c++book.

mif Page v Saturday, July 12, 2003 10:53 AM

Contents

Preface

xv

Chapter 0 Introduction
0.1 A History Lesson
0.2 High Level Differences
0.2.1 Compiled vs. Interpreted Code
0.2.2 Security and Robustness
0.2.3 Multithreading
0.2.4 API Differences
0.3 Ten Reasons To Use C++
0.3.1 C++ Is Still Widely Used
0.3.2 Templates
0.3.3 Operator Overloading
0.3.4 Standard Template Library
0.3.5 Automatic Reclamation of Resources
0.3.6 Conditional Compilation
0.3.7 Distinctions Between Accessor and Mutator
0.3.8 Multiple Implementation Inheritance
0.3.9 Space Efficiency
0.3.10 Private Inheritance
0.4 Key Points
0.5 Exercises

1
1
4
4
4
5
5
6
6
6
6
6
6
8
8
8
8
8
8
9

c++book.mif Page vi Saturday, July 12, 2003 10:53 AM

vi

Contents

Chapter 1 Basic Types and Control Structures

1.1 First Program
1.1.1 main Function
1.1.2 The Preprocessor
1.1.3 The using directive
1.1.4 Output
1.2 Primitive Types
1.2.1 Integer Types
1.2.2 Floating Point Types
1.2.3 Character Type
1.2.4 Boolean Type
1.3 Syntactic Differences
1.3.1 Operators and Expressions
1.3.2 Conditionals
1.3.3 Loops
1.3.4 Definite Assignment
1.4 Additional Syntax
1.4.1 Primitive Type Casts
1.4.2 Labels
1.4.3 typedef Statement
1.5 Key Points
1.6 Exercises

11
11
11
12
13
13
13
13
14
14
14
15
15
15
16
16
16
16
17
17
18
19

Chapter 2 Functions, Arrays, Strings, and Parameter Passing

2.1 Functions
2.1.1 Function definition
2.1.2 Function Invocation
2.1.3 Function Overloading
2.1.4 Function Declarations
2.1.5 Default Parameters
2.1.6 Inline Functions
2.1.7 Separate Compilation
2.2 Arrays and Strings
2.2.1 Using the vector Library Type
2.2.2 Using the string Library Type
2.2.3 Arrays of Objects
2.2.4 Primitive Arrays of Constants

21
21
21
21
22
22
24
24
25
26
27
28
29
30

c++book.mif Page vii Saturday, July 12, 2003 10:53 AM

Contents

2.2.5 Multidimensional Arrays

2.3 Parameter Passing
2.3.1 Call by Reference
2.3.2 Call by Constant Reference
2.3.3 Parameter Passing Summary
2.4 Key Points
2.5 Exercises

vii

30
31
31
33
34
35
35

Chapter 3 Pointers and Reference Variables

3.1 Java vs. C++ Memory Model
3.2 Pointers
3.3 Memory Heap Management
3.3.1 The new Operator
3.3.2 Garbage collection and delete operator
3.3.3 Stale pointers
3.3.4 Double-delete
3.3.5 Functions that Return Pointers
3.4 The -> Operator
3.5 Reference Variables
3.6 Uses and Abuses of Pointer Variables
3.6.1 Parameter Passing: Simulating Call-By-Reference
3.6.2 Arrays and Strings
3.6.3 Avoiding Large Data Movements
3.6.4 Linked Data Structures
3.6.5 Inheritance
3.7 Key Points
3.8 Exercises

37
37
38
43
43
43
45
45
46
48
48
49
49
49
50
50
50
50
51

Chapter 4 Object-Based Programming: Classes

4.1 Similarities, Modulo Syntax
4.2 Accessors vs. Mutators
4.3 The explicit Keyword
4.4 Why C++ Needs More Syntax Than Java
4.5 Initializer Lists
4.6 The Big Three: Destructor, Copy Constructor, and operator=
4.6.1 Destructor
4.6.2 Copy Constructor
4.6.3 operator=

53
54
55
56
58
61
64
64
64
64

c++book.mif Page viii Saturday, July 12, 2003 10:53 AM

viii

Contents

4.6.4 Issues Concerning the Defaults

65
4.6.5 When the Defaults Do Not Work
66
4.6.6 Linked Data Structures
69
4.6.7 Default Constructor
74
4.6.8 Disabling Copying
74
4.7 Friends
74
4.7.1 Classes as friends
75
4.7.2 Methods as friends
75
4.8 Nested Classes
76
4.9 The struct Type
77
4.10 Return By Constant Reference Revisited
77
4.11 Separation of Interface and Implementation
79
4.12 #ifndef and #endif
81
4.13 Static Members
82
4.14 Anonymous Objects
85
4.15 Namespaces
86
4.16 Incomplete Class Declaration
87
4.17 Technical Differences
88
4.17.1 Member Functions and Data Members Cannot Have Same Name
88
4.17.2 Inline Data Member Initialization
88
4.17.3 Default Data Member Initialization
88
4.17.4 Default Visibility
88
4.18 Key Points
88
4.19 Exercises
89
Chapter 5 Operator Overloading
5.1 Basics of Operator Overloading
5.2 Overloading I/O
5.3 Global vs. Class Operators
5.4 What Can Be Overloaded?
5.5 A Rational Number Class
5.5.1 Assignment Operators
5.5.2 Arithmetic Operators
5.5.3 Relational and Equality Operators
5.5.4 Input and Output Operators
5.5.5 Unary Operators
5.6 Matrix Class (for doubles)

91
91
93
94
97
97
103
103
105
105
106
107

c++book.mif Page ix Saturday, July 12, 2003 10:53 AM

Contents

5.7 Liabilities of Operator Overloading

5.8 Key Points
5.9 Exercises

ix

110
110
111

Chapter 6 Object-Oriented Programming: Inheritance

6.1 Basic Syntax
6.2 Dynamic Dispatch
6.3 Constructors, the Big Three and Inheritance
6.3.1 Defaults
6.3.2 Implementing the Big-Three Explicitly
6.3.3 Virtual Destructor
6.4 Abstract Methods and Classes
6.5 Slicing
6.6 Why Inheritance Is Harder In C++ Than Java
6.7 Type Conversions
6.8 Multiple Inheritance and Interfaces
6.9 Protected and Friends
6.10 Technical Differences
6.10.1 No Root Class
6.10.2 Non-Public Inheritance
6.10.3 Reducing Visibility When Overriding
6.10.4 New Methods Hide Originals in Base Classes
6.10.5 Overriding With Different Return Type
6.10.6 Type Compatibility
6.10.7 Reflection
6.10.8 Dynamic Dispatch in Constructors
6.10.9 Default Parameters
6.11 Key Points
6.12 Exercises

113
113
115
116
116
117
118
119
121
122
125
126
128
130
130
130
131
131
132
132
133
133
133
133
134

Chapter 7 Templates
7.1 Review of Generic Programming in Java
7.2 Function Templates
7.3 Class Templates
7.4 Separate Compilation
7.4.1 Everything in the Header
7.4.2 Explicit Instantiation
7.4.3 The export Directive

137
137
138
141
144
145
146
147

c++book.mif Page x Saturday, July 12, 2003 10:53 AM

Contents

7.5 Specialized Templates

7.5.1 Multiple Template Parameters
7.5.2 Template Nontype Parameters
7.5.3 Default Template Parameters
7.5.4 Member Templates
7.6 Templates for Function Objects
7.6.1 The Java Solution
7.6.2 Avoiding Inheritance
7.6.3 Function Objects With operator()
7.7 Key Points
7.8 Exercises

147
147
147
148
148
150
150
152
152
153
154

Chapter 8 Abnormal Control Flow

8.1 The Design of C++
8.2 Non-Exception Error Handling
8.2.1 abort and exit
8.2.2 Assertions
8.2.3 Error states
8.3 Exception Handling
8.3.1 Basic syntax
8.3.2 No finally clause
8.3.3 The Throw List and Unexpected Exceptions
8.3.4 Missing Throw List
8.3.5 What Can Be Thrown
8.3.6 Standard Exception Classes
8.3.7 Rethrowing Exceptions
8.3.8 Catching All Exceptions
8.3.9 Inheritance and Exceptions
8.3.10 Templates and Exceptions
8.4 The auto_ptr Class
8.5 Key Points
8.6 Exercises

155
155
158
158
159
160
160
160
160
160
161
161
162
162
162
162
164
164
165
166

Chapter 9 Input and Output

9.1 The Basic Hierarchy
9.2 Error States
9.3 Output
9.4 Input

167
167
169
171
174

c++book.mif Page xi Saturday, July 12, 2003 10:53 AM

Contents

9.5
9.6
9.7
9.8
9.9
9.10
9.11

xi

Files
Random Access
String Streams
endl

Serialization
Key Points
Exercises

Chapter 10 Collections: The Standard Template Library

10.1 Containers and Iterators
10.1.1 Containers
10.1.2 Iterators
10.1.3 Pairs
10.1.4 Inheritance and Containers
10.1.5 Constructors
10.2 Sequence Containers: vector and list
10.3 Error Checks in the STL
10.4 Other Iterators
10.4.1 Reverse Iterators
10.4.2 Stream Iterators
10.5 Stacks and Queues
10.6 Sets
10.6.1 Standard Function Objects
10.6.2 set Operations
10.6.3 multisets
10.6.4 Using Function Objects to Change Default Ordering
10.7 Maps
10.7.1 Multimaps
10.8 STL Example
10.9 Priority Queue
10.10 Generic Algorithms
10.10.1 Sorting
10.10.2 Searching
10.10.3 Unary Binder Adapters
10.10.4 Copying
10.10.5 Inserter Adapters
10.11 BitSets
10.12 Key Points

175
176
177
179
180
180
181
183
183
183
184
186
187
188
188
191
191
191
192
193
193
194
194
195
195
196
197
197
198
200
200
200
202
202
203
203
204

c++book.mif Page xii Saturday, July 12, 2003 10:53 AM

xii

Contents

10.13 Exercises

205

Chapter 11 Primitive Arrays and Strings

11.1 Primitive Arrays
11.1.1 The C++ Implementation: An Array Name Is a Pointer
11.1.2 Dynamic Allocation of Arrays
11.2 Primitive Strings
11.3 Pointer Hopping
11.3.1 Implications of the Precedence of *, &, and []
11.3.2 What Pointer Arithmetic Means
11.3.3 A Pointer-Hopping Example
11.3.4 Is Pointer Hopping Worthwhile?
11.4 Primitive Arrays and the STL
11.5 Command-Line Arguments
11.6 Primitive Multidimensional Arrays
11.7 Key Points
11.8 Exercises

207
207
208
210
215
219
219
220
222
222
223
223
225
227
227

Chapter 12 C-Style C++

12.1 Preprocessor Macros
12.1.1 Simple Textual Substitution
12.1.2 Parameterized Macros
12.2 Unsafe Casts
12.3 C-Style Parameter Passing
12.4 C Library Routines
12.4.1 printf and scanf
12.4.2 File I/O
12.4.3 malloc, calloc, realloc, and free
12.4.4 atoi, atof, strtol, and strtod
12.4.5 system
12.4.6 qsort
12.4.7 Variable Number of Arguments
12.5 C Programming
12.6 Key Points
12.7 Exercises

229
229
230
231
232
233
235
235
237
242
243
243
243
244
245
246
246

Chapter 13 Using Java and C++: The JNI

13.1 JNI Basics

249
249

c++book.mif Page xiii Saturday, July 12, 2003 10:53 AM

Contents

xiii

13.2 Implementing A Simple Parameter-less, Return-less Method

13.2.1 Compiling a Shared Library
13.3 JNI Types
13.4 Accessing Java Objects
13.4.1 Accessing Fields
13.4.2 Invoking Methods
13.4.3 Invoking Constructors
13.5 Strings and Arrays
13.5.1 Strings
13.5.2 Arrays
13.6 Using Java Exceptions
13.7 Using C Instead of C++
13.8 JNI References
13.9 Java Monitors
13.10 Invocation API
13.11 Key Points
13.12 Exercises

252
253
254
255
256
259
260
260
260
262
265
268
268
270
270
272
273

Bibliography

275

Index

277

c++book.mif Page xiv Saturday, July 12, 2003 10:53 AM

xiv

Contents

c++book.mif Page xv Saturday, July 12, 2003 10:53 AM

Preface

OR many years, C++ was the de-facto language of

choice in introductory CS courses, due largely to its support for object-oriented programming, as
well as its wide adoptance in industy. However, because C++ is arguably the most complex language ever to be widely-used, Java which also supports object-oriented programming, has
recently emerged as the preferred introductory language. Nonetheless, demand for C++ skill is
still high in industry, and most universities require C++ programming at some point in the Computer Science curriculum. Although Java and C++ look similar, programming in C++ is somewhat more challenging, and filled with subtle details. While there are many books that
thoroughly describe C++ (see the bibliography), the vast majority exceed 1,000 pages, and for
the most part are written for either experienced industry programmers or novices.
This book is designed as a quick start guide for students who are knowledgeable in an
object-oriented language (most likely Java) who would like to learn C++. Throughout the text,
we compare and contrast Java and C++, and show C++ substitutes for Java equivalents. We do
not describe in detail basic concepts (such as inheritance) that are common in C++; rather we
describe how the concepts are implemented in C++. This helps achieve one of the important
goals of this book, which is to keep the page count reasonably low. Consequently this book is
not appropriate for students with limited or no prior programming experience.

Organization
The book begins with a brief overview of C++ in Chapter 0. In Chapter 1, we describe some of
the basic expressions and statements in C++, which mostly mirrors simple Java syntax. Functions, arrays, strings, and parameter passing are discussed in Chapter 2. We use the modern
alternative of introducing and using the standard vector and string classes in the C++ library,

xv

c++book.mif Page xvi Saturday, July 12, 2003 10:53 AM

xvi

Preface

rather than the older-style built-in array constructs.

Chapter 3 describes pointers and reference variables, paying particular attention to the
host of pitfalls that await a C++ programmer. Chapter 4 is a long chapter that describes the
basics of classes.
Two C++ features that are not part of Java are operator overloading and templates. Operator overloading is discussed in Chapter 5, and templates are discussed in Chapter 7. In between,
we cover inheritance in Chapter 6. In Chapter 8, we examine exceptions in C++, as well as older
library routines such as abort.
The next two chapters discuss some C++ libraries. Chapter 9 deals with I/O, and
Chapter 10 is an introduction to the Standard Template Library, which is the C++ equivalent to
the Collections API. Both libraries are complex enough to require an entire book; we provide the
most important highlights, rather than attempting to list every detail.
Many courses will require C programming in addition to C++ programming. Because C++
builds on C, many C programming techniques are also found in C++, although in C++ they are
not always the preferred way of doing things. Chapter 11 covers primitive arrays and strings. As
we mentioned earlier, modern C++ uses library classes as an alternative. Nonetheless, primitve
arrays and strings will almost certainly be encountered by any C++ programmer, and the techniques are essentially the same in C++and C. Chapter 12 covers C idioms that are occasionally
used in C++, but probably should be avoided. Finally, in Chapter 13, we describe the Java
Native Interface (again a whole book can be written on this one topic), which allows Java programmers to implement their methods in C++ or C.

Acknowledgements
For this text, I would like to thank my editor Alan Apt and his assistants Toni Holm,
Patrick Linder, and Jake Warde.
<<Add acks for copy-edits, cover, production, marketing>>
I also thank the following reviewers, who provided valuable comments, many of which
have been incorporated into the text:
XXXX, University of YYY
XXXX, University of YYY
XXXX, University of YYY
XXXX, University of YYY
Some of the material in this text (especially Chapters 1, 2, 3, 11, and 12) is adapted from
my textbook Efficient C Programming: A Practical Approach (Prentice-Hall, 1995).
My World Wide Web page http://www.cs.fiu.edu/~weiss will contain updated
source code, an errata list, and a link for receiving bug reports.
M.A.W
Miami, Florida
June, 2003

c++book.mif Page 1 Saturday, July 12, 2003 10:53 AM

H A P T E R

Introduction

++ and Java look similar syntactically. Many of the

basic constructs in Java, such as basic expressions, conditionals, and loop constructs are indeed
identical except for minor technical differences. Other Java features, such as support for objectbased programming via classes have a similar look and feel in C++, as does much of syntax
related to inheritance. One could (more or less) mechanically translate Java code into C++ code
without great difficulty, since most Java constructs have at least one corresponding C++ construct. Some have even renamed Java as C++-lite. Yet, as we will see, while there are many similarities, the languages have significant differences.
In this chapter, we begin by discussing the progression of languages from C to C++ to
Java. Then we discuss some of the major high-level differences between C++ and Java.

0.1 A History Lesson

In 1972, Dennis Ritchie designed C and implemented it on a PDP-11. The initial design of C
reflected the fact that it was used to implement an operating system. Until then, all operating
systems were written in an assembly language because compiler technology did not generate
sufficiently efficient code from higher-level languages. C provided constructs, such as pointer
arithmetic, direct memory access, increment operators, bit operations, and hexadecimal constants that mimicked the PDP-11s instruction set, while at the same time keeping the language
small but high level. Consequently, the compiler was able to generate reasonably efficient code.
Eventually, Unix was written mostly in C, with only a few parts requiring assembly language.
In early 2002, a top-of-the-line home computer that sells for less than $2,000 executes
several hundred million instructions per second. Home computers can be purchased with 1

c++book.mif Page 2 Saturday, July 12, 2003 10:53 AM

Chapter 0 Introduction

Gigabyte of main memory, and 120 Gigabytes of hard drive space. By the time you read this, the
computer just described may well be a relic.
On the other hand, a PDP-11/45, which this author actually used as a college undergraduate, sold for well over $10,000 (in 1970 dollars). Our model had 128 Kilobytes of main memory,
and executed only several thousand instructions per second. The PDP-11/45 was a 16 bit
machine, so not surprisingly, the int type was 16 bits. But other PDP models, had 12, or 18, or
even 36 bits. The C compiler on the PDP-11/45 would typically take about 30 seconds to compile a trivial one hundred line program. But since it used 32K of memory to do so, compilations
were queued, like printer jobs, to avoid compiling two programs at once and crashing the system! Our PDP-11/45 supported over 20 simultaneous users, all connected via old-style dumb
terminals. Most of the terminals operated at 110 baud, though there were a few fast ones that
went at 300 baud. At 110 baud, approximately 10 characters per second are transmitted. So
screen editors were not widely used; instead editing was done a line at a time, using the Unix
editor ed (which still exists!). C thus provided rather terse syntax to minimize typing, and more
importantly displaying.
In this environment, it is not surprising that the number one goal of the compiler was to
compile correct programs as fast as possible into code that was as fast as possible. And since the
main users were typically experts, dealing with incorrect programs was left to the programmer,
rather than the compiler (or runtime system). No checks were performed to ensure that a variable was assigned a value prior to use of its value. After all, who wanted to wait any longer for
the program to compile? Arrays were implemented in the same manner as in an assembly language, with no bounds checking to consume precious CPU cycles. A host of programming practices that involved the use of pointers to achieve performance benefits emerged. There was even
a reserved word, register, that was used to suggest to the compiler that a particular local
variable should be stored in a machine register, rather than on the runtime stack, to make the
program run faster, since the compiler would not try to do any flow analysis on its own.
Soon, Unix became popular in the academic community, and with it, the C language grew.
At one point, C was known as the great portable language, suitable for systems use on many
machines. Unix itself was ported to a host of platforms. Eventually a host of software vendors
started producing C compilers, adding their own extra features, but in so doing, made the language less portable, in part because the language specification was vague in places, allow competing interpretations. Also, a host of clever but nonetheless unacceptable and unportable
programming tricks had emerged; it became clear that these tricks should be disallowed. Since
compiler technology had become better and computers had become faster, it also because feasible to add helpful features to the language, in the hopes of enhancing portability, and reducing
the chances of undetected errors. Eventually, in 1989, this resulted in ANSI C, which is now the
gold standard of C that most programmer adhere to. In 1999, C99 was adopted, adding some
new features, but since C99 is not yet widely implemented in compilers, few programmers find
a need to use those features.

c++book.mif Page 3 Saturday, July 12, 2003 10:53 AM

A History Lesson

In parallel with the standardization of C was an effort by Bjarne Stroustrup to create a

more object-oriented approach to using C. Eventually C++ emerged, and then eventually C++
underwent a standardization process.
C++ should be viewed as a completely new language, based largely on C. It includes modern constructs to support object-oriented programming. Several design issues have greatly influenced C++.
First and foremost, C++ was intended to be upward compatible with C, as much as possible. Thus legal C programs should be legal C++ programs, though this is not 100% true. (Obvious exceptions include C programs that use as identifiers any reserved word that was added to
C++, such as class.) As a result, numerous constructs from C that perhaps we could live without are part of C++, as is the entire idea of using pointers to do arrays. Additionally, because
each new reserved word breaks existing code, the C++ community is reluctant to add new
reserved words, preferring instead to overuse existing words, such as const and static. So
whereas Java has throw and throws, in C++ we simply use throw both to throw an exception and to list exceptions.
The second crucial philosophy of C++ was in part political: C++ programs should be as
fast as C programs if they do the same thing. Otherwise, it was thought, nobody would want to
learn a more complicated language, just to write slower code. As part of this philosophy, in C++,
if you do not use a feature, you should not have to pay a cost a runtime. Among other things, this
still means, no runtime bounds checks, no dynamic dispatch by default, and no garbage collection.
The third thing to know about C++ is that it has continually evolved. The first widely-used
version of C++ supported classes and inheritance. Templates, discussed in Chapter 7, were
added later. Exceptions were added even later. At another time came namespaces. The templates
got fancier, with templates inside of templates. This has lead to compilers being perennially
behind the official specification, buggy with new features, and especially bug-prone when several new features are used at the same time, much to the chagrin of programmers and textbook
authors. Fortunately, C++ has more or less stabilized and the latest compilers are quite good at
meeting (and in some cases doing more than required by) the language specification.
Java was developed by Sun Microsystems, and marketed as a modern alternative to C++.
Thus, cosmetically, there are many similarities. However, the designers of Java had (at least)
two main goals. First, they attempted to allow the programmer to write the same set of programs
that C++ allows, using simpler syntax. In cases where C++ has several ways to do the same
thing, Java attempts to pick the idiom most widely endorsed by the C++ community and support
it with a language feature. In cases where C++ features are abused, Java does not allow the feature. Thus some C++ features are not found in Java. Second, they have attempted to make it
harder for incorrect code to run, requiring the compiler to detect many sets of errors when it first
compiles the program, and then the Virtual Machine to throw an exception when bad things happen at runtime. By detecting most occurrences of bad behavior, not only are honest program-

c++book.mif Page 4 Saturday, July 12, 2003 10:53 AM

Chapter 0 Introduction

ming errors avoided, system security against hacking is enhanced, since hacking generally
works by having the system do something it is not designed to do.
The Java designers have several advantages. First, in designing the language, they can
expect the compiler to work hard. We no longer use PDP-11s, and compiler research has
advanced greatly to the point that modern optimizing compilers can do a terrific job of creating
code, without requiring the programmer to resort to various tricks common in the good old days.
Second, Java designers specified most of the language (classes, inheritance, exceptions) at once,
adding only a second minor revision (inner classes), with most changes after Java 1.1 dealing
with the library. In doing so, they were able to have language features that dont clash.
Although Java is the new kid on the block (unless C# becomes the next new kid), and has
lots of nice features, it is certainly not true that Java is better than C++, Nor would we say that
C++ is better than Java. Instead, a modern programmer should be able to use both languages, as
each language has applications that can make it the logical choice.

0.2 High Level Differences

As we just described, Java and C++ have different views of the world. C++s main priority is
getting correct programs to run as fast as it can; incorrect programs are on their own. Javas main
priority is not allowing incorrect programs to run; hopefully correct programs run reasonably
fast, and the language makes it easier to generate correct programs by restricting some bad programming constructs.
0.2.1

Compiled vs. Interpreted Code

Java consists of both a compiler and a Virtual Machine. The compiler generates Java bytecode,
and the bytecode is then interpreted at runtime. As a result, not only is Java code portable, the
compiled bytecode is portable and in theory can be run on any platform.
On the other hand, a C++ compiler generates native code. The C++ language specification
provides no guidance on the specifics of what the native code looks like, so not only is the result
of the compilation not transferable to a different type of computer, parts of a single program that
are compiled on the same computer by different compilers are almost always not compatible
with each other.
Originally, the difference between compiled code and interpreted code meant that C++
was as much as 50 times faster than Java. Recent improvements in the Java compiler, and more
importantly, the Virtual Machine have dramatically closed the gap. In some cases, Java code
may actually be faster than C++, but generally speaking, one would expect that since C++ does
less runtime checks than Java, equally skilled compiler writers should be able to generate somewhat faster C++ code than Java code, assuming similar coding styles.
0.2.2

Security and Robustness

We have already seen that Java is very concerned with not allowing unsafe code to compile, and
with throwing exceptions at the first sign of trouble. C++ is unfortunately, somewhat lax in this

c++book.mif Page 5 Saturday, July 12, 2003 10:53 AM

High Level Differences

regard. C++ suffers several problems that can never occur in pure Java code. Four that stand out
are the following:
First, it is possible in C++ to have a pointer or reference to an object that has been returned
back to the memory heap, which is a sure disaster. This is because standard C++ does not do garbage collection; instead the programmer must manage memory themselves, and programmers
are surprisingly bad at doing so. However, some C++ systems include garbage collection and
add runtime checks to avoid using stale pointers. These systems are quite close to the Java standard of avoiding memory problems.
Second, Standard C++ does not check array indexes, and a common hacker attack is to
find an input routine that reads a string, but doesnt check that there is enough space for a very
very long string. By judiciously passing a huge string, the hacker can overflow the buffer, writing replacement values onto variables that are stored in memory adjacent to the buffer. Although
this could never happen in Java, since the C++ specification does not disallow bounds checks,
there is no reason that a safe C++ system couldnt check array bounds. It would simply be
slower (but safer) than a competitor, and so it is not widely done by default.
Third, old C++ typecasts allow type confusion, in which a type is cast to an unrelated type.
This can never happen in Java, but is allowed in C++.
Fourth, in Java, all variables have a definite assigned value prior to use of their value. This
is because variables that are not local to a method by default are initialized to zero for primitives
and null for references. For local variables, an entire chapter of the Java Language Specification is devoted to definite assignment, whereby the compiler is required to perform a flow analysis and produce an error message if a local variable cannot be proven (under a long set of rules)
to have been definitely assigned to through all flows of the method. In C++, this behavior is not
required. Rather, a program that uses an uninitialized variable is said to be incorrect, but the
compiler is not required to take any particular action. Most C++ compilers will print warning
messages if uninitialized variables are detected. But the program will still compile. A similar
story occurs in the case of having a flow that fails to return a value in a non-void function.
0.2.3

Multithreading

C++ does not support multithreading as part of the language. Instead, one must use a set of
library routines that are native to the particular platform. Although Java supports multithreading,
the Java memory model and threading specification has recently been discovered to be inadequate and is undergoing revision.
0.2.4

API Differences

The Java API is huge. The core library includes, among other things, the Swing package
javax.swing for designing (mostly) portable GUIs, a networking package in java.net,
database connectivity in java.sql, built-in compression, serialization, and other I/O in
java.util.zip and java.io, reflection in java.lang.reflect, and even support for
remote methods, servlets, and XML.

c++book.mif Page 6 Saturday, July 12, 2003 10:53 AM

Chapter 0 Introduction

Standard C++ has a very small API, containing little more than some I/O support, a complex number package, and a Collections API, known as the Standard Template Library (STL).
Of course, compiler vendors augment Standard C++ with huge libraries, but each vendor has
different versions rather than implementing a single standard.

0.3 Ten Reasons To Use C++

The preceding section might leave one to conclude that C++ has little to offer. This is not true;
there are lots of reason why one would want to learn and use C++. In this section we list ten
advantages of using C++, including features found in C++ that are not found in Java. Figure 0-1
summarizes the differences described in both Section 0.2 and this section, as well as a few differences that are described later in this text, but not otherwise mentioned in this chapter.
0.3.1

C++ Is Still Widely Used

Perhaps the most important reason to learn C++ is that it is still very-widely used, and since it is
a more complex and difficult language than C++, it is likely that there will always be a strong
market for knowledgeable C++ programmers.
0.3.2

Templates

Perhaps the most requested missing feature in Java is the equivalent of the C++ template. In
C++, templates allow the writing of generic, type-independent code, such as sorting algorithms
and generic data structures that work for any type. In Java, this is done by using inheritance, but
the downside is that many errors are not detected at compile time, but instead linger until run
time. Further, using templates in C++ seems to lead to faster code than the inheritance-based
alternative in Java. Generics are under consideration for Java 1.5. Template are discussed in
Chapter 7.
0.3.3

Operator Overloading

Java does not allow the user to defined operators for class types. C++ does, and this is known as
operator overloading. We discuss operator overloading in Chapter 5.
0.3.4

Standard Template Library

C++ provides a large library, known as the Standard Template Library (STL) for data structures
and generic algorithms. The STL has several advantages compared to the Java Collections API,
and we discuss these differences in Chapter 10.
0.3.5

Automatic Reclamation of Resources

Although Java provides garbage collection, which is a fantastic feature, it is hard to control the
management of other resources. For instance, in Java, the programmer must remember to close
files and database connections when they are no longer in use, dispose of graphics contexts, and

c++book.mif Page 7 Saturday, July 12, 2003 10:53 AM

Ten Reasons To Use C++

so on. Although many (non-memory) resources are released by object finalization, relying on
object finalization in Java is a poor idea, since objects need not be reclaimed if the garbage collector deems that memory is not low.

Figure 0-1

Differences between Java and C++

Feature

Java

C++

Compiler output

Virtual machine bytecode

native code

Portability

Hi

Moderate

Garbage collection

Yes

No

Checks array indices

Yes

No

Multithreading

Yes

Platform specific

API

Huge

Small

Security checks

Yes

No

Compiler checks

Numerous

Some

Operator overloading

No

Yes

Templates

No

Yes

Multiple Inheritance

Interface only

Yes

Standard GUI library

Yes

No

Data Structures

Collections API

STL

Exception Handling

Yes

Yes, but poorly integrated

Conditional Compilation

No

Yes

Global Functions

No

Yes

Class grouping

Packages

Namespaces

Pointer variables

No

Yes

Class documentation

Javadoc

Not in standard

Reflection

Yes

No

Space efficient

Not really

Yes

c++book.mif Page 8 Saturday, July 12, 2003 10:53 AM

Chapter 0 Introduction

In C++, each class can provide a special method known as the destructor, which will automatically be invoked when an object is no longer active. Careful C++ programmers need not
remember to release non-memory resources, and memory resource can often be released by layering the memory allocations inside of classes. Many Java programmers who have prior C++
experience lament the lack of destructors. We describe destructors in Section 4.6.
0.3.6

Conditional Compilation

In C++, it is possible to write code that is compiled only if certain conditions are met at compiletime. This is achieved by use of the preprocessor and is a useful feature during debugging. Java
has only a limited way of doing this. The preprocessor is described in Section 12.1.
0.3.7

Distinctions Between Accessor and Mutator

In Java, when a reference variable is marked as final, it simply means that the reference variable
cannot change; there is no easy way to signal that the state of the object being referenced cannot
be changed (i.e. that the object is immutable). C++ provides syntax to distinguish between methods that are accessors and mutators, and marking an object as const (the equivalent of final)
would allow only the accessors of the object to be invoked. Many Java programmers who have
prior C++ experience lament the difficulty of enforcing immutability. We describe accessors and
mutators in Section 4.2.
0.3.8

Multiple Implementation Inheritance

Java does not allow multiple implementation inheritance because it is notoriously difficult to use
generally. However, there are always some cases when it is very useful, and C++ does allow
multiple implementation inheritance. We discuss this in Section 6.8.
0.3.9

Space Efficiency

Java programs are notoriously space inefficient. For instance, a String of length 16 uses
roughly 76 bytes under Suns Java 1.3 compiler. C++ programs are often more space efficient
that Java programs.
0.3.10 Private Inheritance
Java supports only public inheritance, via the extends clause. C++ supports private inheritance, which is occasionally useful for changing visible interfaces, and implementing adapter
patterns. Private inheritance is discussed in Section 6.10.2.

0.4 Key Points

C++ is based on C

c++book.mif Page 9 Saturday, July 12, 2003 10:53 AM

Exercises

The most important consideration in the design of C++ is to make correct programs run as
fast as possible.
Compile-time checks in C++ are not as rigid as in Java, but now many compilers perform
some of the same checks as a Java compiler, yield warning messages.
Run time checks in C++ are not as rigid in Java. In C++, bad array indexes, bad type casts,
and bad pointers do not automatically cause a runtime error. Some compilers will do more
than the minimum, and issue runtime errors for you, but this cannot be relied on.
Compiled units are not compatible across different types of machines, nor are the compatible on the same machine, when generated by different compilers.
Although the STL in C++ is excellent, the remainder of the Standard C++ library pales in
comparison to Java. But many non-standard additions are available.

0.5 Exercises
1. Compare and contrast the basic design goals of C++ and Java.
2. What are the different versions of C++?
3. What are the basic differences between compiled and interpreted languages?
4. What is a buffer overflow problem?
5. List features of Java that are not part of Standard C++ and result in C++ being a less safe
language.
6. Describe some features of C++ that make it more attractive than Java.

c++book.mif Page 10 Saturday, July 12, 2003 10:53 AM

10

Chapter 0 Introduction

c++book.mif Page 11 Saturday, July 12, 2003 10:53 AM

H A P T E R

Basic Types and Control

Structures

ANY Java constructs have direct analogs in C++.

In both languages, execution starts at main, basic control flow constructs are mostly identical,
and both languages have similar support for basic object-based concepts of encapsulation and
information hiding, via the class. Object-oriented programming is supported in both languages
via inheritance. Both languages support strings as a library class type, provide primitive array
support, and provide a growable array as part of the standard library. Both languages have
library support for data structures collections. However, there are some organizational differences even in the simplest part of the language.
In this chapter, we look at the basics of compiling the first C++ program, and discuss the
basic types and control structures that are common in both languages.

1.1 First Program

The typical first C++ program is shown in Figure 1-1. By way of comparison, the equivalent
Java code is shown in Figure 1-2. The C++ program should be placed in a file that has an appropriate suffix. C++ does not specify a particular suffix, but most compilers recognize .cpp.
1.1.1

main Function

Like Java, control starts at main. Unlike Java, main is not part of any class; instead, it is a nonclass method. In C++ methods are called member functions, and we will adopt the convention
that a function that is not declared as part of a class is simply a function (with the adjective member conveniently omitted). C++ also allows global variables.

11

c++book.mif Page 12 Saturday, July 12, 2003 10:53 AM

12

Chapter 1 Basic Types and Control Structures

1
2
3
4
5
6
7
8

#include <iostream>
using namespace std;
int main( )
{
cout << "Hello world" << endl;
return 0;
}

Figure 1-1

1
2
3
4
5
6
7

First C++ program

public class FirstProgram

{
public static void main( String [ ] args )
{
System.out.println( "Hello world" );
}
}

Figure 1-2

First Java program

main must always be declared in global scope, it must have return type int, which by
convention will be 0 unless a non-zero error code is to be transmitted back to the invoking process, in a manner that is similar to calling System.exit (we will always return 0). Although
some programmers prefer to use a void return type, the language specification is clear that the
return type of main should be int.
main can take additional parameters for command-line arguments (we will discuss this in
Section 11.5). Because main is in global scope, there can be only one version of main. This
contrasts with Java, which allows one main per class.
1.1.2

The Preprocessor

Line 1 is an include directive that reads the declarations for the standard I/O routines. The
include directive has the effect of having the compiler logically insert the source code taken
from another file. Specifically, the file name that resides in between < and > refers to a system
file. Alternatively, a pair of double quotes can be used to specify a user-defined (non-system)
file. Thus in our example, the entire contents of the file iostream, which resides in a systemdependent location, is substituted for line 1. In Section 2.1.7 we discuss how the include directive is typically used to enable faster compilations.
In C++, lines that begin with the # are preprocessor directives. We will see an important
use of preprocessor directives in Section 4.12.

c++book.mif Page 13 Saturday, July 12, 2003 10:53 AM

Primitive Types

1.1.3

13

The using directive

The declaration at line 2 is a using directive, and is the moral equivalent of an import directive in
Java. Whereas classes in package java.lang in Java are automatically known without requiring the full class name that includes their package, in C++, classes in the equivalent namespace,
std, require a full class name unless the using directive is supplied. Thus this using directive is
the C++ equivalent of
import java.lang.*;

We discuss the using directive and the C++ equivalent of packages in Section 4.15.
1.1.4

Output

Ignoring the return statement, the simple program in Figure 1-1 consists of a single statement.
This statement, shown at line 6 is the output mechanism in C++. Here a constant string is placed
on the standard output stream cout. Simple terminal input is similar to output, with cin
replacing cout, and >> replacing <<. As an example:
int x;
cout << "Enter a value of x: ";
cin >> x;

reads the next series of characters (skipping whitespace), interpreting them as an integer. Of
course, this example fails to discuss handling input errors, which is crucial in any serious application. Input and output is discussed in more detail in Chapter 9.

1.2 Primitive Types

While Java has only eight primitive types, and these types have precise sets of possible values
on all platforms, C++ has a host of primitive types, with unspecified ranges.
1.2.1

Integer Types

The basic integer type in C++, like Java, is int. However, whereas Java specifies the precise
range of int, C++ makes no such guarantees, and an int can be 16, 32, or 64 bits, depending
on the platform. Even for 32 bit ints, the precise range can vary from machine to machine.
The int type can be augmented with the reserved words short or long. Alternatively
short and long can be used by themselves without int. Once again, the language specification makes few guarantees about how large short and long are, except to state that an int is
never shorter than a short, and never longer than a long. Historically, a short has been 16
bits, a long has been 32 bits, and an int has been whatever fits best on the machine (currently
32 bits). A long constant includes an L at the end, as in 1000L (the L is never needed if the constant is larger than what could be stored in an int).
Two modifiers to the int type are signed or unsigned. These modifiers can be
applied to any of the integer types (or char). An unsigned int will never be negative, and
in effect instructs the compiler to interpret what would normally be the sign bit as an extra bit.

c++book.mif Page 14 Saturday, July 12, 2003 10:53 AM

14

Chapter 1 Basic Types and Control Structures

Thus, whereas a short might store values in the range -32768 to 32767, an unsigned
short could store values in the range 0 to 65535. This use of unsigned to double the set of
range of positive integers carries the significant danger that mixing unsigned and signed values
can produce surprising results, so it is best to use unsigned types judiciously. An unsigned constant includes a U at the end, as in 1000UL. By default, the integer types are signed.
C++ does not define an equivalent to the byte data type. Historically, C++ programmers
have used signed char, which in C++ is simply eight bits, for this purpose.
1.2.2

Floating Point Types

Floating point types in C++ are represented with float and double. Like Java, double provides more significant digits than a float and its use is preferred to minimize roundoff errors.
Java specifies constants for not a number and both positive and negative infinity. Though the
C++ standard does not require it, some C++ implementations will be able to store and print such
values when they occur.
1.2.3

Character Type

The char type, as in Java, is used to store characters. However, unlike Java, where a character
is a 16-bit Unicode, the C++ standard does not specify a size, and C++ compilers generally use
only 8 bits, and store ASCII characters. Additionally, whether a char is signed or unsigned by
default is unspecified.
As character constant is represented in a pair of single quotes, as is done in Java. Additionally, there is a rich set of escape sequences that can be used to express unprintable characters,
quotes, backslashes, and so forth. The Java escape sequences are part of C++, as well some addition sequences including \a for the bell, \? for the question mark, and \v for the vertical tab.
The escape sequence that allows up to three octal characters is also part of C++. The escape
sequence that uses four hexadecimal characters is not part of Standard C++.
The standard header file cctype contains routines such isupper, islower, and
isalpha for determining if a character has a certain property. Also present are toupper and
tolower that return upper and lower-case versions of a character, or the character itself, if it is
not a lower case letter. In effect these routines are versions of the static methods in
java.lang.Character.
Recent C++ compilers have adopted the Java style of supporting Unicode characters with
the addition of the wchar_t type and the wide-character literal. Wide character literals begin
with an L prior to the opening quote, as in Lx or L"hello". As one may expect by now, the
specific behavior of the wchar_t type is implementation dependent.
1.2.4

Boolean Type

Like Java, C++ has a Boolean type. In C++, this type is bool, and has values true or false.
However, unlike Java, this Boolean type was not originally part of C++, which can result in
erroneous code that is legal. We discuss this in Section 1.3.2.

c++book.mif Page 15 Saturday, July 12, 2003 10:53 AM

Syntactic Differences

15

1.3 Syntactic Differences

Many of the basic operators, expressions, and statements are identical in C++ and Java, with
only minor differences.
1.3.1

Operators and Expressions

Operators in C++ are for the most part identical to their Java counterparts, though in many cases,
dubious Java code has well-defined semantics, whereas the same code in C++ can be implementation dependent. Classic examples of this include the overuse of the ++ operator, as in
int x = 5;
x = x++;

In Java this code will set x to be 5 (since the ++ is executed prior to the assignment of the original value of x). In C++, the result of this code is implementation dependent. In both languages
the code is poor style. More generally, Java specifies that arguments to methods and operands of
an operator (except for assignment operators) are evaluated in left-to-right order, whereas C++
makes no such guarantees. Thus, if function readInt is intended to read one integer from standard input, the result of readInt()-readInt() when 4 and 3 are placed on standard input
is guaranteed to be 1 in Java, but can be 1 or -1 in C++.
In Java, integer division rounds down, so 8/-5 is always -1. In C++, the result is implementation dependent and can be -1 or -2, depending on whether the division truncates or rounds
to the nearest integer.
C++ adds the comma operator, but the only good use is the use allowed in Java in the
expressions that are part of a for loop. The comma operator takes two expressions, evaluates the
first, then the second, and the result of the comma operator is the second expression. This is
more trouble than its worth, as illustrated by the fact that mistyping 3.14 as 3,14 does not
generate a compiler error (it simply evaluates to 14). Similarly, the call pow(3,4) returns 81
(three to the fourth power), whereas pow((3,4)) does not compile because (3,4) evaluates
to 4, and then pow is short a parameter. Yuk.
Java has the >>> shift operator, which does bit shifting (to the right), filling in high bits
with 0s. The >> shift operator in Java fills in high bits with whatever the high bit of the left operand was originally. In C++, only the >> shift operator is available, and its semantics depend on
whether the left operand is signed or unsigned. If it is unsigned, high bits are filled with 0; if it is
signed, high bits are filled in with the high bit of the left operand.
1.3.2

Conditionals

The if statement in C++ is identical to Java, except that in C++ the condition of the if statement
can be either an int or a bool. This stems from the fact that historically, early implementations of C++ did not have a bool type, but instead interpreted 0 as false, and anything non-zero
as true. The unfortunate consequence of this decision is that code such as
if( x = 0 )

c++book.mif Page 16 Saturday, July 12, 2003 10:53 AM

16

Chapter 1 Basic Types and Control Structures

which uses = instead of the intended == is not flagged as a compiler error, but instead sets x to
zero and evaluates the condition to false. This is possibly the most common trivial C++ programming error.
Occasionally, especially when reading old C++ code, shorthands that make use of the fact
that nonzero evaluates to true are placed in the conditional expression. Thus it is not uncommon
to see tests that should be
if( i != 0 )

rewritten as
if( i )

Newly written code should avoid these types of shortcuts since they tend to make the code less
readable, and no more efficient.
1.3.3

Loops

Like Java, C++ has the for loop, the while loop, and the do loop, along with the break and continue statements. However in C++, the use of the labelled break and labelled continue to affect
flow of an outer loop is not allowed.
C++ allows the goto, but its use is strongly discouraged.
1.3.4

Definite Assignment

An entire section of the Java Language Specification deals with errors that a Java compiler is
required to detect; such errors cause the compilation to fail. For instance, a Java compiler must
apply a conservative flow analysis is to ensure that every local variable has been definitely
assigned a value, regardless of the flow of control. Although occasionally this forces the programmer to rewrite valid code, much more often, it finds programming errors at compilation
time. A Java compiler is required to verify that all flows of control return a value in a non-void
method. C++ compilers are not required to perform such analyses, though many do if requested
to with suitable compilation options. It wont take long until you will write a C++ program that
compiles and uses an uninitialized variable, causing you runtime grief; so pay attention to warning messages and turn on the compilers ability to generate extra warnings.

1.4 Additional Syntax

This section discusses the relatively straightforward issues of primitive type casts, labels, and
the typedef statement.
1.4.1

Primitive Type Casts

In Java, the typecast operator has two basic uses: downcasting a reference variable to a subclass
reference type, and interpreting a primitive type as a different type (for instance, casting an int
to a byte). C++ has several different type casts. In this section we discuss the typecasts that are
used for primitive types.

c++book.mif Page 17 Saturday, July 12, 2003 10:53 AM

Additional Syntax

17

First, let us mention that C++ is more liberal than Java in accepting code without a cast. In
Java,
double x = 6.0;
int y;
y = x;

does not compile, but the code does compile (possibly with a warning about losing precision) in
C++. In some cases, typecasting is essential in order to produce a correct answer, as in
double quotient;
int x = 6;
int y = 10;
quotient = x / y;

In this example, in both Java and C++, quotient evaluates to 0.0 rather than 0.6, because
the operands are both of type int, and so division is performed with truncation to an int.
There are several syntactical ways to handle the typecast. The Java style, which also works in
C++,
quotient = (double) x / (double) y;

in which we cast both operands is easy to read. Since only one operand needs to be a double,
quotient = (double) x / y;

also works. However, this code is hard to read because one must be aware that the precedence of
the typecast operator is higher than the precedence of division, which is why the typecast applies
to x, and not x/y. An alternative in C++ is preferred:
quotient = double( x ) / y;

Here, the parentheses surround the expression to be casted, rather than the new type. (Note this
does not work for complex types, such as unsigned int).
Although this second style is preferable, it is hard to find the typecasts that are being used
in a program, since the syntax does not stand out. A third form, a late addition to C++, is to use
the static cast:
quotient = static_cast<double>( x ) / y;

Though this is more typing, it is easy to find the casts mechanically.

A different type of typecast, useful for downcasting in inheritance hierarchies is the
dynamic_cast. We discuss this in Section 6.7.
1.4.2

Labels

C++ allows labels prior to virtually any statement, for use with the goto statement. Since gotos
typically are symptomatic of poorly designed code, one would rarely expect to see labels in C++
code.
1.4.3

typedef Statement

The typedef statement is provided by C++ to allow the programmer to assign meaningful names

c++book.mif Page 18 Saturday, July 12, 2003 10:53 AM

18

Chapter 1 Basic Types and Control Structures

to existing types. As an example, suppose that we need to declare objects of type 32-bit
unsigned integer. On some machines this might be unsigned long, while on other, perhaps an
unsigned int is the most appropriate type. On the first machine, we would use
typedef unsigned long uint32;

whereas on the second machine, we would use

typedef unsigned int uint32;

Now, in the rest of the code, either machine can declare variables of type uint32:
uint32 x, y, z;

meaning that the non-portable aspect of the code is confined to one line.

1.5 Key Points

In C++, methods are known as member functions.
In C++, functions can be written outside of any class.
Flow in C++ begins at main. Function main returns an int. Each program can have
only one main.
Lines that begin with # are preprocessor directives. The include (preprocessor) directive
causes the contents of the file referenced by the directive to be logically placed in the
source code in place of the include directive.
Simple output is performed by using cout and the << operator.
Simple input is performed by using cin and the >> operator.
The C++ primitive types include short, int, long. These types may be signed or
unsigned. The byte type in Java can be simulated with signed char.
The C++ char is typically only eight bits and is used to store ASCII characters. The
newly introduced wchar_t type is 16 bits and can be used to store Unicode characters.
The standard header file cctype contains routines that are similar to the static methods
in java.lang.Character.
The bool type is a late addition to C++. Conditions in C++ use the C-style convention of
zero for false and non-zero for true, leading to programming errors when = is used instead
of ==.
C++ adds the comma operator, which is best avoided except in the typical Java use in for
loop expressions.
C++ does not support the labelled break or labelled continue.
C++ does not precisely define the semantics of some operations that can be reasonably
implemented in different ways on various machines.
A C++ compiler is not required to check that variables are definitely assigned values prior
to their use nor that all flows return a value in functions that have non-void return types.
Many compilers do so, but only if requested.
C++ has three forms of typecast, one of which looks like the Java style. For casting
between primitive types, the static_cast is recommended.

c++book.mif Page 19 Saturday, July 12, 2003 10:53 AM

Exercises

19

The typedef statement allows the programmer to assign meaningful names to existing
types.

1.6 Exercises
1. Compile, link, and run the program in Figure 1-1.
2. Write a program to print out the values of the largest and smallest int on your system by
printing the appropriate values in the standard header file <climits>.
3. What is the result of
if( x = y )
cout << x << endl;
else
cout << y << endl;

4. What is the result of

if( x < y < z )
cout << "Sorted" << endl;
else
cout << "Not sorted" << endl;

5. How does C++ handle uninitialized variables?

6. What does the typedef statement do?

c++book.mif Page 20 Saturday, July 12, 2003 10:53 AM

20

Chapter 1 Basic Types and Control Structures

c++book.mif Page 21 Saturday, July 12, 2003 10:53 AM

H A P T E R

Functions, Arrays, Strings,

and Parameter Passing

N this continuation of the discussion in Chapter 1, we

examine how non-class methods are written, discuss the library string and growable array types,
and see several parameter passing mechanisms that are available in C++.

