nuclear physics

CONCEPTS OF PROGRAMMING LANGUAGES TWELFTH EDITION

CONCEPTS OF PROGRAMMING LANGUAGES TWELFTH EDITION

ROBERT W. SEBESTA

University of Colorado at Colorado Springs

330 Hudson Street, NY NY 10013

Senior Vice President, Courseware Portfolio Management: Marcia Horton

Director, Portfolio Management: Engineering, Computer Science & Global Editions: Julian Partridge

Specialist, Higher Ed Portfolio Management: Matt Goldstein

Portfolio Management Assistant: Meghan Jacoby

Managing Content Producer: Scott Disanno

Content Producer: Carole Snyder

Web Developer: Steve Wright

Rights and Permissions Manager: Ben Ferrini

Manufacturing Buyer, Higher Ed, Lake Side ​Communications Inc (LSC): Maura Zaldivar-Garcia

Inventory Manager: Ann Lam

Product Marketing Manager: Yvonne Vannatta

Field Marketing Manager: Demetrius Hall

Marketing Assistant: Jon Bryant

Cover Designer: Joyce Wells, jWellsDesign

Full-Service Project Manager: Prathiba Rajagopal, SPi Global

Composition: SPi Global

Copyright © 2019, 2016, 2013, 2010 Pearson Education, Inc. All rights reserved. Manufactured in the United States of America. This publication is protected by copyright, and permission should be obtained from the publisher prior to any prohibited reproduction, storage in a retrieval system, or

transmission in any form or by any means, electronic, mechanical, photocopying, recording, or likewise. For information regarding permissions, request forms and the appropriate contacts within the Pearson Education Global Rights & Permissions department, please visit http:// www.pearsoned.com/permissions.

Many of the designations by manufacturers and sellers to distinguish their products are claimed as trademarks. Where those designations appear in this book, and the publisher was aware of a trademark claim, the designations have been printed in initial caps or all caps.

The author and publisher of this book have used their best efforts in preparing this book. These efforts include the development, research, and testing of theories and programs to determine their effectiveness. The author and publisher make no warranty of any kind, expressed or implied, with regard to these programs or the documentation contained in this book. The author and publisher shall not be liable in any event for incidental or consequential damages with, or arising out of, the furnishing, performance, or use of these programs.

Library of Congress Cataloging-in-Publication Data

Names: Sebesta, Robert W., author.

Title: Concepts of programming languages / Robert W. Sebesta, University of Colorado at Colorado Springs.

Description: Twelfth edition. | Pearson, [2019]

Identifiers: LCCN 2017059077| ISBN 9780134997186 (alk. paper) | ISBN 0134997182 (alk. paper)

Subjects: LCSH: Programming languages (Electronic computers)

Classification: LCC QA76.7 .S43 2019 | DDC 005.13--dc23 LC record available at https://lccn.loc.gov/2017059077

1 18

ISBN 10: 0-13-499718-2

ISBN 13: 978-0-13-499718-6

Changes for the Twelfth Edition of Concepts of Programming Languages

Chapter 2: Added Section 2.16.4 A Replacement for Objective-C: Swift

Added Section 2.16.5 Another Related Language: Delphi

Deleted Section 2.18.6 Origins and Characteristics of Lua

Chapter 5: Rewrote several paragraphs in Section 5.5.3 to correct and clarify

Chapter 6: Added a paragraph to Section 6.3.2 to describe support for strings in Swift

Added a paragraph to Section 6.4.2 to describe support the enumeration types in Swift

Added a paragraph to Section 6.5.3 to describe support for arrays in Swift

Added a paragraph to Section 6.6.1 to describe support for associative arrays in Swift

Deleted the interview in Section 6.6.1

Added Section 6.12 Optional Types

Chapter 8: Added a design issue and a brief discussion of it to Section 8.3.1.1

Added several paragraphs to Section 8.3.4 that describe iterators in

Python

Chapter 9: Added a paragraph to Section 9.5.4 on Swift parameters

Chapter 11: Deleted Section 11.4.2 (Abstract Data Types in Objective- C)

Chapter 12: Deleted Section 12.4.5 (Objective-C)

Deleted Objective-C column from Table 12.1

Added a paragraph in the Summary on reflection

Preface

Changes for the Twelfth Edition The goals, overall structure, and approach of this twelfth edition of Concepts of Programming Languages remains the same as those of the eleven previous editions. The principal goals are to introduce the fundamental constructs of contemporary programming languages and to provide the reader with the tools necessary for the critical evaluation of existing and future programming languages. A secondary goal is to prepare the reader for the study of compiler design, by providing an in-depth discussion of programming language structures, presenting a formal method of describing syntax, and introducing approaches to lexical and syntax analysis.

The twelfth edition evolved from the eleventh through several different kinds of changes. To maintain the currency of the material, nearly all discussion of some programming languages, specifically Lua and Objective-C, has been removed. Material on the newer language, Swift, was added to several chapters.

In addition, a new section on optional types was added to Chapter 6. Material was added to Section 8.3.4 to describe iterators in Python. In numerous places in the manuscript small changes were made to correct and/or clarify the discussion.

The Vision This book describes the fundamental concepts of programming languages by discussing the design issues of the various language constructs, examining the design choices for these constructs in some of the most common languages, and critically comparing design alternatives.

Any serious study of programming languages requires an examination of some related topics, among which are formal methods of describing the syntax and semantics of programming languages, which are covered in Chapter 3. Also, implementation techniques for various language constructs must be considered: Lexical and syntax analysis are discussed in Chapter 4, and implementation of subprogram linkage is covered in Chapter 10. Implementation of some other language constructs is discussed in various other parts of the book.

The following paragraphs outline the contents of the twelfth edition.

Chapter Outlines Chapter 1 begins with a rationale for studying programming languages. It then discusses the criteria used for evaluating programming languages and language constructs. The primary influences on language design, common design trade-offs, and the basic approaches to implementation are also examined.

Chapter 2 outlines the evolution of the languages that are discussed in this book. Although no attempt is made to describe any language completely, the origins, purposes, and contributions of each are discussed. This historical overview is valuable, because it provides the background necessary to understanding the practical and theoretical basis for contemporary language design. It also motivates further study of language design and evaluation. Because none of the remainder of the book depends on Chapter 2, it can be read on its own, independent of the other chapters.

Chapter 3 describes the primary formal method for describing the syntax of programming language—BNF. This is followed by a description of attribute grammars, which describe both the syntax and static semantics of languages. The difficult task of semantic description is then explored, including brief introductions to the three most common methods: operational, denotational, and axiomatic semantics.

Chapter 4 introduces lexical and syntax analysis. This chapter is targeted to those Computer Science departments that no longer require a compiler design course in their curricula. Similar to Chapter 2, this chapter stands alone and can be studied independently of the rest of the book, except for Chapter 3, on which it depends.

Chapters 5 through 14 describe in detail the design issues for the primary constructs of programming languages. In each case, the design choices for several example languages are presented and evaluated. Specifically, Chapter 5 covers the many characteristics of variables, Chapter 6 covers data types, and Chapter 7 explains expressions and assignment statements. Chapter 8

describes control statements, and Chapters 9 and 10 discuss subprograms and their implementation. Chapter 11 examines data abstraction facilities. Chapter 12 provides an in-depth discussion of language features that support object- oriented programming (inheritance and dynamic method binding), Chapter 13 discusses concurrent program units, and Chapter 14 is about exception handling, along with a brief discussion of event handling.

Chapters 15 and 16 describe two of the most important alternative programming paradigms: functional programming and logic programming. However, some of the data structures and control constructs of functional programming languages are discussed in Chapters 6 and 8. Chapter 15 presents an introduction to Scheme, including descriptions of some of its primitive functions, special forms, and functional forms, as well as some examples of simple functions written in Scheme. Brief introductions to ML, Haskell, and F# are given to illustrate some different directions in functional language design. Chapter 16 introduces logic programming and the logic programming language, Prolog.

To the Instructor Chapters 1 and 3 are typically covered in detail, and though students find it interesting and beneficial reading, Chapter 2 receives little lecture time due to its lack of hard technical content. Because no material in subsequent chapters depends on Chapter 2, as noted earlier, it can be skipped entirely. If a course in compiler design is required, Chapter 4 is not covered.

Chapters 5 through 9 should be relatively easy for students with extensive programming experience in C++, Java, or C#. Chapters 10 through 14 are more challenging and require more detailed lectures.

Chapters 15 and 16 are entirely new to most students at the junior level. Ideally, language processors for Scheme and Prolog should be available for students required to learn the material in these chapters. Sufficient material is included to allow students to dabble with some simple programs.

Undergraduate courses will probably not be able to cover all of the material

in the last two chapters. Graduate courses, however, should be able to completely discuss the material in those chapters by skipping over some parts of the early chapters on imperative languages.

Supplemental Materials The following supplements are available to all readers of this book at www.pearson.com/cs-resources.

A set of lecture note slides. PowerPoint slides are available for each chapter in the book.

All of the figures from the book.

A companion Web site to the book is available at www.pearson.com/cs- resources. This site contains mini-manuals (approximately 100-page tutorials) on a handful of languages.

Solutions to many of the problem sets are available to qualified ​instructors in our Instructor Resource Center at www.pearson.com. Please ​contact your school’s Pearson representative or visit www.pearson.com to register.

Language Processor Availability Processors for and information about some of the programming languages discussed in this book can be found at the following Web sites:

JavaScript is included in virtually all browsers; PHP is included in virtually all Web servers.

All this information is also included on the companion Web site.

Acknowledgments The suggestions from outstanding reviewers contributed greatly to this book’s present form and contents. In alphabetical order, they are:

Numerous other people provided input for the previous editions of Concepts of Programming Languages at various stages of its development. All of their comments were useful and greatly appreciated. In alphabetical order, they are: Vicki Allan, Henry Bauer, Carter Bays, Manuel E. Bermudez, Peter Brouwer, Margaret Burnett, Paosheng Chang, Liang Cheng, John Crenshaw, Charles Dana, Barbara Ann Griem, Mary Lou Haag, John V. Harrison, Eileen Head, Ralph C. Hilzer, Eric Joanis, Leon Jololian, Hikyoo Koh, Jiang B. Liu, Meiliu Lu, Jon Mauney, Robert McCoard, Dennis L. Mumaugh, Michael G. Murphy, Andrew Oldroyd, Young Park, Rebecca Parsons, Steve J. Phelps, Jeffery Popyack, Steven Rapkin, Hamilton Richard, Tom Sager, Raghvinder ​- Sangwan, Joseph Schell, Sibylle Schupp, Mary Louise Soffa, Neelam Soundarajan, Ryan Stansifer, Steve Stevenson, Virginia Teller, Yang Wang, John M. Weiss, Franck Xia, and Salih Yurnas.

Matt Goldstein, Portfolio Management Specialist; Meghan Jacoby, ​Portfolio Management Assistant; Managing Content Producer, Scott Disanno; and Prathiba Rajagopal, all deserve my gratitude for their efforts to produce the twelfth edition both quickly and carefully.

About the Author Robert Sebesta is an Associate Professor Emeritus in the Computer Science Department at the University of Colorado–Colorado Springs. Professor Sebesta received a BS in applied mathematics from the University of Colorado in Boulder and MS and PhD degrees in computer science from Pennsylvania State University. He has taught computer science for more than 40 years. His professional interests are the design and evaluation of programming languages and Web programming.

Contents 1. Chapter 1 Preliminaries 1

1. 1.1 Reasons for Studying Concepts of Programming Languages 2

2. 1.2 Programming Domains 5

3. 1.3 Language Evaluation Criteria 7

4. 1.4 Influences on Language Design 17

5. 1.5 Language Categories 20

6. 1.6 Language Design Trade-Offs 21

7. 1.7 Implementation Methods 22

8. 1.8 Programming Environments 29

1. Summary • Review Questions • Problem Set 30

2. Chapter 2 Evolution of the Major Programming Languages 33

1. 2.1 Zuse’s Plankalkül 36

2. 2.2 Pseudocodes 37

3. 2.3 The IBM 704 and Fortran 40

4. 2.4 Functional Programming: Lisp 45

5. 2.5 The First Step Toward Sophistication: ALGOL 60 50

6. 2.6 Computerizing Business Records: COBOL 56

7. 2.7 The Beginnings of Timesharing: Basic 61

1. Interview: ALAN COOPER—User Design and Language Design 64

8. 2.8 Everything for Everybody: PL/I 66

9. 2.9 Two Early Dynamic Languages: APL and SNOBOL 69

10. 2.10 The Beginnings of Data Abstraction: SIMULA 67 70

11. 2.11 Orthogonal Design: ALGOL 68 71

12. 2.12 Some Early Descendants of the ALGOLs 73

13. 2.13 Programming Based on Logic: Prolog 77

14. 2.14 History’s Largest Design Effort: Ada 79

15. 2.15 Object-Oriented Programming: Smalltalk 83

16. 2.16 Combining Imperative and Object-Oriented Features: C++ 85

17. 2.17 An Imperative-Based Object-Oriented Language: Java 89

18. 2.18 Scripting Languages 92

19. 2.19 The Flagship .NET Language: C# 98

20. 2.20 Markup-Programming Hybrid Languages 100

1. Summary • Bibliographic Notes • Review Questions • Problem Set • Programming Exercises 102

3. Chapter 3 Describing Syntax and Semantics 109

1. 3.1 Introduction 110

2. 3.2 The General Problem of Describing Syntax 111

3. 3.3 Formal Methods of Describing Syntax 113

4. 3.4 Attribute Grammars 128

1. History Note 128

5. 3.5 Describing the Meanings of Programs: Dynamic Semantics 134

1. History Note 142

1. Summary • Bibliographic Notes • Review Questions • Problem Set 155

4. Chapter 4 Lexical and Syntax Analysis 161

1. 4.1 Introduction 162

2. 4.2 Lexical Analysis 163

3. 4.3 The Parsing Problem 171

4. 4.4 Recursive-Descent Parsing 175

5. 4.5 Bottom-Up Parsing 183

1. Summary • Review Questions • Problem Set • Programming Exercises 191

5. Chapter 5 Names, Bindings, and Scopes 197

1. 5.1 Introduction 198

2. 5.2 Names 199

1. History Note 199

3. 5.3 Variables 200

4. 5.4 The Concept of Binding 203

5. 5.5 Scope 211

6. 5.6 Scope and Lifetime 222

7. 5.7 Referencing Environments 223

8. 5.8 Named Constants 224

1. Summary • Review Questions • Problem Set • Programming Exercises 227

6. Chapter 6 Data Types 235

1. 6.1 Introduction 236

2. 6.2 Primitive Data Types 238

3. 6.3 Character String Types 242

1. History Note 243

4. 6.4 Enumeration Types 247

5. 6.5 Array Types 250

1. History Note 251

2. History Note 251

6. 6.6 Associative Arrays 261

7. 6.7 Record Types 263

8. 6.8 Tuple Types 266

9. 6.9 List Types 268

10. 6.10 Union Types 270

11. 6.11 Pointer and Reference Types 273

1. History Note 276

12. 6.12 Optional Types 285

13. 6.13 Type Checking 286

14. 6.14 Strong Typing 287

15. 6.15 Type Equivalence 288

16. 6.16 Theory and Data Types 292

1. Summary • Bibliographic Notes • Review Questions • Problem Set • Programming Exercises 294

7. Chapter 7 Expressions and Assignment Statements 301

1. 7.1 Introduction 302

2. 7.2 Arithmetic Expressions 302

3. 7.3 Overloaded Operators 311

4. 7.4 Type Conversions 313

1. History Note 315

5. 7.5 Relational and Boolean Expressions 316

1. History Note 316

6. 7.6 Short-Circuit Evaluation 318

7. 7.7 Assignment Statements 319

1. History Note 323

8. 7.8 Mixed-Mode Assignment 324

1. Summary • Review Questions • Problem Set • Programming Exercises 324

8. Chapter 8 Statement-Level Control Structures 329

1. 8.1 Introduction 330

2. 8.2 Selection Statements 332

3. 8.3 Iterative Statements 343

4. 8.4 Unconditional Branching 355

1. History Note 356

5. 8.5 Guarded Commands 356

6. 8.6 Conclusions 359

1. Summary • Review Questions • Problem Set • Programming Exercises 360

9. Chapter 9 Subprograms 365

1. 9.1 Introduction 366

2. 9.2 Fundamentals of Subprograms 366

3. 9.3 Design Issues for Subprograms 374

4. 9.4 Local Referencing Environments 375

5. 9.5 Parameter-Passing Methods 376

1. History Note 384

6. 9.6 Parameters That Are Subprograms 392

1. History Note 394

7. 9.7 Calling Subprograms Indirectly 394

8. 9.8 Design Issues for Functions 396

9. 9.9 Overloaded Subprograms 397

10. 9.10 Generic Subprograms 398

11. 9.11 User-Defined Overloaded Operators 404

12. 9.12 Closures 405

13. 9.13 Coroutines 407

1. Summary • Review Questions • Problem Set • Programming Exercises 410

10. Chapter 10 Implementing Subprograms 417

1. 10.1 The General Semantics of Calls and Returns 418

2. 10.2 Implementing “Simple” Subprograms 419

3. 10.3 Implementing Subprograms with Stack-Dynamic Local Variables 421

4. 10.4 Nested Subprograms 429

5. 10.5 Blocks 436

6. 10.6 Implementing Dynamic Scoping 437

1. Summary • Review Questions • Problem Set • Programming Exercises 441

11. Chapter 11 Abstract Data Types and Encapsulation Constructs 447

1. 11.1 The Concept of Abstraction 448

2. 11.2 Introduction to Data Abstraction 449

3. 11.3 Design Issues for Abstract Data Types 452

4. 11.4 Language Examples 453

1. Interview: BJARNE STROUSTRUP—C++: Its Birth, Its Ubiquitousness, and Common Criticisms 454

5. 11.5 Parameterized Abstract Data Types 466

6. 11.6 Encapsulation Constructs 471

7. 11.7 Naming Encapsulations 474

1. Summary • Review Questions • Problem Set • Programming Exercises 478

12. Chapter 12 Support for Object-Oriented Programming 483

1. 12.1 Introduction 484

2. 12.2 Object-Oriented Programming 485

3. 12.3 Design Issues for Object-Oriented Languages 489

4. 12.4 Support for Object-Oriented Programming in Specific Languages 494

1. Interview: BJARNE STROUSTRUP—On Paradigms and Better Programming 498

5. 12.5 Implementation of Object-Oriented Constructs 519

6. 12.6 Reflection 522

1. Summary • Review Questions • Problem Set • Programming Exercises 528

13. Chapter 13 Concurrency 533

1. 13.1 Introduction 534

2. 13.2 Introduction to Subprogram-Level Concurrency 539

3. 13.3 Semaphores 544

4. 13.4 Monitors 549

5. 13.5 Message Passing 551

6. 13.6 Ada Support for Concurrency 552

7. 13.7 Java Threads 560

8. 13.8 C# Threads 570

9. 13.9 Concurrency in Functional Languages 575

10. 13.10 Statement-Level Concurrency 578

1. Summary • Bibliographic Notes • Review Questions • Problem Set • Programming Exercises 580

14. Chapter 14 Exception Handling and Event Handling 587

1. 14.1 Introduction to Exception Handling 588

1. History Note 592

2. 14.2 Exception Handling in C++ 594

3. 14.3 Exception Handling in Java 598

4. 14.4 Exception Handling in Python and Ruby 605

5. 14.5 Introduction to Event Handling 608

6. 14.6 Event Handling with Java 609

7. 14.7 Event Handling in C# 613

1. Summary • Bibliographic Notes • Review Questions • Problem Set

• Programming Exercises 616

15. Chapter 15 Functional Programming Languages 623

1. 15.1 Introduction 624

2. 15.2 Mathematical Functions 625

3. 15.3 Fundamentals of Functional Programming Languages 628

4. 15.4 The First Functional Programming Language: Lisp 629

5. 15.5 An Introduction to Scheme 633

6. 15.6 Common Lisp 651

7. 15.7 ML 653

8. 15.8 Haskell 658

9. 15.9 F# 663

10. 15.10 Support for Functional Programming in Primarily Imperative Languages 666

11. 15.11 A Comparison of Functional and Imperative Languages 669

1. Summary • Bibliographic Notes • Review Questions • Problem Set • Programming Exercises 671

16. Chapter 16 Logic Programming Languages 679

1. 16.1 Introduction 680

2. 16.2 A Brief Introduction to Predicate Calculus 680

3. 16.3 Predicate Calculus and Proving Theorems 684

4. 16.4 An Overview of Logic Programming 686

5. 16.5 The Origins of Prolog 688

6. 16.6 The Basic Elements of Prolog 688

7. 16.7 Deficiencies of Prolog 703

8. 16.8 Applications of Logic Programming 709

1. Summary • Bibliographic Notes • Review Questions • Problem Set • Programming Exercises 710

1. Bibliography 715

2. Index 725

CONCEPTS OF PROGRAMMING LANGUAGES TWELFTH EDITION

1 Preliminaries 1. 1.1 Reasons for Studying Concepts of Programming Languages

2. 1.2 Programming Domains

3. 1.3 Language Evaluation Criteria

4. 1.4 Influences on Language Design

5. 1.5 Language Categories

6. 1.6 Language Design Trade-Offs

7. 1.7 Implementation Methods

8. 1.8 Programming Environments

Before we begin discussing the concepts of programming languages, we must consider a few preliminaries. First, we explain some reasons why computer science students and professional software developers should study general approaches to language design and evaluation. This discussion is especially valuable for those who believe that a working knowledge of one or two programming languages is sufficient for computer scientists. Then, we briefly describe the major programming domains. Next, because the book evaluates language constructs and features, we ​present a list of criteria that can serve as a basis for such judgments. Then, we discuss the two major influences on language design: machine architecture and program design methodologies. After that, we introduce the major categories of programming languages. Next, we describe a few of the most important trade-offs that must be ​- considered during language design.

Because this book is also about the implementation of programming ​- languages, this chapter includes an overview of the most common general approaches to ​implementation. Finally, we briefly describe a few examples of programming ​environments and discuss their impact on software production.

1.1 Reasons for Studying Concepts of Programming Languages It is natural for students to wonder how they will benefit from the study of programming language concepts. After all, many other topics in computer ​- science are worthy of serious study. In fact, many now believe that there are more important areas of computing for study than can be covered in a four- year college curriculum. The following is what we believe to be a compelling list of potential benefits of studying concepts of programming languages:

Increased capacity to express ideas. It is widely believed that the depth at which people can think is influenced by the expressive power of the language in which they communicate their thoughts. Those with only a weak understanding of natural language are limited in the complexity of their thoughts, particularly in depth of abstraction. In other words, it is difficult for people to conceptualize structures they cannot describe, verbally or in writing.

Programmers, in the process of developing software, are similarly constrained. The language in which they develop software places limits on the kinds of control structures, data structures, and abstractions they can use; thus, the forms of algorithms they can construct are likewise limited. Awareness of a wider variety of programming language features can reduce such limitations in software development. Programmers can increase the range of their software development thought processes by learning new language constructs.

It might be argued that learning the capabilities of other languages does not help a programmer who is forced to use a language that lacks those capabilities. That argument does not hold up, however, because often, ​- language constructs can be simulated in other languages that do not support those constructs directly. For example, a C (Harbison and Steele, 2002) programmer who had learned the structure and uses of associative arrays in Perl (Christianson et al., 2013) might design

structures that ​simulate associative arrays in that language. In other words, the study of programming language concepts builds an appreciation for valuable language features and constructs and encourages programmers to use them, even when the language they are using does not directly support such features and constructs.

Improved background for choosing appropriate languages. Some professional programmers have had little formal education in computer science; rather, they have developed their programming skills independently or through in-house training programs. Such training programs often limit instruction to one or two languages that are directly relevant to the current projects of the organization. Other programmers received their formal training years ago. The languages they learned then are no longer widely used, and many features now available in programming languages were not commonly known at the time. The result is that many programmers, when given a choice of languages for a new project, use the language with which they are most familiar, even if it is poorly suited for the project at hand. If these programmers were familiar with a wider range of languages and language constructs, they would be better able to choose the language with the features that best address the problem.

Some of the features of one language often can be simulated in another language. However, it is preferable to use a feature whose design has been integrated into a language than to use a simulation of that feature, which is often less elegant, more cumbersome, and less safe.

Increased ability to learn new languages. Computer programming is still a relatively young discipline, and design methodologies, software development tools, and programming languages are still in a state of continuous evolution. This makes software development an exciting profession, but it also means that continuous learning is essential. The process of learning a new programming language can be lengthy and difficult, especially for someone who is comfortable with only one or two languages and has never examined programming language concepts in general. Once a thorough understanding of the fundamental concepts of languages is acquired, it becomes far easier to see how these concepts

are incorporated into the design of the language being learned. For example, programmers who understand the concepts of object-oriented programming will have a much easier time learning Ruby (Thomas et al., 2013) than those who have never used those concepts.

The same phenomenon occurs in natural languages. The better you know the grammar of your native language, the easier it is to learn a second language. Furthermore, learning a second language has the benefit of teaching you more about your first language.

The TIOBE Programming Community issues an index (http:// www.tiobe.com/index.php/content/paperinfo/tpci/index.htm) that is an indicator of the relative popularity of programming languages. For example, according to the index, Java, C, C++ (Lippman et al., 2012), and C# (Albahari and Abrahari, 2012) were the four most popular languages in use in ​February 2017.1 However, dozens of other languages were widely used at the time. The index data also show that the distribution of usage of programming languages is always changing. The number of languages in use and the dynamic nature of the statistics imply that every software developer must be prepared to learn different languages.

1. Note that this index is only one measure of the popularity of programming languages, and its accuracy is not universally accepted.

Finally, it is essential that practicing programmers know the vocabulary and fundamental concepts of programming languages so they can read and understand programming language descriptions and evaluations, as well as promotional literature for languages and compilers. These are the sources of information needed in order to choose and learn a language.

Better understanding of the significance of implementation. In learning the concepts of programming languages, it is both interesting and necessary to touch on the implementation issues that affect those concepts. In some cases, an understanding of implementation issues leads to an understanding of why languages are designed the way they are. In turn, this knowledge leads to the ability to use a language more intelligently, as it was designed to be used. We can become better

programmers by understanding the choices among programming language constructs and the consequences of those choices.

Certain kinds of program bugs can be found and fixed only by a programmer who knows some related implementation details. Another benefit of understanding implementation issues is that it allows us to visualize how a computer executes various language constructs. In some cases, some knowledge of implementation issues provides hints about the relative ​efficiency of alternative constructs that may be chosen for a program. For example, programmers who know little about the complexity of the implementation of subprogram calls often do not realize that a small subprogram that is frequently called can be a highly inefficient design choice.

Because this book touches on only a few of the issues of implementation, the previous two paragraphs also serve well as rationale for studying compiler design.

Better use of languages that are already known. Most contemporary programming languages are large and complex. Accordingly, it is uncommon for a programmer to be familiar with and use all of the features of a language he or she uses. By studying the concepts of programming languages, programmers can learn about previously unknown and unused parts of the languages they already use and begin to use those features.

Overall advancement of computing. Finally, there is a global view of computing that can justify the study of programming language concepts. Although it is usually possible to determine why a particular programming language became popular, many believe, at least in retrospect, that the most popular languages are not always the best available. In some cases, it might be concluded that a language became widely used, at least in part, because those in positions to choose languages were not sufficiently familiar with programming language concepts.

For example, many people believe it would have been better if ALGOL 60 (Backus et al., 1963) had displaced Fortran (ISO/IEC 1539-1, 2010)

in the early 1960s, because it was more elegant and had much better control statements, among other reasons. That it did not, is due partly to the programmers and software development managers of that time, many of whom did not clearly understand the conceptual design of ALGOL 60. They found its description difficult to read (which it was) and even more difficult to understand. They did not appreciate the benefits of block structure, recursion, and well-structured control statements, so they failed to see the benefits of ALGOL 60 over Fortran.

Of course, many other factors contributed to the lack of acceptance of ALGOL 60, as we will see in Chapter 2. However, the fact that computer users were generally unaware of the benefits of the language played a ​significant role.

In general, if those who choose languages were well informed, perhaps better languages would eventually squeeze out poorer ones.

1.2 Programming Domains Computers have been applied to a myriad of different areas, from controlling nuclear power plants to providing video games in mobile phones. Because of this great diversity in computer use, programming languages with very different goals have been developed. In this section, we briefly discuss a few of the most common areas of computer applications and their associated languages.

1.2.1 Scientific Applications The first digital computers, which appeared in the late 1940s and early 1950s, were invented and used for scientific applications. Typically, the scientific applications of that time used relatively simple data structures, but required large numbers of floating-point arithmetic computations. The most common data structures were arrays and matrices; the most common control structures were counting loops and selections. The early high-level programming languages invented for scientific applications were designed to provide for those needs. Their competition was assembly language, so efficiency was a primary concern. The first language for scientific applications was Fortran. ALGOL 60 and most of its descendants were also intended to be used in this area, although they were designed to be used in related areas as well. For some scientific applications where efficiency is the primary concern, such as those that were common in the 1950s and 1960s, no subsequent language is significantly better than Fortran, which explains why Fortran is still used.

1.2.2 Business Applications The use of computers for business applications began in the 1950s. Special computers were developed for this purpose, along with special languages. The first successful high-level language for business was COBOL (ISO/IEC, 2002), the initial version of which appeared in 1960. It probably still is the

most ​commonly used language for these applications. Business languages are characterized by facilities for producing elaborate reports, precise ways of describing and storing decimal numbers and character data, and the ability to specify decimal arithmetic operations.

There have been few developments in business application languages outside the development and evolution of COBOL. Therefore, this book includes only limited discussions of the structures in COBOL.

1.2.3 Artificial Intelligence Artificial intelligence (AI) is a broad area of computer applications characterized by the use of symbolic rather than numeric computations. Symbolic computation means that symbols, consisting of names rather than numbers, are manipulated. Also, symbolic computation is more conveniently done with linked lists of data rather than arrays. This kind of programming sometimes requires more flexibility than other programming domains. For example, in some AI applications the ability to create and execute code segments during execution is convenient.

The first widely used programming language developed for AI applications was the functional language Lisp (McCarthy et al., 1965), which appeared in 1959. Most AI applications developed prior to 1990 were written in Lisp or one of its close relatives. During the early 1970s, however, an alternative approach to some of these applications appeared—logic programming using the Prolog (Clocksin and Mellish, 2013) language. More recently, some AI applications have been written in systems languages such as Python (Lutz, 2013). Scheme (Dybvig, 2011), a dialect of Lisp, and Prolog are introduced in Chapters 15 and 16, respectively.

1.2.4 Web Software The World Wide Web is supported by an eclectic collection of languages, ranging from markup languages, such as HTML, which is not a programming

language, to general-purpose programming languages, such as Java. Because of the pervasive need for dynamic Web content, some computation capability is often included in the technology of content presentation. This functionality can be provided by embedding programming code in an HTML document. Such code is often in the form of a scripting language, such as JavaScript (Flanagan, 2011) or PHP (Tatroe et al., 2013). There are also some markup- like languages that have been extended to include constructs that control document processing, which are discussed in Section 1.5 and in Chapter 2.

1.3 Language Evaluation Criteria As noted previously, the purpose of this book is to examine carefully the underlying concepts of the various constructs and capabilities of programming languages. We will also evaluate these features, focusing on their impact on the software development process, including maintenance. To do this, we need a set of evaluation criteria. Such a list of criteria is necessarily controversial, because it is difficult to get even two computer scientists to agree on the value of some given language characteristic relative to others. In spite of these differences, most would agree that the criteria discussed in the following subsections are important.

Some of the characteristics that influence three of the four most important of these criteria are shown in Table 1.1, and the criteria themselves are discussed in the following sections.2 Note that only the most important characteristics are included in the table, mirroring the discussion in the following subsections. One could probably make the case that if one considered less important characteristics, virtually all table positions could include “bullets.”

2. The fourth primary criterion is cost, which is not included in the table because it is only slightly related to the other criteria and the characteristics that influence them.

Table 1.1 Language evaluation criteria and the characteristics that affect them

Note that some of these characteristics are broad and somewhat vague, such as writability, whereas others are specific language constructs, such as exception handling. Furthermore, although the discussion might seem to imply that the criteria have equal importance, that implication is not intended, and it is clearly not the case.

1.3.1 Readability One of the most important criteria for judging a programming language is the ease with which programs can be read and understood. Before 1970, software development was largely thought of in terms of writing code. The primary positive characteristic of programming languages was efficiency. Language constructs were designed more from the point of view of the computer than of the computer users. In the 1970s, however, the software life-cycle concept (Booch, 1987) was developed; coding was relegated to a much smaller role, and maintenance was recognized as a major part of the cycle, particularly in terms of cost. Because ease of maintenance is determined in large part by the readability of programs, readability became an important measure of the quality of programs and programming languages. This was an important juncture in the evolution of programming languages. There was a distinct crossover from a focus on machine orientation to a focus on human orientation.

Readability must be considered in the context of the problem domain. For example, if a program that describes a computation is written in a language not designed for such use, the program may be unnatural and convoluted, making it unusually difficult to read.

The following subsections describe characteristics that contribute to the readability of a programming language.

1.3.1.1 Overall Simplicity The overall simplicity of a programming language strongly affects its readability. A language with a large number of basic constructs is more difficult to learn than one with a smaller number. Programmers who must use a large language often learn a subset of the language and ignore its other features. This learning pattern is sometimes used to excuse the large number of language constructs, but that argument is not valid. Readability problems occur whenever the program’s author has learned a different subset from that subset with which the reader is familiar.

A second complicating characteristic of a programming language is feature multiplicity—that is, having more than one way to accomplish a particular operation. For example, in Java, a user can increment a simple integer variable in four different ways:

count = count + 1

count += 1

count++

++count

Although the last two statements have slightly different meanings from each other and from the others in some contexts, all of them have the same meaning when used as stand-alone expressions. These variations are discussed in Chapter 7.

A third potential problem is operator overloading, in which a single operator symbol has more than one meaning. Although this is often useful, it can lead to reduced readability if users are allowed to create their own

overloading and do not do it sensibly. For example, it is clearly acceptable to overload + to use it for both integer and floating-point addition. In fact, this overloading simplifies a language by reducing the number of operators. However, suppose the programmer defined + used between single- dimensioned array operands to mean the sum of all elements of both arrays. Because the usual meaning of vector addition is quite different from this, this unusual meaning could confuse both the author and the program’s readers. An even more extreme example of program confusion would be a user defining + between two vector operands to mean the difference between their respective first elements. Operator overloading is further discussed in Chapter 7.

Simplicity in languages can, of course, be carried too far. For example, the form and meaning of most assembly language statements are models of simplicity, as you can see when you consider the statements that appear in the next section. This very simplicity, however, makes assembly language programs less readable. Because they lack more complex control statements, program structure is less obvious; because the statements are simple, far more of them are required than in equivalent programs in a high-level language. These same arguments apply to the less extreme case of high-level languages with inadequate control and data-structuring constructs.

1.3.1.2 Orthogonality Orthogonality in a programming language means that a relatively small set of primitive constructs can be combined in a relatively small number of ways to build the control and data structures of the language. Furthermore, every possible combination of primitives is legal and meaningful. For example, consider data types. Suppose a language has four primitive data types (integer, float, double, and character) and two type operators (array and pointer). If the two type operators can be applied to themselves and the four primitive data types, a large number of data structures can be defined.

The meaning of an orthogonal language feature is independent of the context of its appearance in a program. (The word orthogonal comes from the mathematical concept of orthogonal vectors, which are independent of each

other.) Orthogonality follows from a symmetry of relationships among primitives. A lack of orthogonality leads to exceptions to the rules of the language. For example, in a programming language that supports pointers, it should be possible to define a pointer to point to any specific type defined in the language. However, if pointers are not allowed to point to arrays, many potentially useful user-defined data structures cannot be defined.

We can illustrate the use of orthogonality as a design concept by comparing one aspect of the assembly languages of the IBM mainframe computers and the VAX series of minicomputers. We consider a single simple situation: adding two 32-bit integer values that reside in either memory or registers and replacing one of the two values with the sum. The IBM mainframes have two instructions for this purpose, which have the forms

A Reg1, memory_cell

AR Reg1, Reg2

where Reg1 and Reg2 represent registers. The semantics of these are

Reg1 ← contents(Reg1) + contents(memory_cell)

Reg1 ← contents(Reg1) + contents(Reg2)

The VAX addition instruction for 32-bit integer values is

ADDL operand_1, operand_2

whose semantics is

operand_2 ← contents(operand_1) + contents(operand_2)

In this case, either operand can be a register or a memory cell.

The VAX instruction design is orthogonal in that a single instruction can use either registers or memory cells as its operands. There are two ways to specify operands, which can be combined in all possible ways. The IBM design is not orthogonal. Only two out of four operand combinations possibilities are legal, and the two require different instructions, A and AR. The IBM design is more restricted and therefore less writable. For example, you cannot add two values and store the sum in a memory location. Furthermore, the IBM design is more difficult to learn because of the

restrictions and the additional instruction.

Orthogonality is closely related to simplicity: The more orthogonal the design of a language, the fewer exceptions the language rules require. Fewer exceptions mean a higher degree of regularity in the design, which makes the language easier to learn, read, and understand. Anyone who has learned a significant part of the English language can testify to the difficulty of learning its many rule exceptions (for example, i before e except after c).

As examples of the lack of orthogonality in a high-level language, consider the following rules and exceptions in C. Although C has two kinds of structured data types, arrays and records (structs), records can be returned from functions but arrays cannot. A member of a structure can be any data type except void or a structure of the same type. An array element can be any data type except void or a function. Parameters are passed by value, unless they are arrays, in which case they are, in effect, passed by reference (because the appearance of an array name without a subscript in a C program is interpreted to be the address of the array’s first element).

As an example of context dependence, consider the C expression

a + b

This expression often means that the values of a and b are fetched and added together. However, if a happens to be a pointer and b is an integer, it affects the value of b. For example, if a points to a float value that occupies four bytes, then the value of b must be scaled—in this case multiplied by 4— before it is added to a. Therefore, the type of a affects the treatment of the value of b. The context of b affects its meaning.

Too much orthogonality can also cause problems. Perhaps the most orthogonal programming language is ALGOL 68 (van Wijngaarden et al., 1969). Every language construct in ALGOL 68 has a type, and there are no restrictions on those types. In addition, most constructs produce values. This combinational freedom allows extremely complex constructs. For example, a conditional can appear as the left side of an assignment, along with declarations and other assorted statements, as long as the result is an address. This extreme form of orthogonality leads to unnecessary complexity.

Furthermore, because languages require a large number of primitives, a high degree of orthogonality results in an explosion of combinations. So, even if the combinations are simple, their sheer numbers lead to complexity.

Simplicity in a language, therefore, is at least in part the result of a combination of a relatively small number of primitive constructs and a limited use of the concept of orthogonality.

Some believe that functional languages offer a good combination of simplicity and orthogonality. A functional language, such as Lisp, is one in which computations are made primarily by applying functions to given parameters. In contrast, in imperative languages such as C, C++, and Java, computations are usually specified with variables and assignment statements. Functional languages offer potentially the greatest overall simplicity, because they can accomplish everything with a single construct, the function call, which can be combined simply with other function calls. This simple elegance is the reason why some language researchers are attracted to functional languages as the primary alternative to complex nonfunctional languages such as Java. Other factors, the most important of which is probably efficiency, however, have prevented functional languages from becoming more widely used.

1.3.1.3 Data Types The presence of adequate facilities for defining data types and data structures in a language is another significant aid to readability. For example, suppose a numeric type is used for an indicator flag because there is no Boolean type in the language. In such a language, for example, in the original version of C, we might have an assignment such as the following:

timeout = 1

The meaning of this statement is unclear, whereas in a language that includes Boolean types, we would have the following:

timeout = true

The meaning of this statement is perfectly clear.

1.3.1.4 Syntax Design The syntax, or form, of the elements of a language has a significant effect on the readability of programs. Following are some examples of syntactic design choices that affect readability:

Special words. Program appearance and thus program readability are strongly influenced by the forms of a language’s special words (for example, while, class, and for). Especially important is the method of forming compound statements, or statement groups, primarily in control constructs. Some languages have used matching pairs of special words or symbols to form groups. C and its descendants use braces to specify compound statements. All of these languages have diminished readability because statement groups are always terminated in the same way, which makes it difficult to determine which group is being ended when an end or a right brace appears. Fortran 95 and Ada (ISO/IEC, 2014) make this clearer by using a distinct closing syntax for each type of statement group. For example, Ada uses end if to terminate a selection construct and end loop to terminate a loop construct. This is an example of the conflict between simplicity that results in fewer reserved words, as in Java, and the greater readability that can result from using more reserved words, as in Ada.

Another important issue is whether the special words of a language can be used as names for program variables. If so, then the resulting programs can be very confusing. For example, in Fortran 95, special words, such as Do and End, are legal variable names, so the appearance of these words in a program may or may not connote something special.

Form and meaning. Designing statements so that their appearance at least partially indicates their purpose is an obvious aid to readability. Semantics, or meaning, should follow directly from syntax, or form. In some cases, this principle is violated by two language constructs that are identical or similar in appearance but have different meanings,

depending perhaps on context. In C, for example, the meaning of the reserved word static depends on the context of its appearance. If used on the definition of a variable inside a function, it means the variable is created at compile time. If used on the definition of a variable that is outside all functions, then it means the variable is visible only in the file in which its definition appears; that is, it is not exported from that file.

One of the primary complaints about the shell commands of UNIX (Robbins, 2005) is that their appearance does no t always suggest their function. For example, the meaning of the UNIX command grep can be deciphered only through prior knowledge, or perhaps cleverness and familiarity with the UNIX editor, ed. The appearance of grep connotes nothing to UNIX beginners. (In ed, the command /regular_expression/ searches for a substring that matches the regular expression. Preceding this with g makes it a global command, specifying that the scope of the search is the whole file being edited. Following the command with p specifies that lines with the matching substring are to be printed. So g/regular_expression/p, which can obviously be abbreviated as grep, prints all lines in a file that contain substrings that match its operand, which is a regular expression.)

1.3.2 Writability Writability is a measure of how easily a language can be used to create programs for a chosen problem domain. Most of the language characteristics that affect readability also affect writability. This follows directly from the fact that the process of writing a program requires the programmer frequently to reread the part of the program that is already written.

As is the case with readability, writability must be considered in the context of the target problem domain of a language. It simply is not fair to compare the writability of two languages in the realm of a particular application when one was designed for that application and the other was not. For example, the writabilities of Visual BASIC (VB) (Halvorson, 2013) and C are dramatically different for creating a program that has a graphical user interface (GUI), for which VB is ideal. Their writabilities are also quite different for writing

systems programs, such as an operating system, for which C was designed.

The following subsections describe the most important characteristics influencing the writability of a language.

1.3.2.1 Simplicity and Orthogonality If a language has a large number of different constructs, some programmers who use the language might not be familiar with all of them. This situation can lead to a misuse of some features and a disuse of others that may be either more elegant or more efficient, or both, than those that are used. It may even be possible, as noted by Hoare (1973), to use unknown features accidentally, with bizarre results. Therefore, a smaller number of primitive constructs and a consistent set of rules for combining them (that is, orthogonality) is much better than simply having a large number of primitives. A programmer can design a solution to a complex problem after learning only a simple set of primitive constructs.

On the other hand, too much orthogonality can be a detriment to writability. Errors in programs can go undetected when nearly any combination of primitives is legal. This can lead to code absurdities that cannot be discovered by the compiler.

1.3.2.2 Expressivity Expressivity in a language can refer to several different characteristics. In a language such as APL (Gilman and Rose, 1983), it means that there are very powerful operators that allow a great deal of computation to be accomplished with a very small program. More commonly, it means that a language has relatively convenient, rather than cumbersome, ways of specifying computations. For example, in C, the notation count++ is more convenient and shorter than count = count + 1. Also, the and then Boolean operator in Ada is a convenient way of specifying short-circuit evaluation of a Boolean expression. The inclusion of the for statement in Java makes writing

counting loops easier than with the use of while, which is also possible. All of these increase the writability of a language.

1.3.3 Reliability A program is said to be reliable if it performs to its specifications under all conditions. The following subsections describe several language features that have a significant effect on the reliability of programs in a given language.

1.3.3.1 Type Checking Type checking is simply testing for type errors in a given program, either by the compiler or during program execution. Type checking is an ​important factor in language reliability. Because run-time type checking is expensive, compile-time type checking is more desirable. Furthermore, the earlier errors in programs are detected, the less expensive it is to make the required repairs. The design of Java requires checks of the types of nearly all variables and expressions at compile time. This virtually eliminates type errors at run time in Java programs. Types and type checking are discussed in depth in Chapter 6.

One example of how failure to type check, at either compile time or run time, has led to countless program errors is the use of subprogram parameters in the original C language (Kernighan and Ritchie, 1978). In this language, the type of an actual parameter in a function call was not checked to determine whether its type matched that of the corresponding formal parameter in the function. An int type variable could be used as an actual parameter in a call to a function that expected a float type as its formal parameter, and neither the compiler nor the run-time system would detect the inconsistency. For example, because the bit string that represents the integer 23 is essentially unrelated to the bit string that represents a floating-point 23, if an integer 23 is sent to a function that expects a floating-point parameter, any uses of the parameter in the function will produce nonsense. Furthermore, such problems are often difficult to diagnose.3 The current version of C has eliminated this

problem by requiring all parameters to be type checked. Subprograms and parameter-​passing techniques are discussed in Chapter 9.

3. In response to this and other similar problems, UNIX systems include a utility program named lint that checks C programs for such problems.

1.3.3.2 Exception Handling The ability of a program to intercept run-time errors (as well as other unusual conditions detectable by the program), take corrective measures, and then continue is an obvious aid to reliability. This language facility is called exception handling. Ada, C++, Java, and C# include extensive capabilities for exception handling, but such facilities are practically nonexistent in some widely used languages, for example, C. Exception handling is discussed in Chapter 14.

1.3.3.3 Aliasing Loosely defined, aliasing is having two or more distinct names in a program that can be used to access the same memory cell. It is now generally accepted that aliasing is a dangerous feature in a programming language. Most programming languages allow some kind of aliasing—for example, two pointers (or references) set to point to the same variable, which is possible in most languages. In such a program, the programmer must always remember that changing the value pointed to by one of the two changes the value referenced by the other. Some kinds of aliasing, as described in Chapters 5 and 9, can be prohibited by the design of a language.

In some languages, aliasing is used to overcome deficiencies in the language’s data abstraction facilities. Other languages greatly restrict aliasing to increase their reliability.

1.3.3.4 Readability and Writability

Both readability and writability influence reliability. A program written in a language that does not support natural ways to express the required algorithms will necessarily use unnatural approaches. Unnatural approaches are less likely to be correct for all possible situations. The easier a program is to write, the more likely it is to be correct.

Readability affects reliability in both the writing and maintenance phases of the life cycle. Programs that are difficult to read are difficult both to write and to modify.

1.3.4 Cost The total cost of a programming language is a function of many of its characteristics.

First, there is the cost of training programmers to use the language, which is a function of the simplicity and orthogonality of the language and the experience of the programmers. Although more powerful languages are not necessarily more difficult to learn, they often are.

Second, there is the cost of writing programs in the language. This is a function of the writability of the language, which depends in part on its closeness in purpose to the particular application. The original efforts to design and implement high-level languages were driven by the desire to lower the costs of creating software.

Both the cost of training programmers and the cost of writing programs in a language can be significantly reduced in a good programming environment. Programming environments are discussed in Section 1.8.

Third, the cost of executing programs written in a language is greatly influenced by that language’s design. A language that requires many run-time type checks will prohibit fast code execution, regardless of the quality of the compiler. Although execution efficiency was the foremost concern in the design of early languages, it is now considered to be less important.

A simple trade-off can be made between compilation cost and execution speed of the compiled code. Optimization is the name given to the collection of techniques that compilers may use to decrease the size and/or increase the execution speed of the code they produce. If little or no optimization is done, compilation can be done much faster than if a significant effort is made to produce optimized code. The choice between the two alternatives is influenced by the environment in which the compiler will be used. In a laboratory for beginning programming students, who often compile their programs many times during development but use little execution time (their programs are small and they must execute correctly only once), little or no optimization should be done. In a production environment, where compiled programs are executed many times after development, it is better to pay the extra cost to optimize the code.

Fourth, there is the cost of poor reliability. If the software fails in a critical system, such as a nuclear power plant or an X-ray machine for medical use, the cost could be very high. The failures of noncritical systems can also be very expensive in terms of lost future business or lawsuits over defective software systems.

The final consideration is the cost of maintaining programs, which includes both corrections and modifications to add new functionality. The cost of software maintenance depends on a number of language characteristics, primarily readability. Because maintenance is often done by individuals other than the original author of the software, poor readability can make the task extremely challenging.

The importance of software maintainability cannot be overstated. It has been estimated that for large software systems with relatively long lifetimes, maintenance costs can be as high as two to four times as much as development costs (Sommerville, 2010).

Of all the contributors to language costs, three are most important: program development, maintenance, and reliability. Because these are functions of writability and readability, these two evaluation criteria are, in turn, the most important.

Of course, a number of other criteria could be used for evaluating

programming languages. One example is portability, or the ease with which programs can be moved from one implementation to another. Portability is most strongly influenced by the degree of standardization of the language. Some languages are not standardized at all, making programs in these languages very difficult to move from one implementation to another. This problem is alleviated in some cases by the fact that implementations for some languages now have single sources. Standardization is a time-consuming and difficult process. A committee began work on producing a standard version of C++ in 1989. It was approved in 1998.

Generality (the applicability to a wide range of applications) and well-​- definedness (the completeness and precision of the language’s official defining document) are two other criteria.

Most criteria, particularly readability, writability, and reliability, are neither precisely defined nor accurately measurable. Nevertheless, they are useful concepts and they provide valuable insight into the design and evaluation of programming languages.

A final note on evaluation criteria: language design criteria are weighed differently from different perspectives. Language implementors are concerned primarily with the difficulty of implementing the constructs and features of the language. Language users are worried about writability first and readability later. Language designers are likely to emphasize elegance and the ability to attract widespread use. These characteristics often conflict with one another.

1.4 Influences on Language Design In addition to those factors described in Section 1.3, several other factors influence the basic design of programming languages. The most important of these are computer architecture and programming design methodologies.

1.4.1 Computer Architecture The basic architecture of computers has had a profound effect on language design. Most of the popular languages of the past 60 years have been designed around the prevalent computer architecture, called the von Neumann ​architecture, after one of its originators, John von Neumann (pronounced “von Noyman”). These languages are called imperative languages. In a von Neumann computer, both data and programs are stored in the same memory. The central processing unit (CPU), which executes instructions, is separate from the memory. Therefore, instructions and data must be transmitted, or piped, from memory to the CPU. Results of operations in the CPU must be moved back to memory. Nearly all digital computers built since the 1940s have been based on the von Neumann architecture. The overall structure of a von Neumann computer is shown in Figure 1.1.

Figure 1.1 The von Neumann computer architecture

Figure 1.1 Full Alternative Text

Because of the von Neumann architecture, the central features of imperative languages are variables, which model the memory cells; assignment statements, which are based on the piping operation; and the iterative form of repetition, which is the most efficient way to implement repetition on this architecture. Operands in expressions are piped from memory to the CPU, and the result of evaluating the expression is piped back to the memory cell represented by the left side of the assignment. Iteration is fast on von Neumann computers because instructions are stored in adjacent cells of memory and repeating the execution of a section of code requires only a branch instruction. This efficiency discourages the use of recursion for repetition, although recursion is sometimes more natural.

The execution of a machine code program on a von Neumann architecture computer occurs in a process called the fetch-execute cycle. As stated earlier, programs reside in memory but are executed in the CPU. Each instruction to be executed must be moved from memory to the processor. The address of the next instruction to be executed is maintained in a register called the program counter. The fetch-execute cycle can be simply described by the following algorithm:

initialize the program counter

repeat forever

  fetch the instruction pointed to by the program counter

  increment the program counter to point at the next instruction

  decode the instruction

  execute the instruction

end repeat

The “decode the instruction” step in the algorithm means the instruction is examined to determine what action it specifies. Program execution terminates when a stop instruction is encountered, although on an actual computer a stop instruction is rarely executed. Rather, control transfers from the operating system to a user program for its execution and then back to the operating system when the user program execution is complete. In a computer system in which more than one user program may be in memory at a given time, this process is far more complex.

As stated earlier, a functional, or applicative, language is one in which the primary means of computation is applying functions to given parameters. Programming can be done in a functional language without the kind of variables that are used in imperative languages, without assignment statements, and without iteration. Although many computer scientists have expounded on the myriad benefits of functional languages, such as Scheme, it is unlikely that they will displace the imperative languages until a non–von

Neumann computer is designed that allows efficient execution of programs in functional languages. Among those who have bemoaned this fact, perhaps the most eloquent was John Backus (1978), the principal designer of the original version of Fortran.

In spite of the fact that the structure of imperative programming languages is modeled on a machine architecture, rather than on the abilities and inclinations of the users of programming languages, some believe that using imperative languages is somehow more natural than using a functional language. So, these people believe that even if functional programs were as efficient as imperative programs, the use of imperative programming languages would still dominate.

1.4.2 Programming Design Methodologies The late 1960s and early 1970s brought an intense analysis, begun in large part by the structured-programming movement, of both the software development process and programming language design.

An important reason for this research was the shift in the major cost of computing from hardware to software, as hardware costs decreased and programmer costs increased. Increases in programmer productivity were relatively small. In addition, progressively larger and more complex problems were being solved by computers. Rather than simply solving sets of equations to simulate satellite tracks, as in the early 1960s, programs were being written for large and complex tasks, such as controlling large petroleum-refining facilities and providing worldwide airline reservation systems.

The new software development methodologies that emerged as a result of the research of the 1970s were called top-down design and stepwise refinement. The primary programming language deficiencies that were discovered were incompleteness of type checking and inadequacy of control statements (requiring the extensive use of gotos).

In the late 1970s, a shift from procedure-oriented to data-oriented ​program design methodologies began. Simply put, data-oriented methods emphasize data design, focusing on the use of abstract data types to solve problems.

For data abstraction to be used effectively in software system design, it must be supported by the languages used for implementation. The first language to provide even limited support for data abstraction was SIMULA 67 (Birtwistle et al., 1973), although that language certainly was not propelled to popularity because of it. The benefits of data abstraction were not widely recognized until the early 1970s. However, most languages designed since the late 1970s support data abstraction, which is discussed in detail in Chapter 11.

The latest step in the evolution of data-oriented software development, which began in the early 1980s, is object-oriented design. Object-oriented methodology begins with data abstraction, which encapsulates processing with data objects and controls access to data, and adds inheritance and dynamic method binding. Inheritance is a powerful concept that greatly enhances the potential reuse of existing software, thereby providing the possibility of significant increases in software development productivity. This is an important factor in the increase in popularity of object-oriented languages. Dynamic (run-time) method binding allows more flexible use of inheritance.

Object-oriented programming developed along with a language that supported its concepts: Smalltalk (Goldberg and Robson, 1989). Although Smalltalk never became as widely used as many other languages, support for object-oriented programming is now part of most popular imperative languages, including Java, C++, and C#. Object-oriented concepts have also found their way into functional programming in CLOS (Bobrow et al., 1988) and F# (Syme et al., 2010), as well as logic programming in Prolog++ (Moss, 1994). Language support for object-oriented programming is discussed in detail in Chapter 12.

Procedure-oriented programming is, in a sense, the opposite of data-​oriented programming. Although data-oriented methods now dominate software development, procedure-oriented methods have not been abandoned. On the contrary, in recent years, a good deal of research has occurred in ​procedure- oriented programming, especially in the area of concurrency. These research

efforts brought with them the need for language facilities for creating and controlling concurrent program units. Java and C# include such capabilities. Concurrency is discussed in detail in Chapter 13.

All of these evolutionary steps in software development methodologies led to new language constructs to support them.

1.5 Language Categories Programming languages are often categorized into four bins: imperative, functional, logic, and object oriented. However, we do not consider languages that support object-oriented programming to form a separate category of languages. We have described how the most popular languages that support object-oriented programming grew out of imperative languages. Although the object-oriented software development paradigm differs significantly from the procedure-oriented paradigm usually used with imperative languages, the extensions to an imperative language required to support object-oriented programming are not intensive. For example, the expressions, assignment statements, and control statements of C and Java are nearly identical. (On the other hand, the arrays, subprograms, and semantics of Java are very different from those of C.) Similar statements can be made for functional languages that support object-oriented programming.

Some authors refer to scripting languages as a separate category of programming languages. However, languages in this category are bound together more by their implementation method, partial or full interpretation, than by a common language design. The languages that are typically called scripting languages, among them Perl, JavaScript, and Ruby (Flanagan and Matsumoto, 2008), are imperative languages in every sense.

A logic programming language is an example of a rule-based language. In an imperative language, an algorithm is specified in great detail, and the specific order of execution of the instructions or statements must be included. In a rule-based language, however, rules are specified in no particular order, and the language implementation system must choose an order in which the rules are used to produce the desired result. This approach to software development is radically different from those used with the other two categories of languages and clearly requires a completely different kind of language. Prolog, the most commonly used logic programming language, and logic programming are discussed in Chapter 16.

In recent years, a new category of languages has emerged, the

markup/programming hybrid languages. Markup languages are not programming languages. For instance, HTML, the most widely used markup language, is used to specify the layout of information in Web documents. However, some programming capability has crept into some extensions to HTML and XML. Among these are the Java Server Pages Standard Tag Library (JSTL) and eXtensible Stylesheet Language Transformations (XSLT). Both of these are briefly introduced in Chapter 2. Those languages cannot be compared to any of the complete programming languages and therefore will not be discussed after Chapter 2.

1.6 Language Design Trade-Offs The programming language evaluation criteria described in Section 1.3 ​- provide a framework for language design. Unfortunately, that framework is self-​contradictory. In his insightful paper on language design, Hoare (1973) stated that “there are so many important but conflicting criteria, that their reconciliation and satisfaction is a major engineering task.”

Two criteria that conflict are reliability and cost of execution. For example, the Java language definition demands that all references to array elements be checked to ensure that the index or indices are in their legal ranges. This step adds a great deal to the cost of execution of Java programs that contain large numbers of references to array elements. C does not require index range checking, so C programs execute faster than semantically equivalent Java programs, although Java programs are more reliable. The designers of Java traded execution efficiency for reliability.

As another example of conflicting criteria that leads directly to design trade- offs, consider the case of APL. APL includes a powerful set of operators for array operands. Because of the large number of operators, a significant number of new symbols had to be included in APL to represent the operators. Also, many APL operators can be used in a single, long, complex expression. One result of this high degree of expressivity is that, for applications involving many array operations, APL is very writable. Indeed, a huge amount of computation can be specified in a very small program. Another result is that APL programs have very poor readability. A compact and concise expression has a certain mathematical beauty but it is difficult for anyone other than the programmer to understand. Well-known author Daniel McCracken (1970) once noted that it took him four hours to read and understand a four-line APL program. The designer of APL traded readability for writability.

The conflict between writability and reliability is a common one in ​language design. The pointers of C++ can be manipulated in a variety of ways, which supports highly flexible addressing of data. Because of the potential

reliability problems with pointers, they are not included in Java.

Examples of conflicts among language design (and evaluation) criteria abound; some are subtle, others are obvious. It is therefore clear that the task of choosing constructs and features when designing a programming language requires many compromises and trade-offs.

1.7 Implementation Methods As described in Section 1.4.1, two of the primary components of a computer are its internal memory and its processor. The internal memory is used to store programs and data. The processor is a collection of circuits that provides a realization of a set of primitive operations, or machine instructions, such as those for arithmetic and logic operations. In most computers, some of these instructions, which are sometimes called macroinstructions, are actually implemented with a set of instructions called microinstructions, which are defined at an even lower level. Because microinstructions are never seen by software, they will not be discussed further here.

The machine language of the computer is its set of instructions. In the absence of other supporting software, its own machine language is the only language that most hardware computers “understand.” Theoretically, a computer could be designed and built with a particular high-level language as its machine language, but it would be very complex and expensive. Furthermore, it would be highly inflexible, because it would be difficult (but not impossible) to use it with other high-level languages. The more practical machine design choice implements in hardware a very low-level language that provides the most commonly needed primitive operations and requires system software to create an interface to programs in higher-level languages.

A language implementation system cannot be the only software on a computer. Also required is a large collection of programs, called the operating system, which supplies higher-level primitives than those of the machine language. These primitives provide system resource management, input and output operations, a file management system, text and/or program editors, and a variety of other commonly needed functions. Because language implementation systems need many of the operating system facilities, they interface with the operating system rather than directly with the processor (in machine language).

The operating system and language implementations are layered over the

machine language interface of a computer. These layers can be thought of as virtual computers, providing interfaces to the user at higher levels. For example, an operating system and a C compiler provide a virtual C computer. With other compilers, a machine can become other kinds of virtual computers. Most computer systems provide several different virtual computers. User programs form another layer over the top of the layer of virtual computers. The layered view of a computer is shown in Figure 1.2.

Figure 1.2 Layered interface of

virtual computers, provided by a typical computer system

Figure 1.2 Full Alternative Text

The implementation systems of the first high-level programming languages, constructed in the late 1950s, were among the most complex software systems of that time. In the 1960s, intensive research efforts were made to understand and formalize the process of constructing these high-level language implementations. The greatest success of those efforts was in the area of syntax analysis, primarily because that part of the implementation process is an application of parts of automata theory and formal language theory that were then well understood.

1.7.1 Compilation Programming languages can be implemented by any of three general methods. At one extreme, programs can be translated into machine language, which can be executed directly on the computer. This method is called a compiler implementation and has the advantage of very fast program execution, once the translation process is complete. Most production implementations of languages, such as C, COBOL, and C++, are by compilers.

The language that a compiler translates is called the source language. The process of compilation and program execution takes place in several phases, the most important of which are shown in Figure 1.3.

Figure 1.3 The compilation process

Figure 1.3 Full Alternative Text

The lexical analyzer gathers the characters of the source program into ​lexical units. The lexical units of a program are identifiers, special words, ​operators, and punctuation symbols. The lexical analyzer ignores comments in the source program because the compiler has no use for them.

The syntax analyzer takes the lexical units from the lexical analyzer and uses them to construct hierarchical structures called parse trees. These parse trees represent the syntactic structure of the program. In many cases, no actual parse tree structure is constructed; rather, the information that would be required to build a tree is generated and used directly. Both lexical units and parse trees are discussed further in Chapter 3. Lexical analysis and syntax analysis, or parsing, are discussed in Chapter 4.

The intermediate code generator produces a program in a different ​language, at an intermediate level between the source program and the final output of the compiler: the machine language program.4 Intermediate ​languages sometimes look very much like assembly languages, and in fact, sometimes are actual assembly languages. In other cases, the intermediate code is at a level somewhat higher than an assembly language. The semantic analyzer is an integral part of the intermediate code generator. The semantic analyzer checks for errors, such as type errors, that are difficult, if not impossible, to detect during syntax analysis.

4. Note that the words program and code are often used interchangeably.

Optimization, which improves programs (usually in their intermediate code version) by making them smaller or faster or both. Because many kinds of optimization are difficult to do on machine language, most optimization is done on the intermediate code.

The code generator translates the optimized intermediate code version of the

program into an equivalent machine language program.

The symbol table serves as a database for the compilation process. The primary contents of the symbol table are the type and attribute information of each user-defined name in the program. This information is placed in the symbol table by the lexical and syntax analyzers and is used by the semantic analyzer and the code generator.

As stated previously, although the machine language generated by a compiler can be executed directly on the hardware, it must nearly always be run along with some other code. Most user programs also require programs from the operating system. Among the most common of these are programs for input and output. The compiler builds calls to required system programs when they are needed by the user program. Before the machine language programs produced by a compiler can be executed, the required programs from the operating system must be found and linked to the user program. The linking operation connects the user program to the system programs by placing the addresses of the entry points of the system programs in the calls to them in the user program. The user and system code together are sometimes called a load module, or executable image. The process of collecting system programs and linking them to user programs is called linking and loading, or sometimes just ​linking. It is accomplished by a systems program called a linker.

In addition to systems programs, user programs must often be linked to previously compiled programs that reside in libraries. So the linker not only links a given program to system programs, but also it may link it to other user or system-supplied programs.

The speed of the connection between a computer’s memory and its processor often determines the speed of the computer, because instructions often can be executed faster than they can be moved to the processor for execution. This connection is called the von Neumann bottleneck; it is the primary limiting factor in the speed of von Neumann architecture computers. The von Neumann bottleneck has been one of the primary motivations for the research and development of parallel computers.

1.7.2 Pure Interpretation Pure interpretation lies at the opposite end (from compilation) among implementation methods. With this approach, programs are interpreted by another program called an interpreter, with no translation whatever. The interpreter program acts as a software simulation of a machine whose fetch- execute cycle deals with high-level language program statements rather than machine instructions. This software simulation obviously provides a virtual machine for the language.

Pure interpretation has the advantage of allowing easy implementation of many source-level debugging operations, because all run-time error messages can refer to source-level units. For example, if an array index is found to be out of range, the error message can easily indicate the source line of the error and the name of the array. On the other hand, this method has the serious disadvantage that execution is 10 to 100 times slower than in compiled systems. The primary source of this slowness is the decoding of the high- level ​language statements, which are far more complex than machine language instructions (although there may be fewer statements than instructions in equivalent machine code). Furthermore, regardless of how many times a statement is executed, it must be decoded every time. Therefore, statement decoding, rather than the connection between the processor and memory, is the bottleneck of a pure interpreter.

Another disadvantage of pure interpretation is that it often requires more space. In addition to the source program, the symbol table must be present during interpretation. Furthermore, the source program may be stored in a form designed for easy access and modification rather than one that provides for minimal size.

Although some simple early languages of the 1960s (APL, SNOBOL (Griswold et al., 1971), and Lisp) were purely interpreted, by the 1980s, the approach was rarely used on high-level languages. However, in recent years, pure interpretation has made a significant comeback with some Web scripting languages, such as JavaScript and PHP, which are now widely used. The process of pure interpretation is shown in Figure 1.4.

Figure 1.4 Pure interpretation

1.7.3 Hybrid Implementation

Systems Some language implementation systems are a compromise between compilers and pure interpreters; they translate high-level language programs to an intermediate language designed to allow easy interpretation. This method is faster than pure interpretation because the source language statements are decoded only once. Such implementations are called hybrid implementation systems.

The process used in a hybrid implementation system is shown in Figure 1.5. Instead of translating intermediate language code to machine code, it simply interprets the intermediate code.

Figure 1.5 Hybrid implementation system

Figure 1.5 Full Alternative Text

Perl is implemented with a hybrid system. Perl programs are partially ​- compiled to detect errors before interpretation and to simplify the interpreter.

Initial implementations of Java were all hybrid. Its intermediate form, called byte code, provides portability to any machine that has a byte code interpreter and an associated run-time system. Together, these are called the Java Virtual Machine. There are now systems that translate Java byte code into machine code for faster execution.

A Just-in-Time (JIT) implementation system initially translates programs to an intermediate language. Then, during execution, it compiles intermediate language methods into machine code when they are called. The machine code version is kept for subsequent calls. JIT systems now are widely used for Java programs. Also, the .NET languages are all implemented with a JIT system.

Sometimes an implementor may provide both compiled and interpreted implementations for a language. In these cases, the interpreter is used to develop and debug programs. Then, after a (relatively) bug-free state is reached, the programs are compiled to increase their execution speed.

1.7.4 Preprocessors A preprocessor is a program that processes a program just before the program is compiled. Preprocessor instructions are embedded in programs. The preprocessor is essentially a macro expander. Preprocessor instructions are commonly used to specify that the code from another file is to be included. For example, the C preprocessor instruction

#include "myLib.h"

causes the preprocessor to copy the contents of myLib.h into the program at the position of the #include.

Other preprocessor instructions are used to define symbols to represent expressions. For example, one could use

#define max(A, B) ((A) > (B) ? (A) : (B))

to determine the largest of two given expressions. For example, the expression

x = max(2 * y, z / 1.73);

would be expanded by the preprocessor to

x = ((2 * y) > (z / 1.73) ? (2 * y) : (z / 1.73);

Notice that this is one of those cases where expression side effects can cause trouble. For example, if either of the expressions given to the max macro have side effects—such as z++—it could cause a problem. Because one of the two expression parameters is evaluated twice, this could result in z being ​- incremented twice by the code produced by the macro expansion.

1.8 Programming Environments A programming environment is the collection of tools used in the development of software. This collection may consist of only a file system, a text editor, a linker, and a compiler. Or it may include a large collection of integrated tools, each accessed through a uniform user interface. In the latter case, the process of the development and maintenance of software is greatly enhanced. Therefore, the characteristics of a programming language are not the only measure of the software development capability of a system. We now briefly describe several programming environments.

UNIX is an older programming environment, first distributed in the middle 1970s, built around a portable multiprogramming operating system. It provides a wide array of powerful support tools for software production and maintenance in a variety of languages. In the past, the most important feature absent from UNIX was a uniform interface among its tools. This made it more difficult to learn and to use. However, UNIX is now often used through a GUI that runs on top of UNIX. Examples of UNIX GUIs are the Solaris Common Desktop Environment (CDE), GNOME, and KDE. These GUIs make the interface to UNIX appear similar to that of Windows and Macintosh systems.

Borland JBuilder is a programming environment that provides an integrated compiler, editor, debugger, and file system for Java development, where all four are accessed through a graphical interface. JBuilder is a complex and powerful system for creating Java software.

Microsoft Visual Studio .NET is a relatively recent step in the evolution of software development environments. It is a large and elaborate collection of software development tools, all used through a windowed interface. This system can be used to develop software in any one of the five .NET languages: C#, Visual Basic.NET, JScript (Microsoft’s version of JavaScript), F# (a functional language), and C++/CLI.

NetBeans is a development environment that is primarily used for Java

application development but also supports JavaScript, Ruby, and PHP. Both Visual Studio and NetBeans are more than development environments—they are also frameworks, which means they actually provide common parts of the code of the application.

SUMMARY The study of programming languages is valuable for some important reasons: It increases our capacity to use different constructs in writing programs, enables us to choose languages for projects more intelligently, and makes learning new languages easier.

Computers are used in a wide variety of problem-solving domains. The design and evaluation of a particular programming language is highly ​- dependent on the domain in which it is to be used.

Among the most important criteria for evaluating languages are readability, writability, reliability, and overall cost. These will be the basis on which we examine and judge the various language features discussed in the remainder of the book.

The major influences on language design have been machine architecture and software design methodologies.

Designing a programming language is primarily an engineering feat, in which a long list of trade-offs must be made among features, constructs, and capabilities.

The major methods of implementing programming languages are compilation, pure interpretation, and hybrid implementation.

Programming environments have become important parts of software development systems, in which the language is just one of the components.

REVIEW QUESTIONS 1. Why is it useful for a programmer to have some background in language

design, even though he or she may never actually design a programming language?

2. How can knowledge of programming language characteristics benefit the whole computing community?

3. What programming language has dominated scientific computing over the past 60 years?

4. What programming language has dominated business applications over the past 60 years?

5. What programming language has dominated artificial intelligence over the past 60 years?

6. In what language is most of UNIX written?

7. What is the disadvantage of having too many features in a language?

8. How can user-defined operator overloading harm the readability of a program?

9. What is one example of a lack of orthogonality in the design of C?

10. What language used orthogonality as a primary design criterion?

11. What primitive control statement is used to build more complicated control statements in languages that lack them?

12. What does it mean for a program to be reliable?

13. Why is type checking the parameters of a subprogram important?

14. What is aliasing?

15. What is exception handling?

16. Why is readability important to writability?

17. How is the cost of compilers for a given language related to the design of that language?

18. What have been the strongest influences on programming language design over the past 60 years?

19. What is the name of the category of programming languages whose structure is dictated by the von Neumann computer architecture?

20. What two programming language deficiencies were discovered as a result of the research in software development in the 1970s?

21. What are the three fundamental features of an object-oriented programming language?

22. What language was the first to support the three fundamental features of object-oriented programming?

23. What is an example of two language design criteria that are in direct conflict with each other?

24. What are the three general methods of implementing a programming language?

25. Which produces faster program execution, a compiler or a pure interpreter?

26. What role does the symbol table play in a compiler?

27. What does a linker do?

28. Why is the von Neumann bottleneck important?

29. What are the advantages in implementing a language with a pure interpreter?

PROBLEM SET 1. Do you believe our capacity for abstract thought is influenced by our

language skills? Support your opinion.

2. What are some features of specific programming languages you know whose rationales are a mystery to you?

3. What arguments can you make for the idea of a single language for all programming domains?

4. What arguments can you make against the idea of a single language for all programming domains?

5. Name and explain another criterion by which languages can be judged (in addition to those discussed in this chapter).

6. What common programming language statement, in your opinion, is most detrimental to readability?

7. Java uses a right brace to mark the end of all compound statements. What are the arguments for and against this design?

8. Many languages distinguish between uppercase and lowercase letters in user-defined names. What are the pros and cons of this design decision?

9. Explain the different aspects of the cost of a programming language.

10. What are the arguments for writing efficient programs even though hardware is relatively inexpensive?

11. Describe some design trade-offs between efficiency and safety in some language you know.

12. In your opinion, what major features would a perfect programming language include?

13. Was the first high-level programming language you learned implemented with a pure interpreter, a hybrid implementation system, or a compiler? (You may have to research this.)

14. Describe the advantages and disadvantages of some programming environment you have used.

15. How do type declaration statements for simple variables affect the readability of a language, considering that some languages do not require them?

16. Write an evaluation of some programming language you know, using the criteria described in this chapter.

17. Some programming languages—for example, Pascal—have used the semicolon to separate statements, while Java uses it to terminate statements. Which of these, in your opinion, is most natural and least likely to result in syntax errors? Support your answer.

18. Many contemporary languages allow two kinds of comments: one in which delimiters are used on both ends (multiple-line comments), and one in which a delimiter marks only the beginning of the comment (one- line comments). Discuss the advantages and disadvantages of each of these with respect to our criteria.

2 Evolution of the Major Programming Languages

1. 2.1 Zuse’s Plankalkül

2. 2.2 Pseudocodes

3. 2.3 The IBM 704 and Fortran

4. 2.4 Functional Programming: Lisp

5. 2.5 The First Step Toward Sophistication: ALGOL 60

6. 2.6 Computerizing Business Records: COBOL

7. 2.7 The Beginnings of Timesharing: Basic

8. 2.8 Everything for Everybody: PL/I

9. 2.9 Two Early Dynamic Languages: APL and SNOBOL

10. 2.10 The Beginnings of Data Abstraction: SIMULA 67

11. 2.11 Orthogonal Design: ALGOL 68

12. 2.12 Some Early Descendants of the ALGOLs

13. 2.13 Programming Based on Logic: Prolog

14. 2.14 History’s Largest Design Effort: Ada

15. 2.15 Object-Oriented Programming: Smalltalk

16. 2.16 Combining Imperative and Object-Oriented Features: C++

17. 2.17 An Imperative-Based Object-Oriented Language: Java

18. 2.18 Scripting Languages

19. 2.19 The Flagship .NET Language: C#

20. 2.20 Markup-Programming Hybrid Languages

This chapter describes the development of a collection of programming ​- languages. It explores the environment in which each was designed and focuses on the ​contributions of the language and the motivation for its development. ​Overall language descriptions are not included; rather, we discuss only some of the new ​features introduced by each language. Of particular interest are the features that most influenced subsequent languages or the field of computer science.

This chapter does not include an in-depth discussion of any language feature or concept; that is left for later chapters. Brief, informal explanations of features will ​suffice for our trek through the development of these languages.

We discuss a wide variety of languages and language concepts that will not be familiar to many readers. These topics are described in detail only in later chapters. Those who find this unsettling may prefer to delay reading this chapter until the rest of the book has been studied.

The choice as to which languages to discuss here was subjective, and some ​- readers will unhappily note the absence of one or more of their favorites. However, to keep this historical coverage to a reasonable size, it was necessary to leave out some languages that some regard highly. The choices were based on our estimate of each language’s importance to language development and the computing world as a whole. We also include brief discussions of some other languages that are referenced later in the book.

The organization of this chapter is as follows: The initial versions of languages generally are discussed in chronological order. However, subsequent versions of ​languages appear with their initial version, rather than in later sections. For example, Fortran 2003 is discussed in the section with Fortran I (1956). Also, in some cases, ​languages of secondary importance that are related to a language that has its own ​section appear in that section.

This chapter includes listings of 14 complete example programs, each in a different language. These programs are not described in this chapter; they are meant to illustrate the appearance of programs in these languages. Readers familiar with any of the common imperative languages should be able to read and understand most of the code in these programs, except those in Lisp, COBOL, and Smalltalk. (A Scheme function similar to the Lisp example is discussed in Chapter 15.) The same problem is solved by the Fortran, ALGOL 60, PL/I, Basic, Pascal, C, Perl, Ada, Java, JavaScript, and C# programs. Note that most of the contemporary languages in this list support dynamic arrays, but because of the simplicity of the example problem, we did not use them in the example programs. Also, in the Fortran 95 program, we avoided using the features that could have avoided the use of loops altogether, in part to keep the program simple and readable and in part just to illustrate the basic loop structure of the language.

Figure 2.1 is a chart of the genealogy of the high-level languages discussed in this chapter.

Figure 2.1 Genealogy of common high-level programming languages

Figure 2.1 Full Alternative Text

2.1 Zuse’s Plankalkül The first programming language discussed in this chapter is highly unusual in several respects. For one thing, it was never implemented. Furthermore, although developed in 1945, its description was not published until 1972. Because so few people were familiar with the language, some of its capabilities did not appear in other languages until 15 years after its development.

2.1.1 Historical Background Between 1936 and 1945, German scientist Konrad Zuse (pronounced “Tsoo- zuh”) built a series of complex and sophisticated computers from electromechanical relays. By early 1945, Allied bombing had destroyed all but one of his latest models, the Z4, so he moved to a remote Bavarian village, Hinterstein, and his research group members went their separate ways.

Working alone, Zuse embarked on an effort to develop a language for expressing computations for the Z4, a project he had begun in 1943 as a proposal for his Ph.D. dissertation. He named this language Plankalkül, which means program calculus. In a lengthy manuscript dated 1945 but not published until 1972 (Zuse, 1972), Zuse defined Plankalkül and wrote algorithms in the language to solve a wide variety of problems.

2.1.2 Language Overview Plankalkül was remarkably complete, with some of its most advanced features in the area of data structures. The simplest data type in Plankalkül was the single bit. Integer and floating-point numeric types were built from the bit type. The floating-point type used twos-complement notation and the “hidden bit” scheme currently used to avoid storing the most significant bit of

the normalized fraction part of a floating-point value.

In addition to the usual scalar types, Plankalkül included arrays and records (called structs in the C-based languages). The records could include nested records.

Although the language had no explicit goto, it did include an iterative statement similar to the Ada for. It also had the command Fin with a superscript that specified an exit out of a given number of iteration loop nestings or to the beginning of a new iteration cycle. Plankalkül included a selection statement, but it did not allow an else clause.

One of the most interesting features of Zuse’s programs was the inclusion of mathematical expressions showing the current relationships between program variables. These expressions stated what would be true during execution at the points in the code where they appeared. These are very similar to the assertions of Java and in those in axiomatic semantics, which is discussed in Chapter 3.

Zuse’s manuscript contained programs of far greater complexity than any written prior to 1945. Included were programs to sort arrays of numbers; test the connectivity of a given graph; carry out integer and floating-point operations, including square root; and perform syntax analysis on logic formulas that had parentheses and operators in six different levels of precedence. Perhaps most remarkable were his 49 pages of algorithms for playing chess, a game in which he was not an expert.

If a computer scientist had found Zuse’s description of Plankalkül in the early 1950s, the single aspect of the language that would have hindered its implementation as defined would have been the notation. Each statement consisted of either two or three lines of code. The first line was most like the statements of current languages. The second line, which was optional, contained the subscripts of the array references in the first line. The same method of indicating subscripts was used by Charles Babbage in programs for his ​Analytical Engine in the middle of the nineteenth century. The last line of each Plankalkül statement contained the type names for the variables mentioned in the first line. This notation is quite intimidating when first seen.

The following example assignment statement, which assigns the value of the expression A[4} + 1 to A[5], illustrates this notation. The row labeled V is for subscripts, and the row labeled S is for the data types. In this example, 1.n means an integer of n bits:

| A + 1 => A V | 4  5 S | 1.n  1.n

We can only speculate on the direction that programming language design might have taken if Zuse’s work had been widely known in 1945 or even 1950. It is also interesting to consider how his work might have been different had he done it in a peaceful environment surrounded by other scientists, rather than in Germany in 1945 in virtual isolation.

2.2 Pseudocodes First, note that the word pseudocode is used here in a different sense than its contemporary meaning. We call the languages discussed in this section pseudocodes because that’s what they were named at the time they were developed and used (the late 1940s and early 1950s). However, they are clearly not pseudocodes in the contemporary sense.

The computers that became available in the late 1940s and early 1950s were far less usable than those of today. In addition to being slow, unreliable, expensive, and having extremely small memories, the machines of that time were difficult to program because of the lack of supporting software.

There were no high-level programming languages or even assembly languages, so programming was done in machine code, which is both tedious and error prone. Among its problems is the use of numeric codes for specifying instructions. For example, an ADD instruction might be specified by the code 14 rather than a connotative textual name, even if only a single letter. This makes programs very difficult to read. A more serious problem is absolute addressing, which makes program modification tedious and error prone. For example, suppose we have a machine language program stored in memory. Many of the instructions in such a program refer to other locations within the program, usually to reference data or to indicate the targets of branch instructions. Inserting an instruction at any position in the program other than at the end invalidates the correctness of all instructions that refer to addresses beyond the insertion point, because those addresses must be increased to make room for the new instruction. To make the addition correctly, all instructions that refer to addresses that follow the addition must be found and modified. A similar problem occurs with deletion of an instruction. In this case, however, machine languages often include a “no operation” instruction that can replace deleted instructions, thereby avoiding the problem.

These are standard problems with all machine languages and were the primary motivations for inventing assemblers and assembly languages. In

addition, most programming problems of that time were numerical and required floating-point arithmetic operations and indexing of some sort to allow the convenient use of arrays. Neither of these capabilities, however, was included in the architecture of the computers of the late 1940s and early 1950s. These deficiencies naturally led to the development of somewhat higher-level languages.

2.2.1 Short Code The first of these new languages, named Short Code, was developed by John Mauchly in 1949 for the BINAC computer, which was one of the first successful stored-program electronic computers. Short Code was later transferred to a UNIVAC I computer (the first commercial electronic computer sold in the United States) and, for several years, was one of the primary means of programming those machines. Although little is known of the original Short Code because its complete description was never published, a programming manual for the UNIVAC I version did survive (Remington-Rand, 1952). It is safe to assume that the two versions were very similar.

The words of the UNIVAC I’s memory had 72 bits, grouped as 12 six-bit bytes. Short Code consisted of coded versions of mathematical expressions that were to be evaluated. The codes were byte-pair values, and many equations could be coded in a word. The following operation codes were included:

01 -  06 abs value 1n (n+2)nd power 02 )  07 + 2n (n+2)nd root 03 =  08 pause 4n if <= n 04 /  09 ( 58 print and tab

Variables were named with byte-pair codes, as were locations to be used as constants. For example, X0 and Y0 could be variables. The statement

X0 = SQRT(ABS(Y0))

would be coded in a word as 00 X0 03 20 06 Y0. The initial 00 was used as

padding to fill the word. Interestingly, there was no multiplication code; multiplication was indicated by simply placing the two operands next to each other, as in algebra.

Short Code was not translated to machine code; rather, it was implemented with a pure interpreter. At the time, this process was called automatic programming. It clearly simplified the programming process, but at the expense of execution time. Short Code interpretation was approximately 50 times slower than machine code.

2.2.2 Speedcoding In other places, interpretive systems were being developed that extended machine languages to include floating-point operations. The Speedcoding system developed by John Backus for the IBM 701 is an example of such a system (Backus, 1954). The Speedcoding interpreter effectively converted the 701 to a virtual three-address floating-point calculator. The system included pseudoinstructions for the four arithmetic operations on floating-point data, as well as operations such as square root, sine, arc tangent, exponent, and logarithm. Conditional and unconditional branches and input/output conversions were also part of the virtual architecture. To get an idea of the limitations of such systems, consider that the remaining usable memory after loading the interpreter was only 700 words and that the add instruction took 4.2 milliseconds to execute. On the other hand, Speedcoding included the novel facility of automatically incrementing address registers. This facility did not appear in hardware until the UNIVAC 1107 computers of 1962. Because of such features, matrix multiplication could be done in 12 Speedcoding instructions. Backus claimed that problems that could take two weeks to program in machine code could be programmed in a few hours using Speedcoding.

