4 5}]
= {with {z y}
{with {x 4}
z}} [x -> {+ 4 5}, y -> {+ x x}]
73
74 CHAPTER 8. IMPLEMENTING LAZINESS
= {with {x 4}
z} [x -> {+ 4 5}, y -> {+ x x}, z -> y]
= z [x -> 4, y -> {+ x x}, z -> y]
= y [x -> 4, y -> {+ x x}, z -> y]
= {+ x x} [x -> 4, y -> {+ x x}, z -> y]
= {+ 4 4}
= 8
In contrast, suppose we used substitution instead of environments:
{with {x {+ 4 5}}
{with {y {+ x x}}
{with {z y}
{with {x 4}
z}}}}
= {with {y {+ {+ 4 5} {+ 4 5}}}
{with {z y}
{with {x 4}
z}}}
= {with {z {+ {+ 4 5} {+ 4 5}}}
{with {x 4}
z}}
= {with {x 4}
{+ {+ 4 5} {+ 4 5}}}
= {+ {+ 4 5} {+ 4 5}}
= {+ 9 9}
= 18
We perform substitution, which means we replace identifiers whenever we encounter bindings for them,
but we don’t replace them only with values: sometimes we replace them with entire expressions. Those
expressions have themselves already had all identifiers substituted.
This situation should look very familiar: this is the very same problem we encountered when switching
from substitution to environments. Substitution defines a program’s value; because environments merely
defer substitution, they should not change that value.
We addressed this problem before using closures. That is, the text of a function was closed over (i.e.,
wrapped in a structure containing) its environment at the point of definition, which was then used when
evaluating the function’s body. The difference here is that we must create closures for all expressions that
are not immediately reduced to values, so their environments can be used when the reduction to a value
actually happens.
We shall refer to these new kinds of values as expression closures. Since they can be the result of
evaluating an expression (as we will soon see), it makes sense to extend the set of values with this new kind
of value. We will also assume that our language has conditionals (since they help illustrate some interesting
points about laziness). Thus we will define the language CFAE/L (where the /L will denote “laziness”) with
the following grammar:
8.1. IMPLEMENTING LAZINESS 75
<CFAE/L> ::= <num>

| {+ <CFAE/L> <CFAE/L>}
| <id>
| {fun {<id>} <CFAE/L>}
| {<CFAE/L> <CFAE/L>}
Observe that the eager counterpart of this language would have the same syntax. The difference lies en-
tirely in its interpretation. As before, we will continue to assume that with expressions are converted into
immediate function applications by the parser or by a pre-processor.
For this language, we define an extended set of values:
(define-type CFAE/L-Value
[numV (n number?)]
(body CFAE/L?)
(env Env?)]
[exprV (expr CFAE/L?)
(env Env?)])
That is, a exprV is just a wrapper that holds an expression and the environment of its definition.
What needs to change in the interpreter? Obviously, procedure application must change. By definition,
we should not evaluate the argument expression; furthermore, to preserve static scope, we should close it
over its environment:1
(local ([define fun-val (interp fun-expr env)]
[define arg-val (exprV arg-expr env) ])
arg-val
(closureV-env fun-val))))]
As a consequence, an expression such as
{with {x 3}
x}
will evaluate to some expression closure value, such as

(exprV (num 3) (mtSub))
This says that the representation of the 3 is closed over the empty environment.
That may be an acceptable output for a particularly simple program, but what happens when we evaluate
this one?
1 The argument expression results in an expression closure, which we then bind to the function’s formal parameter. Since
parameters are bound to values, it becomes natural to regard the expression closure as a kind of value.
{with {x 3}
{+ x x}}
The interpreter evaluates each x in the body to an expression closure (because that’s what is bound to x
in the environment), but the addition procedure cannot handle these: it (and similarly any other arithmetic
primitive) needs to know exactly which number the expression closure corresponds to. The interpreter
must therefore “force” the expression closure to reduce to an actual value. Indeed, we must do so in other
positions as well: the function position of an application, for instance, needs to know which procedure to
invoke. If we do not force evaluation at these points, then even a simple expression such as
{+ {double 5}
{double 10}}}
cannot be evaluated (since at the points of application, double is bound to an expression closure, not a
procedural closure with an identifiable parameter name and body).
Because we need to force expression closures to values in several places in the interpreter, it makes sense
to write the code to do this only once:
;; strict : CFAE/L-Value → CFAE/L-Value [excluding exprV]
(define (strict e)
(type-case CFAE/L-Value e
[exprV (expr env)
(strict (interp expr env))]
[else e]))
Now we can use this for numbers,
(numV (+ (numV-n (strict n1) ) (numV-n (strict n2) ))))
and similarly in other arithmetic primitives, and also for applications:
(local ([define fun-val (strict (interp fun-expr env)) ]
[define arg-val (exprV arg-expr env)])
arg-val
The points where the implementation of a lazy language forces an expression to reduce to a value (if
any) are called the strictness points of the language; hence the perhaps odd name, strict, for the procedure
that annotates these points of the interpreter.
Let’s now exercise (so to speak) the interpreter’s laziness. Consider the following simple example:
{with {f {undef x}}
4}
8.2. CACHING COMPUTATION 77
Had the language been strict, it would have evaluated the named expression, halting with an error (that
undef is not defined). In contrast, out interpreter yields the value 4.
There is actually one more strictness point in our language: the evaluation of the conditional. It needs
to know the precise value that the test expression evaluates to so it can determine which branch to proceed
evaluating. This highlights a benefit of studying languages through interpreters: assuming we had good
test cases, we would quickly discover this problem. (In practice, we might bury the strictness requirement
in a helper function such as num-zero?, just as the arithmetic primitives’ strictness is buried in procedures
such as num+. We therefore need to trace which expression evaluations invoke strict primitives to truly
understand the language’s strictness positions.)
Figure 8.1 and Figure 8.2 present the heart of the interpreter.
Exercise 8.1.1 Obtaining an exprV value as the result of evaluating a program isn’t very useful, because it
doesn’t correspond to an answer as we traditionally know it. Modify the interpreter to always yield answers
of the same sort as the eager interpreter.
Hint: You may find it useful to write a wrapper procedure instead of directly modifying the interpreter.
Exercise 8.1.2 Does laziness give us conditionals “for free”? A conditional serves two purposes: to make
a decision about a value, and to avoid evaluating an unnecessary expression. Which of these does laziness
encompass? Explain your thoughts with a modified interpreter.
Exercise 8.1.3 Interactive Haskell environments usually have one other, extra-lingual strictness point: the
top-level of the Interaction window. Is this reflected here?
8.2 Caching Computation

Evaluating an expression like
{with {x {+ 4 5}}
{with {y {+ x x}}
{with {z y}
{with {x 4}
z}}}
can be rather wasteful: we see in the hand-evaluation, for instance, that we reduce the same expression,
{+ 4 5}, to 9 two times. The waste arises because we bind identifiers to expressions, rather than to their
values. So whereas one of our justifications for laziness was that it helped us avoid evaluating unnecessary
expressions, laziness has had a very unfortunate (and unforeseen) effect: it has has forced the re-evaluation
of necessary expressions.
Let’s make a small change to the interpreter to study the effect of repeated evaluation. Concretely, we
should modify strict to notify us every time it reduces an expression closure to a value:
(define (strict e)
(type-case CFAL-Value e
[exprV (expr env)
(local ([define the-value (strict (interp expr env))])

(begin
(printf ”Forcing exprV to ãñ” the-value)
the-value))]
[else e]))
This will let us track the the amount of computation being performed by the interpreter on account of
laziness. (How many times for our running example? Determine the answer by hand, then modify strict in
the interpreter to check your answer!)
Can we do better? Of course: once we have computed the value of an identifier, instead of only using it,
we can also cache it for future use. Where should we store it? The expression closure is a natural container:
the next time we attempt to evaluate that closure, if we find a value in the cache, we can simply use that
value instead.
To implement caching, we modify the interpreter as follows. First, we have to create a field for the
value of the expression closure. What’s the value of this field? Initially it needs to hold a dummy value, to
eventually be replaced by the actual one. “Replaced” means its value needs to change; therefore, it needs to
be a box. Concretely, we’ll use the boolean value false as the initial value.
[numV (n number?)]
(body CFAE/L?)
(env Env?)]
(env Env?)
(cache boxed-boolean/CFAE/L-Value?)])
We define the cache’s field predicate as follows:
(define (boxed-boolean/CFAL-Value? v)
(and (box? v)
(or (boolean? (unbox v))
(numV? (unbox v))
(closureV? (unbox v)))))
Notice that we carefully exclude exprV values from residing in the box. The box is meant to cache the result
of strictness, which by definition and construction cannot result in a exprV. Therefore, this exclusion should
never result in an error (and an indication to the contrary should be investigated).
Having changed the number of fields, we must modify all uses of the constructor. There’s only one: in
function application.
(local ([define fun-val (strict (interp fun-expr env))]
[define arg-val (exprV arg-expr env (box false) )])
8.3. CACHING COMPUTATIONS SAFELY 79
arg-val
That leaves only the definition of strict. This is where we actually use the cache:
(define (strict e)
[exprV (expr env cache)
(if (boolean? (unbox cache))
(local [(define the-value (strict (interp expr env)))]
(begin
(printf ”Forcing exprV ã to ãñ” expr the-value)
(set-box! cache the-value)
the-value))
(begin
(printf ”Using cached valueñ”)
(unbox cache)))]
[else e]))
With these changes, we see that interpreting the running example needs to force an expression closure
fewer times (how many?). The other instances reuse the value of a prior reduction. Figure 8.3 and Figure 8.4
present the heart of the interpreter. Haskell uses the value cache we have just studied, so it combines the
benefit of laziness (not evaluating unnecessary arguments) with reasonable performance (evaluating the
necessary ones only once).
Exercise 8.2.1 An expression closure is extremely similar to a regular (function) closure. Indeed, if should
be possible to replace the former with the latter. When doing so, however, we don’t really need all the pieces
of function closures: there are no arguments, so only the body and environment matter. Such a closure is
called a thunk, a name borrowed from a reminiscent technique used in Algol 60. Implement laziness entirely
using thunks, getting rid of expression closures.
Exercise 8.2.2 We could have achieved the same effect as using thunks (see Exercise 8.2.1) by simply using
one-argument procedures with a dummy argument value. Why didn’t we propose this? Put otherwise, what
benefit do we derive by keeping expression closures as a different kind of value?
Exercise 8.2.3 Extend this language with recursion and list primitives so you can run the equivalent of the
programs we saw in Section 7.1. In this extended language, implement Fibonacci, run it with and without
value caching, and arrive at a conclusion about the time complexity of the two versions.
8.3 Caching Computations Safely

Any language that caches computation (whether in an eager or lazy regime) is making a very strong tacit
assumption: that an expression computes the same value every time it evaluates. If an expression can yield a
different value in a later evaluation, then the value in the cache is corrupt, and using it in place of the correct
value can cause the computation to go awry. So we must examine this evaluation decision of Haskell.
This assumption cannot be applied to most programs written in traditional languages, because of the use
of side-effects. A method invocation in Java can, for instance, depend on the values of fields (directly, or
indirectly via method accesses) in numerous other objects, any one of which may later change, which will
almost certainly invalidate a cache of the method invocation’s computed value. To avoid having to track
this complex web of dependencies, languages like Java avoid caching values altogether in the general case
(though an optimizing compiler may introduce a cache under certain circumstances, when it can ensure the
cache’s consistency).
Haskell implementations can cache values because Haskell does not provide explicit mutation opera-
tions. Haskell instead forces programmers to perform all computations by composing functions. While this
may seem an onerous style to those unaccustomed to it, the resulting programs are in fact extremely elegant,
and Haskell provides a powerful collection of primitives to enable their construction; we caught a glimpse
of both the style and the primitives in Section 7.1. Furthermore, the lack of side-effects makes it possible for
Haskell compilers to perform some very powerful optimizations not available to traditional language com-
pilers, so what seems like an inefficient style on the surface (such as the creation of numerous intermediate
tuples, lists and other data structures) often has little run-time impact.
Of course, no useful Haskell program is an island; programs must eventually interact with the world,
which itself has true side-effects (at least in practice). Haskell therefore provides a set of “unsafe” operators
that conduct input-output and other operations. Computations that depend on the results of unsafe operations
cannot be cached. Haskell does, however, have a sophisticated type system (featuring quite a bit more,
in fact, than we saw in Section 7.1) that makes it possible to distinguish between the unsafe and “safe”
operations, thereby restoring the benefits of caching to at least portions of a program’s computation. In
practice, Haskell programmers exploit this by limiting unsafe computations to a small portion of a program,
leaving the remainder in the pure style espoused by the language.
The absence of side-effects benefits not only the compiler but, for related reasons, the programmer
also. It greatly simplifies reasoning about programs, because to understand what a particular function is
doing a programmer doesn’t need to be aware of the global flow of control of the program. In particular,
programmers can study a program through equational reasoning, using the process of reduction we have
studied in high-school algebra. The extent to which we can apply equational reasoning depends on the
number of expressions we can reasonably substitute with other, equivalent expressions (including answers).
We have argued that caching computation is safe in the absence of side-effects. But the eager version of
our interpreted language doesn’t have side-effects either! We didn’t need to cache computation in the same
way we have just studied, because by definition an eager language associates identifiers with values in the
environment, eliminating the possibility of re-computation on use. There is, however, a slightly different
notion of caching that applies in an eager language called memoization.
Of course, to use memoization safely, the programmer or implementation would have to establish that
the function’s body does not depend on side-effects—or invalidate the cache when a relevant effect happens.
Memoization is sometimes introduced automatically as a compiler optimization.
Exercise 8.3.1 There are no lazy languages that permit mutation. Why not? Is there a deeper reason beyond
the invaldation of several compiler optimizations?
Exercise 8.3.2 Why do you think there are no lazy languages without type systems?
Hint: This is related to Exercise 8.3.1.
8.3. CACHING COMPUTATIONS SAFELY 81
Referential Transparency
People sometimes refer to the lack of mutation as “referential transparency”, as in, “Haskell is
referentially transparent” (and, by implicit contrast, languages like Java, C++, Scheme and ML are
not). What do they really mean?
Referential transparency is commonly translated as the ability to “replace equals with equals”. For
example, we can always replace 1 + 2 with 3. Now think about that (very loose) definition for
a moment: when can you not replace something with something else that the original thing is
equal to? Never, of course—you always can. So by that definition, every language is “referentially
transparent”, and the term becomes meaningless.
Referential transparency really describes a relation: it relates pairs of terms exactly when they can
be considered equivalent in all contexts.
√ Thus, in most languages, 1 + 2 is referentially transparent
to 3 (assuming no overflow), and 4 (written in the appropriate notation) is referentially transparent
to 2 (assuming the square root function returns only the positive root).
Given this understanding, we can now ask the following question: what is the size of the referential
transparency relation for a program in a given language? While even a language like C subscribes a
referential transparency relation, and some C programs have larger relations (because they minimize
side-effects), the size of this relation is inherently larger for programs written in a language without
mutation. This larger relation enables a much greater use of equational reasoning.
As a programmer, you should strive to make this relation as large as possible, no matter what lan-
guage you program in: this has a positive impact on long-term program maintenance (for instance,
when other programmers need to modify your code). As a student of programming languages,
however, please use this term with care; in particular, always remember that it describes a relation
between phrases in a program, and is rarely meaningful when applied to languages as a whole.
Memoization
Memoization associates a cache with each function. The cache tracks actual argument tuples and
their corresponding return values. When the program invokes a “memoized” function, the evaluator
first tries to find the function’s value in the cache, and only invokes the function proper if that
argument tuple hadn’t been cached before. If the function is recursive, the recursive calls might also
go through the memoized version. Memoization in this instance reduces the exponential number
of calls in computing Fibonacci numbers to a linear number, without altering the natural recursive
definition.
It’s important to note that what we have implemented for lazy languages is not memoization. While
we do cache the value of each expression closure, this is different from caching the value of all
expression closures that contain the same expression closed over the same environment. In our
implementation, if a program contains the same source expression (such as a function invocation)
twice, each use of that expression results in a separate evaluation.
8.4 Scope and Evaluation Regimes

Students of programming languages often confuse the notions of scope (static versus dynamic) and evalua-
tion regimes (eager versus lazy). In particular, readers often engage in the following fallacious reasoning:
Because lazy evaluation substitutes expressions, not values, and because substituting expres-
sions (naively) results in variables getting their values from the point of use rather than the
point of definition, therefore lazy evaluation must result in dynamic scope.
It is very important to not be trapped by this line of thought. The scoping rules of a language are determined
a priori by the language designer. (For the reasons we have discussed in Section 6.5, this should almost
always be static scope.) It is up to the language implementor to faithfully enforce them. Likewise, the
language designer determines the reduction regime, perhaps based on some domain constraints. Again, the
implementor must determine how to correctly implement the chosen regime. We have seen how the use of
appropriate closure values can properly enforce static scope in both eager and lazy evaluation regimes.
8.4. SCOPE AND EVALUATION REGIMES 83
(define-type CFAE/L
[num (n number?)]
[add (lhs CFAE/L?) (rhs CFAE/L?)]
[id (name symbol?)]
[fun (param symbol?) (body CFAE/L?)]
[app (fun-expr CFAE/L?) (arg-expr CFAE/L?)])
[numV (n number?)]
(body CFAE/L?)
(env Env?)]
(env Env?)])
(define-type Env
[mtSub]
[aSub (name symbol?) (value CFAE/L-Value?) (env Env?)])
;; num+ : CFAE/L-Value CFAE/L-Value −→ numV

(numV (+ (numV-n (strict n1)) (numV-n (strict n2)))))
;; num-zero? : CFAE/L-Value → boolean

(define (num-zero? n)
(zero? (numV-n (strict n))))

(define (strict e)
[exprV (expr env)
(local ([define the-value (strict (interp expr env))])
(begin
(printf ”Forcing exprV to ãñ” the-value)
the-value))]
[else e]))
Figure 8.1: Implementation of Laziness: Support Code

;; interp : CFAE/L Env → CFAE/L-Value

(define (interp expr env)
(type-case CFAE/L expr
[num (n) (numV n)]
[add (l r) (num+ (interp l env) (interp r env))]
[id (v) (lookup v env)]
(closureV bound-id bound-body env)]
[define arg-val (exprV arg-expr env)])
arg-val
(closureV-env fun-val))))]))
Figure 8.2: Implementation of Laziness: Interpreter

8.4. SCOPE AND EVALUATION REGIMES 85
[numV (n number?)]
(body CFAE/L?)
(env Env?)]
(env Env?)
(cache boxed-boolean/CFAE/L-Value?)])
(define (boxed-boolean/CFAE/L-Value? v)
(and (box? v)
(or (boolean? (unbox v))
(numV? (unbox v))
(closureV? (unbox v)))))

(define (strict e)
[exprV (expr env cache)
(if (boolean? (unbox cache))
(local [(define the-value (strict (interp expr env)))]
(begin
(printf ”Forcing exprV ã to ãñ” expr the-value)
(set-box! cache the-value)
the-value))
(begin
(printf ”Using cached valueñ”)
(unbox cache)))]
[else e]))
Figure 8.3: Implementation of Laziness with Caching: Support Code

;; interp : CFAE/L Env → CFAE/L-Value

(type-case CFAE/L expr
[num (n) (numV n)]
[define arg-val (exprV arg-expr env (box false))])
arg-val
(closureV-env fun-val))))]))
Figure 8.4: Implementation of Laziness with Caching: Interpreter

Part IV
Recursion
87
Chapter 9
Understanding Recursion
Can we write the factorial function in FAE? We currently don’t have subtraction or multiplication, or a way
of making choices in our FAE code. But those two are easy to address. At this point adding subtraction and
multiplication is trivial, while to make choices, we can add a simple conditional construct, leading to this
language:
<CFAE> ::= <num>
| {+ <CFAE> <CFAE>}
| {* <CFAE> <CFAE>}
| <id>
| {fun {<id>} <CFAE>}
| {<CFAE> <CFAE>}
| {if0 <CFAE> <CFAE> <CFAE>}
An if0 evaluates its first sub-expression. If this yields the value 0 it evaluates the second, otherwise it
evaluates the third. For example,
{if0 {+ 5 -5}
1
2}
evaluates to 1.
Given CFAE, we’re ready to write factorial (recall from Section 6.3 that with can be handled by a
pre-processor or by the parser):
{with {fac {fun {n}
{if0 n
1
{* n {fac {+ n -1}}}}}}
{fac 5}}
What does this evaluate to? 120? No. Consider the following simpler expression, which you were asked to
contemplate when we studied substitution:
89
90 CHAPTER 9. UNDERSTANDING RECURSION
{with {x x} x}
In this program, the x in the named expression position of the with has no binding. Similarly, the environ-
ment in the closure bound to fac binds no identifiers. Therefore, only the environment of the first invocation
(the body of the with) has a binding for fac. When fac is applied to 5, the interpreter evaluates the body
of fac in the closure environment, which has no bound identifiers, so interpretation stops on the intended
recursive call to fac with an unbound identifier error. (As an aside, notice that this problem disappears with
dynamic scope! This is why dynamic scope persisted for as long as it did.)
Before you continue reading, please pause for a moment, study the program carefully, write down the
environments at each stage, step by hand through the interpreter, even run the program if you wish, to
convince yourself that this error will occur. Understanding the error thoroughly will be essential to following
the remainder of theis section.
9.1 A Recursion Construct

It’s clear that the problem arises from the scope rules of with: it makes the new binding available only in its
body. In contrast, we need a construct that will make the new binding available to the named expression also.
Different intents, so different names: Rather than change with, let’s add a new construct to our language,
rec.
<RCFAE> ::= <num>

| {+ <RCFAE> <RCFAE>}
| {* <RCFAE> <RCFAE>}
| <id>
| {fun {<id>} <RCFAE>}
| {<RCFAE> <RCFAE>}
| {if0 <RCFAE> <RCFAE> <RCFAE>}
| {rec {<id> <RCFAE>} <RCFAE>}
RCFAE is CFAE extended with a construct for recursive binding. We can use rec to write a description of
factorial as follows:
{rec {fac {fun {n}

{if0 n
1
{* n {fac {+ n -1}}}}}}
{fac 5}}
Simply defining a new syntactic construct isn’t enough; we must also describe what it means. Indeed,
notice that syntactically, there is nothing but the keyword distinguishing with from rec. The interesting
work lies in the interpreter. But before we get there, we’ll first need to think hard about the semantics at a
more abstract level.
9.2. ENVIRONMENTS FOR RECURSION 91
9.2 Environments for Recursion

It’s clear from the analysis of our failed attempt at writing fac using with that the problem has something
to do with environments. Let’s try to make this intuition more precise.
One way to think about constructs such as with is as environment transformers. That is, they are
functions that consume an environment and transform it into one for each of their sub-expressions. We will
call the environment they consume—which is the one active outside the use of the construct—the ambient
environment.
There are two transformers associated with with: one for its named expression, and the other for its
body. Let’s write them both explicitly.
ρwith,named (e) = e
In other words, whatever the ambient environment for a with, that’s the environment used for the named
expression. In contrast,
ρwith,body (e) = (aSub bound-id
bound-value
e)
where bound-id and bound-value have to be replaced with the corresponding identifier name and value,
respectively.
Now let’s try to construct the intended transformers for rec in the factorial definition above. Since rec
has two sub-expressions, just like with, we will need to describe two transformers. The body seems easier
to tackle, so let’s try it first. At first blush, we might assume that the body transformer is the same as it was
for with, so:
ρrec,body (e) = (aSub ’fac
(closureV · · ·)
e)
Actually, we should be a bit more specific than that: we must specify the environment contained in the clo-
sure. Once again, if we had a with instead of a rec, the closure would close over the ambient environment:
ρrec,body (e) = (aSub ’fac
(closureV ’n ;; bound id
(if0 · · ·) ;; body
e)
e)
But this is no good! When the fac procedure is invoked, the interpreter is going to evaluate its body in the
environment bound to e, which doesn’t have a binding for fac. So this environment is only good for the
first invocation of fac; it leads to an error on subsequent invocations.
Let’s understand how this closure came to have that environment. The closure simply closes over what-
ever environment was active at the point of the procedure’s definition. Therefore, the real problem is making
sure we have the right environment for the named-expression portion of the rec. If we can do that, then
the procedure in that position would close over the right environment, and everything would be set up right
when we get to the body.
We must therefore shift our attention to the environment transformer for the named expression. If we
evaluate an invocation of fac in the ambient environment, it fails immediately because fac isn’t bound. If
we evaluate it in (aSub ’fac (closureV · · · e) e), we can perform one function application before we halt with
an error. What if we wanted to be able to perform two function calls (i.e., one to initiate the computation,
and one recursive call)? Then the following environment would suffice:
ρrec,named (e) =
(aSub ’fac
(closureV ’n
(if0 · · ·) ;; body
(aSub ’fac
(closureV ’n
(if0 · · ·) ;; body
e)
e))
e)
That is, when the body of fac begins to evaluate, it does so in the environment
(aSub ’fac
(closureV ’n
(if0 · · ·) ;; body
e)
e)
which contains the “seed” for one more invocation of fac. That second invocation, however, evaluates its
body in the environment bound to e, which has no bindings for fac, so any further invocations would halt
with an error.
Let’s try this one more time: the following environment will suffice for one initial and two recursive
invocations of fac:
ρrec,named (e) =
(aSub ’fac
(closureV ’n
(if0 · · ·) ;; body
(aSub ’fac
(closureV ’n
(if0 · · ·) ;; body
(aSub ’fac
(closureV ’n
(if0 · · ·) ;; body
e)
e))
e))
e)
9.2. ENVIRONMENTS FOR RECURSION 93
There’s a pattern forming here. To get true recursion, we need to not “bottom out”, which happens when we
run out of extended environments. This would not happen if the “bottom-most” environment were somehow
to refer back to the one enclosing it. If it could do so, we wouldn’t even need to go to three levels; we only
need one level of environment extension: the place of the boxed e in
ρrec,named (e) =
(aSub ’fac
(closureV ’n
(if0 · · ·) ;; body
e )
e)
should instead be a reference to the entire right-hand side of that definition. The environment must be a
cyclic data structure, or one that refers back to itself.
We don’t seem to have an easy way to represent this environment transformer, because we can’t formally
just draw an arrow from the box to the right-hand side. However, in such cases we can use a variable to name
the box’s location, and specify the constraint externally (that is, once again, name and conquer). Concretely,
we can write
ρrec,named (e) =
(aSub ’fac
(closureV ’n
(if0 · · ·) ;; body
E)
e)
But this has introduced an unbound (free) identifier, E. It’s easy to make this bound, by introducing a new
(helper) function:
ρ 0 (e) =
λE.
(aSub ’fac
(closureV ’n
(if0 · · ·) ;; body
E)
e)
We’ll call this function, ρ 0 , a pre-transformer, because it consumes both e, the ambient environment, and E,
the environment to put in the closure. For some ambient environment e0 , let’s set
Fe0 = ρ 0 (e0 )
Observe that Fe0 is a procedure ready to consume an E, the environment to put in the closure. What does it
return? It returns an environment that extends the ambient environment. If we feed the right environment
for E, then recursion can proceed forever. What E will enable this?
Whatever we get by feeding some E0 to Fe0 —that is, Fe0 (E0 )—is precisely the environment that will
be bound in the closure in the named expression of the rec, by definition of Fe0 . We also want that the
environment we get back one that extends the ambient environment with a suitable binding for fac, i.e., the
same environment. In short, we want
E0 = Fe0 (E0 )
That is, the environment E0 we need to feed Fe0 needs to be the same as the environment we will get from
applying Fe0 to it. We’re being asked to supply the very answer we want to produce!
We call such a value—one such that the function’s output is the same as its input—a fixed-point of a
function. In this particular case, the fixed-point of ρ 0 is an environment that extends the ambient environment
with a binding of a name to a closure whose environment is . . . itself.
Exercise 9.2.1 This discussion about recursion has taken place in the context of environments, i.e., deferred
substitutions. How would it differ if we were performing explicit substitution (i.e., without deferral)?
Fixed-Points
Consider functions over the real numbers. The function f (x) = 0 has exactly one fixed point, be-
cause f (n) = n only when n = 0. But not all functions over the reals have fixed points: consider
f (x) = x + 1. A function can have two fixed points: f (x) = x2 has fixed points at 0 and 1 (but not,
say, at −1). And because a fixed point occurs whenever the graph of the function intersects the line
y = x, the function f (x) = x has infinitely many fixed points.
The study of fixed points over topological spaces is fascinating, and yields many rich and surprising
theorems. One of these is the Brouwer fixed point theorem. The theorem says that every continuous
function from the unit n-ball to itself must have a fixed point. A famous consequence of this theorem
is the following result. Take two instances of the same map, align them, and lay them flat on a table.
Now crumple the upper copy, and lay it atop the smooth map any way you like (but entirely fitting
within it). No matter how you place it, at least one point of the crumpled map lies directly above its
equivalent point on the smooth map!
The mathematics we must use to carefully define this fixed-point is not trivial. Fortunately for us, we’re
using programming instead of mathematics! In the world of programming, the solution will be to generate
a cyclic environment.
Recursiveness and Cyclicity
It is important to distinguish between recursive and cyclic data. A recursive object contains refer-
ences to instances of objects of the same kind as it. A cyclic object doesn’t just contains references
to objects of the same kind as itself: it contains references to itself.
To easily recall the distinction, think of a typical family tree as a canonical recursive structure.
Each person in the tree refers to two more family trees, one each representing the lineage of their
mother and father. However, nobody is (usually) their own ancestor, so a family tree is never cyclic.
Therefore, structural recursion over a family tree will always terminate. In contrast, the Web is
not merely recursive, it’s cyclic: a Web page can refer to another page which can refer back to the
first one (or, a page can refer to itself). Naı̈ve recursion over a cyclic datum will potentially not
terminate: the recursor needs to either not try traversing the entire object, or must track which nodes
it has already visited and accordingly prune its traversal. Web search engines face the challenge of
doing this efficiently.
9.3. AN ENVIRONMENTAL HAZARD 95
9.3 An Environmental Hazard

When programming with rec, we have to be extremely careful to avoid using bound values prematurely.
Specifically, consider this program:
{rec {f f}
f}
What should this evaluate to? The f in the body has whatever value the f did in the named expression of
the rec—whose value is unclear, because we’re in the midst of a recursive definition. An implementation
could give you an internal value that indicates that a recursive definition is in progress; it could even go into
an infinite loop, as f tries to look up the definition of f, which depends on the definition of f, which. . . .
There is a safe way out of this pickle. The problem arises because the named expression can be any
complex expression, including the identifier we are trying to bind. But recall that we went to all this trouble
to create recursive procedures. If we merely wanted to bind, say, a number, we have no need to write
{rec {n 5}
{+ n 10}}
when we could write
{with {n 5}
{+ n 10}}
just as well instead. Therefore, instead of the liberal syntax for RCFAE above, we could use a more conserva-
tive syntax that restricts the named expression in a rec to syntactically be a procedure (i.e., the programmer
may only write named expressions of the form {proc ...}). Then, interpretation of the named expres-
sion immediately constructs a closure, and the closure can’t be applied until we interpret the body—by
which time the environment is in a stable state.
Exercise 9.3.1 Are there any other expressions we can allow, beyond just syntactic procedures, that would
not compromise the safety of this conservative recursion regime?
Exercise 9.3.2 Can you write a useful or reasonable program that is permitted in the liberal syntax, and
that safely evaluates to a legitimate value, that the conservative syntax prevents?
Chapter 10
Implementing Recursion
We have now reduced the problem of creating recursive functions to that of creating cyclic environments.
The interpreter’s rule for with in Figure 5.2 was as follows:
(interp bound-body
(aSub bound-id
(interp named-expr
ds)
ds))]
It is tempting to write something similar for rec, perhaps making a concession for the recursive environment
by using a different constructor:
[rec (bound-id named-expr bound-body)
(interp bound-body
( recSub bound-id
(interp named-expr
ds)
ds))]
Unfortunately, this suffers from a fatal flaw. The problem is that it interprets the named expression in the
environment ds. We have decided in Section 9.3 that the named expression must syntactically be a fun
(using, say, the parser to enforce this restriction), which means its value is going to be a closure. That
closure is going to capture its environment, which in this case will be ds, the ambient environment. But
ds doesn’t have a binding for the identifier being bound by the rec expression, which means the function
won’t be recursive. So this attempt cannot succeed.
Rather than hasten to evaluate the named expression, we could pass the pieces of the function to the
procedure that will create the recursive environment. When it creates the recursive environment, it can
generate a closure for the named expression that closes over this recursive environment. In code,
(interp bound-body
97
98 CHAPTER 10. IMPLEMENTING RECURSION
(cyclically-bind-and-interp bound-id
named-expr
env))]
(Recall that ds is the old name for env.) This puts the onus on cyclically-bind-and-interp, but hopefully
also gives it the pieces it needs to address the problem. That procedure is expected to create and return
the appropriate environment, which associates the bound identifier with a closure whose environment is the
containing environment.
Let’s turn our attention to cyclically-bind-and-interp. First, let’s make a note of its contract:
;; cyclically-bind-and-interp : symbol fun env → env
(Section 9.3 explains why the second argument should be a fun and not any other kind of expression.)
Before we can create a cyclic environment, we must first extend it with a binding for the new identifier.
At this point we know the identifier’s name but not necessarily the value bound to it, so we’ll place a dummy
value in the environment:
(define (cyclically-bind-and-interp bound-id named-expr env)
(local ([define value-holder (numV 1729)]
[define new-env (aSub bound-id value-holder env)])
· · ·))
If the program uses the identifier being bound before it has its real value, it’ll get the dummy value as the
result. But because we have assumed that the named expression is syntactically a function, this can’t happen.
Now that we have this extended environment, we can interpret the named expression in it:
(local ([define value-holder (numV 1729)]
[define new-env (aSub bound-id value-holder env)]
[define named-expr-val (interp named-expr new-env)])
· · ·))
The named expression evaluates to a closure, which will close over the extended environment (new-env).
Notice that this environment is half-right and half-wrong: it has the right names bound, but the newest
addition is bound to the wrong (dummy) value.
Now comes the critical step. The value we get from evaluating the named expression is the same value
we want to get on all subsequent references to the name being bound. (We didn’t have this value before,
which is why we had to place a dummy value in the environment.) Therefore, the dummy value—the one
bound to the identifier named in the rec—needs to be replaced with the new value.
To perform this replacement, we need to ensure that the environment is mutable. To make it mutable,
we must use a different kind of value in the environment, ideally a Scheme box.1 Unfortunately, using a
box rather than a RCFAE-Value? would violate the contract for aSub. Therefore, we will need to add a
new variant to the datatype for environments; let’s call it an aRecSub, to commemorate the construct that
motivated its introduction. An aRecSub is just like an aSub, except its value must satisfy this predicate:
1AScheme box is a mutable cell. Boxes have three operations: box : Value → box, which creates a fresh cell containing the
argument value; unbox : box → Value, which returns the value stored in a box; and set-box! : box Value → void, which changes the
value held in a box but returns no value of interest.
99
(define (boxed-RCFAE-Value? v)
(and (box? v)
(RCFAE-Value? (unbox v))))
Consequently, the proper definition of an environment is
(define-type Env
[mtSub]
[aSub (name symbol?)
(value RCFAE-Value?)
(env Env?)]
[aRecSub (name symbol?)
(value boxed-RCFAE-Value?)
(env Env?)])
Now we have to rewrite the code we’ve written so far, using aRecSub instead of aSub, and boxing the
dummy value:
(local ([define value-holder (box (numV 1729)) ]
[define new-env ( aRecSub bound-id value-holder env)]
· · ·))
Now that we have a box in the environment, it’s ripe for mutation:
(local ([define value-holder (box (numV 1729))]
[define new-env (aRecSub bound-id value-holder env)]
(set-box! value-holder named-expr-val)))
Since any closures in the value expression share the same binding, they automatically avail of this update.
Finally, we must remember that cyclically-bind-and-interp has to actually return the updated environment
for the interpreter to use when evaluating the body:
(begin
(set-box! value-holder named-expr-val)
new-env)))
There’s one last thing we need to do. Because we have introduced a new kind of environment, we must
update the environment lookup procedure to recognize it.
[aRecSub (bound-name boxed-bound-value rest-env)

(unbox boxed-bound-value)
(lookup name rest-env))]
This only differs from the rule for aSub in that we must remember that the actual value is encapsulated
within a box. Figure 10.1 and Figure 10.2 present the resulting interpreter.
Working through our factorial example from earlier, the ambient environment is (mtSub), so the value
bound to new-env in cyclically-bind-and-interp is
(aRecSub ’fac
(box (numV 1729))
(mtSub))
Next, named-expr-val is bound to
(closureV ’n
(if0 · · ·)
(aRecSub ’fac
(box (numV 1729))
(mtSub)))
Now the mutation happens. This has the effect of changing the value bound to ’fac in the environment:
(aRecSub ’fac
(box (closureV · · ·))
(mtSub))
But we really should be writing the closure out in full. Now recall that this is the same environment contained
in the closure bound to ’fac. So the environment is really
(aRecSub ’fac
(box (closureV ’n
(if0 · · ·)
))
(mtSub))
where is a reference back to this very same environment! In other words, we have a cyclic environment
that addresses the needs of recursion. The cyclicity ensures that there is always “one more binding” for fac
when we need it.
Exercise 10.0.3 Were we able to implement recursive definitions in F1WAE (Section 4)? If so, how was
that possible without all this machinery?
Exercise 10.0.4 Lift this restriction on the named expression. Introduce a special kind of value that desig-
nates “there’s no value here (yet)”; when a computation produces that value, the evaluator should halt with
an error.
101
(define-type RCFAE-Value
[numV (n number?)]
(body RCFAE?)
(env Env?)])
(define (boxed-RCFAE-Value? v)
(and (box? v)
(RCFAE-Value? (unbox v))))
(define-type Env
[mtSub]
(value RCFAE-Value?)
(env Env?)]
[aRecSub (name symbol?)
(value boxed-RCFAE-Value?)
(env Env?)])
;; lookup : symbol env → RCFAE-Value

(define (lookup name env)
(type-case Env env
[aSub (bound-name bound-value rest-env)
bound-value
(lookup name rest-env))]
[aRecSub (bound-name boxed-bound-value rest-env)
(unbox boxed-bound-value)
(lookup name rest-env))]))
;; cyclically-bind-and-interp : symbol RCFAE env → env

(begin
(set-box! value-holder named-expr-val)
new-env)))
Figure 10.1: Recursion: Support Code

;; interp : RCFAE env → RCFAE-Value

(type-case RCFAE expr
[num (n) (numV n)]
[mult (l r) (num∗ (interp l env) (interp r env))]
[if0 (test truth falsity)
(if (num-zero? (interp test env))
(interp truth env)
(interp falsity env))]
(local ([define fun-val (interp fun-expr env)])
(interp arg-expr env)
(interp bound-body
(cyclically-bind-and-interp bound-id
named-expr
env))]))
Figure 10.2: Recursion: Interpreter

Part V
Intermezzo
103
Chapter 11
Representation Choices
Having grown comfortable with environments, let’s try to get to the essence of what an environment is. This
will have an interesting (and perhaps surprising) implication for their representation, which in turn will lead
us to a deeper investigation of representations in general.
11.1 Representing Environments

We have seen one way of implementing the environment, which is as a list- or stack-like datatype:
(define-type Env
[mtSub]
(value FAE-Value?)
(env Env?)])
The environment, as we’ve noted, is a mapping from identifiers to values. But it’s a particular kind of
mapping: whenever we look up the value of an identifier, we want to get at most one value for it. That is,
the environment is just a (partial) function.
In a language like Scheme, we can implement a partial function directly using Scheme’s functions,
without having to go through a data structure representation. First, let’s define a predicate for our new
representation:
(define (Env? x)
(procedure? x))
This predicate is not exact, but it’ll suffice for our purposes. Using this representation, we have a different
way of implementing aSub (the contract stays the same):
(define (aSub bound-name bound-value env)
(lambda (want-name)
(cond
[(symbol=? want-name bound-name) bound-value]
[else (lookup want-name env)])))
105
106 CHAPTER 11. REPRESENTATION CHOICES
The function aSub must return an environment and, since we’ve chosen to represent environments by
Scheme functions (lambdas), aSub must return a function. This explains the lambda.
An environment is a function that consumes one argument, a want-name, the name we’re trying to look
up. It checks whether the name that name is the one bound by the current procedure. If it is, it returns the
bound value, otherwise it continues the lookup process. How does that work?
(env name))
A environment is just a procedure expecting an identifier’s name, so to look up a name, we simply apply it
to the name we’re looking for.
The implementation of the initial value, mtSub, is simply a function that consumes an identifier name
and halts in error:
(define (mtSub)
(lambda (name)
(error ’lookup ”no binding for identifier”)))
These changes are summarized in Figure 11.1. Given these changes, the core interpreter remains un-
changed from Figure 6.2.
11.2 Representing Numbers

Let’s consider our representation of numbers. We made the decision to represent FAE numbers as Scheme
numbers. Scheme numbers handle overflow automatically by growing as large as necessary. If we want
to have FAE numbers behave differently—for instance, by overflowing like Java’s numbers do—we would
need to use modular arithmetic that captures our desired overflow modes, instead of directly mapping the
operators in FAE to those in Scheme.
Because numbers are not as interesting as some of the other features we’ll be studying, we won’t be
conducting such an exercise. The relevant point is that when writing an interpreter, we get the power to make
these kinds of choices. A related choice, which is relevant to this text, is the representation of functions.
11.3 Representing Functions

What other representations are available for FAE functions (i.e., fun expressions)? Currently, our interpreter
uses a datatype. We might try to use strings or vectors; vectors would gain little over a datatype, and it’s
not quite clear how to use a string. One Scheme type that ought to be useful, however, is Scheme’s own
procedure mechanism, lambda. Let’s consider how that might work.
First, we need to change our representation of function values. We will continue to use a datatype, but
only to serve as a wrapper of the actual function representation (just like the numV clause only wraps the
actual number). That is,
[numV (n number?)]
[closureV (p procedure?)])
11.4. TYPES OF INTERPRETERS 107
We will need to modify the fun clause of the interpreter. When we implemented environments with
procedures, we embedded a variant of the original lookup code in the redefined aSub. Here we do a similar
thing: we want FAE function application to be implemented with Scheme procedure application, so we
embed the original app code inside the Scheme procedure representing a FAE function.
(closureV (lambda (arg-val)
(interp bound-body
(aSub bound-id arg-val env))))]
That is, we construct a closureV that wraps a real Scheme closure. That closure takes a single value, which
is the value of the actual parameter. It then interprets the body in an extended environment that binds the
parameter to the argument’s value.
These changes should immediately provoke two important questions:
1. Which environment will the interpreter extend when evaluating the body? Because Scheme itself
obeys static scoping, the interpreter will automatically employ the environment active at procedure
creation. That is, Scheme’s lambda does the hard work so we can be sure to get the right cache.
2. Doesn’t the body get interpreted when we define the function? No, it doesn’t. It only gets evaluated
when something—hopefully the application clause of the interpreter—extracts the Scheme procedure
from the closureV value and applies it to the value of the actual parameter.
Correspondingly, application becomes

[define arg-val (interp arg-expr env)])
( (closureV-p fun-val)
arg-val))]
Having reduced the function and argument positions to values, the interpreter extracts the Scheme procedure
that represents the function (the boxed expression), and applies it to the argument value.
In short, a fun expression now evaluates to a Scheme procedure that takes a FAE value as its argument.
Function application in FAE is now just procedure application in Scheme. Figure 11.2 presents the entire
revised interpreter.
11.4 Types of Interpreters

We have seen a few different implementations of interpreters that are quite different in flavor. They suggest
the following taxonomy.
Definition 13 (syntactic interpreter) A syntactic interpreter is one that uses the interpreting language to
represent only terms of the interpreted language, implementing all the corresponding behavior explicitly.
Definition 14 (meta interpreter) A meta interpreter is an interpreter that uses language features of the
interpreting language to directly implement behavior of the interpreted language.
While our substitution-based FAE interpreter was very nearly a syntactic interpreter, we haven’t written
any purely syntactic interpreters so far: even that interpreter directly relied on Scheme’s implementation of
numbers. The interpreter in Figure 11.2, which employs Scheme’s lambda (and its attendant static scope)
to represent fun, is distinctly a meta interpreter.
With a good match between the interpreted language and the interpreting language, writing a meta
interpreter can be very easy. With a bad match, though, it can be very hard. With a syntactic interpreter,
implementing each semantic feature will be somewhat hard,1 but in return you don’t have to worry as much
about how well the interpreting and interpreted languages correspond. In particular, if there is a particularly
strong mismatch between the interpreting and interpreted language, it may take less effort to write a syntactic
interpreter than a meta interpreter.
As an example, consider the implementation of laziness. Suppose we use Scheme closures as the repre-
sentation of functions, as in Figure 11.2. Function application in this language automatically becomes eager,
“inheriting” this behavior from Scheme’s eager evaluation semantics. If we instead wanted lazy evaluation,
we would have to expend some effort to “undo” the behavior inherited from Scheme and make application
lazy. Worse, we would be inheriting any subtleties that Scheme’s closure and application semantics might
posess.2 In contrast, the relatively syntactic interpreter given for laziness in Section 8 does not suffer from
this peril.
Based on this discussion, we can now understand when it is and isn’t reasonable to exploit Scheme’s
representations, which may have seemed arbitrary until now. It is reasonable for features not under study
(such as numbers), but unreasonable for features directly under examination (such as function application,
when we’re studying functions—whether eager or lazy). Once we’ve provided a syntactic interpreter to
explain a feature, such as application and recursion, we can then exploit that feature in Scheme to build the
next level of complexity.
As an exercise, we can build upon our latest interpreter to remove the encapsulation of the interpreter’s
response in the FAE-Value type. The resulting interpreter is shown in Figure 11.3. This is a true meta
interpreter: it uses Scheme closures to implement FAE closures, Scheme procedure application for FAE
function application, Scheme numbers for FAE numbers, and Scheme arithmetic for FAE arithmetic. In fact,
ignoring some small syntactic differences between Scheme and FAE, this latest interpreter can be classified
as something more specific than a meta interpreter:
Definition 15 (meta-circular interpreter) A meta-circular interpreter is a meta interpreter in which the

interpreting and interpreted language are the same.
(Put differently, the trivial nature of the interpreter clues us in to the deep connection between the two
languages, whatever their syntactic differences may be.)
1 Though a poor choice of meta language can make this much harder than necessary! We choose Scheme in part because it has
so many powerful features to draw upon in a meta interpreter.