2.1 Functions
In Section 1.1.1 we mentioned that C++ allows methods that are not members of a class. Typically we refer to these as functions. This section discusses functions; member functions are discussed in Chapter 4.
2.1.1

Function definition

Functions, not being part of a class, simply consist of a return type, function name, parameter
list, and body. This complete set, in which the function body is included, is often called a function definition. A function can be invoked from any other function simply by providing an
appropriate set of parameters. Figure 2-1 illustrates a function definition of max2 that looks
similar to Java code, except that both a visibility modifier and the static reserved word are
omitted. (We use max2 to avoid an accidental match with a Standard Library routine named
max).
2.1.2

Function Invocation

A function is invoked by simply providing arguments of type that are compatible with the formal parameters. In C++, there is more latitude granted when the arguments are not of the exact
type required. For instance, a long can be an acceptable argument (possibly with a warning)
even if the formal parameter is an int. When parameters are passed as in Figure 2-1, call-by-

21

c++book.mif Page 22 Saturday, July 12, 2003 10:53 AM

22

Chapter 2 Functions, Arrays, Strings, and Parameter Passing

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

#include <iostream>
using namespace std;
int max2( int a, int b )
{
return a > b ? a : b;
}
int main( )
{
int x = 37;
int y = 52;
cout << "Max is " << max2( x, y ) << endl;
return 0;
}

Figure 2-1

Invoking already-defined method

value is used. Thus the formal parameters become copies of the actual arguments, and as in Java,
the values of the actual arguments cannot be changed as a result of the function invocation. C++
also provides additional ways of passing parameters, including call-by-reference, which does
allow changes to the values of the actual arguments. This is discussed in Section 2.3.
2.1.3

Function Overloading

C++ allows function overloading along the same lines as Java. Several functions can have the
same name as long as they have different signatures. As with Java, the signature of a function
includes the function name and the number and types of the parameters. (It also includes the
parameter passing mechanism, which we discuss in Section 2.3). Like Java, the signature does
not include the return type. C++ has a complicated set of rules that are used to resolve an overloaded call in the case where there are several candidates.
2.1.4

Function Declarations

When a function is invoked, the C++ compiler will check actual parameters against the signatures of all the functions that it has seen thus far in order to resolve overloading. However,
because historically a C++ compiler processes the source code from top to bottom, if a function
definition has not already been seen, a compiler would not include its signature in the list of candidates. As a specific example, the code in Figure 2-2, in which the max2 method is defined
after the invocation is attempted does not compile.
To resolve this, the programmer would have to ensure that every function is defined prior
to its invocation. This is tedious at best, and possibly impossible in the case of mutual recursion
(in which two functions call each other). Thus C++ introduced the notion of the function prototype.

c++book.mif Page 23 Saturday, July 12, 2003 10:53 AM

Functions

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

23

#include <iostream>
using namespace std;
int main( )
{
int x = 37;
int y = 52;
cout << "Max is " << max2( x, y ) << endl;
return 0;
}
int max2( int a, int b )
{
return a > b ? a : b;
}

Figure 2-2

Invoking unseen method does not compile

The syntax of the function prototype is for all intents and purposes identical to the listing
of a method in a Java interface. For instance, the prototype for max2 is:
int max2( int a, int b );

The prototype does not include the body, but instead terminates the declaration with a
semicolon. The prototype, which is also known as a function declaration, allows the max2 function, which is presumed to be defined elsewhere, to be a candidate when max2 is invoked. Typ1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

#include <iostream>
using namespace std;
int max2( int a, int b );
int main( )
{
int x = 37;
int y = 52;
cout << "Max is " << max2( x, y ) << endl;
return 0;
}
int max2( int a, int b )
{
return a > b ? a : b;
}

Figure 2-3

Invoking method after function prototype

c++book.mif Page 24 Saturday, July 12, 2003 10:53 AM

24

Chapter 2 Functions, Arrays, Strings, and Parameter Passing

ical strategy would thus involve listing all the prototypes prior to the first function definition,
thus assuring that all functions can call all other functions. Figure 2-3 illustrates the use of the
function prototype.
2.1.5

Default Parameters

C++ allows the user to specify default values for formal parameters. Typically the default values
are included in the function declaration. As an example, the following declaration for
printInt specifies that by default integers should be printed in decimal:
void printInt( int n, int base = 10 );

Given this declaration, both of the following invocations are legal:

printInt( 50 );
printInt( 50, 8 );

// Outputs 50
// Outputs 62 (50 in octal)

If a default value is specified for a formal parameter, all subsequent formal parameters
must have default values too. In the example above, for instance, we could not specify
void printInt( int base = 10, int n );

Consequently, parameters that might be omitted are arranged so that the most likely to assume
default values will go last.
A default value cannot be specified in both the function declaration and function definition. A default value can be specified in the function definition instead of the declaration, but
this is considered bad practice because it requires that the function definition be placed prior to
the function invocation, in order for the signature that does not include parameters that have
defaults to be considered as a candidate during the overload resolution.
Overuse of default values can lead to ambiguities. For instance,
void printInt( int n, int base = 10 );
void printInt( int n );

is ambiguous if printInt is invoked with one parameter.

2.1.6

Inline Functions

In some situations, the overhead of making a function call can be significant. For instance, the
max2 routine is trivial, and so one might be tempted to simply replace the function invocation in
main with the code that max2 is logically performing:
cout << "Max is " << ( x > y ? x : y ) << endl;

Of course, this would be sacrificing good programming practice for speed. To avoid this
one can use the inline directive.
The inline directive suggests to the compiler that it should generate code that avoids the
overhead of a function call but is nonetheless semantically equivalent. The inline directive can
appear in either a function declaration or definition; however, the function definition must be
available at the point of invocation. Its use is illustrated in Figure 2-4.

c++book.mif Page 25 Saturday, July 12, 2003 10:53 AM

Functions

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

25

#include <iostream>
using namespace std;
inline int max2( int a, int b )
{
return a > b ? a : b;
}
int main( )
{
int x = 37;
int y = 52;
cout << "Max is " << max2( x, y ) << endl;
return 0;
}

Figure 2-4

Inline method definition must be visible prior to call to be expanded

Modern compilers have very sophisticated techniques that are used decide if honoring the
directive produces better code. The compiler may well refuse to perform the optimization for
functions that are too long. Additionally, the directive is likely to be ignored for recursive functions.
2.1.7

Separate Compilation

Often we would like to split a program up into several source files. In such a case, we expect to
be able to invoke a function that is defined in one file from a point that is not in the same file.
This is acceptable as long as a prototype is visible.
The typical scenario to allow this is that instead of having each file list its function declarations at the top, each file creates a corresponding .h file with the function declarations. Then,
any file that needs these declarations can provide an appropriate include directive.
In our example this gives three files all shown in Figure 2-5. The file max2.h simply lists
the function declarations for all functions defined in max2.cpp. The main program is defined
is a separate file and provides the include directive. Recall that this directive replaces line 5 with
the contents of the file max2.h. Finally, max2.cpp provides an include directive also. This
include directive is not needed, but would be typical in the case in which max2.cpp had several functions that were calling each other. Section 4.12 discusses one other issue that is common with this technique.
Compiling this program depends on the platform. Most IDEs perform two steps: compilation and linking. The compilation stage verifies the syntax and generates object code for each of
the files. The linking stage is used to resolve the function invocations with the actual definitions
(sometimes this occurs at runtime).

c++book.mif Page 26 Saturday, July 12, 2003 10:53 AM

26

Chapter 2 Functions, Arrays, Strings, and Parameter Passing

1
2

// Source file: max2.h

int max2( int a, int b );

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

// Source file: main.cpp

#include <iostream>
using namespace std;

1
2
3
4
5
6
7

// Source file: max2.cpp

#include "max2.h"

#include "max2.h"
int main( )
{
int x = 37;
int y = 52;
cout << "Max is " << max2( x, y ) << endl;
return 0;
}

int max2( int a, int b )

{
return a > b ? a : b;
}

Figure 2-5

Separate compilation: Shows max2.h, main.cpp, max2.cpp

In a typical IDE, the .cpp files would be made part of the project; the .h files would not.
Since the contents of the .h files are logically copied into the .cpp files by the include directives, the .h files should not define any functions or global variables, since this would result in
an error message concerning multiple definitions during the linking stage. One exception to this
rule is that inline functions should be placed in the .h files; the compiler will make arrangements to avoid the multiple definition error message, even if the inline directive is not honored.
If a function is declared, and it is invoked, but there is no definition available anywhere,
then during the linking stage an error will occur, stating that the function is undefined. Often this
means either that your function declaration is not the same as the function definition or that your
project does not include the file that contains the missing function definition. Most compilers
will tell you not only the name, but the signature of the function that is either undefined or multiply defined.

2.2 Arrays and Strings

In C++, we can declare and use arrays in two basic ways. The primitive method is to use the
built-in array. The alternative is to use a vector. The differences between the two is somewhat

c++book.mif Page 27 Saturday, July 12, 2003 10:53 AM

Arrays and Strings

27

comparable as the difference between Javas built-in array type and ArrayList library type.
Actually, this is probably an understatement, because using built-in arrays in C++ is much more
difficult than using the Java counterpart, whereas vector and ArrayList are for all intents
and purposes identical. The reason is that using C++ built-in arrays may require you to independently maintain the array size and also provide additional code to reclaim memory. vector has
neither of these problems. The built-in C++ array type is discussed in Chapter 11. If you would
like to keep your blood pressure low, its a good idea to stick with the vector type as much as
possible.
Similarly, strings in C++ come in two basic flavors. The primitive string is simply a builtin array of characters, and are exasperating and dangerous to use. See Chapter 11. The library
type, string, is a full-fledged string class and is easy to use.
2.2.1

Using the vector Library Type

To use the standard vector, your program must include the standard header file vector, as in
#include <vector>

A using directive may be needed if one has not already been provided.
As we will discuss in more detail in Chapter 3, C++ deals with objects differently than
Java. Specifically, in Java, objects are accessed by reference variables, and are treated, by
design, differently from primitive types. In C++, objects are meant to look just like primitive
types. Thus, they can be declared as local variables on the runtime stack, rather than allocated
from the memory heap by calling new.
A vector is created by giving it a name, telling the compiler what type the elements are,
and optionally providing a size (that defaults to zero). The vector maintains its current size
and current capacity in the same manner as an ArrayList. If additions to the vector cause
its size to exceed capacity, than internally the capacity is expanded. Thus in the following:
vector<int> arr1( 3 );
vector<int> arr2;

arr1 is a vector that stores ints; the valid indices are currently 0, 1, 2, and an attempt to use
any other index is incorrect. arr2 is a vector that also stores ints. There are no valid indices. A vector is indexed by using [].
The size of a vector can always be obtained by invoking member function size. It can
be resized by invoking member function resize.
The add method in ArrayList, which adds to the end of the collection of objects is
renamed as push_back in the vector. Thus, the result of arr1.push_back(2) is that
arr1.size() now returns 4, and arr1[3] is now 2.
Figure 2-6 shows a vector that is used to store the first 100 perfect squares. In this code,
replacing line 10 with squares[i]=i*i does not work; the array index would be out-ofbounds. Unfortunately, unlike Java, this does not cause an exception; rather the program runs
and some part of memory gets unintentionally overwritten.
Java programmers might be tempted to create a vector of size 0 by using code such as

c++book.mif Page 28 Saturday, July 12, 2003 10:53 AM

28

Chapter 2 Functions, Arrays, Strings, and Parameter Passing

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

#include <iostream>
#include <vector>
using namespace std;
int main( )
{
vector<int> squares;
for( int i = 0; i < 100; i++ )
squares.push_back( i * i );
for( int j = 0; j < squares.size( ); j++ )
cout << j << " squared is " << squares[ j ] << endl;
return 0;
}

Figure 2-6

Using the vector class

vector<int> arr3( );

thinking that the default parameter would use size 0. Unfortunately this is wrong! The declaration above does not create a vector; instead, it states that arr3 is a function that takes no
parameters and returns a vector<int>. Ouch! The result would almost certainly be a bizarre
string of unintelligible error messages when arr3 was used later.
2.2.2

Using the string Library Type

To use the string library type, you must have the include directive (and possibly the usual
using namespace std directive if it is not already present):
#include <string>

Because objects in Java are meant to look like primitive types, strings in C++ are also
easier to use than in Java. Most importantly, the normal equality and relational operators ==, !=,
<, <=, >, >= all work for the C++ string type. No more bugs because you forgot to use
.equals. When = is applied to a string, a copy is made; changes to the original do not
affect the copy. The length of a string can always be obtained by calling the length member function.
Java strings are immutable: once created a Java strings state cannot be changed. In contrast, the C++ string is mutable. This has two consequences. First, using the array indexing
operator [], not only can individual characters be accessed, they easily can be changed (however, as in vector, no bounds checking is performed). Second, the += operator is efficient; no
more StringBuffers! Figure 2-7 illustrates the use of string concatenation;
makeLongString takes quadratic time in Java, but is linear in C++.

c++book.mif Page 29 Saturday, July 12, 2003 10:53 AM

Arrays and Strings

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

29

#include <iostream>
#include <string>
using namespace std;
// return a string that contains n As
// In Java this code takes forever
string makeLongString( int n )
{
string result = "";
for( int i = 0; i < n; i++ )
result += "A";
return result;
}
int main( )
{
string manyAs = makeLongString( 250000 );
cout << "Short string is " << makeLongString( 20 ) << endl;
cout << "Length is " << manyAs.length( ) << endl;
return 0;
}

Figure 2-7

String concatenation is efficient in C++ because strings are mutable

Two member functions deserve special attention. First, the member function to get substrings is substr. However, unlike Java, the parameters in C++ represent the starting position
and length of the substring, rather than the starting position and first non-included position. Thus
in
string s = "hello";
string sub = s.substr( 2, 2 ); // gives "ll"

sub is a string of length 2, whose first character is s[2].

Second, some legacy code (and C++ libraries) expect primitive strings, not the string
library type. Resist the temptation to use primitive strings. Instead, do all your dirty work with
the string library type, and when a primitive string is required, use the c_str string member
function that is part of string to obtain the primitive equivalent.
2.2.3

Arrays of Objects

In Java, there is no such thing as an array of objects. Instead, what you have is an array of references; when the array is created, by default all the references are null. In C++, when one creates an array of objects, either using a primitive array or a vector, we really get an array of

c++book.mif Page 30 Saturday, July 12, 2003 10:53 AM

30

Chapter 2 Functions, Arrays, Strings, and Parameter Passing

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

#include <iostream>
using namespace std;
// Incorrect implementation of swap2
void swap2( int val1, int val2 )
{
int tmp = val1;
val1 = val2;
val2 = tmp;
}
int main( )
{
int x = 37;
int y = 52;
swap2( x, y );
cout << x << " " << y << endl;
return 0;
}

Figure 2-8

Swap function that does not work because of call-by-value

objects. Specifically, all the items in the array are objects that have been created by calling an
appropriate zero-parameter constructor. In C++, this does not work well when the objects are
different types (for instance they are related via inheritance), but if they are the same type, e.g.
string, it works fine and no special syntax is required.
2.2.4

Primitive Arrays of Constants

Occasionally, we revert to primitive arrays when we have a global constants. The reason is the
convenient array initialization, that is similar to Java and illustrated as:
const int DAYS_INT_MONTH[ ] = { 31, 28, 31, 30, 31, 30,
31, 31, 30, 31, 30, 31 };

Unlike Java, the [] follows the identifier. However, like Java, this creates an array of 12
items. The number of items can be determined by
const int NUM_MONTHS = sizeof( DAYS_IN_MONTH ) /
sizeof( DAYS_IN_MONTH[0] );

2.2.5

Multidimensional Arrays

The primitive implementation of multidimensional arrays is truly exasperating to use, and is discussed briefly in Section 11.6. There is no library type available, but it is fairly trivial to write
one ourselves, and we do so in Section 7.3.

c++book.mif Page 31 Saturday, July 12, 2003 10:53 AM

Parameter Passing

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

31

#include <iostream>
using namespace std;
// Correct implementation of swap2
void swap2( int & a, int & b )
{
int tmp = a;
a = b;
b = tmp;
}
int main( )
{
int x = 37;
int y = 52;
swap2( x, y );
cout << x << " " << y << endl;
return 0;
}

Figure 2-9

Swap function that works using call-by-reference

2.3 Parameter Passing

In Java, all parameters are passed using call-by-value: the value of the actual arguments are copied into the formal parameters.
Parameter passing in C++ is also call-by-value by default. Although this is often desirable,
there are two separate situations where an alternative is useful.
2.3.1

Call by Reference

Figure 2-8 illustrates a function, swap2, that attempts to swap the values of its parameters. But
if main is run, the values of x and y are not swapped, because of call-by-value. The code swaps
val1 and val2, but since these are simply copies of x and y, it is not possible to write a swap
routine that changes x and y when x and y are passed using call-by-value.
C++ allows call-by-reference. To pass a parameter using call-by-reference, we simply
place an & prior to each of the parameters that are to be passed using this mechanism. Thus,
some parameters can be passed using call-by-value, and others using call-by-reference. The
modified version of swap2, which is now correct, is shown in Figure 2-9. Observe that main is
unchanged no special syntax is used to invoke the method.

c++book.mif Page 32 Saturday, July 12, 2003 10:53 AM

32

Chapter 2 Functions, Arrays, Strings, and Parameter Passing

Because formal arguments that are passed using call-by-reference are modifiable in the
invoked function, it is illegal to pass a constant using call by reference. Thus the code
swap2(x,3) would not compile.
Actual arguments must be type-compatible with the formal arguments, without the use of
a typecast. This is required because a typecast generates a temporary variable, and the temporary
variable would become the actual argument, and then changes to the formal parameter in the
invoked function would change the temporary variable (instead of the original), leading to hardto-find bugs.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36

#include <iostream>
#include <vector>
using namespace std;
// Broken binarySearch because of call-by-value
int binarySearch( vector<int> arr, int x )
{
int low = 0, high = arr.size( ) - 1;
while( low <= high )
{
int mid = ( low + high ) / 2;
if( arr[ mid ] == x )
return mid;
else if( x < arr[ mid ] )
high = mid - 1;
else
low = mid + 1;
}
return -1; // not found
}
int main( )
{
vector<int> v;
for( int i = 0; i < 30000; i++ )
v.push_back( i * i );
for( int j = 100000; j < 105000; j++ )
if( binarySearch( v, j ) >= 0 )
cout << j << " is a perfect square" << endl;
return 0;
}

Figure 2-10

Slow binary search method because of call-by-value

c++book.mif Page 33 Saturday, July 12, 2003 10:53 AM

Parameter Passing

33

Finally, we mention that the parameter passing mechanism is part of the signature of the
method and must be included in function declarations.
2.3.2

Call by Constant Reference

Figure 2-10 illustrates a second problem of call-by-value. Here we have a binary search algorithm that returns the index of an item in a sorted vector, and a main that repeatedly calls
binary search. In theory, binary search is an efficient algorithm, requiring only a logarithmic
number of steps, so on a modern machine, 5000 binary searches in a 30,000 item array should
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36

#include <iostream>
#include <vector>
using namespace std;
// Fixed binarySearch uses call-by-constant reference
int binarySearch( const vector<int> & arr, int x )
{
int low = 0, high = arr.size( ) - 1;
while( low <= high )
{
int mid = ( low + high ) / 2;
if( arr[ mid ] == x )
return mid;
else if( x < arr[ mid ] )
high = mid - 1;
else
low = mid + 1;
}
return -1; // not found
}
int main( )
{
vector<int> v;
for( int i = 0; i < 30000; i++ )
v.push_back( i * i );
for( int j = 100000; j < 105000; j++ )
if( binarySearch( v, j ) >= 0 )
cout << j << " is a perfect square" << endl;
return 0;
}

Figure 2-11

Correct binary search method using call-by-constant reference

c++book.mif Page 34 Saturday, July 12, 2003 10:53 AM

34

Chapter 2 Functions, Arrays, Strings, and Parameter Passing

execute in less than a millisecond. However, the code takes noticeable longer, using up seconds
of CPU time. Here the problem is not the binary search, but the parameter passing: because we
are using call-by-value, each call to binarySearch makes a complete copy of vector v.
Needless to say, 5,000 copies of a 30,000 element vector doesnt come cheap. This
problem never occurs in Java because all objects (non-primitive entities) are accessed using Java
reference variables, and so objects are always shared, and never copied by using =.
But in C++, a variable of type vector<int> stores the entire state of the object, and =
copies an entire object to another object. Similarly call-by-value copies the entire state of the
actual argument to the formal parameter.
Certainly one way to avoid this problem would be to use call-by-reference. Then the formal parameter is just another name for the actual argument; no copy is made. Although this
would significantly increase the speed of the program, and solve the problem, there are two serious drawbacks. First, using call-by-reference changes the semantics of the function in that the
caller no longer knows that the actual argument will be unchanged after the call. Second, as we
mentioned in Section 2.3.1, constants or actual arguments requiring typecasts would no longer
be acceptable parameters.
The solution, shown in Figure 2-11 is to augment the call-by-reference with the reserved
word const, signifying that the parameters are to be passed by reference, but that the function
promises not to make any changes to the formal parameter. Thus the actual argument is protected from being changed. If the function implementation attempts to make a change to a
const formal parameter, the compiler will complain and the function will not compile (it is
possible to cast away the const-ness and subvert this rule, but thats life with C++). We denote
this parameter passing mechanism as call-by-constant reference (even though the more verbose
call-by-reference to a constant is more accurate)
When a parameter is passed using call-by-constant reference, it is acceptable to supply a
constant or an actual argument that requires a typecast. In effect, as far as the caller is concerned,
call-by-constant reference has the same semantics as call-by-value, except that copying is
avoided.
2.3.3

Parameter Passing Summary

In C++ choosing a parameter-passing mechanism is an easily overlooked chore of the programmer that can affect correctness and efficiency. The rules are actually relatively simple:
Call by reference is required for any object that may be altered by the function.
Call by value is appropriate for small objects that should not be altered by the function.
This generally includes primitive types and also function objects (Section 7.6.3).
Call by constant reference is appropriate for large objects that should not be altered by the
function. This generally includes library containers such as vector, general class types,
and even string.

c++book.mif Page 35 Saturday, July 12, 2003 10:53 AM

Key Points

35

2.4 Key Points

A function declaration consists of the function minus the body and looks like a method
declaration in a Java interface. A function definition includes the body.
When invoking a function, only those functions who declarations or definitions have
already been seen are eligible to be candidates.
C++ allows default parameters; the default parameters must be the last parameters.
In C++, objects can be created without calling new, and accessed and copied as if they
were primitive entities. See Chapter 3 for more information about this.
C++ supports arrays using both a primitive array and a vector library type. Similarly,
strings can be supported using both primitive arrays of char and a string library type.
The vector in C++ has functionality that is similar to the ArrayList class in Java.
The string class in C++ has functionality that is similar to the String class in Java.
The += operator is efficient, the substring function substr takes a different second
parameter, and all the relational operators work, so methods such as equals and
compareTo are not required.
C++ supports three modes of parameter passing: call-by-value, call-by-reference, and callby-constant reference.

2.5 Exercises
1. Grab the most recent tax table. Write a function that takes an adjusted gross income and
filing status (married, single, etc.) and returns the amount of tax owed. Write a test program to verify that your function behaves correctly.
2. Write a function to compute X N for nonegative integers N. Assume that X 0 = 1.
3. What is the difference between a function declaration and a function definition?
4. Describe default parameters in C++.
5. What is an inline directive in C++?
6. Describe how C++ supports separate compilation.
7. Write a function that accepts a vector of strings and returns a vector containing the strings
in the vector parameter that have the longest length (in other words, if there are ten strings
that are tied for being the longest length, the return value is a vector of size ten containing
those strings). Then write a test program that reads an arbitrary number of strings, invokes
the function, and outputs the strings returned by the vector.
8. What are the basic differences between the C++ string and the Java String classes?
9. What are the different parameters passing mechainsms in C++, and when are they used?

c++book.mif Page 36 Saturday, July 12, 2003 10:53 AM

36

Chapter 2 Functions, Arrays, Strings, and Parameter Passing

c++book.mif Page 37 Saturday, July 12, 2003 10:53 AM

H A P T E R

Pointers and Reference

Variables

OSSIBLY the most significant difference between

Java and C++ concerns how objects are managed by the system. Java has a simple model that
guarantees, among other things, that once an object has been created, the Virtual Machine will
never attempt to destroy the objects memory as long as the object is being actively referenced,
either directly or indirectly. C++, on the other hand, allows objects to be accessed in many different ways, making coding very tricky, and leading to many subtle and hard-to-find runtime
errors.
In this chapter, we review the Java typing system, briefly describe the C++ object model,
and introduce the C++ concepts of pointer variables and reference variables. We will also
describe how the C++ programmer destroys objects that are no longer needed.

3.1 Java vs. C++ Memory Model

In Java, life is relatively simple. Except for instances of the eight primitive types, everything in
Java is an object. Furthermore, objects in Java can only be created on the memory heap. This is
typically done by calling new (the reflection API allows an alternative, but the object still
resides on the memory heap). Such objects are always accessed by reference variables. The Java
programmer cannot force the Virtual Machine to reclaim an unused object. Instead, if an object
no longer has any references to it, either directly or indirectly, the Virtual Machine may, at its
discretion declare the object inactive and reclaim the object. This process is known as garbage
collection. The Virtual Machine furthermore guarantees that it will never attempt to reclaim an
active object.
In Java, primitive entities can never be accessed by reference variables, and can never be
created by invoking new. Thus when a local variable in Java is a primitive type, it is allocated on

37

c++book.mif Page 38 Saturday, July 12, 2003 10:53 AM

38

Chapter 3 Pointers and Reference Variables

the runtime stack, and the local primitive variable is automatically invalidated when the block
that it was created in ends.
The C++ memory model is significantly more complicated. As we have already seen,
local variables, including objects in C++ (such as vectors and strings) can be created without calling new. In such as case, these local objects are allocated on the runtime stack and have
the same semantics as primitive types. Two important consequences of this are as follows:
First, when = is applied to objects, the state of one object is copied. This contrasts to Java,
where objects are accessed indirectly, by reference variables, and an = copies the value of the
reference variable, rather than the state of the object.
Second, when the block in which a local object is created ends, the object will automatically be reclaimed by the system.
Thus in the following code:
void silly( )
{
vector<int> arr1;
vector<int> arr2;
...
// other code not shown
arr2 = arr1;
...
// other code not shown
}

arr1 and arr2 are separate vector<int> objects. The statement arr2=arr1 copies the
entire contents of vector<int> arr1 into vector<int> arr2. When silly returns,
arr1 and arr2 will be destroyed, and their memory reclaimed, as part of the function return
sequence.
This example illustrates a signficant difference between Java and C++. In Java, objects are
shared by several reference variables and are rarely copied. C++ encourages, by default, the
copying of objects. However, copying objects can take time, and often we need to avoid this
expense. For instance, we have already seen in Section 2.3 additional syntax to allow copies to
be avoided.
In a more general setting inherited from C, C++ allows the programmer to obtain a variable that stores the memory address where the object is being kept. Such a variable is called a
pointer variable in C++. Pointer variables in C++ have many of the semantics as reference variables in Java, with extra flexibility (that implies extra dangers). C++ also has another type of
variable called the reference variable, which despite its name is not similar to the reference variable in Java. In the remainder of this chapter, we will discuss both types of variables, and several
tricky C++ issues that are associated with their use.

3.2 Pointers
A pointer variable in C++ is a variable that stores the memory address of any other entity. The
entity can be a primitive type or a class type; however, using pointer variables for primitive
types in good C++ code is rarely needed. Since it does simplify our examples, we will make use

c++book.mif Page 39 Saturday, July 12, 2003 10:53 AM

Pointers

39

of pointers to integers in this section.

To have a pointer point at an object, we need to know that target objects memory address
(that is, where it is stored). For object obj, its memory address is given by applying the unary
address-of operator &. Thus &obj is the memory location that stores obj.
We can declare that an object ptr points at an int object by saying
int *ptr;

The value represented by ptr is an address. As with integer objects, this declaration does not
initialize ptr to any particular value, so using ptr before assigning to it invariably produces
bad results (such as a program crash). Suppose we also have the following declarations:
int x = 5;
int y = 7;

We can make ptr point at x by assigning to ptr the memory location where x is stored. Thus
ptr = &x;

// LEGAL

sets ptr to point at x. Figure 3-1 illustrates this in two ways. On the left a memory model shows
where each object is stored. The figure on the right uses an arrow to indicate pointing.

(&x) 1000

x=5

(&y) 1004

y=7

(&ptr) 1200

1000

ptr

Figure 3-1

x = 10

(&y) 1004

y=7
ptr

Figure 3-2

10

Pointer illustration

(&x) 1000

(&ptr) 1200

ptr = &x = 1000

Result of *ptr=10

c++book.mif Page 40 Saturday, July 12, 2003 10:53 AM

40

Chapter 3 Pointers and Reference Variables

The data being pointed at is obtained by the unary dereferencing operator *. In Figure 3-1
*ptr will evaluate to 5, which is the value of the pointed-at variable x. It is illegal to dereference something that is not a pointer. The * operator is the inverse of & (for example, *&x=5 is
the same as x=5 as long as &x is legal). Dereferencing works not only for reading values from
an object but also for writing new values to the object. Thus, if we say
*ptr = 10;

// LEGAL

we have changed the value of x to 10. Figure 3-2 shows the changes that result. This shows the
problem with pointers: Unrestricted alterations are possible, and a runaway pointer can overwrite all sorts of variables unintentionally.
We could also have initialized ptr at declaration time by having it point to x:
int x = 5;
int y = 7;
int *ptr = &x;

// LEGAL

The declaration says that x is an int initialized to 5, y is an int initialized to 7, and ptr is a
pointer to an int and is initialized to point at x. Let us look at what could have gone wrong.
The following declaration sequence is incorrect:
int *ptr = &x;
int x = 5;
int y = 7;

// ILLEGAL: x is not declared yet

Here we are using x before it has been declared, so the compiler will complain. Here is another
common error:
int x = 5;
int y = 7;
int *ptr = x;

// ILLEGAL: x is not an address

In this case we are trying to have ptr point at x, but we have forgotten that a pointer holds an
address. Thus we need an address on the right side of the assignment. The compiler will complain that we have forgotten the &, but its error message may initially appear cryptic.

(&x) 1000

x=5

(&y) 1004

y=7
ptr

(&ptr) 1200

Figure 3-3

ptr = ?

Uninitialized pointer

c++book.mif Page 41 Saturday, July 12, 2003 10:53 AM

Pointers

41

Continuing with this example, suppose that we have the correct declaration but with ptr
uninitialized:
int x = 5;
int y = 7;
int *ptr;

// LEGAL but ptr is uninitialized

What is the value of ptr? As Figure 3-3 shows, the value is undefined because it was never initialized. Thus the value of *ptr is also undefined. However, using *ptr when ptr is undefined is worse because ptr could hold an address that makes absolutely no sense at all, thus
causing a program crash if it is dereferenced. Even worse, ptr could be pointing at an address
that is accessible, in which case the program will not immediately crash but will be erroneous. If
*ptr is the target of an assignment, then we would be accidentally changing some other data,
which could result in a crash at a later point. This is a tough error to detect because the cause and
symptom may be widely separated in time.
We have already seen the correct syntax for the assignment:
ptr = &x;

// LEGAL

Suppose that we forget the address-of operator. Then the assignment

ptr = x;

// ILLEGAL: x is not an address

rightly generates a compiler error. There are two ways to make the compiler shut up. One is to
take the address on the right side, as in the correct syntax. The other method is erroneous:
*ptr = x;

// Semantically incorrect

The compiler is quiet because the statement says that the int to which ptr is pointing should
get the value of x. For instance, if ptr is &y, then y is assigned the value of x. This assignment
is perfectly legal, but it does not make ptr point at x. Moreover, if ptr is uninitialized, dereferencing it is likely to cause a run-time error, as discussed above. This error is obvious from Figure 3-3. The moral is to always draw a picture at the first sign of pointer trouble.
Using *ptr=x instead of ptr=&x is a common error for two reasons. First, since it
makes the compiler quiet, programmers feel comfortable about using the incorrect semantics.
Second, it looks somewhat like the syntax used for initialization at declaration time. The difference is that the * at declaration time is not a dereferencing * but rather just an indication that the
object is a pointer type.
Some final words before we get to some substantive uses: First, sometimes we want to
state explicitly that a pointer is pointing nowhere, as opposed to an undefined location. The
NULL pointer points at a memory location that is guaranteed to be incapable of holding anything. Consequently, a NULL pointer cannot be dereferenced. The symbolic constant NULL was
prevalent in C, but is being phased out in favor of an explicit 0. But many still feel that NULL is
more readable, and so we use NULL, assuming that the following declaration exists:
const int NULL = 0;

Pointers are best initialized to the NULL pointer because in many cases they have no
default initial values (these rules apply to other predefined types as well).

c++book.mif Page 42 Saturday, July 12, 2003 10:53 AM

42

Chapter 3 Pointers and Reference Variables

Second, a dereferenced pointer behaves just like the object that it is pointing at. Thus, after
the following three statements, the value stored in x is 15:
x = 5;
ptr = &x;
*ptr += 10;

However, we must be cognizant of precedence rules because (as we discuss in Section 11.3) it is
possible to perform arithmetic not only on the dereferenced values but also on the (un-dereferenced) pointers themselves.1 As an example, the following two statements are very different:
*ptr += 1;
*ptr++;

In the first statement the += operator is applied to *ptr, but in the second statement the ++
operator is applied to ptr. The result of applying the ++ operator to ptr is that ptr will be
changed to point at a memory location one memory unit larger than it used to. (We discuss why
this might be useful in Section 11.3.)
Third, if ptr1 and ptr2 are pointers to the same type, then
ptr1 = ptr2;

sets ptr1 to point to the same location as ptr2, while

*ptr1 = *ptr2;

assigns the dereferenced ptr1 the value of the dereferenced ptr2. Figure 3-4 shows that these
statements are quite different. Moreover, when the wrong form is used mistakenly, the consequences might not be obvious immediately. In the previous examples, after the assignment,
*ptr1 and *ptr2 are both 7. Similarly, the expression
ptr1 == ptr2

is true if the two pointers are pointing at the same memory location, while
*ptr1 == *ptr2

is true if the values stored at the two indicated addresses are equal. It is a common mistake to use
the wrong form.
The requirement that ptr1 and ptr2 point to the same type is a consequence of the fact
that C++ is strongly typed: We cannot mix different types of pointers without an explicit type
conversion, unless the user has provided an implicit type conversion.
If several pointers are declared in one statement, the * must precede each variable:
int *ptr1, *ptr2;
int *ptr1, ptr2;

// Correct ptr1 and ptr2 are both pointer to int

// Wrong!! ptr2 is an int

Finally, when pointers are declared, the white space that surrounds the * is unimportant to
the compiler. Pick a style that you like.
1.

This is an unfortunate consequence of C++s very liberal rules that allow arithmetic on pointers, making use of
the fact that pointers are internally stored as integers. We discuss the reasoning for this in Section 11.3.

c++book.mif Page 43 Saturday, July 12, 2003 10:53 AM

Memory Heap Management

43

5
ptr1

5
ptr1

7
ptr2

ptr1

7
ptr2

(a)
Figure 3-4

7
ptr2

(b)

y
(c)

(a) Initial state; (b) ptr1=ptr2 starting from initial state; (c) *ptr1=*ptr2
starting from initial state

3.3 Memory Heap Management

Often in C++, local variables can be allocated on the runtime stack. We know that stack-allocated local variables are created when they are reached in the function, and destroyed when they
are no longer in scope (for instance, when the function returns). Like Java, C++ allows entities
to be created from the memory heap. In C++ this introduces several complications.
3.3.1

The new Operator

Like Java, objects can be created from the memory heap by calling new. The result of new is a
pointer to the newly created object, allocated from the memory heap, rather than the runtime
stack. new behaves the same in C++ and Java, and parameters can be provided to control initialization of the newly created object.
Figure 3-5 illustrates the issues involved in memory heap allocation, which is often called
dynamic memory allocation in C++. It must be emphasized that this example is a poor use of
dynamic memory and an automatic string should be used instead. We do it only to illustrate
dynamic memory allocation in a simple context.
In Figure 3-5, line 9 creates a new string object dynamically. Note that strPtr is a
pointer to a string, so the string itself is accessed by *strPtr, as shown on lines 10 to
14. The parentheses are needed at line 11 because of precedence rules.
3.3.2

Garbage collection and delete operator

In Java, when an object is no longer referenced, it is subject to automatic garbage collection. The
programmer does not have to worry about it. C++ does not have garbage collection. When an
object that is allocated by new is no longer referenced, the delete operator must be applied to
the object (through a pointer). Otherwise, the memory that it consumes is lost (until the program
terminates). This is known as a memory leak. Memory leaks are, unfortunately, common occur-

c++book.mif Page 44 Saturday, July 12, 2003 10:53 AM

44

Chapter 3 Pointers and Reference Variables

rences in many C++ programs. The delete operator is illustrated at line 16.
An example of a memory leak is shown in Figure 3-6, in which we return at line 9 without
calling delete. Fortunately, many sources of memory leaks can be automatically removed
with care. One important rule is to not use new when a stack-allocated variable can be used
instead. A stack-allocated variable is automatically cleaned up (hence it is also known as an
automatic variable in C++). Thus in this code, it would make sense to allocate the vector on the
runtime stack (and avoid the pointer) instead of using new to create it on the heap.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

#include <iostream>
#include <string>
using namespace std;
int main( )
{
string *strPtr;
strPtr = new string( "hello" );
cout << "The string is: " << *strPtr << endl;
cout << "Its length is: " << (*strPtr).length( ) << endl;
*strPtr += " world";
cout << "Now the string is " << *strPtr << endl;
delete strPtr;
return 0;
}

Figure 3-5

1
2
3
4
5
6
7
8
9
10
11
12

Illustration of dynamic memory allocation

#include <iostream>
#include <vector>
using namespace std;
void leak( int i )
{
vector<int> *ptrToVector = new vector<int>( i );
if( i % 2 == 1 )
return;
// some other code not shown ...
delete ptrToVector;
}

Figure 3-6

Illustration of memory leak in program that does nothing useful

c++book.mif Page 45 Saturday, July 12, 2003 10:53 AM

Memory Heap Management

3.3.3

45

Stale pointers

One reason that programmers can get in trouble when using pointers is that it is possible, and
generally expected, that one object may have several pointers pointing at it. Consider the following code:
string *s = new string( "hello" ); // s points at new string
string *t = s;
// t points there, too
delete t;
// The object is gone

Nobody would deliberately write these three lines of code next to each other; instead,
assume that they are scattered in a complex function. Prior to the call to delete, we have one
dynamically allocated object that has two pointers pointing to it.
After the call to delete, the values of s and t (that is, where they are pointing) are
unchanged. However, as illustrated in Figure 3-7, they are now stale. A stale pointer is a pointer
whose value no longer refers to a valid object. Dereferencing s and t can lead to unpredictable
results. What makes things especially difficult is that although it is obvious that t is stale, the
fact that s is stale is much less obvious, if, as assumed, these statements are scattered in a complex function. Furthermore, it is possible that in some situations, the memory that was occupied
by the object is unchanged until a later call to new claims the memory, which can give the illusion that there is no problem.
3.3.4

Double-delete

A second problem is the so-called double-delete. A double-delete occurs when we attempt to

call delete on the same object more than once. This would occur if we now made the call
delete s;

// Oops -- double delete

Since s is stale, the object that it points to is no longer valid. Trouble in the form of a runtime
error is likely to result.
Thus, we see the perils of dynamic memory allocation. We must be certain to never call
delete more than once on an object, and then only after we no longer need it. But if we dont
call delete at all, we get a memory leak. And if we have a pointer variable and intend to call
delete, we must be certain that the object being pointed at was created by a call to new. When
we have functions calling functions calling other functions, it is hard to keep track of everything.

s
t
Figure 3-7

"hello"
Stale pointers: because of the call to delete t, pointers s and t are now
pointing at an object that no longer exists; a call to delete s would now be
an illegal double-deletion

c++book.mif Page 46 Saturday, July 12, 2003 10:53 AM

46

Chapter 3 Pointers and Reference Variables

3.3.5

Functions that Return Pointers

If we allocate a large object as local variable and then return it, then we incur the overhead of
large copy. For this reason, it is not uncommon to see functions that return pointers instead of
objects themselves. Returning pointers is also common in the implementation of linked data
structures (in this case, we are returning pointers to nodes). However, when the return type of a
function is a pointer, one must be extremely careful to avoid stale pointers, memory leaks, and
invalid deletes.
Pointers can go stale even if no dynamic allocation is performed. Consider the code in Figure 3-8. For no good reason (except to illustrate the error), function dup returns a pointer to a
string. If dup calls new to create a string, then the caller will be responsible for calling
delete. Rather than burdening the caller, the programmer has mistakenly decided to have dup
use a stack-allocated string, and return its address. The program compiles but may or may not
work. It has an error. The problem is that the value that dup returns is a pointer. But the pointer
is pointing at ret, which no longer exists, because it is an stack-allocated variable and dup has
returned. One possible scenario, is that the first print at line 10 works, but in invoking the print
routine, the runtime stack overwrites the part of the runtime stack where the local variable ret
was being stored. If so, the second print at line 11 fails to produce the same answer, which can
be quite shocking. Another possibility is that the code works fine, and you have a latent bug.
When returning pointers, make sure that you have something to point to, and that the something
exists after the return is complete.
Suppose that it is important for dup to return a pointer to a string. How can we fix the
stale pointer problem? The easiest way would appear to be to create the new string at line 3 on
the memory heap, as shown in Figure 3-9. However, by doing so, we now force the caller to
clean up memory by calling delete when the caller is done using the return value. If the caller
doesnt do so, a memory leak has been created. In this situation, one would expect to see comments accompanying the dup method instructing callers of their obligations.
1
2
3
4
5
6
7
8
9
10
11
12
13

string * dup( const string & str )

{
string ret = str + str;
return &ret;
}
int main( )
{
string *result = dup( "call" );
cout << *result << endl;
cout << "Now the string is " << *result << endl;
return 0;
}

Figure 3-8

A stale pointer: the pointee, ret, does not exist after dup returns

c++book.mif Page 47 Saturday, July 12, 2003 10:53 AM

Memory Heap Management

1
2
3
4
5

string * dup( const string & str )

{
string *ret = new string( str + str );
return ret;
}

Figure 3-9

1
2
3
4
5
6

47

Safer code, but the caller must call delete or there may be a memory leak

string * dup( const string & str )

{
static string ret;
ret = str + str;
return &ret;
}

Figure 3-10

Using static local variable frees caller from having to call delete

A third option that is sometimes used frees the caller from reclaiming memory. Here, we
return a pointer to a variable that is not allocated from the heap, yet is not allocated on the runtime stack. Two such entities qualify: a global variable, or a static local variable. A static local
variable is essentially the same as a global variable, except that it is only visible inside the function in which it was declared. The static local variable is created once (the first time the function
is invoked), and the variable retains its values between calls to the same function. Thus like a
static class variable in Java, a static local variable could be used to keep track of the number of
times a function has been invoked. Figure 3-10 illustrates this version of dup.
When a function returns a pointer to a static local variable, the caller no longer has to manage memory. However, now the caller must use the return value and, in particular, the object
being pointed at by the return value, prior to making another call to the function. Otherwise, in
the following code fragment,
string *s1 = dup( "hello" );
string *s2 = dup( "world" );
cout << *s1 << " " << *s2 << endl;

the string worldworld is printed twice, because both s1 and s2 are pointing at the same
static object (ret), which is storing the result of the last call to dup. Thus once again, it is
incumbent on the programmer to document that the return value is a pointer to a static variable,
and that the return value must be quickly used.
You should never use delete on an object that was not created by new; if you do, runtime havoc is likely to result. For instance, an attempt to call delete on s1 or s2 in the last
example may lead to a disaster on some C++ implementations. This shows the largest problem
with pointers in C++: when you receive a pointer variable, if the implementation of the function
that is sending you the pointer is hidden (which we would typically expect), unless there are
comments, you have no way of knowing if you are responsible for calling delete, or if you

c++book.mif Page 48 Saturday, July 12, 2003 10:53 AM

48

Chapter 3 Pointers and Reference Variables

should not call delete, and making the wrong decision results in either a memory leak or perhaps an invalid delete. Furthermore, if you have an array of pointers, and the objects being
pointed at were allocated from the memory heap, then you may need to call delete on these
objects when you are done with the array. If any object is being pointed at twice, the programmer must write extra code to avoid double-deletions.

3.4 The -> Operator

In Figure 3-5, at line 11, we see that if strPtr is a pointer to a string, then *strPtr is a
string and thus the length of the string is (*strPtr).length(). The (leading) parentheses are needed because of precedence rules. However, this use of parentheses is annoying. To
avoid typing the parentheses, C++ provides the -> operator, which accesses members of a
pointed-at class type. Thus strPtr->length() is equivalent to (*strPtr).length(),
but is easier to read and write. In this shorthand, strPtr->length() is the length of the object
strPtr is pointing at.

3.5 Reference Variables

In addition to the pointer type, C++ has the reference type. A reference type is a pointer constant
that is always dereferenced implicitly. A reference type can be viewed as an alias for another
object. For instance, in the following code, count becomes a synonym for a longer, hard-totype variable:
int longVariableName = 0;
int & count = longVariableName;
count += 3;

Reference variables must be initialized when they are declared and cannot be changed to
reference another variable. This is because an attempted reassignment via
count = someOtherObject;

assigns to the object longVariableName the value of someOtherObject. This example

is a poor use of reference variables but accurately reflects how they are used in a more general
setting in which the scope of the reference variable is different than that of the object being referenced. One important case is that a reference variable can be used as a formal parameter,
which acts as an alias for an actual argument. We have previously discussed this in the context of
using call-by-reference and call-by-constant reference passing vectors (Section 2.3).
Reference variables are like pointer constants in that the value they store is the address of
the object they refer to. They are different in that an automatic invisible dereference operator is
applied to the reference variable.
Many C++ libraries are based on functions that were written in C, where reference variables are not available. In C, pointer variables are used to achieve call-by-reference, and C++
programs are likely to run into older code that uses this technique. We discuss use pointers to
simulate call-by-reference semantics in Section 12.3.

c++book.mif Page 49 Saturday, July 12, 2003 10:53 AM

Uses and Abuses of Pointer Variables

49

Using C++ reference variables instead of pointer variables translates into a notational convenience, especially because it allows parameters to be passed by reference without the excess
baggage of the & operator on the actual arguments and the * operator that tends to clutter up C
programs.
By the way, pointers can be passed by reference. This is used to allow a function to change
where a pointer, passed as a parameter, is pointing at. A pointer that is passed using call-byvalue cannot be changed to point to a new location (because the formal parameter stores only a
copy of the where value).
Because a reference variable must be initialized at the moment it is declared, it is illegal to
have an array of reference variables.
Finally, we mention that a function can return by reference (or constant reference) in order
to possibly avoid the overhead of a copy. In such a case, the expression in the return statement
must be an object whose lifetime extends past the end of the function. In other words, an object
allocated on the runtime stack should not be returned by reference. We will discuss returning by
reference in Chapter 4. Returning by reference (or constant reference) can sometimes make the
program more efficient by avoiding the overhead of a copy. But doing so is fairly tricky, and
modern optimizing C++ compilers have ways of avoid the overhead of a copy anyway. As a
result, except for a few standard places where the return by reference idiom is used, we do not
recommend it.

3.6 Uses and Abuses of Pointer Variables

Pointer variables are present in much C++ code, and since the vast majority of C++ problems
can be traced to memory management issues, it is worth avoiding their use whenever possible.
The section describes some of the typical occurrences of pointer variables, mentioning a few
instances in which reasonable alternatives exist.
3.6.1

Parameter Passing: Simulating Call-By-Reference

As we have mentioned, libraries that are written to be compatible with C will use pointer variables to achieve either call-by-reference or call-by-constant reference. The idiom is discussed in
Section 12.3. In newly written C++ code, you should not need to use pointers for this purpose.
3.6.2

Arrays and Strings

Primitive arrays and primitive strings in C++ are implemented using pointer variables. This
implementation is a remnant of C, and has been made obsolete by the vector and string
library types.
For strings, the primitive variable type that you will often see is char * (or const
char *, for strings that are not to be changed). Here the pointer variable points at the first character in the string, and the characters are all stored in an array, with a special character \0 signifying the end of the string (this character would not be included in determining the string
length). If you need a primitive string (e.g. as a filename to perform I/O), use the string class

c++book.mif Page 50 Saturday, July 12, 2003 10:53 AM

50

Chapter 3 Pointers and Reference Variables

to construct the string, and then invoke the c_str member function from the string class
to obtain the primitive string.
Other than this, and command line-arguments (Chapter 12) you should not need to use
pointer variables for arrays and strings, but if you must interact with legacy code, see
Chapter 11.
3.6.3

Avoiding Large Data Movements

If presented with an array of objects that need to be rearranged often, you may find it better to
follow the Java style and store an array of pointers to objects, rather than the array of objects.
You may be able to avoid using new, depending on the particular situation by allocating the
objects in a permanent array, and then using the addresses of those objects. If you cannot guarantee the lifetime of your objects, then you will need allocate them from the memory heap using
new.
3.6.4

Linked Data Structures

Classic linked data structures, such as linked lists and binary search trees use pointer variables.
Since these data structures are typically implemented as a class, with proper design we can
lessen the chances of memory problems with reasonable care. We will discuss implementation
of linked data structures later in the text in the context of classes and class templates.
3.6.5

Inheritance

Any serious use of inheritance will require pointer variables and heap memory allocation. The
code will look much like Java code, except that the programmer will have the significant burden
of reclaiming inactive objects. This issue is unavoidable and is a major reason why many people
claim Java is an easier-to-use language for object-oriented programming than C++. We will discuss this in more detail in Chapter 6.

3.7 Key Points

In C++, objects can be allocated both on the runtime stack and on the memory heap.
Objects allocated from the runtime stack are automatically reclaimed when the block in
which they were allocated terminates.
A pointer variable stores the address where another variable resides. A C++ pointer variable has similar semantics (especially =, ==, !=, and NULL) as the Java reference variable.
Objects allocated from the memory heap must eventually be returned back to the memory
heap when they are no longer needed. Otherwise, it is possible to create memory leaks that
may eventually cause your program to run out of memory. An object is returned to the
memory heap by calling delete, with the address of the object.
Never call delete on an object that was not allocated by new.

c++book.mif Page 51 Saturday, July 12, 2003 10:53 AM

Exercises

51

After calling delete on an object, all pointers to that object become stale and should not
be used.
Never attempt to delete an object twice.
A static local variable is created once per program, but the variable is only visible from
inside the function in which it is declared. Each invocation of the function reuses the same
variable, and its value is retained between function invocations.
The -> operator is used to access members of a pointed at class type.
A reference variable in C++ is a pointer constant that is always dereferenced implicitly.
One can view it is a synonym for another name of an object. Reference variables must be
initialized when they are declared and cannot be changed to reference other variables.
Arrays of reference variables are illegal.
Using pointer variables to simulate call-by-reference, or to implement arrays or strings
should be avoided if possible.
Pointer variables will be used to avoid large data moves, implementing linked data structures, and especially in inheritance.