2.2.3 The UNIVAC “Compiling” System

Between 1951 and 1953, a team led by Grace Hopper at UNIVAC developed a series of “compiling” systems named A-0, A-1, and A-2 that expanded a pseudocode into machine code subprograms in the same way as macros are expanded into assembly language. The pseudocode source for these “compilers” was still quite primitive, although even this was a great improvement over machine code because it made source programs much shorter. Wilkes (1952) independently suggested a similar process.

2.2.4 Related Work Other means of easing the task of programming were being developed at about the same time. At Cambridge University, David J. Wheeler (1950) developed a method of using blocks of relocatable addresses to solve, at least partially, the problem of absolute addressing, and later, Maurice V. Wilkes (also at Cambridge) extended the idea to design an assembly program that could combine chosen subroutines and allocate storage (Wilkes et al., 1951, 1957). This was indeed an important and fundamental advance.

We should also mention that assembly languages, which are quite different from the pseudocodes discussed, evolved during the early 1950s. However, they had little impact on the design of high-level languages.

2.3 The IBM 704 and Fortran Certainly one of the greatest single advances in computing came with the introduction of the IBM 704 in 1954, in large measure because its capabilities prompted the development of Fortran. One could argue that if it had not been IBM with the 704 and Fortran, it would soon thereafter have been some other organization with a similar computer and related high-level language. However, IBM was the first with both the foresight and the resources to undertake these developments.

2.3.1 Historical Background One of the primary reasons why the slowness of interpretive systems was tolerated from the late 1940s to the mid-1950s was the lack of floating-point hardware in the available computers. All floating-point operations had to be simulated in software, a very time-consuming process. Because so much processor time was spent in software floating-point processing, the overhead of interpretation and the simulation of indexing were relatively insignificant. As long as floating-point had to be done by software, interpretation was an acceptable expense. However, many programmers of that time never used interpretive systems, preferring the efficiency of hand-coded machine (or assembly) language. The announcement of the IBM 704 system, with both indexing and floating-point instructions in hardware, heralded the end of the interpretive era, at least for scientific computation. The inclusion of floating- point hardware removed the hiding place for the cost of interpretation.

Although Fortran is often credited with being the first compiled high-level language, the question of who deserves credit for implementing the first such language is somewhat open. Knuth and Pardo (1977) give the credit to Alick E. Glennie for his Autocode compiler for the Manchester Mark I computer. Glennie developed the compiler at Fort Halstead, Royal Armaments Research Establishment, in England. The compiler was operational by ​September 1952. However, according to John Backus (Wexelblat, 1981, p. 26),

Glennie’s Autocode was so low level and machine oriented that it should not be considered a compiled system. Backus gives the credit to Laning and Zierler at the ​Massachusetts Institute of Technology.

The Laning and Zierler system (Laning and Zierler, 1954) was the first algebraic translation system to be implemented. By algebraic, we mean that it translated arithmetic expressions, used separately coded subprograms to compute transcendental functions (e.g., sine and logarithm), and included arrays. The system was implemented on the MIT Whirlwind computer, in experimental prototype form, in the summer of 1952 and in a more usable form by May 1953. The translator generated a subroutine call to code each formula, or expression, in the program. The source language was easy to read, and the only actual machine instructions included were for branching. Although this work preceded the work on Fortran, it never escaped MIT.

In spite of these earlier works, the first widely accepted compiled high-level language was Fortran. The following subsections chronicle this important development.

2.3.2 Design Process Even before the 704 system was announced in May 1954, plans were begun for Fortran. By November 1954, John Backus and his group at IBM had produced the report titled “The IBM Mathematical FORmula TRANslating System: FORTRAN” (IBM, 1954). This document described the first version of Fortran, which we refer to as Fortran 0, prior to its implementation. It also boldly stated that Fortran would provide the efficiency of hand-coded programs and the ease of programming of the interpretive pseudocode systems. In another burst of optimism, the document stated that Fortran would eliminate coding errors and the debugging process. Based on this premise, the first Fortran compiler included little syntax error checking.

The environment in which Fortran was developed was as follows: (1) Computers had small memories and were slow and relatively unreliable; (2) the primary use of computers was for scientific computations; (3) there were no existing efficient and effective ways to program computers; and (4)

because of the high cost of computers compared to the cost of programmers, speed of the generated object code was the primary goal of the first Fortran compilers. The characteristics of the early versions of Fortran follow directly from this environment.

2.3.3 Fortran I Overview Fortran 0 was modified during the implementation period, which began in January 1955 and continued until the release of the compiler in April 1957. The implemented language, which we call Fortran I, is described in the first Fortran Programmer’s Reference Manual, published in October 1956 (IBM, 1956). Fortran I included input/output formatting, variable names of up to six characters (it had been just two in Fortran 0), user-defined subroutines, although they could not be separately compiled, the If selection statement, and the Do loop statement.

All of Fortran I’s control statements were based on 704 instructions. It is not clear whether the 704 designers dictated the control statement design of Fortran I or whether the designers of Fortran I suggested these instructions to the 704 designers.

There were no data-typing statements in the Fortran I language. Variables whose names began with I, J, K, L, M, and N were implicitly integer type, and all others were implicitly floating-point. The choice of the letters for this convention was based on the fact that at that time scientists and engineers used letters as variable subscripts, usually i, j, and k. In a gesture of generosity, Fortran’s designers threw in the three additional letters.

The most audacious claim made by the Fortran development group during the design of the language was that the machine code produced by the compiler would be about half as efficient as what could be produced by hand.1 This, more than anything else, made skeptics of potential users and prevented a great deal of interest in Fortran before its actual release. To almost everyone’s surprise, however, the Fortran development group nearly achieved its goal in efficiency. The largest part of the 18 worker-years of effort used to construct the first compiler had been spent on optimization, and

the results were remarkably effective.

1. In fact, the Fortran team believed that the code generated by their compiler could be no less than half as fast as handwritten machine code, or the language would not be adopted by users.

The early success of Fortran is shown by the results of a survey made in April 1958. At that time, roughly half of the code being written for 704s was being written in Fortran, in spite of the skepticism of most of the programming world only a year earlier.

2.3.4 Fortran II The Fortran II compiler was distributed in the spring of 1958. It fixed many of the bugs in the Fortran I compilation system and added some significant features to the language, the most important being the independent compilation of subroutines. Without independent compilation, any change in a program required that the entire program be recompiled. Fortran I’s lack of independent-compilation capability, coupled with the poor reliability of the 704, placed a practical restriction on the length of programs to about 300 to 400 lines (Wexelblat, 1981, p. 68). Longer programs had a poor chance of being compiled completely before a machine failure occurred. The capability of including precompiled machine language versions of subprograms shortened the compilation process considerably and made it practical to develop much larger programs.

2.3.5 Fortrans IV, 77, 90, 95, 2003, and 2008 A Fortran III was developed, but it was never widely distributed. Fortran IV, however, became one of the most widely used programming languages of its time. It evolved over the period 1960 to 1962 and was standardized as Fortran 66 (ANSI, 1966), although that name was rarely used. Fortran IV was

an improvement over Fortran II in many ways. Among its most important additions were explicit type declarations for variables, a logical If construct, and the capability of passing subprograms as parameters to other subprograms.

Fortran IV was replaced by Fortran 77, which became the new standard in 1978 (ANSI, 1978a). Fortran 77 retained most of the features of Fortran IV and added character string handling, logical loop control statements, and an If with an optional Else clause.

Fortran 90 (ANSI, 1992) was dramatically different from Fortran 77. The most significant additions were dynamic arrays, records, pointers, a multiple selection statement, and modules. In addition, Fortran 90 subprograms could be recursively called.

A new concept that was included in the Fortran 90 definition was that of removing some language features from earlier versions. While Fortran 90 included all of the features of Fortran 77, the language definition included a list of constructs that were recommended for removal in the next version of the language.

Fortran 90 included two simple syntactic changes that altered the appearance of both programs and the literature describing the language. First, the required fixed format of code, which required the use of specific character positions for specific parts of statements, was dropped. For example, statement labels could appear only in the first five positions and statements could not begin before the seventh position. This rigid formatting of code was designed around the use of punch cards. The second change was that the official spelling of FORTRAN became Fortran. This change was accompanied by the change in convention of using all uppercase letters for keywords and identifiers in Fortran programs. The new convention was that only the first letter of keywords and identifiers would be uppercase.

Fortran 95 (INCITS/ISO/IEC, 1997) continued the evolution of the language, but only a few changes were made. Among other things, a new iteration construct, Forall, was added to ease the task of parallelizing Fortran programs.

Fortran 2003 (Metcalf et al., 2004) added support for object-oriented programming, parameterized derived types, procedure pointers, and interoperability with the C programming language.

The latest version of Fortran, Fortran 2008 (ISO/IEC 1539-1, 2010), added support for blocks to define local scopes, co-arrays, which provide a parallel execution model, and the DO CONCURRENT construct, to specify loops without interdependencies. A new version, Fortran 2015, is under development and is scheduled for release in 2018.

2.3.6 Evaluation The original Fortran design team thought of language design only as a necessary prelude to the critical task of designing the translator. Furthermore, it never occurred to them that Fortran would be used on computers not manufactured by IBM. However, they were forced to consider building Fortran compilers for other IBM machines only because the successor to the 704, the 709, was announced before the 704 Fortran compiler was released. The effect that Fortran has had on the use of computers, along with the fact that all ​subsequent programming languages owe a debt to Fortran, is indeed impressive in light of the modest goals of its designers.

One of the features of Fortran I, and all of its successors before 90, that allows highly optimizing compilers was that the types and storage for all ​- variables are fixed before run time. No new variables or space could be allocated during execution. This was a sacrifice of flexibility to simplicity and efficiency. It eliminated the possibility of recursive subprograms and made it difficult to implement data structures that grow or change shape dynamically. Of course, the kinds of programs that were being built at the time of the development of the early versions of Fortran were primarily numerical in nature and were simple in comparison with more recent software projects. Therefore, the sacrifice was not a great one.

The overall success of Fortran is difficult to overstate: It dramatically changed the way computers are used. This is, of course, in large part due to its being the first widely used high-level language. In comparison with

concepts and languages developed later, early versions of Fortran suffer in a variety of ways, as should be expected. After all, it would not be fair to compare the performance and comfort of a 1910 Model T Ford with the performance and comfort of a 2017 Ford Mustang. Nevertheless, in spite of the inadequacies of Fortran, the momentum of the huge investment in Fortran software, among other factors, has kept it in use for 60 years.

Alan Perlis, one of the designers of ALGOL 60, said of Fortran in 1978, “Fortran is the lingua franca of the computing world. It is the language of the streets in the best sense of the word, not in the prostitutional sense of the word. And it has survived and will survive because it has turned out to be a remarkably useful part of a very vital commerce” (Wexelblat, 1981, p. 161).

The following is an example of a Fortran 95 program:

! Fortran 95 Example program

! Input: An integer, List_Len, where List_Len is less

! than 100, followed by List_Len-Integer values

! Output:  The number of input values that are greater !  than the average of all input values Implicit none

Integer Dimension(99) :: Int_List

Integer :: List_Len, Counter, Sum, Average, Result

Result= 0

Sum = 0

Read *, List_Len

If ((List_Len > 0) .AND. (List_Len < 100)) Then

! Read input data into an array and compute its sum

Do Counter = 1, List_Len

Read *, Int_List(Counter)

Sum = Sum + Int_List(Counter)

End Do

! Compute the average

Average = Sum / List_Len

! Count the values that are greater than the average

Do Counter = 1, List_Len

If (Int_List(Counter) > Average) Then

  Result = Result + 1 End If

End Do

! Print the result

Print *, 'Number of values > Average is:', Result

Else

Print *, 'Error - list length value is not legal'

End If

End Program Example

2.4 Functional Programming: Lisp The first functional programming language was invented to provide language features for list processing, the need for which grew out of the first applications in the area of artificial intelligence (AI).

2.4.1 The Beginnings of Artificial Intelligence (AI) and List Processing Interest in AI appeared in the mid-1950s in a number of places. Some of this interest grew out of linguistics, some from psychology, and some from mathematics. Linguists were concerned with natural language processing. ​- Psychologists were interested in modeling human information storage and retrieval, as well as other fundamental processes of the brain. Mathematicians were interested in mechanizing certain intelligent processes, such as theorem proving. All of these investigations arrived at the same conclusion: Some method must be developed to allow computers to process symbolic data in linked lists. At the time, nearly all computation was on numeric data in arrays.

The concept of list processing was developed by Allen Newell, J. C. Shaw, and Herbert Simon at the RAND Corporation. It was first published in a classic paper that describes one of the first AI programs, the Logic Theorist,2 and a language in which it could be implemented (Newell and Simon, 1956). The language, named IPL-I (Information Processing Language I), was never implemented. The next version, IPL-II, was implemented on a RAND Johnniac computer. Development of IPL continued until 1960, when the description of IPL-V was published (Newell and Tonge, 1960). The low level of the IPL languages prevented their widespread use. They were actually assembly languages for a hypothetical computer, implemented with an interpreter, in which list-processing instructions were included. Another factor that kept the IPL languages from becoming popular was their

implementation on the obscure Johnniac machine.

2. Logic Theorist discovered proofs for theorems in propositional calculus.

The contributions of the IPL languages were in their list design and their demonstration that list processing was feasible and useful.

IBM became interested in AI in the mid-1950s and chose theorem proving as a demonstration area. At the time, the Fortran project was still underway. The high cost of the Fortran I compiler convinced IBM that their list processing should be attached to Fortran, rather than in the form of a new language. Thus, the Fortran list processing language (FLPL) was designed and implemented as an extension to Fortran. FLPL was used to construct a theorem prover for plane geometry, which was then considered the easiest area for mechanical theorem proving.

2.4.2 Lisp Design Process John McCarthy of MIT took a summer position at the IBM Information Research Department in 1958. His goal for the summer was to investigate symbolic computations and to develop a set of requirements for doing such computations. As a pilot example problem area, he chose differentiation of algebraic expressions. From this study came a list of language requirements. Among them were the control flow methods of mathematical functions: recursion and conditional expressions. The only available high-level language of the time, Fortran I, had neither of these.

Another requirement that grew from the symbolic-differentiation investigation was the need for dynamically allocated linked lists and some kind of implicit deallocation of abandoned lists. McCarthy simply would not allow his elegant algorithm for differentiation to be cluttered with explicit deallocation statements.

Because FLPL did not support recursion, conditional expressions, dynamic storage allocation, or implicit deallocation, it was clear to McCarthy that a new language was needed.

When McCarthy returned to MIT in the fall of 1958, he and Marvin Minsky formed the MIT AI Project, with funding from the Research Laboratory for Electronics at MIT. The first important effort of the project was to produce a software system for list processing. It was to be used initially to implement a program proposed by McCarthy called the Advice Taker.3 This application became the impetus for the development of the list-processing language Lisp. The first version of Lisp is sometimes called “pure Lisp” because it is a purely functional language. In the following section, we describe the development of pure Lisp.

3. Advice Taker represented information with sentences written in a formal language and used a logical inferencing process to decide what to do.

2.4.3 Language Overview

2.4.3.1 Data Structures Pure Lisp has only two kinds of data structures: atoms and lists. Atoms are either symbols, which have the form of identifiers, or numeric literals. The concept of storing symbolic information in linked lists is natural and was used in IPL-II. Such structures allow insertions and deletions at any point, operations that were then thought to be a necessary part of list processing. It was eventually determined, however, that Lisp programs rarely require these operations.

Lists are specified by delimiting their elements with parentheses. Simple lists, in which elements are restricted to atoms, have the form

(A B C D)

Nested list structures are also specified by parentheses. For example, the list

(A (B C) D (E (F G)))

is composed of four elements. The first is the atom A; the second is the sublist (B C); the third is the atom D; the fourth is the sublist (E (F G)), which has

as its second element the sublist (F G).

Internally, lists are stored as single-linked list structures, in which each node has two pointers and represents a list element. A node containing an atom has its first pointer pointing to some representation of the atom, such as its symbol or numeric value, or a pointer to a sublist. A node for a sublist element has its first pointer pointing to the first node of the sublist. In both cases, the second pointer of a node points to the next element of the list. A list is referenced by a pointer to its first element.

The internal representations of the two lists shown earlier are depicted in Figure 2.2. Note that the elements of a list are shown horizontally. The last element of a list has no successor, so its link is NIL, which is represented in Figure 2.2 as a diagonal line in the element. Sublists are shown with the same structure.

Figure 2.2 Internal

representation of two Lisp lists Figure 2.2 Full Alternative Text

2.4.3.2 Processes in Functional Programming Lisp was designed as a functional programming language. All computation in a purely functional program is accomplished by applying functions to arguments. Neither the assignment statements nor the variables that abound in imperative language programs are necessary in functional language programs. Furthermore, repetitive processes can be specified with recursive function calls, making iteration (loops) unnecessary. These basic concepts of functional programming make it significantly different from programming in an imperative language.

2.4.3.3 The Syntax of Lisp Lisp is very different from the imperative languages, both because it is a functional programming language and because the appearance of Lisp programs is so different from those in languages like Java or C++. For example, the syntax of Java is a complicated mixture of English and algebra, while Lisp’s syntax is a model of simplicity. Program code and data have exactly the same form: parenthesized lists. Consider again the list

(A B C D)

When interpreted as data, it is a list of four elements. When viewed as code, it is the application of the function named A to the three parameters B, C, and D.

2.4.4 Evaluation

Lisp completely dominated AI applications for a quarter century. Much of the cause of Lisp’s reputation for being highly inefficient has been eliminated. Many contemporary implementations are compiled, and the resulting code is much faster than running the source code on an interpreter. In addition to its success in AI, Lisp pioneered functional programming, which has proven to be a lively area of research in programming languages. As stated in Chapter 1, many programming language researchers believe functional programming is a much better approach to software development than procedural programming using imperative languages.

The following is an example of a Lisp program:

; Lisp Example function

; The following code defines a Lisp predicate function

; that takes two lists as arguments and returns True

; if the two lists are equal, and NIL (false) otherwise

(DEFUN equal_lists (lis1 lis2)

(COND

((ATOM lis1) (EQ lis1 lis2))

((ATOM lis2) NIL)

((equal_lists (CAR lis1) (CAR lis2))

   (equal_lists (CDR lis1) (CDR lis2))) (T NIL)

)

)

2.4.5 Two Descendants of Lisp Two dialects of Lisp are now widely used, Scheme and Common Lisp. These are briefly discussed in the following subsections.

2.4.5.1 Scheme The Scheme language emerged from MIT in the mid-1970s. It is characterized by its small size, its exclusive use of static scoping (discussed in Chapter 5), and its treatment of functions as first-class entities. As first- class entities, Scheme functions can be assigned to variables, passed as

parameters, and returned as the values of function applications. They can also be the elements of lists. Early versions of Lisp did not provide all of these capabilities, nor did they use static scoping.

As a small language with simple syntax and semantics, Scheme is well suited to educational applications, such as courses in functional programming and general introductions to programming. Scheme is described in some detail in Chapter 15.

2.4.5.2 Common Lisp During the 1970s and early 1980s, a large number of different dialects of Lisp were developed and used. This led to the familiar problem of lack of portability among programs written in the various dialects. Common Lisp ​- (Graham, 1996) was created in an effort to rectify this situation. Common Lisp was designed by combining the features of several dialects of Lisp developed in the early 1980s, including Scheme, into a single language. Being such an amalgam, Common Lisp is a relatively large and complex language. Its basis, however, is pure Lisp, so its syntax, primitive functions, and fundamental nature come from that language.

Recognizing the flexibility provided by dynamic scoping as well as the simplicity of static scoping, Common Lisp allows both. The default scoping for variables is static, but by declaring a variable to be special, that variable becomes dynamically scoped.

Common Lisp has a large number of data types and structures, including records, arrays, complex numbers, and character strings. It also has a form of packages for modularizing collections of functions and data providing access control.

Common Lisp is further described in Chapter 15.

2.4.6 Related Languages

ML (MetaLanguage; Ullman, 1998) was originally designed in the 1980s by Robin Milner at the University of Edinburgh as a metalanguage for a program verification system named Logic for Computable Functions (LCF; Milner et al., 1997). ML is primarily a functional language, but it also supports imperative programming. Unlike Lisp and Scheme, the type of every variable and expression in ML can be determined at compile time. Types are associated with objects rather than names. Types of names and expressions are inferred from their context.

Unlike Lisp and Scheme, ML does not use the parenthesized functional syntax that originated with lambda expressions. Rather, the syntax of ML resembles that of the imperative languages, such as Java and C++.

Miranda was developed by David Turner (1986) at the University of Kent in Canterbury, England, in the early 1980s. Miranda is based partly on the languages ML, SASL, and KRC. Haskell (Hudak and Fasel, 1992) is based in large part on Miranda. Like Miranda, it is a purely functional language, having no variables and no assignment statement. Another distinguishing characteristic of Haskell is its use of lazy evaluation. This means that no expression is evaluated until its value is required. This leads to some surprising capabilities in the language.

Caml (Cousineau et al., 1998) and its dialect that supports object-oriented programming, OCaml (Smith, 2006), descended from ML and Haskell. Finally, F# is a relatively new typed language based directly on OCaml. F# (Syme et al., 2010) is a .NET language with direct access to the whole .NET library. Being a .NET language also means it can smoothly interoperate with any other .NET language. F# supports both functional programming and procedural programming. It also fully supports object-oriented programming.

ML, Haskell, and F# are further discussed in Chapter 15.

2.5 The First Step Toward Sophistication: ALGOL 60 ALGOL 60 strongly influenced subsequent programming languages and is therefore of central importance in any historical study of languages.

2.5.1 Historical Background ALGOL 60 was the result of efforts to design a universal programming language for scientific applications. By late 1954, the Laning and Zierler algebraic system had been in operation for over a year, and the first report on Fortran had been published. Fortran became a reality in 1957, and several other high-level languages were being developed. Most notable among them were IT, which was designed by Alan Perlis at Carnegie Tech, and two languages for the UNIVAC computers, MATH-MATIC and UNICODE. The proliferation of languages made program sharing among users difficult. Furthermore, the new languages were all growing up around single architectures, some for UNIVAC computers and some for IBM 700-series machines. In response to this blossoming of machine-dependent languages, several major computer user groups in the United States, including SHARE (the IBM scientific user group) and USE (UNIVAC Scientific Exchange, the large-scale UNIVAC scientific user group), submitted a petition to the Association for Computing Machinery (ACM) on May 10, 1957, to form a committee to study and recommend action to create a machine-independent scientific programming language. Although Fortran might have been a candidate, it could not become a universal language, because at the time it was solely owned by IBM.

Previously, in 1955, GAMM (a German acronym for Society for Applied Mathematics and Mechanics) had formed a committee to design one universal, machine-independent algorithmic language. The desire for this new language was in part due to the Europeans’ fear of being dominated by IBM.

By late 1957, however, the appearance of several high-level languages in the United States convinced the GAMM subcommittee that their effort had to be widened to include the Americans, and a letter of invitation was sent to ACM. In April 1958, after Fritz Bauer of GAMM presented the formal proposal to ACM, the two groups officially agreed to a joint language design project.

2.5.2 Early Design Process GAMM and ACM each sent four members to the first design meeting. The meeting, which was held in Zurich from May 27 to June 1, 1958, began with the following goals for the new language:

The syntax of the language should be as close as possible to standard mathematical notation, and programs written in it should be readable with little further explanation.

It should be possible to use the language for the description of algorithms in printed publications.

Programs in the new language must be mechanically translatable into machine language.

The first goal indicated that the new language was to be used for scientific programming, which was the primary computer application area at that time. The second was something entirely new to the computing business. The last goal is an obvious necessity for any programming language.

The Zurich meeting succeeded in producing a language that met the stated goals, but the design process required innumerable compromises, both among individuals and between the two sides of the Atlantic. In some cases, the compromises were not so much over great issues as they were over spheres of influence. The question of whether to use a comma (the European method) or a period (the American method) for a decimal point is one example.

2.5.3 ALGOL 58 Overview The language designed at the Zurich meeting was named the International Algorithmic Language (IAL). It was suggested during the design that the language be named ALGOL, for ALGOrithmic Language, but the name was rejected because it did not reflect the international scope of the committee. During the following year, however, the name was changed to ALGOL, and the language subsequently became known as ALGOL 58.

In many ways, ALGOL 58 was a descendant of Fortran, which is quite natural. It generalized many of Fortran’s features and added several new constructs and concepts. Some of the generalizations had to do with the goal of not tying the language to any particular machine, and others were attempts to make the language more flexible and powerful. A rare combination of simplicity and elegance emerged from the effort.

ALGOL 58 formalized the concept of data type, although only variables that were not floating-point required explicit declaration. It added the idea of compound statements, which most subsequent languages incorporated. Some features of Fortran that were generalized were the following: Identifiers were allowed to have any length, as opposed to Fortran I’s restriction to six or fewer characters; any number of array dimensions was allowed, unlike Fortran I’s limitation to no more than three; the lower bound of arrays could be specified by the programmer, whereas in Fortran it was implicitly 1; nested selection statements were allowed, which was not the case in Fortran I.

ALGOL 58 acquired the assignment operator in a rather unusual way. Zuse used the form

expression => variable

for the assignment statement in Plankalkül. Although Plankalkül had not yet been published, some of the European members of the ALGOL 58 committee were familiar with the language. The committee dabbled with the Plankalkül assignment form but, because of arguments about character set limitations,4 the greater-than symbol was changed to a colon. Then, largely at the

insistence of the Americans, the whole statement was turned around to the Fortran form

4. The card punches of that time did not include the greater-than symbol.

variable := expression

The Europeans preferred the opposite form, but that would be the reverse of Fortran.

2.5.4 Reception of the ALGOL 58 Report In December 1958, publication of the ALGOL 58 report (Perlis and Samelson, 1958) was greeted with enthusiasm. In the United States, the new language was viewed more as a collection of ideas for programming language design than as a universal standard language. Actually, the ALGOL 58 report was not meant to be a finished product but rather a preliminary document for international discussion. Nevertheless, three major design and implementation efforts used the report as their basis. At the University of Michigan, the MAD language was born (Arden et al., 1961). The U.S. Naval Electronics Group produced the NELIAC language (Huskey et al., 1963). At System Development Corporation, JOVIAL was designed and implemented (Shaw, 1963). JOVIAL, an acronym for Jules’ Own Version of the International Algebraic Language, represents the only language based on ALGOL 58 to achieve widespread use (Jules was Jules I. Schwartz, one of JOVIAL’s designers). JOVIAL became widely used because it was the official scientific language for the U.S. Air Force for a quarter century.

The rest of the U.S. computing community was not so kind to the new language. At first, both IBM and its major scientific user group, SHARE, seemed to embrace ALGOL 58. IBM began an implementation shortly after the report was published, and SHARE formed a subcommittee, SHARE IAL, to study the language. The subcommittee subsequently recommended that ACM standardize ALGOL 58 and that IBM implement it for all of the 700-

series computers. The enthusiasm was short-lived, however. By the spring of 1959, both IBM and SHARE, through their Fortran experience, had had enough of the pain and expense of getting a new language started, both in terms of developing and using the first-generation compilers and in terms of training users in the new language and persuading them to use it. By the middle of 1959, both IBM and SHARE had developed such a vested interest in Fortran that they decided to retain it as the scientific language for the IBM 700-series machines, thereby abandoning ALGOL 58.

2.5.5 ALGOL 60 Design Process During 1959, ALGOL 58 was furiously debated in both Europe and the United States. Large numbers of suggested modifications and additions were published in the European ALGOL Bulletin and in Communications of the ACM. One of the most important events of 1959 was the presentation of the work of the Zurich committee to the International Conference on Information Processing, for there Backus introduced his new notation for describing the syntax of programming languages, which later became known as BNF (Backus-Naur form). BNF is described in detail in Chapter 3.

In January 1960, the second ALGOL meeting was held, this time in Paris. The purpose of the meeting was to debate the 80 suggestions that had been formally submitted for consideration. Peter Naur of Denmark had become heavily involved in the development of ALGOL, even though he had not been a member of the Zurich group. It was Naur who created and published the ALGOL Bulletin. He spent a good deal of time studying Backus’s paper that introduced BNF and decided that BNF should be used to describe formally the results of the 1960 meeting. After making a few relatively minor changes to BNF, he wrote a description of the new proposed language in BNF and handed it out to the members of the 1960 group at the beginning of the meeting.

2.5.6 ALGOL 60 Overview

Although the 1960 meeting lasted only six days, the modifications made to ALGOL 58 were dramatic. Among the most important new developments were the following:

The concept of block structure was introduced. This allowed the programmer to localize parts of programs by introducing new data environments, or scopes.

Two different means of passing parameters to subprograms were allowed: pass by value and pass by name.

Procedures were allowed to be recursive. The ALGOL 58 description was unclear on this issue. Note that although recursion was new for the imperative languages, Lisp had already provided recursive functions in 1959.

Stack-dynamic arrays were allowed. A stack-dynamic array is one for which the subscript range or ranges are specified by variables, so that the size of the array is set at the time storage is allocated to the array, which happens when the declaration is reached during execution. Stack- dynamic arrays are described in detail in Chapter 6.

Several features that might have had a dramatic impact on the success or failure of the language were proposed and rejected. Most important among these were input and output statements with formatting, which were omitted because they were thought to be machine-dependent.

The ALGOL 60 report was published in May 1960 (Naur, 1960). A number of ambiguities still remained in the language description, and a third meeting was scheduled for April 1962 in Rome to address the problems. At this meeting the group dealt only with problems; no additions to the language were allowed. The results of this meeting were published under the title “Revised Report on the Algorithmic Language ALGOL 60” (Backus et al., 1963).

2.5.7 Evaluation

In some ways, ALGOL 60 was a great success; in other ways, it was a dismal failure. It succeeded in becoming, almost immediately, the only acceptable formal means of communicating algorithms in computing literature, and it remained that for more than 20 years. Every imperative programming language designed since 1960 owes something to ALGOL 60. In fact, most are direct or indirect descendants; examples include PL/I, SIMULA 67, ALGOL 68, C, Pascal, Ada, C++, Java, and C#.

The ALGOL 58/ALGOL 60 design effort included a long list of firsts. It was the first time that an international group attempted to design a programming language. It was the first language that was designed to be machine independent. It was also the first language whose syntax was formally described. This successful use of the BNF formalism initiated several important fields of computer science: formal languages, parsing theory, and BNF-based compiler design. Finally, the structure of ALGOL 60 affected machine architecture. In the most striking example of this, an extension of the language was used as the systems language of a series of large-scale computers, the Burroughs B5000, B6000, and B7000 machines, which were designed with a hardware stack to implement efficiently the block structure and recursive subprograms of the language.

On the other side of the coin, ALGOL 60 never achieved widespread use in the United States. Even in Europe, where it was more popular than in the United States, it never became the dominant language. There are a number of reasons for its lack of acceptance. For one thing, some of the features of ALGOL 60 turned out to be too flexible; they made understanding difficult and implementation inefficient. The best example of this is the pass-by-name method of passing parameters to subprograms, which is explained in Chapter 9. The difficulties of implementing ALGOL 60 are evidenced by Rutishauser’s statement in 1967 that few, if any, implementations included the full ALGOL 60 language (Rutishauser, 1967, p. 8).

The lack of input and output statements in the language was another major reason for its lack of acceptance. Implementation-dependent input/output made programs difficult to port to other computers.

Ironically, one of the most important contributions to computer science associated with ALGOL 60, BNF, was also a factor in its lack of acceptance.

Although BNF is now considered a simple and elegant means of syntax description, in 1960 it seemed strange and complicated.

Finally, although there were many other problems, the entrenchment of Fortran among users and the lack of support by IBM were probably the most important factors in ALGOL 60’s failure to gain widespread use.

The ALGOL 60 effort was never really complete, in the sense that ​- ambiguities and obscurities were always a part of the language description (Knuth, 1967).

The following is an example of an ALGOL 60 program:

comment ALGOL 60 Example Program

Input: An integer, listlen, where listlen is less than

100, followed by listlen-integer values

Output: The number of input values that are greater than

the average of all the input values ;

begin

integer array intlist [1:99];

integer listlen, counter, sum, average, result;

sum := 0;

result := 0;

readint (listlen);

if (listlen > 0) ^ (listlen < 100) then

begin

comment Read input into an array and compute the average;

for counter := 1 step 1 until listlen do

begin

readint (intlist[counter]);

sum := sum + intlist[counter]

end;

comment Compute the average;

average := sum / listlen;

comment Count the input values that are > average;

for counter := 1 step 1 until listlen do

if intlist[counter] > average

then result := result + 1;

comment Print result;

printstring("The number of values > average is:");

printint (result)

end

else

printstring ("Error-input list length is not legal";

end

2.6 Computerizing Business Records: COBOL The story of COBOL is, in a sense, the opposite of that of ALGOL 60. Although it has been used for nearly 60 years, COBOL has had little effect on the design of subsequent languages, except for PL/I. It may still be the most widely used language,5 although it is difficult to be sure one way or the other. Perhaps the most important reason why COBOL has had little influence is that few have attempted to design a new language for business applications since it appeared. That is due in part to how well COBOL’s capabilities meet the needs of its application area. Another reason is that a great deal of growth in business computing over the past 30 years has occurred in small businesses. In these businesses, very little software development has taken place. Instead, most of the software used is purchased as off-the-shelf packages for various general business applications.

5. In the late 1990s, in a study associated with the Y2K problem, it was estimated that there were approximately 800 million lines of COBOL in regular use in the 22 square miles of Manhattan.

2.6.1 Historical Background The beginning of COBOL is somewhat similar to that of ALGOL 60, in the sense that the language was designed by a committee of people meeting for relatively short periods of time. At the time, in 1959, the state of business computing was similar to the state of scientific computing several years earlier, when Fortran was being designed. One compiled language for business applications, FLOW-MATIC, had been implemented in 1957, but it belonged to one manufacturer, UNIVAC, and was designed for that company’s computers. Another language, AIMACO, was being used by the U.S. Air Force, but it was only a minor variation of FLOW-MATIC. IBM had designed a programming language for business applications, COMTRAN

(COMmercial TRANslator), but it had not yet been implemented. Several other language design projects were being planned.

2.6.2 FLOW-MATIC The origins of FLOW-MATIC are worth at least a brief discussion, because it was the primary progenitor of COBOL. In December 1953, Grace Hopper at Remington-Rand UNIVAC wrote a proposal that was indeed prophetic. It suggested that “mathematical programs should be written in mathematical notation, data processing programs should be written in English statements” (Wexelblat, 1981, p. 16). Unfortunately, in 1953, it was impossible to convince nonprogrammers that a computer could be made to understand English words. It was not until 1955 that a similar proposal had some hope of being funded by UNIVAC management, and even then it took a prototype system to do the final convincing. Part of this selling process involved compiling and running a small program, first using English keywords, then using French keywords, and then using German keywords. This demonstration was considered remarkable by UNIVAC management and was instrumental in their acceptance of Hopper’s proposal.

2.6.3 COBOL Design Process The first formal meeting on the subject of a common language for business applications, which was sponsored by the Department of Defense, was held at the Pentagon on May 28 and 29, 1959 (exactly one year after the Zurich ALGOL meeting). The consensus of the group was that the language, then named CBL (Common Business Language), should have the following general characteristics: Most agreed that it should use English as much as possible, although a few argued for a more mathematical notation. The language must be easy to use, even at the expense of being less powerful, in order to broaden the base of those who could program computers. In addition to making the language easy to use, it was believed that the use of English would allow managers to read programs. Finally, the design should not be overly restricted by the problems of its implementation.

One of the overriding concerns at the meeting was that steps to create this universal language should be taken quickly, as a lot of work was already being done to create other business languages. In addition to the existing languages, RCA and Sylvania were working on their own business applications languages. It was clear that the longer it took to produce a universal language, the more difficult it would be for the language to become widely used. On this basis, it was decided that there should be a quick study of existing languages. For this task, the Short Range Committee was formed.

There were early decisions to separate the statements of the language into two categories—data description and executable operations—and to have statements in these two categories be in different parts of programs. One of the debates of the Short Range Committee was over the inclusion of subscripts. Many committee members argued that subscripts were too complex for the people in data processing, who were thought to be uncomfortable with mathematical notation. Similar arguments revolved around whether arithmetic expressions should be included. The final report of the Short Range Committee, which was completed in December 1959, described the language that was later named COBOL 60.

The language specification for COBOL 60, published by the Government Printing Office in April 1960 (Department of Defense, 1960), was described as “initial.” Revised versions were published in 1961 and 1962 (Department of Defense, 1961, 1962). The language was standardized by the American National Standards Institute (ANSI) group in 1968. The next three revisions were standardized by ANSI in 1974, 1985, and 2002. The language continues to evolve today.

2.6.4 Evaluation The COBOL language originated a number of novel concepts, some of which eventually appeared in other languages. For example, the DEFINE verb of COBOL 60 was the first high-level language construct for macros. More important, hierarchical data structures (records), which first appeared in Plankalkül, were first implemented in COBOL. They have been included in most of the imperative languages designed since then. COBOL was also the

first language that allowed names to be truly connotative, because it allowed both long names (up to 30 characters) and word-connector characters (hyphens).

Overall, the data division is the strong part of COBOL’s design, whereas the procedure division is relatively weak. Every variable is defined in detail in the data division, including the number of decimal digits and the location of the implied decimal point. File records are also described with this level of detail, as are lines to be output to a printer, which makes COBOL ideal for printing accounting reports. Perhaps the most important weakness of the original procedure division was in its lack of functions. Versions of COBOL prior to the 1974 standard also did not allow subprograms with parameters.

Our final comment on COBOL: It was the first programming language whose use was mandated by the Department of Defense (DoD). This mandate came after its initial development, because COBOL was not designed specifically for the DoD. In spite of its merits, COBOL probably would not have survived without that mandate. The poor performance of the early compilers simply made the language too expensive to use. Eventually, of course, compilers became more efficient and computers became much faster and cheaper and had much larger memories. Together, these factors allowed COBOL to succeed, inside and outside DoD. Its appearance led to the electronic mechanization of accounting, an important development by any measure.

The following is an example of a COBOL program. This program reads a file named BAL-FWD-File that contains inventory information about a certain collection of items. Among other things, each item record includes the number currently on hand (BAL-ON-HAND) and the item’s reorder point (BAL- REORDER-POINT). The reorder point is the threshold number of items on hand at which more must be ordered. The program produces a list of items that must be reordered as a file named REORDER-LISTING.

IDENTIFICATION DIVISION.

PROGRAM-ID. PRODUCE-REORDER-LISTING.

ENVIRONMENT DIVISION.

CONFIGURATION SECTION.

SOURCE-COMPUTER. DEC-VAX.

OBJECT-COMPUTER. DEC-VAX.

INPUT-OUTPUT SECTION.

FILE-CONTROL.

SELECT BAL-FWD-FILE ASSIGN TO READER.

SELECT REORDER-LISTING ASSIGN TO LOCAL-PRINTER.

DATA DIVISION.

FILE SECTION.

FD BAL-FWD-FILE

LABEL RECORDS ARE STANDARD

RECORD CONTAINS 80 CHARACTERS.

01 BAL-FWD-CARD.

02 BAL-ITEM-NO PICTURE IS 9(5).

02 BAL-ITEM-DESC PICTURE IS X(20).

02 FILLER PICTURE IS X(5).

02 BAL-UNIT-PRICE PICTURE IS 999V99.

02 BAL-REORDER-POINT PICTURE IS 9(5).

02 BAL-ON-HAND PICTURE IS 9(5).

02 BAL-ON-ORDER PICTURE IS 9(5).

02 FILLER PICTURE IS X(30).

FD REORDER-LISTING

LABEL RECORDS ARE STANDARD

RECORD CONTAINS 132 CHARACTERS.

01 REORDER-LINE.

02 RL-ITEM-NO PICTURE IS Z(5).

02 FILLER PICTURE IS X(5).

02 RL-ITEM-DESC PICTURE IS X(20).

02 FILLER PICTURE IS X(5).

02 RL-UNIT-PRICE PICTURE IS ZZZ.99.

02 FILLER PICTURE IS X(5).

02 RL-AVAILABLE-STOCK PICTURE IS Z(5).

02 FILLER PICTURE IS X(5).

02 RL-REORDER-POINT PICTURE IS Z(5).

02 FILLER PICTURE IS X(71).

WORKING-STORAGE SECTION.

01 SWITCHES.

02 CARD-EOF-SWITCH PICTURE IS X.

01 WORK-FIELDS.

02 AVAILABLE-STOCK PICTURE IS 9(5).

PROCEDURE DIVISION.

000-PRODUCE-REORDER-LISTING.

OPEN INPUT BAL-FWD-FILE.

OPEN OUTPUT REORDER-LISTING.

MOVE "N" TO CARD-EOF-SWITCH.

PERFORM 100-PRODUCE-REORDER-LINE

UNTIL CARD-EOF-SWITCH IS EQUAL TO "Y".

CLOSE BAL-FWD-File.

CLOSE REORDER-LISTING.

STOP RUN.

100-PRODUCE-REORDER-LINE.

PERFORM 110-READ-INVENTORY-RECORD.

IF CARD-EOF-SWITCH IS NOT EQUAL TO "Y"]

PERFORM 120-CALCULATE-AVAILABLE-STOCK

IF AVAILABLE-STOCK IS LESS THAN BAL-REORDER-POINT

PERFORM 130-PRINT-REORDER-LINE.

110-READ-INVENTORY-RECORD.

READ BAL-FWD-FILE RECORD

AT END

MOVE "Y" TO CARD-EOF-SWITCH.

120-CALCULATE-AVAILABLE-STOCK.

ADD BAL-ON-HAND BAL-ON-ORDER

GIVING AVAILABLE-STOCK.

130-PRINT-REORDER-LINE.

MOVE SPACE TO REORDER-LINE.

MOVE BAL-ITEM-NO TO RL-ITEM-NO.

MOVE BAL-ITEM-DESC TO RL-ITEM-DESC.

MOVE BAL-UNIT-PRICE TO RL-UNIT-PRICE.

MOVE AVAILABLE-STOCK TO RL-AVAILABLE-STOCK.

MOVE BAL-REORDER-POINT TO RL-REORDER-POINT.

WRITE REORDER-LINE.

2.7 The Beginnings of Timesharing: Basic Basic (Mather and Waite, 1971) is another programming language that has enjoyed widespread use but has gotten little respect. Like COBOL, it has largely been ignored by computer scientists. Also, like COBOL, in its earliest versions it was inelegant and included only a meager set of control statements.

Basic was very popular on microcomputers in the late 1970s and early 1980s. This followed directly from two of the main characteristics of early versions of Basic. It was easy for beginners to learn, especially those who were not science oriented, and its smaller dialects could be implemented on computers with very small memories.6 When the capabilities of microcomputers grew and other languages were implemented, the use of Basic waned. A strong resurgence in the use of Basic began with the appearance of Visual Basic (Microsoft, 1991) in the early 1990s.

6. Some early microcomputers included Basic interpreters that resided in 4096 bytes of ROM.

2.7.1 Design Process Basic (Beginner’s All-purpose Symbolic Instruction Code) was originally designed at Dartmouth College (now Dartmouth University) in New Hampshire by two mathematicians, John Kemeny and Thomas Kurtz, who, in the early 1960s, developed compilers for a variety of dialects of Fortran and ALGOL 60. Their science students generally had little trouble learning or using those languages in their studies. However, Dartmouth was primarily a liberal arts institution, where science and engineering students made up only about 25 percent of the student body. It was decided in the spring of 1963 to design a new language especially for liberal arts students. This new language

would use terminals as the method of computer access. The goals of the system were as follows:

1. It must be easy for nonscience students to learn and use.

2. It must be “pleasant and friendly.”

3. It must provide fast turnaround for homework.

4. It must allow free and private access.

5. It must consider user time more important than computer time.

The last goal was indeed a revolutionary concept. It was based at least partly on the belief that computers would become significantly cheaper as time went on, which they certainly did.

The combination of the second, third, and fourth goals led to the time-shared aspect of Basic. Only with individual access through terminals by numerous simultaneous users could these goals be met in the early 1960s.

In the summer of 1963, Kemeny began work on the compiler for the first version of Basic, using remote access to a GE 225 computer. Design and coding of the operating system for Basic began in the fall of 1963. At 4:00 a.m. on May 1, 1964, the first program using the timeshared Basic was typed in and run. In June, the number of terminals on the system grew to 11, and by the fall it had ballooned to 20.

2.7.2 Language Overview The original version of Basic was very small and, oddly, was not interactive: There was no way for an executing program to get input data from the user. Programs were typed in, compiled, and run, in a sort of batch-oriented way. The original Basic had only 14 different statement types and a single data type—floating-point. Because it was believed that few of the targeted users would appreciate the difference between integer and floating-point types, the type was referred to as “numbers.” Overall, it was a very limited language,

though quite easy to learn.

2.7.3 Evaluation The most important aspect of the original Basic was that it was the first widely used language that was used through terminals connected to a remote computer.7 Terminals had just begun to be available at that time. Before then, most programs were entered into computers through either punched cards or paper tape.

7. Lisp initially was used through terminals, but it was not widely used in the early 1960s.

Much of the design of Basic came from Fortran, with some minor influence from the syntax of ALGOL 60. Later, it grew in a variety of ways, with little or no effort made to standardize it. The American National Standards Institute issued a Minimal Basic standard (ANSI, 1978b), but this represented only the bare minimum of language features. In fact, the original Basic was very similar to Minimal Basic.

Although it may seem surprising, Digital Equipment Corporation used a rather elaborate version of Basic named Basic-PLUS to write significant portions of their largest operating system for the PDP-11 minicomputers, RSTS, in the 1970s.

Basic has been criticized for the poor structure of programs written in it, among other things. By the evaluation criteria discussed in Chapter 1, specifically readability and reliability, the language does indeed fare very poorly. Clearly, the early versions of the language were not meant for and should not have been used for serious programs of any significant size. Later versions are much better suited to such tasks.

The resurgence of Basic in the 1990s was driven by the appearance of Visual Basic (VB). VB became widely used in large part because it provided a simple way of building graphical user interfaces (GUIs), hence the name Visual Basic. When .NET appeared, a new version of VB came with it,

VB.NET. Although it was a significant departure from earlier versions of VB, it quickly displaced the older language. Perhaps the most important difference between VB and the .NET version is that the later version fully supports object-oriented programming.

The following is an example of a Basic program:

REM Basic Example Program

REM Input: An integer, listlen, where listlen is less

REM than 100, followed by listlen-integer values

REM Output: The number of input values that are greater

REM than the average of all input values

DIM intlist(99)

result = 0

sum = 0

INPUT listlen

IF listlen > 0 AND listlen < 100 THEN

REM Read input into an array and compute the sum

FOR counter = 1 TO listlen

INPUT intlist(counter)

sum = sum + intlist(counter)

NEXT counter

REM Compute the average

average = sum / listlen

REM Count the number of input values that are > average

FOR counter = 1 TO listlen

IF intlist(counter) > average

THEN result = result + 1

NEXT counter

REM Print the result

PRINT "The number of values that are > average is:";

result

ELSE

PRINT "Error-input list length is not legal"

END IF

END

interview User Design and Language Design

ALAN COOPER

Best-selling author of About Face: The Essentials of User Interface Design, Alan Cooper also had a large hand in designing what can be touted as the language with the most concern for user interface design, Visual Basic. For him, it all comes down to a vision for humanizing technology.

SOME INFORMATION ON THE BASICS How did you get started in all of this? I’m a high school dropout with an associate degree in programming from a California community college. My first job was as a programmer for American President Lines (one of the United States’ oldest ocean transportation companies) in San Francisco. Except for a few months here and there, I’ve remained self-employed.

What is your current job? Founder and chairman of Cooper, the company that humanizes technology (www.cooper.com).

What is or was your favorite job? Interaction design consultant.

You are very well known in the fields of language design and user interface design. Any thoughts on designing languages versus designing software, versus designing anything else? It’s pretty much the same in the world of software: Know your user.

ABOUT THAT EARLY WINDOWS RELEASE In the 1980s, you started using Windows and have talked about being lured by its plusses: the graphical user interface support and the dynamically linked library that let you create tools that configured themselves. What about the parts of Windows that you eventually helped shape? I was very impressed by Microsoft’s inclusion of support for practical multitasking in Windows. This included dynamic relocation and interprocess communications.

MSDOS.exe was the shell program for the first few releases of Windows. It was a terrible program, and I believed that it could be improved dramatically, and I was the guy to do it. In my spare time, I immediately began to write a better shell program than the one Windows came with. I called it Tripod. Microsoft’s original shell, MSDOS.exe, was one of the main stumbling blocks to the initial success of Windows. Tripod attempted to solve the problem by being easier to use and to configure.

When was that “Aha!” moment? It wasn’t until late in 1987, when I was interviewing a corporate client, that the key design strategy for Tripod popped into my head. As the IS manager explained to me his need to create and publish a wide range of shell solutions to his disparate user base, I realized the conundrum that there is no such thing as an ideal shell. Every user would need their own personal shell, configured to their own needs and skill levels. In an instant, I perceived the solution to the shell design problem: It would be a shell construction set; a tool where each user would be able to construct exactly the shell that he or she needed for a unique mix of applications and training.

What is so compelling about the idea of a shell that can be individualized? Instead of me telling the users what the ideal shell was, they could design their own, personalized ideal shell. With a customizable shell, a programmer would create a shell that was powerful and wide ranging but also somewhat dangerous, whereas an IT manager would create a shell that could be given to a desk clerk that exposed only those few application-specific tools that the clerk used.

“MSDOS.exe was the shell program for the first few releases of Windows. It was a terrible program, and I believed that it could be improved dramatically, and I was the guy to do it. In my spare time, I immediately began to write a better shell program than the one Windows came with.”

How did you get from writing a shell program to collaborating with Microsoft? Tripod and Ruby are the same thing. After I signed a deal with Bill Gates, I changed the name of the prototype from Tripod to Ruby. I then used the Ruby prototype as prototypes should be used: as a disposable model for constructing release-quality code. Which is what I did. MS took the release version of Ruby and added QuickBasic to it, creating VB. All of those original innovations were in Tripod/Ruby.

RUBY AS THE INCUBATOR FOR VISUAL BASIC Let’s revisit your interest in early Windows and that DLL feature. The DLL wasn’t a thing, it was a facility in the OS. It allowed a programmer to build code objects that could be linked to at run time as opposed to only at compile time. This is what allowed me to invent the dynamically extensible parts of VB, where controls can be added by third-party vendors.

The Ruby product embodied many significant advances in software design, but two of them stand out as exceptionally successful. As I mentioned, the dynamic linking capability of Windows had always intrigued me, but having the tools and knowing what to do with them were two different things. With

Ruby, I finally found two practical uses for dynamic linking, and the original program contained both. First, the language was both installable and could be extended dynamically. Second, the palette of gizmos could be added to dynamically.

Was your language in Ruby the first to have a dynamic linked library and to be linked to a visual front end? As far as I know, yes.

Using a simple example, what would this enable a programmer to do with his or her program? Purchase a control, such as a grid control, from a third-party vendor, install it on his or her computer, and have the grid control appear as an integral part of the language, including the visual programming front end.

Why do they call you “the father of Visual Basic”? Ruby came with a small language, one suited only for executing the dozen or so simple commands that a shell program needs. However, this language was implemented as a chain of DLLs, any number of which could be installed at run time. The internal parser would identify a verb and then pass it along the chain of DLLs until one of them acknowledged that it knew how to process the verb. If all of the DLLs passed, there was a syntax error. From our earliest discussions, both Microsoft and I had entertained the idea of growing the language, possibly even replacing it altogether with a “real” language. C was the candidate most frequently mentioned, but eventually, Microsoft took advantage of this dynamic interface to unplug our little shell language and replace it entirely with Quick-​Basic. This new marriage of language to visual front end was static and permanent, and although the original dynamic interface made the coupling possible, it was lost in the process.

SOME FINAL COMMENTS ON NEW IDEAS In the world of programming and programming tools, including languages and environments, what projects most interest you? I’m interested in creating programming tools that are designed to help users

instead of programmers.

What’s the most critical rule, famous quote, or design idea to keep in mind? Bridges are not built by engineers. They are built by ironworkers.

Similarly, software programs are not built by engineers. They are built by programmers.

2.8 Everything for Everybody: PL/I PL/I represents the first large-scale attempt to design a language that could be used for a broad spectrum of application areas. All previous and most subsequent languages have focused on one particular application area, such as science, artificial intelligence, or business.

2.8.1 Historical Background Like Fortran, PL/I was developed as an IBM product. By the early 1960s, the users of computers in industry had settled into two separate and quite different camps: scientific and business. From the IBM point of view, scientific programmers could use either the large-scale 7090 or the small- scale 1620 IBM computers. This group used floating-point data and arrays extensively. Fortran was the primary language, although some assembly language also was used. They had their own user group, SHARE, and had little contact with anyone who worked on business applications.

For business applications, people used the large 7080 or the small 1401 IBM computers. They needed the decimal and character string data types, as well as elaborate and efficient input and output facilities. They used COBOL, although in early 1963 when the PL/I story begins, the conversion from assembly language to COBOL was far from complete. This category of users also had its own user group, GUIDE, and seldom had contact with scientific users.

In early 1963, IBM planners perceived the beginnings of a change in this situation. The two widely separated computer user groups were moving toward each other in ways that were thought certain to create problems. Scientists began to gather large files of data to be processed. This data required more sophisticated and more efficient input and output facilities. Business applications people began to use regression analysis to build management information systems, which required floating-point data and

arrays. It began to appear that computing installations would soon require two separate computers and technical staffs, supporting two very different programming languages.8

8. At the time, large computer installations required both full-time hardware and full-time system software maintenance staff.

These perceptions naturally led to the concept of designing a single universal computer that would be capable of doing both floating-point and decimal arithmetic, and therefore both scientific and business applications. Thus was born the concept of the IBM System/360 line of computers. Along with this came the idea of a programming language that could be used for both business and scientific applications. For good measure, features to support systems programming and list processing were thrown in. Therefore, the new language was to replace Fortran, COBOL, Lisp, and the systems applications of assembly language.

2.8.2 Design Process The design effort began when IBM and SHARE formed the Advanced Language Development Committee of the SHARE Fortran Project in October 1963. This new committee quickly met and formed a subcommittee called the 3 ⨯ 3 Committee, so named because it had three members from IBM and three from SHARE. The 3 ⨯ 3 Committee met for three or four days every other week to design the language.

As with the Short Range Committee for COBOL, the initial design was scheduled for completion in a remarkably short time. Apparently, regardless of the scope of a language design effort, in the early 1960s the prevailing belief was that it could be done in three months. The first version of PL/I, which was then named Fortran VI, was supposed to be completed by December, less than three months after the committee was formed. The committee pleaded successfully on two different occasions for extensions, moving the due date back to January and then to late February 1964.

The initial design concept was that the new language would be an extension

of Fortran IV, maintaining compatibility, but that goal was dropped quickly along with the name Fortran VI. Until 1965, the language was known as NPL (New Programming Language). The first published report on NPL was given at the SHARE meeting in March 1964. A more complete description followed in April, and the version that would actually be implemented was published in December 1964 (IBM, 1964) by the compiler group at the IBM Hursley Laboratory in England, which was chosen to do the implementation. In 1965, the name was changed to PL/I to avoid the confusion of the name NPL with the National Physical Laboratory in England. If the compiler had been developed outside the United Kingdom, the name might have remained NPL.

2.8.3 Language Overview Perhaps the best single-sentence description of PL/I is that it included what were then considered the best parts of ALGOL 60 (recursion and block structure), Fortran IV (separate compilation with communication through global data), and COBOL 60 (data structures, input/output, and report- generating facilities), along with an extensive collection of new constructs, all somehow cobbled together. Because PL/I is no longer a popular language, we will not attempt, even briefly, to discuss all the features of the language, or even its most controversial constructs. Instead, we will mention some of the language’s contributions to the pool of knowledge of programming languages.

PL/I was the first programming language to have the following facilities:

Programs were allowed to create concurrently executing subprograms. Although this was a good idea, it was poorly developed in PL/I.

It was possible to detect and handle 23 different types of exceptions, or run-time errors.

Subprograms were allowed to be used recursively, but the capability could be disabled, allowing more efficient linkage for nonrecursive subprograms.

Pointers were included as a data type.

Cross-sections of arrays could be referenced. For example, the third row of a matrix could be referenced as if it were a single-dimensioned array.

2.8.4 Evaluation Any evaluation of PL/I must begin by recognizing the ambitiousness of the design effort. In retrospect, it appears naive to think that so many constructs could have been combined successfully. However, that judgment must be tempered by acknowledging that there was little language design experience at the time. Overall, the design of PL/I was based on the premise that any construct that was useful and could be implemented should be included, with insufficient concern about how a programmer could understand and make effective use of such a collection of constructs and features. Edsger Dijkstra, in his Turing Award Lecture (Dijkstra, 1972), made one of the strongest criticisms of the complexity of PL/I: “I absolutely fail to see how we can keep our growing programs firmly within our intellectual grip when by its sheer baroqueness the programming language—our basic tool, mind you!— already escapes our intellectual control.”

In addition to the problem with the complexity due to its large size, PL/I suffered from a number of what are now considered to be poorly designed constructs. Among these were pointers, exception handling, and concurrency, although we must point out that in all cases, these constructs had not appeared in any previous language.

In terms of usage, PL/I must be considered at least a partial success. In the 1970s, it enjoyed significant use in both business and scientific applications. It was also widely used during that time as an instructional vehicle in colleges, primarily in several subset forms, such as PL/C (Cornell, 1977) and PL/CS (Conway and Constable, 1976).

The following is an example of a PL/I program:

/* PL/I PROGRAM EXAMPLE

INPUT: AN INTEGER, LISTLEN, WHERE LISTLEN IS LESS THAN

100, FOLLOWED BY LISTLEN-INTEGER VALUES

OUTPUT: THE NUMBER OF INPUT VALUES THAT ARE GREATER THAN

THE AVERAGE OF ALL INPUT VALUES */

PLIEX: PROCEDURE OPTIONS (MAIN);

DECLARE INTLIST (1:99) FIXED.

DECLARE (LISTLEN, COUNTER, SUM, AVERAGE, RESULT) FIXED;

SUM = 0;

RESULT = 0;

GET LIST (LISTLEN);

IF (LISTLEN > 0) & (LISTLEN < 100) THEN

DO;

/* READ INPUT DATA INTO AN ARRAY AND COMPUTE THE SUM */

DO COUNTER = 1 TO LISTLEN;

GET LIST (INTLIST (COUNTER));

SUM = SUM + INTLIST (COUNTER);

END;

/* COMPUTE THE AVERAGE */

AVERAGE = SUM / LISTLEN;

/* COUNT THE NUMBER OF VALUES THAT ARE > AVERAGE */

DO COUNTER = 1 TO LISTLEN;

IF INTLIST (COUNTER) > AVERAGE THEN

RESULT = RESULT + 1;

END;

/* PRINT RESULT */

PUT SKIP LIST ('THE NUMBER OF VALUES > AVERAGE IS:');

PUT LIST (RESULT);

END;

ELSE

PUT SKIP LIST ('ERROR-INPUT LIST LENGTH IS ILLEGAL');

END PLIEX;

2.9 Two Early Dynamic Languages: APL and SNOBOL The structure of this section is different from that of the other sections because the languages discussed here are very different. Neither APL nor SNOBOL had much influence on later mainstream languages.9 Some of the interesting features of APL are discussed later in the book.

9. However, they have had some influence on some nonmainstream languages (J is based on APL, ICON is based on SNOBOL, and AWK is partially based on SNOBOL).

In appearance and in purpose, APL and SNOBOL are quite different. They share two fundamental characteristics, however: dynamic typing and dynamic storage allocation. Variables in both languages are essentially untyped. A variable acquires a type when it is assigned a value, at which time it assumes the type of the value assigned. Storage is allocated to a variable only when it is assigned a value, because before that there is no way to know the amount of storage that will be needed.

2.9.1 Origins and Characteristics of APL APL (Brown et al., 1988) was designed around 1960 by Kenneth E. Iverson at IBM. It was not originally designed to be an implemented programming language but rather was intended to be a vehicle for describing computer architecture. APL was first described in the book from which it gets its name, A Programming Language (Iverson, 1962). In the mid-1960s, the first implementation of APL was developed at IBM.

APL has a large number of powerful operators that are specified with a large

number of symbols, which created a problem for implementors. ​Initially, APL was used through IBM printing terminals. These terminals had special optional print balls that provided the odd character set required by the ​- language. One reason APL has so many operators is that it provides a large number of unit operations on arrays. For example, the transpose of any matrix is done with a single operator. The large collection of operators provides very high expressivity but also makes APL programs difficult to read. Therefore, people think of APL as a language that is best used for “throw-away” programming. Although programs can be written quickly, they should be discarded after use because they are difficult to maintain.

APL has been around for over 55 years and is still used today, although not widely. Furthermore, it has not changed a great deal over its lifetime.

2.9.2 Origins and Characteristics of SNOBOL SNOBOL (pronounced “snowball”; Griswold et al., 1971) was designed in the early 1960s by three people at Bell Laboratories: D. J. Farber, R. E. Griswold, and I. P. Polonsky (Farber et al., 1964). It was designed specifically for text processing. The heart of SNOBOL is a collection of powerful operations for string pattern matching. One of the early applications of SNOBOL was for writing text editors. Because the dynamic nature of SNOBOL makes it slower than alternative languages, it is no longer used for such programs. However, SNOBOL is still a live and supported language that is used for a variety of text-processing tasks in several different application areas.

2.10 The Beginnings of Data Abstraction: SIMULA 67 Although SIMULA 67 never achieved widespread use and had little impact on the programmers and computing of its time, some of the constructs it introduced make it historically important.

2.10.1 Design Process Two Norwegians, Kristen Nygaard and Ole-Johan Dahl, developed the language SIMULA I between 1962 and 1964 at the Norwegian Computing Center (NCC) in Oslo. They were primarily interested in using computers for simulation but also worked in operations research. SIMULA I was designed exclusively for system simulation and was first implemented in late 1964 on a UNIVAC 1107 computer.

As soon as the SIMULA I implementation was completed, Nygaard and Dahl began efforts to extend the language by adding new features and modifying some existing constructs in order to make the language useful for ​general- purpose applications. The result of this work was SIMULA 67, whose design was first presented publicly in March 1967 (Dahl and Nygaard, 1967). We will discuss only SIMULA 67, although some of the features of interest in SIMULA 67 are also in SIMULA I.

2.10.2 Language Overview SIMULA 67 is an extension of ALGOL 60, taking both block structure and the control statements from that language. The primary deficiency of ALGOL 60 (and other languages at that time) for simulation applications was the design of its subprograms. Simulation requires subprograms that are allowed to restart at the position where they previously stopped. Subprograms with

this kind of control are known as coroutines because the caller and called subprograms have a somewhat equal relationship with each other, rather than the rigid master/slave relationship they have in most imperative languages.

To provide support for coroutines in SIMULA 67, the class construct was developed. This was an important development because the concept of data abstraction began with it and data abstraction provides the foundation for object-oriented programming.

It is interesting to note that the important concept of data abstraction was not developed and attributed to the class construct until 1972, when Hoare (1972) recognized the connection.

2.11 Orthogonal Design: ALGOL 68 ALGOL 68 was the source of several new ideas in language design, some of which were subsequently adopted by other languages. We include it here for that reason, even though it never achieved widespread use in either Europe or the United States.

2.11.1 Design Process The development of the ALGOL family did not end when the revised report on ALGOL 60 appeared in 1962, although it was six years until the next design iteration was published. The resulting language, ALGOL 68 (van Wijngaarden et al., 1969), was dramatically different from its predecessor.

One of the most interesting innovations of ALGOL 68 was one of its primary design criteria: orthogonality. Recall our discussion of orthogonality in Chapter 1. The use of orthogonality resulted in several innovative features of ALGOL 68, one of which is described in the following section.

2.11.2 Language Overview One important result of orthogonality in ALGOL 68 was its inclusion of user- defined data types. Earlier languages, such as Fortran, included only a few basic data structures. PL/I included a larger number of data structures, which made it harder to learn and difficult to implement, but it still could not provide an appropriate data structure for every need.

The approach of ALGOL 68 to data structures was to provide a few primitive types and structures and allow the user to combine those primitives to define a large number of different structures. This provision for user-defined data types was carried over to some extent into all of the major imperative languages designed since then. User-defined data types are valuable because

they allow the user to design data abstractions that fit particular problems very closely. All aspects of data types are discussed in Chapter 6.

As another first in the area of data types, ALGOL 68 introduced the kind of dynamic arrays that will be termed implicit heap-dynamic in Chapter 5. A dynamic array is one in which the declaration does not specify subscript bounds. Assignments to a dynamic array cause allocation of required storage. In ALGOL 68, dynamic arrays are called flex arrays.

2.11.3 Evaluation ALGOL 68 includes a significant number of features that had not been previously used. Its use of orthogonality, which some may argue was overdone, was nevertheless revolutionary.

ALGOL 68 repeated one of the sins of ALGOL 60, however, and it was an important factor in its limited popularity. The language was described using an elegant and concise but also unknown metalanguage. Before one could read the language-describing document (van Wijngaarden et al., 1969), he or she had to learn the new metalanguage, called van Wijngaarden grammars, which were far more complex than BNF. To make matters worse, the designers invented a collection of words to explain the grammar and the language. For example, keywords were called indicants, substring extraction was called trimming, and the process of a subprogram execution was called a coercion of deproceduring, which might be meek, firm, or something else.

It is natural to contrast the design of PL/I with that of ALGOL 68, because they appeared only a few years apart. ALGOL 68 achieved writability by the principle of orthogonality: a few primitive concepts and the unrestricted use of a few combining mechanisms. PL/I achieved writability by including a large number of fixed constructs. ALGOL 68 extended the elegant simplicity of ALGOL 60, whereas PL/I simply threw together the features of several languages to attain its goals. Of course, it must be remembered that the goal of PL/I was to provide a unified tool for a broad class of problems, whereas ALGOL 68 was targeted to a single class: scientific applications.

PL/I achieved far greater acceptance than ALGOL 68, due largely to IBM’s promotional efforts and the problems of understanding and implementing ALGOL 68. Implementation was a difficult problem for both, but PL/I had the resources of IBM to apply to building a compiler. ALGOL 68 enjoyed no such benefactor.

2.12 Some Early Descendants of the ALGOLs All imperative languages owe some of their design to ALGOL 60 and/or ALGOL 68. This section discusses some of the early descendants of these languages.

2.12.1 Simplicity by Design: Pascal

2.12.1.1 Historical Background Niklaus Wirth (Wirth is pronounced “Virt”) was a member of the International Federation of Information Processing (IFIP) Working Group 2.1, which was created to continue the development of ALGOL in the mid- 1960s. In August 1965, Wirth and C. A. R. (“Tony”) Hoare contributed to that effort by presenting to the group a somewhat modest proposal for additions and modifications to ALGOL 60 (Wirth and Hoare, 1966). The majority of the group rejected the proposal as being too small an improvement over ALGOL 60. Instead, a much more complex revision was developed, which eventually became ALGOL 68. Wirth, along with a few other group members, did not believe that the ALGOL 68 report should have been released, based on the complexity of both the language and the metalanguage used to describe it. This position later proved to have some validity because the ALGOL 68 documents, and therefore the language, were indeed found to be challenging by the computing community.

The Wirth and Hoare version of ALGOL 60 was named ALGOL-W. It was implemented at Stanford University and was used primarily as an instructional vehicle, but only at a few universities. The primary contributions of ALGOL-W were the value-result method of passing parameters and the case statement for multiple selection. The value-result

method is an alternative to ALGOL 60’s pass-by-name method. Both are discussed in Chapter 9. The case statement is discussed in Chapter 8.

Wirth’s next major design effort, again based on ALGOL 60, was his most successful: Pascal.10 The original published definition of Pascal appeared in 1971 (Wirth, 1971). This version was modified somewhat in the implementation process and is described in Wirth (1973). The features that are often ascribed to Pascal in fact came from earlier languages. For example, user-defined data types were introduced in ALGOL 68, the case statement in ALGOL-W, and Pascal’s records are similar to those of COBOL and PL/I.

10. Pascal is named after Blaise Pascal, a seventeenth-century French philosopher and mathematician who invented the first mechanical adding machine in 1642 (among other things).

2.12.1.2 Evaluation The largest impact of Pascal was on the teaching of programming. In 1970, most students of computer science, engineering, and science were introduced to programming with Fortran, although some universities used PL/I, languages based on PL/I, and ALGOL-W. By the mid-1970s, Pascal had become the most widely used language for this purpose. This was quite natural, because Pascal was designed specifically for teaching programming. It was not until the late 1990s that Pascal was no longer the most commonly used language for teaching programming in colleges and universities.

Because Pascal was designed as a teaching language, it lacks several features that are essential for many kinds of applications. The best example of this is the impossibility of writing a subprogram that takes as a parameter an array of variable length. Another example is the lack of any separate compilation capability. These deficiencies naturally led to many nonstandard dialects, such as Turbo Pascal.

Pascal’s popularity, for both teaching programming and other applications, was based primarily on its remarkable combination of simplicity and expressivity. Although there are some insecurities in Pascal, it is still a

relatively safe language, particularly when compared with Fortran or C. By the mid-1990s, the popularity of Pascal was on the decline, both in industry and in universities, primarily due to the rise of Modula-2, Ada, and C++, all of which included features not available in Pascal.

The following is an example of a Pascal program:

{Pascal Example Program

Input: An integer, listlen, where listlen is less than

100, followed by listlen-integer values

Output: The number of input values that are greater than

the average of all input values }

program pasex (input, output);

type intlisttype = array [1..99] of integer;

var

intlist : intlisttype;

listlen, counter, sum, average, result : integer;

begin

result := 0;

sum := 0;

readln (listlen);

if ((listlen > 0) and (listlen < 100)) then

begin

{ Read input into an array and compute the sum }

for counter := 1 to listlen do

begin

readln (intlist[counter]);

sum := sum + intlist[counter]

end;

{ Compute the average }

average := sum / listlen;

{ Count the number of input values that are > average }

for counter := 1 to listlen do

if (intlist[counter] > average) then

result := result + 1;

{ Print the result }

writeln ('The number of values > average is:',

result)

end { of the then clause of if (( listlen > 0 ... }

else

writeln ('Error-input list length is not legal')

end.

2.12.2 A Portable Systems

Language: C Like Pascal, C contributed little to the previously known collection of language features, but it has been widely used over a long period of time. Although originally designed for systems programming, C is well suited for a wide variety of applications.

2.12.2.1 Historical Background C’s ancestors include CPL, BCPL, B, and ALGOL 68. CPL was developed at Cambridge University in the early 1960s. BCPL is a simple systems language, also developed at Cambridge, this time by Martin Richards in 1967 (Richards, 1969).

The first work on the UNIX operating system was done in the late 1960s by Ken Thompson at Bell Laboratories. The first version was written in assembly language. The first high-level language implemented under UNIX was B, which was based on BCPL. B was designed and implemented by Thompson in 1970.

Neither BCPL nor B is a typed language, which is an oddity among high- level languages, although both are much lower-level than a language such as Java. Being untyped means that all data are considered machine words, which, although simple, leads to many complications and insecurities. For example, there is the problem of specifying floating-point rather than integer arithmetic in an expression. In one implementation of BCPL, the variable operands of a floating-point operation were preceded by periods. Variable operands not preceded by periods were considered to be integers. An alternative to this would have been to use different symbols for the floating- point operators.

This problem, along with several others, led to the development of a new typed language based on B. Originally called NB but later named C, it was designed and implemented by Dennis Ritchie at Bell Laboratories in 1972 (Kernighan and Ritchie, 1978). In some cases through BCPL, and in other

cases directly, C was influenced by ALGOL 68. This is seen in its for and switch statements, in its assigning operators, and in its treatment of pointers.

The only “standard” for C in its first decade and a half was the book by Kernighan and Ritchie (1978).11 Over that time span, the language slowly evolved, with different implementors adding different features. In 1989, ANSI produced an official description of C (ANSI, 1989), which included many of the features that implementors had already incorporated into the language. This standard was updated in 1999 (ISO, 1999). This later version includes a few significant changes to the language. Among these are a complex data type, a Boolean data type, and C++-style comments (//). We will refer to the 1989 version, which has long been called ANSI C, as C89; we will refer to the 1999 version as C99.

11. This language is often referred to as “K & R C.”

2.12.2.2 Evaluation C has adequate control statements and data-structuring facilities to allow its use in many application areas. It also has a rich set of operators that provide a high degree of expressiveness.

One of the most important reasons why C is both liked and disliked is its lack of complete type checking. For example, in versions before C99, functions could be written for which parameters were not type checked. Those who like C appreciate the flexibility; those who do not like it find it too insecure. A major reason for its great increase in popularity in the 1980s was that a compiler for it was part of the widely used UNIX operating system. This inclusion in UNIX provided an essentially free and quite good compiler that was available to programmers on many different kinds of computers.

The following is an example of a C program:

/* C Example Program

Input: An integer, listlen, where listlen is less than

100, followed by listlen-integer values

Output: The number of input values that are greater than

the average of all input values */

int main (){

int intlist[99], listlen, counter, sum, average, result;

sum = 0;

result = 0;

scanf("%d", &listlen);

if ((listlen > 0) && (listlen < 100)) {

/* Read input into an array and compute the sum */

for (counter = 0; counter < listlen; counter++) {

scanf("%d", &intlist[counter]);

sum += intlist[counter];

}

/* Compute the average */

average = sum / listlen;

/* Count the input values that are > average */

for (counter = 0; counter < listlen; counter++)

if (intlist[counter] > average) result++;

/* Print result */

printf("Number of values > average is:%d\n", result);

}

else

printf("Error-input list length is not legal\n");

}

2.13 Programming Based on Logic: Prolog Simply put, logic programming is the use of a formal logic notation to communicate computational processes to a computer. Predicate calculus is the notation used in current logic programming languages.

Programming in logic programming languages is nonprocedural. ​Programs in such languages do not state exactly how a result is to be computed but rather describe the necessary form and/or characteristics of the result. What is needed to provide this capability in logic programming languages is a concise means of supplying the computer with both the relevant information and an inferencing process for computing desired results. Predicate calculus supplies the basic form of communication to the computer, and the proof method, named resolution, developed first by Robinson (1965), supplies the inferencing technique.

2.13.1 Design Process During the early 1970s, Alain Colmerauer and Phillippe Roussel in the ​- Artificial Intelligence Group at the University of Aix-Marseille, together with Robert ​Kowalski of the Department of Artificial Intelligence at the University of Edinburgh, developed the fundamental design of Prolog. The primary components of Prolog are a method for specifying predicate calculus propositions and an implementation of a restricted form of resolution. Both predicate calculus and resolution are described in Chapter 16. The first Prolog interpreter was developed at Marseille in 1972. The version of the language that was implemented is described in Roussel (1975). The name Prolog is from programming logic.

2.13.2 Language Overview

Prolog programs consist of collections of statements. Prolog has only a few kinds of statements, but they can be complex.

One common use of Prolog is as a kind of intelligent database. This application provides a simple framework for discussing the Prolog language.

The database of a Prolog program consists of two kinds of statements: facts and rules. The following are examples of fact statements:

mother(joanne, jake).

father(vern, joanne).

These state that joanne is the mother of jake, and vern is the father of joanne.

An example of a rule statement is

grandparent(X, Z) :- parent(X, Y), parent(Y, Z).

This states that it can be deduced that X is the grandparent of Z if it is true that X is the parent of Y and Y is the parent of Z, for some specific values for the variables X, Y, and Z.

The Prolog database can be interactively queried with goal statements, an example of which is

father(bob, darcie).

This asks if bob is the father of darcie. When such a query, or goal, is presented to the Prolog system, it uses its resolution process to attempt to determine the truth of the statement. If it can conclude that the goal is true, it displays “true.” If it cannot prove it, it displays “false.”

2.13.3 Evaluation In the 1980s, there was a relatively small group of computer scientists who believed that logic programming provided the best hope for escaping from the complexity of imperative languages, and also from the enormous problem

of producing the large amount of reliable software that was needed. So far, however, there are two major reasons why logic programming has not become more widely used. First, as with some other nonimperative approaches, programs written in logic languages thus far have proven to be highly inefficient relative to equivalent imperative programs. Second, it has been determined that it is an effective approach for only a few relatively small areas of application: certain kinds of database management systems and some areas of AI.

There is a dialect of Prolog that supports object-oriented programming: Prolog++ (Moss, 1994). Logic programming and Prolog are described in greater detail in Chapter 16.

2.14 History’s Largest Design Effort: Ada The Ada language is the result of the most extensive and expensive language design effort ever undertaken. The following paragraphs briefly describe the evolution of Ada.

2.14.1 Historical Background The Ada language was developed for the Department of Defense (DoD), so the state of their computing environment was instrumental in determining its form. By 1974, over half of the applications of computers in DoD were embedded systems. An embedded system is one in which the computer hardware is embedded in the device it controls or for which it provides services. Software costs were rising ​rapidly, primarily because of the increasing complexity of systems. More than 450 different programming languages were in use for DoD projects, and none of them was standardized by DoD. Every defense contractor could define a new and different language for every contract.12 Because of this language proliferation, application software was rarely reused. Furthermore, no software development tools were created (because they are usually language dependent). A great many languages were in use, but none was actually suitable for embedded systems applications. For these reasons, in 1974, the Army, Navy, and Air Force each independently proposed the development of a single high-level language for embedded systems.

12. This result was largely due to the widespread use of assembly language for embedded systems, along with the fact that most embedded systems used specialized processors.

2.14.2 Design Process

Noting this widespread interest, in January 1975, Malcolm Currie, director of Defense Research and Engineering, formed the High-Order Language Working Group (HOLWG), initially headed by Lt. Col. William Whitaker of the Air Force. The HOLWG had representatives from all of the military services and liaisons with Great Britain, France, and what was then West Germany. Its initial charter was to do the following:

Identify the requirements for a new DoD high-level language.

Evaluate existing languages to determine whether there was a viable candidate.

Recommend adoption or implementation of a minimal set of programming languages.

In April 1975, the HOLWG produced the Strawman requirements document for the new language (Department of Defense, 1975a). This was distributed to military branches, federal agencies, selected industrial and university representatives, and interested parties in Europe.

The Strawman document was followed by Woodenman (Department of Defense, 1975b) in August 1975, Tinman (Department of Defense, 1976) in January 1976, Ironman (Department of Defense, 1977) in January 1977, and finally Steelman (Department of Defense, 1978) in June 1978.

After a tedious process, the many submitted proposals for the language were narrowed down to four finalists, all of which were based on Pascal. In May 1979, the Cii Honeywell/Bull language design proposal was chosen from the four finalists as the design that would be used. The Cii Honeywell/Bull design team in France, the only foreign competitor among the final four, was led by Jean Ichbiah.

In the spring of 1979, Jack Cooper of the Navy Materiel Command recommended the name for the new language, Ada, which was then adopted. The name commemorates Augusta Ada Byron (1815–1851), countess of Lovelace, mathematician, and daughter of poet Lord Byron. She is generally recognized as being the world’s first programmer. She worked with Charles Babbage on his mechanical computers, the Difference and Analytical

Engines, writing programs for several numerical processes.

The design and the rationale for Ada were published by ACM in its SIGPLAN Notices (ACM, 1979) and distributed to a readership of more than 10,000 people. A public test and evaluation conference was held in October 1979 in Boston, with representatives from over 100 organizations from the United States and Europe. By November, more than 500 language reports had been received from 15 different countries. Most of the reports suggested small modifications rather than drastic changes and outright rejections. Based on the language reports, the next version of the requirements specification, the Stoneman document (Department of Defense, 1980), was released in February 1980.

A revised version of the language design was completed in July 1980 and was accepted as MIL-STD 1815, the standard Ada Language Reference Manual. The number 1815 was chosen because it was the year of the birth of Augusta Ada Byron. Another revised version of the Ada Language Reference Manual was released in July 1982. In 1983, the American National Standards Institute standardized Ada. This “final” official version is described in Goos and Hartmanis (1983). The Ada language design was then frozen for a minimum of five years.

2.14.3 Language Overview This subsection briefly describes four of the major contributions of the Ada language.

Packages in the Ada language provide the means for encapsulating data objects, specifications for data types, and procedures. This, in turn, provides the support for the use of data abstraction in program design.

The Ada language includes extensive facilities for exception handling, which allow the programmer to gain control after any one of a wide variety of exceptions, or run-time errors, has been detected.

Program units can be generic in Ada. For example, it is possible to write a

sort procedure that uses an unspecified type for the data to be sorted. Such a generic procedure must be instantiated for a specified type before it can be used, which is done with a statement that causes the compiler to generate a version of the procedure with the given type. The availability of such generic units increases the range of program units that might be reused, rather than duplicated, by programmers.

The Ada language also provides for concurrent execution of special program units, named tasks, using the rendezvous mechanism. Rendezvous is the name of a method of intertask communication and synchronization.

2.14.4 Evaluation Perhaps the most important aspects of the design of the Ada language to consider are the following:

Because the design was competitive, there were no limits on participation.

The Ada language embodies most of the concepts of software engineering and language design of the late 1970s. Although one can question the actual approaches used to incorporate these features, as well as the wisdom of including such a large number of features in a language, most agree that the features are valuable.

Although most people did not anticipate it, the development of a compiler for the Ada language was a difficult task. Only in 1985, almost four years after the language design was completed, did truly usable Ada compilers begin to appear.

The most serious criticism of Ada in its first few years was that it was too large and too complex. In particular, Hoare (1981) stated that it should not be used for any application where reliability is critical, which is precisely the type of application for which it was designed. On the other hand, others have praised it as the epitome of language design for its time. In fact, even Hoare eventually softened his view of the language.

The following is an example of an Ada program:

-- Ada Example Program

-- Input: An integer, List_Len, where List_Len is less

-- than 100, followed by List_Len-integer values

-- Output: The number of input values that are greater

-- than the average of all input values

with Ada.Text_IO, Ada.Integer.Text_IO;

use Ada.Text_IO, Ada.Integer.Text_IO;

procedure Ada_Ex is

type Int_List_Type is array (1..99) of Integer;

Int_List : Int_List_Type;

List_Len, Sum, Average, Result : Integer;

begin

Result:= 0;

Sum := 0;

Get (List_Len);

if (List_Len > 0) and (List_Len < 100) then

-- Read input data into an array and compute the sum

for Counter := 1 .. List_Len loop

Get (Int_List(Counter));

Sum := Sum + Int_List(Counter);

end loop;

-- Compute the average

Average := Sum / List_Len;

-- Count the number of values that are > average

for Counter := 1 .. List_Len loop

if Int_List(Counter) > Average then

Result:= Result+ 1;

end if;

end loop;

-- Print result

Put ("The number of values > average is:");

Put (Result);

New_Line;

else

Put_Line ("Error-input list length is not legal");

end if;

end Ada_Ex;

2.14.5 Ada 95 and Ada 2005 Two of the most important new features of Ada 95 are described briefly in the following paragraphs. In the remainder of the book, we will use the name

Ada 83 for the original version and Ada 95 (its actual name) for the later version when it is important to distinguish between the two versions. In discussions of language features common to both versions, we will use the name Ada. The Ada 95 standard language is defined in ARM (1995).

The type derivation mechanism of Ada 83 is extended in Ada 95 to allow adding new components to those inherited from a base class. This provides for inheritance, a key ingredient in object-oriented programming languages. Dynamic binding of subprogram calls to subprogram definitions is accomplished through subprogram dispatching, which is based on the tag value of derived types through classwide types. This feature provides for polymorphism, another principal feature of object-oriented programming.

The rendezvous mechanism of Ada 83 provided only a cumbersome and inefficient means of sharing data among concurrent processes. It was necessary to introduce a new task to control access to the shared data. The protected objects of Ada 95 offer an attractive alternative to this. The shared data is encapsulated in a syntactic structure that controls all access to the data, either by rendezvous or by subprogram call.

It is widely believed that the popularity of Ada 95 suffered because the Department of Defense stopped requiring its use in military software systems. There were, of course, other factors that hindered its growth in popularity. Most important among these was the widespread acceptance of C++ for object-oriented programming, which occurred before Ada 95 was released.