2 While Scheme has few dark corners, some versions of popular scripting languages have non-standard semantics for standard
constructs such as first-class functions. A meta interpreter that used these constructs directly would then inherit these dark corners,
probably inadvertently.
11.5. PROCEDURAL REPRESENTATION OF RECURSIVE ENVIRONMENTS 109
Meta-circular interpreters do very little to promote understanding. To gain an understanding of the

language being interpreted—one of the reasons we set out to write interpreters—we must already thoroughly
understand that language so we can write or understand the interpreter! Therefore, meta-circular interpreters
are elegant but not very effective educational tools (except, perhaps, when teaching one particular language).
In this text, therefore, we will write interpreters that balance between syntactic and meta elements, using
only those meta features that we have already understood well (such as Scheme closures). This is the only
meta-circular interpreter we will write.
That said, meta-circular interpreters do serve in one very important role: they’re good at finding weak-
nesses in language definitions! For instance, if you define a new scripting language, no doubt you will put
great effort into the design of its domain-specific features, such as those to parse data files or communicate
over a network. But will you get the domain-independent parts—procedures, scope, etc.—right also? And
how can you be sure? One good way is to try and write a meta, then meta-circular interpreter using the
language. You will probably soon discover all sorts of deficiencies in the core language. The failure to
apply this simple but effective experiment is partially responsible for the messy state of so many scripting
languages (Tcl, Perl, JavaScript, Python, etc.) for so long; only now are they getting powerful enough to
actually support effective meta-circular interpreters.
In short, by writing a meta-circular interpreter, you are likely to find problems, inconsistencies and, in
particular, weaknesses that you hadn’t considered before. In fact, some people would argue that a truly
powerful language is one that makes it easy to write its meta-circular interpreter.
Exercise 11.4.1 It is instructive to extend the Haskell interpreter given in Section 7.1 to implement recur-
sion. Use the data structure representation of the environment. In Section 10, this required mutation. Haskell
does not provide a mutation operation. Without it, are you able to implement recursion?
11.5 Procedural Representation of Recursive Environments

Section 11.1 introduced a second, procedural, representation of environments. Section 10 discussed the
implementation of recursion using a data structure representation of environments. We should therefore
consider whether we can implement recursion using the procedural representation.
Figure 10.2 presents a core interpreter that is relatively independent of the representation of environ-
ments. To enable recursion, we simply need to provide an implementation of the key helper function:
;; cyclically-bind-and-interp : symbol fun env → env
which must begin as follows:
(define (cyclically-bind-and-interp bound-name named-expr env)
· · ·)
We know that the following code pattern must exist because of the nature of the procedural representation
of environments:
..
.
(lambda (want-name)
(cond
[(symbol=? want-name bound-name)
· · ·]
[else (lookup want-name env)]))
· · ·)
If the symbols match, what do we want to return? Looking up an identifier in an environment produces
values. Recall that the named expression must be a function, so its value must be a closure. Thus, the
response if the symbols match must yield a closure:
..
.
(lambda (want-name)
(cond
(closureV (fun-param named-expr)
(fun-body named-expr)
· · ·)]
[else (lookup want-name env)]))
· · ·)
What’s not yet clear is what environment to close over. It clearly can’t be just env; it must also contain this
additional binding. So how about we give a name to this new environment that knows about the binding for
bound-name?
(local ([define rec-ext-env
(lambda (want-name)
(cond
· · ·)]
[else (lookup want-name env)]))])
· · ·))
Having named it, it’s now easy to fill in the two ellipses. What environment do we want to close over
in the closure? One that binds the function named in bound-name to the appropriate closure. This is
the environment rec-ext-env. What do we want to return from this procedure? The recursively extended
environment. This is also rec-ext-env. Thus, ignoring the box momentarily,
(lambda (want-name)
(cond

rec-ext-env )]
rec-ext-env))
The relevant portions of the interpreter are in Figure 10.2 and Figure 11.4. Notice that all the code common
to Figure 11.4 and Figure 11.1 is identical.
Exercise 11.5.1 One difference between Figure 11.1 and Figure 11.4 is the addition of recursive environ-
ment bindings. When we added support for recursive bindings in Section 10, we had to modify lookup. Why
didn’t lookup change between Figure 11.1 and Figure 11.4?
This definition raises two natural questions:

1. Is this really a recursive environment? Yes it is, though you’ll just have to take the word of the authors
of DrScheme that local does indeed define rec-ext-env as a recursive procedure, so references to that
name in the procedure’s body will indeed refer back to the same procedure.
2. Doesn’t the boxed reference to rec-ext-env have the same problem we were trying to avoid with ex-
pressions such as {rec {x x} x}? Actually, it doesn’t. The reference here is “under a lambda”,
that is, it is separated from the binding instance by a procedure declaration. Therefore, when the
named expression portion of the local is evaluated, it associates a closure with rec-ext-env that doesn’t
get invoked until much later—by which time the recursive environment of the local is safely defined.
This is the same issue we discussed in Section 9.3.
Reassuring as these responses may be, there is still something deeply unsatisfying about this solution.
We set out to add recursive functions to RCFAE. We reduced this to the problem of defining recursive
environments, which is legitimate (and, arguably, recursive environments are easier to think about than
recursive functions themselves). But we then implemented recursive environments by falling right back on
Scheme’s recursive functions: an abuse of meta-interpretive power, if ever there was any! For this reason,
the syntactic interpreter given in Section 10 is superior: it doesn’t rely on advanced knowledge of Scheme
(or, at least, no knowledge of features that we don’t also find in more mainstream programming languages).
As an aside, this discussion highlights both a power and peril of meta-interpretive choices. The power
of choosing the procedural representation is that we can add recursion to the language very easily. If our
goal is to add it as quickly as possible, while minimizing error, it makes sense to exploit the effort put into
implementing recursion for Scheme. But the peril is that this implementation does not hold descriptive
power: it still begs the question of how to implement recursion from scratch.
Exercise 11.5.2 Is it possible to implement recursive environments using the procedural representation
without employing Scheme’s constructs for creating recursive procedures? That is, can FAE alone express
recursive functions?
Exercise 11.5.3 The two implementations differ slightly in the way they treat illegal named expressions (i.e.,
ones that are not syntactic procedures). Do you see why? How would you make them behave identically?
(define (Env? x)
(procedure? x))
;; mtSub : () → Env
(define (mtSub)
(lambda (name)
;; aSub: symbol FAE-Value Env → Env

(lambda (want-name)
(cond
;; lookup : symbol Env → FAE-Value

(env name))
Figure 11.1: Procedural Representation of Environments
[numV (n number?)]
;; interp : FAE Env → FAE-Value

(type-case FAE expr
[num (n) (numV n)]
(closureV (lambda (arg-val)
(interp bound-body
(aSub bound-id arg-val env))))]
((closureV-p fun-val)
arg-val))]))
Figure 11.2: Procedural Representation of Functions

(define (number-or-procedure? v)
(or (number? v)
(procedure? v)))
(define-type Env
[mtSub]
[aSub (name symbol?) (value number-or-procedure?) (env Env?)])
;; lookup : symbol Env → number-or-procedure

(type-case Env env
[aSub (bound-name bound-value rest-env)
bound-value
(lookup name rest-env))]))
;; interp : FAE Env → number-or-procedure

(type-case FAE expr
[num (n) n]
[add (l r) (+ (interp l env) (interp r env))]
(lambda (arg-val)
(interp bound-body
(aSub bound-id arg-val env)))]
(fun-val arg-val))]))
Figure 11.3: Meta-Circular Interpreter

(define-type RCFAE-Value
[numV (n number?)]
(body RCFAE?)
(env Env?)])
(define (Env? x)
(procedure? x))
(define (mtSub)
(lambda (name)

(lambda (want-name)
(cond

(lambda (want-name)
(cond
rec-ext-env)]
rec-ext-env))

(env name))
Figure 11.4: Recursion: Support Code with Procedural Representation of Environments

Part VI
State
115
Chapter 12
Church and State
In Section 10 we freely employed Scheme boxes, but we haven’t yet given an account of how they work
(beyond an informal intuition). We therefore begin an investigation of boxes and other forms of mutation—
the changing of values associated with names—to endow a language with state.
Mutation is a standard feature in most programming languages. The programs we have written in
Scheme have, however, been largely devoid of state. Indeed, Haskell has no mutation operations at all.
It is, therefore, possible to design and use languages—even quite powerful ones—that have no explicit no-
tion of state. Simply because the idea that one can program without state hasn’t caught on in the mainstream
is no reason to reject it.
That said, state does have its place in computation. If we create programs to model the real world, then
some of those programs are going to have to accommodate the fact that there the real world has events that
truly alter it. For instance, cars really do consume fuel as they run, so a program that models a fuel tank
might best use state to record changes in fuel level.
Despite that, it makes sense to eschew state where possible because state makes it harder to reason
about programs. Once a language has mutable entities, it becomes necessary to talk about the program
before each mutation happens and similarly after that mutation (i.e., the different “states” of the program).
Consequently, it becomes much harder to determine what a program actually does, because any such answer
becomes dependent on when one is asking: that is, it becomes dependent on time.
Because of this complexity, programmers should use care when introducing state into their programs. A
legitimate use of state is when it models a real world entity that really is itself changing: that is, it models a
temporal or time-variant entity. In contrast, many uses of state to, for instance, manage the counter of a loop,
are inessential, and these invariably lead to errors when programs are used in a multi-threaded context. The
moral is that it’s important to understand what state means and how to evaluate programs that use it (which
is what we’re about to do), but it’s equally important to use it with great care, especially in an increasingly
concurrent world.
117
118 CHAPTER 12. CHURCH AND STATE
Chapter 13
Mutable Data Structures
Let’s extend our source language to support boxes. Once again, we’ll rewind to a simple language so we
can study the effect of adding boxes without too much else in the way. That is, we’ll define BCFAE, the
combination of boxes, conditionals, functions and arithmetic expressions. We’ll continue to use with
expressions with the assumption that the parser converts these into function applications. In particular, we
will introduce four new constructs:
<BCFAE> ::= ...

| {newbox <BCFAE>}
| {setbox <BCFAE> <BCFAE>}
| {openbox <BCFAE>}
| {seqn <BCFAE> <BCFAE>}
We can implement BCFAE by exploiting boxes in Scheme. This would, however, sheds little light on
the nature of boxes. We should instead try to model boxes more explicitly.
What other means have we? If we can’t use boxes, or any other notion of state, then we’ll have to stick
to mutation-free programs to define boxes. Well! It seems clear that this won’t be straightforward.
Let’s first understand boxes better. Suppose we write
(define b1 (box 5))
(define b2 (box 5))
(set-box! b1 6)
(unbox b2)
What response do we get?
This suggests that whatever is bound to b1 and to b2 must inherently be different. That is, we can
think of each value being held in a different place, so changes to one don’t affect the other.1 The natural
representation of a “place” in a modern computer is, of course, a memory cell.
1 Here’s a parable adapted from one I’ve heard ascribed to Guy Steele. Say you and I have gone on a trip. Over dinner, you say,
“You know, I have a Thomas Jefferson $2 note at home!” That’s funny, I say; so do I! We wonder whether it’s actually the same $2
bill that we both think is ours alone. When I get home that night, I call my spouse and ask her to tear my $2 bill in half. You then
call your spouse and ask, “Is our $2 bill intact?” Guy Steele is Solomonic.
119
120 CHAPTER 13. MUTABLE DATA STRUCTURES
13.1 Implementation Constraints

Before we get into the details of memory, let’s first better understand the operational behavior of boxes.
Examine this program:
{with {b {newbox 0}}

{seqn {setbox b {+ 1 {openbox b}}}
{openbox b}}}
which is intended to be equivalent to this Scheme program:

(local ([define b (box 0)])
(begin
(set-box! b (+ 1 (unbox b)))
(unbox b)))
which evaluates to 1, that is, the mutation in the first operation in the sequence has an effect on the output
of the second (which would otherwise have evaluated to 0). Now let’s consider a naive interpreter for seqn
statements. It’s going to interpret the first term in the sequence in the environment given to the interpreter,
then evaluate the second term in the same environment:
[seqn (e1 e2)
(begin
(interp e1 env)
(interp e2 env))]
Besides the fact that this simply punts to Scheme’s begin form, this can’t possibly be correct! Why not?
Because the environment is the only term common to the interpretation of e1 and e2. If the environment is
immutable—that is, it doesn’t contain boxes—and if we don’t employ any global mutation, then the outcome
of interpreting the first sub-expression can’t possibly have any effect on interpreting the second!2 Therefore,
something more complex needs to happen.
One possibility is that we update the environment, and the interpreter always returns both the value of
an expression and the updated environment. The updated environment can then reflect the changes wrought
by mutation. The interpretation of seqn would then use the environment resulting from evaluating the first
sequent to interpret the second.
While this is tempting, it can significantly alter the intended meaning of a program. For instance,
consider this expression:
{with {a {newbox 1}}

{seqn {with {b 3}
b}
b}}
2 Depends on what we mean by “effect”. The first branch of the sequence could, of course, fail to terminate or could result in an
error, which are observable effects. But they are not effects that permit the evaluation of the second branch of the sequence.
13.2. INSIGHT 121
This program should halt with an error, because static scope dictates that the second sequent (b) contains an
unbound identifier. But passing the environment from the first sequent to the second would bind b. In other
words, this strategy destroys static scope.
Even if we were to devise a sophisticated form of this environment-passing strategy (such as removing
all new bindings introduced in a sub-expression), it still wouldn’t be satisfactory. Consider this example:
{with {a {newbox 1}}
{with {f {fun {x} {+ x {openbox a}}}}
{seqn
{setbox a 2}
{f 5}}}}
We want the mutation to affect the box stored in the closure bound to f. But that closure already closes over
the environment present at the time of evaluating the named expression—an environment that still reflects
that a is bound to 1. Even if we update the environment after the setbox operation, we cannot use the
updated environment to evaluate the closure’s body, at least not without (again!) introducing the potential to
violate static scope.
As an aside, notice that in the program fragment above, changing the value of a is not a violation of
static scope! The scoping rule only tells us where each identifier is bound; it does not (in the presence of
mutation) fix the value bound to that identifier. To be pedantic, the value bound to the identifier does in fact
remain the same: it’s the same box for all time. The content of the box can, however, change over time.
We thus face an implementation quandary. There are two possible evaluation strategies for this last code
fragment, both flawed:
• Use the environment (which maps a to 1) stored in the closure for f when evaluating {f 5}. This
will, however, ignore the mutation in the sequencing statement. The program will evaluate to 6 rather
than 7.
• Use the environment present at the time of procedure invocation: {f 5}. This will certainly record
the change to a (assuming a reasonable adaptation of the environment), but this reintroduces dynamic
scope.
To see the latter, we don’t even need a program that uses mutation or sequencing statements. Even a program
such as
{with {x 3}
{with {f {fun {y} {+ x y}}}
{with {x 5}
{f 10}}}}
which should evaluate to 13 evaluates to 15 instead.
13.2 Insight
The preceding discussion does, however, give us some insight into a solution. It tells us that we need to
have two repositories of information. One, the environment, is the guardian of static scope. The other
will become responsible for tracking dynamic changes. This latter entity is known in the parlance as the
store. Determining the value bound to an identifier will become a two-step process: we will first use the
environment to map the identifier to something that the store will then map to a value. What kind of thing is
this intermediary? It’s the index that identifies a mutable cell of memory—that is, it’s a memory location.
Using this insight, we slightly alter our environments:
(define-type Env
[mtSub]
(location number?)
(env Env?)])
The store is really a partial function from address locations to values. This, too, we shall implement as a
data structure.
(define-type Store
[mtSto]
[aSto (location number?)
(value BCFAE-Value?)
(store Store?)])
Correspondingly, we need two lookup procedures:
;; env-lookup : symbol Env → location
(define (env-lookup name env)

(type-case Env env
[mtSub () (error ’env-lookup ”no binding for identifier”)]
[aSub (bound-name bound-location rest-env)
bound-location
(env-lookup name rest-env))]))
;; store-lookup : location Store → BCFAE-Value
(define (store-lookup loc-index sto)

(type-case Store sto
[mtSto () (error ’store-lookup ”no value at location”)]
[aSto (location value rest-store)
(if (= location loc-index)
value
(store-lookup loc-index rest-store))]))
Notice that the types of the two procedures compose to yield a mapping from identifiers to values, just like
the erstwhile environment.
Let’s now dive into the terms of the interpreter. We’ll assume that two identifiers, env and store, are
bound to values of the appropriate type. Some cases are easy: for instance,
13.3. AN INTERPRETER FOR MUTABLE BOXES 123
[num (n) (numV n)]

[id (v) (store-lookup (env-lookup v env) store)]
would all appear to be unchanged. Now consider the conditional:
(if (numV-zero? (interp test env store))
(interp truth env store)
(interp falsity env store))]
Suppose, with this implementation, we evaluate the following program:

{if0 {seqn {setbox b 5}
{openbox b}}
1
{openbox b}}}
We would want this to evaluate to 5. However, the implementation does not accomplish this, because
mutations performed while evaluating the test expression are not propagated to the conditional branches.
In short, what we really want is a (potentially) modified store to result from evaluating the condition’s
test expression. It is this store that we must use to evaluate the branches of the conditional. But the ultimate
goal of the interpreter is to produce answers, not just stores. What this means is that the interpreter must
now return two results: the value corresponding to the expression, and a store that reflects modifications
made in the course of evaluating that expression.
13.3 An Interpreter for Mutable Boxes

To implement state without relying on Scheme’s mutation operations, we have seen that we must modify
the interpreter significantly. The environment must map names to locations in memory, while the store maps
these locations to the values they contain. Furthermore, we have seen that we must force the interpreter to
return not only the value of each expression but also an updated store that reflects mutations made in the
process of computing that value.
To capture this pair of values, we introduce a new datatype,3
(define-type Value×Store
[v×s (value BCFAE-Value?) (store Store?)])
and reflect it in the type of the interpreter:
;; interp : BCFAE Env Store → Value×Store
Before defining the interpreter, let’s look at how evaluation proceeds on a simple program involving boxes.
3 In Scheme source programs, we would write Value×Store as Value*Store and v×s as v*s.
13.3.1 The Evaluation Pattern

Now that we have boxes in our language, we can model objects that have state. For example, let’s look at
a simple stateful object: a light switch. We’ll use number to represent the state of the light switch, where
0 means off and 1 means on. The identifier switch is bound to a box initially containing 0; the function
toggle flips the light switch by mutating the value inside this box:
{with {switch {newbox 0}}

{with {toggle {fun {dum}
{if0 {openbox switch}
{seqn
{setbox switch 1}
1}
{seqn
{setbox switch 0}
0}}}}
...}}
(Since toggle doesn’t require a useful argument, we call its parameter dum.) The interesting property of
toggle is that it can have different behavior on two invocations with the same input. In other words, the
function has memory. That is, if we apply the function twice to the same (dummy) argument, it produces
different values:

{seqn
{setbox switch 1}
1}
{seqn
{setbox switch 0}
0}}}}
{+ {toggle 1729}
{toggle 1729}}}}
This expression should return 1—the first application of toggle returns 1, and the second returns 0. To
see why, let’s write down the environment and store at each step.
The first with expression:

...}
does two things: it allocates the number 0 at some store location (say 100), and then binds the identifier
switch to this location. We’ll assume locations are represented using the boxV constructor, defined below.
This gives the following environment and store:
env = [switch → 101]

store = [101 → (boxV 100), 100 → (numV 0)]
Notice that composing the environment and store maps switch to a box containing 0.
After the second with expression:

{seqn
{setbox switch 1}
1}
{seqn
{setbox switch 0}
0}}}}
...}}
the environment and store are:

env = [toggle → 102, switch → 101]
store = [102 → (closureV ’{fun ...} [switch → 101]),
101 → (boxV 100),
100 → (numV 0)]
Now we come to the two applications of toggle. Let’s examine the first call. Recall the type of interp:
it consumes an expression, an environment, and a store. Thus, the interpretation of the first application is:
(interp ’{toggle 1729}
[toggle → 102, switch → 101]
[102 → (closureV ’{fun ...} [switch → 101]),
101 → (boxV 100),
100 → (numV 0)])
Interpreting switch results in the value (boxV 100), so interpreting {openbox switch} reduces to a
store dereference of location 100, yielding the value (numV 0).
The successful branch of the if0 expression:
{seqn
{setbox switch 1}
1}}
modifies the store; after the setbox, the environment and store are:
env = [toggle → 102, switch → 101]
store = [102 → (closureV ’{fun ...} [switch → 101]),
101 → (boxV 100),
100 → (numV 1)]
For the first application of toggle, the interpreter returns a Value×Store where the value is (numV 1) and
the store is as above.
Now consider the second application of toggle. It uses the store returned from the first application, so
its interpretation is:
(interp ’{toggle 1729}
[toggle → 102, switch → 101]
[102 → (closureV ’{fun ...} [switch → 101]),
101 → (boxV 100),
100 → (numV 1)])
This time, in the body of toggle, the expression {openbox switch} evaluates to 1, so we follow the
failing branch of the conditional. The interpreter returns the value (numV 0) and a store whose location 100
maps to (numV 0).
Look carefully at the two (interp · · ·) lines above that evaluate the two invocations of toggle. Although
both invocations took the same expression and environment, they were evaluated in different stores; that is
the difference that led to the different results. Notice how the interpreter passed the store through the
computation: it passed the original store from addition to the first toggle application, which return a
modified store; it then passed the modified store to the second toggle application, which returned yet
another store. The interpreter returned this final store with the sum of 0 and 1. Therefore, the result of the
entire expression is 1.
13.3.2 The Interpreter

This style of passing the current store in and updated store out of every expression’s evaluation is called
store-passing style. We must now update the CFAE interpreter to use this style, and then extend it to support
the operations on boxes.
Terms that are already syntactically values do not affect the store (since they require no further evalua-
tion). Therefore, they return the store unaltered:
[num (n) (v×s (numV n) store)]
[id (v) (v×s (store-lookup (env-lookup v env) store) store)]
(v×s (closureV bound-id bound-body env) store)]
The interpreter for conditionals reflects a pattern that will soon become very familiar:
(type-case Value×Store (interp test env store)
[v×s (test-value test-store)
(if (num-zero? test-value)
(interp truth env test-store )
(interp falsity env test-store ))])]
In particular, note the store used to interpret the branches: It’s the store that results from evaluating the
condition. The store bound to test-store is “newer” than that bound to store, because it reflects mutations
made while evaluating the test expression.
Exercise 13.3.1 Modify this interpreter to use the wrong store—in this case, store rather than test-store in
the success and failure branches—and then write a program that actually catches the interpreter producing
faulty output. Until you can do this, you have not truly understood how programs that use state should
evaluate!
When we get to arithmetic expressions and function evaluation, we have a choice to make: in which
order do we evaluate the sub-expressions? Given the program
{+ {openbox b}
{with {dummy {setbox b 5}}
{openbox b}}}}
evaluating from left-to-right yields 9 while evaluating from right-to-left produces 10! We’ll fix a left-to-
right order for binary operations, and function-before-argument (also a kind of “left-to-right”) for applica-
tions. Thus, the rule for addition is
[add (l r)
(type-case Value×Store (interp l env store )
[v×s (l-value l-store)
(type-case Value×Store (interp r env l-store )
[v×s (r-value r-store)
(v×s (num+ l-value r-value)
r-store )])])]
Carefully observe the stores used in the two invocations of the interpreter as well as the one returned with
the resulting value. It’s easy to make a mistake!
To upgrade a CFAE interpreter to store-passing style, we must also adapt the rule for applications. This
looks more complex, but for the most part it’s really just the same pattern carried through:
(type-case Value×Store (interp fun-expr env store )
[v×s (fun-value fun-store)
(type-case Value×Store (interp arg-expr env fun-store )
[v×s (arg-value arg-store)
(local ([define new-loc (next-location arg-store)])
(interp (closureV-body fun-value)
(aSub (closureV-param fun-value)
new-loc
(closureV-env fun-value))
(aSto new-loc
arg-value
arg-store )))])])]
Notice that every time we extend the environment, we map the binding to a new location (using next-
location, defined below). This new location is then bound to the result of evaluating the argument. As a
result, tracing the formal parameter through both the environment and the store still yields the same result.
Finally, we need to demonstrate the interpretation of boxes. First, we must extend our notion of values:
(define-type BCFAE-Value
[numV (n number?)]
(body BCFAE?)
(env Env?)]
[boxV (location number?)])
Given this new kind of value, let’s study the interpretation of the four new constructs.
Sequences are easy. The interpreter evaluates the first sub-expression, ignores the resulting value (thus
having evaluated the sub-expression only for the effect it might have on the store)—notice that e1-value is
bound but never used—and returns the result of evaluating the second expression in the store (potentially)
modified by evaluating the first sub-expression:
[seqn (e1 e2)
(type-case Value×Store (interp e1 env store)
[v×s (e1-value e1-store)
(interp e2 env e1-store)])]
The essence of newbox is to obtain a new storage location, wrap its address in a boxV, and return the
boxV as the value portion of the response accompanied by an extended store.
[newbox (value-expr)
(type-case Value×Store (interp value-expr env store)
[v×s (expr-value expr-store)
(local ([define new-loc (next-location expr-store)])
(v×s (boxV new-loc)
(aSto new-loc expr-value expr-store)))])]
To modify the content of a box, the interpreter first evaluates the first sub-expression to a location, then
updates the store with a new value for the same location. Because all expressions must return a value,
setbox chooses to return the new value put in the box as the value of the entire expression.
[setbox (box-expr value-expr)
(type-case Value×Store (interp box-expr env store)
[v×s (box-value box-store)
(type-case Value×Store (interp value-expr env box-store)
[v×s (value-value value-store)
(v×s value-value
(aSto (boxV-location box-value)
value-value
value-store))])])]
Opening a box is straightforward: get a location, look it up in the store, and return the resulting value.
[openbox (box-expr)
13.4. SCOPE VERSUS EXTENT 129

(v×s (store-lookup (boxV-location box-value)
box-store)
box-store)])]
Of course, the term “the store” is ambiguous, because there are different stores before and after evaluating
the sub-expression. Which store should we use? Does this interpreter do the right thing?
All that remains is to implement next-location. Here’s one implementation:
(define next-location
(local ([define last-loc (box −1)])
(lambda (store)
(begin
(set-box! last-loc (+ 1 (unbox last-loc)))
(unbox last-loc)))))
This is an extremely unsatisfying way to implement next-location, because it ultimately relies on a box!
However, this box is not essential. Can you get rid of it?
The core of the interpreter is in Figure 13.1 and Figure 13.2.
Exercise 13.3.2 Define next-location so it does not have side-effects.
Hint: You may need to modify the interpreter to do this.
13.4 Scope versus Extent

Notice that while closures refer to the environment of definition, they do not refer to the corresponding
store. The store is therefore a global record of changes made during execution. As a result, stores and
environments have different patterns of flow. Whereas the interpreter employs the same environment for
both arms of an addition, for instance, it cascades the store from one arm to the next and then back out
alongside the resulting value. This latter kind of flow is sometimes called threading, since it resembles the
action of a needle through cloth.
These two flows of values through the interpreter correspond to a deep difference between names and
values. A value persists in the store long after the name that introduced it has disappeared from the envi-
ronment. This is not inherently a problem, because the value may have been the result of a computation,
and some other name may have since come to be associated with that value. In other words, identifiers have
lexical scope; values themselves, however, potentially have indefinite, dynamic extent.
Some languages confuse these two ideas. As a result, when an identifier ceases to be in scope, they
remove the value corresponding to the identifier. That value may be the result of the computation, however,
and some other identifier may still have a reference to it. This premature removal of the value will, therefore,
inevitably lead to a system crash. Depending on how the implementation “removes” the value, however, the
system may crash later instead of sooner, leading to extremely vexing bugs. This is a common problem in
languages like C and C++.
In most cases, garbage collection (Section 21) lets languages dissociate scope from the reclamation of
space consumed by values. The performance of garbage collectors is often far better than we might naı̈vely
imagine (especially in comparison to the alternative).
In terms of language design, there are many reasons why C and C++ have adopted the broken policy of
not distinguishing between scope and extent. These reasons roughly fall into the following categories:
• Justified concerns about fine-grained performance control.
• Mistakes arising from misconceptions about performance.
• History (we understand things better now than we did then).
• Ignorance of concepts that were known even at that time.
Whatever their reasons, these language design flaws have genuine and expensive consequences: they cause
both errors and poor performance in programs. These errors, in particular, can lead to serious security
problems, which have serious financial and social consequences. Therefore, the questions we raise here are
not merely academic.
While programmers who are experts in these languages have evolved a series of ad hoc techniques for
contending with these problems, we students of programming languages should know better. We should
recognize their techniques for what they are, namely symptoms of a broken programming language design
rather than proper solutions to a problem. Serious students of languages and related computer science
technologies take these flaws as a starting point for exploring new and better designs.
Exercise 13.4.1 Modify the interpreter to evaluate addition from right to left instead of left-to-right. Con-
struct a test case that should yield different answers in the two cases, and show that your implementation
returns the right value on your test case.
Exercise 13.4.2 Modify seqn to permit an arbitrary number of sub-expressions, not just two. They should
evaluate in left-to-right order.
Exercise 13.4.3 New assignments to a location currently mask the old thanks to the way we’ve defined
store-lookup, but the data structure still has a record of the old assignments. Modify the implementation of
stores so that they have at most one assignment for each location.
Exercise 13.4.4 Use Scheme procedures to implement the store as a partial function.
13.4. SCOPE VERSUS EXTENT 131
;; interp : BCFAE Env Store → Value×Store

(define (interp expr env store)
(type-case BCFAE expr
[add (l r)
(type-case Value×Store (interp l env store)
(type-case Value×Store (interp r env l-store)
r-store)])])]
(type-case Value×Store (interp fun-expr env store)
(type-case Value×Store (interp arg-expr env fun-store)
new-loc
(aSto new-loc
arg-value
arg-store)))])])]
(interp truth env test-store)
(interp falsity env test-store))])]
..
.
Figure 13.1: Implementing Mutable Data Structures , Part 1

..
.
[newbox (value-expr)
(type-case Value×Store (interp value-expr env store)
[v×s (expr-value expr-store)
(local ([define new-loc (next-location expr-store)])
(v×s (boxV new-loc)
(aSto new-loc expr-value expr-store)))])]
(type-case Value×Store (interp value-expr env box-store)
(v×s value-value
(aSto (boxV-location box-value)
value-value
value-store))])])]
[openbox (box-expr)
(v×s (store-lookup (boxV-location box-value)
box-store)
box-store)])]
[seqn (e1 e2)
(interp e2 env e1-store)])]))
Figure 13.2: Implementing Mutable Data Structures, Part 2

Chapter 14
Variables
In Section 13 we studied the implementation of mutable data structures. The boxes we studied there could
just as well have been vectors or other container types, such as objects with fields.
In traditional languages like C and Java, there are actually two forms of mutation. One is mutating the
value in a container, such as an object (in Java). The expression
o.f = e
evaluates o to an object, e to some value, and changes the content of field f of o to hold the value of e. Note
that o can be an arbitrary expression (for instance, it can look up an object in some other data structure) that
is evaluated to a value. In contrast, a programmer can also write a method such as
void m (int i) {
i = 4;
}
Here, i must literally be an identifier; it cannot be an arbitrary expression that evaluates to an identifier.
That is, we are not mutating the value contained within a box (or position in a vector, or a field); rather, we
are mutating the value bound to an identifier itself. That makes the identifier a variable. A more interesting
example would be a pattern that repeatedly occurs in object-oriented programming:
private int x;
void set_x (int y) {
x = y;
}
void get_x () {
return x;
}
Here, the variable x is private to the object, and can be accessed only through the getter and setter methods.
The setter assigns a new value to x.
133
134 CHAPTER 14. VARIABLES
14.1 Implementing Variables

First, let’s extend our language to include variables:
<VCFAE> ::= ...

| {set <id> <VCFAE>}
| {seqn <VCFAE> <VCFAE>}
Observe that the set expression expects a literal identifier after the keyword.
Implementing variables is a little different from implementing boxes. In the latter case, we first evaluate
the position that identifies the box:
..
.])]
In contrast, in a language with variables, identifiers do not represent boxes. Therefore, the corresponding
code:
[set (var value)
(type-case Value×Store (interp var env store)
[v×s (var-value var-store)
..
.])]
would be counter-productive. Evaluating the identifier would result in a value, which we cannot mutate, as
opposed to a location, whose content we can modify by updating the store. This immediately suggests a
slightly different evaluation strategy:
[set (var value)
(type-case Value×Store (interp value env store)
(local ([define the-loc (env-lookup var env)])
..
.)])]
That is, we evaluate the expression that represents the new value to be stored in the interpreter. Instead of
evaluating the identifier, however, we only look it up in the environment. This results in a location where
the new value should be stored. In particular, notice an unusual pattern: the interpreter dereferences the
identifier in the environment, but does not dereference the result (the identifier’s location) in the store. We
have not seen this pattern before, and will never see it again after this material.
Many languages make a distinction between mutable data structures and mutable identifiers. When a
mutable identifier appears in the assignment position of an assignment operator (many languages use the
same syntactic operator, = or :=, to represent both operations), the language implementation only partially
resolves the identifier. This special kind of value—the location of an identifier—is traditionally known as
an l-value (pronounced “ell-value”).
Whence this unusual name? Consider the following two statements in C:
14.2. INTERACTION BETWEEN VARIABLES AND FUNCTION APPLICATION 135
x = 2;
y = x;
In the second statement, x must be reduced to a value—i.e., store-lookup and env-lookup must be composed
and applied to its content—whereas in the first statement, x must only be reduced to a location, not to a
value. In languages where locations are not values (more on that below), this odd kind of “value” is known
as an “l-value”, since it appears only on the left-hand-side of assignment statements.
Given this insight, we can now easily complete the definition of the assignment statement:
[set (var value)
(v×s value-value
(aSto the-loc value-value value-store)))])]
The rest of the interpreter remains unchanged. Note, in particular, that it still employs store-passing style.
Figure 14.1 and Figure 14.2 present the core of the interpreter.
14.2 Interaction Between Variables and Function Application

Variables and function application appear to be two independent language features, but perhaps they are not.
Consider the following program:
{with {v 0}
{with {f {fun {y}
{set y 5}}}
{seqn {f v}
v}}}
What do we expect it to evaluate to? There are two different, reasonable answers: 0 and 5. The first assumes
that the mutation is to the formal variable, y, and does not affect the actual argument, v; the second assumes
that this mutation does have the effect of modifying the actual argument.
Our current implementation yields the value 0. This is because the act of invoking the function binds
the formal parameter to a new location:
new-loc
(aSto new-loc
arg-value
arg-store)))
The evaluated argument is held in this new location. Therefore, changes to the content of that location in the
store do not affect the actual parameter. This is the standard form of eager evaluation, traditionally called
call-by-value.
Let’s now explore the alternative. This form of evaluation is called call-by-reference. This new technique
gets its name because we will pass a reference to the actual argument, rather than merely its value. Thus,
updates to the reference within the called procedure will become visible to the calling context, too.
To explore this design, let’s extend our language further so we have two kinds of procedures: call-by-
value (fun) and call-by-reference (refun):
<RVCFAE> ::= ...

| {refun {<id>} <RVCFAE>}
| {set <id> <RVCFAE>}
| {seqn <RVCFAE> <RVCFAE>}
That is, syntactically a call-by-reference procedure looks the same as a call-by-value procedure other than
the distinguishing keyword. It is their interpretation that will distinguish them.
All the code we have developed thus far remains the same for call-by-value procedure invocation. In
particular, with expressions should continue to expand into immediate applications of fun-defined proce-
dures. Let us proceed to defining the interpretation of call-by-reference procedures.
The first step is to evaluate a reference procedure definition. This is straightforward:
[refun (bound-id bound-body)
(v×s (refclosV bound-id bound-body env) store)]
We create a new kind of closure so we can later distinguish what kind of procedure we are about to apply,
but its fields are the same:
(define-type RVCFAE-Value
[numV (n number?)]
(body RVCFAE?)
(sc SubCache?)]
[refclosV (param symbol?)
(body RVCFAE?)
(sc SubCache?)])
Now let us study the interpretation of application. After evaluating the procedure position, we must check
which kind of procedure it is before evaluating the argument. If it’s a call-by-value procedure, we proceed
as before:
(type-case RVCFAE-Value fun-value
[closureV (cl-param cl-body cl-env)
14.3. PERSPECTIVE 137

(interp cl-body
(aSub cl-param
new-loc
cl-env)
(aSto new-loc
arg-value
arg-store)))])]
[refclosV (cl-param cl-body cl-env)
..
.]
[numV ( ) (error ’interp ”trying to apply a number”)])])]
We can thus focus our attention on the interpretation of function applications where the function position
evaluates to a reference procedure closure.
When applying a call-by-reference procedure, we must supply it with the location of the actual argument.
This presupposes that the actual argument will be a variable. We put this down to an implicit constraint of
the language, namely that whenever applying a reference procedure, we assume that the argument expres-
sion is syntactically a variable. Given this, we can easily determine its location, and extend the closure’s
environment with the formal parameter bound to this location:
(local ([define arg-loc (env-lookup (id-name arg-expr) env)])
(interp cl-body
(aSub cl-param
arg-loc
cl-env)
fun-store))]
Notice the recurrence of the l-value pattern: an environment lookup without a corresponding store lookup.
(This is why we dispatched on the type of the closure without first evaluating the argument: had we not done
so, the argument expression would have been reduced to a value, which would be either useless or incorrect
in the case of reference procedures.) As a result, any mutations to the formal parameter are now changes
to the same location as the actual parameter, and are thus effectively mutations to the actual parameter also.
Thus, the example that inaugurated this section will yield the result 5.
Figure 14.3 and Figure 14.4 present the core of the interpreter.
14.3 Perspective
Should languages have reference procedures? Passing references to procedures has the following dangerous
property: the formal parameter becomes an alias of the actual parameter, as all changes to the formal
manifest as changes to the actual also. This is especially insidiuous because the programmer may not
know he is about to apply a reference procedure: Some languages like C offer the ability to mark specific
parameters of multi-parameter procedures with keywords such as & and ref, meaning they alone should be
passed by reference (these are known as reference parameters). The client of such a procedure may thus find
that, mysteriously, the act of invoking this procedure has changed the value of his identifiers. This aliasing
effect can lead to errors that are particularly difficult to detect and diagnose.
This phenomenon cannot occur with call-by-value: changes to the variable in the called procedure do
not affect the caller. There is, therefore, nearly universal agreement in modern languages that arguments
should be passed by value. If the called procedure intends to mutate a value, it must consume a box (or other
container data structure); the caller must, in turn, signal acceptance of this behavior by passing a box as the
actual argument. The caller then has the freedom to inspect the content of the (possibly mutated) box and
determine whether to accept this mutation in the remainder of the program, or to reject it by ignoring the
altered content of the box.
Why did languages introduce reference parameters? For one thing, they are “cheaper”: they do not
require additional allocation. (We can see this difference clearly when we contrast the two kinds of procedure
application.) However, the problems they introduce arguably far outweigh this small savings in memory.
Reference parameters do, however, also confer a small expressiveness benefit. Without reference pa-
rameters, we cannot define a procedure that swaps the content of two variables. In the following code,
{with {swap {fun {x}
{fun {y}
{with {z x}
{seqn {set x y}
{set y z}}}}}}
{with {a 3}
{with {b 2}
{seqn {{swap a} b}
b}}}}
the result of the computation is still 2, because the mutations to x and y inside the procedure do not affect
a and b. In contrast,
{with {swap {refun {x}
{refun {y}
{with {z x}
{seqn {set x y}
{set y z}}}}}}
{with {a 3}
{with {b 2}
{seqn {{swap a} b}
b}}}}
results in the value 3: since x and y are just aliases to a and b, mutations to the former are reflected as
mutations to the latter. (Note that both procedures must be refuns and not funs, else the swap is at best
partial.)
This example also, however, illustrates why aliasing can cause problems. The implementor of the pro-
cedure may have used mutation accidentally, without meaning to affect the caller. The procedure boundary
abstraction has, however, been compromised by the aliasing, and accidental side-effects can leak into the
calling contexts, exposing unnecessary implementation details of the procedure.
In the early days of programming language design, before programs were particularly sophisticated, the
ability to write simple abstractions such as swap was considered valuable (since it is used, for instance, in
the implementation of some sorting algorithms). Today, however, we recognize that such abstractions are
rather meager in the face of the needs of modern systems. We pay greater attention, instead, to the need
for creating useful abstraction boundaries between units of modularity such as procedures: the fewer hidden
interactions they have, and the less they interfere with one another, the more easily we can reason about their
behavior in isolation.
Exercise 14.3.1 While call-by-value preserves the value of variables in the calling context, it does not
protect all values. In particular, in many call-by-value languages, a composite data structure (such as a
vector) passed as an argument may be mutated by the callee, with the effects visible to the caller.
1. Does this behavior contradict the claim that the language is passing “values” as arguments? Use our
investigation of mutable data structures in Section 13 to make your argument rigorous.
Hint: Implement an interpreter for a language with both boxes and call-by-reference application,
then argue about similarities and differences.
2. Languages like ML tackle this problem by forcing programmers to annotate all mutable data struc-
tures using references, the ML counterpart to boxes. Any data structure not so mutated is considered
immutable. What trade-offs does ML’s design introduce?
Exercise 14.3.2 There appears to be a neutral ground between call-by-value and call-by-reference. Con-
sider the following proposed syntax:
{with {swap {fun {x}
{fun {y}
{with {z x}
{seqn {set x y}
{set y z}}}}}}
{with {a 3}
{with {b 2}
{seqn {{swap {ref a}} {ref b}}
b}}}}
The ref notation is an indicator to the interpreter to pass the variable’s location rather than its value; that
is, by using {ref a} and {ref b}, the invoker of the procedure indicates his willingness to have his
variables be aliased and thus, potentially, be mutated.
1. Modify the interpreter to support the use of ref for procedure arguments.
2. Does this proposal result in a procedural abstraction of the process of swapping the values of two
variables? If it does, this would reconcile the design tension between the two invocation techniques:
it avoids the difficulty of call-by-value (the inability to write a swap procedure) as well as that of
call-by-reference (aliasing of parameters without the caller’s knowledge). Discuss.
3. Suppose programmers are allowed to apply ref to variables elsewhere in the program. What type
should the interpreter use to represent the resulting value? How does this compare to an l-value?
Does this introduce the need for additional operators in the language? How does this relate to the &
operator in C?
(define-type VCFAE-Value
[numV (n number?)]
(body VCFAE?)
(env Env?)])
;; interp : VCFAE Env Store → Value×Store

(type-case VCFAE expr
[add (l r)
r-store)])])]
..
.
Figure 14.1: Implementing Variables, Part 1

..
.
new-loc
(aSto new-loc
arg-value
arg-store)))])])]
(interp truth env test-store)
(interp falsity env test-store))])]
[set (var value)
(v×s value-value
[seqn (e1 e2)
Figure 14.2: Implementing Variables, Part 2

(define-type RVCFAE-Value
[numV (n number?)]
(body RVCFAE?)
(env Env?)]
[refclosV (param symbol?)
(body RVCFAE?)
(env Env?)])
;; interp : RVCFAE Env Store → Value×Store

(type-case RVCFAE expr
[add (l r)
r-store)])])]
[if0 (test pass fail)
(interp pass env test-store)
(interp fail env test-store))])]
[refun (bound-id bound-body)
(v×s (refclosV bound-id bound-body env) store)]
..
.
Figure 14.3: Implementing Call-by-Reference, Part 1

..
.
(type-case RVCFAE-Value fun-value
[closureV (cl-param cl-body cl-env)
(interp cl-body
(aSub cl-param
new-loc
cl-env)
(aSto new-loc
arg-value
arg-store)))])]
(local ([define arg-loc (env-lookup (id-name arg-expr) env)])
(interp cl-body
(aSub cl-param
arg-loc
cl-env)
fun-store))]
[numV ( ) (error ’interp ”trying to apply a number”)])])]
[set (var value)
(v×s value-value
[seqn (e1 e2)
Figure 14.4: Implementing Call-by-Reference, Part 2

Part VII
Continuations
145
Chapter 15
Some Problems with Web Programs
Web programs can be awfully buggy. For instance, consider the following interaction with a popular com-
mercial travel Web site.
1. Choose the option to search for a hotel room, and enter the corresponding information.
2. Suppose the Web site response with two hotels, options A and B.
3. Using your browser’s interactive facilities, open the link to hotel A in a separate window.
4. Suppose you find A reasonable, but are curious about the details of B. You therefore return to the
window listing all the hotels and open the details for B in a separate window.
5. Having scanned the details of B, you find A a more attractive option. Since the window for A is still
on-screen, you switch to it and click the reservation link.
6. The travel site makes your reservation at hotel B.

If an error like this were isolated to a single Web page, or even to a single site, we can put it down to
programmer error. But when such errors occur on numerous sites, as indeed they do, it forces us to sys-
tematically investigate their cause and to more carefully consider the design and implementation of Web
programs.
Before we investigate the problem in general, it helps to understand its breadth. The following is an
uncontroversial property that we would expect of a travel reservation site:
The user should receive a reservation at the hotel that was displayed on the page he submitted.
Can we generalize this? That is, should a user receive information based strictly on the information displayed
on the page on which the user clicked a button?
Consider an on-line bookstore. Conduct the same sequence of interactions as above, except with books
instead of hotels. Upon examining choice B, suppose you clicked to add it to your “shopping cart”. Now
when you go to the page for book A and add it, too, to your shopping cart, what do you expect to find in it?
Certainly, the bookseller hopes you have both A and B in the cart (since, after all, they are in the business of
selling as many books as possible). This is a clear violation of the property we elucidated above.
147
148 CHAPTER 15. SOME PROBLEMS WITH WEB PROGRAMS
The problem is compounded by the number of interaction operations supported by modern Web browsers.
In addition to opening Web pages in new windows, browsers offer the ability to clone the currently-visible
page, to go back to a previous page, to go forward to a page (from which the user had previously gone
back), to create bookmarks, and so on. Worse, most of these operations are silent: the browser does not
notify the Web application that they have been executed, so the application must reconstruct events based
on the submissions it receives.
Many of the problems with Web programs trace back to their structure. The Web’s architecture dictates
that every time a Web program sends an Web page to a user, it is forced to terminate; this is because the Web
implements a stateless protocol. If and when the user chooses to resume the computation (by clicking on a
link or button), some other program must resume the computation. This forces a rather perverse program
structure on the programmer. We will now study the implications of this structure in some detail.
Stateful and Stateless Protocols

Suppose a client-server computation performs multiple interactions. In a stateful protocol, the server
maintains some state information recording its context in the dialog. A well-known example of a
stateful protocol is FTP, the Internet file-transfer protocol. In an FTP session, the user can enter
multiple commands, and the interpretation of each command is relative to the history of past com-
mands. That is, two invocations of ls (to list files in a directory) will produce different answers
if the user has invoked cd (to change the directory) betwixt. (The context information here is the
current directory.)
In contrast to many traditional Internet protocols, the Web implements a stateless protocol, meaning
it does not retain any record of prior communication. As a result, in principle the Web application is
responsible for completely restoring the state of the computation on each interaction. By analogy,
suppose you were executing an editor such as Emacs within an SSH session (which is also stateful:
this state is lost when the connection dies). In a stateless SSH, after every unit of output the con-
nection would close. When the user entered another keystroke, the communication would have to
carry with it the entire state of running applications (indeed, the entire history, to enable Undo op-
erations), the server would have to invoke Emacs afresh, run all the commnands entered so far and,
having restored the application to its past state. . . enter one new keystroke. (In practice, therefore,
Web applications are instead designed to communicate with coarse granularity.)
Stateful protocols are easier to program, because the developer is not responsible for setup and
breakdown of the state at each interaction. So why use a stateless protocol? They confer the advan-
tage that the server can tolerate far higher loads. If a server can maintain only 1,000 connections at
any instant, a stateful protocol that keeps connections open until the transaction terminates would
not be able to service more than 1,000 users at a time. Worse, it would need a policy for determin-
ing when to terminate connections that appear to no longer be active (e.g., users who have neither
logged out nor completed a purchase). A stateless protocol avoids this; the server can serve many
more clients in rapid order, and can ignore clients who are not interested in completing a computa-
tion. It pays the price of transmitting enough data to resume the compuation.
Chapter 16
The Structure of Web Programs
Suppose we are trying to implement the following simple Web program. The program presents the user with
a prompt for a number. Given an input, it presents a prompt for a second number. Given a second input, it
displays the sum of the two numbers in a Web page:1
(web-display
(+ (web-read ”First number: ”)
(web-read ”Second number: ”)))
While this is an extremely simple application, it is sufficient for demonstrating many concepts. Furthermore,
it is a microcosm of a Web application that accepts information in multiple stages, such as the outward flight
choice on one page and the return choice on the next.
Even this “addition server” is difficult to implement:
1. The Web developer must turn this application into three programs:
(a) The first program displays the first form.
(b) The second program consumes the form values from the first form, and generates the second
form.
(c) The third program consumes the form values from the second form, computes the output, and
generates the result.
2. Because the value entered in the first form is needed by the third program to compute its output, this
value must somehow be transmitted between from the first program to the third. This is typically done
by using the hidden field mechanism of HTML.
3. Suppose, instead of using a hidden field, the application developer used a Java Servlet session object,
or a database field, to store the first number. (Application developers are often pushed to do this
because that is the feature most conveniently supported by the language and API they are employing.)
Then, if the developer were to exploratorily open multiple windows, as we discussed in Section 15,
the application can compute the wrong answer.
1 We are assuming the existence of some simple primitives that mask the necessary but, for now, irrelevant complexity of
generating HTML forms, and so on.
149
150 CHAPTER 16. THE STRUCTURE OF WEB PROGRAMS
In particular, the program we have written above, which runs perfectly well on a display console, cannot
run on the Web: the moment web-read dispatches its Web form to the user, the Web protocol forces the com-
putation to terminate, taking with it all memory of what had to happen next, i.e., the pending computation.
Where is this pending computation specified? The system resumes execution at the URL specified in
the “action” field of the generated form. The developer is therefore responsible for making sure that the
application that resides at that URL is capable of resuming the computation in its entirety. We have entirely
neglected this problem by assuming the existence of a web-read procedure, but in fact the entire problem is
that we cannot implement it without a more explicit handle on the pending computation.
16.1 Explicating the Pending Computation

For our motivating example, what is the pending computation at the point of the first interaction? In words, it
is to consume the result from the form (the first number), generate a form for the second number, add them,
then display their result. Since natural language is unwieldy, we would benefit from writing this pending
computation in code instead:
(web-display
(+ •
where we use • to represent the result from the user’s form submission. What is •, exactly? It appears to be
an invented notation that we must then explain formally. Instead, we can treat it as an identifier, binding it
in the traditional way:
(lambda (•)
(web-display
(+ •
(web-read ”Second number: ”))))
This procedure, then, represents the computation pending at the point of the first interaction. Applying this
procedure to the result of that interaction (the user’s input) resumes the computation. Similarly, the pending
computation at the point of the second interaction is
(lambda (•2 )
(web-display
(+ •
•2 )))
where • is the user’s response to the first prompt, which should be in the closure of this procedure.
16.2 A Better Server Primitive

Suppose, therefore, that we had a modified version of web-read that we’ll call web-read/k. This new pro-
cedure takes two arguments. The first is a string that it converts into a form, as before. The second is a
procedure of one argument representing the pending computation, which we’ll henceforth call the receiver.
16.2. A BETTER SERVER PRIMITIVE 151
Figure 16.1: Form generated by web-read/k
Every time web-read/k is invoked, it creates a fresh entry in a hash table. It stores the receiver in this
entry, and generates a form action URL that contains the hash table key for this procedure. The hash table is
kept in memory by the Web server (which, we’ll assume, doesn’t terminate). web-read/k generates a page
and then terminates the Web application’s execution, in conformance with the Web protocol.
This generated page is shown in Figure 16.1. The image shows, in outline, the Web page generated by
invoking web-read/k with the first argument reading ”First”. This string becomes the prompt. Next to the
prompt is a text box where the user can provide input. The action field of the HTML form has a reference to
the hash table key of the corresponding fresh entry (in this instance, k2592).
When the user submits a response, the server invokes the application named launch. This application
does two things. First, it uses the key associated with the id argument to obtain a receiver closure from the
hash table. Second, it extracts the input typed by the user in the text box. The receiver is then applied to this
extracted value. This resumes the computation.
Assuming the existence of such a primitive, we might try to rewrite our running application as
(web-display
(+ (web-read/k ”First number: ”
(lambda (•)
•))
but this won’t work! Recall that at every Web interaction, the Web application entirely terminates. That
means, any computation that has not been included in the receiver is lost forever. As a consequence, when
this application resumes, the only “remembered” computation is that in the receiver, which is just the identity
procedure: the second Web input, as well as the ensuing computation and the display of the result, have all
been irretrievably lost.
In other words, any computation that isn’t explicitly mentioned in the receiver simply never gets per-
formed, because of the program’s termination after each interaction. This forces us to move all pending
computation into the receiver. Here’s what we might try:
(web-read/k ”First number: ”
(lambda (•)
(web-display
(+ •
(web-read ”Second number: ”)))))
This, however, is subject to the same analysis: it still uses the hypothetical web-read procedure, which we’ve
conceded we don’t quite know how to implement. We must, therefore, instead employ web-read/k again, as
follows:
(lambda (•)
(web-read/k ”Second number: ”
(lambda (•)
(web-display
(+ • •))))))
Oh, not quite: we want to add the first number to the second, not just compute twice the second number.
Therefore:
(lambda (•1 )
(lambda (•2 )
(web-display
(+ •1 •2 ))))))
Now, when the program finally generates the sum, it can safely halt without having registered any receivers,
because there aren’t any computations left to perform. Relative to the original source program, however, the
structure of this application is considerably more intricate.
16.3. TESTING WEB TRANSFORMATIONS 153
Exercise 16.2.1 To be entirely pedantic, there is one thing left to do, which is to explicitly halt the program.
Extend the program to do this, then transform it to correctly employ web-read/k.
16.3 Testing Web Transformations