3.8 Exercises
1. How is the C++ memory model different from the Java memory model?
2. What is a pointer variable?
3. What is a stale pointer?
4. What is a memory leak?
5. Which objects have their memory automatically reclaimed?
6. What does the delete operator do?
7. What happens if delete is invoked twice on the same object?
8. What happens if delete is invoked on an object that is not heap-allocated?
9. If a function returns a pointer to an object, why cant the object be a stack-based local variable?
10. If a function returns a pointer to static data, what must the user be sure to do?
11. If a function returns a pointer to heap-allocated data, what must the user be sure to do?
12. What does the -> operator do?
13. What is a reference variable in C++?
14. Why must C++ reference variables be initialized when declared?
15. Consider
int a, b;
int *ptr;
int *ptrptr;
ptr = &a;
ptrptr = &ptr;

c++book.mif Page 52 Saturday, July 12, 2003 10:53 AM

52

Chapter 3 Pointers and Reference Variables

a. Is this legal?
b. What are the values of *ptr and **ptrptr?
c. Is ptrptr=ptr legal?
16. Is *&x always equal to x? If not, give an example.
17. Is &*x always equal to x? If not, give an example.
18. For the declaration
int a = 5;
int *ptr = &a;

What are the values of the following?

a. ptr
b. *ptr
c. ptr == a
d. ptr == &a
e. &ptr
f. *a
g. *&a
h. **&ptr

c++book.mif Page 53 Saturday, July 12, 2003 10:53 AM

H A P T E R

Object-Based
Programming: Classes

IKE Java, C++ uses the class to support object-based

programming, and uses inheritance to support object-oriented programming. Classes in C++
operate in much the same way as in Java, except that there is significant additional syntax
because of C++s attempt to make class types look exactly like primitive types.
In this chapter, we begin by examining the similarities, and then we look at a host of
important syntactic constructs found in the implementation of classes in C++. This chapter does
not discuss two interesting topics: operator overloading (Chapter 5) and inheritance (Chapter 6).
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

public class IntCell

{
public IntCell( )
{ this( 0 ); }
public IntCell( int initialValue )
{ storedValue = initialValue; }
public int getValue( )
{ return storedValue; }
public void setValue( int val )
{ storedValue = val; }
private int storedValue;
}

Figure 4-1

Java class that stores an int value

53

c++book.mif Page 54 Saturday, July 12, 2003 10:53 AM

54

Chapter 4 Object-Based Programming: Classes

4.1 Similarities, Modulo Syntax

We begin by rewriting the Java class shown in Figure 4-1. This class wraps an int variable and
allows access to the wrapped int via getValue and setValue. The C++ equivalent of this
class is shown in Figure 4-2.
In C++, top-level classes cannot be declared with any visibility specifier. Also notice that
the class declaration in C++ ends with a semicolon. Forgetting the semicolon is a common error
that almost always produces an error message, but the message can be hard to decipher, and if
the class is declared in a .h file, the error might point to the .cpp file that includes this .h file.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

class IntCell
{
public:
IntCell( int initialValue = 0 )
{ storedValue = initialValue; }
int getValue( )
{ return storedValue; }
void setValue( int val )
{ storedValue = val; }
private:
int storedValue;
};

Figure 4-2

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

Initial version of C++ class that stores an int value (needs more work)

int main( )
{
IntCell m1;
IntCell m2 = 37;
IntCell m3( 55 );
cout << m1.getValue( ) << " " << m2.getValue( )
<< " " << m3.getValue( ) << endl;
m1 = m2;
m2.setValue( 40 );
cout << m1.getValue( ) << " " << m2.getValue( ) << endl;
return 0
}

Figure 4-3

Example showing the use of IntCell

c++book.mif Page 55 Saturday, July 12, 2003 10:53 AM

Accessors vs. Mutators

1
2
3
4
5
6
7
8

55

bool containsZero( const vector<IntCell> & arr )

{
for( int i = 0; i < arr.size( ); i++ )
if( arr[ i ].getValue( ) == 0 )
return true;
return false;
}

Figure 4-4

Routine to return true if a vector of IntCells contains at least one zero. Not
compatible with original version of IntCell

Two other difference are immediate. First, instead of supplying a visibility modifier for
each member, we simply provide visibility modifiers for sections of the class. Thus in Figure 42, the constructors and methods are public, while the data is private. There is no package visible
specifier in C++. In a C++ class, visibility is private until a public modifier is seen.
Second, in Java, constructors can invoke each other using a call to this. In C++, this is
not allowed. Instead, there are two typical alternatives. The first alternative, shown in Figure 42, is to use default parameters in the constructor. Thus an IntCell can be constructed with
either an int or no parameters. The default parameter of 0 signals that if no parameter is provided, the parameter defaults to 0. Another alternative that works if the constructors are too different to be expressed with default parameters is to declare a private member function that can
be invoked by all the constructors. (Although the initialization routine can be a public member
function, like Java, it is best to avoid doing so, because of considerations that come into play
with inheritance). Default parameters must be compile-time constants.
Figure 4-3 illustrates the use of the IntCell class. The class declaration of IntCell
must be placed prior to using the IntCell type name. Typically, one would place it in a .h
file, and the main program would use an include directive. Observe that in main we create
IntCell objects on the runtime stack, using both no parameters and one parameter constructors. Also observe that if the constructor accepts parameters, they can be placed in parentheses
(if there are two or more parameters, they must be placed in parentheses). Line 9 shows that
objects can be copied. As a result of this statement, the contents of the IntCell object m2 are
copied into IntCell object m1. This would be the equivalent of cloning in Java. Thus the second output statement prints 37 and then 40.

4.2 Accessors vs. Mutators

The IntCell example in Figure 4-2 is relatively simple, yet by itself it shows some complications. For instance, suppose we have a vector of IntCell objects, and we want to count
how many of these objects are storing the value 0. The code shown in Figure 4-4, which should
be straightforward, does not compile.
This code does not compile since arr is supposed to be constant and the compiler is
unsure that invoking the getValue method on one of the array elements wont cause a change

c++book.mif Page 56 Saturday, July 12, 2003 10:53 AM

56

Chapter 4 Object-Based Programming: Classes

to one of the array elements, and thus the array. Though the compiler could look at the implementation of getValue and see that no changes are made to the IntCell, typically this
implementation might not be available in C++ form (in the most general case, it may be invoking other member functions that are already compiled into a library). Thus we need some syntax
to tell the compiler that getValue is not going to change the state of the IntCell.
A member function that looks at an object but promises not to changed the state of the
object is known as an accessor. A member function that might change the state of the object is
known as a mutator. In the IntCell class, getValue is logically an accessor, and
setValue is logically a mutator.
In C++, a member function is assumed to be a mutator unless it is explicitly marked as an
accessor. To mark a member function as an accessor, we place a const at the end of its signature as shown:
int getValue( ) const
{ return storedValue; }

Whether a member function is an accessor or a mutator is considered part of its signature.

When invoking a member function on a constant object, only the accessors are candidates; as
weve already seen, attempts to invoke the mutator fail.
Many less-experienced C++ programmers view the const as an annoyance, or little more
than a comment. In fact, deciding whether a member function is an accessor or a mutator is an
important part of the class design, the const that signifies an accessor should never be omitted.
If it is, you wind up having to remove the const in other places, making your code less readable and robust. The ability to mark an object as constant, and expect that this objects state not
change (at least in the context of a method) allows the compiler to perform aggressive optimization and avoids the need for locking in a multithreaded environment in which several threads
access the object simultaneously. Experienced C++ programmers who move to Java often find
the lack of Java support for immutability (the ability to mark the state of an object as unchangeable) to be a liability.
If a member function is marked as an accessor, then the compiler will not allow the member function to compile if it attempts to make any changes to the values of its data members.
This includes assignment to the data members as well as invoking other mutators of the class
(since invoking the mutator could change the data members).
Occasionally an accessor needs to change the values of one of its data members. For
instance, if a data member is being used to count the number of calls to getValue, we might
want to adjust it. This would not normally be allowed if getValue is an accessor, but one can
mark a data member as mutable to exempt it from the normal restrictions.

4.3 The explicit Keyword

To see the second problem with the IntCell class, consider the code in Figure 4-5. Here at line
5, we see an assignment statement that would never be allowed in Java, and that looks a bit dubious: we are copying an int into an IntCell, in spite of the fact that they have different types.

c++book.mif Page 57 Saturday, July 12, 2003 10:53 AM

The explicit Keyword

1
2
3
4
5
6
7
8
9

57

int main( )
{
IntCell m;
m = 3;
cout << m.getValue( ) << endl;
return 0;
}

Figure 4-5

Example of implicit type conversion, which may or may not be desirable depending on context.

However this code compiles. The reason for this is that C++s type compatibility rules are
somewhat lenient in places. The thinking of the compiler is that since m is an IntCell, it
should (by default) expect to see an IntCell on the right-hand side of the assignment operator.
What it sees is not an IntCell. However, the compiler is willing to do a type conversion. So
the question it must answer is whether it can fabricate a temporary variable of type IntCell in
place of the 3. That requires constructing the temporary, and since there is an IntCell constructor that takes an int, the compiler will use that constructor to fabricate the temporary and
then copy that value into m. Whether this is a good idea or not depends on the application. For
instance, in a class RationalNumber, if there is a constructor that takes a single int, then a
statement such as r=0, where r is of type RationalNumber would be convenient, and would
in fact compile.
On the other hand, since the vector type has a one parameter constructor, this would
open the door for allowing:
vector<double> arr(20);
...
arr = 5;

which would create a temporary vector of size 5, copy its contents into arr, and then free up the
temporary, instead of doing the sensible thing and reporting an error.
Faced with this dilemma, C++ adopts the following rule: A one-parameter constructor
automatically implies the ability to perform a type conversion. Further, by default the compiler
will apply the type conversion if needed, even if not requested, to satisfy assignments and
parameter matches (but not call-by-reference, which requires exact matches). This is known as
an implicit type conversion. The reason for this is that C++ is trying to treat objects in the same
fashion as it treats primitives. To disallow implicit type conversions, the programmer can mark
the constructor as explicit.
The explicit directive is meaningless for constructors that do not take one parameter. If a
constructor is marked as explicit, then it will not be considered in object creations in which
the initialization uses =; instead you must explicitly place the one parameter in parentheses.

c++book.mif Page 58 Saturday, July 12, 2003 10:53 AM

58

Chapter 4 Object-Based Programming: Classes

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

class IntCell
{
public:
explicit IntCell( int initialValue = 0 )
{ storedValue = initialValue; }
int getValue( ) const
{ return storedValue; }
void setValue( int val )
{ storedValue = val; }
private:
int storedValue;
};

Figure 4-6

Second version of C++ class that stores an int value (still needs more work)

What this means is that if IntCell is marked explicit, then in Figure 4-3, the declaration
of m1 and m3 are acceptable, but the declaration of m2 is in error.
As a general rule, one parameter constructors should be marked explicit unless it
makes sense to allow implicit type conversions. Thus in the vector type, the single parameter
constructor has been marked explicit. In the string type, in which a const char *
(primitive string constant) can be passed as a parameter, the constructor is not marked as
explicit to facilitate the mixing of const char * and string library types.
The revised version of IntCell, with the changes made to signify that getValue is an
accessor and the constructor is marked explicit is shown in Figure 4-6.

4.4 Why C++ Needs More Syntax Than Java

Already we have seen two examples where C++ adds syntax that is not needed at all in Java.
Both examples are similar in that the syntax is added in order to allow C++ class types to behave
like primitive types, which is in marked contrast to the design decisions in Java.
In Java, primitive entities can be marked final, and certain operations are not allowed
on them (assignments, autoincrement, etc.). But it is impossible to mark an object as final. Only
the reference can be marked final, but that does not prohibit changes to the state of the object.
The only way to enforce object immutability in Java is to declare data private, and not provide
any mutators. Thus Java is not concerned with extending the immutability concept from primitive entities to objects. C++ is, but this adds syntax.
In Java, a typecast for reference variables does not create a new object; instead it simply
allows the object to be accessed by a different, yet compatible, reference variable. Since objects
are accessed by reference variables, creating new objects implicitly is never needed for nonprimitive entities.

c++book.mif Page 59 Saturday, July 12, 2003 10:53 AM

Why C++ Needs More Syntax Than Java

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

59

// Java version: see Figure 4-8

class Student
{
private String name;
private Date
birthDate;
private double gpa;
}
// C++ version: see Figure 4-9
class Student
{
private:
string name;
Date
birthDate;
double gpa;
};
// C++ version with non-primitive members accessed by pointers
// Same memory layout as Java version: see Figure 4-8
class Student
{
private:
string * name;
Date
* birthDate;
double
gpa;
};

Figure 4-7

Data members in a Java class and two C++ classes

When a class includes data members that are not primitive entities, a whole new set of
considerations are introduced. For example, the code in Figure 4-7 shows a Java class with three

name

"Jane"

birthDate

Jan 1, 1984

gpa

Figure 4-8

3.9

Memory layout in Java, and also in C++ when data members include pointers

c++book.mif Page 60 Saturday, July 12, 2003 10:53 AM

60

Chapter 4 Object-Based Programming: Classes

name

"Jane"

birthDate

Jan 1, 1984

gpa

3.9

Figure 4-9

Memory layout in C++, when objects are declared as non-pointer data members

private data members, the syntactically similar C++ class (which has different semantics), and a
second C++ class whose semantics more closely mirror the Java class.
First, Figure 4-8 shows the layout in the Virtual Machine for Java objects of type
Student. Since name and birthDate are objects, they are accessed by reference variables.
The intent of the picture is that the data fields that are stored as part of a Student instance are
simply pointer variables.
Figure 4-9 shows the layout in C++ for the first declaration of type Student. The declaration is cosmetically identical to the Java declaration, but as is evident from Figure 4-9, complete instances of the data members are stored as part of the Student entity. We can expect that
a copy of Student objects copies all the data members, and this looks good in C++, except of
course, that the copy can be expensive if the data members are large. Recall also that by default
parameter passing and returning makes copies, which is why it can be important to pass objects
using call-by-constant reference instead of call-by-value.
Figure 4-9 shows one new issue that C++ must deal with. In Java, in the constructor, each
of the reference data members is initialized to null, and then assigned to reference an object of
the appropriate type. In C++, the first step of initializing members to null does not work. The
alternative, initializing each member with a default constructor might work, but there are limitations. For instance, there might not be a zero-parameter constructor for type Date. Thus, we
need some syntax to specify how each of the data members is initialized. We discuss this in
Section 4.5.
Although the first piece of C++ in Figure 4-7 looks similar to Java code, we have seen that
it is quite different semantically. Further, often it is the case that the data members are pointer
variables; as we mentioned in Section 3.6, this may be needed in linked data structures and programs that involve inheritance. The last piece of code in Figure 4-7 shows the data members
declared as pointer variables. With this declaration, the memory layout mirrors the Java layout.
However, this layout creates numerous subtleties.
First, a copy is no longer a real copy; when data members are copied, the result is that the
birthDate and name are shared amongst two instances of Student. This is known as a

c++book.mif Page 61 Saturday, July 12, 2003 10:53 AM

Initializer Lists

61

shallow copy. Typically with C++, we expect a deep copy. Certainly the visible semantics of
C++ should not depend on the implementation details, and the default visible semantics that we
would expect to see should be a deep copy. Thus when data members are pointer variables, if we
expect deep copy semantics, we must redefine the assignment operation to ensure correct behavior. We discuss this in Section 4.6.
A related problem concerns how objects of type Student are returned. By default, a
copy is made, but the copy creates a new temporary object. This is not an assignment operator,
but instead a different kind of constructor, known as the copy constructor. We discuss this in
Section 4.6. The default can be time-consuming, so in some cases it is worth trying to avoid it by
returning using constant reference. We discuss this in Section 4.10.
Another related problem is that in the first implementation, when objects of type
Student are created and destroyed, all of its constituent components, which are part of
Student are automatically destroyed too. In the second implementation this is no longer true
by default, because name and birthDate are simply pointer variables. The objects they are
pointing at are not reclaimed unless delete is applied to them; and as these are private variables it is difficult for the user to do so. Instead we need to provide a routine, called the destructor, that ensures that private heap objects allocated by the Student class are reclaimed when
the Student is itself reclaimed. We discuss this in Section 4.6.
Another complication deals with the fact that in Java, the language consists of both the
Java Language Specification and the Java Virtual Machine Specification. Specifically, a Java
compiler can look at Java bytecode to decide the valid methods for a class. The C++ setup places
the class declaration in a .h file, whose contents are logically copied into all source files that
provide an appropriate include directive. This can make compilation slow. In C++, we can specify the declaration of the class, which lists its member functions and memory layout, and provide an implementation separately, thus reducing the size of the .h file. We discuss this in
Sections 4.11 and 4.12.

4.5 Initializer Lists

The initializer list is the C++ construct that is used to specify non-default initialization of each
data member in an object. As an example, the initializer list can be used to rewrite the IntCell
constructor in Figure 4-6 as
explicit IntCell( int initialValue = 0 )
: storedValue( initialValue )
{ }

Syntactically, the initializer list is part of the constructor and appears prior to the opening
brace that signifies the body of the constructor. Although the behavior appears the same, the
revised constructor shown above is different from the original.
In the new form, in which an initializer list is used, the storedValue data member is
immediately initialized to using initialValue when the memory for it is set aside in the

c++book.mif Page 62 Saturday, July 12, 2003 10:53 AM

62

Chapter 4 Object-Based Programming: Classes

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

// Inferior constructor: uses copy assignment

class Student
{
public:
Student( const string & n, const Date & b, double g )
{
name = n;
birthDate = b;
gpa = g;
}
private:
string name;
Date
birthDate;
double gpa;
};

Figure 4-10

Inferior constructor uses copy assignment

newly constructed IntCell object. Because the body of the constructor is empty, no further
operations are performed.
The original form behaves slightly differently. When the memory for storedValue is
set aside for the newly constructed IntCell object, storedValue is initialized by using the
default initialization for int. Then in the body of the constructor, initialValue is copied
into storedValue using the copy assignment operator (for int).
The difference between these two is minor when data members are primitives. However,
when data members are not primitives, failing to use initializer lists can cause inferior performance, and may even cause code to not compile.
1
2
3
4
5
6
7
8
9
10
11
12
13
14

// Better version: uses initializer lists

class Student
{
public:
Student( const string & n, const Date & b, double g )
: name( n ), birthDate( b ), gpa( g )
{
}
private:
string name;
Date
birthDate;
double gpa;
};

Figure 4-11

Constructor uses initializer list

c++book.mif Page 63 Saturday, July 12, 2003 10:53 AM

Initializer Lists

63

Consider, for instance, two alternatives for a Student class constructor. The first version
is shown in Figure 4-10 without an initializer list, while an improved version is shown in
Figure 4-11 with an initializer list.
Consider the initialization of the birthDate data member. In this instance, using an initializer list allows us to initialize the Date data member directly. Without an initializer list, one
must first create a default Date. Since a default Date must represent a valid Date, this Date
may well be initialized to the current Date (today); it certainly will not be some random memory values. Then in the body of the constructor, the intended Date must be copied into to the
Date data member, overwriting the initialized state. Obviously this means that we have wasted
CPU cycles in initializing the Date to todays date. Depending on how complex the Date class
itself is, and how often Student objects are constructed, this could be nontrivial, as it could
involve creation of strings to store months, and so on.
Because initialization of each class member should usually be done with its own constructor, when possible you should use explicit initializer lists. Note however, that this form is
intended for relatively simple cases only. If the initialization is not simple (e.g. if error checks
are needed, or the initialization of one data member depends on another), perhaps the body of
the constructor should be used for more complex logic.
It is important to note that the order of evaluation of the initializer list is given by the order
in which class members are listed. This is one reason why it is bad style to have the initialization
of a data member depend on another data member. If your code depends on the order of initialization, its probably dubious code, and should be avoided; if this is impossible, at least comment the fact that there is an order dependency, so a future programmer does not change the
order of the data members.
An initializer list is required in four common situations.
1. If any data member does not have a zero-parameter constructor, the data member must be
initialized in the initializer list.
2. If a superclass does not have a constructor, the subclass must use an initializer list to initialize the inherited component. (Chapter 6). One can view this as being the same as the
first situation.
3. Constant data members must be initialized in the initializer list. A constant data member
can never be altered after the data member is constructed. This means you could not apply
the copy assignment operator in the body of the class constructor to set the value of the
constant data member. An example of a constant data member could be the identification
number in a Student class. Each Student has his or own unique identification number, but presumably the identification number never changes. This is similar to final data
members in Java, except that in Java, it is the reference variable that is final, not the
object. In C++ it is the data member.
4. A data member that is a reference variable (for instance an ostream &) must be initialized in the constructor.

c++book.mif Page 64 Saturday, July 12, 2003 10:53 AM

64

Chapter 4 Object-Based Programming: Classes

4.6 The Big Three: Destructor, Copy Constructor, and operator=

In C++, classes come with three special functions that are already written for you. These are the
destructor, copy constructor, and operator=, collectively known as the Big-Three. In many
cases you can accept the default behavior provided by the compiler. Sometimes you cannot.
4.6.1

Destructor

The destructor is called whenever an object goes out of scope or is subjected to a delete. Typically, the only responsibility of the destructor is to free up any resources that were allocated
during the use of the object. This includes calling delete for any corresponding news, closing
any files that were opened, and so on. The default applies the destructor on each data member.
4.6.2

Copy Constructor

There is a special constructor that is required to construct a new object, initialized to a copy of
the same type of object. This is the copy constructor. For any object, such as an IntCell
object, a copy constructor is called in the following instances:
a declaration with initialization, such as
IntCell copy = original;
IntCell copy( original );

but not
copy = original; // Assignment operator, discussed later

an object passed using call-by-value (instead of by & or const &), which, as mentioned
in Section 2.3, should usually not be done anyway for large objects.
an object returned by value (instead of by & or const &)
The first case is the simplest to understand because the constructed objects were explicitly
requested. The second and third cases construct temporary objects that are never seen by the
user. In both cases we are constructing new objects as copies of existing objects, so certainly the
copy constructor is applicable.
By default the copy constructor is implemented by applying copy constructors to each
data member in turn. For data members that are primitive types (for instance, int, double, or
pointers), simple assignment is done. This would be the case for the storedValue data member in our IntCell class. For data members that are themselves class objects, the copy constructor for each data members class is applied to that data member.
4.6.3

operator=

The copy assignment operator, operator=, is called when = is applied to two already-constructed objects. lhs=rhs is intended to copy the state of rhs into lhs. By default
operator= is implemented by applying operator= to each data member in turn.

c++book.mif Page 65 Saturday, July 12, 2003 10:53 AM

The Big Three: Destructor, Copy Constructor, and operator=

4.6.4

65

Issues Concerning the Defaults

If we examine the IntCell class, we see that the defaults are perfectly acceptable, and so we
do not have to do anything. This is often the case. If a class consists of data members that are
exclusively primitive types and objects for which the defaults make sense, the class defaults will
usually make sense. Thus a class whose data members are int, double, vector<int>,
string, and even vector<string> can accept the defaults.
The main problem occurs in a class that contains a data member that is a pointer. Suppose
the class contains a single data member that is a pointer. This pointer points at a heap-allocated
object. The default destructor for pointers does nothing (recall that we must delete ourselves),
so there may be a memory leak if the default is used. Furthermore, the copy constructor and
operator= both copy not the objects being pointed at, but simply the value of the pointer.
Thus we will simply have two class instances that contain pointers that point to the same object.
As discussed in Section 4.4, this is a so-called shallow copy. Typically, we would expect a deep
copy, in which a clone of the entire object is made. Thus, when a class contains pointers as data
members, and deep semantics are important, we typically must implement the destructor,
operator=, and copy constructor (in other words, the Big-Three) ourselves.
For IntCell, the signatures of these operations are:
~IntCell( );
// destructor
IntCell( const IntCell & rhs );
// copy constructor
const IntCell & operator=( const IntCell & rhs );

Although the defaults for IntCell are acceptable, we can write the implementations
anyway, as shown in Figure 4-12. For the destructor, after the body is executed, the destructors
are called for the data members. So the default is an empty body. For the copy constructor, the
default is an initializer list of copy constructors, followed by execution of the body.
operator= is the most interesting. Line 18 is an alias test, to make sure we are not copying to ourselves. Assuming we are not, we apply operator= to each data member (at line 19).
We then return a reference to the current object, at line 20, so assignments can be chained, as in
a=b=c. (The return is actually a constant reference so that the nonsensical (a=b)=c is disallowed by the compiler; however, it seems that returning a reference instead of a constant reference is more common, and is the default).
Let us look at the uses of the keyword this in more detail. In C++, the pointer this is
defined to point at the current object in exactly the same way as this references the current
object in Java. Consequently, *this is the current object, and returning *this achieves the
desired result. As in Java, under no circumstances will the compiler knowingly allow you to
modify this. As we see, the return at line 20 uses *this. The other use of this is at line 18.
The expression a=a is logically a non-operation (no-op). In some cases, although not
here, failing to treat this as a special case can result in the destruction of a. As an example, consider a program that copies one file to another. A normal algorithm begins by truncating the target file to zero length. If no check is performed to make sure the source and target file are indeed

c++book.mif Page 66 Saturday, July 12, 2003 10:53 AM

66

Chapter 4 Object-Based Programming: Classes

different, then the source file will be truncated, hardly a desirable feature. When performing
copies, the first thing we should do is check for this special case, known as aliasing.
In the routines that we write, if the defaults make sense, we will always accept them. However, if the defaults do not make sense, we will need to implement the destructor, and
operator=, and the copy constructor. When the default does not work, the copy constructor
can generally be implemented by mimicking normal construction and then calling operator=.
Another often-used option is to give a reasonable working implementation of the copy constructor, but then place it in the private section, to disallow call-by-value.
4.6.5

When the Defaults Do Not Work

The most common situation in which the defaults do not work occurs when a data member is a
pointer type, and the pointee is heap-allocated by some object member function (such as the
constructor). As an example, suppose we implement the IntCell by dynamically allocating an
int, as shown in Figure 4-13.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27

class IntCell
{
public:
~IntCell( )
{
// Does nothing since IntCell contains only an int data
// member. If IntCell contained any class objects their
// destructors would be called.
}
IntCell( const IntCell & rhs )
: storedValue( rhs.storedValue )
{
}
IntCell & IntCell::operator=( const IntCell & rhs )
{
if( this != &rhs )
// Standard alias test
storedValue = rhs.storedValue;
return *this;
}
...
private:
int storedValue;
};

Figure 4-12

The defaults that are generated automatically be the compiler

c++book.mif Page 67 Saturday, July 12, 2003 10:53 AM

The Big Three: Destructor, Copy Constructor, and operator=

1
2
3
4
5
6
7
8
9
10
11
12
13

class IntCell
{
public:
explicit IntCell( int initialValue = 0 )
{ storedValue = new int( initialValue ); }
int getValue( ) const
{ return *storedValue; }
void setValue( int val )
{ *storedValue = val; }
private:
int *storedValue;
};

Figure 4-13

1
2
3
4
5
6
7
8
9
10
11
12
13

67

Data member is a pointer; defaults are no good

int f( )
{
IntCell a( 2 );
IntCell b = a;
IntCell c;
c = b;
a.setValue( 4 );
cout << a.getValue( ) << endl << b.getValue( ) << endl
<< c.getValue( ) << endl;
return 0;
}

Figure 4-14

Simple function that exposes problems in Figure 4-13

There are now numerous problems that are exposed in Figure 4-14. First, the output is
three 4s, even though logically only a should be 4. The problem is that the default copy constructor and operator= copy the pointer storedValue. Thus a.storedValue,
b.storedValue, and c.storedValue all point at the same int value. These copies are
a.storedValue

b.storedValue
c.storedValue
Figure 4-15

After line 5 in Figure 4-14, default copy constructor generates shallow copies

c++book.mif Page 68 Saturday, July 12, 2003 10:53 AM

68

Chapter 4 Object-Based Programming: Classes

shallow: the pointers, rather than the pointees are copied. A second less obvious problem is a
memory leak. The int initially allocated by as constructor remains allocated and needs to be
reclaimed. The int allocated by cs constructor is no longer referenced by any pointer variable.
It also needs to be reclaimed, but we no longer have a pointer to it. These problems are illustrated in Figures 4-15 and 4-16.

a.storedValue

b.storedValue
c.storedValue
Figure 4-16

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27

After line 7 in Figure 4-14, default operator= generates shallow copy, and
leaks memory

class IntCell
{
public:
explicit IntCell( int initialValue = 0 )
{ storedValue = new int( initialValue ); }
IntCell( const IntCell & rhs )
{ storedValue = new int( *rhs.storedValue ); }
~IntCell( )
{ delete storedValue; }
IntCell & operator=( const IntCell & rhs )
{
if( this != &rhs )
*storedValue = *rhs.storedValue;
return *this;
}
int getValue( ) const
{ return *storedValue; }
void setValue( int val )
{ *storedValue = val; }
private:
int *storedValue;
};

Figure 4-17

Data member is a pointer; Big-Three needs to be written.

c++book.mif Page 69 Saturday, July 12, 2003 10:53 AM

The Big Three: Destructor, Copy Constructor, and operator=

69

To fix these problems, we implement the Big-Three. The result is shown in Figure 4-17.
Generally speaking, if a destructor is necessary to reclaim memory, then the defaults for copy
assignment and copy construction are not acceptable.
4.6.6

Linked Data Structures

Linked data structures, such as linked lists and binary search trees provide classic examples in
which the Big-Three need to be written. Although the Standard Library provides implementations of stacks, queues, lists, sets, and maps, you may on occasion find that you need to implement your own. For instance, the linked list (and thus queue) implementations in the Standard
Library use doubly-linked lists, and a singly-linked list can be implemented faster and with less
space. In this section we illustrate a singly-linked list implementation of a queue, with minor
syntactical improvements in sections that follow. Our interest in this implementation is concerned mostly with memory management and the Big-Three.
Recall that a queue supports insertion at one end (the back), and deletion at the other end
(the front). These operations are enqueue and dequeue, respectively. As Figure 4-18 shows,
a singly-linked list in which we store a pointer to both the front and the back of the list can be
used to represent a queue. In an empty queue, the data member front is NULL. In this case, the
linked list contains list nodes that each store the data and a pointer to the next node in the list and
this is implemented in a ListNode class shown in Figure 4-19. The class uses public data
because the data members need to be accessed from the queue class. There are several alternate
solutions, including the use of nested classes, that we will discuss in Sections 4.7 4.9.
Figure 4-20 shows that to enqueue a new integer x, we create a new ListNode and
attach it after the last node in the list, in the process updating back. Figure 4-21 shows that to
dequeue the front item, we simply advance front, after saving the data in the front node so it
can be returned. In Java, when we advance front, the node that was formerly referenced by
front becomes unreferenced and eligible for garbage collection. In C++, we must clean up the
memory ourselves.

front

Figure 4-18

back

Linked list implementation of the queue

c++book.mif Page 70 Saturday, July 12, 2003 10:53 AM

70

Chapter 4 Object-Based Programming: Classes

1
2
3
4
5
6
7
8
9

class ListNode
{
public:
int
element;
ListNode *next;
ListNode( int theElement, ListNode * n = NULL )
: element( theElement ), next( n ) { }
};

Figure 4-19

ListNode class

back

Before

...

back

After

Figure 4-20

...

enqueue operation for linked-list-based implementation

c++book.mif Page 71 Saturday, July 12, 2003 10:53 AM

The Big Three: Destructor, Copy Constructor, and operator=

71

front

Before

...

...

front
After

Figure 4-21

dequeue operation for linked-list-based implementation

In fact, we can now see that several other routines require that we clean up memory: two
obvious candidates are makeEmpty and the destructor, both of which must clean up all the
nodes in the list. Furthermore, if we write a destructor, we must write operator=, and if we
copy a large queue into a small queue, it is evident that nodes will have to be reclaimed.

c++book.mif Page 72 Saturday, July 12, 2003 10:53 AM

72

Chapter 4 Object-Based Programming: Classes

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36

class UnderflowException { };
class IntQueue
{
private:
ListNode *front;
ListNode *back;
public:
IntQueue( ) : front( NULL ), back( NULL )
{ }
IntQueue( const IntQueue & rhs )
: front( NULL ), back( NULL )
{ *this = rhs; }
~IntQueue( )
{ makeEmpty( ); }
const IntQueue & operator= ( const IntQueue & rhs )
{
if( this != &rhs )
{
makeEmpty( );
ListNode *rptr = rhs.front;
for( ; rptr != NULL; rptr = rptr->next )
enqueue( rptr->element );
}
return *this;
}
void makeEmpty( )
{
while( !isEmpty( ) )
dequeue( );
}

Figure 4-22

Linked list based queue class Part I

The implementation of the queue is provided in the IntQueue class that begins in
Figure 4-22. UnderflowException is simply a class that is used if getFront or
dequeue is attempted on an empty queue. We will discuss exceptions in more detail in
Chapter 8, but for now we remark that UnderflowException will behave exactly like a
Java runtime exception that is uncaught. We show the two data members at the top of the
IntQueue class declaration instead of the bottom simply so we can see them in this discussion.
Normally, the private section is placed at the end of a C++ class declaration.

c++book.mif Page 73 Saturday, July 12, 2003 10:53 AM

The Big Three: Destructor, Copy Constructor, and operator=

73

The zero-parameter constructor, shown at lines 10 and 11 does no more than initialize
front and back to NULL; this is NOT done by default. The copy constructor at lines 13 to 15
first makes the queue empty, and then copies rhs into it using operator=. Since
operator= might try to reclaim nodes in this queue prior to the copy, it is important that this
queue be placed into a respectable state prior to invoking the copy assignment operator. The
destructor is shown at lines 17 and 18. Instead of reclaiming the memory itself, it delegates the
dirty work to makeEmpty, which presumably cleans up the memory. The copy assignment
operator operator=, is shown at lines 20 to 30. After the alias test, it empties out this queue,
and then steps through rhs, enqueueing each item it sees. The alias test is crucial here, since
otherwise, a self-assignment makes the queue empty!
makeEmpty is shown at lines 32 to 36. Recall that the destructor was relying on
makeEmpty to reclaim all the list nodes. makeEmpty is implemented simply by calling
dequeue until the queue is empty, thus delegating the dirty work of memory management yet
again.
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63

bool isEmpty( ) const

{ return front == NULL; }
int getFront( ) const
{
if( isEmpty( ) )
throw UnderflowException( );
return front->element;
}
void enqueue( int x )
{
if( isEmpty( ) )
back = front = new ListNode( x );
else
back = back->next = new ListNode( x );
}
int dequeue( )
{
int frontItem = getFront( );
ListNode *old = front;
front = front->next;
delete old;
return frontItem;
}
};

Figure 4-23

Linked-list-based queue class part II

c++book.mif Page 74 Saturday, July 12, 2003 10:53 AM

74

Chapter 4 Object-Based Programming: Classes

isEmpty, getFront, and enqueue are all shown in Figure 4-23, and are relatively
straightforward, (in some cases trivial), as they are only a few lines of code each and do not
involve reclaiming memory. The reclaiming of memory however, can only be deferred for so
long, and finally in dequeue, shown in lines 55 to 62 we have to bite the bullet. As the code
shows, a call at line 57 to getFront gives us the frontItem that can be returned at line 61
(or throws an UnderflowException that is unhandled). Prior to advancing front at line 59,
we save a pointer to it (line 58) so we can reclaim the node (line 60) by invoking the delete
operator. Thus we have managed to funnel all memory management to three lines of code,
which is generally your best strategy, since memory management is so bug-prone.
4.6.7

Default Constructor

If no constructors (not including the special copy constructor) are provided for a class, then a
default constructor is automatically generated. The default takes no parameters. Each data member that is a class type is initialized by its zero-parameter constructor. However the primitive
members of objects allocated on the runtime stack or from the memory heap are not guaranteed
to be initialized.
4.6.8

Disabling Copying

Both operator= and the copy constructor can be disabled (independently) by placing their
declarations in the private section of the class. It is still worthwhile to provide an implementation of these methods. Similarly, as in Java, constructors can be private; as in Java, this is common in the case of factory classes (such as InetAddress) that provide static methods to
create object instances. C++ supports the same idiom. Destructors should generally not be private.
If the class contains data members that do not have the ability to copy themselves, then the
default operator= will not work.

4.7 Friends
Typically we would like to make data members (and occasionally member functions) private.
However as we saw in Section 4.6.6, it often is inconvenient to do so, because there may be one
other class (or possibly a few select classes) that needs access to implementation details. But
making the members public grants access to everyone.
Java has an intermediate visibility modifier, package visibility, in which specific members
are marked as being accessible to other classes. However, while this allows access only to
classes that happen to be in the same package (typically not a severe limitation), the access is
granted to all such classes.
In C++, there is no such visibility modifier. Instead, the class can grant waivers to others
of the normal privacy restrictions. Such a waiver would apply to the access of all of the class
private members. This waiver can granted only by the class that is willing to allow access to its

c++book.mif Page 75 Saturday, July 12, 2003 10:53 AM

Friends

1
2
3
4
5
6
7
8
9
10
11

75

class ListNode
{
private:
int
element;
ListNode *next;
ListNode( int theElement, ListNode * n = NULL )
: element( theElement ), next( n ) { }
friend class IntQueue;
};

Figure 4-24

ListNode class with friend declaration

private members. The recipient of the waiver can be either an entire class, or a specific function.
There is no limit to the number of waivers that a class can grant.
A waiver is known as a friend declaration.
4.7.1

Classes as friends

Figure 4-24 offers an alternative to the ListNode class that was previously declared in
Figure 4-19. Notice that here the class contains private members only, and thus with no friend
declarations, is unusable (even instances cannot be created). The friend declaration at line 10
allows any of the IntQueue member functions to access the private members.
4.7.2

Methods as friends

A class can also grant access to a specific function. For instance, though it makes little sense,
friend int main( );

grants main access to private data.

A member function can be granted access by using the :: operator (we discuss the ::
operator in Section 4.11) to specify the class that it is part of. For instance,
friend void IntQueue::enqueue( int x );

grants access to IntQueues enqueue member function that takes an int. The friend declaration is very specific:
friend int IntQueue::getFront( );

does not grant access to the getFront member function implemented in Figure 4-23, because
the getFront method is an accessor, and the declaration that was provided signals that the
access was granted to a mutator. The friend declaration
friend int IntQueue::getFront( ) const;

fixes that problem.

As a matter of style, just as package visibility is often an inferior design decision, friend
declarations are generally best avoided, as they weaken the encapsulation and information hid-

c++book.mif Page 76 Saturday, July 12, 2003 10:53 AM

76

Chapter 4 Object-Based Programming: Classes

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

class IntQueue
{
public:
...
private:
class ListNode
{
public:
int
element;
ListNode *next;
ListNode( int theElement, ListNode * n = NULL )
: element( theElement ), next( n ) { }
};
ListNode *front;
ListNode *back;
};

Figure 4-25

Nested ListNode class

ing benefits that the private section offers. An alternative approach to our original problem of
granting IntCell access to ListNode internals is to use nested classes, and this is discussed
in the next section.

4.8 Nested Classes

Like Java, C++ supports nested classes. Specifically, the nested class in C++ behaves in a manner that is similar to the static nested class in Java. Figure 4-25 shows that we can hide the
ListNode class inside of the IntCell class.
If the ListNode class was public, outside of IntCell it would be known as
IntCell::ListNode. But the point of the idiom is to hide the class, so it is placed in the private section.
Unlike Java, private members of the nested classes are not visible from instances of the
enclosing class. So the members in ListNode need to be public; however, this is of little consequence, since the ListNode type itself is hidden. Alternatively, the IntCell class could
declare ListNode as a friend, and then ListNode could have private members.
Similarly, unlike Java, private members of the enclosing class are not visible from
instances of the nested class. Here the most sensible way to remedy this limitation would be for
the enclosing class to use a friend declaration to grant access to instances of the nested class.
C++ does not support Java style inner classes in which instances of the inner class are constructed with the hidden reference to an outer object that caused its creation. Thus the C++
nested class is simply used as a convenience of hiding one type inside another type.

c++book.mif Page 77 Saturday, July 12, 2003 10:53 AM

The struct Type

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21

77

const string & findMax( const vector<string> & arr )

{
int maxIndex = 0;
for( int i = 1; i < arr.size( ); i++ )
if( arr[ maxIndex ] < arr[ i ] )
maxIndex = i;
return arr[ maxIndex ];
}
const string & findMaxWrong( const vector<string> & arr )
{
string maxValue = arr[ 0 ];
for( int i = 1; i < arr.size( ); i++ )
if( maxValue < arr[ i ] )
maxValue = arr[ i ];
return maxValue;
}

Figure 4-26

Two versions to find the maximum string (alphabetically); only the first version
is correct

C++ allows local classes in which a class is declared inside a function. However, their utility is dubious because unlike Java, automatic local variables in the enclosing function cannot be
accessed (static local variables can be accessed, but this hardly seems like sufficient justification
to introduce the added complexity of using a local class).
C++ does not allow anonymous classes.

4.9 The struct Type

In C++, the struct type is a relic from the C programming language. A struct in C++ is
essentially a class in which the members default to public. Recall that in a class, the members
default to private. Thus in Figure 4-25 we can use struct instead of class at line 7, and then
omit line 9. Clearly the struct type is not needed, but you will often see it, and it is commonly
used by programmer to signify a type that contains mostly data that is accessed directly, rather
than through member functions.

4.10 Return By Constant Reference Revisited

Let us return to an issue raised earlier about how objects are returned. Recall that the default
return mechanism is return by value. In return by value, the value of the expression is copied
back to the caller. If the return type is an object that is expensive to copy, then this can be undesirable. So there are occasions in which it is worth returning by constant reference, in an attempt

c++book.mif Page 78 Saturday, July 12, 2003 10:53 AM

78

Chapter 4 Object-Based Programming: Classes

to avoid the copy of the return value. But this is tricky, and there are two important issues.
First, as Figure 4-26 shows, if you are returning a constant reference, then the expression
that is being returned must have lifetime that extends past the end of the function. Thus
findMaxWrong is incorrect because it returns a reference to maxValue, and maxValue
does not exist once this function returns. The first implementation is correct, since it is guaranteed that the array item exists when the function returns. Notice that a return expression such as
return arr[ maxIndex ] + "";

does not work. The result of the string concatenation is an unnamed temporary string whose
destructor will be called as soon as the function terminates.
Ensuring that the return expression has long lifetime is only half of the task. Consider the
following three calls to findMax.
string s1 = findMax( arr );
const string & s2 = findMax( arr );
string & s3 = findMax( arr );

// copies
// no copy
// illegal

The first call is legal, but defeats the entire purpose of returning by constant reference.
Specifically, the object s1 is a string, it is being created, its initial value is another string,
so this statement causes execution of a string copy constructor. The similar code
string s1;
s1 = findMax( arr );

is no better; it creates a default string, and then the second line causes execution of the
string copy assignment operator.
The second call, in which we initialize s2 is the correct way to avoid the copy. The declaration says that s2 is a reference variable. Thus, it is not a new string. The initialization
states that s2 references the same string object as the return value of findMax. Since
findMax returns by (constant) reference, its return value references the maximum string that is
actually contained in arr. Thus s2 references the maximum string that is actually contained
in arr.
If findMax returned by value, then this would appear to be dubious code, because s2
would be referencing an unnamed temporary, whose destructor could be called as soon as the
statement terminated. Using s2 at the next line could result in an attempt to access an already
destructed string. However, the language specification has specifically contemplated this scenario and it is guaranteed that the temporary variable will not be destroyed while the reference
variable is active. Even so, in this situation, the return by value causes a copy to create the temporary variable. Thus, to avoid the copy, we must both:
return by (const, if appropriate) reference
have the caller use a (const, if appropriate) reference variable to access the return value
The third call should not compile, because it attempts to throw away the const-ness of the
reference variable returned by findMax. Specifically, findMax was returning a reference that

c++book.mif Page 79 Saturday, July 12, 2003 10:53 AM

Separation of Interface and Implementation

1
2
3
4
5
6
7
8
9
10
11

79

class IntCell
{
public:
explicit IntCell( int initialValue = 0 );
int getValue( ) const;
void setValue( int val );
private:
int storedValue;
};

Figure 4-27

Third version of C++ class that stores an int value, now with class declaration
separate from implementation (still needs more work)

could now be used to modify the referenced object, and if this declaration were to be allowed by
the compiler, then s3 could be used to modify the referenced object, in violation of the expectation of findMax.
However, the const-ness can be cast away using the special const_cast:
string & s3 = const_cast<string &> ( findMax( arr ) );

Now an attempt to change the object that was being referenced compiles (though if the
object was actually stored in read-only-memory, the code may fail at runtime). Needless to say,
using a const_cast to circumvent const-ness is almost always inadvisable and is best
avoided.

4.11 Separation of Interface and Implementation

If a program consists of many separate source files that use a class type, each source file must be
able to see the class declaration. The easiest way to do this is to put the class declaration in a
header file (.h) and then have each .cpp file that uses a class type provide include directives.
Because the include directive causes the contents of an entire file to logically replace the
include directive, two serious downsides emerge:
First, normally, in a separate compilation model, only the files that have changed are
recompiled, but this clearly is not the case here, since header files are never directly compiled
themselves. As a result, any change to the implementation of any of the member functions in a
class requires the recompilation of every file that uses the class.
Second, the recompilation is expensive, because if the class is used in numerous .cpp
source files, the compiler must repeatedly parse the entire contents of the header file.
To solve these problems, C++ allows the separation of the class specification from its
implementation. The specification lists the class and its members (data and functions). The
implementation provides implementations of the member functions. Often the specification by
itself is called a class declaration, while the complete class, with member functions imple-

c++book.mif Page 80 Saturday, July 12, 2003 10:53 AM

80

Chapter 4 Object-Based Programming: Classes

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

#include "IntCell.h"
IntCell::IntCell( int initialValue )
: storedValue( initialValue )
{
}
int IntCell::getValue( ) const
{
return storedValue;
}
void IntCell::setValue( int val )
{
storedValue = val;
}

Figure 4-28

Third version of C++ class that stores an int value; class implementation (final version)

mented, is the class implementation. The class declaration is also known as the class specification, and is sometimes known as the class interface.
Figure 4-27 shows the class declaration for IntCell, which we last saw in Figure 4-6.
The class implementation is shown in Figure 4-28. Observe that we use the :: scoping operator
to signify that we are implementing member functions, rather than plain (non-member) functions. The main program is shown in Figure 4-29.
The same syntax is used whether these member functions are public or private. The signature of the member functions in the implementation file must exactly match the signature in the
class specification, including parameter passing mechanism, return mechanism, and all uses of
const. The names of the formal parameters do not have to match, and in fact, may be omitted
in the class declaration.
It is illegal to provide an implementation of a member function that was not listed in the
class declaration. If a member function is listed in the class declaration, but is not implemented,
the program will still compile. It will link and run if the missing member function is never actually invoked. This allows the programmer to implement and debug the class in stages.
The compilation mechanism is the same as described in Section 2.1.7. The .cpp files
(main.cpp and IntCell.cpp) would be compiled as part of the project, and the .h files
would automatically be handled by the include directives.
If the implementation of a member function in IntCell.cpp changes, then only
IntCell.cpp needs to be recompiled. If the class declaration in IntCell.h changes, then
all .cpp files that have referenced the class declaration must be recompiled.
We do remark that implementing a member function inside the class declaration does have
the advantage that an aggressive compiler can perform inline optimization. Thus, often trivial

c++book.mif Page 81 Saturday, July 12, 2003 10:53 AM

#ifndef and #endif

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

81

#include "IntCell.h"
#include <iostream>
using namespace std;
int main( )
{
IntCell m1;
IntCell m2 = 37;
IntCell m3( 55 );
cout << m1.getValue( ) << " " << m2.getValue( )
<< " " << m3.getValue( ) << endl;
m1 = m2;
m2.setValue( 40 );
cout << m1.getValue( ) << " " << m2.getValue( ) << endl;
return 0;
}

Figure 4-29

Example showing the use of IntCell when class is separated (no changes)

one-liners that are not likely to undergo changes in future versions are implemented in the class
declaration. Most notably, this often includes constructors and destructors.

4.12 #ifndef and #endif

Suppose the code in Figure 4-29 is changed to contain a duplicate of the include directive at line
1. In other words, the altered code begins with:
#include "IntCell.h"
#include "IntCell.h"

Whereas function declarations can be repeated, and the declarations subsequent to the first
are simply ignored, a class declaration cannot be repeated within a single .cpp file. Thus this
revision does not compile.
Of course, nobody would deliberately write this kind of code. Unfortunately, in a complicated project, we often require that some of the .h files have include directives that reference
other .h files. In that case, it is possible that the effect of a series of include directives is to
include the same .h file more than once, and it may be impossible to avoid this scenario.
Figure 4-30 shows the C++ idiom that is used to avoid this problem. This idiom is placed
inside the header file, and the idea is that the compiler should be directed to read the .h file only
once per compilation of a .cpp file. At line 2, prior to reading the class declaration for
IntCell, the preprocessor defines the symbol INTCELL_H. Presumably, this is a unique symbol based on the name of the header file for which no other attempts have ever been made to
define. Thus, at line 1, if the INTCELL_H symbol is NOT defined, it is safe to read the class

c++book.mif Page 82 Saturday, July 12, 2003 10:53 AM

82

Chapter 4 Object-Based Programming: Classes

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

#ifndef INTCELL_H
#define INTCELL_H
class IntCell
{
public:
explicit IntCell( int initialValue = 0 );
int getValue( ) const;
void setValue( int val );
private:
int storedValue;
};
#endif

Figure 4-30

Final version of C++ class that stores an int value, with class declaration separate from implementation, includes ifndef/endif idiom

declaration. So #ifndef is a preprocessor directive and stands for not defined. The #endif
at line 15 closes the body of the #ifndef. As a matter of safe programming, this idiom, which
we refer to as the ifndef/endif idiom, should be used in all header files. Even if there is no class
declaration in the header file, using this idiom allows the compiler to avoid some parsing effort
if the file appears a second time in a chain of include directives.

4.13 Static Members

Like Java, C++ allows static members. A static data member of a class is not part of any
instance, but instead is part of the class. Whether there are 0, 1, or 20 instances of the class there
is exactly one shared instance of the static data member. In this regard static data members in
C++ are identical to their Java counterpart.
Member functions can access static data. If such a function access only static data, and no
instance data, then the member function should be declared static, since even in the case where
there are 0 instances of the class, it still makes sense to call the member function. And clearly if
there are 0 instances of the class, it would be impossible to invoke a non-static member function.
Again, this is exactly the same as in Java.
The main difference between C++ and Java with regard to static members concerns their
initialization. Java has precise rules on when the initialization occurs and the initialization can
be done both inline or in a static initializer block. Generally speaking C++ allows neither.
An example that illustrates both static data members and static member functions is the
Ticket class shown in Figure 4-31. For brevity, we do not separate interface and implementation. Each ticket stores its id, and we would like ids to be unique. One way to do this is to
assign the ids sequentially, keeping track of the number of tickets. When a new ticket is constructed, we increment the number of tickets and use the new number as the ticket id.

c++book.mif Page 83 Saturday, July 12, 2003 10:53 AM

Static Members

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30

83

class Ticket
{
public:
Ticket( ) : id( ++ticketCount )
{ }
int getID( ) const
{ return id; }
static int getTicketCount( )
{ return ticketCount; }
private:
int id;
static int ticketCount;
Ticket( const Ticket & rhs )
{
id = ++ticketCount;
}
const Ticket & operator= ( const Ticket & rhs )
{
if( this != &rhs )
id = ++ticketCount;
return *this;
}
};
int Ticket::ticketCount = 0;

Figure 4-31

// place this line in Ticket.cpp

Static data member defined in implementation file

In this scenario, the ticket id is a data member that is part of each unique ticket instance,
but the ticket count is shared data, and is thus static. ticketCount is declared at line 15 as
being a static variable. Unfortunately, this does not cause the creation of the shared
ticketCount object; this must be provided separately as shown on line 30. Note that this definition cannot be placed in the header file, because if the header file is included in separate
.cpp files, ticketCount will be multiply defined. (The ifndef/endif idiom only avoids multiple reading of the file for each .cpp file, not for separate .cpp files). ticketCount is initialized when it is defined.
The constructor, shown at lines 4 and 5 increments ticketCount and uses it to initialize id. getTicketCount is a static method, since it could theoretically return 0 if invoked
prior to the first creation of a ticket. Also, observe the idiom of disabling the copy constructor
and copy assignment operator at lines 17 to 27 by placing them in the private section. Allowing

c++book.mif Page 84 Saturday, July 12, 2003 10:53 AM

84

Chapter 4 Object-Based Programming: Classes

1
2
3
4
5
6
7
8
9
10

int main( )
{
cout << Ticket::getTicketCount( ) << " tickets" << endl;
Ticket t1, t2, t3;
cout << Ticket::getTicketCount( ) << " tickets" << endl;
cout << "t2 is " << t2.getID( ) << endl;
return 0;
}

Figure 4-32

Invoking static member functions

a real copy would create duplicate tickets with identical ids, but using the implementation that
is provided is not really a copy, so in this case it seems reasonable to disable both forms of copying.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

class MathUtils
{
...
private:
static vector<int> primes;
static bool forceStaticInit;
static bool staticInit( int n )
{
// Use Sieve of Erastothenis to eliminate non-primes
vector<bool> nums( n + 1, true );
for( int i = 2; i * i <= n; i++ )
for( int j = i * 2; j <= n; j += i )
nums[ j ] = false;
for( int k = 2; k <= n; k++ )
if( nums[ k ] )
primes.push_back( k );
return true;
}
};
// In implementation file
vector<int> MathUtils::primes;
bool
MathUtils::forceStaticInit = staticInit( 1000 );

Figure 4-33

Simulating the static initilizer in C++: initialize array with prime numbers less
than or equal to 1000

c++book.mif Page 85 Saturday, July 12, 2003 10:53 AM

Anonymous Objects

1
2
3
4
5
6
7
8
9

85

class Utilities
{
public:
static const int BITS_PER_BYTES = 8;
...
};
// In Utilities.cpp
const int Utilities::BITS_PER_BYTES;

Figure 4-34

Constant static integral members initialized in class declaration

Disabling the copy constructor means, among other things, that a Ticket object may not
be passed using call-by-value, nor returned using return-by-value.
To invoke the getTicketCount member function, we once again use the :: scoping
operator, as shown on both lines 3 and 6 in Figure 4-32.
If a static data member involves a complex initialization, Java would process the initialization in a static initializer block. C++ does not have a static initializer block, but the effect can be
achieved as follows: Define a private static member function staticInit, and use a call to
this static member function to initialize the static data member. If the static data member does
not have appropriate copy semantics that would easily allow this, fabricate a second private
static data member of type bool, and have it invoke staticInit. Inside staticInit, you
can then explicitly initialize all the static data members of the class. An example of this strategy
is shown in Figure 4-33.
Because a static member function does not affect any particular instance of a class, static
member functions are never marked as accessors. Normally, static member functions would be
implemented in the .cpp file along with the non-static member functions.
If a static data member is a constant integer type (int, short, long, char, etc.) its initialization can be performed in the class declaration. However, it must still be defined in the
.cpp file, and the definition must not have any initialization. This hardly seems worth the
effort, except that initializing this way makes the static data member a constant integral expression, and qualifies it to be used as a case in a switch statement, and a few other places where
constant integral expressions are specifically required. Figure 4-34 illustrates the syntax.