There were several additions to Ada 95 to get Ada 2005. Among these were interfaces, similar to those of Java, more control of scheduling algorithms, and synchronized interfaces.

Ada is widely used in both commercial and defense avionics, air traffic control, and rail transportation, as well as in other areas.

2.15 Object-Oriented Programming: Smalltalk Smalltalk was the first programming language that fully supported object-​- oriented programming. It is therefore an important part of any discussion of the evolution of programming languages.

2.15.1 Design Process The concepts that led to the development of Smalltalk originated in the Ph.D. dissertation work of Alan Kay in the late 1960s at the University of Utah (Kay, 1969). Kay had the remarkable foresight to predict the future availability of powerful desktop computers. Recall that the first microcomputer systems were not marketed until the mid-1970s, and they were only remotely related to the machines envisioned by Kay, which were seen to execute a million or more instructions per second and contain several megabytes of memory. Such machines, in the form of workstations, became widely available only in the early 1980s.

Kay believed that desktop computers would be used by nonprogrammers and thus would need very powerful human-interfacing capabilities. The computers of the late 1960s were largely batch oriented and were used exclusively by professional programmers and scientists. For use by nonprogrammers, Kay determined, a computer would have to be highly interactive and use sophisticated graphics in its user interface. Some of the graphics concepts came from the LOGO experience of Seymour Papert, in which graphics were used to aid children in the use of computers (Papert, 1980).

Kay originally envisioned a system he called Dynabook, which was meant to be a general information processor. It was based in part on the Flex language, which he had helped design. Flex was based primarily on SIMULA 67.

Dynabook used the paradigm of the typical desk, on which there are a number of papers, some partially covered. The top sheet is often the focus of attention, with the others temporarily out of focus. The display of Dynabook would model this scene, using screen windows to represent various sheets of paper on the desktop. The user would interact with such a display both through keystrokes and by touching the screen with his or her fingers. After the preliminary design of Dynabook earned him a Ph.D., Kay’s goal became to see such a machine constructed.

Kay found his way to the Xerox Palo Alto Research Center (Xerox PARC) and presented his ideas on Dynabook. This led to his employment there and the subsequent birth of the Learning Research Group at Xerox. The first charge of the group was to design a language to support Kay’s programming paradigm and implement it on the best personal computer then available. These efforts resulted in an “Interim” Dynabook, consisting of a Xerox Alto workstation and Smalltalk-72 software. Together, they formed a research tool for further development. A number of research projects were conducted with this system, including several experiments to teach programming to children. Along with the experiments came further developments, leading to a sequence of languages that ended with Smalltalk-80. As the language grew, so did the power of the hardware on which it resided. By 1980, both the language and the Xerox hardware nearly matched the early vision of Alan Kay.

2.15.2 Language Overview The Smalltalk world is populated by nothing but objects, from integer constants to large complex software systems. All computing in Smalltalk is done by the same uniform technique: sending a message to an object to invoke one of its methods. A reply to a message is an object, which either returns the requested information or simply notifies the sender that the requested processing has been completed. The fundamental difference between a message and a subprogram call is this: A message is sent to a data object, specifically to one of the methods defined for the object. The called method is then executed, often modifying the data of the object to which the message was sent; a subprogram call is a message to the code of a

subprogram. Usually the data to be processed by the subprogram is sent to it as a parameter.13

13. Of course, a method call can also pass data to be processed by the called method.

In Smalltalk, object abstractions are classes, which are very similar to the classes of SIMULA 67. Instances of the class can be created and are then the objects of the program.

The syntax of Smalltalk is unlike that of most other programming language, in large part because of the use of messages, rather than arithmetic and logic expressions and conventional control statements. One of the Smalltalk control constructs is illustrated in the example in the next subsection.

2.15.3 Evaluation Smalltalk has done a great deal to promote two separate aspects of computing: graphical user interfaces and object-oriented programming. The windowing systems that are now the dominant method of user interfaces to software systems grew out of Smalltalk. Today, the most significant software design methodologies and programming languages are object oriented. Although the origin of some of the ideas of object-oriented languages came from SIMULA 67, they reached maturation in Smalltalk. It is clear that Smalltalk’s impact on the computing world is extensive and will be long- lived.

The following is an example of a Smalltalk class definition:

"Smalltalk Example Program"

"The following is a class definition, instantiations of which can draw equilateral polygons of any number of sides"

class name Polygon

superclass Object

instance variable names ourPen

numSides

sideLength

"Class methods"

"Create an instance"

new

^ super new getPen

"Get a pen for drawing polygons"

getPen

ourPen <- Pen new defaultNib: 2

"Instance methods"

"Draw a polygon"

draw

numSides timesRepeat: [ourPen go: sideLength;

turn: 360 // numSides]

"Set length of sides"

length: len

sideLength <- len

"Set number of sides"

sides: num

numSides <- num

2.16 Combining Imperative and Object-Oriented Features: C++ The origins of C were discussed in Section 2.12; the origins of Simula 67 were discussed in Section 2.10; the origins of Smalltalk were discussed in ​ Section 2.15. C++ builds language facilities, borrowed from Simula 67, on top of C to support much of what Smalltalk pioneered. C++ has evolved from C through a sequence of modifications to improve its imperative features and to add constructs to support object-oriented programming.

2.16.1 Design Process The first step from C toward C++ was made by Bjarne Stroustrup at Bell Laboratories in 1980. The initial modifications to C included the addition of function parameter type checking and conversion and, more significantly, classes, which are related to those of SIMULA 67 and Smalltalk. Also included were derived classes, public/private access control of inherited components, constructor and destructor methods, and friend classes. During 1981, inline functions, default parameters, and overloading of the assignment operator were added. The resulting language was called C with Classes and is described in Stroustrup (1983).

It is useful to consider some goals of C with Classes. The primary goal was to provide a language in which programs could be organized as they could be organized in SIMULA 67—that is, with classes and inheritance. A second important goal was that there should be little or no performance penalty relative to C. For example, array index range checking was not even considered because a significant performance disadvantage, relative to C, would result. A third goal of C with Classes was that it could be used for every application for which C could be used, so virtually none of the features of C would be removed, not even those considered to be unsafe.

By 1984, this language was extended by the inclusion of virtual methods, which provide dynamic binding of method calls to specific method definitions, method name and operator overloading, and reference types. This version of the language was called C++. It is described in Stroustrup (1984).

In 1985, the first available implementation appeared: a system named Cfront, which translated C++ programs into C programs. This version of Cfront and the version of C++ it implemented were named Release 1.0. It is described in Stroustrup (1986).

Between 1985 and 1989, C++ continued to evolve, based largely on user reactions to the first distributed implementation. This next version was named Release 2.0. Its Cfront implementation was released in June 1989. The most important features added to C++ Release 2.0 were support for multiple inheritance (classes with more than one parent class) and abstract classes, along with some other enhancements. Abstract classes are described in Chapter 12.

Release 3.0 of C++ evolved between 1989 and 1990. It added templates, which provide parameterized types, and exception handling. The current version of C++, which was standardized in 1998, is described in ISO (1998).

In 2002, Microsoft released its .NET computing platform, which included a new version of C++, named Managed C++, or MC++. MC++ extends C++ to provide access to the functionality of the .NET Framework. The additions include properties, delegates, interfaces, and a reference type for garbage-​- collected objects. Properties are discussed in Chapter 11. Delegates are briefly discussed in the introduction to C# in Section 2.19. Because .NET does not support multiple inheritance, neither does MC++.

2.16.2 Language Overview Because C++ has both functions and methods, it supports both procedural and object-oriented programming.

Operators in C++ can be overloaded, meaning the user can create operators

for existing operators on user-defined types. C++ methods can also be overloaded, meaning the user can define more than one method with the same name, provided either the numbers or types of their parameters are different.

Dynamic binding in C++ is provided by virtual methods. These methods define type-dependent operations, using overloaded methods, within a collection of classes that are related through inheritance. A pointer to an object of class A can also point to objects of classes that have class A as an ancestor. When this pointer points to an overloaded virtual method, the method of the current type is chosen dynamically.

Both methods and classes can be templated, which means that they can be parameterized. For example, a method can be written as a templated method to allow it to have versions for a variety of parameter types. Classes enjoy the same flexibility.

C++ supports multiple inheritance. The exception-handling constructs of C++ are discussed in Chapter 14.

2.16.3 Evaluation C++ rapidly became and remains a widely used language. One factor in its popularity is the availability of good and inexpensive compilers. Another factor is that it is almost completely backward compatible with C (meaning that C programs, with few changes, can be compiled as C++ programs), and in most implementations it is possible to link C++ code with C code—and thus relatively easy for the many C programmers to learn C++. Finally, at the time C++ first appeared, when object-oriented programming began to receive widespread interest, C++ was the only available language that was suitable for large commercial software projects.

On the negative side, because C++ is a very large and complex language, it clearly suffers drawbacks similar to those of PL/I. It inherited most of the insecurities of C, which make it less safe than languages such as Ada and Java.

2.16.4 A Replacement for Objective-C, Swift Beginning with MAC OS X in 2002, Apple systems software was written in Objective-C. Swift was developed by Apple as an improved replacement for Objective-C. Work on Swift began in 2010 by Chris Lattner. The language was introduced in 2014, with version 2 being introduced in 2015. The first version was proprietary, but the second is open source. It is currently implemented under all of Apple’s operating systems, as well as on Linux.

Among the features of Swift are a tuple data type, an option type ​(variables of this type can have a special no-value value), protocols (similar to Java ​- interfaces), two categories of types, class and struct, which support reference types and value types, as in C#, generic types, no pointers, a safer switch ​- construct, in which the default is to not fall through to the next option, and all statement collections, including a single statement, must be enclosed in braces in all control constructs.

Statements need not be terminated with semicolons, unless there are two or more statements on the same line. Types of data need not be declared, as type inferencing is used. Unlike C and C++, assignment statements do not return a value, so using x = 0 is not legal in a Boolean expression. This eliminates a common error in C and C++ programs, in which x = 0 is typed instead of x == 0. Heap allocated objects are automatically reclaimed, using reference counters.

Swift programs can interact with Objective-C code and Swift uses the same libraries as that language. Swift is already the tenth most popular programming language, according to the TIOBE Community Report.

2.16.5 Another Related Language: Delphi

Delphi (Lischner, 2000) is a hybrid language, similar to C++ and ​Objective- C, in that it was created by adding object-oriented support, among other things, to an existing imperative language, in this case Pascal. Apple designed an object-oriented version of Pascal, named Object Pascal, but subsequently dropped the project. Borland, which had developed Turbo Pascal for ​Windows, also designed an object-oriented version of Pascal, based on Turbo Pascal, also named Object Pascal. For several reasons, Borland renamed Object Pascal as Delphi. The first product with that name, which included an integrated development environment (IDE), was released in 1995. Some consider the IDE to be Delphi and the underlying programming language to be Object Pascal. Other suppliers of similar products continue to refer to the language as Object Pascal.

Many of the differences between C++ and Delphi are a result of the predecessor language and the surrounding programming cultures from which they are derived. Because C is a powerful but potentially unsafe language, C++ also fits that description, at least in the areas of array subscript range checking, pointer arithmetic, and its numerous type coercions. Similarly, because Pascal is more elegant and safer than C, Delphi is more elegant and safer than C++. Delphi is also less complex than C++. For example, Delphi does not include user-defined operator overloading, generic subprograms, and parameterized classes, all of which are part of C++.

Delphi was designed by Anders Hejlsberg, who had previously developed the Turbo Pascal system. Hejlsberg, who moved to Microsoft in 1996, was the lead designer of C#.

2.17 An Imperative-Based Object- Oriented Language: Java Java’s designers started with C++, removed some constructs, changed some, and added a few others. The resulting language provides much of the power and flexibility of C++, but in a smaller, simpler, and safer language. Since that initial design, Java has grown considerably.

2.17.1 Design Process Java, like many programming languages, was designed for an application for which there appeared to be no satisfactory existing language. In 1990, Sun Microsystems determined there was a need for a programming language for embedded consumer electronic devices, such as toasters, microwave ovens, and interactive TV systems. Reliability was one of the primary goals for such a language. It may not seem that reliability would be an important factor in the software for a microwave oven. If an oven had malfunctioning software, it probably would not pose a grave danger to anyone and most likely would not lead to large legal settlements. However, if the software in a particular model was found to be erroneous after a million units had been manufactured and sold, their recall would entail significant cost. Therefore, reliability is an important characteristic of the software in consumer electronic products.

After considering C and C++, it was decided that neither would be satisfactory for developing software for consumer electronic devices. Although C was relatively small, it did not provide support for object- oriented programming, which they deemed a necessity. C++ supported object-oriented programming, but it was judged to be too large and complex, in part because it also supported procedure-oriented programming. It was also believed that neither C nor C++ provided the necessary level of reliability. So, a new language, later named Java, was designed. Its design was guided by the fundamental goal of providing greater simplicity and reliability than

C++ was believed to provide.

Although the initial impetus for Java was consumer electronics, none of the products with which it was used in its early years were ever marketed. Starting in 1993, when the World Wide Web became widely used, and largely because of the new graphical browsers, Java was found to be a useful tool for Web programming. In particular, Java applets, which are relatively small Java programs that are interpreted in Web browsers and whose output can be included in displayed Web documents, quickly became very popular in the middle to late 1990s. In the first few years of Java popularity, the Web was its most common application.

The Java design team was headed by James Gosling, who had previously designed the UNIX emacs editor and the NeWS windowing system.

2.17.2 Language Overview As we stated previously, Java is based on C++ but it was specifically designed to be smaller, simpler, and more reliable. Like C++, Java has both classes and primitive types. Java arrays are instances of a predefined class, whereas in C++ they are not, although many C++ users build wrapper classes for arrays to add features like index range checking, which is implicit in Java.

Java does not have pointers, but its reference types provide some of the capabilities of pointers. These references are used to point to class instances. All objects are allocated on the heap. References are always implicitly dereferenced, when necessary. So they behave more like ordinary scalar variables.

Java has a primitive Boolean type named boolean, used mainly for the control expressions of its control statements (such as if and while). Unlike C and C++, arithmetic expressions cannot be used for control expressions.

One significant difference between Java and many of its predecessors that support object-oriented programming, including C++, is that it is not possible to write stand-alone subprograms in Java. All Java subprograms are methods

and are defined in classes. Furthermore, methods can be called through a class or object only. One consequence of this is that while C++ supports both procedural and object-oriented programming, Java supports object-oriented programming only.

Another important difference between C++ and Java is that C++ supports multiple inheritance directly in its class definitions. Java supports only single inheritance of classes, although some of the benefits of multiple inheritance can be gained by using its interface construct.

Among the C++ constructs that were not copied into Java are structs and unions.

Java includes a relatively simple form of concurrency control through its synchronized modifier, which can appear on methods and blocks. In either case, it causes a lock to be attached. The lock ensures mutually exclusive access or execution. In Java, it is relatively easy to create concurrent processes, which in Java are called threads.

Java uses implicit storage deallocation for its objects, often called garbage collection. This frees the programmer from needing to delete objects explicitly when they are no longer needed. Programs written in languages that do not have garbage collection often suffer from what is sometimes called memory leakage, which means that storage is allocated but never deallocated. This can obviously lead to eventual depletion of all available storage. Object deallocation is discussed in detail in Chapter 6.

Unlike C and C++, Java includes assignment type coercions (implicit type conversions) only if they are widening (from a “smaller” type to a “larger” type). So int to float coercions are done across the assignment operator, but float to int coercions are not.

2.17.3 Evaluation The designers of Java did well at trimming the excess and/or unsafe features of C++. For example, the elimination of half of the assignment coercions that

are done in C++ was clearly a step toward higher reliability. Index range checking of array accesses also makes the language safer. The addition of concurrency enhances the range of applications that can be written in the language, as do the class libraries for graphical user interfaces, database access, and networking.

Java’s portability, at least in intermediate form, has often been attributed to the design of the language, but it is not. Any language can be translated to an intermediate form and “run” on any platform that has a virtual machine for that intermediate form. The price of this kind of portability is the cost of interpretation, which traditionally has been about an order of magnitude more than execution of machine code. The initial version of the Java interpreter, called the Java Virtual Machine (JVM), indeed was at least 10 times slower than equivalent compiled C programs. However, many Java programs are now translated to machine code before being executed, using Just-in-Time (JIT) compilers. This makes the efficiency of Java programs competitive with that of programs in conventionally compiled languages such as C++, at least when array index range checking is not considered.

The use of Java increased faster than that of any other programming language. Initially, this was due to its value in programming dynamic Web documents. Clearly, one of the reasons for Java’s rapid rise to prominence is simply that programmers like its design. Some developers thought C++ was too large and complex to be practical and safe. Java offered them an alternative that has much of the power of C++, but in a simpler, safer language. Another reason is that the compiler/interpreter system for Java is free and easily obtained on the Web. Java is now widely used in a variety of different applications areas.

The most recent version of Java, Java SE8, appeared in 2014. Since the first version, significant features have been added to the language. Among these are an enumeration class, generics, a new iteration construct, lambda expressions, and numerous class libraries.

The following is an example of a Java program:

// Java Example Program

// Input: An integer, listlen, where listlen is less

// than 100, followed by length-integer values

// Output: The number of input data that are greater than

// the average of all input values

import java.io.*;

class IntSort {

public static void main(String args[]) throws IOException {

DataInputStream in = new DataInputStream(System.in);

int listlen,

counter,

sum = 0,

average,

result = 0;

int[] intlist = new int[99];

listlen = Integer.parseInt(in.readLine());

if ((listlen > 0) && (listlen < 100)) {

/* Read input into an array and compute the sum */

for (counter = 0; counter < listlen; counter++) {

intlist[counter] =

Integer.valueOf(in.readLine()).intValue();

sum += intlist[counter];

}

/* Compute the average */

average = sum / listlen;

/* Count the input values that are > average */

for (counter = 0; counter < listlen; counter++)

if (intlist[counter] > average) result++;

/* Print result */

System.out.println(

"\nNumber of values > average is:" + result);

} //** end of then clause of if ((listlen > 0) ...

else System.out.println(

"Error-input list length is not legal\n");

} //** end of method main

} //** end of class IntSort

2.18 Scripting Languages Scripting languages have evolved over the past 35 years. Early scripting ​- languages were used by putting a list of commands, called a script, in a file to be interpreted. The first of these languages, named sh (for shell), began as a small collection of commands that were interpreted as calls to system subprograms that performed utility functions, such as file management and simple file filtering. To this were added variables, control flow statements, functions, and various other capabilities, and the result is a complete programming language. One of the most powerful and widely known of these is ksh (Bolsky and Korn, 1995), which was developed by David Korn at Bell Laboratories.

Another scripting language is awk, developed by Al Aho, Brian Kernighan, and Peter Weinberger at Bell Laboratories (Aho et al., 1988). awk began as a report-generation language but later became a more general-purpose language.

2.18.1 Origins and Characteristics of Perl The Perl language, developed by Larry Wall, was originally a combination of sh and awk. Perl has grown significantly since its beginnings and is now a powerful, although still somewhat primitive, programming language. Although it is still often called a scripting language, it is actually more similar to a typical imperative language, since it is always compiled, at least into an intermediate language, before it is executed. Furthermore, it has all the constructs to make it applicable to a wide variety of areas of computational problems.

Perl has a number of interesting features, only a few of which are mentioned in this chapter and discussed later in the book.

Variables in Perl are statically typed and implicitly declared. There are three distinctive namespaces for variables, denoted by the first character of the variables’ names. All scalar variable names begin with dollar signs ($), all array names begin with at signs (@), and all hash names (hashes are briefly described below) begin with percent signs (%). This convention makes variable names in programs more readable than those of most other programming languages.

Perl includes a large number of implicit variables. Some of them are used to store Perl parameters, such as the particular form of newline character or characters that are used in the implementation. Implicit variables are commonly used as default parameters to built-in functions and default operands for some operators. The implicit variables have distinctive— although cryptic—names, such as $! and @_. The implicit variables’ names, like the user-defined variable names, use the three namespaces, so $! is the name of a scalar variable.

Perl’s arrays have two characteristics that set them apart from the arrays of the common imperative languages. First, they have dynamic length, meaning that they can grow and shrink as needed during execution. Second, arrays can be sparse, meaning that there can be gaps between the elements. These gaps do not take space in memory, and the iteration statement used for arrays, foreach, iterates over the missing elements.

Perl includes associative arrays, which are called hashes. These data structures are indexed by strings and are implicitly controlled hash tables. The Perl system supplies the hash function and increases the size of the structure when necessary.

Perl is a powerful, but somewhat dangerous, language. Its scalar type stores both strings and numbers, which are normally stored in double-​precision ​- floating-point form. Depending on the context, numbers may be coerced to strings and vice versa. If a string is used in numeric context and the string cannot be converted to a number, zero is used and there is no warning or error message provided for the user. This can lead to errors that are not detected by the compiler or run-time system. Array indexing cannot be checked, because there is no set subscript range for any array. References to nonexistent elements return undef, which is interpreted as zero in numeric

context. So, there is also no error detection in array element access.

Perl’s initial use was as a UNIX utility for processing text files. It was and still is widely used as a UNIX system administration tool. When the World Wide Web appeared, Perl achieved widespread use as a common gateway interface language for use with the Web, although it is now rarely used for that purpose. Perl is used as a general-purpose language for a variety of applications, such as computational biology and artificial intelligence.

The following is an example of a Perl program:

# Perl Example Program

# Input: An integer, $listlen, where $listlen is less

# than 100, followed by $listlen-integer values.

# Output: The number of input values that are greater than

# the average of all input values.

($sum, $result) = (0, 0);

$listlen = <STDIN>;

if (($listlen > 0) && ($listlen < 100)) {

# Read input into an array and compute the sum

for ($counter = 0; $counter < $listlen; $counter++) {

$intlist[$counter] = <STDIN>;

} #- end of for (counter ...

# Compute the average

$average = $sum / $listlen;

# Count the input values that are > average

foreach $num (@intlist) {

if ($num > $average) { $result++; }

} #- end of foreach $num ...

# Print result

print "Number of values > average is: $result \n";

} #- end of if (($listlen ...

else {

print "Error--input list length is not legal \n";

}

2.18.2 Origins and Characteristics of JavaScript Use of the Web exploded in the mid-1990s after the first graphical browsers

appeared. The need for computation associated with HTML documents, which by themselves are completely static, quickly became critical. Computation on the server side was made possible with the common gateway interface (CGI), which allowed HTML documents to request the execution of programs on the server, with the results of such computations returned to the browser in the form of HTML documents. Computation on the browser end became available with the advent of Java applets. Both of these approaches have now been replaced for the most part by newer technologies, primarily scripting languages.

JavaScript was originally developed by Brendan Eich at Netscape. Its original name was Mocha. It was later renamed LiveScript. In late 1995, LiveScript became a joint venture of Netscape and Sun Microsystems and its name was changed to JavaScript. JavaScript has gone through extensive evolution, moving from version 1.0 to version 1.5 by adding many new features and capabilities. A language standard for JavaScript was developed in the late 1990s by the European Computer Manufacturers Association (ECMA) as ECMA-262. This standard has also been approved by the International Standards Organization (ISO) as ISO-16262. Microsoft’s version of JavaScript is named JScript .NET.

Although a JavaScript interpreter could be embedded in many different applications, its most common use is embedded in Web browsers. JavaScript code is embedded in HTML documents and interpreted by the browser when the documents are displayed. The primary uses of JavaScript in Web programming are to validate form input data and create dynamic HTML documents.

In spite of its name, JavaScript is related to Java only through the use of similar syntax. Java is strongly typed, but JavaScript is dynamically typed (see Chapter 5). JavaScript’s character strings and its arrays have dynamic length. Because of this, array indices are not checked for validity, although this is required in Java. Java fully supports object-oriented programming, but JavaScript supports neither inheritance nor dynamic binding of method calls to methods.

One of the most important uses of JavaScript is for dynamically creating and modifying HTML documents. JavaScript defines an object hierarchy that

matches a hierarchical model of an HTML document, which is defined by the Document Object Model. Elements of an HTML document are accessed through these objects, providing the basis for dynamic control of the elements of documents.

Following is a JavaScript script for the problem previously solved in several languages in this chapter. Note that it is assumed that this script will be called from an HTML document and interpreted by a Web browser.

// example.js

// Input: An integer, listLen, where listLen is less

// than 100, followed by listLen-numeric values

// Output: The number of input values that are greater

// than the average of all input values

var intList = new Array(99);

var listLen, counter, sum = 0, result = 0;

listLen = prompt (

"Please type the length of the input list", "");

if ((listLen > 0) && (listLen < 100)) {

// Get the input and compute its sum

for (counter = 0; counter < listLen; counter++) {

intList[counter] = prompt (

"Please type the next number", "");

sum += parseInt(intList[counter]);

}

// Compute the average

average = sum / listLen;

// Count the input values that are > average

for (counter = 0; counter < listLen; counter++)

if (intList[counter] > average) result++;

// Display the results

document.write("Number of values > average is: ",

result, "<br />");

} else

document.write(

"Error - input list length is not legal <br />");

2.18.3 Origins and Characteristics of PHP PHP (Tatroe et al., 2013) was developed by Rasmus Lerdorf, an employee of

the Apache Group, in 1994. His initial motivation was to provide a tool to help track visitors to his personal Web site. In 1995, he developed a package called Personal Home Page Tools, which became the first publicly distributed version of PHP. Originally, PHP was an abbreviation for Personal Home Page. Later, its user community began using the recursive name PHP: Hypertext ​Preprocessor, which subsequently forced the original name into obscurity. PHP is now developed, distributed, and supported as an open- source product. PHP processors are resident on most Web servers.

PHP is an HTML-embedded server-side scripting language specifically designed for Web applications. PHP code is interpreted on the Web server when an HTML document in which it is embedded has been requested by a browser. PHP code usually produces HTML code as output, which replaces the PHP code in the HTML document. Therefore, a Web browser never sees PHP code.

PHP is similar to JavaScript in its syntactic appearance, the dynamic nature of its strings and arrays, and its use of dynamic typing. PHP’s arrays are a combination of JavaScript’s arrays and Perl’s hashes.

The original version of PHP did not support object-oriented programming. Abstract classes, interfaces, destructors, and access controls for class members have since been added to the language.

PHP allows simple access to HTML form data, so form processing is easy with PHP. PHP provides support for many different database management systems. This makes it a useful language for building programs that need Web access to databases.

The current version of PHP is 7, released in 2015.

2.18.4 Origins and Characteristics of Python Python (Lutz, 2013) is an object-oriented interpreted scripting language. Its

initial design was by Guido van Rossum at Stichting Mathematisch Centrum in the Netherlands in the early 1990s. Its development is being continued by the Python Software Foundation. Python is being used for the same kinds of applications as Perl: system administration and other relatively small computing tasks. Python is an open-source system that is available for most common computing platforms. The Python implementation is available at www.python.org, which also has extensive information regarding Python.

Python’s syntax is not based directly on any commonly used language. It is type checked, but dynamically typed. Instead of arrays, Python includes three kinds of data structures: lists; immutable lists, which are called tuples; and hashes, which are called dictionaries. There is a collection of list methods, such as append, insert, remove, and sort, as well as a collection of methods for dictionaries, such as keys, values, copy, and has_key. Python also supports list comprehensions, which originated with the Haskell language. List comprehensions are discussed in Section 15.8.

Python is object oriented, includes the pattern-matching capabilities of Perl, and has exception handling. Garbage collection is used to reclaim objects when they are no longer needed.

Support for form processing is provided by the cgi module. Modules that support cookies, networking, and database access are also available.

Python includes support for concurrency with its threads, as well as support for network programming with its sockets. It also has more support for functional programming than other nonfunctional programming languages.

One of the more interesting features of Python is that it can be easily extended by any user. The modules that support the extensions can be written in any compiled language. Extensions can add functions, variables, and object types. These extensions are implemented as additions to the Python interpreter.

2.18.5 Origins and Characteristics

of Ruby Ruby (Thomas et al., 2005) was designed by Yukihiro Matsumoto (aka Matz) in the early 1990s and released in 1996. Since then it has continually evolved. The motivation for Ruby was dissatisfaction of its designer with Perl and Python. Although both Perl and Python support object-oriented programming,14 neither is a pure object-oriented language, at least in the sense that each has primitive (nonobject) types and each supports functions.

14. Actually, Python’s support for object-oriented programming is partial.

The primary characterizing feature of Ruby is that it is a pure object-​oriented language, just as is Smalltalk. Every data value is an object and all operations are via method calls. The operators in Ruby are only syntactic mechanisms to specify method calls for the corresponding operations. Because they are methods, they can be redefined. All classes, predefined or user defined, can be subclassed.

Both classes and objects in Ruby are dynamic in the sense that methods can be dynamically added to either. This means that both classes and objects can have different sets of methods at different times during execution. So, different instantiations of the same class can behave differently. Collections of methods, data, and constants can be included in the definition of a class.

The syntax of Ruby is related to that of Eiffel and Ada. There is no need to declare variables, because dynamic typing is used. The scope of a variable is specified in its name: A variable whose name begins with a letter has local scope; one that begins with @ is an instance variable; one that begins with $ has global scope. A number of features of Perl are present in Ruby, including implicit variables with silly names, such as $_.

As is the case with Python, any user can extend and/or modify Ruby. Ruby is culturally interesting because it is the first programming language designed in Japan that has achieved relatively widespread use in the United States.

2.19 The Flagship .NET Language: C# C#, along with the development platform .NET,15 was announced by Microsoft in 2000. In January 2002, production versions of both were released.

15. The .NET development system is briefly discussed in Chapter 1.

2.19.1 Design Process C# is based on C++ and Java but includes some ideas from Delphi and Visual Basic. Its lead designer, Anders Hejlsberg, also designed Turbo Pascal and ​- Delphi, which explains the Delphi parts of the heritage of C#.

The purpose of C# is to provide a language for component-based software development, specifically for such development in the .NET Framework. In this environment, components from a variety of languages easily can be combined to form systems. All of the .NET languages, which include C#, VB.NET, Managed C++, F#, and JScript .NET,16 use the common type system (CTS). The CTS provides a common class library. All types in all five .NET languages inherit from a single class root, System.Object. Compilers that conform to the CTS specification create objects that can be combined into software systems. All .NET languages are compiled into the same intermediate form, Intermediate Language (IL).17 Unlike Java, however, the IL is never interpreted. A Just-in-Time compiler is used to translate IL into machine code before it is executed.

16. Many other languages have been modified to be .NET languages.

17. Initially, IL was called MSIL (Microsoft Intermediate Language), but apparently many ​people thought that name was too long.

2.19.2 Language Overview Many believe that one of Java’s most important advances over C++ lies in the fact that it excludes some of C++’s features. For example, C++ supports multiple inheritance, pointers, structs, enum types, operator overloading, and a goto statement, but Java includes none of these. The designers of C# obviously disagreed with this wholesale removal of features, because all of these except multiple inheritance have been included in C#.

To the credit of C#’s designers, however, in several cases, the C# version of a C++ feature has been improved. For example, the enum types of C# are safer than those of C++, because they are never implicitly converted to integers. This allows them to be more type safe. The struct type was changed significantly, resulting in a truly useful construct, whereas in C++ it serves virtually no purpose. C#’s structs are discussed in Chapter 12. C# takes a stab at improving the switch statement that is used in C, C++, and Java. C#’s switch is discussed in Chapter 8.

Although C++ includes function pointers, they share the lack of safety that is inherent in C++’s pointers to variables. C# includes a new type, delegates, which are both object-oriented and type-safe references to subprograms. Delegates are used for implementing event handlers, controlling the execution of threads, and callbacks.18 Callbacks are implemented in Java with interfaces; in C++, method pointers are used.

18. When an object calls a method of another object and needs to be notified when that method has completed its task, the called method calls its caller back. This is known as a callback.

In C#, methods can take a variable number of parameters, as long as they are all the same type. This is specified by the use of a formal parameter of array type, preceded by the params reserved word.

Both C++ and Java use two distinct typing systems: one for primitives and one for objects. In addition to being confusing, this leads to a frequent need to convert values between the two systems—for example, to put a primitive

value into a collection that stores objects. C# makes the conversion between values of the two typing systems partially implicit through the implicit boxing and unboxing operations, which are discussed in detail in Chapter 12.19

19. This feature was added to Java in Java 5.0.

Among the other features of C# are rectangular arrays, which are not supported in most programming languages, and a foreach statement, which is an iterator for arrays and collection objects. A similar foreach statement is found in Perl, PHP, and Java 5.0. Also, C# includes properties, which are an alternative to public data members. Properties are specified as data members with get and set methods, which are implicitly called when references and assignments are made to the associated data members.

C# has evolved continuously and quickly from its initial release in 2002. The most recent version is C# 7.0. New in C# 7.0 are tuples and a form of pattern matching.

2.19.3 Evaluation C# was meant to be an improvement over both C++ and Java as a general-​- purpose programming language. Although it can be argued that some of its features are a step backward, C# includes some constructs that move it beyond its predecessors. Some of its features will surely be adopted by other programming languages of the future.

The following is an example of a C# program:

// C# Example Program

// Input: An integer, listlen, where listlen is less than

// 100, followed by listlen-integer values.

// Output: The number of input values that are greater

// than the average of all input values.

using System;

public class Ch2example {

static void Main() {

int[] intlist;

int listlen,

counter,

sum = 0,

average,

result = 0;

intList = new int[99];

listlen = Int32.Parse(Console.readLine());

if ((listlen > 0) && (listlen < 100)) {

// Read input into an array and compute the sum

for (counter = 0; counter < listlen; counter++) {

intList[counter] =

Int32.Parse(Console.readLine());

sum += intList[counter];

} //- end of for (counter ...

// Compute the average

average = sum / listlen;

// Count the input values that are > average

foreach (int num in intList)

if (num > average) result++;

// Print result

Console.WriteLine(

"Number of values > average is:" + result);

} //- end of if ((listlen ...

else

Console.WriteLine(

"Error--input list length is not legal");

} //- end of method Main

} //- end of class Ch2example

2.20 Markup-Programming Hybrid Languages A markup-programming hybrid language is a markup language in which some of the elements can specify programming actions, such as control flow and computation. The following subsections introduce two such hybrid languages, XSLT and JSP.

2.20.1 XSLT eXtensible markup language (XML) is a metamarkup language. Such a language is used to define markup languages. XML-derived markup languages are used to define XML data documents. Although XML documents are human readable, they are processed by computers. This processing sometimes consists only of transformations to alternative forms that can be effectively displayed or printed. In many cases, such transformations are to HTML, which can be displayed by a Web browser. In other cases, the data in the document is processed, just as with other forms of data files.

The transformation of XML documents to HTML documents is specified in another markup language, eXtensible stylesheet language transformations (XSLT) (www.w3.org/TR/XSLT). XSLT can specify programming-like operations. Therefore, XSLT is a markup-programming hybrid language. XSLT was defined by the World Wide Web Consortium (W3C) in the late 1990s.

An XSLT processor is a program that takes as input an XML data document and an XSLT document (which is also in the form of an XML document). In this processing, the XML data document is transformed to another XML document,20 using the transformations described in the XSLT document. The XSLT document specifies transformations by defining templates, which are

data patterns that could be found by the XSLT processor in the XML input file. Associated with each template in the XSLT document are its transformation instructions, which specify how the matching data is to be transformed before being put in the output document. So, the templates (and their associated processing) act as subprograms, which are “executed” when the XSLT processor finds a pattern match in the data of the XML document.

20. The output document of the XSLT processor could also be in HTML or plain text.

XSLT also has programming constructs at a lower level. For example, a looping construct is included, which allows repeated parts of the XML document to be selected. There is also a sort process. These lower-level constructs are specified with XSLT tags, such as <for-each>.

2.20.2 JSP The “core” part of the Java Server Pages Standard Tag Library (JSTL) is another markup-programming hybrid language, although its form and purpose are different from those of XSLT. Before discussing JSTL, it is necessary to introduce the ideas of servlets and Java Server Pages (JSP). A servlet is an instance of a Java class that resides on and is executed on a Web server system. The execution of a servlet is requested by a markup document being displayed by a Web browser. The servlet’s output, which is in the form of an HTML document, is returned to the requesting browser. A program that runs in the Web server process, called a servlet container, controls the execution of servlets. Servlets are commonly used for form processing and for database access.

JSP is a collection of technologies designed to support dynamic Web documents and provide other processing needs of Web documents. When a JSP document, which is often a mixture of HTML and Java, is requested by a browser, the JSP processor program, which resides on a Web server system, converts the document to a servlet. The document’s embedded Java code is copied to the servlet. The plain HTML is copied into Java print statements that output it as is. The JSTL markup in the JSP document is processed, as

discussed in the following paragraph. The servlet produced by the JSP processor is run by the servlet container.

The JSTL defines a collection of XML action elements that control the processing of the JSP document on the Web server. These elements have the same form as other elements of HTML and XML. One of the most commonly used JSTL control action elements is if, which specifies a Boolean expression as an attribute.21 The content of the if element (the text between the opening tag (<if>) and its closing tag (</if>)) is HTML code that will be included in the output document only if the Boolean expression evaluates to true. The if element is related to the C/C++ #if preprocessor command. The JSP container processes the JSTL parts of JSP documents in a way that is similar to how the C/C++ preprocessor processes C and C++ programs. The preprocessor commands are instructions for the preprocessor to specify how the output file is to be constructed from the input file. Similarly, JSTL control action elements are instructions for the JSP processor to specify how to build the XML output file from the XML input file.

21. An attribute in HTML, which is embedded in the opening tag of an element, provides ​further information about that element.

One common use of the if element is for the validation of form data submitted by a browser user. Form data is accessible by the JSP processor and can be tested with the if element to ensure that it is sensible data. If not, the if element can insert an error message for the user in the output document.

For multiple selection control, JSTL has choose, when, and otherwise elements. JSTL also includes a forEach element, which iterates over collections, which typically are form values from a client. The forEach element can include begin, end, and step attributes to control its iterations.

SUMMARY We have investigated the development of a number of programming languages. This chapter gives the reader a good perspective on current issues in language design. We have set the stage for an in-depth discussion of the important ​features of contemporary languages.

BIBLIOGRAPHIC NOTES Perhaps the most important source of historical information about the development of early programming languages is History of Programming Languages, edited by Richard Wexelblat (1981). It contains the developmental background and environment of 13 important programming languages, as told by the designers themselves. A similar work resulted from a second “history” conference, published as a special issue of ACM SIGPLAN Notices (ACM, 1993a). In this work, the history and evolution of 13 more programming languages are discussed.

The paper “Early Development of Programming Languages” (Knuth and Pardo, 1977), which is part of the Encyclopedia of Computer Science and Technology, is an excellent 85-page work that details the development of languages up to and including Fortran. The paper includes example programs to demonstrate the features of many of those languages.

Another book of great interest is Programming Languages: History and ​- Fundamentals, by Jean Sammet (1969). It is a 785-page work filled with details of 80 programming languages of the 1950s and 1960s. Sammet has also published several updates to her book, such as Roster of Programming Languages for 1974–75 (1976).

REVIEW QUESTIONS 1. In what year was Plankalkül designed? In what year was that design

published?

2. What two common data structures were included in Plankalkül?

3. How were the pseudocodes of the early 1950s implemented?

4. Speedcoding was invented to overcome two significant shortcomings of the computer hardware of the early 1950s. What were they?

5. Why was the slowness of interpretation of programs acceptable in the early 1950s?

6. What hardware capability that first appeared in the IBM 704 computer strongly affected the evolution of programming languages? Explain why.

7. In what year was the Fortran design project begun?

8. What was the primary application area of computers at the time Fortran was designed?

9. What was the source of all of the control flow statements of Fortran I?

10. What was the most significant feature added to Fortran I to get Fortran II?

11. What control flow statements were added to Fortran IV to get Fortran 77?

12. Which version of Fortran was the first to have any sort of dynamic variables?

13. Which version of Fortran was the first to have character string handling?

14. Why were linguists interested in artificial intelligence in the late 1950s?

15. Where was Lisp developed? By whom?

16. In what way are Scheme and Common Lisp opposites of each other?

17. What dialect of Lisp is used for introductory programming courses at some universities?

18. What two professional organizations together designed ALGOL 60?

19. In what version of ALGOL did block structure appear?

20. What missing language element of ALGOL 60 damaged its chances for widespread use?

21. What language was designed to describe the syntax of ALGOL 60?

22. On what programming language was COBOL based?

23. In what year did the COBOL design process begin?

24. What data structure that appeared in COBOL originated with Plankalkül?

25. What organization was most responsible for the early success of COBOL (in terms of extent of use)?

26. What user group was the target of the first version of Basic?

27. Why was Basic an important language in the early 1980s?

28. PL/I was designed to replace what two languages?

29. For what new line of computers was PL/I designed?

30. What features of SIMULA 67 are now important parts of some object- oriented languages?

31. What innovation of data structuring was introduced in ALGOL 68 but is often credited to Pascal?

32. What design criterion was used extensively in ALGOL 68?

33. What language introduced the case statement?

34. What operators in C were modeled on similar operators in ALGOL 68?

35. What are two characteristics of C that make it less safe than Pascal?

36. What is a nonprocedural language?

37. What are the two kinds of statements that populate a Prolog database?

38. What is the primary application area for which Ada was designed?

39. What are the concurrent program units of Ada called?

40. What Ada construct provides support for abstract data types?

41. What populates the Smalltalk world?

42. What three concepts are the basis for object-oriented programming?

43. Why does C++ include the features of C that are known to be unsafe?

44. What language was Swift designed to replace?

45. What do the Ada and COBOL languages have in common?

46. What was the first application for Java?

47. What characteristic of Java is most evident in JavaScript?

48. How does the typing system of PHP and JavaScript differ from that of Java?

49. What array structure is included in C# but not in C, C++, or Java?

50. What two languages was the original version of Perl meant to replace?

51. For what application area is JavaScript most widely used?

52. What is the relationship between JavaScript and PHP, in terms of their use?

53. PHP’s primary data structure is a combination of what two data structures from other languages?

54. What data structure does Python use in place of arrays?

55. What characteristic does Ruby share with Smalltalk?

56. What characteristic of Ruby’s arithmetic operators makes them unique among those of other languages?

57. What deficiency of the switch statement of C is addressed with the changes made by C# to that statement?

58. What is the primary platform on which C# is used?

59. What are the inputs to an XSLT processor?

60. What is the output of an XSLT processor?

61. What element of the JSTL is related to a subprogram?

62. To what is a JSP document converted by a JSP processor?

63. Where are servlets executed?

PROBLEM SET 1. What features of Plankalkül do you think would have had the greatest

influence on Fortran 0 if the Fortran designers had been familiar with Plankalkül?

2. Determine the capabilities of Backus’s 701 Speedcoding system, and compare them with those of a contemporary programmable hand calculator.

3. Write a short history of the A-0, A-1, and A-2 systems designed by Grace Hopper and her associates.

4. Compare the facilities of Fortran 0 with those of the Laning and Zierler system.

5. Which of the three original goals of the ALGOL design committee, in your opinion, was most difficult to achieve at that time?

6. Make an educated guess as to the most common syntax error in Lisp programs.

7. Lisp began as a pure functional language but gradually acquired more and more imperative features. Why?

8. Describe in detail the three most important reasons, in your opinion, why ALGOL 60 did not become a very widely used language.

9. Why, in your opinion, did COBOL allow long identifiers when Fortran and ALGOL did not?

10. Outline the major motivation of IBM in developing PL/I.

11. Was IBM’s assumption, on which it based its decision to develop PL/I, correct, given the history of computers and language developments since 1964?

12. Describe, in your own words, the concept of orthogonality in programming language design.

13. What is the primary reason why PL/I became more widely used than ALGOL 68?

14. What are the arguments both for and against the idea of a typeless language?

15. Are there any logic programming languages other than Prolog?

16. What is your opinion of the argument that languages that are too complex are too dangerous to use, and we should therefore keep all languages small and simple?

17. Do you think language design by committee is a good idea? Support your opinion.

18. Languages continually evolve. What sort of restrictions do you think are appropriate for changes in programming languages? Compare your answers with the evolution of Fortran.

19. Build a table identifying all of the major language developments, together with when they occurred, in what language they first appeared, and the identities of the developers.

20. There have been some public interchanges between Microsoft and Sun concerning the design of Microsoft’s J++ and C# and Sun’s Java. Read some of these documents, which are available on their respective Web sites, and write an analysis of the disagreements concerning the delegates.

21. In recent years data structures have evolved within scripting languages to replace traditional arrays. Explain the chronological sequence of these developments.

22. Explain two reasons why pure interpretation is an acceptable implementation method for several recent scripting languages.

23. Why, in your opinion, do new scripting languages appear more frequently than new compiled languages?

24. Give a brief general description of a markup-programming hybrid language.

PROGRAMMING EXERCISES 1. To understand the value of records in a programming language, write a

small program in a C-based language that uses an array of structs that store student information, including name, age, GPA as a float, and grade level as a string (e.g., “freshmen,” etc.). Also, write the same program in the same language without using structs.

2. To understand the value of recursion in a programming language, write a program that implements quicksort, first using recursion and then without recursion.

3. To understand the value of counting loops, write a program that implements matrix multiplication using counting loop constructs. Then write the same program using only logical loops—for example, while loops.

3 Describing Syntax and Semantics 1. 3.1 Introduction

2. 3.2 The General Problem of Describing Syntax

3. 3.3 Formal Methods of Describing Syntax

4. 3.4 Attribute Grammars

5. 3.5 Describing the Meanings of Programs: Dynamic Semantics

This chapter begins by defining the terms syntax and semantics. Then, a detailed discussion of the most common method of describing syntax, context-free grammars (also known as Backus-Naur Form), is presented. Included in this ​discussion are derivations, parse trees, ambiguity, descriptions of operator precedence and associativity, and extended Backus- Naur Form. Attribute grammars, which can be used to describe both the syntax and static semantics of programming languages, are discussed next. In the last section, three formal methods of describing semantics—operational, axiomatic, and denotational semantics—are introduced. Because of the inherent complexity of the semantics description methods, our discussion of them is brief. One could easily write an entire book on just one of the three (as several authors have).

3.1 Introduction The task of providing a concise yet understandable description of a programming language is difficult but essential to the language’s success. ALGOL 60 and ALGOL 68 were first presented using concise formal descriptions; in both cases, however, the descriptions were not easily understandable, partly because each used a new notation. The levels of acceptance of both languages suffered as a result. On the other hand, some languages have suffered the problem of having many slightly different dialects, a result of a simple but informal and imprecise definition.

One of the problems in describing a language is the diversity of the ​people who must understand the description. Among these are initial evaluators, implementors, and users. Most new programming languages are subjected to a period of scrutiny by potential users, often people within the organization that employs the language’s designer, before their designs are completed. These are the initial evaluators. The success of this feedback cycle depends heavily on the clarity of the description.

Programming language implementors obviously must be able to determine how the expressions, statements, and program units of a language are formed, and also their intended effect when executed. The difficulty of the implementors’ job is, in part, determined by the completeness and precision of the language description.

Finally, language users must be able to determine how to encode software solutions by referring to a language reference manual. Textbooks and courses enter into this process, but language manuals are usually the only authoritative printed information source about a language.

The study of programming languages, like the study of natural languages, can be divided into examinations of syntax and semantics. The syntax of a programming language is the form of its expressions, statements, and ​- program units. Its semantics is the meaning of those expressions, statements, and ​program units. For example, the syntax of a Java while statement is

while (boolean_expr) statement

The semantics of this statement form is that when the current value of the Boolean expression is true, the embedded statement is executed. Then control implicitly returns to the Boolean expression to repeat the process. If the Boolean expression is false, control transfers to the statement following the while construct.

Although they are often separated for discussion purposes, syntax and semantics are closely related. In a well-designed programming language, semantics should follow directly from syntax; that is, the appearance of a statement should strongly suggest what the statement is meant to accomplish.

Describing syntax is easier than describing semantics, partly because a concise and universally accepted notation is available for syntax description, but none has yet been developed for semantics.

3.2 The General Problem of Describing Syntax A language, whether natural (such as English) or artificial (such as Java), is a set of strings of characters from some alphabet. The strings of a language are called sentences or statements. The syntax rules of a language specify which strings of characters from the language’s alphabet are in the language. English, for example, has a large and complex collection of rules for specifying the syntax of its sentences. By comparison, even the largest and most complex programming languages are syntactically very simple.

Formal descriptions of the syntax of programming languages, for simplicity’s sake, often do not include descriptions of the lowest-level syntactic units. These small units are called lexemes. The description of lexemes can be given by a lexical specification, which is usually separate from the syntactic description of the language. The lexemes of a programming language include its numeric literals, operators, and special words, among others. One can think of programs as strings of lexemes rather than of characters.

Lexemes are partitioned into groups—for example, the names of variables, methods, classes, and so forth in a programming language form a group called identifiers. Each lexeme group is represented by a name, or token. So, a token of a language is a category of its lexemes. For example, an identifier is a token that can have lexemes, or instances, such as sum and total. In some cases, a token has only a single possible lexeme. For example, the token for the arithmetic operator symbol + has just one possible lexeme. Consider the following Java statement:

index = 2 * count + 17;

The lexemes and tokens of this statement are

The example language descriptions in this chapter are very simple, and most include lexeme descriptions.

3.2.1 Language Recognizers In general, languages can be formally defined in two distinct ways: by recognition and by generation (although neither provides a definition that is practical by itself for people trying to learn or use a programming language). Suppose we have a language L that uses an alphabet ∑ of characters. To define L formally using the recognition method, we would need to construct a mechanism R, called a recognition device, capable of reading strings of characters from the alphabet ∑. R would indicate whether a given input string was or was not in L. In effect, R would either accept or reject the given string. Such devices are like filters, separating legal sentences from those that are incorrectly formed. If R, when fed any string of characters over ∑, accepts it only if it is in L, then R is a description of L. Because most useful languages are, for all practical purposes, infinite, this might seem like a

lengthy and ineffective process. Recognition devices, however, are not used to enumerate all of the sentences of a language—they have a different purpose.

The syntax analysis part of a compiler is a recognizer for the language the compiler translates. In this role, the recognizer need not test all possible strings of characters from some set to determine whether each is in the language. Rather, it need only determine whether given programs are in the language. In effect then, the syntax analyzer determines whether the given programs are syntactically correct. The structure of syntax analyzers, also known as parsers, is discussed in Chapter 4.

3.2.2 Language Generators A language generator is a device that can be used to generate the sentences of a language. We can think of the generator as having a button that produces a sentence of the language every time it is pushed. Because the particular sentence that is produced by a generator when its button is pushed is unpredictable, a generator seems to be a device of limited usefulness as a language descriptor. However, people prefer certain forms of generators over recognizers because they can more easily read and understand them. By contrast, the syntax-​checking portion of a compiler (a language recognizer) is not as useful a language description for a programmer because it can be used only in trial-and-error mode. For example, to determine the correct syntax of a particular statement using a compiler, the programmer can only submit a speculated version and note whether the compiler accepts it. On the other hand, it is often possible to determine whether the syntax of a particular statement is correct by comparing it with the structure of the generator.

There is a close connection between formal generation and recognition devices for the same language. This was one of the seminal discoveries in computer science, and it led to much of what is now known about formal languages and compiler design theory. We return to the relationship of generators and recognizers in the next section.

3.3 Formal Methods of Describing Syntax This section discusses the formal language-generation mechanisms, usually called grammars, that are commonly used to describe the syntax of programming languages.

3.3.1 Backus-Naur Form and Context-Free Grammars In the middle to late 1950s, two men, Noam Chomsky and John Backus, in unrelated research efforts, developed the same syntax description formalism, which subsequently became the most widely used method for programming language syntax.

3.3.1.1 Context-Free Grammars In the mid-1950s, Noam Chomsky, a noted linguist (among other things), described four classes of generative devices or grammars that define four classes of languages (Chomsky, 1956, 1959). Two of these grammar classes, named context-free and regular, turned out to be useful for describing the syntax of programming languages. The forms of the tokens of programming languages can be described by regular grammars. The syntax of whole programming languages, with minor exceptions, can be described by context- free grammars. Because Chomsky was a linguist, his primary interest was the theoretical nature of natural languages. He had no interest at the time in the artificial languages used to communicate with computers. So it was not until later that his work was applied to programming languages.

3.3.1.2 Origins of Backus-Naur Form Shortly after Chomsky’s work on language classes, the ACM-GAMM group began designing ALGOL 58. A landmark paper describing ALGOL 58 was presented by John Backus, a prominent member of the ACM-GAMM group, at an international conference in 1959 (Backus, 1959). This paper introduced a new formal notation for specifying programming language syntax. The new notation was later modified slightly by Peter Naur for the description of ALGOL 60 (Naur, 1960). This revised method of syntax description became known as Backus-Naur Form, or simply BNF.

BNF is a natural notation for describing syntax. In fact, something similar to BNF was used by Panini to describe the syntax of Sanskrit several hundred years before Christ (Ingerman, 1967).

Although the use of BNF in the ALGOL 60 report was not immediately accepted by computer users, it soon became and is still the most popular method of concisely describing programming language syntax.

It is remarkable that BNF is nearly identical to Chomsky’s generative devices for context-free languages, called context-free grammars. In the remainder of the chapter, we refer to context-free grammars simply as ​grammars. ​- Furthermore, the terms BNF and grammar are used interchangeably.

3.3.1.3 Fundamentals A metalanguage is a language that is used to describe another language. BNF is a metalanguage for programming languages.

BNF uses abstractions for syntactic structures. A simple Java assignment statement, for example, might be represented by the abstraction <assign> (pointed brackets are often used to delimit names of abstractions). The actual definition of <assign> can be given by

<assign> → <var> = <expression>

The text on the left side of the arrow, which is aptly called the left-hand side (LHS), is the abstraction being defined. The text to the right of the arrow is the definition of the LHS. It is called the right-hand side (RHS) and consists of some mixture of tokens, lexemes, and references to other abstractions. (Actually, tokens are also abstractions.) Altogether, the definition is called a rule, or production. In the example rule just given, the abstractions <var> and <expression> obviously must be defined for the <assign> definition to be useful.

This particular rule specifies that the abstraction <assign> is defined as an instance of the abstraction <var>, followed by the lexeme =, followed by an instance of the abstraction <expression>. One example sentence whose syntactic structure is described by the rule is

total = subtotal1 + subtotal2

The abstractions in a BNF description, or grammar, are often called nonterminal symbols, or simply nonterminals, and the lexemes and tokens of the rules are called terminal symbols, or simply terminals. A BNF description, or grammar, is a collection of rules.

Nonterminal symbols can have two or more distinct definitions, representing two or more possible syntactic forms in the language. Multiple definitions can be written as a single rule, with the different definitions separated by the symbol |, meaning logical OR. For example, a Java if statement can be described with the rules

<if_stmt> → if ( <logic_expr> ) <stmt>

<if_stmt> → if ( <logic_expr> ) <stmt> else <stmt>

or with the rule

<if_stmt> → if ( <logic_expr> ) <stmt>

| if ( <logic_expr> ) <stmt> else <stmt>

In these rules, <stmt> represents either a single statement or a compound statement.

Although BNF is simple, it is sufficiently powerful to describe nearly all of the syntax of programming languages. In particular, it can describe lists of similar constructs, the order in which different constructs must appear, and nested structures to any depth, and even imply operator precedence and operator associativity.

3.3.1.4 Describing Lists Variable-length lists in mathematics are often written using an ellipsis ( . . . ); 1, 2, . . . is an example. BNF does not include the ellipsis, so an alternative method is required for describing lists of syntactic elements in programming languages (for example, a list of identifiers appearing on a data declaration statement). For BNF, the alternative is recursion. A rule is recursive if its LHS appears in its RHS. The following rules illustrate how recursion is used to describe lists:

<ident_list> → identifier

| identifier, <ident_list>

This defines <ident_list> as either a single token (identifier) or an identifier followed by a comma and another instance of <ident_list>. Recursion is used to describe lists in many of the example grammars in the remainder of this chapter.

3.3.1.5 Grammars and Derivations A grammar is a generative device for defining languages. The sentences of the language are generated through a sequence of applications of the rules, beginning with a special nonterminal of the grammar called the start symbol. This sequence of rule applications is called a derivation. In a grammar for a complete programming language, the start symbol represents a complete

program and is often named <program>. The simple grammar shown in Example 3.1 is used to illustrate derivations.

EXAMPLE 3.1 A Grammar for a Small Language

<program> → begin <stmt_list> end

<stmt_list> → <stmt>

| <stmt> ; <stmt_list>

<stmt> → <var> = <expression>

<var> → A | B | C

<expression> → <var> + <var>

| <var> - <var>

| <var>

The language described by the grammar of Example 3.1 has only one statement form: assignment. A program consists of the special word begin, followed by a list of statements separated by semicolons, followed by the special word end. An expression is either a single variable or two variables separated by either a + or - operator. The only variable names in this language are A, B, and C.

A derivation of a program in this language follows:

<program> => begin <stmt_list> end

     => begin <stmt> ; <stmt_list> end

     => begin <var> = <expression> ; <stmt_list> end

     => begin A = <expression> ; <stmt_list> end

     => begin A = <var> + <var> ; <stmt_list> end

     => begin A = B + <var> ; <stmt_list> end

     => begin A = B + C ; <stmt_list> end

     => begin A = B + ; <stmt> end

     => begin A = B + C ; <var> = <expression> end

     => begin A = B + C ; B = <expression> end

     => begin A = B + C ; B = <var> end

     => begin A = B + C ; B = C end

This derivation, like all derivations, begins with the start symbol, in this case <program>. The symbol => is read “derives.” Each successive string in the sequence is derived from the previous string by replacing one of the nonterminals with one of that nonterminal’s definitions. Each of the strings in the derivation, including <program>, is called a sentential form.

In this derivation, the replaced nonterminal is always the leftmost nonterminal in the previous sentential form. Derivations that use this order of replacement are called leftmost derivations. The derivation continues until the sentential form contains no nonterminals. That sentential form, consisting of only terminals, or lexemes, is the generated sentence.

In addition to leftmost, a derivation may be rightmost or in an order that is neither leftmost nor rightmost. Derivation order has no effect on the language generated by a grammar.

By choosing alternative RHSs of rules with which to replace nonterminals in the derivation, different sentences in the language can be generated. By exhaustively choosing all combinations of choices, the entire language can be generated. This language, like most others, is infinite, so one cannot generate all the sentences in the language in finite time.

Example 3.2 is another example of a grammar for part of a typical ​- programming language.

EXAMPLE 3.2 A Grammar for Simple Assignment Statements

<assign> → <id> = <expr>

<id> → A| B | C

<expr> → <id> + <expr>

| <id> * <expr>

| ( <expr>)

| <id>

The grammar of Example 3.2 describes assignment statements whose right sides are arithmetic expressions with multiplication and addition operators and parentheses. For example, the statement

A = B * ( A + C )

is generated by the leftmost derivation:

<assign> => <id> = <expr>

    => A = <expr>

    => A = <id> * <expr>

    => A = B * <expr>

    => A = B * ( <expr>)

    => A = B * ( <id> + <expr>)

    => A = B * ( A + <expr>)

    => A = B * ( A + <id>)

    => A = B * ( A + C )

3.3.1.6 Parse Trees One of the most attractive features of grammars is that they naturally describe the hierarchical syntactic structure of the sentences of the languages they define. These hierarchical structures are called parse trees. For example, the parse tree in Figure 3.1 shows the structure of the assignment statement derived previously.

Figure 3.1 A parse tree for the simple statement A = B * (A + C)

Figure 3.1 Full Alternative Text

Every internal node of a parse tree is labeled with a nonterminal ​symbol; every leaf is labeled with a terminal symbol. Every subtree of a parse tree describes one instance of an abstraction in the sentence.

3.3.1.7 Ambiguity A grammar that generates a sentential form for which there are two or more distinct parse trees is said to be ambiguous. Consider the grammar shown in Example 3.3, which is a minor variation of the grammar shown in Example 3.2.

EXAMPLE 3.3 An Ambiguous Grammar for Simple Assignment Statements

<assign> → <id> = <expr>

<id> → A | B | C

<expr> → <expr> + <expr>

| <expr> * <expr>

| ( <expr> )

| <id>

The grammar of Example 3.3 is ambiguous because the sentence

A = B + C * A

has two distinct parse trees, as shown in Figure 3.2. The ambiguity occurs because the grammar specifies slightly less syntactic structure than does the grammar of Example 3.2. Rather than allowing the parse tree of an expression to grow only on the right, this grammar allows growth on both the left and the right.

Figure 3.2 Two distinct parse trees for the same sentence, A = B + C * A

Figure 3.2 Full Alternative Text

Syntactic ambiguity of language structures is a problem because compilers often base the semantics of those structures on their syntactic form.

Specifically, the compiler chooses the code to be generated for a statement by examining its parse tree. If a language structure has more than one parse tree, then the meaning of the structure cannot be determined uniquely. This problem is discussed in two specific examples in the following subsections.

There are several other characteristics of a grammar that are sometimes useful in determining whether a grammar is ambiguous.1 They include the following: (1) if the grammar generates a sentence with more than one leftmost derivation and (2) if the grammar generates a sentence with more than one rightmost derivation.

1. Note that it is mathematically impossible to determine whether an arbitrary grammar is ambiguous.

Some parsing algorithms can be based on ambiguous grammars. When such a parser encounters an ambiguous construct, it uses nongrammatical information provided by the designer to construct the correct parse tree. In many cases, an ambiguous grammar can be rewritten to be unambiguous but still generate the desired language.

3.3.1.8 Operator Precedence When an expression includes two different operators, for example, x + y * z, one obvious semantic issue is the order of evaluation of the two operators (for example, in this expression is it add and then multiply, or vice versa?). This semantic question can be answered by assigning different precedence levels to operators. For example, if * has been assigned higher precedence than + (by the language designer), multiplication will be done first, regardless of the order of appearance of the two operators in the expression.

As stated previously, a grammar can describe a certain syntactic structure so that part of the meaning of the structure can be determined from its parse tree. In particular, the fact that an operator in an arithmetic expression is generated lower in the parse tree (and therefore must be evaluated first) can be used to indicate that it has precedence over an operator produced higher up in the tree. In the first parse tree of Figure 3.2, for example, the multiplication

operator is generated lower in the tree, which could indicate that it has precedence over the addition operator in the expression. The second parse tree, however, indicates just the opposite. It appears, therefore, that the two parse trees indicate conflicting precedence information.

Notice that although the grammar of Example 3.2 is not ambiguous, the precedence order of its operators is not the usual one. In this grammar, a parse tree of a sentence with multiple operators, regardless of the particular operators involved, has the rightmost operator in the expression at the lowest point in the parse tree, with the other operators in the tree moving progressively higher as one moves to the left in the expression. For example, in the expression A + B * C, * is the lowest in the tree, indicating it is to be done first. However, in the expression A * B + C, + is the lowest, indicating it is to be done first.

A grammar can be written for the simple expressions we have been discussing that is both unambiguous and specifies a consistent precedence of the + and * operators, regardless of the order in which the operators appear in an expression. The correct ordering is specified by using separate nonterminal symbols to represent the operands of the operators that have different precedence. This requires additional nonterminals and some new rules. Instead of using <expr> for both operands of both + and *, we could use three nonterminals to represent operands, which allows the grammar to force different operators to different levels in the parse tree. If <expr> is the root symbol for expressions, + can be forced to the top of the parse tree by having <expr> directly generate only + operators, using the new nonterminal, <term>, as the right operand of +. Next, we can define <term> to generate * operators, using <term> as the left operand and a new nonterminal, <factor>, as its right operand. Now, * will always be lower in the parse tree, simply because it is farther from the start symbol than + in every derivation. The grammar of Example 3.4 is such a grammar.

EXAMPLE 3.4 An Unambiguous Grammar for Expressions

<assign> → <id> = <expr>

<id> → A | B | C

<expr> → <expr> + <term>

| <term>

<term> → <term> * <factor>

| <factor>

<factor> → ( <expr> )

| <id>

The grammar in Example 3.4 generates the same language as the ​grammars of Examples 3.2 and 3.3, but it is unambiguous and it specifies the usual precedence order of multiplication and addition operators. The following derivation of the sentence A = B + C * A uses the grammar of Example 3.4:

<assign> => <id> = <expr>

    => A = <expr>

    => A = <expr> + <term>

    => A = <term> + <term>

    => A = <factor> + <term>

    => A = <id> + <term>

    => A = B + <term>

    => A = B + <term> * <factor>

    => A = B + <factor> * <factor>

    => A = B + <id> * <factor>

    => A = B + C * <factor>

    => A = B + C * <id>

    => A = B + C * A

The unique parse tree for this sentence, using the grammar of ​Example 3.4, is shown in Figure 3.3.

Figure 3.3 The unique parse

tree for A = B + C * A using an unambiguous grammar

Figure 3.3 Full Alternative Text

The connection between parse trees and derivations is very close: Either can easily be constructed from the other. Every derivation with an unambiguous grammar has a unique parse tree, although that tree can be represented by different derivations. For example, the following derivation of the sentence A = B + C * A is different from the derivation of the same sentence given previously. This is a rightmost derivation, whereas the previous one is leftmost. Both of these derivations, however, are represented by the same parse tree.

<assign> => <id> = <expr>

    => <id> = <expr> + <term>

    => <id> = <expr> + <term> * <factor>

    => <id> = <expr> + <term> * <id>

    => <id> = <expr> + <term> * A

    => <id> = <expr> + <factor> * A

    => <id> = <expr> + <id> * A

    => <id> = <expr> + C * A

    => <id> = <term> + C * A

    => <id> = <factor> + C * A

    => <id> = <id> + C * A

    => <id> = B + C * A

    => A = B + C * A

3.3.1.9 Associativity of Operators When an expression includes two operators that have the same precedence (as * and / usually have)—for example, A / B * C—a semantic rule is required to specify which should have precedence.2 This rule is named associativity.

2. An expression with two occurrences of the same operator has the same issue; for example, A / B / C.

As was the case with precedence, a grammar for expressions may correctly imply operator associativity. Consider the following example of an assignment statement:

A = B + C + A

The parse tree for this sentence, as defined with the grammar of Example 3.4, is shown in Figure 3.4.

Figure 3.4 A parse tree for A = B + C + A illustrating the

associativity of addition Figure 3.4 Full Alternative Text

The parse tree of Figure 3.4 shows the left addition operator lower than the right addition operator. This is the correct order if addition is meant to be left associative, which is typical. In most cases, the associativity of addition in a computer is irrelevant. In mathematics, addition is associative, which means that left and right associative orders of evaluation mean the same thing. That is, (A + B) + C = A + (B + C). Floating-point addition in a computer, however, is not necessarily associative. For example, suppose floating-point values store seven digits of accuracy. Consider the problem of adding 11 numbers together, where one of the numbers is 107 and the other ten are 1. If the small numbers (the 1’s) are each added to the large number, one at a time, there is no effect on that number, because the small numbers occur in the eighth digit of the large number. However, if the small numbers are first added together and the result is added to the large number, the result in seven-digit accuracy is 1.000001 * 107. Subtraction and division are not associative, whether in mathematics or in a computer. Therefore, correct associativity may be essential for an expression that contains either of them.

When a grammar rule has its LHS also appearing at the beginning of its RHS, the rule is said to be left recursive. This left recursion specifies left associativity. For example, the left recursion of the rules of the grammar of Example 3.4 causes it to make both addition and multiplication left associative. Unfortunately, left recursion disallows the use of some important syntax analysis algorithms. When one of these algorithms is to be used, the grammar must be modified to remove the left recursion. This, in turn, disallows the grammar from precisely specifying that certain operators are left associative. Fortunately, left associativity can be enforced by the compiler, even though the grammar does not dictate it.

In most languages that provide it, the exponentiation operator is right associative. To indicate right associativity, right recursion can be used. A grammar rule is right recursive if the LHS appears at the right end of the RHS. Rules such as

<factor> → <exp> ** <factor>

|<exp>

<exp> → ( <expr> )

|id

could be used to describe exponentiation as a right-associative operator.

3.3.1.10 An Unambiguous Grammar for if-else The BNF rules for a Java if-else statement are as follows:

<if_stmt> → if (<logic_expr>) <stmt>

if (<logic_expr>) <stmt> else <stmt>

If we also have <stmt> → <if_stmt>, this grammar is ambiguous. The simplest sentential form that illustrates this ambiguity is

if (<logic_expr>) if (<logic_expr>) <stmt> else <stmt>

The two parse trees in Figure 3.5 show the ambiguity of this sentential form. Consider the following example of this construct:

Figure 3.5 Two distinct parse trees for the same sentential form

Figure 3.5 Full Alternative Text

if (done == true)

if (denom == 0)

  quotient = 0;

  else quotient = num / denom;

The problem is that if the upper parse tree in Figure 3.5 is used as the basis for translation, the else clause would be executed when done is not true, which probably is not what was intended by the author of the construct. We will examine the practical problems associated with this else-association problem in Chapter 8.

We will now develop an unambiguous grammar that describes this if statement. The rule for if constructs in many languages is that an else clause, when present, is matched with the nearest previous unmatched then clause. Therefore, there cannot be an if statement without an else between a then clause and its matching else. So, for this situation, statements must be distinguished between those that are matched and those that are unmatched, where unmatched statements are else-less ifs and all other statements are matched. The problem with the earlier grammar is that it treats all statements as if they had equal syntactic significance—that is, as if they were all matched.

To reflect the different categories of statements, different abstractions, or nonterminals, must be used. The unambiguous grammar based on these ideas follows:

<stmt> → <matched> | <unmatched>

<matched> → if (<logic_expr>) <matched> else <matched>

|any non-if statement

<unmatched> → if (<logic_expr>) <stmt>

|if (<logic_expr>) <matched> else <unmatched>

There is just one possible parse tree, using this grammar, for the following sentential form:

if (<logic_expr>) if (<logic_expr>) <stmt> else <stmt>

3.3.2 Extended BNF Because of a few minor inconveniences in BNF, it has been extended in several ways. Most extended versions are called Extended BNF, or simply EBNF, even though they are not all exactly the same. The extensions do not enhance the descriptive power of BNF; they only increase its readability and writability.

Three extensions are commonly included in the various versions of EBNF. The first of these denotes an optional part of an RHS, which is delimited by brackets. For example, a C if-else statement can be described as

<if_stmt> → if (<expression>) <statement> [else <statement>]

Without the use of the brackets, the syntactic description of this statement would require the following two rules:

<if_stmt> → if (<expression>) <statement>

|if (<expression>) <statement> else <statement>

The second extension is the use of braces in a RHS to indicate that the enclosed part can be repeated indefinitely or left out altogether. This extension allows lists to be built with a single rule, instead of using recursion and two rules. For example, lists of identifiers separated by commas can be described by the following rule:

<ident_list> → <identifier> {, <identifier>}

This is a replacement of the recursion by a form of implied iteration; the part

enclosed within braces can be iterated any number of times.

The third common extension deals with multiple-choice options. When a single element must be chosen from a group, the options are placed in parentheses and separated by the OR operator, |. For example,

<term> → <term> (* | / | %) <factor>

In BNF, a description of this <term> would require the following three rules:

<term> → <term> * <factor>

| <term> / <factor>

| <term> % <factor>

The brackets, braces, and parentheses in the EBNF extensions are metasymbols, which means they are notational tools and not terminal symbols in the syntactic entities they help describe. In cases where these metasymbols are also terminal symbols in the language being described, the instances that are terminal symbols can be underlined or quoted. Example 3.5 illustrates the use of braces and multiple choices in an EBNF grammar.

EXAMPLE 3.5 BNF and EBNF Versions of an Expression Grammar BNF:

<expr> → <expr> + <term>

| <expr> - <term>

| <term>

<term> → <term> * <factor>

| <term> / <factor>

| <factor>

<factor> → <exp> ** <factor>

<exp>

<exp> → (<expr>)

| id

EBNF:

<expr> → <term> {(+ | -) <term>}

<term> → <factor> {(* | /) <factor>}

<factor> → <exp> { ** <exp>}

<exp> → (<expr>)

| id

The BNF rule

<expr> → <expr> + <term>

clearly specifies—in fact forces—the + operator to be left associative. However, the EBNF version,

<expr> → <term> {+ <term>}

does not imply the direction of associativity. This problem is overcome in a syntax analyzer based on an EBNF grammar for expressions by designing the syntax analysis process to enforce the correct associativity. This is discussed further in Chapter 4.

Some versions of EBNF allow a numeric superscript to be attached to the right brace to indicate an upper limit to the number of times the enclosed part can be repeated. Also, some versions use a plus (+) superscript to indicate one or more repetitions. For example,

<compound> → begin <stmt> {<stmt>} end

and

<compound> → begin {<stmt>}+ end

are equivalent.

In recent years, some variations on BNF and EBNF have appeared. Among these are the following:

In place of the arrow, a colon is used and the RHS is placed on the next line.

Instead of a vertical bar to separate alternative RHSs, they are simply placed on separate lines.

In place of square brackets to indicate something being optional, the subscript opt is used. For example,

Constructor Declarator→SimpleName (FormalParameterListopt)

Rather than using the | symbol in a parenthesized list of elements to indicate a choice, the words “one of” are used. For example,

AssignmentOperator → one of = *= /= %= += -=

<<= >>= &= ^= |=

There is a standard for EBNF, ISO/IEC 14977:1996 (1996), but it is rarely used. The standard uses the equal sign (=) instead of an arrow in rules, terminates each RHS with a semicolon, and requires quotes on all terminal symbols. It also specifies a host of other notational rules.

3.3.3 Grammars and Recognizers Earlier in this chapter, we suggested that there is a close relationship between generation and recognition devices for a given language. In fact, given a ​- context-free grammar, a recognizer for the language generated by the grammar can be algorithmically constructed. A number of software systems have been developed that perform this construction. Such systems allow the quick creation of the syntax analysis part of a compiler for a new language and are therefore quite valuable. One of the first of these syntax analyzer generators is named yacc (yet another compiler compiler) (Johnson, 1975). There are now many such systems available.

3.4 Attribute Grammars An attribute grammar is a device used to describe more of the structure of a programming language than can be described with a context-free grammar. An attribute grammar is an extension to a context-free grammar. The extension allows certain language rules to be conveniently described, such as type compatibility. Before we formally define the form of attribute grammars, we must clarify the concept of static semantics.

history note Attribute grammars have been used in a wide variety of applications. They have been used to provide complete descriptions of the syntax and static semantics of programming languages (Watt, 1979); they have been used as the formal definition of a language that can be input to a compiler generation system (Farrow, 1982); and they have been used as the basis of several syntax-directed editing systems (Teitelbaum and Reps, 1981; Fischer et al., 1984). In addition, attribute grammars have been used in natural-language processing systems (Correa, 1992).

3.4.1 Static Semantics There are some characteristics of programming languages that are difficult to describe with BNF, and some that are impossible. As an example of a syntax rule that is difficult to specify with BNF, consider type compatibility rules. In Java, for example, a floating-point value cannot be assigned to an integer type variable, although the opposite is legal. Although this restriction can be specified in BNF, it requires additional nonterminal symbols and rules. If all of the typing rules of Java were specified in BNF, the grammar would become too large to be useful, because the size of the grammar determines the size of the syntax analyzer.

As an example of a syntax rule that cannot be specified in BNF, consider the common rule that all variables must be declared before they are referenced. It has been proven that this rule cannot be specified in BNF.

These problems exemplify the categories of language rules called static semantics rules. The static semantics of a language is only indirectly related to the meaning of programs during execution; rather, it has to do with the legal forms of programs (syntax rather than semantics). Many static semantic rules of a language state its type constraints. Static semantics is so named because the analysis required to check these specifications can be done at compile time.

Because of the problems of describing static semantics with BNF, a variety of more powerful mechanisms has been devised for that task. One such mechanism, attribute grammars, was designed by Knuth (1968) to describe both the syntax and the static semantics of programs.

Attribute grammars are a formal approach both to describing and checking the correctness of the static semantics rules of a program. Although they are not always used in a formal way in compiler design, the basic concepts of attribute grammars are at least informally used in every compiler (see Aho et al., 1988).

Dynamic semantics, which is the meaning of expressions, statements, and program units, is discussed in Section 3.5.

3.4.2 Basic Concepts Attribute grammars are context-free grammars to which have been added attributes, attribute computation functions, and predicate functions. Attributes, which are associated with grammar symbols (the terminal and nonterminal symbols), are similar to variables in the sense that they can have values assigned to them. Attribute computation functions, sometimes called semantic functions, are associated with grammar rules. They are used to specify how attribute values are computed. Predicate functions, which state the static semantic rules of the language, are associated with grammar

rules.

These concepts will become clearer after we formally define attribute grammars and provide an example.

3.4.3 Attribute Grammars Defined An attribute grammar is a grammar with the following additional features:

Associated with each grammar symbol X is a set of attributes A(X). The set A(X) consists of two disjoint sets S(X) and I(X), called synthesized and inherited attributes, respectively. Synthesized attributes are used to pass semantic information up a parse tree, while inherited attributes pass semantic information down and across a tree.

Associated with each grammar rule is a set of semantic functions and a possibly empty set of predicate functions over the attributes of the symbols in the grammar rule. For a rule the X0→X1 … Xn, synthesized attributes of X0 are computed with semantic functions of the form S(X0)=f(A(X1), …, A(Xn)). So the value of a synthesized attribute on a parse tree node depends only on the values of the attributes on that node’s children nodes. Inherited attributes of symbols Xj, 1≤j≤n (in the rule above), are computed with a semantic function of the form I(Xj=f(A(X0), …, A(Xn)). So the value of an inherited attribute on a parse tree node depends on the attribute values of that node’s parent node and those of its sibling nodes. Note that, to avoid circularity, inherited attributes are often restricted to functions of the form I(Xj)=f(A(X0), …, A(X(_{j−1}))). This form prevents an inherited attribute from depending on itself or on attributes to the right in the parse tree.

A predicate function has the form of a Boolean expression on the union of the attribute set {A(X0), …, A(Xn)} and a set of literal attribute values. The only derivations allowed with an attribute grammar are those in which every predicate associated with every nonterminal is true. A false predicate function value indicates a violation of the syntax or

static semantics rules of the language.

A parse tree of an attribute grammar is the parse tree based on its ​underlying BNF grammar, with a possibly empty set of attribute values attached to each node. If all the attribute values in a parse tree have been computed, the tree is said to be fully attributed. Although in practice it is not always done this way, it is convenient to think of attribute values as being computed after the complete unattributed parse tree has been constructed by the compiler.

3.4.4 Intrinsic Attributes Intrinsic attributes are synthesized attributes of leaf nodes whose values are determined outside the parse tree. For example, the type of an instance of a variable in a program could come from the symbol table, which is used to store variable names and their types. The contents of the symbol table are set based on earlier declaration statements. Initially, assuming that an unattributed parse tree has been constructed and that attribute values are needed, the only attributes with values are the intrinsic attributes of the leaf nodes. Given the intrinsic attribute values on a parse tree, the semantic functions can be used to compute the remaining attribute values.

3.4.5 Examples of Attribute Grammars As a very simple example of how attribute grammars can be used to describe static semantics, consider the following fragment of an attribute grammar that describes the rule that the name on the end of an Ada procedure must match the procedure’s name. (This rule cannot be stated in BNF.) The string attribute of <proc_name>, denoted by <proc_name>.string, is the actual string of characters that were found immediately following the reserved word procedure by the compiler. Notice that when there is more than one occurrence of a nonterminal in a syntax rule in an attribute grammar, the nonterminals are subscripted with brackets to distinguish them. Neither the

subscripts nor the brackets are part of the described language.

Syntax rule: <proc_def> → procedure <proc_name>[1]

               <proc_body> end <proc_name>[2];

Predicate: <proc_name>[1]string == <proc_name>[2].string

In this example, the predicate rule states that the name string attribute of the <proc_name> nonterminal in the subprogram header must match the name string attribute of the <proc_name> nonterminal following the end of the subprogram.

Next, we consider a larger example of an attribute grammar. In this case, the example illustrates how an attribute grammar can be used to check the type rules of a simple assignment statement. The syntax and static semantics of this assignment statement are as follows: The only variable names are A, B, and C. The right side of the assignments can be either a variable or an expression in the form of a variable added to another variable. The variables can be one of two types: int or real. When there are two variables on the right side of an assignment, they need not be the same type. The type of the expression when the operand types are not the same is always real. When they are the same, the expression type is that of the operands. The type of the left side of the assignment must match the type of the right side. So the types of operands in the right side can be mixed, but the assignment is valid only if the target and the value resulting from evaluating the right side have the same type. The attribute grammar specifies these static semantic rules.

The syntax portion of our example attribute grammar is

<assign> → <var> = <expr>

<expr> → <var> + <var>

| <var>

<var> → A | B | C

The attributes for the nonterminals in the example attribute grammar are

described in the following paragraphs:

actual_type—A synthesized attribute associated with the nonterminals <var> and <expr>. It is used to store the actual type, int or real, of a variable or expression. In the case of a variable, the actual type is intrinsic. In the case of an expression, it is determined from the actual types of the child node or children nodes of the <expr> nonterminal.

expected_type—An inherited attribute associated with the nonterminal <expr>. It is used to store the type, either int or real, that is expected for the expression, as determined by the type of the variable on the left side of the assignment statement.

The complete attribute grammar follows in Example 3.6.

EXAMPLE 3.6 An Attribute Grammar for Simple Assignment Statements

1. Syntax rule: <assign> → <var> = <expr>

Semantic rule: <expr>.expected_type ← <var>.actual_type

2. Syntax rule: <expr> → <var>[2] + <var>[3]

Semantic rule: <expr>.actual_type ←

            if (<var>[2].actual_type = int) and

              (<var>[3].actual_type = int)

            then int

           else real

           end if

Predicate: <expr>.actual_type == <expr>.expected_type

3. Syntax rule: <expr> → <var>

Semantic rule: <expr>.actual_type ← <var>.actual_type

Predicate: <expr>.actual_type == <expr>.expected_type

4. Syntax rule: <var> → A | B | C

Semantic rule: <var>.actual_type ← look-up(<var>.string)

The look-up function looks up a given variable name in the symbol table and returns the variable’s type.

A parse tree of the sentence A = A + B generated by the grammar in Example 3.6 is shown in Figure 3.6. As in the grammar, bracketed numbers are added after the repeated node labels in the tree so they can be referenced unambiguously.

Figure 3.6 A parse tree for A = A + B

Figure 3.6 Full Alternative Text

3.4.6 Computing Attribute Values Now, consider the process of computing the attribute values of a parse tree, which is sometimes called decorating the parse tree. If all attributes were inherited, this could proceed in a completely top-down order, from the root to the leaves. Alternatively, it could proceed in a completely bottom-up order, from the leaves to the root, if all the attributes were synthesized. Because our

grammar has both synthesized and inherited attributes, the evaluation process cannot be in any single direction. The following is an evaluation of the attributes, in an order in which it is possible to compute them:

1. <var>.actual_type ← look-up(A) (Rule 4)

2. <expr>.expected_type ← <var>.actual_type (Rule 1)

3. <var>[2].actual_type ← look-up(A) (Rule 4)

<var>[3].actual_type ← look-up(B) (Rule 4)

4. <expr>.actual_type ← either int or real (Rule 2)

5. <expr>.expected_type == <expr>.actual_type is either

            TRUE or FALSE (Rule 2)

The tree in Figure 3.7 shows the flow of attribute values in the example of Figure 3.6. Solid lines show the parse tree; dashed lines show attribute flow in the tree.

Figure 3.7 The flow of attributes in the tree

Figure 3.7 Full Alternative Text

The tree in Figure 3.8 shows the final attribute values on the nodes. In this example, A is defined as a real and B is defined as an int.

Figure 3.8 A fully attributed parse tree

Figure 3.8 Full Alternative Text

Determining attribute evaluation order for the general case of an attribute grammar is a complex problem, requiring the construction of a dependency graph to show all attribute dependencies.

3.4.7 Evaluation Checking the static semantic rules of a language is an essential part of all compilers. Even if a compiler writer has never heard of an attribute grammar, he or she would need to use the fundamental ideas of attribute grammars to design the checks of static semantics rules for his or her compiler.

One of the main difficulties in using an attribute grammar to describe all of the syntax and static semantics of a real contemporary programming ​- language is the size and complexity of the attribute grammar. The large number of attributes and semantic rules required for a complete programming

language make such grammars difficult to write and read. Furthermore, the attribute values on a large parse tree are costly to evaluate. On the other hand, less formal attribute grammars are a powerful and commonly used tool for compiler writers, who are more interested in the process of producing a compiler than they are in formalism.