One of the subtle problems with transforming interactive programs for the Web is that they are difficult to
test. This difficulty has at least two facets. First, the use of HTML makes programs unwieldy, so we would
rather defer its use until the end, but without it we cannot interact with a Web browser. Second, testing a
program at the console can be misleading: a computation may not have been properly moved into a receiver
but, because Scheme programs do not terminate after every interaction, we would never notice this problem
until we ran the program on the Web.
Fortunately, it is easy to simulate the Web’s behavior at the console with the following code. The fol-
lowing implementation of web-read/k stores the receiver and prompt in a box, and terminates the program’s
execution using error:
(define the-receiver (box ’dummy-value))
(define receiver-prompt (box ’dummy-value))
(define (web-display n)
(printf ”Web output: ãñ” n))
(define (web-read/k p k)
(begin
(set-box! receiver-prompt p)
(set-box! the-receiver k)
(error ’web-read/k ”run (resume) to enter number and simulate clicking Submit”)))
The procedure resume uses the values in these boxes to resume the computation:
(define (resume)
(begin
(display (unbox receiver-prompt))
((unbox the-receiver) (read))))
We can therefore test a program such as the addition application as follows:
web-read/k: run (resume) to enter number and simulate clicking Submit
>
This means the program has arrived at web-read/k for the first time. We run
> (resume)
which prompts us for the first input. Providing an input results in the same terminating “error” message,
corresponding to the next interaction point. Running (resume) prompts for the second input. When we
provide the second, we see the sum of the two numbers printed to the console.
Exercise 16.3.1 Use this testing harness to execute the incorrectly transformed versions of the addition
program to explore their behavior.
16.4 Executing Programs on a Traditional Server

Suppose we must run our Web application on a traditional Web server, which does not provide support for
the hash table used by web-read/k. This doesn’t mean we must waste the effort we expended transforming
the program: that effort was a direct consequence of the Web’s protocol, which the traditional server also
obeys (even more slavishly!).
What’s the problem with executing this program on a traditional server?
(lambda (•1 )
(lambda (•2 )
(web-display
(+ •1 •2 ))))))
If web-read/k cannot behave in a privileged fashion, then its receiver argument will not be invoked automat-
ically by the server. Instead, the entire computation will terminate with the first interaction.
To reflect this problem, let us use a different primitive, web-read/r in place of web-read/k. The suffix
indicates that it will be given the name of a receiver as a second argument. web-read/r uses this name
in the URL inserted in the action field of the generated form. To do so, however, each receiver must be a
named Web application that the server can invoke directly (whereas the receivers are currently anonymous
procedures nested within other procedures).
The process of making nested procedures into top-level ones is known as lifting. That is, each anony-
mous procedure is moved to the top-level and given an unique name. In the example program above, the
innermost procedure might become
(define (f2 •2 )
(web-display
(+ •1
•2 )))
which the outer procedure can refer to:

(define (f1 •1 )
(web-read/r ”Second number: ”
”f2”))
The main program then becomes

(web-read/r ”First number: ”
”f1”)
16.4. EXECUTING PROGRAMS ON A TRADITIONAL SERVER 155
Because of this transformation, web-read/r can safely terminate after using the procedure named in the
second argument in the action field’s URL. All the remaining work must be completed by this named top-
level procedure. Each top-level procedure consumes one argument, which is the data provided by the user.
Unfortunately, by sloppily lifting the procedures to the top-level, we’ve created a problem: •1 is a free
identifier in f2! The problem is that we were simplistic in the way we lifted the procedures. (A different sim-
plistic method—failing to keep the two instances of • separate—would, of course, have created a different
problem, namely that f2 would have just added • to itself, ignoring the user’s first input.)
In general, when lifting we must add parameters for all the free variables of the procedure being lifted,
then pass along the values for parameters from the point of closure creation. In general, procedure lifting re-
quires the computation of a transitive closure (because lifting one procedure may render another procedure’s
previously-bound identifiers free). That is, the Web program ought to become:
(define (f2 •1 •2 )
(web-display
(+ •1
•2 )))
(define (f1 •1 )
(web-read/r ”Second number: ”
”f2”))
”f1”)
But how is f2 to get this extra argument? Recall that each top-level procedure takes only one argument: the
user’s input. The (traditional) Web server can’t know that it has to hold on to this value and communicate it.
In practice, these values become the source of values to be stored in hidden fields. Every top-level
receiver has to be sensitive to creating and extracting these form values. Specifically, the converted Web
application has the following form:
(define (f2 user-input)
(local ([define •1 (get-form-field user-input ’n1 )]
[define •2 (get-form-field user-input ’n2 )])
(web-display
(+ •1
•2 ))))
(define (f1 user-input)
(web-read/r/fields ”Second number: ”
”f2”
user-input
(list ’n1 )))
”f1”)
where n1 and n2 are the names used in the form. The procedure web-read/r/fields takes the same first
two arguments as web-read/r. The third argument is the data structure representing the user’s input. This is
followed by a list of field names; these fields are extracted from the user input and inserted into the generated
HTML form using hidden fields.
How would f1 know which parameters must be passed to f2 using the hidden fields? These are precisely
those identifiers that are free in the receivers.
Exercise 16.4.1 Automate this transformation, i.e., write a program that implements it without the need for
human intervention.
Chapter 17
More Web Transformation
We have already seen how the application

(web-display
must be transformed into

(lambda (•1 )
(lambda (•2 )
(web-display
(+ •1 •2 ))))))
to execute on the Web. Let us now examine some more applications of a more complex flavor.
17.1 Transforming Library and Recursive Code

Suppose we have the procedure tally. It consumes a list of items and prompts the user for the cost of each
item. When done, it generates the sum of these items, which a programmer could display on the Web. The
code for tally is as follows:
(define (tally item-list)
(if (empty? item-list)
0
(+ (web-read (generate-item-cost-prompt (first item-list)))
(tally (rest item-list)))))
This version of tally is clearly not Web-friendly, due to the use of web-read, which we do not know how to
implement. We must therefore transform this code.
157
158 CHAPTER 17. MORE WEB TRANSFORMATION
The first thing to observe is that on its own, tally is not a complete program: it doesn’t do anything!
Instead, it is a library procedure that may be used in many different contexts. Because it has a Web inter-
action, however, there is the danger that at the point of interaction, the rest of the computation—i.e., the
computation that invoked tally—will be lost. To prevent this, tally must consume an extra argument, a re-
ceiver, that represents the rest of the pending computation. To signify this change in contract, we will use
the convention of appending /k to the name of the procedure and k to name the receiver parameter.
(define (tally/k item-list k)
0
(+ (web-read (generate-item-cost-prompt (first item-list)))
(tally (rest item-list)))))
What is the first thing this procedure does? It checks for whether the list is empty. Does this involve any
Web interaction? Clearly not; all the data and primitives are available locally. If the list is not empty, then
tally prompts for a user input through the Web. This must happen through web-read/k. What is the receiver
of this Web invocation? That is, what computation remains to be done? Clearly the recursive invocation of
tally; but there is also the receiver, k, which represents the rest of the waiting computation. Therefore, the
Web-friendly version of tally appears to become
0
(web-read/k (generate-item-cost-prompt (first item-list))
(lambda (v)
(+ v
(tally/k (rest item-list)
k))))))
We can read the second argument to web-read/k as saying: “Consume the value provided by the user and add
it to the value generated by the recursion. The receiver in the recursive invocation is the same k as before,
because the computation pending outside the procedure has not changed.”
This may look reasonable, but it suffers from an immediate problem. When the recursive call occurs, if
the list had two or more elements, then there will immediately be another Web interaction. Because this will
terminate the program, the pending addition will be lost! Therefore, the addition of v has to move into the
receiver fed to tally/k. In code:
0
(lambda (first-item-cost)
(lambda (tally-of-remaining-costs)
(k (+ first-item-cost
17.2. TRANSFORMING MULTIPLE PROCEDURES 159
tally-of-remaining-costs))))
That is, the receiver of the Web interaction is invoked with the cost of the first item. When tally/k is invoked
recursively, it is applied to the rest of the list. Its receiver must therefore receive the tally of costs of the
remaining items. That explains the pattern in the receiver.
The only problem is, where does a receiver ever get a value? We create larger-and-larger receivers on
each recursive invocation, but what ever invokes them?
Here is the same problem from a different angle (that also answers the question above). Notice that each
recursive invocation of tally/k takes place in the aftermath of a Web interaction. We have already seen how
the act of Web interaction terminates the pending computation. Therefore, when the list empties, where
is the value 0 going? Presumably to the pending computation—but there is none. Any computation that
would have been pending has now been recorded in k, which is expecting a value. Therefore, the correct
transformation of this procedure is
(k 0)
(lambda (first-item-cost)
(lambda (tally-of-remaining-costs)
(k (+ first-item-cost
tally-of-remaining-costs))))
Now we have a truly reusable abstraction. Whatever the computation pending outside the invocation of
tally/k, its proper Web transformation yields a receiver. If this receiver is fed as the second parameter to
tally/k, then it is guaranteed to be invoked with the value that tally would have produced in a non-Web (e.g.,
console) interaction. The pattern of receiver creation within tally/k ensures that no pending computation
gets lost due to the behavior of the Web protocol.
Exercise 17.1.1 There is a strong formal claim hidden behind this manual transformation: that the value
given to the initial k fed to tally/k is the same as that returned by tally in the non-Web version. Prove this.
17.2 Transforming Multiple Procedures

Suppose we have the procedure total+s&h. It consumes a list of items to purchase, queries the user for the
cost of each item, then generates another prompt for the corresponding shipping-and-handling cost,1 and
finally prints the result of adding these together. The procedure total+s&h relies on tally to compute the
sum of the goods alone.
(define (total+s&h item-list)
1 The term shipping and handling refers to a cost levied in the USA by companies that handle long-distance product orders
placed by the mail, phone and Internet. It is ostensibly the price of materials to package and labor to dispatch the ordered goods.
This rate is usually a (step) procedure of the cost of items ordered, and must hence be calculated at the end of the transaction.
(local ([define total (tally item-list)])

(+ (web-read (generate-s&h-prompt total))
total)))
Just as we argued in the transformation of tally, this procedure alone does not constitute a computation. It
must therefore consume an extra parameter, representing a receiver that will consume its result. Likewise,
it cannot invoke tally, because the latter performs a Web interaction; it must instead invoke tally/k, passing
along a suitable receiver to ensure no computation is lost.
(define (total+s&h/k item-list k)
(local ([define total (tally/k item-list ??? )])
(+ (web-read (generate-s&h-prompt total))
total)))
Reasoning as before, what is the first thing total+s&h/k does? It invokes a procedure to compute the
tally. Because this procedure involves a Web interaction, it must be invoked appropriately. That is, the
transformed procedure must take the form
(tally/k item-list
(lambda (tally-of-items)
??? )))
What is the pending computation? It is to bind the resulting value to total, then perform another Web
interaction:
(tally/k item-list
(local ([define total tally-of-items])
??? ))))
(Notice that the Web transformation has forced us to give names to intermediate results, thereby rendering
the name total unnecessary. We will, however, leave it in the transformed program so that the transformation
appears as mechanical as possible.) With the pending computation, this is
(tally/k item-list
(local ([define total tally-of-items])
(web-read/k (generate-s&h-prompt total)
(lambda (s&h-amount)
(k (+ s&h-amount
total))))))))
Notice how total+s&h/k had to create a receiver to pass to tally/k, the transformed version of tally. Reading
this receiver, it says to consume the value computed by tally/k (which it binds to tally-of-items), ask the user
17.3. TRANSFORMING STATE 161
to enter the shipping-and-handling amount, compute the final total, and convey this amount to the initial
receiver.
It’s easy to forget this last step: to apply k, the initial receiver supplied to total+s&h/k, to the final value.
Doing so would effectively “forget” all the computation that was waiting for the result of total+s&h/k, i.e.,
the computation awaiting the result of total+s&h in the original program. This is obviously undesirable.
You might worry that the local might be “forgotten” by the web-read/k that follows. But all we care
about is that the name total be associated with its value, and the receiver will take care of that (since it is a
closure, it must be closed over the value of total).
17.3 Transforming State

Suppose we want to write a program that keeps track of an account’s balance. On every invocation it presents
the current balance and asks the user for a change (i.e., deposit or withdrawal, represented respectively by
positive and negative numbers). We would like the Web application to look like this:
(define account
(local ([define balance 0])
(lambda ()
(begin
(set! balance (+ balance
(web-read
(format ”Balance: ã; Change” balance))))
(account)))))
Note that account is bound to a closure, which holds a reference to balance. Recall that mutable variables
introduce a distinction between their location and the value at that location. The closure closes over the
location, while the store is free to mutate underneath. Thus, even though balance always refers to the same
location, its value (the actual account balance) changes with each interaction.
How do we transform this program? Clearly the procedure bound to account must take an additional
argument to represent the remainder of the computation:
(define account/k
(lambda (k)
(begin
(set! balance (+ balance
(web-read
(format ”Balance: ã; Change” balance))))
(account/k ??? )))))
More importantly, we must move the web-read to be the first action in the procedure:
(define account/k
(lambda (k)
(begin
(web-read/k (format ”Balance: ã; Change” balance)
(lambda (v)
(begin
(set! balance (+ balance v))
(account/k ??? ))))))))
What’s left is to determine what argument to pass as the receiver in the recursive call. What new pending
activity have we created? The only thing the procedure does on each recursion is to mutate balance, which
is already being done in the receiver to the Web interaction primitive. Therefore, the only pending work is
whatever was waiting to be done before invoking account/k. This results in the following code:
(define account/k
(lambda (k)
(begin
(web-read/k (format ”Balance: ã; Change” balance)
(lambda (v)
(begin
(set! balance (+ balance v))
(account/k k))))))))
The closure created as the receiver for the Web interaction has a key property: it closes over the location
of balance, not the value. The value itself is stored in the heap memory that is kept alive by the Web server.
Exercise 17.3.1 If we wanted to run this application without any reliance on a custom server (Section 16.4),
we would have to put these heap data somewhere else. Can we put them in hidden fields, as we discussed
in Section 16? If so, test and make sure this works correctly, even in the face of the user interactions we
discussed in Section 15. If it fails, find an alternative to hidden fields that does work appropriately!
17.4 The Essence of the Transformation

By now we have performed the Web transformation often enough that we should begin to see a clear pattern
between the original and transformed procedure. Essentially, given a procedure f and its transformed version
f/k, we have that if the application
(f v1 . . . vm )
would have computed the answer a, then executing

(f/k v1 . . . vm k)
results in k being applied to the value a. This invariant ensures that, provided k is properly stored somewhere,
applying it to the value when it becomes available will leave the result of the computation unaffected.
17.5. TRANSFORMING HIGHER-ORDER PROCEDURES 163
17.5 Transforming Higher-Order Procedures

Suppose our Web program were the following:
(define (get-one-temp c)
(web-read (format ”Temperature in city ã” c)))
(web-display
(average
(map get-one-temp
(list ”Bangalore” ”Budapest” ”Houston” ”Providence”))))
(Assume we’ve define average elsewhere, and that it performs no Web input-output.) In principle, converting
this program is merely an application of what we studied in Section 17.1 and Section 17.2, but we’ll work
through the details to reinforced what you read earlier.
Transforming get-one-temp is straightforward:
(define (get-one-temp/k c k)
(web-read/k (format ”Temperature in city ã” c)
k))
This means we must invoke this modified procedure in the map. We might thus try
(web-display
(average
(map get-one-temp/k
Unfortunately, map is expecting its first argument, the procedure, to consume only the elements of the list;
it does not provide the second argument that get-one-temp/k needs. So Scheme reports
map: arity mismatch for procedure get-one-temp/k: expects 2 arguments, given 1
It therefore becomes clear that we must modify map also. Let’s first write map in full:
(define (map f l)
(if (empty? l)
empty
(cons (f (first l))
(map f (rest l)))))
Clearly we must somehow modify the invocation of f . What can we pass as a second argument? Here’s one
attempt:
(define (map f l)
(if (empty? l)
empty
(cons (f (first l) (lambda (x) x))
(map f (rest l)))))
That is, we’ll pass along the identity procedure. Does that work? Think about this for a moment.
Let’s try testing it. We get the following interaction:

web-read/k: run (resume) to enter number and simulate clicking Submit
>
This means the program has arrived at web-read/k for the first time. We run
> (resume)
which prompts us for an input. Suppose we enter 25. We then see
Temperature in city Bangalore: 25

25
>
It stopped: the program terminated without ever giving us a second Web prompt and asking us for the
temperature in another city!
Why? Because the value of the receiver stored in the hash table or box is the identity procedure. When
computation resumes (on the user’s submission), we expect to find the closure representing the rest of the
computation. Since the stored closure is instead just the identity procedure, the program terminates thinking
its task is done.
This gives us a pretty strong hint: the receiver we pass had better make some reference to map, and
indeed, had better continue the iteration. In fact, let’s think about where we get the first value for cons. This
value is the temperature for a city. It must therefore come from web-read/k. But that is exactly the value
that web-read/k supplies to its receiver. Therefore, everything starting with the cons onward must move to
within the closure:
(define (map f/k l)
(if (empty? l)
empty
(f/k (first l)
(lambda (v)
(cons v
(map f (rest l)))))))
This version is still not quite okay. This is because the recursive call invokes map, which suffers from
the same problem we have just discussed above. Indeed, running this version teminates after reading the
temperature for the second city, and returns just a list containing the second city’s temperature! Instead, it
must invoke a modified version of map, namely map/k, with an appropriate additional argument:
(define (map/k f/k l k)
(if (empty? l)
empty
(f/k (first l)
17.5. TRANSFORMING HIGHER-ORDER PROCEDURES 165
(lambda (v)
(cons v
(map/k f/k (rest l) ??? ))))))
We must determine what to pass as an argument in the recursive call. But before we do that, let’s study what
we have written carefully. When the first Web interaction results in a response, the server will invoke the
(lambda (v) · · ·). This conses the input temperature to the value of the recursive call. The recursive call will,
however, eventually result in an invocation of web-read/k. That invocation will halt the program. Once the
program halts, we lose record of the cons. So this program can’t work either! We must instead move the
cons inside the receiver, where it won’t be “forgotten”.
Given what we’ve learned in Section 17.4, it makes sense to think of the value given on invoking map/k
on the rest of the list as the list of temperatures for the remaining cities. Therefore, we simply need to cons
the temperature for the first city onto this result:
(if (empty? l)
empty
(f/k (first l)
(lambda (v)
(map/k f/k (rest l)
(lambda (v-rest)
(cons v v-rest)))))))
Now we’re ready to modify the main program. We had previously written
(web-display
(average
(map get-one-temp/k
We have to convert the invocation of map to one of map/k and, in turn, determine what to pass as the second
argument to map/k. Using the same reasoning we have employed before (in particular that, as written,
the web-display and average procedures will never execute, since they will be forgotten when the server
terminates the program), we know to write this:
(map/k get-one-temp/k
(list ”Bangalore” ”Budapest” ”Houston” ”Providence”)
(lambda (v)
(web-display
(average v))))
This program now runs through the four cities, accepts the temperatures in order, and produces . . . the empty
list.
What went wrong here? We can reason this way. The empty list cannot result from average (which must
produce a number), so we can reason that the initial receiver must never have been invoked at all. (We can
verify this by commenting out the definition of average and noticing that this doesn’t cause a problem: the
procedure is never invoked.) So it must be the case that the receiver supplied to map/k never made it any
further.
Studying map/k, we see the problem. Though the procedure consumes a receiver, that receiver never
gets used anywhere in its body. In fact, we should be passing the result of the cons to this procedure:
(if (empty? l)
empty
(f/k (first l)
(lambda (v)
(map/k f/k (rest l)
(lambda (v-rest)
(k (cons v v-rest))))))))
Everything now looks hunky-dory, so we run the program, enter the four temperatures, and still get . . . the
empty list!
Since there is really only one place in the program where we explicitly mention the empty list, we might
suspect it now. Indeed, the first branch in the conditional of map/k is indeed the culprit. When a value
becomes available, we should not return it. Why not? Because we know no procedure is awaiting it directly.
Why not? Because according to the Web protocol, any waiting procedures would have terminated when
the whole program terminated at the previous interaction point! Therefore, to return a value, a procedure
must instead hand the value to the receiver. (Of course, this merely reiterates the invariant we established
in Section 17.4.) That is, we want
(if (empty? l)
(k empty)
(f/k (first l)
(lambda (v)
(map/k f/k (rest l)
(lambda (v-rest)
(k (cons v v-rest))))))))
The moral of this lengthy story is that, to make a program Web-ready, we must (a) generate receivers that
capture pending computations, and (b) pass values to receivers instead of returning them. In rare cases, a
procedure will neither return a value nor generate additional pending computation—get-one-temp is a good
example—in which case, its transformed version will consume a receiver and pass along the same receiver
to other computations (as get-one-temp/k does).
Exercise 17.5.1 Why did we not transform average? In general, what principle guides whether or not we
transform a given procedure? (Make sure your principle also applies to map!)
17.6 Perspective on the Web Transformation

Notice three implications of the transformation the Web forces us to employ:
17.6. PERSPECTIVE ON THE WEB TRANSFORMATION 167
1. We have had to make decisions about the order of evaluation. That is, we had to choose whether
to evaluate the left or the right argument of addition first. This was an issue we had specified only
implicitly earlier; if our evaluator had chosen to evaluate arguments right-to-left, the Web program at
the beginning of this document would have asked for the second argument before the first! We have
made this left-to-right order of evaluation explicit in our transformation.
2. The transformation we use is global, namely it (potentially) affects all the procedures in the program
by forcing them all to consume an extra receiver as an argument. We usually don’t have a choice as
to whether or not to transform a procedure. Suppose f invokes g and g invokes h, and we transform
f to f /k but don’t transform g or h. Now when f /k invokes g and g invokes h, suppose h consumes
input from the Web. At this point the program terminates, but the last receiver procedure (necessary
to resume the computation when the user supplies an input) is the one given to f /k, with all record of
g and h erased.2
3. This transformation sequentializes the program. Given a nested expression, it forces the programmer
to choose which sub-expression to evaluate first (a consequence of the first point above); further, every
subsequent operation lies in the receiver, which in turn picks the first expression to evaluate, pushing
all other operations into its receiver; and so forth. The net result is a program that looks an awful lot
like a traditional procedural program. This suggests that this series of transformations can be used to
compile a program in a language like Scheme into one in a language like C!
Exercise 17.6.1 This presentation has intentionally left out the contracts on the procedures. Add contracts
to all the procedures—both the original programs and the Web versions.
Exercise 17.6.2 Adding contracts to the Web versions (Exercise 17.6.1) reveals a very interesting pattern
in the types of the receivers. Do you see a connection between this pattern and the behavior of the Web?
2 Indeed, we would have encountered an error even earlier, when the transformed version of f , namely f /k, tried to invoke g
with an extra receiver argument that g was not transformed to accept. In this even simpler way, therefore, the transformation process
has a cascading effect.
Chapter 18
Conversion into Continuation-Passing Style
Given how much the Web distorts the structure of an application, we would benefit from automating the
Web transformation. Then we could program in the lucid style of the initial versions of the programs, and
leave it to a “Web compiler” to transform these into their corresponding Web forms. This compiler would
be responsible for creating the receivers automatically. With a few further steps we could also implement
lifting and translation into the Web protocol (e.g., employing hidden fields).
To build such a compiler, we must better understand the receiver. Each expression’s receiver is a proce-
dure representing the rest of the computation to be performed when this expression completes. Furthermore,
we have seen that once we have converted applications use the receiver, it is useless to return values in the
traditional fashion, because all the code that would have otherwise received this return value is now encap-
sulated in the receiver. Traditionally, the return value propagates using the control stack, which represents
the remainder of the computation. The receiver is therefore a procedural representation of the stack.
Programs that obey this stylized form of generating and using receivers are said to be in continuation-
passing style, or CPS. Note that CPS is a style of program; many programs can be in CPS. Any program that
satisfies the programming pattern that its stack is represented procedurally can be said to be in CPS. More
to the point, any program that is not in CPS can be converted into CPS. This is in fact precisely the essence
of the Web compiler: it is a program that converts an arbitrary source program into CPS.1
18.1 The Transformation, Informally

The translations we have done by hand have been rather informal in nature. If we wish to automate this trans-
formation, we must adhere to more rigid rules. For instance, we have to agree upon a uniform representation
for all CPS terms, so any program processing them knows what to both generate and expect.
Consider our running example of addition. Given this term,
(web-read ”Second number: ”))
1 We will often find it convenient to have a verb form to represent the act of performing this transformation. It is common, par
abus de langage, to use CPS itself as a verb (as in, “CPS the following program”). Remember, though, that in proper usage, CPS
itself names a form of program, not an algorithm or procedure used to convert programs into that form.
169
170 CHAPTER 18. CONVERSION INTO CONTINUATION-PASSING STYLE
we hand-translated it to the following:

(lambda (l-val)
(lambda (r-val)
(+ l-val r-val)))))
This hand-translation is sufficient if this expression is the entire program. If, however, we wish to use it
in a broader context (e.g., as a sub-expression in a larger program), this does not suffice, since it does not
recognize that there may be a pending computation outside its own evaluation. How can we make this
translated expression reflect that fact? We have to introduce a procedure that consumes a receiver, and uses
that receiver to communicate the computed value:
(lambda (k)
(lambda (l-val)
(lambda (r-val)
( k (+ l-val r-val)))))))
In particular, if a transformer were (recursively) invoked upon the sub-expression
it would need to return
(lambda (k)
(lambda (l-val)
(lambda (r-val)
(k (+ l-val r-val)))))))
which can then be employed in the transformation of a larger program. (Observe that in the special case
where this is the entire program, applying this transformed term to the identity procedure results in the same
result as our original manual transformation.)
The following convention is therefore particularly useful: Every term, when converted to CPS, will be
represented as a procedure. This procedure will consume one argument, its receiver. The converted body
will communicate any value it computes to this receiver; if the body requires further computation, this will
be done using an augmented version of the receiver (i.e., no prior computation will be lost).
Let us similarly consider the transformation of a procedure application. For simplicity, we will assume
that procedures have only one parameter. Recall that not only the parameter position, but even the procedure
position, of an application can be an arbitrary expression and must therefore be reduced to a value. For
example, this expression could be an excerpt from the core of a parser:
((extract-from-table next-char)
18.1. THE TRANSFORMATION, INFORMALLY 171
(get-lookahead LOOKAHEAD-CONSTANT))
The next character is used as an index into a table, and the procedure obtained is applied to a constant
number of lookahead characters. The transformation of the procedure position should be
(extract-from-table/k next-char
(lambda (f-val)
??? ))
Similarly, the transformation of the argument position would be
(get-lookahead/k LOOKAHEAD-CONSTANT
(lambda (a-val)
??? ))
Given these two values (f-val, the procedure, and a-val, the actual parameter), we can now perform the
application. The following looks reasonable:
(k (f-val a-val))
Unfortunately, if the procedure itself performs a Web interaction, then that will halt the computation, erasing
any record of returning the value to k. Instead, therefore, k must be given as an argument to the procedure,
which can then use it appropriately. Putting together the pieces, the transformation into CPS of the procedure
application above is
(lambda (k)
(extract-from-table/k next-char
(lambda (f-val)
(get-lookahead/k LOOKAHEAD-CONSTANT
(lambda (a-val)
(f-val a-val k))))))
Reading this sequentially, it says to evaluate the procedure expression, and store its value in f-val; then
evaluate the argument, and store its value in a-val; and finally invoke the procedure on the argument. This
procedure’s receiver is the same as that of the procedure application itself.
What do we do with variables and simple constants? Recall that every term in CPS must be a procedure
that consumes a receiver. Since there is no computation to perform, the constant is simply communicated to
the receiver. Thus, the CPS representation of 3 is just
(lambda (k)
(k 3))
Suppose we are given the procedure (lambda (x) x). What does it look like in CPS? Since every lambda
expression is also a constant, it is tempting to use the same rule above for these expressions also, i.e.:
(lambda (k)
(k (lambda (x) x)))
However, the transformation is more subtle than that. Observe that a procedure application invokes the
procedure on two arguments, not one, whereas the procedure (lambda (x) x) consumes only one. What is
the second argument? It’s the dynamic receiver: i.e., the receiver at the time of the procedure application
(as opposed to its definition). Furthermore, we don’t want to ignore this receiver: it’s the stack active at
the point of procedure invocation, so we want to preserve it. This is in direct contrast to what we did with
environments—there we wanted the static environment—and more in line with our treatment of the store,
where we wanted the current store at the time of procedure application, and therefore did not close over
the store. Similarly, we do not close over the receiver at the point of procedure creation. The transformed
version instead reads
(lambda (k)
(k (lambda (x dyn-k)
( (lambda (k) (k x))
dyn-k))))
where the boxed expression is the result of transforming the body, namely, x. This is equivalent (when the
inner application finally happens) to
(lambda (k)
(k (lambda (x dyn-k)
(dyn-k x))))
That is, it’s a procedure that accepts a value and a dynamic receiver and sends the value to that receiver.
This, of course, is precisely the behavior we would expect of the identity function.
That leaves only one kind of expression to translate, namely web-read. We are in a happy position: to
translate web-read into web-read/k we need access to the receiver, which is precisely what our transforma-
tion has given us access to. Therefore, an expression such as
(web-read ”First number: ”)
simply becomes
(lambda (k)
k))
18.2 The Transformation, Formally

For this material, please switch to the PLAI - Pretty Big language level.
We are now ready to write this transformation formally.

In general, every program transformer is a kind of compiler. The transformer of a program into CPS
is, however, a peculiar kind of compiler: it is a source-to-source transformation. That is, the transformer
consumes a program in a source language and generates a new program in the same language (assuming the
language has closures and some other properties copacetic to the use of CPS).
Scheme provides a particularly convenient mechanism, called a macro, for writing source-to-source
transformers. In this section, we present the transformer into CPS using macros but without much elaboration
of the macro mechanism. We will discuss macros in considerably more detail in Section 36.
18.2. THE TRANSFORMATION, FORMALLY 173
Macros are triggered by a keyword that signals the transformation. We will use the keyword cps, so that
for every term of the form (cps e), the expression e will be transformed into CPS. Every macro definition
consists of a preamble naming the new keyword being defined and a list of rules dictating how to trans-
form source expressions. The rules employ pattern-matching to extract portions of the input expression for
insertion into the output.
The cps macro’s header has the following form:
(define-syntax cps
(syntax-rules (+ lambda web-read)
This says we are defining a macro for the keyword cps, and that this macro will treat +, lambda and web-
read specially: they must match exactly against an input for the corresponding rule to apply. The macro’s
rules are as follows. Note that these are simply the generalization of the instances we studied above.
First, the rule for addition:
[(cps (+ e1 e2))
(lambda (k)
((cps e1) (lambda (l-val)
((cps e2) (lambda (r-val)
(k (+ l-val r-val)))))))]
This says that whenever the term has the form
(cps (+ e1 e2))
where + is expected to match precisely, the sub-expressions (which can be arbitrary code) are named e1 and
e2 respectively. The corresponding transformed term has the form
(lambda (k)
(k (+ l-val r-val)))))))
where e1 and e2 are inserted into the transformed term. Since they are within a (cps ·) expression, they will
be transformed into CPS recursively by the same macro.
The transformer for a procedure application is as follows:
[(cps (f a))
(lambda (k)
((cps f ) (lambda (f-val)
((cps a) (lambda (a-val)
(f-val a-val k))))))]
following the pattern: convert the procedure expression and apply it to a receiver expecting the procedure
value; do likewise for the argument; and when both values are handy, apply the procedure to the current
receiver and the argument value.
The corresponding transformation for a procedure is therefore
[(cps (lambda (a) body))
(lambda (k)
(k (lambda (a dyn-k)
((cps body) dyn-k))))]
Recall that every user procedure must now consume the dynamic receiver in addition to its formal parameter.
As we have discussed, the transformation of web-read is especially simple:
[(cps (web-read prompt))
(lambda (k)
(web-read/k prompt k))]
Finally, if all other terms fail to match, we assume that the source program is a simple value (namely an
identifier or a non-procedural constant such as a string or number). The corresponding term in CPS consumes
a receiver (to maintain the consistency of the interface) and immediately sends the value to the receiver:
[(cps v)
(lambda (k) (k v))]
Figure 18.1 presents the entire CPS macro.
Example Given these rules, our transformer will convert

into
(lambda (k1)
((lambda (k2)
(web-read/k ”First number: ” k2))
(lambda (l-val)
((lambda (k3)
(web-read/k ”Second number: ” k3))
(lambda (r-val)
(k1 (+ l-val r-val)))))))
This may look rather more complicated than we are used to. However, we merely need to perform the inner
procedure applications, substituting the known receivers for k2 and k3. Doing this yields:
(lambda (k1)
(lambda (l-val)
(lambda (r-val)
(k1 (+ l-val r-val)))))))
which is exactly what we expect!
18.3. TESTING 175
The Fischer CPS Transformation

The CPS transformer we have studied here is one of the very oldest, due to Michael Fischer. In
the three decades since Fischer defined this transformation, there has been considerable research
into building a better CPS transformation. Why? This version, while easy to understand, introduces
a considerable amount of overhead: look at all the procedures in the output that weren’t in the
source program! In principle, we would like a transformer that approximates the cleanliness of
hand-translation: this would be useful both to humans who must read the output and to programs
that must process it. This is what the better transformers strive to achieve.
18.3 Testing
The code in Figure 18.1 also defines define-cps so you can write top-level definitions that are translated into
CPS . Thus, you can write the following program, reminiscent of the cascading transformation we discussed
in Section 17.6:
(define-cps (g x) (+ x x))
(define-cps (h f ) (lambda (x) (f x)))
(define-cps (dummy x) ((h g) 10))
It also defines the run construct, which is useful for testing code converted to CPS. You can use it to execute
(run (dummy 1729))
and observe the computation terminate with the value 20.
Exercise 18.3.1 Suppose, instead, the CPS rule for a procedure were
(lambda (k)
((cps body) k))))]
i.e., the transformed procedure ignored the dynamic receiver and used the static one instead. What impact
would this have on program behavior? Predict, then run and check!
(define-syntax define-cps
(syntax-rules ()
[(define-cps (f arg) body)
(define-cps f (lambda (arg) body))]
[(define-cps v val)
(define v ((cps val) (lambda (x) x)))]))
(define-syntax cps
(syntax-rules (+ lambda web-read)
[(cps (+ e1 e2))
(lambda (k)
(k (+ l-val r-val)))))))]
(lambda (k)
((cps body) dyn-k))))]
[(cps (web-read prompt))
(lambda (k)
(web-read/k prompt k))]
[(cps (f a))
(lambda (k)
((cps f ) (lambda (f-val)
((cps a) (lambda (a-val)
(f-val a-val k))))))]
[(cps v)
(lambda (k) (k v))]))
(define-syntax run
(syntax-rules ()
[(run e) ((cps e)
(lambda (x)
(error ”terminating with value” x)))]))
Figure 18.1: Implementation of CPS Converter

Chapter 19
Programming with Continuations
For this material, please switch to the PLAI - Pretty Big language level.
In Section 18 we saw how conversion to CPS restores the Web programming interface we desire: pro-
grammers can use web-read and not have to “invert” the structure of the program. While in principle this
accomplishes the task, in practice conversion to CPS has several disadvantages:
1. It requires access to the source of the entire program. If a procedure is defined in a library for which
we don’t have access to the source, or is perhaps written in a different language (as map often is), then
the CPS translator will either fail to run or will produce potentially erroneous output (i.e., code that
does not properly restore the state of the computation).
2. By replacing the machine’s stack with an explicit representation in the form of receivers, it inhibits
optimizations built into compilers and microprocessor architectures.
3. As we will see in Section 20.4, executing a program in CPS also assumes that the run-time system
will not needlessly create stack frames (since the stack is entirely represented by the receiver). Since
many languages (such as C and Java) do anyway, the program consumes memory unnecessarily. In
an extreme case, a Java or C program that would have executed without exhausting memory will run
out of memory after conversion into CPS.
The first of these problems is particularly compelling, since it affects not only performance but even cor-
rectness. We would benefit from an operation that automatically constructs the receiver at any point during
the program’s execution, instead of expecting it to have already been created through a static compilation
process.
Some programming languages, notably Scheme, have such an operation. This operation creates a rep-
resentation of the “rest of the computation” (which is what the receiver represented) as a procedure of one
argument. Giving that procedure a value causes the remaining computation to resume with that value. This
procedure is called a continuation. This explains where CPS obtains its name, but note that the program does
not need to be transformed a priori; the continuation is created automatically and on-the-fly. As you might
imagine, creating (or “capturing”) a continuation simply involves copying the stack, though there are less
and more efficient ways of obtaining the same effect.
177
178 CHAPTER 19. PROGRAMMING WITH CONTINUATIONS
Adding continuations to a language makes it easy to create a better Web programming protocol, as we
shall see. But just as laziness—which was already present in the shell but was essentially an extra-lingual
feature (since programmers could not explicitly control it)—, once exposed as a feature in a programming
language, gave programmers immense power in numerous contexts, so do continuations. We will explore
this power in greater detail.
19.1 Capturing Continuations

In Scheme, we create a value representing the continuation using one of two related constructs. The tradi-
tional form is called call/cc, short for “call with current continuation”. call/cc consumes a procedure as an
argument, and invokes this procedure with a continuation. That is, uses of call/cc typically look like
(call/cc
(lambda (k)
· · · k · · ·)) ;; k is the continuation
Because the extra lambda is extremely tiresome, however, Scheme provides a nicer interface to capturing
the current continuation: you may instead equivalently write
(let/cc k
· · · k · · ·)) ;; k is bound to the continuation
Note that let/cc is a binding construct: it introduces a new scope, binding the named identifier (k, above) in
that context. For the rest of this material, we’ll use let/cc rather than call/cc.1
19.2 Escapers
Let’s write some programs using continuations. What is the value of this program?
(let/cc k
(k 3))
We must first determine the continuation bound to k. This is the same procedure as the value of the receiver
in CPS. Since in this case, there is no computation waiting to be done outside the let/cc expression, the
receiver would be the initial receiver, namely the identity function. Therefore, this receiver is
(lambda (•)
•)
Applying this to 3 produces the answer 3.
Consider this program:
(+ 1
(let/cc k
1 So why does Scheme offer call/cc, ghastly as it looks? Historically, the original standards writers were loath to add new binding
forms, and the use of lambda meant they didn’t need to create one. Also, call/cc lets us create some incredibly clever programming
puzzles that we can’t write quite as nicely with let/cc alone! Ask us for some.
19.3. EXCEPTIONS 179
(k 3)))
What is the continuation this time? It’s
(lambda (•)
(+ 1 •))
(all the code “outside the parens”). This procedure, when invoked, yields the value 4. Because the contin-
uation is the rest of the computation, we want to halt with the value 4. But this looks confusing, because
substitution gives
(+ 1
(k 3)) ;; where k is bound to (lambda (•) (+ 1 •))
= (+ 1
((lambda (•)
(+ 1 •))
3))
= (+ 1
(+ 1 3))
which performs the addition twice, producing the answer 5. The problem is that we’re effectively applying
the continuation twice, whereas computation should halt after it has been applied once. We will use a special
notation to reflect this: lambda↑ will represent a procedure that, when its body finishes computing, halts
the entire computation. We’ll call these escaper procedures, for obvious reasons. That is, the continuation
is really
(lambda↑ (•)
(+ 1 •))
so the expression
(+ 1
((lambda↑ (•)
(+ 1 •))
3))
evaluates to 4, with the outermost addition ignored (because we invoked an escaper).
19.3 Exceptions
Let’s consider a similar, but slightly more involved, example. Suppose we are deep in the midst of a com-
putation when we realize we are about to divide by zero. At that point, we realize that we want the value of
the entire expression to be one. We can use continuations to represent this pattern of code:
(define (f n)
(+ 10
(∗ 5
(let/cc k
(/ 1 n)))))
(+ 3 (f 0))
The continuation bound to k is
(lambda↑ (•)
(+ 3
(+ 10
(∗ 5
•))))
but oops, we’re about to divide by zero! Instead, we want the entire division expression to evaluate to one.
Here’s how we can do it:
(define (f n)
(+ 10
(∗ 5
(let/cc k
(/ 1 (if (zero? n)
(k 1)
n))))))
so that • in the continuation is substituted with 1, we bypass the division entirely, and the program can
continue to evaluate.
Have you seen such a pattern of programming before? But of course: k here is acting as an exception
handler, and the invocation of k is raising the exception. A better name for k might be esc:
(define (f n)
(+ 10
(∗ 5
(let/cc esc
(/ 1 (if (zero? n)
(esc 1)
n))))))
which makes pretty clear what’s happening: when you invoke the continuation, it’s as if the entire let/cc
expression that binds esc should be cut out of the program and replaced with the value passed to esc, i.e., its
as if the actual code for f is really this:
(define (f n)
(+ 10
(∗ 5
1)))
In general, this “cut-and-paste” semantics for continuations is the simplest way (in conjunction with escaper
procedures) of understanding a program that uses continuations.
There was, incidentally, something sneaky about the program above: it featured an expression in the
body of a let/cc that did not invoke the continuation. That is, if you can be sure the user of f will never pass
an argument of 0, it’s as if the body of the procedure is really
19.4. WEB PROGRAMMING 181
(define (f n)
(+ 10
(∗ 5
(let/cc esc
(/ 1 n)))))
but we haven’t talked about what happens in programs that don’t invoke the continuation. In fact, that’s
quite easy: the value of the entire let/cc expression is exactly that of the value of its body, just as when you
don’t actually raise an exception.
19.4 Web Programming

Now that we’re getting a handle on continuations, it’s easy to see how they apply to the Web. We no longer
need the procedure web-read/k; now, web-read can be implemented directly to do the same thing that web-
read/k was expected to do. web-read captures the current continuation, which corresponds to the second
argument supplied to web-read/k (except the continuation is now captured automatically, instead of having
to be supplied as an explicit second argument).
The rest of the implementation is just as before: it stores the continuation in a fresh hash table entry,
and generates a URL containing that hash table entry’s key. The launcher extracts the continuation from the
hash table and applies it to the user’s input. As a result, all the programs we have written using web-read
are now directly executable, without the need for the CPS transformation.
19.5 Producers and Consumers

A number of programs follow a producer-consumer metaphor: one process generates (possibly an infinite
number of) values, while another consumes them as it needs new ones. We saw several examples of this
form in Haskell, for instance. Many client-server programs are like this. Web programs have this form (we,
the user, are the supplier of inputs—and on some Web sites, the number really does seem quite close to
infinite . . . ). I/O, in general, works this way. So it’s worth understanding these processes at a deeper level.
To avoid wrangling with the complexities of these APIs, we’ll reduce this to a simpler problem. We’d
like to define a producer of symbols (representing a sequence of cities to visit) that takes one argument,
send, which masks the details of the API, and sends a sequence of city names on demand:
(define (route-producer send)
(begin
(send ’providence)
(send ’houston)
(send ’bangalore)))
That is, when we first invoke route-producer, it invokes send with the value ’providence—and halts. When
we invoke it again, it sends ’houston. The third time we invoke it, it sends the value ’bangalore. (For now,
let’s not worry about what happens if we invoke it additional times.)
What do we supply as a parameter to elicit this behavior? Clearly, we can’t simply supply an argument
such as (lambda (x) x). This would cause all three send operations to happen in succession, returning the
value of the third one, namely ’bangalore. Not only is this the wrong value on the first invocation, it also
produces the same answer no matter how many times we invoke route-producer—no good.
In fact, if we want to somehow suspend the computation, it’s clear we need one of these exception-like
operations so that the computation halts prematurely. So a simple thing to try would be this:
(let/cc k (route-producer k))
What does this do? The continuation bound to k is
(lambda↑ (•) •)
Substituting it in the body results in the following program:
(begin
((lambda↑ (•) •)
’providence)
((lambda↑ (•) •)
’houston)
((lambda↑ (•) •)
’bangalore))
(all we’ve done is substitute send three times). When we evaluate the first escaper, the entire computation
reduces to ’providence—and computation halts! In other words, we see the following interaction:
> (let/cc k (route-producer k))
’providence
This is great—we’ve managed to get the program to suspend after the first send! So let’s try doing this a
few more times:
’providence
’providence
’providence
Hmm—that should dampen our euphoria a little.
What’s the problem? We really want route-producer to “remember” where it was when it sent a value
so it can send us the next value when we invoke it again. But to resume, we must first capture a snapshot of
our computation before we sent a value. That sounds familiar. . . .
Going back to route-producer, let’s think about what the continuation is at the point of invoking send
the first time. The continuation is
(lambda↑ (•)
(begin
•
(send ’houston)
19.5. PRODUCERS AND CONSUMERS 183
(with send substituted appropriately). Well, this looks promising! If we could somehow hang on to this
continuation, we could use it to resume the producer by sending ’houston, and so forth.
To hang on to the computation, we need to store it somewhere (say in a box), and invoke it when we run
the procedure the next time. What is the initial value to place in the box? Well, initially we don’t know what
the continuation is going to be, and anyway we don’t need to worry because we know how to get a value out
the first time (we just saw how, above). So we’ll make the box initially contain a flag value, such as false.
Examining what’s in the box will let us decide whether we’re coming through the first time or not. If we are
we proceed as before, otherwise we need to capture continuations and what-not.
Based on this, we can already tell that the procedure is going to look roughly like this:
(define route-producer
(local ([define resume (box false)])
(lambda (send)
(if (unbox resume)
;; then
there’s a continuation present—do something!
;; else
(begin
(send ’houston)
(send ’bangalore)))))))
where the box bound to resume stores the continuation. Note that resume is outside the lambda but in its
scope, so the identifier is defined only once and all invocations of the procedure share it.
If (unbox resume) doesn’t evaluate to false,2 that means the box contains a continuation. What can we
do with continuations? Well, really only one thing: invoke them. So we must have something like this if the
test succeeds:
((unbox resume) · · ·)
But what is that continuation? It’s one that goes into the midst of the begin (recall we wrote it down above).
Since we really don’t care about the value, we may as well pass the continuation some kind of dummy value.
Better to pass something like ’dummy, which can’t be confused for the name of a city. So the procedure
now looks like
(lambda (send)
(if (unbox resume)
((unbox resume) ’dummy)
(begin
(send ’houston)
(send ’bangalore))))))
2 In Scheme, only false fails a conditional test; all other values succeed.
Of course, we haven’t actually done any of the hard work yet. Recall that we said we want to capture
the continuation of send, but because send is given as a parameter by the user, we can’t be sure it’ll do the
right thing. Instead, we’ll locally define a version of send that does the right thing. That is, we’ll rename the
send given by the user to real-send, and define a send of our own that invokes real-send.
What does our send need to do? Obviously it needs to capture its continuation. Thus:
(lambda (real-send)
(local ([define send (lambda (value-to-send)
(let/cc k
· · ·))])
(if (unbox resume)
(begin
(send ’houston)
What do we do with k? It’s the continuation that we want to store in resume:
(set-box! resume k)
But we also want to pass a value off to real-send:
(real-send value-to-send)
So we just want to do this in sequence (observe that we can’t do these in the other order):
(begin
(set-box! resume k)
(real-send value-to-send))
When the client next invokes route-producer, the continuation stored in the box bound to resume is that of
invoking (our locally defined) send within the body. . . which is exactly where we want to resume computa-
tion! Here’s our entire procedure:
(lambda (real-send)
(local ([define send (lambda (value-to-send)
(let/cc k
(begin
(set-box! resume k)
(real-send value-to-send))))])
(if (unbox resume)
(begin
19.5. PRODUCERS AND CONSUMERS 185
(send ’houston)
Here’s the interaction we wanted:3

’providence
’houston
’bangalore
It’s a bit unwieldy to keep writing these let/ccs, so we can make our lives easier as follows:
(define (get producer)

(let/cc k (producer k)))
(we could equivalently just define this as (define get call/cc)) so that
> (get route-producer)

’providence
’houston
’bangalore
Exercise 19.5.1 How would you define the same operators in Haskell?
Exercise 19.5.2 Before you read further: do you see the subtle bug lurking in the definitions above?
Hint: We would expect to be able to invoke (get route-producer) three times and combine the result into a
list.
3 It’s critical that you work through the actual continuations by hand.
Continuations and “Goto” Statements

What we’re doing here has already gone far beyond the simple exception pattern we saw earlier.
There, we used continuations only to ignore partial computations. Now we’re doing something
much richer. We’re first executing part of a computation in a producer. At some point, we’re
binding a continuation and storing it in a persistent data structure. Then we switch to performing
some completely different computation (in this case, the top-level requests for the next city). At
this point, the partial computation of the producer has seemingly gone away, because we invoked
an escaper to return to the top-level. But by accessing the continuation in the data structure, we are
able to resume a prior computation: that is, we can not only jump “out”, as in exceptions, but we
can even jump back “in”! We do “jump back in” in the Web computations, but that’s a much tamer
form of resumption. What we’re doing here is resuming between two separate computations!
Some people compare continuations to “goto” statements, but we should think this through. Goto
statements can usually jump to an arbitrary line, which means they may continue execution with
nonsensical state (e.g., variables may not be properly initialized). In contrast, a continuation only
allows you to resume a computational state you have visited before. By being more controlled in
that sense, continuations avoid the worst perils of gotos.
On the other hand, continuations are far more powerful than typical goto statements. Usually, a
goto statement can only transfer control to a lexically proximate statement (due to how compilers
work). In contrast, a computation can represent any arbitrary prior computation and, irrespective of
lexical proximity, the programmer can resurrect this computation. In short, continuations are more
structured, yet much more powerful, than goto statements.
19.6 A Better Producer

The producer shown above is pretty neat—indeed, we’d like to be able to use this in more complex compu-
tations. For instance, we’d like to initialize the producer, then write
(list (get route-producer)
(get route-producer)
(get route-producer))
Evaluating this in DrScheme produces. . . an infinite loop!
What went wrong? It’s revealing to use this version of the program instead:
(list (let/cc k (route-producer k))
(let/cc k (route-producer k)))
After a while, click on Break in DrScheme. If you try it a few times, sooner or later you’ll find that the break
happens at the second let/cc—but never the third! In fact, if you ran this program instead:
(begin
(printf ”got here!ñ”)
19.6. A BETTER PRODUCER 187