4.14 Anonymous Objects

Occasionally, we have need to construct anonymous objects (often temporaries). In that case, the
name of the class, with parameters appropriate for its constructors can be used. For instance,
suppose we wanted to generalize the Student constructor at line 5 in Figure 4-11. Recall that
the constructor is implemented as
Student( const string & n, const Date & b, double g )
: name( n ), birthDate( b ), gpa( g )
{ }

c++book.mif Page 86 Saturday, July 12, 2003 10:53 AM

86

Chapter 4 Object-Based Programming: Classes

Suppose we want to add defaults. The default string is "", the default double is 0.0,
but we need a default Date. Presumably there is an appropriate constructor, so assuming the
existence of a zero-parameter constructor, Date() represents a default date. Thus the constructor is:
Student( const string & n = "", const Date & b = Date( ),
double g = 0.0 )
: name( n ), birthDate( b ), gpa( g )
{ }

4.15 Namespaces
The C++ equivalent of packages is the namespace. To declare a namespace, we simply write
namespace namespaceName
{
...
}

where namespaceName is the name of the namespace. Inside the braces can be functions,
objects, and class declarations. A class ClassName declared in namespace namespaceName
is formally known as namespaceName::ClassName.
As with Java packages, entities inside of a namespace can be accessed without specifying
the namespace. Like packages, namespaces are open-ended, so one can have several separate
namespace declarations.
As in Java, it can be inconvenient to write the complete class name, that includes the
namespace name. A using directive is the equivalent of an import directive. The first form,
using namespaceName::ClassName;

allows ClassName to be used as a shorthand for namespaceName::ClassName. The second form,

using namespace namespaceName;

is the equivalent of the Java wild-card import directive (that ends .*), and allows all entities in
the namespace namespaceName to be known by their shorthands. (Older compilers handle
this wild-card form better than the more specific using directive above, so C++ code tends to use
the wild-card using directive).
In Java, classes can be declared as public or package visible. No such syntax exists in
C++; all (top-level) classes in the namespace are visible outside of the namespace.
Namespaces can be nested, with the :: operator used to access the nested namespace.
Namespace names are normal C++ identifiers and do not include the dot (.) that is typical
in Java namespaces.
Classes, functions, and objects that are declared outside of any namespace are considered
to be in the global namespace. These can always be accessed with a leading ::, as in
::IntCell. This may be necessary from code that is inside another namespace that also contains a class called IntCell.

c++book.mif Page 87 Saturday, July 12, 2003 10:53 AM

Incomplete Class Declaration

87

Classes, functions, and objects can be declared in an anonymous namespace. Such entities
are not visible outside of the compilation unit (i.e. .cpp file) in which they are declared. This
allows us to declare classes without fear of conflict.

4.16 Incomplete Class Declaration

Consider the following situation:
class A
{
...
B *data;
};
class B
{
...
A *data;
};

Here, both class A and class B contain a data member that is a pointer to an object of the
other class type. The declaration of class A will generate an error, because the compiler does not
know that B is a class type. Clearly switching the order of declaration for classes A and B wont
work. The solution is an incomplete class declaration that serves solely to allow the compiler to
know the existence of the class type. In our example:
class B;
// incomplete class declaration
class A
{
...
B *data;
};
class B
{
...
A *data;
};

Note that if a class makes a more active use of another class, an incomplete class declaration may not be sufficient. For instance, if class A contained a data member of type B, rather than
simply a pointer, the compiler would still complain. In this case, the complete class declaration
of B (the implementation is not needed, but the memory layout is) would have to precede A.
Class B could not then have a data member of type A (since this would imply A and B are infinitely large). Class B could have a data member that is a pointer to A.

c++book.mif Page 88 Saturday, July 12, 2003 10:53 AM

88

Chapter 4 Object-Based Programming: Classes

4.17 Technical Differences

Although many of the differences are simply syntactic, and of course the entire object model is
different, there are some differences that are subtle and worth mentioning, as they can cause
confusion for Java programmers.
4.17.1 Member Functions and Data Members Cannot Have Same Name
In Java, a class can use a name for both methods and data at the same time. This is not allowed
in C++. In Java the name of class can be used as a method in that class (although usually this just
means that the programmer intending to write a constructor mistakenly provides a void return
type, which becomes legal). In C++ the compiler will not allow a member function to have the
same name as the class in which it is declared.
4.17.2 Inline Data Member Initialization
In Java, by default data members can be initialized inline. In C++ this is illegal, and a compiler
error will result. Data members in C++ can only be initialized in a constructor.
4.17.3 Default Data Member Initialization
In Java, data members will by default be initialized to zero for primitive types, and null for
reference variables. No such guarantee is provided in C++. Primitive types and pointer variables
are not guaranteed default initialization unless they are global or static. Non-primitive data
members are guaranteed initialization by a zero parameter constructor by default, and if no such
constructor is present for the data member, then the lack of a constructor for the class will generate a compiler error.
4.17.4 Default Visibility
In Java, the default visibility is package visibility. In C++ the default visibility is private for
classes, and public for structs.

4.18 Key Points

C++ classes consist of the declaration and definition. Sometimes they are combined but
for large applications, it is common to place the declaration in a .h header file and the
definition in a .cpp implementation file.
The header files should use the ifndef/endif idiom to avoid being processed more than
once per implementation file.
Like Java, constructors are used to create objects. However, primitive data members are
not initialized by default.

c++book.mif Page 89 Saturday, July 12, 2003 10:53 AM

Exercises

89

A copy constructor is always called whenever a brand new object is created that is initialized to be a copy of an existing object. This includes formal parameters that are passed
using call-by-value.
A copy assignment (operator=) is always called whenever an already existing objects
state is changed to be a copy of another already existing object.
When an object goes out of scope, its destructor is invoked. A destructor is needed for any
class that allocates resources (heap memory, file handles, etc.).
If a destructor is written, then typically a copy constructor and copy assignment operator
also need to be written to provide deep copy semantics.
C++ requires that the class designer differentiate accessor member functions (const
member functions) from mutator member functions (non-const member functions). The
const-ness of a member function is part of its signature.
A one-parameter constructor implies the existence of a type-cast operator. The type-cast
operator may be used by the system implicitly, unless the constructor is declared
explicit.
Initializer lists are used by the constructor to initialize data members by invoking their
constructors. A constant data member, or data member without a zero-parameter constructor must be initialized in the initializer list.
A friend of a class is able to access private members of the class. Typically a class grants
friendship to an entire other class, but it can also grant friendship on a function-by-function basis.
C++ supports nested classes but not inner classes or anonymous classes. Although local
classes are allowed, because automatic variables in the function are not accessible, they
have less utility than their Java counterpart.
C++ allows static members. Although there is no static initializer, one can relatively easily
write a function that simulates the behavior of the static initializer. Static data members
must be declared in the class declaration and defined in the implementation file.
Member functions can return parameters by constant reference to avoid copying but it is
fairly tricky to do so.
C++ namespaces are the rough equivalent of packages. The using directive is the equivalent of the import directive.
Incomplete class declarations are used to inform the compiler of the existence of a class
and is used if two or more classes refer to each other circularly.

4.19 Exercises
1. How are public and private class members listed in C++?
2. What is the difference between the class interface and implementation in C++?
3. In Java, it is legal to declare a method with the same name as the class; the result is not a
constructor. Is this also legal C++?

c++book.mif Page 90 Saturday, July 12, 2003 10:53 AM

90

Chapter 4 Object-Based Programming: Classes

4. What is the C++ alternative to using this to invoke another constructor?

5. How are accessors and mutators differentiated in C++?
6. What is an implicit type conversion?
7. What does the explicit reserved word do?
8. What is an initializer list and when must it be used?
9. What is the Big-Three, what does it default to, and when is the default unacceptable?
10. How can copying be disabled for a class type?
11. What is a friend declaration?
12. What are the semantics of nested classes in C++?
13. How does the C++ programmer achieve separate compilation?
14. What is the ifndef/endif idiom and why is it used?
15. How are static members declared and defined in C++?
16. What is an anonymous object?
17. What Java construct most closely mirrors the C++ namespace?
18. What is an incomplete class declaration and why is it used?
19. Implement a Stack class that stores ints and supports size, isEmpty, clear,
push, pop, and peek (as in Java). Provide two different implementations: an array and a
singly-linked list. Implement the Big-Three if appropriate.
20. Implement a Set class that stores strings and supports size, isEmpty, clear,
add, remove, and contains (as in Java). Provide five different implementations: a
sorted array, an unsorted array, a sorted singly-linked list, an unsorted singly-linked list,
and an unbalanced binary search tree. Implement the Big-Three if appropriate.

c++book.mif Page 91 Saturday, July 12, 2003 10:53 AM

H A P T E R

Operator Overloading

N contrast with Java, C++ is designed around making

objects appear to be as similar as possible to primitives. In Chapter 4, we saw that this decision
introduces significant complications that Java does not have. In this Chapter we look at a feature
of C++ that is not part of Java at all, namely operator overloading.
In this chapter, we describe operator overloading, which is the ability to define a meaning
for the existing operators when applied to new class types. In doing so, we can write classes
such as Rational, ComplexNumber, and BigDecimal without resorting to named methods such as add. This makes these types look as if they were primitive types built in to the language from day 1. Operator overloading has a few subtleties, but for the most part with
reasonable care, it is a useful feature that is not at all that complex. Indeed, more than a few programmers list operator overloading as a feature they would like added to Java.

5.1 Basics of Operator Overloading

We begin our discussion with a Person class sketched in Figure 5-1. Once again, we will not
separate the class declaration and implementation except when needed later to illustrate syntax.
Our minimal Person class provides a constructor, that initializes a name and social security
number (a presumably unique identifier of the person), as well as accessors for those data members and a print routine. Obviously we would expect more data members, accessors, and perhaps some mutators, but this class by itself is sufficient to illustrate some basic principles.
In Java, a class such as this would normally be expected to provide an equals method
(and also a hashCode method). So we have provided an equals method that compares two
Person objects (we will declare two Persons equal if they have the same ssn). Given two

91

c++book.mif Page 92 Saturday, July 12, 2003 10:53 AM

92

Chapter 5 Operator Overloading

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23

class Person
{
public:
Person( int s, const string & n = "" )
: ssn( s ), name( n )
{ }
const string & getName( ) const
{ return name; }
int getSsn( ) const
{ return ssn; }
void print( ostream & out = cout ) const
{ out << "[ " << ssn << ", " << name << " ]"; }
bool equals( const Person & rhs ) const
{ return ssn == rhs.ssn; }
private:
const int ssn;
string name;
};

Figure 5-1

Person class with print and equals

Person objects p1 and p2, the result of p1.equals(p2) is the same in this method as it
would be in Java.
However, primitive types are compared with ==. It would be nice if we could write
p1==p2. This is what operator overloading allows us to do. The syntax in this case is easy: we
replace the member function name equals with the name operator==, and voila, it works.
There is no reason to remove equals, we could in fact keep both names, and this is shown in
Figure 5-2.
It is clear from this syntax, that if lhs and rhs are of type Person, that lhs==rhs is
simply a shorthand for lhs.operator==(rhs) that the compiler winds up calling behind
the scenes. And in fact, this is true. Both lhs==rhs and lhs.operator==(rhs) will compile, though obviously using the latter form defeats the whole purpose of defining
operator== in the first place.
In the declaration for both equals and operator==, observe that rhs, which is the
second operand, is passed using constant reference, to signify that it is for input only. The member functions are declared as accessors, signifying that lhs, which is the first parameter, is to
not be altered.
Not only can we overload operator==, we can overload all six of the relational and
equality operators. Note that we have already overloaded operator=, the copy assignment

c++book.mif Page 93 Saturday, July 12, 2003 10:53 AM

Overloading I/O

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

93

class Person
{
public:
Person( int s, const string & n = "" )
: ssn( s ), name( n )
{ }
const string & getName( ) const
{ return name; }
int getSsn( ) const
{ return ssn; }
void print( ostream & out = cout ) const
{ out << "[ " << ssn << ", " << name << " ]"; }
bool equals( const Person & rhs ) const
{ return ssn == rhs.ssn; }
bool operator==( const Person & rhs ) const
{ return equals( rhs ); }
private:
const int ssn;
string name;
};

Figure 5-2

Person class with overloaded operator==

operator. And if it made sense, we could overload operators such as +, +=, <<, >>, etc. Of most
interest is operator<<, which we discuss in the next section.

5.2 Overloading I/O

Given the Person class as defined in Figure 5-2, we can send an instance to any output stream
by invoking the print method. If the print method has no parameters, the output is sent to
the standard output stream cout. With parameters, we could send it elsewhere, such as cerr,
or a file stream. Although we do not discuss I/O until Chapter 9, it is not too much to admit that
istream and ostream, like the Java InputStream and OutputStream, are base classes
in an inheritance hierarchy that covers all input streams and output streams.
Again, examining the parameters to print, we see that the ostream is passed by reference, signalling that after the Person is processed, the state of the ostream can (and must) be
changed. The print routine is an accessor, signifying that the Person itself does not change.
So given a Person p, p.print(cout), or simply p.print() outputs the p to
cout. However, for primitive types, we are used to invoking the << operator, as shown in the

c++book.mif Page 94 Saturday, July 12, 2003 10:53 AM

94

Chapter 5 Operator Overloading

1
2
3
4
5

ostream & operator<< ( ostream & out, const Person & p )

{
p.print( out );
return out;
}

Figure 5-3

Implementation of operator<< as a non-member function

implementation of print at line 15. And since we have operator overloading, shouldnt we be
able to overload operator<<?
The answer to the question is yes, however if we attempt to place operator<< in the
Person class, we see a problem. If the operator signature takes one parameter, as in
operator<< ( ostream & out ) const

then it must be invoked as p<<cout, in order for the compiler to rewrite it as

p.operator<<(cout). But of course, we write the expression as cout<<p, suggesting
that we need to go into the ostream class and add
operator<< ( const Person & p ) const

Needless to say, changing the ostream class is not going to be a viable solution. Instead,
C++ provides an alternative: we can define operator<< as a non-class function. In that case,
a respectable implementation is shown in Figure 5-3.
Here we see that operator<< takes two parameters: first, the ostream, passed by reference, as expected, and second, a Person, passed by constant reference, since it is not to be
changed and can be a large object that is expensive to copy. Since calls to << are typically
chained, we want operator<< to return a reference to out, so that
cout << p1 << " " << p2;

evaluates to
( cout << p1 ) << " " << p2;

and then the left-hand operand of the second << is still an ostream.
The implementation of operator<< simply invokes the print member function of
Person p. (Note that if this member function was not declared as an accessor in Figure 5-2,
then this invocation would not compile.) We then return the ostream.
This code is boilerplate: any class can provide an implementation of a print member
function and an overloaded operator<< that invokes it. If all classes do so, then we can
always print objects using operator<<. This convention in effect replaces Javas convention
of every class providing a toString method.

5.3 Global vs. Class Operators

The discussion thus far suggests that an operator can be overloaded either as a member function
or as a non-member function. It also suggests that if it is overloaded as a member function, then
the first operand of the operator must be of the class type, since it becomes *this. In other

c++book.mif Page 95 Saturday, July 12, 2003 10:53 AM

Global vs. Class Operators

95

words, in lhs==rhs, when operator== is a member function *this is lhs. So if the first
operand cannot be of the class type in which the overloaded operator would be a member, as in
the case of operator<< where the first operand is ostream, then we must use a non-member function.
Overloading an operator as a non-member function has a significant disadvantage: since it
is not a member of the class, the implementation cannot access any private data, unless it is
made a friend of the class. Thus, the following code does not compile unless Person has
declared that operator<< is a friend:
ostream & operator<< ( ostream & out, const Person & p )
{
out << "[ " << p.ssn << ", " << p.name << " ]";
return out;
}

On the other hand, the implementation of member function operator== can access private data, since it is a member function of class Person:
bool operator==( const Person & rhs ) const
{ return ssn == rhs.ssn; }

In reality, the fact that non-member functions cannot access private data is rarely a significant liability because the class designer often can implement non-member functions by invoking
public member functions of the class. Furthermore, the member function implementation of
operator== has its own subtle problem. Recall that in Java, we expect equals to be symmetric: a.equals(b) and b.equals(a) should give the same result if both are not null.
The implementation of operator== given above fails this requirement. Specifically,
because the Person constructor is not declared explicit, and because a one-parameter constructor is available (a default parameter can be used for the name), the following code compiles, and returns true:
Person p1( 123, "Joe" );
cout << ( p1 == 123 ) << endl;

In trying to invoke p1==Person, the compiler finds an inexact match, but can use the
one-parameter constructor to generate a temporary Person object from the int 123. However,
cout << ( 123 == p1 ) << endl;

does not compile: when using operator overloading, if the implementation is a class member
function, the first parameter MUST be exact. Implicit type conversion are not allowed.
If instead of implementing operator== as a member function, it was implemented as a
non-member function:
bool operator== ( const Person & lhs, const Person & rhs )
{
return lhs.equals( rhs );
}

both calls to == above would compile and yield true. In the non-member definition, we now

c++book.mif Page 96 Saturday, July 12, 2003 10:53 AM

96

Chapter 5 Operator Overloading

have two parameters. Both parameters are input parameters only, and being potentially large are
thus passed using call-by-constant reference. Since the function is not a member function, it cannot be an accessor function. Observe that the fact that this member function cannot access private data does not limit us.
It is illegal to provide both versions of operator==, since invoking p1==p2 would be
ambiguous. In our particular situation, the most sensible approach would be to make the constructor explicit, in which case we would not need to worry about this.
However, there are many cases where implicit type conversions make sense. For instance,
if we have a BigInteger class, it would be nice to be able to write the following code:
int log2( const BigInteger & x )
{
int result = 0;
for( BigInteger b = x; b != 1; b /= 2 )
result++;
}

Certainly, we would expect that using 1!=b in the for loop would also work. (Well presume that ints are 32 bits, and that a BigInteger has less than 2,000,000,000 ( 2 31 ) bits.
Given that this would require 250 Megabytes to represent, thats probably not too unreasonable
a limit). Looking at BigInteger, to support b!=1 and 1!=b we have several choices.
Choice #1 is to provide a single non-member function:
bool operator!= ( const BigInteger & lhs, const BigInteger & rhs )

both b!=1 and 1!=b will generate a temporary BigInteger and in either case, the original
BigInteger and the temporary BigInteger will be compared.
An alternative is to provide a member function, and a non-member function. While they
cannot clash, we can still overload, yielding both:
bool operator!= ( int lhs, const BigInteger & rhs )
bool BigInteger::operator!= ( const BigInteger & rhs ) const

Of course, the second function would also be declared in the BigInteger class declaration.
With this approach, two BigIntegers are compared with operator!=; b!=1 generates a temporary and then the member version of operator!= is used, and finally, 1!=b is an
exact match for the non-member function. Although we can always provide the following
implementation for the non-member function:
bool operator!= ( int lhs, const BigInteger & rhs )
{
return rhs != lhs;
}

we might be able to do better with sufficient effort, based on the fact that lhs must be small. If
this is not the case, we are probably better of with the first choice above. If it is the case, then it
stands to reason that it is worth defining three overloaded functions:
bool operator!= ( int lhs, const BigInteger & rhs )
bool BigInteger::operator!= ( const BigInteger & rhs ) const

c++book.mif Page 97 Saturday, July 12, 2003 10:53 AM

What Can Be Overloaded?

97

bool BigInteger::operator!= ( int rhs ) const

since any slick algorithm for the non-member version, could also be used to implement the
member function that takes int rhs as a parameter.
Which of these design decisions is best depends on the particular application. The
string class, for instance, provides a host of overloaded operators that handle primitive
strings, characters, and other string objects, rather than relying on the temporaries that are
created in an implicit conversion. The cost of the temporaries is not only the cost to construct
them, but also the cost to invoke their destructors, and in the case of string (and
BigInteger) this is certain to involve calls to new and delete.

5.4 What Can Be Overloaded?

In C++, all but four operators can be overloaded. The four nonoverloadable operators are
the dot operator., the rarely-used .*, the ternary operator ?:, and sizeof. Additionally,
though it is legal to overload && and ||, doing so is of dubious utility, since those operations
guarantee left-to-right short circuit evaluation, and when overloaded that guarantee is impossible to replicate. Also the comma operator, which is dubious in the first place, is best not overloaded.
When overloading operators, the operator precedence and associativity cannot be
changed. That is, a+b/c/d is always a+((b/c)/d). Thus overloading ^ in an attempt to
implement exponentiation is dubious because exponentiation associates right to left and should
have higher precedence than multiplication, neither of which is true for operator^.
Arity (whether an operator is unary or binary) cannot be changed, so for example we cannot write a unary / operator or a binary ~. Finally, only existing operators can be overloaded;
new operators cannot be created. This disallows, for instance, overloading ** to achieve exponentiation. Foiled again!
The rules above tells us that almost all of the sensible operators are overloadable. For
classes such as BigDecimal or Rational, this would include basic arithmetic operators +, , *, etc.; assignment operators =, +=, -=, etc.; the bit operators &, ^, and |; the equality and
relational operators; the shifting operators (both for bit shifting and I/O); the unary + and - operators; ++ and --;. We will see all of these operators overloaded when we discuss the
Rational class in Section 5.5
There are other operators that surprisingly can be overloaded. The most obvious of those
is the array indexing operator [], since we know it is overloaded in vector. We will discuss
the issues involved in overloading operator[] in Section 5.6.
Other operators that can be overloaded include the unary pointer dereferencing
operator*, operator->, operator(), and the typecast operator. We will eventually see
uses for all of these, but not in this chapter.

5.5 A Rational Number Class

The classic use of operator overloading is implementing a numeric type such as BigInteger,

c++book.mif Page 98 Saturday, July 12, 2003 10:53 AM

98

Chapter 5 Operator Overloading

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21

#include "Rational.h"
#include <iostream>
using namespace std;
// Rational number test program
int main( )
{
Rational x;
Rational sum = 0;
int n = 0;
cout << "Type as many rational numbers as you want" << endl;
for( sum = 0, n = 0; cin >> x; sum += x, n++ )
cout << "Read " << x << endl;
cout << "Read " << n << " rationals" << endl;
cout << "Average is " << ( sum / n ) << endl;
return 0;
}

Figure 5-4

Using the Rational class

BigDecimal, ComplexNumber, or Rational. The code in Figure 5-4 illustrates a program that uses a rational number class, Rational. To run the program, simply type in rational
numbers, separated by whitespace. Terminate the input with either the end-of-file marker (control-Z on Windows, control-D on Unix), or simply type something that is not a Rational.
The program simply computes the average of its input numbers. If the input numbers are
typed as
3 8/6 -2/5

the program replies with 59/45. If no numbers are typed (prior to the end-of-file marker), the
program replies with indeterminate. If we examine the program, we see that the Rational
type behaves exactly as any other primitive numeric type.
Figure 5-5 shows the first half of the Rational class declaration, which contains all of
the public members. Figure 5-6 shows the second half and also includes operators that are nonmembers. The contents of these figures reside entirely in Rational.h.
Many of the operations are similar (e.g < and >), so only a representative set of operators
are implemented in the code that follows. The implementation of the class follows in a host of
separate code fragments that are all placed in Rational.cpp.
A rational number consists of a numerator and a denominator. The constructors allow,
among other things, a single int as the initializer, and this constructor is not marked
explicit, so that we can take advantage of mixing Rational and int types.

c++book.mif Page 99 Saturday, July 12, 2003 10:53 AM

A Rational Number Class

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42

99

#ifndef _RATIONAL_H
#define _RATIONAL_H
#include <iostream>
using namespace std;
class Rational
{
public:
// Constructors
Rational( int numerator = 0 )
: numer( numerator ), denom( 1 ) { }
Rational( int numerator, int denominator )
: numer( numerator ), denom( denominator )
{ fixSigns( ); reduce( ); }

Figure 5-5

// Assignment Ops
const Rational & operator+=(
const Rational & operator-=(
const Rational & operator/=(
const Rational & operator*=(

const
const
const
const

Rational
Rational
Rational
Rational

// Unary Operators
const Rational & operator++( );
Rational operator++( int );
const Rational & operator--( );
Rational operator--( int );
const Rational & operator+( ) const;
Rational operator-( ) const;
bool operator!( ) const;

//
//
//
//

&
&
&
&

rhs
rhs
rhs
rhs

);
);
);
);

Prefix
Postfix
Prefix
Postfix

// Named Member Functions

double toDouble( ) const
// Do the division
{ return static_cast<double>( numer ) / denom; }
int toInt( ) const
// Do the division
{ return numer >= 0 ? numer / denom : - ( -numer / denom ); }
bool isPositive( ) const
{ return numer > 0; }
bool isNegative( ) const
{ return numer < 0; }
bool isZero( ) const
{ return numer == 0; }
void print( ostream & out = cout ) const;
Rational declaration (public section)

We maintain the invariant that the denominator is never negative and that the rational
number is expressed in lowest form. Thus the Rational 9/-6 would be represented with

c++book.mif Page 100 Saturday, July 12, 2003 10:53 AM

100

Chapter 5 Operator Overloading

43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70

private:
// A rational number is represented by a numerator and
// denominator in reduced form
int numer;
// The numerator
int denom;
// The denominator
void fixSigns( );
void reduce( );

// Ensures denom >= 0

// Ensures lowest form

};
// Math Binary Ops
Rational operator+( const
Rational operator-( const
Rational operator/( const
Rational operator*( const
//
bool
bool
bool
bool
bool
bool

Rational
Rational
Rational
Rational

Relational & Equality Ops

operator< ( const Rational
operator<=( const Rational
operator> ( const Rational
operator>=( const Rational
operator==( const Rational
operator!=( const Rational

&
&
&
&
&
&

&
&
&
&

lhs,
lhs,
lhs,
lhs,

lhs,
lhs,
lhs,
lhs,
lhs,
lhs,

const
const
const
const

const
const
const
const
const
const

Rational
Rational
Rational
Rational

Rational
Rational
Rational
Rational
Rational
Rational

&
&
&
&
&
&

&
&
&
&

rhs
rhs
rhs
rhs

rhs
rhs
rhs
rhs
rhs
rhs

);
);
);
);
);
);

);
);
);
);

// I/O
ostream & operator<< ( ostream & out, const Rational & value );
istream & operator>> ( istream & in, Rational & value );
#endif
Figure 5-6
Rational declaration (private section and non-members)

numerator -3 and denominator 3. We allow the denominator to be zero, thus representing infinity, negative infinity, or (if the numerator is also zero, indeterminate). These invariants are maintained internally by applying fixSigns and reduce as appropriate. Thus, not only is the data
representation private, but so are fixSigns and reduce, as are shown in Figure 5-7.
reduce uses a greatest-common-divisor algorithm, which illustrates the most interesting facet
of Figure 5-7. Although we could make this routine a private static member function, we have
chosen to simply make it a non-member function, but to avoid conflicting with others with the
same name, we limit its scope to the .cpp file by placing it in the anonymous namespace
(Section 4.15). Also, since the behavior of the % operator is undefined if the operands are negative, we insure that they are not by having gcd switch the sign of a negative numerator prior to
calling gcdRec.
Our class defines several named member functions that are all trivial. toDouble and
toInt are used to create a double or int from the Rational. An implicit type conversion
is not supported. We could have tried to allow it by using operator overloading to implement
type conversion operators:

c++book.mif Page 101 Saturday, July 12, 2003 10:53 AM

A Rational Number Class

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41

101

operator double ( ) const

// Allows type cast to double
{ return toDouble( ); }
operator int ( ) const
// Allows type cast to int
{ return toInt( ); }
namespace
{
// n is guaranteed non-negative
int gcdRec( int n, int m )
{
if( n % m == 0 )
return m;
else
return gcdRec( m, n % m );
}
int gcd( int m, int n )
{
if( m > 0 )
return gcdRec( n, m )
else
return gcdRec( n, -m );
}
}
void Rational::fixSigns( )
{
if( denom < 0 )
{
denom = -denom;
numer = -numer;
}
}
void Rational::reduce( )
{
int d = 1;
if( denom !=
d = gcd(
if( d > 1 )
{
numer /=
denom /=
}

0 && numer != 0 )
numer, denom );

d;
d;

Figure 5-7

Private member routines and local gcd to keep Rationals in normalized form

c++book.mif Page 102 Saturday, July 12, 2003 10:53 AM

102

Chapter 5 Operator Overloading

1
2
3
4
5
6
7
8
9
10
11
12
13

const Rational & Rational::operator/=( const Rational & rhs )

{
if( this == & rhs )
numer = denom = 1;
else
{
numer *= rhs.denom;
denom *= rhs.numer;
fixSigns( );
reduce( );
}
return *this;
}

1
2
3
4
5
6
7
8
9
10
11

const Rational & Rational::operator/=( const Rational & rhs )

{
int newNumer = numer *= rhs.denom;
int newDenom = denom *= rhs.numer;
numer = newNumer;
denom = newDenom;
fixSigns( );
reduce( );
return *this;
}

Figure 5-8

Two possible implementations of operator/=

However, this is a bad idea because it can result in ambiguities. Specifically, if r is a

Rational, then r==0 becomes ambiguous given the set of operators that we have defined.
This is because the Rational constructor allows 0 to be typecast up to a Rational (invoking the constructor would create a Rational), while at the same time operator int allows
r to be typecast down to an int. The compiler has no way of deciding which cast is better, so in
the absence of an exact match, it declares an ambiguity. This is known as a dual-direction
implicit conversion, and we recommend against allowing dual-direction in implicit conversions
because it can eventually result in an ambiguity, and even worse, the ambiguity is not detected
unless it actually occurs, meaning that you might not realize that your class has potential problems.
The remainder of this section examines the operator overloading issues. We will not be
overly concerned with efficiency issues, and instead will focus on keeping the code simple.

c++book.mif Page 103 Saturday, July 12, 2003 10:53 AM

A Rational Number Class

1
2
3
4
5
6
7
8

103

const Rational & Rational::operator+=( const Rational & rhs )

{
numer = numer * rhs.denom + rhs.numer * denom;
denom = denom * rhs.denom;
reduce( );
return *this;

}
Figure 5-9

5.5.1

Assignment operator

Assignment Operators

The various assignment operators use concepts that are identical to the copy assignment operator
(operator=), discussed in Section 4.6. operator+= is representative and is shown in
Figure 5-9. Although it is not an issue here, it is always important to consider the possible
effects of aliasing prior to making changes to data members.
As an example, the implementation of operator/= shown below fails to work correctly
when it is invoked using r/r.
const Rational & Rational::operator/=( const Rational & rhs )
{
numer *= rhs.denom;
denom *= rhs.numer;
fixSigns( );
reduce( );
return *this;
}

To fix the problem, either a special aliasing test can be used, or temporaries can be used.
Both ideas are implemented in Figure 5-8.
It is important to note that the existence of operator+= does not imply anything about
operator+.
5.5.2

Arithmetic Operators

Arithmetic operators such as addition and subtraction can be implemented as either member
functions or outside of the class. Recall that as member functions they have access to private
implementation details. However, when implemented as a non-member function, the first oper1
2
3
4
5
6

Rational operator+( const Rational & lhs, const Rational & rhs )
{
Rational answer( lhs );
// Initialize answer with lhs
answer += rhs;
// Add the second operand
return answer;
// Return answer by copy
}

Figure 5-10

Addition operator

c++book.mif Page 104 Saturday, July 12, 2003 10:53 AM

104

Chapter 5 Operator Overloading

1
2
3
4

bool operator<=( const Rational & lhs, const Rational & rhs )
{
return !(lhs - rhs).isPositive( );
}

Figure 5-11

Relational operator

and does not have to be an exact match if an implicit type conversion is available. Thus we
choose to use a nonmember function for maximum flexibility.
The existence of operator+ does not imply anything about operator+=. However, it
is customary to ensure that these two operations have consistent semantics. The implementation
in Figure 5-10 shows the typical idiom that works whether or not we implement operator+ as
a member function, does not require access to any private details, and guarantees that
operator+ is implemented consistently with respect to operator+=. Because neither operand is mutable, there are no aliasing issues. Also, observe that the return type is by value,
because we are returning an object (the sum) that did not exist prior to the call and which is created as a stack-allocated local variable.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

void Rational::print( ostream & out ) const

{
if( denom != 0 )
{
out << numer;
if( denom != 1 )
out << '/' << denom;
}
// Messy code for denom == 0
else if( numer == 0 )
out << "indeterminate";
else
{
if( numer < 0 )
out << '-';
out << "infinity";
}
}
ostream & operator<<( ostream & out, const Rational & value )
{
value.print( out );
return out;
}

Figure 5-12

Output operator

c++book.mif Page 105 Saturday, July 12, 2003 10:53 AM

A Rational Number Class

5.5.3

105

Relational and Equality Operators

As was the case for the binary arithmetic operators, we choose to implement the relational and
equality operators as nonmembers. Importantly, the existence of == does not imply anything
about !=, and similarly each of the relational operators must be implemented. We implement the
most complicated of the group in Figure 5-11.
5.5.4

Input and Output Operators

Figure 5-12 shows the output routines using the idiom we discussed in Section 5.2. For input,
we use istream instead of ostream, >> instead of <<, and the second parameter is passed
by reference, since the point of the operator is to change the state of the Rational parameter.
The implementation is shown in Figure 5-13. Since I/O is not discussed until Chapter 9,
we will simply remark that the basic algorithm attempts to read an integer. Then it tries to read a
/. If it reads a character and it is a /, then it reads the denominator, otherwise it puts the character back on the input stream, and assumes the denominator is 1. A anonymous temporary
Rational object is created and then assigned to value at line 17, and the stream is returned at
line 18.
Observe that this implementation accesses no private information from the Rational
class. The creation of the temporary Rational at line 17 could have been avoided if this function was made a friend of the Rational class, or if a setValue member function would
have been provided. In that case, values numerator and denominator could have been set once
num and den were known.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

istream & operator>>( istream & in, Rational & value )

{
int num;
int den = 1;
in >> num;
char ch;
in.get( ch );

// try to read the /

if( !in.eof( ) )
// if read has hit EOF, eof is true
if( ch == '/' )
in >> den;
else
in.putback( ch );
value = Rational( num, den );
return in;
}

Figure 5-13

Input operator

c++book.mif Page 106 Saturday, July 12, 2003 10:53 AM

106

Chapter 5 Operator Overloading

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27

const Rational & Rational::operator++( )

{
numer += denom;
return *this;
}

// Prefix form

Rational Rational::operator++( int )

{
Rational tmp = *this;
numer += denom;
return tmp;
}

// Postfix form

bool Rational::operator!( ) const

{
return !numer;
}
const Rational & Rational::operator+( ) const
{
return *this;
}
Rational Rational::operator-( ) const
{
return Rational( -numer, denom );
}

Figure 5-14

Unary operators

5.5.5

Unary Operators

We continue with the ++ and -- operators. We will examine the auto-increment operator. As in
Java, there are two flavors: prefix (before the operand) and postfix (after the operand). Both add
1 to an object, but the result of the expression (which is meaningful if used in a larger expression) is the new value in the prefix form and the original value in the postfix form. They are
completely different in semantics and precedence. Consequently, we need to write separate routines for each form. Since they have the same name, they must have different signatures to be
distinguished. This is done in C++ by specifying an empty parameter list for the prefix form and
a single (anonymous and unused) int parameter for the postfix form. ++x calls the zeroparameter operator++; x++ calls the one-parameter operator++. This int parameter is
never used; it is present only to give a different signature.
The prefix and postfix forms shown in Figure 5-14 add 1 by increasing numer by the
value of denom. In the prefix form we can then return *this by constant reference, as done for
the assignment operators. The postfix form requires that we return the initial value of *this,

c++book.mif Page 107 Saturday, July 12, 2003 10:53 AM

Matrix Class (for doubles)

107

and thus we use a temporary. Because of the temporary, we have to return by value instead of
reference. Even if the copy constructor for the return is optimized away, the use of the temporary
suggests that, in many cases, the prefix form will be faster than the postfix form.
The three remaining unary operators have straightforward implementations, as shown in
Figure 5-14. operator! returns true if the object is zero; this is done by applying ! to the
numerator. Unary operator+ evaluates to the current object; a constant reference return can
be used here. operator- returns the negative of the current object by creating a new object
whose numerator is the negative of the current object. The return must be by copy because the
new object is a local variable. However, there is a trap lurking in operator-. If the word
Rational is omitted on line 26, then the comma operator evaluates (-numer,denom) as
denom, and then an implicit conversion gives the Rational denom/1, which is returned.
Have we had enough of the comma operator yet?

5.6 Matrix Class (for doubles)

C++ does not provide a library type for two-dimensional arrays. However, it is easy to create a
class that supports two-dimensional arrays in the same style as Java. We will call this class
MatrixOfDouble. In Chapter 7, we discuss how to generalize this to support two-dimensional arrays of any type.
Recall that in Java, in the declaration
double [ ][ ] mat = { { 3, 2 }, { 4, 1, 3 }, { 5, 2 } };

mat.length is 3 (because there are three rows), and each row is itself a one-dimensional
array. Thus, mat[0] is of type double[], and mat[0].length is 2. Similarly
mat[1].length is 3, and mat[2].length is 2.
As shown in Figure 5-15, our class will store the two-dimensional array using a vector
of vectors (at line 33). Note that in the declaration of array, white space must separate the
two > characters; otherwise the compiler will interpret the >> token as shift operator. In other
words, we must write
vector<vector< double> > array;

// white space needed

and not
vector<vector<double>> array;

// oops!

The constructor first constructs array as having rows entries each of type
vector<double>. Since each entry of array is constructed with the zero-parameter constructor, it follows that each entry of array is a vector<double> object of size 0. Thus we
have rows zero-length vectors of double. The body of the constructor is then entered and
each row is resized to have cols columns. Thus the constructor terminates with what appears to
be a two-dimensional array. (Note that the doubles themselves are not guaranteed any initialization)

c++book.mif Page 108 Saturday, July 12, 2003 10:53 AM

108

Chapter 5 Operator Overloading

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36

#ifndef _MATRIX_OF_DOUBLE_H
#define _MATRIX_OF_DOUBLE_H
#include <vector>
using namespace std;
class MatrixOfDouble
{
public:
MatrixOfDouble( int rows, int cols ) : array( rows )
{ setNumCols( cols ); }
int
{
int
{

numrows( ) const
return array.size( ); }
numcols( ) const
return numrows( ) > 0 ? array[ 0 ].size( ) : 0; }

void setNumRows( int rows )

{ array.resize( rows ); }
void setNumCols( int cols )
{ for( int i = 0; i < rows; i++ )
array[ i ].resize( cols );
}
void setDimensions( int rows, int cols )
{ setNumRows( rows ); setNumCols( cols ); }
const vector<double> & operator[]( int row ) const
{ return array[ row ]; }
vector<double> & operator[]( int row )
{ return array[ row ]; }
private:
vector< vector<double> > array;
};
#endif

Figure 5-15

Java class to represent two-dimensional array of double

Because vectors know how to clean up their own memory, we do not need to worry
about a destructor, copy constructor, or copy assignment operator.
The numrows and numcols accessors are easily implemented as shown. As we will see,
it is possible for the user to make the two dimensional array non-rectangular, in which case, the
result from numcols is meaningless.

c++book.mif Page 109 Saturday, July 12, 2003 10:53 AM

Matrix Class (for doubles)

109

We also provide member functions to change the number of rows and columns. Note that
when the number of rows is changed by resizing the array, any additional rows have length 0
(i.e. no columns).
As our Java example illustrated, the key operation is [], which in C++ can be overloaded
as operator[]. The result of mat[r] is the invocation of mat.operator[](r), and
returns a vector corresponding to row r of matrix mat. Thus we have a skeleton:
vector<double> operator [] ( int row )
{ return array[ row ]; }

The main question is whether this is an accessor or a mutator, and what the return mechanism should be. If we consider the following routine:
void copy( MatrixOfDouble & to, const MatrixOfDouble & from )
{
for( int r = 0; r < to.numrows( ); r++ )
to[ r ] = from[ r ];
}

in which we copy each row of from into the corresponding row of to, we see contradictions.
If operator[] returns a vector<double> by value, then to[r] cannot appear on
the left-hand side of the assignment. The only way to affect a change of the elements stored in
to is if operator[] returns a reference to a vector. Unfortunately, doing so would allow
from[ r ] = to[ r ];

to compile, in violation of froms const-ness. That cannot be allowed in a good design.

Thus we see that we need operator[] to return a constant reference for from, but a
plain reference for to. In other words, we need two versions of operator[], which differ
only in the return types. That is not allowed for two members that otherwise have the same signature. And therein lies the trick: whether a member function is an accessor or a mutator is part
of the signature. Thus the version that returns a constant reference will be marked as an accessor,
and the version that returns a reference will be marked as a mutator. This is shown at lines 27 to
30.
It is worth emphasizing two points. First, in resolving the call, the compiler will always
choose the mutator unless it is not a candidate (because the object it is acting upon is a constant).
Second, any member function that returns a reference to private data is explicitly allowing the
private data to be changed later on by anyone who retains that reference. As such, the technique
should generally be avoided except in cases such as this one where this is a desired outcome, and
member functions that return by reference should not be marked as accessors, even though the
call itself does not changed the state of the object.
To see how this works, the expression mat[0][1]=3.14 consists of:
vector<double> & row0 = mat.operator[] ( 0 );
double
& item01 = row0.operator[] ( 1 );
item01 = 3.14;

// matrix [] mutator
// vector [] mutator

c++book.mif Page 110 Saturday, July 12, 2003 10:53 AM

110

Chapter 5 Operator Overloading

Here row0 is a reference to (i.e. another name for) mat.array[0] that is stored internally in mat, representing row 0. item00 is a reference to mat.array[0][1], stored internally. So changing item01 to 3.14 changes the entry in mat.array[0][1].
Notice that although we are invoking operator[] twice, we are invoking two different
versions of operator[]. Notice that both versions are mutators, but do not make any changes
on their own to the objects they are acting upon. But by returning references to the object, they
open the door for later changes.

5.7 Liabilities of Operator Overloading

When used properly, operator overloading can provide spectacular results, as can be seen by the
Rational class. So why is it not part of Java? The main reason is that programmers tend to
abuse a good thing.
For instance, suppose we have a Company class and a Person class. Occasionally we
want to add or remove a Person from a Company. The sensible thing to do is to define add
and remove member functions in the Company class. But the overly aggressive operator overloader will instead overload operator+= and operator-= in place of add and remove,
and then we get code such as:
Company c;
Person p;
...
c += p;

which is not intuitive and can be hard to read. Pretty soon operators are used everywhere and
nobody can understand the code. Operator overloading should never be used in place of named
members when the operators do not provide intuitive semantics.

5.8 Key Points

Operator overloading allows us to define meanings for the existing operators when
applied to non-primitive types.
When an operator is overloaded as a member function, the first operand is accessible as
*this, and subsequent operands are parameters. The first operand must be an exact
match of the class type; implicit type conversions are not acceptable.
When an operator is overloaded as a non-member function, all operands are parameters,
but the implementation does not, by default, have access to private members of any class.
All but a few operators can be overloaded. The most common operators to overload
include the assignment operator(s), equality and relational operators, and the input and
output stream operators. Later we will see that the function call operator can also be overloaded.
operator<< should be overloaded for all class types. Typically this is done by providing a companion print member function.

c++book.mif Page 111 Saturday, July 12, 2003 10:53 AM

Exercises

111

operator+= and operator+ must be overloaded separately, and consistently. Typically operator+ simply invokes operator+=.
When operator++ is overloaded, typically both a prefix and a postfix version are provided.
When operator[] is overloaded, typically both an accessor and a mutator version are
provided.

5.9 Exercises
1. Which operators cannot be overloaded?
2. Under what circumstances can the precedence, associativity, or arity of an operator be
changed?
3. Explain the difference between overloading operator= as a member function versus a
non-member function.
4. When must an operator be overloaded as a non-member function?
5. Under what circumstances is it a dangerous idea to overload the type conversion operator?
6. Explain how this is used in the implementation of assignment operators.
7. How are operator+= and operator+ implemented to ensure compatible semantics?
8. Implement a Complex class that supports complex numbers. Provide as many operations
as you can.
9. Implement a BigInteger class that supports arbitrary precision integers, providing as
many operations as you can with reasonable efficiency. Provide two implementations: one
maintains the digits in an array, and a second maintains the digits in a linked list.
10. Implement a Polynomial class that supports single-variable polynomials. Support
operator+, operator-, and operator*, as well as the corresponding assignment
operators, operator==, operator!=, as well as an overloaded input and output operator. Add a public method, eval, that takes evaluates the polynomial at a particular point.
You should use a good separation of interface and implementation. The polynomial is
implemented as a sorted vector of pairs, in which each pair represents a coefficient and
an exponent. For instance, the polynomial x^2+4 is represented by a vector of size 2,
with index zero containing (4,0) and index one containing (1,2). If this polynomial is represented by p, then p.eval(2) returns 8. You may assume that all coefficients are of
type double and exponents are non-negative integers. A polynomial can be constructed
by passing a string. For simplicity, you may assume that there are no spaces in the
string. This makes it easy to implement the required operator<<. An example of a
call to the constructor is
Polynomial p( "x^2+8x+15" );

11. Implement a Map class that stores keys and values that are both strings. Like the Java
Map, your map will support isEmpty, size, clear, get, put, and remove. Additionally, support operator[] which takes a key as a parameter, and returns a reference

c++book.mif Page 112 Saturday, July 12, 2003 10:53 AM

112

Chapter 5 Operator Overloading

to the corresponding value in the map. If the key is not present, operator[] will insert
it with a default value, and return a reference to the newly inserted value. This implies that
operator[] is a mutator. Add an accessor version with similar semantics, but have it
throw an exception if the key is not found. Provide two separate implementations: one
maintains the items in an array, and a second maintains the items in a linked list. In both
cases, each key and value is stored together in a Pair, which is a nested class that you
should define.

c++book.mif Page 113 Tuesday, April 29, 2003 2:13 PM

H A P T E R

Object-Oriented
Programming: Inheritance

IKE Java, C++ supports inheritance. Most of the features associated with inheritance that are found in Java have equivalent implementations in C++.
In some cases, these features are not the default behavior, and thus the C++ programmer must be
more careful than a Java programmer. In other cases, the behavior of similar constructs is
slightly different. And additionally, C++ supports some techniques, such as multiple inheritance,
that Java does not allow.
In this chapter, we describe the basics of Java inheritance, see some of the extra code that
the C++ programmer must write to avoid subtle errors, discuss multiple inheritance in C++, and
examine the differences between similar C++ and Java constructs.

6.1 Basic Syntax

To illustrate the basics of inheritance, we begin with a simple Person class in Figure 6-1 and
then a simple Student class in Figure 6-2 that extends Person. The Person class duplicates Figure 5-1, but has some subtle errors that eventually must be fixed. But first, lets examine the basic syntax in the Student class.
First, instead of the extends clause that is found in Java, an IS-A relationship is signalled
in C++ by using the syntax seen at line 1 in Figure 6-2:
class Student : public Person

The reserved word public is required; otherwise we get private inheritance, which does
not model an IS-A relationship (see Section 6.10.2) and is typically not what we want.

113

c++book.mif Page 114 Tuesday, April 29, 2003 2:13 PM

114

Chapter 6 Object-Oriented Programming: Inheritance

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

class Person
{
public:
Person( int s, const string & n = "" )
: ssn( s ), name( n )
{ }
const string & getName( ) const
{ return name; }
int getSsn( ) const
{ return ssn; }
void print( ostream & out = cout ) const
{ out << ssn << ", " << name; }
private:
int ssn;
string name;
};
ostream & operator<< ( ostream & out, const Person & p )
{
p.print( out );
return out;
}

Figure 6-1

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

Person as a base class (version #1 slightly flawed)

class Student : public Person

{
public:
Student( int s, const string & n = "", double g = 0.0 )
: Person( s, n ), gpa( g )
{ }
double getGpa( ) const
{ return gpa; }
void print( ostream & out = cout ) const
{ Person::print( out ); out << ", " << gpa; }
private:
double gpa;
};

Figure 6-2

Student as a derived class (not flawed)

c++book.mif Page 115 Tuesday, April 29, 2003 2:13 PM

Dynamic Dispatch

115

Initialization of the subclass requires a call to the superclass constructor. As in Java, by

default this initialization is a call to the zero-parameter superclass constructor. By using an initializer list, we can specify a call to any superclass constructor, and as in Java, this is required if
the superclass does not have a zero-parameter constructor. In our example, the initializer list at
line 5 in Figure 6-2 illustrates the syntax.
At line 12, we see the equivalent to Javas super.print: we simply use the :: scoping
operator to specify the call to the Persons implementation of print.
Otherwise, the rest of the class is similar to what would be seen in Java. The print routine in Student intends to override the print routine in Person. The getName and
getSsn member functions from Person are inherited by Student. The memory layout of
Student includes the data members specified in Person, but these data members cannot be
accessed from member functions implemented in class Student.
Thus,
Person p( 987654321, "Bob" );
Student s( 123456789, "Jane", 4.0 );
p.print( );
s.print( );
cout << s.getName( ) << " " << s.getGpa( ) << endl;

produces the expected output.

Additionally, we implement operator<< for Person, and expect that a Student can
be passed as a parameter because a Student IS-A Person. Indeed this is true; however, there
is a subtle bug in Persons print member function that causes the call to operator<< to
print the wrong information when a Student is passed as a parameter. We can easily fix this
problem in the next section.

6.2 Dynamic Dispatch

As we mentioned, the print routine in Person has a serious, yet easily fixable, error that is
illustrated by the following code:
Student s( 123456789, "Jane", 4.0 );
const Person & p = s;
// p and s are same object
s.print( );
// calls Student::print
p.print( );
// calls wrong print!!