3.5 Describing the Meanings of Programs: Dynamic Semantics We now turn to the difficult task of describing the dynamic semantics, or meaning, of the expressions, statements, and program units of a programming language. Because of the power and naturalness of the available notation, describing syntax is a relatively simple matter. On the other hand, no universally accepted notation or approach has been devised for dynamic semantics. In this section, we briefly describe several of the methods that have been developed. For the remainder of this section, when we use the term semantics, we mean dynamic semantics.

There are several different reasons underlying the need for a methodology and notation for describing semantics. Programmers obviously need to know precisely what the statements of a language do before they can use them effectively in their programs. Compiler writers must know exactly what language constructs mean to design implementations for them correctly. If there were a precise semantics specification of a programming language, programs written in the language potentially could be proven correct without testing. Also, compilers could be shown to produce programs that exhibited exactly the behavior given in the language definition; that is, their correctness could be verified. A complete specification of the syntax and semantics of a programming language could be used by a tool to generate a compiler for the language automatically. Finally, language designers, who would develop the semantic descriptions of their languages, could in the process discover ambiguities and inconsistencies in their designs.

Software developers and compiler designers typically determine the semantics of programming languages by reading English explanations in language manuals. Because such explanations are often imprecise and incomplete, this approach is clearly unsatisfactory. Due to the lack of complete semantics specifications of programming languages, programs are rarely proven correct without testing, and commercial compilers are never generated automatically from language descriptions.

Scheme, a functional language described in Chapter 15, is one of only a few programming languages whose definition includes a formal semantics description. However, the method used is not one described in this chapter, as this chapter is focused on approaches that are suitable for imperative languages.

3.5.1 Operational Semantics The idea behind operational semantics is to describe the meaning of a statement or program by specifying the effects of running it on a machine. The effects on the machine are viewed as the sequence of changes in its state, where the machine’s state is the collection of the values in its storage. An obvious operational semantics description, then, is given by executing a compiled version of the program on a computer. Most programmers have, on at least one occasion, written a small test program to determine the meaning of some programming language construct, often while learning the language. Essentially, what such a programmer is doing is using operational semantics to determine the meaning of the construct.

There are several problems with using this approach for complete formal semantics descriptions. First, the individual steps in the execution of machine language and the resulting changes to the state of the machine are too small and too numerous. Second, the storage of a real computer is too large and complex. There are usually several levels of memory devices, as well as connections to enumerable other computers and memory devices through networks. Therefore, machine languages and real computers are not used for formal operational semantics. Rather, intermediate-level languages and interpreters for idealized computers are designed specifically for the process.

There are different levels of uses of operational semantics. At the highest level, the interest is in the final result of the execution of a complete program. This is sometimes called natural operational semantics. At the lowest level, operational semantics can be used to determine the precise meaning of a program through an examination of the complete sequence of state changes that occur when the program is executed. This use is sometimes called structural operational semantics.

3.5.1.1 The Basic Process The first step in creating an operational semantics description of a language is to design an appropriate intermediate language, where the primary desired characteristic of the language is clarity. Every construct of the intermediate language must have an obvious and unambiguous meaning. This language is at the intermediate level, because machine language is too low-level to be easily understood and another high-level language is obviously not suitable. If the semantics description is to be used for natural operational semantics, a virtual machine (an interpreter) must be constructed for the intermediate language. The virtual machine can be used to execute either single statements, code segments, or whole programs. The semantics description can be used without a virtual machine if the meaning of a single statement is all that is required. In this use, which is structural operational semantics, the intermediate code can be visually inspected.

The basic process of operational semantics is not unusual. In fact, the concept is frequently used in programming textbooks and programming language reference manuals. For example, the semantics of the C for construct can be described in terms of simpler statements, as in

The human reader of such a description is the virtual computer and is assumed to be able to “execute” the instructions in the definition correctly and recognize the effects of the “execution.”

The intermediate language and its associated virtual machine used for formal operational semantics descriptions are often highly abstract. The intermediate

language is meant to be convenient for the virtual machine, rather than for human readers. For our purposes, however, a more human-oriented intermediate language could be used. As such an example, consider the following list of statements, which would be adequate for describing the semantics of the simple control statements of a typical programming language:

ident = var

ident = ident + 1

ident = ident – 1

goto label

if var relop var goto label

In these statements, relop is one of the relational operators from the set {=, <>, >, <, <=, <=}, ident is an identifier, and var is either an identifier or a constant. These statements are all simple and therefore easy to understand and implement.

A slight generalization of these three assignment statements allows more general arithmetic expressions and assignment statements to be described. The new statements are

ident = var bin_op var

ident = un_op var

where bin_op is a binary arithmetic operator and un_op is a unary operator. Multiple arithmetic data types and automatic type conversions, of course, complicate this generalization. Adding just a few more relatively simple instructions would allow the semantics of arrays, records, pointers, and subprograms to be described.

In Chapter 8, the semantics of various control statements are described using this intermediate language.

3.5.1.2 Evaluation The first and most significant use of formal operational semantics was to describe the semantics of PL/I (Wegner, 1972). That particular abstract machine and the translation rules for PL/I were together named the Vienna Definition Language (VDL), after the city where IBM designed it.

Operational semantics provides an effective means of describing semantics for language users and language implementors, as long as the descriptions are kept simple and informal. The VDL description of PL/I, unfortunately, is so complex that it serves no practical purpose.

Operational semantics depends on programming languages of lower levels, not mathematics. The statements of one programming language are described in terms of the statements of a lower-level programming language. This approach can lead to circularities, in which concepts are indirectly defined in terms of themselves. The methods described in the following two sections are much more formal, in the sense that they are based on mathematics and logic, not programming languages.

3.5.2 Denotational Semantics Denotational semantics is the most rigorous and most widely known formal method for describing the meaning of programs. It is solidly based on recursive function theory. A thorough discussion of the use of denotational semantics to describe the semantics of programming languages is necessarily long and complex. It is our intent to provide the reader with an introduction to the central concepts of denotational semantics, along with a few simple examples that are relevant to programming language specifications.

The process of constructing a denotational semantics specification for a programming language requires one to define for each language entity both a mathematical object and a function that maps instances of that language entity onto instances of the mathematical object. Because the objects are rigorously defined, they model the exact meaning of their corresponding

entities. The idea is based on the fact that there are rigorous ways of manipulating mathematical objects but not programming language constructs. The difficulty with this method lies in creating the objects and the mapping functions. The method is named denotational because the mathematical objects denote the meaning of their corresponding syntactic entities.

The mapping functions of a denotational semantics programming language specification, like all functions in mathematics, have a domain and a range. The domain is the collection of values that are legitimate parameters to the function; the range is the collection of objects to which the parameters are mapped. In denotational semantics, the domain is called the syntactic domain, because it is syntactic structures that are mapped. The range is called the semantic domain.

Denotational semantics is related to operational semantics. In operational semantics, programming language constructs are translated into simpler programming language constructs, which become the basis of the meaning of the construct. In denotational semantics, programming language constructs are mapped to mathematical objects, either sets or, more often, functions. However, unlike operational semantics, denotational semantics does not model the step-by-step computational processing of programs.

3.5.2.1 Two Simple Examples We use a very simple language construct, character string representations of binary numbers, to introduce the denotational method. The syntax of such binary numbers can be described by the following grammar rules:

<bin_num> → '0'

      | '1'

      | <bin_num> '0'

      | <bin_num> '1'

A parse tree for the example binary number, 110, is shown in Figure 3.9.

Notice that we put apostrophes around the syntactic digits to show they are not mathematical digits. This is similar to the relationship between ASCII coded digits and mathematical digits. When a program reads a number as a string, it must be converted to a mathematical number before it can be used as a value in the program.

Figure 3.9 A parse tree of the binary number 110

Figure 3.9 Full Alternative Text

The syntactic domain of the mapping function for binary numbers is the set of all character string representations of binary numbers. The semantic domain is the set of nonnegative decimal numbers, symbolized by N.

To describe the meaning of binary numbers using denotational semantics, we associate the actual meaning (a decimal number) with each rule that has a single terminal symbol as its RHS.

In our example, decimal numbers must be associated with the first two grammar rules. The other two grammar rules are, in a sense, computational rules, because they combine a terminal symbol, to which an object can be associated, with a nonterminal, which can be expected to represent some construct. Presuming an evaluation that progresses upward in the parse tree, the nonterminal in the right side would already have its meaning attached. So, a syntax rule with a nonterminal as its RHS would require a function that computed the meaning of the LHS, which represents the meaning of the complete RHS.

The semantic function, named Mbin, maps the syntactic objects, as described in the previous grammar rules, to the objects in N, the set of non-negative decimal numbers. The function Mbin is defined as follows:

Mbin ('0') = 0

Mbin ('1') = 1

Mbin (<bin_num> '0') = 2 * Mbin (<bin_num>)

Mbin (<bin_num>'1') = 2 * Mbin (<bin_num>) + 1

The meanings, or denoted objects (which in this case are decimal numbers), can be attached to the nodes of the parse tree shown on the previous page, yielding the tree in Figure 3.10. This is syntax-directed semantics. Syntactic entities are mapped to mathematical objects with concrete meaning.

Figure 3.10 A parse tree with denoted objects for 110

Figure 3.10 Full Alternative Text

In part because we need it later, we now show a similar example for describing the meaning of syntactic decimal literals. In this case, the syntactic domain is the set of character string representations of decimal numbers. The semantic domain is once again the set N.

<dec_num> → '0'|'1'|'2'|'3'|'4'|'5'|'6'|'7''8'|'9'

    |<dec_num> ('0'|'1'|'2'|'3'|'4'|'5'|'6'|'7'|'8'|'9')

The denotational mappings for these syntax rules are

Mdec ('0') = 0, Mdec ('1') = 1, Mdec ('2') = 2, . . ., Mdec ('9') = 9

Mdec (<dec_num> '0') = 10 * Mdec (<dec_num>)

Mdec (<dec_num> '1') = 10 * Mdec (<dec_num>) +1

. . .

Mdec (<dec_num> '9') = 10 * Mdec (<dec_num>) +9

In the following sections, we present the denotational semantics descriptions of a few simple constructs. The most important simplifying assumption made here is that both the syntax and static semantics of the constructs are correct. In addition, we assume that only two scalar types are included: integer and Boolean.

3.5.2.2 The State of a Program The denotational semantics of a program could be defined in terms of state changes in an ideal computer. Operational semantics are defined in this way, and denotational semantics are defined in nearly the same way. In a further simplification, however, denotational semantics is defined in terms of only the values of all of the program’s variables. So, denotational semantics uses the state of the program to describe meaning, whereas operational semantics uses the state of a machine. The key difference between operational semantics and denotational semantics is that state changes in operational semantics are defined by coded algorithms, written in some programming language, whereas in denotational semantics, state changes are defined by mathematical functions.

Let the state s of a program be represented as a set of ordered pairs, as follows:

s = { <i1, v1>, <i2, v2>, …, <in, vn>}

Each i is the name of a variable, and the associated v’s are the current values of those variables. Any of the v’s can have the special value undef, which indicates that its associated variable is currently undefined. Let VARMAP be a function of two parameters: a variable name and the program state. The

value of ​VARMAP ( ij, s) is vj (the value paired with ij in state s). Most semantics mapping functions for programs and program constructs map states to states. These state changes are used to define the meanings of programs and program constructs. Some language constructs—for example, expressions—are mapped to values, not states.

3.5.2.3 Expressions Expressions are fundamental to most programming languages. We assume here that expressions have no side effects. Furthermore, we deal with only very simple expressions: The only operators are + and *, and an expression can have at most one operator; the only operands are scalar integer variables and integer literals; there are no parentheses; and the value of an expression is an integer. Following is the BNF description of these expressions:

<expr> → <dec_num> | <var> | <binary_expr>

<binary_expr> → <left_expr> <operator> <right_expr>

<left_expr> → <dec_num> | <var>

<right_expr> → <dec_num> | <var>

<operator> → + | *

The only error we consider in expressions is a variable having an undefined value. Obviously, other errors can occur, but most of them are machine-​- dependent. Let Z be the set of integers, and let error be the error value. Then Z∪ {error} is the semantic domain for the denotational specification for our expressions.

The mapping function for a given expression E and state s follows. To distinguish between mathematical function definitions and the assignment statements of programming languages, we use the symbol Δ = to define mathematical functions. The implication symbol, =>, used in this definition connects the form of an operand with its associated case (or switch) construct. Dot notation is used to refer to the child nodes of a node. For

example, <binary_expr>.<left_expr> refers to the left child node of <binary_expr>.

Me(<expr>, s) Δ= case <expr> of

         <dec_num>=> Mdec (<dec_num>, s)

         <var> =>if VARMAP (<var>, s) == undef

               then error

               else VARMAP(<var>, s)

         <binary_expr> =>

          if( Me (<binary_expr>.<left_expr>,s) == undef OR

            Me (<binary_expr>.<right_expr>, s) == undef)

          then error

          else if (<binary_expr>.<operator> == '+')

              then Me (<binary_expr>.<left_expr>, s) +

                  Me (<binary_expr>.<right_expr>, s)

              else Me (<binary_expr>.<left_expr>, s) *

                 Me (<binary_expr>.<right_expr>, s)

3.5.2.4 Assignment Statements

An assignment statement is an expression evaluation plus the setting of the target variable to the expression’s value. In this case, the meaning function maps a state to a state. This function can be described with the following:

Ma (x = E, s) Δ= if Me (E, s) == error

            then error

            else s' = {< i1, v1′ >, < i2, v2′>, . . . , < in, vn′>}, where

                for j = 1, 2, . . . , n

                 if ij == x

                then vj′ = Me (E, s)

                else vj′ = VARMAP( ij,, s)

Note that the comparison in the third last line above, ij == x, is of names, not values.

3.5.2.5 Logical Pretest Loops The denotational semantics of a logical pretest loop is deceptively simple. To expedite the discussion, we assume that there are two other existing mapping functions, Msl and Mb, that map statement lists and states to states and Boolean expressions to Boolean values (or error), respectively. The function is

M1 (while B do L, s) Δ= if Mb (B, s) == undef

            then error

            else if Mb (B, s) == false

              then s

              else if Msl (L, s) == error

                 then error

                 else M1 (while B do L, Msl (L, s))

The meaning of the loop is simply the value of the program variables after the statements in the loop have been executed the prescribed number of times, assuming there have been no errors. In essence, the loop has been converted from iteration to recursion, where the recursion control is mathematically defined by other recursive state mapping functions. Recursion is easier to describe with mathematical rigor than iteration.

One significant observation at this point is that this definition, like actual program loops, may compute nothing because of nontermination.

3.5.2.6 Evaluation Objects and functions, such as those used in the earlier constructs, can be defined for the other syntactic entities of programming languages. When a complete system has been defined for a given language, it can be used to determine the meaning of complete programs in that language. This provides a framework for thinking about programming in a highly rigorous way.

history note A significant amount of work has been done on the possibility of using denotational language descriptions to generate compilers automatically (Jones, 1980; Milos et al., 1984; ​Bodwin et al., 1982). These efforts have shown that the method is feasible, but the work has never progressed to the point where it can be used to generate useful compilers.

As stated previously, denotational semantics can be used as an aid to language design. For example, statements for which the denotational

semantic description is complex and difficult may indicate to the designer that such statements may also be difficult for language users to understand and that an alternative design may be in order.

Because of the complexity of denotational descriptions, they are of little use to language users. On the other hand, they provide an excellent way to describe a language concisely.

Although the use of denotational semantics is normally attributed to Scott and Strachey (1971), the general denotational approach to language description can be traced to the nineteenth century (Frege, 1892).

3.5.3 Axiomatic Semantics Axiomatic semantics, thus named because it is based on mathematical logic, is the most abstract approach to semantics specification discussed in this chapter. Rather than directly specifying the meaning of a program, axiomatic semantics specifies what can be proven about the program. Recall that one of the possible uses of semantic specifications is to prove the correctness of programs.

In axiomatic semantics, there is no model of the state of a machine or program or model of state changes that take place when the program is executed. The meaning of a program is based on relationships among program variables and constants, which are the same for every execution of the program.

Axiomatic semantics has two distinct applications: program verification and program semantics specification. This section focuses on program verification in its description of axiomatic semantics.

Axiomatic semantics was defined in conjunction with the development of an approach to proving the correctness of programs. Such correctness proofs, when they can be constructed, show that a program performs the computation described by its specification. In a proof, each statement of a program is both preceded and followed by a logical expression that specifies constraints on

program variables. These, rather than the entire state of an abstract machine (as with operational semantics), are used to specify the meaning of the statement. The notation used to describe constraints—indeed, the language of axiomatic semantics—is predicate calculus. Although simple Boolean expressions are often adequate to express constraints, in some cases they are not.

When axiomatic semantics is used to specify formally the meaning of a statement, the meaning is defined by the statement’s effect on assertions about the data affected by the statement.

3.5.3.1 Assertions The logical expressions used in axiomatic semantics are called predicates, or assertions. An assertion immediately preceding a program statement describes the constraints on the program variables at that point in the program. An assertion immediately following a statement describes the new constraints on those variables (and possibly others) after execution of the statement. These assertions are called the precondition and postcondition, respectively, of the statement. For two adjacent statements, the postcondition of the first serves as the precondition of the second. Developing an axiomatic description or proof of a given program requires that every statement in the program has both a precondition and a postcondition.

In the following sections, we examine assertions from the point of view that preconditions for statements are computed from given postconditions, although it is possible to consider these in the opposite sense. We assume all variables are integer type. As a simple example, consider the following assignment statement and postcondition:

sum = 2 * x + 1 {sum > 1}

Precondition and postcondition assertions are presented in braces to distinguish them from parts of program statements. One possible precondition for this statement is {x > 10}.

In axiomatic semantics, the meaning of a specific statement is defined by its

precondition and its postcondition. In effect, the two assertions specify precisely the effect of executing the statement.

In the following subsections, we focus on correctness proofs of statements and programs, which is a common use of axiomatic semantics. The more general concept of axiomatic semantics is to state precisely the meaning of statements and programs in terms of logic expressions. Program verification is one application of axiomatic descriptions of languages.

3.5.3.2 Weakest Preconditions The weakest precondition is the least restrictive precondition that will guarantee the validity of the associated postcondition. For example, in the statement and postcondition given in Section 3.5.3.1, {x > 10}, {x > 50}, and {x > 1000} are all valid preconditions. The weakest of all preconditions in this case is {x > 0}.

If the weakest precondition can be computed from the most general postcondition for each of the statement types of a language, then the processes used to compute these preconditions provide a concise description of the semantics of that language. Furthermore, correctness proofs can be constructed for programs in that language. A program proof is begun by using the characteristics of the results of the program’s execution as the postcondition of the last statement of the program. This postcondition, along with the last statement, is used to compute the weakest precondition for the last statement. This precondition is then used as the postcondition for the second last statement. This process continues until the beginning of the program is reached. At that point, the precondition of the first statement states the conditions under which the program will compute the desired results. If these conditions are implied by the input specification of the program, the program has been verified to be correct.

An inference rule is a method of inferring the truth of one assertion on the basis of the values of other assertions. The general form of an inference rule is as follows:

S1, S2, …, SnS

This rule states that if S1, S2, . . . , and Sn are true, then the truth of S can be inferred. The top part of an inference rule is called its antecedent; the bottom part is called its consequent.

An axiom is a logical statement that is assumed to be true. Therefore, an axiom is an inference rule without an antecedent.

For some program statements, the computation of a weakest precondition from the statement and a postcondition is simple and can be specified by an axiom. In most cases, however, the weakest precondition can be specified only by an inference rule.

To use axiomatic semantics with a given programming language, whether for correctness proofs or for formal semantics specifications, either an axiom or an inference rule must exist for each kind of statement in the language. In the following subsections, we present an axiom for assignment statements and inference rules for statement sequences, selection statements, and logical pretest loop statements. Note that we assume that neither arithmetic nor ​- Boolean expressions have side effects.

3.5.3.3 Assignment Statements The precondition and postcondition of an assignment statement together define its meaning. To define the meaning of an assignment statement there must be a way to compute its precondition from its postcondition.

Let x = E be a general assignment statement and Q be its postcondition. Then, its weakest precondition, P, is defined by the axiom

P = QxSE

which means that P is computed as Q with all instances of x replaced by E. For example, if we have the assignment statement and postcondition

a = b / 2 - 1 {a < 10}

the weakest precondition is computed by substituting b / 2 - 1 for a in the postcondition {a < 10}, as follows:

b / 2 - 1 < 10

b < 22

Thus, the weakest precondition for the given assignment statement and postcondition is {b < 22}. Remember that the assignment axiom is guaranteed to be correct only in the absence of side effects. An assignment statement has a side effect if it changes some variable other than its target.

The usual notation for specifying the axiomatic semantics of a given statement form is

{P} S {Q}

where P is the precondition, Q is the postcondition, and S is the statement form. In the case of the assignment statement, the notation is

{QxSE} x = E{Q}

As another example of computing a precondition for an assignment statement, consider the following:

x = 2 * y - 3 {x > 25}

The precondition is computed as follows:

2 * y - 3 > 25

y > 14

So {y > 14} is the weakest precondition for this assignment statement and postcondition.

Note that the appearance of the left side of the assignment statement in its right side does not affect the process of computing the weakest precondition. For example, for

x = x + y - 3 {x > 10}

the weakest precondition is

x + y - 3 > 10

y > 13 - x

Recall that axiomatic semantics was developed to prove the correctness of programs. In light of that, it is natural at this point to wonder how the axiom for assignment statements can be used to prove anything. Here is how: A given assignment statement with both a precondition and a postcondition can be considered a logical statement, or theorem. If the assignment axiom, when applied to the postcondition and the assignment statement, produces the given precondition, the theorem is proved. For example, consider the following logical statement:

{x > 3} x = x - 3 {x > 0}

Using the assignment axiom on the statement and its postcondition produces {x > 3}, which is the given precondition. Therefore, we have proven the example logical statement.

Next, consider the following logical statement:

{x > 5} x = x - 3 {x > 0}

In this case, the given precondition, {x > 5}, is not the same as the assertion produced by the axiom. However, it is obvious that {x > 5} implies {x > 3}. To use this in a proof, an inference rule named the rule of consequence is needed. The form of the rule of consequence is

{P} S {Q}, P′ => P, Q => Q′{P′} S{Q′}

The => symbol means “implies,” and S can be any program statement. The rule can be stated as follows: If the logical statement {P} S{Q} is true, the assertion P′ implies the assertion P, and the assertion Q implies the assertion Q′, then it can be inferred that {P′} S{Q′}. In other words, the rule of consequence says that a postcondition can always be weakened and a precondition can always be strengthened. This is quite useful in program proofs. For example, it allows the completion of the proof of the last logical statement example above. If we let P be {x > 3}, Q and Q′ be {x > 0}, and P

′ be {x > 5}, we have

{x > 3}x = x − 3{x > 0},(x > 5)=>{x > 3},(x > 0)=>(x > 0) {x > 5}x = x − 3{x > 0}

The first term of the antecedent ({x > 3} x = x - 3 {x > 0}) was proven with the assignment axiom. The second and third terms are obvious. Therefore, by the rule of consequence, the consequent is true.

3.5.3.4 Sequences The weakest precondition for a sequence of statements cannot be described by an axiom, because the precondition depends on the particular kinds of statements in the sequence. In this case, the precondition can only be described with an inference rule. Let S1 and S2 be adjacent program statements. If S1 and S2 have the following pre- and postconditions

{P1} S1 {P2} {P2} S2 {P3}

the inference rule for such a two-statement sequence is

{P1} S1 {P2}, {P2} S2 {P3}{P1} S1, S2 {P3}

So, for our example, {P1} S1; S2 {P3} describes the axiomatic semantics of the sequence S1; S2. The inference rule states that to get the sequence precondition, the precondition of the second statement is computed. This new assertion is then used as the postcondition of the first statement, which can then be used to compute the precondition of the first statement, which is also the precondition of the whole sequence. If S1 and S2 are the assignment statements

x1=E1

and

x2=E2

then we have

{P3x2→E2} x2=E2 {P3} {(P3x2→E2)x1→E1} x1=E1 {P3x2→E2}

Therefore, the weakest precondition for the sequence x1 = E1; x2 = E2 with postcondition P3 is {(P3x2→E2)x1→E1}.

For example, consider the following sequence and postcondition:

y = 3 * x + 1;

x = y + 3;

{x < 10}

The precondition for the second assignment statement is

y < 7

which is used as the postcondition for the first statement. The precondition for the first assignment statement can now be computed:

3 * x + 1 < 7

x < 2

So, {x < 2} is the precondition of both the first statement and the two-​- statement sequence.

3.5.3.5 Selection We next consider the inference rule for selection statements, the general form of which is

if B then S1 else S2

We consider only selections that include else clauses. The inference rule is

{B and P} S1 {Q}, {(not B) and P} S2 {Q} {P} if B bb{then} S1 bb{else} S2 {Q}}

This rule specifies that selection statements must be proven both when the

Boolean control expression is true and when it is false. The first logical statement above the line represents the then clause; the second represents the else clause. According to the inference rule, we need a precondition P that can be used in the precondition of both the then and else clauses.

Consider the following example of the computation of the precondition using the selection inference rule. The example selection statement is

if x > 0 then

y = y - 1

else

y = y + 1

Suppose the postcondition, Q, for this selection statement is {y > 0}. We can use the axiom for assignment on the then clause

y = y - 1 {y > 0}

This produces {y - 1 > 0} or {y > 1}. It can be used as the P part of the precondition for the then clause. Now we apply the same axiom to the else clause

y = y + 1 {y > 0}

which produces the precondition {y + 1 > 0} or {y > -1}. Because {y > 1} => {y > -1}, the rule of consequence allows us to use {y > 1} for the precondition of the whole selection statement.

3.5.3.6 Logical Pretest Loops Another essential construct of imperative programming languages is the logical pretest, or while loop. Computing the weakest precondition for a while loop is inherently more difficult than for a sequence, because the number of iterations cannot always be predetermined. In a case where the number of iterations is known, the loop can be unrolled and treated as a sequence.

The problem of computing the weakest precondition for loops is similar to

the problem of proving a theorem about all positive integers. In the latter case, induction is normally used, and the same inductive method can be used for some loops. The principal step in induction is finding an inductive hypothesis. The corresponding step in the axiomatic semantics of a while loop is finding an assertion called a loop invariant, which is crucial to finding the weakest precondition.

The inference rule for computing the precondition for a while loop is as follows:

{I and B} S {I}{I} while B do S end {I and (not B)}

In this rule, I is the loop invariant. This seems simple, but it is not. The complexity lies in finding an appropriate loop invariant.

The axiomatic description of a while loop is written as

{P} while B do S end {Q}

The loop invariant must satisfy a number of requirements to be useful. First, the weakest precondition for the while loop must guarantee the truth of the loop invariant. In turn, the loop invariant must guarantee the truth of the postcondition upon loop termination. These constraints move us from the inference rule to the axiomatic description. During execution of the loop, the truth of the loop invariant must be unaffected by the evaluation of the loop- controlling Boolean expression and the loop body statements. Hence, the name invariant.

Another complicating factor for while loops is the question of loop termination. A loop that does not terminate cannot be correct, and in fact computes nothing. If Q is the postcondition that holds immediately after loop exit, then a precondition P for the loop is one that guarantees Q at loop exit and also guarantees that the loop terminates.

The complete axiomatic description of a while construct requires all of the following to be true, in which I is the loop invariant:

P => I

{I and B} S { I }

(I and (not B)) => Q

the loop terminates

If a loop computes a sequence of numeric values, it may be possible to find a loop invariant using an approach that is used for determining the inductive hypothesis when mathematical induction is used to prove a statement about a mathematical sequence. The relationship between the number of iterations and the precondition for the loop body is computed for a few cases, with the hope that a pattern emerges that will apply to the general case. It is helpful to treat the process of producing a weakest precondition as a function, wp. In general

wp(statement, postcondition)=precondition

A wp function is often called a predicate transformer, because it takes a predicate, or assertion, as a parameter and returns another predicate.

To find I, the loop postcondition Q is used to compute preconditions for several different numbers of iterations of the loop body, starting with none. If the loop body contains a single assignment statement, the axiom for assignment statements can be used to compute these cases. Consider the example loop:

while y <> x do y = y + 1 end {y = x}

Remember that the equal sign is being used for two different purposes here. In assertions, it means mathematical equality; outside assertions, it means the assignment operator.

For zero iterations, the weakest precondition is, obviously,

{y = x}

For one iteration, it is

wp(y = y + 1, {y = x}) = {y + 1 = x}, or {y = x - 1}

For two iterations, it is

wp(y = y + 1, {y = x - 1})={y + 1 = x - 1}, or {y = x - 2}

For three iterations, it is

wp(y = y + 1, {y = x - 2})={y + 1 = x - 2}, or {y = x – 3}

It is now obvious that {y < x} will suffice for cases of one or more iterations. Combining this with {y = x} for the zero iterations case, we get {y <= x}, which can be used for the loop invariant. A precondition for the while statement can be determined from the loop invariant. In fact, I can be used as the precondition, P.

We must ensure that our choice satisfies the four criteria for I for our example loop. First, because P = I, P => I. The second requirement is that the following must be true:

{ I and B } S { I }

In our example, we have

{y <= x and y <> x} y = y + 1 {y <= x}

Applying the assignment axiom to

y = y + 1 {y <= x}

we get {y + 1 <= x}, which is equivalent to {y < x}, which is implied by {y <= x and y <> x}. So, the earlier statement is proven.

Next, we must have

{I and (not B)} => Q

In our example, we have

{(y <= x) and not (y <> x)} => {y = x}

{(y <= x) and (y = x)} => {y = x}

{y = x} => {y = x}

So, this is obviously true. Next, loop termination must be considered. In this example, the question is whether the loop

{y <= x} while y <> x do y = y + 1 end {y = x}

terminates. Recalling that x and y are assumed to be integer variables, it is easy to see that this loop does terminate. The precondition guarantees that y initially is not larger than x. The loop body increments y with each iteration, until y is equal to x. No matter how much smaller y is than x initially, it will eventually become equal to x. So the loop will terminate. Because our choice of I satisfies all four criteria, it is a satisfactory loop invariant and loop precondition.

The previous process used to compute the invariant for a loop does not always produce an assertion that is the weakest precondition (although it does in the example).

As another example of finding a loop invariant using the approach used in mathematical induction, consider the following loop statement:

while s > 1 do s = s / 2 end {s = 1}

As before, we use the assignment axiom to try to find a loop invariant and a precondition for the loop. For zero iterations, the weakest precondition is {s = 1}.

For one iteration, it is

wp(s = s / 2, {s = 1}) = {s / 2 = 1}, or {s = 2}

For two iterations, it is

wp(s = s / 2, {s = 2}) = {s / 2 = 2}, or {s = 4}

For three iterations, it is

wp(s = s / 2, {s = 4}) = {s / 2 = 4}, or {s = 8}

From these cases, we can see clearly that the invariant is

{s is a nonnegative power of 2}

Once again, the computed I can serve as P, and I passes the four requirements. Unlike our earlier example of finding a loop precondition, this one clearly is not a weakest precondition. Consider using the precondition {s > 1}. The logical statement

{s > 1} while s > 1 do s = s / 2 end {s = 1}

can easily be proven, and this precondition is significantly broader than the one computed earlier. The loop and precondition are satisfied for any positive value for s, not just powers of 2, as the process indicates. Because of the rule of consequence, using a precondition that is stronger than the weakest precondition does not invalidate a proof.

Finding loop invariants is not always easy. It is helpful to understand the nature of these invariants. First, a loop invariant is a weakened version of the loop postcondition and also a precondition for the loop. So, I must be weak enough to be satisfied prior to the beginning of loop execution, but when combined with the loop exit condition, it must be strong enough to force the truth of the postcondition.

Because of the difficulty of proving loop termination, that requirement is often ignored. If loop termination can be shown, the axiomatic description of the loop is called total correctness. If the other conditions can be met but termination is not guaranteed, it is called partial correctness.

In more complex loops, finding a suitable loop invariant, even for partial correctness, requires a good deal of ingenuity. Because computing the precondition for a while loop depends on finding a loop invariant, proving the correctness of programs with while loops using axiomatic semantics can be difficult.

3.5.3.7 Program Proofs

This section provides validations for two simple programs. The first example of a correctness proof is for a very short program, consisting of a sequence of three assignment statements that interchange the values of two variables.

{x = A AND y = B}

t = x;

x = y;

y = t;

{x = B AND y = A}

Because the program consists entirely of assignment statements in a sequence, the assignment axiom and the inference rule for sequences can be used to prove its correctness. The first step is to use the assignment axiom on the last statement and the postcondition for the whole program. This yields the precondition

{x = B AND t = A}

Next, we use this new precondition as a postcondition on the middle statement and compute its precondition, which is

{y = B AND t = A}

Next, we use this new assertion as the postcondition on the first statement and apply the assignment axiom, which yields

{y = B AND x = A}

which is the same as the precondition on the program, except for the order of operands on the AND operator. Because AND is a symmetric operator, our proof is complete.

The following example is a proof of correctness of a pseudocode program that computes the factorial function.

{n >= 0}

count = n;

fact = 1;

while count <> 0 do

fact = fact * count;

count = count - 1;

end

{fact = n!}

The method described earlier for finding the loop invariant does not work for the loop in this example. Some ingenuity is required here, which can be aided by a brief study of the code. The loop computes the factorial function in order of the last multiplication first; that is, (n - 1) * n is done first, assuming n is greater than 1. So, part of the invariant can be the following:

fact = (count + 1) * (count + 2) * . . . * (n - 1) * n

But we also must ensure that count is always nonnegative, which we can do by adding that to the assertion above, to get the following:

I = (fact = (count + 1) * . . . * n) AND (count >= 0)

Next, we must confirm that this I meets the requirements for invariants. Once again we let I also be used for P, so P clearly implies I. The next question is

{ I and B } S { I }

I and B is the following:

((fact = (count + 1) * . . . * n) AND (count >= 0)) AND

(count <> 0)

This reduces to

(fact = (count + 1) * . . . * n) AND (count > 0)

In our case, we must compute the precondition of the body of the loop, using the invariant for the postcondition. For

{P} count = count - 1 {I}

we compute P to be

{(fact = count * (count + 1) * . . . * n) AND

(count >= 1)}

Using this as the postcondition for the first assignment in the loop body,

{P} fact = fact * count {(fact = count * (count + 1)

                        * . . . * n) AND (count >= 1)}

In this case, P is

{(fact = (count + 1) * . . . * n) AND (count >= 1)}

It is clear that I and B implies this P, so by the rule of consequence,

{I AND B} S {I}

is true. Finally, the last test of I is

I AND (NOT B) => Q

For our example, this is

((fact = (count + 1) * . . . * n) AND (count >= 0)) AND

(count = 0)) => fact = n!

This is clearly true, for when count = 0, the first part is precisely the definition of factorial. So, our choice of I meets the requirements for a loop invariant. Now we can use our P (which is the same as I) from the while as the postcondition on the second assignment of the program

{P} fact = 1 {(fact = (count + 1) * . . . * n) AND

(count >= 0)}

which yields for P

(1 = (count + 1) * . . . * n) AND (count >= 0))

Using this as the postcondition for the first assignment in the code

{P} count = n {(1 = (count + 1) * . . . * n) AND

(count >= 0))}

produces for P

{(n + 1) * . . . * n = 1) AND (n >= 0)}

The left operand of the AND operator is true (because 1 = 1) and the right operand is exactly the precondition of the whole code segment, {n >= 0}. Therefore, the program has been proven to be correct.

3.5.3.8 Evaluation As stated previously, to define the semantics of a complete programming language using the axiomatic method, there must be an axiom or an inference rule for each statement type in the language. Defining axioms or inference rules for some of the statements of programming languages has proven to be a difficult task. An obvious solution to this problem is to design the language with the axiomatic method in mind, so that only statements for which axioms or inference rules can be written are included. Unfortunately, such a language would necessarily leave out some useful and powerful statements.

Axiomatic semantics is a powerful tool for research into program correctness proofs, and it provides an excellent framework in which to reason about programs, both during their construction and later. Its usefulness in describing the meaning of programming languages to language users and compiler writers is, however, highly limited.

SUMMARY Backus-Naur Form and context-free grammars are equivalent metalanguages that are well suited for the task of describing the syntax of programming languages. Not only are they concise descriptive tools, but also the parse trees that can be associated with their generative actions give graphical evidence of the underlying syntactic structures. Furthermore, they are naturally related to recognition devices for the languages they generate, which leads to the relatively easy construction of syntax analyzers for compilers for these languages.

An attribute grammar is a descriptive formalism that can describe both the syntax and static semantics of a language. Attribute grammars are extensions to context-free grammars. An attribute grammar consists of a grammar, a set of attributes, a set of attribute computation functions, and a set of predicates that describe static semantics rules.

This chapter provides brief introductions to three methods of semantic description: operational, denotational, and axiomatic. Operational semantics is a method of describing the meaning of language constructs in terms of their effects on an ideal machine. In denotational semantics, mathematical objects are used to represent the meanings of language constructs. Language entities are converted to these mathematical objects with recursive functions. Axiomatic semantics, which is based on formal logic, was devised as a tool for proving the correctness of programs.

BIBLIOGRAPHIC NOTES Syntax description using context-free grammars and BNF are thoroughly ​- discussed in Cleaveland and Uzgalis (1976).

Research in axiomatic semantics was begun by Floyd (1967) and further developed by Hoare (1969). The semantics of a large part of Pascal was described by Hoare and Wirth (1973) using this method. The parts they did not complete involved functional side effects and goto statements. These were found to be the most difficult to describe.

The technique of using preconditions and postconditions during the development of programs is described (and advocated) by Dijkstra (1976) and also discussed in detail in Gries (1981).

Good introductions to denotational semantics can be found in Gordon (1979) and Stoy (1977). Introductions to all of the semantics description methods discussed in this chapter can be found in Marcotty et al. (1976). Another good reference for much of the chapter material is Pagan (1981). The form of the denotational semantic functions in this chapter is similar to that found in Meyer (1990).

REVIEW QUESTIONS 1. Define syntax and semantics.

2. Who are language descriptions for?

3. Describe the operation of a general language generator.

4. Describe the operation of a general language recognizer.

5. What is the difference between a sentence and a sentential form?

6. Define a left-recursive grammar rule.

7. What three extensions are common to most EBNFs?

8. Distinguish between static and dynamic semantics.

9. What purpose do predicates serve in an attribute grammar?

10. What is the difference between a synthesized and an inherited attribute?

11. How is the order of evaluation of attributes determined for the trees of a given attribute grammar?

12. What is the primary use of attribute grammars?

13. Explain the primary uses of a methodology and notation for describing the semantics of programming languages.

14. Why can machine languages not be used to define statements in operational semantics?

15. Describe the two levels of uses of operational semantics.

16. In denotational semantics, what are the syntactic and semantic domains?

17. What is stored in the state of a program for denotational semantics?

18. Which semantics approach is most widely known?

19. What two things must be defined for each language entity in order to construct a denotational description of the language?

20. Which part of an inference rule is the antecedent?

21. What is a predicate transformer function?

22. What does partial correctness mean for a loop construct?

23. On what branch of mathematics is axiomatic semantics based?

24. On what branch of mathematics is denotational semantics based?

25. What is the problem with using a software pure interpreter for operational semantics?

26. Explain what the preconditions and postconditions of a given statement mean in axiomatic semantics.

27. Describe the approach of using axiomatic semantics to prove the correctness of a given program.

28. Describe the basic concept of denotational semantics.

29. In what fundamental way do operational semantics and denotational semantics differ?

PROBLEM SET 1. The two mathematical models of language description are generation

and recognition. Describe how each can define the syntax of a programming language.

2. Write EBNF descriptions for the following:

1. A Java class definition header statement

2. A Java method call statement

3. A C switch statement

4. A C union definition

5. C float literals

3. Rewrite the BNF of Example 3.4 to give + precedence over * and force + to be right associative.

4. Rewrite the BNF of Example 3.4 to add the ++ and -- unary operators of Java.

5. Write a BNF description of the Boolean expressions of Java, including the three operators &&, ||, and ! and the relational expressions.

6. Using the grammar in Example 3.2, show a parse tree and a leftmost derivation for each of the following statements:

1. A = A * (B + (C * A))

2. B = C * (A * C + B)

3. A = A * (B + (C))

7. Using the grammar in Example 3.4, show a parse tree and a leftmost derivation for each of the following statements:

1. A = (A + B) * C

2. A = B + C + A

3. A = A * (B + C)

4. A = B * (C * (A + B))

8. Prove that the following grammar is ambiguous:

<S> → <A>

<A> → <A> + <A> | <id>

<id> → a | b | c

9. Modify the grammar of Example 3.4 to add a unary minus operator that has higher precedence than either + or *.

10. Describe, in English, the language defined by the following grammar:

<S> → <A> <B> <C>

<A> → a <A> | a

<B> → b <B> | b

<C> → c <C> | c

11. Consider the following grammar:

<S> → <A> a <B> b

<A> → <A> b | b

<B> → a <B> | a

Which of the following sentences are in the language generated by this grammar?

1. baab

2. bbbab

3. bbaaaaaS

4. bbaab

12. Consider the following grammar:

<S> → a <S> c <B> | <A> | b

<A> → c <A> | c

<B> → d | <A>

Which of the following sentences are in the language generated by this grammar?

1. abcd

2. acccbd

3. acccbcc

4. acd

5. accc

13. Write a grammar for the language consisting of strings that have n copies of the letter a followed by the same number of copies of the letter b, where n > 0. For example, the strings ab, aaaabbbb, and aaaaaaaabbbbbbbb are in the language but a, abb, ba, and aaabb are not.

14. Draw parse trees for the sentences aabb and aaaabbbb, as derived from the grammar of Problem 13.

15. Convert the BNF of Example 3.1 to EBNF.

16. Convert the BNF of Example 3.3 to EBNF.

17. Convert the following EBNF to BNF:

S→A{bA}

A→a[b]A

18. What is the difference between an intrinsic attribute and a nonintrinsic synthesized attribute?

19. Write an attribute grammar whose BNF basis is that of Example 3.6 in Section 3.4.5 but whose language rules are as follows: Data types cannot be mixed in expressions, but assignment statements need not have the same types on both sides of the assignment operator.

20. Write an attribute grammar whose base BNF is that of Example 3.2 and whose type rules are the same as for the assignment statement example of Section 3.4.5.

21. Using the virtual machine instructions given in Section 3.5.1.1, give an operational semantic definition of the following:

1. Java do-while

2. Ada for

3. C++ if-then-else

4. C for

5. C switch

22. Write a denotational semantics mapping function for the following statements:

1. Ada for

2. Java do-while

3. Java Boolean expressions

4. Java for

5. C switch

23. Compute the weakest precondition for each of the following assignment statements and postconditions:

1. a = 2 * (b - 1) - 1 {a > 0}

2. b = (c + 10) / 3 {b > 6}

3. a = a + 2 * b - 1 {a > 1}

4. x = 2 * y + x - 1 {x > 11}

24. Compute the weakest precondition for each of the following sequences of assignment statements and their postconditions:

1. a = 2 * b + 1;

b = a - 3

{b < 0}

2. a = 3 * (2 * b + a);

b = 2 * a - 1

{b > 5}

25. Compute the weakest precondition for each of the following selection constructs and their postconditions:

1. if (a == b)

 b = 2 * a + 1

else

 b = 2 * a;

{b > 1}

2. if (x < y)

 x = x + 1

else

 x = 3 * x

{x < 0}

3. if (x > y)

  y = 2 * x + 1

else

  y = 3 * x - 1;

{y > 3}

26. Explain the four criteria for proving the correctness of a logical pretest loop construct of the form while B do S end.

27. Prove that (n+1) *…* n=1.

28. Prove the following program is correct:

{n > 0}

count = n;

sum = 0;

while count <> 0 do

sum = sum + count;

count = count - 1;

end

{sum = 1 + 2 + . . . + n}

4 Lexical and Syntax Analysis 1. 4.1 Introduction

2. 4.2 Lexical Analysis

3. 4.3 The Parsing Problem

4. 4.4 Recursive-Descent Parsing

5. 4.5 Bottom-Up Parsing

This chapter begins with an introduction to lexical analysis, along with a ​- simple example. Next, the general parsing problem is discussed, including the two ​primary approaches to parsing, and the complexity of parsing. Then, we ​- introduce the recursive-descent implementation technique for top-down parsers, including examples of parts of a recursive-descent parser and a trace of a parse using one. The last section discusses bottom-up parsing and the LR parsing algorithm. This section includes an example of a small LR parsing table and the parse of a string using the LR parsing process.

4.1 Introduction A serious investigation of compiler design requires at least a semester of intensive study, including the design and implementation of a compiler for a small but realistic programming language. The first part of such a course is devoted to lexical and syntax analyses. The syntax analyzer is the heart of a compiler, because several other important components, including the semantic analyzer and the intermediate code generator, are driven by the actions of the syntax analyzer.

Some readers may wonder why a chapter on any part of a compiler would be included in a book on programming languages. There are at least two reasons to include a discussion of lexical and syntax analyses in this text: First, syntax analyzers are based directly on the grammars discussed in Chapter 3, so it is natural to discuss them as an application of grammars. Second, lexical and syntax analyzers are needed in numerous situations outside compiler design. Many applications, among them program listing formatters, programs that compute the complexity of programs, and programs that must analyze and react to the contents of a configuration file, all need to do lexical and syntax analyses. Therefore, lexical and syntax analyses are important topics for software developers, even if they never need to write a compiler. Furthermore, some computer science programs no longer require students to take a compiler design course, which leaves students with no instruction in lexical or syntax analysis. In those cases, this chapter can be covered in the programming language course. In degree programs that require a compiler design course, this chapter can be skipped.

Three different approaches to implementing programming languages are introduced in Chapter 1: compilation, pure interpretation, and hybrid implementation. The compilation approach uses a program called a compiler, which translates programs written in a high-level programming language into machine code. Compilation is typically used to implement programming languages that are used for large applications, often written in languages such as C++ and COBOL. Pure interpretation systems perform no translation; rather, programs are interpreted in their original form by a software

interpreter. Pure interpretation is usually used for smaller systems in which execution efficiency is not critical, such as scripts embedded in HTML documents, written in ​languages such as JavaScript. Hybrid implementation systems translate programs written in high-level languages into intermediate forms, which are interpreted. These systems are now more widely used than ever, thanks in large part to the popularity of scripting languages. Traditionally, hybrid systems have resulted in much slower program execution than compiler systems. However, in recent years the use of Just-in- Time (JIT) compilers has become widespread, particularly for Java programs and programs written for the Microsoft .NET system. A JIT compiler, which translates intermediate code to machine code, is used on methods at the time they are first called. In effect, a JIT compiler transforms a hybrid system to a delayed compiler system.

All three of the implementation approaches just discussed use both lexical and syntax analyzers.

Syntax analyzers, or parsers, are nearly always based on a formal description of the syntax of programs. The most commonly used syntax-description formalism is context-free grammars, or BNF, which is introduced in Chapter 3. Using BNF, as opposed to using some informal syntax description, has at least three compelling advantages. First, BNF descriptions of the syntax of programs are clear and concise, both for humans and for software systems that use them. Second, the BNF description can be used as the direct basis for the syntax analyzer. Third, implementations based on BNF are relatively easy to maintain because of their modularity.

Nearly all compilers separate the task of analyzing syntax into two distinct parts, lexical analysis and syntax analysis, although this terminology is confusing. The lexical analyzer deals with small-scale language constructs, such as names and numeric literals. The syntax analyzer deals with the large- scale constructs, such as expressions, statements, and program units. Section 4.2 introduces lexical analyzers. Sections 4.3, 4.4, and 4.5 discuss syntax analyzers.

There are three reasons why lexical analysis is separated from syntax analysis:

1. Simplicity—Techniques for lexical analysis are less complex than those required for syntax analysis, so the lexical-analysis process can be simpler if it is separate. Also, removing the low-level details of lexical analysis from the syntax analyzer makes the syntax analyzer both smaller and less complex.

2. Efficiency—Although it pays to optimize the lexical analyzer, because lexical analysis requires a significant portion of total compilation time, it is not fruitful to optimize the syntax analyzer. Separation facilitates this selective optimization.

3. Portability—Because the lexical analyzer reads input program files and often includes buffering of that input, it is somewhat platform dependent. However, the syntax analyzer can be platform independent. It is always good to isolate machine-dependent parts of any software system.

4.2 Lexical Analysis A lexical analyzer is essentially a pattern matcher. A pattern matcher attempts to find a substring of a given string of characters that matches a given character pattern. Pattern matching is a traditional part of computing. One of the earliest uses of pattern matching was with text editors, such as the ed line editor, which was introduced in an early version of UNIX. Since then, pattern matching has found its way into some programming languages—for example, Perl and JavaScript. It is also available through the standard class libraries of Java, C++, and C#.

A lexical analyzer serves as the front end of a syntax analyzer. Technically, lexical analysis is a part of syntax analysis. A lexical analyzer performs syntax analysis at the lowest level of program structure. An input program appears to a compiler as a single string of characters. The lexical analyzer collects characters into logical groupings and assigns internal codes to the groupings according to their structure. In Chapter 3, these logical groupings are named lexemes, and the internal codes for categories of these groupings are named tokens. Lexemes are recognized by matching the input character string against character string patterns. Although tokens are usually represented as integer values, for the sake of readability of lexical and syntax analyzers, they are often referenced through named constants.

Consider the following example of an assignment statement:

result = oldsum - value / 100;

Following are the tokens and lexemes of this statement:

Lexical analyzers extract lexemes from a given input string and produce the corresponding tokens. In the early days of compilers, lexical analyzers often processed an entire source program file and produced a file of tokens and lexemes. Now, however, most lexical analyzers are subprograms that locate the next lexeme in the input, determine its associated token code, and return them to the caller, which is the syntax analyzer. So, each call to the lexical analyzer returns a single lexeme and its token. The only view of the input program seen by the syntax analyzer is the output of the lexical analyzer, one token at a time.

The lexical-analysis process includes skipping comments and white space outside lexemes, as they are not relevant to the meaning of the program. Also, the lexical analyzer inserts lexemes for user-defined names into the symbol table, which is used by later phases of the compiler. Finally, lexical analyzers detect syntactic errors in tokens, such as ill-formed floating-point literals, and report such errors to the user.

There are three approaches for building a lexical analyzer:

1. Write a formal description of the token patterns of the language using a

descriptive language related to regular expressions.1 These descriptions are used as input to a software tool that automatically generates a lexical analyzer. There are many such tools available for this. The oldest of these, named lex, is commonly included as part of UNIX systems.

1. These regular expressions are the basis for the pattern-matching facilities now part of many programming languages, either directly or through a class library.

2. Design a state transition diagram that describes the token patterns of the language and write a program that implements the diagram.

3. Design a state transition diagram that describes the token patterns of the language and hand-construct a table-driven implementation of the state diagram.

A state transition diagram, or just state diagram, is a directed graph. The nodes of a state diagram are labeled with state names. The arcs are labeled with the input characters that cause the transitions among the states. An arc may also include actions the lexical analyzer must perform when the transition is taken.

State diagrams of the form used for lexical analyzers are representations of a class of mathematical machines called finite automata. Finite automata can be designed to recognize members of a class of languages called regular languages. Regular grammars are generative devices for regular languages. The tokens of a programming language are a regular language, and a lexical analyzer is a finite automaton.

We now illustrate lexical-analyzer construction with a state diagram and the code that implements it. The state diagram could simply include states and transitions for each and every token pattern. However, that approach results in a very large and complex diagram, because every node in the state diagram would need a transition for every character in the character set of the language being analyzed. We therefore consider ways to simplify it.

Suppose we need a lexical analyzer that recognizes only arithmetic expressions, including variable names and integer literals as operands.

Assume that the variable names consist of strings of uppercase letters, lowercase letters, and digits but must begin with a letter. Names have no length limitation. The first thing to observe is that there are 52 different characters (any uppercase or lowercase letter) that can begin a name, which would require 52 transitions from the transition diagram’s initial state. However, a lexical analyzer is interested only in determining that it is a name and is not concerned with which specific name it happens to be. Therefore, we define a character class named LETTER for all 52 letters and use a single transition on the first letter of any name.

Another opportunity for simplifying the transition diagram is with the integer literal tokens. There are 10 different characters that could begin an integer literal lexeme. This would require 10 transitions from the start state of the state diagram. Because specific digits are not a concern of the lexical analyzer, we can build a much more compact state diagram if we define a character class named DIGIT for digits and use a single transition on any character in this character class to a state that collects integer literals.

Because our names can include digits, the transition from the node following the first character of a name can use a single transition on LETTER or DIGIT to continue collecting the characters of a name.

Next, we define some utility subprograms for the common tasks inside the lexical analyzer. First, we need a subprogram, which we can name getChar, that has several duties. When called, getChar gets the next character of input from the input program and puts it in the global variable nextChar. getChar also must determine the character class of the input character and put it in the global variable charClass. The lexeme being built by the lexical analyzer, which could be implemented as a character string or an array, will be named lexeme.

We implement the process of putting the character in nextChar into the string array lexeme in a subprogram named addChar. This subprogram must be explicitly called because programs include some characters that need not be put in lexeme, for example the white-space characters between lexemes. In a more realistic lexical analyzer, comments also would not be placed in lexeme.

When the lexical analyzer is called, it is convenient if the next character of input is the first character of the next lexeme. Because of this, a function named getNonBlank is used to skip white space every time the analyzer is called.

Finally, a subprogram named lookup is needed to compute the token code for the single-character tokens. In our example, these are parentheses and the arithmetic operators. Token codes are numbers arbitrarily assigned to tokens by the compiler writer.

The state diagram in Figure 4.1 describes the patterns for our tokens. It includes the actions required on each transition of the state diagram.

Figure 4.1 A state diagram to

recognize names, ​parentheses, and ​arithmetic operators

Figure 4.1 Full Alternative Text

The following is a C implementation of a lexical analyzer specified in the state diagram of Figure 4.1, including a main driver function for testing purposes:

/* front.c - a lexical analyzer system for simple

arithmetic expressions */

#include <stdio.h>

#include <ctype.h>

/* Global declarations */

/* Variables */

int charClass;

char lexeme [100];

char nextChar;

int lexLen;

int token;

int nextToken;

FILE *in_fp, *fopen();

/* Function declarations */

void addChar();

void getChar();

void getNonBlank();

int lex();

/* Character classes */

#define LETTER 0

#define DIGIT 1

#define UNKNOWN 99

/* Token codes */

#define INT_LIT 10

#define IDENT 11

#define ASSIGN_OP 20

#define ADD_OP 21

#define SUB_OP 22

#define MULT_OP 23

#define DIV_OP 24

#define LEFT_PAREN 25

#define RIGHT_PAREN 26

/******************************************************/

/* main driver */

main() {

/* Open the input data file and process its contents */

if ((in_fp = fopen("front.in", "r")) == NULL)

printf("ERROR - cannot open front.in \n");

else {

getChar();

do {

lex();

} while (nextToken! = EOF);

}

}

/*****************************************************/

/* lookup - a function to lookup operators and parentheses

and return the token */

int lookup(char ch) {

switch (ch) {

case '(':

addChar();

nextToken = LEFT_PAREN;

break;

case ')':

addChar();

nextToken = RIGHT_PAREN;

break;

case '+':

addChar();

nextToken = ADD_OP;

break;

case '-':

addChar();

nextToken = SUB_OP;

break;

case '*':

addChar();

nextToken = MULT_OP;

break;

case '/':

addChar();

nextToken = DIV_OP;

break;

default:

addChar();

nextToken = EOF;

break;

}

return nextToken;

}

/*****************************************************/

/* addChar - a function to add nextChar to lexeme */

void addChar() {

if (lexLen <= 98) {

lexeme[lexLen++] = nextChar;

lexeme[lexLen] = 0;

}

else

printf("Error - lexeme is too long \n");

}

/*****************************************************/

/* getChar - a function to get the next character of

input and determine its character class */

void getChar() {

if ((nextChar = getc(in_fp)) = EOF) {

if (isalpha(nextChar))

charClass = LETTER;

else if (isdigit(nextChar))

charClass = DIGIT;

else charClass = UNKNOWN;

}

else

charClass = EOF;

}

/*****************************************************/

/* getNonBlank - a function to call getChar until it

returns a non-whitespace character */

void getNonBlank() {

while (isspace(nextChar))

getChar();

}

/

*****************************************************/

/* lex - a simple lexical analyzer for arithmetic

expressions */

int lex() {

lexLen = 0;

getNonBlank();

switch (charClass) {

/* Parse identifiers */

case LETTER:

addChar();

getChar();

while (charClass == LETTER || charClass == DIGIT) {

addChar();

getChar();

}

nextToken = IDENT;

break;

/* Parse integer literals */

case DIGIT:

addChar();

getChar();

while (charClass == DIGIT) {

addChar();

getChar();

}

nextToken = INT_LIT;

break;

/* Parentheses and operators */

case UNKNOWN:

lookup(nextChar);

getChar();

break;

/* EOF */

case EOF:

nextToken = EOF;

lexeme[0] = 'E';

lexeme[1] = 'O';

lexeme[2] = 'F';

lexeme[3] = 0;

break;

} /* End of switch */

printf("Next token is: %d, Next lexeme is %s\n",

nextToken, lexeme);

return nextToken;

} /* End of function lex */

This code illustrates the relative simplicity of lexical analyzers. Of course, we have left out input buffering, as well as some other important details. ​- Furthermore, we have dealt with a very small and simple input language.

Consider the following expression:

(sum + 47) / total

Following is the output of the lexical analyzer of front.c when used on this expression:

Next token is: 25 Next lexeme is (

Next token is: 11 Next lexeme is sum

Next token is: 21 Next lexeme is +

Next token is: 10 Next lexeme is 47

Next token is: 26 Next lexeme is )

Next token is: 24 Next lexeme is /

Next token is: 11 Next lexeme is total

Next token is: -1 Next lexeme is EOF

Names and reserved words in programs have similar patterns. Although it is possible to build a state diagram to recognize every specific reserved word of a programming language, that would result in a prohibitively large state diagram. It is much simpler and faster to have the lexical analyzer recognize names and reserved words with the same pattern and use a lookup in a table of reserved words to determine which names are reserved words. Using this approach considers reserved words to be exceptions in the names token category.

A lexical analyzer often is responsible for the initial construction of the symbol table, which acts as a database of names for the compiler. The entries in the symbol table store information about user-defined names, as well as the attributes of the names. For example, if the name is that of a variable, the variable’s type is one of its attributes that will be stored in the symbol table. Names are usually placed in the symbol table by the lexical analyzer. The attributes of a name are usually put in the symbol table by some part of the compiler that is subsequent to the actions of the lexical analyzer.

4.3 The Parsing Problem The part of the process of analyzing syntax that is referred to as syntax analysis is often called parsing. We will use these two interchangeably.

This section discusses the general parsing problem and introduces the two main categories of parsing algorithms, top-down and bottom-up, as well as the complexity of the parsing process.

4.3.1 Introduction to Parsing Parsers for programming languages construct parse trees for given programs. In some cases, the parse tree is only implicitly constructed, meaning that perhaps only a traversal of the tree is generated. But in all cases, the information required to build the parse tree is created during the parse. Both parse trees and derivations include all of the syntactic information needed by a language processor.

There are two distinct goals of syntax analysis: First, the syntax analyzer must check the input program to determine whether it is syntactically correct. When an error is found, the analyzer must produce a diagnostic message and recover. In this case, recovery means it must get back to a normal state and continue its analysis of the input program. This step is required so that the compiler finds as many errors as possible during a single analysis of the input program. If it is not done well, error recovery may create more errors, or at least more error messages. The second goal of syntax analysis is to produce a complete parse tree, or at least trace the structure of the complete parse tree, for syntactically correct input. The parse tree (or its trace) is used as the basis for translation.

Parsers are categorized according to the direction in which they build parse trees. The two broad classes of parsers are top-down, in which the tree is built from the root downward to the leaves, and bottom-up, in which the parse tree is built from the leaves upward to the root.

In this chapter, we use a small set of notational conventions for grammar symbols and strings to make the discussion less cluttered. For formal languages, they are as follows:

1. Terminal symbols—lowercase letters at the beginning of the alphabet (a, b, . . .)

2. Nonterminal symbols—uppercase letters at the beginning of the alphabet (A, B, . . .)

3. Terminals or nonterminals—uppercase letters at the end of the alphabet (W, X, Y, Z)

4. Strings of terminals—lowercase letters at the end of the alphabet (w, x, y, z)

5. Mixed strings (terminals and/or nonterminals)—lowercase Greek letters (α, β, δ, γ)

For programming languages, terminal symbols are the small-scale syntactic constructs of the language, what we have referred to as lexemes. The nonterminal symbols of programming languages are usually connotative names or abbreviations, surrounded by angle brackets—for example, <while_​statement>, <expr>, and <function_def>. The sentences of a language (​programs, in the case of a programming language) are strings of terminals. Mixed strings describe right-hand sides (RHSs) of grammar rules and are used in parsing algorithms.

4.3.2 Top-Down Parsers A top-down parser traces or builds a parse tree in preorder. A preorder traversal of a parse tree begins with the root. Each node is visited before its branches are followed. Branches from a particular node are followed in left- to-right order. This corresponds to a leftmost derivation.

In terms of the derivation, a top-down parser can be described as follows: Given a sentential form that is part of a leftmost derivation, the parser’s task

is to find the next sentential form in that leftmost derivation. The general form of a left sentential form is xAα, whereby our notational conventions x is a string of terminal symbols, A is a nonterminal, and α is a mixed string. Because x contains only terminals, A is the leftmost nonterminal in the sentential form, so it is the one that must be expanded to get the next sentential form in a leftmost derivation. Determining the next sentential form is a matter of choosing the correct grammar rule that has A as its LHS. For example, if the current sentential form is xAα and the A-rules are A→bB, A→cBb, and A→a, a top-down parser must choose among these three rules to get the next sentential form, which could be xbBα, xcBbα, or xaα. This is the parsing decision problem for top-down parsers.

Different top-down parsing algorithms use different information to make parsing decisions. The most common top-down parsers choose the correct RHS for the leftmost nonterminal in the current sentential form by comparing the next token of input with the first symbols that can be generated by the RHSs of those rules. Whichever RHS has that token at the left end of the string it generates is the correct one. So, in the sentential form xAα, the parser would use whatever token would be the first generated by A to determine which A-rule should be used to get the next sentential form. In the example above, the three RHSs of the A-rules all begin with different terminal symbols. The parser can easily choose the correct RHS based on the next token of input, which must be a, b, or c in this example. In general, choosing the correct RHS is not so straightforward, because some of the RHSs of the leftmost nonterminal in the current sentential form may begin with a nonterminal.

The most common top-down parsing algorithms are closely related. A recursive-descent parser is a coded version of a syntax analyzer based directly on the BNF description of the syntax of language. The most common alternative to recursive descent is to use a parsing table, rather than code, to implement the BNF rules. Both of these, which are called LL ​algorithms, are equally powerful, meaning they work on the same subset of all context-free grammars. The first L in LL specifies a left-to-right scan of the input; the second L specifies that a leftmost derivation is generated. Section 4.4 introduces the recursive-descent approach to implementing an LL parser.

4.3.3 Bottom-Up Parsers A bottom-up parser constructs a parse tree by beginning at the leaves and progressing toward the root. This parse order corresponds to the reverse of a rightmost derivation. That is, the sentential forms of the derivation are produced in order of last to first. In terms of the derivation, a bottom-up parser can be described as follows: Given a right sentential form α, the parser must determine what substring of α is the RHS of the rule in the grammar that must be reduced to its LHS to produce the previous sentential form in the rightmost derivation. For example, the first step for a bottom-up parser is to determine which substring of the initial given sentence is the RHS to be reduced to its corresponding LHS to get the second last sentential form in the derivation. The process of finding the correct RHS to reduce is complicated by the fact that a given right sentential form may include more than one RHS from the grammar of the language being parsed. The correct RHS is called the ​handle. A right sentential form is a sentential form that appears in a rightmost derivation.

Consider the following grammar and derivation:

S → aAc

A→aA | b

S => aAc => aaAc => aabc

A bottom-up parser of this sentence, aabc, starts with the sentence and must find the handle in it. In this example, this is an easy task, for the string contains only one RHS, b. When the parser replaces b with its LHS, A, it gets the second to last sentential form in the derivation, aaAc. In the general case, as stated previously, finding the handle is much more difficult, because a sentential form may include several different RHSs.

A bottom-up parser finds the handle of a given right sentential form by examining the symbols on one or both sides of a possible handle. Symbols to the right of the possible handle are usually tokens in the input that have not yet been analyzed.

The most common bottom-up parsing algorithms are in the LR family, where the L specifies a left-to-right scan of the input and the R specifies that a rightmost derivation is generated.

4.3.4 The Complexity of Parsing Parsing algorithms that work for any unambiguous grammar are complicated and inefficient. In fact, the complexity of such algorithms is O(n3), which means the amount of time they take is on the order of the cube of the length of the string to be parsed. This relatively large amount of time is required because these algorithms frequently must back up and reparse part of the sentence being analyzed. Reparsing is required when the parser has made a mistake in the parsing process. Backing up the parser also requires that part of the parse tree being constructed (or its trace) must be dismantled and rebuilt. O(n3) algorithms are normally not useful for practical processes, such as syntax analysis for a compiler, because they are far too slow. In situations such as this, computer scientists often search for algorithms that are faster, though less general. Generality is traded for efficiency. In terms of parsing, faster algorithms have been found that work for only a subset of the set of all possible grammars. These algorithms are acceptable as long as the subset includes grammars that describe programming languages. (Actually, as discussed in Chapter 3, the whole class of context-free grammars is not adequate to describe all of the syntax of most programming languages.)

All algorithms used for the syntax analyzers of commercial compilers have complexity O(n), which means the time they take is linearly related to the length of the string to be parsed. This is vastly more efficient than O(n3) algorithms.

4.4 Recursive-Descent Parsing This section introduces the recursive-descent top-down parser implementation process.

4.4.1 The Recursive-Descent Parsing Process A recursive-descent parser is so named because it consists of a collection of subprograms, many of which are recursive, and it produces a parse tree in top-down order. This recursion is a reflection of the nature of programming languages, which include several different kinds of nested structures. For example, statements are often nested in other statements. Also, parentheses in expressions must be properly nested. The syntax of these structures is naturally described with recursive grammar rules.

EBNF is ideally suited for recursive-descent parsers. Recall from ​Chapter 3 that the primary EBNF extensions are braces, which specify that what they enclose can appear zero or more times, and brackets, which specify that what they enclose can appear once or not at all. Note that in both cases, the enclosed symbols are optional. Consider the following examples:

<if_statement> → if <logic_expr> <statement> [else <statement>]

<ident_list> → ident {, ident}

In the first rule, the else clause of an if statement is optional. In the second, an <ident_list> is an identifier, followed by zero or more repetitions of a comma and an identifier.

A recursive-descent parser has a subprogram for each nonterminal in its associated grammar. The responsibility of the subprogram associated with a particular nonterminal is as follows: When given an input string, it traces out

the parse tree that can be rooted at that nonterminal and whose leaves match the input string. In effect, a recursive-descent parsing subprogram is a parser for the language (set of strings) that is generated by its associated nonterminal.

Consider the following EBNF description of simple arithmetic expressions:

<expr> → <term> {(+ | -) <term>}

<term> → <factor> {(* | /) <factor>}

<factor> → id | int_constant | ( <expr> )

Recall from Chapter 3 that an EBNF grammar for arithmetic expressions, such as this one, does not force any associativity rule. Therefore, when using such a grammar as the basis for a compiler, one must take care to ensure that the code generation process, which is normally driven by syntax analysis, produces code that adheres to the associativity rules of the language. This can be done easily when recursive-descent parsing is used.

In the following recursive-descent function, expr, the lexical analyzer is the function that is implemented in Section 4.2. It gets the next lexeme and puts its token code in the global variable nextToken. The token codes are defined as named constants, as in Section 4.2.

A recursive-descent subprogram for a rule with a single RHS is relatively simple. For each terminal symbol in the RHS, that terminal symbol is compared with nextToken. If they do not match, it is a syntax error. If they match, the lexical analyzer is called to get the next input token. For each nonterminal, the parsing subprogram for that nonterminal is called.

The recursive-descent subprogram for the first rule in the previous example grammar, written in C, is

/* expr

Parses strings in the language generated by the rule:

<expr> -> <term> {(+ | -) <term>}

*/

void expr() {

printf("Enter <expr>\n");

/* Parse the first term */

term();

/* As long as the next token is + or -, get

the next token and parse the next term */

while (nextToken == ADD_OP || nextToken == SUB_OP) {

lex();

term();

}

printf("Exit <expr>\n");

} /* End of function expr */

Notice that the expr function includes tracing output statements, which are included to produce the example output shown later in this section.

Recursive-descent parsing subprograms are written with the convention that each one leaves the next token of input in nextToken. So, whenever a parsing function begins, it assumes that nextToken has the code for the leftmost token of the input that has not yet been used in the parsing process.

The part of the language that the expr function parses consists of one or more terms, separated by either plus or minus operators. This is the language generated by the nonterminal <expr>. Therefore, first it calls the function that parses terms (term). Then it continues to call that function as long as it finds ADD_OP or SUB_OP tokens (which it passes over by calling lex). This recursive-descent function is simpler than most, because its associated rule has only one RHS. Furthermore, it does not include any code for syntax error detection or recovery, because there are no detectable errors associated with the grammar rule.

A recursive-descent parsing subprogram for a nonterminal whose rule has more than one RHS begins with code to determine which RHS is to be parsed. Each RHS is examined (at compiler construction time) to determine the set of terminal symbols that can appear at the beginning of sentences it can generate. By matching these sets against the next token of input, the parser can choose the correct RHS.

The parsing subprogram for <term> is similar to that for <expr>:

/* term

Parses strings in the language generated by the rule:

<term> -> <factor> {(* | /) <factor>)

*/

void term() {

printf("Enter <term>\n");

/* Parse the first factor */

factor();

/* As long as the next token is * or /, get the

next token and parse the next factor */

while (nextToken == MULT_OP || nextToken == DIV_OP) {

lex();

factor();

}

printf("Exit <term>\n");

} /* End of function term */

The function for the <factor> nonterminal of our arithmetic expression grammar must choose between its two RHSs. It also includes error detection. In the function for <factor>, the reaction to detecting a syntax error is simply to call the error function. In a real parser, a diagnostic message must be produced when an error is detected. Furthermore, parsers must recover from the error so that the parsing process can continue.

/* factor

Parses strings in the language generated by the rule:

<factor> -> id | int_constant | ( <expr )

*/

void factor() {

printf("Enter <factor>\n");

/* Determine which RHS */

if (nextToken == IDENT || nextToken == INT_LIT)

/* Get the next token */

lex();

/* If the RHS is ( <expr> ), call lex to pass over the

left parenthesis, call expr, and check for the right

parenthesis */

else {

if (nextToken == LEFT_PAREN) {

lex();

expr();

if (nextToken == RIGHT_PAREN)

lex();

else

error();

} /* End of if (nextToken == ... */

/* It was not an id, an integer literal, or a left

parenthesis */

else error();

} /* End of else */

printf("Exit <factor>\n");;

} /* End of function factor */

Following is the trace of the parse of the example expression (sum + 47) / total, using the parsing functions expr, term, and factor, and the function lex from Section 4.2. Note that the parse begins by calling lex and the start symbol routine, in this case, expr.

Next token is: 25 Next lexeme is (

Enter <expr>

Enter <term>

Enter <factor>

Next token is: 11 Next lexeme is sum

Enter <expr>

Enter <term>

Enter <factor>

Next token is: 21 Next lexeme is +

Exit <factor>

Exit <term>

Next token is: 10 Next lexeme is 47

Enter <term>

Enter <factor>

Next token is: 26 Next lexeme is )

Exit <factor>

Exit <term>

Exit <expr>

Next token is: 24 Next lexeme is /

Exit <factor>

Next token is: 11 Next lexeme is total

Enter <factor>

Next token is: -1 Next lexeme is EOF

Exit <factor>

Exit <term>

Exit <expr>

The parse tree traced by the parser for the preceding expression is shown in Figure 4.2.

Figure 4.2 Parse tree for (sum + 47)/ total

Figure 4.2 Full Alternative Text

One more example grammar rule and parsing function should help solidify the reader’s understanding of recursive-descent parsing. Following is a grammatical description of the Java if statement:

<ifstmt> → if (<boolexpr>) <statement> [else <statement>]

The recursive-descent subprogram for this rule follows:

/* Function ifstmt

Parses strings in the language generated by the rule:

<ifstmt> -> if (<boolexpr>) <statement>

[else <statement>]

*/

void ifstmt() {

/* Be sure the first token is 'if' */

if (nextToken = IF_CODE)

error();

else {

/* Call lex to get to the next token */

lex();

/* Check for the left parenthesis */

if (nextToken = LEFT_PAREN)

error();

else {

/* Parse the Boolean expression */

boolexpr();

/* Check for the right parenthesis */

if (nextToken = RIGHT_PAREN)

error();

else {

/* Parse the then clause */

statement();

/* If an else is next, parse the else clause */

if (nextToken == ELSE_CODE) {

/* Call lex to get over the else */

lex();

statement();

} /* end of if (nextToken == ELSE_CODE ... */

} /* end of else of if (nextToken != RIGHT ... */

} /* end of else of if (nextToken != LEFT ... */

} /* end of else of if (nextToken != IF_CODE ... */

} /* end of ifstmt */

Notice that this function uses parser functions for statements and Boolean expressions that are not given in this section.

The objective of these examples is to convince you that a recursive-descent parser can be easily written if an appropriate grammar is available for the language. The characteristics of a grammar that allows a recursive-descent parser to be built are discussed in the following subsection.

4.4.2 The LL Grammar Class Before choosing to use recursive descent as a parsing strategy for a compiler or other program analysis tool, one must consider the limitations of the approach, in terms of grammar restrictions. This section discusses these restrictions and their possible solutions.

One simple grammar characteristic that causes a catastrophic problem for LL parsers is left recursion. For example, consider the following rule:

A→A+B

A recursive-descent parser subprogram for A immediately calls itself to parse the first symbol in its RHS. That activation of the A parser subprogram then immediately calls itself again, and again, and so forth. It is easy to see that this leads nowhere (except to stack overflow).

The left recursion in the rule A→A+B is called direct left recursion, because it occurs in one rule. Direct left recursion can be eliminated from a grammar by the following process:

For each nonterminal, A,

1. Group the A-rules as A→Aα1, | ⋯ | Aαm | β1 | β2 | ⋯ | βn where none of the β's begins with A

2. Replace the original A-rules with

A→β1A′ | β2A′ | ⋯ | βnA′ A′→α1A′ | α2A′ | αmA′ | ε

Note that ε specifies the empty string. A rule that has ε as its RHS is called an erasure rule, because its use in a derivation effectively erases its LHS from the sentential form.

Consider the following example grammar and the application of the above process:

E→E+T | T T→T * F | F F→(E )+id

For the E-rules, we have α1 = + T and β = T, so we replace the E-rules with

E→T E′ E′→+T E′ | ε

For the T-rules, we have α1 = *F and β = F, so we replace the T-rules with

T→F T′ T  ′→*F T′ | ε

Because there is no left recursion in the F-rules, they remain the same, so the complete replacement grammar is

E→T E′ E′→+T E′ | ε T→F T′ T′→ * F T′ | ε F→(E ) | id

This grammar generates the same language as the original grammar but is not left recursive.

As was the case with the expression grammar written using EBNF in Section 4.4.1, this grammar does not specify left associativity of operators. However, it is relatively easy to design the code generation based on this grammar so that the addition and multiplication operators will have left associativity.

Indirect left recursion poses the same problem as direct left recursion. For example, suppose we have

A → B a A B → A b

A recursive-descent parser for these rules would have the A subprogram immediately call the subprogram for B, which immediately calls the A subprogram. So, the problem is the same as for direct left recursion. The problem of left recursion is not confined to the recursive-descent approach to building top-down parsers. It is a problem for all top-down parsing algorithms. Fortunately, left recursion is not a problem for bottom-up parsing algorithms.

There is an algorithm to modify a given grammar to remove indirect left recursion (Aho et al., 2006), but it is not covered here. When writing a ​- grammar for a programming language, one can usually avoid including left

recursion, both direct and indirect.

Left recursion is not the only grammar trait that disallows top-down parsing. Another is whether the parser can always choose the correct RHS on the basis of the next token of input, using only the first token generated by the leftmost nonterminal in the current sentential form. There is a relatively simple test of a non-left recursive grammar that indicates whether this can be done, called the pairwise disjointness test. This test requires the ability to compute a set based on the RHSs of a given nonterminal symbol in a grammar. These sets, which are called FIRST, are defined as

FIRST (α) = {a | α =>* aβ} (IF α =>* ε, ε is in FIRST (α))

in which =>* means 0 or more derivation steps.

An algorithm to compute FIRST for any mixed string α can be found in Aho et al. (2006). For our purposes, FIRST can usually be computed by inspection of the grammar.

The pairwise disjointness test is as follows:

For each nonterminal, A, in the grammar that has more than one RHS, for each pair of rules, A→αi and A→αj, it must be true that

FIRST(αi)∩FIRST(αj)=ϕ

(The intersection of the two sets, FIRST(αi) and FIRST(αj), must be empty.)

In other words, if a nonterminal A has more than one RHS, the first terminal symbol that can be generated in a derivation for each of them must be unique to that RHS. Consider the following rules:

A→aB | bAb | Bb B→cB | d

The FIRST sets for the RHSs of the A-rules are {a}, {b}, and {c}, {d}, which are clearly disjoint. Therefore, these rules pass the pairwise disjointness test. What this means, in terms of a recursive-descent parser, is that the code of the subprogram for parsing the nonterminal A can choose which RHS it is dealing with by seeing only the first terminal symbol of input (token) that is

generated by the nonterminal. Now consider the rules

A→aB | BAb B→aA | b

The FIRST sets for the RHSs in the A-rules are {a} and {a}, {b} which are clearly not disjoint. So, these rules fail the pairwise disjointness test. In terms of the parser, the subprogram for A could not determine which RHS was being parsed by looking at the next symbol of input, because if it were an a, it could be either RHS. This issue is of course more complex if one or more of the RHSs begin with nonterminals.

In many cases, a grammar that fails the pairwise disjointness test can be modified so that it will pass the test. For example, consider the rule

<variable> → identifier | identifier [<expression>]

This states that a <variable> is either an identifier or an identifier followed by an expression in brackets (a subscript). These rules clearly do not pass the pairwise disjointness test, because both RHSs begin with the same terminal, identifier. This problem can be alleviated through a process called left factoring.

We now take an informal look at left factoring. Consider our rules for <variable>. Both RHSs begin with identifier. The parts that follow identifier in the two RHSs are ε (the empty string) and [<expression>]. The two rules can be replaced by the following two rules:

<variable> → identifier <new>

<new> →ε | [<expression>]

It is not difficult to see that together, these two rules generate the same language as the two rules with which we began. However, these two pass the pairwise disjointness test.

If the grammar is being used as the basis for a recursive-descent parser, an alternative to left factoring is available. With an EBNF extension, the problem disappears in a way that is very similar to the left factoring solution.

Consider the original rules above for <variable>. The subscript can be made optional by placing it in square brackets, as in

<variable> → identifier [[<expression>]]

In this rule, the outer brackets are metasymbols that indicate that what is inside is optional. The inner brackets are terminal symbols of the programming language being described. The point is that we replaced two rules with a single rule that generates the same language but passes the pairwise disjointness test.

A formal algorithm for left factoring can be found in Aho et al. (2006). Left factoring cannot solve all pairwise disjointness problems of grammars. In some cases, rules must be rewritten in other ways to eliminate the problem.

4.5 Bottom-Up Parsing This section introduces the general process of bottom-up parsing and includes a description of the LR parsing algorithm.

4.5.1 The Parsing Problem for Bottom-Up Parsers Consider the following grammar for arithmetic expressions:

E→E + T | T

T→T * F | F

F→ (E) | id

Notice that this grammar generates the same arithmetic expressions as the example in Section 4.4. The difference is that this grammar is left recursive, which is acceptable to bottom-up parsers. Also note that grammars for bottom-up parsers normally do not include metasymbols such as those used to specify extensions to BNF. The following rightmost derivation illustrates this grammar:

E => E+T

  => E + T * F

  => E + T * id

  => E + F * id

  => E + id * id

  => T + id * id

  => F + id * id

  => id + id * id

The underlined part of each sentential form in this derivation is the RHS that is rewritten as its corresponding LHS to get the previous sentential form. The process of bottom-up parsing produces the reverse of a rightmost derivation. So, in the example derivation, a bottom-up parser starts with the last sentential form (the input sentence) and produces the sequence of sentential forms from there until all that remains is the start symbol, which in this grammar is E. In each step, the task of the bottom-up parser is to find the specific RHS, the handle, in the sentential form that must be rewritten to get the next (previous) sentential form. As mentioned earlier, a right sentential form may include more than one RHS. For example, the right sentential form

E + T * id

includes three RHSs, E+T, T and id. Only one of these is the handle. For example, if the RHS E+T were chosen to be rewritten in this sentential form, the resulting sentential form would be E * id, but E * id is not a legal right sentential form for the given grammar.

The handle of a right sentential form is unique. The task of a bottom-up parser is to find the handle of any given right sentential form that can be generated by its associated grammar. Formally, handle is defined as follows:

Definition: β is the handle of the right sentential form γ=αβw if and only if S => *rm αAw =>rmαβw

In this definition, =>rm specifies a rightmost derivation step, and => *rm specifies zero or more rightmost derivation steps. Although the definition of a handle is mathematically concise, it provides little help in finding the handle of a given right sentential form. In the following, we provide the definitions of several substrings of sentential forms that are related to handles. The purpose of these is to provide some intuition about handles.

Definition: β is a phrase of the right sentential form γ if and only if S =>* γ=α1Aα2 =>+α1βα2

In this definition, =>+ means one or more derivation steps.

Definition: β is a simple phrase of the right sentential form γ if and only if S =>* γ=α1Aα2 =>+α1βα2

If these two definitions are compared carefully, it is clear that they differ only in the last derivation specification. The definition of phrase uses one or more steps, while the definition of simple phrase uses exactly one step.

The definitions of phrase and simple phrase may appear to have the same lack of practical value as that of a handle, but that is not true. Consider what a phrase is relative to a parse tree. It is the string of all of the leaves of the partial parse tree that is rooted at one particular internal node of the whole parse tree. A simple phrase is just a phrase that takes a single derivation step from its root nonterminal node. In terms of a parse tree, a phrase can be derived from a single nonterminal in one or more tree levels, but a simple phrase can be derived in just a single tree level. Consider the parse tree shown in Figure 4.3.

Figure 4.3 A parse tree for E+T * id

The leaves of the parse tree in Figure 4.3 comprise the sentential form E+T * id. Because there are three internal nodes, there are three phrases. Each internal node is the root of a subtree, whose leaves are a phrase. The root node of the whole parse tree, E, generates all of the resulting sentential form, E+T * id, which is a phrase. The internal node, T, generates the leaves T * id, which is another phrase. Finally, the internal node, F, generates id, which is also a phrase. So, the phrases of the sentential form E+T * id are E+T * id, and id. Notice that phrases are not necessarily RHSs in the underlying grammar.

The simple phrases are a subset of the phrases. In the previous example, the only simple phrase is id. A simple phrase is always a RHS in the grammar.

The reason for discussing phrases and simple phrases is this: The handle of

any rightmost sentential form is its leftmost simple phrase. So now we have a highly intuitive way to find the handle of any right sentential form, assuming we have the grammar and can draw a parse tree. This approach to finding handles is of course not practical for a parser. (If you already have a parse tree, why do you need a parser?) Its only purpose is to provide the reader with some intuitive feel for what a handle is, relative to a parse tree, which is easier than trying to think about handles in terms of sentential forms.

We can now consider bottom-up parsing in terms of parse trees, although the purpose of a parser is to produce a parse tree. Given the parse tree for an entire sentence, you easily can find the handle, which is the first thing to rewrite in the sentence to get the previous sentential form. Then the handle can be pruned from the parse tree and the process repeated. Continuing to the root of the parse tree, the entire rightmost derivation can be constructed.