(begin
(printf ”here too!ñ”)
(let/cc k (route-producer k))))
you’d find several got heres in the in the Interactions window, but no here toos. That’s revealing!
What’s going on? The first continuation bound to real-send in route-producer is
(lambda↑ (•)
(list •
Since we are invoking route-producer for the first time, in its body we eventually invoke send on ’providence.
What’s the computation that we capture and store in resume? Let’s compute this step-by-step. The top-level
computation is
Substituting the body of route-producer in place of the first invocation, and evaluating the conditional, we
get
(list (begin
(send ’houston)
(send ’bangalore))
We’re playing fast-and-loose with scope—technically, we should write the entire local that binds send be-
tween the first let/cc and begin, as well as all the values bound in the closure—but let’s be a bit sloppy for
the sake of readability. Just remember what real-send is bound to: it’ll be relevant in a little while!
When send now captures its continuation, what it really captures is
(lambda↑ (•)
(list (begin
•
(send ’houston)
send stores this continuation in the box, then supplies ’providence to the continuation passed as real-send.
This reduces the entire computation to
(list ’providence

The second continuation therefore becomes
(lambda↑ (•)
(list ’providence
•
which becomes the new value of real-send. The second time into route-producer, however, we do have a
value in the box bound to resume, so we have to extract and invoke it. We do, resulting in
((lambda↑ (•)
(list (begin
•
(send ’houston)
’dummy)
which makes
(list (begin
’dummy
(send ’houston)
the entire computation (because we invoked a lambda↑). This looks like it should work fine. When we
invoke send on ’houston, we capture the second computation, and march down the line.
Unfortunately, a very subtle bug is lurking here. The problem is that the send captured in the continuation
is closed over the old value of real-send, because the real-send that the continuation closes over is from the
previous invocation of route-producer. Executing the begin inside send first stores the new continuation in
the box bound to resume:
(lambda↑ (•)
(list (begin
•
but then executes the old real-send:
((lambda↑ (•)
(list •
’houston)
In other words, the entire computation now becomes
(list ’houston
which is basically the same thing we had before! (We’ve got ’houston instead of ’providence, but the
important part is what follows it.) As you can see, we’re now stuck in a vicious cycle. Now we see why it’s
the second sub-expression in the list where the user break occurs—it’s the one that keeps evaluating over
and over (the first is over too quickly to notice, while the computation never gets to the third).
This analysis gives us a good idea of what’s going wrong. Even though we’re passing in a fresh, correct
value for real-send, the closure still holds the old and, by now, wrong value. We need the new value to
somehow replace the value in the old closure.
This sounds like the task of an assignment statement, to mutate the value bound to real-send. So we’ll
first box and refer to the value:
(lambda (real-send)
(local ([define send-to (box real-send)]
[define send (lambda (value-to-send)
(let/cc k
(begin
(set-box! resume k)
((unbox send-to) value-to-send))))])
(if (unbox resume)
(begin
(send ’houston)
Now we have a box whose content we can replace. We should replace it with the new value for real-send.
Where is that new value available? It’s available at the time of invoking the continuation—that is, we could
pass that new value along instead of passing ’dummy. Where does this value go? Using our “replacing text”
semantics, it replaces the entire (let/cc · · ·) expression in the definition of send. Therefore, that expression
evaluates to the new value to be put into the box. All we need to do is actually update the box:
(lambda (real-send)

(set-box! send-to
(let/cc k
(begin
(set-box! resume k)
((unbox send-to) value-to-send)))))])
(if (unbox resume)
((unbox resume) real-send)
(begin
(send ’houston)
This time, even though send invokes the old closure, the box in that closure’s environment now has the
continuation for the new resumption point. Therefore,
> (list (get route-producer)
(get route-producer)
(get route-producer))
(providence houston bangalore)
There’s just one problem left with this code: it deeply intertwines two very separate concerns, one of
which is sending out the actual values (which is specific to whatever domain we are computing in) and the
other of which is erecting and managing the continuation scaffold. It would be nice to separate these two.
This is actually a lot easier than it looks. Simply factor out the body into a separate procedure:
(define (route-producer-body send)
(begin
(send ’houston)
(This, you recall, is what we set out to write in the first place!) Everything else remains in a general
procedure that simply consumes (what used to be) “the body”:
(define (general-producer body)
(lambda (real-send)
(set-box! send-to
(let/cc k
(begin
(set-box! resume k)
((unbox send-to) value-to-send)))))])
(if (unbox resume)

((unbox resume) real-send)
(body send))))))
Introducing this separation makes it easy to write a host of other value generators. They can even
generate a potentially infinite number of values! For instance, here’s one that generates all the odd positive
intgers:
(define (odds-producer-body send)

(local ([define (loop n)
(begin
(send n)
(loop (+ n 2)))])
(loop 1)))
We can make an actual generator out of this as follows:
(define odds-producer (general-producer odds-producer-body))
and then invoke it in a number of different ways:
> (get odds-producer)

1
3
5
7
9
or (assuming an initialized program)
> (+ (get odds-producer)

(get odds-producer)
(get odds-producer)
(get odds-producer)
(get odds-producer))
25
Exercise 19.6.1 Are you sure we’re done? If we try to extract more items than a producer has (for instance,
invoking (get route-producer) a fourth time), what happens?
19.7 Why Continuations Matter

Continuations are valuable because they enable programmers to create new control operators. That is, if a
language did not already have (say) a producer-consumer construct, continuations make it possible for the
programmer to build them manually and concisely. More importantly, these additions can be encapsulated
in a library, so the rest of the program does not need to be aware of them and can use them as if they were
built into the language. If the language did not offer continuations, defining some of these libraries would
require a whole-program conversion into CPS. This would not only place an onerous burden on the users of
these operations, in some cases this may not even be possible (for instance, when some of the source is not
available). This is, of course, the same argument we made for not wanting to rely on CPS for the Web, but
the other examples illustrate that this argument applies in several other contexts as well.
Chapter 20
Implementing Continuations
Now that we’ve seen how continuations work, let’s study how to implement them in an interpreter.
The first thing we’ll do is change our representation of closures. Instead of using a structure to hold the
pieces, we’ll use Scheme procedures instead. This will make the rest of this implementation a lot easier.1
First, we modify the datatype of values. Only two rules in the interpreter need to change: that which
creates procedure values and that which consumes them. These are the fun and app cases, respectively.
Both rules are very straightforward: one creates a procedure and wraps it in an instance of closureV, while
the other extracts the procedure in the closureV structure and applies it to the argument. The code is in
Figure 11.2. Note that this change is so easy only because functions in the interpreted language closely
match those of Scheme: both are eager, and both obey static scope. (As we’ve discussed before, this is our
the usual benefit of using meta interpreters that match the interpreted language.)
Recall that any program can be converted to CPS—therefore, so can the interpreter. Performing this
conversion gives us access to the continuation at every stage, which we can then make available to the
programmer through a new language construct. We will therefore approach the implementation of continua-
tions in two steps: first making them explicit in the interpreter, then providing access to them in an extended
language.
20.1 Representing Continuations

We’ll assume that the interpreter takes an extra argument k, a receiver. The receiver expects the answer from
each expression’s interpretation. Thus, if the interpreter already has a value handy, it supplies that value
to the receiver, otherwise it passes a (possibly augmented) receiver along to eventually receive the value of
that expression. The cardinal rule is this: We never want to use an invocation of interp as a sub-expression
of some bigger expression. Instead, we want interp to communicate its answer by passing it to the given
k. (Recall the Web situation: every time you invoke a procedure that interacts, the program terminates.
Suppose interp were executing on the Web. Then any non-trivial context that invokes interp will disappear
when the interpreter halts. Therefore, all pending computation must be bundled into the receiver.)
1 Yes,we’re switching to a more meta interpreter, but this is acceptable for two reasons: (1) by now, we understand procedures
well, and (2) the purpose of this lecture is to implement continuations, and so long as we accomplish this without using Scheme
continuations, we won’t have cheated.
193
194 CHAPTER 20. IMPLEMENTING CONTINUATIONS
Let’s consider the simplest case, namely numbers. A number needs no further evaluation: it is already a
value. Therefore, we can feed the number to the awaiting receiver.
[num (n) (k (numV n))]
Identifiers and closures, already being values, look equally easy:
[id (v) (k (lookup v env))]
[fun (param body)
(k (closureV (lambda (arg-val)
(interp body (aSub param arg-val env)))))]
Now let’s tackle addition. The rule traditionally looks like this:
[add (l r) (numV+ (interp l env) (interp r env))]
The naive solution might be to transform it as follows:
[add (l r) (k (numV+ (interp l env) (interp r env)))]
but do we have a value immediately handy to pass off to k? We will after interpreting both sub-expressions
and adding their result, but we don’t just yet. Recall that we can’t invoke interp in the midst of some
larger computation. Therefore, we need to bundle that remaining computation into a receiver. What is that
remaining computation?
We can calculate the remaining computation as follows. In the naive version, what’s the first thing the
interpreter needs to do? It must evaluate the left sub-expression.2 So we compute that first, and move all the
remaining computation into the receiver of that invocation of the interpreter:
[add (l r) (interp l env
(lambda (lv)
(k (num+ lv (interp r env)))))]
In other words, in the new receiver, we record the computation waiting to complete after reducing the
left sub-expression to a value. However, this receiver is not quite right either. It has two problems: the
invocation of interp on r has the wrong arity (it supplies only two arguments, while the interpreter now
consumes three), and we still have an invocation of the interpreter in a sub-expression position. We can
eliminate both problems by performing the same transformation again:
(lambda (lv)
(interp r env
(lambda (rv)
(k (num+ lv rv))))))]
That is, the first thing to do in the receiver of the value of the left sub-expression is to interpret the right
sub-expression; the first thing to do with its value is to add them, and so on.
Can we stop transforming now? It is true that interp is no longer in a sub-expression—it’s always the
first thing that happens in a receiver. What about the invocation of numV+, though? Do we have to transform
it the same way?
2 Notice that once again, we’ve been forced to choose an order of evaluation, just as we had to do to implement state.
20.1. REPRESENTING CONTINUATIONS 195
It depends. When we perform this transformation, we have to decide which procedures are primitive
and which ones are not. The interpreter clearly isn’t. Usually, we treat simple, built-in procedures such
as arithmetic operators as primitive, so that’s what we’ll do here (since num+ is just a wrapper around
addition).3 Had we chosen to transform its invocation also, we’d have to add another argument to it, and so
on. As an exercise, you should consider performing this transformation.
Now let’s tackle the conditional. Clearly the interpretation of the test expression takes place in a sub-
expression position, so we’ll need to lift it out. An initial transformation would yield this:
(interp test env
(lambda (tv)
(if (num-zero? tv)
(interp truth env ??? )
(interp falsity env ??? ))))]
Do we need to transform the subsequent invocations of interp? No we don’t! Once we perform the test, we
interpret one branch or the other, but no code in this rule is awaiting the result of interpretation to perform
any further computation—the result of the rule is the same as the result of interpreting the chosen branch.
Okay, so what receivers do they use? The computation they should invoke is the same computation that
was awaiting the result of evaluating the conditional. The receiver k represents exactly this computation.
Therefore, we can replace both sets of ellipses with k:
(interp test env
(lambda (tv)
(if (num-zero? tv)
(interp truth env k)
(interp falsity env k))))]
That leaves only the rule for application. The first few lines of the transformed version will look familiar,
since we applied the same transformation in the case of addition:
(interp fun-expr env
(lambda (fun-val)
(interp arg-expr env
(lambda (arg-val)
((closureV-p fun-val) arg-val) ))))]
All we have to determine is what to write in place of the box.
Is the code in the box still valid? Well, the reason we write interpreters is so that we can experiment!
How about we just try it on a few expressions and see what happens?
3 Using our Web analogy,the question is which primitives might arguably invoke a Web interaction. Ones that use the Web must
be transformed and be given receivers to stash on the server, while ones that don’t can remain unmolested. Arithmetic, clearly,
computes entirely locally.
> (interp-test ’5 5)
#t
> (interp-test ’{+ 5 5} 10)
#t
> (interp-test ’{with {x {+ 5 5}} {+ x x}} 20)
procedure interp: expects 3 arguments, given 2 ...
Oops! DrScheme highlights the interpretation of the body in the rule for fun.
Well, of course! The interpreter expects three arguments, and we’re supplying it only two. What should
the third argument be? It needs to be a receiver, but which one? In fact, it has to be whatever receiver is
active at the time of the procedure invocation. This is eerily reminiscent of the store: while the environment
stays static, we have to pass this extra value that reflects the current state of the dynamic computation. That
is, we really want the rule for functions to read
[fun (param body)
(k (closureV (lambda (arg-val dyn-k)
(interp body (aSub param arg-val env) dyn-k))))]
(What happens if we use k instead of dyn-k in the invocation of the interpreter? Try it and find out!)
Correspondingly, application becomes
(lambda (fun-val)
(lambda (arg-val)
arg-val k)))))]
The core of the interpreter is in Figure 20.1.
What, incidentally, is the type of the interpreter? Obviously it now has one extra argument. More
interestingly, what is its return type? It used to return values, but now. . . it doesn’t return! That’s right:
whenever the interpreter has a value, it passes the value off to a receiver.
20.2 Adding Continuations to the Language

At this point, we have most of the machinery we need to add continuations explicitly as values in the
language. The receivers we have been implementing are quite similar to the actual continuations we need.
They appear to differ in two ways:
1. They capture what’s left to be done in the interpreter, not in the user’s program.
2. They are regular Scheme procedures, not lambda↑ procedures.
However,
20.2. ADDING CONTINUATIONS TO THE LANGUAGE 197
1. They capture what’s left to be done in the interpreter, not in the user’s program. Because the interpreter
closes over the expressions of the user’s program, invocations of the interpreter simulate the execution
of the user’s program. Therefore, the receivers effectively do capture the continuations of the user’s
program.
2. They are regular Scheme procedures, not lambda↑ procedures. We have taken care of this through
the judicious use of a programming pattern. Recall our discussion of the type of the revised inter-
preter? The interpreter never returns—thereby making sure that no computation awaits the value of
the receiver, which is effectively the same as computation terminating when the receiver is done.
In other words, we’ve very carefully set up the interpreter to truly represent the continuations, making it
easy to add continuations to the language.
Adding Continuations to Languages

Different languages take different approaches to adding continuations. Scheme’s is the most spar-
tan. It add just one primitive procedure, call/cc. The resulting continuations can be treated as if
they were procedures, so that procedure application does double duty. DrScheme slightly enriches
the language by also providing let/cc, which is a binding construct, but it continues to overload
procedure application.
The language SML uses callcc (which is not a binding construct) to capture continuations, and adds
a throw construct to invoke continuations. Consequently, in SML, procedure application invokes
only procedures, and throw invokes only continuations, making it possible for a type-checker to
distinguish between the two cases.
It’s possible that a language could have both a binding construct like let/cc and a separate throw-like
construct for continuation invocation, but there don’t appear to be any.
To implement continuations, we will take the DrScheme approach of adding a binding construct but
overloading procedural application:
<KCFAE> ::= ...
| {<KCFAE> <KCFAE>}
| {bindcc <id> <KCFAE>}
(This grammar is just for purposes of illustration. It would be easy to add a throw construct instead of
overloading application.)
To implement continuations, we need to add one new rule to the interpreter, and update the existing rule
for application. We’ll also add a new kind of type, called contV.
How does bindcc evaluate? Clearly, we must interpret the body in an extended environment:
[bindcc (cont-var body)
(interp body
(aSub cont-var
(contV ??? )
env)
k)]
The receiver used for the body is the same as the receiver of the bindcc expression. This captures the
intended behavior when the continuation is not used: namely, that evaluation proceeds as if the continuation
were never captured.
What kind of value should represent a continuation? Clearly it needs to be a Scheme procedure, so we
can apply it later. Functions are represented by procedures of two values: the parameter and the receiver
of the application. Clearly a continuation must also take the value of the parameter. However, the whole
point of having continuations in the language is to ignore the receiver at the point of invocation and instead
employ the one stored in the continuation value. Therefore, we can just ignore the receiver at the point
of application. The invoked continuation instead uses, as its own receiver, that captured at the point of its
creation:
(interp body
(aSub cont-var
(contV (lambda (val)
(k val)))
env)
k)]
(Note again the reliance on Scheme’s static scope to close over the value of k.) This makes the modification
to the application clause very easy:
(lambda (fun-val)
(lambda (arg-val)
(type-case KCFAE-Value fun-val
[closureV (c) (c arg-val k)]
[contV (c) (c arg-val)]
[else (error ”not an applicable value”)])))))]
Notice the very clear contrast between function and continuation application: function application employs
the receiver at the point of application, whereas continuation application employs the receiver at the point
of creation. This difference is dramatically highlighted by this code.
One last matter: what is the initial value of k? If we want to be utterly pedantic, it should be all
the computation we want to perform with the result of interpretation—i.e., a representation of “the rest of
DrScheme”. In practice, it’s perfectly okay to use the identity function. Then, when the interpreter finishes
its task, it invokes the identity function, which returns the value to DrScheme. For the purpose of testing,
it’s even better to use a procedure that prints a value and then halts the program entirely (as we discussed in
Section 16.3)—that lets us test whether we converted the interpreter into CPS properly.
And that’s it! In these few lines, we have captured the essence of continuations. (The heart of the
interpreter is in Figure 20.2.) Note in particular two properties of continuations that are reflected by, but
perhaps not obvious from, this implementation:
20.3. ON STACKS 199
• To reiterate: we ignore the continuation at the point of application, and instead use the continuation
from the point of creation. This is the semantic representation of the intuition we gave earlier for
understanding continuation programs: “replace the entire let/cc expression with the value supplied
to the continuation”. Note, however, that the captured continuation is itself a dynamic object—it
depends on the entire history of calls—and thus cannot be computed purely from the program source
without evaluation. In this sense, it is different from the environment in a closure, which can partially
be determined entirely statically (that is, we can determine which identifiers are in the environment,
though it is undecidable what their values will be.)
• The continuation closes over the environment; in particular, its body is scoped statically, not dynami-
cally.
Exercise 20.2.1 Add web-read to KCFAE.
20.3 On Stacks
Let’s re-examine the procedure application rule in the interpreter:
(lambda (fun-val)
(lambda (arg-val)
arg-val k)))))]
Observe how the receiver of the value of the function position, and that of the value of the argument position,
are both new procedures, whereas when the procedure application finally happens (in the latter receiver),
the third argument is just k.
Given that the receivers represent the stack at every point, this should seem rather strange: wasn’t the
stack meant to hold a record of procedure invocations? And if so, why is the receiver “growing” when
evaluating the function and the argument, but “shrinking” back to its original size when the application
actually happens?
We have said that the receiver corresponds directly to the stack. In particular, the receiver is a procedure
that may refer to another procedure (that it closes over), which may refer to another procedure (that it closes
over), and so on. Each of these procedures represents one stack frame (sometimes called an activation
record, because it records an extant activation of a procedure in the running program). Returning a value
“pops” the stack; since we have made the stack explicit, the equivalent operation is to pass the value to be
returned to the receiver.
We therefore see, from this pattern, that the stack is used solely to store intermediate results. It plays no
part in the actual invocation of a function. This probably sets on its head everything you have been taught
about stacks until now. This is an important and, perhaps, startling point:
Stacks are not necessary for invoking functions.
The stack only plays a role in evaluating the argument to the function (and, in a language with first-class
functions, evaluating the function position itself); once that argument has been reduced to a value (in an
eager language), the stack has no further role with respect to invoking a function. The actual invocation is
merely a jump to an address holding the code of the function: it’s a goto.
Converting a program to CPS thus accomplishes two things. First, it exposes something—the stack—
normally hidden during the evaluation process; this is an instance of reflection. The transformation also
makes this a value that a programmer can manipulate directly (even changing the meaning of the program
as a result); this is known as reification.
Reflection and Reification

Reflection and reification are very powerful programming concepts. Most programmers encounter
them only very informally or in a fairly weak setting. For instance, Java offers a very limited
form of reflection (a programmer can, for instance, query the names of methods of an object),
and some languages reify very low-level implementation details (such as memory addresses in C).
Few languages reify truly powerful computational features; the ones that do enable entirely new
programming patterns that programmers accustomed only to more traditional languages usually
can’t imagine. Truly smart programmers sometimes create their own languages with the express
purpose of reifying some useful hidden element, and implement their solution in the new language,
to create a very powerful kind of reusable abstraction. A classic instance of this is Web programmers
who have reified stacks to enable a more direct programming style.
20.4 Tail Calls

Converting the program to CPS helps us clearly see which calls are just gotos, and which ones need stack
build-up. The ones that are just gotos are those invocations that use the same receiver argument as the one
they received. Those that build a more complex receiver are relying on the stack.
Procedure calls that do not place any burden on the stack are called tail calls. Converting a program
to CPS helps us identify tail calls, though it’s possible to identify them from the program source itself. An
invocation of g in a procedure f is a tail call if, in the control path that leads to the invocation of g, the
value of f is determined by the invocation of g. In that case, g can send its value directly to whoever is
expecting f ’s value; this verbal description is captured precisely in the CPSed version (since f passes along
its receiver to g, which sends its value to that receiver). This insight is employed by compilers to perform
tail call optimization (abbreviated TCO, and sometimes referred to as last call optimization), whereby they
ensure that tail calls incur no stack growth.
Here are some consequences of TCO:
• With TCO, it no longer becomes necessary for a language to provide looping constructs. Whatever
was previously written using a custom-purpose loop can now be written as a recursive procedure. So
long as all recursive calls are tail calls, the compiler will convert the calls into gotos, accomplishing
the same efficiency as the loop version. For instance, here’s a very simple version of a for loop,
written using tail calls:
20.4. TAIL CALLS 201
(define (for init condition change body result)

(if (condition init)
(for (change init)
condition
change
body
(body init result))
result))
By factoring out the invariant arguments, we can write this more readably as
(define (for init condition change body result)
(local [(define (loop init result)
(if (condition init)
(loop (change init)
(body init result))
result))]
(loop init result)))
To use this as a loop, write

(for 10 positive? sub1 + 0)
which evaluates to 55. It’s possible to make this look more like a traditional for loop using macros,
which we will discuss in Section 36. That aside, notice how similar this is to a fold operator! Indeed,
foldl employs a tail call in its recursion, meaning it is just as efficient as looping constructs in more
traditional languages.
• Put differently, thanks to TCO, the set of looping constructs is extensible, not limited by the imagina-
tion of the language designer. In particular, with care it becomes easy to create loops (or iterators)
over new data structures without suffering an undue performance penalty.
• While TCO is traditionally associated with languages such as Scheme and ML, there’s no reason they
must be. It’s perfectly possible to have TCO in any language. Indeed, as our analysis above has
demonstrated, TCO is the natural consequence of understanding the true meaning of function calls.
A languages that deprives you of TCO is cheating you of what is rightfully yours—stand up for your
rights! Because so many language designers and implementors habitually mistreat their users by
failing to support TCO, however, programmers have become conditioned to think of all function calls
as inherently expensive, even when they are not.
• A special case of a tail call is known as tail recursion, which occurs when the tail call within a
procedure is to the same procedure. This is the behavior we see in the procedure for above. Keep in
mind, however, that tail recursion optimization is only a special case. While it is an important special
case (since it enables the implementation of linear loops), it is neither the most interesting case nor,
more importantly, the only useful one.
Sometimes, programmers will find it natural to split a computation across two procedures, and use tail
calls to communicate between them.4 This leads to very natural program structures. A programmer
using a language like Java is, however, forced into an unpleasant decision. If they split code across
methods, they pay the penalty of method invocations that use the stack needlessly. But even if they
combine the code into a single procedure, it’s not clear that they can easily turn the two code bodies
into a single loop. Even if they do, the structure of the code has now been altered irrevocably. Consider
the following example:
(define (even? n)
(if (zero? n)
true
(odd? (sub1 n))))
(define (odd? n)
(if (zero? n)
false
(even? (sub1 n))))
Try writing this entirely through loops and compare the resulting structure.
Therefore, even if a language gives you tail recursion optimization, remember that you are getting less
than you deserve. Indeed, it sometimes suggests an implementor who realized that the true nature of
function calls permitted calls that consumed no new stack space but, due to ignorance or a failure of
imagination, restricted this power to tail recursion only. The primitive you really want a language to
support is tail call optimization. With it, you can express solutions more naturally, and can also build
very interesting abstractions of control flow patterns.
• Note that CPS converts every program into a form where every call is a tail call!
Exercise 20.4.1 If, in CPS, every call is a tail call, and the underlying language supports TCO (as Scheme
does), does the CPS version of a program run in constant stack space even if the original does not? Discuss.
Exercise 20.4.2 Java does not support TCO. Investigate why not.
20.5 Testing
You might think, from last time’s extended example of continuation use, that it’s absolutely necessary to
have state to write any interesting continuation programs. While it’s true that most practical uses of the full
power of continuations (as opposed to merely exceptions, say) do use state, it’s possible to write some fairly
complicated continuation programs without state for the purposes of testing our interpreter. Here are some
4 They may not even communicate mutually. In the second version of the loop above, for invokes loop to initiate the loop. That
call is a tail call, and well it should be, otherwise the entire loop will have consumed stack space. Because Scheme has tail calls,
notice how effortlessly we were able to create this abstraction. If the language supprted only tail recursion optimization, the latter
version of the loop, which is more pleasant to read and maintain, would actually consume stack space against our will.
20.5. TESTING 203
such programs. You should, of course, first determine their value by hand (by writing the continuations in
the form of lambda↑ procedures, substituting, and evaluating).
First, a few old favorites, just to make sure the easy cases work correctly:
1. {bindcc k 3}
2. {bindcc k {k 3}}
3. {bindcc k {+ 1 {k 3}}}
4. {+ 1 {bindcc k {+ 1 {k 3}}}}
And now for some classic examples from the continuations lore:
1. {{bindcc k
{k {fun {dummy}
3}}}
1729}
2. {bindcc k
{k
{k
{k 3}}}}
3. {{{bindcc k k}
{fun {x} x}}
3}
4. {{{{bindcc k k}
{fun {x} x}}
{fun {x} x}}
3}
The answer in each case is fairly obvious, but you would be cheating yourself if you didn’t hand-evaluate
each of these first. This is painful, but there’s no royal road to understanding! (If in doubt, of course, run
their counterpart in Scheme.)
(define-type CFAE-Value
[numV (n number?)]
;; interp : CFAE Env receiver → doesn’t return

(define (interp expr env k)
(type-case CFAE expr
(lambda (lv)
(interp r env
(lambda (rv)
(k (num+ lv rv))))))]
(interp test env
(lambda (tv)
(if (num-zero? tv)
[fun (param body)
(lambda (fun-val)
(lambda (arg-val)
arg-val k)))))]))
Figure 20.1: Making Continuations Explicit

20.5. TESTING 205
(define-type KCFAE-Value
[numV (n number?)]
[closureV (p procedure?)]
[contV (c procedure?)])
;; interp : KCFAE Env receiver → doesn’t return

(define (interp expr env k)
(type-case KCFAE expr
(lambda (lv)
(interp r env
(lambda (rv)
(k (num+ lv rv))))))]
(interp test env
(lambda (tv)
(if (num-zero? tv)
[fun (param body)
(lambda (fun-val)
(lambda (arg-val)
(type-case KCFAE-Value fun-val
[closureV (c) (c arg-val k)]
[contV (c) (c arg-val)]
[else (error ”not an applicable value”)])))))]
(interp body
(aSub cont-var
(contV (lambda (val)
(k val)))
env)
k)]))
Figure 20.2: Adding Continuations as Language Constructs

Part VIII
Memory Management
207
Chapter 21
Automatic Memory Management
21.1 Motivation
We have seen in Section 13 that, when making data structures explicit in the store, the store (or heap) has
allocated data that are no longer necessary. Ideally, we would like this space reclaimed automatically. This
would enable us to program as if we had an infinite amount of memory yet, so long as we never exceed the
actual virtual memory capacity at any instant, the program would never halt with an out-of-memory error.
Whose responsibility is it to reclaim these locations? It can’t be the responsibility of the allocating
procedure. For instance, suppose we had written a procedure filter-pos that filters out the positive numbers
in a given list. This procedure cannot know know whether or not its caller needs the argument list again.
That procedure may pass the same list to some other procedure, and so on.
Even if it the chain of responsibility is clear, memory reclamation is often frustrating because it interferes
with the flow of control in the program, mixing high-level algorithmic description with low-level resource
management. Let’s say we knew for sure that the input list would not be used any longer. The procedure
filter-pos could then attempt to reclaim the list’s elements as follows:
(define (filter-pos l)
(cond
[(empty? l) empty]
[else
(begin
(reclaim-memory! (first l))
(if (> (first l) 0)
(cons (first l) (filter-pos (rest l)))
(filter-pos (rest l))))]))
There is a subtle bug in this program, but let’s focus on a simpler problem with it: while this reclaims each
first element, it doesn’t reclaim the conses that constitute the input list. We might therefore try
(cond
[(empty? l) empty]
209
210 CHAPTER 21. AUTOMATIC MEMORY MANAGEMENT
[else
(begin
(reclaim-memory! (rest l))
(if (> (first l) 0)
(filter-pos (rest l))))]))
Unfortunately, this version duplicates the bug! Once we reclaim the first and rest of the list, we can no
longer refer to those elements. In particular, in a concurrent system (and most software today is concurrent),
the moment we reclaim the memory, another process might write into it, so if we access the memory we
might get nonsensical output. And even otherwise, in general, if we reclaim and then perform a procedure
call (in this case, a recursive one), when we return (as we do in the first branch, to perform the cons) that
heap location may have since been overridden with other values. So this is a problem even in the absence of
concurrency. We must therefore instead write
(cond
[(empty? l) empty]
[else
(local ([define result
(if (> (first l) 0)
(filter-pos (rest l)))])
(begin
(reclaim-memory! (rest l))
result))]))
While this version is no longer succeptible to the problems we discussed earlier, it has introduced a signif-
icant new problem. Whereas earlier filter-pos was tail-recursive in cases when the list element not positive,
now filter-pos is never tail recursive. In fact, the problem we see here is a common problem with loop-like
programs: we must hold on to the value being passed in the recursive call so we can reclaim it after the call
completes, which forces us to destroy any potential for tail-call optimizations.
In short, even when we know who is responsible for reclaiming data, we face several problems:
• The program structure may be altered significantly.
• Concurrency, or even just other function calls, can expose very subtle reclamation errors.
• Loops often lose tail-calling behavior.
• It becomes much harder to define simple abstractions. For example, we would need two versions of
a filter-pos procedure, one that does and one that doesn’t reclaim its argument list. In turn, every
procedure that wants to invoke filter-pos must choose which version to invoke. And so on up the
21.2. TRUTH AND PROVABILITY 211
abstraction hierarchy. (Can you see how the number of possible options can grow exponentially in the
number of arguments?)
These problems, furthermore, assume we can even know which procedure is responsible for managing
every datum, which is a very strong assumption. Sometimes, two procedures may share a common resource
(think of the pasteboard in a typical windowing system, which is shared between multiple applications),
which means it’s no single unit’s responsibility in particular.
At any rate, reasoning about these chains of ownership is hard, and making the wrong decisions leads
to numerous insidious errors. Therefore, it would be better if we could make this the responsibility of
the run-time system: that is, whatever is responsible for allocating memory should also be responsible for
reclaiming memory when it is no longer necessary. That is, we usually prefer to program with automated
memory management, colloquially referred to by the much more colorful term, garbage collection.
21.2 Truth and Provability

In the previous paragraph, we have given the garbage collector the responsibility of reclaiming memory
“when it is no longer necessary”. This puts a very significant pressure on the garbage collector: the collector
must know whether or not a programmer is going to use a datum again or not. In other words, garbage
collection becomes an artificial intelligence problem.
This highlights a common tension that arises in computer science, and especially in programming lan-
guage design: that between truth and provability. This might sound like a very profound philosophical
issue—and it is—but you are already very familiar with it from math courses, where a professor asked you
to prove something she knew to be true, but you were unable to construct an actual line of reasoning for it!
We see this tension in several other places, too: for example, the type checker may not know whether or not
a given operation will succeed, while the programmer has a complex line of reasoning that justifies it; and
the optimizer in a compiler might not be able to prove that an expensive expression is equivalent to a less
expensive one (you might notice that this goes back to our discussion about referential transparency).
Anyway, the garbage collector obviously cannot know a programmer’s intent, so it needs to approxi-
mate her intent as best as possible. Furthermore, this approximation must meet a few tightly intertwined
properties. To understand these, let us consider a few extreme implementations of collectors.
The first collector reclaims absolutely no garbage. Obviously it runs very quickly, and it never acciden-
tally reclaims something that it should not reclaim. However, this is obviously useless. This suggests that a
collector must demonstrate
utility The collector’s approximation must identify enough garbage to actually help computation continue.
Another collector avoids this problem by reclaiming all data in memory, irrespective of whether or not they
are necessary for future computation. This, too, is obviously not very useful, because the computation would
soon crash. Therefore, a collector must exhibit
soundness The collector must never reclaim a datum that is used by a subsequent computation.
A third collector, wanting to avoid both of the above perils, halts at every datum and computes a very
complex simulation of the program’s execution to determine whether or not the program will access this
datum again. It has to consider all execution paths, both branches of each conditional, and so on. This, too,
would be unacceptable: a collector must manifest
efficiency The collector should run sufficiently quickly so that programmers do not get frustrated (and
therefore turn back to manual memory management).
In practice, garbage collectors reconcile these demands thanks to a very useful approximation of truth:
reachability. That is, a collector begins with a set of memory locations called the root set; this typically
includes all the heap references on the stack and in the current registers. From the root set, the collector
sweeps the heap to determine which objects are reachable: if object o is reachable, then all objects that o
refers to in its fields are also reachable—and so on, recursively. All reachable objects are called live, and
survive garbage collection; the collector reclaims the storage allocated to all other objects.
With a little reflection, we realize that reachability is an excellent approximation of truth. If an object is
reachable, then there is (in effect) some sequence of field dereferences and function or method invocations
that can use its value. Since the programmer may have written exactly such a sequence of invocations, the
collector should not reclaim the object. If, on the other hand, an object is not reachable in this fashion, no
sequence of dereferences and invocations can use its value.1 Therefore, the garbage collector can safely
reclaim its space.
Reachability is, of course, not always a strong enough approximation to truth. For instance, consider the
following program fragment:
(define v (make-vector 1000))
(define k (vector-length v))
;; rest of program
Suppose the rest of the program never references v.2 In that case, after k has been given its value the
space consumed by the vector bound to v should be reclaimed; but since v is a global variable, it is always
reachable, so the collector cannot reclaim it. In general, large data structures bound to global variables are
invariably candidates for space leakage, which is what we call the phenomenon of a collector not reclaiming
space that we know is no longer necessary. (Notice the difference between truth and provability coming into
play very strongly.) Tracing space leaks is sometimes subtle, but it is often as simple as looking at values
bound to global and static variables and, when those values are no longer necessary, mutating the variable
to a value like null (in Java) or (void) (in Scheme).
Notice, by the way, the asymmetry in our justification for why tracing is a reasonable approximation
to truth. Unreachable objects will not be used so they can always be reclaimed safely, whereas reachable
objects may be used again so we must allow them to persist. In fact, a collector can sometimes reason about
these “maybe” cases. For instance, consider the following program:
(local ([define v (make-vector 1000)]
[define k (vector-length v)])
· · ·)
1 This claim makes a certain important assumption about the underlying programming language that is not always valid: it
applies to languages like Java and Scheme but not to C and C++. Do you see it?
2 This assumes that the rest of the program text is known. Modern languages support features such as dynamic loading, which is
the ability to extend the program during its execution.

21.2. TRUTH AND PROVABILITY 213
Now suppose the body of this expression doesn’t reference v. Because v is not global, as soon as the value of
k has been computed, the implementation can safely set the value of v to a null or void value, thus making the
vector formerly bound to v a candidate for reclamation immediately. Many implementations for languages
that employ garbage collection do in fact perform such “safe for space” optimizations.
Part IX
Semantics
215
Chapter 22
Shrinking the Language
How small can a programming language be? We’ve already seen that Scheme is sufficient for expressing
a large number of computations quite concisely. The version of Scheme we’ve seen so far is nevertheless
quite small; here are most of the features we’ve used:
x y z · · · ;; variables
’a ’b ’c · · · ;; symbols
0 1 2 · · · ;; numbers
+ - ∗ · · · ;; arithmetic operators
define-type type-case
cons first rest
cond if
true false zero? and or
() ;; function application
local
let/cc call/cc
define lambda
We have seen in Section 18 that we can express continuations using just procedures and applications, in the
form of CPS. Similarly, it is easy to see that we can express various type constructors with lists by being
disciplined about their use. (To wit, we would put a tag at the beginning of the list indicating what type we
are representing, and check the tag before every access to avoid ill-typed operations.) We have also seen in
Section 6.3 that local definitions can be expressed in terms of function definitions and application.
That still leaves several constructs, which we can organize into groups related by purpose:
• variables, procedure definitions, applications
• numbers and arithmetic
• Boolean constants and operations
• lists and other aggregate data structures
217
218 CHAPTER 22. SHRINKING THE LANGUAGE
In what follows, we will methodically eliminate most of these features until we are left with a minimal set
that is surprisingly small. In particular, we will demonstrate that the first group—variables, procedures and
application—can encode all the remaining language constructs.
As a motivating example, consider the following definition of the factorial function:
(define (fact n)
(if (zero? n)
1
(∗ n (fact (sub1 n)))))
This contains most of the features under discussion.
22.1 Encoding Lists

Let’s first consider lists. Though we don’t need lists to define factorial, they are a useful indicator of how
we would handle compound data structures.
We will consider a data structure even simpler than lists, namely pairs. Once we have pairs, we can
construct lists quite easily: each pair’s second element is another pair, representing the next “link” in the
list. Therefore, we must demonstrate how to implement pair, left and right without using any Scheme data
structure primitives (such as lists or new types).
How can we do this? The pair constructor must consume two values and return something. What can
we return? Since we are not trying to eliminate procedures, perhaps it can return a procedure of some sort.
That is, every instance of
(pair A B)
in the program becomes
(lambda . . .)
We will write this as
(pair A B) ≡ (lambda . . .)
where ≡ represents textual substitution. So what should the argument and body of the procedure be? Should
it return A? Sometimes, yes, if the program wants the first value in the pair; at other times, it should return
B. In other words, the response needs to be parameterized to depend on the selector. We can express this
neatly thus:
(pair A B) ≡ (lambda (selector) (selector A B))
This defers the problem to defining the selectors, but there is hope now. The operator that chooses the left is
simply
left ≡ (lambda (A B) A)
and analogously,
right ≡ (lambda (A B) B)
22.2. ENCODING BOOLEAN CONSTANTS AND OPERATIONS 219
Finally, we must rewrite every use of these primitives to be in object-verb rather than in verb-object
form: that is, because the pair is itself a procedure that consumes the selector as an argument, we must write
(e left)
in place of
(first e)
for every e that will evaluate to a pair, and analogously replacing rest with right.1
Exercise 22.1.1 Our knowledge of computer science tells us that the left and right fields must consume
space somewhere; yet in this encoding, they appear to consume no space at all. Is that true (and, if so, how
is that possible)?
22.2 Encoding Boolean Constants and Operations

What is a conditional? At its simplest, it makes a choice between two computations based on some criterion.
We are used to a construct, such as if, as the operator that makes that choice, based on some Boolean value.
But what if, instead, the choice were made by the Boolean value itself?
Here’s an analogous situation. In a pure object-oriented language, every value is an object, including
the Boolean values. That is, we can think of values true and false as subtypes of a Boolean type,
each endowed with a choose method. The choose method takes two objects as arguments, each with
a run method that represents the rest of the computation. A true object invokes the run method of its
first argument object, while a false object invokes run in its second argument. Thus, effectively, the
conditional is implemented by the process of dynamic dispatch (which chooses which object to run when
choose is invoked on a Boolean value).
We can apply this insight into encoding conditionals entirely in terms of procedures. Every instance of
(if C T F)
is rewritten as
(C T F)
which we will henceforth write as
(if C T F) ≡ (C T F)
(once again, read ≡ as textual replacement). where we assume that C will always evaluate to one of the two
Boolean values. We therefore reduce the problem to one of defining Boolean values that correctly implement
the choice operation.
Defining the Boolean values is quite easy given our preceding discussion of objects. The value repre-
senting truth must consume the two options, ignore (by convention) the second and use only the first:
yes ≡ (lambda (T F) T)
1 This shift from algebra’s verb-object convention is, of course, familiar to object-oriented programmers.
Likewise,
no ≡ (lambda (T F) F)
Exercise 22.2.1 This encoding of Booleans assume the use of lazy evaluation. Provide an example that
illustrates this, and demonstrate how we can remove this dependency.
Exercise 22.2.2 Define and and or in terms of our limited set of primitives. Do your definitions perform
short-circuiting?
Exercise 22.2.3 Is it purely coincidental that left and yes have the same definition?
22.3 Encoding Numbers and Arithmetic

Having dispatched of lists and Booleans, we are now ready to tackle numbers. Let’s agree to limit our
attention to the natural numbers (an integer no smaller than zero).
What is the essence of a natural number? It is a counting object: it tells us how many instances there are
of some discrete entity. While it is conventional to use “Arabic” numerals (0, 1, 2, 3, . . . ) to represent these
numbers, there are many other representations available (for instance, the whole numbers—natural numbers
strictly bigger than zero—can be represented using Roman numerals: I, II, III, IV, . . . ). Even compared with
this variety of notations, though, the representation we will define here is truly striking.
Let’s think about “one-ness”. The number one represents many things. It captures the number objects
in a collection of one book, of one maple leaf, of one walrus, of one cricket ball. It also represents the act of
applying a function to a value.
Which function? Which value? Any function and any value will do just fine; indeed, to avoid having to
decide, we can simply make them parameters. That is, we can represent one as
one ≡ (lambda (f ) (lambda (x) (f x)))
This is the most abstract way we have of saying “the act of applying some function to some argument once”.
If we supply the concrete arguments add1 and 0 for f and x, respectively, we get the expected numeral 1
from Scheme. But we can also supply square and 5, respectively, to get the numeral 25, and so on.
If that’s what represents one-ness, what represents two-ness and three-ness? Why, the same kind of
thing: respectively, two applications and three applications of some function to some argument:
two ≡ (lambda (f ) (lambda (x) (f (f x))))
three ≡ (lambda (f ) (lambda (x) (f (f (f x)))))
and so on. Indeed, supplying add1 and 0 to each of these numerals yields expected Scheme numeral. We
should therefore intuitively think of f as an “add one” operation and x as the “zero” constant, but in fact
we can supply any operation and base constant we want (and, in what follows, we will in fact exploit this
abstract representation).
Since we want to represent the natural numbers, we must be able to represent zero, too. The pattern
above suggests the following numeral:
zero ≡ (lambda (f ) (lambda (x) x))
22.3. ENCODING NUMBERS AND ARITHMETIC 221
and indeed, that is a numeral that represents zero applications of f to x. (As we said, think of x as “zero”, so
a procedure that returns x unmolested is a good representative for 0.)
Now that we have numerals to represent the natural numbers, it’s time to define operations on them.
Let’s begin with incrementing a number by one. That is, we expect
(succ one) ≡ (succ (lambda (f ) (lambda (x) (f x))))
to yield the equivalent of

(lambda (f ) (lambda (x) (f (f x))))
This looks nasty: it appears we must perform “surgery” on the procedure to “insert” another application of f .
This is impossible since the procedures are opaque objects (in a computer, represented by the implementation
as some sequence of bits we may not even understand).
It’s important to note that what we want is the equivalent of the representation of two: that is, we want
a numeral that represents two-ness. Here is another term that has the same effect:
(lambda (f )
(lambda (x)
(f ((one f ) x))))
That is, it applies one to f and x, obtaining the effect of applying f to x once. It then applies f again to the
result. This has the same net effect as the more concise representation of two. By the same line of reasoning,
we can see that this pattern always represents the act of incrementing a number:
succ ≡
(lambda (n)
(lambda (f )
(lambda (x)
(f ((n f ) x)))))
Now we can tackle addition. Observe the following arithmetic result:
n times
z }| {
m + n = 1 + 1 + · · · + 1 +m
Let’s try putting this in words. To add m and n, we add one n times to m. That is, we apply an operation
that adds one, n times, to a base value of m. How do we iterate anything n times? That’s exactly what
the numeral for n represents: the act of performing n applications. The numeral expects two values: the
operation to apply, and the base value. The operation we want is the addition of one, which we’ve just
defined; the base value is the other addend. Therefore:
sum ≡
(lambda (m)
(lambda (n)
((n succ) m)))
A similar insight gives us multiplication:

n times
z }| {
m × n = m + m + · · · + m +0
From this, and employing a similar line of reasoning, we can define

prod ≡
(lambda (m)
(lambda (n)
((n (sum m)) zero)))
(We can see from this definition the wisdom of having the binary operators accept one argument at a time.)
It’s easy to see that we can define other additive operators inductively. But how about subtraction?
This seems to create an entirely new level of difficulty. Addition seemed to need the ability to modify the
numeral to apply the first argument one more time, but we found a clever way of applying it “from the
outside”. Subtraction, on the other hand, requires that a procedure not be applied, which seems harder still.
The solution to this problem is to make the following observation. Consider the pair h0, 0i. Now apply
the following algorithm. Given such a pair of numbers, create a new pair. The new pair’s left component is
the old pair’s right component; the new pair’s right component is the old pair’s right component incremented
by one. Visually,
initial value : h0, 0i

after 1 iteration : h0, 1i
after 2 iterations : h1, 2i
and so on.
You might find this procedure rather strange, in that it entirely ignores the left component of each
preceding pair to create the next one in the sequence. Notice, however, that after n iterations, the left
component holds the value n − 1. Furthermore, observe the operations that we used to obtain these pairs:
creation of an initial pair, pair deconstruction, increment by one, and new pair construction. That is, the
following procedure represents the algorithm applied at each step:
(lambda (p)
(pair (right p)
(succ (right p))))
The following represents the initial value:
(pair zero zero)
If we apply this n times, then read out the left value of the resulting pair, we get. . . n − 1!
pred ≡
(lambda (n)
22.4. ELIMINATING RECURSION 223
(left
(lambda (p)
(pair (right p)
((n (succ (right p)))) )
(pair zero zero) )))
Once we have subtraction by one, we can implement regular subtraction and other such operations.
That leaves only one arithmetic-related primitive we need to implement factorial, namely Scheme’s
zero?. What does this operator do? Given a representation for zero, it returns true, otherwise false. What
is the one characteristic that distinguishes the numeral for zero from that for all non-zero numbers? The
latter all apply their first argument to their second at least once, while the former does not. Therefore, the
following defines zero?:
iszero ≡
(lambda (n)
((n (lambda (ignore) no)) yes))
If the first argument is applied at all, no matter how many times it’s applied, it returns the representation of
false; if it never is, then the “zero” value, the representation of true, results.
Historical aside: These numerals are known as the Church numerals, in honor of their inventor, Alonzo
Church.
Exercise 22.3.1 Can you extend this encoding to other kinds of numbers: negative integers, rationals, reals,
complex numbers, . . . ?
Hint: Some of these are easy to encode using pairs of other kinds of numbers.
Exercise 22.3.2 Here are two alternate representations of the exponentiation operation. Which one is
faster?
(define exp1
(lambda (m)
(lambda (n)
((n (prod m)) one))))
(define exp2
(lambda (m)
(lambda (n)
(n m))))
22.4 Eliminating Recursion

The define construct of Scheme is surprisingly powerful. It not only assigns values to names, it also en-
ables the construction of recursive procedures (otherwise the definition of factorial given above would not
function). To eliminate define, therefore, we must create a way of defining recursive procedures. . . without
recursion!
As we do this, we will sometimes encounter expressions that we’re not yet sure how to write. We will
use the symbol • to represent a special value: if the computation ever tries to apply it, it halts immediately.
Think of it as a landmine that the computation should never apply.2
Let’s now study the recursion in factorial. Let’s begin with the following skeletal definition:
fact ≡
(lambda (n)
(if (zero? n)
1
(∗ n (• (sub1 n)))))
This definition is not entirely useless. Indeed, it correctly computes the factorial on the input 0. On any
input greater than 0, however, the computation terminates uselessly.
We can make a more useful definition by including a copy of fact as follows:
fact ≡
(lambda (n)
(if (zero? n)
1
(lambda (n)
(if (zero? n)
1
(∗ n ( (∗ n (• (sub1 n)))))
(sub1 n)))))
This definition works perfectly well on inputs of 0 and 1, but not greater. We can repeat this process
endlessly—a process, not at all coincidentally, reminiscent of creating the proper environment for recursion
in Section 9—but obviously, we will never get the true definition of factorial. We’ll have to do better.
While we’re trying to generate a spark of insight, let’s try to clean up the code above. Instead of relying
on this unspecified • operation, we might as well just parameterize over that location in the program:
mk-fact ≡
(lambda (f )
(lambda (n)
(if (zero? n)
1
(∗ n (f (sub1 n))))))
The resulting procedure isn’t quite factorial itself, but rather a factorial-maker: given the right value for
the parameter, it will yield the proper factorial procedure. That still begs the question, however, of what to
supply as a parameter.
Let’s go back to our doomed attempt to nest copies of the factorial procedure. This has the advantage
that, until the copies run out, there is always another copy available. So we have a clearer handle on the
problem now: we need to provide as a parameter something that will create another copy upon demand.
2A good approximation of • is the Scheme procedure exit.
It would seem that mk-fact is just such a creator of copies. So what happens if we feed mk-fact as the
argument to mk-fact—
(mk-fact mk-fact)
—to fill the hole where we need a factorial generator? This application results in the procedure
(lambda (f )
(lambda (n)
(if (zero? n)
1
(∗ n (mk-fact (sub1 n))))))
(We’ve just substituted mk-fact for f in the body of mk-fact.) We can safely apply this procedure to 0 and
obtain 1, but if we apply it to any larger input, we get an error: mk-fact is expecting a procedure as its
argument, but here we’re applying it to a number.
Okay, so we cannot apply mk-fact to a number. To gain some insight into what we can apply it to, let’s
apply it to •, so the new definition of mk-fact is:
(lambda (f )
(lambda (n)
(if (zero? n)
1
(∗ n ((f •) (sub1 n))))))
and apply this to mk-fact. Upon substitution, this evaluates to
(lambda (n)
(if (zero? n)
1
(∗ n ((mk-fact •) (sub1 n))))))
This procedure clearly works correctly on the argument value 0. For argument 1, evaluation results in:
(∗ 1
(if (zero? 0)
1
(∗ 0 ((• •) (sub1 n)))))
which also works correctly. For initial arguments of 2 and greater, however, evaluation halts on trying to
apply •. In short, we have a definition of mk-fact that, when applied to itself, works correctly for values of
0 and 1, but no higher.
By itself, this does not seem like any kind of progress at all. In fact, however, we have come most of the
way to a solution. Recall that earlier, to define a factorial procedure for values of 0 and 1, we had to copy the
definition of the procedure, resulting in a definition that was roughly twice as large as the one we wanted.
Now we have one that is roughly the same size as the original, i.e., it involves no copying.
It is not surprising that the computation eventually terminated, when we supplied • as the argument.
But what else could we have supplied? Observe that when we supplied mk-fact as an argument, the term
in the “recursive” position evaluated to (mk-fact •), which continued computation correctly (for one more
step); whereas supplying • made that term (• •), which terminated it. Therefore, we should ensure that
the function position of that application is always mk-fact. This means precisely that instead of supplying
• as the argument, we should supply mk-fact! This should not be surprising: just as applying mk-fact to
itself outside the body of mk-fact was useful in initiating the computation, doing the same inside is useful in
continuing it.
That is, we must define
mk-fact ≡
(lambda (f )
(lambda (n)
(if (zero? n)
1
(∗ n ((f f ) (sub1 n))))))
such that factorial is ostensibly defined by
(mk-fact mk-fact)
Does that work? Substituting as before, we get
(lambda (n)
(if (zero? n)
1
(∗ n ( (mk-fact mk-fact) (sub1 n)))))
This of course means we can substitute the inner application also:
(lambda (n)
(if (zero? n)
1
(lambda (n)
(if (zero? n)
1
(∗ n ( (∗ n ( (mk-fact mk-fact) (sub1 n))))) (sub1 n)))))
and now we can see the recursion unwind: as we need another copy of factorial, the application of mk-fact to
itself generates a fresh copy. Thus, we have a satifactory solution to the problem of defining the “recursive”
factorial function without any use of recursion!
To summarize, we have the following definition,
mk-fact ≡
(lambda (f )
(lambda (n)
(if (zero? n)
1
(∗ n ((f f ) (sub1 n))))))
with factorial defined by

fact ≡
(mk-fact mk-fact)
That is, factorial is
((lambda (mk-fact)
(mk-fact mk-fact))
(lambda (f )
(lambda (n)
(if (zero? n)
1
(∗ n ((f f ) (sub1 n)))))))
(Test this! In a fresh Scheme session, apply this expression directly to numeric values and make sure you
get the factorial of the input as a result. Pretty amazing, huh?)
While this is a correct implementation of factorial, we seem to be writing a lot of code relative to the
recursive version defined using define. Furthermore, we would like to know how much of this solution can
apply to other functions also. With that in mind, let’s try to refactor this code a little. What would we like to
write? As programmers, we would rather not have to keep track of the self-application in the body; that is,
we would rather write
(make-recursive-procedure
(lambda (fact)
(lambda (n)
(if (zero? n)
1
(∗ n (fact (sub1 n)))))))
which looks almost exactly like the definition created using define. So, how do we get the self-application
out?
Observe that the definition of factorial above is equivalent to this one:
((lambda (mk-fact)
(mk-fact mk-fact))
(lambda (f )
(lambda (g)
(lambda (n)
(if (zero? n)
1
( (∗ n (g (sub1 n))))))
(f f ))))
All we have done is introduce a new level of abstraction, binding g to the self-application of f . Note,
however, that the boxed expression is precisely the definition of factorial that we wanted to write, except
that fact is called g. One more level of abstraction separates the factorial-specific code from the recursive
function generator:
make-recursive-procedure ≡
(lambda (p)
((lambda (mk-fact)
(mk-fact mk-fact))
(lambda (f )
(p
(f f )))))
In fact, because this code has nothing to do with defining factorial at all, we can rename mk-fact:
make-recursive-procedure ≡
(lambda (p)
((lambda (f )
(f f ))
(lambda (f )
(p (f f )))))
This is now a generic procedure that creates recursive procedures out of its argument! It is remarkable that
such a procedure even exists; its structure is daunting at first sight, but relatively easy to understand once
you grasp the need for “copies” of a procedure, and that self-application generates as many of these copies
as necessary.
In the literature, the procedure make-recursive-procedure is known as the Y combinator. It is sometimes
also known as a “fixed-point combinator”, because it computes the fixed-point of its argument procedure.
The Lambda Calculus

With the definition of the Y combinator, we have reduced all of Scheme to just three primitives:
procedure definition, procedure application, and variables. With just those three, we have provided
an encoding of all of the rest of the language. This compact little language is the core of what is
known, also for historical reasons, as the lambda calculus. The “calculus” part of the language is
beyond the scope of our study.
In the 1930s, several mathematicians were asking fundamental questions about what could be com-
puted procedurally, and about the relative power of different formalisms. While Alan Turing was
defining his Turing machine formalism, Alonzo Church and several others created an alternate for-
malism: the lambda calculus. These mathematicians were able to demonstrate that several of their
formalisms—particularly these two—were equivalent in expressive power, so theoreticians could
choose one or the other based on convenience and suitability, without worrying about expressive
constraints. (To this day, many choose to use the lambda calculus and its variants since it offers so
much more expressive power than a Turing machine.) Indeed, the fact that so many independently-
derived formalisms had the same expressive power led to the formulation of the Church-Turing
thesis: that no formal language is more powerful than those defined by Church and Turing (the
lambda calculus and the Turing machine, respectively).
Exercise 22.4.1 Type the definition of make-recursive-procedure into Scheme and use it to create a recursive
factorial procedure:
(make-recursive-procedure
(lambda (fact)
(lambda (n)
(if (zero? n)
1
(∗ n (fact (sub1 n)))))))
What do you observe? Explain and correct.