Clearly p and s are the same object. So if we invoke the print member function on
either p or s, we expect the same answer; certainly this basic polymorphic behavior is what happens in Java. However, in our code, s.print invokes Student::print on our student,
while p.print invokes Person::print on our student.
The problem is that it in the call to p.print, the compile-time type of p is Person,
while the run-time type (the type of the object p actually references when the program is run) is
Student. In C++, by default, the decision on which member function to invoke is made on the
basis of the compile-time type of p. This is known as static dispatch. (For instance methods)
Java always uses the runtime type, which is known as dynamic dispatch.

c++book.mif Page 116 Tuesday, April 29, 2003 2:13 PM

116

Chapter 6 Object-Oriented Programming: Inheritance

Dynamic dispatch is almost always the preferred course of action. However, dynamic dispatch incurs some run-time overhead because it requires that the program maintain extra information and that the compiler generate code to determine which member function to invoke. This
overhead was once thought to be significant, and would be incurred even if no inheritance was
actually used, and so C++ did not make it the default. This is unfortunate, because we now know
that the overhead of dynamic dispatch is relatively minor.
In order to achieve dynamic dispatch, the C++ programmer must mark the base class
method with the reserved word virtual. A virtual function uses dynamic dispatch. A non-virtual function uses static dispatch. Thus if we rewrite the print member function in class
Person as:
virtual void print( ostream & out = cout ) const
{ out << ssn << ", " << name; }

the code at the start of this section behaves as expected. This also is the minor change that fixes
operator<<.
As a general rule, if a method is overridden in a derived class, it should be declared virtual
in the base class to ensure that the correct method is always selected. In fact, a stronger statement applies: if a method might reasonably be expected to be overridden in a derived class, it
should be declared virtual in the base class. Once a method is marked virtual, it is virtual from
that point down in the inheritance hierarchy.
Only if a member function is intended to be invariant in an inheritance hierarchy (or if an
entire class is not intended to be extended) does it make sense to not mark it as virtual. Thus
getName and getSsn are not marked as virtual. In Java, these would be marked final to
signify that it is illegal to attempt to override. C++ does not have final methods or final classes,
so the lack of virtual in a method declaration is a signal to the reader that the method is
intended to be final, and those semantics are guaranteed if invoked through a base class reference (or pointer).
When the class declaration and implementation are separate, the virtual declaration must
be in the member function declaration and should not be in the separate member function definition.

6.3 Constructors, the Big Three and Inheritance

In the subclass, like all classes, we need to consider whether we can accept defaults for the
destructor, copy constructor, and copy assignment operator, or whether we need to write our
own.
6.3.1

Defaults

First, we need to decide what the defaults are. For all three, the inherited component is considered to be a data member. Thus, by default:
the copy constructor is implemented by invoking a copy constructor on the base class(es),

c++book.mif Page 117 Tuesday, April 29, 2003 2:13 PM

Constructors, the Big Three and Inheritance

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23

117

class Student
{
...
Student( const Student & rhs )
: Person( rhs ), gpa( rhs.gpa )
{ }
~Student( )
{
// automatically chains up
}
const Student & operator= ( const Student & rhs )
{
if( this != & rhs )
{
Person::operator= ( rhs );
gpa = rhs.gpa;
}
return *this;
}
...
};

Figure 6-3

Explicit implementation of the Big-Three defaults for Student

followed by invoking copy constructors on each of the newly added data members. If any
of the required copy constructors cannot be called (because they are private), then an
attempt to invoke a default copy assignment operator will generate a compiler error.
the copy assignment operator is implemented by invoking a copy assignment operator on
the base class(es), followed by invoking copy assignment operators on each of the newly
added data members. If any of the required copy assignment operators cannot be called
(because they are private, or the data member is constant or otherwise not assignable),
then an attempt to invoke a default copy assignment operator will generate a compiler
error.
the destructor is implemented by invoking destructors on each of the newly added data
members, followed by invoking the destructor on the base class(es).
6.3.2

Implementing the Big-Three Explicitly

If any default is unacceptable, then it can be implemented explicitly. Generally, if one of the
defaults is unacceptable, all three are. The typical scenario is that the base class operations are
fine, because if the base class does not support copying, it generally makes little sense to allow
the subclasses to do so. In this scenario, the newly added data members allocate memory from
the heap without cleaning it up automatically,

c++book.mif Page 118 Tuesday, April 29, 2003 2:13 PM

118

Chapter 6 Object-Oriented Programming: Inheritance

Figure 6-3 shows an explicit implementation of the defaults for the Student class. To
implement the copy constructor, we need to make sure to include a call to the base class copy
constructor in the initializer list. To implement the destructor, we simply list the additional
actions that must be taken in addition to the default. To implement the copy assignment operator,
we need to make sure that we chain up to the base class.
In short, to implement copy semantics, we must chain up to the base class, while in the
destructor the chaining up is automatic.
6.3.3

Virtual Destructor

A much trickier aspect of the Big-Three concerns polymorphism. Consider the following code:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29

Student *s1 = new Student( "123456789", "Jane", 4.0 );

Person *p1 = s1;
delete p1;
class Person
{
public:
Person( int s, const string & n = "" )
: ssn( s ), name( n )
{ }
virtual ~Person( )
{ }
const string & getName( ) const
{ return name; }
int getSsn( ) const
{ return ssn; }
virtual void print( ostream & out = cout ) const
{ out << ssn << ", " << name; }
private:
int ssn;
string name;
};
ostream & operator<< ( ostream & out, const Person & p )
{
p.print( out );
return out;
}

Figure 6-4

Person as a base class (correct version)

c++book.mif Page 119 Tuesday, April 29, 2003 2:13 PM

Abstract Methods and Classes

119

Here, s1 points at a Student object, and certainly the second assignment is legal, since
Student IS-A Person. So when delete is invoked, whose destructor is used?
Since the destructor is a member function, the answer depends on whether or not the
destructor is declared virtual. In our original code in Figure 6-1, it is not, so we invoke the
Person destructor instead of the Student destructor. This means that any memory in the
Student data fields that was allocated from the memory heap either directly (via calling new),
or indirectly (e.g. if there is a string, or vector as a data member), is never reclaimed, and
we have a subtle memory leak.
Thus we see an important rule: In a base class, a destructor should always be declared virtual to ensure polymorphic destruction. As with all virtual methods, this costs some space and
time, so classes that are intended to be final can avoid declaring their destructors virtual. The
reason it is easy to do this incorrectly is that the base class destructor should be virtual even if
defaults are used in the entire hierarchy. Otherwise, as we just mentioned, memory that was allocated indirectly, and which otherwise would have been released will leak.
Figure 6-4 shows the Person class with correct virtual declarations. The Student class
in Figure 6-1 needs no changes from the original.

6.4 Abstract Methods and Classes

In Java, abstract methods are methods that serve as placeholders, with implementations deferred
to concrete subclasses. Such methods are declared with the reserved word abstract. A class
with at least one abstract method must also be marked as abstract, and an abstract class cannot
be instantiated; instead it serves as a common superclass.
In C++, a method is abstract if:
1. It is declared virtual
2. The declaration is followed by =0
For instance, the member function area is declared abstract as follows:
virtual double area( ) const = 0;

In C++, a class is abstract if it has at least one abstract method, and like Java, abstract
classes cannot be instantiated. A C++ class with no abstract methods is not abstract. This rule is
different than Java: abstract classes in Java do not have to have an abstract method (e.g.
java.awt.event.WindowAdapter). C++ escapes from this rule by allowing a virtual
method that has an implementation to be declared abstract. The sole purpose of doing so seems
to be to allow the class to be abstract.
A common place where this trick is used is the destructor, because an abstract class must
be extended, and thus must have a virtual destructor. The destructor must have an implementation (since the derived classes chain up to it), but if we mark it as abstract, then the class is now
automatically abstract, even if no other abstract methods are present.
Abstract methods in C++ are often denoted as pure virtual methods.

c++book.mif Page 120 Tuesday, April 29, 2003 2:13 PM

120

Chapter 6 Object-Oriented Programming: Inheritance

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40

class Shape
{
public:
Shape( const string & s ) : shapeType( s ) { }
virtual ~Shape( ) { }
const string & getType( ) const
{ return shapeType; }
virtual double getArea( ) const = 0;
virtual void print( ostream & out ) const
{ out << getType( ) << " of area " << getArea( ); }
private:
string shapeType;
};
ostream & operator<< ( ostream & out, const Shape & s )
{
s.print( out );
return out;
}
class Circle : public Shape
{
public:
Circle( double r ) : Shape( "Circle" ), radius( r )
{ }
double getArea( ) const
{ return 3.14 * radius * radius; }
private:
double radius;
};
class Square : public Shape
{
// similar implementation as Circle not shown
};

Figure 6-5

Shape hierarchy with abstract base class

Figure 6-5 shows an abstract Shape class and a Circle class that implements all the
abstract methods. A Square class would be similar and is not shown (we use all three classes
later). A Shape stores its type information as private data, so this data must be initialized in the
constructor (since the subclasses do not have access to private data). Thus like Java, abstract

c++book.mif Page 121 Tuesday, April 29, 2003 2:13 PM

Slicing

121

classes do provide constructors that are used by the subclasses. The destructor at line 5 is
marked virtual (we do not make it abstract, since we have other abstract methods).
getType, shown at lines 7 and 8 is not marked virtual, to signify that it is intended as a
final method. Although subclasses can provide different implementations, if getType is
invoked through a Shape reference or pointer, we will always invoke the Shapes getType
method, which is similar to the behavior we would get in Java.
Abstract method getArea is declared at line 10, and is overridden in the Circle class.
The print method is virtual, signalling that we expect that the default implementation we have
provided may need to be overridden. The fact that Circle and Square do not override
print in our implementation does not justify removing the virtual declaration. The declaration
is there to express that the print method is not invariant and may need to be overridden.
In the Circle class, we see that the Circle constructor invokes the Shape constructor
at line 27 to initialize its inherited components.

6.5 Slicing
In our discussion of virtual member functions, we have talked about accessing derived classes
by pointers and reference variables. So why not by direct base class objects? The reason is relatively simple: it doesnt work!
Consider the following example:
Student s( 123456789, "Jane", 4.0 );
Person p( 987654321", "Bob" );
p = s;
p.print( );

Clearly the first two lines create two objects. The first object is of type Student, and the
second is of type Person, as shown in Figure 6-6. And equally, clearly, the third statement copies s into the already existing object p. As Figure 6-6 shows, p only has room for the name and

"Bob"

"Jane"

987654321

123456789

4.0

Figure 6-6

Memory layout of base class and derived class object; shaded portions are inherited data and might not be visible

c++book.mif Page 122 Tuesday, April 29, 2003 2:13 PM

122

Chapter 6 Object-Oriented Programming: Inheritance

"Jane"

"Jane"

123456789

123456789

4.0

Figure 6-7

Result of copying Person into Student

ssn data members, and so those are the only members that can be copied. As shown in
Figure 6-7, ss gpa cannot be copied into p because the p has no room for it.
As this example shows, if a derived class object is copied into a base class object, only the
base class portion is actually copied. The derived class object has been sliced; and this effect is
known in C++ as slicing.
Slicing occurs in all forms of object copying. Thus, if s is a Student
Person p = s;

constructs a new Person object, but only using a slice of s.

Slicing also occurs when objects are passed using call-by-value. For instance, even if
print is declared virtual in class Person, if we mistakenly write operator<< passing
Person using call-by-value:
ostream & operator<< ( ostream & out, Person p )
{
p.print( out );
return out;
}

then the call

cout << s << endl;

does not work, because s is passed using call-by-value, and only a slice of s is copied into p.
Inside operator<<, the print method is thus invoked on a Person object. As a result, callby-value should never be used in conjunction with inheritance: call-by-value and inheritance
dont mix.

6.6 Why Inheritance Is Harder In C++ Than Java

It is common to hear that inheritance is harder in C++ because methods are not virtual by
default. But that is simply not true. The reason that inheritance is harder in C++ than Java is
because of the slicing problem.

c++book.mif Page 123 Tuesday, April 29, 2003 2:13 PM

Why Inheritance Is Harder In C++ Than Java

123

In Java, if we want to store a collection of shapes, we can simply throw them in an array of
Shape, and then safely invoke the print and area methods, with automatic dynamic dispatch. But what we actually have is not an array of Shape objects, but rather, an array of reference variables that all reference objects that are (subtypes of) Shapes.
In C++, we cannot store an array of Shape, because of slicing. As soon as we would try
to place a derived class object into the array, it would be sliced. All calls would be on Shape
objects. Of course Shape is abstract, so that is another problem, but the abstractness of Shape
is not really the critical issue. It is the slicing problem.
We can use an array of pointers to Shapes, and then the code looks exactly like Java. In
fact if:
1. we declare every method as virtual
2. only allocate objects using new
3. access all objects by pointers
4. use collections of pointers
we can for the most part do everything that we do in Java, with one important exception: Java
does garbage collection, while in C++ we must eventually call delete. And thats the hard
part.
Using inheritance in C++ implies that we must make significant use of pointers, and allocate objects on the memory heap. Which means we are stuck with the thorny issue of cleaning
up memory, which is quite a nuisance in C++ and notoriously error-prone.
As an example of how we use pointers with inheritance to achieve polymorphic behavior,
the code in Figure 6-8 shows how we store Circles and Squares (and in general any kind of
Shape) in a single collection.
The main routine, shown at lines 36 to 40 simply invokes testShapes mostly so we
can make clear that any allocated memory in testShapes must be cleaned up. In
testShapes, line 25 is the C++ equivalent of creating an empty ArrayList, and then lines
27 to 29 is the equivalent of calling add. We do not have any slicing problems since the
vector is storing pointers, so we are never copying Shape objects themselves. We can then
pass the vector to methods such as printArray and totalArea, which will scan through
the vector, and invoke the appropriate print and area methods on the Shapes being
pointed at.
In testShapes, arr is a local variable allocated on the runtime stack. This means that
when testShapes returns, arrs destructor will be called. This will free the vector, but
not the Shapes that were being pointed at. Since testShapes created objects from the memory heap, it must reclaim them, and this is the job of cleanup.
cleanup simply steps through the array invoking delete. Observe that this would not
work if the Shape class did not correctly declare a virtual destructor. printShapes and
totalArea are similar, and simply illustrate different syntax. Since operator<< accepts
any Shape, and *a[i] is an object that is a (subtype of) Shape we can pass it to

c++book.mif Page 124 Tuesday, April 29, 2003 2:13 PM

124

Chapter 6 Object-Oriented Programming: Inheritance

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40

void printShapes( const vector<Shape*> & a )

{
for( int i = 0; i < a.size( ); i++ )
cout << *a[ i ] << endl;
}
double totalArea( const vector<Shape*> & a )
{
double sum = 0.0;
for( int i = 0; i < a.size( ); i++ )
sum += a[ i ]->getArea( );
return sum;
}
void cleanup( vector<Shape *> & a )
{
for( int i = 0; i < a.size( ); i++ )
delete a[ i ];
}
void testShapes( )
{
vector<Shape *> arr;
arr.push_back( new Circle( 2.0 ) );
arr.push_back( new Square( 2.0 ) );
arr.push_back( new Square( 4.0 ) );
printShapes( arr );
cout << "Total area is " << totalArea( arr ) << endl;
cleanup( arr );
}
int main( )
{
testShapes( );
return 0;
}

Figure 6-8

Heterogeneous collection of shapes

operator<< at line 4. The getArea method can be invoked using the normal -> operator,
as shown at line 12. The code here is not much different than the equivalent Java code, except
that we must write cleanup.

c++book.mif Page 125 Tuesday, April 29, 2003 2:13 PM

Type Conversions

125

6.7 Type Conversions

In Java, the typecast operator is used in conjunction with inheritance to cast down the inheritance hierarchy (i.e. from a superclass or interface reference down to a subclass reference). This
is used to invoke methods on the actual object that were not defined by the compile-time reference type. For instance, in C++, if we have the following line of code,
Person *p = new Student( "123456789", "Jane", 4.0 );

then invoking p->gpa() is not legal.

In C++, one can use a typecast to cast down the inheritance hierarchy. This makes most
sense when casting a pointer type or a reference type, since as weve already seen, accessing
objects directly in C++ doesnt work well with inheritance because of slicing. So in the code
above, after
Student *s = (Student *) p;

s->gpa() is legal, as would be ((Student *) p)->gpa().

Although this code works, it is not the preferred way of doing a downcast. This is because
it would run even if p was actually pointing at a Person at runtime. In other words, unlike
Java, this typecast is not checked at runtime. In Java this code would generate a
ClassCastException, but in C++ the code may well run, producing garbage.
To alleviate this problem, C++ introduced the dynamic cast. If we use
Student *s = dynamic_cast<Student *>( p );

then dynamic_cast returns NULL if the runtime type of p is not compatible with s.
dynamic_cast can also be used to cast references:
Student jane( "123456789", "Jane", 4.0 );
Person & pref = jane;
Student & sref = dynamic_cast<Student &>( pref );

If the cast fails because the object is of the wrong type, or if the cast is not applied to a
pointer or a reference, then a bad_cast exception is thrown.
The dynamic_cast only works with types that are polymorphic, meaning that the class
type must have at least one virtual method.
Casting in C++ is much less common than in Java, because most uses of Java casting
involves casting down from Object in generic collections (e.g. the java.util classes). As
we will see in Chapter 7, C++ uses a different mechanism, known as templates, to support
generic algorithms, removing almost all uses of casting.
C++ does not provide an instanceof operator. With pointer variables and the dynamic
cast, the instanceof operator can be simulating by invoking a dynamic cast and verifying
that the return value is not NULL. However, using instanceof in Java is relatively rare in
well-designed object-oriented programs, and the idiom in C++ is even more rare.

c++book.mif Page 126 Tuesday, April 29, 2003 2:13 PM

126

Chapter 6 Object-Oriented Programming: Inheritance

1
2
3
4
5
6
7
8
9
10
11
12
13
14

class Printable
{
public:
virtual ~Printable( ) { }
virtual void print( ostream & out = cout ) const = 0;
};
class Serializable
{
public:
virtual ~Serializable( ) { }
virtual void readMe( istream & in = cin ) = 0;
virtual void writeMe( ostream & out = cout ) const = 0;
};

Figure 6-9

Example of declaring interfaces

6.8 Multiple Inheritance and Interfaces

C++ does not provide a special interface type. Instead, one simply implements pure abstract
base classes which are classes that contain only abstract member functions. For instance,
Figure 6-9 shows two interfaces, Printable and Serializable. We can now have a class
that implements both of these interfaces, as shown in Figure 6-10. As in Java, Person is now
type compatible with Printable and Serializable, and of course, Person could also
easily extend a class that was not a Java-style interface.
Since C++ does not distinguish between interfaces and non-interfaces, Person could
extend several classes that had implementations. When properly done, this can be very useful.
For instance, C++ defines an istream class, an ostream class, and then an iostream class
that supports both input and output at the same time.
However, multiple implementation inheritance, in which two or more base class are not
done in the interface style creates significant complications because the implementations that are
inherited could be conflicting.
1
2
3
4
5
6
7
8
9
10

class Person : public Printable, public Serializable

{
public:
...
// Complicated, so probably implement in .cpp
void print( ostream & out = cout ) const;
void readMe( istream & in = cin );
void writeMe( ostream & out = cout ) const;
...
};

Figure 6-10

Example of implementing multiple interfaces

c++book.mif Page 127 Tuesday, April 29, 2003 2:13 PM

Multiple Inheritance and Interfaces

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44

127

class Person
{
public:
Person( const string & n, const string & t )
: name( n ), ptype( t ) { }
virtual ~Person( )
{ }
private:
string name;
string ptype;
};
class Student : public Person
{
public:
Student( const string & n )
: Person( "Student", n ) { ... }
int getHours( ) const;
// number of credit hours taken
private:
int hours;
};
class Employee : public Person
{
public:
Employee( const string & n )
: Person( "Employee", n ) { ... }
int getHours( ) const;
// number of vacation hours left
private:
int hours;
};
class StudentEmployee : public Student, public Employee
{
public:
// Constructor sets type to StudentEmployee???
// ??? getHours ???
private:
// ??? name ???
// ??? hours ???
};

Figure 6-11

Example of conflicting implementations (needs fixing)

For instance, consider the scenario in Figure 6-11. The Person class has subclasses

c++book.mif Page 128 Tuesday, April 29, 2003 2:13 PM

128

Chapter 6 Object-Oriented Programming: Inheritance

Student and Employee, and then StudentEmployee extends both Student and
Employee, thus having the functionality of both.
But now consider the problems that have to be resolved in StudentEmployee.
Since getHours is defined in both classes, which is inherited? Since there is an ambiguity, either both are inherited, but then invoker must make plain which of the methods is to be
used, or both are overridden in the StudentEmployee class. The first possibility leads to
ugly code; in the second, StudentEmployee will not be computing the same information as
both Student and Employee, violating at least one IS-A relationship. Most likely, we would
want to just change the method name to avoid this conflict.
More tricky is the memory layout. Since Student and Employee both define an
hours data member, it seems that StudentEmployee needs two copies of hours. This is
certainly true, as seen if we change the name of one of the getHours, but leave the data member in tact. A basic tenet of inheritance is that the subclasses inherit the superclass data.
But now what about name? Both Student and Employee have a name data member,
so according to our logic, there should be two copies. But thats no good, since Student and
Employee inherited name from Person. By default however, we get two copies. To get only
one copy, we must use virtual inheritance. When Person is extended, the subclasses use
virtual to signal that any inherited data from Person is stored in a different manner. When
multiple such subclasses are themselves extended, the compiler will be able to distinguish
between data members like hours that were created in the subclasses and data members like
name that are really all part of a single ancestor class.
If the ancestor class supplies some of the data, it makes sense that its constructor should be
invocable. Thus with virtual inheritance, not only can a superclass constructor be invoked, but
also the constructor of an ancestor class that is virtually extended can be invoked in the initializer list. All other initializations of the ancestors data by other initializers in the initializer list
are ignored.
The result of all these changes, with an illustration of the syntax is shown in Figure 6-12.
Observe the deficiency of this approach: although the conflict is at StudentEmployee, it is
the Student and Employee classes that are responsible for using virtual inheritance. Thus if
the base classes have not already declared their use of virtual inheritance, it is difficult to inherit
from both of them without having to disturb existing code.

6.9 Protected and Friends

Under normal public inheritance, protected members in C++ are visible only in the class and any
subclasses. The visibility rules do not use the arcane rules from Java that depend on the invocation being through a subtype reference. As with Java, use of protected data members is often
poor design, as it encourages subclasses to access these data members directly, and results in
more changes being required should their be changes to the data members. Protected member
functions are used to provide an implementation of a helper function that is logically private, but
can be used by subclasses.

c++book.mif Page 129 Tuesday, April 29, 2003 2:13 PM

Protected and Friends

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44

class Person
{
public:
Person( const string & t, const string & n )
: ptype( t ), name( n ) { }
virtual ~Person( )
{ }
private:
string ptype;
string name;
};
class Student : virtual public Person
{
public:
Student( const string & n, int h )
: Person( "Student", n ), hours( h ) { ... }
int getCreditHours( ) const;
// credit hours taken
private:
int hours;
};
class Employee : virtual public Person
{
public:
Employee( const string & n, int h )
: Person( "Employee", n ), hours( h ) { ... }
int getVacationHours( ) const;
// vacation hours left
private:
int hours;
};
class StudentEmployee : public Student, public Employee
{
public:
StudentEmployee( const string & n, int ch, int vh )
: Person( "StudentEmployee", n ),
Student( "ignored", ch ), Employee( "ignored", vh )
{ }
private:
// one name, two hours
};

Figure 6-12

Example of implementing multiple interfaces

129

c++book.mif Page 130 Tuesday, April 29, 2003 2:13 PM

130

Chapter 6 Object-Oriented Programming: Inheritance

1
2
3
4
5
6
7
8
9
10
11
12

class NewCell : private IntCell

{
public:
explicit NewCell( int initialValue = 0 )
: IntCell( initialValue ) { }
int get( ) const
{ return getValue( ); }
void put( int value )
{ setValue( value ); }
};

Figure 6-13

Private inheritance to change an interface. Avoids delegation

Friendship in C++ is not inherited. If a class or method is a friend of class Base, and class
Derived extends Base, then the class or method can only access the portions of Derived
that were inherited from Base. It cannot access the private details of Derived.

6.10 Technical Differences

Our discussion mentioned several differences between inheritance in Java and C++. In addition
to those differences, we mention several technical points that sometimes are of interest to C++
programmers. (There are also a host of complex and subtle points that we avoid discussing; after
all these are two different languages.)
6.10.1 No Root Class
C++ does not have a root base class that is similar to Object. As we discuss in Chapter 7, C++
uses templates in many of the places where Object is used in Java.
6.10.2 Non-Public Inheritance
C++ allows not only public inheritance, but also private (and protected) inheritance. The
default is private inheritance, which you will certainly run into when you forget to type the
reserved word public.
In private inheritance, all of the inherited base components become private in the derived
class. An IS-A relationship no longer holds (except inside the implementation of the derived
class), so a derived class pointer or reference can no longer be passed as an argument that
requires a base class pointer or reference.
Private inheritance sometimes models a HAS-A relationship, and is typically used to
change the visible interface of an existing object. For instance, if in the IntCell class in
Chapter 4 we decided that we did not like the names getValue and setValue, but instead
liked get and put, we could use private inheritance instead of the Java style of delegation. This
is shown in Figure 6-13. If we used public inheritance, then NewCell would be type compati-

c++book.mif Page 131 Tuesday, April 29, 2003 2:13 PM

Technical Differences

131

ble with IntCell and would still have getValue and setValue. With private inheritance,
getValue and setValue are only visible inside of NewCell, though they can be used in
the implementation of get and put, but not outside of the NewCell class.
6.10.3 Reducing Visibility When Overriding
C++ allows you to reduce the visibility of a method. If a method is public in a base class and private in a derived class, then it is visible if invoked through a base class reference and invisible
otherwise. This is not a pretty picture because it violates the contract of an IS-A relationship.
Specifically, what would happen if a public virtual function in a base class is overridden by a
private function, and then the function is invoked through a base class pointer that is actually
pointing at a derived class object? In C++ the method is invoked using dynamic dispatch and so
the private method is used. This seems like bad semantics.
6.10.4 New Methods Hide Originals in Base Classes
The overloading algorithms in C++ and Java are different. Consider the code in Figure 6-14. In
the code, we have a base class and a derived class. The base class declares member function
foo with no parameters, and the derived class declares member function foo with an int
parameter.
In Java, only the zero-parameter foo is available when foo is invoked by a Base reference. The same is true in C++. However, in Java, both versions of foo are available when foo
is invoked by a derived class reference. However, in C+ this is not true. Instead, when a method
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

class Base
{
public:
virtual void foo( );
};
class Derived : public Base
{
public:
void foo( int x );
};
void test( Base & arg1, Derived & arg2 )
{
arg1.foo( );
// OK
arg1.foo( 4 );
// Illegal, as expected
arg2.foo( 4 );
// Legal, as expected
arg2.foo( );
// Illegal; not like Java
}

Figure 6-14

Derived class methods hide base class methods with same name

c++book.mif Page 132 Tuesday, April 29, 2003 2:13 PM

132

Chapter 6 Object-Oriented Programming: Inheritance

is declared in a derived class, it hides all methods of the same name in the base class. Thus foo
with no parameters is no longer accessible through a derived class reference, even though it
would be accessible through a base class reference.
There are two ways around this problem. The first is to override all of the hidden methods
and redefine them in the derived class with an implementation that chains to the base class. Thus
in class Derived, we add
void foo( ) { Base::foo( ); }

// place in class Derived

The other alternative is to provide a using declaration in the derived class. In class
Derived, we add
using Base::bar;

// place in class Derived

One reason why this rule is important is that an accessor hides a mutator. Most likely this
was unintentional, but many compilers will warn you, and what they are saying, in effect, is that
you made a slight change to the signature when you overrode the original member function. Pay
attention to those warnings.
6.10.5 Overriding With Different Return Type
C++ allows the return type of an overriding method to be slightly changed. Specifically, if the
original return type is a pointer (or reference) to a class type B, the new return type may be a
pointer (or reference) to class D, provided D is a publicly derived class of B. This condition simply states that you can change the return type to be a subclass of the original return type, as long
as an IS-A relationship holds. This is useful for overriding methods such as
virtual const Base & Base::operator++( );

with
virtual const Derived & Derived::operator++( );

6.10.6 Type Compatibility

In Java, a variable that is an array of subclass types is type compatible with an array of superclass types.
For instance,
Object [ ] arr = new Employee[ 10 ];

// Legal Java

This type compatibility doesnt make much sense, when one sees that
arr[ 0 ] = new Basketball( );

// Compiles in Java

must be allowed to compile, since at compile time arr[0] is of type Object and
BasketBall is an Object. Of course at runtime this throws an exception, but the whole
point of the typing system is to try to avoid runtime problems.
In C++, the equivalent scenario would be an array of pointers, and C++ does not allow an
array of subclass pointers to be type-compatible with an array of superclass pointers.

c++book.mif Page 133 Tuesday, April 29, 2003 2:13 PM

Key Points

133

6.10.7 Reflection
C++ has little support for reflection. About all you can do is invoke the typeid method by
passing an expression. It returns an object of type type_info representing the runtime type of
the expression. Standard C++ guarantees that this object contains a public data member called
name, and that you can compare type_info objects with == and !=. Thus,
Person *p = new Student( 123456789, "Jane", 4.0 );
Student *s = p;
cout << typeid( *s ).name << endl;
// prints name
cout << ( typeid( *s ) == typeid( Person ) ) << endl; // false (0)

6.10.8 Dynamic Dispatch in Constructors

In Java, when an instance method is invoked (via this) in a constructor, dynamic dispatch is
always used. In C++, dynamic dispatch is not used. This is really of minimal consequence, since
invoking non-final instance methods in Java is known to be a dangerous practice.
6.10.9 Default Parameters
Java does not allow default parameters. The most likely reason is that in C++ default parameters
are bound at compile time. Changing the default value in a derived class is unsafe, because
doing so can create an inconsistency with virtual functions, which are bound at run time. Thus if
you provide default parameters in a base class method, keep them unchanged in the overridden
derived class method.

6.11 Key Points

The IS-A relationship is modelled in C++ with public inheritance.
Private inheritance is the (probably wrong) default and can be used to model a HAS-A
relationship.
Because of slicing, polymorphism requires that objects be accessed with either base class
pointers or references.
Dynamic dispatch is performed only if a member function is declared virtual.
Constructors are never declared virtual.
Destructors are always declared virtual in a base class.
Abstract methods are implemented as virtual member functions with =0 at the end of their
declarations. A class with at least one abstract method is an abstract class.
An initializer list call to a base class constructor replaces the Java call to super.
Chaining to a base class method is achieved using the :: scope operator.
If operator= is overridden for the derived class, the base class operator= should be
invoked as part of the implementation.
The dynamic cast is the preferred form of downcasting in C++. It can be used to implement instanceof, though it is rarely needed in C++.

c++book.mif Page 134 Tuesday, April 29, 2003 2:13 PM

134

Chapter 6 Object-Oriented Programming: Inheritance

C++ does not have interfaces, but the effect can be achieved with an abstract base class
that contains only abstract methods and a virtual destructor.
C++ allows multiple inheritance. If two implementations conflict, both of those implementations should have been declared using virtual inheritance to avoid replication of data
members.
Protected members are visible in derived class implementations, regardless of how they
are accessed.
Friendship is not inherited.
There is no root class that is equivalent to Object. Instead templates are used to implement generic algorithms.
Although it is bad style, C++ allows the reduction of visibility of a method in a derived
class.
Methods declared in the derived class hide the base class methods with the same name.
The return type of a derived class method can be changed from the base class method if it
is replaced with a subclass of the original return type.
An array of a derived type is not type compatible with an array of the base type.
C++ does not support reflection.

6.12 Exercises
1. What is the C++ equivalent of the extends clause?
2. What does virtual do?
3. What is slicing?
4. Why is it bad to declare a base class parameter using call-by-value?
5. How is the superclass constructor invoked in C++?
6. What are the defaults for the Big-Three in a derived class?
7. Why is the default base class destructor unacceptable?
8. How are abstract classes declared in C++?
9. What is a pure virtual function?
10. Why is inheritance harder to do in C++ than in Java?
11. How are heterogeneous collections stored in C++?
12. What are the semantics of dynamic_cast?
13. How is the Java interface programmed in C++?
14. What are the rules for protected and friends with respect to inheritance?
15. What is virtual inheritance?
16. What is private inheritance?
17. What does it mean for a derived class method to hide a base class method?
18. Add a Rectangle class to the hierarchy in Figure 6-5 and modify Figure 6-8 appropriately to include objects of type Rectangle.

c++book.mif Page 135 Tuesday, April 29, 2003 2:13 PM

Exercises

135

19. Define an abstract base class called Employee that contains a name (string), a social
security number (string), and the respective accessor functions and contains an abstract
method that gets and sets a salary. It also contains a method called print whose task is to
output the name and social security number. You should not use protected members.
Include a two-parameter constructor, using initializer lists, and give all parameters
defaults. Carefully decide which members should be virtual. Next, derive a class called
Hourly that adds a new data member to store an hourly wage (double). Its print
method must print the name, social security number, and salary (with the phrase
"per hour"). It will certainly want to call the base class print. Provide an accessor
and mutator for the salary, and make sure that its constructor initializes a salary. Next,
derive another class called Salaried that adds a new data member to store a yearly salary (double). Its print method must print the name, social security number, and salary
(with phrase "annual"). Provide an accessor and mutator for the salary, and make sure
that its constructor initializes a salary. Exercise the classes by declaring objects of type
Salaried and Hourly via constructors and calling their print methods. Provide a
single operator<< that prints an Employee (by calling print). This method will
automatically work for anything in the Employee hierarchy. In order to hold all employees, create a class called Roster that is able to hold a variable number of Employee *
objects. Roster should have a vector of Employee *. Provide the capability to add
an employee, and print the entire roster of employees. To add an employee, Roster::add is passed a pointer to an Employee object and calls vector::push_back. Don't worry about error checks. To summarize, Roster has public
methods named add and print. Write a short test program, in which you create a Roster object, call new for both kinds of Employee, sending the result to Roster::add,
and output the Roster via a call to print.
20. Define a hierarchy that includes Person, Student, Athlete, StudentAthlete,
FootballPlayer, BasketballPlayer, StudentFootballPlayer,
StudentBasketballPlayer, StudentFootballAndBasketballPlayer.
Give a Person a name, a Student a GPA, an Athlete a uniform number, a
FootballPlayer a boolean representing true for offense, and false for defense, a
BasketballPlayer a scoring average (as a double). Define appropriate data representations, constructors, destructors, and methods, and make use of virtual inheritance.
Observe that StudentFootballAndBasketballPlayer will have two uniform
numbers. Write a test program that obtains both numbers.

c++book.mif Page 136 Tuesday, April 29, 2003 2:13 PM

136

Chapter 6 Object-Oriented Programming: Inheritance

c++book.mif Page 137 Tuesday, April 29, 2003 2:13 PM

H A P T E R

Templates

N Java, inheritance is used to write type-independent

code. In C++, however, a different alternative is used: the template.
In this chapter, we begin by reviewing how generic algorithms are implemented in Java,
and then see the basics of how templates are written for functions and classes. We will also see
how templates are used to implement function objects, which in Java are implemented via inheritance and interfaces. We will discuss how templates affect separate compilation, and finally, we
will briefly mention some of the advanced uses of templates.

7.1 Review of Generic Programming in Java

Java uses inheritance to implement generic programming. For instance, all of the Collections
API classes, for instance, are written in terms of the Object class, and since all objects in Java
are instances of this class, these collections are usable by all class types. In other parts of the language, a generic algorithm is implemented in terms of an interface type, and thus works for any
class that implements the interface.
However, this approach has some limitations. For instance, since primitive types are not
objects, they cannot be directly added into a standard Java collection. Instead, wrapper classes
such as Integer must be used. Java 1.5 will fix this deficiency by automatically wrapping
primitives when they are passed as arguments to parameters of type Object (a process known
as boxing). A second problem is that using inheritance leads to excessive downcasting because
the return type of the generic methods tends to be Object or an interface type. The downcasting is not only time consuming, it tends to delay typing errors from compile-time until runtime,
which is less desirable. Java 1.5 adds generics to the language, which will detect many typing
errors at compile time.

137

c++book.mif Page 138 Tuesday, April 29, 2003 2:13 PM

138

Chapter 7 Templates

1
2
3
4
5
6
7
8
9
10
11
12
13
14

// Return the maximum item in array a.

// Assume a.size( ) > 0.
// Comparable objects must provide operator<
template <typename Comparable>
const Comparable & findMax( const vector<Comparable> & a )
{
int maxIndex = 0;
for( int i = 1; i < a.size( ); i++ )
if( a[ maxIndex ] < a[ i ] )
maxIndex = i;
return a[ maxIndex ];
}

Figure 7-1

findMax function template

7.2 Function Templates

C++ provides function templates. A function template is not an actual function, but instead is a
pattern for what could become a function. Figure 7-1 illustrates a function template findMax
that is virtually identical to the (correct) routine for string shown in Figure 4-26. Line 4 indi1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

#include "IntCell.h"
#include <string>
using namespace std;
int main( )
{
vector<int>
vector<double>
vector<string>
vector<IntCell>
vector<int>

v1(
v2(
v3(
v4(
v5(

37
40
80
75
75

);
);
);
);
);

// Additional code to fill in the vectors not shown

cout
cout
cout
cout
cout

<<
<<
<<
<<
<<

findMax(
findMax(
findMax(
findMax(
findMax(

v1
v2
v3
v4
v5

)
)
)
)
)

<<
<<
<<
<<
<<

endl;
endl;
endl;
endl;
endl;

return 0;
}

Figure 7-2

Using the findMax function template

//
//
//
//
//

OK: Comparable = int

OK: Comparable = double
OK: Comparable = string
Illegal: no operator<
OK: uses findMax, line 15

c++book.mif Page 139 Tuesday, April 29, 2003 2:13 PM

Function Templates

139

cates that Comparable is the template argument: it can be replaced by any type to generate a
function (both class and typename can be used interchangeably here). For instance, if a call
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43

const int & findMax( const vector<int> & a )

{
int maxIndex = 0;
for( int i = 1; i < a.size( ); i++ )
if( a[ maxIndex ] < a[ i ] )
maxIndex = i;
return a[ maxIndex ];
}
const double & findMax( const vector<double> & a )
{
int maxIndex = 0;
for( int i = 1; i < a.size( ); i++ )
if( a[ maxIndex ] < a[ i ] )
maxIndex = i;
return a[ maxIndex ];
}
const string & findMax( const vector<string> & a )
{
int maxIndex = 0;
for( int i = 1; i < a.size( ); i++ )
if( a[ maxIndex ] < a[ i ] )
maxIndex = i;
return a[ maxIndex ];
}
const IntCell & findMax( const vector<IntCell> & a )
{
int maxIndex = 0;
for( int i = 1; i < a.size( ); i++ )
if( a[ maxIndex ] < a[ i ] )
maxIndex = i;
return a[ maxIndex ];
}

Figure 7-3

findMax function template expanded (including a bad expansion for Int-

c++book.mif Page 140 Tuesday, April 29, 2003 2:13 PM

140

Chapter 7 Templates

to findMax is made with a vector<string> as parameter, then a function will be generated by replacing Comparable with string. A general rule of thumb is that a function template is not a function; rather, it is simply a function wannabee. When it is expanded, the result is
a function. Figures 7-2 and 7-3 illustrate that function templates are expanded automatically as
needed. Specifically, Figure 7-2 shows the code that the programmer would write, and then
Figure 7-3 shows the code that the compiler generates internally to expand the function template
into real functions. An expansion for each new type generates additional code; this is one example of code bloat, when it occurs in large projects. However, once the template is expanded for a
particular type, it is not expanded a second time for the same type. Thus internally, in Figure 7-3,
only one function is shown for the expansion with int.
Note also, that in Figure 7-2, the call at line 18 will result in a compile-time error. This is
because when Comparable is replaced by IntCell, in Figure 7-3 line 39 becomes illegal:
there is no operator< defined for IntCell. Since this line (line 39 in Figure 7-3) is internally generated, the compiler will most likely flag the calling code (line 18 in Figure 7-1) and
the corresponding line in the template declaration (line 10 in Figure 7-2) in the error message
that it outputs. To help the users of a template avoid these kinds of errors, it is customary to
include, prior to any template, comments that explain what assumptions are made about the template argument(s). This includes assumptions about what kinds of constructors are required. In
findMax, a zero-parameter constructor and copy constructor are both required by vector,
but it is reasonable to assume that the user has satisfied the requirements of the parameters, and
thus we have not included those requirements in the comments.
When line 18 in Figure 7-1 is commented out, the result is that there are three separate
overloaded versions of findMax that have been expanded from the template.
Because template arguments can assume any class type, when deciding on parameter passing and return passing conventions, it should be assumed that template arguments are not primitive types. That is why we have returned by constant reference.
Error messages relating to function templates are generally of two types. Type independent errors, such as missing semicolons, are detected when the template is compiled. Type sensitive errors, such as undefined operators and methods are detected when a particular expansion
causes a problem.
Because a function template is not a function (it is just a wannabee), function templates
can be placed in header files without causing multiple definitions. In fact, this is probably the
simplest way of compiling function templates.
Not surprisingly, there are many arcane rules that deal with function templates. Most of
the problems occur when the template cannot provide an exact match for the parameters, but can
come close. There must be ways to either declare or resolve and ambiguity. For instance, consider the code in Figure 7-4, which illustrates a max2 function template, and also a specific
max2 function that is not a template. Four calls to max2 are made in main. The first two are
expanded with int and double as Comparable, respectively. The third call, at line 25, uses
the non-template function that is defined at lines 12 to 15. Thus if there is an expandable tem-

c++book.mif Page 141 Tuesday, April 29, 2003 2:13 PM

Class Templates

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29

141

#include "IntCell.h"
#include <string>
using namespace std;
template <typename Comparable>
const Comparable & max2( const Comparable & lhs,
const Comparable & rhs )
{
return lhs > rhs ? lhs : rhs;
}
const string & max2( const string & lhs, const string & rhs )
{
return lhs > rhs ? lhs : rhs;
}
int main( )
{
string s = "hello";
int
a = 37;
double b = 3.14;
cout
cout
cout
cout

<<
<<
<<
<<

max2(
max2(
max2(
max2(

a,
b,
s,
a,

a
b
s
b

)
)
)
)

<<
<<
<<
<<

endl;
endl;
endl;
endl;

//
//
//
//

OK: expand with int

OK: expand with double
OK: not a template
Ambiguous

return 0;
}

Figure 7-4

max2 function might be ambiguous

plate and a non-template, and both are equivalent matches, then the non-template wins. The
fourth call, at line 26, is ambiguous because whether we expand with either an int or double,
in both cases, one parameter is an exact match, and one is an approximate match.

7.3 Class Templates

In the simplest version, a class template works much like a function template. Figure 7-5 shows
an ObjectCell class template, that is virtually identical to the IntCell class seen earlier in
Figure 4-6, but works for any type, Object, that has a zero-parameter constructor, a copy constructor, and an assignment operator. Since we presume that objects have these requirements
normally, we have omitted the comments, letting the typename Object speak for itself. Note
that Object is passed by constant reference, since it could be large. Also, notice that the
default parameter for the constructor is not 0, because 0 might not be a valid Object. Instead,
the default parameter is an Object created with its zero-parameter constructor.

c++book.mif Page 142 Tuesday, April 29, 2003 2:13 PM

142

Chapter 7 Templates

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20

#ifndef _OBJECTCELL_H
#define _OBJECTCELL_H
template <typename Object>
class ObjectCell
{
public:
explicit ObjectCell( const Object & initValue = Object( ) )
: storedValue( initValue )
{ }
const Object & getValue( ) const
{ return storedValue; }
void setValue( const Object & val )
{ storedValue = val; }
private:
Object storedValue;
};
#endif

Figure 7-5

Class template that stores any Object value

Figure 7-6 shows how ObjectCell can be used to store objects of several types. Notice
that ObjectCell is not a class; it is only a class template. Like function templates, class templates are wannabee classes. In Figure 7-6, the actual classes are ObjectCell<int> and
ObjectCell<double>.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17

#include "ObjectCell.h"
#include <iostream>
using namespace std;
int main( )
{
ObjectCell<int>
m1;
ObjectCell<double> m2( 3.14 );
m1.setValue( 37 );
m2.setValue( m2.getValue( ) * 2 );
cout << m1.getValue( ) << endl;
cout << m2.getValue( ) << endl;
return 0;
}

Figure 7-6

Using the ObjectCell class template

c++book.mif Page 143 Tuesday, April 29, 2003 2:13 PM

Class Templates

143

It should be clear that the vector class is in reality a class template. It turns out that the
string class is an instantiated class template (the class template is basic_string). A third
class template is complex, which is most often instantiated as complex<double> and is
found in the standard header file complex. And finally, ostream and istream are actually
instantiations of class template basic_ostream and basic_istream.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36

#ifndef _MATRIX_H
#define _MATRIX_H
#include <vector>
using namespace std;
template <typename Object>
class Matrix
{
public:
Matrix( int rows, int cols ) : array( rows )
{ setNumCols( cols ); }
int
{
int
{

numrows( ) const
return array.size( ); }
numcols( ) const
return numrows( ) > 0 ? array[ 0 ].size( ) : 0; }

void setNumRows( int rows )

{ array.resize( rows ); }
void setNumCols( int cols )
{ for( int i = 0; i < numrows( ); i++ )
array[ i ].resize( cols );
}
void setDimensions( int rows, int cols )
{ setNumRows( rows ); setNumCols( cols ); }
const vector<Object> & operator[]( int row ) const
{ return array[ row ]; }
vector<Object> & operator[]( int row )
{ return array[ row ]; }
private:
vector< vector<Object> > array;
};
#endif

Figure 7-7

Java class to represent two-dimensional array of double

c++book.mif Page 144 Tuesday, April 29, 2003 2:13 PM

144

Chapter 7 Templates

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

#include <iostream>
#include "Matrix.h"
using namespace std;
int main( )
{
Matrix<int> m( 2, 2 );
m[ 0 ][ 0 ] = 1; m[ 0 ][ 1 ] = 2;
m[ 1 ][ 0 ] = 3; m[ 1 ][ 1 ] = 4;
cout << "m has " << m.numrows( ) << " rows and "
<< m.numcols( ) << " cols." << endl;
cout << m[ 0 ][ 0 ] << " " << m[ 0 ][ 1 ] << endl <<
m[ 1 ][ 0 ] << " " << m[ 1 ][ 1 ] << endl;
return 0;
}

Figure 7-8

Using the Matrix class

Many of the classes that we have seen earlier are excellent candidates to be class templates. Two obvious examples are the matrix class from Section 5.6 and the queue class in
Section 4.6.6. Figure 7-7 shows a Matrix class template, that is virtually identical to the
MatrixOfDouble class seen earlier in Figure 5-15, but works for any type, Object, that has
the requisites of a vector. The Matrix class template is easily used, as shown in Figure 7-8.

7.4 Separate Compilation

If we implement a class template entirely in its declaration, then as we have seen there is very
little syntax. Many class templates are, in fact, implemented this way because historically compiler support for separate compilation of templates has been weak and platform specific. Thus,
in many cases, the entire class with its implementation is placed in a single header file. Popular
implementations of the standard library follow this strategy in implementing its class templates.
The recent standard has attempted to rectify this situation, and so separation of the class
templates declaration from its implementation, which has long been part of the language, is
becoming more appealing. The ideas are the same for templates as for non templates.
Figure 7-9 shows the declaration for the ObjectCell class template. This part is, of
course, simple enough, since it is just a subset of the entire class template that we have already
seen.
For the implementation, we have a collection of function templates. This means that each
function must include the template declaration, and when using the scope operator, the name of
the class must be instantiated with the template argument. Thus in Figure 7-10, the name of the
class is ObjectCell<Object>. Although the syntax seems innocuous enough, it can get

c++book.mif Page 145 Tuesday, April 29, 2003 2:13 PM

Separate Compilation

145

quite cumbersome. For instance, to define operator= in the specification requires no extra
baggage. In the implementation, we would have the brutal:
template <typename Object>
const ObjectCell<Object> &
ObjectCell<Object>::operator= ( const ObjectCell<Object> & rhs )
{
if( this != &rhs )
storedValue = rhs.storedValue;
return *this;
}

Even with this, the issue now becomes how to organize the class template declaration and
the member function template definitions. The main problem is that the implementations in
Figure 7-10 are not actually functions; they are still wannabees. They are not even expanded
when the ObjectCell template is instantiated. Each member function template is expanded
only when it is invoked.
7.4.1

Everything in the Header

The first option is to put both the declaration and implementation in the header file. This would
not work for classes, since we could get multiply defined functions if several different source
files had include directives that processed this header, but since everything here is a wannabee,
there is no problem.
With this strategy, it might be a little easier for reading purposes to simply have the header
file issue an include directive (prior to the #endif) to automatically read in the implementation
file. With this strategy, the .cpp files that store the templates are not compiled directly.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

#ifndef _OBJECTCELL_H
#define _OBJECTCELL_H
template <typename Object>
class ObjectCell
{
public:
explicit ObjectCell( const Object & initValue = Object( ) );
const Object & getValue( ) const;
void setValue( const Object & val );
private:
Object storedValue;
};
#endif

Figure 7-9

Class template that stores any Object value

c++book.mif Page 146 Tuesday, April 29, 2003 2:13 PM

146

Chapter 7 Templates

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

#include "ObjectCell.h"
template <typename Object>
ObjectCell<Object>::ObjectCell( const Object & initValue )
: storedValue( initValue )
{
}
template <typename Object>
const Object & ObjectCell<Object>::getValue( ) const
{
return storedValue;
}
template <typename Object>
void ObjectCell<Object>::setValue( const Object & val )
{
storedValue = val;
}

Figure 7-10

7.4.2

Implementation of ObjectCell class template

Explicit Instantiation

On some compilers we can achieve many of the benefits of separate compilation if we use
explicit instantiation. In this scenario, we set up the .h and .cpp files as would normally be
done for classes. Thus both Figure 7-9 and Figure 7-10 would be exactly as currently shown.
The header file would NOT have an include directive reading the implementation. The main
routine would only have an include directive for the header file. So Figure 7-6 would be
unchanged also. If we compile both .cpp files, we find that the instantiated member functions
are not found. We fix the problem by creating a separate file containing explicit instantiations of
the ObjectCell for all the types we use. An example of these explicit instantiations is shown
in Figure 7-11. This file is compiled as part of the project. We have had success using this technique with several older compilers. The downside is that all of the template expansions have to
be listed by the programmer, and sometimes if the class template uses other class templates,
those have to be listed too. The advantage is that if the implementation of the member functions
in ObjectCell changes, only ObjectCellExpand.cpp needs to be recompiled.
1
2
3
4

#include "ObjectCell.cpp"
template class ObjectCell<int>;
template class ObjectCell<double>;

Figure 7-11

Instantiation file ObjectCellExpand.cpp

c++book.mif Page 147 Tuesday, April 29, 2003 2:13 PM

Specialized Templates

7.4.3

147

The export Directive

Newer compilers support the export directive, which is a recent language addition. If we issue an
export directive prior to the class template declaration, as shown in Figure 7-12, then the extra
file that we added in Section 7.4.2 becomes unnecessary, and the program links as we would
have expected. Dont expect to see this correctly implemented on all compilers. The export
directive also works for function templates.