4.5.2 Shift-Reduce Algorithms Bottom-up parsers are often called shift-reduce algorithms, because shift and reduce are the two most common actions they specify. An integral part of every bottom-up parser is a stack. As with other parsers, the input to a bottom-up parser is the stream of tokens of a program and the output is a sequence of grammar rules. The shift action moves the next input token onto the parser’s stack. A reduce action replaces an RHS (the handle) on top of the parser’s stack by its corresponding LHS. Every parser for a programming language is a pushdown automaton (PDA), because a PDA is a recognizer for a context-free language. You need not be intimate with PDAs to understand how a bottom-up parser works, although it helps. A PDA is a very simple mathematical machine that scans strings of symbols from left to right. A PDA is so named because it uses a pushdown stack as its memory. PDAs can be used as recognizers for context-free languages. Given a string of symbols over the alphabet of a context-free language, a PDA that is designed for the purpose can determine whether the string is or is not a sentence in the language. In the process, the PDA can produce the information needed to construct a parse tree for the sentence.

With a PDA, the input string is examined, one symbol at a time, left to right.

The input is treated very much as if it were stored in another stack, because the PDA never sees more than the leftmost symbol of the input.

Note that a recursive-descent parser is also a PDA. In that case, the stack is that of the run-time system, which records subprogram calls (among other things), which correspond to the nonterminals of the grammar.

4.5.3 LR Parsers Many different bottom-up parsing algorithms have been devised. Most of them are variations of a process called LR. LR parsers use a relatively small program and a parsing table that is built for a specific programming language. The original LR algorithm was designed by Donald Knuth (Knuth, 1965). This algorithm, which is sometimes called canonical LR, was not used in the years immediately following its publication because producing the required parsing table required large amounts of computer time and memory. Subsequently, several variations on the canonical LR table construction process were developed (DeRemer, 1971; DeRemer and Pennello, 1982). These are characterized by two properties: (1) They require far less computer resources to produce the required parsing table than the canonical LR algorithm, and (2) they work on smaller classes of grammars than the canonical LR algorithm.

There are three advantages to LR parsers:

1. They can be built for all programming languages.

2. They can detect syntax errors as soon as it is possible in a left-to-right scan.

3. The LR class of grammars is a proper superset of the class parsable by LL parsers (for example, many left recursive grammars are LR, but none are LL).

The only disadvantage of LR parsing is that it is difficult to produce by hand the parsing table for a given grammar for a complete programming language.

This is not a serious disadvantage, however, for there are several programs available that take a grammar as input and produce the parsing table, as ​- discussed later in this section.

Prior to the appearance of the LR parsing algorithm, there were a number of parsing algorithms that found handles of right sentential forms by looking both to the left and to the right of the substring of the sentential form that was suspected of being the handle. Knuth’s insight was that one could effectively look to the left of the suspected handle all the way to the bottom of the parse stack to determine whether it was the handle. But all of the information in the parse stack that was relevant to the parsing process could be represented by a single state, which could be stored on the top of the stack. In other words, Knuth discovered that regardless of the length of the input string, the length of the sentential form, or the depth of the parse stack, there were only a relatively small number of different situations, as far as the parsing process is concerned. Each situation could be represented by a state and stored in the parse stack, one state symbol for each grammar symbol on the stack. At the top of the stack would always be a state symbol, which represented the relevant information from the entire history of the parse, up to the current time. We will use subscripted uppercase S’s to represent the parser states.

Figure 4.4 shows the structure of an LR parser. The contents of the parse stack for an LR parser have the following form:

Figure 4.4 The structure of an LR parser

Figure 4.4 Full Alternative Text

S0X1S1X2…XmSm (top)

where the S’s are state symbols and the X’s are grammar symbols. An LR parser configuration is a pair of strings (stack, input), with the detailed form

(S0X1S1X2S2…XmSm, aiAi+1…an$)

Notice that the input string has a dollar sign at its right end. This sign is put there during initialization of the parser. It is used for normal termination of the parser. Using this parser configuration, we can formally define the LR parser process, which is based on the parsing table.

An LR parsing table has two parts, named ACTION and GOTO. The ACTION part of the table specifies most of what the parser does. It has state symbols as its row labels and the terminal symbols of the grammar as its column labels. Given a current parser state, which is represented by the state symbol on top of the parse stack, and the next symbol (token) of input, the parse table specifies what the parser should do. The two primary parser actions are shift and reduce. Either the parser shifts the next input symbol onto the parse stack, along with a state symbol, or it already has the handle on top of the stack, which it reduces to the LHS of the rule whose RHS is the same as the handle. Two other actions are possible: accept, which means the parser has successfully completed the parse of the input, and error, which means the parser has detected a syntax error.

The rows of the GOTO part of the LR parsing table have state symbols as labels. This part of the table has nonterminals as column labels. The values in the GOTO part of the table indicate which state symbol should be pushed onto the parse stack after a reduction has been completed, which means the handle has been removed from the parse stack and the new nonterminal has been pushed onto the parse stack. The specific symbol is found at the row

whose label is the state symbol on top of the parse stack after the handle and its associated state symbols have been removed. The column of the GOTO table that is used is the one with the label, that is the LHS of the rule used in the reduction.

Consider the traditional grammar for arithmetic expressions that follows:

1. E→E + T

2. E→T

3. T→T * F

4. T→F

5. F→ (E)

6. F→id

The rules of this grammar are numbered to provide a simple way to reference them in a parsing table.

Figure 4.5 shows the LR parsing table for this grammar. Abbreviations are used for the actions: R for reduce and S for shift. R4 means reduce using rule 4; S6 means shift the next symbol of input onto the stack and push state 6 onto the stack. Empty positions in the ACTION table indicate syntax errors. In a complete parser, these could have calls to error-handling routines.

Figure 4.5 The LR parsing table for an arithmetic expression grammar

Figure 4.5 Full Alternative Text

LR parsing tables can easily be constructed using a software tool, such as yacc (Johnson, 1975), which takes the grammar as input. Although LR parsing tables can be produced by hand, for a grammar of a real programming language, the task would be lengthy, tedious, and error prone. For real compilers, LR parsing tables are always generated with software

tools.

The initial configuration of an LR parser is

(S0, a1…an$)

The parser actions are informally defined as follows:

1. The Shift process is simple: The next symbol of input is pushed onto the stack, along with the state symbol that is part of the Shift specification in the ACTION table.

2. For a Reduce action, the handle must be removed from the stack. Because for every grammar symbol on the stack there is a state symbol, the number of symbols removed from the stack is twice the number of symbols in the handle. After removing the handle and its associated state symbols, the LHS of the rule is pushed onto the stack. Finally, the GOTO table is used, with the row label being the symbol that was exposed when the handle and its state symbols were removed from the stack, and the column label being the nonterminal that is the LHS of the rule used in the reduction.

3. When the action is Accept, the parse is complete and no errors were found.

4. When the action is Error, the parser calls an error-handling routine.

Although there are many parsing algorithms based on the LR concept, they differ only in the construction of the parsing table. All LR parsers use this same parsing algorithm.

Perhaps the best way to become familiar with the LR parsing process is through an example. Initially, the parse stack has the single symbol 0, which represents state 0 of the parser. The input contains the input string with an end marker, in this case a dollar sign, attached to its right end. At each step, the parser actions are dictated by the top (rightmost in Figure 4.4) symbol of the parse stack and the next (leftmost in Figure 4.4) token of input. The correct action is chosen from the corresponding cell of the ACTION part of

the parse table. The GOTO part of the parse table is used after a reduction action. Recall that GOTO is used to determine which state symbol is placed on the parse stack after a reduction.

Following is a trace of a parse of the string id + id * id, using the LR parsing algorithm and the parsing table shown in Figure 4.5.

The algorithms to generate LR parsing tables from given grammars, which are described in Aho et al. (2006), are not overly complex but are beyond the scope of a book on programming languages. As stated previously, there are a number of different software systems available to generate LR parsing tables.

SUMMARY Syntax analysis is a common part of language implementation, regardless of the implementation approach used. Syntax analysis is normally based on a formal syntax description of the language being implemented. A context-free grammar, which is also called BNF, is the most common approach for describing syntax. The task of syntax analysis is usually divided into two parts: lexical analysis and syntax analysis. There are several reasons for separating lexical analysis—namely, simplicity, efficiency, and portability.

A lexical analyzer is a pattern matcher that isolates the small-scale parts of a program, which are called lexemes. Lexemes occur in categories, such as integer literals and names. These categories are called tokens. Each token is assigned a numeric code, which along with the lexeme is what the lexical analyzer produces. There are three distinct approaches to constructing a lexical analyzer: using a software tool to generate a table for a table-driven analyzer, building such a table by hand, and writing code to implement a state diagram description of the tokens of the language being implemented. The state diagram for tokens can be reasonably small if character classes are used for transitions, rather than having transitions for every possible character from every state node. Also, the state diagram can be simplified by using a table lookup to recognize reserved words.

Syntax analyzers have two goals: to detect syntax errors in a given program and to produce a parse tree, or possibly only the information required to build such a tree, for a given program. Syntax analyzers are either top-down, meaning they construct leftmost derivations and a parse tree in top-down order, or bottom-up, in which case they construct the reverse of a rightmost derivation and a parse tree in bottom-up order. Parsers that work for all unambiguous grammars have complexity O(n3). However, parsers used for implementing syntax analyzers for programming languages work on subclasses of unambiguous grammars and have complexity O(n).

A recursive-descent parser is an LL parser that is implemented by writing code directly from the grammar of the source language. EBNF is ideal as the

basis for recursive-descent parsers. A recursive-descent parser has a subprogram for each nonterminal in the grammar. The code for a given grammar rule is simple if the rule has a single RHS. The RHS is examined left to right. For each nonterminal, the code calls the associated subprogram for that nonterminal, which parses whatever the nonterminal generates. For each terminal, the code compares the terminal with the next token of input. If they match, the code simply calls the lexical analyzer to get the next token. If they do not, the subprogram reports a syntax error. If a rule has more than one RHS, the subprogram must first determine which RHS it should parse. It must be possible to make this determination on the basis of the next token of input.

Two distinct grammar characteristics prevent the construction of a ​recursive- descent parser based on the grammar. One of these is left recursion. The process of eliminating direct left recursion from a grammar is relatively simple. Although we do not cover it, an algorithm exists to remove both direct and indirect left recursion from a grammar. The other problem is detected with the pairwise disjointness test, which tests whether a parsing subprogram can determine which RHS is being parsed on the basis of the next token of input. Some grammars that fail the pairwise disjointness test often can be modified to pass it, using left factoring.

The parsing problem for bottom-up parsers is to find the substring of the current sentential form that must be reduced to its associated LHS to get the next (previous) sentential form in the rightmost derivation. This substring is called the handle of the sentential form. A parse tree can provide an intuitive basis for recognizing a handle. A bottom-up parser is a shift-reduce algorithm, because in most cases it either shifts the next lexeme of input onto the parse stack or reduces the handle that is on top of the stack.

The LR family of shift-reduce parsers is the most commonly used ​bottom-up parsing approach for programming languages, because parsers in this family have several advantages over alternatives. An LR parser uses a parse stack, which contains grammar symbols and state symbols to maintain the state of the parser. The top symbol on the parse stack is always a state symbol that represents all of the information in the parse stack that is relevant to the parsing process. LR parsers use two parsing tables: ACTION and GOTO.

The ACTION part specifies what the parser should do, given the state symbol on top of the parse stack and the next token of input. The GOTO table is used to determine which state symbol should be placed on the parse stack after a reduction has been done.

REVIEW QUESTIONS 1. What are three reasons why syntax analyzers are based on grammars?

2. Explain the three reasons why lexical analysis is separated from syntax analysis.

3. Define lexeme and token.

4. What are the primary tasks of a lexical analyzer?

5. Describe briefly the three approaches to building a lexical analyzer.

6. What is a state transition diagram?

7. Why are character classes used, rather than individual characters, for the letter and digit transitions of a state diagram for a lexical analyzer?

8. What are the two distinct goals of syntax analysis?

9. Describe the differences between top-down and bottom-up parsers.

10. Describe the parsing problem for a top-down parser.

11. Describe the parsing problem for a bottom-up parser.

12. Explain why compilers use parsing algorithms that work on only a subset of all grammars.

13. Why are named constants used, rather than numbers, for token codes?

14. Describe how a recursive-descent parsing subprogram is written for a rule with a single RHS.

15. Explain the two grammar characteristics that prohibit them from being used as the basis for a top-down parser.

16. What is the FIRST set for a given grammar and sentential form?

17. Describe the pairwise disjointness test.

18. What is left factoring?

19. What is a phrase of a sentential form?

20. What is a simple phrase of a sentential form?

21. What is the handle of a sentential form?

22. What is the mathematical machine on which both top-down and bottom- up parsers are based?

23. Describe three advantages of LR parsers.

24. What was Knuth’s insight in developing the LR parsing technique?

25. Describe the purpose of the ACTION table of an LR parser.

26. Describe the purpose of the GOTO table of an LR parser.

27. Is left recursion a problem for LR parsers?

PROBLEM SET 1. Perform the pairwise disjointness test for the following grammar rules.

1. A→aB | b | cBB

2. B→aB | bA | aBb

3. C→aaA | b | caB

2. Perform the pairwise disjointness test for the following grammar rules.

1. S→aSb | bAA

2. A→B{aB} | a

3. B→aB | a

3. Show a trace of the recursive descent parser given in Section 4.4.1 for the string a + b * c.

4. Show a trace of the recursive descent parser given in Section 4.4.1 for the string a * (b + c).

5. Given the following grammar and the right sentential form, draw a parse tree and show the phrases and simple phrases, as well as the handle.

S→aAb | bBAA→ab | aABB→aB | b

1. aaAbb

2. bBab

3. aaAbBb

6. Given the following grammar and the right sentential form, draw a parse tree and show the phrases and simple phrases, as well as the handle.

S→AbB | bAcA→Ab | aBBB→Ac | cBb | c

1. aAcccbbc

2. AbcaBccb

3. baBcBbbc

7. Show a complete parse, including the parse stack contents, input string, and action for the string id * (id + id), using the grammar and parse table in Section 4.5.3.

8. Show a complete parse, including the parse stack contents, input string, and action for the string (id + id) * id, using the grammar and parse table in Section 4.5.3.

9. Get the algorithm to remove the indirect left recursion from a grammar from Aho et al. (2006). Use this algorithm to remove all left recursion from the following grammar: S→Aa | BbA→Aa | Abc | c | SbB→bb

PROGRAMMING EXERCISES 1. Design a state diagram to recognize one form of the comments of the C-

based programming languages, those that begin with /* and end with */.

2. Design a state diagram to recognize the floating-point literals of your favorite programming language.

3. Write and test the code to implement the state diagram of Problem 1.

4. Write and test the code to implement the state diagram of Problem 2.

5. Modify the lexical analyzer given in Section 4.2 to recognize the following list of reserved words and return their respective token codes: for (FOR_CODE, 30), if (IF_CODE, 31), else (ELSE_CODE, 32), while (WHILE_CODE, 33), do (DO_CODE, 34), int (INT_CODE, 35), float (FLOAT_CODE, 36), switch (SWITCH_CODE, 37).

6. Convert the lexical analyzer (which is written in C) given in Section 4.2 to Java.

7. Convert the recursive descent parser routines for <expr>, <term>, and <factor> given in Section 4.4.1 to Java.

8. For those rules that pass the test in Problem 1, write a recursive-descent parsing subprogram that parses the language generated by the rules. Assume you have a lexical analyzer named lex and an error-handling subprogram named error, which is called whenever a syntax error is detected.

9. For those rules that pass the test in Problem 2, write a recursive-descent parsing subprogram that parses the language generated by the rules. Assume you have a lexical analyzer named lex and an error-handling subprogram named error, which is called whenever a syntax error is detected.

10. Implement and test the LR parsing algorithm given in Section 4.5.3.

11. Write an EBNF rule that describes the while statement of Java or C++. Write the recursive-descent subprogram in Java or C++ for this rule.

12. Write an EBNF rule that describes the for statement of Java or C++. Write the recursive-descent subprogram in Java or C++ for this rule.

5 Names, Bindings, and Scopes 1. 5.1 Introduction

2. 5.2 Names

3. 5.3 Variables

4. 5.4 The Concept of Binding

5. 5.5 Scope

6. 5.6 Scope and Lifetime

7. 5.7 Referencing Environments

8. 5.8 Named Constants

This chapter introduces the fundamental semantic issues of variables. It begins by describing the nature of names and special words in programming ​languages. The attributes of variables, including type, address, and value, are then ​discussed, including the issue of aliases. The important concepts of binding and ​binding times are introduced next, including the different possible binding times for variable attributes and how they define four different categories of variables. Following that, two very different scoping rules for names, static and dynamic, are described, along with the concept of a referencing environment of a statement. Finally, named constants and variable initialization are discussed.

5.1 Introduction Imperative programming languages are, to varying degrees, abstractions of the underlying von Neumann computer architecture. The architecture’s two primary components are its memory, which stores both instructions and data, and its processor, which provides operations for modifying the contents of the memory. The abstractions in a language for the memory cells of the machine are variables. In some cases, the characteristics of the abstractions are very close to the characteristics of the cells; an example of this is an integer variable, which is usually represented directly in one or more bytes of memory. In other cases, the abstractions are far removed from the organization of the hardware memory, as with a three-dimensional array, which requires a software mapping function to support the abstraction.

A variable can be characterized by a collection of properties, or attributes, the most important of which is type, a fundamental concept in programming languages. Designing the data types of a language requires that a variety of issues be considered. (Data types are discussed in Chapter 6.) Among the most important of these issues are the scope and lifetime of variables.

Functional programming languages allow expressions to be named. These named expressions appear like assignments to variable names in imperative languages, but are fundamentally different in that they cannot be changed. So, they are like the named constants of the imperative languages. Pure functional languages do not have variables that are like those of the imperative languages. However, many functional languages do include such variables.

In the remainder of this book, families of languages will often be referred to as if they were single languages. For example, Fortran will mean all of the versions of Fortran. This is also the case for Ada. Likewise, a reference to C will mean the original version of C, as well as C89 and C99. When a specific version of a language is named, it is because it is different from the other family members within the topic being discussed. If we add a plus sign (+) to the name of a version of a language, we mean all versions of the language

beginning with the one named. For example, Fortran 95+ means all versions of Fortran beginning with Fortran 95. The phrase C-based languages will be used to refer to C, Objective-C, C++, Java, and C#.1

1. We were tempted to include the scripting languages JavaScript and PHP as C-based languages, but decided they were just a bit too different from their ancestors.

5.2 Names Before beginning our discussion of variables, the design of one of the fundamental attributes of variables, names, must be covered. Names are also associated with subprograms, formal parameters, and other program constructs. The term identifier is often used interchangeably with name.

5.2.1 Design Issues The following are the primary design issues for names:

Are names case sensitive?

Are the special words of the language reserved words or keywords?

These issues are discussed in the following two subsections, which also include examples of several design choices.

5.2.2 Name Forms A name is a string of characters used to identify some entity in a program.

C99 has no length limitation on its internal names, but only the first 63 are significant. External names in C99 (those defined outside functions, which must be handled by the linker) are restricted to 31 characters. Names in Java and C# have no length limit, and all characters in them are significant. C++ does not specify a length limit on names, although implementors sometimes do.

Names in most programming languages have the same form: a letter followed by a string consisting of letters, digits, and underscore characters ( _ ). Although the use of underscore characters to form names was widely used in

the 1970s and 1980s, that practice is now far less popular. In the C-based languages, it has to a large extent been replaced by the so-called camel notation, in which all of the words of a multiple-word name except the first are capitalized, as in myStack.2 Note that the use of underscores and mixed case in names is a programming style issue, not a language design issue.

2. It is called “camel” because words written in it often have embedded uppercase letters, which look like a camel’s humps.

history note The earliest programming languages used single-character names. This notation was natural because early programming was primarily mathematical, and mathematicians have long used single-character names for unknowns in their formal notations.

Fortran I broke with the tradition of the single-character name, allowing up to six characters in its names.

All variable names in PHP must begin with a dollar sign. In Perl, the special character at the beginning of a variable’s name, $, @, or %, specifies its type (although in a different sense than in other languages). In Ruby, special characters at the beginning of a variable’s name, @ or @@, indicate that the variable is an instance or a class variable, respectively.

In many languages, notably the C-based languages, uppercase and lowercase letters in names are distinct; that is, names in these languages are case sensitive. For example, the following three names are distinct in C++: rose, ROSE, and Rose. To some people, this is a serious detriment to readability, because names that look very similar in fact denote different entities. In that sense, case sensitivity violates the design principle that language constructs that look similar should have similar meanings. But in languages whose variable names are case-sensitive, although Rose and rose look similar, there is no connection between them.

Obviously, not everyone agrees that case sensitivity is bad for names. In C,

the problems of case sensitivity are avoided by the convention that variable names do not include uppercase letters. In Java and C#, however, the problem cannot be escaped because many of the predefined names include both uppercase and lowercase letters. For example, the Java method for converting a string to an integer value is parseInt, and spellings such as ParseInt and parseint are not recognized. This is a problem of writability rather than readability, because the need to remember specific case usage makes it more difficult to write correct programs. It is a kind of intolerance on the part of the language designer, which is enforced by the compiler.

5.2.3 Special Words Special words in programming languages are used to make programs more readable by naming actions to be performed. They also are used to separate the syntactic parts of statements and programs. In most languages, special words are classified as reserved words, which means they cannot be redefined by programmers. But in some, such as Fortran, they are only keywords, which means they can be redefined.

A reserved word is a special word of a programming language that cannot be used as a name. There is one potential problem with reserved words: If the language includes a large number of reserved words, the user may have difficulty making up names that are not reserved. The best example of this is COBOL, which has 300 reserved words. Unfortunately, some of the most commonly chosen names by programmers are in the list of reserved words— for example, LENGTH, BOTTOM, DESTINATION, and COUNT.

In program code examples in this book, reserved words are presented in boldface.

In most languages, names that are defined in other program units, such as Java packages and C and C++ libraries, can be made visible to a program. These names are predefined, but visible only if explicitly imported. Once imported, they cannot be redefined.

5.3 Variables A program variable is an abstraction of a computer memory cell or collection of cells. Programmers often think of variables as names for memory locations, but there is much more to a variable than just a name.

The move from machine languages to assembly languages was largely one of replacing absolute numeric memory addresses for data with names, making programs far more readable and therefore easier to write and maintain. Assembly languages also provided an escape from the problem of manual absolute addressing, because the translator that converted the names to actual addresses also chose those addresses.

A variable can be characterized as a sextuple of attributes: (name, address, value, type, lifetime, and scope). Although this may seem too complicated for such an apparently simple concept, it provides the clearest way to explain the various aspects of variables.

Our discussion of variable attributes will lead to examinations of the important related concepts of aliases, binding, binding times, declarations, scoping rules, and referencing environments.

The name, address, type, and value attributes of variables are discussed in the following subsections. The lifetime and scope attributes are discussed in Sections 5.4.3 and 5.5, respectively.

5.3.1 Name Variable names are the most common names in programs. They were discussed at length in Section 5.2 in the general context of entity names in programs. Most variables have names. The ones that do not are discussed in Section 5.4.3.3.

5.3.2 Address The address of a variable is the machine memory address with which it is associated. This association is not as simple as it may appear. In many languages, it is possible for the same variable to be associated with different addresses at different times during the execution of the program. For example, if a subprogram has a local variable that is allocated from the run- time stack when the subprogram is called, different calls may result in that variable having different addresses. These are in a sense different instantiations of the same variable.

The process of associating variables with addresses is further discussed in Section 5.4.3. An implementation model for subprograms and their activations is discussed in Chapter 10.

The address of a variable is sometimes called its l-value, because the address is what is required when the name of a variable appears in the left side of an assignment.

It is possible to have multiple variables that have the same address. When more than one variable name can be used to access the same memory location, the variables are called aliases. Aliasing is a hindrance to readability because it allows a variable to have its value changed by an assignment to a different variable. For example, if variables named total and sum are aliases, any change to the value of total also changes the value of sum and vice versa. A reader of the program must always remember that total and sum are different names for the same memory cell. Because there can be any number of aliases in a program, this can be very difficult in practice. Aliasing also makes program verification more difficult.

Aliases can be created in programs in several different ways. One common way in C and C++ is with their union types. Unions are discussed at length in Chapter 6.

Two pointer variables are aliases when they point to the same memory location. The same is true for reference variables. This kind of aliasing is

simply a side effect of the nature of pointers and references. When a C++ pointer is set to point at a named variable, the pointer, when dereferenced, and the variable’s name are aliases.

Aliasing can be created in many languages through subprogram parameters. These kinds of aliases are discussed in Chapter 9.

The time when a variable becomes associated with an address is very important to an understanding of programming languages. This subject is discussed in Section 5.4.3.

5.3.3 Type The type of a variable determines the range of values the variable can store and the set of operations that are defined for values of the type. For example, the int type in Java specifies a value range of −2147483648 to 2147483647 and arithmetic operations for addition, subtraction, multiplication, division, and modulus.

5.3.4 Value The value of a variable is the contents of the memory cell or cells associated with the variable. It is convenient to think of computer memory in terms of abstract cells, rather than physical cells. The physical cells, or individually addressable units, of most contemporary computer memories are eight-bit units called bytes. A byte size is too small for most program variables. An abstract memory cell has the size required by the variable with which it is associated. For example, although floating-point values may occupy four physical bytes in a particular implementation of a particular language, a floating-point value is thought of as occupying a single abstract memory cell. The value of each simple nonstructured type is considered to occupy a single abstract cell. Henceforth, the term memory cell will mean abstract memory cell.

A variable’s value is sometimes called its r-value because it is what is

required when the name of the variable appears in the right side of an assignment statement. To access the r-value, the l-value must be determined first. Such determinations are not always simple. For example, scoping rules can greatly complicate matters, as is discussed in Section 5.5.

5.4 The Concept of Binding A binding is an association between an attribute and an entity, such as between a variable and its type or value, or between an operation and a symbol. The time at which a binding takes place is called binding time. Binding and binding times are prominent concepts in the semantics of programming languages. Bindings can take place at language design time, language implementation time, compile time, load time, link time, or run time. For example, the asterisk symbol (*) is usually bound to the multiplication operation at language design time. A data type, such as int in C, is bound to a range of possible values at language implementation time. At compile time, a variable in a Java program is bound to a particular data type. A variable may be bound to a storage cell when the program is loaded into memory. That same binding does not happen until run time in some cases, as with variables declared in Java methods. A call to a library subprogram is bound to the subprogram code at link time.

Consider the following C++ assignment statement:

count = count + 5;

Some of the bindings and their binding times for the parts of this assignment statement are as follows:

The type of count is bound at compile time.

The set of possible values of count is bound at compiler design time.

The meaning of the operator symbol + is bound at compile time, when the types of its operands have been determined.

The internal representation of the literal 5 is bound at compiler design time.

The value of count is bound at execution time with this statement.

A complete understanding of the binding times for the attributes of program entities is a prerequisite for understanding the semantics of a programming language. For example, to understand what a subprogram does, one must understand how the actual parameters in a call are bound to the formal parameters in its definition. To determine the current value of a variable, it may be necessary to know when the variable was bound to storage and with which statement or statements.

5.4.1 Binding of Attributes to Variables A binding is static if it first occurs before run time begins and remains unchanged throughout program execution. If the binding first occurs during run time or can change in the course of program execution, it is called dynamic. The physical binding of a variable to a storage cell in a virtual memory environment is complex, because the page or segment of the address space in which the cell resides may be moved in and out of memory many times during program execution. In a sense, such variables are bound and unbound repeatedly. These bindings, however, are maintained by computer hardware, and the changes are invisible to the program and the user. Because they are not important to the discussion, we are not concerned with these hardware bindings. The essential point is to distinguish between static and dynamic bindings.

5.4.2 Type Bindings Before a variable can be referenced in a program, it must be bound to a data type. The two important aspects of this binding are how the type is specified and when the binding takes place. Types can be specified statically through some form of explicit or implicit declaration.

5.4.2.1 Static Type Binding

An explicit declaration is a statement in a program that lists variable names and specifies that they are a particular type. An implicit declaration is a means of associating variables with types through default conventions, rather than declaration statements. In this case, the first appearance of a variable name in a program constitutes its implicit declaration. Both explicit and implicit declarations create static bindings to types.

Most widely used programming languages that use static type binding exclusively and were designed since the mid-1960s require explicit declarations of all variables (Visual Basic, ML, C#, and Swift are some exceptions).

Implicit variable type binding is done by the language processor, either a compiler or an interpreter. There are several different bases for implicit variable type bindings. The simplest of these is naming conventions. In this case, the compiler or interpreter binds a variable to a type based on the syntactic form of the variable’s name.

Although they are a minor convenience to programmers, implicit declarations can be detrimental to reliability because they prevent the compilation process from detecting some typographical and programmer errors.

Some of the problems with implicit declarations can be avoided by requiring names for specific types to begin with particular special characters. For example, in Perl any name that begins with $ is a scalar, which can store either a string or a numeric value. If a name begins with @, it is an array; if it begins with a %, it is a hash structure.3 This creates different namespaces for different type variables. In this scenario, the names @apple and %apple are unrelated, because each is from a different namespace. Furthermore, a program reader always knows the type of a variable when reading its name.

3. Both arrays and hashes are considered types—both can store any scalar in their elements.

Another kind of implicit type declarations uses context. This is sometimes called type inference. In the simpler case, the context is the type of the value assigned to the variable in a declaration statement. For example, in C# a var declaration of a variable must include an initial value, whose type is taken as

the type of the variable. Consider the following declarations:

var sum = 0;

var total = 0.0;

var name = "Fred";

The types of sum, total, and name are int, float, and string, respectively. Keep in mind that these are statically typed variables—their types are fixed for the lifetime of the unit in which they are declared.

Visual Basic, Swift, and the functional languages ML, Haskell, OCaml, and F# also use type inferencing.

5.4.2.2 Dynamic Type Binding With dynamic type binding, the type of a variable is not specified by a declaration statement, nor can it be determined by the spelling of its name. Instead, the variable is bound to a type when it is assigned a value in an assignment statement. When the assignment statement is executed, the variable being assigned is bound to the type of the value of the expression on the right side of the assignment. Such an assignment may also bind the variable to an address and a memory cell, because different type values may require different amounts of storage. Any variable can be assigned any type value. Furthermore, a variable’s type can change any number of times during program execution. It is important to realize that the type of a variable whose type is dynamically bound may be temporary.

When the type of a variable is statically bound, the name of the variable can be thought of being bound to a type, in the sense that the type and name of a variable are simultaneously bound. However, when a variable’s type is dynamically bound, its name can be thought of as being only temporarily bound to a type. In reality, the names of variables are never bound to types. Names can be bound to variables and variables can be bound to types.

Languages in which types are dynamically bound are dramatically different from those in which types are statically bound. The primary advantage of dynamic binding of variables to types is that it provides more programming

flexibility. For example, a program to process numeric data in a language that uses dynamic type binding can be written as a generic program, meaning that it is capable of dealing with data of any numeric type. Whatever type data is input will be acceptable, because the variables in which the data are to be stored can be bound to the correct type when the data is assigned to the variables after input. By contrast, because of static binding of types, one cannot write a C program to process data without knowing the type of that data.

Before the mid-1990s, the most commonly used programming languages used static type binding, the primary exceptions being some functional languages such as Lisp. However, since then there has been a significant shift to languages that use dynamic type binding. In Python, Ruby, JavaScript, and PHP, type binding is dynamic. For example, a JavaScript script may contain the following statement:

list = [10.2, 3.5];

Regardless of the previous type of the variable named list, this assignment causes it to become the name of a single-dimensioned array of length 2. If the statement

list = 47;

followed the previous example assignment, list would become the name of a scalar variable.

The option of dynamic type binding was included in C# 2010. A variable can be declared to use dynamic type binding by including the dynamic reserved word in its declaration, as in the following example:

dynamic any;

This is similar, although also different from declaring any to have type object. It is similar in that any can be assigned a value of any type, just as if it were declared object. It is different in that it is not useful for several different situations of interoperation; for example, with dynamically typed languages such as IronPython and IronRuby (.NET versions of Python and Ruby, respectively). However, it is useful when data of unknown type come

into a program from an external source. Class members, properties, method parameters, method return values, and local variables can all be declared dynamic.

In pure object-oriented languages—for example, Ruby—all variables are references and do not have types; all data are objects and any variable can reference any object. Variables in such languages are, in a sense, all the same type—they are references. However, unlike the references in Java, which are restricted to referencing one specific type of value, variables in Ruby can reference any object.

There are two disadvantages to dynamic type binding. First, it causes programs to be less reliable, because the error-detection capability of the compiler is diminished relative to a compiler for a language with static type bindings. Dynamic type binding allows any variable to be assigned a value of any type. Incorrect types of right sides of assignments are not detected as errors; rather, the type of the left side is simply changed to the incorrect type. For example, suppose that in a particular JavaScript program, i and x are currently the names of scalar numeric variables and y is currently the name of an array. Furthermore, suppose that the program needs the assignment statement

i = x;

but because of a keying error, it has the assignment statement

i = y;

In JavaScript (or any other language that uses dynamic type binding), no error is detected in this statement by the interpreter—the type of the variable named i is simply changed to an array. But later uses of i will expect it to be a scalar, and correct results will be impossible. In a language with static type binding, such as Java, the compiler would detect the error in the assignment i = y, and the program would not get to execution.

Note that this disadvantage is also present to some extent in some languages that use static type binding, such as C and C++, which in many cases automatically convert the type of the RHS of an assignment to the type of the

LHS.

Perhaps the greatest disadvantage of dynamic type binding is cost. The cost of implementing dynamic attribute binding is considerable, particularly in execution time. Type checking must be done at run time. Furthermore, every variable must have a run-time descriptor associated with it to maintain the current type. The storage used for the value of a variable must be of varying size, because different type values require different amounts of storage.

Finally, languages that have dynamic type binding for variables are usually implemented using pure interpreters rather than compilers. Computers do not have instructions whose operand types are not known at compile time. Therefore, a compiler cannot build machine instructions for the expression a + b if the types of a and b are not known at compile time. Pure interpretation typically takes at least 10 times as long as it does to execute equivalent machine code. Of course, if a language is implemented with a pure interpreter, the time to do dynamic type binding is hidden by the overall time of interpretation, so it seems less costly in that environment. On the other hand, languages with static type bindings are seldom implemented by pure interpretation, because programs in these languages can be easily translated to very efficient machine code versions.

5.4.3 Storage Bindings and Lifetime The fundamental character of an imperative programming language is in large part determined by the design of the storage bindings for its variables. It is therefore important to have a clear understanding of these bindings.

The memory cell to which a variable is bound somehow must be taken from a pool of available memory. This process is called allocation. Deallocation is the process of placing a memory cell that has been unbound from a variable back into the pool of available memory.

The lifetime of a variable is the time during which the variable is bound to a

specific memory location. So, the lifetime of a variable begins when it is bound to a specific cell and ends when it is unbound from that cell. To investigate storage bindings of variables, it is convenient to separate scalar (unstructured) variables into four categories, according to their lifetimes. These categories are named static, stack-dynamic, explicit heap-dynamic, and implicit heap-dynamic. In the following sections, we discuss the definitions of these four categories, along with their purposes, advantages, and disadvantages.

5.4.3.1 Static Variables A static variable is one that is bound to a memory cell before program execution begins and remains bound to that same memory cell until program execution terminates. Statically bound variables have several valuable applications in programming. Globally accessible variables are often used throughout the execution of a program, thus making it necessary to have them bound to the same storage during that execution. Sometimes it is ​convenient to have subprograms that are history sensitive. Such a subprogram must have local static variables.

One advantage of static variables is efficiency. All addressing of static variables can be direct;4 other kinds of variables often require indirect addressing, which is slower. Also, no run-time overhead is incurred for allocation and deallocation of static variables, although this time is often negligible.

4. In some implementations, static variables are addressed through a base register, making accesses to them as costly as for stack-allocated variables.

One disadvantage of static binding to storage is reduced flexibility; in particular, a language that has only static variables cannot support recursive subprograms. Another disadvantage is that storage cannot be shared among variables. For example, suppose a program has two subprograms, both of which require large arrays. Furthermore, suppose that the two subprograms are never active at the same time. If the arrays are static, they cannot share the same storage.

C and C++ allow programmers to include the static specifier on a variable definition in a function, making the variables it defines static. Note that when the static modifier appears in the declaration of a variable in a class definition in C++, Java, and C#, it also implies that the variable is a class variable, rather than an instance variable. Class variables are created statically some time before the class is first instantiated.

5.4.3.2 Stack-Dynamic Variables Stack-dynamic variables are those whose storage bindings are created when their declaration statements are elaborated, but whose types are statically bound. Elaboration of such a declaration refers to the storage allocation and binding process indicated by the declaration, which takes place when execution reaches the code to which the declaration is attached. Therefore, elaboration occurs during run time. For example, the variable declarations that appear at the beginning of a Java method are elaborated when the method is called and the variables defined by those declarations are deallocated when the method completes its execution.

As their name indicates, stack-dynamic variables are allocated from the run- time stack.

Some languages—for example, C++ and Java—allow variable declarations to occur anywhere a statement can appear. In some implementations of these languages, all of the stack-dynamic variables declared in a function or method (not including those declared in nested blocks) may be bound to storage at the beginning of execution of the function or method, even though the declarations of some of these variables do not appear at the beginning. In such cases, the variable becomes visible at the declaration, but the storage binding (and initialization, if it is specified in the declaration) occurs when the function or method begins execution. The fact that storage binding of a variable takes place before it becomes visible does not affect the semantics of the language.

The advantages of stack-dynamic variables are as follows: To be useful, at least in most cases, recursive subprograms require some form of dynamic

local storage so that each active copy of the recursive subprogram has its own version of the local variables. These needs are conveniently met by stack- dynamic variables. Even in the absence of recursion, having stack-dynamic local storage for subprograms is not without merit, because all subprograms share the same memory space for their locals.

The disadvantages, relative to static variables, of stack-dynamic variables are the run-time overhead of allocation and deallocation, possibly slower accesses because indirect addressing is required, and the fact that subprograms cannot be history sensitive. The time required to allocate and deallocate stack-dynamic variables is not significant, because all of the stack- dynamic variables that are declared at the beginning of a subprogram are allocated and deallocated together, rather than by separate operations.

In Java, C++, and C#, variables defined in methods are by default stack dynamic.

All attributes other than storage are statically bound to stack-dynamic scalar variables. That is not the case for some structured types, as is discussed in Chapter 6. Implementation of allocation/deallocation processes for stack-​- dynamic variables is discussed in Chapter 10.

5.4.3.3 Explicit Heap-Dynamic Variables Explicit heap-dynamic variables are nameless (abstract) memory cells that are allocated and deallocated by explicit run-time instructions written by the programmer. These variables, which are allocated from and deallocated to the heap, can only be referenced through pointer or reference variables. The heap is a collection of storage cells whose organization is highly disorganized due to the unpredictability of its use. The pointer or reference variable that is used to access an explicit heap-dynamic variable is created as any other scalar variable. An explicit heap-dynamic variable is created by either an operator (for example, in C++) or a call to a system subprogram provided for that purpose (for example, in C).

In C++, the allocation operator, named new, uses a type name as its operand. When executed, an explicit heap-dynamic variable of the operand type is created and its address is returned. Because an explicit heap-dynamic variable is bound to a type at compile time, that binding is static. However, such variables are bound to storage at the time they are created, which is during run time.

In addition to a subprogram or operator for creating explicit heap-​dynamic variables, some languages include a subprogram or operator for explicitly destroying them.

As an example of explicit heap-dynamic variables, consider the following C++ code segment:

int *intnode; // Create a pointer

intnode = new int; // Create the heap-dynamic variable

. . .

delete intnode; // Deallocate the heap-dynamic variable

// to which intnode points

In this example, an explicit heap-dynamic variable of int type is created by the new operator. This variable can then be referenced through the pointer, intnode. Later, the variable is deallocated by the delete operator. C++ requires the explicit deallocation operator delete, because it does not use implicit storage reclamation, such as garbage collection.

In Java, all data except the primitive scalars are objects. Java objects are explicitly heap dynamic and are accessed through reference variables. Java has no way of explicitly destroying a heap-dynamic variable; rather, implicit garbage collection is used. Garbage collection is discussed in Chapter 6.

C# has both explicit heap-dynamic and stack-dynamic objects, all of which are implicitly deallocated. In addition, C# supports C++-style pointers. Such pointers are used to reference heap, stack, and even static variables and objects. These pointers have the same dangers as those of C++, and the objects they reference on the heap are not implicitly deallocated. Pointers are included in C# to allow C# components to interoperate with C and C++ components. To discourage their use, and also to make clear to any program reader that the code uses pointers, the header of any method that defines a

pointer must include the reserved word unsafe.

Explicit heap-dynamic variables are often used to construct dynamic structures, such as linked lists and trees, that need to grow and/or shrink during execution. Such structures can be built conveniently using pointers or references and explicit heap-dynamic variables.

The disadvantages of explicit heap-dynamic variables are the difficulty of using pointer and reference variables correctly, the cost of references to the variables, and the complexity of the required storage management implementation. This is essentially the problem of heap management, which is costly and complicated. Implementation methods for explicit heap-dynamic variables are discussed at length in Chapter 6.

5.4.3.4 Implicit Heap-Dynamic Variables Implicit heap-dynamic variables are bound to heap storage only when they are assigned values. In fact, all their attributes are bound every time they are assigned. For example, consider the following JavaScript assignment statement:

highs = [74, 84, 86, 90, 71];

Regardless of whether the variable named highs was previously used in the program or what it was used for, it is now an array of five numeric values.

The advantage of such variables is that they have the highest degree of flexibility, allowing highly generic code to be written. One disadvantage of implicit heap-dynamic variables is the run-time overhead of maintaining all the dynamic attributes, which could include array subscript types and ranges, among others. Another disadvantage is the loss of some error detection by the compiler, as discussed in Section 5.4.2.2.

5.5 Scope One of the important factors in understanding variables is scope. The scope of a variable is the range of statements in which the variable is visible. A variable is visible in a statement if it can be referenced or assigned in that statement.

The scope rules of a language determine how a particular occurrence of a name is associated with a variable, or in the case of a functional language, how a name is associated with an expression. In particular, scope rules determine how references to variables declared outside the currently executing subprogram or block are associated with their declarations and thus their attributes (blocks are discussed in Section 5.5.2). A clear understanding of these rules for a language is therefore essential to the ability to write or read programs in that language.

A variable is local in a program unit or block if it is declared there. The nonlocal variables of a program unit or block are those that are visible within the program unit or block but are not declared there. Global variables are a special category of nonlocal variables, which are discussed in Section 5.5.4.

Scoping issues of classes, packages, and namespaces are discussed in ​ Chapter 11.

5.5.1 Static Scope ALGOL 60 introduced the method of binding names to nonlocal variables called static scoping,5 which has been copied by many subsequent imperative languages and many nonimperative languages as well. Static scoping is so named because the scope of a variable can be statically determined—that is, prior to execution. This permits a human program reader (and a compiler) to determine the type of every variable in the program simply by examining its source code.

5. Static scoping is sometimes called lexical scoping.

There are two categories of static-scoped languages: those in which subprograms can be nested, which creates nested static scopes, and those in which subprograms cannot be nested. In the latter category, static scopes are also created by subprograms but nested scopes are created only by nested class definitions and blocks.

Ada, JavaScript, Common Lisp, Scheme, Fortran 2003+, F#, and Python allow nested subprograms, but the C-based languages do not.

Our discussion of static scoping in this section focuses on those languages that allow nested subprograms. Initially, we assume that all scopes are associated with program units and that all referenced nonlocal variables are declared in other program units.6 In this chapter, it is assumed that scoping is the only method of accessing nonlocal variables in the languages under discussion. This is not true for all languages. It is not even true for all languages that use static scoping, but the assumption simplifies the discussion here.

6. Nonlocal variables not defined in other program units are discussed in Section 5.5.4.

When the reader of a program finds a reference to a variable, the attributes of the variable can be determined by finding the statement in which it is declared (either explicitly or implicitly). In static-scoped languages with nested subprograms, this process can be thought of in the following way. Suppose a reference is made to a variable x in subprogram sub1. The correct declaration is found by first searching the declarations of subprogram sub1. If no declaration is found for the variable there, the search continues in the declarations of the subprogram that declared subprogram sub1, which is called its static parent. If a declaration of x is not found there, the search continues to the next-larger enclosing unit (the unit that declared sub1’s parent), and so forth, until a declaration for x is found or the largest unit’s declarations have been searched without success. In that case, an undeclared variable error is reported. The static parent of subprogram sub1, and its static parent, and so forth up to and including the largest enclosing subprogram, are called the static ancestors of sub1. Actual implementation techniques for

static scoping, which are discussed in Chapter 10, are usually much more efficient than the process just described.

Consider the following JavaScript function, big, in which the two functions sub1 and sub2 are nested:

function big() {

function sub1() {

var x = 7;

sub2();

}

function sub2() {

var y = x;

}

var x = 3;

sub1();

}

Under static scoping, the reference to the variable x in sub2 is to the x declared in the procedure big. This is true because the search for x begins in the procedure in which the reference occurs, sub2, but no declaration for x is found there. The search continues in the static parent of sub2, big, where the declaration of x is found. The x declared in sub1 is ignored, because it is not in the static ancestry of sub2.

In some languages that use static scoping, regardless of whether nested subprograms are allowed, some variable declarations can be hidden from some other code segments. For example, consider again the JavaScript function big. The variable x is declared in both big and in sub1, which is nested inside big. Within sub1, every simple reference to x is to the local x. Therefore, the outer x is hidden from sub1.

5.5.2 Blocks Many languages allow new static scopes to be defined in the midst of executable code. This powerful concept, introduced in ALGOL 60, allows a section of code to have its own local variables whose scope is minimized. Such variables are typically stack dynamic, so their storage is allocated when the section is entered and deallocated when the section is exited. Such a

section of code is called a block. Blocks provide the origin of the phrase block-structured language.

The C-based languages allow any compound statement (a statement sequence surrounded by matched braces) to have declarations and thereby define a new scope. Such compound statements are called blocks. For example, if list were an integer array, one could write the following:

if (list[i] < list[j]) {

int temp;

temp = list[i];

list[i] = list[j];

list[j] = temp;

}

The scopes created by blocks, which could be nested in larger blocks, are treated exactly like those created by subprograms. References to variables in a block that are not declared there are connected to declarations by searching enclosing scopes (blocks or subprograms) in order of increasing size.

Consider the following skeletal C function:

void sub() {

int count;

. . .

while (. . .) {

int count;

count++;

. . .

}

. . .

}

The reference to count in the while loop is to that loop’s local count. In this case, the count of sub is hidden from the code inside the while loop. In general, a declaration for a variable effectively hides any declaration of a variable with the same name in a larger enclosing scope.7 Note that this code is legal in C and C++ but illegal in Java and C#. The designers of Java and C# believed that the reuse of names in nested blocks was too error prone to be allowed.

7. As discussed in Section 5.5.4, in C++, such hidden global variables can be accessed in the inner scope using the scope operator (::).

Although JavaScript uses static scoping for its nested functions, nonfunction blocks cannot be defined in the language.

Most functional programming languages include a construct that is related to the blocks of the imperative languages, usually named let. These constructs have two parts, the first of which is to bind names to values, usually specified as expressions. The second part is an expression that uses the names defined in the first part. Programs in functional languages are comprised of expressions, rather than statements. Therefore, the final part of a let construct is an expression, rather than a statement. In Scheme, a let construct is a call to the function LET with the following form:

(LET (

 ( name1 expression1)

 . . .

 ( namen expressionn))

 expression

)

The semantics of the call to LET is as follows: The first n expressions are evaluated and the values are assigned to the associated names. Then, the final expression is evaluated and the return value of LET is that value. This differs from a block in an imperative language in that the names are of values; they are not variables in the imperative sense. Once set, they cannot be changed. However, they are like local variables in a block in an imperative language in that their scope is local to the call to LET. Consider the following call to LET:

(LET (

(top (+ a b))

(bottom (- c d)))

(/ top bottom)

)

This call computes and returns the value of the expression (a + b) / (c - d).

In ML, the form of a let construct is as follows:

let

 val name1 = expression1

 . . .

 val namen = expressionn

in

 expression

end;

Each val statement binds a name to an expression. As with Scheme, the names in the first part are like the named constants of imperative languages; once set, they cannot be changed.8 Consider the following let construct:

8. In Chapter 15, we will see that they can be reset, but that the process actually creates a new name.

let

val top = a + b

val bottom = c - d

in

top / bottom

end;

The general form of a let construct in F# is as follows:

let left_side = expression

The left_side of let can be a name or a tuple pattern (a sequence of names separated by commas).

The scope of a name defined with let inside a function definition is from the end of the defining expression to the end of the function. The scope of let can be limited by indenting the following code, which creates a new local scope. Although any indentation will work, the convention is that the indentation is four spaces. Consider the following code:

let n1 =

let n2 = 7

let n3 = n2 + 3

n3;;

let n4 = n3 + n1;;

The scope of n1 extends over all of the code. However, the scope of n2 and n3 ends when the indentation ends. So, the use of n3 in the last let causes an error. The last line of the let n1 scope is the value bound to n1; it could be any expression.

Chapter 15 includes more details of the let constructs in Scheme, ML, Haskell, and F#.

5.5.3 Declaration Order In C89, as well as in some other languages, all data declarations in a function except those in nested blocks must appear at the beginning of the function. However, some languages—for example, C99, C++, Java, JavaScript, and C# —allow variable declarations to appear anywhere a statement can appear in a program unit. Declarations may create scopes that are not associated with compound statements or subprograms. For example, in C99, C++, and Java, the scope of all local variables is from their declarations to the ends of the blocks in which those declarations appear.

In the official documentation for C#, the scope of any variable declared in a block is said to be the whole block, regardless of the position of the declaration in the block, as long as it is not in a nested block. The same is true for methods. However, this is misleading, because the C# language definition requires that all variables be declared before they are used. Therefore, although the scope of a variable is said to extend from the declaration to the

top of the block or subprogram in which that declaration appears, the variable still cannot be used above its declaration.

Recall that C# does not allow the declaration of a variable in a nested block to have the same name as a variable in a nesting scope. This, together with the rule that the scope of a declaration is the whole block, makes the following nested declaration of x illegal:

{

{int x; // Illegal

...

}

int x;

}

Note that C# still requires that all be declared before they are used. Therefore, although the scope of a variable extends from the declaration to the top of the block or subprogram in which that declaration appears, the variable still cannot be used above its declaration.

In JavaScript, local variables can be declared anywhere in a function, but the scope of such a variable is always the entire function. If used before its declaration in the function, such a variable has the value undefined. The reference is not illegal.

The for statements of C++, Java, and C# allow variable definitions in their initialization expressions. In early versions of C++, the scope of such a variable was from its definition to the end of the smallest enclosing block. In the standard version, however, the scope is restricted to the for construct, as is the case with Java and C#. Consider the following skeletal method:

void fun() {

. . .

for (int count = 0; count < 10; count++){

. . .

}

. . .

}

In later versions of C++, as well as in Java and C#, the scope of count is from the for statement to the end of its body (the right brace).

5.5.4 Global Scope Some languages, including C, C++, PHP, JavaScript, and Python, allow a program structure that is a sequence of function definitions, in which variable definitions can appear outside the functions. Definitions outside functions in a file create global variables, which potentially can be visible to those functions.

C and C++ have both declarations and definitions of global data. Declarations specify types and other attributes but do not cause allocation of storage. Definitions specify attributes and cause storage allocation. For a specific global name, a C program can have any number of compatible declarations, but only a single definition.

A declaration of a variable outside function definitions specifies that the variable is defined in a different file. A global variable in C is implicitly visible in all subsequent functions in the file, except those that include a declaration of a local variable with the same name. A global variable that is defined after a function can be made visible in the function by declaring it to be external, as in the following:

extern int sum;

In C99, definitions of global variables usually have initial values. Declarations of global variables never have initial values. If the declaration is outside function definitions, it need not include the extern qualifier.

This idea of declarations and definitions carries over to the functions of C and C++, where prototypes declare names and interfaces of functions but do not provide their code. Function definitions, on the other hand, are complete.

In C++, a global variable that is hidden by a local with the same name can be accessed using the scope operator (::). For example, if x is a global that is hidden in a function by a local named x, the global could be referenced as ::x.

PHP statements can be interspersed with function definitions. Variables in

PHP are implicitly declared when they appear in statements. Any variable that is implicitly declared outside any function is a global variable; variables implicitly declared in functions are local variables. The scope of global variables extends from their declarations to the end of the program but skips over any subsequent function definitions. So, global variables are not implicitly visible in any function. Global variables can be made visible in functions in their scope in two ways: (1) If the function includes a local variable with the same name as a global, that global can be accessed through the $GLOBALS array, using the name of the global as a string literal subscript, and (2) if there is no local variable in the function with the same name as the global, the global can be made visible by including it in a global declaration statement. Consider the following example:

$day = "Monday";

$month = "January";

function calendar() {

$day = "Tuesday";

global $month;

print "local day is $day ";

$gday = $GLOBALS['day'];

print "global day is $gday <br \>";

print "global month is $month ";

}

calendar();

Interpretation of this code produces the following:

local day is Tuesday

global day is Monday

global month is January

The global variables of JavaScript are very similar to those of PHP, except that there is no way to access a global variable in a function that has declared a local variable with the same name.

The visibility rules for global variables in Python are unusual. Variables are not normally declared, as in PHP. They are implicitly declared when they appear as the targets of assignment statements. A global variable can be referenced in a function, but a global variable can be assigned in a function only if it has been declared to be global in the function. Consider the following examples:

day = "Monday"

def tester():

print "The global day is:", day

tester()

The output of this script, because globals can be referenced directly in ​- functions, is as follows:

The global day is: Monday

The following script attempts to assign a new value to the global day:

day = "Monday"

def tester():

print "The global day is:", day

day = "Tuesday"

print "The new value of day is:", day

tester()

This script creates an UnboundLocalError error message, because the assignment to day in the second line of the body of the function makes day a local variable, which makes the reference to day in the first line of the body of the function an illegal forward reference to the local.

The assignment to day can be to the global variable if day is declared to be global at the beginning of the function. This prevents the assignment to day from creating a local variable. This is shown in the following script:

day = "Monday"

def tester():

global day

print "The global day is:", day

day = "Tuesday"

print "The new value of day is:", day

tester()

The output of this script is as follows:

The global day is: Monday

The new value of day is: Tuesday

Functions can be nested in Python. Variables defined in nesting functions are accessible in a nested function through static scoping, but such variables must

be declared nonlocal in the nested function.9 An example skeletal ​program in Section 5.7 illustrates accesses to nonlocal variables.

9. The nonlocal reserved word was introduced in Python 3.

All names defined outside function definitions in F# are globals. Their scope extends from their definitions to the end of the file.

Declaration order and global variables are also issues in the class and member declarations in object-oriented languages. These are discussed in Chapter 12.

5.5.5 Evaluation of Static Scoping Static scoping provides a method of nonlocal access that works well in many situations. However, it is not without its problems. First, in most cases it allows more access to both variables and subprograms than is necessary. It is simply too crude a tool for concisely specifying such restrictions. Second, and perhaps more important, is a problem related to program evolution. Software is highly dynamic—programs that are used regularly continually change. These changes often result in restructuring, thereby destroying the initial structure that restricted variable and subprogram access in a static- scoped language. To avoid the complexity of maintaining these access restrictions, developers often discard structure when it gets in the way. Thus, getting around the restrictions of static scoping can lead to program designs that bear little resemblance to the original, even in areas of the program in which changes have not been made. Designers are encouraged to use far more globals than are necessary. All subprograms can end up being nested at the same level, in the main program, using globals instead of deeper levels of nesting.10 Moreover, the final design may be awkward and contrived, and it may not reflect the underlying conceptual design. These and other defects of static scoping are discussed in detail in Clarke, Wileden, and Wolf (1980). An alternative to the use of static scoping to control access to variables and subprograms is an encapsulation construct, which is included in many newer languages. Encapsulation constructs are ​discussed in Chapter 11.

10. Sounds like the structure of a C program, doesn’t it?

5.5.6 Dynamic Scope The scope of variables in APL, SNOBOL4, and the early versions of Lisp is dynamic. Perl and Common Lisp also allow variables to be declared to have dynamic scope, although the default scoping mechanism in these languages is static. Dynamic scoping is based on the calling sequence of subprograms, not on their spatial relationship to each other. Thus, the scope can be determined only at run time.

Consider again the function big from Section 5.5.1, which is reproduced here, minus the function calls:

function big() {

function sub1() {

var x = 7;

}

function sub2() {

var y = x;

var z = 3;

}

var x = 3;

}

Assume that dynamic-scoping rules apply to nonlocal references. The meaning of the identifier x referenced in sub2 is dynamic—it cannot be determined at compile time. It may reference the variable from either declaration of x, depending on the calling sequence.

One way the correct meaning of x can be determined during execution is to begin the search with the local declarations. This is also the way the process begins with static scoping, but that is where the similarity between the two techniques ends. When the search of local declarations fails, the declarations of the dynamic parent, or calling function, are searched. If a declaration for x is not found there, the search continues in that function’s dynamic parent, and so forth, until a declaration for x is found. If none is found in any dynamic ancestor, it is a run-time error.

Consider the two different call sequences for sub2 in the earlier example. First, big calls sub1, which calls sub2. In this case, the search proceeds from

the local procedure, sub2, to its caller, sub1, where a declaration for x is found. So, the reference to x in sub2 in this case is to the x declared in sub1. Next, sub2 is called directly from big. In this case, the dynamic parent of sub2 is big, and the reference is to the x declared in big.

Note that if static scoping were used, in either calling sequence discussed, the reference to x in sub2 would be to big’s x.

Perl’s dynamic scoping is unusual—in fact, it is not exactly like that discussed in this section, although the semantics are often that of traditional dynamic scoping (see Programming Exercise 1).

5.5.7 Evaluation of Dynamic Scoping The effect of dynamic scoping on programming is profound. When dynamic scoping is used, the correct attributes of nonlocal variables visible to a program statement cannot be determined statically. Furthermore, a reference to the name of such a variable is not always to the same variable. A statement in a subprogram that contains a reference to a nonlocal variable can refer to different nonlocal variables during different executions of the subprogam. Several kinds of programming problems follow directly from dynamic scoping.

First, during the time span beginning when a subprogram begins its execution and ending when that execution ends, the local variables of the subprogram are all visible to any other executing subprogram, regardless of its textual proximity or how execution got to the currently executing subprogram. There is no way to protect local variables from this accessibility. Subprograms are always executed in the environment of all previously called subprograms that have not yet completed their executions. As a result, dynamic scoping results in less reliable programs than static scoping.

A second problem with dynamic scoping is the inability to type check references to nonlocals statically. This problem results from the inability to

statically find the declaration for a variable referenced as a nonlocal.

Dynamic scoping also makes programs much more difficult to read, because the calling sequence of subprograms must be known to determine the meaning of references to nonlocal variables. This task can be virtually impossible for a human reader.

Finally, accesses to nonlocal variables in dynamic-scoped languages take far longer than accesses to nonlocals when static scoping is used. The reason for this is explained in Chapter 10.

On the other hand, dynamic scoping is not without merit. In many cases, the parameters passed from one subprogram to another are variables that are defined in the caller. None of these needs to be passed in a dynamically scoped language, because they are implicitly visible in the called subprogram.

It is not difficult to understand why dynamic scoping is not as widely used as static scoping. Programs in static-scoped languages are easier to read, are more reliable, and execute faster than equivalent programs in dynamic- scoped languages. It was precisely for these reasons that dynamic scoping was replaced by static scoping in most current dialects of Lisp. Implementation methods for both static and dynamic scoping are discussed in Chapter 10.

5.6 Scope and Lifetime Sometimes the scope and lifetime of a variable appear to be related. For example, consider a variable that is declared in a Java method that contains no method calls. The scope of such a variable is from its declaration to the end of the method. The lifetime of that variable is the period of time beginning when the method is entered and ending when execution of the method terminates. Although the scope and lifetime of the variable are clearly not the same, because static scope is a textual, or spatial, concept whereas lifetime is a temporal concept, they at least appear to be related in this case.

This apparent relationship between scope and lifetime does not hold in other situations. In C and C++, for example, a variable that is declared in a function using the specifier static is statically bound to the scope of that function and is also statically bound to storage. So, its scope is static and local to the function, but its lifetime extends over the entire execution of the program of which it is a part.

Scope and lifetime are also unrelated when subprogram calls are involved. Consider the following C++ functions:

void printheader() {

. . .

} /* end of printheader */

void compute() {

int sum;

. . .

printheader();

} /* end of compute */

The scope of the variable sum is completely contained within the compute function. It does not extend to the body of the function printheader, although printheader executes in the midst of the execution of compute. However, the lifetime of sum extends over the time during which printheader executes. Whatever storage location sum is bound to before the call to printheader, that binding will continue during and after the execution of printheader.

5.7 Referencing Environments The referencing environment of a statement is the collection of all variables that are visible in the statement. The referencing environment of a statement in a static-scoped language is the variables declared in its local scope plus the collection of all variables of its ancestor scopes that are visible. In such a language, the referencing environment of a statement is needed while that statement is being compiled, so code and data structures can be created to allow references to variables from other scopes during run time. Techniques for implementing references to nonlocal variables in both static- and dynamic-scoped languages are discussed in Chapter 10.

In Python, scopes can be created by function definitions. The referencing environment of a statement includes the local variables, plus all of the variables declared in the functions in which the statement is nested (excluding variables in nonlocal scopes that are hidden by declarations in nearer functions). Each function definition creates a new scope and thus a new environment. Consider the following Python skeletal program:

g = 3; # A global

def sub1():

a = 5; # Creates a local

b = 7; # Creates another local

. . . <------------------------------ 1

def sub2():

global g; # Global g is now assignable here

c = 9; # Creates a new local

. . . <------------------------------ 2

def sub3():

nonlocal c: # Makes nonlocal c visible here

g = 11; # Creates a new local

. . . <------------------------------ 3

The referencing environments of the indicated program points are as follows:

Now consider the variable declarations of this skeletal program. First, note that, although the scope of sub1 is at a higher level (it is less deeply nested) than sub3, the scope of sub1 is not a static ancestor of sub3, so sub3 does not have access to the variables declared in sub1. There is a good reason for this. The variables declared in sub1 are stack dynamic, so they are not bound to storage if sub1 is not in execution. Because sub3 can be in execution when sub1 is not, it cannot be allowed to access variables in sub1, which would not necessarily be bound to storage during the execution of sub3.

A subprogram is active if its execution has begun but has not yet terminated. The referencing environment of a statement in a dynamically scoped language is the locally declared variables, plus the variables of all other ​- subprograms that are currently active. Once again, some variables in active subprograms can be hidden from the referencing environment. Recent subprogram activations can have declarations for variables that hide variables with the same names in previous subprogram activations.

Consider the following example program. Assume that the only function calls are the following: main calls sub2, which calls sub1.

void sub1() {

int a, b;

. . . <------------ 1

} /* end of sub1 */

void sub2() {

int b, c;

. . . . <------------ 2

sub1();

} /* end of sub2 */

void main() {

int c, d;

. . . <------------ 3

sub2();

} /* end of main */

The referencing environments of the indicated program points are as follows:

5.8 Named Constants A named constant is a variable that is bound to a value only once. Named constants are useful as aids to readability and program reliability. Readability can be improved, for example, using the name pi instead of the constant 3.14159265.

Another important use of named constants is to parameterize a program. For example, consider a program that processes a fixed number of data values, say 100. Such a program usually uses the constant 100 in a number of locations for declaring array subscript ranges and for loop control limits. Consider the following skeletal Java program segment:

void example() {

int[] intList = new int[100];

String[] strList = new String[100];

. . .

for (index = 0; index < 100; index++) {

. . .

}

. . .

for (index = 0; index < 100; index++) {

. . .

}

. . .

average = sum / 100;

. . .

}

When this program must be modified to deal with a different number of data values, all occurrences of 100 must be found and changed. On a large program, this can be tedious and error prone. An easier and more reliable method is to use a named constant as a program parameter, as follows:

void example() {

final int len = 100;

int[] intList = new int[len];

String[] strList = new String[len];

. . .

for (index = 0; index < len; index++) {

. . .

}

. . .

for (index = 0; index < len; index++) {

. . .

}

. . .

average = sum / len;

. . .

}

Now, when the length must be changed, only one line must be changed (the variable len), regardless of the number of times it is used in the program. This is another example of the benefits of abstraction. The name len is an abstraction for the number of elements in some arrays and the number of iterations in some loops. This illustrates how named constants can aid modifiability.

C++ allows dynamic binding of values to named constants. This allows expressions containing variables to be assigned to constants in the declarations. For example, the C++ statement

const int result = 2 * width + 1;

declares result to be an integer type named constant whose value is set to the value of the expression 2 * width + 1, where the value of the variable width must be visible when result is allocated and bound to its value.

Java also allows dynamic binding of values to named constants. In Java, named constants are defined with the final reserved word (as in the earlier example). The initial value can be given in the declaration statement or in a subsequent assignment statement. The assigned value can be specified with any expression.

C# has two kinds of named constants: those defined with const and those defined with readonly. The const named constants, which are implicitly static, are statically bound to values; that is, they are bound to values at compile time, which means those values can be specified only with literals or other const members. The readonly named constants, which are dynamically bound to values, can be assigned in the declaration or with a

static constructor.11 So, if a program needs a constant-valued object whose value is the same on every use of the program, a const constant is used. However, if a program needs a constant-valued object whose value is determined only when the object is created and can be different for different executions of the program, then a readonly constant is used.

11. Static constructors in C# run at some indeterminate time before the class is instantiated.

The discussion of binding values to named constants naturally leads to the topic of initialization, because binding a value to a named constant is the same process, except it is permanent.

In many instances, it is convenient for variables to have values before the code of the program or subprogram in which they are declared begins executing. The binding of a variable to a value at the time it is bound to storage is called initialization. If the variable is statically bound to storage, binding and initialization occur before run time. In these cases, the initial value must be specified as a literal or an expression whose only nonliteral operands are named constants that have already been defined. If the storage binding is dynamic, initialization is also dynamic and the initial values can be any expression.

In most languages, initialization is specified on the declaration that creates the variable. For example, in C++, we could have

int sum = 0;

int* ptrSum = &sum;

char name[] = "George Washington Carver";

SUMMARY Case sensitivity and the use of underscores are the design issues for names.

Variables can be characterized by the sextuple of attributes: name, address, value, type, lifetime, and scope.

Aliases are two or more variables bound to the same storage address. They are regarded as detrimental to reliability but are difficult to eliminate entirely from a language.

Binding is the association of attributes with program entities. Knowledge of the binding times of attributes to entities is essential to understanding the semantics of programming languages. Binding can be static or dynamic. ​- Declarations, either explicit or implicit, provide a means of specifying the static binding of variables to types. In general, dynamic binding allows greater flexibility but at the expense of readability, efficiency, and reliability.

Scalar variables can be separated into four categories by considering their lifetimes: static, stack dynamic, explicit heap dynamic, and implicit heap dynamic.

Static scoping is a central feature of ALGOL 60 and some of its descendants. It provides a simple, reliable, and efficient method of allowing visibility of nonlocal variables in subprograms. Dynamic scoping provides more flexibility than static scoping but, again, at the expense of readability, reliability, and efficiency.

Most functional languages allow the user to create local scopes with let constructs, which limit the scope of their defined names.

The referencing environment of a statement is the collection of all of the variables that are visible to that statement.

Named constants are simply variables that are bound to values only once.

REVIEW QUESTIONS 1. What are the design issues for names?

2. What is the potential danger of case-sensitive names?

3. What is an alias?

4. Which category of C++ reference variables always produces aliases?

5. What is the l-value of a variable? What is the r-value?

6. Define binding and binding time.

7. After language design and implementation, what are the four times bindings can take place in a program?

8. Define static binding and dynamic binding.

9. What are the advantages and disadvantages of implicit declarations?

10. What are the advantages and disadvantages of dynamic type binding?

11. Define static, stack-dynamic, explicit heap-dynamic, and implicit heap- dynamic variables. What are their advantages and disadvantages?

12. Define lifetime, scope, static scope, and dynamic scope.

13. How is a reference to a nonlocal variable in a static-scoped program connected to its definition?

14. What is the general problem with static scoping?

15. What is the referencing environment of a statement?

16. What is a static ancestor of a subprogram? What is a dynamic ancestor of a subprogram?

17. What is a block?

18. What is the purpose of the let constructs in functional languages?

19. What is the difference between the names defined in an ML let construct from the variables declared in a C block?

20. Describe the encapsulation of an F# let inside a function and outside all functions.

21. What are the advantages and disadvantages of dynamic scoping?

22. What are the advantages of named constants?

PROBLEM SET 1. Which of the following identifier forms is most readable? Support your

decision.

SumOfSales

sum_of_sales

SUMOFSALES

2. Some programming languages are typeless. What are the obvious advantages and disadvantages of having no types in a language?

3. Write a simple assignment statement with one arithmetic operator in some language you know. For each component of the statement, list the various bindings that are required to determine the semantics when the statement is executed. For each binding, indicate the binding time used for the language.

4. Dynamic type binding is closely related to implicit heap-dynamic variables. Explain this relationship.

5. Describe a situation when a history-sensitive variable in a subprogram is useful.

6. Consider the following JavaScript skeletal program:

// The main program

var x;

function sub1() {

var x;

function sub2() {

. . .

}

}

function sub3() {

. . .

}

Assume that the execution of this program is in the following unit order:

main calls sub1

sub1 calls sub2

sub2 calls sub3

1. Assuming static scoping, in the following, which declaration of x is the correct one for a reference to x?

1. sub1

2. sub2

3. sub3

2. Repeat part a, but assume dynamic scoping.

7. Assume the following JavaScript program was interpreted using static- scoping rules. What value of x is displayed in function sub1? Under dynamic-scoping rules, what value of x is displayed in function sub1?

var x;

function sub1() {

document.write("x = " + x + "");

}

function sub2() {

var x;

x = 10;

sub1();

}

x = 5;

sub2();

8. Consider the following JavaScript program:

var x, y, z;

function sub1() {

var a, y, z;

function sub2() {

var a, b, z;

. . .

}

. . .

}

function sub3() {

var a, x, w;

. . .

}

List all the variables, along with the program units where they are declared, that are visible in the bodies of sub1, sub2, and sub3, assuming static scoping is used.

9. Consider the following Python program:

x = 1;

y = 3;

z = 5;

def sub1():

a = 7;

y = 9;

z = 11;

. . .

def sub2():

global x;

a = 13;

x = 15;

w = 17;

. . .

def sub3():

nonlocal a;

a = 19;

b = 21;

z = 23;

. . .

. . .

List all the variables, along with the program units where they are declared, that are visible in the bodies of sub1, sub2, and sub3, assuming static scoping is used.

10. Consider the following C program:

void fun(void) {

int a, b, c; /* definition 1 */

. . .

while (. . .) {

int b, c, d; /*definition 2 */

. . . <------------- 1

while (. . .) {

int c, d, e; /* definition 3 */

. . . <------------- 2

}

. . . <-------------- 3

}

. . . <---------------- 4

}

For each of the four marked points in this function, list each visible ​- variable, along with the number of the definition statement that defines it.

11. Consider the following skeletal C program:

void fun1(void); /* prototype */

void fun2(void); /* prototype */

void fun3(void); /* prototype */

void main() {

int a, b, c;

. . .

}

void fun1(void) {

int b, c, d;

. . .

}

void fun2(void) {

int c, d, e;

. . .

}

void fun3(void) {

int d, e, f;

. . .

}

Given the following calling sequences and assuming that dynamic scoping is used, what variables are visible during execution of the last function called? Include with each visible variable the name of the function in which it was defined.

1. main calls fun1; fun1 calls fun2; fun2 calls fun3.

2. main calls fun1; fun1 calls fun3.

3. main calls fun2; fun2 calls fun3; fun3 calls fun1.

4. main calls fun3; fun3 calls fun1.

5. main calls fun1; fun1 calls fun3; fun3 calls fun2.

6. main calls fun3; fun3 calls fun2; fun2 calls fun1.

12. Consider the following program, written in JavaScript-like syntax:

// main program

var x, y, z;

function sub1() {

var a, y, z;

. . .

}

function sub2() {

var a, b, z;

. . .

}

function sub3() {

var a, x, w;

. . .

}

Given the following calling sequences and assuming that dynamic scoping is used, what variables are visible during execution of the last subprogram activated? Include with each visible variable the name of the unit where it is declared.

1. main calls sub1; sub1 calls sub2; sub2 calls sub3.

2. main calls sub1; sub1 calls sub3.

3. main calls sub2; sub2 calls sub3; sub3 calls sub1.

4. main calls sub3; sub3 calls sub1.

5. main calls sub1; sub1 calls sub3; sub3 calls sub2.

6. main calls sub3; sub3 calls sub2; sub2 calls sub1.

PROGRAMMING EXERCISES 1. Perl allows both static and a kind of dynamic scoping. Write a Perl

program that uses both and clearly shows the difference in effect of the two. Explain clearly the difference between the dynamic scoping described in this chapter and that implemented in Perl.

2. Write a Common Lisp program that clearly shows the difference between static and dynamic scoping.

3. Write a JavaScript script that has subprograms nested three deep and in which each nested subprogram references variables defined in all of its enclosing subprograms.

4. Repeat Programming Exercise 3 with Python.

5. Write a C99 function that includes the following sequence of statements:

x = 21;

int x;

x = 42;

Run the program and explain the results. Rewrite the same code in C++ and Java and compare the results.

6. Write test programs in C++, Java, and C# to determine the scope of a variable declared in a for statement. Specifically, the code must determine whether such a variable is visible after the body of the for statement.

7. Write three functions in C or C++: one that declares a large array statically, one that declares the same large array on the stack, and one that creates the same large array from the heap. Call each of the subprograms a large number of times (at least 100,000) and output the time required by each. Explain the results.

6 Data Types 1. 6.1 Introduction

2. 6.2 Primitive Data Types

3. 6.3 Character String Types

4. 6.4 Enumeration Types

5. 6.5 Array Types

6. 6.6 Associative Arrays

7. 6.7 Record Types

8. 6.8 Tuple Types

9. 6.9 List Types

10. 6.10 Union Types

11. 6.11 Pointer and Reference Types

12. 6.12 Optional Types

13. 6.13 Type Checking

14. 6.14 Strong Typing

15. 6.15 Type Equivalence

16. 6.16 Theory and Data Types

This chapter first introduces the concept of a data type and the characteristics of the common primitive data types. Then, the designs of enumeration and subrange types are discussed. Next, the details of structured data types—

specifically arrays, associative arrays, records, tuples, lists, and unions—are investigated. This section is followed by an in-depth look at pointers and references. The last category of types discussed are the optional types.

For each of the various categories of data types, the design issues are stated and the design choices made by the designers of some common languages are described. These designs are then evaluated.

The next three sections provide a thorough investigation of type checking, strong typing, and type equivalence rules. The last section of the chapter briefly introduces the fundamentals of the theory of data types.

Implementation methods for data types sometimes have a significant impact on their design. Therefore, implementation of the various data types is another important part of this chapter, especially arrays.

6.1 Introduction A data type defines a collection of data values and a set of predefined operations on those values. Computer programs produce results by manipulating data. An important factor in determining the ease with which they can perform this task is how well the data types available in the language being used match the objects in the real world of the problem being addressed. Therefore, it is crucial that a language supports an appropriate collection of data types and structures.

The contemporary concepts of data typing have evolved over the last 60 years. In the earliest languages, all problem space data structures had to be modeled with only a few basic language-supported data structures. For example, in pre-90 Fortrans, linked lists and binary trees were implemented with arrays.

The data structures of COBOL took the first step away from the Fortran I model by allowing programmers to specify the accuracy of decimal data values, and also by providing a structured data type for records of information. PL/I extended the capability of accuracy specification to integer and floating-point types. The designers of PL/I included many data types, with the intent of supporting a large range of applications. A better approach, introduced in ALGOL 68, is to provide a few basic types and a few flexible structure-defining operators that allow a programmer to design a data structure for each need. Clearly, this was one of the most important advances in the evolution of data type design. User-defined types also provide improved readability through the use of meaningful names for types. They allow type checking of the variables of a special category of use, which would otherwise not be possible. User-defined types also aid modifiability: A programmer can change the type of a category of variables in a program by changing a type definition statement only.

Taking the concept of a user-defined type a step further, we arrive at abstract data types, which are supported by most programming languages designed since the mid-1980s. The fundamental idea of an abstract data type is that the

interface of a type, which is visible to the user, is separated from the representation and set of operations on values of that type, which are hidden from the user. All of the types provided by a high-level programming language are abstract data types. User-defined abstract data types are discussed in detail in Chapter 11.

There are a number of uses of the type system of a programming language. The most practical of these is error detection. The process and value of type checking, which is directed by the type system of the language, are discussed in Section 6.12. A second use of a type system is the assistance it provides for program modularization. This results from the cross-module type checking that ensures the consistency of the interfaces among modules. Another use of a type system is documentation. The type declarations in a program document information about its data, which provides clues about the program’s behavior.

The type system of a programming language defines how a type is associated with each expression in the language and includes its rules for type equivalence and type compatibility. Certainly, one of the most important parts of understanding the semantics of a programming language is understanding its type system.

The two most common structured (nonscalar) data types in the imperative languages are arrays and records, although the popularity of associative arrays has increased significantly in recent years. Lists have been a central part of functional programming languages since the first such language appeared in 1959 (Lisp). Over the last decade, the increasing popularity of functional programming has led to lists being added to primarily imperative languages, such as Python and C#.

The structured data types are defined with type operators, or constructors, which are used to form type expressions. For example, C uses brackets and asterisks as type operators to specify arrays and pointers.

It is convenient, both logically and concretely, to think of variables in terms of descriptors. A descriptor is the collection of the attributes of a variable. In an implementation, a descriptor is an area of memory that stores the attributes of a variable. If the attributes are all static, descriptors are required only at

compile time. These descriptors are built by the compiler, usually as a part of the symbol table, and are used during compilation. For dynamic attributes, however, part or all of the descriptor must be maintained during execution. In this case, the descriptor is used by the run-time system. In all cases, descriptors are used for type checking and building the code for the allocation and deallocation operations.

Care must be taken when using the term variable. One who uses only traditional imperative languages may think of identifiers as variables, but that can lead to confusion when considering data types. Identifiers do not have data types in some programming languages. It is wise to remember that identifiers are just one of the attributes of a variable.

The word object is often associated with the value of a variable and the space it occupies. In this book, however, we reserve object exclusively for instances of user-defined and language-defined abstract data types, rather than for the values of all program variables of predefined types. Objects are discussed in detail in Chapters 11 and 12.

In the following sections, many common data types are discussed. For most, design issues particular to the type are stated. For all, one or more example designs are described. One design issue is fundamental to all data types: What operations are provided for variables of the type, and how are they specified?

6.2 Primitive Data Types Data types that are not defined in terms of other types are called primitive data types. Nearly all programming languages provide a set of primitive data types. Some of the primitive types are merely reflections of the hardware— for example, most integer types. Others require only a little nonhardware support for their implementation.

To specify the structured types, the primitive data types of a language are used, along with one or more type constructors.

6.2.1 Numeric Types Some early programming languages only had numeric primitive types. Numeric types still play a central role among the collections of types supported by contemporary languages.

6.2.1.1 Integer The most common primitive numeric data type is integer. The hardware of many computers supports several sizes of integers. These sizes of integers, and often a few others, are supported by some programming languages. For example, Java includes four signed integer sizes: byte, short, int, and long. Some languages, for example, C++ and C#, include unsigned integer types, which are types for integer values without signs. Unsigned types are often used for binary data.

A signed integer value is represented in a computer by a string of bits, with one of the bits (typically the leftmost) representing the sign. Most integer types are supported directly by the hardware. One example of an integer type that is not supported directly by the hardware is the long integer type of Python (F# also provides such integers). Values of this type can have

unlimited length. Long integer values can be specified as literals, as in the following example:

243725839182756281923L

Integer arithmetic operations in Python that produce values too large to be represented with int type store them as long integer type values.

A negative integer could be stored in sign-magnitude notation, in which the sign bit is set to indicate negative and the remainder of the bit string represents the absolute value of the number. Sign-magnitude notation, however, does not lend itself to computer arithmetic. Most computers now use a notation called twos complement to store negative integers, which is convenient for addition and subtraction. In twos-complement notation, the representation of a negative integer is formed by taking the logical complement of the positive version of the number and adding one. Ones- complement notation is still used by some computers. In ones-complement notation, the negative of an integer is stored as the logical complement of its absolute value. Ones-complement notation has the disadvantage that it has two representations of zero. See any book on assembly language programming for details of integer representations.

6.2.1.2 Floating-Point Floating-point data types model real numbers, but the representations are only approximations for many real values. For example, neither of the fundamental numbers π or e (the base for the natural logarithms) can be correctly represented in floating-point notation. Of course, neither of these numbers can be precisely represented in any finite amount of computer memory. On most computers, floating-point numbers are stored in binary, which exacerbates the problem. For example, even the value 0.1 in decimal cannot be represented by a finite number of binary digits.1 Another problem with floating-point types is the loss of accuracy through arithmetic operations. For more information on the problems of floating-point notation, see any book on numerical analysis.

1. 0.1 in decimal is 0.0001100110011 . . . in binary.

Floating-point values are represented as fractions and exponents, a form that is borrowed from scientific notation. Older computers used a variety of different representations for floating-point values. However, most newer machines use the IEEE Floating-Point Standard 754 format. Language implementors use whatever representation is supported by the hardware. Most languages include two floating-point types, often called float and double. The float type is the standard size, usually stored in four bytes of memory. The double type is provided for situations where larger fractional parts and/or a larger range of exponents is needed. Double-precision variables usually occupy twice as much storage as float variables and provide at least twice the number of bits of fraction.

The collection of values that can be represented by a floating-point type is defined in terms of precision and range. Precision is the accuracy of the fractional part of a value, measured as the number of bits. Range is a combination of the range of fractions and, more important, the range of exponents.

Figure 6.1 shows the IEEE Floating-Point Standard 754 format for single- and double-precision representation (IEEE, 1985). Details of the IEEE formats can be found in Tanenbaum (2005).

Figure 6.1 IEEE floating-point formats: (a) single precision, (b) double precision

Figure 6.1 Full Alternative Text

6.2.1.3 Complex Some programming languages support a complex data type—for example, Fortran and Python. Complex values are represented as ordered pairs of floating-point values. In Python, the imaginary part of a complex literal is specified by following it with a j or J—for example,

(7 + 3j)

Languages that support a complex type include operations for arithmetic on complex values.

6.2.1.4 Decimal Most larger computers that are designed to support business systems applications have hardware support for decimal data types. Decimal data types store a fixed number of decimal digits, with the implied decimal point at a fixed position in the value. These are the primary data types for business data processing and are therefore essential to COBOL. C# and F# also have decimal data types.

Decimal types have the advantage of being able to precisely store decimal values, at least those within a restricted range, which cannot be done with ​- floating-point. For example, the number 0.1 (in decimal) can be exactly represented in a decimal type, but not in a floating-point type, as is noted in

Section 6.2.1.2. The disadvantages of decimal types are that the range of values is restricted because no exponents are allowed, and their representation in memory is mildly wasteful, for reasons discussed in the following paragraph.

Decimal types are stored very much like character strings, using binary codes for the decimal digits. These representations are called binary coded decimal (BCD). In some cases, they are stored one digit per byte, but in others, they are packed two digits per byte. Either way, they take more storage than binary representations. It takes at least four bits to code a decimal digit. Therefore, to store a six-digit coded decimal number requires 24 bits of memory. However, it takes only 20 bits to store the same number in binary.2 The operations on decimal values are done in hardware on machines that have such capabilities; otherwise, they are simulated in software.

2. Of course, unless a program needs to maintain a large number of large decimal values, the difference is insignificant.

6.2.2 Boolean Types Boolean types are perhaps the simplest of all types. Their range of values has only two elements: one for true and one for false. They were introduced in ALGOL 60 and have been included in most general-purpose languages designed since 1960. One popular exception is C89, in which numeric expressions are used as conditionals. In such expressions, all operands with nonzero values are considered true, and zero is considered false. Although C99 and C++ have a Boolean type, they also allow numeric expressions to be used as if they were Boolean. This is not the case in the subsequent languages, Java and C#.

Boolean types are often used to represent switches or flags in programs. Although other types, such as integers, can be used for these purposes, the use of Boolean types is more readable.