Chapter 23
Semantics
We have been writing interpreters in Scheme in order to understand various features of programming lan-
guages. What if we want to explain our interpreter to someone else? If that person doesn’t know Scheme,
we can’t communicate how our interpreter works. It would be convenient to have some common language
for explaining interpreters. We already have one: math!
Let’s try some simple examples. If our program is a number n, it just evaluates to some mathematical
representation of n. We’ll use a nb to represent this number, whereas n itself will hold the numeral. For
instance, the numeral 5 is represented by the number b5 (note the subtle differences in typesetting!). In other
words, we will write
n ⇒ nb
where we read ⇒ as “reduces to”. Numbers are already values, so they don’t need further reduction.
How about addition? We might be tempted to write
{+ l r} ⇒ ld
+r
In particular, the addition to the left of the ⇒ is in the programming language, while the one on the right
happens in mathematics and results in a number. That is, the addition symbol on the left is syntactic. It could
map to any mathematical operation. A particularly perverse language might map it to multiplication, but
more realistically, it is likely to map to addition modulo some base to reflect fixed-precision arithmetic. It
is the expression on the right that gives it meaning, and in this case it assigns the meaning we would expect
(corresponding, say, to DrScheme’s use of unlimited-precision numbers for integers and rationals).
That said, this definition is unsatisfactory. Mathematical addition only works on numbers, but l and r
might each be complex expressions in need of reduction to a value (in particular, a number) so they can be
added together. We denote this as follows:
l ⇒ lbv r ⇒ rbv
{+ l r} ⇒ lvd
+ rv
The terms above the bar are called the antecedents, and those below are the consequents. This rule is just a
convenient way of writing an “if . . . then” expression: it says that if the conditions in the antecedent hold,
then those in the consequent hold. If there are multiple conditions in the antecedent, they must all hold for
231
232 CHAPTER 23. SEMANTICS
the rule to hold. So we read the rule above as: if l reduces to lv , and if r reduces to rv , then adding the
respective expressions results in the sum of their values. (In particular, it makes sense to add lv and rv , since
each is now a number.) A rule of this form is called a judgment, because based on the truth of the conditions
in the antecedent, it issues a judgment in the consequent (in this case, that the sum will be a particular value).
These rules subtly also bind names to values. That is, a different way of reading the rule is not as an “if
. . . then” but rather as an imperative: it says “reduce l, call the result lv ; reduce r, call its result rv ; if these
two succeed, then add lv and rv , and declare the sum the result for the entire expression”. Seen this way, l
and r are bound in the consequent to the sub-expressions of the addition term, while lv and rv are bound in
the antecedent to the results of evaluation (or reduction). This representation truly is an abstract description
of the interpreter.
Let’s turn our attention to functions. We want them to evaluate to closures, which consist of a name, a
body and an environment. How do we represent a structure in mathematics? A structure is simply a tuple,
in this case a triple. (If we had multiple kinds of tuples, we might use tags to distinguish between them, but
for now that won’t be necessary.) We would like to write
{fun {i} b} ⇒ hi, b, ???i
but the problem is we don’t have a value for the environment to store in the closure. So we’ll have to make
the environment explicit. From now on, ⇒ will always have a term and an environment on the left, and a
value on the right. We first rewrite our two existing reduction rules:
n, E ⇒ nb
l, E ⇒ lbv r, E ⇒ rbv
{+ l r}, E ⇒ lvd+ rv
Now we can define a reduction rule for functions:
{fun {i} b}, E ⇒ hi, b, E i
Given an environment, we can also look up the value of identifiers:
i, E ⇒ E (i)
All that remains is application. As with addition, application must first evaluate its subexpressions, so
the general form of an application must be as follows:
f , E ⇒??? a, E ⇒???
{ f a}, E ⇒???
What kind of value must f reduce to? A closure, naturally:
f , E ⇒ hi, b, E 0 i a, E ⇒???
{ f a}, E ⇒???
233
(We’ll use E 0 to represent to closure environment to make clear that it may be different from E .) We don’t
particularly care what kind of value a reduces to; we’re just going to substitute it:
f , E ⇒ hi, b, E 0 i a, E ⇒ av
{ f a}, E ⇒???
But what do we write below? We have to evaluate the body, b, in the extended environment; whatever value
it returns is the value of the application. So the evaluation of b also moves into the antecedent:
f , E ⇒ hi, b, E 0 i a, E ⇒ av b, ??? ⇒ bv
{ f a}, E ⇒ bv
In what environment do we reduce the body? It has to be the environment in the closure; if we use E instead
of E 0 , we introduce dynamic rather than static scoping! But additionally, we must extend E 0 with a binding
for the identifier named by i; in particular, it must be bound to the value of the argument. We can write all
this concisely as
f , E ⇒ hi, b, E 0 i a, E ⇒ av b, E 0 [i←av ] ⇒ bv
{ f a}, E ⇒ bv
where E 0 [i←av ] means “the environment E 0 extended with the identifier i bound to the value av ”. If E 0
already has a binding for i, this extension shadows that binding.
The judicious use of names conveys information here. We’re demanding that the value used to extend
the environment must be the same as that resulting from evaluating a: the use of av in both places indicates
that. It also places an ordering on operations: clearly the environment can’t be extended until av is available,
so the argument must evaluate before application can proceed with the function’s body. The choice of two
different names for environments—E and E 0 —denotes that the two environments need not be the same.
We call this a big-step operational semantics. It’s a semantics because it ascribes meanings to programs.
(We can see how a small change can result in dynamic instead of static scope and, more mundanely, that the
meaning of + is given to be addition, not some other binary operation.) It’s operational because evaluation
largely proceeds in a mechanical fashion; we aren’t compiling the entire program into a mathematical object
and using fancy math to reduce it to an answer. Finally, it’s big-step because ⇒ reduces expressions down
to irreducible answers. In contrast, a small-step semantics performs one atomic reduction at a time, rather
like a stepper in a programming environment.
Exercise 23.0.2 Extend the semantics to capture conditionals.
Exercise 23.0.3 Extend the semantics to capture lists.

Hint: You may want to consider tagging tuples.
Exercise 23.0.4 Extend the semantics to capture recursion.
Exercise 23.0.5 Alter the semantics to reflect lazy evaluation instead.

234 CHAPTER 23. SEMANTICS
Part X
Types
235
Chapter 24
Introduction
Until now, we’ve largely ignored the problem of program errors. We haven’t done so entirely: if a program-
mer writes
{fun {x}}
we do reject this program, because it isn’t syntactically legal—every function must have a body. But what
if, instead, he were to write
{+ 3
{fun {x} x}}
? Right now, our interpreter might produce an error such as
num-n: not a number
A check deep in the bowels of our interpreter is flagging the use of a non-numeric value in a position
expecting a number.
At this point, we can make the same distinction between the syntactic and meta levels about errors as
we did about representations. The error above is an error at the syntactic level,1 because the interpreter
is checking for the correct use of its internal representation. Suppose we had division in the interpreted
language, and the corresponding num/ procedure failed to check that the denominator was non-zero; then
the interpreter’s behavior would be that of Scheme’s on division-by-zero. If we had expected an error
and Scheme did not flag one (or vice versa), then the interpreter would be unfaithful to the intent of the
interpreted language.
Of course, this discussion about the source of error messages somewhat misses the point: we really
ought to reject this program without ever executing it. But rejecting it is difficult because this program is
legitimate from the perspective of the parser. It’s only illegal from the semantic viewpoint, it is the meaning,
as opposed to the syntax, of + that does not accept functions as arguments. Therefore, we clearly need a
more sophisticated layer that checks for the validity of programs.
1 Not to be confused with a syntax error!
237
238 CHAPTER 24. INTRODUCTION
How hard is this? Rejecting the example above seems pretty trivial: indeed, it’s so easy, we could almost
build this into the parser (to not accept programs that have syntactic functions as arguments to arithmetic
primitives). Let’s think more broadly. Sometimes it does not seem much harder: for instance,
{with {f {fun {x} {+ x 1}}}

{+ 3
{f 5}}}
is clearly legal, whereas
{with {f {fun {x}

{fun {y} {+ x y}}}}
{+ 3
{f 5}}}
is not. Here, simply substituting f in the body seems to be enough. The problem does not quite reduce
to the parsing problem that we had earlier—a function application is necessary to determine the program’s
validity. But consider this program:
{fun {f}
{+ 3
{f 5}}}
Is this program valid? Clearly, it depends on whether or not f, when applied to 5, evaluates to a number.
Since this expression may be used in many different contexts, we cannot know whether or not this is legal
without examining each application, which in turn may depend on other substitutions, and so on. In short, it
appears that we will need to run the program just to determine whether f is always bound to a function, and
one that can accept numbers—but running the program is precisely what we’re trying to avoid!
We now commence the study of types and type systems, which are designed to identify the abuse of types
before executing a program. First, we need to build an intuition for the problems that types can address, and
the obstacles that they face. Consider the following program:
{+ 3
{if0 mystery
5
{fun {x} x}}}
This program executes successfully (and evaluates to 8) if mystery is bound to 0, otherwise it results in
an error. The value of mystery might arise from any number of sources. For instance, it may be bound to
0 only if some mathematical statement, such as the Collatz conjecture, is true.2 In fact, we don’t even need
to explore something quite so exotic: our program may simply be
2 Consider the function f (n) defined as follows: If n is even, divide n by 2; if odd, compute 3n + 1. The Collatz conjecture posits
that, for every positive integer n, there exists some k such that f k (n) = 1. (The sequences demonstrating convergence to 1 are often
quite long, even for small numbers! For instance: 7 → 22 → 11 → 34 → 17 → 52 → 26 → 13 → 40 → 20 → 10 → 5 → 16 → 8
→ 4 → 2 → 1.)
24.1. WHAT ARE TYPES? 239
{+ 3
{if0 {read-number}
5
{fun {x} x}}}
Unless we can read the user’s mind, we have no way of knowing whether this program will execute without
error. In general, even without involving the mystery of mathematical conjectures or the vicissitudes of
users, we cannot statically determine whether a program will halt with an error, because of the Halting
Problem.
This highlights an important moral:
Type systems are always prey to the Halting Problem. Consequently, a type system for a
general-purpose language must always either over- or under-approximate: either it must re-
ject programs that might have run without an error, or it must accept programs that will error
when executed.
While this is a problem in theory, what impact does this have on practice? Quite a bit, it turns out. In
languages like Java, programmers think they have the benefit of a type system, but in fact many common
programming patterns force programmers to employ casts instead. Casts intentionally subvert the type
system and leave checking for execution time. This indicates that Java’s evolution is far from complete. In
contrast, most of the type problems of Java are not manifest in a language like ML, but its type systems still
holds a few (subtler) lurking problems. In short, there is still much to do before we can consider type system
design a solved problem.
24.1 What Are Types?

A type is any property of a program that we can establish without executing the program. In particular,
types capture the intuition above that we would like to predict a program’s behavior without executing it.
Of course, given a general-purpose programming language, we cannot predict its behavior entirely without
execution (think of the user input example, for instance). So any static prediction of behavior must neces-
sarily be an approximation of what happens. People conventionally use the term type to refer not just to any
approximation, but one that is an abstraction of the set of values.
A type labels every expression in the language, recording what kind of value evaluating that expression
will yield. That is, types describe invariants that hold for all executions of a program. They approximate this
information in that they typically record only what kind of value the expression yields, not the precise value
itself. For instance, types for the language we have seen so far might include number and function. The
operator + consumes only values of type number, thereby rejecting a program of the form
{+ 3
{fun {x} x}}
To reject this program, we did not need to know precisely which function was the second argument to
+, be it {fun {x} x} or {fun {x} {fun {y} {+ x y}}}. Since we can easily infer that 3 has
type number and {fun {x} x} has type function, we have all the information we need to reject the
program without executing it.
Note that we are careful to refer to valid programs, but never correct ones. Types do not ensure the
correctness of a program. They only guarantee that the program does not make certain kinds of errors.
Many errors lie beyond the ambit of a type system, however, and are therefore not caught by it. Many type
systems will not, for instance, distinguish between a program that sorts values in ascending order from one
that sorts them in descending order, yet the difference between those two is usually critical for a program’s
overall correctness.
24.2 Type System Design Forces

Designing a type system involves finding a careful balance between two competing forces:
1. Having more information makes it possible to draw richer conclusions about a program’s behavior,
thereby rejecting fewer valid programs or permitting fewer buggy ones.
2. Acquiring more information is difficult:
• It may place unacceptable restrictions on the programming language.

• It may incur greater computational expense.
• It may force the user to annotate parts of a program. Many programmers (sometimes unfairly)
balk at writing anything beyond executable code, and may thus view the annotations as onerous.
• It may ultimately hit the limits of computability, an unsurpassable barrier. (Often, designers can
surpass this barrier by changing the problem slightly, though this usually moves the task into
one of the three categories above.)
24.3 Why Types?

Types form a very valuable first line of defense against program errors. Of course, a poorly-designed type
system can be quite frustrating: Java programming sometimes has this flavor. A powerful type system such
as that of ML, however, is a pleasure to use. ML programmers, for instance, claim that programs that type
correctly often work correctly within very few development iterations.
Types that have not been subverted (by, for instance, casts in Java) perform several valuable roles:
• When type systems detect legitimate program errors, they help reduce the time spent debugging.
• Type systems catch errors in code that is not executed by the programmer. This matters because if a
programmer constructs a weak test suite, many parts of the system may receive no testing. The system
may thus fail after deployment rather than during the testing stage. (Dually, however, passing a type
checker makes many programmers construct poorer test suites—a most undesirable and unfortunate
consequence!)
24.3. WHY TYPES? 241
• Types help document the program. As we discussed above, a type is an abstraction of the values that
an expression will hold. Explicit type declarations therefore provide an approximate description of
code’s behavior.
• Compilers can exploit types to make programs execute faster, consume less space, spend less time in
garbage collection, and so on.
• While no language can eliminate arbitrarily ugly code, a type system imposes a baseline of order that
prevents at least a few truly impenetrable programs—or, at least, prohibits certain kinds of terrible
coding styles.
Chapter 25
Type Judgments
25.1 What They Are

First, we must agree on a language of types. Recall that types need to abstract over sets of values; earlier, we
suggested two possible types, number and function. Since those are the only kinds of values we have
for now, let’s use those as our types.
We present a type system as a collection of rules, known formally as type judgments, which describe
how to determine the type of an expression.1 There must be at least one type rule for every kind of syntactic
construct so that, given a program, at least one type rule applies to every sub-term. Judgments are often
recursive, determining an expression’s type from the types of its parts.
The type of any numeral is number:
n : number
(read this as saying “any numeral n has type number”) and of any function is function:
{fun {i} b} : function
but what is the type of an identifier? Clearly, we need a type environment (a mapping from identifiers to
types). It’s conventional to use Γ (the upper-case Greek “gamma”) for the type environment. As with the
value environment, the type environment must appear on the left of every type judgment. All type judgments
will have the following form:
Γè : t
where e is an expression and t is a type, which we read as “Γ proves that e has type t”. Thus,
Γ`n : number
{fun {i} b} : function

Γì : Γ(i)
1A type system for us is a collection of types, the corresponding judgments that ascribe types to expressions, and an algorithm
for performing this ascription. For many languages a simple algorithm suffices, but as languages get more sophisticated, devising
this algorithm can become quite difficult, as we will see in Section 30.
243
244 CHAPTER 25. TYPE JUDGMENTS
The last rule simply says that the type of identifier i is whatever type it is bound to in the environment.
This leaves only addition and application. Addition is quite easy:
Γ`l : number Γ`r : number
Γ`{+ l r} : number
All this leaves is the rule for application. We know it must have roughly the following form:
Γ` f : function Γà : τa ···
Γ`{ f a} :???
where τa is the type of the expression a (we will often use τ to name an unknown type).
What’s missing? Compare this against the semantic rule for applications. There, the representation of a
function held an environment to ensure we implemented static scoping. Do we need to do something similar
here?
For now, we’ll take a much simpler route. We’ll demand that the programmer annotate each function
with the type it consumes and the type it returns. This will become part of a modified function syntax. That
is, the language becomes
<TFWAE> ::= ...
| {fun {<id> : <type>} : <type> <TFWAE>}
(the “T”, naturally, stands for “typed”) where the two type annotations are now required: the one immedi-
ately after the argument dictates what type of value the function consumes, while that after the argument but
before the body dictates what type it returns. An example of a function definition in this language is
{fun {x : number} : number
{+ x x}}
We must also change our type grammar; to represent function types we conventionally use an arrow, where
the type at the tail of the arrow represents the type of the argument and that at the arrow’s head represents
the type of the function’s return value:
<type> ::= number
| (<type> -> <type>)
(notice that we have dropped the overly naive type function from our type language). Thus, the type of
the function above would be (number -> number). The type of the outer function below
{fun {x : number} : (number -> number)
{fun {y : number} : number
{+ x y}}}
is (number -> (number -> number)), while the inner function has type (number -> number).
Equipped with these types, the problem of checking applications becomes easy:
Γ` f : (τ1 →τ2 ) Γà : τ1
Γ`{ f a} : τ2
25.1. WHAT THEY ARE 245
That is, if you provide an argument of the type the function is expecting, it will provide a value of the type
it promises. Notice how the judicious use of the same type name τ1 and τ2 accurately captures the sharing
we desire.
There is one final bit to the introductory type puzzle: how can we be sure the programmer will not lie?
That is, a programmer might annotate a function with a type that is completely wrong (or even malicious).
(A different way to look at this is, having rid ourselves of the type function, we must revisit the typing
rule for a function declaration.) Fortunately, we can guard against cheating and mistakes quite easily: instead
of blindly accepting the programmer’s type annotation, we check it:
Γ[i←τ1 ]`b : τ2
Γ`fun {i : τ1 } : τ2 b} : (τ1 →τ2 )
This rule says that we will believe the programmer’s annotation if the body has type τ2 when we extend the
environment with i bound to τ1 .
There is an important relationship between the type judgments for function declaration and for applica-
tion:
• When typing the function declaration, we assume the argument will have the right type and guarantee
that the body, or result, will have the promised type.
• When typing a function application, we guarantee the argument has the type the function demands,
and assume the result will have the type the function promises.
This interplay between assumptions and guarantees is quite crucial to typing functions. The two “sides”
are carefully balanced against each other to avoid fallacious reasoning about program behavior. In addition,
just as number does not specify which number will be used, a function type does not limit which of many
functions will be used. If, for instance, the type of a function is (number -> number), the function
could be either increment or decrement (or a lot else, besides). The type checker is able to reject misuse of
any function that has this type without needing to know which actual function the programmer will use.
By the way, it would help to understand the status of terms like i and b and n in these judgments. They are
“variable” in the sense that they will be replaced by some program term: for instance, {fun {i : τ1 } : τ2 b}
may be instantiated to {fun {x : number} : number x}, with i replaced by x, and so forth. But
they are not program variables; rather, they are variables that stand for program text (including program
variables). They are therefore called metavariables.
Exercise 25.1.1 It’s possible to elide the return type annotation on a function declaration, leaving only the
argument type annotation. Do you see how?
Exercise 25.1.2 Because functions can be nested within each other, a function body may not be closed at
the time of type-checking it. But we don’t seem to capture the definition environment for types the way we
did for procedures. So how does such a function definition type check? For instance, how does the second
example of a typed procedure above pass this type system?
25.2 How Type Judgments Work

Let’s see how the set of type judgments described above accepts and rejects programs.
1. Let’s take a simple program,
{+ 2
{+ 5 7}}
We stack type judgments for this term as follows:
0`5
/ : number 0`7
/ : number
0`2
/ : number
0`{+
/ 5 7} : number
0`{+
/ 2 {+ 5 7}} : number
This is a type judgment tree.2 Each node in the tree uses one of the type judgments to determine the
type of an expression. At the leaves (the “tops”) are, obviously, the judgments that do not have an
antecedent (technically known as the axioms); in this program, we only use the axiom that judges
numbers. The other two nodes in the tree both use the judgment on addition. The metavariables in the
judgments (such as l and r for addition) are replaced here by actual expressions (such as 2, 5, 7 and
{+ 5 7}): we can employ a judgment only when the pattern matches consistently. Just as we begin
evaluation in the empty environment, we begin type checking in the empty type environment; hence
we have 0/ in place of the generic Γ.
Observe that at the end, the result is the type number, not the value 14.
2. Now let’s examine a program that contains a function:
{{fun {x : number} : number

{+ x 3}}
5}
The type judgment tree looks as follows:
[x←number]`x : number [x←number]`3 : number

[x←number]`{+ x 3} : number
0`5
/ : number
0`{fun
/ {x : number} : number {+ x 3}} : (number→number)
0`{{fun
/ {x : number} : number {+ x 3}} 5} : number
When matching the sub-tree at the top-left, where we have just Γ in the type judgment, we have the
extended environment in the actual derivation tree. We must use the same (extended) environment
consistently, otherwise the type judgment for addition cannot be applied. The set of judgments used
2 If it doesn’t look like a tree to you, it’s because you’ve been in computer science too long and have forgotten that real trees
grow upward, not downward. Botanically, however, most of these “trees” are really shrubs.
25.2. HOW TYPE JUDGMENTS WORK 247
to assign this type is quite different from the set of judgments we would use to evaluate the program:
in particular, we type “under the fun”, i.e., we go into the body of the fun even if the function
is never applied. In contrast, we would never evaluate the body of a function unless and until the
function was applied to an actual parameter.
3. Finally, let’s see what the type judgments do with a program that we know to contain a type error:
{+ 3
{fun {x : number} : number
x}}
The type judgment tree begins as follows:
???
0`{+
/ 3 {fun {x : number} : number x}}} : ???
We don’t yet know what type (if any) we will be able to ascribe to the program, but let’s forge
on: hopefully it’ll become clear soon. Since the expression is an addition, we should discharge the
obligation that each sub-expression must have numeric type. First for the left child:
0`3
/ : number ???
0`{+
Now for the right sub-expression. First let’s write out the sub-expression, then determine its type:
0`3
/ : number 0`{fun
/ {x : number} : number x}} :???
0`{+
As per the judgments we have defined, any function expression must have an arrow type:
0`3
/ : number 0`{fun
/ {x : number} : number x}} : (???→???)
0`{+
This does the type checker no good, however, because arrow types are distinct from numeric types,
so the resulting tree above does not match the form of the addition judgment (no matter what goes in
place of the two ???’s). To match the addition judgment the tree must have the form
0`3
/ : number 0`{fun
/ {x : number} : number x}} : number
0`{+
Unfortunately, we do not have any judgments that let us conclude that a syntactic function term can
have a numeric type. So this doesn’t work either.
In short, we cannot construct a legal type derivation tree for the original term. Notice that this is not
the same as saying that the tree directly identifies an error: it does not. A type error occurs when we
are unable to construct a type judgment tree.
This is subtle enough to bear repeating: To flag a program as erroneous, we must prove that no type
derivation tree can possibly exist for that term. But perhaps some sequence of judgments that we haven’t
thought of exists that (a) is legal and (b) correctly ascribes a type to the term! To avoid this we may need
to employ quite a sophisticated proof technique, even human knowledge. (In the third example above, for
instance, we say, “we do not have any judgments that let us conclude that a syntactic function term can have
a numeric type”. But how do we know this is true? We can only conclude this by carefully studying the
structure of the judgments. A computer program might not be so lucky, and in fact may get stuck endlessly
trying judgments!)
This is why a set of type judgments alone does not suffice: what we’re really interested in is a type system
that includes an algorithm for type-checking. For the set of judgments we’ve written here, and indeed for the
ones we’ll study initially, a simple top-down, syntax-directed algorithm suffices for (a) determining the type
of each expression, and (b) concluding that some expressions manifest type errors. As our type judgments
get more sophisticated, we will need to develop more complex algorithms to continue producing tractable
and useful type systems.
Chapter 26
Typing Control
26.1 Conditionals
Let’s expand our language with a conditional construct. We can use if0 like before, but for generality it’s
going to be more convenient to have a proper conditional and a language of predicates. The type judgment
for the conditional must have the following form:
Γ`c :??? Γ`t :??? Γè :???

Γ`{if c t e} :???
where c is the conditional, t the “then”-expression, and e the “else”-expression.
Let’s begin with the type for c. What should it be? In a language like Scheme we permit any value, but
in a stricter, typed language, we might demand that the expression always evaluate to a boolean. (After all,
if the point is to detect errors sooner, then it does us no good to be overly lax in our type rules.) However,
we don’t yet have such a type in our type language, so we must first extend that language:
<type> ::= number
| boolean
| (<type> -> <type>)
Armed with the new type, we can now ascribe a type to the conditional expression:
Γ`c : boolean Γ`t :??? Γè :???

Γ`{if c t e} :???
Now what of the other two, and of the result of the expression? One option is, naturally, to allow both arms
of the conditional to have whatever types the programmer wants:
Γ`c : boolean Γ`t : τ1 Γè : τ2

Γ`{if c t e} :???
By using two distinct type variables, we do not demand any conformity between the actual types of the
arms. By permitting this flexibility, however, we encounter two problems. The first is that it isn’t clear what
249
250 CHAPTER 26. TYPING CONTROL
type to ascribe to the expression overall.1 Second, it reduces our ability to trap program errors. Consider a
program like this:
{+ 3
{if {is-zero mystery}
5
{fun {x} x}}}
Because we know nothing about mystery, we must conservatively conclude that it might be non-zero,
which means eventually we are going to see a type error that we only catch at run-time. But why permit the
programmer to write such a program at all? We might as well prevent it from ever executing. Therefore, we
use the following rule to type conditionals:
Γ`c : boolean Γ`t : τ Γè : τ

Γ`{if c t e} : τ
Notice that by forcing the two arms to have the same type, we can assign that common type to the entire
expression, so the type system does not need to know which branch was chosen on a given execution: the
type remains the same.
Having added conditionals and the type boolean isn’t very useful yet, because we haven’t yet intro-
duced predicates to use in the test position of the conditional. Indeed, we can easily see that this is true
becuase we have not yet written a function type with boolean on the right-hand side of the arrow. You
can, however, easily imagine adding procedures such as is-zero, with type number -> boolean.
26.2 Recursion
Now that we have conditionals, if we can also implement recursion, we would have a Turing-complete
language (that could, for instance, with a little more arithmetic support, enable writing factorial). So the
next major piece of the puzzle is typing recursion.
Given the language TFAE (typed FAE), can we write a recursive program? Let’s try to write an infinite
loop. Our first attempt might be this FAE program
{with {f {fun {i}

{f i}}}
{f 10}}
which, expanded out, becomes
{{fun {f}
{f 10}}
{fun {i}
{f i}}}
1 It’s tempting to create a new kind of type, a union type, so that the type of the expression is τ ∪ τ . This has far-reaching
1 2
consequences, however, including a significant reduction in type-based guarantee of program reliability.
26.2. RECURSION 251
When we place type annotations on this program, we get
{{fun {f : (num -> num)} : num

{f 10}}
{fun {i : num} : num
{f i}}}
These last two steps don’t matter, of course. This program doesn’t result in an infinite loop, because the f
in the body of the function isn’t bound, so after the first iteration, the program halts with an error.
As an aside, this error is easier to see in the typed program: when the type checker tries to check the
type of the annotated program, it finds no type for f on the last line. Therefore, it would halt with a type
error, preventing this erroneous program from ever executing.2
Okay, that didn’t work, but we knew about that problem: we saw it in Section 9 when introducing
recursion. At the time, we asked you to consider whether it was possible to write a recursive function
without an explicit recursion construct, and Section 22 shows that it is indeed possible. The essence of the
solution presented there is to use self-application:
{with {omega {fun {x}

{x x}}}
{omega omega}}
How does this work? Simply substituting omega with the function, we get
{{fun {x} {x x}}

{fun {x} {x x}}}
Substituting again, we get
{{fun {x} {x x}}

{fun {x} {x x}}}
and so on. In other words, this program executes forever. It is conventional to call the function ω (lower-case
Greek “omega”), and the entire expression Ω (upper-case Greek “omega”).3
Okay, so Ω seems to be our ticket. This is clearly an infinite loop in FAE. All we need to do is convert
it to TFAE, which is simply a matter of annotating all procedures. Since there’s only one, ω, this should be
especially easy.
To annotate ω, we must provide a type for the argument and one for the result. Let’s call the argument
type, namely the type of x, τa and that of the result τr , so that ω : τa →τr . The body of ω is {x x}. From
this, we can conclude that τa must be a function (arrow) type, since we use x in the function position of an
application. That is, τa has the form τ1 →τ2 , for some τ1 and τ2 yet to be determined.
2 In this particular case, of course, a simpler check would prevent the erroneous program from starting to execute, namely
checking to ensure there are no free variables.
3 Strictly speaking, it seems anachronistic to refer to the lower and upper “case” for the Greek alphabet, since the language
predates moveable type in the West by nearly two millennia.

What can we say about τ1 and τ2 ? τ1 must be whatever type x’s argument has. Since x’s argument is
itself x, τ1 must be the same as the type of x. We just said that x has type τa . This immediately implies that
τa = τ1 →τ2 = τa →τ2
In other words,
τa = τa →τ2
What type can we write that satisfies this equation? In fact, no types in our type language can satisify
it, because this type is recursive without a base case. Any type we try to write will end up being infinitely
long. Since we cannot write an infinitely long type (recall that we’re trying to annotate ω, so if the type is
infinitely long, we’d never get around to finishing the text of the program), it follows by contradiction4 that
ω and Ω cannot be typed in our type system, and therefore their corresponding programs are not programs
in TFAE. (We are being rather lax here—what we’ve provided is informal reasoning, not a proof—but such
a proof does exist.)
26.3 Termination
We concluded our exploration of the type of Ω by saying that the annotation on the argument of ω must be
infinitely long. A curious reader ought to ask, is there any connection between the boundlessness of the type
and the fact that we’re trying to perform a non-terminating computation? Or is it mere coincidence?
TFAE, which is a first cousin of a language you’ll sometimes see referred to as the simply-typed lambda
calculus,5 enjoys a rather interesting property: it is said to be strongly normalizing. This intimidating term
says of a programming language that no matter what program you write in the language, it will always
terminate!
To understand why this property holds, think about our type language. The only way to create compound
types is through the function constructor. But every time we apply a function, we discharge one function
constructor: that is, we “erase an arrow”. Therefore, after a finite number of function invocations, the
computation must “run out of arrows”.6 Because only function applications can keep a computation running,
the computation is forced to terminate.
This is a very informal argument for why this property holds—it is cetainly far from a proof (though,
again, a formal proof of this property does exist). However, it does help us see why we must inevitably have
bumped into an infinitely long type while trying to annotate the infinite loop.
What good is a language without infinite loops? There are in fact lots of programs that we would like to
ensure will not run forever. These include:
• inner-loops of real-time systems
• program linkers
4 We implicitly assumed it would be possible to annotate ω and explored what that type annotation would be. The contradiction
is that no such annotation is possible.
5 Why “simply”? You’ll see what other options there are next week.
6 Oddly, this never happens to mythological heroes.
26.4. TYPED RECURSIVE PROGRAMMING 253
• packet filters in network stacks
• client-side Web scripts
• network routers
• device (such as photocopier) initialization
• configuration files (such as Makefiles)
and so on. That’s what makes the simply-typed lambda calculus so wonderful: instead of pondering and
testing endlessly (so to speak), we get mathematical certitude that, with a correct implementation of the type
checker, no infinite loops can sneak past us. In fact, the module system of the SML programming language
is effectively an implementation of the simply-typed lambda calculus, thereby guaranteeing that no matter
how complex a linking specification we write, the linking phase of the compiler will always terminate.
Exercise 26.3.1 We’ve been told that the Halting Problem is undecidable. Yet here we have a language
accompanied by a theorem that proves that all programs will terminate. In particular, then, the Halting
Problem is not only very decidable, it’s actually quite simple: In response to the question “Does this program
halt”, the answer is always “Yes!” Reconcile.
Exercise 26.3.2 While the simply-typed lambda calculus is fun to discuss, it may not be the most pliant
programming language, even as the target of a compiler (much less something programmers write explicitly).
Partly this is because it doesn’t quite focus on the right problem. To a Web browsing user, for instance, what
matters is whether a downloaded program runs immediately; five minutes isn’t really distinguishable from
non-termination.
Consequently, a better variant of the lambda calculus might be one whose types reflect resources, such
as time and space. The “type” checker would then ask the user running the program for resource bounds,
then determine whether the program can actually execute within the provided resources. Can you design
and implement such a language? Can you write useful programs in it?
26.4 Typed Recursive Programming

Strong normalization says we must provide an explicit recursion construct. To do this, we’ll simply reintro-
duce our rec construct to define the language TRCFAE. The BNF for the language is
<TRCFAE> ::= ...

| {rec {<id> : <type> <TRCFAE>} <TRCFAE>}
with the same type language. Note that the rec construct needs an explicit type annotation also.
What is the type judgment for rec? It must be of the form
???
Γ`{rec {i : τi v} b} : τ
since we want to conclude something about the entire term. What goes in the antecedent? We can determine
this more easily by realizing that a rec is a bit like an immediate function application. So just as with
functions, we’re going to have assumptions and guarantees—just both in the same rule.
We want to assume that τi is a legal annotation, and use that to check the body; but we also want to
guarantee that τi is a legal annotation. Let’s do them in that order. The former is relatively easy:
Γ[i←τi ]`b : τ ???

Now let’s hazard a guess about the form of the latter:
Γ[i←τi ]`b : τ Γ`v : τ

But what is the structure of the term named by v? Surely it has references to the identifier named by i in
it, but i is almost certainly not bound in Γ (and even if it is, it’s not bound to the value we want for i).
Therefore, we’ll have to extend Γ with a binding for i—not surprising, if you think about the scope of i in a
rec term—to check v also:
Γ[i←τi ]`b : τ Γ[i←τi ]`v : τ
Is that right? Do we want v to have type τ, the type of the entire expression? Not quite: we want it to have
the type we promised it would have, namely τi :
Γ[i←τi ]`b : τ Γ[i←τi ]`v : τi

Now we can understand how the typing of recursion works. We extend the environment not once, but
twice. The extension to type b is the one that initiates the recursion; the extension to type v is the one that
sustains it. Both extensions are therefore necessary. And because a type checker doesn’t actually run the
program, it doesn’t need an infinite number of arrows. When type checking is done and execution begins,
the run-time system does, in some sense, need “an infinite quiver of arrows”, but we’ve already seen how to
implement that in Section 10.
Exercise 26.4.1 Define the BNF entry and generate a type judgment for with in the typed language.
Exercise 26.4.2 Typing recursion looks simple, but it’s actually worth studying in detail. Take a simple
example such as Ω and work through the rules:
• Write Ω with type annotations so it passes the type checker. Draw the type judgment tree to make sure
you understand why this version of Ω types.
• Does the expression named by v in rec have to be a procedure? Do the typing rules for rec depend
on this?
Chapter 27
Typing Data
27.1 Recursive Types

27.1.1 Declaring Recursive Types
We saw in the previous lecture how rec was necessary to write recursive programs. But what about defining
recursive types? Recursive types are fundamental to computer science: even basic data structures like lists
and trees are recursive (since the rest of a list is also a list, and each sub-tree is itself a tree).
Suppose we try to type the program
{rec {length : ???

{fun {l : ???} : number
{if {empty? l}
0
{+ 1 {length {rest l}}}}}}
{length {numCons 1 {numCons 2 {numCons 3 numEmpty}}}}}
What should we write in place of the question marks?

Let’s consider the type of l. What kind of value can be an argument to l? Clearly a numeric cons,
because that’s the argument supplied in the first invocation of length. But eventually, a numeric empty is
passed to l also. This means l needs to have two types: (numeric) cons and empty.
In languages like ML (and Java), procedures do not consume arguments of more than one distinct type.
Instead, they force programmers to define a new type that encompasses all the possible arguments. This is
precisely what a datatype definition, of the kind we have been writing in Scheme, permits us to do. So let’s
try to write down such a datatype in a hypothetical extension to our (typed) implemented language:
{datatype numList
{[numEmpty]
[numCons {fst : number}
{rst : ???}]}
{rec {length : (numList -> number) ...}
{length ...}}}
255
256 CHAPTER 27. TYPING DATA
We assume that a datatype declaration introduces a collection of variants, followed by an actual body that
uses the datatype. What type annotation should we place on rst? This should be precisely the new type we
are introducing, namely numList.
A datatype declaration therefore enables us to do a few distinct things all in one notation:
1. Give names to new types.
2. Introduce conditionally-defined types (variants).
3. Permit recursive definitions.
If these are truly distinct, we should consider whether there are more primitive operators that we may provide
so a programmer can mix-and-match them as necessary.1
But how distinct are these three operations, really? Giving a type a new name would be only so useful
if the type were simple (for instance, creating the name bool as an alias for boolean may be convenient,
but it’s certainly not conceptually significant), so this capability is most useful when the name is assigned to
a complex type. Recursion needs a name to use for declaring self-references, so it depends on the ability to
introduce a new name. Finally, well-founded recursion depends on having both recursive and non-recursive
cases, meaning the recursive type must be defined as a collection of variants (of which at least one is not
self-referential). So the three capabilities coalesce very nicely.
As you may have noticed above, the datatypes we have introduced in our typed language are a bit
different from those we’re using in Scheme. Our Scheme datatypes are defined at the top-level, while those
in the implemented language enclose the expressions that refer to them. This is primarily to make it easier
to deal with the scope of the introduced types. Obviously, a full-fledged language (like ML and Haskell)
permits apparently top-level datatype declarations, but we’ll make this simplifying assumption here.
27.1.2 Judgments for Recursive Types

Let’s consider another example of a recursive type: a family tree.
{datatype FamilyTree
{[unknown]
[person {name : string}
{mother : FamilyTree}
{father : FamilyTree}]}
...}
This data definition allows us to describe as much of the genealogy as we know, and terminate the construc-
tion when we reach an unknown person. What type declarations ensue from this definition?
unknown : →FamilyTree
person : string × FamilyTree × FamilyTree→FamilyTree

1 “Programming languages should be designed not by piling feature on top of feature, but by removing the weaknesses and
restrictions that make additional features appear necessary”.

27.1. RECURSIVE TYPES 257
This doesn’t yet give us a way of distinguishing between the two variants, and of selecting the fields in each
variant. In Scheme, we use type-case to perform both of these operations. A corresponding case dispatcher
for the above datatype might look like
{FamilyTree-cases v
[{unknown} ...]
[{person n m f} ...]}
Its pieces would be typed as follows:
Γ`v : FamilyTree Γè1 : τ Γ[n←string, m←FamilyTree, f ←FamilyTree]è2 : τ

Γ`{FamilyTree-cases v {[unknown] e1 } {[person n m f ] e2 }} : τ
In other words, to determine the type of the entire FamilyTree-cases expression, τ, we first ensure that
the value being dispatched is of the right type. Then we must make sure each branch of the switch returns a
τ.2 We can ensure that by checking each of the bodies in the right type environment. Because unknown has
no fields, its cases branch binds no variables, so we check e1 in Γ. In the branch for person, however,
we bind three variables, so we must check the type of e2 in a suitably extended Γ.
Though the judgment above is for a very specific type declaration, the general principle should be clear
from what we’ve written. Effectively, the type checker introduces a new type rule for each typed cases
statement based on the type declaration at the time it sees the declaration. Writing the judgment above in
terms of subscripted parameters is tedious but easy.
Given the type rules above, consider the following program:
{[unknown]
{person "Mitochondrial Eve" {unknown} {unknown}}}
What is the type of the expression in the body of the datatype declaration? It’s FamilyTree. But when the
value escapes from the body of the declaration, how can we access it any longer? (We assume that the type
checker renames types consistently, so FamilyTree in one scope is different from FamilyTree in another
scope—just because the names are the same, the types should not conflate.) It basically becomes an opaque
type that is no longer usable. This does not appear to be very useful at all!3
At any rate, the type checker permitted a program that is quite useless, and we might want to prevent
this. Therefore, we could place the restriction that the type defined in the datatype (in this case, FamilyTree)
should be different from the type of the expression body τ. This prevents programmers from inadvertently
returning values that nobody else can use.
2 Based on the preceding discussion, if the two cases needed to return different types of values, how would you address this need
in a language that enforced the type judgment above?

3 Actually, you could use this to define the essence of a module or object system. These are called existential types. But we
won’t study them further in this course.