7.5 Specialized Templates

There are numerous other features of templates that can be used to produce fairly bizarre code.
Rather than list all of those, we concentrate on a few features that are commonly used.
7.5.1

Multiple Template Parameters

The template parameter list can have more than one parameter. For instance, a map, in which we
store keys and values can be declared as
template <typename KeyType, typename ValueType>
class Map
{
...
};

An instantiation of the Map to store names and birthdates, in which a name is a string and a
birthdate is of type Date would be
Map<string,Date> birthdays;

In fact, map (the class template name is all lower case) is part of the Standard Library and mimics the TreeMap found in java.util.
7.5.2

Template Nontype Parameters

Template parameter lists can include nontype parameters. For instance,

template <typename Object, int size>
class Buffer
{
...
};
1
#ifndef _OBJECTCELL_H
2
#define _OBJECTCELL_H
3
4
5
6
7

template <typename Object>

export class ObjectCell
{
...
Figure 7-12 The export directive

c++book.mif Page 148 Tuesday, April 29, 2003 2:13 PM

148

Chapter 7 Templates

A nontype parameter must be a compile-time constant. Legal instantiations of the

Buffer template would then include
Buffer<string,1024> buf1;
Buffer<string,2048> buf2;

The advantage of using a nontype parameter is that size can be used in places that
require constant expressions (such as primitive arrays), so a data member such as
Object buf[ size ];

// Primitive array instead of vector

would be legal in this class template, but would not be legal if size was simply a parameter to
a Buffer constructor, in which only Object was in the template parameter list. The disadvantage, of course, is the potential for code bloat.
Additionally, note that buf1 and buf2 above have different types. This may be an
advantage or a disadvantage.
7.5.3

Default Template Parameters

Template parameter lists can have default values, for both type and nontype parameters. For
instance,
template <typename Object=char, int size=4096>
class Buffer
{
...
};

A nontype parameter must be a compile-time constant. Legal instantiations of the

Buffer template would then include
Buffer<int>
Buffer<>

buf1;
buf2;

// Buffer<int,4096>
// Buffer<char,4096>

Default template parameters are relatively new and not implemented on many compilers.
7.5.4

Member Templates

Member templates are member functions that are themselves templates; this is basically a template inside of a template. As an example of its use, consider the code in Figure 7-13. Even
though a float can be used in place of a double, and an int can be used in place of a long,
if we have
Pair<int,float>
p1( 3, 4.0 );
Pair<long,double> p2 = p1;

the code does not compile because p1 and p2 are not type compatible, and so the default copy
constructor is not applicable. Member templates allow us to solve this problem by providing a
constructor that accepts any Pair, instead of the specific Pair.

c++book.mif Page 149 Tuesday, April 29, 2003 2:13 PM

Specialized Templates

1
2
3
4
5
6
7
8
9
10
11
12
13

149

template <typename Type1, typename Type2>

class Pair
{
public:
Pair( const Type1 & f, const Type2 & s )
: first( f ), second( s )
{ }
...
Type1 first;
Type2 second;
};

Figure 7-13

Pair class template that is very picky about types

This strategy is illustrated in Figure 7-14. Lines 9 to 12 contain the member template.
Observe that the constructor for Pair accepts a Pair with arbitrary types and as long as the
types are compatible the constructor template will expand. If the types are not compatible, we
get a compile-time error due to line 11, which is perfect.
Member templates are also a relatively new addition to C++, so many compilers do not
support them. The use shown here in extending type compatibility among different template
instantiations is probably its most common.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

template <typename Type1, typename Type2>

class Pair
{
public:
Pair( const Type1 & f, const Type2 & s )
: first( f ), second( s )
{ }
template <typename Type3, typename Type4>
Pair( const Pair<Type3,Type4> & rhs )
: first( rhs.first ), second( rhs.second )
{ }
...
Type1 first;
Type2 second;
};

Figure 7-14

Pair class template that is allows construction from any type-compatible

Pair using member template

c++book.mif Page 150 Tuesday, April 29, 2003 2:13 PM

150

Chapter 7 Templates

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

class Rectangle
{
public:
explicit Rectangle( int len = 0, int wid = 0 )
: length( len ), width( wid ) { }
int getLength( ) const
{ return length; }
int getWidth( ) const
{ return width; }
void print( ostream & out = cout ) const
{ out << "Rectangle " << getLength( ) << " by "
<< getWidth( ); }
private:
int length;
int width;
};
ostream & operator<< ( ostream & out, const Rectangle & rhs )
{
rhs.print( out );
return out;
}

Figure 7-15

Simple rectangle class used in function object examples

7.6 Templates for Function Objects

In Java, inheritance can used to implement function objects. In fact, we can use inheritance in
C++ in almost the same way. Our example begins with a basic rectangle class shown in
Figure 7-15.
If we want to find the maximum rectangle using the findMax function template in
Figure 7-1, then we need to have operator< defined for rectangles. This might not be easy to
do. For instance, it might already be defined to return the maximum area, but we are interested
in the maximum perimeter or maximum length. Or worse, we want to find the maximum based
on several different criteria.
7.6.1

The Java Solution

In Java, we typically solve this problem by passing a second parameter to findMax, which is a
function object. The function object is an instance of a class that implements some interface. In
this case, the interface is Comparator (in java.util), which specifies a method that can

c++book.mif Page 151 Tuesday, April 29, 2003 2:13 PM

Templates for Function Objects

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21

151

template <typename Object>

class Comparator
{
public:
virtual bool isLessThan( const Object & lhs,
const Object & rhs ) const = 0;
};
// Generic findMax, with a Java-style function object.
template <typename Object, typename Comparator>
const Object &
findMax( const vector<Object> & a, Comparator cmp )
{
int maxIndex = 0;
for( int i = 1; i < a.size( ); i++ )
if( cmp.isLessThan( a[ maxIndex ], a[ i ] ) )
maxIndex = i;
return a[ maxIndex ];
}

Figure 7-16

Java-style comparator and findMax routine

compare any two objects. Since we do not have a root class of Object in C++, we can simulate
the Java idea by declaring that Comparator is a class template, and have the findMax template accept a Comparator<Object> as a second parameter. But a better plan is to make the
Comparator a template parameter, so that a Comparator that compares base class objects
can be used on a vector of derived class objects. This resulting Comparator class template
and findMax function template is shown in Figure 7-16. Since we expect the Comparator to
be a base class, it probably should have a virtual destructor and be passed using call-by-constant
reference, instead of call-by-value. But as we will see, this turns out not to be necessary.
1
2
3
4
5
6
7
8
9
10
11
12

int main( )
{
vector<Rectangle> a;
a.push_back( Rectangle( 1, 10 ) );
a.push_back( Rectangle( 10, 1 ) );
a.push_back( Rectangle( 5, 5 ) );
cout << findMax( a, LessThanByLength( ) ) << endl;
return 0;
}

Figure 7-17

main that invokes findMax with a comparator

c++book.mif Page 152 Tuesday, April 29, 2003 2:13 PM

152

Chapter 7 Templates

1
2
3
4
5
6
7

class LessThanByLength : public Comparator<Rectangle>

{
public:
bool isLessThan( const Rectangle & lhs,
const Rectangle & rhs ) const
{ return lhs.getLength( ) < rhs.getLength( ); }
};

Figure 7-18
1
2
3
4
5
6
7

Class that implements comparator interface for rectangles (preliminary)

class LessThanByLength
{
public:
bool isLessThan( const Rectangle & lhs,
const Rectangle & rhs ) const
{ return lhs.getLength( ) < rhs.getLength( ); }
};

Figure 7-19

Class that implements comparator for rectangles, without comparator

Now main can invoke findMax if it passes a vector of Rectangles and an appropriate function object. main is shown in Figure 7-17. The comparator is an anonymous instance
of class LessThanByLength, which implements the Comparator interface and is shown
in Figure 7-18. Specifically, it extends an instantiated Comparator template by implementing
isLessThan.
Except for the template baggage that replaces the use of Object as a superclass, this
implementation is exactly equivalent to the Java idiom for function objects.
7.6.2

Avoiding Inheritance

If we examine the code carefully, we see that the Comparator class template is not used at all!
Thus we can simplify Figure 7-18, and remove the Comparator class template completely.
The revised version is shown in Figure 7-19. Otherwise findMax and main are unchanged.
7.6.3

Function Objects With operator()

We can use operator overloading to make the implementation of the function object look slicker.
To do this, we replace isLessThan with operator(), as shown in Figure 7-20.

c++book.mif Page 153 Tuesday, April 29, 2003 2:13 PM

Key Points

1
2
3
4
5
6
7

153

class LessThanByLength
{
public:
bool operator( ) ( const Rectangle & lhs,
const Rectangle & rhs ) const
{ return lhs.getLength( ) < rhs.getLength( ); }
};

Figure 7-20

Overloading of the function call operator for function object

In the implementation of findMax, the call at line 17 becomes

if( cmp.operator() ( a[ maxIndex ], a[ i ] ) )

Since operator() is the function call operator, this can be rewritten as

if( cmp( a[ maxIndex ], a[ i ] ) )

At this point it makes sense to change the name of the parameter to findMax from cmp
to isLessThan. Thus we get the code in Figure 7-21.
To summarize, the function object is declared in Figure 7-20 and provides an implementation of the function call operator. The routine that uses the function object is a template, and type
of the function object is a template parameter. Syntactically, when the function object is used, it
looks like a normal global function call. The template routine can itself be invoked as in Java by
passing an instance of the function object. The compiler will do all the type resolution.
Function objects are used extensively in the STL, which is the C++ equivalent of the Collections API. The STL is discussed in Chapter 10.

7.7 Key Points

Templates in C++ replace the use of inheritance in Java to implement type-independent
algorithms. Templates allow many mismatched types to be detected at compile time.
Function templates are not functions; they are function wannabees.
When a function template is expanded, the result is a function.
1
2
3
4
5
6
7
8
9
10
11
12

template <typename Object, typename Comparator>

const Object & findMax( const vector<Object> & a,
Comparator isLessThan )
{
int maxIndex = 0;
for( int i = 1; i < a.size( ); i++ )
if( isLessThan( a[ maxIndex ], a[ i ] ) )
maxIndex = i;
return a[ maxIndex ];
}

Figure 7-21

Implementation of findMax with operator() (final version)

c++book.mif Page 154 Tuesday, April 29, 2003 2:13 PM

154

Chapter 7 Templates

Class templates are wannabee classes.

Separate compilation of templates is complicated by the fact that the templates are not real
functions or member functions. Various mechanisms can be used, depending on the compiler.
Parameters to a function that are template parameters should rarely be passed using callby-value. An exception is function objects.
Templates can be instantiated with multiple parameters.
Templates can be instantiated with non-type parameters, such as an int.
Template parameters can have defaults.
Member templates are member function templates inside of templates.
Function objects are implemented in C++ by overloading the function call operator.

7.8 Exercises
1. When using templates, what kinds of errors are detected at compile time that are not
detected until runtime in Java?
2. When are errors in a template definition detected by the compiler?
3. What is code bloat?
4. What strategies are used for separate compilation of templates?
5. What is a member template?
6. How are templates used for function objects?
7. Implement a function template that takes a single vector as a parameter, and sorts the vector (using any simple sorting algorithm).
8. Implement a function template that takes a single vector as a parameter and a function
object that represents the comparator as a second parameter and sorts the vector using the
comparator as the basis for ordering (using any simple sorting algorithm).
9. Reimplement Exercise 4.19 using templates to generalize the types of the objects in the
stack.
10. Reimplement Exercise 4.20 using templates to generalize the types of the objects in the
set.
11. Reimplement Exercise 5.11 using templates to generalize the types of the keys and values
in the map.

c++book.mif Page 155 Tuesday, April 29, 2003 2:13 PM

H A P T E R

Abnormal Control Flow

AVA is notoriously concerned with not allowing unsafe

programs to execute. The compiler will catch many programming errors, and the runtime system
will throw exceptions when invalid operations occur. C++ was designed with a different mentality.
In this chapter, we discuss how C++ programs typically handle errors, and then examine
exceptions in C++. Although the syntax has similarities (try and catch), C++ exception handling is a shell of what Java provides.

8.1 The Design of C++

From day 1, Java was defined with code safety in mind. C++ was designed with code efficiency
in mind, which was certainly justifiable given the speed of computers in 1980s. Although error
handling in C++ is important, the language specification does not go out of its way to detect programming errors at either compile time or runtime; often a program will terminate due to an
error only if the underlying hardware complains.
As an example, we know that in Java, all variables have a definite value prior to use. So
the Java equivalent of the code in Figure 8-1 would be illegal and would not compile. Older C++
compilers say nothing about this code, while newer compilers issue warnings about the two
uninitialized variables. However most C++ compilers will still compile the code and let the user
run it.
A second example is shown in Figure 8-2. Here we see the old-style cast used to convert
between two completely unrelated pointers types. This code compiles, probably with a warning,
and runs, producing some bizarre output. Of course the dynamic_cast can be used to solve
this problem, but Java does not allow type confusion of the sort shown here at all. Similar non-

155

c++book.mif Page 156 Tuesday, April 29, 2003 2:13 PM

156

Chapter 8 Abnormal Control Flow

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

class Account
{
public:
Account( int b = 0 )
: balance( b ) { }
int getBalance( ) const
{ return balance; }
void deposit( int d )
{ balance += d; }
private:
int balance;
};
int main( )
{
Account *acc1 = new Account( );
Account *acc2;
acc1->deposit( 50 );
cout << acc1->getBalance( ) << endl;
cout << acc2->getBalance( ) << endl;
return 0;
}

Figure 8-1

C++ code that compiles (probably with a warning)

sense can be seen in Figure 8-3: Here we allow the changing of a constant object. Though C++
puts in a good faith effort to disallow it, if we use the const_cast, we can change an object
that we promised not to change.
But certainly the most evil of C++ issues is shown in Figure 8-4. Here we have an obvious
array index that is out-of-bounds. Almost any language would detect this at runtime. But not
C++. Instead, it uses the four bytes that follow arr[9] as its guess for arr[10]. If arr[10]
were on the left-hand side of the assignment operator, we could even change its value. Almost
certainly this would be some other variable in the program, leading to hard-to-find bugs.
In C, on which primitive C++ arrays are based, the lack of bounds checking was justified
on the grounds that the implementation simply stored a pointer variable. And being a language
of the 1970s, that was used instead of assembly language to implement operating systems, certainly speed was an important criteria.
But the vector class was added in the mid 1990s, and not including bounds checking as
an automatic part of the indexing operator seems inexcusable. Yet it accurately reflects the C++
philosophy of not forcing the user to pay at runtime for anything more than bare necessities, and
certainly not paying for a feature that was free in C.

c++book.mif Page 157 Tuesday, April 29, 2003 2:13 PM

The Design of C++

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

157

class Barbell
{
public:
Barbell( double b ) : weight( b ) { }
double getWeight( ) const
{ return weight; }
private:
double weight;
};
int main( )
{
Barbell *bb = new Barbell( 15.6 );
cout << bb->getWeight( ) << endl;
Account *acc = (Account *) bb;
cout << acc->getBalance( ) << endl;
acc->deposit( 40 );
cout << bb->getWeight( ) << endl;
return 0;
}

Figure 8-2

C++ code that might generate a warning but compiles, allowing type confusion

In C++ error handling in general seems to follow this trend. The compiler and runtime
systems are less likely to signal errors than in Java. And even when such errors are signalled, in
many cases the programmer can avoid acknowledging that the error has occurred, and can thus
continue executing code while the program is no longer in a good state.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

int main( )
{
const string h = "hello";
// two failed attempts to change h
h[ 0 ] = 'j';
// does not compile, thankfully
string & href = h; // does not compile, thankfully
// third time is a charm
string & ref = const_cast<string &> ( h );
ref[ 0 ] = 'j';
cout << h << endl;
// prints jello
return 0;
}

Figure 8-3

C++ code that allows changes to constants

c++book.mif Page 158 Tuesday, April 29, 2003 2:13 PM

158

Chapter 8 Abnormal Control Flow

1
2
3
4
5
6
7
8
9
10

int main( )
{
int i;
vector<int> arr( 10, 37 );

// 10 items, all with val 37

for( i = 0; i <= 10; i++)

cout << i << " " << arr[ i ] << endl;
return 0;
}

Figure 8-4

Array bounds are not checked in C++

This contrasts with Java, where if the Virtual Machine signals an exception, the Virtual
Machine will shut down (the erroneous thread) if the exception is not handled, and where the
programmer is explicitly required to write some code that acknowledges that a checked exception might occur (even if that code simply declares a throws list that allows the exception to
propagate).

8.2 Non-Exception Error Handling

In this section, we discuss two widely-used error handling techniques that predate exception
handling.
8.2.1

abort and exit

If the programmer has detected an unrecoverable error, there are plausible solutions. The best is
to throw an exception and not handle it (or possibly handle the exception in main, and let main
return with a nonzero return value). In a pre-exception world, there are two common ways to
abnormally terminate a program from inside any function.
The functions abort and exit, both declared in the standard header file cstdlib,
cause the program to terminate. exit can provide a parameter that will be used in place of a
return value for main. The signatures of abort and exit are:
void abort( );
void exit( int status );

abort causes the program to terminate abruptly. No destructors are invoked in the process.
exit causes the destructors of static objects to be called, but not any local objects that are in
pending function calls. Do not call exit inside the destructor of a static object. (why?)
Because few destructors are invoked, both abort and exit are inferior to throwing an
uncaught exception. Further, exit should never be used outside of main to handle a normal
termination, since a normal termination should guarantee that all appropriate destructors are
invoked.
As part of the termination process, exit invokes functions that are registered with
atexit in reverse order of their registrations (the functions are also invoked if termination is

c++book.mif Page 159 Tuesday, April 29, 2003 2:13 PM

Non-Exception Error Handling

1
2
3
4
5
6
7
8
9
10

159

void printExit( )
{
cout << "Invoking the printExit method..." << endl;
}
int main( )
{
atexit( printExit );
foo( );
}

Figure 8-5

Example of calling atexit

normal). Such functions accept no parameters and return void, and the number of registered
functions is implementation dependent. The signature for atexit is:
int atexit( void f( void ) );

with atexit returning 0 if the limit of registered functions has already been reached. Figure 85 illustrates the use of atexit. If foo returns normally, without calling abort (or throwing an
unhandled exception), printExit will be invoked as the program terminates. printExit will
also be invoked if foo calls exit.
8.2.2

Assertions

The assert preprocessor macro (see Section 12.1 for a discussion of preprocessor macros) is
made available by including the standard header file cassert.
If NDEBUG is defined, calls to assert are ignored. Otherwise, if the parameter to
assert is zero, an error message is printed and the program is aborted, by calling abort.
Recall that no destructors are invoked, so this is rather drastic. The error message includes the
actual expression as well as the source code file name and line number.
For instance, in Figure 8-6 if a call to foo is made with a NULL pointer, an error message
similar to
1
2
3
4
5
6
7
8
9
10

Assertion failed: s != NULL, file Fig8-6.cpp, line 8

#include <iostream>
#include <string>
#include <cassert>
using namespace std;
void foo( string *s )
{
assert( s != NULL );
...
}

Figure 8-6

Example of calling assert

c++book.mif Page 160 Tuesday, April 29, 2003 2:13 PM

160

Chapter 8 Abnormal Control Flow

will appear (on the standard error), and the program will terminate abnormally. If NDEBUG is
defined prior to line 8 via
#define NDEBUG

assert will be ignored.

8.2.3

Error states

Many older C++ libraries, including the I/O library use the notion of an error state. If an error
occurs, either it is returned in an error code, a global error code is set, or the object records that it
is in an error state (in an internal data member). In all cases, the error state can be tested, and
possibly cleared if the user invokes appropriate functions. We will discuss how this works with
I/O in Chapter 9. In general, this solution is inferior to exceptions.

8.3 Exception Handling

Exception handling in C++ is based on a similar model popularized in a language called Ada. In
the 1980s and 1990s, Ada was the language of choice for U.S. Government applications
because, unlike C and C++, but like Java that followed, it attempted to catch many errors at
compile time, and do extensive runtime checking in attempt to avoid programming errors in
mission-critical systems. The Java exception handling is a refinement of C++s model.
8.3.1

Basic syntax

The basic C++ syntax is similar to Java, in that a member function can throw an exception, and
a caller can either ignore the exception and let it propagate to its caller, or attempt to handle the
exception with a try/catch block.
8.3.2

No finally clause

C++ does not have a finally clause. However, it is guaranteed that if a function returns either
normally with no active exception, or abnormally with an active exception, all its destructors
will be invoked. Thus local objects that allocated heap memory indirectly will free the heap
memory back to the system automatically, assuming their destructors are correctly written. This
leaves only heap memory that was directly allocated and is accessed by local pointer variables.
In such a case the auto_ptr, which we discuss in Section 8.4 can be used. The basic idea of
the auto_ptr is that it wraps a pointer variable and has a destructor that calls delete, so if
we create an auto_ptr, the heap memory can be freed by its destructor.
8.3.3

The Throw List and Unexpected Exceptions

In C++, we can specify a throw list, cosmetically like Java. Notice that the reserved word is
throw, rather than throws, to avoid an additional reserved word in the language. We say that
this is cosmetic because unlike Java, the throw list in C++ is not examined at compile time. So if
the method throws an exception that is not in the throw list, the program will still compile. How-

c++book.mif Page 161 Tuesday, April 29, 2003 2:13 PM

Exception Handling

161

ever, at run time, if an exception is thrown that was not in the throw list, then a call to
std::unexpected will occur, which normally calls abort. It is possible to change the
behavior of std::unexpected by invoking set_unexpected.
The throw list syntax in C++ is slightly different from Java:
void foo( ) throw( UnderflowException, OverflowException );

An empty (as opposed to missing) throw list signals that no exceptions are expected to be
thrown. If an expected exception is thrown but not caught, the function std:terminate will
be invoked. This function is also called if the exception handling mechanism determines that the
runtime stack has been corrupted, or if a destructor that is executed as part of the handling of an
exception itself throws an exception (to avoid infinite recursion). Calling set_terminate
allows the programmer to supply a function that contains different behavior for the
terminate function.
8.3.4

Missing Throw List

A missing (as opposed to empty) throw list signals that any exception might be thrown. Thus,
nothing is unexpected. Missing throw lists are part of C++ in order to preserve backward compatibility when exceptions were added. Of course, this is why compile-time checking of throw
lists is not feasible. Most C++ code omits the throw list.
8.3.5

What Can Be Thrown

In Java, only objects that are subclasses of Throwable (or Throwable itself) can be thrown
as exceptions. In C++, any object can be thrown as an exception, but clearly, it is better to create
exception

runtime_error

logic_error
bad_alloc

range_error

out_of_range

length_error

bad_cast
overflow_error
underflow_error

invalid_argument domain_error
bad_typeid
bad_exception
ios_base::failure

Figure 8-7

Standard exceptons

c++book.mif Page 162 Tuesday, April 29, 2003 2:13 PM

162

Chapter 8 Abnormal Control Flow

a class type that can store information about why the exception was thrown.
In Java, Throwable objects contain a stack trace as part of their data. Since there is no
root exception class in C++, you cannot count on much information from the exception object
besides its type.
8.3.6

Standard Exception Classes

Although there is no root exception class, C++ attempts to provide an inheritance hierarchy for
its exceptions. These exceptions are shown in Figure 8-7. In Java, most of these exceptions
would be considered runtime exceptions and would not be caught by the programmer. The only
method in class exception, besides the Big-Three and a zero-parameter constructor is the
what method, which returns a primitive string giving you some information. exception and
its subclasses promise not to throw any exception (specifically bad_alloc) when their methods are invoked.
8.3.7

Rethrowing Exceptions

An exception can be rethrown simply by providing a throw statement in a catch block:

catch( domain_error e )
{
// do some cleanup (not shown)
throw;
// rethrow e
}

8.3.8

Catching All Exceptions

Because there is no root hierarchy, special syntax is needed to catch all exceptions. If ... is
used in a catch handler, it matches any exception. However, the exception object will not be
available (since there is no way to provide its type). As in Java, if there are multiple catch
clauses, the first match wins. Thus ... has to be the last catch block in a series.
8.3.9

Inheritance and Exceptions

When inheritance is used, an exception object of a type DerivedException can be caught

by a handler for BaseException. To achieve polymorphism, call by reference (or constant
reference) should be used, since call-by-value slices the exception object.
Figure 8-8 illustrates truly nonsensical code, including a throw list for each method. Since
the call at line 22 could cause either of two kinds of exceptions, but both are subtypes of
logic_error, we can simply catch logic_error, in the same way we would in Java. Most
of the rules that concern inheritance are similar in C++ and Java.
The throw list in an overriding method cannot add exceptions. The catch clauses are tried
in sequence so a derived exception class must be listed before a base exception class, and all
must be listed before ... (the compiler will check this). If the exceptions form an inheritance

c++book.mif Page 163 Tuesday, April 29, 2003 2:13 PM

Exception Handling

163

hierarchies and the methods are marked virtual, dynamic dispatch will occur when the
exceptions methods are invoked, as long as the exception is not passed by value.
The alert reader, may be wondering why the reference to the exception, e, at line 26 is not
stale. After all, in both cases, a local temporary was storing the exception object, and the function in which that temporary was created has terminated. Clearly if this was true, exceptions
would have to be passed using call-by-value, which would be impossible with inheritance (since
call-by-value and inheritance do not mix). Thus as a special case, the exception handling mechanism guarantees that the exception object will not have its destructor called until it is no longer
an active exception. This means the object is valid until it is caught, and the catch block that it is
caught in terminates, unless the catch block rethrows the exception, in which case it retains its
validity as if the catch block was never executed.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29

double mysqrt( double x ) throw( domain_error )

{
if( x < 0 )
throw domain_error( "sqrt of neg number" );
else
return sqrt( x );
}
double myexp( double x ) throw( out_of_range )
{
double result = exp( x );
// call routine in cmath
if( result == HUGE_VAL )
throw out_of_range( "exp too large" );
else
return result;
}
double f( double x )
{
try
{
return mysqrt( x ) + myexp( x );
}
catch( logic_error & e )
{
cout << e.what( ) << endl; // invoke correct what
return -1.0;
// cant be answer for any other x
}
}

Figure 8-8

Nonsense example to illustrate inheritance

c++book.mif Page 164 Tuesday, April 29, 2003 2:13 PM

164

Chapter 8 Abnormal Control Flow

8.3.10 Templates and Exceptions

Templates and exceptions dont mix. Specifically, do not put a throw list on any function template. The reason is that you have to worry not only about the logic in your function, but whether
the various operations on the instantiated types might themselves throw an exception. This
becomes an impossible task and is typical of what happens when two separate features are added
to a language at a late stage. The features often are not 100% compatible.

8.4 The auto_ptr Class

When a pointer variable is allocated locally, it can hard to reclaim its memory. For instance
void foo( )
{
Person *p = new Student( 123456789, "Jane", 4.0 );
p->print( );
...
delete p;
}

Even if we have done all the correct things, such as declaring a virtual destructor for
Person, this code could leak memory if foo abruptly terminates with an unhandled exception.
Because C++ does not have a finally block, it is difficult to guarantee that the delete statement is executed. However, it is guaranteed that all local variables will have their destructors
called.
Thus, C++ provides a wrapper class template, the auto_ptr, declared in the standard
header file memory. The idea of auto_ptr is that it wraps a pointer variable, provides sufficient operator overloading so that the wrapped variable can be accessed without any additional
baggage, and since its instances are objects, we can expect a destructor to be invoked. The
destructor can call delete on the pointer variable that it is wrapping. The auto_ptr also
provides a get routine, in case access to the wrapped pointer is needed and relying on an
implicit type conversion is not appropriate. In a simple example;
void foo( )
{
auto_ptr<Person> p( new Student( 123456789, "Jane", 4.0 ) );
p->print( );
...
}

Now, even if print, or some other routine throws an exception that is not handled, the
Students destructor is invoked. The call to print works because auto_ptr overloads
operator->. Also overloaded are operator*, the type conversion operator (a cast to
Person *, in our example), copying, and copy construction.
If an auto_ptr is constructed to point at an object that was not heap-allocated, disaster
can ensue, so dont do that. If two auto_ptrs are constructed to point at the same heap-allocated object, disaster can ensue, because of double-deletion. So dont do that either.

c++book.mif Page 165 Tuesday, April 29, 2003 2:13 PM

Key Points

165

If an auto_ptr is copied into another auto_ptr, either by copy construction or copy

assignment, then we have double-deletion potential. To avoid double-deletion, the auto_ptrs
destructor only calls delete if it is the owner of the pointer. When the auto_ptr is initialized with a pointer it becomes the owner. If it is copied to another auto_ptr, ownership transfers. The first auto_ptr will not longer attempt to call delete. This owner transfership
allows two other uses of the auto_ptr.
In the first additional use, a function allocates an object from the heap memory, and
returns a pointer to the object back to the caller. Now the caller is responsible for reclaiming the
memory. With an auto_ptr, the function wraps the pointer in an auto_ptr and returns it
using return by value. The caller saves a copy of the return value in a local auto_ptr, transferring ownership, and thus the memory will be automatically reclaimed.
In the second additional use, we pass an auto_ptr to a function that we invoke, and let
that function clean up the memory, because we have no use for it after the return. Though this
works, it also implies that we need to be very careful when we pass auto_ptrs to other routines, since on the return, the allocated memory is likely to have been reclaimed.
An important consequence is that placing auto_ptrs into standard containers, such as
vectors, is a terrible idea (and is in fact illegal in Standard C++). Once it is in a vector, it
will tend to get copied over and over again, possibly into a temporary variable. If the temporary
variable winds up with ownership, then when it is destroyed, it will reclaim the dynamically
memory.

8.5 Key Points

For robust code, avoid calling exit outside of main. Try to use exceptions instead.
assert can be used for debugging code, but might be inappropriate for production code.
However, a simple replacement for assert that substitutes a call to abort with throwing
of an exception can easily be written (Section 12.1).
Exceptions in Java are based on exceptions in C++, but the C++ implementation doesnt
work well mainly because of the need for backward compatibility with earlier versions of
C++.
Few C++ library routines signal errors by throwing an exception. Instead, many incorporate error states into either return codes, global variables, or object state.
C++ does not have a finally block. Instead, attempt to ensure that all of your heap-allocated objects are wrapped inside a stack-allocated object, so that they can be freed by calls
to destructors. One way of doing this is to use the auto_ptr.
A throw list specifies exceptions that can be expected to occur. If an exception is thrown
from a method that is not in the methods throw list, function std::unexpected is
invoked.
A missing throw list indicates that any exception can occur.
Any object can be thrown in C++.

c++book.mif Page 166 Tuesday, April 29, 2003 2:13 PM

166

Chapter 8 Abnormal Control Flow

The standard library defines a small hierarchy of what would be Java errors and runtime
exceptions that can be extended by the programmer. This hierarchy is rooted at class
exception.
Inside a catch block, we can rethrow an exception simply by issuing throw. The exception is not required in the throw statement.
To catch all exceptions use ... .
Many of the Java rules relating to exceptions are in effect implemented in C++.
Exceptions should always be caught using call-by-reference or call-by-constant reference.
Catching by reference allows the changing of the state of the exception object prior to
rethrowing it.
The exception object is always valid up to the end of the catch clause that last handles the
exception.
Templates and exceptions dont mix. Avoid using throw lists in function templates.
auto_ptr is a class template that wraps a pointer variable. The pointer variable should
be viewing a heap-allocated object. When the auto_ptr is destroyed, its destructor will
delete the object it is wrapping if the auto_ptr still enjoys ownership. Ownership is
transferred if the auto_ptr is copied into another auto_ptr.
Avoid using auto_ptr objects in container classes such as vector.

8.6 Exercises
1. Describe the semantics of abort and exit.
2. What does assert do?
3. List some standard C++ exceptions.
4. What does a missing throw list mean?
5. What objects can be thrown as exceptions?
6. How is an exception rethrown?
7. How can you catch all exceptions?
8. Why are exceptions never caught using call-by-value?
9. What is an auto_ptr?
10. How long is the exception object valid?
11. Why is it dangerous to include a throw specifier in a method template?
12. What happens if an exception that is not listed in the throw list is thrown?
13. Implement the auto_ptr class template. For an implementation that allows compatability between derived and base class auto_ptrs, you should use member templates.

c++book.mif Page 167 Tuesday, April 29, 2003 2:13 PM

H A P T E R

Input and Output

N Java, I/O is supported by a large library that resides

mostly in package java.io. The package makes use of inheritance, by defining four abstract
classes (InputStream, OutputStream, Reader, and Writer), which can then be
extended to target different sources of data (files, sockets, arrays), and which can be wrapped in
a classic decorator pattern to achieve different functionality (buffering, compression, encryption,
serialization, pushback operators, and so on). In C++, inheritance is also used to define a hierarchy, though the C++ hierarchy is considerably smaller and the use of the decorator pattern is not
adopted.
In this chapter, we begin by discussing the I/O hierarchy in C++. Because exceptions are
not part of I/O, we see how the error state is encoded in a stream. Then we look at basic output
and input, and see an equivalent to the use of StringTokenizer that is seen in Java.

9.1 The Basic Hierarchy

The basic inheritance hierarchy is shown in Figure 9-1.
At the root of the hierarchy is class ios_base. This class contains operations and information that is independent of whether the I/O operations are intended to work in ASCII (char),
Unicode (wchar_t), or some other unspecified character type. It contains data members that
represent the state of the stream, and constants that can control whether a file is opened in binary
or non-binary mode, and whether reading/writing begins at the start, end, or some specific point
inside a file. Such information does not depend on whether operations are input or output, nor do
they depend on the type of characters being used, so it makes sense to factor them out into a
common superclass.

167

c++book.mif Page 168 Tuesday, April 29, 2003 2:13 PM

168

Chapter 9 Input and Output

ios_base

ios

istream

ifstream

ostream

istringstream

iostream

fstream
Figure 9-1

ostringstream

ofstream

stringstream

Basic inheritance hierarchy of streams. Any class c in shaded area is really a

template instantiation basic_c<char>, and there is a similar instantiation for
wchar_t (e.g. wifstream is used for wide characters)

Subclasses of ios_base are all templates that are instantiated with the type of characters
that are used in the implementation. Standard instantiations include one for char, and another
for the wide character type wchar_t.
The basic_ios class encapsulates some information that is common to both input and
output. The instantiation basic_ios<char> is defined with a typedef as ios; and is similar
instantiation basic_ios<wchar_t> is defined with a typedef as wios. In other words, we
have:
template <class chartype>
class basic_ios : public ios_base { ... }
typedef basic_ios<char>
ios;
typedef basic_ios<wchar_t> wios;

Subclasses of basic_ios are basic_istream and basic_ostream; a typical

simplified implementation of istream is:
template <class chartype>
class basic_istream : virtual public basic_ios<chartype> { ... }
typedef basic_istream<char>
istream;
typedef basic_istream<wchar_t> wistream;

Note the use of virtual inheritance since basic_iostream uses multiple inheritance. In
the remainder of this chapter, we will ignore the fact that these are class templates, and instead
refer directly to the instantiations with char.

c++book.mif Page 169 Tuesday, April 29, 2003 2:13 PM

Error States

169

Subclasses of istream include ifstream, for input from files, and

istringstream, for input from a string. istringstream is like StringReader, in that
it implements all of the Reader operations, taking its data from a String, istringstream
also is like StringTokenizer, in that StringReader has little more than the ability to
read one character, while StringTokenizer can separate its string into tokens.
C++ does not define a socketistream class, but there are third-party implementations
such as the GNU library that do, and such classes easily fit into the inheritance hierarchy.
The iostream class extends both istream and ostream, and allows reading and
writing to a stream simultaneously. Often doing so is fairly tricky, and so we recommend following the Java style of having each stream do reading or writing exclusively. But the declaration is
interesting, at least so much as it provides a practical example of multiple inheritance:
template <class chartype>
class basic_iostream : public basic_istream<chartype>,
public basic_ostream<chartype>
{ ... }

Eight streams are predefined: cin, cout, cerr, and clog which are in order: standard
input, output, error, and an error stream that is automatically flushed. Also predefined are the
wide character counterparts wcin, wcout, wcerr, and wclog.

9.2 Error States

A stream can interrogate its state by invoking one of eof, fail, bad, or good. If an I/O operation fails because of eof, the stream state will be set to indicate eof. For instance
cin >> x;
if( cin.eof( ) )
cout << "Oops ... eof was reached" << endl;

Note carefully, that

while( !cin.eof( ) )
{
cin >> x;
cout << "Read " << x << endl;
}

does not work, because the end of file stream state is not set until an I/O operation actually fails.
Thus the loop above goes around one time too many, printing the last successfully read value
twice (since x wont change on the unsuccessful read). This is an extremely common error that
could never happen in Java, because an exception would cause the flow to leave prior to the
cout. But in C++, if you do not test the stream state, flow continues as if nothing happened.
If the stream is in an error state, then the expression cin, when used as a condition of an if
or loop, returns false. Thus, commonly seen code is:
while( cin >> x )
cout << "Read " << x << endl;

c++book.mif Page 170 Tuesday, April 29, 2003 2:13 PM

170

Chapter 9 Input and Output

which correctly outputs all successfully read data. However, if the read fails (for instance x is an
int, and there is a non-integer character on the input), this loop terminates at that point, rather
than attempting error recovery.
For more robust code, one can use a test that attempts recovery. An example of this strategy is shown in Figure 9-2. Here we have a routine that reads items, separated by white space, of
arbitrary type (as long as the type has correctly overloaded operator<<).
This code illustrates three important points. First, fail will be true if the I/O operation
has failed for some reason, but that the stream is still in good shape, and so the error is correctable. If, for instance, we are trying to read an integer, fail could indicate that non-integer data,
such as the word "Joe" was found on the input stream. Thus, second, we can solve the problem
by skipping over the next token. However, third, whenever the stream is in an error state, all I/O
on the stream will continue to fail. So before we attempt to read the string that we will discard,
we must first clear the error state by invoking clear. (That read cannot normally fail, since
there must be data, and the stream is in good shape otherwise, but extra critical code would
check an error state after the read also).
A technical point about this example that is not related to I/O, but instead deals with templates. The code can be invoked as:
vector<int> arr;
readData( cin, arr );

If the call readData returned the vector instead of using it as a parameter, then the instantiation of the function template must be explicit, as in
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

vector<int> arr = readData<int>( cin );

template <typename Object>
void readData( istream & in, vector<Object> & items )
{
items.resize( 0 );
Object x;
string junk;
// to skip over bad data
while( !( in >> x ).eof( ) )
{
if( in.fail( ) )
{
in.clear( );
// clear the error state
in >> junk;
// skip over junk
cerr << "Skipping " << junk << endl;
}
else
items.push_back( x );
}
}

Figure 9-2

Function template to read data, with error checking

c++book.mif Page 171 Tuesday, April 29, 2003 2:13 PM

Output

171

Figure 9-3

Methods to test stream error state

Method

Result is true ...

eof

if the stream has ALREADY encounted end-of-file

bad

for an invalid operation; stream is probably corrupt

fail

for an unsuccessful operation, such as reading data in the wrong format

good

if none of the other conditions are true

since a function templates return type is not considered in determining a template expansion,
and thus the compiler would have no way to deduce what Object should be.
The method good returns true if the stream is not in an error state. The method bad is
like fail, except it is more severe: the stream has been corrupted for some reason, and so it is
not worth attempting recovery. Figure 9-3 summarizes the methods that can test the state of a
stream.

9.3 Output
As we have already seen, the vast majority of output statements simply overload operator<<.
In addition to operator<<, single characters can be output by invoking the put member
function. The most interesting part of output in C++ is probably the technique that is used to
finely tune how the output is formatted, especially since we might not be happy with the
defaults.
For instance, if we run the code in Figure 9-4, the output that is produced is:
Pat 40000.1
Sandy 125443

Here we see two deficiencies. First, by default doubles are only output to six significant
digits. Second, integer and string types print only the minimum number of characters needed,
making it hard to align output. We would prefer output such as:
Pat
Sandy

40000.11
125443.10

in which we force two decimal places, and require that both the string and double be padded with at least a few spaces. But strings should be placed on the left, followed with padding, while double should be placed on the right, preceded with padding.
The number of significant digits and digits after the decimal point, as well as how much
and where padding of output is done is part of the state of a stream. Specifically, it is part of the
format state. Thus we can invoke methods on the stream to examine and possibly change the format state, on a stream-by-stream basis.
The easiest way to do this is to use manipulators. For instance, some of the manipulators,
of concern to us, with examples of their use on a specific output stream cout, could include

c++book.mif Page 172 Tuesday, April 29, 2003 2:13 PM

172

Chapter 9 Input and Output

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30

class Person
{
public:
Person( const string & n = "", double s = 0.0 )
: name( n ), salary( s ) { }
void print( ostream & out = cout ) const
{ out << name << " " << salary; }
private:
string name;
double salary;
};
ostream & operator<< ( ostream & out, const Person & p )
{
p.print( out );
return out;
}
int main( )
{
vector<Person> arr;
arr.push_back( Person( "Pat", 40000.11 ) );
arr.push_back( Person( "Sandy", 125443.10 ) );
for( int i = 0; i < arr.size( ); i++ )
cout << arr[ i ] << endl;
return 0;
}

Figure 9-4
cout
cout
cout
cout
cout
cout
1
2
3
4
5
6

Stream output defaults are not pretty

<< setw( 15 );
// set next field width to 15
<< left;
// field on left, padding on right
<< right;
// field on right, padding on left
<< setprecision( 2 ); // two decimal places
<< scientific:
// output double w/ scientific notation
<< fixed;
// output double no scientific notation
void print( ostream & out = cout ) const
{
out << left << setw( 15 ) << name << " "
<< right << fixed
<< setprecision( 2 ) << setw( 12 ) << salary;
}

Figure 9-5

Revised print that uses manipulators

c++book.mif Page 173 Tuesday, April 29, 2003 2:13 PM

Output

173

Most of these change the format state for all subsequent operations (until overridden by a contradictory manipulator), except for setw which only applies to the next field. The manipulators
that accept a parameter are available by including the standard header file iomanip. Figure 9-5
shows how we can use these manipulators to generate aligned, nicely-formatted output.
There are a host of manipulators that are available.
Manipulator boolalpha and noboolalpha are used to control whether bools are
printed as false and true or 0 and 1. The latter is the default (for backward compatibility).
Thus
cout << boolalpha << true << " " << noboolalpha << true << endl;

prints
true 1

oct, dec, and hex are manipulators used to control how numbers are output. Alternatively, setbase(b) can be used. The base by default is not printed, but this can be changed by
manipulator showbase; the default is noshowbase. Thus,
cout << 37 << " " << oct << 37 << " "
<< hex << 37 << " " << setbase( 10 ) << 37 << endl;

prints
37 45 25 37

If we have
cout << showbase;
cout << 37 << " " << oct << 37 << " "
<< hex << 37 << " " << setbase( 10 ) << 37 << endl;

then the output is

37 045 0x25 37

uppercase and nouppercase controls whether the x in 0x and e in scientific notation are printed in lower case or upper case. left and right control the positioning of the data
relative to the padding. internal puts fill characters between the sign and the value.
setprecision(n) sets the floating point precision. setw(w) sets the width of the next output only to w. fixed and scientific control whether scientific notation is output.
setfill(ch) sets the fill white space to ch. For instance, setfill(*) can be
used to fill with *, as is commonly done on cashiers checks to prevent fraud. The following
code
cout << setprecision( 2 ) << setfill( * ) << fixed << right;
cout << setw( 8 ) << 12.49 << endl;
cout << setw( 8 ) << 3.1 << endl;

outputs
***12.49
****3.10

c++book.mif Page 174 Tuesday, April 29, 2003 2:13 PM

174

Chapter 9 Input and Output

showpoint and noshowpoint control whether or not a decimal point is printed if

there are no digits following the decimal point. showpos and noshowpos control whether or
not a leading + is printed in nonnegative numbers.
flush flushes an output stream, causing any pending writes to be completed.
endl inserts a newline and flushes an ostream. ends inserts a null (this is used for
null-terminating primitive strings) and flushes an ostream.
skipws and noskipws controls whether leading white space is skipped on input operators. ws on an input stream skips leading white space for the next read only.

9.4 Input
We have already seen that input streams make use of overloaded sets of operator<< to do
significant work and in Section 9.2 we saw how the error state of a stream can be accessed and
cleared. In this section we discuss some additional input operations.
Often we want to perform character at a time input. Although operator<< is overloaded to accept a character, using it can be tedious because operator<< skips whitespace by
default. Although the manipulator skipws can change this setting for future reads, and ws can
change the state for the next read, at best this is tedious, and at worse it potentially is time-consuming, if we repeatedly have to set the format state and then reset because the character-at-atime input is interspersed with other input.
For this reason, istreams provide a get method. There are several versions of get, but
the easiest to use is the one with signature
istream & get( char & ch );

Thus, in
char ch;
if( cin.get( ch ) )
cout << "Read " << ch << endl;
else
cout << "Read error" << endl;

(or for wide characters)

wchar_t ch;
if( wcin.get( ch ) )
wcout << L"Read " << ch << endl;
else
wcout << L"Read error" << endl;

a single character is read, and if the read fails the stream is put in a bad state. The unget
method is used to undo a get. The peek method is used to examine the next character in the
input stream without digesting it. The declarations of these methods are:
istream & unget( );
int peek( );

c++book.mif Page 175 Tuesday, April 29, 2003 2:13 PM

Files

1
2
3
4
5
6
7
8
9
10

175

istream & getline( istream & in, string & str, char delim )
{
char ch;
str = "";
// empty string, will build one char at-a-time
while( in.get( ch ) && ch != delim )
str += ch;
return in;
}

Figure 9-6

getline routine implementation

The return type of peek is an int, in much the same way as the read method in
java.io.Reader returns a char in an int variable. If peek sees the end of input, it
returns EOF. Unlike Java, if the return value of peek is assigned to a variable of type char
prior to checking if it is EOF, the compiler will not complain, and the code is likely in error, particularly for unformatted data streams.
The function getline, which is not a member of istream, can be used to read a line of
input. The signature is
istream & getline( istream & in, string & str, char delim = '\n' );

getline reads characters from an input stream and forms a string str. Reading stops when
either delim is encountered or the end-of-file is reached. The delim character is not included
in the string but is removed from the input stream. Some older compilers have broken implementations of getline. The implementation in Figure 9-6 shows a plausible implementation
of getline (and not being a template, one that should get preferential consideration by the
compiler), and serves to illustrate the use of get.
Finally, member function ignore is used to skip characters. ignores signature is:
istream & ignore( int n = 1, int delim = EOF );

ignore reads and discards n characters from the istream, or all characters up to and including the delim, or until the end-of-file is encountered.

9.5 Files
Files are modelled by either an ifstream for input, or an ofstream for output (again there
are corresponding class template expansions for wide-character implementations). An
fstream can be used for both input and output, but we do not recommend it. To perform file I/
O, the standard header fstream should be included.
File streams can be constructed with a primitive string (either a string constant, null terminated array of character, or a result of c_str on a string object), and optionally a mode that
describes how the file is to be used. Some examples include:
ifstream file1( "data.txt" );

c++book.mif Page 176 Tuesday, April 29, 2003 2:13 PM

176

Chapter 9 Input and Output

char name2[ 100 ];

get( name2, 100 );
// reads a whole line or 99 chars
ofstream file2( name2 );
string name3
cin >> name3;
ofstream file3( name3.c_str( ), ios_base::out | ios_base::trunc );
string name4;
getline( cin, name4 );
ofstream file4( name4.c_str( ), ios_base::out | ios_base::app );

In these examples, first we see that an ifstream can be constructed with a file name. The second example constructs an ofstream with a primitive string (an array of character), and the
default output mode of truncation. Note that although cin>>name2 compiles, using it is very
dangerous, since it can lead to a buffer overflow. Invoking an overloaded version of get that
works with character arrays is a much safer solution. Option number 3 uses a string class
object, and we see that we can invoke c_str to obtain the primitive string that the ofstream
constructor requires. Also, we see an explicit use of a mode, in which we bitwise-or out and
trunc. This is the default. Finally we see a constructor that opens a file for appending. (Older
versions of C++ use ios instead of ios_base). Alternatively, we can simply declare the
stream object and use member function open later. The state of the stream should be tested after
it is opened, as in
if( file4 )
// ok
else
// not ok

When a stream goes out of scope, it is guaranteed that its destructor is called, thereby closing the stream automatically. The user can invoke close if it is desired to close the stream
sooner, perhaps to reopen a different file with the same stream object.

9.6 Random Access

Both the istream and ostream classes allow random access in the stream. This is most useful for file streams. The member function tellg (tellp for ostreams) returns the current
position in the stream, measured in characters. seekg and seekp allow the changing of the
current position.
seekg is used for istreams, and seekp for ostreams (g stands for get, p for put,
and we can presume that the different names are a result of avoiding ambiguity in the case of
multiple inheritance from both istream and ostream). The declaration for member function
seekg is:
istream & seekg( int offset, int whence = ios_base::beg );

As a result of seekg, the current position in the stream is changed to either

c++book.mif Page 177 Tuesday, April 29, 2003 2:13 PM

String Streams

177

1. offset characters from the beginning of the stream, if whence is ios_base::beg

2. offset characters from the current position in the stream, if whence is
ios_base::cur
3. offset characters from the end position in the stream, if whence is ios_base::end
offset may be negative. Generally, it is an error to attempt to seekg to before the beginning
of the stream. On some systems, notably Unix, a seekg past the end of the file is supported by
extending the file with undefined contents.
As an example, the routine in Figure 9-7 prints the last howMany characters in the
(binary) file fileName. After opening the file for reading, and checking for errors, we invoke
seekg at line 9 to go to the end. We then back up howMany characters, taking care to avoid
backing up to before the beginning. We do this by calling tellg at line 10 to see how large the
file is and use the smaller of the file size and howMany as the new value of howMany. Then at
line 13 we back up howMany characters, and finally, we can read characters with get and output them with put.

9.7 String Streams

The typical problem with operator>> for istreams is that all whitespace blanks, tabs,
and in particular, newlines are treated equally. So if we are expecting each line of a data file
to have two integers, and we write code such as

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

while( fin >> x >> y )

...
void lastChars( const string & fileName, int howMany )
{
ifstream fin( fileName.c_str( ), ios_base::binary );
if( !fin )
throw io_exception( );
// made this one up
else if( howMany <= 0 )
throw invalid_argument( "howMany is negative" );
fin.seekg( 0, ios_base::end );
int fileSize = fin.tellg( );
if( fileSize < howMany )
howMany = fileSize;
fin.seekg( -howMany, ios_base::cur );
char ch;
while( fin.get( ch ) )
cout.put( ch );
}

Figure 9-7

Routine that prints last howMany characters from fileName

c++book.mif Page 178 Tuesday, April 29, 2003 2:13 PM

178

Chapter 9 Input and Output

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27

public static void twoInts( BufferedReader fin )

throws IOException
{
String oneLine = null;
while( ( oneLine = fin.readLine( ) ) != null )
{
int x = 0, y = 0;
StringTokenizer st = new StringTokenizer( oneLine );
if( st.countTokens( ) != 2 )
{
System.err.println( "Skipping line: " + oneLine );
continue;
}
try
{
x = Integer.parseInt( st.nextToken( ) );
y = Integer.parseInt( st.nextToken( ) );
}
catch( NumberFormatException e )
{
System.err.println( "Skipping line: " + oneLine );
continue;
}
System.out.println( "Successfully read " + x + " " + y );
}
}

Figure 9-8

Java code to distinguish lines containing exactly two integers

a file that had only one int on each of the first two lines would be processed without error. If
we wanted to insist that every line had two and only two integers, we would need to do more
work.
In Java, we would read one line at a time into a String. Once we had the string, we
could parse it with a StringTokenizer, using code such as Figure 9-8.
This is exactly the behavior that can be implemented with an istringstream, which
like its companion ostringstream, is available by including the standard header sstream.
An istringstream is constructed by passing a string as a parameter. At that point, all of
the basic istream operators, including operator>> and testing of error states are available.
Note that the error states apply to the istringstream, and not the fstream.
Figure 9-9 shows the C++ implementation of twoInts, which is only cosmetically different from the Java code. Observe, first, that fin is a reference to an istream instead of an
ifstream. As with Java, it is always best to use the most generic type. Once we have the
istringstream at line 6, we can do two input operations, and then test the error state, skipping a line. Since there is no equivalent to countTokens to check if there are exactly two

c++book.mif Page 179 Tuesday, April 29, 2003 2:13 PM

179

endl

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

void twoInts( istream & fin ) throw( io_exception )

{
string oneLine;
while( getline( fin, oneLine ) )
{
istringstream st( oneLine );
int x, y;
st >> x;
st >> y;
if( st.fail( ) )
{
cerr << "Skipping line " << oneLine << endl;
continue;
}
string junk;
if( st >> junk )
{
cerr << "Skipping line " << oneLine << endl;
continue;
}
cout << "Successfully read " << x << " " << y << endl;
}
if( !fin.eof( ) )
throw io_exception( );
}

Figure 9-9

C++ code to distinguish lines containing exactly two integers

tokens, we attempt to read a string, which should fail. If it succeeds, we print an error message and go on to the next line. Note carefully that in this code, each iteration of the loop creates
a new istringstream object (on the runtime stack), destroying the original.
Declaring istringstream inside the while loop has the advantage that since each iteration creates a fresh istringstream, we do not have to clear the error state. The obvious disadvantage is the repeated calls to constructors and destructors. An alternative is to use the str
method of istringstream to change str. Then we can put the istringstream object
outside the loop, uninitialized, and set its string each time around the loop. However, now we
must clear the error state. Figure 9-10 shows this approach.
For ostringstream, writes can be directed to a string instead of standard output, files,
or other places. To extract the string from the ostringstream, invoke the str method. A
classic example is the conversion of any (printable) type to a string as shown in Figure 9-11.