A Boolean value could be represented by a single bit, but because a single bit of memory cannot be accessed efficiently on many machines, they are often

stored in the smallest efficiently addressable cell of memory, typically a byte.

6.2.3 Character Types Character data are stored in computers as numeric codings. Traditionally, the most commonly used coding was the 8-bit code ASCII (American Standard Code for Information Interchange), which uses the values 0 to 127 to code 128 different characters. ISO 8859-1 is another 8-bit character code, but it allows 256 different characters.

Because of the globalization of business and the need for computers to communicate with other computers around the world, the ASCII character set became inadequate. In response, in 1991, the Unicode Consortium published the UCS-2 standard, a 16-bit character set. This character code is often called Unicode. Unicode includes the characters from most of the world’s natural languages. For example, Unicode includes the Cyrillic alphabet, as used in Serbia, and the Thai digits. The first 128 characters of Unicode are identical to those of ASCII. Java was the first widely used language to use the Unicode character set. Since then, it has found its way into JavaScript, Python, Perl, C#, F#, and Swift.

After 1991, the Unicode Consortium, in cooperation with the International Standards Organization (ISO), developed a 4-byte character code named UCS-4, or UTF-32, which is described in the ISO/IEC 10646 Standard, published in 2000.

To provide the means of processing codings of single characters, most programming languages include a primitive type for them. However, Python supports single characters only as character strings of length 1.

6.3 Character String Types A character string type is one in which the values consist of sequences of characters. Character string constants are used to label output, and the input and output of all kinds of data are often done in terms of strings. Of course, character strings also are an essential type for all programs that do character manipulation.

6.3.1 Design Issues The two most important design issues that are specific to character string types are the following:

Should strings be a special kind of character array or a primitive type?

Should strings have static or dynamic length?

6.3.2 Strings and Their Operations The most common string operations are assignment, catenation, substring reference, comparison, and pattern matching.

A substring reference is a reference to a substring of a given string. Substring references are discussed in the more general context of arrays, where the substring references are called slices.

In general, both assignment and comparison operations on character strings are complicated by the possibility of string operands of different lengths. For example, what happens when a longer string is assigned to a shorter string, or vice versa? Usually, simple and sensible choices are made for these situations, although programmers often have trouble remembering them.

In some languages, pattern matching is supported directly in the language. In others, it is provided by a function or class library.

If strings are not defined as a primitive type, string data is usually stored in arrays of single characters and referenced as such in the language. This is the approach taken by C and C++, which use char arrays to store character strings. These languages provide a collection of string operations through standard libraries. Many users of strings and many of the library functions use the convention that character strings are terminated with a special character, null, which is represented with zero. This is an alternative to maintaining the length of string variables. The library operations simply carry out their operations until the null character appears in the string being operated on. Library functions that produce strings often supply the null character. The character string literals that are built by the compiler also have the null character. For example, consider the following declaration:

char str[] = "apples";

In this example, str represents an array of char elements, specifically apples0, where 0 is the null character.

Some of the most commonly used library functions for character strings in C and C++ are strcpy, which moves strings; strcat, which catenates one given string onto another; strcmp, which lexicographically compares (by the order of their character codes) two given strings; and strlen, which returns the number of characters, not counting the null character, in the given string. The parameters and return values for most of the string manipulation functions are char pointers that point to arrays of char. Parameters can also be string literals.

The string manipulation functions of the C standard library, which are also available in C++, are inherently unsafe and have led to numerous programming errors. The problem is that the functions in this library that move string data do not guard against overflowing the destination. For example, consider the following call to strcpy:

strcpy(dest, src);

If the length of dest is 20 and the length of src is 50, strcpy will write over the 30 bytes that follow dest. The point is that strcpy does not know the length of dest, so it cannot ensure that the memory following it will not be overwritten. The same problem can occur with several of the other functions in the C string library. In addition to C-style strings, C++ also supports strings through its standard class library, which is also similar to that of Java. Because of the insecurities of the C string library, C++ programmers should use the string class from the standard library, rather than char arrays and the C string library.

In Java, strings are supported by the String class, whose values are constant strings, and the StringBuffer class, whose values are changeable and are more like arrays of single characters. These values are specified with methods of the StringBuffer class. C# and Ruby include string classes that are similar to those of Java.

Python includes strings as a primitive type and has operations for substring reference, catenation, indexing to access individual characters, as well as methods for searching and replacement. There is also an operation for character membership in a string. So, even though Python’s strings are primitive types, for character and substring references, they act very much like arrays of characters. However, Python strings are immutable, similar to the String class objects of Java.

In F#, strings are a class. Individual characters, which are represented in Unicode UTF-16, can be accessed, but not changed. Strings can be catenated with the + operator. In ML, string is a primitive immutable type. It uses ^ for its catenation operator and includes functions for substring referencing and getting the size of a string.

In Swift, the String class supports its character strings. String objects can be either constants or variables. The binary + operator catenates String variables. The append method is used to add a Character object to a String object. The characters method of String is used to examine individual characters of a String object.

history note SNOBOL 4 was the first widely known language to support pattern matching.

Perl, JavaScript, Ruby, and PHP include built-in pattern-matching operations. In these languages, the pattern-matching expressions are somewhat loosely based on mathematical regular expressions. In fact, they are often called regular expressions. They evolved from the early UNIX line editor, ed, to become part of the UNIX shell languages. Eventually, they grew to their current complex form. There is at least one complete book on this kind of ​- pattern-matching expressions (Friedl, 2006). In this section, we provide a brief look at the style of these expressions through two relatively simple examples.

Consider the following pattern expression:

/[A-Za-z][A-Za-z\d]+/

This pattern matches (or describes) the typical name form in programming languages. The brackets enclose character classes. The first character class specifies all letters; the second specifies all letters and digits (a digit is specified with the abbreviation \d). If only the second character class were included, we could not prevent a name from beginning with a digit. The plus operator following the second category specifies that there must be one or more of what is in the category. So, the whole pattern matches strings that begin with a letter, followed by one or more letters or digits.

Next, consider the following pattern expression:

/\d+\.?\d*|\.\d+/

This pattern matches numeric literals. The \. specifies a literal decimal point.3 The question mark quantifies what it follows to have zero or one appearance. The vertical bar (|) separates two alternatives in the whole pattern. The first alternative matches strings of one or more digits, possibly followed by a decimal point, followed by zero or more digits; the second

alternative matches strings that begin with a decimal point, followed by one or more digits.

3. The period must be “escaped” with the backslash because period has special meaning in a regular expression.

Pattern-matching capabilities using regular expressions are included in the class libraries of C++, Java, Python, C#, and F#.

6.3.3 String Length Options There are several design choices regarding the length of string values. First, the length can be static and set when the string is created. Such a string is called a static length string. This is the choice for the strings of Python, the immutable objects of Java’s String class, as well as similar classes in the C++ standard class library, Ruby’s built-in String class, and the .NET class library available to C# and F#.

The second option is to allow strings to have varying length up to a declared and fixed maximum set by the variable’s definition, as exemplified by the strings in C and the C-style strings of C++. These are called limited dynamic length strings. Such string variables can store any number of characters between zero and the maximum. Recall that strings in C use a special character to indicate the end of the string’s characters, rather than maintaining the string length.

The third option is to allow strings to have varying length with no maximum, as in JavaScript, Perl, and the standard C++ library. These are called dynamic length strings. This option requires the overhead of dynamic storage allocation and deallocation but provides maximum flexibility.

6.3.4 Evaluation String types are important to the writability of a language. Dealing with strings as arrays can be more cumbersome than dealing with a primitive

string type. For example, consider a language that treats strings as arrays of characters and does not have a predefined function that does what strcpy in C does. Then, a simple assignment of one string to another would require a loop. The addition of strings as a primitive type to a language is not costly in terms of either language or compiler complexity. Therefore, it is difficult to justify the omission of primitive string types in some contemporary languages. Of course, providing strings through a standard library is nearly as convenient as having them as a primitive type.

String operations such as simple pattern matching and catenation are essential and should be included for string type values. Although dynamic length strings are obviously the most flexible, the overhead of their implementation must be weighed against that additional flexibility.

6.3.5 Implementation of Character String Types Character string types could be supported directly in hardware; but in most cases, software is used to implement string storage, retrieval, and manipulation. When character string types are represented as character arrays, the language often supplies few operations.

A descriptor for a static character string type, which is required only during compilation, has three fields. The first field of every descriptor is the name of the type. In the case of static character strings, the second field is the type’s length (in characters). The third field is the address of the first character. This descriptor is shown in Figure 6.2. Limited dynamic strings require a run-time descriptor to store the fixed maximum length, the current length, and the address, as shown in Figure 6.3. Dynamic length strings require a simpler run-time descriptor because only the current length and the address need to be stored. Although we depict descriptors as independent blocks of storage, in most cases, they are stored in the symbol table.

Figure 6.2 Compile-time descriptor for static strings

Figure 6.3 Run-time descriptor

for limited dynamic strings The limited dynamic strings of C and C++ do not require run-time descriptors, because the end of a string is marked with the null character. They do not need the maximum length, because index values in array references are not range checked in these languages.

Static length and limited dynamic length strings require no special dynamic storage allocation. In the case of limited dynamic length strings, sufficient storage for the maximum length is allocated when the string variable is bound to storage, so only a single allocation process is involved.

Dynamic length strings require more complex storage management. The length of a string, and therefore the storage to which it is bound, must grow and shrink dynamically.

There are three approaches to supporting the dynamic allocation and deallocation that is required for dynamic length strings. First, strings can be stored in a linked list, so that when a string grows, the newly required cells can come from anywhere in the heap. The drawbacks to this method are the extra storage occupied by the links in the list representation and the necessary complexity of string operations.

The second approach is to store strings as arrays of pointers to individual characters allocated in the heap. This method still uses extra memory, but string processing can be faster than with the linked-list approach.

The third alternative is to store complete strings in adjacent storage cells. The problem with this method arises when a string grows: How can storage that is adjacent to the existing cells continue to be allocated for the string variable? Frequently, such storage is not available. Instead, a new area of memory is found that can store the complete new string, and the old part is moved to this area. Then, the memory cells used for the old string are deallocated. This latter approach is the one typically used. The general problem of managing allocation and deallocation of variable-size segments is discussed in Section 6.11.7.3.

Although the linked-list method requires more storage, the associated allocation and deallocation processes are simple. However, some string operations are slowed by the required pointer chasing. On the other hand, using adjacent memory for complete strings results in faster string operations and requires significantly less storage, but the allocation and deallocation processes are slower.

6.4 Enumeration Types An enumeration type is one in which all of the possible values, which are named constants, are provided, or enumerated, in the definition. Enumeration types provide a way of defining and grouping collections of named constants, which are called enumeration constants. The definition of a typical enumeration type is shown in the following C# example:

enum days {Mon, Tue, Wed, Thu, Fri, Sat, Sun};

The enumeration constants are typically implicitly assigned the integer values, 0, 1, . . . but can be explicitly assigned any integer literal in the type’s definition.

6.4.1 Design Issues The design issues for enumeration types are as follows:

Is an enumeration constant allowed to appear in more than one type definition, and if so, how is the type of an occurrence of that constant in the program checked?

Are enumeration values coerced to integer?

Are any other types coerced to an enumeration type?

All of these design issues are related to type checking. If an enumeration variable is coerced to a numeric type, then there is little control over its range of legal operations or its range of values. If an int type value is coerced to an enumeration type, then an enumeration type variable could be assigned any integer value, whether it represented an enumeration constant or not.

6.4.2 Designs

In languages that do not have enumeration types, programmers usually simulate them with integer values. For example, suppose we needed to represent colors in a C program and C did not have an enumeration type. We might use 0 to represent blue, 1 to represent red, and so forth. These values could be defined as follows:

int red = 0, blue = 1;

Now, in the program, we could use red and blue as if they were of a color type. The problem with this approach is that because we have not defined a type for our colors, there is no type checking when they are used. For example, it would be legal to add the two together, although that would rarely be an intended operation. They could also be combined with any other numeric type operand using any arithmetic operator, which would also rarely be useful. Furthermore, because they are just variables, they could be assigned any integer value, thereby destroying the relationship with the colors. This latter problem could be prevented by making them named constants.

C and Pascal were the first widely used languages to include an enumeration data type. C++ includes C’s enumeration types. In C++, we could have the following:

enum colors {red, blue, green, yellow, black};

colors myColor = blue, yourColor = red;

The colors type uses the default internal values for the enumeration constants, 0, 1, . . . , although the constants could have been specifically assigned any integer literal (or any constant-valued expression) by the programmer. The enumeration values are coerced to int when they are put in integer context. This allows their use in any numeric expression. For example, if the current value of myColor is blue, then the expression

myColor++

would assign the integer code for green to myColor.

C++ also allows enumeration constants to be assigned to variables of any numeric type, though that would likely be an error. However, no other type

value is coerced to an enumeration type in C++. For example,

myColor = 4;

is illegal in C++. This assignment would be legal if the right side had been cast to colors type. This prevents some potential errors.

C++ enumeration constants can appear in only one enumeration type in the same referencing environment.

In 2004, an enumeration type was added to Java in Java 5.0. All enumeration types in Java are implicitly subclasses of the predefined class Enum. Because enumeration types are classes, they can have instance data fields, constructors, and methods. Syntactically, Java enumeration type definitions appear like those of C++, except that they can include fields, constructors, and methods. The possible values of an enumeration are the only possible instances of the class. All enumeration types inherit toString, as well as a few other methods. An array of the instances of an enumeration type can be fetched with the static method values. The internal numeric value of an enumeration variable can be fetched with the ordinal method. No expression of any other type can be assigned to an enumeration variable. Also, an enumeration variable is never coerced to any other type.

C# enumeration types are like those of C++, except that they are never coerced to integer. So, operations on enumeration types are restricted to those that make sense. Also, the range of values is restricted to that of the particular enumeration type.

In ML, enumeration types are defined as new types with datatype declarations. For example, we could have the following:

datatype weekdays = Monday | Tuesday | Wednesday |

Thursday | Friday

The type of the elements of weekdays is integer.

F# has enumeration types that are similar to those of ML, except the reserved word type is used instead of datatype and the first value is preceded by an OR operator (|).

Swift has an enumeration type in which the enumeration values are names, which represent themselves, rather than having internal integer values. An enumeration type is defined in a structure that is similar to a switch structure, as in:

enum fruit {

case orange

case apple

case banana

}

Dot notation is used to reference enumeration values, so in our example, the value of apple is referenced as fruit.apple.

Interestingly, none of the relatively recent scripting languages include enumeration types. These include Perl, JavaScript, PHP, Python, and Ruby. Even Java was a decade old before enumeration types were added.

6.4.3 Evaluation Enumeration types can provide advantages in both readability and reliability. Readability is enhanced very directly: Named values are easily recognized, whereas coded values are not.

In the area of reliability, the enumeration types of C#, F#, Java 5.0, and Swift provide two advantages: (1) No arithmetic operations are legal on enumeration types; this prevents adding days of the week, for example, and (2) second, no enumeration variable can be assigned a value outside its defined range.4 If the colors enumeration type has 10 enumeration constants and uses 0..9 as its internal values, no number greater than 9 can be assigned to a ​colors type variable.

4. In C# and F#, an integer value can be cast to an enumeration type and assigned to the name of an enumeration variable. Such values must be tested with Enum.IsDefined method before assigning them to the name of an enumeration variable.

Because C treats enumeration variables like integer variables, it does not

provide either of these two advantages.

C++ is a little better. Numeric values can be assigned to enumeration type variables only if they are cast to the type of the assigned variable. Numeric values assigned to enumeration type variables are checked to determine whether they are in the range of the internal values of the enumeration type. Unfortunately, if the user uses a wide range of explicitly assigned values, this checking is not effective. For example,

enum colors {red = 1, blue = 1000, green = 100000}

In this example, a value assigned to a variable of colors type will only be checked to determine whether it is in the range of 1..100000.

6.5 Array Types An array is a homogeneous aggregate of data elements in which an individual element is identified by its position in the aggregate, relative to the first element. The individual data elements of an array are of the same type. References to individual array elements are specified using subscript expressions. If any of the subscript expressions in a reference include variables, then the reference will require an additional run-time calculation to determine the address of the memory location being referenced.

In many languages, such as C, C++, Java, and C#, all of the elements of an array are required to be of the same type. In these languages, pointers and references are restricted to point to or reference a single type. So the objects or data values being pointed to or referenced are also of a single type. In some other languages, such as JavaScript, Python, and Ruby, variables are typeless references to objects or data values. In these cases, arrays still consist of elements of a single type, but the elements can reference objects or data values of different types. Such arrays are still homogeneous, because the array elements are of the same type. In Swift, arrays can be typed, that is, they will contain values only of a single type, or untyped, which means they can contain values of any type.

C# and Java 5.0 provide generic arrays, that is, arrays whose elements are references to objects, through their class libraries. These are discussed in Section 6.5.3.

6.5.1 Design Issues The primary design issues specific to arrays are the following:

What types are legal for subscripts?

Are subscripting expressions in element references range checked?

When are subscript ranges bound?

When does array allocation take place?

Are ragged or rectangular multidimensioned arrays allowed, or both?

Can arrays be initialized when they have their storage allocated?

What kinds of slices are allowed, if any?

In the following sections, examples of the design choices made for the arrays of the most common programming languages are discussed.

6.5.2 Arrays and Indices Specific elements of an array are referenced by means of a two-level syntactic mechanism, where the first part is the aggregate name, and the second part is a possibly dynamic selector consisting of one or more items known as subscripts or indices. If all of the subscripts in a reference are constants, the selector is static; otherwise, it is dynamic. The selection operation can be thought of as a mapping from the array name and the set of subscript values to an element in the aggregate. Indeed, arrays are sometimes called finite mappings. Symbolically, this mapping can be shown as

history note The designers of pre-90 Fortrans and PL/I chose parentheses for array subscripts because no other suitable characters were available at the time. Card punches did not include bracket characters.

array\_name(subscript\_value\_list)→element

The syntax of array references is fairly universal: The array name is followed by the list of subscripts, which is surrounded by either parentheses or brackets. In some languages that provide multidimensioned arrays as arrays

of arrays, each subscript appears in its own brackets. A problem with using parentheses to enclose subscript expressions is that they often are also used to enclose the parameters in subprogram calls; this use makes references to arrays appear exactly like those calls. For example, consider the following Ada assignment statement:

Sum := Sum + B(I);

Because parentheses are used for both subprogram parameters and array subscripts in Ada, both program readers and compilers are forced to use other information to determine whether B(I) in this assignment is a function call or a reference to an array element. This results in reduced readability.

history note Fortran I limited the number of array subscripts to three, because at the time of the design, execution efficiency was a primary concern. Fortran I designers had developed a very fast method for accessing the elements of arrays of up to three dimensions, using the three index registers of the IBM 704. Fortran IV was first implemented on an IBM 7094, which had seven index registers. This allowed Fortran IV’s designers to allow arrays with up to seven subscripts. Most other contemporary languages enforce no such limits.

The designers of Ada specifically chose parentheses to enclose subscripts so there would be uniformity between array references and function calls in expressions, in spite of potential readability problems. They made this choice in part because both array element references and function calls are mappings. Array element references map the subscripts to a particular element of the array. Function calls map the actual parameters to the function definition and, eventually, a functional value.

Most languages other than Fortran and Ada use brackets to delimit their array indices.

Two distinct types are involved in an array type: the element type and the type of the subscripts. The type of the subscripts is often integer.

Early programming languages did not specify that subscript ranges must be implicitly checked. Range errors in subscripts are common in programs, so requiring range checking is an important factor in the reliability of languages. Many contemporary languages also do not specify range checking of subscripts, but Java, ML, and C# do.

Subscripting in Perl is a bit unusual in that although the names of all arrays begin with at signs (@), because array elements are always scalars and the names of scalars always begin with dollar signs ($), references to array elements use dollar signs rather than at signs in their names. For example, for the array @list, the second element is referenced with $list[1].

One can reference an array element in Perl with a negative subscript, in which case the subscript value is an offset from the end of the array. For example, if the array @list has five elements with the subscripts 0..4, $list [−2] references the element with the subscript 3. A reference to a nonexistent element in Perl yields undef, but no error is reported.

6.5.3 Subscript Bindings and Array Categories The binding of the subscript type to an array variable is usually static, but the subscript value ranges are sometimes dynamically bound.

In some languages, the lower bound of the subscript range is implicit. For example, in the C-based languages, the lower bound of all subscript ranges is fixed at 0. In some other languages, the lower bounds of the subscript ranges must be specified by the programmer.

There are four categories of arrays, based on the binding to subscript ranges, the binding to storage, and from where the storage is allocated. The category names indicate the design choices of these three. In the first three of these categories, once the subscript ranges are bound and the storage is allocated, they remain fixed for the lifetime of the variable. Of course, when the subscript ranges are fixed, the array cannot change size.

A static array is one in which the subscript ranges are statically bound and storage allocation is static (done before run time). The advantage of static arrays is efficiency: No dynamic allocation or deallocation is required. The disadvantage is that the storage for the array is fixed for the entire execution time of the program.

A fixed stack-dynamic array is one in which the subscript ranges are statically bound, but the allocation is done at declaration elaboration time during execution. The advantage of fixed stack-dynamic arrays over static arrays is space efficiency. A large array in one subprogram can use the same space as a large array in a different subprogram, as long as both subprograms are not active at the same time. The same is true if the two arrays are in different blocks that are not active at the same time. The disadvantage is the required allocation and deallocation time.

A fixed heap-dynamic array is similar to a fixed stack-dynamic array, in that the subscript ranges and the storage binding are both fixed after storage is allocated. The differences are that both the subscript ranges and storage bindings are done when the user program requests them during execution, and the storage is allocated from the heap, rather than the stack. The advantage of fixed heap-dynamic arrays is flexibility—the array’s size always fits the problem. The disadvantage is allocation time from the heap, which is longer than allocation time from the stack.

A heap-dynamic array is one in which the binding of subscript ranges and storage allocation is dynamic and can change any number of times during the array’s lifetime. The advantage of heap-dynamic arrays over the others is flexibility: Arrays can grow and shrink during program execution as the need for space changes. The disadvantage is that allocation and deallocation take longer and may happen many times during execution of the program. Examples of the four categories are given in the following paragraphs.

Arrays declared in C and C++ functions that include the static modifier are static.

Arrays that are declared in C and C++ functions without the static specifier are examples of fixed stack-dynamic arrays.

C and C++ also provide fixed heap-dynamic arrays. The standard C library functions malloc and free, which are general heap allocation and deallocation operations, respectively, can be used for C arrays. C++ uses the operators new and delete to manage heap storage. An array is treated as a pointer to a collection of storage cells, where the pointer can be indexed, as discussed in Section 6.11.5.

In Java, all non-generic arrays are fixed heap-dynamic. Once created, these arrays keep the same subscript ranges and storage. C# also provides fixed heap-dynamic arrays.

Objects of the C# List class are generic heap-dynamic arrays. These array objects are created without any elements, as in

List<String> stringList = new List<String>();

Elements are added to this object with the Add method, as in

stringList.Add("Michael");

Access to elements of these arrays is through subscripting.

Java includes a generic class similar to C#’s List, named ArrayList. It is different from C#’s List in that subscripting is not supported—get and set methods must be used to access the elements.

A Perl array can be made to grow by using the push (puts one or more new elements on the end of the array) and unshift (puts one or more new elements on the beginning of the array), or by assigning a value to the array specifying a subscript beyond the highest current subscript of the array. An array can be made to shrink to no elements by assigning it the empty list, (). The length of an array is defined to be the largest subscript plus one.

Like Perl, JavaScript allows arrays to grow with the push and unshift methods and shrink by setting them to the empty list. However, negative subscripts are not supported.

JavaScript arrays can be sparse, meaning the subscript values need not be contiguous. For example, suppose we have an array named list that has 10

elements with the subscripts 0..9.5 Consider the following assignment statement:

5. The subscript range could just as easily have been 1000 .. 1009.

list[50] = 42;

Now, list has 11 elements and length 51. The elements with subscripts 11..49 are not defined and therefore do not require storage. A reference to a nonexistent element in a JavaScript array yields undefined.

Arrays in Python and Ruby can be made to grow only through methods to add elements or catenate other arrays. Ruby and Perl support negative subscripts, but Python does not. In Python an element or slice of an array can be deleted. A reference to a nonexistent element in Python results in a run- time error, whereas a similar reference in Ruby yields nil and no error is reported.

Swift dynamic arrays are objects that use integer subscripts, beginning at zero, and include several useful methods. The append method adds an element to the end of an array. The insert method inserts a new element at any position in the array, but results in an error if the insertion is at a subscript beyond the current length of the array. Elements can be removed from an array with the removeAtIndex method. There are also reverse and count methods.

Although the ML definition does not include arrays, its widely used implementation, SML/NJ, does.

The only predefined collection type that is part of F# is the array (other collection types are provided through the .NET Framework Library). These arrays are like those of C#. A foreach statement is included in the language for array processing.

6.5.4 Array Initialization Some languages provide the means to initialize arrays at the time their

storage is allocated. C, C++, Java, Swift, and C# allow initialization of their arrays. Consider the following C declaration:

int list [] = {4, 5, 7, 83};

The array list is created and initialized with the values 4, 5, 7, and 83. The compiler also sets the length of the array. This is meant to be a convenience but is not without cost. It effectively removes the possibility that the system could detect some kinds of programmer errors, such as mistakenly leaving a value out of the list.

As discussed in Section 6.3.2, character strings in C and C++ are implemented as arrays of char. These arrays can be initialized to string constants, as in

char name [] = "freddie";

The array name will have eight elements, because all strings are terminated with a null character (zero), which is implicitly supplied by the system for string constants.

Arrays of strings in C and C++ can also be initialized with string literals. For example,

char *names [] = {"Bob", "Jake", "Darcie"};

This example illustrates the nature of character literals in C and C++. In the previous example of a string literal being used to initialize the char array name, the literal is taken to be a char array. But in the latter example (names), the literals are taken to be pointers to characters, so the array is an array of pointers to characters. For example, names[0] is a pointer to the letter 'B' in the literal character array that contains the characters 'B', 'o', 'b', and the null character.

In Java, similar syntax is used to define and initialize an array of references to String objects. For example,

String[] names = ["Bob", "Jake", "Darcie"];

6.5.5 Array Operations An array operation is one that operates on an array as a unit. The most common array operations are assignment, catenation, comparison for equality and inequality, and slices, which are discussed separately in Section 6.5.5.

The C-based languages do not provide any array operations, except through the methods of Java, C++, and C#. Perl supports array assignments but does not support comparisons.

Python’s arrays are called lists, although they have all the characteristics of dynamic arrays. Because the objects can be of any types, these arrays are heterogeneous. Python provides array assignment, although it is only a reference change. Python also has operations for array catenation (+) and element membership (in). It includes two different comparison operators: one that determines whether the two variables reference the same object (is) and one that compares all corresponding objects in the referenced objects, regardless of how deeply they are nested, for equality (==).

Like Python, the elements of Ruby’s arrays are references to objects. And like Python, when a == operator is used between two arrays, the result is true only if the two arrays have the same length and the corresponding elements are equal. Ruby’s arrays can be catenated with an Array method.

F# includes many array operators in its Array module. Among these are Array.append, Array.copy, and Array.length.

Arrays and their operations are the heart of APL; it is the most powerful array-processing language ever devised. Because of its relative obscurity and its lack of effect on subsequent languages, however, we present here only a glimpse into its array operations.

In APL, the four basic arithmetic operations are defined for vectors (single- dimensioned arrays) and matrices, as well as scalar operands. For example,

A + B

is a valid expression, whether A and B are scalar variables, vectors, or matrices.

APL includes a collection of unary operators for vectors and matrices, some of which are as follows (where V is a vector and M is a matrix):

APL also includes several special operators that take other operators as operands. One of these is the inner product operator, which is specified with a period (.). It takes two operands, which are binary operators. For example,

+.×

is a new operator that takes two arguments, either vectors or matrices. It first multiplies the corresponding elements of two arguments, and then it sums the results. For example, if A and B are vectors,

A × B

is the mathematical inner product of A and B (a vector of the products of the corresponding elements of A and B). The statement

A +.× B

is the sum of the inner product of A and B. If A and B are matrices, this expression specifies the matrix multiplication of A and B.

The special operators of APL are actually functional forms, which are described in Chapter 15.

6.5.6 Rectangular and Jagged

Arrays A rectangular array is a multidimensioned array in which all of the rows have the same number of elements and all of the columns have the same number of elements. Rectangular arrays model rectangular tables exactly.

A jagged array is one in which the lengths of the rows need not be the same. For example, a jagged matrix may consist of three rows, one with 5 elements, one with 7 elements, and one with 12 elements. This also applies to the columns and higher dimensions. So, if there is a third dimension (layers), each layer can have a different number of elements. Jagged arrays are made possible when multidimensioned arrays are actually arrays of arrays. For example, a matrix would appear as an array of single-​dimensioned arrays.

C, C++, and Java support jagged arrays but not rectangular arrays. In those languages, a reference to an element of a multidimensioned array uses a separate pair of brackets for each dimension. For example,

myArray[3][7]

C# and F# support both rectangular and jagged arrays. For rectangular arrays, all subscript expressions in references to elements are placed in a single pair of brackets. For example,

myArray[3, 7]

6.5.7 Slices A slice of an array is some substructure of that array. For example, if A is a matrix, then the first row of A is one possible slice, as are the last row and the first column. It is important to realize that a slice is not a new data type. Rather, it is a mechanism for referencing part of an array as a unit. If arrays cannot be manipulated as units in a language, that language has no use for slices.

Consider the following Python declarations:

vector = [2, 4, 6, 8, 10, 12, 14, 16]

mat = [[1, 2, 3],[4, 5, 6],[7, 8, 9]]

Recall that the default lower bound for Python arrays is 0. The syntax of a Python slice reference is a pair of numeric expressions separated by a colon. The first is the first subscript of the slice; the second is the first subscript after the last subscript in the slice. Therefore, vector[3:6] is a three-element array with the fourth through sixth elements of vector (those elements with the subscripts 3, 4, and 5). A row of a matrix is specified by giving just one subscript. For example, mat[1] refers to the second row of mat; a part of a row can be specified with the same syntax as a part of a single-dimensioned array. For example, mat[0][0:2] refers to the first and second element of the first row of mat, which is [1, 2].

Python also supports more complex slices of arrays. For example, vector[0:7:2] references every other element of vector, up to but not including the element with the subscript 7, starting with the subscript 0, which is [2, 6, 10, 14].

Perl supports slices of two forms, a list of specific subscripts or a range of subscripts. For example,

@list[1..5] = @list2[3, 5, 7, 9, 13];

Notice that slice references use array names, not scalar names, because slices are arrays (not scalars).

Ruby supports slices with the slice method of its Array object, which can take three forms of parameters. A single integer expression parameter is interpreted as a subscript, in which case slice returns the element with the given subscript. If slice is given two integer expression parameters, the first is interpreted as a beginning subscript and the second is interpreted as the number of elements in the slice. For example, suppose list is defined as follows:

list = [2, 4, 6, 8, 10]

list.slice(2, 2) returns [6, 8]. The third parameter form for slice is a range, which has the form of an integer expression, two periods, and a second

integer expression. With a range parameter, slice returns an array of the element with the given range of subscripts. For example, list.slice (1..3) returns [4, 6, 8].

6.5.8 Evaluation Arrays have been included in virtually all programming languages. The primary advances since their introduction in Fortran I have been slices and dynamic arrays. As discussed in Section 6.6, the latest advances in arrays have been in associative arrays.

6.5.9 Implementation of Array Types Implementing arrays requires considerably more compile-time effort than does implementing primitive types. The code to allow accessing of array elements must be generated at compile time. At run time, this code must be executed to produce element addresses. There is no way to precompute the address to be accessed by a reference such as

list[k]

A single-dimensioned array is implemented as a list of adjacent memory cells. Suppose the array list is defined to have a subscript range lower bound of 0. The access function for list is often of the form

address (list[k]) = address (list[0]) + k * element_size

where the first operand of the addition is the constant part of the access function, and the second is the variable part.

If the element type is statically bound and the array is statically bound to storage, then the value of the constant part can be computed before run time. However, the addition and multiplication operations must be done at run

time.

The generalization of this access function for an arbitrary lower bound is

address (list[k]) = address (list[lower_bound]) +

          ((k −lower_bound) * element_size)

The compile-time descriptor for single-dimensioned arrays can have the form shown in Figure 6.4. The descriptor includes information required to construct the access function. If run-time checking of index ranges is not done and the attributes are all static, then only the access function is required during execution; no descriptor is needed. If run-time checking of index ranges is done, then those index ranges may need to be stored in a run-time descriptor. If the subscript ranges of a particular array type are static, then the ranges may be incorporated into the code that does the checking, thus eliminating the need for the run-time descriptor. If any of the descriptor entries are dynamically bound, then those parts of the descriptor must be maintained at run time.

Figure 6.4 Compile-time descriptor for single- dimensioned arrays

True multidimensional arrays, that is, those that are not arrays of arrays, are more complex to implement than single-dimensioned arrays, although the

extension to more dimensions is straightforward. Hardware memory is linear —just a simple sequence of bytes. So values of data types that have two or more dimensions must be mapped onto the single-dimensioned memory. There are two ways in which multidimensional arrays can be mapped to one dimension: row major order and column major order (not used in any widely used language). In row major order, the elements of the array that have as their first subscript the lower bound value of that subscript are stored first, followed by the elements of the second value of the first subscript, and so forth. If the array is a matrix, it is stored by rows. For example, if the matrix had the values

3 4 7

6 2 5

1 3 8

it would be stored in row major order as

3, 4, 7, 6, 2, 5, 1, 3, 8

The access function for a multidimensional array is the mapping of its base address and a set of index values to the address in memory of the element specified by the index values. The access function for two-dimensional arrays stored in row major order can be developed as follows. In general, the address of an element is the base address of the structure plus the element size times the number of elements that precede it in the structure. For a matrix in row major order, the number of elements that precede an element is the number of rows above the element times the size of a row, plus the number of elements to the left of the element in its row. This is illustrated in Figure 6.5, in which we assume that subscript lower bounds are all zero.

Figure 6.5 The location of the [i,j] element in a matrix

Figure 6.5 Full Alternative Text

To get an actual address value, the number of elements that precede the desired element must be multiplied by the element size. Now, the access function can be written as

location(a[i,j]) = address of a[0, 0]

        + ((((number of rows above the ith row) * (size of a row))

        + (number of elements left of the jth column)) *

         element size)

Because the number of rows above the ith row is i and the number of elements to the left of the jth column is j, we have

location(a[i, j]) = address of a[0, 0] + (((i * n) + j) *

           element_size)

where n is the number of elements per row. The first term is the constant part and the last is the variable part.

The generalization to arbitrary lower bounds results in the following access function:

location(a[i, j]) = address of a[row_lb, col_lb]

          = (((i − row_lb) * n) + (j − col_lb)) * element_size

where row_lb is the lower bound of the rows and col_lb is the lower bound of the columns. This can be rearranged to the form

location(a[i, j]) = address of a[row_lb, col_lb]

          − (((row_lb * n) + col_lb) * element_size)

             + (((i * n) + j) * element_size)

where the first two terms are the constant part and the last is the variable part. This can be generalized rather easily to an arbitrary number of dimensions.

For each dimension of an array, one add and one multiply instruction are required for the access function. Therefore, accesses to elements of arrays with several subscripts are costly. The compile-time descriptor for a multidimensional array is shown in Figure 6.6.

Figure 6.6 A compile-time ​- descriptor for a ​- multidimensional array

6.6 Associative Arrays An associative array is an unordered collection of data elements that are indexed by an equal number of values called keys. In the case of non-​- associative arrays, the indices never need to be stored (because of their regularity). In an associative array, however, the user-defined keys must be stored in the structure. So each element of an associative array is in fact a pair of entities, a key and a value. We use Perl’s design of associative arrays to illustrate this data structure. Associative arrays are also supported directly by Python, Ruby, and Swift and by the standard class libraries of Java, C++, C#, and F#.

The only design issue that is specific for associative arrays is the form of references to their elements.

6.6.1 Structure and Operations In Perl, associative arrays are called hashes, because in the implementation their elements are stored and retrieved with hash functions. The namespace for Perl hashes is distinct: Every hash variable name must begin with a percent sign (%). Each hash element consists of two parts: a key, which is a string, and a value, which is a scalar (number, string, or reference). Hashes can be set to literal values with the assignment statement, as in

%salaries = ("Gary" => 75000, "Perry" => 57000,

"Mary" => 55750, "Cedric" => 47850);

Individual element values are referenced using notation that is similar to that used for Perl arrays. The key value is placed in braces and the hash name is replaced by a scalar variable name that is the same except for the first character. Although hashes are not scalars, the value parts of hash elements are scalars, so references to hash element values use scalar names. Recall that scalar variable names begin with dollar signs ($). So, an assignment of 58850 to the element of %salaries with the key "Perry" would appear as follows:

$salaries{"Perry"} = 58850;

A new element is added using the same assignment statement form. An element can be removed from the hash with the delete operator, as in the following:

delete $salaries{"Gary"};

The entire hash can be emptied by assigning the empty literal to it, as in the following:

@salaries = ();

The size of a Perl hash is dynamic: It grows when an element is added and shrinks when an element is deleted, and also when it is emptied by assignment of the empty literal. The exists operator returns true or false, depending on whether its operand key is an element in the hash. For example,

if (exists $salaries{"Shelly"}) . . .

The keys operator, when applied to a hash, returns an array of the keys of the hash. The values operator does the same for the values of the hash. The each operator iterates over the element pairs of a hash.

Python’s associative arrays, which are called dictionaries, are similar to those of Perl, except the values are all references to objects. The associative arrays supported by Ruby are similar to those of Python, except that the keys can be any object,6 rather than just strings. There is a progression from Perl’s hashes, in which the keys must be strings, to PHP’s arrays, in which the keys can be integers or strings, to Ruby’s hashes, in which any type object can be a key.

6. Objects that change do not make good keys, because the changes could change the hash function value. Therefore, arrays and hashes are never used as keys.

PHP’s arrays are both normal arrays and associative arrays. They can be treated as either. The language provides functions that allow both indexed and hashed access to elements. An array can have elements that are created

with simple numeric indices and elements that are created with string hash keys.

Swift’s associative arrays are called dictionaries. The keys can be of one specific type, but the values can be of mixed types, in which case they are objects.

An associative array is much better than an array if searches of the elements are required, because the implicit hashing operation used to access elements is very efficient. Furthermore, associative arrays are ideal when the data to be stored is paired, as with employee names and their salaries. On the other hand, if every element of a list must be processed, it is more efficient to use an array.

6.6.2 Implementing Associative Arrays The implementation of Perl’s associative arrays is optimized for fast lookups, but it also provides relatively fast reorganization when array growth requires it. A 32-bit hash value is computed for each entry and is stored with the entry, although an associative array initially uses only a small part of the hash value. When an associative array must be expanded beyond its initial size, the hash function need not be changed; rather, more bits of the hash value are used. Only half of the entries must be moved when this happens. So, although expansion of an associative array is not free, it is not as costly as might be expected.

The elements in PHP’s arrays are placed in memory through a hash function. However, all elements are linked together in the order in which they were created. The links are used to support iterative access to elements through the current and next functions.

6.7 Record Types A record is an aggregate of data elements in which the individual elements are identified by names and accessed through offsets from the beginning of the structure.

There is frequently a need in programs to model a collection of data in which the individual elements are not of the same type or size. For example, information about a college student might include name, student number, grade point average, and so forth. A data type for such a collection might use a character string for the name, an integer for the student number, a ​floating- point for the grade point average, and so forth. Records are designed for this kind of need.

It may appear that records and heterogeneous arrays are the same, but that is not the case. The elements of a heterogeneous array are all references to data objects that reside in scattered locations, often on the heap. The elements of a record are of potentially different sizes and reside in adjacent memory locations.

Records have been part of all of the most popular programming languages, except pre-90 versions of Fortran, since the early 1960s, when they were introduced by COBOL. In some languages that support object-oriented programming, data classes serve as records.

In C, C++, C#, and Swift, records are supported with the struct data type. In C++, structures are a minor variation on classes. In C#, structs also are related to classes, but are quite different from them. C# structs are stack-​- allocated value types, as opposed to class objects, which are heap-allocated reference types. Structs in C++ and C# are normally used as encapsulation structures, rather than data structures. They are further discussed in this capacity in Chapter 11. Structs are also included in ML and F#.

In Python and Ruby, records can be implemented as hashes, which themselves can be elements of arrays.

The following sections describe how records are declared or defined, how references to fields within records are made, and the common record operations.

The following design issues are specific to records:

What is the syntactic form of references to fields?

Are elliptical references allowed?

6.7.1 Definitions of Records The fundamental difference between a record and an array is that record elements, or fields, are not referenced by indices. Instead, the fields are named with identifiers, and references to the fields are made using these identifiers. Another difference between arrays and records is that records in some languages are allowed to include unions, which are discussed in Section 6.10.

The COBOL form of a record declaration, which is part of the data division of a COBOL program, is illustrated in the following example:

01 EMPLOYEE-RECORD.

02 EMPLOYEE-NAME.

05 FIRST PICTURE IS X(20).

05 Middle PICTURE IS X(10).

05 LAST PICTURE IS X(20).

02 HOURLY-RATE PICTURE IS 99V99.

The EMPLOYEE-RECORD record consists of the EMPLOYEE-NAME record and the HOURLY-RATE field. The numerals 01, 02, and 05 that begin the lines of the record declaration are level numbers, which indicate by their relative values the hierarchical structure of the record. Any line that is followed by a line with a higher-level number is itself a record. The PICTURE clauses show the formats of the field storage locations, with X(20) specifying 20 alphanumeric characters and 99V99 specifying four decimal digits with the decimal point in the middle.

In Java, records can be defined as data classes, with nested records defined as nested classes. Data members of such classes serve as the record fields.

6.7.2 References to Record Fields References to the individual fields of records are syntactically specified by several different methods, two of which name the desired field and its enclosing records. COBOL field references have the form

field_name OF record_name_1 OF . . . OF record_name_n

where the first record named is the smallest or innermost record that contains the field. The next record name in the sequence is that of the record that contains the previous record, and so forth. For example, the Middle field in the COBOL record example above can be referenced with

Middle OF EMPLOYEE-NAME OF EMPLOYEE-RECORD

Most of the other languages use dot notation for field references, where the components of the reference are connected with periods. Names in dot notation have the opposite order of COBOL references: They use the name of the largest enclosing record first and the field name last. For example, if Middle is a field in the Employee_Name record which is embedded in the Employee_Record record, it would be referenced with the following:

Employee_Record.Employee_Name.Middle

A fully qualified reference to a record field is one in which all intermediate record names, from the largest enclosing record to the specific field, are named in the reference. In the COBOL example above the field reference is fully qualified. As an alternative to fully qualified references, COBOL allows elliptical references to record fields. In an elliptical reference, the field is named, but any or all of the enclosing record names can be omitted, as long as the resulting reference is unambiguous in the referencing environment. For example, FIRST, FIRST OF EMPLOYEE-NAME, and FIRST OF EMPLOYEE-RECORD are elliptical references to the employee’s first name in the COBOL record declared above. Although elliptical references are a programmer

convenience, they require a compiler to have elaborate data structures and procedures in order to correctly identify the referenced field. They are also somewhat detrimental to readability.

6.7.3 Evaluation Records are frequently valuable data types in programming languages. The design of record types is straightforward, and their use is safe.

Records and arrays are closely related structural forms, and therefore it is interesting to compare them. Arrays are used when all the data values have the same type and/or are processed in the same way. This processing is easily done when there is a systematic way of sequencing through the structure. Such processing is well supported by using dynamic subscripting as the addressing method.

Records are used when the collection of data values is heterogeneous and the different fields are not processed in the same way. Also, the fields of a record often need not be processed in a particular order. Field names are like literal, or constant, subscripts. Because they are static, they provide very efficient access to the fields. Dynamic subscripts could be used to access record fields, but it would disallow type checking and would also be slower.

Records and arrays represent thoughtful and efficient methods of fulfilling two separate but related applications of data structures.

6.7.4 Implementation of Record Types The fields of records are stored in adjacent memory locations. But because the sizes of the fields are not necessarily the same, the access method used for arrays is not used for records. Instead, the offset address, relative to the beginning of the record, is associated with each field. Field accesses are all handled using these offsets. The compile-time descriptor for a record has the

general form shown in Figure 6.7. Run-time descriptors for records are unnecessary.

Figure 6.7 A compile-time ​- descriptor for a record

Figure 6.7 Full Alternative Text

6.8 Tuple Types A tuple is a data type that is similar to a record, except that the elements are not named.

Python includes an immutable tuple type. If a tuple needs to be changed, it can be converted to an array with the list function. After the change, it can be converted back to a tuple with the tuple function. One use of tuples is when an array must be write protected, such as when it is sent as a parameter to an external function and the user does not want the function to be able to modify the parameter.

Python’s tuples are closely related to its lists, except that tuples are immutable. A tuple is created by assigning a tuple literal, as in the following example:

myTuple = (3, 5.8, 'apple')

Notice that the elements of a tuple need not be of the same type.

The elements of a tuple can be referenced with indexing in brackets, as in the following:

myTuple[1]

This references the first element of the tuple, because tuple indexing begins at 1.

Tuples can be catenated with the plus (+) operator. They can be deleted with the del statement. There are also other operators and functions that operate on tuples.

ML includes a tuple data type. An ML tuple must have at least two elements, whereas Python’s tuples can be empty or contain one element. As in Python, an ML tuple can include elements of mixed types. The following statement creates a tuple:

val myTuple = (3, 5.8, 'apple');

The syntax of a tuple element access is as follows:

#1(myTuple);

This references the first element of the tuple.

A new tuple type can be defined in ML with a type declaration, such as the following:

type intReal = int * real;

Values of this type consist of an integer and a real. The asterisk can be misleading. It is used to separate the tuple components, indicating a type product, and has nothing to do with arithmetic.

F# also has tuples. A tuple is created by assigning a tuple value, which is a list of expressions separated by commas and delimited by parentheses, to a name in a let statement. If a tuple has two elements, they can be referenced with the functions fst and snd, respectively. The elements of a tuple with more than two elements are often referenced with a tuple pattern on the left side of a let statement. A tuple pattern is simply a sequence of names, one for each element of the tuple, with or without the delimiting parentheses. When a tuple pattern is the left side of a let construct, it is a multiple assignment. For example, consider the following let constructs:

let tup = (3, 5, 7);;

let a, b, c = tup;;

This assigns 3 to a, 5 to b, and 7 to c.

Tuples are used in Python, ML, and F# to allow functions to return multiple values. In Swift, tuples are passed by value, so they are sometimes used to pass data to a function when the function is not to change that data.

6.9 List Types Lists were first supported in the first functional programming language, Lisp. They have always been part of the functional languages, but in recent years they have found their way into some imperative languages.

Lists in Scheme and Common Lisp are delimited by parentheses and the elements are not separated by any punctuation. For example,

(A B C D)

Nested lists have the same form, so we could have

(A (B C) D)

In this list, (B C) is a list nested inside the outer list.

Data and code have the same syntactic form in Lisp and its descendants. If the list (A B C) is interpreted as code, it is a call to the function A with parameters B and C.

The fundamental list operations in Scheme are two functions that take lists apart and two that build lists. The CAR function returns the first element of its list parameter. For example, consider the following example:

(CAR '(A B C))

The quote before the parameter list is to prevent the interpreter from considering the list a call to the A function with the parameters B and C, in which case it would interpret it. This call to CAR returns A.

The CDR function returns its parameter list minus its first element. For example, consider the following example:

(CDR '(A B C))

This function call returns the list (B C).

Common Lisp also has the functions FIRST (same as CAR), SECOND, . . . , TENTH, which return the element of their list parameters that is specified by their names.

In Scheme and Common Lisp, new lists are constructed with the CONS and LIST functions. The function CONS takes two parameters and returns a new list with its first parameter as the first element and its second parameter as the remainder of that list. For example, consider the following:

(CONS 'A '(B C))

This call returns the new list (A B C).

The LIST function takes any number of parameters and returns a new list with the parameters as its elements. For example, consider the following call to LIST:

(LIST 'A 'B '(C D))

This call returns the new list (A B (C D)).

ML has lists and list operations, although their appearance is not like those of Scheme. Lists are specified in square brackets, with the elements separated by commas, as in the following list of integers:

[5, 7, 9]

[] is the empty list, which could also be specified with nil.

The Scheme CONS function is implemented as a binary infix operator in ML, represented as ::. For example,

3 :: [5, 7, 9]

returns the following new list: [3, 5, 7, 9].

The elements of a list must be of the same type, so the following list would be illegal:

[5, 7.3, 9]

ML has functions that correspond to Scheme’s CAR and CDR, named hd (head) and tl (tail). For example,

hd [5, 7, 9] is 5

tl [5, 7, 9] is [7, 9]

Lists and list operations in Scheme and ML are more fully discussed in Chapter 15.

Lists in F# are related to those of ML with a few notable differences. ​- Elements of a list in F# are separated by semicolons, rather than the commas of ML. The operations hd and tl are the same, but they are called as methods of the List class, as in List.hd [1; 3; 5; 7], which returns 1. The CONS operation of F# is specified as two colons, as in ML.

Python includes a list data type, which also serves as Python’s arrays. Unlike the lists of Scheme, Common Lisp, ML, and F#, the lists of Python are mutable. They can contain any data value or object. A Python list is created with an assignment of a list value to a name. A list value is a sequence of expressions that are separated by commas and delimited with brackets. For example, consider the following statement:

myList = [3, 5.8, "grape"]

The elements of a list are referenced with subscripts in brackets, as in the following example:

x = myList[1]

This statement assigns 5.8 to x. The elements of a list are indexed starting at zero. List elements also can be updated by assignment. A list element can be deleted with del, as in the following statement:

del myList[1]

This statement removes the second element of myList.

Python includes a powerful mechanism for creating arrays called list comprehensions. A list comprehension is an idea derived from set notation. It first appeared in the functional programming language Haskell (see

Chapter 15). The mechanics of a list comprehension is that a function is applied to each of the elements of a given array and a new array is constructed from the results. The syntax of a Python list comprehension is as follows:

[expression for iterate_var in array if condition]

Consider the following example:

[x * x for x in range(12) if x % 3 == 0]

The range function creates the array [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]. The conditional filters out all numbers in the array that are not evenly divisible by 3. Then, the expression squares the remaining numbers. The results of the squaring are collected in an array, which is returned. This list comprehension returns the following array:

[0, 9, 36, 81]

Slices of lists are also supported in Python.

Haskell’s list comprehensions have the following form:

[body | qualifiers]

For example, consider the following definition of a list:

[n * n | n <- [1..10]]

This defines a list of the squares of the numbers from 1 to 10.

F# includes list comprehensions, which in that language can also be used to create arrays. For example, consider the following statement:

let myArray = [|for i in 1 .. 5 -> (i * i) |];;

This statement creates the array [1; 4; 9; 16; 25] and names it myArray.

Recall from Section 6.5 that C# and Java support generic heap-dynamic collection classes, List and ArrayList, respectively. These structures are

actually lists.

6.10 Union Types A union is a type whose variables may store different type values at different times during program execution. As an example of the need for a union type, consider a table of constants for a compiler, which is used to store the constants found in a program being compiled. One field of each table entry is for the value of the constant. Suppose that for a particular language being compiled, the types of constants were integer, floating-point, and Boolean. In terms of table management, it would be convenient if the same location, a table field, could store a value of any of these three types. Then all constant values could be addressed in the same way. The type of such a location is, in a sense, the union of the three value types it can store.

6.10.1 Design Issues The problem of type checking union types, which is discussed in Section 6.12, is their major design issue.

6.10.2 Discriminated Versus Free Unions C and C++ provide union constructs in which there is no language support for type checking. In C and C++, the union construct is used to specify union structures. The unions in these languages are called free unions, because programmers are allowed complete freedom from type checking in their use. For example, consider the following C union:

union flexType {

int intEl;

float floatEl;

};

union flexType el1;

float x;

. . .

el1.intEl = 27;

x = el1.floatEl;

This last assignment is not type checked, because the system cannot determine the current type of the current value of el1, so it assigns the bit string representation of 27 to the float variable x, which, of course, is nonsense.

Type checking of unions requires that each union construct include a type indicator. Such an indicator is called a tag, or discriminant, and a union with a discriminant is called a discriminated union. The first language to provide discriminated unions was ALGOL 68. They are now supported by ML, Haskell, and F#.

6.10.3 Unions in F# A union is declared in F# with a type statement using OR operators (|) to define the components. For example, we could have the following:

type intReal =

| IntValue of int

| RealValue of float;;

In this example, intReal is the union type. IntValue and RealValue are constructors. Values of type intReal can be created using the constructors as if they were a function, as in the following examples:7

7. The let statement is used to assign values to names and to create a static scope; the double semicolons are used to terminate statements when the F# interactive interpreter is being used.

let ir1 = IntValue 17;;

let ir2 = RealValue 3.4;;

Accessing the value of a union is done with a pattern-matching structure. Pattern matching in F# is specified with the match reserved word. The general form of the construct is as follows:

match pattern with

 | expression_list1−>expression1

 | . . .

 | expression_listn−>expressionn

The pattern can be any data type. The expression list can include wild card characters (_) or be solely a wild card character. For example, consider the following match construct:

let a = 7;;

let b = "grape";;

let x = match (a, b) with

| 4, "apple" -> apple

| _, "grape" -> grape

| _ -> fruit;;

To display the type of the intReal union, the following function could be used:

let printType value =

match value with

| IntValue value -> printfn "It is an integer"

| RealValue value -> printfn "It is a float";;

The following lines show calls to this function and the output:

printType ir1;;

It is an integer

printType ir2;;

It is a float

6.10.4 Evaluation Unions are potentially unsafe constructs in some languages. They are one of the reasons why C and C++ are not strongly typed: These languages do not allow type checking of references to their unions. On the other hand, unions can be safely used, as in their design in ML, Haskell, and F#.

Neither Java nor C# includes unions, which may be reflective of the growing concern for safety in some programming languages.

6.10.5 Implementation of Union Types Unions are implemented by simply using the same address for every possible variant. Sufficient storage for the largest variant is allocated.

6.11 Pointer and Reference Types A pointer type is one in which the variables have a range of values that consists of memory addresses and a special value, nil. The value nil is not a valid address and is used to indicate that a pointer cannot currently be used to reference a memory cell.

Pointers are designed for two distinct kinds of uses. First, pointers provide some of the power of indirect addressing, which is frequently used in assembly language programming. Second, pointers provide a way to manage dynamic storage. A pointer can be used to access a location in an area where storage is dynamically allocated called a heap.

Variables that are dynamically allocated from the heap are called heap-​- dynamic variables. They often do not have identifiers associated with them and thus can be referenced only by pointer or reference type variables. ​- Variables without names are called anonymous variables. It is in this latter application area of pointers that the most important design issues arise.

Pointers, unlike arrays and records, are not structured types, although they are defined using a type operator (* in C and C++). Furthermore, they are also different from scalar variables because they are used to reference some other variable, rather than being used to store data. These two categories of variables are called reference types and value types, respectively.

Both kinds of uses of pointers add writability to a language. For example, suppose it is necessary to implement a dynamic structure like a binary tree in a language that does not have pointers or dynamic storage. This would require the programmer to provide and maintain a pool of available tree nodes, which would probably be implemented in parallel arrays. Also, it would be necessary for the programmer to guess the maximum number of required nodes. This is clearly an awkward and error-prone way to deal with binary trees.

Reference variables, which are discussed in Section 6.11.6, are closely related

to pointers.

6.11.1 Design Issues The primary design issues particular to pointers are the following:

What are the scope and lifetime of a pointer variable?

What is the lifetime of a heap-dynamic variable (the value a pointer references)?

Are pointers restricted as to the type of value to which they can point?

Are pointers used for dynamic storage management, indirect addressing, or both?

Should the language support pointer types, reference types, or both?

6.11.2 Pointer Operations Languages that provide a pointer type usually include two fundamental pointer operations: assignment and dereferencing. The first operation sets a pointer variable’s value to some useful address. If pointer variables are used only to manage dynamic storage, then the allocation mechanism, whether by operator or built-in subprogram, serves to initialize the pointer variable. If pointers are used for indirect addressing to variables that are not heap dynamic, then there must be an explicit operator or built-in subprogram for fetching the address of a variable, which can then be assigned to the pointer variable.

An occurrence of a pointer variable in an expression can be interpreted in two distinct ways. First, it could be interpreted as a reference to the contents of the memory cell to which it is bound, which in the case of a pointer is an address. This is exactly how a nonpointer variable in an expression would be interpreted, although in that case its value likely would not be an address.

However, the pointer also could be interpreted as a reference to the value in the memory cell pointed to by the memory cell to which the pointer variable is bound. In this case, the pointer is interpreted as an indirect reference. The former case is a normal pointer reference; the latter is the result of dereferencing the pointer. Dereferencing, which takes a reference through one level of indirection, is the second fundamental pointer operation.

Dereferencing of pointers can be either explicit or implicit. In many contemporary languages, it occurs only when explicitly specified. In C++, it is explicitly specified with the asterisk (*) as a prefix unary operator. Consider the following example of dereferencing: If ptr is a pointer variable with the value 7080 and the cell whose address is 7080 has the value 206, then the assignment

j = *ptr

sets j to 206. This process is shown in Figure 6.8.

Figure 6.8 The assignment ​-

operation j = *ptr Figure 6.8 Full Alternative Text

When pointers point to records, the syntax of the references to the fields of these records varies among languages. In C and C++, there are two ways a pointer to a record can be used to reference a field in that record. If a pointer variable p points to a record with a field named age, (*p).age can be used to refer to that field. The operator ->, when used between a pointer to a struct and a field of that struct, combines dereferencing and field reference. For example, the expression p -> age is equivalent to (*p).age.

Languages that provide pointers for the management of a heap must include an explicit allocation operation. Allocation is sometimes specified with a subprogram, such as malloc in C. In languages that support object-oriented programming, allocation of heap objects is often specified with the new operator. C++, which does not provide implicit deallocation, uses delete as its deallocation operator.

6.11.3 Pointer Problems The first high-level programming language to include pointer variables was PL/I, in which pointers could be used to refer to both heap-dynamic variables and other program variables. The pointers of PL/I were highly flexible, but their use could lead to several kinds of programming errors. Some of the problems of PL/I pointers are also present in the pointers of subsequent languages. Some recent languages, such as Java, have replaced pointers completely with reference types, which, along with implicit deallocation, minimize the primary problems with pointers. A reference type is really only a pointer with restricted operations.

6.11.3.1 Dangling Pointers

A dangling pointer, or dangling reference, is a pointer that contains the address of a heap-dynamic variable that has been deallocated. Dangling pointers are dangerous for several reasons. First, the location being pointed to may have been reallocated to some new heap-dynamic variable. If the new variable is not the same type as the old one, type checks of uses of the dangling pointer are invalid. Even if the new dynamic variable is the same type, its new value will have no relationship to the old pointer’s dereferenced value. Furthermore, if the dangling pointer is used to change the heap- dynamic variable, the value of the new heap-dynamic variable will be destroyed. Finally, it is possible that the location now is being temporarily used by the storage management system, possibly as a pointer in a chain of available blocks of storage, thereby allowing a change to the location to cause the storage manager to fail.

The following sequence of operations creates a dangling pointer in many languages:

1. A new heap-dynamic variable is created and pointer p1 is set to point to it.

2. Pointer p2 is assigned p1’s value.

3. The heap-dynamic variable pointed to by p1 is explicitly deallocated (possibly setting p1 to nil), but p2 is not changed by the operation. p2 is now a dangling pointer. If the deallocation operation did not change p1, both p1 and p2 would be dangling. (Of course, this is a problem of aliasing—p1 and p2 are aliases.)

For example, in C++ we could have the following:

int * arrayPtr1;

int * arrayPtr2 = new int[100];

arrayPtr1 = arrayPtr2;

delete [] arrayPtr2;

// Now, arrayPtr1 is dangling, because the heap storage

// to which it was pointing has been deallocated.

In C++, both arrayPtr1 and arrayPtr2 are now dangling pointers, because the C++ delete operator has no effect on the value of its operand pointer. In

C++, it is common (and safe) to follow a delete operator with an assignment of zero, which represents null, to the pointer whose pointed-to value has been deallocated.

Notice that the explicit deallocation of dynamic variables is the cause of dangling pointers.

history note Pascal included an explicit deallocate operator: dispose. Because of the problem of dangling pointers caused by dispose, some Pascal implementations simply ignored dispose when it appeared in a program. Although this effectively prevents dangling pointers, it also disallows the reuse of heap storage that the program no longer needs. Recall that Pascal initially was designed as a teaching language, rather than as an industrial tool.

6.11.3.2 Lost Heap-Dynamic Variables A lost heap-dynamic variable is an allocated heap-dynamic variable that is no longer accessible to the user program. Such variables are often called garbage, because they are not useful for their original purpose, and they also cannot be reallocated for some new use in the program. Lost heap-dynamic variables are most often created by the following sequence of operations:

1. Pointer p1 is set to point to a newly created heap-​dynamic variable.

2. p1 is later set to point to another newly created heap-​dynamic variable.

The first heap-dynamic variable is now inaccessible, or lost. This is sometimes called memory leakage. Memory leakage is a problem, regardless of whether the language uses implicit or explicit deallocation. In the following sections, we investigate how language designers have dealt with the problems of dangling pointers and lost heap-dynamic variables.

6.11.4 Pointers in C and C++ In C and C++, pointers can be used in the same ways as addresses are used in assembly languages. This means they are extremely flexible but must be used with great care. This design offers no solutions to the dangling pointer or lost heap-dynamic variable problems. However, the fact that pointer arithmetic is possible in C and C++ makes their pointers more interesting than those of the other programming languages.

C and C++ pointers can point at any variable, regardless of where it is allocated. In fact, they can point anywhere in memory, whether there is a variable there or not, which is one of the dangers of such pointers.

In C and C++, the asterisk (*) denotes the dereferencing operation and the ampersand (&) denotes the operator for producing the address of a variable. For example, consider the following code:

int *ptr;

int count, init;

. . .

ptr = &init;

count = *ptr;

The assignment to the variable ptr sets it to the address of init. The assignment to count dereferences ptr to produce the value at init, which is then assigned to count. So, the effect of the two assignment statements is to assign the value of init to count. Notice that the declaration of a pointer specifies its domain type.

Pointers can be assigned the address value of any variable of the correct domain type, or they can be assigned the constant zero, which is used for nil.

Pointer arithmetic is also possible in some restricted forms. For example, if ptr is a pointer variable that is declared to point at some variable of some data type, then

ptr + index

is a legal expression. The semantics of such an expression is as follows. Instead of simply adding the value of index to ptr, the value of index is first scaled by the size of the memory cell (in memory units) to which ptr is pointing (its base type). For example, if ptr points to a memory cell for a type that is four memory units in size, then index is multiplied by 4, and the result is added to ptr. The primary purpose of this sort of address arithmetic is array manipulation. The following discussion is related to single-​- dimensioned arrays only.

In C and C++, all arrays use zero as the lower bound of their subscript ranges, and array names without subscripts always refer to the address of the first element. Consider the following declarations:

int list [10];

int *ptr;

Now consider the assignment

ptr = list;

This assigns the address of list[0] to ptr. Given this assignment, the following are true:

*(ptr + 1) is equivalent to list[1].

*(ptr + index) is equivalent to list[index].

ptr[index] is equivalent to list[index].

It is clear from these statements that the pointer operations include the same scaling that is used in indexing operations. Furthermore, pointers to arrays can be indexed as if they were array names.

Pointers in C and C++ can point to functions. This feature is used to pass functions as parameters to other functions. Pointers are also used for parameter passing, as discussed in Chapter 9.

C and C++ include pointers of type void *, which can point at values of any type. In effect they are generic pointers. However, type checking is not a

problem with void * pointers, because these languages disallow dereferencing them. One common use of void * pointers is as the types of parameters of functions that operate on memory. For example, suppose we wanted a function to move a sequence of bytes of data from one place in memory to another. It would be most general if it could be passed two pointers of any type. This would be legal if the corresponding formal parameters in the function were void * type. The function could then convert them to char * type and do the operation, regardless of what type pointers were sent as actual parameters.

6.11.5 Reference Types A reference type variable is similar to a pointer, with one important and fundamental difference: A pointer refers to an address in memory, while a reference refers to an object or a value in memory. As a result, although it is natural to perform arithmetic on addresses, it is not sensible to do arithmetic on references.

C++ includes a special kind of reference type that is used primarily for the formal parameters in function definitions. A C++ reference type variable is a constant pointer that is always implicitly dereferenced. Because a C++ reference type variable is a constant, it must be initialized with the address of some variable in its definition, and after initialization a reference type variable can never be set to reference any other variable. The implicit dereference prevents assignment to the address value of a reference variable.

Reference type variables are specified in definitions by preceding their names with ampersands (&). For example,

int result = 0;

int &ref_result = result;

. . .

ref_result = 100;

In this code segment, result and ref_result are aliases.

When used as formal parameters in function definitions, C++ reference types

provide for two-way communication between the caller function and the called function. This is not possible with nonpointer primitive parameter types, because C++ parameters are passed by value. Passing a pointer as a parameter accomplishes the same two-way communication, but pointer formal parameters require explicit dereferencing, making the code less readable and less safe. Reference parameters are referenced in the called function exactly as are other parameters. The calling function need not specify that a parameter whose corresponding formal parameter is a reference type is anything unusual. The compiler passes addresses, rather than values, to reference parameters.

In their quest for increased safety over C++, the designers of Java removed C++-style pointers altogether. Unlike C++ reference variables, Java reference variables can be assigned to refer to different class instances; they are not constants. All Java class instances are referenced by reference variables. That is, in fact, the only use of reference variables in Java. These issues are discussed further in Chapter 12.

In the following, String is a standard Java class:

String str1;

. . .

str1 = "This is a Java literal string";

In this code, str1 is defined to be a reference to a String class instance or object. It is initially set to null. The subsequent assignment sets str1 to reference the String object, "This is a Java literal string".

Because Java class instances are implicitly deallocated (there is no explicit deallocation operator), there cannot be dangling references in Java.

C# includes both the references of Java and the pointers of C++. However, the use of pointers is strongly discouraged. In fact, any subprogram that uses pointers must include the unsafe modifier. Note that although objects pointed to by references are implicitly deallocated, that is not true for objects pointed to by pointers. Pointers were included in C# primarily to allow C# programs to interoperate with C and C++ code.

All variables in the object-oriented languages Smalltalk, Python, and Ruby

are references. They are always implicitly dereferenced. Furthermore, the direct values of these variables cannot be accessed.

6.11.6 Evaluation The problems of dangling pointers and garbage have already been discussed at length. The problems of heap management are discussed in Section 6.11.7.3.

Pointers have been compared with the goto. The goto statement widens the range of statements that can be executed next. Pointer variables widen the range of memory cells that can be referenced by a variable. Perhaps the most damning statement about pointers was made by Hoare (1973): “Their introduction into high-level languages has been a step backward from which we may never recover.”

On the other hand, pointers are essential in some kinds of programming applications. For example, pointers are necessary to write device drivers, in which specific absolute addresses must be accessed.

The references of Java and C# provide some of the flexibility and the capabilities of pointers, without the hazards. It remains to be seen whether programmers will be willing to trade the full power of C and C++ pointers for the greater safety of references. The extent to which C# programs use pointers will be one measure of this.

6.11.7 Implementation of Pointer and Reference Types In most languages, pointers are used in heap management. The same is true for Java and C# references, as well as the variables in Smalltalk and Ruby, so we cannot treat pointers and references separately. First, we briefly describe how pointers and references are represented internally. We then discuss two possible solutions to the dangling pointer problem. Finally, we describe the

major problems with heap-management techniques.

6.11.7.1 Representations of Pointers and References In most larger computers, pointers and references are single values stored in memory cells. However, in early microcomputers based on Intel microprocessors, addresses have two parts: a segment and an offset. So, pointers and references are implemented in these systems as pairs of 16-bit cells, one for each of the two parts of an address.

6.11.7.2 Solutions to the Dangling- Pointer Problem There have been several proposed solutions to the dangling-pointer problem. Among these are tombstones (Lomet, 1975), in which every heap-dynamic variable includes a special cell, called a tombstone, that is itself a pointer to the heap-dynamic variable. The actual pointer variable points only at tombstones and never to heap-dynamic variables. When a heap-dynamic variable is deallocated, the tombstone remains but is set to nil, indicating that the heap-dynamic variable no longer exists. This approach prevents a pointer from ever pointing to a deallocated variable. Any reference to any pointer that points to a nil tombstone can be detected as an error.

Tombstones are costly in both time and space. Because tombstones are never deallocated, their storage is never reclaimed. Every access to a heap-​dynamic variable through a tombstone requires one more level of indirection, which requires an additional machine cycle on most computers. Apparently none of the designers of the more popular languages have found the additional safety to be worth this additional cost, because no widely used language uses tombstones.

An alternative to tombstones is the locks-and-keys approach used in the implementation of UW-Pascal (Fischer and LeBlanc, 1977, 1980). In this compiler, pointer values are represented as ordered pairs (key, address), where the key is an integer value. Heap-dynamic variables are represented as the storage for the variable plus a header cell that stores an integer lock value. When a heap-dynamic variable is allocated, a lock value is created and placed both in the lock cell of the heap-dynamic variable and in the key cell of the pointer that is specified in the call to new. Every access to the dereferenced pointer compares the key value of the pointer to the lock value in the heap- dynamic variable. If they match, the access is legal; otherwise the access is treated as a run-time error. Any copies of the pointer value to other pointers must copy the key value. Therefore, any number of pointers can reference a given heap-​dynamic variable. When a heap-dynamic variable is deallocated with dispose, its lock value is cleared to an illegal lock value. Then, if a pointer other than the one specified in the dispose is dereferenced, its address value will still be intact, but its key value will no longer match the lock, so the access will not be allowed.

Of course, the best solution to the dangling-pointer problem is to take deallocation of heap-dynamic variables out of the hands of programmers. If programs cannot explicitly deallocate heap-dynamic variables, there will be no dangling pointers. To do this, the run-time system must implicitly deallocate heap-dynamic variables when they are no longer useful. Lisp systems have always done this. Both Java and C# also use this approach for their reference variables. Recall that C#’s pointers do not include implicit deallocation.

6.11.7.3 Heap Management Heap management can be a very complex run-time process. We examine the process in two separate situations: one in which all heap storage is allocated and deallocated in units of a single size, and one in which variable-size segments are allocated and deallocated. Note that for deallocation, we discuss only implicit approaches. Our discussion will be brief and far from comprehensive, since a thorough analysis of these processes and their associated problems is not so much a language design issue as it is an

implementation issue.

Single-Size Cells  The simplest situation is when all allocation and deallocation is of single-size cells. It is further simplified when every cell already contains a pointer. This is the scenario of many implementations of Lisp, where the problems of dynamic storage allocation were first encountered on a large scale. All Lisp programs and most Lisp data consist of cells in linked lists.

In a single-size allocation heap, all available cells are linked together using the pointers in the cells, forming a list of available space. Allocation is a ​- simple matter of taking the required number of cells from this list when they are needed. Deallocation is a much more complex process. A heap-dynamic ​- variable can be pointed to by more than one pointer, making it difficult to determine when the variable is no longer useful to the program. Simply because one pointer is disconnected from a cell obviously does not make it garbage; there could be several other pointers still pointing to the cell.

In Lisp, several of the most frequent operations in programs create collections of cells that are no longer accessible to the program and therefore should be deallocated (put back on the list of available space). One of the fundamental design goals of Lisp was to ensure that reclamation of unused cells would not be the task of the programmer but rather that of the run-time system. This goal left Lisp implementors with the fundamental design question: When should deallocation be performed?

There are several different approaches to garbage collection. The two most common traditional techniques are in some ways opposite processes. These are named reference counters, in which reclamation is incremental and is done when inaccessible cells are created, and mark-sweep, in which reclamation occurs only when the list of available space becomes empty. These two methods are sometimes called the eager approach and the lazy approach, respectively. Many variations of these two approaches have been developed. In this section, however, we discuss only the basic processes.

The reference counter method of storage reclamation accomplishes its goal by maintaining in every cell a counter that stores the number of pointers that are currently pointing at the cell. Embedded in the decrement operation for the reference counters, which occurs when a pointer is disconnected from the cell, is a check for a zero value. If the reference counter reaches zero, it means that no program pointers are pointing at the cell, and it has thus become garbage and can be returned to the list of available space.

There are three distinct problems with the reference counter method. First, if storage cells are relatively small, the space required for the counters is significant. Second, some execution time is obviously required to maintain the counter values. Every time a pointer value is changed, the cell to which it was pointing must have its counter decremented, and the cell to which it is now pointing must have its counter incremented. In a language like Lisp, in which nearly every action involves changing pointers, that can be a significant portion of the total execution time of a program. Of course, if pointer changes are not too frequent, this is not a problem. Some of the inefficiency of reference counters can be eliminated by an approach named deferred reference ​counting, which avoids reference counters for some pointers. The third problem is that complications arise when a collection of cells is connected circularly. The problem here is that each cell in the circular list has a reference counter value of at least 1, which prevents it from being collected and placed back on the list of available space. A solution to this problem can be found in Friedman and Wise (1979).

The advantage of the reference counter approach is that it is intrinsically incremental. Its actions are interleaved with those of the application, so it never causes significant delays in the execution of the application.

The original mark-sweep process of garbage collection operates as follows: The run-time system allocates storage cells as requested and disconnects pointers from cells as necessary, without regard for storage reclamation (allowing garbage to accumulate), until it has allocated all available cells. At this point, a mark-sweep process is begun to gather all the garbage left floating around in the heap. To facilitate the process, every heap cell has an extra indicator bit or field that is used by the collection algorithm.

The mark-sweep process consists of three distinct phases. First, all cells in

the heap have their indicators set to indicate they are garbage. This is, of course, a correct assumption for only some of the cells. The second part, called the marking phase, is the most difficult. Every pointer in the program is traced into the heap, and all reachable cells are marked as not being garbage. After this, the third phase, called the sweep phase, is executed: All cells in the heap that have not been specifically marked as still being used are returned to the list of available space.

To illustrate the flavor of algorithms used to mark the cells that are currently in use, we provide the following simple version of a marking algorithm. We assume that all heap-dynamic variables, or heap cells, consist of an information part; a part for the mark, named marker; and two pointers named llink and rlink. These cells are used to build directed graphs with at most two edges leading from any node. The marking algorithm traverses all spanning trees of the graphs, marking all cells that are found. Like other graph traversals, the marking algorithm uses recursion.

for every pointer r do

mark(r)

void mark(void * ptr) {

if (ptr != 0)

if (*ptr.marker is not marked) {

set *ptr.marker

mark(*ptr.llink)

mark(*ptr.rlink)

    } }

An example of the actions of this procedure on a given graph is shown in Figure 6.9. This simple marking algorithm requires a great deal of storage (for stack space to support recursion). A marking process that does not require additional stack space was developed by Schorr and Waite (1967). Their method reverses pointers as it traces out linked structures. Then, when the end of a list is reached, the process can follow the pointers back out of the structure.

Figure 6.9 An example of the actions of the marking algorithm

Figure 6.9 Full Alternative Text

The most serious problem with the original version of mark-sweep was that it was done too infrequently—only when a program had used all or nearly all of the heap storage. Mark-sweep in that situation takes a good deal of time, because most of the cells must be traced and marked as being currently used. This causes a significant delay in the progress of the application. Furthermore, the process may yield only a small number of cells that can be placed on the list of available space. This problem has been addressed in a

variety of improvements. For example, incremental mark-sweep garbage collection occurs more frequently, long before memory is exhausted, making the process more effective in terms of the amount of storage that is reclaimed. Also, the time required for each run of the process is obviously shorter, thus reducing the delay in application execution. Another alternative is to perform the mark-sweep process on parts, rather than all of the memory associated with the application, at different times. This provides the same kinds of improvements as incremental mark-sweep.

Both the marking algorithms for the mark-sweep method and the processes required by the reference counter method can be made more efficient by use of the pointer rotation and slide operations that are described by Suzuki (1982).

Variable-Size Cells  Managing a heap from which variable-size cells8 are allocated has all the difficulties of managing one for single-size cells, but also has additional problems. Unfortunately, variable-size cells are required by most programming languages. The additional problems posed by variable-size cell management depend on the method used. If mark-sweep is used, the following additional problems occur:

8. The cells have variable sizes because these are abstract cells, which store the values of ​variables, regardless of their types. Furthermore, a variable could be a structured type.

The initial setting of the indicators of all cells in the heap to indicate that they are garbage is difficult. Because the cells are different sizes, scanning them is a problem. One solution is to require each cell to have the cell size as its first field. Then the scanning can be done, although it takes slightly more space and somewhat more time than its counterpart for fixed-size cells.

The marking process is nontrivial. How can a chain be followed from a pointer if there is no predefined location for the pointer in the pointed-to

cell? Cells that do not contain pointers at all are also a problem. Adding an internal pointer to each cell, which is maintained in the background by the run-time system, will work. However, this background maintenance processing adds both space and execution time overhead to the cost of running the program.

Maintaining the list of available space is another source of overhead. The list can begin with a single cell consisting of all available space. Requests for segments simply reduce the size of this block. Reclaimed cells are added to the list. The problem is that before long, the list becomes a long list of various-size segments, or blocks. This slows allocation because requests cause the list to be searched for sufficiently large blocks. Eventually, the list may consist of a large number of very small blocks, which are not large enough for most requests. At this point, adjacent blocks may need to be collapsed into larger blocks. Alternatives to using the first sufficiently large block on the list can shorten the search but require the list to be ordered by block size. In either case, maintaining the list is additional overhead.

If reference counters are used, the first two problems are avoided, but the available-space list-maintenance problem remains.

For a comprehensive study of memory management problems, see Wilson (2005).

6.12 Optional Types There are situations in programming when there is a need to be able to indicate that a variable does not currently have a value. Some older languages use zero as a nonvalue for numeric variables. This approach has the disadvantage of not being able to distinguish between when the variable is supposed to have the zero value and when the zero indicates that it has no value. Some newer languages provide types that can have a normal value or a special value to indicate that their variables have no value. Variables that have this capability are called optional types. Optional types are now directly supported in C#, F#, and Swift, among others.

C# has two categories of variables, value and reference types. Reference types, which are classes, are optional types by their nature. The null value indicates that a reference type has no value. Value types, which are all struct types, can be declared to be optional types, which allows them to have the value null. A variable is declared to be an optional type by following its type name with a question mark (?), as in

int? x;

To determine whether a variable has a normal value, it can be tested against null, as in

int? x;

. . .

if(x == null)

Console.WriteLine("x has no value");

else

Console.WriteLine("The value of x is: {0}", x);

Swift’s optional types are similar to those of C#, except that the nonvalue is named nil, instead of null. The Swift version of the above code is:

var Int? x;

. . .

if  x  ==  nil   print("x has no value")

else

  print("The value of x is: \(x)")

6.13 Type Checking For our discussion of type checking, the concept of operands and operators is generalized to include subprograms and assignment statements. Subprograms will be thought of as operators whose operands are their parameters. The assignment symbol will be thought of as a binary operator, with its target variable and its expression being the operands.

Type checking is the activity of ensuring that the operands of an operator are of compatible types. A compatible type is one that either is legal for the operator or is allowed under language rules to be implicitly converted by compiler-​generated code (or the interpreter) to a legal type. This automatic conversion is called a coercion. For example, if an int variable and a float variable are added in Java, the value of the int variable is coerced to float and a floating-point add is done.

A type error is the application of an operator to an operand of an inappropriate type. For example, in the original version of C, if an int value was passed to a function that expected a float value, a type error would occur (because compilers for that language did not check the types of parameters).

If all bindings of variables to types are static in a language, then type checking can nearly always be done statically. Dynamic type binding requires type checking at run time, which is called dynamic type checking.

Some languages, such as JavaScript and PHP, because of their dynamic type binding, allow only dynamic type checking. It is better to detect errors at compile time than at run time, because the earlier correction is usually less costly. The penalty for static checking is reduced programmer flexibility. Fewer shortcuts and tricks are possible. Such techniques, though, are now generally recognized to be error prone and detrimental to readability.

Type checking is complicated when a language allows a memory cell to store values of different types at different times during execution. Such memory

cells can be created with C and C++ unions and the discriminated unions of ML, Haskell, and F#. In these cases, type checking, if done, must be dynamic and requires the run-time system to maintain the type of the current value of such memory cells. So, even though all variables are statically bound to types in languages such as C++, not all type errors can be detected by static type checking.

6.14 Strong Typing One of the ideas in language design that became prominent in the so-called structured-programming revolution of the 1970s was strong typing. Strong typing is widely acknowledged as being a highly valuable language characteristic. Unfortunately, it is often loosely defined, and it is sometimes used in computing literature without being defined at all.

A programming language is strongly typed if type errors are always detected. This requires that the types of all operands can be determined, either at compile time or at run time. The importance of strong typing lies in its ability to detect all misuses of variables that result in type errors. A strongly typed language also allows the detection, at run time, of uses of the incorrect type values in variables that can store values of more than one type.

C and C++ are not strongly typed languages because both include union types, which are not type checked.

ML is strongly typed, even though the types of some function parameters may not be known at compile time. F# is strongly typed.

Java and C#, although they are based on C++, are nearly strongly typed. Types can be explicitly cast, which could result in a type error. However, there are no implicit ways type errors can go undetected.

The coercion rules of a language have an important effect on the value of type checking. For example, expressions are strongly typed in Java. However, an arithmetic operator with one floating-point operand and one integer operand is legal. The value of the integer operand is coerced to floating-point, and a floating-point operation takes place. This is what is usually intended by the programmer. However, the coercion also results in a loss of one of the benefits of strong typing—error detection. For example, suppose a program had the int variables a and b and the float variable d. Now, if a programmer meant to type a + b, but mistakenly typed a + d, the error would not be detected by the compiler. The value of a would simply be

coerced to float. So, the value of strong typing is weakened by coercion. Languages with a great deal of coercion, like C, and C++, are less reliable than those with no coercion, such as ML and F#. Java and C# have half as many assignment type coercions as C++, so their error detection is better than that of C++, but still not nearly as effective as that of ML and F#. The issue of coercion is examined in detail in Chapter 7.

6.15 Type Equivalence The idea of type compatibility was defined when the issue of type checking was introduced. The compatibility rules dictate the types of operands that are acceptable for each of the operators and thereby specify the possible type errors of the language.9 The rules are called compatibility because in some cases the type of an operand can be implicitly converted by the compiler or run-time system to make it acceptable for the operator.

9. Type compatibility is also an issue in the relationship between the actual parameters in a subprogram call and the formal parameters of the subprogram definition. This issue is ​discussed in Chapter 9.

The type compatibility rules are simple and rigid for the predefined scalar types. However, in the cases of structured types, such as arrays and records and some user-defined types, the rules are more complex. Coercion of these types is rare, so the issue is not type compatibility, but type equivalence. That is, two types are equivalent if an operand of one type in an expression can be substituted for one of the other type, without coercion. Type equivalence is a strict form of type compatibility—compatibility without coercion. The central issue here is how type equivalence is defined.

The design of the type equivalence rules of a language is important, because it influences the design of the data types and the operations provided for values of those types. With the types discussed here, there are very few predefined operations. Perhaps the most important result of two variables being of equivalent types is that either one can have its value assigned to the other.

There are two approaches to defining type equivalence: name type equivalence and structure type equivalence. Name type equivalence means that two variables have equivalent types if they are defined either in the same declaration or in declarations that use the same type name. Structure type equivalence means that two variables have equivalent types if their types have identical structures. There are some variations of these two approaches,

and many languages use combinations of them.

Name type equivalence is easy to implement but is more restrictive. Under a strict interpretation, a variable whose type is a subrange of the integers would not be equivalent to an integer type variable. For example, supposing Ada used strict name type equivalence, consider the following Ada code:

type Indextype is 1..100;

count : Integer;

index : Indextype;

The types of the variables count and index would not be equivalent; count could not be assigned to index or vice versa.

Another problem with name type equivalence arises when a structured or user-defined type is passed among subprograms through parameters. Such a type must be defined only once, globally. A subprogram cannot state the type of such formal parameters in local terms. This was the case with the original version of Pascal.

Note that to use name type equivalence, all types must have names. Most languages allow users to define types that are anonymous—they do not have names. For a language to use name type equivalence, such types must implicitly be given internal names by the compiler.