Obviously, this restriction doesn’t reach far enough. Returning a vector of FamilyTree values avoids the
restriction above, but the effect is the same: no part of the program outside the scope of the datatype can use
these values. So we may want a more stringent restriction: the type being different should not appear free
in τ.
This restriction may be overreaching, however. For instance, a programmer might define a new type,
and return a package (a vector, say) consisting of two values: an instance of the new type, and a procedure
that accesses the instances. For instance,
{[unknown]
{with {unknown-person : FamilyTree {unknown}}
{vector
{person "Mitochondrial Eve"
unknown-person
unknown-person}
{fun {v : FamilyTree} : string
{FamilyTree-cases v
[{unknown} {error ...}]
[{person n m f} n]}}}}}}
In this vector, the first value is an instance of FamilyTree, while the second value is a procedure of type
FamilyTree→string
Other values, such as unknown-person, are safely hidden from access. If we lift the restriction of the
previous paragraph, this becomes a legal pair of values to return from an expression. Notice that the pair in
effect forms an object: you can’t look into it, so the only way to access it is with the “public” procedure.
Indeed, this kind of type definition sees use in defining object systems.
That said, we still don’t have a clear description of what restriction to affix on the type judgment for
datatypes. Modern programming languages address this quandary by affixing no restriction at all. Instead,
they effectively force all type declarations to be at the “top” level. Consequently, no type name is ever
unbound, so the issues of this section do not arise. When we do need to restrict access, we employ module
systems to delimit the scope of type bindings.
27.1.3 Space for Datatype Variant Tags

One of the benefits programmers incur from using datatypes—beyond the error checking—is slightly better
space consumption. (Note: “better space consumption” = “using less space”.) Whereas without static type
checking we would need tags that indicate both the type and the variant, we now need to store only the
variant. Why? Because the type checker statically ensures that we won’t pass the wrong kind of value to
procedures! Therefore, the run-time system needs to use only as many bits as are necessary to distinguish
27.1. RECURSIVE TYPES 259
between all the variants of a type, rather than all the datatypes as well (in addition). Since the number of
variants is usually quite small, of the order of 3–4, the number of bits necessary for the tags is usually small
also.
We are now taking a big risk, however. In the liberal tagging regime, where we use both type and
variant tags, we can be sure a program will never execute on the wrong kind of data. But if we switch to a
more liberal tagging regime—one that doesn’t store type tags also—we run a huge risk. If we perform an
operation on a value of the wrong type, we may completely destroy our data. For instance, suppose we can
somehow pass a NumList to a procedure expecting a FamilyTree. If the FamilyTree-cases operation
looks only at the variant bits, it could end up accessing a numCons as if it were a person. But a numCons
has only two fields; when the program accesses the third field of this variant, it is essentially getting junk
values. Therefore, we have to be very careful performing these kinds of optimizations. How can we be sure
they are safe?
Chapter 28
Type Soundness
We would like a guarantee that a program that passes a type checker will never exhibit certain kinds of errors
when it runs. In particular, we would like to know that the type system did indeed abstract over values: that
running the type checker correctly predicted (up to the limits of the abstraction) what the program would
do. We call this property of a type system type soundness:1
For all programs p, if the type of p is τ, then p will evaluate to a value that has type τ.
Note that the statement of type soundness connects types with execution. This tells the user that the type
system is not some airy abstraction: what it predicts has bearing on practice, namely on the program’s
behavior when it eventually executes.
We have to be more careful about how we define type soundness. For instance, we say above (emphasis
added) “p will evaluate to a value such that . . . ”. But what if the program doesn’t terminate? So we must
recast this statement to say
For all programs p, if the type of p is τ and p evaluates to v, then v : τ.2
Actually, this isn’t quite true either. What if the program executes an expression like (first empty)? There
are a few options open to the language designer:
• Return a value such as −1. We hope you cringe at this idea! It means a program that fails to properly
check for return values at every single place will potentially produce nonsensical results. (Such errors
are common in C programs, where operators like malloc and fopen return special values but pro-
grammers routinely forget to check them.3 Indeed, many of these errors lead to expensive, frustrating
and threatening security violations.)
1 The term “soundness” comes from mathematical logic.
2 We could write this more explicitly as: For all programs p, if the type checker assigns p the type τ, and if the semantics says
that p evaluates to a value v, then the type checker will also assign v the type τ.
3 Ian Barland: “In Ginsberg’s Howl, the name of the ancient god Moloch is used to exemplify society’s embracing of soulless
machines, machines for the sake of machines rather than the elegance they can embody. I do not consider it coincidental that this
name sounds like ‘malloc’.”
261
262 CHAPTER 28. TYPE SOUNDNESS
• Diverge, i.e., go into an infinite loop. This approach is used by theoreticians (study the statement of
type soundness carefully and you can see why), but as software engineers we should soundly (ahem)
reject this.
• Raise an exception. This is the preferred modern solution.
Raising exceptions means the program does not terminate with a value, nor does it not terminate. We must
therefore refine this statement still further:
For all programs p, if the type of p is τ, p will, if it terminates, either evaluate to a value v such
that v : τ, or raise one of a well-defined set of exceptions.
The exceptions are a bit of a cop-out, because we can move arbitrarily many errors into that space. In
Scheme, for instance, the trivial type checker rejects no programs, and all errors fall under the exceptions.
In contrast, researchers work on very sophisticated languages where some traditional actions that would
raise an exception (such as violating array bounds) instead become type errors. This last phrase of the type
soundness statement therefore leaves lots of room for type system design.
As software engineers, we should care deeply about type soundness. To paraphrase Robin Milner, who
first proved a modern language’s soundness (specifically, for ML),
Well-typed programs do not go wrong.
That is, a program that passes the type checker (and is thus “well-typed”) absolutely cannot exhibit certain
classes of mistakes. 4
Why is type soundness not obvious? Consider the following simple program (the details of the numbers
aren’t relevant):
{if0 {+ 1 2}
{{fun {x : number} : number {+ 1 x}} 7}
{{fun {x : number} : number {+ 1 {+ 2 x}}} 1}}
During execution, the program will explore only one branch of the conditional:
1, 0/ ⇒ 1 2, 0/ ⇒ 2 ...
{+ 1 2}, 0/ ⇒ 3 {{fun ...} 1}}, 0/ ⇒ 4
{if0 {+ 1 2} {{fun ...} 7} {{fun ...} 1}}, 0/ ⇒ 4
but the type checker must explore both:
0`1
/ : number 0`2
/ : number ... ...
0`{+
/ 1 2} : number 0`{{fun
/ ...} 7} : number 0`{{fun
/ ...} 1} : number
0`{if0
/ {+ 1 2} {{fun ...} 7} {{fun ...} 1}} : number
4 The term “wrong” here is misleading. It refers to a particular kind of value, representing an erroneous configuration, in the
semantics Milner was using; in that context, this slogan is tongue-in-cheek. Taken out of context it is misleading, because a
well-typed program can still go wrong in the sense of producing erroneous output.
263
Furthermore, even for each expression, the proof trees in the semantics and the type world will be quite
different (imagine if one of them contains recursion: the evaluator must iterate as many times as necessary
to produce a value, while the type checker examines each expression only once). As a result, it is far from
obvious that the two systems will have any relationship in their answers. This is why a theorem is not only
necessary, but sometimes also difficult to prove.
Type soundness is, then, really a claim that the type system and run-time system (as represented by the
semantics) are in sync. The type system erects certain abstractions, and the theorem states that the run-time
system mirrors those abstractions. Most modern languages, like ML and Java, have this flavor.
In contrast, C and C++ do not have sound type systems. That is, the type system may define certain
abstractions, but the run-time system does not honor and protect these. (In C++ it largely does for object
types, but not for types inherited from C.) This is a particularly insidious kind of language, because the static
type system lulls the programmer into thinking it will detect certain kinds of errors, but it fails to deliver on
that promise during execution.
Actually, the reality of C is much more complex: C has two different type systems. There is one type
system (with types such as int, double and even function types) at the level of the program, and a
different type system, defined solely by lengths of bitstrings, at the level of execution. This is a kind of
“bait-and-switch” operation on the part of the language. As a result, it isn’t even meaningful to talk about
soundness for C, because the static types and dynamic type representations simply don’t agree. Instead,
the C run-time system simply interprets bit sequences according to specified static types. (Procedures like
printf are notorious for this: if you ask to print using the specifier %s, printf will simply print a
sequence of characters until it hits a null-teriminator: never mind that the value you were pointing to was
actually a double! This is of course why C is very powerful at low-level programming tasks, but how often
do you actually need such power?)
To summarize all this, we introduce the notion of type safety:
Type safety is the property that no primitive operation ever applies to values of the wrong type.
By primitive operation we mean not only addition and so forth, but also procedure application. A safe
language honors the abstraction boundaries it erects. Since abstractions are crucial for designing and main-
taining large systems, safety is a key software engineering attribute in a language. (Even most C++ libraries
are safe, but the problem is you have to be sure no legacy C library isn’t performing unsafe operations, too.)
Using this concept, we can construct the following table:
statically checked not statically checked
type safe ML, Java Scheme
type unsafe C, C++ assembly
The important thing to remember is, due to the Halting Problem, some checks simply can never be per-
formed statically; something must always be deferred to execution time. The trade-off in type design is to
maximize the number of these decisions statically without overly restricting the power of the programmer.
The designers of different languages have divergent views on the powers a programmer should have.
So what is “strong typing”? This appears to be a meaningless phrase, and people often use it in a non-
sensical fashion. To some it seems to mean “The language has a type checker”. To others it means “The
264 CHAPTER 28. TYPE SOUNDNESS
language is sound” (that is, the type checker and run-time system are related). To most, it seems to just
mean, “A language like Pascal, C or Java, related in a way I can’t quite make precise”. If someone uses this
phrase, be sure to ask them to define it for you. (For amusement, watch them squirm.)
Chapter 29
Explicit Polymorphism
29.1 Motivation
Earlier, we looked at examples like the length procedure (from now on, we’ll switch to Scheme with imagi-
nary type annotations):
(define lengthNum
(lambda (l : numlist) : number
(cond
[(numEmpty? l) 0]
[(numCons? l) (add1 (lengthNum (numRest l)))])))
If we invoke lengthNum on (list 1 2 3), we would get 3 as the response.
Now suppose we apply lengthNum to (list ’a ’b ’c). What do we expect as a response? We might expect
it to evaluate to 3, but that’s not what we’re going to get! Instead, we are going to get a type error (before
invocation can even happen), because we are applying a procedure expecting a numlist to a value of type
symlist (a list of symbols).
We can, of course, define another procedure for computing the length of lists of symbols:
(define lengthSym
(lambda (l : symlist) : number
(cond
[(symEmpty? l) 0]
[(symCons? l) (add1 (lengthSym (symRest l)))])))
Invoking lengthSym on (list ’a ’b ’c) will indeed return 3. But look closely at the difference between
lengthNum and lengthSym: what changed in the code? Very little. The changes are almost all in the type
annotations, not in the code that executes. This is not really surprising, because there is only one length
procedure in Scheme, and it operates on all lists, no matter what values they might hold.
This is an unfortunate consequence of the type system we have studied. We introduced types to reduce
the number of errors in our program (and for other reasons we’ve discussed, such as documentation), but in
the process we’ve actually made it more difficult to write some programs. This is a constant tension in the
265
266 CHAPTER 29. EXPLICIT POLYMORPHISM
design of typed programming languages. Introducing new type mechanisms proscribes certain programs,1
but in return it invalidates some reasonable programs, making them harder to write. The length example is
a case in point.
Clearly computing the length of a list is very useful, so we might be tempted to somehow add length
as a primitive in the language, and devise special type rules for it so that the type checker doesn’t mind
what kind of list is in use. This is a bad idea! There’s a principle of language design that says it’s generally
unadvisable for language designers to retain special rights for themselves that they deny programmers who
use their language. It’s unadvisable because its condescending and paternalistic. It suggests the language
designer somehow “knows better” than the programmer: trust us, we’ll build you just the primitives you
need. In fact, programmers tend to always exceed the creative bounds of the language designer. We can
already see this in this simple example: Why length and not reverse? Why length and reverse but not
append? Why all three and not map? Or filter or foldl and foldr or. . . . Nor is this restricted to lists: what
about trees, graphs, and so forth? In short, special cases are a bad idea. Let’s try to do this right.
29.2 Solution
To do this right, we fall back on an old idea: abstraction. The two length functions are nearly the same
except for small differences; that means we should be able to parameterize over the differences, define a
procedure once, and instantiate the abstraction as often as necessary. Let’s do this one step at a time.
Before we can abstract, we should identify the differences clearly. Here they are, boxed:
(define length Num
(lambda (l : num list) : number
(cond
[( num Empty? l) 0]
[( num Cons? l) (add1 (length Num ( num Rest l)))])))
(define length Sym

(lambda (l : sym list) : number
(cond
[( sym Empty? l) 0]
[( sym Cons? l) (add1 (length Sym ( sym Rest l)))])))
Because we want only one length procedure, we’ll drop the suffixes on the two names. We’ll also
abstract over the num and sym by using the parameter τ, which will stand (of course) for a type:
(define length
(lambda (l : τ list) : number
(cond
[( τ Empty? l) 0]
[( τ Cons? l) (add1 (length ( τ Rest l)))])))
1 It had better: if it didn’t prevent some programs, it wouldn’t catch any errors!
29.2. SOLUTION 267
It’s cleaner to think of list as a type constructor, analogous to how variants define value constructors:
that is, list is a constructor in the type language whose argument is a type. We’ll use an applicative notation
for constructors in keeping with the convention in type theory. This avoids the odd “concatenation” style of
writing types that our abstraction process has foisted upon us. This change yields
(define length
(lambda (l : list(τ)) : number
(cond
[( τ Empty? l) 0]
[( τ Cons? l) (add1 (length ( τ Rest l)))])))
At this point, we’re still using concatenation for the list operators; it seems to make more sense to make
those also parameters to Empty and Cons. To keep the syntax less cluttered, we’ll write the type argument
as a subscript:
(define length
(cond
[(Empty?τ l) 0]
[(Cons?τ l) (add1 (length (Restτ l)))])))
The resulting procedure declaration says that length consumes a list of any type, and returns a single number.
For a given type of list, length uses the type-specific empty and non-empty list predicates and rest-of-the-list
selector.
All this syntactic manipulation is hiding a great flaw, which is that we haven’t actually defined τ any-
where! As of now, τ is just a free (type) variable. Without binding it to specific types, we have no way of
actually providing different (type) values for τ and thereby instantiating different typed versions of length.
Usually, we have a simple technique for eliminating unbound identifiers, which is to bind them using a
procedure. This would suggest that we define length as follows:
(define length
(lambda (τ)
(cond
[(Empty?τ l) 0]
[(Cons?τ l) (add1 (length (Restτ l)))]))))
but this is horribly flawed! To wit:
1. The procedure length now has the wrong form: instead of consuming a list as an argument, it con-
sumes a value that it will bind to τ, returning a procedure that consumes a list as an argument.
2. The program isn’t even syntactically valid: there is no designation of argument and return type for the
procedure that binds τ!2
2 You might wonder why we don’t create a new type, call it type, and use this as the type of the type arguments. This is trickier
than it seems: is type also a type? What are the consequences of this?
3. The procedure bound to length expects one argument which is a type. This violates our separation
of the static and dynamic: types are supposed to be static, whereas procedure arguments are values,
which are dynamic.
So on the one hand, this seems like the right sort of idea—to introduce an abstraction—but on the other
hand, we clearly can’t do it the way we did above. We’ll have to be smarter.
The last complaint above is actually the most significant, both because it is the most insurmountable and
because it points the way to a resolution. There’s a contradiction here: we want to have a type parameter,
but we can’t have the type be a value. So how about we create procedures that bind types, and execute these
procedures during type checking, not execution time?
As always, name and conquer. We don’t want to use lambda for these type procedures, because lambda
already has a well-defined meaning: it creates procedures that evaluate during execution. Instead, we’ll
introduce a notion of a type-checking-time procedure, denoted by Λ (capital λ ). A Λ procedure takes only
types as arguments, and its arguments do not have further type annotations. We’ll use angles rather than
parentheses to denote their body. Thus, we might write the length function as follows:
(define length
<Λ (τ)
(cond
[(Empty?τ l) 0]
[(Cons?τ l) (add1 (length (Restτ l)))]))>)
This is a lot better than the previous code fragment, but it’s still not quite there. The definition of length
binds it to a type procedure of one argument, which evaluates to a run-time procedure that consumes a list.
Yet length is applied in its own body to a list, not to a type.
To remedy this, we’ll need to apply the type procedure to an argument (type). We’ll again use the angle
notation to denote application:
(define length
<Λ (τ)
(cond
[(Empty?τ l) 0]
[(Cons?τ l) (add1 (length<τ> (Restτ l)))]))>)
If we’re going to apply length to τ, we might as well assume Empty?, Cons? and Rest are also type-
procedures, and supply τ explicitly through type application rather than through the clandestine subscript
currently in use:
(define length
<Λ (τ)
(cond
[(Empty?<τ> l) 0]
[(Cons?<τ> l) (add1 (length<τ> (Rest<τ> l)))]))>)
29.3. THE TYPE LANGUAGE 269
Thus, an expression like (Rest<τ> l) first applies Rest to τ, resulting in an actual rest procedure that applies
to lists of values of type τ; this procedure consumes l as an argument and proceeds as it would in the type-
system-free case. In other words, every type-parameterized procedure, such as Rest or length, is a generator
of infinitely many procedures that each operate on specific types. The use of the procedure becomes
(length<num> (list 1 2 3))
(length<sym> (list ’a ’b ’c))
We call this language parametrically polymorphic with explicit type parameters. The term polymor-
phism means “having many forms”; in this case, the polymorphism is induced by the type parameters,
where each of our type-parameterized procedures is really a representative of an infinite number of func-
tions that differ only in the type parameter. The “explicitly” comes from the fact that our language forces
the programmer to write the Λ’s and type application.
29.3 The Type Language

As a result of these ideas, our type language has grown considerably richer. In particular, we now permit
type variables as part of the type language. These type variables are introduced by type procedures (Λ), and
discharged by type applications. How shall we write such types? We may be tempted to write
length : type → (list(type) → number)
but this has two problems: first, it doesn’t distinguish between the two kinds of arrows (“type arrows” and
“value arrows”, corresponding to Λ and lambda, respectively), and secondly, it doesn’t really make clear
which type is which, a problem if there are multiple type parameters:
map : type, type → list(type) × (type → type) → list(type)
Instead, we adopt the following notation:
length : ∀α. list(α) → number
where it’s understood that every ∀ parameter is introduced by a type procedure (Λ).3 Here are the types for
a few other well-known polymorphic functions:
filter : ∀α. list(α) × (α → boolean) → list(α)

map : ∀α, β . list(α) × (α → β ) → list(β )
The type of map, in particular, makes this type notation is superior to our previous proposal: when multiple
types are involved, we must give each one a name to distinguish between them.
3 It’sconventional to use the beginning of the Greek alphabet—α, β and so on—as the canonical names of polymorphic types,
rather than begin from τ. This has two reasons. First, τ is conventionally a meta-variable, whereas α and β are type variables.
Second, not many people know what Greek letter comes after τ. . . .
29.4 Evaluation Semantics and Efficiency

While we have introduced a convenient notation, we haven’t entirely clarified its meaning. In particular,
it appears that every type function application actually happens during program execution. This seems
extremely undesirable for two reasons:
• it’ll slow down the program, in comparison to both the typed but non-polymorphic programs (that we
wrote at the beginning of the section) and the non-statically-typed version, which Scheme provides;
• it means the types must exist as values at run-time.

Attractive as it may seem to students who see this for the first time, we really do not want to permit types
to be ordinary values. A type is an abstraction of a value; conceptually, therefore, it does not make any sense
for the two to live in the same universe. If the types were not supplied until execution, the type checker not
be able to detect errors until program execution time, thereby defeating the most important benefit that types
confer.
It is therefore clear that the type procedures must accept arguments and evaluate their bodies before the
type checker even begins execution. By that time, if all the type applications are over, it suffices to use the
type checker built earlier, since what remains is a language with no type variables remaining. We call the
phase that performs these type applications the type elaborator.
The problem with any static procedure applications is to ensure they will lead to terminating processes!
If they don’t, we can’t even begin the next phase, which is traditional type checking. In the case of using
length, the first application (from the procedure use) is on the type num. This in turn inspires a recursive
invocation of length also on type num. Because this latter procedure application is no different from the
initial invocation, the type expander does not need to perform the application. (Remember, if the language
has no side-effects, computations will return the same result every time. Type application has no side-
effects.)
This informal argument suggests that only one pass over the body is necessary. We can formalize this
with the following type judgments:
Γè : ∀α.τ
Γè < τ 0 >: τ[α←τ 0 ]
This judgment says that on encountering a type application, we substitute the quantified type with the type
argument replacing the type variable. The program source contains only a fixed number of type applications
(even if each of these can execute arbitrarily many times), so the type checker performs this application only
once. The corresponding rule for a type abstraction is
Γ[α]è : τ
Γ` < Λ (α) e >: ∀α.τ
This says that we extend Γ with a binding for the type variable α, but leave the associated type unspecified
so it is chosen nondeterministically. If the choice of type actually matters, then the program must not type-
check.
Observe that the type expander conceptually creates many monomorphically typed procedures, but we
don’t really want most of them during execution. Having checked types, it’s fine if the length function that
actually runs is essentially the same as Scheme’s length. This is in fact what most evaluators do. The static
type system ensures that the program does not violate types, so the program that runs doesn’t need type
checks.
29.5 Perspective
Explicit polymorphism seems extremely unwieldy: why would anyone want to program with it? There
are two possible reasons. The first is that it’s the only mechanism that the language designer gives for
introducing parameterized types, which aid in code reuse. The second is that the language includes some
additional machinery so you don’t have to write all the types every time. In fact, C++ introduces a little of
both (though much more of the former), so programmers are, in effect, manually programming with explicit
polymorphism virtually every time they use the STL (Standard Template Library). Similarly, the Java 1.5 and
C# languages support explicit polymorphism. But we can possibly also do better than foist this notational
overhead on the programmer.
Chapter 30
Type Inference
30.1 Inferring Types

We’ve seen the value of having explicit polymorphism in our language—it lets us write programs that work
on may different types of values. Even mainstream languages like C++ and, more recently, Java have recog-
nized the value of this form of parameteric polymorphism, and they have noticed that it complements and is
not subsumed by the polymorphism common to object-oriented languages (called subtype polymorphism).
Despite its benefits, it’s very unwieldy to use explicit parametric polymorphism to write programs such
as this:
(define length
<Λ (τ)
(cond
[(Empty?<τ> l) 0]
[(Cons?<τ> l) (add1 (length<τ> (Rest<τ> l)))]))>)
when we could instead write
(define length
(lambda (l)
(cond
[(empty? l) 0]
[(cons? l) (add1 (length (rest l)))])))
As computer scientists, we should ask: Is it possible for a programming environment to convert the latter
into the former? That is, can the environment automatically fill in the types necessary for the former? This
would be the best of both worlds, because the programmer would avoid the trouble of the typing while still
getting the benefit of the typing.
While this would be nice, it also seems nearly magical. It seems hard enough for humans to get this
right; can a program (the environment) do better? Still, we should not despair too much. We’ve already seen
several instances such as closure creation, garbage collection, and so on, where the language implementation
273
274 CHAPTER 30. TYPE INFERENCE
was able to do a more accurate job than a human could have done, thereby providing a valuable feature while
reducing the programmer’s burden. Maybe inserting type annotations could be another of those tasks.
Because this is obviously challenging, let’s try to make the problem easier. Let’s ignore the polymor-
phism, and just focus on generating types for monomorphic programs (i.e., those that don’t employ poly-
morphism). In fact, just to make life really simple, let’s just consider a program that operates over numbers,
such as factorial.
30.1.1 Example: Factorial

Suppose we’re given the following program:
(define fact
(lambda (n)
(cond
[(zero? n) 1]
[true (∗ n (fact (sub1 n)))])))
We’ve purposely written true instead of else for reasons we’ll soon see. It should be clear that using true
doesn’t affect the meaning of the program (in general, else is just a more readable way of writing true).
If we were asked to determine the type of this function and had never seen it before, our reasoning might
proceed roughly along these lines. First, we would name each expression:
(define fact
1 (lambda (n)
2 (cond
[ 3 (zero? n) 4 1]
[ 5 true 6 (∗ n 7 (fact 8 (sub1 n)))])))
We would now reason as follows. We’ll use the notation [[ · ]] to mean the type of the expression within the
brackets.
• The type of the expression labeled 1 1 is clearly a function type (since the expression is a lambda).
The function’s argument type is that of n, and it computes a value with the type of 2 . In other words,
[[ 1 ]] = [[n]]→[[ 2 ]]
• Because 2 is a conditional, we want to ensure the following:
– The first and second conditional expressions evaluate to boolean values. That is, we would like
the following to hold:
[[ 3 ]] = boolean
[[ 5 ]] = boolean
1 We’ll need to use this phrase repeatedly, and it’s quite a mouthful. Therefore, we will henceforth say “the type of n ” when
we mean “the type of the expression labeled by n ”.

30.1. INFERRING TYPES 275
– We would like both branches of the conditional to evaluate to a value of the same type, so we
can assign a meaningful type to the entire conditional expression:
[[ 2 ]] = [[ 4 ]] = [[ 6 ]]
• What is the type of 3 ? We have a constraint on what it can be:
[[zero?]] = [[n]]→[[ 3 ]]
Because we know the type of zero?, we know that the right-hand-side of the above equality must be:
[[n]]→[[ 3 ]] = number→boolean
which immediately tells us that [[n]] = number.
The first response in the cond tells us that [[ 4 ]] = number, which immediately resolves the type of 2 and
determines the type of 1 in atomic terms. That is, the type of fact must be number → number. However,
it’s worthwhile to continue with this process as an illustration:
• We have a constraint on the type of 6 : it must be the same as the result type of multiplication.
Concretely,
[[n]] × [[ 7 ]]→[[ 6 ]] = number × number→number
• The type of 7 must be whatever fact returns, while 8 must be the type that fact consumes:
[[ 1 ]] = [[ 8 ]]→[[ 7 ]]
• Finally, the type of 8 must be the return type of sub1:
[[sub1]] = [[n]]→[[ 8 ]] = number→number
30.1.2 Example: Numeric-List Length

Now let’s look at a second example:
(define nlength
(lambda (l)
(cond
[(nempty? l) 0]
[(ncons? l) (add1 (nlength (nrest l)))])))
First, we annotate it:
(define nlength
1 (lambda (l)
2 (cond
[ 3 (nempty? l) 4 0]
[ 5 (ncons? l) 6 (add1 7 (nlength 8 (nrest l)))])))
We can begin by deriving the following constraints:
[[ 1 ]] = [[l]]→[[ 2 ]]
[[ 2 ]] = [[ 4 ]] = [[ 6 ]]
[[ 3 ]] = [[ 5 ]] = boolean
Because 3 and 5 are each applications, we derive some constraints from them:
[[nempty?]] = [[l]]→[[ 3 ]] = numlist→boolean
[[ncons?]] = [[l]]→[[ 5 ]] = numlist→boolean

The first conditional’s response is not very interesting:2
[[ 4 ]] = [[0]] = number
Finally, we get to the second conditional’s response, which yields several constraints:
[[add1]] = [[ 7 ]]→[[ 6 ]] = number→number
[[ 1 ]] = [[ 8 ]]→[[ 7 ]]
[[nrest]] = [[l]]→[[ 8 ]] = numlist→numlist

Notice that in the first and third set of constraints above, because the program applies a primitive, we can
generate an extra constraint which is the type of the primitive itself. In the second set, because the function
is user-defined, we cannot generate any other meaningful constraint just by looking at that one expression.
Solving all these constraints, it’s easy to see both that the constraints are compatible with one another,
and that each expression receives a monomorphic type. In particular, the type of 1 is numlist → number,
which is therefore the type of nlength also (and proves to be compatible with the use of nlength in expression
7 ).
30.2 Formalizing Constraint Generation

What we’ve done so far is extremely informal. Let’s formalize it.
Constraints relate different portions of the program by determining how they should be compatible
for the program to execute without error. Consequently, a single program point may result in multiple
constraints. Each set of constraints represents a “wish list” about that particular point in the program.
Consequently, a program may lead to contradictory constraints; hopefully we will be able to find these later.
One slightly confusing aspect of constraints is that we write them to look like equations, but they reflect
2 Note that the 0 inside the [[ · ]] is an expression itself, not a number labeling an expression.
30.3. ERRORS 277
what we hope will be true, not what we know is true. Specifically, they represent what suffices for safe
program execution.3
For each expression n in the program’s abstract syntax tree, we introduce a variable of the form [[n]].
That is, if the program has the form (foo 1 2), we would want to introduce variables for 1, 2 and (foo 1 2).
Because abstract syntax tree nodes are unwieldy to write down explicitly, we will associate the node with
the expression at that node. We use [[ · ]] to represent the type of a node, so the types of the expressions in
the example above would be [[1]], [[2]] and [[(foo 1 2)]].
Each expression type generates different constraints. We present below a table that relates the type of
expression at a node to the (set of) constraints generated for that node. Remember to always read [[ · ]] as
“the type of the expression” (within the brackets):
Expression at Node Generated Constraints

n, where n is a numeral [[n]] = number
true [[true]]=boolean
false [[false]]=boolean
(add1 e) [[(add1 e)]]=number [[e]]=number
(+ e1 e2) [[(+ e1 e2)]]=number [[e1]] = number [[e2]] = number
(zero? e) [[(zero? e)]]=boolean [[e]] = number
(ncons e1 e2) [[(ncons e1 e2)]]=list(num) [[e1]] = number [[e2]] = list(num)
(nfirst e) [[(nfirst e)]]=number [[e]] = list(num)
(nrest e) [[(nrest e)]]=list(num) [[e]] = list(num)
(nempty? e) [[(nempty? e)]]=boolean [[e]] = list(num)
nempty [[nempty]]=list(num)
(if c t e) [[(if c t e)]]=[[t]] [[(if c t e)]]=[[e]] [[c]]=boolean
(lambda (x) b) [[(lambda (x) b)]]= [[x]]→[[b]]
(f a) [[f ]] = [[a]]→[[(f a)]]
Notice that in the two prior examples, we did not create new node numbers for those expressions that
consisted of just a program identifier; correspondingly, we have not given a rule for identifiers. We could
have done this, for consistency, but it would have just created more (unnecessary) variables.
Exercise 30.2.1 What is the complexity of constraint generation?
Exercise 30.2.2 Using the expression at the node, rather than the node itself, introduces a subtle ambiguity.
Do you see why?
30.3 Errors
Here’s an erroneous program:
3 We use the term “suffices” advisedly: these constraints are sufficient but not necessary. They may reject some programs that
might have run without error had the type system not intervened. This is inherent in the desire to statically approximate dynamic
behavior: the Halting Problem is an insurmountable obstacle. An important constraint on good type system design is to maximize
the set of legal problems while still not permitting errors: balancing programmer liberty with execution safety.
(define nlsum
(lambda (l)
(cond
[(nempty? l) 0]
[(ncons? l) (+ (nrest l)
(nlsum (nrest l)))])))
Can you spot the problem?
First, we’ll annotate the sub-expressions:
(define nlsum
1 (lambda (l)
2 (cond
[ 3 (nempty? l) 4 0]
[ 5 (ncons? l) 6 (+ 7 (nrest l)
8 (nlsum 9 (nrest l)))])))
Generating constraints as usual, we get the following (amongst others):

[[ 8 ]] = number
because the function returns a number in both branches of the conditional, and
[[ 9 ]] = numlist
from the type of nrest. Consequently, it appears we can infer that the value bound to nlsum has the type
numlist → number. This is indeed the type we expect for this procedure.
We should not, however, annotate any types before we’re generated, examined and resolved all the
constraints: we must make sure there are no inconsistencies. Completing the generation and solution process
does, in fact, result in an inconsistency for this program. In particular, we have
[[ 7 ]] = numlist
from the type of nrest, while
[[ 7 ]] = number
from the type of +. Indeed, the latter is the type we want: the numlist only materializes because of the faulty
use of nrest. Had the programmer used nfirst instead of nrest in the left-hand-side argument to the addition,
the entire program would have checked correctly. Instead, the type inference engine must recognize that
there is an error resulting from a type conflict: the expression (nrest l) is expected to have both the type
number and the type numlist. Because these are not compatible types, the type “checker” must present the
user with this error.
We use quotes around “checker” because it has, in some sense, disappeared. Instead of checking types
annotated by the programmer, the type system now tries to fill in the programmer’s annotations. If it suc-
ceeds, it can do so only by respecting the types of operations, so there is no checking left to be done. Failure
to annotate the program completely and unambiguously is now the indication of a type error.
30.4. EXAMPLE: USING FIRST-CLASS FUNCTIONS 279
30.4 Example: Using First-Class Functions

We will consider one final example of constraint generation, to show that the process scales in the presence
of functions as arguments. Consider the following program:
(define nmap
1 (lambda (f l)
2 (cond
[ 3 (nempty? l) 4 nempty]
[ 5 (ncons? l) 6 (ncons 7 (f 8 (nfirst l))
9 (nmap f 10 (nrest l)))])))
This program generates the following constraints:
[[ 1 ]] = [[f ]] × [[l]]→[[ 2 ]]
We get the usual constraints about boolean conditional tests and the type equality of the branches (both must
be of type numlist due to the first response). From the second response, we derive
[[ncons]] = [[ 7 ]] × [[ 9 ]]→[[ 6 ]] = number × numlist→numlist
The most interesting constraint is this one:
[[f ]] = [[ 8 ]]→[[ 7 ]]
In other words, we don’t need a sophisticated extension to handle first-class functions: the constraint gener-
ation phase we described before suffices.
Continuing, we obtain the following three constraints also:
[[nfirst]] = [[l]]→[[ 8 ]] = numlist→number
[[nmap]] = [[f ]] × [[ 10 ]]→[[ 9 ]]

[[nrest]] = [[l]]→[[ 10 ]] = numlist→numlist
Since l is of type numlist, we can substitute and solve to learn that f has type number → number. Conse-
qently, nmap has type
(number → number) × numlist → numlist
which is the type we would desire and expect!

30.5 Solving Type Constraints

30.5.1 The Unification Algorithm
To solve type constraints, we turn to a classic algorithm: unification. Unification consumes a set of con-
straints and either
• identifies inconsistencies amongst the constraints, or
• generates a substitution that represents the solution of the constraints.
A substitution associates each identifier (for which we are trying to solve) with a constant or another iden-
tifier. Our identifiers represent the types of expressions (thus 4 is a funny notation for an identifier that
represents the type of the expression labeled 4). In our universe, inconsistencies indicate type errors, and
the constants are terms in the type language (such as number and number → boolean).
The unification algorithm is extremely simple. Begin with an empty substitution. Push all the constraints
onto a stack. If the stack is empty, return the substitution; otherwise, pop the constraint X = Y off the stack:
1. If X and Y are identical identifiers, do nothing.
2. If X is an identifier, replace all occurrences of X by Y both on the stack and in the substitution, and
add X 7→ Y to the substitution.
3. If Y is an identifier, replace all occurrences of Y by X both on the stack and in the substitution, and
add Y 7→ X to the substitution.
4. If X is of the form C(X1 , . . . , Xn ) for some constructor C,4 and Y is of the form C(Y1 , . . . ,Yn ) (i.e., it
has the same constructor), then push Xi = Yi for all 1 ≤ i ≤ n onto the stack.
5. Otherwise, X and Y do not unify. Report an error.
Does this this algorithm terminate? On every iteration of the main loop, it pops a constraint off the stack. In
some cases, however, we push new constraints on. The size of each of these constraints is, however, smaller
than the constraint just popped. Therefore, the total number of iterations cannot be greater than the sum of
the sizes of the initial constraint set. The stack must therefore eventually become empty.
Exercise 30.5.1 What are the space and time complexity of this algorithm?
30.5.2 Example of Unification at Work

Let’s consider the following example:
1 ( 2 (lambda (x) x)
3 7)
4 In our type language, the type constructors are → and the base types (which are constructors of arity zero). More on this in
Section 30.5.3.
30.5. SOLVING TYPE CONSTRAINTS 281
This generates the following constraints:
[[ 2 ]] = [[ 3 ]]→[[ 1 ]]
[[ 2 ]] = [[x]]→[[x]]
[[ 3 ]] = number
The unification algorithm works as follows:
Action Stack Substitution

Initialize [[ 2 ]] = [[ 3 ]]→[[ 1 ]] empty
[[ 2 ]] = [[x]]→[[x]]
[[ 3 ]] = number
Step 2 [[ 3 ]]→[[ 1 ]] = [[x]]→[[x]] [[ 2 ]] 7→ [[ 3 ]]→[[ 1 ]]
[[ 3 ]] = number
Step 4 [[ 3 ]] = [[x]] [[ 2 ]] 7→ [[ 3 ]]→[[ 1 ]]
[[ 1 ]] = [[x]]
[[ 3 ]] = number
Step 2 [[ 1 ]] = [[x]] [[ 2 ]] 7→ [[x]]→[[ 1 ]]
[[x]] = number [[ 3 ]] 7→ [[x]]
Step 2 [[x]] = number [[ 2 ]] 7→ [[x]]→[[x]]
[[ 3 ]] 7→ [[x]]
[[ 1 ]] 7→ [[x]]
Step 2 empty [[ 2 ]] 7→ number→number
[[ 3 ]] 7→ number
[[ 1 ]] 7→ number
[[x]] 7→ number
At this point, we have solutions for all the sub-expressions and we know that the constraint set is consistent.
Writing these in detail is painstaking, but it’s usually easy to simulate this algorithm on paper by just
crossing out old values when performing a substitution. Be sure to work through our examples for more
practice with unification!
30.5.3 Parameterized Types

In the presentation of unification above, we saw only one type constructor with positive arity: →. A regular
programming language will typically have many more constuctors. A common source of parameterized
types is containers: lists, trees, queues, stacks, and so forth. For instance, it is common to think of lists
as parameterized over their content, thus yielding list(number), list(symbol), list(list(number)), list(number
→ symbol) and so on. Identifying list as one of the type constructors for the unification algorithm suffices
for typing unityped lists.
30.5.4 The “Occurs” Check
Suppose we generate the following type constraint:
list([[x]]) = list(list([[x]]))
By Step 4, we should push [[x]] = list([[x]]) onto the stack. Eventually, obeying to Step 2, we will add the
mapping [[x]] 7→ list([[x]]) to the substition but, in the process, attempt to replace all instances of [[x]] in the
substitution with the right-hand side, which does not terminate.
This is a familiar problem: we saw it earlier when trying to define substitution in the presence of recur-
sion. Because these are problematic to handle, a traditional unification algorithm checks in Steps 2 and 3
whether the identifier about to be (re)bound in the substitution occurs in the term that will take its place. If
the identifier does occur, the unifier halts with an error.5 Otherwise, the algorithm proceeds as before.
Exercise 30.5.2 Write a program that will generate the above constraint!
30.6 Underconstrained Systems

We have seen earlier that if the system has too many competing constraints—for instance, forcing an iden-
tifier to have both type number and boolean—there can be no satisfying type assignment, so the system
should halt with an error. We saw this informally earlier; Step 5 of the unification algorithm confirms that
the implementation matches this informal behavior.
But what if the system is under-constrained? This is interesting, because some of the program identifiers
never get assigned a type! In a procedure such as map, for instance:
(define (map f l)
(cond
[(empty? l) empty]
[(cons? l) (cons (f (first l))
(map f (rest l)))]))
This is an instructive example. Solving the constraints reveals that there is no constraint on the type passed
as a parameter to the list type constructor. Working through the steps, we get a type for map of this form:
(α→β ) × list(α)→list(β )
where α and β are unconstrained type identifiers. This is the same type we obtained through explicit
parametric polymorphism. . . except that the unification algorithm has found it for us automatically!
5 This is not the only reasonable behavior! It is possible to define fixed-point types, which are solutions to the circular constraint
equations above. This topic is, however, beyond the scope of this text.
30.7. PRINCIPAL TYPES 283
30.7 Principal Types

The type generated by this Hindley-Milner6 system has a particularly pleasing property: it is a principal
type. What does that mean? For a term t, consider a type τ. τ is a principal type of t if, for any other type τ 0
that types t, there exists a substitution (perhaps empty) that, when applied to τ, yields τ 0 .
There are a few ways of re-phrasing the above:
• The Hindley-Milner type system infers the “most general” type for a term.
• The type generated by the Hindler-Milner type system imposes fewest constraints on the program’s
behavior. In particular, it imposes constraints necessary for type soundness, but no more.
From a software engineering perspective, this is very attractive: it means a programmer could not possi-
bly annotate a procedure with a more general type than the type inference algorithm would derive. Thus,
using the algorithm instead of performing manual annotation will not restrict the reusability of code, and
may even increase it (because the programmer’s annotation may mistakenly overconstrain the type). Of
course, there are other good reasons for manual annotation, such as documentation and readability, so a
good programming style will mix annotation and inference judiciously.
6 Named for Roger Hindley and Robin Milner, who independently discovered this in the late 1960s and early 1970s.
Chapter 31
Implicit Polymorphism
31.1 The Problem

Consider the function
(lambda (x) x)
The type inference engine would infer that this function has type
α→α
(or some other type, modulo renaming the type variable).

Now consider a program of the form (the let construct is similar to the with in our interpreted language):
(let ([id (lambda (x) x)])
(+ (id 5)
(id 6)))
First we need a type judgment for let. Here is a reasonable one: it is exactly what one gets by using the
existing rules for functions and applications, since we have consistently defined with as an application of
an immediate function:
Γ`v : τ 0 Γ[x←τ 0 ]`b : τ
Γ`(let([xv])b) : τ
Given this judgment, the type variable α in the type inferred for id would unify with the type of 5
(namely, number) at one application and with the type of 6 (also number) at the other application. Since
these are consistent, the unification algorithm would conclude that id is being used as if it had the type
number→number
in this program.
Now suppose we use id in a context where we apply it to values of different types. The following
program is certainly legal in Scheme:
285
286 CHAPTER 31. IMPLICIT POLYMORPHISM
(if (id true)

(id 5) ;; then
(id 6))) ;; else
This should be legal even in typed Scheme, because we’re returning the same type of value in both branches
of the conditional. But what happens when we supply this program to our type system? It infers that id
has type α→α. But it then unifies α with the type of each of the arguments. Two of these are the same
(number) but the first application is to a value of type boolean. This forces the type inference algorithm
to try and unify number with boolean. Since these are distinct base types, type inference fails with a type
error!
We definitely do not want this program to be declared erroneous. The problem is not with the program
itself, but with the algorithm we employ to infer types. That suggests that we should try to improve the
algorithm.
31.2 A Solution
What’s the underlying problem in the type algorithm? We infer that id has type α→α; we are then stuck
with that type for every use of id. It must therefore be either number → number or boolean → boolean—but
it can’t be both. That is, we cannot use it in a truly polymorphic manner!
This analysis makes clear that the problem has something to do with type variables and their unification.
We arrive at a contradiction because α must unify with both number and boolean. But what if α didn’t
need to do that? What if we didn’t use the same type variable every time? Then perhaps we could avoid the
problem entirely.
One way to get fresh type variables for each application of id is to literally substitute the uses of id with
their value. That is, instead of type checking the program above, suppose we were to check the following
program:
(if ((lambda (x) x) true)
((lambda (x) x) 5)
((lambda (x) x) 6)))
We don’t want to have to write this program, of course, but that’s okay: a simple pre-processor can substitute
every let-bound identifier in the body before type-checking. If we did that, we get a different result from
type-checking:
(if ( 1 (lambda (x) x) true)
( 2 (lambda (x) x) 5)
( 3 (lambda (x) x) 6)))
Each use of id results in a different type; for instance, the id procedure at 1 might have type α→α, 2
might have type β →β and 3 might have γ→γ. Then α could unify with type boolean, β with type number
and γ with type number. Because these are distinct type variables, they need not unify with one another.
31.3. A BETTER SOLUTION 287
Each application would succeed, and the entire program would successfully pass the type checker. This in
fact corresponds more accurately with what happens during execution, because on the first invocation the
identifier x in id holds a value of boolean type, and on the subsequent invocation (in the first branch of the
conditional, in this case) it holds a number. The separate type variables accurately reflect this behavior.
31.3 A Better Solution

The solution we’ve presented above has two problems:
1. It can lead to considerable code size explosion. For instance, consider this program:
(let ([x
(let ([y
(let ([z 3])
(+ z z))])
(+ y y))])
(+ x x))
Expand it in full. In general, how big can a program grow upon expansion?
2. Since let does not permit recursion, consider letrec or local, the Scheme analog of rec. What happens
if we substitute code in a recursive definition?
In short, the code substitution solution is not workable, but it does contain the germ of a good idea. We see
that what it does is generate fresh type variables at every use: this is the essence of the solution. So perhaps
we can preserve the essence while dispensing with that particular implementation.
Indeed, we can build further on the intuition we have developed. A closure has only one name for an
identifier, but a closure can be used multiple times, even recursively, without confusion. This is because,
in effect, each application consistently renames the bound identifier(s) so they cannot be confused across
instances. Working backwards, since we want fresh identifiers that cannot be confused across instances, we
want to create an analogous type closure that we instantiate at every use of a polymorphic function.
We will therefore use a modified rule for typing let:
Γ`v : τ 0 Γ[x←CLOSE(τ 0 )]`b : τ
That is, we bind x to a “closed” type when we check the body. The idea is, whenever we encounter this
special type in the body, we instantiate its type variables with fresh type variables:
Γè : CLOSE(τ 0 )
Γè : τ
where τ is the same as τ 0 , except all type variables have been renamed consistently to unused type variables.
Returning to the identity procedure, the type inference algorithm binds id to the type CLOSE(α→α).
At each use of id, the type checker renames the type variables, generating types such as α1 →α1 , α2 →α2 ,
and so on. As we have seen before, these types permit the body to successfully type check. Therefore, we
have successfully captured the intuition behind code-copying without the difficulties associated with it.
31.4 Recursion
We have provided a rule for let above, but in fact a similar rule can apply to letrec also. There are some
subtleties that we must defer to a more advanced setting, but safe uses of letrec (namely, those where the
right-hand side is syntactically a procedure) can safely employ the type closure mechanism described above
to infer polymorphic types.1
31.5 A Significant Subtlety

Alas something is still rotten in the state of inferring polymorphic types. When we rename all type variables
in a CLOSE type, we may rename variables that were not bound in the let or letrec expression: for instance,
(lambda (y)
(let ([f (lambda (x) y)])
(if (f true)
(+ (f true) 5)
6)))
Our algorithm would infer the type CLOSE(α→β ) (or the equivalent under renaming) for f . (Because x
and y are not linked in the body, the inference process assigns them potentially different types; hence the
presence of both α and β in the type.)
At the first application, in the test of the conditional, we generate fresh type names, α1 and β1 . The
type α1 unifies with boolean, and β1 unifies with boolean (since it’s used in a conditional context). At the
second application, the algorithm generates two fresh names, α2 and β2 . α2 will unify with boolean (since
that is the type of the argument to f ), while β2 unifies with number, because the entire expression is the first
argument to addition. Reasoning thus, we can see that the program successfully passes the type checker.
But this program should fail! Simply looking at it, it’s obvious that f can return either a boolean or a
numeric value, but not both. Indeed, if we apply the entire expression to true, there will be a type error at
the addition; if we apply it to 42, the type error will occur at the conditional. Sure enough, in our earlier
type systems, it would have failed with an error while unifying the constraints on the return types of f . So
how did it slip through?
The program successfully passed the type checker because of our use of type closures. We did not,
however, correctly apply our intuition about closures. When we apply a closure, we only get new identifiers
for those bound by the closures—not those in its lexical scope. The variables in the closure’s lexical scope
are shared between all applications of the closure. So should it be in the case of type closures. We should
only generate fresh type variables for the types introduced by the let or letrec.
Concretely, we must modify our rule for let (and correspondingly that for letrec) so the type closures
track which identifiers must be renamed:
Γ`v : τ 0 Γ[x←CLOSE(τ 0 , Γ)]`b : τ

1 In particular, we are no longer using code-copying, which encounters an obvious difficulty in the presence of recursion.
31.6. WHY LET AND NOT LAMBDA? 289
That is, a type closure tracks the environment of closure creation. Correspondingly,
Γè : CLOSE(τ 0 , Γ0 )
Γè : τ
where τ is the same as τ 0 , except the renaming applies only to type variables in τ 0 that are not bound by Γ0 .
Applying these rules to the example above, we rename the α’s but not β , so the first use of f gets type
α1 →β and the second use α2 →β . This forces β = number = boolean, which results in a type error during
unification.
31.6 Why Let and not Lambda?

The kind of polymorphism we have seen above is called let-based polymorphism, in honor of the ML
programming language, which introduced this concept. Note that let in ML is recursive (so it corresponds
to Scheme’s letrec or local, and the rec we have studied in this class). In particular, ML treats let as a
primitive construct, rather than expanding it into an immediate function application as Scheme does (and as
we did with with in our interpreters).
The natural question is to wonder why we would have a rule that makes let-bound identifiers polymor-
phic, but not admit the same polymorphic power for lambda-bound identifiers. The reason goes back to our
initial approach to polymorphism, which was to substitute the body for the identifier. When we have access
to the body, we can successfully perform this substitution, and check for the absence of errors. (Later we
saw how type closures achieve the same effect while offering several advantages, but the principle remains
the same.)
The last example above shows the danger in generalizing the type of lambda-bound identifiers: without
knowing what they will actually receive as a value (which we cannot know until run-time), we cannot be
sure that they are in fact polymorphic. Because we have to decide at type-checking time whether or not to
treat an identifier polymorphically, we are forced to treat them monomorphically, and extend the privilege
of polymorphism only to let-bound identifiers. Knowing exactly which value will be substituted turns out
to be a gigantic advantage for the type system!
31.7 The Structure of ML Programs

While our type inference algorithm inferred types with type variables, we could not actually exploit this
power directly. We could use such a value several times in the same type contexts, and the same expression
elsewhere several times in a different type context, but not combine the two copies of the code through a
binding. Let-based polymorphism earned us this power of abstraction.
Let-based polymorphism depends fundamentally on having access to the bound value when checking
the scope of the binding. As a result, an ML program is typically written as a series of let expressions; the
ML evaluator interprets this as a sequence of nested lets. It similarly treats the initial environment as one
long sequence of lets so, for instance, if a programmer uses map in a top-level expression, the evaluator
effectively treats the use of map as being in the body of the binding of map. Therefore, the uses of map
benefit from the polymorphic nature of that function.
Exercise 31.7.1 What is the time and space complexity of the polymorphic type inference algorithm that
uses type closures?
31.8 Interaction with Effects

Suppose we add polymorphically-typed boxes to the language:
;; box : alpha −→ ref(alpha)
;; unbox : ref(alpha) −→ alpha
;; set-box! : ref(alpha) alpha −→ ref(alpha)
(We’re assuming here that set-box! returns the box as its result.) On their own, they look harmless.
Now consider the following program:
(let ([f (box (lambda (x) x))])
(begin
(set-box! f (lambda (y) (+ y 5)))
((unbox f ) true)))
When run, this program will yield a run-time error because y is bound to the value true, then used in an
addition. A sound type system should, therefore, flag this program as erroneous.
In fact, however, this program type checks without yielding an error. Notice that f has the closed type
ref (α → α) in the empty type environment. This type is renamed at each use, which means the function
applied to true has type (say) α 2 → α 2 , even though the value in the box has been re-bound to number →
number. In fact, this bug resulting from the unexpected interaction between state and polymorphism lay
dormant in ML for many years, and this brief program could crash the system.
What has happened here is that we’ve destoyed the semantics of boxes. The whole point of introducing
the box is to introduce sharing; the implementation of the type system has, however, lost that very sharing.
One solution to this problem would be to prohibit the use of boxes on the right-hand side of let (and
letrec) expressions, or at least not polymorphically generalize them. The problem is actually more general,
however: any potential effect (such as mutation, continuation capture, and so on) runs into similar prob-
lems. Studies of large bodies of ML code have shown that programmers don’t in fact need the power of
polymorphic generalization for these effects. Therefore, rather than create a vastly more complicated type
system, a simple, practical solution is to simply prohibit such effects in locations that the type system will
automatically treat as polymorphic.
Part XI
Programming by Searching
291
Chapter 32
Introduction
While we have already seen a wide variety of languages, in some ways they are all—Scheme, Haskell, and
so on—still quite conventional. But we have also seen what a language can do for us grow steadily: laziness
gave us the ability to express infinite computations; continuations gave us the ability to better structure
interactive applications; garbage collection gave us the ability to pretend we had an infinite amount of
memory; and type inference gave us the ability to drop type annotations. Progressing, we now find that we
don’t even have to view computations as functions. A program defines its data, and then defines relations
over those data; the act of computing an answer is replaced with the act of searching for an answer amongst
those data and their interrelationships. The canonical language with this power is Prolog.
Some people colloquiually describe Prolog’s implementation by saying it absolves programmers of hav-
ing to specify algorithms. In an abstract sense, this is true: the programmer only specifies the data, and
Prolog searches these exhaustively. In practice, however, this strategy fails for many reasons, mostly having
to do with efficiency. Nevertheless, it is helpful to understand Prolog’s behavior, because many sub-problems
can often benefit from this kind of search.
293
Chapter 33
Programming in Prolog
33.1 Example: Academic Family Trees

A Prolog program consists of a set of facts and a collection of rules. Given a program, a user can ask the
Prolog evaluator whether a particular fact is true or not. The query may be true atomically (because it’s in
the set of facts); if it’s not, the evaluator needs to apply the rules to determine truth (and, if no collection of
rules does the trick, then the query is judged false). With this little, Prolog accomplishes a lot!
Let’s plunge into a Prolog program. It’s inevitable, in the course of studying Prolog, to encounter a ge-
nealogy example. We’ll look at the genealogy of a particular, remarkable mathematician, where parenthood
is determined by PhD advisors.1
We’ll first list a few facts:
advisor(barwise,feferman).
advisor(feferman,tarski).
advisor(tarski,lesniewski).
advisor(lesniewski,twardowski).
advisor(twardowski,brentano).
advisor(brentano,clemens).
All facts are described by name of a relation (here, advisor) followed by a tuple of values in the relation.
In this case, we will assume that the person in the first position was advised by the person in the second posi-
tion. Prolog does not ask us to declare relations formally before providing their contents; nor does it provide
a means (beyond comments) of formally describing the relationship. Therefore, Prolog programmers must
be careful to keep track of how to interpret each relation.
Facts relate constants. In the example above, barwise and feferman are both constants. A constant
is, in fact, just a relation of arity zero, but understanding this isn’t necessary for the rest of this material.
In this example, all relations (including constants) are named by lower-case initial letters. This is not a
coincidence; Prolog requires this. Accidentally writing Barwise instead of barwise would change the
meaning of code quite a bit, because an initial capital denotes a variable, which we will study soon. Just
keep in mind that the case of the initial letter matters.
1 Thanks to the Mathematics Genealogy Project at http://www.genealogy.ams.org/.
295
296 CHAPTER 33. PROGRAMMING IN PROLOG
Given this input, we can ask Prolog to confirm a few basic facts:
:- advisor(barwise,feferman).
yes
This asks Prolog at its prompt whether the constants barwise and feferman are in the advisor rela-
tionship. Prolog responds affirmatively. In contrast,
:- advisor(feferman,barwise).
no
Prolog responds negatively because it has no such fact in the database specified by the program.
So far, Prolog has done nothing interesting at all. But with rules, we can begin to explore this universe.
A standard genealogical question is whether one person is another’s ancestor. Let’s build up this rule
one step at a time:
ancestor(x,y):-advisor(x,y).
This says that y is x’s (academic) ancestor if y is x’s advisor. But this isn’t very interesting either: it just
sets up ancestor to be an alias for advisor. Just to be sure, however, let’s make sure Prolog recognizes
it as such:
:- ancestor(barwise,feferman).
no
What?!? Oh, that’s right: x and y are constants, not variables. This means Prolog currently knows how
to relate only the constants x and y, not the constants barwise and feferman, or indeed any other
constants. This isn’t what we mean at all! What we should have written is
ancestor(X,Y):-advisor(X,Y).
Now, sure enough,
yes
So far, so good. There is another way for one person to be another’s academic ancestor: by transitivity.
We can describe this verbally, but it’s at least as concise, and just as readable, to do so in Prolog:2
ancestor(X,Y):-
advisor(X,Z),
ancestor(Z,Y).
Read the ‘,’ as “and”, while the multiple definitions for ancestor combine with “or” (i.e., each represents
a valid way to be an ancestor, so to be in the ancestor relation it’s enough to satisfy one rule or the other).
All Prolog rules are written in this “or of and’s” form (Disjunctive Normal Form). Notice the use of Z twice
on the right-hand side. This is intentional: this is what captures the fact that the same person must be both
the immediate advisor and herself a descendant.
Armed with this extended rule, we can ask more interesting queries of Prolog:
2 (Careful with those capital letters!)
33.1. EXAMPLE: ACADEMIC FAMILY TREES 297
:- advisor(barwise,tarski).
yes
so we can confirm that Barwise was advised by a legend. But it’s always a good idea to write tests that
ensure we didn’t write a faulty rule that made the relation too large:
:- advisor(tarski,barwise).
no
By the way, here’s an easy kind of mistake to make in Prolog: suppose you write
advisor(tawrdowski,brentano).
instead of
advisor(twardowski,brentano).
Then you get
:- advisor(barwise,clemens).
no
Prolog doesn’t have any way of knowing about slight misspellings of Polish names. It accepts your facts
as truth; garbage in, garbage out. This is another important pitfall (along with capitalization and making a
relation too large) to keep in mind.
Now let’s expand the relation with a few more facts. Franz Brentano actually had two advisors, of whom
we’ve given credit to only one above. So we should add the fact
advisor(brentano,trendelenburg).
We can now ask Prolog
:- advisor(barwise,clemens).
yes
:- ancestor(barwise,trendelenburg).
yes
and it shows that the relationships Prolog tracks really are relations, not functions: the mapping truly can
be one-to-many. (We could already have guessed this from the rule for ancestor, where we provided
multiple ways of determining whether or not a pair of constants was in that relation.)
Now let’s add some more rules. The simplest is to ask questions in the other direction:
descendant(X,Y):-ancestor(Y,X).
As of now, each person has only one immediate descendant. But most of these people produced many
students. Tarski, one of the great logicians and mathematicians, not only generated a rich corpus of material,
but also trained a string of remarkable students. We already know of one, Feferman. Let’s add a few more:
advisee(tarski,montague).
advisee(tarski,mostowski).
advisee(tarski,robinson).
Now, clearly, there are more ways of being one’s descendant. There are two ways of fixing the descendant
relation. One, which involves a lot of work, is to add more facts and rules. A much easier fix is to note that
every advisee relationship subscribes a corresponding advisor relationship:
advisor(X,Y):-advisee(Y,X).
And sure enough,
:- descendant(clemens,montague).
yes
:- descendant(trendelenburg,montague).
yes
:- descendant(feferman,montague).
no
:- descendant(barwise,montague).
no
We haven’t at all explained how Prolog evaluates, and that’s largely because it seems so very intuitive
(though once we have multiple clauses, as in descendant, it may be a little less than obvious). But then,
we also haven’t seen Prolog do anything truly superlative. Let’s explore some more.
Let’s first assume we’ve removed Trendelenburg from the database, so Brentano has only one advisor.
(We can do this in a source file by using C-style comments, delimiting text in /* and */.) Then let’s ask
Prolog the following query:
:- ancestor(barwise,X).
What does this mean? We know all the parts: advisor is a relation we’ve defined (by both facts and rules);
barwise is a constant; and X is a variable. We should interpret this as a query, asking Prolog whether there
is a value for X that would satisfy (make true) this query. In fact, we know there is (clemens). But Prolog’s
response is worth studying. This particular Prolog system3 prints
SOLUTION:
X=feferman
So not only did Prolog establish that the query was valid, it also found a solution for X! Now this isn’t the
solution we expected above, but if you think about it for a moment, it’s clear that the query has multiple
solutions, and Prolog has picked one of them. In fact, at the bottom of the window (in this interface), Prolog
says Press cancel to stop, or continue for more solutions. Clicking on the Continue button provides one
more solution, then another, then another, and so on until there are no more, so the final output is
3 Trinc-Prolog R3. In many textual Prolog systems, it’s conventional to print a caret to indicate that another solution is available.
The user types a semi-colon to ask Prolog to present it.