9.8 endl
Because endl writes a newline and flushes the stream, using endl can be time-consuming

c++book.mif Page 180 Tuesday, April 29, 2003 2:13 PM

180

Chapter 9 Input and Output

1
2
3
4
5
6
7
8
9
10
11
12

void twoInts( istream & fin ) throw( io_exception )

{
string oneLine;
istringstream st;
while( getline( fin, oneLine ) )
{
st.str( oneLine );
st.clear( );
...
}
}

Figure 9-10

Revised code avoids excessive creation of istringstream objects

when there is significant disk-bound (or network bound I/O). In such a case, writing the "\n"
character directly can be more efficient.
One one of our machines, we copied from one file to another a line at a time, and measured the time spent writing. In this program which is almost exclusively I/O, we observed that
using endl is four times slower for writing, The files were approximately 2,000,000 lines, with
78,000,000 characters. When each line was written using endl, the time spent writing was
approximately 40 seconds. Ending the line with "\n" reduced the time to 10 seconds. Using the
character \n instead of a string "\n" did not affect the running time.
Needless to say, this consideration is important only if a significant portion of the running
time is spend performing I/O.

9.9 Serialization
Serialization is not part of standard C++. Each implementation might provide some customized
support for serialization, but certainly objects written by the implementation could only be read
in the same implementation.

9.10 Key Points

The C++ I/O library combines inheritance and templates. The templates are used to spec1
2
3
4
5
6
7

template <typename Object>

string toString( const Object & x )
{
ostringstream os;
os << x;
return os.str( );
}

Figure 9-11

Generic toString

c++book.mif Page 181 Tuesday, April 29, 2003 2:13 PM

Exercises

181

ify the underlying implementation of characters.

Four standard streams are defined for each character type: cin, cout, cerr, and clog
for char; wcin, wcout, wcerr, and wclog for wchar_t.
ifstream and istringstream are subclasses of istream.
ofstream and ostringstream are subclasses of ostream.
iostream extends both istream and ostream.
Error handling is performed by checking the state of the stream.
The most common I/O operations involve operator<< and operator>>.
Manipulators are used to control formatting options for output and digestion of white
space for input.
To repair an istream error, typically the stream state must be cleared first and then some
input must be digested (perhaps by reading a string).
The istream and ostream hierarchies allow for random access in the stream. Most
often this is used for file streams.
Files are closed automatically by destructors.
An istringstream can be used in the same way as a Java StringTokenizer.
An ostringstream can be used to print data to a string.
Sockets are not part of standard C++, but there is GNU sockets class that contains
TCPstream, which is part of the stream hierarchy.
Serialization is not part of standard C++.

9.11 Exercises
1. Describe the I/O hierarchy in C++.
2. What are the standard predefined I/O streams?
3. How are errors handled in the C++ I/O library?
4. What is wrong with the while(!fin.eof()) idiom?
5. What happens if a file is not closed by the user?
6. List some of the I/O manipulators.
7. What are istringstream and ostringstream?
8. Using Figure 9-11, implement a function template
string operator+( const string & lhs, const Object & rhs )

that mimics Javas string concatenation.

9. Write a routine that gives the character that is in the middle of a file (if the file has an even
size, return the character in position fileLength/2).
10. Suppose a data file consists of lines in which each line stores a string, another
string, and an integer. Write a routine that reads the file, returning a vector of objects
(each object is a class type representing a line). Any ill-formatted lines will generate an
error message and will be skipped.

c++book.mif Page 182 Tuesday, April 29, 2003 2:13 PM

182

Chapter 9 Input and Output

11. Write a program that processes include directives. Since an included file may itself contain
include directives, the basic algorithm is recursive.
12. Write a method that takes the name of a file as a parameter, and reverses the contents of
the file.

c++book.mif Page 183 Tuesday, April 29, 2003 2:13 PM

H A P T E R

1 0

Collections: The Standard

Template Library

N Java, the Collections API in package java.util

implements standard data structures such as lists, sets, and maps. C++ has a package that provides similar functionality, namely the Standard Template Library, which is known simply as the
STL. As the name suggests, the STL makes heavy use of templates.
In this chapter, we describe how the STL is organized and cover the basic containers and
iterators, as well as a small collection of algorithms such as sorting and searching.

10.1 Containers and Iterators

In Java, the Collections API makes heavy use of inheritance. Recall that the Collections API
defines an interface, Collection, and then subinterfaces List and Set, and Set itself has
subinterface SortedSet. These interface have various implementations. The Collections API
also defines an interface, Iterator, that can be used to traverse any collection. Each collection class is responsible for defining an appropriate iterator class. The Java idiom is that the iterator class is declared as a private inner class of the collection, and that each container provides a
factory method, iterator, that returns, using the Iterator interface type, a newly created
and initialized instance of the iterator it has defined. The Iterator interface in Java is limited
in that it contains only three methods: next, hasNext, and remove.
C++ has the same ideas of collections and iterators. However, the specific methods are
slightly different. C++ iterators are more flexible than Java iterators, and not surprisingly, operator overloading is used to replace named methods with operators.
10.1.1 Containers
C++ defines several container templates. Like Java, some collections are unordered; others, are

183

c++book.mif Page 184 Tuesday, April 29, 2003 2:13 PM

184

Chapter 10 Collections: The Standard Template Library

ordered. Some collections allow duplicates; others do not. All containers support the following
operations:
int size( ) const
void clear( )
bool empty( ) const

size returns the number of elements in the container; empty returns true if the container contains no elements and false otherwise.
Unlike Java, there is no universal add method; different containers use different names.
Some of the container class templates are vector, deque, list, set, multiset, map,
multimap, and priority_queue.
vector is the equivalent of an ArrayList. The add operation for vector is named
push_back. vector supports operator[]. list is the equivalent of LinkedList. Its
add operation is also named push_back, but list does not support operator[]. However, list does support push_front. deque is an array-based data structure that supports
efficient indexing with operator[], and both push_front and push_back, all in constant time per operation.
set is the equivalent of TreeSet. The add operator for set is insert. multiset
allows duplicates, whereas set does not. map is the equivalent of TreeMap. The add operation is insert, but one must pass the key and value in a single pair object. However, map also
provides an overloaded operator[] that makes the map look just like an array. A
multimap allows duplicate keys.
The STL also contains a priority_queue class. Its operation is known as push.
Comparing these collections with Java, we see that STL supports sets and maps that contain
duplicates, as well as the priority queue, but does not support searching with hash tables.
10.1.2 Iterators
In Java, each container defines an internal iterator type, but exports it through the Iterator
interface type. In C++, each container defines several iterator types, and these specific iterator
types are used by the programmer instead of an abstract type.
For instance, if we have a vector<int>, the basic iterator type is
vector<int>::iterator. Another iterator type, vector<int>::const_iterator,
does not allow changes to the container on which the iterator is operating. This implies that the
basic iterator can be used to change the container.
All iterators are guaranteed to have at least the following set of operations:
++itr and itr++ advance the iterator itr to the next location. Both the prefix and
postfix forms are available. This does not cause any change to container. Some iterators support
--itr and itr--. Those iterators are called bidirectional iterators. Some iterators support
both itr+=k and itr+k. Those iterators are called random-access iterators. itr+=k
advances the iterator k positions. itr+k returns a new iterator that is k positions ahead of itr.

c++book.mif Page 185 Tuesday, April 29, 2003 2:13 PM

Containers and Iterators

1
2
3
4
5
6
7
8

185

template <typename Container>

void print( const Container & c, ostream & out = cout )
{
typename Container::const_iterator itr;
for( itr = c.begin( ); itr != c.end( ); ++itr )
out << *itr << " ";
out << endl;
}

Figure 10-1

Routine to print any container

*itr returns a reference to the container object that itr is currently representing. The
reference that is returned is modifiable for basic iterators, but is not modifiable (i.e. a constant
reference) for const_iterators.
itr1==itr2 returns true if iterators itr1 and itr2 refer to the same position in the
same container and false otherwise.
itr1!=itr2 returns true if iterators itr1 and itr2 refer to different positions or different containers and false otherwise.
In order to use an iterator, we must obtain one from a container. The way this is done in
C++ is that the container has two methods, begin and end that return iterators. Each collection
defines four methods:
iterator begin( );
const_iterator begin( ) const;
iterator end( );
const_iterator end( ) const;

begin returns an iterator that is positioned at the first item in the container. end returns
an iterator that is position at the endmarker, which represents a position one past the last element
in the container. For instance, on an empty container, begin and end return the same position.
begin and end both make use of the fact that identical-looking methods can be overloaded if one is an accessor and one is a mutator. So if begin is invoked on a constant container, we will get a const_iterator, which wont support any changes to the container. If
begin is invoked on a mutable container, we will get an iterator, which can be used to
change the container.
Typically we initialize a local iterator to be a copy of the begin iterator, and have it step
through the container, stopping as soon as it hits the endmarker. As an example, Figure 10-1
shows a print function that prints the elements of any container, provided that the elements in
the container have provided an operator<<. If the container is a set, its elements are output
in sorted order. Figure 10-2 illustrates four different containers that invoke the print function,
along with the expected output (in comments). Observe that both set and multiset output in
sorted order, with multiset allowing the second insertion of foo.

c++book.mif Page 186 Tuesday, April 29, 2003 2:13 PM

186

Chapter 10 Collections: The Standard Template Library

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28

#include <iostream>
#include <vector>
#include <list>
#include <set>
#include <string>
using namespace std;
int main( )
{
vector<int> vec;
vec.push_back( 3 ); vec.push_back( 4 );
list<double> lst;
lst.push_back( 3.14 ); lst.push_front( 6.28 );
set<string> s;
s.insert( "foo" ); s.insert( "bar" ); s.insert( "foo" );
multiset<string> ms;
ms.insert( "foo" ); ms.insert( "bar" ); ms.insert( "foo" );
print(
print(
print(
print(

vec );
lst );
s );
ms );

//
//
//
//

3 4
6.28 3.14
bar foo
bar foo foo

return 0;
}

Figure 10-2

Example of invoking print with four different containers

10.1.3 Pairs
If we try to print a map, the program will not compile immediately because the elements of a
map are pairs of keys and values. If operator<< is overloaded for pair, then we can in fact
use the print routine. Figure 10-3 illustrates the general strategy.
As expected, pair is a class template, and stores two data members first and
second, which can be directly accessed, without invoking methods. So we can easily overload
operator<< to output a pair, assuming its components first and second have done so
too.
In Figure 10-4 we can create a map that stores the name of a city and the zip code, both as
strings. (The zip code cannot be an int, since many zip codes begin with 0). Lines 10 and 11
show that a map stores pair objects, and the pair objects can be added by calling insert. Line
12 shows the much more natural equivalent that makes use of operator overloading. We describe
maps in more detail in Section 10.7.

c++book.mif Page 187 Tuesday, April 29, 2003 2:13 PM

Containers and Iterators

1
2
3
4
5
6
7
8
9

187

#include <iostream>
#include <map>
using namespace std;
template <typename Type1, typename Type2>
ostream & operator<<( ostream & out, const pair<Type1,Type2> & p )
{
return out << "[" << p.first << "," << p.second << "]";
}

Figure 10-3

Overloading operator<< for pairs.

10.1.4 Inheritance and Containers

Recall that because of slicing, derived class objects cannot be copied into base class objects.
Unlike Java, the STL stores copies of objects, not simply references to the objects. Thus, heterogeneous containers storing multiple types of compatible objects should store pointers to the
objects, rather than the objects themselves.
Figure 10-5 shows a vector that stores pointers to both Student and Person (these
classes were defined in Figure 6-2 and Figure 6-4, respectively). Observe that we must again
overload operator<<, because if we do not, the existing operator<< that outputs the value
of the pointer (the memory address of the object it is pointing at) will be used. By providing our
own, with the base class type as a parameter, we have a better match than the existing version
that accepts a generic void * as a parameter.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

#include <iostream>
#include <map>
#include <string>
using namespace std;
int main( )
{
map<string,string> zip;
zip.insert( pair<string,string>( "Miami", "33199" ) );
zip.insert( pair<string,string>( "Princeton", "08544" ) );
zip[ "Boston" ] = "02134";
// Prints: [Boston,02134] [Miami,33199] [Princeton,08544]
print( zip );
return 0;
}

Figure 10-4

Example of invoking print with maps

c++book.mif Page 188 Tuesday, April 29, 2003 2:13 PM

188

Chapter 10 Collections: The Standard Template Library

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

ostream & operator<< ( ostream & out, const Person *p )

{
return out << *p;
}
int main( )
{
vector<Person *> vec;
vec.push_back( new Person( 987654321, "Bob" ) );
vec.push_back( new Student( 123456789, "Jane", 4.0 ) );
print( vec );
return 0;
}

Figure 10-5

Heterogeneous containers

It is tempting to make operator<< a function template, as in:

template <typename Object>
ostream & operator<< ( ostream & out, const Object *p )
{ return out << *p; }

However, doing so is dangerous because operator<< is already a template, so ambiguities

(for instance, with char *, the primitive string) may result.
10.1.5 Constructors
All containers can be constructed from other containers. However, instead of a constructor that
accepts another container, the constructors will accept a pair of iterators representing the first
item from the other container and the first non-included item from the other container. Thus, for
instance,
vector<int> clone( original.begin( ), original.end( ) );

constructs a new vector clone with the same elements as any container original.

10.2 Sequence Containers: vector and list

The vector is the STL equivalent of ArrayList and list is the STL equivalent of
LinkedList. deque has no Java counterpart, and stores items in an array that appears to
grow in both directions (several plausible implementations are known to accomplish this, one of
which maintains the deque as a pair of vectors that grow in opposite directions, and which
reallocates the elements evenly if removal causes one of the vectors to empty).
One of the issues that we consider in C++ is that although expansion of an array is relatively cheap when it needs to occur, in Java references are relocated to a larger array, whereas in
C++ entire objects are relocated. If these are large objects rather than pointers, this can be more

c++book.mif Page 189 Tuesday, April 29, 2003 2:13 PM

Sequence Containers: vector and list

189

costly than expected. Thus lists in C++ have the additional C++ advantage of generally
requiring less data movement compared to arrays when objects are large and a reasonable estimate of the vector capacity is not available at the start.
The basic operations that are supported by both containers are:
void push_back( const Object & x );
Object & back( );
void pop_back( );
Object & front( );
iterator insert( iterator pos, const Object & x );
iterator erase( iterator pos );
iterator erase( iterator start, iterator end );

push_back adds x to the end of the container. back returns the object at the end of the
container; an accessor is also defined that returns a constant reference. pop_back removes the
object at the end of the container. front returns the object at the front of the container.
insert adds x into the container, prior to the position given by the iterator. This is a constant
time operation for list, but not for vector or deque. insert returns an iterator representing the position of the inserted item.
Adding x to the front of c could be implemented as
c.insert( c.begin( ), x );

Adding x to the back of c could be implemented as

c.insert( c.end( ), x );

since end returns the endmarker.

The one-parameter erase removes the object at the position given by the iterator, and is
constant time for list, but not vector or deque. It returns the position of the element that
followed pos prior to the call to erase. Most importantly, this operation invalidates pos,
which is now stale. Typically pos is reset to the return value of erase. Removing the first and
last elements of container c can be done with
c.erase( c.begin( ) );
c.erase( --c.end( ) );

In the second call, observe that the return value from c.end() is an unnamed temporary whose
position represents the endmarker. Thus the -- operator changes the state of the unnamed temporary to view the last item in the container. After the erase method is called, the unnamed
temporarys destructor is invoked.
Two-parameter erase removes all items beginning at position start, up to but not
including end. The idea of a range being half-open-ended is similar to substring operations in
java.util.String. It means that an entire container can be erased by the call:
c.erase( c.begin( ), c.end( ) );

A possible implementation of erase is:

c++book.mif Page 190 Tuesday, April 29, 2003 2:13 PM

190

Chapter 10 Collections: The Standard Template Library

template <typename Object>

list<Object>::iterator list<Object>::erase( iterator from,
iterator to )
{
for( iterator itr = from; itr != to; )
itr = erase( itr );
return to;
}

This implementation is inefficient for vector, in which it is better to remove elements in

reverse order. For instance, in the worst-case, this algorithm applied to empty the entire
vector would consume quadratic time. An alternate algorithm is
template <typename Object>
list<Object>::iterator list<Object>::erase( iterator from,
iterator to )
{
--from;
for( iterator itr = --to; itr != from; )
itr = --erase( itr );
return ++from;
}

This code makes use of the fact that corresponding to the endmarker is the beginmarker,
which is a valid position, but whose contents should not be accessed. One can see that traversing
in reverse order is messy. We will revisit this in Section 10.4.
For deque and list, two additional operations are available with expected semantics:
void push_front( const Object & x );
void pop_front( );

The list also provides a slice operation that allows the transfer of a sublist to somewhere else.
For vector and deque, additional operations include
Object & operator[] ( int idx );
Object & at( int idx );
int capacity( ) const;
void reserve( int newCapacity );

// vector only

operator[] and at come in both accessor and mutator versions and support array
indexing. at does bounds checking. capacity returns the internal capacity, and reserve
can be used to set the new capacity. reserve is only available for vector. If a good estimate
is available, it can be used to avoid array expansion.
In addition to the usual set of constructors, a vector can be constructed with either an
initial size, or an initial size and an initial value for all the elements. For instance,
vector<int> v1( 20 );
vector<int> v2( 30, 37 );

// 20 ints
// 30 ints, all with value 37

c++book.mif Page 191 Tuesday, April 29, 2003 2:13 PM

Error Checks in the STL

191

10.3 Error Checks in the STL

The compiler does significant type-checking because of the template expansion mechanism.
Thus, many type errors that are detected only at runtime in the Java Collections API (bad downcasts, etc.) are detected at compile time in C++.
However, except for member function at, virtually no other STL members do any kind of
error checking at runtime. Thus applying * to an iterator that is out of bounds (for instance at the
beginmarker or endmarker) or is stale can have dire consequences, or fail in an undetectable
manner (which is probably worse!). Comparing iterators that are in different containers will
always produce false, rather than throwing an exception. Finally, certain mutating operations on
containers cause the associated iterators to become invalid.
Javas rule is that any structural modification by a container method invalidates all iterators viewing the container, any structural modification by the iterator invalidates all other iterators, and a ConcurrentModificationException will be thrown on an attempt to use an
invalidated iterator.
C++s rule is that some, but not all, operations invalidate an iterator, but the result is simply that the iterator is stale. For instance, push_back on a list does not invalidate any iterators. push_back on a vector invalidates all iterators. erase on a list invalidates all
iterators that were viewing erased elements. erase on a vector invalidates all iterators that
were at or past any erased element. The logic in both cases is that an iterator may be implemented internally as a pointer variable. In the case of a list, it is a pointer to a list node. In the
case of the vector, it is a pointer into the underlying primitive array at the particular item.
Because adding to an array could cause a resize, the pointer could become stale. On the other
hand, adding to a linked list would never invalidate the pointer to a list node. In C++, you must
be extremely careful when using iterators, since the runtime system will not provide much of a
safety net.

10.4 Other Iterators

Two types of iterations get special treatment in C++. First, as we saw in Section 10.2, often we
want to traverse a container backwards. This is usually awkward, because of the asymmetry
between begin and end: begin is the first element, but end is the endmarker. Another type
of iteration concerns the idea of repeatedly reading or writing to a file. We briefly discuss some
C++ goodies to handle both reverse iteration and stream iterations.
10.4.1 Reverse Iterators
Suppose we want to print any collection backwards. Using iterators, the code is surprisingly
messy, as shown in Figure 10-6, and uses a bidirectional iterator. At line 7, we initialize itr to
the last position in the container (the position prior to the endmarker), and we loop until we are
at the first position. Then we print the item in the first position (if there is one). What a mess.
A reverse iterator allows us to do to reverse traversal more naturally. First, each of the
standard container defines member function rbegin and rend that return reverse iterators rep-

c++book.mif Page 192 Tuesday, April 29, 2003 2:13 PM

192

Chapter 10 Collections: The Standard Template Library

1
2
3
4
5
6
7
8
9
10
11
12

// Print the contents of Container c in reverse

template <typename Container>
void printReverse( const Container & c, ostream & out = cout )
{
typename Container::const_iterator itr;
for( itr = c.end( ), --itr; itr != c.begin( ); --itr )
out << *itr << " ";
if( !c.empty( ) )
out << *c.begin( );
out << endl;
}

Figure 10-6

Awkward routine to print a container in reverse with normal iterator

resenting the last position (not the endmarker), and the beginmarker (not the first position). The
reverse iterator is reverse_iterator or const_reverse_iterator, as appropriate,
and for a reverse iterator, ++ moves toward the front while -- moves toward the rear, opposite
to normal iterator semantics.
10.4.2 Stream Iterators
A stream iterator allows us to repeatedly read items or write items. Outputting items is discussed in Section 10.10 in conjunction with the copy algorithm. Use of the input stream iterator
is best done with an example.
Suppose we want to read words from a file "dict.txt" into a set. Assume we are not
concerned with errors. Then we already know we can use the following code:
ifstream fin( "dict.txt" );
string x;
set<string> s;
while( fin >> x )
1
2
3
4
5
6
7
8
9
10

s.insert( x );
// Print the contents of Container c in reverse
template <typename Container>
void printReverse( const Container & c, ostream & out = cout )
{
typename Container::const_reverse_iterator itr;
for( itr = c.rbegin( ); itr != c.rend( ); ++itr )
out << *itr << " ";
out << endl;
}

Figure 10-7

Routine to print a container in reverse with reverse iterator

c++book.mif Page 193 Tuesday, April 29, 2003 2:13 PM

Stacks and Queues

1
2
3
4
5
6
7
8
9
10
11
12
13
14

193

#include <queue>
#include <iostream>
#include <list>
using namespace std;
int main( )
{
queue<int,list<int> > q;
q.push( 37 ); q.push( 111 );
for( ; !q.empty( ); q.pop( ) )
cout << q.front( ) << endl;
return 0;
}

Figure 10-8

Using the queue adapter

Recall that the containers have constructors that take two iterators. The stream iterator
allows us to rewrite this code, passing to it a pair of istream_iterator<string> objects.
The first object represents the start of the file, the second represents the end-of-file marker. The
code becomes:
ifstream fin( "dict.txt" )
set<string> s( istream_iterator<string>( fin ),
istream_iterator<string>( ) );

10.5 Stacks and Queues

The STL contains class templates named stack and queue. Clearly these are not needed,
since vector, list, and deque all implement stack operations with push_back, back,
and pop_back, while list and deque implement queue operations with push_back,
front, and pop_front.
Thus what the STL provides are adapters, that add nicer names to the interface. For
stack, the member functions are push, pop, and top. For queue, we get push, front,
and pop (which are not clear improvements).
The stack and queue templates require that the user specify the type of objects and the
underlying representation. By default, a vector is used for stack and a deque for queue.
The standard header stack must be included for stacks, while the standard header queue
must be included for queues. Figure 10-8 shows how to use the queue adapter.

10.6 Sets
The set class template in C++ behaves in the same manner as Java. A set does not allow
duplicates, and by default, iteration of a set views items in the default order. However, sets
can use a function object to override the default ordering.

c++book.mif Page 194 Tuesday, April 29, 2003 2:13 PM

194

Chapter 10 Collections: The Standard Template Library

10.6.1 Standard Function Objects

C++ defines several class templates in the standard header functional that implement function objects. These function object templates are less, greater, equal_to,
not_equal_to, great_equal and less_equal. The default for a set is an appropriately instantiated less template. The less template is easy to code, as shown in Figure 10-9
and illustrates that by default, a set expects to have an operator< available.
10.6.2 set Operations
The set provides several member functions in addition to the usual suspects, including:
pair<iterator,bool> insert( const Object & x );
pair<iterator,bool> insert( iterator hint, const Object & x );
iterator find( const Object & x ) const;
int erase( const Object & x );
iterator erase( iterator itr );
iterator erase( iterator start, iterator end );
iterator lower_bound( const Object & x );
iterator upper_bound( const Object & x );
pair<iterator,iterator> equal_range( const Object & x );

insert adds x to the set. Since duplicates are not allowed, if x is not present, the
returned pair will contain the iterator representing the already contained x, and false. Otherwise, it will contain the iterator representing the newly inserted x, and true. The two-parameter insert allows specification of a hint, representing the position where x should go. If the
hint is accurate, the insertion is fast. If not, the insertion still has performance that is comparable
to the one-parameter insert. For instance, the following code might be faster using the twoparameter insert than the one-parameter insert:
set<int> s;
for( int i = 0; i < 1000000; i++ )
s.insert( s.end( ), i );

find returns an iterator representing the position of x in the set. If x is not found, the
endmarker is returned. The various erase routines behave in the same manner as for the
sequence containers in Section 10.2, except that erase that takes Object x returns the number of items removed. In a set this is 0 or 1, but could be larger in a multiset.
1
2
3
4
5
6
7
8

template <typename Object>

class less
{
public:
bool operator( ) ( const Object & lhs,
const Object & rhs ) const
{ return lhs < rhs; }
};

Figure 10-9

Implementation of the less function template

c++book.mif Page 195 Tuesday, April 29, 2003 2:13 PM

Sets

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

195

class PtrToPersonLess
{
public:
bool operator() ( const Person *lhs, const Person *rhs ) const
{ return lhs->getSsn( ) < rhs->getSsn( ); }
};
int main( )
{
set<Person *, PtrToPersonLess> s;
s.insert( new Person( 987654321, "Bob" ) );
s.insert( new Student( 123456789, "Jane", 4.0 ) );
print( s );
return 0;
}

Figure 10-10 Heterogeneous set

lower_bound returns an iterator to the first element in the set with a key that is greater
than or equal to x. upper_bound returns an iterator to the first element in the set with a key
that is greater than x. equal_range returns a pair of iterators representing lower_bound
and upper_bound. These routines are typically most useful in multisets.
10.6.3 multisets
A multiset is like a set except that duplicates are allowed. The return type of insert is
modified to indicate that the insert always succeeds. As a result, we no longer need a pair,
but can simply return an iterator representing the newly inserted x.
iterator insert( const Object & x );
iterator insert( iterator hint, const Object & x );

For the multiset, the erase member function that takes an Object x removes all
occurrences of x. To simply remove one occurrence, use the erase member function that takes
an iterator. To find all occurrences of x, we cannot simply call find; that returns an iterator referencing one occurrence (if there is one), but which specific occurrence is returned is not guaranteed. Instead, the range returned by lower_bound and upper_bound (with
upper_bound not included) contains all of occurrences of x; typically this is obtained by a
call to equal_range.
10.6.4 Using Function Objects to Change Default Ordering
Just as a Java TreeSet can be specified to use either the default ordering or an ordering given
by a comparator, a C++ set uses either a default ordering (operator<) or one provided by a

c++book.mif Page 196 Tuesday, April 29, 2003 2:13 PM

196

Chapter 10 Collections: The Standard Template Library

map<string,double> salaries;

1
2
3
4
5
6
7
8
9
10
11
12

salaries[ "Pat" ] = 75000.00;

cout << salaries[ "Pat" ] << endl;
cout << salaries[ "Jan" ] << endl;
map<string,double>::const_iterator itr;
itr = salaries.find( "Chris" );
if( itr == salaries.end( ) )
cout << "Not an employee of this company!" << endl;
else
cout << itr->second << endl;

Figure 10-11 Accessing values in a map

function object. Recall from Section 7.6.3 that the function object idiom in C++ is implemented
by providing a class that contains an overloaded operator(), and then instantiating a template with the class name as a template parameter. Figure 10-10 illustrates the idiom, in which
the code seen earlier in Figure 10-5 is adapted to use a set instead of a vector.

10.7 Maps
As we have already seen, a map behaves like a set instantiated with a pair representing a key
and value, with a comparison function that refers only to the key. Thus it supports all of the set
operations, including insert, but as we saw in Figure 10-4, we must insert a properly instantiated pair. The find operation for maps requires only a key, but the iterator that it returns references a pair. Similarly, erase requires only a key, and otherwise behaves like the sets
erase.
Most importantly, the map overloads the array indexing operator[]:
ValueType & operator[] ( const KeyType & key )

The semantics of operator[] are as follows. If the key is present in the map, a reference to the value is returned. If the key is not present in the map, it is inserted with a default
value into the map, and then a reference to the inserted default value is returned. The default
value is obtained by applying a zero-parameter constructor, or is zero for the primitive types.
These semantics do not allow an accessor version of operator[], and so operator[] cannot be used on a map that is constant. For instance, if a map is passed by constant reference,
inside the routine, operator[] is unusable. This could be a case where casting away constness is useful.
The code snippet in Figure 10-11 illustrates two techniques to access items in a map. First
observe that at line 3, the left-hand-side invokes operator[], thus inserting "Pat" and a
double of value 0 into the map, returning a reference to that double. Then the assignment
changes that double, inside the map, to 75000. Line 4 outputs 75000. Unfortunately, line 5
inserts "Jan" and a salary of 0.0 into the map, and then prints it. This may or may not be the

c++book.mif Page 197 Tuesday, April 29, 2003 2:13 PM

STL Example

197

proper thing to, depending on the application. If it is important to distinguish between items that
are in the map and not in the map, or if it is important to not insert into the map (because it is
immutable), then an alternate approach shown at lines 7 to 12 can be used. There we see a call to
find. If the key is not found, the iterator is the endmarker, and can be tested. If the key is
found, we can access the second item in pair referenced by the iterator, which is the value for
the key. We could make a change to itr->second if instead of a const_iterator, itr is
an iterator.
10.7.1 Multimaps
A multimap is a map in which duplicate keys are allowed. In Java, the effect of a multimap is
achieved by using a map whose values are Lists.
Otherwise, multimaps behave like maps but do not support operator[].

10.8 STL Example

As an example of the use of STL, we write a function that reads from an istream a sequence
of words and outputs onto an ostream the list of words along with the line numbers on which
the words occur. Thus the call
printConcordance( ifstream( "data.txt" ), cout );

will read from file data.txt, and send the list of words to standard output. For simplicity, we
assume that a word is any sequence of non-white space characters, and thus can easily be
extracted by operator>>.
The idea is to use a map, in which the keys are the words and the line numbers are stored
in a vector<int> or list<int>. Well use vector<int>, but typedef
vector<int> as LList, so we can change it later without much effort. When we see an
existing word, we add its line number to the list that is stored in the map. If the word is new, we
add its line number to the empty list that is created when the word is inserted into the map. After
all the words have been read, we can iterate through the map, and output the words and their line
numbers, and since the keys are sorted, the output will be sorted by words. The line numbers
themselves are output with an iterator.
Complete code is shown in Figure 10-12. The previously mentioned typedef is shown at
line 10. We discuss printConcordance first. The loop from lines 25 to 32 populates the
map. We read one line at a time and then create an istreamstring object at line 27. The
loop at lines 30 and 31 steps though each word on the line and adds it to the map. The tricky
code is clearly line 31. If the word is already seen, then operator[] returns a reference to the
list that is in the map, and thus push_back simply adds a new line number to the end of the
list. If the word is new, operator[] adds it into the map, with an empty list, and returns a reference to the empty list that is now in the map. push_back makes this a one-element list.
We can then step through the map with the standard loop at lines 34 to 36. operator<<
for the pair that is in the map is overloaded at lines 11 to 17.

c++book.mif Page 198 Tuesday, April 29, 2003 2:13 PM

198

Chapter 10 Collections: The Standard Template Library

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37

#include <iostream>
#include <fstream>
#include <sstream>
#include <map>
#include <string>
#include <vector>
#include <iomanip>
using namespace std;
typedef vector<int> LList;
ostream & operator<<( ostream & out,
const pair<string,LList> & rhs )
{
out << left << setw( 20 ) << rhs.first;
print( rhs.second, out );
// Figure 10-1
return out;
}
void printConcordance( istream & in, ostream & out )
{
string
oneLine;
map<string,LList> wordMap;
// Read the words; add them to wordMap
for( int lineNum = 1; getline( in, oneLine ); lineNum++ )
{
istringstream st( oneLine );
string word;
while( st >> word )
wordMap[ word ].push_back( lineNum );
}
map<string,LList>::iterator itr;
for( itr = wordMap.begin( ); itr != wordMap.end( ); ++itr )
out << *itr << endl;
}

Figure 10-12 The concordance routine

First, we print the word which is the first data member, and second we output the list of
line numbers, which is the second data member.

10.9 Priority Queue

The priority_queue template, found in the standard header queue contains several member functions, most importantly:

c++book.mif Page 199 Tuesday, April 29, 2003 2:13 PM

Priority Queue

199

void push( const Object & x );

const Object & top( ) const;
void pop( );

push adds x to the priority queue, top returns the largest element in the priority queue,
and pop removes the largest element from the priority queue. Duplicates are allowed; if there
are several largest elements, only one of them is removed.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39

#include <vector>
#include <queue>
#include <functional>
#include <string>
#include <iostream>
using namespace std;
// Empty the priority queue and print its contents.
template <typename PQueue>
void dump( const string & msg, PQueue & pq )
{
if( pq.empty( ) )
cout << msg << " is empty" << endl;
else
{
cout << msg << ": " << pq.top( );
pq.pop( );
while( !pq.empty( ) )
{
cout << " " << pq.top( );
pq.pop( );
}
cout << endl;
}
}
int main( )
{
priority_queue<int>
maxpq;
priority_queue<int,vector<int>,greater<int> > minpq;
minpq.push( 8 ); minpq.push( 5 ); minpq.push( 5 );
maxpq.push( 8 ); maxpq.push( 5 ); maxpq.push( 5 );
dump( "minpq", minpq );
dump( "maxpq", maxpq );
return 0;
}

Figure 10-13 Priority queues in C++

// minpq: 5 5 8
// maxpq: 8 5 5

c++book.mif Page 200 Tuesday, April 29, 2003 2:13 PM

200

Chapter 10 Collections: The Standard Template Library

Sometimes priority queues are set up to remove and access the smallest item instead of the
largest item. In such a case, the priority queue can be instantiated with an appropriate greater
function object to override the default ordering.
The priority queue template is instantiated with an item type, the container type (as in
stack and queue), and the comparator, with defaults allowed for the last two parameters. In
Figure 10-13, line 29 shows the default instantiation of priority_queue, that allows access
to the largest items, while line 30 shows an instantiation that allows access to the smallest item.

10.10 Generic Algorithms

The Standard Library includes a rich set of functions that can be applied to the standard containers. These are called the generic algorithms, and are used by including the standard header
algorithm. Some of the algorithms include routines for sorting, searching, and copying (possibly with substitutions). In all there are over sixty generic algorithms. In this section we highlight a few of the most useful algorithms.
10.10.1 Sorting
Sorting in C++ is accomplished by use of function template sort. The parameters to sort represent the start and endmarker of a (range in a) container, and an optional comparator:
void sort( Iterator begin, Iterator end );
void sort( Iterator begin, Iterator end, Comparator cmp );

The iterators must support random access. The sort algorithm does not guarantee that equal
items retain their original order. For that, we can use stable_sort instead of sort.
As an example, in
sort( v.begin( ), v.end( ) );
sort( v.begin( ), v.end( ), greater<int>( ) );
sort( v.begin( ), v.begin( ) + ( v.end( ) - v.begin( ) ) / 2 );

the first call sorts the entire container, v, in non-decreasing order. The second call sorts the entire
container in non-increasing order. The third call sorts the first half of the container in nondecreasing order. Note that (v.begin()+v.end())/2 is not allowed; instead we can compute a separation distance, halve it, and add it to the begin iterator.
The sorting algorithm is generally quicksort, which yields an O( N log N ) algorithm on
average. However, O( N log N ) worst-case performance is not guaranteed. In addition to sorting, there are also algorithms for selection, shuffling, partitioning, partial sorting, reversing,
rotating, and merging,
10.10.2 Searching
Several generic searching algorithms are available for containers. The two most basic are:
Iterator find( Iterator begin, Iterator end, const Object & x );
Iterator find_if( Iterator begin, Iterator end, Predicate pred );

c++book.mif Page 201 Tuesday, April 29, 2003 2:13 PM

Generic Algorithms

201

find returns an iterator representing the first occurrence of x in the range specified by begin
and end, or end if x is not found. find_if returns an iterator representing the first occurrence of an object for which the function object pred would return true, or end if no match is
found.
For instance, suppose we want to find the first occurrence of a string of length exactly 9 in
a vector<string>. First, we define a function object that expresses this condition:
class StringLengthComp
{
public:
bool operator() ( const string & s ) const
{ return s.length( ) == 9; }
};

However, for additional code reuse, a class template might be better:

template <int len>
class StringLength
{
public:
bool operator() ( const string & s ) const
{ return s.length( ) == len; }
};

Then, at the end of the following code fragment, in which v is of type

vector<string>, and itr is of type vector<string>::const_iterator will
either be v.end() or will be located at a string of length 9:
itr = find_if( v.begin( ), v.end( ), StringLength<9>( ) );

Thus, find is implemented as a call to find_if with an appropriately instantiated

equal_to as the third parameter.
There are a host of generic algorithms that are available. We list a few of the common
ones.
binary_search is used to search a sorted range for an object. A comparator can be
provided, or the default ordering can be used. equal_range, lower_bound and
upper_bound search sorted ranges and behave with the same semantics as the identically
named member functions in set. min_element can be used to find the smallest item in a
range, and can be invoked with or without a comparator. count returns the number of occurrences of an object in a range delimited by a pair of iterators. count_if returns the number of
objects in a range delimited by a pair of iterators that are true according to a predicate.
adjacent_find returns an iterator referring to the first element such that a predicate
returns true when applied to the element and its predecessor. The default predicate is
equal_to, in which case adjacent_find finds the first occurrence of an element that
equals the next element.

c++book.mif Page 202 Tuesday, April 29, 2003 2:13 PM

202

Chapter 10 Collections: The Standard Template Library

find_first_of takes four iterators representing two sequences (the first sequence and
the second sequence). It returns an iterator representing the first occurrence of any of the elements in the second sequence. For instance,
vector<int> wins; //
wins.push_back( 37337
wins.push_back( 46521
wins.push_back( 53810

store winning numbers

);
),
);

vector<int> myNumbers;
...
// populate myNumbers
vector<int>::const_iterator itr;
itr = find_first_of( myNumbers.begin( ), myNumbers.end( ),
wins.begin( ), wins.end( ) );

searches the vector myNumbers for any of the numbers in wins and returns an iterator representing the first occurrence of such a number in myNumbers.
A fifth parameter, a predicate, can be used to decide if an item in the second sequence is a
match for an item in the first sequence. Of course, the default is equal_to.
10.10.3 Unary Binder Adapters
The find_if algorithm requires that we pass a function object that returns a Boolean. This is
called a predicate. As we know, the library defines several templates that are predicates: less,
equal_to, etc. However these predicates are not directly suitable for find_if, because their
operator() requires two parameters. find_if requires unary predicates, which are predicates that take only one parameter.
A unary binder adapter is function that takes a binary predicate, such as less (appropriately instantiated) and a value, and constructs a unary predicate in which the value is used as one
of the two parameters to the binary predicate. This is called binding a parameter. We can bind
the first parameter with bind1st or we can bind the second parameter with bind2nd.
For instance, the result of bind2nd(less<int>(),0) is a new unary predicate. If
this new function object is f1, then any call f1(x) is interpreted as less<int>()(x,0).
Similarly, if f2 represents the function object given by bind1st(0,less<int>()), then
any call f2(x) is interpreted as less<int>()(0,x).
If we want to find the first negative number in a vector, v, of ints, we can use:
itr = find_if( v.begin( ), v.end( ), bind2nd( less<int>( ), 0 ) );

10.10.4 Copying
There are several algorithms that deal with copying. These include copy, copy_backwards,
remove, remove_copy, remove_if, remove_copy_if, replace, replace_copy,
replace_if, replace_copy_if, unique, unique_copy. Method names that have
copy and non-copy versions differ in whether they change the original, or leave the original
unchanged and produce a new sequence. A typical routine is copy:

c++book.mif Page 203 Tuesday, April 29, 2003 2:13 PM

BitSets

203

Iterator copy( Iterator begin, Iterator end, Iterator target );

The copy algorithm copies the range specified by begin and end to the target. The target
must be modifiable and be large enough to store the result. For instance, the code
vector<int> source( 10, 37 );
vector<int> target( 10 );
ostream_iterator<int> out( cout, "\n" );
copy( source.begin( ), source.end( ), target.begin( ) );
copy( source.begin( ), source.end( ), out );

copies the contents of source to target, changing all those values to 37, and then copies the
contents to the standard output stream, printing newlines after each item.
On the other hand, if we had
vector<int> smallArray;
// size is 0
copy( source.begin( ), source.end( ), smallArray.begin( ) );

the call to copy generates difficulties at runtime because the target array is not large enough.
10.10.5 Inserter Adapters
The last example in Section 10.10.4 illustrates a problem with copying. Often, we want the target of a copy (for instance replace_copy_if) to be an initially empty container. Rather than
providing another set of functions, C++ provides inserter adapters. The idea of an inserter
adapter is that we create an iterator, whose operator= is altered.
The most common adapter is the back_inserter. The back_inserter is constructed with a sequence container, and generates an iterator whose assignment operator is
replaced by a push_back operation. As a result, the effect of
copy( source.begin( ), source.end( ), back_inserter( target ) );

is to copy the contents of source to the end of target, lengthening target.

Two other adapters are front_inserter, which not surprisingly causes a call to
push_front (and cannot be used for vector), and inserter, which causes a call to
insert, and must be constructed with both a container and an iterator specifying the insertion
point. Thus, an equivalent call to the most recent copy above is the following:
copy( source.begin( ), source.end( ),
inserter( target, target.end( ) ) );

10.11BitSets
Like Java, C++ has bitwise operators that can manipulate a set of bits stored in a primitive type
and a bitset class template.
The class template, which is declared in the standard header bitset, is instantiated with
the number of bits to be stored (this must be a compile-time constant), and indexing starts at 0.
Bits can be accessed with test, or alternatively, with the array indexing operator[]. set
and unset can be used to turn on or off a particular bit; with no parameters these methods

c++book.mif Page 204 Tuesday, April 29, 2003 2:13 PM

204

Chapter 10 Collections: The Standard Template Library

affect all bits. Alternatively, operator[] can be used on the left-hand side of an assignment.
Also overloaded are the standard bitwise operators, that allow bitwise operations on bitset
types. The bitsets involved in those operations must have identical sizes.
The bitset has a look and feel that is similar to both vector and map; but, it can be
expected to be more efficient that vector, set, or map. However, a set or map could be
space-efficient for cases where there are many bits, but only a few are ever set to be on.

10.12Key Points
Standard STL containers include the sequence containers: vector, list, and deque,
and also set, multiset, map, multimap, and priority_queue. multisets
allow duplicates and multimaps allow duplicate keys.
maps store keys and values. operator[] returns the value associated with a key, and if
the key is not present, it is inserted with a default value, that is then returned.
Containers can be accessed by iterators, which are more powerful than their Java counterparts.
Significant compile-time type checking is performed by the STL.
Little runtime error checking is performed by the STL.
There are several general types of iterators: forward iterators, bidirectional iterators, random access iterators, and stream iterators are the most common. Additionally, there are
const_iterators, reverse_iterators, and const_reverse_iterators.
Iterators use operator overloading extensively. The common operators are ++, *, =, ==,
and !=. Bidirectional iterators allow --. Random access iterators allow - and +, = and
+=.
Stream iterators allow repeated iteration over an input or output stream.
Each container has a begin and end member function that yields iterators that represent
the beginning of the container and the endmarker of the container. There are both accessor
and mutator versions of begin and end.
pair is a class template that stores the first item and second item as public data. The
pair is used in the map, which is a set of pairs, and also in the return type of some set
member functions.
If a standard container is storing a heterogeneous collection, it should store pointers to the
objects.
Six function objects are defined as class templates in the standard header functional.
These are less, greater, equal_to, not_equal_to, great_equal, and
less_equal.
The unary adapters allow the conversion of the standard binary predicates to unary predicates, by supplying one of the parameters to the binary predicate.
The inserter adapters allow copying into empty containers by converting the assignment
operator of an iterator into an insertion operation on the container.

c++book.mif Page 205 Tuesday, April 29, 2003 2:13 PM

Exercises

205

The standard library includes over 60 function templates for sorting, searching, copying,
and many other generic algorithms.
C++ has a bitset class template. To use it, the number of bits must be known at compile
time. If this is not possible, alternatives such as vector<bool>, set<int> (containing only the true bits), or map<int,bool> can be used, but might not be as fast.

10.13Exercises
1. How does the STL differ from the Java Collections API?
2. Describe the functionality of iterators in C++.
3. What are the different types of iterators?
4. What does end return?
5. What kind of error checks are performed by STL routines?
6. What is a const_iterator and how is it used?
7. What is a reverse iterator and how is it used?
8. What is a stream iterator and how is it used?
9. What is the difference between a set and a multiset?
10. Why is there no operator[] accessor for map?
11. What is a unary binder adapter?
12. What is an inserter adapter?
13. Describe the general categories of STL algorithms and give an example or an algorithm in
each category.
14. In Exercise 6.19, make two modifications. First, in the Employee class, add
operator< that orders employees by name. Then change the implementation of Roster
to use a multiset of Employee * (ordered by name). Part of the multiset template instantiation includes an appropriate function object. The multiset print routine
should output employees in sorted order (by name).
15. Implement a spelling checker. Prompt the user for the name of a file that stores a dictionary of words. Then prompt for the name of a file that you want to spell-check. Any word
that is not in the dictionary is considered to be misspelled. Output, in sorted order, each
misspelled word and the line number(s) on which it occurs. If a word is misspelled more
than once, it is listed once, but with several line numbers. Of course you should verify that
files open correctly. Use the following rule to determine what a word is: The input is considered to be a sequence of tokens separated by whitespace. Any token that ends with a
single period, question mark, comma, semicolon, or colon should have the punctuation
removed. After doing this, any token that contains letters only is considered a word. Convert this word to lower case.
16. Implement the sort template that takes a pair of iterators and a comparator, using any
simple sorting algorithm. Then implement the sort template that takes a pair of iterators.
In order to do this, and reuse the three-parameter sort, you will need to define a phantom

c++book.mif Page 206 Tuesday, April 29, 2003 2:13 PM

206

Chapter 10 Collections: The Standard Template Library

four-parameter sort template that takes an object of the type to be sorted as the fourth
parameter. The two-parameter sort will invoke the four-parameter sort, which in turn
will invoke a three-parameter sort, with less<Object>() as the third parameter.

c++book.mif Page 207 Tuesday, April 29, 2003 2:13 PM

H A P T E R

1 1

Primitive Arrays and

Strings

N a perfect world, all array and strings manipulations in

C++ would be done with the vector class template and the string class that is part of the
standard library. But life is not perfect. Many programs written prior to the standardization of
vector and string make use of primitive arrays and strings, and some library routines interface with primitive arrays and strings instead of the library classes. Parameters to main, for
instance, are a primitive array of primitive strings.
In this chapter, we discuss primitive arrays and strings in C++. We will see the relationship
between the primitive array and pointer variables, and how this relationship influences the
design of the STL. We will also discuss command-line arguments, and briefly mention multidimensional arrays.

11.1 Primitive Arrays

An array is declared by giving it a name and by telling the compiler what type the elements are.
If we are defining an array (creating a new array object), a size must also be provided. The size
can be omitted if an initialization is present; the compiler will then count the number of initializers and take that as the array size. The declaration
int a[ 3 ];

// Three int objects: a[0], a[1], and a[2]

has the compiler allocate space to store three integers, namely a[0], a[1], and a[2]. No
index range checking is performed in C++, so an access out of the array index bounds is not
caught by the compiler or runtime system; instead, undefined and occasionally mysterious
behavior occurs. Furthermore, if the array is passed as an actual argument to a function, the
function has no idea how large the array is unless an additional parameter is passed. Finally,
arrays cannot be copied by the = operator. In this section we will stick with the core language

207

c++book.mif Page 208 Tuesday, April 29, 2003 2:13 PM

208

Chapter 11 Primitive Arrays and Strings

features of arrays and pointers and discuss why these restrictions come into play.
11.1.1 The C++ Implementation: An Array Name Is a Pointer
When a new array is allocated, the compiler multiplies the size in bytes of the type in the declaration by the array size (the integer constant between the []) to decide how much memory it
needs to set aside. This is essentially the only use for the size component. In fact, after the array
is allocated, with minor exceptions, the size is irrelevant because the name of the array represents a pointer to the beginning of allocated memory for that array. This is illustrated in Figure
11-1.
Suppose we have the declarations
int a[ 3 ];
int i;

The compiler allocates memory as follows: First, three integers are set aside for the array object.
These are referenced by a[0], a[1], and a[2]. The objects in the array are guaranteed to be
stored in one contiguous block of memory. Thus if a[0] is stored at memory location 1000 and
integers require four bytes, it is guaranteed that a[1] is located at memory location 1004 and
a[2] at memory location 1008. (Java does not make this guarantee, in order to allow the garbage collector the option of storing parts of the array separately if needed to minimize memory
fragmentation). Finally, the compiler allocates storage for the object i. One possibility is shown
in Figure 11-1, where i is allocated the next available memory slot.
For any i, we can deduce that a[i] would be stored at memory location 1000 + 4i. The
value stored in a is exactly equal to &a[0]; this equivalence is always guaranteed and tells us
that a is actually a pointer. Note also that &a is not the same as &a[0]. Now we can see that to
access the item a[i], the compiler needs only to fetch the value of a and add to it 4i.