Structure type equivalence is more flexible than name type equivalence, but it is more difficult to implement. Under name type equivalence, only the two type names must be compared to determine equivalence. Under structure type equivalence, however, the entire structures of the two types must be compared. This comparison is not always simple. (Consider a data structure that refers to its own type, such as a linked list.) Other questions can also arise. For example, are two record (or struct) types equivalent if they have the same structure but different field names? Are two single-dimensioned array types in a language that allows lower bounds of array subscript ranges to be set in their declarations equivalent if they have the same element type but have subscript ranges of 0..10 and 1..11? Are two enumeration types equivalent if they have the same number of components but spell the literals differently?

Another difficulty with structure type equivalence is that it disallows differentiating between types with the same structure. For example, consider the following Ada-like declarations:

type Celsius = Float;

Fahrenheit = Float;

The types of variables of these two types are considered equivalent under structure type equivalence, allowing them to be mixed in expressions, which is surely undesirable in this case, considering the difference indicated by the type’s names. In general, types with different names are likely to be abstractions of different categories of problem values and should not be considered equivalent.

Ada uses a restrictive form of name type equivalence but provides two type constructs, subtypes and derived types, that avoid the problems associated with name type equivalence. A derived type is a new type that is based on some previously defined type with which it is not equivalent, although it may have identical structure. Derived types inherit all the properties of their parent types. Consider the following example:

type Celsius is new Float;

type Fahrenheit is new Float;

The types of variables of these two derived types are not equivalent, although their structures are identical. Furthermore, variables of neither type is type equivalent with any other floating-point type. Literals are exempt from the rule. A literal such as 3.0 has the type universal real and is type equivalent to any floating-point type. Derived types can also include range constraints on the parent type, while still inheriting all of the parent’s operations.

An Ada subtype is a possibly range-constrained version of an existing type. A subtype is type equivalent with its parent type. For example, consider the following declaration:

subtype Small_type is Integer range 0..99;

The type Small_type is equivalent to the type Integer.

Note that Ada’s derived types are very different from Ada’s subrange types. For example, consider the following type declarations:

type Derived_Small_Int is new Integer range 1..100;

subtype Subrange_Small_Int is Integer range 1..100;

Variables of both types, Derived_Small_Int and Subrange_Small_Int, have the same range of legal values and both inherit the operations of Integer. However, variables of type Derived_Small_Int are not compatible with any Integer type. On the other hand, variables of type Subrange_Small_Int are compatible with variables and constants of Integer type and any subtype of Integer.

For variables of an Ada unconstrained array type, structure type equivalence is used. For example, consider the following type declaration and two object declarations:

type Vector is array (Integer range <>) of Integer;

Vector_1: Vector (1..10);

Vector_2: Vector (11..20);

The types of these two objects are equivalent, even though they have different names and different subscript ranges, because for objects of unconstrained array types, structure type equivalence rather than name type equivalence is used. Because both types have 10 elements and the elements of both are of type Integer, they are type equivalent.

For constrained anonymous types, Ada uses a highly restrictive form of name type equivalence. Consider the following Ada declarations of constrained anonymous types:

A : array (1..10) of Integer;

In this case, A has an anonymous but unique type assigned by the compiler and unavailable to the program. If we also had

B : array (1..10) of Integer;

A and B would be of anonymous but distinct and not equivalent types, though they are structurally identical. The multiple declaration

C, D : array (1..10) of Integer;

creates two anonymous types, one for C and one for D, which are not equivalent. This declaration is actually treated as if it were the following two declarations:

C : array (1..10) of Integer;

D : array (1..10) of Integer;

Note that Ada’s form of name type equivalence is more restrictive than the name type equivalence that is defined at the beginning of this section. If we had written instead

type List_10 is array (1..10) of Integer;

C, D : List_10;

then the types of C and D would be equivalent.

Name type equivalence works well for Ada, in part because all types, except anonymous arrays, are required to have type names (and anonymous types are given internal names by the compiler).

Type equivalence rules for Ada are more rigid than those for languages that have many coercions among types. For example, the two operands of an addition operator in Java can have virtually any combination of numeric types in the language. One of the operands will simply be coerced to the type of the other. But in Ada, there are no coercions of the operands of an arithmetic operator.

C uses both name and structure type equivalence. Every struct, enum, and union declaration creates a new type that is not equivalent to any other type. So, name type equivalence is used for structure, enumeration, and union types. Other nonscalar types use structure type equivalence. Array types are equivalent if they have the same type components. Also, if an array type has a constant size, it is equivalent either to other arrays with the same constant size or to with those without a constant size. Note that typedef in C and C++ does not introduce a new type; it simply defines a new name for an existing type. So, any type defined with typedef is type equivalent to its parent type. One exception to C using name type equivalence for structures,

enumerations, and unions is if two structures, enumerations, or unions are defined in different files, in which case structural type equivalence is used. This is a loophole in the name type equivalence rule to allow equivalence of structures, enumerations, and unions that are defined in different files.

C++ is like C except there is no exception for structures and unions defined in different files.

In languages that do not allow users to define and name types, such as Fortran and COBOL, name equivalence obviously cannot be used.

Object-oriented languages such as Java and C++ bring another kind of type compatibility issue with them. The issue is object compatibility and its relationship to the inheritance hierarchy, which is discussed in Chapter 12.

Type compatibility in expressions is discussed in Chapter 7; type compatibility for subprogram parameters is discussed in Chapter 9.

6.16 Theory and Data Types Type theory is a broad area of study in mathematics, logic, computer science, and philosophy. It began in mathematics in the early 1900s and later became a standard tool in logic. Any general discussion of type theory is necessarily complex, lengthy, and highly abstract. Even when restricted to computer science, type theory includes such diverse and complex subjects as typed lambda calculus, combinators, the metatheory of bounded quantification, existential types, and higher-order polymorphism. All these topics are far beyond the scope of this book.

In computer science there are two branches of type theory: practical and abstract. The practical branch is concerned with data types in commercial programming languages; the abstract branch primarily focuses on typed lambda calculus, an area of extensive research by theoretical computer scientists over the past half century. This section is restricted to a brief description of some of the mathematical formalisms that underlie data types in programming languages.

A data type defines a set of values and a collection of operations on those values. A type system is a set of types and the rules that govern their use in programs. Obviously, every typed programming language defines a type system. The formal model of a type system of a programming language consists of a set of types and a collection of functions that define the type rules of the language, which are used to determine the type of any expression. A formal system that describes the rules of a type system, attribute grammars, is introduced in Chapter 3.

An alternative model to attribute grammars uses a type map and a collection of functions, not associated with grammar rules, that specify the type rules. A type map is similar to the state of a program used in denotational semantics, consisting of a set of ordered pairs, with the first element of each pair being a variable’s name and the second element being its type. A type map is constructed using the type declarations in the program. In a static typed language, the type map need only be maintained during compilation,

although it changes as the program is analyzed by the compiler. If any type checking is done dynamically, the type map must be maintained during execution. The concrete version of a type map in a compilation system is the symbol table, constructed primarily by the lexical and syntax analyzers. Dynamic types sometimes are maintained with tags attached to values or objects.

As stated previously, a data type is a set of values, although in a data type the elements are often ordered. For example, the elements in all enumeration types are ordered. However, in a mathematical set the elements are not ordered. Despite this difference, set operations can be used on data types to describe new data types. The structured data types of programming languages are defined by type operators, or constructors that correspond to set operations. These set operations/type constructors are briefly introduced in the following paragraphs.

A finite mapping is a function from a finite set of values, the domain set, onto values in the range set. Finite mappings model two different categories of types in programming languages, functions and arrays, although in some languages functions are not types. All languages include arrays, which are defined in terms of a mapping function that maps indices to elements in the array. For traditional arrays, the mapping is simple—integer values are mapped to the addresses of array elements; for associative arrays, the mapping is defined by a function that describes a hashing operation. The hashing function maps the keys of the associate arrays, usually character strings,10 to the addresses of the array elements.

10. In Ruby, the associative array keys need not be character strings—they can be any type.

A Cartesian, or cross product of n sets, S1, S2, …, Sn, is a set denoted S1×S2×…×Sn. Each element of the Cartesian product set has one element from each of the constituent sets. So, S1×S2={(x, y) | x is in S1 and y is in S2}. For example, if S1={1, 2} and S2={a, b}, S1×S2= {(1, a), (1, b), (2, a), (2, b)}. A Cartesian product defines tuples in mathematics, which appear in Python, ML, Swift, and F# as a data type (see Section 6.5). Cartesian products also model records, or structs, although not exactly. Cartesian products do not have element names, but records require

them. For example, consider the following C struct:

struct intFloat {

int myInt;

float myFloat;

};

This struct defines the Cartesian product type int × float. The names of the elements are myInt and myFloat.

The union of two sets, S1 and S2, is defined as S1 ∪ S2={x | x is in S1 or x is in S2}. Set union models the union data types, as described in Section 6.10.

Mathematical subsets are defined by providing a rule that elements must follow. These sets model the subtypes of Ada, although not exactly, because subtypes must consist of contiguous elements of their parent sets. Elements of mathematical sets are unordered, so the model is not perfect.

Notice that pointers, defined with type operators, such as the asterisk in C, are not defined in terms of a set operation.

This concludes our discussion of formalisms in data types, as well as our whole discussion of data types.

SUMMARY The data types of a language are a large part of what determines that language’s style and usefulness. Along with control structures, they form the heart of a language.

The primitive data types of most imperative languages include numeric, character, and Boolean types. The numeric types are often directly supported by hardware.

The user-defined enumeration and subrange types are convenient and add to the readability and reliability of programs.

Arrays are part of most programming languages. The relationship between a reference to an array element and the address of that element is given in an access function, which is an implementation of a mapping. Arrays can be either static, as in C++ arrays whose definition includes the static specifier; fixed stack-dynamic, as in C functions (without the static specifier); fixed heap-dynamic, as with Java’s objects; or heap dynamic, as in Perl’s arrays. Most languages allow only a few operations on complete arrays.

Records are now included in most languages. Fields of records are specified in a variety of ways. In the case of COBOL, they can be referenced without naming all of the enclosing records, although this is messy to implement and harmful to readability. In several languages that support object-oriented programming, records are supported with objects.

Tuples are similar to records, but do not have names for their constituent. parts. They are part of Python, ML, and F#.

Lists are staples of the functional programming languages, but are now also included in Python and C#.

Unions are locations that can store different type values at different times. Discriminated unions include a tag to record the current type value. A free union is one without the tag. Most languages with unions do not have safe

designs for them, the exceptions being ML, Swift, and F#.

Pointers are used for addressing flexibility and to control dynamic storage management. Pointers have some inherent dangers: Dangling pointers are difficult to avoid, and memory leakage can occur.

Reference types, such as those in Java and C#, provide heap management without the dangers of pointers.

Enumeration and record types are relatively easy to implement. Arrays are also straightforward, although array element access is an expensive process when the array has several subscripts. The access function requires one addition and one multiplication for each subscript.

Pointers are relatively easy to implement, if heap management is not considered. Heap management is easy if all cells have the same size but is complicated for variable-size cell allocation and deallocation.

Optional type variables are variables that allow a nonvalue to be stored. This allows a program to indicate when a variable currently has no value.

Strong typing is the concept of requiring that all type errors be detected. The value of strong typing is increased reliability.

The type equivalence rules of a language determine what operations are legal among the structured types of a language. Name type equivalence and structure type equivalence are the two fundamental approaches to defining type equivalence.

Type theories have been developed in many areas. In computer science, the practical branch of type theory defines the types and type rules of programming languages. Set theory can be used to model most of the structured data types in programming languages.

BIBLIOGRAPHIC NOTES A wealth of literature exists that is concerned with data type design, use, and implementation. Hoare gives one of the earliest systematic definitions of structured types in Dahl et al. (1972). A general discussion of a wide variety of data types is given in Cleaveland (1986).

Implementing run-time checks on the possible insecurities of Pascal data types is discussed in Fischer and LeBlanc (1980). Most compiler design books, such as Fischer and LeBlanc (1991) and Aho et al. (1986), describe implementation methods for data types, as do the other programming language texts, such as Pratt and Zelkowitz (2001) and Scott (2009). A detailed discussion of the problems of heap management can be found in Tenenbaum et al. (1990). Garbage-collection methods are developed by Schorr and Waite (1967) and Deutsch and Bobrow (1976). A comprehensive discussion of garbage-collection algorithms can be found in Cohen (1981) and Wilson (2005).

REVIEW QUESTIONS 1. What is a descriptor?

2. What are the advantages and disadvantages of decimal data types?

3. What are the design issues for character string types?

4. Describe the three string length options.

5. Define ordinal, enumeration, and subrange types.

6. What are the advantages of user-defined enumeration types?

7. In what ways are the user-defined enumeration types of C# more reliable than those of C++?

8. What are the design issues for arrays?

9. Define static, fixed stack-dynamic, fixed heap-dynamic, and heap- dynamic arrays. What are the advantages of each?

10. What happens when a nonexistent element of an array is referenced in Perl?

11. How does JavaScript support sparse arrays?

12. What languages support negative subscripts?

13. What languages support array slices with stepsizes?

14. What is an aggregate constant?

15. Define row major order and column major order.

16. What is an access function for an array?

17. What are the required entries in a Java array descriptor, and when must they be stored (at compile time or run time)?

18. What is the structure of an associative array?

19. What is the purpose of level numbers in COBOL records?

20. Define fully qualified and elliptical references to fields in records.

21. What is the primary difference between a record and a tuple?

22. Are the tuples of Python mutable?

23. What is the purpose of an F# tuple pattern?

24. In what primarily imperative language do lists serve as arrays?

25. What is the action of the Scheme function CAR?

26. What is the action of the F# function tl?

27. In what way does Scheme’s CDR function modify its parameter?

28. On what are Python’s list comprehensions based?

29. Define union, free union, and discriminated union.

30. Are the unions of F# discriminated?

31. What are the design issues for pointer types?

32. What are the two common problems with pointers?

33. Why are the pointers of most languages restricted to pointing at a single type variable?

34. What is a C++ reference type, and what is its common use?

35. Why are reference variables in C++ better than pointers for formal

parameters?

36. What advantages do Java and C# reference type variables have over the pointers in other languages?

37. Describe the lazy and eager approaches to reclaiming garbage.

38. Why wouldn’t arithmetic on Java and C# references make sense?

39. What is a compatible type?

40. Define type error.

41. Define strongly typed.

42. Why is Java not strongly typed?

43. What is a nonconverting cast?

44. What languages have no type coercions?

45. Why are C and C++ not strongly typed?

46. What is name type equivalence?

47. What is structure type equivalence?

48. What is the primary advantage of name type equivalence?

49. What is the primary disadvantage to structure type equivalence?

50. For what types does C use structure type equivalence?

51. What set operation models C’s struct data type?

PROBLEM SET 1. What are the arguments for and against representing Boolean values as

single bits in memory?

2. How does a decimal value waste memory space?

3. VAX minicomputers use a format for floating-point numbers that is not the same as the IEEE standard. What is this format, and why was it chosen by the designers of the VAX computers? A reference for VAX floating-point representations is Sebesta (1991).

4. Compare the tombstone and lock-and-key methods of avoiding dangling pointers, from the points of view of safety and implementation cost.

5. What disadvantages are there in implicit dereferencing of pointers, but only in certain contexts?

6. Explain all of the differences between Ada’s subtypes and derived types.

7. What significant justification is there for the -> operator in C and C++?

8. What are all of the differences between the enumeration types of C++ and those of Java?

9. Multidimensional arrays can be stored in row major order, as in C++, or in column major order, as in Fortran. Develop the access functions for both of these arrangements for three-dimensional arrays.

10. In the Burroughs Extended ALGOL language, matrices are stored as a single-dimensioned array of pointers to the rows of the matrix, which are treated as single-dimensioned arrays of values. What are the advantages and disadvantages of such a scheme?

11. Analyze and write a comparison of C’s malloc and free functions with C++’s new and delete operators. Use safety as the primary

consideration in the comparison.

12. Analyze and write a comparison of using C++ pointers and Java reference variables to refer to fixed heap-dynamic variables. Use safety and convenience as the primary considerations in the comparison.

13. Write a short discussion of what was lost and what was gained in Java’s designers’ decision to not include the pointers of C++.

14. What are the arguments for and against Java’s implicit heap storage recovery, when compared with the explicit heap storage recovery required in C++? Consider real-time systems.

15. What are the arguments for the inclusion of enumeration types in C#, although they were not in the first few versions of Java?

16. What would you expect to be the level of use of pointers in C#? How often will they be used when it is not absolutely necessary?

17. Make two lists of applications of matrices, one for those that require jagged matrices and one for those that require rectangular matrices. Now, argue whether just jagged, just rectangular, or both should be included in a programming language.

18. Compare the string manipulation capabilities of the class libraries of C++, Java, and C#.

19. Look up the definition of strongly typed as given in Gehani (1983) and compare it with the definition given in this chapter. How do they differ?

20. In what way is static type checking better than dynamic type checking?

21. Explain how coercion rules can weaken the beneficial effect of strong typing.

PROGRAMMING EXERCISES 1. Design a set of simple test programs to determine the type compatibility

rules of a C compiler to which you have access. Write a report of your findings.

2. Determine whether some C compiler to which you have access implements the free function.

3. Write a program that does matrix multiplication in some language that does subscript range checking and for which you can obtain an assembly language or machine language version from the compiler. Determine the number of instructions required for the subscript range checking and compare it with the total number of instructions for the matrix multiplication process.

4. If you have access to a compiler in which the user can specify whether subscript range checking is desired, write a program that does a large number of matrix accesses and time their execution. Run the program with subscript range checking and without it, and compare the times.

5. Write a simple program in C++ to investigate the safety of its enumeration types. Include at least 10 different operations on enumeration types to determine what incorrect or just silly things are legal. Now, write a C# program that does the same things and run it to determine how many of the incorrect or silly things are legal. Compare your results.

6. Write a program in C++ or C# that includes two different enumeration types and has a significant number of operations using the enumeration types. Also write the same program using only integer variables. Compare the readability and predict the reliability differences between the two programs.

7. Write a C program that does a large number of references to elements of

two-dimensioned arrays, using only subscripting. Write a second program that does the same operations but uses pointers and pointer arithmetic for the storage-mapping function to do the array references. Compare the time efficiency of the two programs. Which of the two programs is likely to be more reliable? Why?

8. Write a Perl program that uses a hash and a large number of operations on the hash. For example, the hash could store people’s names and their ages. A random-number generator could be used to create three- character names and ages, which could be added to the hash. When a duplicate name was generated, it would cause an access to the hash but not add a new element. Rewrite the same program without using hashes. Compare the execution efficiency of the two. Compare the ease of programming and readability of the two.

9. Write a program in the language of your choice that behaves differently if the language used name equivalence than if it used structural equivalence.

10. For what types of A and B is the simple assignment statement A=B legal in C++ but not Java?

7 Expressions and ​Assignment Statements

1. 7.1 Introduction

2. 7.2 Arithmetic Expressions

3. 7.3 Overloaded Operators

4. 7.4 Type Conversions

5. 7.5 Relational and Boolean Expressions

6. 7.6 Short-Circuit Evaluation

7. 7.7 Assignment Statements

8. 7.8 Mixed-Mode Assignment

As the title indicates, the topic of this chapter is expressions and assignment statements. The semantics rules that determine the order of evaluation of operators in expressions are discussed first. This is followed by an explanation of a potential problem with operand evaluation order when functions can have side effects. Overloaded operators, both predefined and user defined, are then discussed, along with their effects on the expressions in programs. Next, mixed-mode expressions are described and evaluated. This leads to the definition and evaluation of widening and narrowing type conversions, both implicit and explicit. Relational and Boolean expressions are then discussed, including the process of short-circuit evaluation. Finally, the assignment statement, from its simplest form to all of its variations, is ​- covered, including assignments as expressions and mixed-mode assignments.

Character string pattern-matching expressions were covered as a part of the ​- material on character strings in Chapter 6, so they are not mentioned in this chapter.

7.1 Introduction Expressions are the fundamental means of specifying computations in a programming language. It is crucial for a programmer to understand both the syntax and semantics of expressions of the language he or she uses. A formal mechanism (BNF) for describing the syntax of expressions was introduced in Chapter 3. In this chapter, the semantics of expressions are discussed.

To understand expression evaluation, it is necessary to be familiar with the orders of operator and operand evaluation. The operator evaluation order of expressions is dictated by the associativity and precedence rules of the language. Although the value of an expression sometimes depends on it, the order of operand evaluation in expressions is often unstated by language designers. This allows implementors to choose the order, which leads to the possibility of programs producing different results in different implementations. Other issues in expression semantics are type mismatches, coercions, and short-circuit evaluation.

The essence of the imperative programming languages is the dominant role of assignment statements. The purpose of these statements is to cause the side effect of changing the values of variables, or the state, of the program. So an integral part of all imperative languages is the concept of variables whose values change during program execution.

Functional languages use variables of a different sort, such as the parameters of functions. These languages also have declaration statements that bind values to names. These declarations are similar to assignment statements, but do not have side effects.

7.2 Arithmetic Expressions Automatic evaluation of arithmetic expressions similar to those found in mathematics, science, and engineering was one of the primary goals of the first high-level programming languages. Most of the characteristics of arithmetic expressions in programming languages were inherited from conventions that had evolved in mathematics. In programming languages, arithmetic expressions consist of operators, operands, parentheses, and function calls. An operator can be unary, meaning it has a single operand, binary, meaning it has two operands, or ternary, meaning it has three operands.

In most programming languages, binary operators are infix, which means they appear between their operands. One exception is Perl, which has some operators that are prefix, which means they precede their operands. In Scheme and Lisp, all operators are prefix. Most unary operators are prefix, but the ++ and -- operators of C-based languages can be either prefix or postfix.

The purpose of an arithmetic expression is to specify an arithmetic computation. An implementation of such a computation must cause two actions: fetching the operands, usually from memory, and executing arithmetic operations on those operands. In the following sections, we investigate the common design details of arithmetic expressions.

Following are the primary design issues for arithmetic expressions, all of which are discussed in this section:

What are the operator precedence rules?

What are the operator associativity rules?

What is the order of operand evaluation?

Are there restrictions on operand evaluation side effects?

Does the language allow user-defined operator overloading?

What type mixing is allowed in expressions?

7.2.1 Operator Evaluation Order The operator precedence and associativity rules of a language dictate the order of evaluation of its operators.

7.2.1.1 Precedence The value of an expression depends at least in part on the order of evaluation of the operators in the expression. Consider the following expression:

a + b * c

Suppose the variables a, b, and c have the values 3, 4, and 5, respectively. If evaluated left to right (the addition first and then the multiplication), the result is 35. If evaluated right to left, the result is 23.

Instead of simply evaluating the operators in an expression from left to right or right to left, mathematicians long ago developed the concept of placing operators in a hierarchy of evaluation priorities and basing the evaluation order of expressions partly on this hierarchy. For example, in mathematics, multiplication is considered to be of higher priority than addition, perhaps due to its higher level of complexity. If that convention were applied in the previous example expression, as would be the case in most programming languages, the multiplication would be done first.

The operator precedence rules for expression evaluation partially define the order in which the operators of different precedence levels are evaluated. The operator precedence rules for expressions are based on the hierarchy of operator priorities, as seen by the language designer. The operator precedence rules of the common imperative languages are nearly all the same, because they are based on those of mathematics. In these languages, exponentiation

has the highest precedence (when it is provided by the language), followed by multiplication and division on the same level, followed by binary addition and subtraction on the same level.

Many languages also include unary versions of addition and subtraction. Unary addition is called the identity operator because it usually has no associated operation and thus has no effect on its operand. Ellis and Stroustrup (1990, p. 56), speaking about C++, call it a historical accident and correctly label it useless. Unary minus, of course, changes the sign of its operand. In Java and C#, unary minus also causes the implicit conversion of short and byte operands to int type.

In all of the common imperative languages, the unary minus operator can appear in an expression either at the beginning or anywhere inside the expression, as long as it is parenthesized to prevent it from being next to another operator. For example,

a + (- b) * c

is legal, but

a + - b * c

usually is not.

Next, consider the following expressions:

- a / b

- a * b

- a ** b

In the first two cases, the relative precedence of the unary minus operator and the binary operator is irrelevant—the order of evaluation of the two operators has no effect on the value of the expression. In the last case, however, it does matter.

Of the common programming languages, only Fortran, Ruby, Visual Basic, and Ada have the exponentiation operator. In all four, exponentiation has higher precedence than unary minus, so

- A ** B

is equivalent to

-(A ** B)

The precedences of the arithmetic operators of Ruby and the C-based languages are as follows:

The ** operator is exponentiation. The % operator takes two integer operands and yields the remainder of the first after division by the second.1 The ++ and -- operators of the C-based languages are described in Section 7.7.4.

1. In versions of C before C99, the % operator was implementation dependent in some ​situations, because division was also implementation dependent.

APL is odd among languages because it has a single level of precedence, as illustrated in the next section.

Precedence accounts for only some of the rules for the order of operator evaluation; associativity rules also affect it.

7.2.1.2 Associativity Consider the following expression:

a - b + c - d

If the addition and subtraction operators have the same level of precedence,

as they do in programming languages, the precedence rules say nothing about the order of evaluation of the operators in this expression.

When an expression contains two adjacent2 occurrences of operators with the same level of precedence, the question of which operator is evaluated first is answered by the associativity rules of the language. An operator can have either left or right associativity, meaning that when there are two adjacent operators with the same precedence, the left operator is evaluated first or the right operator is evaluated first, respectively.

2. We call operators “adjacent” if they are separated by a single operand.

Associativity in common languages is left to right, except that the exponentiation operator (when provided) sometimes associates right to left. In the Java expression

a - b + c

the left operator is evaluated first.

Exponentiation in Fortran and Ruby is right associative, so in the expression

A ** B ** C

the right operator is evaluated first.

In Visual Basic, the exponentiation operator, ^, is left associative.

The associativity rules for a few common languages are given here:

As stated in Section 7.2.1.1, in APL, all operators have the same level of precedence. Thus, the order of evaluation of operators in APL expressions is determined entirely by the associativity rule, which is right to left for all operators. For example, in the expression

A × B + C

the addition operator is evaluated first, followed by the multiplication operator (× is the APL multiplication operator). If A were 3, B were 4, and C were 5, then the value of this APL expression would be 27.

Many compilers for the common languages make use of the fact that some arithmetic operators are mathematically associative, meaning that the associativity rules have no impact on the value of an expression containing only those operators. For example, addition is mathematically associative, so in mathematics the value of the expression

A + B + C

does not depend on the order of operator evaluation. If floating-point operations for mathematically associative operations were also associative, the compiler could use this fact to perform some simple optimizations. Specifically, if the compiler is allowed to reorder the evaluation of operators, it may be able to produce slightly faster code for expression evaluation. Compilers commonly do these kinds of optimizations.

Unfortunately, in a computer, both floating-point representations and floating-point arithmetic operations are only approximations of their mathematical counterparts (because of size limitations). The fact that a mathematical operator is associative does not necessarily imply that the corresponding computer floating-point operation is associative. In fact, only if all the operands and intermediate results can be exactly represented in floating-point notation will the process be precisely associative. For example, there are pathological situations in which integer addition on a computer is not associative. For example, suppose that a program must evaluate the expression

A + B + C + D

and that A and C are very large positive numbers, and B and D are negative numbers with very large absolute values. In this situation, adding B to A does not cause an overflow exception, but adding C to A does. Likewise, adding C to B does not cause overflow, but adding D to B does. Because of the limitations of computer arithmetic, addition is catastrophically nonassociative in this case. Therefore, if the compiler reorders these addition operations, it affects the value of the expression. This problem, of course, can be avoided by the programmer, assuming the approximate values of the variables are known. The programmer can specify the expression in two parts (in two assignment statements), ensuring that overflow is avoided. However, this situation can arise in far more subtle ways, in which the programmer is less likely to notice the order dependence.

7.2.1.3 Parentheses Programmers can alter the precedence and associativity rules by placing parentheses in expressions. A parenthesized part of an expression has precedence over its adjacent unparenthesized parts. For example, although multiplication has precedence over addition, in the expression

(A + B) * C

the addition will be evaluated first. Mathematically, this is perfectly natural. In this expression, the first operand of the multiplication operator is not available until the addition in the parenthesized subexpression is evaluated. Also, the expression from Section 7.2.1.2 could be specified as

(A + B) + (C + D)

to avoid overflow.

Languages that allow parentheses in arithmetic expressions could dispense with all precedence rules and simply associate all operators left to right or right to left. The programmer would specify the desired order of evaluation with parentheses. This approach would be simple because neither the author nor the readers of programs would need to remember any precedence or associativity rules. The disadvantage of this scheme is that it makes writing

expressions more tedious, and it also seriously compromises the readability of the code. Yet this was the choice made by Ken Iverson, the designer of APL.

7.2.1.4 Ruby Expressions Recall that Ruby is a pure object-oriented language, which means, among other things, that every data value, including literals, is an object. Ruby supports the collection of arithmetic and logic operations that are included in the C-based languages. What sets Ruby apart from the C-based languages in the area of expressions is that all of the arithmetic, relational, and assignment operators, as well as array indexing, shifts, and bitwise logic operators, are implemented as methods. For example, the expression a + b is a call to the + method of the object referenced by a, passing the object referenced by b as a parameter.

One interesting result of the implementation of operators as methods is that they can be overridden by application programs. Therefore, these operators can be redefined. While it is often not useful to redefine operators for predefined types, it is useful, as we will see in Section 7.3, to define predefined operators for user-defined types, which can be done with operator overloading in some languages.

In C++ and Ada, operators are actually implemented as function calls.

7.2.1.5 Expressions in Lisp As is the case with Ruby, all arithmetic and logic operations in Lisp are performed by subprograms. But in Lisp, the subprograms must be explicitly called. For example, to specify the C expression a + b * c in Lisp, one must write the following expression:3

3. When a list is interpreted as code in Lisp, the first element is the function name and others are parameters to the function.

(+ a (* b c))

In this expression, + and * are the names of functions.

7.2.1.6 Conditional Expressions if-then-else statements can be used to perform a conditional expression assignment. For example, consider

if (count == 0)

average = 0;

else

average = sum / count;

In the C-based languages, this code can be specified more conveniently in an assignment statement using a conditional expression, which has the following form:

expression_1 ? expression_2 : expression_3

where expression_1 is interpreted as a Boolean expression. If expression_1 evaluates to true, the value of the whole expression is the value of expression_2; otherwise, it is the value of expression_3. For example, the effect of the example if-then-else can be achieved with the following assignment statement, using a conditional expression:

average = (count == 0) ? 0 : sum / count;

In effect, the question mark denotes the beginning of the then clause, and the colon marks the beginning of the else clause. Both clauses are mandatory. Note that ? is used in conditional expressions as a ternary operator.

Conditional expressions can be used anywhere in a program (in a C-based language) where any other expression can be used. In addition to the C-based languages, conditional expressions are provided in Perl, JavaScript, and Ruby.

7.2.2 Operand Evaluation Order A less commonly discussed design characteristic of expressions is the order of evaluation of operands. Variables in expressions are evaluated by fetching their values from memory. Constants are sometimes evaluated the same way. In other cases, a constant may be part of the machine language instruction and not require a memory fetch. If an operand is a parenthesized expression, all of the operators it contains must be evaluated before its value can be used as an operand.

If neither of the operands of an operator has side effects, then operand evaluation order is irrelevant. Therefore, the only interesting case arises when the evaluation of an operand does have side effects.

7.2.2.1 Side Effects A side effect of a function, naturally called a functional side effect, occurs when the function changes either one of its parameters or a global variable. (A global variable is declared outside the function but is accessible in the function.)

Consider the following expression:

a + fun(a)

If fun does not have the side effect of changing a, then the order of evaluation of the two operands, a and fun(a), has no effect on the value of the expression. However, if fun changes a, there is an effect. Consider the following situation: fun returns 10 and changes the value of its parameter to 20. Suppose we have the following:

a = 10;

b = a + fun(a);

Then, if the value of a is fetched first (in the expression evaluation process), its value is 10 and the value of the expression is 20. But if the second operand

is evaluated first, then the value of the first operand is 20 and the value of the expression is 30.

The following C program illustrates the same problem when a function changes a global variable that appears in an expression:

int a = 5;

int fun1() {

a = 17;

return 3;

} /* end of fun1 */

void main() {

a = a + fun1();

} /* end of main */

The value computed for a in main depends on the order of evaluation of the operands in the expression a + fun1(). The value of a will be either 8 (if a is evaluated first) or 20 (if the function call is evaluated first).

Note that functions in mathematics do not have side effects, because there is no notion of variables in mathematics. The same is true for functional programming languages. In both mathematics and functional programming languages, functions are much easier to reason about and understand than those in imperative languages, because their context is irrelevant to their meaning.

There are two possible solutions to the problem of operand evaluation order and side effects. First, the language designer could disallow function evaluation from affecting the value of expressions by simply disallowing functional side effects. Second, the language definition could state that operands in expressions are to be evaluated in a particular order and demand that implementors guarantee that order.

Disallowing functional side effects in the imperative languages is difficult, and it eliminates some flexibility for the programmer. Consider the case of C and C++, which have only functions, meaning that all subprograms return one value. To eliminate the side effects of two-way parameters and still provide subprograms that return more than one value, the values would need to be placed in a struct and the struct returned. Access to globals in functions would also have to be disallowed. However, when efficiency is important,

using access to global variables to avoid parameter passing is an important method of increasing execution speed. In compilers, for example, global access to data such as the symbol table is commonplace.

The problem with having a strict evaluation order is that some code optimization techniques used by compilers involve reordering operand evaluations. A guaranteed order disallows those optimization methods when function calls are involved. There is, therefore, no perfect solution, as is borne out by actual language designs.

The Java language definition guarantees that operands appear to be evaluated in left-to-right order, eliminating the problem discussed in this section.

7.2.2.2 Referential Transparency and Side Effects The concept of referential transparency is related to and affected by functional side effects. A program has the property of referential transparency if any two expressions in the program that have the same value can be substituted for one another anywhere in the program, without affecting the action of the program. The value of a referentially transparent function depends entirely on its parameters.4 The connection of referential transparency and functional side effects is illustrated by the following example:

4. Furthermore, the value of the function cannot depend on the order in which its parameters are evaluated.

result1 = (fun(a) + b) / (fun(a) - c);

temp = fun(a);

result2 = (temp + b) / (temp - c);

If the function fun has no side effects, result1 and result2 will be equal, because the expressions assigned to them are equivalent. However, suppose fun has the side effect of adding 1 to either b or c. Then result1 would not be equal to result2. So, that side effect violates the referential transparency

of the program in which the code appears.

There are several advantages to referentially transparent programs. The most important of these is that the semantics of such programs is much easier to understand than the semantics of programs that are not referentially transparent. Being referentially transparent makes a function equivalent to a mathematical function, in terms of ease of understanding.

Because they do not have variables, programs written in pure functional languages are referentially transparent. Functions in a pure functional language cannot have state, which would be stored in local variables. If such a function uses a value from outside the function, that value must be a constant, since there are no variables. Therefore, the value of the function depends on the values of its parameters.

Referential transparency will be further discussed in Chapter 15.

7.3 Overloaded Operators Arithmetic operators are often used for more than one purpose. For example, + usually is used to specify integer addition and floating-point addition. Some languages—Java, for example—also use it for string catenation. This multiple use of an operator is called operator overloading and is generally thought to be acceptable, as long as neither readability nor reliability suffers.

As an example of the possible dangers of overloading, consider the use of the ampersand (&) in C++. As a binary operator, it specifies a bitwise logical AND operation. As a unary operator, however, its meaning is totally different. As a unary operator with a variable as its operand, the expression value is the address of that variable. In this case, the ampersand is called the address-of operator. For example, the execution of

x = &y;

causes the address of y to be placed in x. There are two problems with this multiple use of the ampersand. First, using the same symbol for two completely unrelated operations is detrimental to readability. Second, the simple keying error of leaving out the first operand for a bitwise AND operation can go undetected by the compiler, because it is interpreted as an address-of operator. Such an error may be difficult to diagnose.

Virtually all programming languages have a less serious but similar problem, which is often due to the overloading of the minus operator. The problem is only that the compiler cannot tell if the operator is meant to be binary or unary.5 So once again, failure to include the first operand when the operator is meant to be binary cannot be detected as an error by the compiler. However, the meanings of the two operations, unary and binary, are at least closely related, so readability is not adversely affected.

5. ML alleviates this problem by using different symbols for unary and binary minus operators, tilde (') for unary and dash for binary.

Some languages that support abstract data types (see Chapter 11), for

example, C++, C#, and F#, allow the programmer to further overload operator symbols. For instance, suppose a user wants to define the * operator between a scalar integer and an integer array to mean that each element of the array is to be multiplied by the scalar. Such an operator could be defined by writing a function subprogram named * that performs this new operation. The compiler will choose the correct meaning when an overloaded operator is specified, based on the types of the operands, as with language-defined overloaded operators. For example, if this new definition for * is defined in a C# program, a C# compiler will use the new definition for * whenever the * operator appears with a simple integer as the left operand and an integer array as the right operand.

When sensibly used, user-defined operator overloading can aid readability. For example, if + and * are overloaded for a matrix abstract data type and A, B, C, and D are variables of that type, then

A * B + C * D

can be used instead of

MatrixAdd(MatrixMult(A, B), MatrixMult(C, D))

On the other hand, user-defined overloading can be harmful to readability. For one thing, nothing prevents a user from defining + to mean multiplication. Furthermore, seeing an * operator in a program, the reader must find both the types of the operands and the definition of the operator to determine its meaning. Any or all of these definitions could be in other files.

Another consideration is the process of building a software system from modules created by different groups. If the different groups overloaded the same operators in different ways, these differences would obviously need to be eliminated before putting the system together.

C++ has a few operators that cannot be overloaded. Among these are the class or structure member operator (.) and the scope resolution operator (::). Interestingly, operator overloading was one of the C++ features that was not copied into Java. However, it did reappear in C#.

The implementation of user-defined operator overloading is discussed in

Chapter 9.

7.4 Type Conversions Type conversions are either narrowing or widening. A narrowing conversion converts a value to a type that cannot store even approximations of all of the values of the original type. For example, converting a double to a float in Java is a narrowing conversion, because the range of double is much larger than that of float. A widening conversion converts a value to a type that can include at least approximations of all of the values of the original type. For example, converting an int to a float in Java is a widening conversion. Widening conversions are nearly always safe, meaning that the approximate magnitude of the converted value is maintained. Narrowing conversions are not always safe—sometimes the magnitude of the converted value is changed in the process. For example, if the floating-point value 1.3E25 is converted to an integer in a Java program, the result will not be in any way related to the original value.

Although widening conversions are usually safe, they can result in reduced accuracy. In many language implementations, although integer-to-floating- point conversions are widening conversions, some precision may be lost. For example, in many cases, integers are stored in 32 bits, which allows at least 9 decimal digits of precision. But floating-point values are also stored in 32 bits, with only about seven decimal digits of precision (because of the space used for the exponent). So, integer-to-floating-point widening can result in the loss of two digits of precision.

Coercions of nonprimitive types are, of course, more complex. In Chapter 5, the complications of assignment compatibility of array and record types were discussed. There is also the question of what parameter types and return types of a method allow it to override a method in a superclass—only when the types are the same, or also some other situations. That issue, as well as the concept of subclasses as subtypes, are discussed in Chapter 12.

Type conversions can be either explicit or implicit. The following two subsections discuss these kinds of type conversions.

7.4.1 Coercion in Expressions One of the design decisions concerning arithmetic expressions is whether an operator can have operands of different types. Languages that allow such expressions, which are called mixed-mode expressions, must define conventions for implicit operand type conversions because computers do not have binary operations that take operands of different types. Recall that in Chapter 5, coercion was defined as an implicit type conversion that is initiated by the compiler or runtime system. Type conversions explicitly requested by the programmer are referred to as explicit conversions, or casts, not coercions.

Although some operator symbols may be overloaded, we assume that a computer system, either in hardware or in some level of software simulation, has an operation for each operand type and operator defined in the language.6 For overloaded operators in a language that uses static type binding, the compiler chooses the correct type of operation on the basis of the types of the operands. When the two operands of an operator are not of the same type and that is legal in the language, the compiler must choose one of them to be coerced and generate the code for that coercion. In the following discussion, the coercion design choices of several common languages are examined.

6. This assumption is not true for many languages. An example is given later in this section.

Language designers are not in agreement on the issue of coercions in arithmetic expressions. Those against a broad range of coercions are concerned with the reliability problems that can result from such coercions, because they reduce the benefits of type checking. Those who would rather include a wide range of coercions are more concerned with the loss in flexibility that results from restrictions. The issue is whether programmers should need to be concerned with this category of errors or whether the compiler should detect them.

As a simple illustration of the problem, consider the following Java code:

int a;

float b, c, d;

. . .

d = b * a;

Assume that the second operand of the multiplication operator was supposed to be c, but because of a keying error it was typed as a. Because mixed-mode expressions are legal in Java, the compiler would not detect this as an error. It would simply insert code to coerce the value of the int operand, a, to float. If mixed-mode expressions were not legal in Java, this keying error would have been detected by the compiler as a type error.

Because error detection is reduced when mixed-mode expressions are allowed, F#, Ada, and ML do not allow them. For example, they do not allow mixing of integer and floating-point operands in expressions.

In most of the other common languages, there are no restrictions on mixed- mode arithmetic expressions.

The C-based languages have integer types that are smaller than the int type. In Java, these are byte and short. Operands of all of these types are coerced to int whenever virtually any operator is applied to them. So, while data can be stored in variables of these types, it cannot be manipulated before conversion to a larger type. For example, consider the following Java code:

history note As a more extreme example of the dangers and costs of too much coercion, consider PL/I’s efforts to achieve flexibility in expressions. In PL/I, a character string variable can be the operand of an arithmetic operator with an integer as the other operand. At run time, the string is scanned for a numeric value. If the value happens to contain a decimal point, the value is assumed to be of ​floating-point type, the other operand is coerced to floating point, and the resulting operation is ​floating-point. This coercion policy is very expensive, because both the type check and the conversion must be done at run time. It also eliminates the possibility of detecting programmer errors in expressions, because a binary operator can combine an operand of any type with an operand of virtually any other type.

byte a, b, c;

. . .

a = b + c;

The values of b and c are coerced to int and an int addition is performed. Then, the sum is converted to byte and put in a. Given the large size of the memories of contemporary computers, there is little incentive to use byte and short, unless a large number of them must be stored.

7.4.2 Explicit Type Conversion Most languages provide some capability for doing explicit conversions, both widening and narrowing. In some cases, warning messages are produced when an explicit narrowing conversion results in a significant change to the value of the object being converted.

In the C-based languages, explicit type conversions are called casts. To specify a cast, the desired type is placed in parentheses just before the expression to be converted, as in

(int)angle

One of the reasons for the parentheses around the type name in these conversions is that the first of these languages, C, has several two-word type names, such as long int.

In ML and F#, the casts have the syntax of function calls. For example, in F# we could have the following:

float(sum)

7.4.3 Errors in Expressions A number of errors can occur during expression evaluation. If the language requires type checking, either static or dynamic, then operand type errors cannot occur. The errors that can occur because of coercions of operands in

expressions have already been discussed. The other kinds of errors are due to the limitations of computer arithmetic and the inherent limitations of arithmetic. The most common error occurs when the result of an operation cannot be represented in the memory cell where it must be stored. This is called overflow or underflow, depending on whether the result was too large or too small. One limitation of arithmetic is that division by zero is disallowed. Of course, the fact that it is not mathematically allowed does not prevent a program from attempting to do it.

Floating-point overflow, underflow, and division by zero are examples of run-time errors, which are sometimes called exceptions. Language facilities that allow programs to detect and deal with exceptions are discussed in Chapter 14.

7.5 Relational and Boolean Expressions In addition to arithmetic expressions, programming languages support relational and Boolean expressions.

7.5.1 Relational Expressions A relational operator is an operator that compares the values of its two operands. A relational expression has two operands and one relational operator. The value of a relational expression is Boolean, except when Boolean is not a type included in the language. The relational operators are often overloaded for a variety of types. The operation that determines the truth or falsehood of a relational expression depends on the operand types. It can be simple, as for integer operands, or complex, as for character string operands. Typically, the types of the operands that can be used for relational operators are numeric types, strings, and enumeration types.

history note The Fortran I designers used English abbreviations for the relational operators because the symbols > and < were not on the card punches at the time of ​Fortran I’s design (mid-1950s).

The syntax of the relational operators for equality and inequality differs among some programming languages. For example, for inequality, the C- based languages use !=, Fortran 95+ uses .NE. or <>, and ML and F# use <>.

JavaScript and PHP have two additional relational operators, === and !==. These are similar to their relatives, == and !=, but prevent their operands from being coerced. For example, the expression

"7" == 7

is true in JavaScript, because when a string and a number are the operands of a relational operator, the string is coerced to a number. However,

"7" === 7

is false, because no coercion is done on the operands of this operator.

Ruby uses == for the equality relational operator that uses coercions, and eql? for equality with no coercions. Ruby uses === only in the when clause of its case statement, as discussed in Chapter 8.

The relational operators always have lower precedence than the arithmetic operators, so that in expressions such as

a + 1 > 2 * b

the arithmetic expressions are evaluated first.

7.5.2 Boolean Expressions Boolean expressions consist of Boolean variables, Boolean constants, relational expressions, and Boolean operators. The operators usually include those for the AND, OR, and NOT operations, and sometimes for exclusive OR and equivalence. Boolean operators usually take only Boolean operands (Boolean variables, Boolean literals, or relational expressions) and produce Boolean values.

In the mathematics of Boolean algebras, the OR and AND operators must have equal precedence. However, the C-based languages assign a higher precedence to AND than OR. Perhaps this resulted from the baseless correlation of multiplication with AND and of addition with OR, which would naturally assign higher precedence to AND.

Because arithmetic expressions can be the operands of relational expressions, and relational expressions can be the operands of Boolean expressions, the

three categories of operators must be placed in different precedence levels, relative to each other.

The precedence of the arithmetic, relational, and Boolean operators in the C- based languages is as follows:

Versions of C prior to C99 are odd among the popular imperative languages in that they have no Boolean type and thus no Boolean values. Instead, numeric values are used to represent Boolean values. In place of Boolean operands, scalar variables (numeric or character) and constants are used, with zero considered false and all nonzero values considered true. The result of evaluating such an expression is an integer, with the value 0 if false and 1 if true. Arithmetic expressions can also be used for Boolean expressions in C99 and C++.

One odd result of C’s design of relational expressions is that the following expression is legal:

a > b > c

The leftmost relational operator is evaluated first because the relational operators of C are left associative, producing either 0 or 1. Then, this result is

compared with the variable c. There is never a comparison between b and c in this expression.

Some languages, including Perl and Ruby, provide two sets of the binary logic operators, && and and for AND and || and or for OR. One difference between && and and (and || and or) is that the spelled versions have lower precedence. Also, and and or have equal precedence, but && has higher precedence than ||.

When the nonarithmetic operators of the C-based languages are included, there are more than 40 operators and at least 14 different levels of precedence. This is clear evidence of the richness of the collections of operators and the complexity of expressions possible in these languages.

Readability dictates that a language should include a Boolean type, as was stated in Chapter 6, rather than simply using numeric types in Boolean expressions. Some error detection is lost in the use of numeric types for Boolean operands, because any numeric expression, whether intended or not, is a legal operand to a Boolean operator. In the other imperative languages, any non-Boolean expression used as an operand of a Boolean operator is detected as an error.

7.6 Short-Circuit Evaluation A short-circuit evaluation of an expression is one in which the result is determined without evaluating all of the operands and/or operators. For example, the value of the arithmetic expression

(13 * a) * (b / 13 - 1)

is independent of the value of (b / 13 - 1) if a is 0, because 0 * x = 0 for any x. So, when a is 0, there is no need to evaluate (b / 13 - 1) or perform the second multiplication. However, in arithmetic expressions, this shortcut is not easily detected during execution, so it is never taken.

The value of the Boolean expression

(a >= 0) && (b < 10)

is independent of the second relational expression if a < 0, because the expression (FALSE && (b < 10)) is FALSE for all values of b. So, when a is less than zero, there is no need to evaluate b, the constant 10, the second relational expression, or the && operation. Unlike the case of arithmetic expressions, this shortcut easily can be discovered during execution.

To illustrate a potential problem with non-short-circuit evaluation of Boolean expressions, suppose Java did not use short-circuit evaluation. A table lookup loop could be written using the while statement. One simple version of Java code for such a lookup, assuming that list, which has listlen elements, is the array to be searched and key is the searched-for value, is

index = 0;

while ((index < listlen) && (list[index] != key))

index = index + 1;

If evaluation is not short-circuit, both relational expressions in the Boolean expression of the while statement are evaluated, regardless of the value of the first. Thus, if key is not in list, the program will terminate with a subscript out-of-range exception. The same iteration that has index == listlen will

reference list[listlen], which causes the indexing error because list is declared to have listlen-1 as an upper-bound subscript value.

If a language provides short-circuit evaluation of Boolean expressions and it is used, this is not a problem. In the preceding example, a short-circuit evaluation scheme would evaluate the first operand of the AND operator, but it would skip the second operand if the first operand is false.

A language that provides short-circuit evaluations of Boolean expressions and also has side effects in expressions allows subtle errors to occur. Suppose that short-circuit evaluation is used on an expression and part of the expression that contains a side effect is not evaluated; then the side effect will occur only in complete evaluations of the whole expression. If program correctness depends on the side effect, short-circuit evaluation can result in a serious error. For example, consider the Java expression

(a > b) || ((b++) / 3)

In this expression, b is changed (in the second arithmetic expression) only when a <= b. If the programmer assumed b would be changed every time this expression is evaluated during execution (and the program’s correctness depends on it), the program will fail.

In the C-based languages, the usual AND and OR operators, && and ||, respectively, are short-circuit. However, these languages also have bitwise AND and OR operators, & and |, respectively, that can be used on Boolean- valued operands and are not short-circuit. Of course, the bitwise operators are only equivalent to the usual Boolean operators if all operands are restricted to being either 0 (for false) or 1 (for true).

All of the logical operators of Ruby, Perl, ML, F#, and Python are short-​- circuit evaluated.

7.7 Assignment Statements As we have previously stated, the assignment statement is one of the central constructs in imperative languages. It provides the mechanism by which the user can dynamically change the bindings of values to variables. In the following section, the simplest form of assignment is discussed. Subsequent sections describe a variety of alternatives.

7.7.1 Simple Assignments Nearly all programming languages currently being used use the equal sign for the assignment operator. All of these must use something different from an equal sign for the equality relational operator to avoid confusion with their assignment operator.

ALGOL 60 pioneered the use of := as the assignment operator, which avoids the confusion of assignment with equality. Ada also uses this assignment operator.

The design choices of how assignments are used in a language have varied widely. In some languages, such as Fortran and Ada, an assignment can appear only as a stand-alone statement, and the destination is restricted to a single variable. There are, however, many alternatives.

7.7.2 Conditional Targets Perl allows conditional targets on assignment statements. For example, consider

($flag ? $count1 : $count2) = 0;

which is equivalent to

if ($flag) {

$count1 = 0;

} else {

$count2 = 0;

}

7.7.3 Compound Assignment Operators A compound assignment operator is a shorthand method of specifying a commonly needed form of assignment. The form of assignment that can be abbreviated with this technique has the destination variable also appearing as the first operand in the expression on the right side, as in

a = a + b

Compound assignment operators were introduced by ALGOL 68, were later adopted in a slightly different form by C, and are part of the other C-based languages, as well as Perl, JavaScript, Python, and Ruby. The syntax of these assignment operators is the catenation of the desired binary operator to the = operator. For example,

sum += value;

is equivalent to

sum = sum + value;

The languages that support compound assignment operators have versions for most of their binary operators.

7.7.4 Unary Assignment Operators The C-based languages, Perl, and JavaScript include two special unary

arithmetic operators that are actually abbreviated assignments. They combine increment and decrement operations with assignment. The operators ++ for increment and -- for decrement can be used either in expressions or to form stand-alone single-operator assignment statements. They can appear either as prefix operators, meaning that they precede the operands, or as postfix operators, meaning that they follow the operands. In the assignment statement

sum = ++ count;

the value of count is incremented by 1 and then assigned to sum. This operation could also be stated as

count = count + 1;

sum = count;

If the same operator is used as a postfix operator, as in

sum = count ++;

the assignment of the value of count to sum occurs first; then count is incremented. The effect is the same as that of the two statements

sum = count;

count = count + 1;

An example of the use of the unary increment operator to form a complete assignment statement is

count ++;

which simply increments count. It does not look like an assignment, but it certainly is one. It is equivalent to the statement

count = count + 1;

When two unary operators apply to the same operand, the association is right to left. For example, in

- count ++

count is first incremented and then negated. So, it is equivalent to

- (count ++)

rather than

(- count) ++

7.7.5 Assignment as an Expression In the C-based languages, Perl, and JavaScript, the assignment statement produces a result, which is the same as the value assigned to the target. It can therefore be used as an expression and as an operand in other expressions. This design treats the assignment operator much like any other binary operator, except that it has the side effect of changing its left operand. For example, in C, it is common to write statements such as

while ((ch = getchar()) != EOF) { ... }

In this statement, the next character from the standard input file, usually the keyboard, is gotten with getchar and assigned to the variable ch. The result, or value assigned, is then compared with the constant EOF. If ch is not equal to EOF, the compound statement {...} is executed. Note that the assignment must be parenthesized—in the languages that support assignment as an expression, the precedence of the assignment operator is lower than that of the relational operators. Without the parentheses, the new character would be compared with EOF first. Then, the result of that comparison, either 0 or 1, would be assigned to ch.

The disadvantage of allowing assignment statements to be operands in expressions is that it provides yet another kind of expression side effect. This type of side effect can lead to expressions that are difficult to read and understand. An expression with any kind of side effect has this disadvantage. Such an expression cannot be read as an expression, which in mathematics is a denotation of a value, but only as a list of instructions with an odd order of execution. For example, the expression

a = b + (c = d / b) - 1

denotes the instructions

Assign d / b to c

Assign b + c to temp

Assign temp - 1 to a

Note that the treatment of the assignment operator as any other binary operator allows the effect of multiple-target assignments, such as

sum = count = 0;

in which count is first assigned the zero, and then count’s value is assigned to sum. This form of multiple-target assignments is also legal in Python.

There is a loss of error detection in the C design of the assignment operation that frequently leads to program errors. In particular, if we type

if (x = y) ...

instead of

if (x == y) ...

which is an easily made mistake, it is not detectable as an error by the compiler. Rather than testing a relational expression, the value that is assigned to x is tested (in this case, it is the value of y that reaches this statement). This is actually a result of two design decisions: allowing assignment to behave like an ordinary binary operator and using two very similar operators, = and ==, to have completely different meanings. This is another example of the safety deficiencies of C and C++ programs. Note that Java and C# allow only boolean expressions in their if statements, disallowing this problem.

7.7.6 Multiple Assignments Several recent programming languages, including Perl and Ruby provide

multiple-target, multiple-source assignment statements. For example, in Perl one can write

($first, $second, $third) = (20, 40, 60);

The semantics is that 20 is assigned to $first, 40 is assigned to $second, and 60 is assigned to $third. If the values of two variables must be interchanged, this can be done with a single assignment, as with

($first, $second) = ($second, $first);

This correctly interchanges the values of $first and $second, without the use of a temporary variable (at least one created and managed by the programmer).

history note The PDP-11 computer, on which C was first implemented, has autoincrement and autodecrement addressing modes, which are hardware versions of the increment and decrement operators of C when they are used as array indices. One might guess from this that the design of these C operators was based on the design of the PDP-11 architecture. That guess would be wrong, however, because the C operators were inherited from the B language, which was designed before the first PDP-11.

The syntax of the simplest form of Ruby’s multiple assignment is similar to that of Perl, except the left and right sides are not parenthesized. Also, Ruby includes a few more elaborate versions of multiple assignments, which are not discussed here.

7.7.7 Assignment in Functional Programming Languages All of the identifiers used in pure functional languages and some of them

used in other functional languages are just names of values. As such, their values never change. For example, in ML, names are bound to values with the val declaration, whose form is exemplified in the following:

val cost = quantity * price;

If cost appears on the left side of a subsequent val declaration, that declaration creates a new version of the name cost, which has no relationship with the previous version, which is then hidden.

F# has a somewhat similar declaration that uses the let reserved word. The difference between F#’s let and ML’s val is that let creates a new scope, whereas val does not. In fact, val declarations are often nested in let constructs in ML. let and val are further discussed in Chapter 15.

7.8 Mixed-Mode Assignment Mixed-mode expressions were discussed in Section 7.4.1. Frequently, assignment statements also are mixed mode. The design question is: Does the type of the expression have to be the same as the type of the variable being assigned, or can coercion be used in some cases of type mismatch?

C, C++, and Perl use coercion rules for mixed-mode assignment that are similar to those they use for mixed-mode expressions; that is, many of the possible type mixes are legal, with coercion freely applied.7

7. Note that in Python and Ruby, types are associated with objects, not variables, so there is no such thing as mixed-mode assignment in those languages.

In a clear departure from C++, Java and C# allow mixed-mode assignment only if the required coercion is widening.8 So, an int value can be assigned to a float variable, but not vice versa. Disallowing half of the possible mixed-mode assignments is a simple but effective way to increase the reliability of Java and C#, relative to C and C++.

8. Not quite true: If an integer literal, which the compiler by default assigns the type int, is assigned to a char, byte, or short variable and the literal is in the range of the type of the variable, the int value is coerced to the type of the variable in a narrowing conversion. This narrowing conversion cannot result in an error.

Of course, in functional languages, where assignments are just used to name values, there is no such thing as a mixed-mode assignment.

SUMMARY Expressions consist of constants, variables, parentheses, function calls, and operators. Assignment statements include target variables, assignment operators, and expressions.

The semantics of an expression is determined in large part by the order of evaluation of operators. The associativity and precedence rules for operators in the expressions of a language determine the order of operator evaluation in those expressions. Operand evaluation order is important if functional side effects are possible. Type conversions can be widening or narrowing. Some narrowing conversions produce erroneous values. Implicit type conversions, or coercions, in expressions are common, although they eliminate the error-​- detection benefit of type checking, thus lowering reliability.

Assignment statements have appeared in a wide variety of forms, including conditional targets, assigning operators, and list assignments.

REVIEW QUESTIONS 1. Define operator precedence and operator associativity.

2. What is a ternary operator?

3. What is a prefix operator?

4. What operator usually has right associativity?

5. What is a nonassociative operator?

6. What associativity rules are used by APL?

7. What is the difference between the way operators are implemented in C++ and Ruby?

8. Define functional side effect.

9. What is a coercion?

10. What is a conditional expression?

11. What is an overloaded operator?

12. Define narrowing and widening conversions.

13. In JavaScript, what is the difference between == and ===?

14. What is a mixed-mode expression?

15. What is referential transparency?

16. What are the advantages of referential transparency?

17. How does operand evaluation order interact with functional side effects?

18. What is short-circuit evaluation?

19. Name a language that always does short-circuit evaluation of Boolean expressions. Name one that never does it.

20. How does C support relational and Boolean expressions?

21. What is the purpose of a compound assignment operator?

22. What is the associativity of C’s unary arithmetic operators?

23. What is one possible disadvantage of treating the assignment operator as if it were an arithmetic operator?

24. What two languages include multiple assignments?

25. What mixed-mode assignments are allowed in Java?

26. What mixed-mode assignments are allowed in ML?

27. What is a cast?

PROBLEM SET 1. When might you want the compiler to ignore type differences in an

expression?

2. State your own arguments for and against allowing mixed-mode arithmetic expressions.

3. Do you think the elimination of overloaded operators in your favorite language would be beneficial? Why or why not?

4. Would it be a good idea to eliminate all operator precedence rules and require parentheses to show the desired precedence in expressions? Why or why not?

5. Should C’s assigning operations (for example, +=) be included in other languages (that do not already have them)? Why or why not?

6. Should C’s single-operand assignment forms (for example, ++count) be included in other languages (that do not already have them)? Why or why not?

7. Describe a situation in which the add operator in a programming language would not be commutative.

8. Describe a situation in which the add operator in a programming language would not be associative.

9. Assume the following rules of associativity and precedence for expressions:

Show the order of evaluation of the following expressions by parenthesizing all subexpressions and placing a superscript on the right parenthesis to indicate order. For example, for the expression

a + b * c + d

the order of evaluation would be represented as

((a + (b * c) 1 ) 2 + d)

3

1. a * b - 1 + c

2. a * (b - 1) / c mod d

3. (a - b) / c & (d * e / a - 3)

4. -a or c = d and e

5. a > b xor c or d <= 17

6. -a + b

10. Show the order of evaluation of the expressions of Problem 9, assuming that there are no precedence rules and all operators associate right to left.

11. Write a BNF description of the precedence and associativity rules defined for the expressions in Problem 9. Assume the only operands are the names a,b,c,d, and e.

12. Using the grammar of Problem 11, draw parse trees for the expressions of Problem 9.

13. Let the function fun be defined as

int fun(int* k) {

*k += 4;

return 3 * (*k) - 1;

}

Suppose fun is used in a program as follows:

void main() {

int i = 10, j = 10, sum1, sum2;

sum1 = (i / 2) + fun(&i);

sum2 = fun(&j) + (j / 2);

}

What are the values of sum1 and sum2

1. operands in the expressions are evaluated left to right?

2. operands in the expressions are evaluated right to left?

14. What is your primary argument against (or for) the operator precedence rules of APL?

15. Explain why it is difficult to eliminate functional side effects in C.

16. For some language of your choice, make up a list of operator symbols that could be used to eliminate all operator overloading.

17. Determine whether the narrowing explicit type conversions in two languages you know provide error messages when a converted value loses its usefulness.

18. Should an optimizing compiler for C or C++ be allowed to change the order of subexpressions in a Boolean expression? Why or why not?

19. Consider the following C program:

int fun(int *i) {

*i += 5;

return 4;

}

void main() {

int x = 3;

x = x + fun(&x);

}

What is the value of x after the assignment statement in main, assuming

1. operands are evaluated left to right.

2. operands are evaluated right to left.

20. Why does Java specify that operands in expressions are all evaluated in left-to-right order?

21. Explain how the coercion rules of a language affect its error detection.

PROGRAMMING EXERCISES 1. Run the code given in Problem 13 (in the Problem Set) on some system

that supports C to determine the values of sum1 and sum2. Explain the results.

2. Rewrite the program of Programming Exercise 1 in C++, Java, and C#, run them, and compare the results.

3. Write a test program in your favorite language that determines and outputs the precedence and associativity of its arithmetic and Boolean operators.

4. Write a Java program that exposes Java’s rule for operand evaluation order when one of the operands is a method call.

5. Repeat Programming Exercise 4 with C++.

6. Repeat Programming Exercise 4 with C#.

7. Write a program in either C++, Java, or C# that illustrates the order of evaluation of expressions used as actual parameters to a method.

8. Write a C program that has the following statements:

int a, b;

a = 10;

b = a + fun();

printf("With the function call on the right, ");

printf(" b is: %d\n", b);

a = 10;

b = fun() + a;

printf("With the function call on the left, ");

printf(" b is: %d\n", b);

and define fun to add 10 to a. Explain the results.

9. Write a program in either Java, C++, or C# that performs a large number

of floating-point operations and an equal number of integer operations and compare the time required.

8 Statement-Level Control Structures

1. 8.1 Introduction

2. 8.2 Selection Statements

3. 8.3 Iterative Statements

4. 8.4 Unconditional Branching

5. 8.5 Guarded Commands

6. 8.6 Conclusions

The flow of control, or execution sequence, in a program can be examined at several levels. In Chapter 7, the flow of control within expressions, which is ​- governed by operator associativity and precedence rules, was discussed. At the highest level is the flow of control among program units, which is discussed in ​Chapters 9 and 13. Between these two extremes is the important issue of the flow of control among statements, which is the subject of this chapter.

We begin by giving an overview of the evolution of control statements. This topic is followed by a thorough examination of selection statements, both those for two-way and those for multiple selection. Next, we discuss the variety of looping statements that have been developed and used in programming languages. Next, we take a brief look at the problems associated with unconditional branch statements. Finally, we describe the guarded command control statements.

8.1 Introduction Computations in imperative-language programs are accomplished by evaluating expressions and assigning the resulting values to variables. However, there are few useful programs that consist entirely of assignment statements. At least two additional linguistic mechanisms are necessary to make the computations in programs flexible and powerful: some means of selecting among alternative control flow paths (of statement execution) and some means of causing the repeated execution of statements or sequences of statements. Statements that provide these kinds of capabilities are called control statements.

Computations in functional programming languages are accomplished by evaluating expressions and applying functions to given parameters. Furthermore, the flow of execution among the expressions and functions is controlled by other expressions and functions, although some of them are similar to the control statements in the imperative languages.

The control statements of the first successful programming language, Fortran, were, in effect, designed by the architects of the IBM 704. All were directly related to machine language instructions, so their capabilities were more the result of instruction design rather than language design. At the time, little was known about the difficulty of programming, and, as a result, the control statements of Fortran in the mid-1950s were thought to be entirely adequate. By today’s standards, however, they are considered seriously lacking.

A great deal of research and discussion was devoted to control statements in the 10 years between the mid-1960s and the mid-1970s. One of the primary conclusions of these efforts was that, although a single control statement (a selectable goto) is minimally sufficient, a language that is designed not to include a goto needs only a small number of different control statements. In fact, it was proven that all algorithms that can be expressed by flowcharts can be coded in a programming language with only two control statements: one for choosing between two control flow paths and one for logically controlled iterations (Böhm and Jacopini, 1966). An important result of this is that the

unconditional branch statement is superfluous—potentially useful but nonessential. This fact, combined with the practical problems of using unconditional branches, or gotos, led to a great deal of debate about the goto, as will be discussed in Section 8.4.

Programmers care less about the results of theoretical research on control statements than they do about writability and readability. All languages that have been widely used include more control statements than the two that are minimally required, because writability is enhanced by a larger number and wider variety of control statements. For example, rather than requiring the use of a logically controlled loop statement for all loops, it is easier to write programs when a counter-controlled loop statement can be used to build loops that are naturally controlled by a counter. The primary factor that restricts the number of control statements in a language is readability, because the presence of a large number of statement forms demands that program readers learn a larger language. Recall that few people learn all of the statements of a relatively large language; instead, they learn the subset they choose to use, which is often a different subset from that used by the programmer who wrote the program they are trying to read. On the other hand, too few control statements can require the use of lower-level statements, such as the goto, which also makes programs less readable.

The question as to the best collection of control statements to provide the required capabilities and the desired writability has been widely debated. It is essentially a question of how much a language should be expanded to increase its writability at the expense of its simplicity, size, and readability.

A control structure is a control statement and the collection of statements whose execution it controls.

There is only one design issue that is relevant to all of the selection and iteration control statements: Should the control structure have multiple entries? All selection and iteration constructs control the execution of code segments, and the question is whether the execution of those code segments always begins with the first statement in the segment. It is now generally believed that multiple entries add little to the flexibility of a control statement, relative to the decrease in readability caused by the increased complexity. Note that multiple entries are possible only in languages that

include gotos and statement labels.

At this point, the reader might wonder why multiple exits from control structures are not considered a design issue. The reason is that all programming languages allow some form of multiple exits from control structures, the rationale being as follows: If all exits from a control structure are restricted to transferring control to the first statement following the structure, where control would flow if the control structure had no explicit exit, there is no harm to readability and also no danger. However, if an exit can have an unrestricted target and therefore can result in a transfer of control to anywhere in the program unit that contains the control structure, the harm to readability is the same as for a goto statement anywhere else in a program. Languages that have a goto statement allow it to appear anywhere, including in a control structure. Therefore, the issue is the inclusion of a goto, not whether multiple exits from control expressions are allowed.

8.2 Selection Statements A selection statement provides the means of choosing between two or more execution paths in a program. Such statements are fundamental and essential parts of all programming languages, as was proven by Böhm and Jacopini.

Selection statements fall into two general categories: two-way and n-way, or multiple selection. Two-way selection statements are discussed in ​ Section 8.2.1; multiple-selection statements are covered in Section 8.2.2.

8.2.1 Two-Way Selection Statements Although the two-way selection statements of contemporary imperative languages are quite similar, there are some variations in their designs. The general form of a two-way selector is as follows:

if control_expression

   then clause

   else clause

8.2.1.1 Design Issues The design issues for two-way selectors can be summarized as follows:

What is the form and type of the expression that controls the selection?

How are the then and else clauses specified?

How should the meaning of nested selectors be specified?

8.2.1.2 The Control Expression Control expressions are specified in parentheses if the then reserved word (or some other syntactic marker) is not used to introduce the then clause. In those cases where the then reserved word (or alternative marker) is used, there is less need for the parentheses, so they are often omitted, as in Ruby.

In C89, which did not have a Boolean data type, arithmetic expressions were used as control expressions. This can also be done in Python, C99, and C++. However, in those languages either arithmetic or Boolean expressions can be used. In other contemporary languages, only Boolean expressions can be used for control expressions.

8.2.1.3 Clause Form In many languages, the then and else clauses appear as either single statements or compound statements. One variation of this is Perl, in which all then and else clauses must be compound statements, even if they have only one statement. Many languages use braces to form compound statements, which serve as the bodies of then and else clauses. In Python and Ruby, the then and else clauses are statement sequences, rather than compound statements. The complete selection statement is terminated in these languages with a reserved word.

Python uses indentation to specify compound statements. For example,

if x > y :

x = y

print "case 1"

All statements equally indented are included in the compound statement.1 Notice that rather than then, a colon is used to introduce the then clause in Python.

1. The statement following the compound statement must have the same indentation as the if.

The variations in clause form have implications for the specification of the meaning of nested selectors, as discussed in the next subsection.

8.2.1.4 Nesting Selectors Recall that in Chapter 3, we discussed the problem of syntactic ambiguity of a straightforward grammar for a two-way selector statement. That ambiguous grammar was as follows:

<if_stmt> → if <logic_expr> then <stmt>

           | if <logic_expr> then <stmt> else <stmt>

The issue is that when a selection statement is nested in the then clause of a selection statement, it is not clear with which if an else clause should be associated. This problem is reflected in the semantics of selection statements. Consider the following Java-like code:

if (sum == 0)

if (count == 0)

result = 0;

else

result = 1;

This statement can be interpreted in two different ways, depending on whether the else clause is matched with the first then clause or the second. Notice that the indentation seems to indicate that the else clause belongs with the first then clause. However, with the exceptions of Python and F#, indentation has no effect on semantics in contemporary languages and is therefore ignored by their compilers.

The crux of the problem in this example is that the else clause follows two then clauses with no intervening else clause, and there is no syntactic indicator to specify a matching of the else clause to one of the then clauses. In Java, as in many other imperative languages, the static semantics of the language specify that the else clause is always paired with the nearest previous unpaired then clause. A static semantics rule, rather than a syntactic entity, is used to provide the disambiguation. So, in the example, the else

clause would be paired with the second then clause. The disadvantage of using a rule rather than some syntactic entity is that although the programmer may have meant the else clause to be the alternative to the first then clause and the compiler found the structure syntactically correct, its semantics is the opposite. To force the alternative semantics in Java, the inner if is put in a compound, as in

if (sum == 0) {

if (count == 0)

result = 0;

}

else

result = 1;

C, C++, and C# have the same problem as Java with selection statement nesting. Because Perl requires that all then and else clauses be compound, it does not. In Perl, the previous code would be written as follows:

if (sum == 0) {

if (count == 0) {

result = 0;

}

} else {

result = 1;

}

If the alternative semantics were needed, it would be

if (sum == 0) {

if (count == 0) {

result = 0;

}

else {

result = 1;

}

}

Another way to avoid the issue of nested selection statements is to use an alternative means of forming compound statements. Consider the syntactic structure of the Java if statement. The then clause follows the control expression and the else clause is introduced by the reserved word else. When the then clause is a single statement and the else clause is present, although

there is no need to mark the end, the else reserved word in fact marks the end of the then clause. When the then clause is a compound, it is terminated by a right brace. However, if the last clause in an if, whether then or else, is not a compound, there is no syntactic entity to mark the end of the whole selection statement. The use of a special word for this purpose resolves the question of the semantics of nested selectors and also adds to the readability of the statement. This is the design of the selection statement in Ruby. For example, consider the following Ruby statement:

if a > b then sum = sum + a

acount = acount + 1

else sum = sum + b

bcount = bcount + 1

end

The design of this statement is more regular than that of the selection statements of the C-based languages, because the form is the same regardless of the number of statements in the then and else clauses. (This is also true for Perl.) Recall that in Ruby, the then and else clauses consist of statement sequences rather than compound statements. The first interpretation of the selector example at the beginning of this section, in which the else clause is matched to the nested if, can be written in Ruby as follows:

if sum == 0 then

if count == 0 then

result = 0

else

result = 1

end

end

Because the end reserved word closes the nested if, it is clear that the else clause is matched to the inner then clause.

The second interpretation of the selection statement at the beginning of this section, in which the else clause is matched to the outer if, can be written in Ruby as follows:

if sum == 0 then

if count == 0 then

result = 0

end

else

result = 1

end

The following statement, written in Python, is semantically equivalent to the last Ruby statement above:

if sum == 0 :

if count == 0 :

result = 0

else:

result = 1

If the line else: were indented to begin in the same column as the nested if, the else clause would be matched with the inner if.

ML does not have a problem with nested selectors because it does not allow else-less if statements.

8.2.1.5 Selector Expressions In the functional languages ML, F#, and LISP, the selector is not a statement; it is an expression that results in a value. Therefore, it can appear anywhere any other expression can appear. Consider the following example selector written in F#:

let y =

if x > 0 then x

else 2 * x;;

This creates the name y and sets it to either x or 2 * x, depending on whether x is greater than zero.

In F#, the type of the value returned by the then clause of an if construct must be the same as that of the value returned by its else clause. If there is no else clause, the then clause cannot return a value of a normal type. In this case, it can only return a unit type, which is a special type that means no value. A unit type is represented in code as ().

8.2.2 Multiple-Selection Statements The multiple-selection statement allows the selection of one of any number of statements or statement groups. It is, therefore, a generalization of a selector. In fact, two-way selectors can be built with a multiple selector.

The need to choose from among more than two control paths in programs is common. Although a multiple selector can be built from two-way selectors and gotos, the resulting structures are cumbersome, unreliable, and difficult to write and read. Therefore, the need for a special structure is clear.

8.2.2.1 Design Issues Some of the design issues for multiple selectors are similar to some of those for two-way selectors. For example, one issue is the question of the type of expression on which the selector is based. In this case, the range of possibilities is larger, in part because the number of possible selections is larger. A two-way selector needs an expression with only two possible values. Another issue is whether single statements, compound statements, or statement sequences may be selected. Next, there is the question of whether only a single selectable segment can be executed when the statement is executed. This is not an issue for two-way selectors, because they always allow only one of the clauses to be on a control path during one execution. As we shall see, the resolution of this issue for multiple selectors is a trade-off between reliability and flexibility. Another issue is the form of the case value specifications. Finally, there is the issue of what should result from the selector expression evaluating to a value that does not select one of the segments. (Such a value would be unrepresented among the selectable segments.) The choice here is between simply disallowing the situation from arising and having the statement do nothing at all when it does arise.

The following is a summary of these design issues:

What is the form and type of the expression that controls the selection?

How are the selectable segments specified?

Is execution flow through the structure restricted to include just a single selectable segment?

How are the case values specified?

How should unrepresented selector expression values be handled, if at all?

8.2.2.2 Examples of Multiple Selectors The C multiple-selector statement, switch, which is also part of C++, Java, and JavaScript, is a relatively primitive design. Its general form is

switch (expression) {

 case constant_ expression1: statement1;

 . . .

 case constantn: statement_n;

 [default: statementn+1]

}

where the control expression and the constant expressions are some discrete type. This includes integer types, as well as characters and enumeration types. The selectable statements can be statement sequences, compound statements, or blocks. The optional default segment is for unrepresented values of the control expression. If the value of the control expression is not represented and no default segment is present, then the statement does

nothing.

The switch statement does not provide implicit branches at the end of its code segments. This allows control to flow through more than one selectable code segment on a single execution. Consider the following example:

switch (index) {

case 1:

case 3: odd += 1;

sumodd += index;

case 2:

case 4: even += 1;

sumeven += index;

default: printf("Error in switch, index = %d\n", index);

}

This code prints the error message on every execution. Likewise, the code for the 2 and 4 constants is executed every time the code at the 1 or 3 constants is executed. To separate these segments logically, an explicit branch must be included. The break statement, which is actually a restricted goto, is normally used for exiting switch statements. break transfers control to the first statement after the compound statement in which it appears.

The following switch statement uses break to restrict each execution to a single selectable segment:

switch (index) {

case 1:

case 3: odd += 1;

sumodd += index;

break;

case 2:

case 4: even += 1;

sumeven += index;

break;

default: printf("Error in switch, index = %d\n", index);

}

Occasionally, it is convenient to allow control to flow from one selectable code segment to another. For example, in the example above, the segments for the case values 1 and 2 are empty, allowing control to flow to the segments for 3 and 4, respectively. This is the reason why there are no

implicit branches in the switch statement. The reliability problem with this design arises when the mistaken absence of a break statement in a segment allows control to flow to the next segment incorrectly. The designers of C’s switch traded a decrease in reliability for an increase in flexibility. Studies have shown, however, that the ability to have control flow from one selectable segment to another is rarely used. C’s switch is modeled on the multiple-selection statement in ALGOL 68, which also does not have implicit branches from selectable segments.

The C switch statement has virtually no restrictions on the placement of the case expressions, which are treated as if they were normal statement labels. This laxness can result in highly complex structure within the switch body. The following example is taken from Harbison and Steele (2002).

switch (x)

default:

if (prime(x))

case 2: case 3: case 5: case 7:

process_prime(x);

else

case 4: case 6: case 8: case 9: case 10:

process_composite(x);

This code may appear to have diabolically complex form, but it was designed for a real problem and works correctly and efficiently to solve that problem.2

2. The problem is to call process_prime when x is prime and process_composite when x is not prime. The design of the switch body resulted from an attempt to optimize based on the knowledge that x was most often in the range of 1 to 10.

The Java switch prevents this sort of complexity by disallowing case expressions from appearing anywhere except the top level of the body of the switch.

The C# switch statement differs from that of its C-based predecessors in two ways. First, C# has a static semantics rule that disallows the implicit execution of more than one segment. The rule is that every selectable segment must end with an explicit unconditional branch statement: either a break, which transfers control out of the switch statement, or a goto, which

can transfer control to one of the selectable segments (or virtually anywhere else). For example,

switch (value) {

case -1:

Negatives++;

break;

case 0:

Zeros++;

goto case 1;

case 1:

Positives++;

default:

Console.WriteLine("Error in switch \n");

}

Note that Console.WriteLine is the method for displaying strings in C#.

The other way C#’s switch differs from that of its predecessors is that the control expression and the case statements can be strings in C#.

PHP’s switch uses the syntax of C’s switch but allows more type flexibility. The case values can be any of the PHP scalar types—string, integer, or double precision. As with C, if there is no break at the end of the selected segment, execution continues into the next segment.

Ruby has two forms of multiple-selection constructs, both of which are called case expressions and both of which yield the value of the last expression evaluated. The only version of Ruby’s case expressions that is described here is semantically similar to a list of nested if statements:

case

when Boolean_expression then expression

. . .

when Boolean_expression then expression

[else expression]

end

The semantics of this case expression is that the Boolean expressions are evaluated one at a time, top to bottom. The value of the case expression is the value of the first then expression whose Boolean expression is true. The else represents true in this statement, and the else clause is optional. For

example,3

3. This example is from Thomas et al. (2013).

leap = case

when year % 400 == 0 then true

when year % 100 == 0 then false

else year % 4 == 0

end

This case expression evaluates to true if year is a leap year.

The other Ruby case expression form is similar to the switch of Java. Perl and Python do not have multiple-selection statements.

8.2.2.3 Implementing Multiple Selection Structures A multiple selection statement is essentially an n-way branch to segments of code, where n is the number of selectable segments. Implementing such a statement must be done with multiple conditional branch instructions. Consider again the general form of the C switch statement, with breaks:

switch (expression) {

 case constant_expression1: statement1;

  break;

 . . .

 case constantn: statementn;

  break;

 [default: statementn+1]

}

One simple translation of this statement follows:

Code to evaluate expression into t

goto branches

label1 : code for statement1

         goto out

. . .

labeln : code for statementn

         goto out

default: code for statementn+1

         goto out

branches: if t = constant_expression1 goto label1

          . . .

          if t = constant_expressionn goto labeln

          goto default

out:

The code for the selectable segments precedes the branches so that the targets of the branches are all known when the branches are generated. An alternative to these coded conditional branches is to put the case values and labels in a table and use a linear search with a loop to find the correct label. This requires less space than the coded conditionals.

The use of conditional branches or a linear search on a table of cases and

labels is a simple but inefficient approach that is acceptable when the number of cases is small, say less than 10. It takes an average of about half as many tests as there are cases to find the right one. For the default case to be chosen, all other cases must be tested. In statements with 10 or more cases, the low efficiency of this form is not justified by its simplicity.

When the number of cases is 10 or greater, the compiler can build a hash table of the segment labels, which would result in approximately equal (and short) times to choose any of the selectable segments. If the language allows ranges of values for case expressions, as in Ruby, the hash method is not suitable. For these situations, a binary search table of case values and segment addresses is better.

If the range of the case values is relatively small and more than half of the whole range of values is represented, an array whose indices are the case values and whose values are the segment labels can be built. Array elements whose indices are not among the represented case values are filled with the default segment’s label. Then finding the correct segment label is found by array indexing, which is very fast.

Of course, choosing among these approaches is an additional burden on the compiler. In many compilers, only two different methods are used. As in other situations, determining and using the most efficient method costs more compiler time.

8.2.2.4 Multiple Selection Using if In many situations, a switch or case statement is inadequate for multiple selection (Ruby’s case is an exception). For example, when selections must be made on the basis of a Boolean expression rather than some ordinal type, nested two-way selectors can be used to simulate a multiple selector. To alleviate the poor readability of deeply nested two-way selectors, some languages, such as Perl and Python, have been extended specifically for this use. The extension allows some of the special words to be left out. In particular, else-if sequences are replaced with a single special word, and the closing special word on the nested if is dropped. The nested selector is then

called an else-if clause. Consider the following Python selector statement (note that else-if is spelled elif in Python):

if count < 10 :

bag1 = True

elif count < 100 :

bag2 = True

elif count < 1000 :

bag3 = True

which is equivalent to the following:

if count < 10 :

bag1 = True

else :

if count < 100 :

bag2 = True

else :

if count < 1000 :

bag3 = True

else :

bag4 = True

The else-if version (the first) is the more readable of the two. Notice that this example is not easily simulated with a switch statement, because each selectable statement is chosen on the basis of a Boolean expression. Therefore, the else-if statement is not a redundant form of switch. In fact, none of the multiple selectors in contemporary languages are as general as the if-then-else-if ​statement. An operational semantics description of a general selector statement with else-if clauses, in which the E’s are logic expressions and the S’s are statements, is given here:

  if E1 goto 1

  if E2 goto 2

  . . .

1: S1

goto out

2: S2

goto out

. . .

out: . . .

From this description, we can see the difference between multiple selection structures and else-if statements: In a multiple selection statement, all the E’s would be restricted to comparisons between the value of a single expression and some other values.

Languages that do not include the else-if statement can use the same control structure, with only slightly more typing.

The Python example if-then-else-if statement above can be written as the Ruby case statement:

case

when count < 10 then bag1 = true

when count < 100 then bag2 = true

when count < 1000 then bag3 = true

end

Else-if statements are based on the common mathematics statement, the conditional expression.

The Scheme multiple selector, which is based on mathematical conditional expressions, is a special form function named COND. COND is a slightly generalized version of the mathematical conditional expression; it allows more than one predicate to be true at the same time. Because different mathematical conditional expressions have different numbers of parameters, COND does not require a fixed number of actual parameters. Each parameter to COND is a pair of expressions in which the first is a predicate (it evaluates to either #T or #F).

The general form of COND is

 (COND

  ( predicate1 expression1)

  ( predicate2 expression2)

  . . .

  ( predicaten expressionn)

  [(ELSE expressionn+1)]

)

where the ELSE clause is optional.