33.1. EXAMPLE: ACADEMIC FAMILY TREES 299
SOLUTION:
X=feferman
SOLUTION:
X=tarski
SOLUTION:
X=lesniewski
SOLUTION:
X=twardowski
SOLUTION:
X=brentano
SOLUTION:
X=clemens
no
Wow! Prolog actually “filled in the blank”. In fact, if we put Trendelenburg back into the picture, Prolog
prints one more solution:
SOLUTION:
X=trendelenburg
We can ask a similar query with the variable in the first position instead:
:- ancestor(X,clemens).
SOLUTION:
X=brentano
SOLUTION:
X=barwise
SOLUTION:
X=feferman
SOLUTION:
X=tarski
SOLUTION:
X=lesniewski
SOLUTION:
X=twardowski
This shows that Prolog isn’t just working as a functional program might, where the last position in the
relation is like a “return” location. Prolog really doesn’t discriminate between different positions where you
might put variables.
Maybe this isn’t so surprising. After all, Prolog is merely listing the same chain of relationship that we
entered as facts at the top of the program. Actually, this isn’t quite true: it had to apply the transitive rule of
ancestor to find all the solutions (and these are indeed all of them). But perhaps a more impressive test
would be to ask a query that runs counter to the facts we entered. For this, we should employ the advisee
relation.
:- descendant(tarski,X).
SOLUTION:
X=feferman
SOLUTION:
X=montague
SOLUTION:
X=mostowski
SOLUTION:
X=robinson
SOLUTION:
X=barwise
Sure enough, Prolog produces the entire part of Alfred Tarski’s family tree that we’ve taught it. Notice that
to get to Barwise it had to recur through Feferman using advisor, and to get to Robinson it had to employ
advisee.
This is pretty impressive already. But we can take this a step further. Why stop with only one variable?
We could in fact ask
:- ancestor(X,Y).
In response, Prolog will actually compute all pairs in the ancestor relation and present them sequentially:
SOLUTION:
X=barwise
Y=feferman
SOLUTION:
X=feferman
Y=tarski
SOLUTION:
X=tarski
Y=lesniewski
SOLUTION:
X=lesniewski
Y=twardowski
and so on and on.

Now let’s explore another relation: academic siblings. We can define a sibling pretty easily: they must
have a common advisor.
sibling(X,Y):-
advisor(X,Z),
advisor(Y,Z).
We can either ask Prolog to confirm relationships (and just to be sure, try them both ways):
33.2. INTERMISSION 301
:- sibling(robinson,montague).
yes
:- sibling(montague,robinson).
yes
or to generate their instances:
:- sibling(robinson,X).
SOLUTION:
X=feferman
SOLUTION:
X=montague
SOLUTION:
X=mostowski
SOLUTION:
X=robinson
How’s that? When Robinson comes along to find out who her academic siblings are, she finds. . . herself!
It’s not very surprising that we got this output. What we meant to say was that different people, X and Y,
are academic siblings if. . . and so on. While we may have mentally made a note that we expect X and Y to be
different people, we didn’t tell Prolog that. And indeed, because Prolog programs can be “run backwards”,
it’s dangerous to not encode such assumptions. Making this assumption explicit is quite easy:4
sibling(X,Y):-
advisor(X,Z),
advisor(Y,Z),
X \== Y.
Now, sure enough, we get the right number of siblings.
33.2 Intermission
At this point, we have seen most of the elements of (the core of) Prolog. We’ve seen fact declarations
and the expression of rules over them to create extended relations. We’ve also seen that Prolog evaluates
programs as simple rule-lookups or as queries (where the former are a special case of the latter). We’ve seen
Prolog’s variables, known as logic variables, which can take on multiple values over time as Prolog boldly
and tirelessly seeks out new solutions until it has exhausted the space of possibilities. And finally, related to
this last step, Prolog backtracks as necessary to find solutions, in accordance with the non-determinism of
the rules.
4 For a very subtle reason, we cannot move the last line earlier. We will understand why better once we’ve implemented Prolog.
33.3 Example: Encoding Type Judgments

Let’s look at another use for Prolog: to encode type judgments. Recall that we had rules of the form
Γè : τ
where some were axioms and the others were conditionally-defined judgments. The former we will turn into
facts, the latter into rules.
First, we must determine a representation for abstract syntax in Prolog. We don’t want to deal with
parsing, and we don’t actually distinguish between individual values of a type, so we’ll assume constants
have been turned into an abstract node that hides the actual value. Thus, we use the constant numConst to
represent all syntactically numeric expressions (i.e., those abstract syntax terms of type numE), boolConst
to represent true and false, and so on.
Given this, we will define a three-place relation, type. The first place will be the type environment,
represented as a list; the second will be the expression; and the third the type of the expression. (When
writing this as a function in a traditional language, we might define it as a two-argument function that
computes the expression’s type. But because Prolog can “run backward”, it doesn’t have a distinguished
“return”. Instead, what we normally think of as the “result” is just another tuple in the relation.) Our axioms
therefore become:
type(_,numConst,num).
type(_,boolConst,bool).
The _ represents that we don’t care what goes in that position. (We could as well have used a fresh logic
variable, but the underscore makes our intent clearer.) That is, no matter what the type environment, numeric
constants will always have type num.
The easiest judgment to tackle is probably that for conditional. It translates very naturally into:
type(TEnv,if(Test,Then,Else),Tau) :-
type(TEnv,Test,bool),
type(TEnv,Then,Tau),
type(TEnv,Else,Tau).
Pay close attention to lower- and upper-case initials! Both type and if are in lower-case: the former
represents the type relation, while the latter is the abstract syntax term’s constructor (the choice of name is
arbitrary). Everything else is a type variable. (Notice, by the way, that Prolog performs pattern-matching on
its input, just as we saw for Haskell.)
Given these two facts and one rule for type, we can ask Prolog to type-check some programs (where
[] denotes the empty list):
:- type([],boolConst,bool).
yes
:- type([],if(boolConst,numConst,numConst),num).
yes
:- type([],if(boolConst,numConst,boolConst),num).
no
33.3. EXAMPLE: ENCODING TYPE JUDGMENTS 303
The implementation of this rule in your type checkers reflected exactly the semantics we gave: if the three
conditions in the antecedent were met, then the consequent holds. In contrast, because Prolog lets us query
relations in any way we please, we can instead use the same implementation to ask what the type of an
expression is (i.e., make Prolog perform type inference):
:- type([],boolConst,T).
T=bool
no
:- type([],if(boolConst,numConst,numConst),T).
T=num
no
:- type([],if(boolConst,numConst,boolConst),T).
no
It should be no surprise that Prolog “inferred” a type in the first case, since the use precisely matches the
axiom/fact. In the second case, however, Prolog used the rule for conditionals to determine solutions to the
type of the first expression and matched these against those for the second, finding the only result. In the
third case, since the program does not have a type, Prolog fails to find any solutions.
We can now turn evaluation around by asking Prolog strage questions, such as “What expression have
type num?”
:- type([],T,num).
Amazingly enough, Prolog responds with:5
SOLUTION:
T=numConst
SOLUTION:
T=if(boolConst, numConst, numConst)
SOLUTION:
T=if(boolConst, numConst, if(boolConst, numConst, numConst))
SOLUTION:
T=if(boolConst, numConst,
if(boolConst, numConst,
if(boolConst, numConst, numConst)))
The output here actually gives us a glimpse into the search order being employed by this implementation
(notice that it depth-first expands the else-clause of the conditional).
Next let’s deal with identifiers. We’ve said that the type environment is a list; we’ll use a two-place
bind relation to track what type each identifier is bound to.6 To look up the type of an identifier in the type
environment, therefore:
5 The output has been indented for readability.
6 We’ll make the simplifying assumption that all bound identifiers in the program are consistently renamed to be distinct.
type([bind(V,T)|_],var(V),T).
type([bind(_,_)|TEnvRest],var(V),T):-
type(TEnvRest,var(V),T).
A quick test:
:- type([bind(w,bool),bind(v,num)],var(v),T).
T=num
Next we’ll specify the rule for functions:
type(TEnv,fun(Var,Body),arrow(T1,T2)) :-
type([bind(Var,T1)|TEnv],Body,T2).
Testing this:
:- type([],fun(x,if(var(x),numConst,boolConst)),T).
no
:- type([],fun(x,if(var(x),numConst,numConst)),T).
T=arrow(bool,num)
Notice that in the second example, Prolog has determined that the bound identifier must be a boolean, since
it’s used in the test expression of a conditional.
Finally, the rule for applications holds no surprises:
type(TEnv,app(Fun,Arg),T2) :-
type(TEnv,Fun,arrow(T1,T2)),
type(TEnv,Arg,T1).
Running it:
:- type([],
app(fun(x,if(var(x),
numConst,
numConst)),
boolConst),
T).
T=num
Now let’s try some more interesting functions:
:- type([],fun(x,var(x)),T).
T=arrow(__2823020, __2823020)
This is Prolog’s way of saying that parts of the answer are indeterminate, i.e., there are no constraints on it.
In short, Prolog is inferring parameterized types!
33.4. FINAL CREDITS 305
:- type([],
app(fun(id,
if(app(var(id),boolConst),
app(var(id),boolConst),
app(var(id),boolConst))),
fun(x,var(x))),
T).
T=bool
:- type([],
app(fun(id,
if(app(var(id),boolConst),
app(var(id),numConst),
app(var(id),numConst))),
fun(x,var(x))),
T).
no
Finally, we have to try:
:- type([], fun(x, app(var(x), var(x))), num).

no
:- type([], fun(x, app(var(x), var(x))), T).

T=arrow(arrow(arrow...
Exercise 33.3.1 Are Prolog’s types truly polymorphic? Do they automatically exhibit let-based polymor-
phism (Section 31)? Write appropriate test expressions and present Prolog’s output to justify your case.
33.4 Final Credits

We’ve now seen even more of Prolog. We’ve encountered the “don’t care” notation. Prolog computes the
most general response it can, so if there are no constraints on some part of the answer, it leaves them unde-
fined (using the same symbol to show sharing constraints, as in the inferred type of the identity function).
Prolog will match patterns as deep as they are nested, and programmers can use the same variable twice
in a rule to indicate that they intend for the values of both to be the same. (Having already seen this with
genealogical trees, we made much more extensive use of it to encode type judgments, mimicking the use of
meta-variables when we wrote the judgments on paper.)
Putting together the pieces, we found that Prolog was a very convenient encoding of the rules of a type
checker. Indeed, for free, we were able to turn our type checker into a type inference engine. Thinking about
how we implemented type inference manually may give us some clues as to how to implement Prolog!
Chapter 34
Implementing Prolog
A Prolog program is a collection of facts and rules. Against these, we ask Prolog a query, also known as
providing a goal that it must satisfy. If a goal contains logic variables, Prolog must find assignments (if
necessary) for those variables consistent with the facts and rules.
It is easy to see that in the typical case, a collection of Prolog facts subscribes a flat relational database.
Searching such a database is not especially hard (there are obviously challenges to doing so efficiently, but
these concerns are handled well by database management systems; they’re outside the scope of our study).
Prolog evaluation becomes interesting with the introduction of rules.
34.1 Implementation
As we have seen, Prolog rules are in disjunctive normal form, i.e., an “or of and’s”. Such formulas naturally
subscribe a tree, where the individual nodes in the tree are labeled by a Boolean connective. Given a goal,
Prolog can construct a disjunctive tree of the facts and rules that match the head term in the goal. For
instance, if the query is
Prolog can construct an or-labeled tree of all the ways of being an ancestor (there are two ways, correspond-
ing to two rules and no facts). Prolog then explores each node in turn: each of these is, reasonably enough,
called a sub-goal. If the sub-goal is a fact, there is nothing further to explore; otherwise it can expand into
further sub-goals that must, in turn, each be satisfied. In short, Prolog naturally constructs and then explores
a classical search tree. Thus we say that programming in Prolog is “programming by searching”. More
precisely, Prolog’s search is non-deterministic: it permits rules to have multiple definitions and searches
them all, looking for any—and, indeed, every—way to satisfy it.
Armed with this background, we can now delve into the pieces. To clarify the execution of Prolog, we
must provide an account of (at least) the following:
• How does Prolog search the space of relations?
• What exactly constitutes satisfaction?
307
308 CHAPTER 34. IMPLEMENTING PROLOG
• What are logic variables? How do logic variables take on a value, and how does this value change
over the course of a search?
• What causes computation to terminate?
34.1.1 Searching
In what order should Prolog search the tree? Both of the canonical answers make sense:
breadth-first search This has the benefit of never getting “stuck” in an infinite expansion exploring one
sub-goal while a different one is satisfiable and might have led to an answer. Unfortunately, it is also
rather expensive to maintain the queue necessary for breadth-first search.
depth-first search This has the disadvantage that it may get “stuck” exploring one non-satisfying disjunct
while another can satisfy the query. It has the advantage of executing relatively efficiency because it
corresponds well to a stack-based execution, which modern systems are tuned well to implement.
Though breadth-first search will therefore produce satisfying answers in situations when depth-first search
will not, Prolog chooses the latter, sacrificing some purity at the altar of efficiency.
Exercise 34.1.1 Write a Prolog program that would yield responses under a breadth-first order but that
fails to terminate in a traditional implementation.
Hint: Even when choosing depth-first search, a Prolog system has the choice of choosing sub-goals left-
to-right, right-to-left, etc. You can glean insight into the order chosen by the Prolog implementation you’re
using by examining the order in which it prints output for queries that have many satisfying answers.
Exercise 34.1.2 Rewrite your example from the previous exercise to retain as much of its structure as pos-
sible, but to provide as many answers as possible in the Prolog system you’re using. Discuss what you had
to change and its implications for programming by searching.
34.1.2 Satisfaction
The example we saw in Section 33.1 helps illustrate numerous concepts:
sibling(X,Y):-
advisor(X,Z),
advisor(Y,Z),
X \== Y.
Given the query
sibling(robinson,feferman}.
34.1. IMPLEMENTATION 309
we can pretend Prolog “binds” X to robinson and Y to feferman; this therefore appears to make two
“recursive calls” to advisor. However, these cannot be function invocations in the traditional sense,
because Z is not bound. In fact, Z is—as we have said—a logic variable.
Prolog searches every pair in the advisor relation to find values that it can associate with Z. None
of the facts about advisor yield a match, but we also have a rule relating advisor to advisee. This
spawns a fresh goal, trying to satisfy advisee(Z,robinson). This eventually finds a satisfying match,
with Z bound to tarski. Now that Prolog has a value of Z, it can proceed with the second clause (recall
that , should be read as “and”, so there is no point proceeding until the first clause has been satisfied).
Now we understand why logic variables are called variables: they change their value. Initially Z had no
value at all, whereas now it holds the value tarski. To satisfy the next goal, Prolog must therefore attempt
to satisfy advisor(feferman,tarski). This is, of course, a fact about advisor, so Prolog imme-
diately satisfies it. Now it proceeds to successfully discharge the inequality test, and return an affirmative
answer—that is, Robinson and Feferman are indeed academic siblings.
This account does not yet explain two things:
• How many times can a logic variable vary? Does it change its value just once (going from unbound
to bound), or does it change more than once?
• Why did we have to place the inequality test at the end of the rule?
To study these questions, let us consider the more general query,
sibling(mostowski,B).
Initially, X is bound to mostowski and Y to the logic variable B. This yields the sub-goal
advisor(mostowski,Z).
which eventually binds Z to tarski (by the reasoning we saw earlier). The first clause having been
satisfied, we now effectively evaluate the sub-goal advisor(B,tarski) (replacing Y with B).
What is the value of B in this sub-goal? Actually, it is not meaningful to ask for the value, because there
are several. First, Prolog assigns B to feferman, because this is the first matching rule in the program.
Evaluation now proceeds as before, with the inequality test succeeding. Now Prolog can not only report
success, it can also report a value for B, namely feferman. This is indeed what Prolog prints.
In fact, Prolog has done more than just assign a value to B: it has also recorded the last successful
satisfying clause. When the user now asks for an additional output, the program resumes execution from
that clause onward. That is, Prolog has the ability to store the state of a computation (specifically, of a
search), and restore it on demand. This sounds remarkably similar to the behavior of a continuation.
When the user asks for resumption, Prolog tries the same sub-goal again. It now finds a different student
who was advised by Tarski, namely montague. Again, we find that the computation can successfully com-
plete, so the user sees montague presented as a binding for B. Observe that in this process, the same logic
variable B has gone from having no value at all to being bound to tarski to being bound to montague.
As the user explores more of the computation tree, the (same) logic variable keeps changing its binding.
Eventually, after more prompting from the user, Prolog finds that mostowski is also a satisfying
assignment for the sub-goal. Program control now returns to the conjunction of clauses in the rule for
sibling to evaluate the third clause. When Prolog now performs the comparison, we find that the values
bound to the two logic variables are the same so the comparison fails, thereby causing the clause to fail.
This explains why we must perform the check at the end. Until the end, we cannot be sure both the logic
variables, X and Y, will actually be bound, because we cannot know for which parameter—or both—the user
will supply a logic variable. Only at the end do they have values. Thus, even though the language appears
“logical”, there is a clear imperative effect in the binding of values to logic variables, so the programmer
must take care to sequence correctly (just as with variables in less peculiar languages).
Having failed the comparison, what does Prolog do next? It does not return to the top-level reporting
failure. Instead, it continues searching by looking for more assignments to the logic variable. It finds such an
assignment: robinson. Therefore B is bound to robinson, an assignment that satisfies the comparison,
resulting in the next displayed bound value. Further continuing the computation yields no more successful
bindings, thereby causing the program to halt with no more successful satisfying assignments.
For Prolog to resume the computation when a clause (specifically, the inequality test) fails, it must have
access to the continuation of the previous satisfying clause. Each clause must also provide the continuation
of its own evaluation, which a failing computation can use to resume computation. That means every Prolog
expression must consume a continuation and return one. It consumes a continuation that dictates where
to resume in case of failure, and it returns its continuation to indicate success (to be used as some other
computation’s resumption point in case of failure).
Computation terminates when no more terms satisfy.
34.1.3 Matching with Logic Variables

Our explanation has focused on matching logic variables against atomic terms, which looks like a simple
assignment process. As we saw in Section 33.3, however, we may need to match against terms with a
complex structure (such as the environment). Given a goal such as
type([bind(V,T)|_],var(V),T).
and a rule
type([bind(x,num)],var(x),num).
how do we perform the matching? Easy—we use unification!1 Recall that unification was a process for
matching terms with constructors and variables, and it computed a substitution as a result (Section 30.5.1).
This substitution is precisely the assignment of values to logic variables. The Prolog notation _ is simply
shorthand for introducing a logic variable with a completely fresh name. Assigning logic variables to atomic
values such as feferman is simply a trivial instance of unification, and terms such as bind and var are
simply more term constructors.
The only difference in Prolog’s unification is its interaction with continuations. When unification fails,
Prolog must undo the assignment of values to logic variables. When it succeeds Prolog leaves the logic vari-
ables assigned, but these assignments must be undone before commencing to compute the next satisfying
assignment for the same variables. Therefore, Prolog undoes the bindings before invoking the failure con-
tinuation, and makes the returned success continuation also undo the bindings upon invocation (i.e., when
the computation is ready to seek another assignment).
1 In short, it isn’t accidental that Prolog automatically computed precisely the same types as we studied in Section 33.3.
34.2. SUBTLETIES AND COMPROMISES 311
34.2 Subtleties and Compromises

This account of Prolog’s implementation hides several subtleties and extensions.
First, as we recall from the discussion of unification, it was possible to have underconstrained variables.
What happens to these in Prolog? In fact, we saw precisely these in action, when we asked Prolog to infer
(for instance) the type of the identity function. Before displaying them, Prolog consistently renames each
unbound logic variable (e.g., __2823020) so we can distinguish them from one another.
Exercise 34.2.1 Write a non-trivial Prolog program whose output contains two distinct unbound logic vari-
ables.
Second, when discussing unification, we also mentioned the occurs check. While it would seem reason-
able to use the occurs check whenever possible, it proves to be very computationally expensive. As a result,
most Prolog systems forego the occurs check by default, usually providing it as a configurable option. This
is another common compromise in Prolog systems.
Third, Prolog users sometimes want to better control either the performance or even the meaning of
a computation by terminating a search at some point. Prolog provides an operator written !, pronounced
“cut”, for indicating this. Cuts are another controversial feature of Prolog because they sacrifice linguistic
purity for pragmatic reasons.
34.3 Future Directions

Programming by searching is arguably still in its infancy. While the influence of Prolog has waxed and
waned, its ideas live on in many interesting control operators, and the notion of specifying connections
between data, and leaving it to the system to search for the answer, remains very attractive.
Because the system performs the search, people sometimes colloquially say that in Prolog, the user does
not specify the “algorithm”. This is a very misleading remark. Prolog programmers exploit the built-in
exhaustive search algorithm. A user might therefore want a more sophisticated search, for instance one
that exploited knowledge about numbers. This is an active area of research that has grown into the field of
constraint programming.
There are also ways to ameliorate the ill-effects of the depth-first search strategy. In many cases, the
sub-goal in the depth-first search repeats. In such cases, in the absence of other side-effects, repeating the
sub-goal is not going to generate a different answer. In such cases, an implementation can use memoization
to check whether it has (in the appropriate context) issued a sub-goal query before and, if it has, terminate
the search as unsatisfiable (i.e., a failure). In such cases, the programmer gets the performance benefit of
depth-first search without the danger of non-termination. Prolog systems call this memoization procedure
tabling. Because it incurs some cost, the programmer is asked to stipulate which rules must be tabled.
Tabled Prolog extends the expressive elegance of Prolog to many interesting cases such as certain kinds of
fixed-point computations.
Part XII
Domain-Specific Languages and

Metaprogramming
313
Chapter 35
Domain-Specific Languages
35.1 Language Design Variables

Programming languages differ in numerous ways:
1. Each uses rather different notations for writing down programs. As we’ve observed, however, syntax
is only partially interesting. (This is, however, less true of languages that are trying to mirror the
notation of a particular domain.)
2. Control constructs: for instance, early languages didn’t even support recursion, while most modern
languages still don’t have continuations.
3. The kinds of data they support. Indeed, sophisticated languages like Scheme blur the distinction
between control and data by making fragments of control into data values (such as first-class functions
and continuations).
4. The means of organizing programs: do they have functions, modules, classes, . . . ?
5. Automation such as memory management, run-time safety checks, and so on.
Each of these items suggests natural questions to ask when you design your own languages in particular
domains.
35.2 Languages as Abstractions

Languages are abstractions: ways of seeing or organizing the world according to certain patterns, so that
a task becomes easier to carry out. More concretely, think about a loop in Java. When you write a loop,
you expect the machine to carry out certain tasks for you automatically: testing the termination condition,
running the loop code if the test passes, exiting the loop if it doesn’t, etc. The loop is an abstraction: a
reusable pattern where the language executes part of the pattern automatically, and you supply the parts that
are different. You could write down all of those steps manually, but then your program would be longer,
harder to read, and more painful to write, debug and maintain. Scheme’s map and filter are also abstractions,
315
316 CHAPTER 35. DOMAIN-SPECIFIC LANGUAGES
which differ from Java’s loops in one significant way: you can define Scheme’s loops as abstractions in user
programs.
35.3 Domain-Specific Languages

Based on the above description, it becomes clear that some domains may be served better by programming
in a language specialized to that domain. While we are familiar with such languages (often bundled with
software packages that blur the boundary between the package and the language) such as Mathematica
and Matlab, this principle is not new. Indeed, study the names of four of the oldest popular programming
languages, and you spot a pattern:
Fortran Stands for “formula translator”.
Algol An “algorithmic language”.
COBOL An abbreviation for “COmmon Business-Oriented Language”.
LISP Short for a “list-processing” language.
Notice the heavy emphasis on very concrete domains (or, in the case of LISP, of a language construct)?
Indeed, it was not until the late 1960s and 1970s that programming languages really became liberated from
their domains, and the era of general-purpose languages (GPL) began. Now that we know so much about
the principles of such languages (as we’ve been seeing all semester long), it is not surprising that language
designers are shifting their sights back to particular domains.
Indeed, I maintain that designing GPLs has become such a specialized task—well, at least designing
good GPLs, without making too many mistakes along the way—that most lay efforts are fraught with peril.
In contrast, most people entering the programming workforce are going to find a need to build languages
specific to the domains they find themselves working in, be they biology, finance or the visual arts. Indeed,
I expect many of you will build one or more “little languages” in your careers.
Before you rush out to design a domain-specific language (DSL), however, you need to understand some
principles that govern their design. Here is my attempt to describe them. These are somewhat abstract; they
will become clearer as we study the example that follows in more detail.
First and foremost—define the domain! If your audience doesn’t understand what the domain is, or (this
is subtly different) why programming for this domain is difficult, they’re not going to pay attention to your
language.
Justify why your language should exist in terms of the current linguistic terrain. In particular, be sure to
explain why your language is better than simply using the most expressive GPLs around. (Small improve-
ments are insufficient, compared with the odds that the considerably greater resources that are probably
going into language implementation, library support, documentation, tutorials and so on for that GPL com-
pared with your language.) In short, be very clear on what your DSL will do that is very difficult in GPLs.
These reasons usually take on one or more of the following forms:
• Notational convenience, usually by providing a syntax that is close to established norms in the domain
but far removed from the syntax of GPLs. (But before you get too wrapped up in fancy visual notations,
35.3. DOMAIN-SPECIFIC LANGUAGES 317
keep in mind that programs are not only written but also edited; how good is your editor compared
with vi or Emacs?)
• Much better performance because the DSL implementation knows something about the domain. For
instance, some toolkits take limited kinds of programs but will, in return, automatically compute
the derivative or integral of a function—a very useful activity in many kinds of high-performance
scientific computing.
• A non-standard semantics: for instance, when neither eager nor lazy evaluation is appropriate.
There are generally two kinds of DSLs, which I refer to as “enveloping” and “embedded”. Enveloping
languages are those that try to control other programs, treating them as components. Good examples are
shell languages, and early uses of languages like Perl.
Enveloping languages work very well when used for simple tasks: imagine the complexity of spawning
processes and chaining ports compared with writing a simple shell directive like ls -l | sort | uniq.
However, they must provide enough abstraction capabilities to express a wide variety of controls, which in
turn brings data structures through the back door (since a language with just functions but without, say, lists
and queues, requires unreasonable encodings through the lambda calculus). Indeed, invariably program-
mers will want mapping and filtering constructs. The net result is that such languages often begin simple,
but grow in an unwieldy way (responding to localized demands rather than proactively conducting global
analysis).
One way to improve the power of an enveloping language without trying to grow it in an ad hoc way
is to embed another language inside it. That is, the enveloping language provides basic functionality, but
when you want something more powerful, you can escape to a more complete (or another domain-specific)
language. For instance, the language of Makefiles has this property: the Makefile language has very limited
power (mainly, the ability to determine whether files are up-to-date and, if not, run some set of commands),
and purposely does not try to grow much richer (though some variants of make do try). Instead, the actual
commands can be written in any language, typically Unix shell, so the make command only needs to know
how to invoke the command language; it does not itself need to implement that language.
The other kinds of languages are embedded in an application, and expose part of the application’s func-
tionality to a programmer who wants to customize it. A canonical example is Emacs Lisp: Emacs functions
as a stand-alone application without it, but it exposes some (most) of its state through Emacs Lisp, so a
programmer can customize the editor in impressive ways. Another example may be the command language
of the sendmail utility, which lets a programmer describe rewriting rules and custom mail handlers.
Any time one language is embedded inside another language (as opposed to an application), there are
some problems with this seemingly happy symbiosis:
1. The plainest, but often most vexing, is syntactic. Languages that have different syntaxes often don’t
nest within one another very nicely (imagine embedding an infix language inside Scheme, or XML
within Java). While the enveloping language may have been defined to have a simple syntax, the act
of escaping into another language can significantly complicate parsing.
2. Can the embedded language access values from the language that encloses it? For example, if you
embed an XML path language inside Java, can the embedded language access Java variables? And
318 CHAPTER 35. DOMAIN-SPECIFIC LANGUAGES
even if it could, what would that mean if the languages treat the same kinds of values very differently?
(For instance, if you embed an eager language inside a lazy one, what are the strictness points?)
3. Often, the DSL is able to make guarantees of performance only because it restricts its language in some
significant way. (One interesting example we have seen is the simply-typed lambda calculus which,
by imposing the restriction of annotations in its type language, is able to deliver unto us the promise
of termination.) If the DSL embeds some other language, then the analysis may become impossible,
because the analyzer doesn’t understand the embedded language. In particular, the guarantees may
not longer hold!
In general, as a DSL developer, be sure to map out a growth route. Anticipate growth and have a concrete
plan for how you will handle it. No DSL designer ever went wrong predicting that her programmers might
someday want (say) closures, and many a designer did go wrong by being sure his programmers wouldn’t.
Don’t fall for this same trap. At the very least, think about all the features you have seen in this course and
have good reasons for rejecting them.
You should, of course, have thought at the very outset about the relationship between your DSL and
GPL s. It doesn’t hurt for you to think about it again. Will your language grow into a GPL ? And if so, would
you be better off leveraging the GPL by just turning your language into a library? Some languages even
come with convenient ways of creating little extension languages (as we will see shortly), which has the
benefit that you can re-use all the effort already being poured into the GPL.
In short, the single most important concept to understand about your DSL is its negative space. Language
designers, not surprisingly, invariably have a tendency to think mostly about what is there. But when you’re
defining a DSL remember that perhaps the most important part of it is what isn’t there. Having a clear def-
inition of your language’s negative space will help you with the design; indeed, it is virtually a prerequisite
for the design process. It’s usually a lot easier to argue about what shouldn’t (and should) be in the negative
space than to contemplate what goes in. And to someone studying your language for the first time, a clear
definition of the negative space will greatly help understand your rationale for building it, and perhaps even
how you built it—all of which is very helpful for deciding whether or not one finds this the right language
for the task, both now and in the future.
Chapter 36
Macros as Compilers
For these notes, please use the P RETTY B IG language level.
36.1 Language Reuse

We have so far implemented languages as interpreters. In the real world, however, programming languages
are defined not only by their implementation but also by their toolkit: think of the times you’ve disliked
programming in a language because you didn’t like the default editor or the debugger or the lack of a
debugger or . . . . Therefore, when we set out to implement a fresh language implementation, we run the risk
that we’ll upset our users if we don’t provide all the programming tools they’re already accustomed to.
One way around this is to not create an entire implementation from scratch. Instead, we could just
compile the new language into an existing language. If we do that, we can be fairly sure of reusing most of
the tools built for the existing language. There is one problem, which is that feedback such as error messages
may not make too much sense to the programmer (since she is expecting messages in terms of the constructs
of the DSL, while the messages are in terms of the constructs of the target language). This is a real concern,
but it is ameliorated some by the tools we will use.
Many languages provide a syntactic preprocessor that translates terms before handing them off to the
evaluator. In languages like C and Scheme they’re called macros, while in C++ they’re called templates.
We will now study the Scheme macro system in some depth. By default, the Scheme macro system permits
programmers to add constructs to Scheme, thereby effectively providing a compiler from Scheme+ (the
extended Scheme language) to Scheme itself.
36.1.1 Example: Measuring Time

Suppose we want to add a construct to the language that measures the time elapsed while evaluating an
expression. That is, we want (my-time e) which returns the time it took (in milliseconds, say) to evaluate
e. (The actual time command in Scheme also returns the value, but this version suffices for now. We’ll use
my-time for our attempts to avoid clashing with the vesion built-in.)
This is easy; here’s the code:
(define (my-time e)
319
320 CHAPTER 36. MACROS AS COMPILERS
(let ([begin-time (current-milliseconds)])

(begin
e
(- (current-milliseconds) begin-time))))
Let’s test it:
> (my-time (+ 1 2))
0
Good; that’s about what we’d expect. Even for slightly more computationally expensive expressions, we get
> (my-time (expt 2 1000))
0
Well, that’s because DrScheme is really fast, see. How about:
> (my-time (expt 2 10000))
0
Hmm. Zero milliseconds? Maybe not. So let’s try
> (my-time (expt 2 1000000))
0
This time DrScheme noticeably gives pause—we can tell from a wristwatch—so something is afoot.
The problem is that we defined my-time to be a procedure, and Scheme is an eager language. Therefore,
the entire expression reduced to a value before the body of my-time began to evaluate. As a result, the
difference in time was always going to be a constant. On different machines we might get different values,
but the value isn’t going to change, no matter what the expression!
How do we define my-time? There are three options.
First would be to introduce lazy evaluation into the language. This may be tempting, but it’s going to
make a mess overall, because it’d be impossible to determine when an expression is going to reduce, and an
expression that has already been reduced to a value may need to have not been reduced later. This is not a
viable solution.
The second is to make my-time take a thunk (recall: a procedure of no arguments). That is, we would
have
(define (my-time e-thunk)
(begin
(e-thunk)
(- (current-milliseconds) begin-time))))
so that
> (my-time (lambda () (expt 2 10000)))
0
36.1. LANGUAGE REUSE 321
60
2023
This may be sufficient, but it’s certainly not satisfactory: we’ve introduced an unnecessary syntactic pattern
into the code for which we have no explanation other than that’s just what the language demands. This is
not an acceptable abstraction.
Finally, another is to accomplish the effect of textual substitution by using. . . textual substitution. In
Scheme, we can instead write
(define-syntax my-time
(syntax-rules ()
[(my-time e)
(begin
e
(- (current-milliseconds) begin-time)))]))
When we test this, we find
> (my-time (expt 2 1000))
0
Hmm! But ever hopeful:
> (my-time (expt 2 10000))
10
> (my-time (expt 2 100000))
70
> (my-time (expt 2 1000000))
2053
which is what we expect.
How does this version of my-time work? The Scheme macro system trawls the program source and
gathers all the syntax definitions. It then substitutes all the uses of these syntax definitions with the bodies,
where each syntax definition is defined by pattern-matching (we’ll see several more examples). Only after
finishing all the substitution does it hand the program to the Scheme evaluator, which therefore doesn’t need
to know anything about the syntax definitions. That is, given the above syntax definition and the program
(my-time (expt 2 10000)), the program that the Scheme evaluator actually sees is
(begin
(expt 2 10000)
(- (current-milliseconds) begin-time)))
This is the right-hand-side of the first (and only) clause in the list of rules, except e has been substituted with
the exponentiation expression. This is now an ordinary Scheme expression that the evaluator can reduce to a
value. Notice that the current time is now measured before and after the expression evaluates, thus ensuring
that we do in fact clock its evaluation.1
Exercise 36.1.1 Do macros merely implement laziness?
36.1.2 Example: Local Definitions

We saw earlier this semester that
{with {var val} body}
could be rewritten as
{{fun {var} body} val}
by a preprocessor, so our core evaluator did not need to implement with directly. The same is true of the
let construct in Scheme. Here’s a simple macro for let (again, we’ll use the my- convention to avoid any
clashes):
(define-syntax my-let-1
(syntax-rules ()
[(my-let-1 (var val) body)
((lambda (var) body) val)]))
Sure enough,
> (my-let-1 (x 3) (+ x 4))
7
> (my-let-1 (x 3) (my-let-1 (y 4) (+ x y)))
7
In full Scheme, however, the let construct is a bit more complex: it permits binding several identifiers at the
same time (as we saw in a homework assignment regarding with). Therefore, the true translation should
be regarded as something along these lines:
(let ([var val] . . .) body) =⇒ ((lambda (var . . .) body) val . . .)
That is, we want each of the variables to remain in the same order, and likewise each of the value expressions—
except we don’t know how many we will encounter, so we use . . . to indicate “zero or more”.
How are we going to define this macro? In fact, it couldn’t be easier. A researcher, Eugene Kohlbecker,
observed that numerous extensions to Scheme had this same “zero or more” form, and noticed that people
always wrote them informally using the stylized notation above. He therefore simply defined a macro system
that processed that notation:
(define-syntax my-let
(syntax-rules ()
[(my-let ([var val] · · ·) body)
((lambda (var · · ·) body) val · · ·)]))
1 Technically, this isn’t exactly the expression that evaluates. We’ll return to this in a bit.
Therefore (my-let ([x 3] [y 4]) (+ x y)) translates into ((lambda (x y) body) 3 4) which, sure enough, reduces
to 7. Notice how the macro system is smart enough to treat the ([var val] · · ·) pattern as being the composite
of the var · · · and val · · · patterns.2 In particular, if no identifiers are bound, then this turns into an immediate
application of a thunk to no arguments, which just evaluates the body.
36.1.3 Example: Nested Local Definitions

In a let, all the named expressions are bound in the same scope, which doesn’t include any of the bound
names. Sometimes, it’s useful to bind names sequentially so later bindings can refer to earlier ones. Scheme
provides the construct let∗ for this task:
(let∗ ([a 5]
[b 12]
[aˆ2 (∗ a a)]
[bˆ2 (∗ b b)]
[aˆ2+bˆ2 (+ aˆ2 bˆ2)])
(sqrt aˆ2+bˆ2))
(Think of what this would evaluate to with let instead of let∗.)
We can implement let∗ very easily by unraveling it into a sequence of lets:
(let∗ ([var val] · · ·) body) =⇒ (let ([var0 val0 ])
(let ([var1 val1 ])
···
(let ([varn valn ])
body)))
There is a stylized way of writing such macros in Scheme, which is to split them into two cases: when the
sequence is empty and when the sequence has one or more elements. When there are no identifiers being
bound, then let∗ does the same thing as let (which is to reduce to the expression itself):
(let∗ () body) =⇒ body
Since each · · · means “zero or more”, we need to use a more refined pattern to indicate “one or more”:
(let∗ ([var0 val0] [var-rest val-rest] · · ·) body)
The rewrite rule then becomes
(let∗ ([var0 val0] [var-rest val-rest] · · ·) body) =⇒ (let ([var0 val0])
(let∗ ([var-rest val-rest]
..
.)
body))
That is, we apply the macro for let∗ recursively. Written in Scheme syntax, this is expressed as (notice the
two cases):
2 The use of brackets versus parentheses is purely stylistic.
(define-syntax my-let∗
(syntax-rules ()
[(my-let∗ () body)
body]
[(my-let∗ ([var0 val0]
[var-rest val-rest] · · ·)
body)
(let ([var0 val0])
(my-let∗ ([var-rest val-rest] · · ·)
body))]))
There is nothing in Scheme that prevents a runaway expansion. Therefore, it’s possible to write a
misbehaving macro that expands forever, so that evaluation never even begins. However, most macros
follow the simple stylistic pattern above, which guarantees termination (the recursion is over the bound
identifiers, and each time through, one more identifier-value pair is taken off).
36.1.4 Example: Simple Conditional

Let’s say we want a simplified form of conditional that has only two branches and one conditional. This is
effectively the same as if:
(cond2 [t e1] [else e2]) =⇒ (if t e1 e2)
We might try the following macro:
(define-syntax cond2
(syntax-rules ()
[(cond2 (t e1) (else e2))
(if t e1 e2)]))
This correctly evaluates expressions such as
(cond2 [(even? (current-seconds)) ’even]
[else ’odd])
Unfortunately, this also permits expressions such as
(cond2 [(even? (current-seconds)) ’even]
[(odd? (current-seconds)) ’odd])
This shouldn’t be syntactically legal, because cond2 permits only one conditional; in place of the second,
we require programmers to write else. We can see that this second expression doesn’t get evaluated at all by
writing something atrocious:
(cond2 [false ’even]
[(/ 1 0) ’odd])
which evaluates to ’odd.
What we want is for the cond2 macro to simply reject any uses that don’t have else in the second question
position. This is where the mystical () after syntax-rules comes in: it lists the keywords in the macro. That
is, we should instead define the macro as
(define-syntax cond2
(syntax-rules (else)
[(cond2 (t e1) (else e2))
(if t e1 e2)]))
Then, we get the following interaction:
> (cond2 [false ’even]
[(/ 1 0) ’odd])
cond2: bad syntax in: (cond2 (false (quote even)) ((/ 1 0) (quote odd)))}
Without the keyword designation, Scheme has no way of knowing that else has a special status; naturally,
it makes no sense to build that knowledge into the macro system. Absent such knowledge, it simply treats
else as a macro variable, and matches it against whatever term is in that position. When we put else in the
keyword list, however, the expander no longer binds it but rather expects to find it in the right position—or
else rejects the program.
36.1.5 Example: Disjunction

Let’s consider one more example from Scheme lore. In Scheme, conditionals like or and and short-circuit:
that is, when they reach a term whose value determines the result of the expression, they do not evaluate the
subsequent terms. Let’s try to implement or.
To begin with, let’s define the two-arm version of or:
(define (my-or2-fun e1 e2)
(if e1
e1
e2))
Sure enough, a very simple example appears to work
> (my-or2-fun false true)
#t
but it fails on a more complex example:
> (let ([x 0])
(my-or2-fun (zero? x)
(zero? (/ 1 x))))
/: division by zero
whereas a short-circuiting evaluator would not have permitted the error to occur. The problem is, once again,
Scheme’s eager evaluation regime, which performs the divison before it ever gets to the body of my-or2-fun.
In contrast, a macro does not have this problem:
(define-syntax my-or2
(syntax-rules ()
[(my-or2 e1 e2)
(if e1 e1 e2)]))
which yields
> (my-or2 false true)
#t
> (let ([x 0])
(my-or2 (zero? x)
(zero? (/ 1 x))))
#t
In particular, the second expression translates into
(let ([x 0])
(if (zero? x)
(zero? x)
(zero? (/ 1 x))))
(just replace e1 and e2 consistently).
As this expansion begins to demonstrate, however, this is an unsatisfying macro. We evaluate the first
expression twice, which has the potential to be inefficient but also downright wrong. (Suppose the first
expression were to output something; then we’d see the output twice. If the expression wrote a value into a
database and returned a code, executing it a second time may produce a different result than the first time.)
Therefore, we’d really like to hold on to the value of the first evaluation and return it directly if it’s not false.
That is, we want
(define-syntax my-or2
(syntax-rules ()
[(my-or2 e1 e2)
(let ([result e1])
(if result
result
e2))]))
This expands the second expression into
(let ([x 0])
(let ([result (zero? x)])
(if result
result
(zero? (/ 1 x)))))
Since Scheme is eager, the expression in the e1 position gets evaluated only once. You should construct test
cases that demonstrate this.
36.2. HYGIENE 327
36.2 Hygiene
Now what if the use of my-or2 really looked like this?
(let ([result true])
(my-or2 false
result))
which should evaluate to true. The expansion, however, is
(let ([result false])
(if result
result
result)))
which evaluates to false!
What happened here? When we look at just the input expression, we do not see only one binding of
result. Reasoning locally to that expression, we assume that my-or2 will evaluate the first expression and,
finding it false, will evaluate the second; since this is result, which is bound to true, the overall response
should also be true. Instead, however, the use of result within the macro definition interferes with result in
the context of its use, resulting in the incorrect result.
The problem we see here should seem awfully familiar: this is exactly the same problem we saw under
a different guise when trying to understand scope. Here, result in the second arm of the disjunction is bound
in the let just outside the disjunction. In contrast, result inside the macro is bound inside the macro. We as
programmers should not need to know about all the names used within macros—just as we don’t need to
know the names of identifiers used within functions! Therefore, macros should be forced to obey the scoping
rules of the language.
Just to be sure, let’s try this expression in our evaluator:
> (let ([result true])
(my-or2 false result))
#t
We get true! This is because Scheme’s macro system is hygienic. That is, it automatically renames identifiers
to avoid accidental name clashes. The expression that actually evaluates is something like
(let ([g1729 false])
(if g1729
g1729
result)))
where g1729 is a uniquely-generated identifier name. Notice that only the results within the macro definition
get renamed. In fact, because let is itself a macro, its identifiers also get renamed (as do those introduced by
lambda and other binding forms), so the real program sent to the evaluator might well be
(let ([g4104 true])
(let ([g1729 false])

(if g1729
g1729
g4104)))
Many macro systems, such as that of C, are not hygienic. Programmers sometimes try to circumvent this by
using hideous identifier names, such as __macro_result_. This is not a solution!
1. Not only is a it painful to have to program this way, small typos would greatly increase development
time, and the macro would be much harder to decipher when a programmer tries to modify or correct
it later.
2. This solution is only as good as the programmer’s imagination; the problem still persists, lying in wait
for just the right (wrong!) identifier to be bound in the context of use. Indeed, while a programmer
may choose a sufficiently obscure name from the perspective of other programmers, not all source is
written by humans. A tool generating C code (such as a Scheme-to-C compiler) may happen to use
exactly this naming convention.
3. This name is only obscure “upto one level”. If the macro definition is recursive, then recursive in-
stances of the macro may interfere with one another.
4. If you use this macro to debug the source that contains the macro (e.g., compiling the C compiler
using itself), then your carefully-chosen “obscure” name is now guaranteed to clash!
In short, to return to a theme of this course: we should view these kinds of contortions by programmers as
a symptom of a problem that must be addressed by better language design. Don’t settle for mediocrity! In
this case, hygiene is that solution.3
Notice, by the way, that we needed hygiene for the proper execution of our very first macro, because
my-time introduced the identifier begin-time. At the time, we never even gave a thought to this identifier,
which means in the absence of hygiene, we had a disaster waiting to happen. With hygiene, we can program
using normal names (like begin-time and result) and not have to worry about the consequencs down the line,
just as with static scope we can use reasonable names for local identifiers.
36.3 More Macrology by Example

Many languages provide a looping construct for iterating through integers sequentially. Scheme doesn’t for
three reasons:
1. Because most such loops are anyway inappropriate: the indices only exist to traverse sequential data
structures. Uses of map or filter over a list accomplish the same thing but at a higher level of abstrac-
tion.
2. Because recursion in the presence of tail calls has the same computational effect.
3 The algorithm, in effect, “paints” each expression on expansion, then consistently renames identifiers that have the same paints.
36.3. MORE MACROLOGY BY EXAMPLE 329
3. Because, if we really crave a more traditional syntax, we can define it using a macro!
We’ll build up a loop macro in three stages.
36.3.1 Loops with Named Iteration Identifiers

Here’s our first attempt at a for loop macro.4 We’ve generously embellished it with keywords to increase
readability:
(define-syntax for0
(syntax-rules (from to in)
[(for0 hvari from hlowi to hhighi in hbodiesi · · ·)
(local ([define loop (lambda (hvari)
(if (> hvari hhighi)
’done
(begin
hbodiesi · · ·
(loop (+ hvari 1)))))])
(loop hlowi))]))
This lets us write programs such as
(for0 x
from 2
to 5
in (display x))
which prints 2, 3, 4 and 5. However, when we try this on a program like this
(for0 x
from 2
to (read)
in (display x))
we notice an unpleasant phenomenon: the program reads the upper-bound of the loop every time through
the loop. To correct it, we should make sure it evaluates the upper-bound expression only once, which we
can do with a small change to the macro:
(define-syntax for1
(local ([define high-value hhighi]
[define loop (lambda (hvari)
(if (> hvari high-value)
’done
(begin
4 We’re using the convention of wrapping macro pattern-variables in h·i to emphasize their relationship to BNF.
hbodiesi · · ·
(loop (+ hvari 1)))))])
(loop hlowi))]))
In general, we must be very careful with macros to ensure expressions are evaluated the right number of
times. In this instance, hlowi is going to be evaluated only once and hvari is only an identifier name, but we
have to make sure hhighi is evaluated only once.
In fact, however, this version is also buggy! If there is a (read) in the hlowi position, that’s going to get
evaluated second instead of first, which is presumably not what we wanted (though notice that we didn’t
formally specify the behavior of for, either). So to get it right, we really need to evaluate hlowi and bind its
value to an identifier first.
In general, it’s safer to bind all expression positions to names. Scheme’s eager evaluation semantics
ensures the expressions will only be evaluted once. We don’t always want this, but we want it so often that
we may as well do it by default. (The times we accidentally bind an expression too early—for instance, the
conditional expression of a while loop—we will usually discover the problem pretty quickly by testing.) In
addition we must be sure to do this binding in the right order, mirroring what the user expects (and what
our documentation for the new language construct specifies). (Observe that the problematic expression in
this example is (read), which has the side-effect of prompting the user. Of course, we may want to limit
evaluation for efficiency reasons also.)
36.3.2 Overriding Hygiene: Loops with Implicit Iteration Identifiers

When we define a loop such as the one above, we often have no real use for the loop variable. It might be
convenient to simply introduce an identifier, say it, that is automatically bound to the current value of the
loop index. Thus, the first loop example above might instead be written as
(for2 from 2
to 5
in (display it))
Here’s a proposed macro that implements this construct:
(define-syntax for2
[(for2 from hlowi to hhighi in hbodiesi · · ·)
[define loop (lambda (it)
(if (> it high-value)
’done
(begin
hbodiesi · · ·
(loop (+ it 1)))))])
(loop hlowi))]))
Notice that in place of hvari, we are now using it. When we run this in DrScheme, we get:
> (for2 from 2 to 5 in (display it))

reference to undefined identifier: it
Oops! What happened?

Actually, the macro system did exactly what it should. Remember hygiene? This was supposed to
prevent inadvertent capture of identifiers across the macro definition/macro use boundary. It just so happens
that in this case, we really do want it written in the macro use to be bound by it in the macro definition.
Clearly, here’s a good example of where we want to “break” hygiene, intentionally.
Unfortunately, the simple syntax-rules mechanism we’ve been using so far isn’t quite up to this task;
we must instead switch to a slightly more complex macro definition mechanism called syntax-case. For the
most part, this looks an awful lot like syntax-rules, with a little more notation. For instance, we can define
for3 to be the same macro as for1, except written using the new macro definition mechanism instead:
(define-syntax (for3 x)
(syntax-case x (from to in)
(syntax
’done
(begin
hbodiesi · · ·
(loop (+ hvari 1)))))])
(loop hlowi)) ) ]))
To convert any syntax-rules macro definition into a corresponding one that uses syntax-case, we must
make the three changes boxed above (adding a parameter to the macro name, providing the parameter as an
explicit argument to syntax-case, and wrapping the entire output expression in (syntax · · ·).
We can similarly define for4:
(syntax
’done
(begin
hbodiesi · · ·
(loop (+ it 1)))))])
(loop hlowi)))]))
This does not solve the hygiene problem; it simply enables it by converting the macro to use syntax-case.
The reason is that syntax-case provides additional capabilities. In particular, it provides a procedure called
datum→syntax-object, which takes an arbitrary Scheme datum and a term in the macro body, and “paints”
the datum with the same colors as those on the macro body. This has the effect of persuading the hygiene
mechanism to treat the introduced term as if it were written by the programmer. As a result, it gets renamed
consistently. Thus, we must write
(with-syntax ([it (datum→syntax-object (syntax for4) ’it)])
(syntax
’done
(begin
hbodiesi · · ·
(loop (+ it 1)))))])
(loop hlowi))))]))
The with-syntax construct introduces new pattern variables for use in the output. The first argument to
datum→syntax-object identifies which expression the identifier the expander must pretend “introduced” the
identifier. The second, in this case, is the symbol that will be painted appropriately. Therefore, the result of
expansion on our running example will look something like
(local ([define high-value 5]
[define loop (lambda (g1729)
(if (> g1729 high-value)
’done
(begin
(display g1729)
(loop (+ g1729 1)))))])
(loop 2))
Observe how the uses of it are all renamed consistently. (In practice, other bound identifiers such as high-
value and even loop will also acquire fresh names, but we don’t show that here to keep the code more
readable.) Indeed, this mechanism is sufficiently robust that it will even do the right thing with nested loops:
(for4 from 2 to 5 in
(for4 from 1 to it in
(display it))
(newline))
generates
12
123
1234
12345
In the inner loop, notice that the it in the loop bound (from 1 to it) is the iteration index for the outer loop,
while the it in (display it) is the index for the inner loop. The macro system associates each it appropriately
because each use of for4 gets a different coat of colors. Unfortunately, we have lost the ability to refer to
the outer iteration in the inner loop.
36.3.3 Combining the Pieces: A Loop for All Seasons

A better design for an iteration construct would be to combine these ways of specifying the iteration identifier
(explicitly and implicitly). This is easy to do: we simply have two rules.5 If an identifier is present, use it as
before, otherwise bind it and recur in the macro.
(with-syntax ([it (datum→syntax-object (syntax for5) ’it)])
(syntax
(for5 it from hlowi to hhighi in hbodiesi · · ·)))]
(syntax
’done
(begin
hbodiesi · · ·
(loop (+ hvari 1)))))])
(loop hlowi)))]))
This passes all the expected tests: both the following expressions print the numbers 2 through 5:
(for5 x from 2 to 5 in (display x))
(for5 from 2 to 5 in (display it))
while this
(for5 x from 2 to 5 in
(for5 from 1 to x in
(printf ”[ã, ã] ” x it))
(newline))
5 When defining such macros, be very sure to test carefully: if an earlier rule subsumes a later rule, the macro system will not
complain, but the code will never get to a later rule! In this case we need not worry since the two rules have truly different structure.
prints
[2, 1] [2, 2]
[3, 1] [3, 2] [3, 3]
[4, 1] [4, 2] [4, 3] [4, 4]
[5, 1] [5, 2] [5, 3] [5, 4] [5, 5]
There are still ways to many ways of improving this macro. First, we might want to make sure hvari is
really a variable. We can use identifier? for this purpose. The syntax-case mechanism also permits guards,
which are predicates that refine the patterns and don’t allow a rule to fire unless the predicates are met.
Finally, the following program does not work:
(for5 x from 2 to 5 in
(for5 from 1 to it in
(printf ”[ã, ã] ” x it))
(newline))
It reports that the boxed it is not bound (why?). Try to improve the macro to bind it in this case.
36.4 Comparison to Macros in C

Macro systems have a bad rap in the minds of many programmers. This is invariably because the only
macro system they have been exposed to is that of C. C’s macros are pretty awful, and indeed used to
be worse: macros could contain parts of lexical tokens, and macro application would glue them together
(e.g., the identifier list-length could be assembled by a macro that generated lis, another generating t-le
and yet another generating ngth). C macros are not hygienic. Because C has no notion of local scope, C
macros could not easily introduce local identifiers. Finally, C macros are defined by the C pre-processor
(cpp), which operates on files a line at a time. Therefore, to apply a macro over a multi-line argument, a C
programmer would have to use a \ at the end of each line to fool the pre-processor into concatenating the
adjacent line with the present one. Failing to remember to use the line-continuation character could lead to
interesting errors.
In contrast, Scheme’s macros operate over parenthesized expressions instead of pestering programmers
with lines. They respect lexical boundaries (to create a new identifier, you must do so explicitly—it cannot
happen by accident). Scheme macros are hygienic. They have many more features that we haven’t discussed
here. In short, they correct just about every mistake that C’s macro system made.
36.5 Abuses of Macros

When shouldn’t a programmer use macros?
As you can see, macros provide a programmer-controlled form of inlining, that is, directly substituting
the body of an abstraction in place of its use. Compilers often inline small procedures to avoid paying
the cost of procedure invocation and return. This permits programmers to define abstractions for simple
operations—such as finding the corresponding matrix element in the next row when the matrix is stored
36.6. USES OF MACROS 335
linearly, or performing some bit-twiddling—without worrying about the overhead of function invocation.
Normally, inlining is done automatically by a compiler, after it has examined the size of the procedure body
and determined whether or not it is cost-effective to inline.
Unfortunately, early compilers were not savvy enough to inline automatically. C programmers, looking
to squeeze the last bit of performance out of their programs, therefore began to replace function definitions
with macro definitions, thereby circumventing the compiler. Invariably, compilers got smarter, architectures
changed, the cost of operations altered, and the hard-coded decisions of programmers came to be invalidated.
Nowadays, we should regard the use of macros to manually implement inlining as a programming error.
Unfortunately, many C programmers still think this is the primary use of macros (and in C, it’s not useful
for a whole lot else), thereby further despoiling their reputation. (The previous sentence is intentionally
ambiguous.)
Another bad use of macros is to implement laziness. Macros do not correspond to lazy evaluation. Lazi-
ness is a property of when the implementation evaluates arguments to functions. Macros are not functions.
For instance, in Scheme, we cannot pass a macro as an argument: try passing or as an argument to map and
see what happens. Indeed, macro expansion (like type-checking) happens in a completely different phase
than evaluation, while laziness is very much a part of evaluation. So please don’t confuse the two.
36.6 Uses of Macros

When should a programmer use macros?
providing cosmetics Obviously, macros can be used to reduce the syntactic burden on programmers. These
are perhaps the least interesting use; at least, a macro that does this should also fulfill one of the other
uses.
introducing binding constructs Macros can be used to implement non-standard binding constructs. We
have seen two examples, let and let∗, above. If these were not already in the language, we could
easily build them using macros. They would be impossible to define as language contructs in most
other languages.
altering the order of evaluation Macros can be used to impose new orders-of-evaluation. For instance, we
saw time suspend evaluation until the clock’s value had been captured. The or construct introduced
short-circuit evaluation. Often, programmers can obtain the same effect by thunking all the sub-
expressions and thawing (the opposite of thunking) them in the desired order, but then the programmer
would be forced to write numerous lambda () · · ·)’s—replacing one intrusive, manual pattern with
another. (In particular, if a programmer fails to obey the pattern faithfully, the behavior may become
quite difficult to predict.)
defining data languages Sometimes the sub-terms of a macro application may not be Scheme expressions
at all. We have seen simple instances of this: for example, in (my-let ([x 3] [y 4]) (+ x y)), neither
the parentheses wrapping the two bindings, nor those surrounding each name-value pair, signify ap-
plications. In general, the terms may have arbitrary structure, even including phrases that would be
meaningless in Scheme, such as (my-macro (lambda (x))) that would be syntactic errors otherwise.
We can get some of the same benefit from using quotations, but those are run-time values, whereas
here the macro can traverse the sub-terms and directly generate code.
In particular, suppose you wish to describe a datum without choosing whether it will be represented
as a structure or as a procedure. In ordinary Scheme, you have to make this decision up front, because
you cannot “introduce a lambda” after the fact. Designating the datum using a macro lets you hide
this decision, deferring the actual representation to the macro.
Chapter 37
Macros and their Impact on Language

Design
37.1 Language Design Philosophy

The Revised5 Report on the Algorithmic Language Scheme famously begins with the following design man-
ifesto:
Programming languages should be designed not by piling feature on top of feature, but by
removing the weaknesses and restrictions that make additional features appear necessary.
Scheme augments a minimal set of features with a powerful macro system, which enable the creation of
higher-level language primitives. This approach can only work, however, with a carefully designed target
language for expansion. Its success in Scheme depends on a potent combination of two forces:
• A set of very powerful core features.
• Very few restrictions on what can appear where (i.e., values in the language are truly first-class, which
in turn means the expressions that generate them can appear nearly anywhere).
The first means many macros can accomplish their tasks with relatively little effort, and the second means
the macros can be written in a fairly natural fashion.
This manner of structuring a language means that even simple programs may, unbenownst to the pro-
grammer, invoke macros, and tools for Scheme must be sensitive to this fact. For instance, DrScheme is
designed to be friendly to beginners. Even simple beginner programs expand into rather complicated and
relatively mangled code, many of whose constructs the beginner will not understand. Therefore, when
reporting errors, DrScheme uses various techniques to make sure this complexity is hidden from the pro-
grammer.
Building a language through macros does more than just complicate error reporting. It also has signifi-
cant impact on the forms of generated code that the target implementation must support. Programmers who
build these implementations make certain assumptions about the kinds of programs they must handle well;
these are invariably based on what “a normal human would write”. Macros, however, breaks these unwritten
337
338 CHAPTER 37. MACROS AND THEIR IMPACT ON LANGUAGE DESIGN
rules. They produce unusual and unexpected code, resulting in correctness and, particularly, performance
errors. Sometimes these are easy to fix; in many other cases they are not. We will study examples to illustrate
instances where macros crucially depend on the target language’s handling of certain key code patterns.
37.2 Example: Pattern Matching

We will now examine a rather unusual construct that a programmer would never write, and explain why
an implementation should nevertheless search for instances of it and handle it efficiently. To set the stage,
consider the Scheme construct let, which binds names to values in a local lexical context. Though this (or
an equivalent way of introducing local scope) would be a language primitive in most languages, in Scheme
this is expressible as a rather simple macro in terms of first-class functions. That is,
(let ((v e) · · ·) b)
can be implemented by expanding into1
((lambda (v · · ·) b) e · · ·)
where (lambda (v · · ·) b) introduces an (anonymous) procedure with argument list v · · · and body b, and the
outer parentheses apply this procedure to the argument expressions e · · ·. The application binds the variables
v · · · to the values of the expressions e · · ·, and in that extended environment evaluates the body b—exactly
what we would intend as the semantics for let. For instance, the program
(let ([x 3]
[y 2])
(+ x y))
which evaluates to 2 + 3, i.e., 5, is transformed into
((lambda (x y)
(+ x y))
3 2)
This macro is, in fact, quite easy to implement: thanks to hygiene and pattern matching, the implementer of
let merely needs to write
(define-syntax let
(syntax-rules ()
[(let ([v e] · · ·) b)
((lambda (v · · ·) b) e · · ·)]))
The Scheme pre-processor finds all bodies of the form (let · · ·), matches them against the input pattern (here,
(let ([v e] · · ·) b)), binds the pattern variables (v, e and b) to the corresponding sub-expressions, and replaces
the body with the output pattern in which the pattern variables have been replaced by the sub-expressions
bound to them.
1 For simplicity, we assume the body has only one expression. In reality, Scheme permits multiple expressions in the body,
which is useful in imperative programs.

37.2. EXAMPLE: PATTERN MATCHING 339
There is, however, a significant performance difference between the two forms. A compiler can im-
plement let by extending the current activation record with one more binding (for which space can be
pre-allocated by the creator of the record). In contrast, the expanded code forces the compiler to both create
a new closure and then apply it—both relatively more expensive operations.
Given this expense, you might think it silly for a Scheme system to implement the let-to-lambda macro:
why take an efficient source-language instruction, whose intent is apparent, and make it less efficient behind
the programmer’s back? Yet at least one Scheme compiler (Chez Scheme) does precisely this. Furthermore,
in the back end, it finds instances of ((lambda · · ·) · · ·) and effectively handles them as it would have let.
Why would a compiler behave so perversely? Surely no human would intentionally write ((lambda · · ·)
· · ·), so how else could these arise? The operative phrase is, of course, “no human”. Scheme programs are
full of program-generating programs, and by treating this odd syntactic pattern as a primitive, all macros
that resolve into it benefit from the compiler’s optimizations.
Consider a simple symbol-based conditional matcher: the user writes a series of symbol and action pairs,
such as
(switch [off 0]
[on 1])
The matcher performs the symbol comparison and, when a symbol matches, executes the corresponding
action (in this case, the actions are already numerical values). The entire (switch · · ·) expression becomes a
function of one argument, which is the datum to compare. Thus, a full program might be
(define m
(switch [off 0]
[on 1]))
with the following interactions with the Scheme evaluator:
> (m ’off)
0
> (m ’on)
1
To implement switch, we need a macro rule when there are one or more cases:
(switch ⇒ (lambda (v)

[sym0 act0] (if (symbol=? v (quote sym0))
[pat-rest act-rest] act0
· · ·) ((switch
[pat-rest act-rest]
· · ·)
v)))
This yields a function that consumes the actual value (v) to match against. The matcher compares v against
the first symbol. If the comparison is successful, it invokes the first action. Otherwise it needs to invoke the
(define-syntax switch
(syntax-rules ()
[(switch) (lambda (v) false)]
[(switch [sym0 act0]
[pat-rest act-rest]
· · ·)
(lambda (v)
(if (symbol=? v (quote sym0))
act0
((switch
[pat-rest act-rest]
· · ·)
v)))]))
Figure 37.1: Simple Pattern Matcher
pattern-matcher on the remaining clauses. Since a matcher is a function, invoking it is a matter of function
application. So applying this function to v will continue the matching process.2
For completeness, we also need a rule when no patterns remain. For simplicity, we define our matcher
to return false3 (a better response might be to raise an exception):
(switch) ⇒ (lambda (v) false)
Combining these two rules gives us the complete macro, shown in Figure 37.1.
Given this macro, the simple use of switch given above generates
(lambda (v0)
(if (symbol=? v0 (quote off))
0
((lambda (v1)
(if (symbol=? v1 (quote on))
1
((lambda (v2)
false)
v1)))
v0)))
(I’ve used different names for each v, as the hygienic expander might, to make it easy to keep them all apart.
Each v is introduced by another application of the switch macro.)
2 The · · · denotes “zero or more”, so the pattern of using one rule followed by a · · · is common in Scheme macros to capture the
potential for an unlimited number of body expressions.
3 In many Scheme systems, true and false are written as #t and #f, respectively.
37.3. EXAMPLE: AUTOMATA 341
While this expanded code is easy to generate, its performance is likely to be terrible: every time one
clause fails to match, the matcher creates and applies another closure. As a result, even if the programmer
wrote a pattern matching sequence that contained no memory-allocating code, the code might yet allocate
memory! That would be most unwelcome behavior.
Fortunately, the compiler comes to the rescue. It immediately notices the ((lambda · · ·) · · ·) pattern and
collapses these, producing effectively the code:
(lambda (v0)
0
(let ([v1 v0])
1
(let ([v2 v1])
false)))))
In fact, since the compiler can now see that these lets are now redundant (all they do is rename a variable),
it can remove them, resulting in this code:
(lambda (v0)
0
1
false)))
This is pretty much exactly what you would have been tempted to write by hand. In fact, read it and it’s
obvious that it implements a simple conditional matcher over symbols. Furthermore, it has a very convenient
interface: a matcher is a first-class function value suitable for application in several contexts, being passed to
other procedures, etc. The macro produced this by recursively generating lots of functions, but a smart choice
of compiler “primitive”—((lambda · · ·) · · ·), in this case—that was sensitive to the needs of macros reduced
the result to taut code. Indeed, it now leaves the code in a state where the compiler can potentially apply
further optimizations (e.g., for large numbers of comparisons, it can convert the cascade of comparisons into
direct branches driven by hashing on the symbol being compared).
37.3 Example: Automata

Next, we examine another optimization that is crucial for capturing the intended behavior of many programs.
As an example, suppose we want to define automata manually. Ideally, we should be able to specify the
automata once and have different interpretations for the same specification; we also want the automata to
be as easy as possible to write (here, we stick to textual notations). In addition, we want the automata to
execute fairly quickly, and to integrate well with the rest of the code (so they can, for instance, be written
in-line in programs).
Concretely, suppose we want to write a simple automaton that accepts only patterns of the form (01)∗ .
We might want to write this textually as
automaton see0
see0 : 0 -> see1
see1 : 1 -> see0
where the state named after the keyword automaton identifies the initial state.
Consider a slightly more complex automaton, one that recognizes the Lisp identifier family car, cdr,
cadr, cddr, cddar and so on. To simplify the example, let’s say it should recognize the regular language
c(ad)∗ r. The corresponding automaton might look like
automaton init
init : c -> more
more : a -> more
d -> more
r -> end
end :
We leave defining a more formal semantics for the automaton language as an exercise for the reader.
It is easy to see that some representation of the textual description suffices for treating the automata
statically. How do we implement them as programs with dynamic behavior? We request you, dear reader,
to pause now and sketch the details of an implementation before proceeding further.
A natural implementation of this language is to create a vector or other random-access data structure to
represent the states. Each state has an association indicating the actions—implemented as an association list,
associative hash table, or other appropriate data structure. The association binds inputs to next states, which
are references or indices into the data structure representing states. Given an actual input stream, a program
would walk this structure based on the input. If the stream ends, it would accept the input; if no next state
is found, it would reject the input; otherwise, it would proceed as per the contents of the data structure. (Of
course, other implementations of acceptance and rejection are possible.)
One Scheme implementation of this program would look like this. First we represent the automaton as
a data structure:
(define machine
’((init (c more))
(more (a more)
(d more)
(r end))
(end)))
The following program is parameterized over machines and inputs:
(define (run machine init-state stream)
(define (walker state stream)
(or (empty? stream) ;; if empty, return true, otherwise . . .
(let ([transitions (cdr (assv state machine))]
[in (first stream)])
(let ([new-state (assv in transitions)])
(if new-state
(walker (cadr new-state) (rest stream))
false)))))
(walker init-state stream))
Here are two instances of running this:
> (run machine ’init ’(c a d a d d r))
true
> (run machine ’init ’(c a d a d d r r))
false
This is not the most efficient implementation we could construct in Scheme, but it is representative of
the general idea.
While this is a correct implementation of the semantics, it takes quite a lot of effort to get right. It’s easy
to make mistakes while querying the data structure, and we have to make several data structure decisions in
the implementation (which we have done only poorly above). Can we do better?
To answer this question affirmatively, let’s ignore the details of data structures and understand the
essence of these implementations.
1. Per state, we need fast conditional dispatch to determine the next state.
2. Each state should be quickly accessible.
3. State transition should have low overhead.
Let’s examine these criteria more closely to see whether we can recast them slightly:
fast conditional dispatch This could just be a conditional statement in a programming language. Compiler
writers have developed numerous techniques for optimizing properly exposed conditionals.
rapid state access Pointers of any sort, including pointers to functions, would offer this.
quick state transition If only function calls were implemented as gotos . . .
In other words, the init state could be represented by

(lambda (stream)
(or (empty? stream)
(case (first stream)
[(c) (more (rest stream)) ]
[else false])))
That is, if the stream is empty, the procedure halts returning a true value; otherwise it dispatches on the first
stream element. Note that the boxed expression is invoking the code corresponding to the more state. The
code for the more state would similarly be
(lambda (stream)
(or (empty? stream)

[(a) (more (rest stream))]
[(d) (more (rest stream))]
[(r) (end (rest stream))]
[else false])))
Each underlined name is a reference to a state: there are two self-references and one to the code for the end
state. Finally, the code for the end state fails to accept the input if there are any characters in it at all. While
there are many ways of writing this, to remain consistent with the code for the other states, we write it as
(lambda (stream)
(or (empty? stream)
(case (first stream) ;; no matching clauses, so always false
[else false])))
The full program is shown in Figure 37.2. This entire definition corresponds to the machine; the definition
of machine is bound to init, which is the function corresponding to the init state, so the resulting value
needs only be applied to the input stream. For instance:
> (machine ’(c a d a d d r))
true
> (machine ’(c a d a d d r r))
false
What we have done is actually somewhat subtle. We can view the first implementation as an interpreter
for the language of automata. This moniker is justified because that implementation has these properties:
1. Its output is an answer (whether or not the automaton recognizes the input), not another program.
2. It has to traverse the program’s source as a data structure (in this case, the description of the automa-
ton) repeatedly across inputs.
3. It consumes both the program and a specific input.
It is, in fact, a very classical interpreter. Modifying it to convert the automaton data structure into some
intermediate representation would eliminate the second overhead in the second clause, but would still leave
open the other criteria.
In contrast, the second implementation given above is the result of compilation, i.e., it is what a compiler
from the automaton language to Scheme might produce. Not only is the result a program, rather than an
answer for a certain input, it also completes the process of transforming the original representation into one
that does not need repeated processing.
While this compiled representation certainly satisfies the automaton language’s semantics, it leaves two
major issues unresolved: efficiency and conciseness. The first owes to the overhead of the function appli-
cations. The second is evident because our description has become much longer; the interpreted solution
required the user to provide only a concise description of the automaton, and reused a generic interpreter
to manipulate that description. What is missing here is the actual compiler that can generate the compiled
version.
(define machine
(letrec ([init
(lambda (stream)
(or (empty? stream)
[(c) (more (rest stream))]
[else false])))]
[more
(lambda (stream)
(or (empty? stream)
[(a) (more (rest stream))]
[(d) (more (rest stream))]
[(r) (end (rest stream))]
[else false])))]
[end
(lambda (stream)
(or (empty? stream)
[else false])))])
init))
Figure 37.2: Alternate Implementation of an Automaton
37.3.1 Concision
First, let us slightly alter the form of the input. We assume that automata are written using the following
syntax (presented informally):
(automaton init
(init : (c → more))
(more : (a → more)
(d → more)
(r → end))
(end : ))
The general transformation we want to implement is quite clear from the result of compilation, above:
(state : (label → target) · · ·) ⇒ (lambda (stream)
(or (empty? stream)
[(label) (target (rest stream))]
···
[else false])))
(define-syntax automaton
(syntax-rules (: →) ;; match ‘:’ and ‘→’ literally, not as pattern variables
[(automaton init-state
(state : (label → target) · · ·)
· · ·)
(letrec ([state
(lambda (stream)
(or (empty? stream)
[(label) (target (rest stream))]
···
[else false])))]
· · ·)
init-state)]))
Figure 37.3: A Macro for Executable Automata
Having handled individual rules, we must make the automaton macro wrap all these procedures into a
collection of mutually-recursive procedures. The result is the macro shown in Figure 37.3. To use the
automata that result from instances of this macro, we simply apply them to the input:
> (define m (automaton init
[init : (c → more)]
[more : (a → more)
(d → more)
(r → end)]
[end : ]))
> (m ’(c a d a d d r))
true
> (m ’(c a d a d d r r))
false
By defining this as a macro, we have made it possible to truly embed automata into Scheme programs.
This is true purely at a syntactic level—since the Scheme macro system respects the lexical structure of
Scheme, it does not face problems that an external syntactic preprocessor might face. In addition, an au-
tomaton is just another applicable Scheme value. By virtue of being first-class, it becomes just another
linguistic element in Scheme, and can participate in all sorts of programming patterns.
In other words, the macro system provides a convenient way of writing compilers from “Scheme+” to
Scheme. More powerful Scheme macro systems allow the programmer to embed languages that are truly
different from Scheme, not merely extensions of it, into Scheme. A useful slogan (due to Matthew Flatt and
quite possibly others) for Scheme’s macro system is that it’s a lightweight compiler API.
37.3.2 Efficiency
The remaining complaint against this implementation is that the cost of a function call adds so much over-
head to the implementation that it negates any benefits the automaton macro might conceivably manifest.
In fact, that’s not what happens here at all, and this section examines why not.
Tony Hoare once famously said, “Pointers are like jumps”4 . What we are seeking here is the reverse of
this phenomenon: what is the goto-like construct that corresponds to a dereference in a data structure? The
answer was given by Guy Steele: the tail call.
Armed with this insight, we can now reexamine the code. Studying the output of compilation, or the
macro, we notice that the conditional dispatcher invokes the function corresponding to the next state on the
rest of the stream—but does not touch the return value. This is no accident: the macro has been carefully
written to only make tail calls.5
In other words, the state transition is hardly more complicated than finding the next state (which is
statically determinate, since the compiler knows the location of all the local functions) and executing the
code that resides there. Indeed, the code generated from this Scheme source looks an awful lot like the
automaton representation we discussed at the beginning of section 37.3: random access for the procedures,
references for state transformation, and some appropriately efficient implementation of the conditional.
The moral of this story is that we get the same representation we would have had to carefully craft by
hand virtually for free from the compiler. In other words, languages represent the ultimate form of reuse,
because we get to reuse everything from the mathematical (semantics) to the practical (libraries), as well as
decades of research and toil in compiler construction.
Tail Calls versus Tail Recursion
This example should help demonstrate the often-confused difference between tail calls and tail recursion.
Many books discuss tail recursion, which is a special case where a function makes tail calls to itself. They
point out that, because implementations must optimize these calls, using recursion to encode a loop results
in an implementation that is really no less efficient than using a looping construct. They use this to justify,
in terms of efficiency, the use of recursion for looping.
These descriptions unfortunately tell only half the story. While their comments on using recursion for
looping are true, they obscure the subtlety and importance of optimizing all tail calls, which permit a family
of functions to invoke one another without experiencing penalty. This leaves programmers free to write read-
able programs without paying a performance penalty—a rare “sweet spot” in the readability-performance
trade-off. Traditional languages that offer only looping constructs and no tail calls force programmers to
artificially combine procedures, or pay via performance.
The functions generated by the automaton macro are a good illustration of this. If the implementation
did not perform tail-call optimization but the programmer needed that level of performance, the macro would
be forced to somehow combine all the three functions into a single one that could then employ a looping
4 The context for the quote is pejorative: “Pointers are like jumps, leading wildly from one part of the data structure to another.
Their introduction into high-level languages has been a step backwards from which we may never recover.”
5 Even if the code did need to perform some operation with the result, it is often easy in practice to convert the calls to tail-calls
using accumulators. In general, as we have seen, the conversion to continuation-passing style converts all calls to tail calls.
construct. This leads to an unnatural mangling of code, making the macro much harder to develop and
maintain.
37.4 Other Uses

Scheme macros can do many more things. datum→syntax-object can be used to manufacture identifiers
from syntax supplied to the macro. Macros can also define other macros! You will find such examples
as you begin to employ macros in your work. Books such as Kent Dybvig’s The Scheme Programming
Language and Paul Graham’s On Lisp elaborate on these programming styles.
37.5 Perspective
We have now seen several examples of Scheme’s macro system at work. In the process, we have seen how
features that would otherwise seem orthogonal, such as macros, first-class procedures and tail-calls, are in
fact intimately wedded together; in particular, the absence of the latter two would greatly complicate use of
the former. In this sense, the language’s design represents a particularly subtle, maximal point in the design
space of languages: removing any feature would greatly compromise what’s left, while what is present is an
especially good notation for describing algorithms.
Part XIII
What’s Next?
349
Chapter 38
Programming Interactive Systems
Consider writing a program that solves a graph algorithm such as finding a minimum spanning tree. Your
implementation of the algorithm may suffer from several problems: even if the program runs to completion,
the output it produces may not be minimal, it may not span, and it may not even be a tree! While debugging
such a program by manual inspection of intermediate data structures can be onerous, often (especially if
the bug is in one of the latter two cases) it is easy to spot the problem visually. Suppose, however, you
were writing a library that was meant to be used in textual applications. It would not make sense to add
graph-drawing code to your library program. But how can you instrument your program from the “outside”
to add this functionality?
In principle, adding this graph-drawing instrumentation is something a debugger should support well. It
is, however, a very domain-specific instrumentation; it doesn’t make sense for a general-purpose debugger
to offer the ability to add graph-drawing code, because this is a property many applications don’t require. In
general, a debugger offers only a general, domain-independent view of a program’s data, while we would
often like to specialize this view to a domain-specific one. Ideally, that is, we would like to make the
debugger scriptable, so each application can install its own programs that run atop a library of debugging
primitives.
In most programming languages, programs expect to be in control of their execution. That is, the
beginning of the program initiates a process of execution, and the end of the program terminates it. The
program may invoke external agents (such as functions in libraries, procedures written in other languages
through a foreign-function interface, or even a method residing on a remote host through a remote procedure
call mechanism), but expect at some point (either synchronously or asynchronously) to get a response that
resumes the computation.
In a debugger scripting language, in contrast, the script is most certainly not in control of the computa-
tion. Two other programs—the target program being debugged, and the debugger itself—control the flow
of computation. The script should have the ability to install an event generator that triggers whenever some
event of interest—such as a method invocation, a variable mutation, a thread spawn, and so forth—happens
in the target program. The script only runs in response to these events. Furthermore, the script does not
return any answers to the target program (which is essentially unaware of the script’s existence) or to the
debugger (which, being general-purpose, may not know what to do with it), but rather preserves information
for its own (possible) subsequent execution.
351
352 CHAPTER 38. PROGRAMMING INTERACTIVE SYSTEMS
In fact, most modern applications share this characteristic: the application must lay dormant waiting for
some behavior from an external source of events. We have already seen this property in the case of Web
applications; programmers needed to invert the structure of their source to expose the points of resumption,
and manually represent the rest of the computation at each point. To remedy this problem, we devised
the send/suspend primitive, and showed that better language support makes Web applications far easier to
construct and maintain.
The Web is a simple interactive medium in the sense that there is only one kind of behavior available
to a user, namely clicking on a link or on a button. Other user actions are masked by the browser and not
conveyed to the Web application. In contrast, a modern GUI application—such as that Web browser itself—
must respond to multiple kinds of behavior, including keystrokes, mouse button pushes and mouse clicks.
Applications may also need to be sensitive to the receipt of network packets or the ticking of the system
clock.
Confronted with the challenge of building complex interactive systems, most programming languages
have taken a rather weak route. Traditional programming languages offer the callback as the primary method
of registering code that will respond to events. A callback is essentially a procedure that consumes arguments
corresponding to the event information—such as which key was pressed—and returns. . . nothing.
Why does a callback not return anything? A callback is a piece of application code that is “installed”
in the underlying system layer, to be invoked when that layer receives an event that the application must
process. That is, the system code invokes the callback. But it is the application that must process the event,
so there is no meaningful information that the callback can return to the system. This explains the return
type.
What do we know about procedures that do not return a value? To be of any use at all, they must
accomplish their action by using side-effects (such as mutation). This explains why side-effects, especially
mutation, are so prevalent in GUI applications. But side-effects, and procedures that don’t return any value,
are difficult to compose, and make it harder to reason about a program’s behavior. Is it surprising that GUIs
are difficult to develop correctly and to test for the absence of errors?
Many programmers see the current design of GUI libraries, with their plethora of callbacks, as a fixture
in the design space. We should know better. The callback-based solution should be especially familiar to us
from our Web programming experience: the continuation generated by CPS a procedure that does not return
any meaningful value (in fact, it doesn’t return at all), and represents the rest of the computation. As such,
it is a special kind of callback. If through better language design we were able to improve the state of Web
programming, perhaps better language design can improve GUI programming (and debugger scripting, and
. . . ) also?
This is precisely the kind of challenge that leads to research in programming languages. For instance,
the FrTime1 language in DrScheme is designed to make developing interactive applications, from debugging
scripts to GUIs, more direct. Though FrTime programs superficially look like Scheme programs, they eval-
uate under a very different semantics: certain expressions are known to yield sequences of events, and any
computation that depends on it in turn gets recomputed every time a new event occurs. Thus, for instance,
the expression
(make-circle mouse-pos 10 ”blue”)
1 Pronounced “father time”.
353
automatically draws a blue circle of radius ten pixels wherever the mouse is currently centered, because
mouse-pos is a value in the language that updates whenever the mouse moves. Because it updates, every
expression that depends on it must also update to remain fresh. There is, therefore, no need for a loop to
ensure the circle is re-drawn: the “loop” is built into the language semantics! Note, of course, that there is
no need for a callback either.
This is just a miniscule example of how a deeper analysis of a problem can lead to better linguistic tools,
which can in turn make a problem much simpler than it was initially. Look up FrTime in DrScheme for
more on this fascinating language.
354 CHAPTER 38. PROGRAMMING INTERACTIVE SYSTEMS
Chapter 39
What Else is Next
At this point, it may be humbling to consider all the topics in programming languages that we haven’t cov-
ered: object-oriented programming, class-oriented programming and type systems for the same; threads;
module systems; middleware; and so on. And that’s just looking at what programmers routinely use in
practice. A bit further afield we find techniques such as mixins (parameterized classes), contracts and
aspect-oriented programming, which are rapidly gaining a foothold in industry as well. And then we get to
topics that are still very much in the realm of research, such as dependent types, type classes, delimited con-
trol, resource accounting or logic metaprogramming. We’ve applied programming language theory to one
important practical problem (Web programming), but numerous ones abound (such as the static validation
of spreadsheets).
The particular set of topics we’ve chosen to study in this book meet one of these critiera: they’re essential
knowledge before we can examine these other topics above, or they’re just plain fun. Fortunately, most topics
are both.
Go forth and populate the world with beautiful designs.
355
Index
BNF ,7 C, 30, 47, 49, 261, 263

CPS ,169, 352 C++, 263, 271
GUI applications, interactive, 352 C#, 271
cache, 78
abstract syntax, 6 cache, invalidation of, 80
activation record, 199 caching computation in lazy evaluation, 78
actual parameter, 28 call-by-reference, 136
address (see “location”), 122 call-by-value, 136
AE, 6 callback, 50, 352
Algol 60, 7, 79 capture, intentional in macros, 331
alias, 137 cast, type, 239, 240
aliasing, 138 Char, Haskell type, 60
antecedent, 231
Church numerals, 223
application, function, 27
Church, Alonzo, 223
assume-guarantee, 245, 254
Clemens, Franz, 295
automated memory management, 211
closure, 46
axiom, 246
closure, expression, 74
axiomatic semantics, 4
Common Lisp, 30, 52
Backus-Naur Form, 7 compiler, 172
Bangalore, India, 163 compiler optimization, 80
Barland, Ian, 261 compiler, for the Web, 169
Barwise, Jon, 295 concrete syntax, 6
bash, 69 consequent, 231
BFS, 308 constraint programming, 311
big-step operational semantics, 233 container type, 133
binding form, judgment as, 232 container types, 281
binding instance, 17 continuation, 177
Bool, Haskell type, 60 continuation, in Prolog, 309
bound instance, 17 continuation-passing style, 169
Bourne Again Shell, 69 control stack, 169
box, 98 Curry, Haskell, 59
breadth-first search, 308 cut, in Prolog, 311
Brentano, Franz, 295 cycle, 65
Budapest, Hungary, 163 cyclic, 93
356
INDEX 357
cyclicity, 94 filtering, as example of static scope, 48

Fischer CPS Transformation, 174
database, 149 Fischer, Michael, 174
databases, and Prolog, 307 fixed-point, 94, 228
de Bruijn index, 24 fixed-point type, 282
de Bruijn, Nicolaas, 24 fixed-points, using tabling, 311
declaration, function, 27 Flanagan, Cormac, 53
denotational semantics, 4 formal parameter, 28
depth-first search, 308 France, Tour de, 65
DFS, 308 free instance, 18
differentiation, 49 FrTime programming language, 352
disjunctive normal form, 307 function, 27
Disjunctive Normal Form, Prolog rules in, 296 function application, 27
DNF, 307 function declaration, 27
DrScheme, 8 function definition, 27
dyanmic, 23 function invocation, 27
dynamic, 172 function, partial, 105
dynamic extent, 129
dynamic loading, 212 Gödel, Kurt, 14
dynamic scope, 36, 82, 90 garbage collection, 211
genealogy, academic, 295
eager, 23 Ginsberg, Allen, 261
eager evaluation, 82 goal, 307
eager evaluation, testing for, 64 goto, 200
elaboration, type, 270 goto statement, 186
encapsulation, 192 Greenberg, Michael, 246
environment, 53
environment, type, 243 Halting Problem, 239, 253
equational reasoning, 80, 81 Haskell, 59, 256
escaper, 179 Haskell Prelude, 59, 61
evaluation regime, 82 heap, 209
exception, 262 hidden field, HTML, 149
exceptions, 179 Hindley, J. Roger, 283
existential type, 257 Hindley-Milner, 283
explicit polymorphism, 269 homogenous lists, 61
expression closure, 74 Houston, TX, USA, 163
extent, 129 hunky-dory, 166
hygiene, in macros, 327
F1WAE, 27
fact, 295 identifier, 15
Feferman, Solomon, 295 index, de Bruijn, 24
Felleisen, Matthias, 3 infix notation, 5
filter, 61 inlining, abuse of macros, 334
358 INDEX
instance, binding, 17 map, 62

instance, bound, 17 memoization, 81, 311
instance, free, 18 memory address (see “memory location”), 122
Int, Haskell type, 60 memory location, 122
interaction, Web browser, 148 meta interpreter, 107
interpreter, 4, 28 meta-circular interpreter, 108
interpreter semantics, 4 meta-representation, 237
interpreter, meta, 107 metavariables, 245
interpreter, meta-circular, 108 Milner, Robin, 262, 283
interpreter, syntactic, 107 ML, 255, 256, 262, 289
invalidating cached values, 80 module system, 257
invocation, function, 27 monomorphism, 274
iterator, 201 Montague, Richard, 298
Mostowski, Andrzej, 297
Java, 106, 255, 263, 273 mutability of environment, 98
Java 1.5, 271
mutation, 117
Java Servlet, 149
mutation, absence in Haskell, 80
judgment, 232
judgment, type, 243 namespace, 30
non-terminal, 7
keywords, in macros, 325
numeral, 220, 231
l-value, 134, 137, 140
lambda calculus, 217, 228 object system, 257, 258
lambda lifting, 153 object-verb, 219
last call optimization, 200 objects, 133
lazy, 23 occurs check in unification, 282
lazy evaluation, 82 occurs check, in Prolog, 311
lazy evaluation, and macros, 335 omega, 251
lazy evaluation, testing for, 64 operational semantics, 4, 233
Leśniewski, Stanisław, 295 optimization by compiler, 80
leakage, 212
length, 62 parameter, actual, 28
let-based polymorphism, 289 parameter, formal, 28
lexical scope, 129 parser, 6
lifting procedures, 153 partial function, 105
Lisp, 52 Perlis, Alan, 72
liveness of objects, 212 philosophy, 5
local, 111 polymorphic types, 63
location, 122 polymorphism, 63
logic variable, 295, 301, 309 polymorphism, explicit, 269
polymorphism, let-based, 289
macro, 172 polymorphism, subtype, 273
macros in Scheme, 319 postfix notation, 5
INDEX 359
prefix notation, 5 semantics, 4, 233

Prelude, Haskell, 59, 61 semantics, axiomatic, 4
primitive, in CPS transformation, 195 semantics, denotational, 4
principal type, 283 semantics, interpreter, 4
product, 61 semantics, operational, 4
production, 7 send/suspend, 352
Prolog, 295 shadowed binding, 18
provability, 211 shipping and handling, 159
Providence, RI, USA, 163 short-circuit evaluation, 325
short-circuited evaluation, 64
quicksort, 48 short-circuiting, 220
quote notation, 8 simply-typed lambda calculus, 252
SML, 197, 253
reachability, 212
SML module language, 253
read, 6
soundness of type system, 261
reading, 7
space safety, 213
reasoning, equational, 80
stack, 199
receiver, 150, 158
stack frame, 199
recursive types, 255
stack, control, 169
reduction regime, 23
state, 117
reference parameters, 138
stateful protocol, 148
referential transparency, 81, 211
stateless protocol, 148
reflection, 200
reification, 200 stateless protocol for the Web, 148
resources as types, 253 static, 23
RnRS, 256 static distance coordinates, 25
Robinson, Julia, 297 static scope, 36, 82
root set, 212 STL (Standard Template Library), 271
rule, 295 store, 122
store-passing style, 123, 126, 135
s-expression, 7 strict, 77
safety, 263 strictness, 76
satisfaction, in Prolog, 307, 308 strong normalization, 252
scanning, 7 style, continuation-passing, 169
Scheme, 52 sub-goal, 307
scope, 17, 82 substitution, 16, 280
scope, dynamic, 36 subtype polymorphism, 273
scope, static, 36 sum, 61
scripting language, 108, 109 swap, 138
scripting languages, 52 syntactic interpreter, 107
scripting, debugger, 351 syntactic representation, 237
search tree, 307 syntax, abstract, 6
self-application, 225, 251 syntax, concrete, 6
360 INDEX
syntax-case, 331 verb-object, 219

syntax-rules, 321
WAE, 16
tabling, in Prolog, 311 Web applications, interactive, 352
tail call, 200 Web compiler, 169
tail call optimization, 200 Web programs, 147
tail recursion, 201 with, 15
take, 65
Tarski, Alfred, 295 XML, 72
templates in C++, 319 Y combinator, 228
TFWAE, 244
threaded store, 129 zip, 67
threading, 129 zipWith, 67
thunk, 79
tokenizing, 7
Tour de France, 65
transitive closure, 154
Trendelenburg, Friedrich, 297
truth, 211
truth and provability, 211
Twardowski, Kazimierz, 295
type, 238, 239
type closure, 287
type conflict, 278
type elaboration, 270
type environment, 243
type error, 247
type inference, 63
type judgment, 243
type judgment tree, 246
type safety, 263
type soundness, 261
type system, 238, 243
type variable, 63, 269
types and correctness, 240
undecidability, 199
unification, 280
unification, in Prolog, 310
union type, 250
variable, 15, 133

variable, logic, 295

Programming Languages: Application and Interpretation

Uploaded by

Programming Languages: Application and Interpretation

Uploaded by

Programming Languages:

Application and Interpretation

This work is licensed under the

This book is available free-of-cost from the author’s Web site.

Background and Prerequisite

In addition, Michael Greenberg instructed me in the rudiments of classifying flora.

5.3 Fixing the Interpreter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

7 Programming with Laziness 59

11 Representation Choices 105

12 Church and State 117

13 Mutable Data Structures 119

VII Continuations 145

15 Some Problems with Web Programs 147

16 The Structure of Web Programs 149

17 More Web Transformation 157

17.6 Perspective on the Web Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166

18 Conversion into Continuation-Passing Style 169

19 Programming with Continuations 177

20 Implementing Continuations 193

VIII Memory Management 207

21 Automatic Memory Management 209

22 Shrinking the Language 217

25 Type Judgments 243

26 Typing Control 249

27 Typing Data 255

28 Type Soundness 261

29 Explicit Polymorphism 265

30 Type Inference 273

30.6 Underconstrained Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282

31 Implicit Polymorphism 285

XI Programming by Searching 291

33 Programming in Prolog 295

34 Implementing Prolog 307

XII Domain-Specific Languages and Metaprogramming 313

35 Domain-Specific Languages 315

36 Macros as Compilers 319

36.1.3 Example: Nested Local Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . 323

37 Macros and their Impact on Language Design 337

XIII What’s Next? 349

38 Programming Interactive Systems 351

39 What Else is Next 355

• some behavior associated with each syntax,

• numerous useful libraries, and

• a collection of idioms that programmers of that language use.

1.1 Modeling Meaning

1.2 Modeling Syntax

Irrespective of whether we write

Indeed, each of these notations is in use by at least one programming language.

1.3 A Primer on Parsers

returns (list 1 ’a), to

returns (list ’+ 3 4), and to

returns (list ’+ (list ’- 3 4) 7).

<AE> ::= <num>

(define (parse sexp)

Language: PLAI - Advanced Student.

Language: PLAI - Advanced Student.

> (parse ’{+ {- 3 4} 7})

1.4 Primus Inter Parsers

(mismatched delimiters) but permits both of

{with {x {+ 5 5}} {+ x x}}

We will want this to evaluate to 20. Here’s a more elaborate example:

<WAE> ::= <num>

3.1 Defining Substitution

f (12, 1) = 123 + 13 = 1728 + 1 = 1729

f (10, 9) = 103 + 93 = 1000 + 729 = 1729 3

Nevertheless, it’s a good idea to pin down this operation precisely.

Beginning with the program

{with {x 5} {+ x {with {x 3} 10}}}

The rules reduce this to

{with {x 5} {+ 5 {with {5 3} 10}}}

{with {x 5} {+ x {with {x 3} 10}}}

{with {x 5} {+ 5 {with {x 3} 10}}}

Let’s consider a closely related expression:

{with {x 5} {+ x {with {x 3} x}}}

{with {x 5} {+ 5 {with {x 3} 5}}}

which, when evaluated, yields 10.