The semantics of COND is as follows: The predicates of the parameters are evaluated one at a time, in order from the first, until one evaluates to #T. The expression that follows the first predicate that is found to be #T is then evaluated and its value is returned as the value of COND. If none of the predicates is true and there is an ELSE, its expression is evaluated and the value is returned. If none of the predicates is true and there is no ELSE, the value of COND is unspecified. Therefore, all CONDs should include an ELSE.

Consider the following example call to COND:

(COND

((> x y) "x is greater than y")

((< x y) "y is greater than x")

(ELSE "x and y are equal")

)

Note that string literals evaluate to themselves, so that when this call to COND is evaluated, it produces a string result.

8.3 Iterative Statements An iterative statement is one that causes a statement or collection of statements to be executed zero, one, or more times. An iterative statement is often called a loop. Every programming language from Plankalkül on has included some method of repeating the execution of segments of code. Iteration is the very essence of the power of the computer. If some means of repetitive execution of a statement or collection of statements were not possible, programmers would be required to state every action in sequence; useful programs would be huge and inflexible and take unacceptably large amounts of time to write and mammoth amounts of memory to store.

The first iterative statements in programming languages were directly related to arrays. This resulted from the fact that in the earliest years of computers, computing was largely numerical in nature, frequently using loops to process data in arrays.

Several categories of iteration control statements have been developed. The primary categories are defined by how designers answered two basic design questions:

How is the iteration controlled?

Where should the control mechanism appear in the loop statement?

The primary possibilities for iteration control are logical, counting, or a combination of the two. The main choices for the location of the control mechanism are the top of the loop or the bottom of the loop. Top and bottom here are logical, rather than physical, denotations. The issue is not the physical placement of the control mechanism; rather, it is whether the mechanism is executed and affects control before or after execution of the statement’s body. A third option, which allows the user to decide where to put the control, at the top, at the bottom, or even within the controlled segment, is discussed in Section 8.3.3.

The body of an iterative statement is the collection of statements whose execution is controlled by the iteration statement. We use the term pretest to mean that the test for loop completion occurs before the loop body is executed and posttest to mean that it occurs after the loop body is executed. The iteration statement and the associated loop body together form an iteration statement.

8.3.1 Counter-Controlled Loops A counting iterative control statement has a variable, called the loop variable, in which the count value is maintained. It also includes some means of specifying the initial and terminal values of the loop variable, and the difference between sequential loop variable values, often called the stepsize. The initial, terminal, and stepsize specifications of a loop are called the loop parameters.

Although logically controlled loops are more general than counter-​controlled loops, they are not necessarily more commonly used. Because ​counter- controlled loops are more complex, their design is more demanding.

Counter-controlled loops are sometimes supported by machine instructions designed for that purpose. Unfortunately, machine architecture can outlive the prevailing approaches to programming at the time of the architecture design. For example, VAX computers have a very convenient instruction for the implementation of posttest counter-controlled loops, which Fortran had at the time of the design of the VAX (mid-1970s). But Fortran no longer had such a loop by the time VAX computers became widely used (it had been replaced by a pretest loop). Furthermore, no other widely used language of the time had a posttest counting loop.

8.3.1.1 Design Issues There are many design issues for iterative counter-controlled statements. The nature of the loop variable and the loop parameters provide a number of

design issues. The type of the loop variable and that of the loop parameters obviously should be the same or at least compatible, but what types should be allowed? One apparent choice is integer, but what about enumeration, character, and floating-point types? Another question is whether the loop variable is a normal variable, in terms of scope, or whether it should have some special scope. Allowing the user to change the loop variable or the loop parameters within the loop can lead to code that is very difficult to understand, so another question is whether the additional flexibility that might be gained by allowing such changes is worth that additional complexity. A similar question arises about the number of times and the specific time when the loop parameters are evaluated: If they are evaluated just once, loops are simple but less flexible. Finally, what is the value of the loop variable after loop termination, if its scope extends beyond the loop?

The following is a summary of these design issues:

What are the type and scope of the loop variable?

Should it be legal for the loop variable or loop parameters to be changed in the loop, and if so, does the change affect loop control?

Should the loop parameters be evaluated only once, or once for every iteration?

What is the value of the loop variable after loop termination?

The issue of the value of the loop variable after loop termination is solved in some languages, such as Fortran 90, by making the loop variable undefined after loop termination. Other languages, such as Ada, make the scope of the loop variable the loop itself.

8.3.1.2 The for Statement of the C- Based Languages The general form of C’s for statement is

for (expression_1; expression_2; expression_3)

 loop body

The loop body can be a single statement, a compound statement, or a null statement.

Because assignment statements in C produce results and thus can be considered expressions, the expressions in a for statement are often assignment statements. The first expression is for initialization and is evaluated only once, when the for statement execution begins. The second expression is the loop control and is evaluated before each execution of the loop body. As is usual in C, a zero value means false and all nonzero values mean true. Therefore, if the value of the second expression is zero, the for is terminated; otherwise, the loop body statements are executed. In C99, the expression also could be a Boolean type. A C99 Boolean type stores only the values 0 or 1. The last expression in the for is executed after each execution of the loop body. It is often used to increment the loop counter. An operational semantics description of the C for statement is shown next. Because C expressions can be used as statements, expression evaluations are shown as statements.

    expression_1

loop:

    if expression_2 = 0 goto out

    [loop body]

    expression_3

    goto loop

out: . . .

Following is an example of a skeletal C for statement:

for (count = 1; count <= 10; count++)

. . .

}

All of the expressions of C’s for are optional. An absent second expression is considered true, so a for without one is potentially an infinite loop. If the first and/or third expressions are absent, no assumptions are made. For example, if the first expression is absent, it simply means that no initialization takes place.

Note that C’s for need not count. It can easily model counting and logical loop structures, as demonstrated in the next section.

The C for design choices are the following: There is no explicit loop variable and no loop parameters. All involved variables can be changed in the loop body. The expressions are evaluated in the order stated previously. Although it can create havoc, it is legal to branch into a C for loop body.

C’s for is one of the most flexible, because each of the expressions can comprise multiple expressions, which in turn allow multiple loop variables that can be of any type. When multiple expressions are used in a single expression of a for statement, they are separated by commas. All C statements have values, and this form of multiple expression is no exception. The value of such a multiple expression is the value of the last component.

Consider the following for statement:

for (count1 = 0, count2 = 1.0;

count1 <= 10 && count2 <= 100.0;

sum = ++count1 + count2, count2 *= 2.5);

The operational semantics description of this is

    count1 = 0

    count2 = 1.0

loop:

    if count1 > 10 goto out

    if count2 > 100.0 goto out

    count1 = count1 + 1

    sum = count1 + count2

    count2 = count2 * 2.5

    goto loop

out: . . .

The example C for statement does not need and thus does not have a loop body. All the desired actions are part of the for statement itself, rather than in its body. The first and third expressions are multiple statements. In both of these cases, the whole expression is evaluated, but the resulting value is not used in the loop control.

The for statement of C99 and C++ differs from that of earlier versions of C in two ways. First, in addition to an arithmetic expression, it can use a Boolean expression for loop control. Second, the first expression can include variable definitions. For example,

for (int count = 0; count < len; count++) { . . . }

The scope of a variable defined in the for statement is from its definition to the end of the loop body.

The for statement of Java and C# is like that of C++, except that the loop control expression is restricted to boolean.

In all of the C-based languages, the last two loop parameters are evaluated with every iteration. Furthermore, variables that appear in the loop parameter expression can be changed in the loop body. Therefore, these loops can be complex and potentially unreliable.

8.3.1.3 The for Statement of Python

The general form of Python’s for is

for loop_variable in object:

 - loop body

[else:

 - else clause]

The loop variable is assigned the value in the object, which is often a range, one for each execution of the loop body. After loop termination, the loop variable has the value last assigned to it. The loop variable can be changed in the loop body, but such a change has no effect on loop operation. The else clause, when present, is executed if the loop terminates normally.

Consider the following example:

for count in [2, 4, 6]:

print count

produces

2

4

6

For most simple counting loops in Python, the range function is used. range takes one, two, or three parameters. The following examples demonstrate the actions of range:

range(5) returns [0, 1, 2, 3, 4]

range(2, 7) returns [2, 3, 4, 5, 6]

range(0, 8, 2) returns [0, 2, 4, 6]

Note that range never returns the highest value in a given parameter range.

8.3.1.4 Counter-Controlled Loops in

Functional Languages Counter-controlled loops in imperative languages use a counter variable, but such variables do not exist in pure functional languages. Rather than iteration to control repetition, functional languages use recursion. Rather than a statement, functional languages use a recursive function. Counting loops can be simulated in functional languages as follows: The counter can be a parameter for a function that repeatedly executes the loop body, which can be specified in a second function sent to the loop function as a parameter. So, such a loop function takes the body function and the number of repetitions as parameters.

The general form of an F# function for simulating counting loops, named forLoop in this case, is as follows:

let rec forLoop loopBody reps =

if reps <= 0 then

()

else

loopBody()

forLoop loopBody, (reps - 1);;

In this function, the parameter loopBody is the function with the body of the loop and the parameter reps is the number of repetitions. The reserved word rec appears before the name of the function to indicate that it is recursive. The empty parentheses do nothing; they are there because in F# an empty statement is illegal and every if must have an else clause.

8.3.2 Logically Controlled Loops In many cases, collections of statements must be repeatedly executed, but the repetition control is based on a Boolean expression rather than a counter. For these situations, a logically controlled loop is convenient. Actually, logically controlled loops are more general than counter-controlled loops. Every counting loop can be built with a logical loop, but the reverse is not true. Also, recall that only selection and logical loops are essential to express the

control structure of any flowchart.

8.3.2.1 Design Issues Because they are much simpler than counter-controlled loops, logically controlled loops have fewer design issues.

Should the control be pretest or posttest?

Should the logically controlled loop be a special form of a counting loop or a separate statement?

8.3.2.2 Examples The C-based programming languages include both pretest and posttest logically controlled loops that are not special forms of their counter- controlled iterative statements. The pretest and posttest logical loops have the following forms:

while (control_expression)

 loop body

and

do

 loop body

while (control_expression);

These two statement forms are exemplified by the following C# code segments:

sum = 0;

indat = Int32.Parse(Console.ReadLine());

while (indat >= 0) {

sum += indat;

indat = Int32.Parse(Console.ReadLine());

}

value = Int32.Parse(Console.ReadLine());

do {

value /= 10;

digits ++;

} while (value > 0);

Note that all variables in these examples are integer type. The ReadLine method of the Console object gets a line of text from the keyboard. Int32.Parse finds the number in its string parameter, converts it to int type, and returns it.

In the pretest version of a logical loop (while), the statement or statement segment is executed as long as the expression evaluates to true. In the posttest version (do), the loop body is executed until the expression evaluates to false. In both cases, the statement can be compound. The operational semantics descriptions of those two statements follow:

while

loop:

 if control_expression is false goto out

 [loop body]

  goto loop

out: . . .

do-while

loop:

  [loop body]

  if control_expression is true goto loop

It is legal in both C and C++ to branch into both while and do loop bodies. The C89 version uses an arithmetic expression for control; in C99 and C++, it may be either arithmetic or Boolean.

Java’s while and do statements are similar to those of C and C++, except the control expression must be boolean type, and because Java does not have a goto, the loop bodies cannot be entered anywhere except at their beginnings.

Posttest loops are infrequently useful and also can be somewhat dangerous, in the sense that programmers sometimes forget that the loop body will always be executed at least once. The syntactic design of placing a posttest control physically after the loop body, where it has its semantic effect, helps avoid such problems by making the logic clear.

A pretest logical loop can be simulated in a purely functional form with a recursive function that is similar to the one used to simulate a counting loop statement in Section 8.3.1.5. In both cases, the loop body is written as a function. Following is the general form of a simulated logical pretest loop, written in F#:

let rec whileLoop test body =

if test() then

body()

whileLoop test body

else

();;

8.3.3 User-Located Loop Control Mechanisms In some situations, it is convenient for a programmer to choose a location for loop control other than the top or bottom of the loop body. As a result, some languages provide this capability. A syntactic mechanism for user-located loop control can be relatively simple, so its design is not difficult. Such loops have the structure of infinite loops but include one or more user-located loop exits. Perhaps the most interesting question is whether a single loop or

several nested loops can be exited. The design issues for such a mechanism are the following:

Should the conditional mechanism be an integral part of the exit?

Should only one loop body be exited, or can enclosing loops also be exited?

C, C++, Python, Ruby, and C# have unconditional unlabeled exits (break). Java and Perl have unconditional labeled exits (break in Java, last in Perl).

Following is an example of nested loops in Java, in which there is a break out of the outer loop from the nested loop:

outerLoop:

for (row = 0; row < numRows; row++)

for (col = 0; col < numCols; col++) {

sum += mat[row][col];

if (sum > 1000.0)

break outerLoop;

}

C, C++, and Python include an unlabeled control statement, continue, that transfers control to the control mechanism of the smallest enclosing loop. This is not an exit but rather a way to skip the rest of the loop statements on the current iteration without terminating the loop construct. For example, consider the following:

while (sum < 1000) {

getnext(value);

if (value < 0) continue;

sum += value;

}

A negative value causes the assignment statement to be skipped, and control is transferred instead to the conditional at the top of the loop. On the other hand, in

while (sum < 1000) {

getnext(value);

if (value < 0) break;

sum += value;

}

a negative value terminates the loop.

Both last and break provide for multiple exits from loops, which may seem to be somewhat of a hindrance to readability. However, unusual conditions that require loop termination are so common that such a statement is justified. Furthermore, readability is not seriously harmed, because the target of all such loop exits is the first statement after the loop (or an enclosing loop) rather than just anywhere in the program. Finally, the alternative of using multiple breaks to leave more than one level of loops is even more detrimental to readability.

The motivation for user-located loop exits is simple: They fulfill a common need for goto statements using a highly restricted branch statement. The target of a goto can be many places in the program, both above and below the goto itself. However, the targets of user-located loop exits must be below the exit and can only follow immediately at the end of a compound statement.

8.3.4 Iteration Based on Data Structures A general data-based iteration statement uses a user-defined data structure and a user-defined function (the iterator) to go through the structure’s elements. The iterator is called at the beginning of each iteration, and each time it is called, the iterator returns an element from a particular data structure in some specific order. For example, suppose a program has a user- defined binary tree of data nodes, and the data in each node must be processed in some particular order. A user-defined iteration statement for the tree would successively set the loop variable to point to the nodes in the tree, one for each iteration. The initial execution of the user-defined iteration statement needs to issue a special call to the iterator to get the first tree element. The iterator must always remember which node it presented last so that it visits all nodes without visiting any node more than once. So an iterator must be history sensitive. A user-defined iteration statement

terminates when the iterator fails to find more elements.

The for statement of the C-based languages, because of its great flexibility, can be used to simulate a user-defined iteration statement. Once again, suppose the nodes of a binary tree are to be processed. If the tree root is pointed to by a variable named root, and if traverse is a function that sets its parameter to point to the next element of a tree in the desired order, the following could be used:

for (ptr = root; ptr == null; ptr = traverse(ptr)) {

. . .

}

In this statement, traverse is the iterator.

User-defined iteration statements are more important in object-oriented programming than they were in earlier software development paradigms, because users of object-oriented programming routinely use abstract data types for data structures, especially collections. In such cases, a user-defined iteration statement and its iterator must be provided by the author of the data abstraction because the representation of the objects of the type is not known to the user.

An enhanced version of the for statement was added to Java in Java 5.0. This statement simplifies iterating through the values in an array or objects in a collection that implements the Iterable interface. (All of the predefined generic collections in Java implement Iterable.) For example, if we had an ArrayList4 collection named myList of strings, the following statement would iterate through all of its elements, setting each to myElement:

4. An ArrayList is a predefined generic collection that is actually a dynamic- length array of whatever type it is declared to store.

for (String myElement : myList) { . . . }

This new statement is referred to as “foreach,” although its reserved word is for.

C# and F# (and the other .NET languages) also have generic library classes

for collections. For example, there are generic collection classes for lists, which are dynamic length arrays, stacks, queues, and dictionaries (hash table). All of these predefined generic collections have built-in iterators that are used implicitly with the foreach statement. Furthermore, users can define their own collections and write their own iterators, which can implement the IEnumerator interface, which enables the use of foreach on these collections.

For example, consider the following C# code:

List<String> names = new List<String>();

names.Add("Bob");

names.Add("Carol");

names.Add("Alice");

. . .

foreach (String name in names)

Console.WriteLine(name);

In Ruby, a block is a sequence of code, delimited by either braces or the do and end reserved words. Blocks can be used with specially written methods to create many useful constructs, including iterators for data structures. This construct consists of a method call followed by a block. A block is actually an anonymous method that is sent to the method (whose call precedes it) as a parameter. The called method can then call the block, which can produce output or objects.

Ruby predefines several iterator methods, such as times and upto for counter-controlled loops, and each for simple iterations of arrays and hashes. For example, consider the following example of using times:

>> 4.times {puts "Hey!"}

Hey!

Hey!

Hey!

Hey!

=> 4

Note that >> is the prompt of the interactive Ruby interpreter and => is used to indicate the return value of the expression. The Ruby puts statement displays its parameter. In this example, the times method is sent to the object 4, with the block sent along as a parameter. The times method calls the block

four times, producing the four lines of output. The destination object, 4, is the return value from times.

The most common Ruby iterator is each, which is often used to go through arrays and apply a block to each element.5 For this purpose, it is convenient to allow blocks to have parameters, which, if present, appear at the beginning of the block, delimited by vertical bars ( | ). The following example, which uses a block parameter, illustrates the use of each:

5. This is similar to the map functions discussed in Chapter 15.

>> list = [2, 4, 6, 8]

=> [2, 4, 6, 8]

>> list.each {|value| puts value}

2

4

6

8

=> [2, 4, 6, 8]

In this example, the block is called for each element of the array to which the each method is sent. The block produces the output, which is a list of the array’s elements. The return value of each is the array to which it is sent.

Instead of a counting loop, Ruby has the upto method. For example, we could have the following:

1.upto(5) {|x| print x, " "}

This produces the following output:

1 2 3 4 5

Syntax that resembles a for loop in other languages could also be used, as in the following:

for x in 1..5

print x, " "

end

Ruby actually has no for statement—constructs like the above are converted

by Ruby into upto method calls.

Now we consider how blocks work. The yield statement is similar to a method call, except that there is no receiver object and the call is a request to execute the block attached to the method call, rather than a call to a method. yield is only called in a method that has been called with a block. If the block has parameters, they are specified in parentheses in the yield statement. The value returned by a block is that of the last expression evaluated in the block. It is this process that is used to implement the built-in iterators, such as times.

Python provides strong support for iteration. Suppose one needs to process the nodes in some user-defined data structure. Further suppose that the structure has a traversal method that goes through the nodes of the structure in the desired order. The following skeletal class definition includes such a traversal method that produces the nodes of an instance of this class, one at a time.

class MyStructure:

# Other method definitions, including a constructor

def traverse(self):

# if there is another node:

# set nod to next node

# else:

# return

yield nod

The traverse method appears to be a regular Python method, but it contains a yield statement, which dramatically changes the semantics of the method. In effect, the method is run in a separate thread of control. The yield statement acts like a return. On the first call to traverse, yield returns the initial node of the structure. However, on the second call, it returns the second node. On all but the first call to traverse, it begins its execution where it left off on the previous execution. Instead of restarting at its beginning, it is resumed. Any local storage in such a method is maintained across its calls. In the case of traverse, subsequent calls begin their execution at the beginning of its code, but in the state that it was in its previous execution. In Python, any method that contains a yield statement is called a generator, because it generates data one element at a time.

Of course, one could also produce all of the nodes of the structure, store them in an array, and process them from the array. However, the number of nodes could be large, requiring a large array to store them. The approach using the iterator is more elegant and is not affected by the size of the data structure.

8.4 Unconditional Branching An unconditional branch statement transfers execution control to a specified location in the program. The most heated debate in language design in the late 1960s was over the issue of whether unconditional branching should be part of any high-level language, and if so, whether its use should be restricted. The unconditional branch, or goto, is the most powerful statement for controlling the flow of execution of a program’s statements. However, careless use of the goto can lead to serious problems. The goto has stunning power and great flexibility (all other control structures can be built with goto and a selector), but it is this power that makes its use dangerous. Without usage restrictions, imposed by either language design or programming standards, goto statements can make programs very difficult to read, and as a result, highly unreliable and costly to maintain.

history note Although several thoughtful people had pointed out the potential problems of gotos earlier, it was Edsger Dijkstra who gave the computing world the first widely read exposé on the dangers of the goto. In his letter he noted, “The goto statement as it stands is just too primitive; it is too much an invitation to make a mess of one’s program” (Dijkstra, 1968a). During the first few years after publication of Dijkstra’s views on the goto, a large number of people argued publicly for either outright banishment or at least restrictions on the use of the goto. Among those who did not favor complete elimination was Donald Knuth (1974), who argued that there were occasions when the efficiency of the goto outweighed its harm to readability.

These problems follow directly from a goto’s ability to force any program statement to follow any other in execution sequence, regardless of whether that statement precedes or follows the previously executed statement in textual order. Readability is best when the execution order of statements in a program is nearly the same as the order in which they appear—in our case,

this would mean top to bottom, which is the order to which we are accustomed. Thus, restricting gotos so they can transfer control only downward in a program partially alleviates the problem. It allows gotos to transfer control around code sections in response to errors or unusual conditions but disallows their use to build any sort of loop.

A few languages have been designed without a goto—for example, Java, Python, and Ruby. However, most currently popular languages include a goto statement. Kernighan and Ritchie (1978) call the goto infinitely abusable, but it is nevertheless included in Ritchie’s language, C. The languages that have eliminated the goto have provided additional control statements, usually in the form of loop exits, to code one of the justifiable applications of the goto.

The relatively new language, C#, includes a goto, even though one of the languages on which it is based, Java, does not. One legitimate use of C#’s goto is in the switch statement, as discussed in Section 8.2.2.2.

All of the loop exit statements discussed in Section 8.3.3 are actually camouflaged goto statements. They are, however, severely restricted gotos and are not harmful to readability. In fact, it can be argued that they improve readability, because to avoid their use results in convoluted and unnatural code that would be much more difficult to understand.

8.5 Guarded Commands Quite different forms of selection and loop structures were suggested by Dijkstra (1975). His primary motivation was to provide control statements that would support a program design methodology that ensured correctness during development rather than when verifying or testing completed programs. This methodology is described in Dijkstra (1976). Another motivation is the increased clarity in reasoning that is possible with guarded commands. Simply put, a selectable segment of a selection statement in a guarded-command statement can be considered independently of any other part of the statement, which is not true for the selection statements of the common programming languages.

Guarded commands are covered in this chapter because they are the basis for the linguistic mechanism developed later for concurrent programming in CSP (Hoare, 1978). Guarded commands are also used to define functions in Haskell, as discussed in Chapter 15.

Dijkstra’s selection statement has the form

if <Boolean expression> -> <statement>

[] <Boolean expression> -> <statement>

[] . . .

[] <Boolean expression> -> <statement>

fi

The closing reserved word, fi, is the opening reserved word spelled backward. This form of closing reserved word is taken from ALGOL 68. The small blocks, called fatbars, are used to separate the guarded clauses and allow the clauses to be statement sequences. Each line in the selection statement, consisting of a Boolean expression (a guard) and a statement or statement sequence, is called a guarded command.

This selection statement has the appearance of a multiple selection, but its semantics is different. All of the Boolean expressions are evaluated each time the statement is reached during execution. If more than one expression is true, one of the corresponding statements can be nondeterministically chosen for execution. An implementation might always choose the statement associated with the first Boolean expression that evaluates to be true. But it may choose any statement associated with a true Boolean expression. So, the correctness of the program cannot depend on which statement is chosen (among those associated with true Boolean expressions). If none of the Boolean expressions are true, a run-time error occurs that causes program termination. This forces the programmer to consider and list all possibilities. Consider the following example:

if i = 0 -> sum := sum + i

[] i > j -> sum := sum + j

[] j > i -> sum := sum + k

fi

If i = 0 and j > i, this statement chooses nondeterministically between the first and third assignment statements. If i is equal to j and is not zero, a run- time error occurs because none of the conditions are true.

This statement can be an elegant way of allowing the programmer to state that the order of execution, in some cases, is irrelevant. For example, to find the largest of two numbers, we can use

if x >= y -> max := x

[] y >= x -> max := y

fi

This computes the desired result without overspecifying the solution. In particular, if x and y are equal, it does not matter which we assign to max. This is a form of abstraction provided by the nondeterministic semantics of the statement.

Now, consider this same process coded in a traditional programming language selector:

if (x >= y)

max = x;

else

max = y;

This could also be coded as follows:

if (x > y)

max = x;

else

max = y;

There is no practical difference between these two statements. The first assigns x to max when x and y are equal; the second assigns y to max in the same circumstance. This choice between the two statements complicates the formal analysis of the code and the correctness proof of it. This is one of the reasons why guarded commands were developed by Dijkstra.

The loop structure proposed by Dijkstra has the form

do <Boolean expression> -> <statement>

[] <Boolean expression> -> <statement>

[] . . .

[] <Boolean expression> -> <statement>

od

The semantics of this statement is that all Boolean expressions are evaluated on each iteration. If more than one is true, one of the associated statements is nondeterministically (perhaps randomly) chosen for execution, after which the expressions are again evaluated. When all expressions are simultaneously false, the loop terminates.

Consider the following problem: Given four integer variables, q1, q2, q3, and q4, rearrange the values of the four so that q1 <= q2 <= q3 <= q4. Without guarded commands, one straightforward solution is to put the four values into an array, sort the array, and then assign the values from the array back into the scalar variables q1, q2, q3, and q4. While this solution is not difficult, it requires a good deal of code, especially if the sort process must be included.

Now, consider the following code, which uses guarded commands to solve the same problem but in a more concise and elegant way.6

6. This code appears in a slightly different form in Dijkstra (1975).

do q1 > q2 -> temp := q1; q1 := q2; q2 := temp;

[] q2 > q3 -> temp := q2; q2 := q3; q3 := temp;

[] q3 > q4 -> temp := q3; q3 := q4; q4 := temp;

od

Dijkstra’s guarded command control statements are interesting, in part because they illustrate how the syntax and semantics of statements can have an impact on program verification and vice versa. Program verification is virtually impossible when goto statements are used. Verification is greatly simplified if (1) only logical loops and selections are used or (2) only guarded commands are used. The axiomatic semantics of guarded commands are conveniently specified (Gries, 1981). It should be obvious, however, that there is considerably increased complexity in the implementation of the guarded commands over their conventional deterministic counterparts.

8.6 Conclusions We have described and discussed a variety of statement-level control structures. A brief evaluation now seems to be in order.

First, we have the theoretical result that only sequence, selection, and ​pretest logical loops are absolutely required to express computations (Böhm and Jacopini, 1966). This result has been used by those who wish to ban unconditional branching altogether. Of course, there are already sufficient practical problems with the goto to condemn it without also using a theoretical reason. One of the main legitimate needs for gotos—premature exits from loops—can be met with restricted branch statements, such as break.

One obvious misuse of the Böhm and Jacopini result is to argue against the inclusion of any control structures beyond selection and pretest logical loops. No widely used language has yet taken that step; furthermore, we doubt that any ever will, because of the negative effect on writability and readability. Programs written with only selection and pretest logical loops are generally less natural in structure, more complex, and therefore harder to write and more difficult to read. For example, the C# multiple selection structure is a great boost to C# writability, with no obvious negatives. Another example is the counting loop structure of many languages, especially when the statement is simple.

It is not so clear that the utility of many of the other control structures that have been proposed is worth their inclusion in languages (Ledgard and Marcotty, 1975). This question rests to a large degree on the fundamental question of whether the size of languages must be minimized. Both Wirth (1975) and Hoare (1973) strongly endorse simplicity in language design. In the case of control structures, simplicity means that only a few control statements should be in a language, and they should be simple.

The rich variety of statement-level control structures that have been invented shows the diversity of opinion among language designers. After all the

invention, discussion, and evaluation, there is still no unanimity of opinion on the precise set of control statements that should be in a language. Most contemporary languages do, of course, have similar control statements, but there is still some variation in the details of their syntax and semantics. Furthermore, there is still disagreement on whether a language should include a goto; C++ and C# do, but Java and Ruby do not.

SUMMARY Control statements occur in several categories: selection, multiple selection, iterative, and unconditional branching.

The switch statement of the C-based languages is representative of ​multiple- selection statements. The C# version eliminates the reliability problem of its predecessors by disallowing the implicit continuation from a selected segment to the following selectable segment.

A large number of different loop statements have been invented for high- level languages. C’s for statement is the most flexible iteration statement, although its flexibility leads to some reliability problems.

Most languages have exit statements for their loops; these statements take the place of one of the most common uses of goto statements.

Data-based iterators are loop statements for processing data structures, such as linked lists, hashes, and trees. The for statement of the C-based languages allows the user to create iterators for user-defined data. The foreach statement of Perl and C# is a predefined iterator for standard data structures. In the contemporary object-oriented languages, iterators for collections are specified with standard interfaces, which are implemented by the designers of the collections.

Ruby includes iterators that are a special form of methods that are sent to various objects. The language predefines iterators for common uses, but also allows user-defined iterators.

The unconditional branch, or goto, has been part of most imperative languages. Its problems have been widely discussed and debated. The current consensus is that it should remain in most languages but that its dangers should be minimized through programming discipline.

Dijkstra’s guarded commands are alternative control statements with positive theoretical characteristics. Although they have not been adopted as the

control statements of a language, part of the semantics appear in the concurrency mechanisms of CSP and the function definitions of Haskell.

REVIEW QUESTIONS 1. What is the definition of control structure?

2. What did Böhm and Jocopini prove about flowcharts?

3. What is the definition of block?

4. What is/are the design issue(s) for all selection and iteration control statements?

5. What are the design issues for selection structures?

6. What is unusual about Python’s design of compound statements?

7. Under what circumstances must an F# selector have an else clause?

8. What are the common solutions to the nesting problem for two-way selectors?

9. What are the design issues for multiple-selection statements?

10. Between what two language characteristics is a trade-off made when deciding whether more than one selectable segment is executed in one execution of a multiple selection statement?

11. What is unusual about C’s multiple-selection statement?

12. On what previous language was C’s switch statement based?

13. Explain how C#’s switch statement is safer than that of C.

14. What are the design issues for all iterative control statements?

15. What are the design issues for counter-controlled loop statements?

16. What is a pretest loop statement? What is a posttest loop statement?

17. What is the difference between the for statement of C++ and that of Java?

18. In what way is C’s for statement more flexible than that of many other languages?

19. What does the range function in Python do?

20. What contemporary languages do not include a goto?

21. What are the design issues for logically controlled loop statements?

22. What is the main reason user-located loop control statements were invented?

23. What are the design issues for user-located loop control mechanisms?

24. What advantage does Java’s break statement have over C’s break statement?

25. What are the differences between the break statement of C++ and that of Java?

26. What is a user-defined iteration control?

27. What Scheme function implements a multiple selection statement?

28. How does a functional language implement repetition?

29. How are iterators implemented in Ruby?

30. What language predefines iterators that can be explicitly called to iterate over its predefined data structures?

31. What common programming language borrows part of its design from Dijkstra’s guarded commands?

PROBLEM SET 1. Describe three situations where a combined counting and logical looping

statement is needed.

2. Study the iterator feature of CLU in Liskov et al. (1981) and determine its advantages and disadvantages.

3. Compare the set of Ada control statements with those of C# and decide which are better and why.

4. What are the pros and cons of using unique closing reserved words on compound statements?

5. What are the arguments, pros and cons, for Python’s use of indentation to specify compound statements in control statements?

6. Analyze the potential readability problems with using closure reserved words for control statements that are the reverse of the corresponding initial reserved words, such as the case-esac reserved words of ALGOL 68. For example, consider common typing errors such as transposing characters.

7. Use the Science Citation Index to find an article that refers to Knuth (1974). Read the article and Knuth’s paper and write a paper that summarizes both sides of the goto issue.

8. In his paper on the goto issue, Knuth (1974) suggests a loop control statement that allows multiple exits. Read the paper and write an operational semantics description of the statement.

9. What are the arguments both for and against the exclusive use of Boolean expressions in the control statements in Java (as opposed to also allowing arithmetic expressions, as in C++)?

10. Describe a programming situation in which the else clause in Python’s

for statement would be convenient.

11. Describe three specific programming situations that require a posttest loop.

12. Speculate as to the reason control can be transferred into a C loop statement.

PROGRAMMING EXERCISES 1. Rewrite the following pseudocode segment using a loop structure in the

specified languages:

k = (j + 13) / 27

loop:

if k > 10 then goto out

k = k + 1

i = 3 * k - 1

goto loop

out: . . .

1. C, C++, Java, or C#

2. Python

3. Ruby

Assume all variables are integer type. Discuss which language, for this code, has the best writability, the best readability, and the best combination of the two.

2. Redo Programming Exercise 1, except this time make all the variables and constants floating-point type, and change the statement

k = k + 1

to

k = k + 1.2

3. Rewrite the following code segment using a multiple-selection statement in the following languages:

if ((k == 1) || (k == 2)) j = 2 * k - 1

if ((k == 3) || (k == 5)) j = 3 * k + 1

if (k == 4) j = 4 * k - 1

if ((k == 6) || (k == 7) || (k == 8)) j = k - 2

1. C, C++, Java, or C#

2. Python

3. Ruby

Assume all variables are integer type. Discuss the relative merits of the use of these languages for this particular code.

4. Consider the following C program segment. Rewrite it using no gotos or breaks.

j = -3;

for (i = 0; i < 3; i++) {

switch (j + 2) {

case 3:

case 2: j--; break;

case 0: j += 2; break;

default: j = 0;

}

if (j > 0) break;

j = 3 - i

}

5. In a letter to the editor of CACM, Rubin (1987) uses the following code segment as evidence that the readability of some code with gotos is better than the equivalent code without gotos. This code finds the first row of an n by n integer matrix named x that has nothing but zero values.

for (i = 1; i <= n; i++) {

for (j = 1; j <= n; j++)

if (x[i][j] != 0)

goto reject;

println ('First all-zero row is:', i);

break;

reject:

}

Rewrite this code without gotos in one of the following languages: C, C++, Java, or C#. Compare the readability of your code to that of the example code.

6. Consider the following programming problem: The values of three integer variables—first, second, and third—must be placed in the three variables max, mid, and min, with the obvious meanings, without using arrays or user-defined or predefined subprograms. Write two solutions to this problem, one that uses nested selections and one that does not. Compare the complexity and expected reliability of the two.

7. Rewrite the C program segment of Programming Exercise 4 using if and goto statements in C.

8. Rewrite the C program segment of Programming Exercise 4 in Java without using a switch statement.

9. Translate the following call to Scheme’s COND to C and set the resulting value to y.

(COND

((> x 10) x)

((< x 5) (* 2 x))

((= x 7) (+ x 10))

)

9 Subprograms 1. 9.1 Introduction

2. 9.2 Fundamentals of Subprograms

3. 9.3 Design Issues for Subprograms

4. 9.4 Local Referencing Environments

5. 9.5 Parameter-Passing Methods

6. 9.6 Parameters That Are Subprograms

7. 9.7 Calling Subprograms Indirectly

8. 9.8 Design Issues for Functions

9. 9.9 Overloaded Subprograms

10. 9.10 Generic Subprograms

11. 9.11 User-Defined Overloaded Operators

12. 9.12 Closures

13. 9.13 Coroutines

Subprograms are the fundamental building blocks of programs and are therefore among the most important concepts in programming language design. We now explore the design of subprograms, including parameter- passing methods, local referencing environments, overloaded subprograms, generic subprograms, and the aliasing and problematic side effects that are associated with subprograms. We also include discussions of indirectly called subprograms, closures, and coroutines.

Implementation methods for subprograms are discussed in Chapter 10.

9.1 Introduction Two fundamental abstraction facilities can be included in a programming language: process abstraction and data abstraction. In the early history of high-level programming languages, only process abstraction was included. Process abstraction, in the form of subprograms, has been a central concept in all programming languages. In the 1980s, however, many people began to believe that data abstraction was equally important. Data abstraction is discussed in detail in Chapter 11.

The first programmable computer, Babbage’s Analytical Engine, built in the 1840s, had the capability of reusing collections of instruction cards at several different places in a program. In a modern programming language, such a collection of statements is written as a subprogram. This reuse results in savings in memory space and coding time. Such reuse is also an abstraction, for the details of the subprogram’s computation are replaced in a program by a statement that calls the subprogram. Instead of describing how some computation is to be done in a program, that description (the collection of statements in the subprogram) is enacted by a call statement, effectively abstracting away the details. This increases the readability of a program by emphasizing its logical structure while hiding its low-level details.

The methods of object-oriented languages are closely related to the subprograms discussed in this chapter. The primary way methods differ from subprograms is the way they are called and their associations with classes and objects. Although these special characteristics of methods are discussed in Chapter 12, the features they share with subprograms, such as parameters and local variables, are discussed in this chapter.

9.2 Fundamentals of Subprograms

9.2.1 General Subprogram Characteristics All subprograms discussed in this chapter, except the coroutines described in Section 9.13, have the following characteristics:

Each subprogram has a single entry point.

The calling program unit is suspended during the execution of the called subprogram, which implies that there is only one subprogram in execution at any given time.

Control always returns to the caller when the subprogram execution terminates.

Alternatives to these result in coroutines and concurrent units (Chapter 13).

Most subprograms have names, although some are anonymous. Section 9.12 has examples of anonymous subprograms in C#.

9.2.2 Basic Definitions A subprogram definition describes the interface to and the actions of the subprogram abstraction. A subprogram call is the explicit request that a specific subprogram be executed. A subprogram is said to be active if, after having been called, it has begun execution but has not yet completed that execution. The two fundamental kinds of subprograms, procedures and functions, are defined and discussed in Section 9.2.4.

A subprogram header, which is the first part of the definition, serves

several purposes. First, it specifies that the following syntactic unit is a subprogram definition of some particular kind.1 In languages that have more than one kind of subprogram, the kind of the subprogram is usually specified with a special word. Second, if the subprogram is not anonymous, the header provides a name for the subprogram. Third, it may specify a list of parameters.

1. Some programming languages include both kinds of subprograms, procedures and functions.

Consider the following header examples:

def adder parameters):

This is the header of a Python subprogram named adder. Ruby subprogram headers also begin with def. The header of a JavaScript subprogram begins with function.

In C, the header of a function named adder might be as follows:

void adder (parameters)

The reserved word void in this header indicates that the subprogram does not return a value.

The body of subprograms defines its actions. In the C-based languages (and some others—for example, JavaScript) the body of a subprogram is delimited by braces. In Ruby, an end statement terminates the body of a subprogram. As with compound statements, the statements in the body of a Python function must be indented and the end of the body is indicated by the first statement that is not indented.

One characteristic of Python functions that sets them apart from the functions of other common programming languages is that function def statements are executable. When a def statement is executed, it assigns the given name to the given function body. Until a function’s def has been executed, the function cannot be called. Consider the following skeletal example:

if . . .

def fun( . . . ):

. . .

else

def fun( . . . ):

. . .

If the then clause of this selection construct is executed, that version of the function fun can be called, but not the version in the else clause. Likewise, if the else clause is chosen, its version of the function can be called but the one in the then clause cannot.

Ruby methods differ from the subprograms of other programming languages in several interesting ways. Ruby methods are often defined in class definitions but can also be defined outside class definitions, in which case they are considered methods of the root object, Object. Such methods can be called without an object receiver, as if they were functions in C or C++. If a Ruby method is called without a receiver, self is assumed. If there is no method by that name in the class, enclosing classes are searched, up to Object, if necessary.

The parameter profile of a subprogram contains the number, order, and types of its formal parameters. The protocol of a subprogram is its parameter profile plus, if it is a function, its return type. In languages in which subprograms have types, those types are defined by the subprogram’s protocol.

Subprograms can have declarations as well as definitions. This form parallels the variable declarations and definitions in C, in which declarations are used to provide type information but not to define variables. Subprogram declarations provide the subprogram’s protocol but do not include their bodies. They are necessary in languages that do not allow forward references to subprograms. In both the cases of variables and subprograms, declarations are needed for static type checking. In the case of subprograms, it is the type of the parameters that must be checked. Function declarations are common in C and C++ programs, where they are called prototypes. Such declarations are often placed in header files.

In most other languages (other than C and C++), subprograms do not need declarations, because there is no requirement that subprograms be defined

before they are called.

9.2.3 Parameters Subprograms typically describe computations. There are two ways that a nonmethod subprogram can gain access to the data that it is to process: through direct access to nonlocal variables (declared elsewhere but visible in the subprogram) or through parameter passing. Data passed through parameters are accessed using names that are local to the subprogram. Parameter passing is more flexible than direct access to nonlocal variables. In essence, a subprogram with parameter access to the data that it is to process is a parameterized computation. It can perform its computation on whatever data it receives through its parameters (presuming the types of the parameters are as expected by the subprogram). If data access is through nonlocal variables, the only way the computation can proceed on different data is to assign new values to those nonlocal variables between calls to the subprogram. Extensive access to nonlocals can reduce reliability. Variables that are visible to the subprogram where access is desired often end up also being visible where access to them is not needed. This problem was discussed in Chapter 5.

Although methods also access external data through nonlocal references and parameters, the primary data to be processed by a method is the object through which the method is called. However, when a method does access nonlocal data, the reliability problems are the same as with nonmethod subprograms. Also, in an object-oriented language, method access to class variables (those associated with the class, rather than an object) is related to the concept of nonlocal data and should be avoided whenever possible. In this case, as well as the case of a C function accessing nonlocal data, the method can have the side effect of changing something other than its parameters or local data. Such changes complicate the semantics of the method and make it less reliable.

Pure functional programming languages, such as Haskell, do not have mutable data, so functions written in them are unable to change memory in any way—they simply perform calculations and return a resulting value (or

function, since functions are values in a pure functional language).

In some situations, it is convenient to be able to transmit computations, rather than data, as parameters to subprograms. In these cases, the name of the subprogram that implements that computation may be used as a parameter. This form of parameter is discussed in Section 9.6. Data parameters are discussed in Section 9.5.

The parameters in the subprogram header are called formal parameters. They are sometimes thought of as dummy variables because they are not variables in the usual sense: In most cases, they are bound to storage only when the subprogram is called, and that binding is often through some other program variables.

Subprogram call statements must include the name of the subprogram and a list of parameters to be bound to the formal parameters of the subprogram. These parameters are called actual parameters.2 They must be distinguished from formal parameters, because the two usually have different restrictions on their forms, and of course, their uses are quite different.

2. Some authors call actual parameters arguments and formal parameters just parameters.

In most programming languages, the correspondence between actual and formal parameters—or the binding of actual parameters to formal parameters —is done by position: The first actual parameter is bound to the first formal parameter and so forth. Such parameters are called positional ​parameters. This is an effective and safe method of relating actual parameters to their corresponding formal parameters, as long as the parameter lists are relatively short.

When parameter lists are long, however, it is easy for a programmer to make mistakes in the order of actual parameters in the list. One solution to this problem is to provide keyword parameters, in which the name of the formal parameter to which an actual parameter is to be bound is specified with the actual parameter in a call. The advantage of keyword parameters is that they can appear in any order in the actual parameter list. Python functions can be called using this technique, as in

sumer(length = my_length,

list = my_array,

sum = my_sum)

where the definition of sumer has the formal parameters length, list, and sum.

The disadvantage to keyword parameters is that the user of the subprogram must know the names of formal parameters.

In addition to keyword parameters, some languages, for example Python, allow positional parameters. Keyword and positional parameters can be mixed in a call, as in

sumer(my_length,

sum = my_sum,

list = my_array)

The only restriction with this approach is that after a keyword parameter appears in the list, all remaining parameters must be keyworded. This restriction is necessary because a position may no longer be well defined after a keyword parameter has appeared.

In Python, Ruby, C++, and PHP, formal parameters can have default values. A default value is used if no actual parameter is passed to the formal parameter in the subprogram header. Consider the following Python function header:

def compute_pay(income, exemptions = 1, tax_rate)

The exemptions formal parameter can be absent in a call to compute_pay; when it is, the value 1 is used. No comma is included for an absent actual parameter in a Python call, because the only value of such a comma would be to indicate the position of the next parameter, which in this case is not necessary because all actual parameters after an absent actual parameter must be keyworded. For example, consider the following call:

pay = compute_pay(20000.0, tax_rate = 0.15)

In C++, which does not support keyword parameters, the rules for default

parameters are necessarily different. The default parameters must appear last, because parameters are positionally associated. Once a default parameter is omitted in a call, all remaining formal parameters must have default values. A C++ function header for the compute_pay function can be written as follows:

float compute_pay(float income, float tax_rate,

int exemptions = 1)

Notice that the parameters are rearranged so that the one with the default value is last. An example call to the C++ compute_pay function is

pay = compute_pay(20000.0, 0.15);

In most languages that do not have default values for formal parameters, the number of actual parameters in a call must match the number of formal parameters in the subprogram definition header. However, in C, C++, Perl, and JavaScript, this is not required. When there are fewer actual parameters in a call than formal parameters in a function definition, it is the programmer’s responsibility to ensure that the parameter correspondence, which is always positional, and the subprogram execution are sensible.

Although this design, which allows a variable number of parameters, is clearly prone to error, it is also sometimes convenient. For example, the printf function of C can print any number of items (data values and/or literals).

C# allows methods to accept a variable number of parameters, as long as they are of the same type. The method specifies its formal parameter with the params modifier. The call can send either an array or a list of expressions, whose values are placed in an array by the compiler and provided to the called method. For example, consider the following method:

public void DisplayList(params int[] list) {

foreach (int next in list) {

Console.WriteLine("Next value {0}", next);

}

}

If DisplayList is defined for the class MyClass and we have the following declarations,

Myclass myObject = new Myclass;

int[] myList = new int[6] {2, 4, 6, 8, 10, 12};

DisplayList could be called with either of the following:

myObject.DisplayList(myList);

myObject.DisplayList(2, 4, 3 * x - 1, 17);

Ruby supports a complicated but highly flexible actual parameter configuration. The initial parameters are expressions, whose value objects are passed to the corresponding formal parameters. The initial parameters can be followed by a list of key => value pairs, which are placed in an anonymous hash and a reference to that hash is passed to the next formal parameter. These are used as a substitute for keyword parameters, which Ruby does not support. The hash item can be followed by a single parameter preceded by an asterisk. This parameter is called the array formal parameter. When the method is called, the array formal parameter is set to reference a new Array object. All remaining actual parameters are assigned to the elements of the new Array object. If the actual parameter that corresponds to the array formal parameter is an array, it must also be preceded by an asterisk, and it must be the last actual parameter.3 So, Ruby allows a variable number of parameters in a way similar to that of C#. Because Ruby arrays can store different types, there is no requirement that the actual parameters passed to the array have the same type.

3. Not quite true, because the array formal parameter can be followed by a method or function reference, which is preceded by an ampersand (&).

The following example skeletal function definition and call illustrate the parameter structure of Ruby:

list = [2, 4, 6, 8]

def tester(p1, p2, p3, *p4)

. . .

end . . .

tester('first', mon => 72, tue => 68, wed => 59, *list)

Inside tester, the values of its formal parameters are as follows:

p1 is 'first'

p2 is {mon => 72, tue => 68, wed => 59}

p3 is 2

p4 is [4, 6, 8]

Python supports parameters that are similar to those of Ruby.

9.2.4 Procedures and Functions There are two distinct categories of subprograms—procedures and ​functions —both of which can be viewed as approaches to extending the language. Subprograms are collections of statements that define parameterized computations. Functions return values and procedures do not. In most languages that do not include procedures as a separate form of subprogram, functions can be defined not to return values and they can be used as procedures. The computations of a procedure are enacted by single call statements. In effect, procedures define new statements. For example, if a particular language does not have a sort statement, a user can build a procedure to sort arrays of data and use a call to that procedure in place of the unavailable sort statement. Only some older languages, such as Fortran and Ada, support procedures.

Procedures can produce results in the calling program unit by two methods: (1) If there are variables that are not formal parameters but are still visible in both the procedure and the calling program unit, the procedure can change them; and (2) if the procedure has formal parameters that allow the transfer of data to the caller, those parameters can be changed.

Functions structurally resemble procedures but are semantically modeled on mathematical functions. If a function is a faithful model, it produces no side effects; that is, it modifies neither its parameters nor any variables defined outside the function. Such a function returns a value—that is its only desired effect. The functions in most programming languages have side effects.

Functions are called by appearances of their names in expressions, along with the required actual parameters. The value produced by a function’s execution is returned to the calling code, effectively replacing the call itself. For example, the value of the expression f(x) is whatever value f produces when called with the parameter x. For a function that does not produce side effects,

the returned value is its only effect.

Functions define new user-defined operators. For example, if a language does not have an exponentiation operator, a function can be written that returns the value of one of its parameters raised to the power of another parameter. Its header in C++ could be

float power(float base, float exp)

which could be called with

result = 3.4 * power(10.0, x)

The standard C++ library includes a similar function named pow. Compare this with the same operation in Perl, in which exponentiation is a built-in operation:

result = 3.4 * 10.0 ** x

In some programming languages, users are permitted to overload operators by defining new functions for operators. User-defined overloaded operators are discussed in Section 9.11.

9.3 Design Issues for Subprograms Subprograms are complex structures, and it follows from this that a lengthy list of issues is involved in their design. One obvious issue is the choice of one or more parameter-passing methods that will be used. The wide variety of approaches that have been used in various languages is a reflection of the diversity of opinion on the subject. A closely related issue is whether the types of actual parameters will be type checked against the types of the corresponding formal parameters.

The nature of the local environment of a subprogram dictates to some degree the nature of the subprogram. The most important question here is whether local variables are statically or dynamically allocated.

Next, there is the question of whether subprogram definitions can be nested. Another issue is whether subprogram names can be passed as parameters. If subprogram names can be passed as parameters and the language allows subprograms to be nested, there is the question of the correct referencing environment of a subprogram that has been passed as a parameter.

As seen in Chapter 5, side effects of functions can cause problems. So, restrictions on side effects are a design issue for functions. The types and number of values that can be returned from functions are other design issues.

Finally, there are the questions of whether subprograms can be overloaded or generic. An overloaded subprogram is one that has the same name as another subprogram in the same referencing environment. A generic subprogram is one whose computation can be done on data of different types in different calls. A closure is a nested subprogram and its referencing environment, which together allow the subprogram to be called from anywhere in a program.

The following is a summary of these design issues for subprograms in general. Additional issues that are specifically associated with functions are discussed in Section 9.10.

Are local variables statically or dynamically allocated?

Can subprogram definitions appear in other subprogram definitions?

What parameter-passing method or methods are used?

Are the types of the actual parameters checked against the types of the formal parameters?

If subprograms can be passed as parameters and subprograms can be nested, what is the referencing environment of a passed subprogram?

Are functional side effects allowed?

What types of values can be returned from functions?

How many values can be returned from functions?

Can subprograms be overloaded?

Can subprograms be generic?

If the language allows nested subprograms, are closures supported?

These issues and example designs are discussed in the following sections.

9.4 Local Referencing Environments This section discusses the issues related to variables that are defined within subprograms. The issue of nested subprogram definitions is also briefly covered.

9.4.1 Local Variables Subprograms can define their own variables, thereby defining local referencing environments. Variables that are defined inside subprograms are called local variables, because their scope is usually the body of the subprogram in which they are defined.

In the terminology of Chapter 5, local variables can be either static or stack dynamic. If local variables are stack dynamic, they are bound to storage when the subprogram begins execution and are unbound from storage when that execution terminates. There are several advantages of stack-dynamic local variables, the primary one being flexibility. It is essential that recursive subprograms have stack-dynamic local variables. Another advantage of stack-​dynamic locals is that the storage for local variables in an active subprogram can be shared with the local variables in all inactive subprograms. This is not as important an advantage as it was when computers had smaller memories.

The main disadvantages of stack-dynamic local variables are the following: First, there is the cost of the time required to allocate, initialize (when necessary), and deallocate such variables for each call to the subprogram. Second, accesses to stack-dynamic local variables must be indirect, whereas accesses to static variables can be direct.4 This indirectness is required because the place in the stack where a particular local variable will reside can be determined only during execution (see Chapter 10). Finally, when all local

variables are stack dynamic, subprograms cannot be history sensitive; that is, they cannot retain data values of local variables between calls. It is sometimes convenient to be able to write history-sensitive subprograms. A common example of a need for a history-sensitive subprogram is one whose task is to generate pseudorandom numbers. Each call to such a subprogram computes one pseudorandom number, using the last one it computed. It must, therefore, store the last one in a static local variable. Coroutines and the subprograms used in iterator loop constructs (discussed in Chapter 8) are other examples of subprograms that need to be history sensitive.

4. In some implementations, static variables are also accessed indirectly, thereby eliminating this disadvantage.

The primary advantage of static local variables over stack-dynamic local variables is that they are slightly more efficient—they require no run-time overhead for allocation and deallocation. Also, if accessed directly, these accesses are obviously more efficient. And, of course, they allow subprograms to be history sensitive. The greatest disadvantage of static local variables is their inability to support recursion. Also, their storage cannot be shared with the local variables of other inactive subprograms.

In most contemporary languages, local variables in a subprogram are by default stack dynamic. In C and C++ functions, locals are stack dynamic unless specifically declared to be static. For example, in the following C (or C++) function, the variable sum is static and count is stack dynamic.

int adder(int list[], int listlen) {

static int sum = 0;

int count;

for (count = 0; count < listlen; count ++)

sum += list [count];

return sum;

}

The methods of C++, Java, and C# have only stack-dynamic local variables.

In Python, the only declarations used in method definitions are for globals. Any variable declared to be global in a method must be a variable defined outside the method. A variable defined outside the method can be referenced

in the method without declaring it to be global, but such a variable cannot be assigned in the method. If the name of a global variable is assigned in a method, it is implicitly declared to be a local and the assignment does not disturb the global. All local variables in Python methods are stack dynamic.

9.4.2 Nested Subprograms The idea of nesting subprograms originated with ALGOL 60. The motivation was to be able to create a hierarchy of both logic and scopes. If a subprogram is needed only within another subprogram, why not place it there and hide it from the rest of the program? Because static scoping is usually used in languages that allow subprograms to be nested, this also provides a highly structured way to grant access to nonlocal variables in enclosing subprograms. Recall that in Chapter 5, the problems introduced by this were discussed. For a long time, the only languages that allowed nested subprograms were those directly descending from ALGOL 60, which were ALGOL 68, Pascal, and Ada. Many other languages, including all of the direct descendants of C, do not allow subprogram nesting. Recently, some new languages again allow it. Among these are JavaScript, Python, and Ruby. Also, most functional programming languages allow subprograms to be nested.

9.5 Parameter-Passing Methods Parameter-passing methods are the ways in which parameters are transmitted to and/or from called subprograms. First, we focus on the different semantics models of parameter-passing methods. Then, we discuss the various implementation models invented by language designers for these semantics models. Next, we survey the design choices of several languages and discuss the actual methods used to implement the implementation models. Finally, we consider the design considerations that face a language designer in choosing among the methods.

9.5.1 Semantics Models of Parameter Passing Formal parameters are characterized by one of three distinct semantics models: (1) They can receive data from the corresponding actual parameter; (2) they can transmit data to the actual parameter; or (3) they can do both. These models are called in mode, out mode, and inout mode, respectively. For example, consider a subprogram that takes two arrays of int values as parameters—list1 and list2. The subprogram must add list1 to list2 and return the result as a revised version of list2. Furthermore, the subprogram must create a new array from the two given arrays and return it. For this subprogram, list1 should be in mode, because it is not to be changed by the subprogram. list2 must be inout mode, because the subprogram needs the given value of the array and must return its new value. The third array should be out mode, because there is no initial value for this array and its computed value must be returned to the caller.

There are two conceptual models of how data transfers take place in parameter transmission: Either an actual value is copied (to the caller, to the called, or both ways) or an access path is transmitted. Most commonly, the access path is a simple pointer or reference. Figure 9.1 illustrates the three

semantics models of parameter passing when values are copied.

Figure 9.1 The three semantics models of parameter passing when physical moves are used

Figure 9.1 Full Alternative Text

9.5.2 Implementation Models of Parameter Passing A variety of models have been developed by language designers to guide the implementation of the three basic parameter transmission modes. In the following sections, we discuss several of these, along with their relative

strengths and weaknesses.

9.5.2.1 Pass-by-Value When a parameter is passed by value, the value of the actual parameter is used to initialize the corresponding formal parameter, which then acts as a local variable in the subprogram, thus implementing in-mode semantics.

Pass-by-value is normally implemented by copy, because accesses often are more efficient with this approach. It could be implemented by transmitting an access path to the value of the actual parameter in the caller, but that would require that the value be in a write-protected cell (one that can only be read). Enforcing the write protection is not always a simple matter. For example, suppose the subprogram to which the parameter was passed passes it in turn to another subprogram. This is another reason to use copy transfer. As we will see in Section 9.5.4, C++ provides a convenient and effective method for specifying write protection on pass-by-value parameters that are transmitted by access path.

The advantage of pass-by-value is that for scalars it is fast, in both linkage cost and access time.

The main disadvantage of the pass-by-value method if copies are used is that additional storage is required for the formal parameter, either in the called subprogram or in some area outside both the caller and the called subprogram. In addition, the actual parameter must be copied to the storage area for the corresponding formal parameter. The storage and the copy operations can be costly if the parameter is large, such as an array with many elements.

9.5.2.2 Pass-by-Result Pass-by-result is an implementation model for out-mode parameters. When a parameter is passed by result, no value is transmitted to the subprogram. The corresponding formal parameter acts as a local variable, but just before

control is transferred back to the caller, its value is transmitted back to the caller’s actual parameter, which obviously must be a variable. (How would the caller reference the computed result if it were a literal or an expression?)

The pass-by-result method has the advantages and disadvantages of pass-by- value, plus some additional disadvantages. If values are returned by copy (as opposed to access paths), as they typically are, pass-by-result also requires the extra storage and the copy operations that are required by pass-by-value. As with pass-by-value, the difficulty of implementing pass-by-result by transmitting an access path usually results in it being implemented by copy. In this case, the problem is in ensuring that the initial value of the actual parameter is not used in the called subprogram.

One additional problem with the pass-by-result model is that there can be an actual parameter collision, such as the one created with the call

sub(p1, p1)

In sub, assuming the two formal parameters have different names, the two can obviously be assigned different values. Then, whichever of the two is copied to their corresponding actual parameter last becomes the value of p1 in the caller. Thus, the order in which the actual parameters are copied determines their value. For example, consider the following C# method, which specifies the pass-by-result method with the out specifier on its formal parameter.5

5. The out specifier must also be specified on the corresponding actual parameter.

void Fixer(out int x, out int y) {

x = 17;

y = 35;

}

. . .

f.Fixer(out a, out a);

If, at the end of the execution of Fixer, the formal parameter x is assigned to its corresponding actual parameter first, then the value of the actual parameter a in the caller will be 35. If y is assigned first, then the value of the actual parameter a in the caller will be 17.

Because the order can be implementation dependent for some languages, different implementations can produce different results.

Calling a subprogram with two identical actual parameters can also lead to different kinds of problems when other parameter-passing methods are used, as discussed in Section 9.5.2.4.

Another problem that can occur with pass-by-result is that the implementor may be able to choose between two different times to evaluate the addresses of the actual parameters: at the time of the call or at the time of the return. For example, consider the following C# method and following code:

void DoIt(out int x, int index){

x = 17;

index = 42;

}

. . .

sub = 21;

f.DoIt(list[sub], sub);

The address of list[sub] changes between the beginning and end of the method. The implementor must choose the time to bind this parameter to an address—at the time of the call or at the time of the return. If the address is computed on entry to the method, the value 17 will be returned to list[21]; if computed just before return, 17 will be returned to list[42]. This makes programs unportable between an implementation that chooses to evaluate the addresses for out-mode parameters at the beginning of a subprogram and one that chooses to do that evaluation at the end. An obvious way to avoid this problem is for the language designer to specify when the address to be used to return the parameter value must be computed.

9.5.2.3 Pass-by-Value-Result Pass-by-value-result is an implementation model for inout-mode parameters in which actual values are copied. It is in effect a combination of pass-by- value and pass-by-result. The value of the actual parameter is used to initialize the corresponding formal parameter, which then acts as a local

variable. In fact, pass-by-value-result formal parameters must have local storage associated with the called subprogram. At subprogram termination, the value of the formal parameter is transmitted back to the actual parameter.

Pass-by-value-result is sometimes called pass-by-copy, because the actual parameter is copied to the formal parameter at subprogram entry and then copied back at subprogram termination.

Pass-by-value-result shares with pass-by-value and pass-by-result the ​- disadvantages of requiring multiple storage for parameters and time for copying values. It shares with pass-by-result the problems associated with the order in which actual parameters are assigned.

The advantages of pass-by-value-result are relative to pass-by-reference, so they are discussed in Section 9.5.2.4.

9.5.2.4 Pass-by-Reference Pass-by-reference is a second implementation model for inout-mode parameters. Rather than copying data values back and forth, however, as in pass-by-value-result, the pass-by-reference method transmits an access path, usually just an address, to the called subprogram. This provides the access path to the cell storing the actual parameter. Thus, the called subprogram is allowed to access the actual parameter in the calling program unit. In effect, the actual parameter is shared with the called subprogram.

The advantage of pass-by-reference is that the passing process itself is efficient, in terms of both time and space. Duplicate space is not required and no copying is required.

There are, however, several disadvantages to the pass-by-reference method. First, access to the formal parameters will be slower than pass-by-value parameters, because of the additional level of indirect addressing that is required.6 Second, if only one-way communication to the called subprogram is required, inadvertent and erroneous changes may be made to the actual parameter. This issue is addressed below.

6. This is further explained in Section 9.5.3.

Another problem of pass-by-reference is that aliases can be created. This problem should be expected, because pass-by-reference makes access paths available to the called subprograms, thereby providing access to nonlocal variables. The problem with these kinds of aliasing is the same as in other circumstances: It is harmful to readability and thus to reliability. It also makes program verification more difficult. Another issue with pass by reference is whether the called subprogram is allowed to change a passed pointer. In C, this is ​possible, but in some other languages, such as Pascal and C++, formal parameters that are addresses are implicitly dereferenced in the called subprogram, which prevents such changes.

There are several ways pass-by-reference parameters can create aliases. First, collisions can occur between actual parameters. Consider a C++ function that has two parameters that are to be passed by reference, as in

void fun(int &first, int &second)

If the call to fun happens to pass the same variable twice, as in

fun(total, total)

then first and second in fun will be aliases.

Second, collisions between array elements can also cause aliases. For example, suppose the function fun is called with two array elements that are ​- specified with variable subscripts, as in

fun(list[i], list[j])

If these two parameters are passed by reference and i happens to be equal to j, then first and second are again aliases.

Third, if two of the formal parameters of a subprogram are an element of an array and the whole array, and both are passed by reference, then a call such as

fun1(list[i], list)

could result in aliasing in fun1, because fun1 can access all elements of list through the second parameter and access a single element through its first parameter.

Still another way to get aliasing with pass-by-reference parameters is through collisions between formal parameters and nonlocal variables that are visible. For example, consider the following C code:

int * global;

void main() {

. . .

sub(global);

. . .

}

void sub(int * param) {

. . .

}

Inside sub, param and global are aliases.

All these possible aliasing situations are eliminated if pass-by-value-result is used instead of pass-by-reference. However, in place of aliasing, other problems sometimes arise, as discussed in Section 9.5.2.3.

9.5.2.5 Pass-by-Name Pass-by-name is an inout-mode parameter transmission method that does not correspond to a single implementation model. When parameters are passed by name, the actual parameter is, in effect, textually substituted for the corresponding formal parameter in all its occurrences in the subprogram. This method is quite different from those discussed thus far; in which case, formal parameters are bound to actual values or addresses at the time of the subprogram call. A pass-by-name formal parameter is bound to an access method at the time of the subprogram call, but the actual binding to a value or an address is delayed until the formal parameter is assigned or referenced. Implementing a pass-by-name parameter requires a subprogram to be passed to the called subprogram to evaluate the address or value of the formal parameter. The referencing environment of the passed subprogram must also

be passed. This subprogram/referencing environment is a closure (see Section 9.12).7 Pass-by-name parameters are both complex to implement and inefficient. They also add significant complexity to the program, thereby lowering its readability and reliability.

7. These closures were originally (in ALGOL 60) called thunks.

Because pass-by-name is not part of any widely used language, it is not discussed further here. However, it is used at compile time by the macros in assembly languages and for the generic parameters of the generic subprograms in C++, Java 5.0, and C# 2005, as discussed in Section 9.9.

9.5.3 Implementing Parameter- Passing Methods We now address the question of how the various implementation models of parameter passing are actually implemented.

In most contemporary languages, parameter communication takes place through the run-time stack. The run-time stack is initialized and maintained by the run-time system, which manages the execution of programs. The run- time stack is used extensively for subprogram control linkage and parameter passing, as discussed in Chapter 10. In the following discussion, we assume that the stack is used for all parameter transmission.

Pass-by-value parameters have their values copied into stack locations. The stack locations then serve as storage for the corresponding formal parameters. Pass-by-result parameters are implemented as the opposite of pass-by-value. The values assigned to the pass-by-result actual parameters are placed in the stack, where they can be retrieved by the calling program unit upon termination of the called subprogram. Pass-by-value-result parameters can be implemented directly from their semantics as a combination of pass-by-value and pass-by-result. The stack location for such a parameter is initialized by the call and is then used like a local variable in the called subprogram.

Pass-by-reference parameters are perhaps the simplest to implement. Most languages only allow variables to be passed by reference. However, Fortran passes all forms of parameters by reference. In Fortran, regardless of the type of the actual parameter, only its address must be placed in the stack. In the case of literals, the address of the literal is put in the stack. In the case of an expression, the compiler must build code to evaluate the expression, which must be executed just before the transfer of control to the called subprogram. The address of the memory cell in which the code places the result of its evaluation is then put in the stack. The Fortran compiler must prevent the called subprogram from changing parameters that are literals or expressions.

Access to the formal parameters in the called subprogram is by indirect addressing from the stack location of the address. The implementation of pass-by-value, -result, -value-result, and -reference, where the run-time stack is used, is shown in Figure 9.2. Subprogram sub is called from main with the call sub(w, x, y, z), where w is passed by value, x is passed by result, y is passed by value-result, and z is passed by reference.

Figure 9.2 One possible stack

implementation of the common parameter-passing methods

Figure 9.2 Full Alternative Text

Function header: void sub (int a, int b, int c, int d)

Function call in main: sub (w,x,y,z)

(pass w by value, x by result, y by value-result, z by reference)

9.5.4 Parameter-Passing Methods of Some Common Languages C uses pass-by-value. Pass-by-reference (inout mode) semantics is achieved by using pointers as parameters. The value of the pointer is made available to the called function and nothing is copied back. However, because what was passed is an access path to the data of the caller, the called function can change the caller’s data. But all references to pointer formal parameter must be explicitly dereferenced in the function. C copied this use of the pass-by- value method from ALGOL 68. In both C and C++, formal parameters can be typed as pointers to constants. The corresponding actual parameters need not be constants, for in such cases they are coerced to constants. This allows pointer parameters to provide the efficiency of pass-by-reference with the one-way semantics of pass-by-value. Write protection of those parameters in the called function is implicitly specified.

C++ includes a special pointer type, called a reference type, as discussed in Chapter 6, which is often used for parameters. Reference parameters are implicitly dereferenced in the function or method, and their semantics is pass- by-reference. C++ also allows reference parameters to be defined to be constants. For example, we could have

void fun(const int &p1, int p2, int &p3) { . . . }

where p1 is pass-by-reference but cannot be changed in the function fun, p2 is pass-by-value, and p3 is pass-by-reference. ​Neither p1 nor p3 need be explicitly dereferenced in fun.

history note ALGOL 60 introduced the pass-by-name method. It also allowed pass-by- value as an option. Primarily because of the difficulty in implementing them, pass-by-name parameters were not carried from ALGOL 60 to any subsequent languages that became popular (other than SIMULA 67).

Constant parameters and in-mode parameters are not exactly alike. Constant parameters clearly implement in mode. However, in all of the common imperative languages except Ada, in-mode parameters can be assigned in the subprogram even though those changes are never reflected in the values of the corresponding actual parameters. In Ada, such an assignment is illegal. Constant parameters can never be assigned.

As with C and C++, all Java parameters are passed by value. ​However, because objects can be accessed only through reference variables, object parameters are in effect passed by reference. Although an object reference passed as a parameter cannot itself be changed in the called subprogram, the referenced object can be changed if a method is available to cause the change. Because reference variables cannot point to scalar variables directly and Java does not have pointers, scalars cannot be passed by reference in Java (although a reference to an object that contains a scalar can). Therefore, if a scalar is passed to a Java method, it cannot be changed by that method.

history note ALGOL W (Wirth and Hoare, 1966) introduced the pass-by-value-result method of parameter passing as an alternative to the inefficiency of pass-by- name and the problems of pass-by-reference.

The default parameter-passing method of C# is pass-by-value. Pass-by- reference can be specified by preceding both a formal parameter and its corresponding actual parameter with ref. For example, consider the following C# skeletal method and call:

void sumer(ref int oldSum, int newOne) { . . . }

. . .

sumer(ref sum, newValue);

The first parameter to sumer is passed by reference; the second is passed by value. All ref parameters must be assigned a value before they are passed to an actual parameter.

C# also supports out-mode parameters, which are pass-by-reference parameters that do not need initial values. Such parameters are specified in the formal parameter list with the out modifier.

PHP’s parameter passing is similar to that of C#, except that either the actual parameter or the formal parameter can specify pass-by-reference. Pass-by- reference is specified by preceding one or both of the parameters with an ampersand.

In Swift, the default parameter passing method is pass by value, and formal parameters passed this way cannot be changed in the called subprogram. Pass-by-reference semantics can be specified by preceding the formal parameter with the reserved word inout.

Perl employs a primitive means of passing parameters. All actual parameters are implicitly placed in a predefined array named @_ (of all things!). The subprogram retrieves the actual parameter values (or addresses) from this array. The most peculiar thing about this array is its magical nature, exposed by the fact that its elements are in effect aliases for the actual parameters. Therefore, if an element of @_ is changed in the called subprogram, that change is reflected in the corresponding actual parameter in the call, assuming there is a corresponding actual parameter (the number of actual parameters need not be the same as the number of formal parameters) and it is a variable.

The parameter-passing method of Python and Ruby is called pass-by-​-

assignment or pass-by-sharing. Because all data values are objects, every variable is a reference to an object. In pass-by-assignment, the actual parameter value is assigned to the formal parameter. Therefore, pass-by- assignment is in effect pass-by-reference, because the value of all actual parameters are references. However, only in certain cases does this result in pass-by-reference semantics. For example, many objects are essentially immutable. In a pure object-oriented language, the process of changing the value of a variable with an assignment statement, as in

x = x + 1

does not change the object referenced by x. Rather, it takes the object referenced by x, increments it by 1, thereby creating a new object (with the value x + 1), and then changes x to reference the new object. So, when a reference to a scalar object is passed to a subprogram, the object being referenced cannot be changed in place. Because the reference is passed by value, even though the formal parameter is changed in the subprogram, that change has no effect on the actual parameter in the caller.

Now, suppose a reference to an array is passed as a parameter. If the corresponding formal parameter is assigned a new array object, there is no effect on the caller. However, if the formal parameter is used to assign a value to an element of the array, as in

list[3] = 47

the actual parameter is affected. So, changing the reference of the formal parameter has no effect on the caller, but changing an element of the array that is passed as a parameter does.

9.5.5 Type Checking Parameters It is now widely accepted that software reliability demands that the types of actual parameters be checked for consistency with the types of the corresponding formal parameters. Without such type checking, small typographical errors can lead to program errors that may be difficult to diagnose because they are not detected by the compiler or the run-time

system. For example, in the function call

result = sub1(1)

the actual parameter is an integer constant. If the formal parameter of sub1 is a floating-point type, no error will be detected without parameter type checking. Although an integer 1 and a floating-point 1 have the same value, the representations of these two are very different. sub1 cannot produce a correct result given an integer actual parameter value if it expects a floating- point value.

Early programming languages, such as Fortran 77 and the original version of C, did not require parameter type checking; most later languages require it. However, the relatively recent languages Perl, JavaScript, and PHP do not.

C and C++ require some special discussion in the matter of parameter type checking. In the original C, neither the number of parameters nor their types were checked. In C89, the formal parameters of functions can be defined in two ways. They can be defined as in the original C; that is, the names of the parameters are listed in parentheses and the type declarations for them follow, as in the following function:

double sin(x)

double x;

{ . . . }

Using this form avoids type checking, thereby allowing calls such as

double value;

int count;

. . .

value = sin(count);

to be legal, although they are never correct.

The alternative to the original C definition approach is called the prototype method, in which the formal parameter types are included in the list, as in

double sin(double x)

{ . . . }

If this version of sin is called with the same call, that is, with the following, it is also legal:

value = sin(count);

The type of the actual parameter (int) is checked against that of the formal parameter (double). Although they do not match, int is coercible to double (it is a widening coercion), so the conversion is done. If the conversion is not possible (for example, if the actual parameter had been an array) or if the number of parameters is wrong, then a semantics error is detected. So in C89, the user chooses whether parameters are to be type checked.

In C99 and C++, all functions must have their formal parameters in prototype form. However, type checking can be avoided for some of the parameters by replacing the last part of the parameter list with an ellipsis, as in

int printf(const char* format_string, . . .);

A call to printf must include at least one parameter, a pointer to a literal character string. Beyond that, anything (including nothing) is legal. The way printf determines whether there are additional parameters is by the presence of format codes in the string parameter. For example, the format code for integer output is %d. This appears as part of the string, as in the following:

printf("The sum is %d\n", sum);

The % tells the printf function that there is one more parameter.

There is one more interesting issue with actual to formal parameter coercions when primitives can be passed by reference, as in C#. Suppose a call to a method passes a float value to a double formal parameter. If this parameter is passed by value, the float value is coerced to double and there is no problem. This particular coercion is very useful, for it allows a library to provide double versions of subprograms that can be used for both float and double values. However, suppose the parameter is passed by reference. When the value of the double formal parameter is returned to the float actual parameter in the caller, the value will overflow its location. To avoid this problem, C# requires the type of a ref actual parameter to match exactly the type of its corresponding formal parameter (no coercion is allowed).

In Python and Ruby, there is no type checking of parameters, because typing in these languages is a different concept. Objects have types, but variables do not, so formal parameters are typeless. This disallows the very idea of type checking parameters.

9.5.6 Multidimensional Arrays as Parameters The storage-mapping functions that are used to map the index values of references to elements of multidimensional arrays to addresses in memory were discussed at length in Chapter 6. In some languages, such as C and C++, when a multidimensional array is passed as a parameter to a subprogram, the compiler must be able to build the mapping function for that array while seeing only the text of the subprogram (not the calling subprogram). This is true because the subprograms can be compiled separately from the programs that call them. Consider the problem of passing a matrix to a function in C. Multidimensional arrays in C are really arrays of arrays, and they are stored in row major order. Following is a storage-mapping function for row major order for matrices when the lower bound of all indices is 0 and the element size is 1:

address (mat[i, j]) = address(mat[0,0]) + i *

           number_of_columns + j

Notice that this mapping function needs the number of columns but not the number of rows. Therefore, in C and C++, when a matrix is passed as a parameter, the formal parameter must include the number of columns in the second pair of brackets. This is illustrated in the following skeletal C program:

void fun(int matrix[][10]) {

. . . }

void main() {

int mat[5][10];

. . .

fun(mat);

. . .

}

The problem with this method of passing matrices as parameters is that it does not allow a programmer to write a function that can accept matrices with different numbers of columns; a new function must be written for every matrix with a different number of columns. This, in effect, disallows writing flexible functions that may be effectively reusable if the functions deal with multidimensional arrays. In C and C++, there is a way around the problem because of their inclusion of pointer arithmetic. The matrix can be passed as a pointer, and the actual dimensions of the matrix also can be passed as parameters. Then, the function can evaluate the user-written storage-mapping function using pointer arithmetic each time an element of the matrix must be referenced. For example, consider the following function prototype:

void fun(float *mat_ptr,

int num_rows,

int num_cols);

The following statement can be used to move the value of the variable x to the [row][col] element of the parameter matrix in fun:

*(mat_ptr + (row * num_cols) + col) = x;

Although this works, it is obviously difficult to read, and because of its complexity, it is error prone. The difficulty with reading this can be alleviated by using a macro to define the storage-mapping function, such as

#define mat_ptr(r,c) (*mat_ptr + ((r) *

(num_cols) + (c)))

With this, the assignment can be written as

mat_ptr(row,col) = x;

Other languages use different approaches to dealing with the problem of passing multidimensional arrays.

In Java and C#, arrays are objects. They are all single dimensioned, but the elements can be arrays. Each array inherits a named constant (length in Java

and Length in C#) that is set to the length of the array when the array object is created. The formal parameter for a matrix appears with two sets of empty brackets, as in the following Java method:

float sumer(float mat[][]) {

float sum = 0.0f;

for (int row = 0; row < mat.length; row++) {

for (int col = 0; col < mat[row].length; col++) {

sum += mat[row][col];

} //** for (int row . . .

} //** for (int col . . .

return sum;

}

Because each array has its own length value, in a matrix the rows can have different lengths.

9.5.7 Design Considerations Two important considerations are involved in choosing parameter-passing methods: efficiency and whether one-way or two-way data transfer is needed.

Contemporary software-engineering principles dictate that access by subprogram code to data outside the subprogram should be minimized. With this goal in mind, in-mode parameters should be used whenever no data are to be returned through parameters to the caller. Out-mode parameters should be used when no data are transferred to the called subprogram but the subprogram must transmit data back to the caller. Finally, inout-mode parameters should be used only when data must move in both directions between the caller and the called subprogram.

There is a practical consideration that is in conflict with this principle. Sometimes it is justifiable to pass access paths for one-way parameter transmission. For example, when a large array is to be passed to a subprogram that does not modify it, a one-way method may be preferred. However, pass-by-value would require that the entire array be moved to a local storage area of the subprogram. This would be costly in both time and space. Because of this, large arrays are often passed by reference. This is

precisely the reason why the Ada 83 definition allowed implementors to choose between the two methods for structured parameters. C++ constant reference parameters offer another solution. Another alternative approach would be to allow the user to choose between the methods.

The choice of a parameter-passing method for functions is related to another design issue: functional side effects. This issue is discussed in Section 9.10.

9.5.8 Examples of Parameter Passing Consider the following C function:

void swap1(int a, int b) {

int temp = a;

a = b;

b = temp;

}

Suppose this function is called with

swap1(c, d);

Recall that C uses pass-by-value. The actions of swap1 can be described by the following pseudocode:

a = c - Move first parameter value in

b = d - Move second parameter value in

temp = a

a = b

b = temp

Although a ends up with d’s value and b ends up with c’s value, the values of c and d are unchanged because nothing is transmitted back to the caller.

We can modify the C swap function to deal with pointer parameters to achieve the effect of pass-by-reference:

void swap2(int *a, int *b) {

int temp = *a;

*a = *b;

*b = temp;

}

swap2 can be called with

swap2(&c, &d);

The actions of swap2 can be described with the following:

a = &c - Move first parameter address in

b = &d - Move second parameter address in

temp = *a

*a = *b

*b = temp

In this case, the swap operation is successful: The values of c and d are in fact interchanged. swap2 can be written in C++ using reference parameters as follows:

void swap2(int &a, int &b) {

int temp = a;

a = b;

b = temp;

}

This simple swap operation is not possible in Java, because it has neither pointers nor C++’s kind of references. In Java, a reference variable can point to only an object, not a scalar value.

The semantics of pass-by-value-result is identical to those of pass-by-​- reference, except when aliasing is involved. Ada uses pass-by-value-result for inout-mode scalar parameters. To explore pass-by-value-result, consider the following function, swap3, which we assume uses pass-by-value-result parameters. It is written in a syntax similar to that of Ada.

procedure swap3(a : in out Integer, b : in out Integer) is

temp : Integer;

begin

temp := a;

a := b;

b := temp;

end swap3

Suppose swap3 is called with

swap3(c, d);

The actions of swap3 with this call are

addr_c = &c - Move first parameter address in

addr_d = &d - Move second parameter address in

a = *addr_c - Move first parameter value in

b = *addr_d - Move second parameter value in

temp = a

a = b

b = temp

*addr_c = a - Move first parameter value out

*addr_d = b - Move second parameter value out

So once again, this swap subprogram operates correctly. Next, consider the call

swap3(i, list[i]);

In this case, the actions are

addr_i = &i - Move first parameter address in

addr_listi= &list[i] - Move second parameter address in

a = *addr_i - Move first parameter value in

b = *addr_listi - Move second parameter value in

temp = a

a = b

b = temp

*addr_i = a - Move first parameter value out

*addr_listi = b - Move second parameter value out

Again, the subprogram operates correctly, in this case because the addresses to which to return the values of the parameters are computed at the time of the call rather than at the time of the return. If the addresses of the actual parameters were computed at the time of the return, the results would be wrong.

Finally, we must explore what happens when aliasing is involved with pass-

by-value-result and pass-by-reference. Consider the following skeletal program written in C-like syntax:

int i = 3; /* i is a global variable */

void fun(int a, int b) {

i = b;

}

void main() {

int list[10];

list[i] = 5;

fun(i, list[i]);

}

In fun, if pass-by-reference is used, i and a are aliases. If pass-by-value- result is used, i and a are not aliases. The actions of fun, assuming pass-by- value-result, are as follows:

addr_i = &i - Move first parameter address in

addr_listi = &list[i] - Move second parameter address in

a = *addr_i - Move first parameter value in

b = *addr_listi - Move second parameter value in

i = b - Sets i to 5

*addr_i = a - Move first parameter value out

*addr_listi = b - Move second parameter value out

In this case, the assignment to the global i in fun changes its value from 3 to 5, but the copy back of the first formal parameter (the second to last line in the example) sets it back to 3. The important observation here is that if pass- by-​reference is used, the result is that the copy back is not part of the semantics, and i remains 5. Also note that because the address of the second parameter is computed at the beginning of fun, any change to the global i has no effect on the address used at the end to return the value of list[i].

9.6 Parameters That Are Subprograms In programming, a number of situations occur that are most conveniently handled if subprogram names can be sent as parameters to other subprograms. One common example of these occurs when a subprogram must sample some mathematical function. For example, a subprogram that does numerical integration estimates the area under the graph of a function by sampling the function at a number of different points. When such a subprogram is written, it should be usable for any given function; it should not need to be rewritten for every function that must be integrated. It is therefore natural that the name of a program function that evaluates the mathematical function to be integrated be sent to the integrating subprogram as a parameter.

Although the idea is natural and seemingly simple, the details of how it works can be confusing. If only the transmission of the subprogram code was necessary, it could be done by passing a single pointer. However, two complications arise.

First, there is the matter of type checking the parameters of the activations of the subprogram that was passed as a parameter. In C and C++, functions cannot be passed as parameters, but pointers to functions can. The type of a pointer to a function includes the function’s protocol. Because the protocol includes all parameter types, such parameters can be completely type checked.

The second complication with parameters that are subprograms appears only with languages that allow nested subprograms. The issue is what referencing environment for executing the passed subprogram should be used. There are three choices:

The environment of the call statement that enacts the passed subprogram (shallow binding)

The environment of the definition of the passed subprogram (deep binding)

The environment of the call statement that passed the subprogram as an actual parameter (ad hoc binding)

The following example program, written with the syntax of JavaScript, illustrates these choices:

function sub1() {

var x;

function sub2() {

alert(x); // Creates a dialog box with the value of x

};

function sub3() {

var x;

x = 3;

sub4(sub2);

};

function sub4(subx) {

var x;

x = 4;

subx();

};

x = 1;

sub3();

};

Consider the execution of sub2 when it is called in sub4. For shallow binding, the referencing environment of that execution is that of sub4, so the reference to x in sub2 is bound to the local x in sub4, and the output of the program is 4. For deep binding, the referencing environment of sub2’s execution is that of sub1, so the reference to x in sub2 is bound to the local x in sub1, and the output is 1. For ad hoc binding, the binding is to the local x in sub3, and the output is 3.

In some cases, the subprogram that declares a subprogram also passes that subprogram as a parameter. In those cases, deep binding and ad hoc binding are the same. Ad hoc binding has never been used because, one might surmise, the environment in which the procedure appears as a parameter has no natural connection to the passed subprogram.