&a[0] (1000)

a[0]

&a[1] (1004)

a[1]

&a[2] (1008)

a[2]

&i

(1012)

i
...

local constant

Figure 11-1

a=1000

Memory model for arrays (assumes 4 byte int); declaration is int a[3]; int i;

c++book.mif Page 209 Tuesday, April 29, 2003 2:13 PM

Primitive Arrays

209

Now that we have seen how arrays are manipulated in C++, we can see why some of the
limitations discussed earlier occur, and we can also see how arrays are passed as function parameters. First we have the problem of checking that the index is in range. Performing the bounds
check would require that we store the array size in an additional parameter. Certainly this is feasible, but it does incur both time and space overhead. In a common application of arrays (short
strings), the overhead could be significant. As we have mentioned in Section 2.2.1 and illustrated in Section 8.1, the lack of range checking can cause serious problems such as off-by-one
errors in array indexing that can lead to bugs that are very difficult to spot. (If index range
checking is crucial, use the vectors at member function).
The second limitation of the basic array (is solved by vector) is array copying. Suppose
that a and b are arrays of the same type. In many languages, if the arrays are also the same size,
the statement a=b would perform an element-by-element copy of the array b into the array a. In
C++ this statement is illegal because a and b represent constant pointers to the start of their
respective arrays, specifically &a[0] and &b[0]. Then a=b is an attempt to change where a
points, rather than copying the contents of array b into array a. What makes the statement illegal, rather than legal but wrong, is that a cannot be reassigned to point somewhere else because
it is essentially a constant object. The only way to copy two arrays is to do it element by element; there is no shorthand. A similar argument shows that the expression a==b does not evaluate to true if and only if each element of a matches the corresponding element of b. Instead,
this expression is legal. It evaluates to true if and only if a and b represent the same memory
location (that is, they refer to the same array).
Finally, an array can be used as a parameter to a function, and the rules follow logically
from our understanding that an array name is little more than a pointer. Suppose we have a function functionCall that accepts one array of int as its parameter. The caller/callee views are
functionCall( actualArray );
// Function Call
functionCall( int formalArray[ ] ) // Function Declaration

Note that in the function declaration, the brackets serve only as a type declaration, in the same
way that int does. Note that the [] must follow the formal parameter, unlike Java where it can
either follow or precede the formal parameter. In the function call only the name of the array is
passed; there are no brackets. In accordance with the call-by-value conventions of C++, the
value of actualArray is copied into formalArray. Because actualArray represents
the memory location where the entire array actualArray is stored, formalArray[i]
accesses actualArray[i]. This means that the variables represented by the indexed array
are modifiable. Thus an array, when considered as an aggregate, is passed by reference. Furthermore, any size component in the formalArray declaration is ignored, and the size of the
actual array is unknown. If the size is needed, it must be passed as an additional parameter.
Note that passing the aggregate by reference means that functionCall can change elements in the array. We can use the const directive to attempt to disallow this (but this technique is not foolproof because of the ability to cast away const-ness):
functionCall( const int formalArray[ ] );

c++book.mif Page 210 Tuesday, April 29, 2003 2:13 PM

210

Chapter 11 Primitive Arrays and Strings

11.1.2 Dynamic Allocation of Arrays

Suppose we want to read a sequence of numbers and store them in an array for processing. The
fundamental property of an array requires us to declare a size so that the compiler can allocate
the correct amount of memory, and we must make this declaration prior to the first access of the
array. If we have no idea how many items to expect, then it is difficult to make a reasonable
choice for the array size. In this section we show how to allocate our arrays dynamically and
expand them if our initial estimate is too small. This technique, dynamic array allocation,
allows us to allocate arbitrary-sized arrays and make them larger or smaller as the program runs,
and is what goes on behind the scenes in vector and in the Java ArrayList.
The allocation method for arrays that we have seen thus far is
int a1[ SIZE ];

// SIZE is a compile-time constant

We also know we can use

int *a2;

like an array, except that no memory is allocated by the compiler for the array. We know that the
new operator allows us to obtain memory from the memory heap as the program runs. An alternative form is the array new operator, new[], which causes the creation of an array of objects
from the memory heap, invoking a zero-parameter constructor for each array item (unfortunately, for ints, this does not always guarantee any initial value). Thus, the expression
new int [ SIZE ]

allocates the array. The expression evaluates to the address where the start of that memory
resides. It may be assigned only to an int * object, as in
int *a2 = new int [ SIZE ];

As a result, a2 is virtually indistinguishable from a1. The new operator is type-safe, meaning
that
int *a2 = new char[ SIZE ];

would be detected at compile time as a type mismatch error.

So what is the difference, if any, between the two forms of array memory allocation? A
minor difference is that a1 cannot appear on the left side of an assignment operator because the
array name is a constant, while a2 can. This is a relatively minor difference, since if we declared
int * const a2 = new int [ SIZE ];

there is no difference with respect to what can apepar on the left-hand side. A more important
difference is that SIZE does not have to be a compile-time constant when we use new.
But the most important difference is that the memory for a1 is taken from a different
source than a2, which is allocated from the memory heap and eventually must be released to
avoid a memory leak. The source of the memory, however, is otherwise transparent to the user.
Thus, when a1 is a local variable and the function in which it is declared returns (that is,
when a1 exits scope), the memory associated with the array is reclaimed automatically by the
system. a1 exits scope when the block in which it is declared is exited. For example, in Figure

c++book.mif Page 211 Tuesday, April 29, 2003 2:13 PM

Primitive Arrays

211

a1

a2

Figure 11-3

Memory reclamation in Figure 11-2

11-2, a1 is a local variable in a function f. When f returns, the entire contents of the a1 object,
including the memory associated with the array, is freed. In contrast, when a2 exits scope only
the memory associated with the pointer is freed; the memory allocated by new is now unreferenced, and we have a memory leak. The memory is claimed as used, but unreferenced, and will
not be used to satisfy future new requests, and if the array contains class type objects, those
objects destructors will not have been invoked. The situation is shown graphically in Figure 113.
To recycle the memory, we must use the delete[] operator. The syntax is
delete [ ] a2;

The [] is absolutely necessary here to ensure that all of the objects in the allocated array
have their destructors called prior to reclaiming of the memory for array a2. Without the [] it is
possible that only a2[0]s destructor is called, and the remaing items in the array do not have
their destructors called, nor will a2s memory be reclaimed, which is hardly what we intend.
1
2
3
4
5
6
7
8
9
10
11
12
13
14

void f( int i )
{
int a1[ 10 ];
int *a2 = new int [ 10 ];
...
g( a1 );
g( a2 );
//
//
//
//

On return, all memory associated with a1 is freed

On return, only the pointer a2 is freed;
10 ints have leaked
delete [ ] a2;
// This would fix the leak

Figure 11-2

Two ways to allocate arrays; one leaks memory

c++book.mif Page 212 Tuesday, April 29, 2003 2:13 PM

212

Chapter 11 Primitive Arrays and Strings

With new and delete we have to manage the memory ourselves rather than allow the compiler to do it for us. Why then, would we be interested in this? The answer is that by managing
memory ourselves, we can build expanding arrays. Suppose, for example, that in Figure 11-2 we
decide, after the declarations but before the calls to g at lines 7 and 8, that we really wanted 12
ints instead of 10. In the case of a1 we are stuck, and the call at line 7 cannot work. However,
with a2 we have an alternative, as illustrated by the following maneuver:
int *original = a2;
// 1. Save pointer to the original
a2 = new int [ 12 ];
// 2. Have a2 point at more memory
for( int i = 0; i < 10; i++ ) // 3. Copy the old data over
a2[ i ] = original[ i ];
delete [ ] original;
// 4. Recycle the original array

Figure 11-4 shows the changes that result. A moments thought will convince you that this is an
expensive operation, because we copy all of the elements from original to a1. If, for
instance, this array expansion is in response to reading input, it would be inefficient to re-expand

(a)

(b)

a2

a2

Original

(c)

a2

Original

(d)

a2

Original

Figure 11-4

Array expansion: (a) starting point: a2 points at 10 integers; (b) after step 1:
original points at the 10 integers; (c) after steps 2 and 3: a2 points at 12 integers, the first 10 of which are copied from original; (d) after step 4: the 10 integers are freed

c++book.mif Page 213 Tuesday, April 29, 2003 2:13 PM

Primitive Arrays

213

every time we read a few elements. Thus when array expansion is implemented, we always
make it some multiplicative constant times as large. For instance, we might expand to make it
twice as large. In this way, when we expand the array from N items to 2N items, the cost of the N
copies can be amortized over the next N items that can be inserted into the array without an
expansion.

c++book.mif Page 214 Tuesday, April 29, 2003 2:13 PM

214

Chapter 11 Primitive Arrays and Strings

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41

#include <iostream>
#include <cstdlib>
using namespace std;
// Read an unlimited number of ints with no attempts at error
// recovery; return a pointer to the data, and set ItemsRead
int * getInts( int & itemsRead )
{
int arraySize = 0;
int inputVal;
int *array = NULL;
// Initialize to NULL pointer
itemsRead = 0;
cout << "Enter any number of integers: ";
while( cin >> inputVal )
{
if( itemsRead == arraySize )
{
// Array doubling code
int *original = array;
array = new int[ arraySize * 2 + 1 ];
for( int i = 0; i < arraySize; i++ )
array[ i ] = original[ i ];
delete [ ] original; // Safe if Original is NULL
arraySize = arraySize * 2 + 1;
}
array[ itemsRead++ ] = inputVal;
}
return array;
}
int main( )
{
int *array;
int numItems;
array = getInts( numItems );
for( int i = 0; i < numItems; i++ )
cout << array[ i ] << endl;
return 0;
}

Figure 11-5

Code to read an unlimited number of ints and write them out

To make things more concrete, Figure 11-5 shows a program that reads an unlimited number of integers from the standard input and stores the result in a dynamically expanding array.
The function declaration for getInts tells us that it returns the address where the array will

c++book.mif Page 215 Tuesday, April 29, 2003 2:13 PM

Primitive Strings

215

reside, and it sets a reference parameter itemsRead to indicate how many items were actually
read.
At the start of getInts, itemsRead is set to 0, as is the initial arraySize. We
repeatedly read new items at line 15. If the array is full, as indicated by a successful test at line
17, then the array is expanded. Lines 19 to 23 perform the array doubling. At line 19 we save a
pointer to the currently allocated block of memory. We have to remember that the first time
through the loop, the pointer will be NULL. At line 20 we allocate a new block of memory,
roughly twice the size of the old. We add one so that the initial doubling converts a zero-sized
array to an array of size one. At line 24 we set the new array size. At line 26, the actual input
item is assigned to the array, and the number of items read is incremented. When the input fails
(for whatever reason), we merely return the pointer to the dynamically allocated memory. Note
carefully that
We do not delete[] the array.
The memory returned is somewhat larger than is actually needed. This can be easily fixed.
The main routine calls getInts, assigning the return value to a pointer.
As we can see, this is lots of work. Thats why modern C++ programmers use vector.

11.2 Primitive Strings

An important use of pointers and arrays is the C++ implementation of strings. The C++ base language provides some minimal support for strings, based entirely on the conventions of C and the
C library. The result is too minimal to be useful in a modern language, so as they do with arrays,
Sane C++ programmers rely on the string class type rather than the predefined language features. Nonetheless, it is still important to know how strings are implemented in the basic C
library because they often occur in older code, and are occassionaly needed even in newer code.
In C++ as well as C, a primitive string is an array of characters. As a result, when passed
to a function, the string has type char * or const char *. At first glance we might assume
that the string "Nina" is an array of four characters: 'N', 'i', 'n', and 'a'. The problem
with this assumption is that if we pass this array to any routine, that routine would not know how
many characters are in the array because, as we have seen, a function that receives an array only
receives a pointer and thus has no idea how large the actual array is. One solution to this problem is to use a slightly larger array with an end marker.
For instance, we can declare an array of five elements, placing a blank in the last spot to
signal that only the first four positions represent significant characters. If all routines are written
to reflect this convention, we have a solution to our problem that requires little alteration of the
language. Because we might actually want to use a blank in the string (for example, to store a
street address), we need to pick an end marker that is not likely to appear elsewhere in the string.
In C++ this special character is the null terminator '\0'. The escape sequence indicates that
the null terminator is always represented internally as zero, which, as we shall see, leads to some

c++book.mif Page 216 Tuesday, April 29, 2003 2:13 PM

216

Chapter 11 Primitive Arrays and Strings

shorthands when the controlling expression is written in an if statement or a loop. (A common

error is to forget the \; this leaves '0', which is the character representation for the digit 0).
Therefore, if an array of six characters has 'N', 'i', 'n', 'a', and '\0', it represents the
string "Nina", no matter what is in the sixth character.
So far, what has C++ provided us in the way of string support? The answer is absolutely
nothing! Furthermore, there are some things it does not directly provide in the language. Suppose we declare two strings str1 and str2:
char str1[ 10 ]; // Max length is nine
char str2[ 10 ]; // Max length is nine

Then the following statements cannot be correct:

str1 = str2;
bool cond = ( str1 == str2 );

// Wrong!
// Wrong!

This follows directly from the facts that str1 and str2 are arrays and array assignment and
comparison are not supported directly by the language. Almost all of the support, in fact, is provided by the C++ library, which specifies routines that work for null-terminated strings. The
prototypes for these routines are given in the standard header cstring. Some routines of interest are shown in Figure 11-6.
strlen(str) gives the length of the string represented by str (not including the null
terminator); the length of "Nina" is four. In this and all routines, if a NULL pointer is passed,
you can expect a program crash. Notice that this is different from passing a pointer to a memory
cell that contains the '\0' character, which represents the empty string of length 0.
strcpy(lhs,rhs) performs the assignment of strings; characters in the array given by rhs
are copied into the array given by lhs until the null terminator is copied. If the string represented by lhs is not large enough to store the copy, then somebody elses memory gets overwritten.
strcpy( lhs, rhs )

is meant to mimic the statement

lhs = rhs;

The return type char * allows strcpy calls to be chained in the same way as assignments: strcpy(a,strcpy(b,c)) is much like a=b=c. strcat(lhs,rhs) appends a
copy of the string represented by rhs to the end of lhs. As with strcpy, it is the programmers responsibility to assure that lhs is pointing at sufficient memory to store the result.
strcmp compares two strings returning a negative number, zero, or a positive number, depending on whether the first string is lexicographically less than, equal to, or greater than the second.
1
2
3
4

size_t
char *
char *
int

Figure 11-6

strlen( const char *str );

strcpy(
char *lhs, const char *rhs );
strcat(
char *lhs, const char *rhs );
strcmp( const char *lhs, const char *rhs );
Some of the string routines in <cstring>

c++book.mif Page 217 Tuesday, April 29, 2003 2:13 PM

Primitive Strings

217

C++, as described so far, provides library routines for strings but no language support. In
fact, the only language support is provided by a string constant. A string constant provides a
shorthand mechanism for specifying a sequence of characters. It automatically includes the null
terminator as an invisible last character. Any character (specified with an escape sequence if
necessary) may appear in the string constant. Thus "Nina" represents a five-character array.
Additionally, a string constant can be used as an initializer for a character array. Thus:
char name1[
] = "Nina"; // name1 is an array of five char
char name2[ 9 ] = "Nina"; // name2 is an array of nine char
char name3[ 4 ] = "Nina"; // name3 is an array of four char

In the first case the size of the array allocated for name1 is determined implicitly, while in the
second case we have over-allocated (which is necessary if we intend later to copy a longer string
into name2). The third case is wrong because we have not allocated enough memory for the
null terminator. Initialization by a string constant is a special exemption; we cannot say
char name4[ 8 ] = name1;

// ILLEGAL!

A string constant can be used in any place that both a string and a constant string can. For
instance, it may be used as the second parameter to strcpy but not as the first parameter. This
is because the declaration for strcpy does not disallow the possibility that the first parameter
might be altered (indeed, we know that it will). Because a string constant can be stored in readonly memory, allowing it to be used as a target of strcpy could result in a hardware error. Note
carefully that we can always send a nonconstant string to a parameter that expects a constant
string. Thus we have
strcpy( name2, "Mark" );
strcpy( "Mark", name2 );
strcpy( name2, name1 );

// LEGAL
// ILLEGAL!
// LEGAL

The declarations for the string routines indicate that the parameters are pointers. This follows from the fact that the name of an array is a pointer. The second parameter to strcpy is a
constant string, meaning that any string can be passed and it is guaranteed to be unchanged. The
first parameter is a non-contant string, and might be changed. Consequently, a constant string
cannot be passed; this includes string constants.
Beginners tend to take the equivalence of arrays and pointers one step too far. Recall that
the fundamental difference between an array and a pointer is that an array definition allocates
enough memory to store the array, while a pointer points to memory that is allocated elsewhere.
Because strings are arrays of characters, this distinction applies to strings. A common error is
declaring a pointer when an array is needed. As examples, consider the following declarations:
char name[ ] = "Nina";
char *name1 = "Nina";
char *name2;

The first declaration allocates five bytes for name, initializing it to a copy of the string constant
"Nina" (including the null terminator). The second declaration states merely that name1
points at the zeroth character of the string constant "Nina". In fact, the declaration is wrong

c++book.mif Page 218 Tuesday, April 29, 2003 2:13 PM

218

Chapter 11 Primitive Arrays and Strings

because we are mixing pointer types: the right side is a const char *, while the left side is
merely a char *. Some compilers will complain. The reason for this is that a subsequent
name1[ 3 ] = 'e';

is an attempt to alter the string constant. Since a string constant is supposed to be constant, this
action should not be allowed. The easiest way for the compiler to do this is to follow the convention that if a is a constant array, then a[i] is a constant also and cannot be assigned to. If the
statement
char *name1

= "Nina";

were allowed, this would be hard to enforce. By enforcing const-ness at each assignment, the
problem becomes manageable. It is legal to use
const char *name1 = "Nina";

but that is hardly the same as declaring an array to store a copy of the actual string; furthermore,
name1[3]='e' is easily seen by the compiler to be illegal in this case. A common example
where this would be used is
const char *message = "Welcome to FIU!";

Another common consequence of declaring a pointer instead of an array object is the following statement (in which we assume that name2 is declared as above):
strcpy( name2, name );

Here the programmer expects to copy name into name2 but is fooled because the declaration
for strcpy indicates that two pointers are to be passed. The call fails because name2 is just a
pointer rather than a pointer to sufficient memory to hold a copy of name. If name2 is a NULL
pointer, points at a string constant stored in read-only memory, or points at an illegal random
location, strcpy is certain to attempt to dereference it, generating an error. If name2 points at
a modifiable array (for instance, name2=name is executed), there is no problem.
All these considerations tell us that using the C++ string is a better option that primitive strings in most cases to safely hide all new uses of primitive strings inside of a string.
A string can be constructed from a const char *, and a const char * can be
extracted from a string via the member function c_str. So when interacting with older code
that expects char * and produces char *, one strategy is to create a string as soon as possible, do the string manuipulations safely with the string class, extract a const char * by
using c_str and pass that to the older code. The return value from the older code can be immediately converted into a string. For instance, suppose there is a routine
const char *getenv( const char *prop );

that expects a primitive string and returns a primitive string. We can pass it a string prop,
and assign the result to a string val, as
string val = getenv( prop.c_str( ) );

making use of the automatic implicit conversion from const char * to string.

c++book.mif Page 219 Tuesday, April 29, 2003 2:13 PM

Pointer Hopping

219

ptr

a[0] a[1] a[2] a[3] a[4] a[5] a[6] a[7] a[8] a[9]
Figure 11-7

Pointer arithmetic: x=&a[3]; y=x+4;

11.3 Pointer Hopping

Many programmers spend lots of time attempting to hand-optimize their code. One common
myth is that pointers can be used to access arrays more quickly than the usual indexing method.
While this is occasionally true and sometimes leads to better or simpler code, in this section we
will show that this is not universally true (and in fact is frequently false). This is interesting for
two reasons. First, it illustrates that with modern compilers, the speedups obtained by low-level
optimizations often do not justify the effort that is put into them. Instead, we should concentrate
our optimizations on larger algorithmic issues. Second, although the tricks that we used reflect
an old way of programming, it is the basis for many constructs in the STL. So when we wonder
why things in the STL are how they are, this section provides some of the answers and also
shows that STL generic algorithms can be applied to primitive arrays.
Our discussion proceeds as follows: First we look at how arithmetic applies to pointers.
We have two issues to consider. First, in an expression such as *x+10 or *x++, is the operator
(+ or ++) being applied to x or *x? The answer to this question is determined by normal precedence rules. In the first case 10 is added to *x, while in the second case the increment is applied
to x (after the value of *x is used). The second issue, then, is to decide what it means to increment or apply various operations to a pointer. Then we will see an application that shows how
pointer math is typically used and whether or not it is a good idea.
11.3.1 Implications of the Precedence of *, &, and []
The dereferencing operator * and the address-of operator & are grouped together in a class of
prefix operators. These operators include the unary minus (-), the not operator (!), the bitwise
complement operator (~), and the prefix increment and decrement operators (++ and --), as
well as new, delete, and sizeof. The prefix unary operators have higher precedence than
almost all other operators, except for the scope operators and the postfix operators, such as the
postfix increment and decrement operators (++ and --), the function call operator (), and the
array access operator []. Consequently, the only arithmetic operators that have higher precedence than a dereferencing operator are the postfix increment and decrement operators. In all of

c++book.mif Page 220 Tuesday, April 29, 2003 2:13 PM

220

Chapter 11 Primitive Arrays and Strings

the following expressions, the operator is applied to the dereferenced value:

*x + 5
*x == 0
*x / *y

// Adds 5 to *x
// True if *x is 0
// Divide *x by *y

Notice carefully that because of precedence rules, *x++ is interpreted as *(x++), not
(*x)++. The precedence of the array indexing operator tells us that if x is a pointer, all of the
following operators are applied to the indexed value of x:
5 + x[ 0 ]
0 == x[ 0 ]
++x[ 0 ]
x++[ 0 ]
x == &x[ 0 ]

//
//
//
//
//

Add x[0] and 5

True if x[0] is 0
Increment x[0]. Same as ++*x (why?)
Same as *x++ (why?)
Always true

In the last example we reiterated that x always stores the memory location of x[0]. The precedence rules are convenient here because we do not need to parenthesize, as in &(x[0]).
11.3.2 What Pointer Arithmetic Means
Suppose that x and y are pointer variables. Now that we have decided on precedence rules, we
need to know what the interpretation is for arithmetic performed on pointers. For instance, what
does it mean to multiply x by 2? The answer in most cases is that arithmetic on pointers would
be totally meaningless and is therefore illegal. Most other languages allow only comparison,
assignment, and dereferencing of pointers. C++ is somewhat more lenient.
Looking at the various operators, we see that none of the multiplicative operators makes
sense. Therefore, a pointer may not be involved in a multiplication. Note carefully that the dereferenced value can, of course, be multiplied, and that what we are restricting is computations
involving addresses.
Equality and logical operators all make sense for pointers, so they are allowed and have
obvious meanings. Two pointers are equal if they both point to NULL or they both point to the
same address. Assignment by = is allowed, as we have seen, but *=, /=, and %= are disallowed.
Therefore, the questionable operators are the additive operators (including +=, -=, ++, --) and
the relational operators (<, <=, >=, >). To make sense, all these operators need to be viewed in
the context of an array.
Figure 11-7 shows an array a, a pointer ptr, and the assignment ptr=a. The figure reinforces the idea that the value stored in a is just the memory location where the zeroth element of
the array is stored and that elements of an array are guaranteed to be stored in consecutive and
increasing memory locations. If the array a is an array of characters, a[1] is stored in memory
location a+1 because characters use one byte. Thus the expression ++ptr would increase ptr
by one, which would equal the memory location of a[1].
We see from this example that adding an integer to a pointer variable can make sense in an
array of characters. If a was an array of four-byte integers, adding 1 to ptr would make only
partial sense under our current interpretation. This is because ptr would not really be pointing

c++book.mif Page 221 Tuesday, April 29, 2003 2:13 PM

Pointer Hopping

1
2
3
4
5
6
7
8
9
10
11
12
13
14

221

// initialize implemented with usual indexing mechanism

void initialize1( int arr[ ], int n, int val )
{
for( int i = 0; i < n; i++ )
arr[ i ] = val;
}
// initialize implemented with pointer hopping
void initialize2( int arr[ ], int n, int val )
{
int *endMarker = arr + n;
for( int *p = arr; p != endMarker; p++ )
*p = val;
}

Figure 11-8

Array initialization coded two ways: first by using indexing, second by using
pointer hopping

at an integer but somewhere in the middle and would be misaligned, generally leading to a hardware fault. Since that interpretation would give erroneous results, C++ uses the following interpretation: ++ptr adds the size of the pointed at object to the address stored in ptr.
This interpretation carries over to other pointer operations. The expression x=&a[3]
makes x point at a[3]. Parentheses are not needed, as mentioned earlier. The expression
y=x+4 makes y point at a[7]. We could thus use a pointer to traverse an array instead of using
the usual index iteration method. We will discuss this in Sections 11.3.3 and 11.3.4.
Although it makes sense to add or subtract an integer type from a pointer type, it does not
make sense to add two pointers. It does, however, make sense to subtract two pointers: y-x
evaluates to 4 in the example above (since subtraction is the inverse of addition). Thus pointers
can be subtracted but not added.
Given two pointers x and y, x<y is true if the object x is pointed at is at a lower address
than the object y is pointing at. Assuming that neither is pointing at NULL, this expression is
almost always meaningless unless both are pointing at elements in the same array. In that case
x<y is true if x is pointing at a lower-indexed element than y because, as we have seen, the elements of an array are guaranteed to be stored in increasing and contiguous parts of memory. This
is the only legitimate use of the relational operator on pointers, and all other uses should be
avoided. To summarize, we have the following pointer operations:
Pointers may be assigned, compared for equality (and inequality), and dereferenced in
C++, as well as almost all other languages. The operators are =, ==, !=, and *.
We can apply the prefix or postfix increment operators to a pointer, can add an integer, and
can subtract either an integer or pointer. The operators are ++, --, +, -, +=, and -=.

c++book.mif Page 222 Tuesday, April 29, 2003 2:13 PM

222

Chapter 11 Primitive Arrays and Strings

We can apply relational operators to pointers, but the result makes sense only if the pointers point to parts of the same array, or one pointer points to NULL. The operators are <,
<=, >, and >=.
We can test against NULL by applying the ! operator (because the NULL pointer is 0).
We can subscript and delete pointers via [], delete, and delete[].
We can apply trivial operators, such as & and sizeof, to find out information about the
pointer (not the object it is pointing at).
We can apply some other operators, such as ->.
11.3.3 A Pointer-Hopping Example
Figure 11-8 illustrates how pointers can be used to traverse arrays. We have written two versions
of initialize.
The first version, initialize1, uses the normal indexing mechanism to step through
the array and is straightforward. The pointer-hopping version is initialize2. At line 11, we
declare a pointer endMarker that is initialized to point one past the array portion we are trying
to initialize; in STL jargon, it is the endmarker. A second pointer p, repeatedly hops through the
array initializing each entry until it hits the endmarker.
A crucial observation, with respect to the discussion of the STL in Chapter 10, is that in
initialize2, p is used to iterate over the collection of integers, and this is the basis for the
selection of the particular operators such as * and ++ in the STL iterators. In particular, as we
discuss in Section 11.4, this makes the generic algorithms in the standard library usable by primitive arrays.
11.3.4 Is Pointer Hopping Worthwhile?
Why might a pointer implementation be faster than an array implementation? Let us consider a
string of length 3. In the array implementation we access the array via s[0], s[1], s[2], and
s[3]. s[i] is accessed by adding one to the previous value i-1 and then adding s and i to
get the required memory location. In our pointer implementation, s[i] is accessed by adding
one to sp, and we never keep a counter i. Thus we save an addition for each character, paying
only an extra two subtractions during the return statement.
The next question is whether or not the trickier code is worth the time savings. The answer
is that in most programs a few subroutines dominate the total running time. Historically, the use
of trickier code for speed has been justified only in those routines that actually account for a significant portion of the programs running time, or in routines used in enough different programs
to make the optimization worthwhile. Thus in the old days C programs that used pointer hopping
judiciously had a large speed advantage over programs written in other high-level languages.
However, good modern compilers can, in many cases, perform this optimization. Thus the use of
pointers to traverse arrays will help some compilers, will be neutral for others, or may even generate slower code than the typical index addressing mechanism.

c++book.mif Page 223 Tuesday, April 29, 2003 2:13 PM

Primitive Arrays and the STL

223

The moral of the story is that, in many cases, it is best to leave minute coding details to the
compiler and concentrate on the larger algorithmic issues and on writing the clearest code possible. Many systems have a profiler tool that will allow you to decide where a program is spending most of its running time. This will tell you where to apply algorithmic improvements, so it is
important to learn how to use the optimizer and profiler on your system.

11.4 Primitive Arrays and the STL

A pointer variable that points at an array element satisfies all of the properties of a random
access iterator. Thus by passing a pair of pointers, primitive arrays can be the subject of an STL
generic algorithm. For instance, the following two lines:
int arr[ ] = { 3, 4, 1, 2, 5 };
sort( arr, arr + 4 );

sorts the first four (not all five) elements of arr. This is because arr represents an iterator for
the beginning of array arr, while arr+4 is an iterator representing index 4 (and since the second parameter in all STL range operations an endmarker of the range), signalling the first item
not to be included in the sort. Indeed, the iterator that is part of vector is often defined as simply a pointer variable, which helps explain why vector iterators do not have any bounds
checking built in to them.
Functions such as copy work, but since primitive arrays do not have push_back operations, primitive arrays that are the target of a copy need to already have enough space to store the
result.

11.5 Command-Line Arguments

In Java, command-line arguments are available in an array of strings passed as the mandatory
parameter to main. In C++, we can access command-line arguments if we provide additional
parameters to main. However, there are two relatively minor differences.
First, the command that invokes the executable is considered a command-line argument.
Thus, the 0th command-line argument is often skipped when processing command-line arguments. Second, the command-line arguments are passed as a primitive array of primitive strings.
Because primitive arrays do not know how large they are, an additional parameter, the number
of command-line arguments (including the executable name) is passed.
The number of command line arguments is the first paramter to main, and is commonly
named argc. The second parameter, which is an array of strings, is historically named argv.
Note the type: array of char *, and this gives a typical signature of:
int main( int argc, char *argv[ ] );

As an example, the echo command, shown in Figure 11-9, prints out its command-line
arguments. Alternatively, we sometimes see
int main( int argc, char **argv );

making use of the fact that the array variable is simply a pointer. However,

c++book.mif Page 224 Tuesday, April 29, 2003 2:13 PM

224

Chapter 11 Primitive Arrays and Strings

1
2
3
4
5
6
7
8
9
10

#include <iostream>
using namespace std;
int main( int argc, char *argv[ ], char *envp[ ] )
{
for( int i = 0; envp[ i ] != NULL; i++ )
cout << envp[ i ] << endl;
return 0;
}

Figure 11-10 List all environment variables

int main( int argc, char argv[ ][ ] );

is not an acceptable alternative, as it signals, illegally, a two-dimensional array (see

Section 11.6).
In addition to command-line arguments, some systems invoke C++ with environment variables. An environment variable can be used to indicate, for instance, the location of a directory
suitable for the placement of temporary files. If a C++ program is being used as a server-side
script (e.g. a script to hande submitted forms), information about the client and the request will
be stored in a host of environment variables.
Environment variables are stored as a list of entries in the form name=value. The variable name is typically upper case. Some systems provide this list as a third parameter to main.
Like argv, the environment list is represented as an array of char *. The end of the environment list is marked with a NULL pointer, so an additional parameter is not needed. (The use of
argc and argv is retained for backward compatability, since the same scenario could have
been used there.) Typically the environment list parameter name is envp.
The environment parameter to main is not part of the C++ standard. Some systems maintain a global variable environ, which is an equivalent representation of envp.
1
2
3
4
5
6
7
8
9
10
11
12

#include <iostream>
using namespace std;
int main( int argc, char *argv[ ] )
{
cout << "Invoking command: " << argv[ 0 ] << endl;
for( int i = 1; i < argc; i++ )
cout << argv[ i ] << " ";
cout << endl;
return 0;
}

Figure 11-9

The echo command (a little more verbose than normal).

c++book.mif Page 225 Tuesday, April 29, 2003 2:13 PM

Primitive Multidimensional Arrays

225

m[0][0]

m[0][1]

m[0][2]

m[1][0]

m[1][1]

m[1][2]

m[0]

m[1]

m[2]

m[3]

m[4]

m[5]

Figure 11-11 Memory layout of multidimensional array

If only one environment variable needs to be examined, that standard library provides
getenv (standard header cstdlib needs to be included), with signature previously discussed
in Section 11.2:
const char *getenv( const char *prop );

Figure 11-10 shows how we list all the environment variables, assuming that the third
parameter to main is being supported.

11.6 Primitive Multidimensional Arrays

Sometimes arrays need to be accessed based on more than one index. A common example of
this is a matrix. A multidimensional array is an array that is accessed by more than one index. It
is allocated by specifying the size of its indices, and each element is accessed by placing each
index in its own pair of brackets. As an example, the declaration
int m[ 2 ][ 3 ];

// x has two rows and three columns

defines the two-dimensional array m, with the first index ranging from 0 to 1 and the second
index ranging from 0 to 2 (for a total of six objects). The compiler sets aside six memory locations for these objects.
An alternate form allows initialization, in a manner similar to Java:
int m[ ][ ] = { { 1, 2, 3 }, { 4, 5, 6 } };

which defines a two-dimensional array m, as before with 1, 2, 3 in row 0, and 4, 5, 6, in row 1.

Figure 11-11 shows the memory layout of either declaration, and helps explain why using
primitive multidimensional arrays is hard to do in C++. In Java, a two-dimensional array is
implemented as an array of arrays. Similarly, a vector of vector or the Matrix class in
Section 7.3 is implemented as a vector of vectors. However, when a multidimensional
array is created, we do not get an array of arrays. Instead we get a single array, arranged in rowmajor order. This means that the array elements proceed from the lowest to highest row.
Evidently, to access m[r][c], the compiler simply accesses the item in row-major index
3*r+c. Because of this, in a multidimensional array, the compiler must always know all the
dimensions of the multidimensional array, except possibly the first dimension. Thus, to declare a
routine to compute the sum of all the elements in the matrix, the following code works:

c++book.mif Page 226 Tuesday, April 29, 2003 2:13 PM

226

Chapter 11 Primitive Arrays and Strings

int sum1( int m[2][3] )

{
int totalSum = 0;
for( int r = 0; r < 2; r++ )
for( int c = 0; c < 3; c++ )
totalSum += m[ r ][ c ];
return totalSum;
}

So does
int sum2( int m[ ][ 3 ], int rows )
{
int totalSum = 0;
for( int r = 0; r < rows; r++ )
for( int c = 0; c < 3; c++ )
totalSum += m[ r ][ c ];
return totalSum;
}

But this does not:

int sum3( int m[ ][ ], int rows, int cols )
{
int totalSum = 0;
for( int r = 0; r < rows; r++ )
for( int c = 0; c < cols; c++ )
totalSum += m[ r ][ c ];
return totalSum;
}

Clearly, the compiler needs to know how many columns are in two-dimension array m in order
to do the double indexing, and this information is not available (the simple presence of a parameter named cols is hardly sufficient). If we want our routine to work for any two-dimensional
array, with arbitrary numbers of rows and columns, we would have to write this nonsense:
int sum4( int m[ ], int rows, int cols )
{
int totalSum = 0;
for( int r = 0; r < rows; r++ )
for( int c = 0; c < cols; c++ )
totalSum += m[ r * cols + c ];
return totalSum;
}

and then pass &m[0][0] as a parameter, but once we do so, whats the point? We might as well
declare a one-dimensional array from the start.
Although these problems can be avoided by creating a primitive array of primitive arrays,
the memory management issues are even more challenging than for a single primitive array. And
like one-dimensional arrays, memory management is required if we do not know the dimensions
at compile time. Thus it is highly recommended that a matrix class be used.

c++book.mif Page 227 Tuesday, April 29, 2003 2:13 PM

Key Points

227

11.7 Key Points

Primitive arrays are represented by a pointer variable rather than a real array object.
Heap-allocated arrays are created by invoking the array new operator, new[].
Heap-allocated arrays are reclaimed by invoking the array delete operator, delete[].
Do not simply invoke delete, as that may only reclaim the object in index 0 (if that
object is itself a pointer variable).
When a primitive array is passed as a parameter to a function, the array when viewed as an
aggregate is being passed by reference.
Primitive strings are implemented as a null-terminated array of characters.
String copies are difficult to do correctly because there is no runtime check that the target
string has enough memory to store the result (plus the null terminator).
A common string error is to simply declare a char*, and assume that a string is created.
The char* has to point at memory that can store the characters.
Avoid primitive string manipulations by using the library string class, the implicit conversion from char* to string, and the c_str member function.
++, when applied to a pointer variable that is pointing at an array element advances the
pointer variable to the next array element.
Pointer hopping used to be widely used to increase program speed. Modern optimizing
compilers reduce the need for pointer hopping.
Pointers can be used as iterators for the STL generic algorithms.
Command-line arguments are available through the argv parameter that can be provided
to main. argv is an array of strings (char *argv[]), and the command itself is
argv[0]. The number of command line arguments is passed as the argc parameter to
main.
Environment variables can often be accessed by a third parameter to main.
Avoid multidimensional arrays, which are simply stored as a flat one-dimensional array by
the compiler. When a primitive multidimensional array is listed as a formal parameter, all
but the first dimension must be provided.

11.8 Exercises
1. How are primitive arrays implemented in C++?
2. How are primtive strings implemented in C++?
3. How is a primitive array created from the memory heap?
4. How is a primitive array returned to the memory heap?
5. What is the difference between delete[] and delete?
6. What is the difference between the declarations int a[3] and int *a?
7. What happens when copying two primitive strings with = is attempted?
8. What is the result of using < on two primitive strings?
9. What is pointer hopping and why is it no longer as useful as it was in the past?

c++book.mif Page 228 Tuesday, April 29, 2003 2:13 PM

228

Chapter 11 Primitive Arrays and Strings

10. What is the result of adding a pointer and an integer?

11. How are command line arguments accessed in C++?
12. How are environment variables accessed in C++?
13. Write a program that reads a sequence of doubles into a primitive array and outputs
them in sorted order. You may not use the vector class template.
14. Write a function that returns true if primitive string str1 is a prefix of primitive string
str2. Do not use any library routines.
15. Write a function that returns a map<string,string> that contains the environment
variables.

c++book.mif Page 229 Tuesday, April 29, 2003 2:13 PM

H A P T E R

1 2

C-Style C++

ECAUSE C++ is based on C, and C++ was designed

to be compatible with most C programs, several C constructs that have much better C++ equivalents are still part of C++. We have already seen, for example, that C-style strings are part of the
language and still are used as parameters to main. Some C constructs are required for thirdparty libraries that were intended to be used by C programs.
In this chapter, we describe C constructs that are commonly seen in C++ programs, particularly in older C++ code, even though better alternatives exist. These constructs include preprocessor macros, using pointers to simulate call-by-reference, and the C libraries that are part of
Standard C++. Although it is a relatively rare occurrence, occasionally program speed can be
improved by judiciously using C constructs in performance-critical sections of code.

12.1 Preprocessor Macros

The original C language included a phase that was executed prior to the compilation. This phase,
known as the preprocessor, conceptually applied simple textual alterations to the input file, passing the result along to the compiler. The two most popular alterations were (and still are) the
inclusion of header files, via #include, and the macro replacement of tokens via #define.
The #define statement can also be used to implement parameterized macros, which look like,
but are quite different from, functions. The preprocessor in current C and also C++ is now part
of the compilation phase, although the apparent behavior is as if the preprocessor was executed
prior to compilation.
We have already seen the preprocessor directives #include, #ifndef, and #endif.
Additional preprocessor directives include #ifdef, #else, #elif, and #if, most of which
are self-explanatory. Both #if and #elif can be used in code such as:

229

c++book.mif Page 230 Tuesday, April 29, 2003 2:13 PM

230

Chapter 12 C-Style C++

#if !defined( __BORLANDC__ ) || __BORLANDC__ >= 0x0530

// code for newer Borland compilers
#elif defined( __GNUC__ )
// code for g++
#endif

defined is part of the preprocessor also, and returns true if the symbol has been the target of a
#define. A symbol can also be undefined by using #undef.
12.1.1 Simple Textual Substitution
The most interesting (and dangerous if abused) preprocessor directive is #define. The directive
#define IDENTIFIER ReplacementText

causes the substitution of all occurrences of the token IDENTIFIER with the sequence of
tokens given in ReplacementText. This form is commonly used to give symbolic meaning
to constants. Typical usage is
#define MAX 50

all subsequent occurrences of MAX when seen as a separate token (characters inside a string constant do not count) are replaced by 50. Thus
cout << "MAX is" << MAX << endl;

becomes
cout << "MAX is" << 50 << endl;

ReplacementText can be essentially anything, so theoretically, one could see uses that
include variations of the following:
#define FOREVER for( ; ; )
#define extends public :

On the other hand, simple errors in the #define yield compilation errors, but the compiler will flag the error at the point of the expansion, often providing little hint that the problem
was at the #define statement. For instance, these three #define statements
#define MAX 50;
#define MAX = 50
#define MAX 50
// the max

if used in
if( i == MAX)

cause an expansion to, respectively,

if( i == 50; )
if( i == = 50 )
if( i == 50 // the max )

all of which have syntax errors. Its hard to get in serious trouble here, since the code doesnt
compile. However, the following directive illustrates more dangerous problems:

c++book.mif Page 231 Tuesday, April 29, 2003 2:13 PM

Preprocessor Macros

231

#define RANGE_SIZE MAX - MIN

The directive is bad because if it used in

int r2 = RANGE_SIZE * RANGE_SIZE;

the statement expands to

int r2 = MAX - MIN * MAX - MIN ;

which is a problem because the precedence is wrong. This means that when using preprocessor
definitions, it is important to parenthesize liberally, as in
#define RANGE_SIZE (MAX-MIN)

The use of #define to provide replacement text (rather than simply the fact that a symbol is defined for the ifndef/endif trick) is obsolete in C++, since symbolic constants can be
declared with const, and textual substitution, such as replacing :public with extends is
probably bad style. It was needed in C because old C did not have a const, and array sizes
needed to be constants.
12.1.2 Parameterized Macros
Parameterized macros were used in C to avoid the overhead of function calls. For instance,
rather than write a function to compute the absolute value, we define a macro:
#define absoluteValue(x)

( (x) >= 0 ? (x) : -(x) )

Of course we parenthesize liberally. There are various syntax rules. One rule is that the parenthesis that follows the macro name cannot be separated with whitespace, since otherwise it looks
like a simple textual replacement.
Note here that the parameters are typeless, which is nice in a pre-template world. Now, the
statement
y = absoluteValue(a-3);

is expanded to
y = ( (a-3) >= 0 ? (a-3) : -(a-3) );

what happens is that the macro is expanded, with a-3 replacing all occurrences of x. The macro
can have several parameters. Also note the use of the ?: operator, which seems to have been
invented just for this, since substitution with a full if statement would not generate legal code.
Although the macro looks good here, it should not be necessary in the modern C++ world,
because optimizing compilers are very good at doing this exact kind of inline optimization. They
are so good at it, that whereas C++ has an inline directive, Java does not even bother, deciding
that the compiler (and possibly the runtime system in the case of Java) doesnt need any hints.
Other uses of the macros such as writing a typeless routine or simulating call by reference (by
putting the changes to the parameters in the substitution text) are better handled by templates or
call-by-reference.
Not only are there better C++ alternatives to the macro, in general macros are not semantically equivalent to a function call, as is demonstrated by

c++book.mif Page 232 Tuesday, April 29, 2003 2:13 PM

232

Chapter 12 C-Style C++

1
2
3
4
5
6
7
8
9
10
11

#define printDebug( expr ) cout << __FILE__ << " [" <<
\
__LINE__ << "] (" << #expr << "): " << ( expr ) << endl;
int main( )
{
int x = 5, y = 7;
printDebug( x + y );
return 0;
}

Figure 12-1

Macro for debugging prints with file and line number included

z = absoluteValue( ++a );

which is expanded to
z = ( (++a) >= 0 ? (++a) : -(++a) );

and produces the absolute value of a+2, and increments a twice.

There are some uses of macros that are reasonable. The assert mechanism is a macro,
and takes advantage of some preprocessor symbols such as __LINE__ and __FILE__ that
expand to the current line number and file name, and the # operator that converts a parameter to
a string. An example of similar macro code is shown in Figure 12-1 which uses \ to continue a
line and which outputs
Fig12.1.cpp [8] (x + y): 12

Similarly, we could write a version of assertWithException that is similar to assert,

but throws an exception instead of calling abort.

12.2 Unsafe Casts

In C, the only typecast is the Java style cast, except that it is unchecked. Thus as we saw in
Section 6.7 to cast down an inheritance hierarchy, the dynamic_cast should be used. Sometimes we need to do totally bizarre casts. This is a very rare occurrence and usually means we
are either writing totally bizarre code, or perhaps device drivers or other low-level code.
In such a case, although the old style cast works, it is better to use
reinterpret_cast, since this cast stands out much more in the code. When you use a
reinterpret_cast, all bets are off, since you are in effect, disregarding the typing information of an object. All you can expect is that the result of a reinterpret_cast will have the
new type, with the same bits as the old. In the case of a pointer, it will be pointing at the same
object, but the object will be viewed as a different type that might not make any sense at all.

c++book.mif Page 233 Tuesday, April 29, 2003 2:13 PM

C-Style Parameter Passing

1
2
3
4
5
6
7
8
9
10
11
12

233

void clear( char [ ] arr, int n )

{
int *iarr = reinterpret_cast<int *>( arr );
int newn = n / sizeof( int );
// number of ints
int extras = n % sizeof( int );
// extra chars leftover
for( int i = 0; i < newn; i++ )
iarr[ i ] = 0;
for( int j = n - extras; j < n; j++ )
arr[ j ] = 0;
}

Figure 12-2

Nonsense example of reinterpret_cast

Heres a poor example: we have a primitive array of char and we want to set all the
entries to zero. We are unhappy with the compilers performance (very unusual) so we decide to
take matters into our own hands. The idea is to treat the array of char as an array of int, and
set the ints to zero. This reduces the number of assignments by a factor of 4, since typically an
int is four chars. The resulting code is shown in Figure 12-2. This code is nonsense because
it is unlikely to be useful in practice.

12.3 C-Style Parameter Passing

Because C does not have call-by-reference, pointer variables are used to achieve the effect. The
idea is identical to what happens in Java. Recall that in Java, when we pass to a method a reference to an object, call-by-value guarantees that the value of the reference (i.e. which object it
references) cannot be changed by the method, but the method can change the state of the object.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

void swap3( int *a, int *b )

{
int tmp = *a;
*a = *b;
*b = tmp;
}
int main( )
{
int x = 5, y = 7;
swap3( &x, &y );
cout << "x=" << x << " y=" << y << endl;
return 0;
}

Figure 12-3

Using pointers as parameters to simulate call-by-reference

c++book.mif Page 234 Tuesday, April 29, 2003 2:13 PM

234

Chapter 12 C-Style C++

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

struct tm
{
int
tm_sec;
int
tm_min;
int
tm_hour;
int
tm_mday;
int
tm_mon;
int
tm_year;
int
tm_wday;
int
tm_yday;
int
tm_isdst;
};

/*
/*
/*
/*
/*
/*
/*
/*
/*

seconds after the minute (0- 61)

minutes after the hour
(0- 59)
hours after midnight
(0- 23)
day of the month
(1- 31)
month since January
(0- 11)
years since 1900
(0)
days since Sunday
(0- 6)
days since January 1
(0-365)
daylight savings time flag

*/
*/
*/
*/
*/
*/
*/
*/
*/

typedef unsigned long time_t;

/* Some functions */
char *asctime(const struct tm *);
time_t mktime(struct tm *);

Figure 12-4

struct tm as defined in <ctime> (and <time.h> for C systems).

So in C, a we pass the address of an object to a function that expects a pointer to the

object. Then the function can change the state of the object (by dereferencing the pointer), and
we have simulated a call-by-reference. The typical idiom is shown in Figure 12-3. Here we see
that function swap3 swaps the values stored in the objects pointed at by a and b, and main
passes the address of the two objects that are to be swapped.
We know that in addition to call-by-reference, C++ uses call-by-constant reference to
avoid the overhead that call-by-value imposes to copy large objects. Although C does not have
classes, the struct in C can be large, especially if it stores a primitive array. Thus, structs
in C are almost always passed by sending the address of a struct to a routine that expects a
pointer, putting a const in front if needed to signal that the struct is not to be altered.
Figure 12-4 shows some type declarations that C++ has inherited from C. All standard
header files in C are part of C++, and are located in both the same .h file as in C, and alternately, a file that does not end in .h, but starts with c, and is in the std namespace.
The tm struct stores a host of ints that attempt to encapsulate a specific point in time.
A time_t is a type representing time; typically it is a 32-bit unsigned entity, storing the number of seconds since midnight Jan 1, 1970. This allows a time_t object to store time until
roughly the year 2037.
We see two functions in time.h. The function asctime takes a pointer to a tm
struct and returns a primitive string representing the time and date of the tm struct that is
being pointed at. Note that the return type is a pointer; looking at the documentation, we find
that the returned string is static data, and must be used prior to the next call to asctime.
mktime takes a pointer to a tm struct and returns the time_t entity that corresponds to it,

c++book.mif Page 235 Tuesday, April 29, 2003 2:13 PM

C Library Routines

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27

235

// Find Friday the 13th birthdays for person born Oct 13, 1937
#include <ctime>
#include <iostream>
using namespace std;
int main( )
{
const int FRIDAY = 6 - 1;
tm theTime = { 0 };
theTime.tm_mon = 10 - 1;
theTime.tm_mday = 13;

// Sunday is 0, etc...
// Set all fields to 0
// January is 0, etc...
// 13th day of the month

for( int year = 1937; year < 2073; year++ )

{
theTime.tm_year = year - 1900; // 1900 is 0, etc...
if( mktime( &theTime ) == -1 )
{
cerr << "mktime failed in " << year << endl;
continue;
}
if( theTime.tm_wday == FRIDAY )
cout << asctime( &theTime );
}
return 0;
}

Figure 12-5

Progam to find all Friday the 13th birthdays for a friend

or -1 if the tm struct is out-of-range. The tm struct is mutable because mktime also

attempts to fill in fields such as tm_wday and tm_yday. Note that in both cases, the tm type
includes the word struct. This is optional in C++, and so we have never done it, but it is
required in C. A person born on Oct. 13, 1937 will naturally have a few birthdays fall on Friday
the 13th. Figure 12-5 shows a simple program to calculate when this occurs.

12.4 C Library Routines

As we have mentioned, the C library is part of C++. This section described some of the library
routines that you may run into. For most, you should run away from these.
12.4.1 printf and scanf
printf is the basic output function in C. It is surprising to see how many C++ programs still
have calls to printf. The basic form of printf is:
int printf( const char *control, val1, val2, ... );

c++book.mif Page 236 Tuesday, April 29, 2003 2:13 PM

236

Chapter 12 C-Style C++

The first parameter is the control string and is output, except that the additional parameters are substituted as appropriate into the control string in places marked by a % conversion
sequence. For instance %d is used to print an integer in decimal, %o prints an integer in octal, %s
prints a (primitive) string, %f prints a float or double, and %% prints a %.
The return value of printf is the number of characters actually written, or -1 if there is
an error. But hardly anybody ever bothers to check the return code. As an example,
int x = 5;
double y = 3.14;
printf( "x is %d, y is %f\n", x, y );

returns 18 (characters, newline included) and prints

x is 5, y is 3.14

After the % and before the character that specifies the type, a host of options control all of the
same things that can be controlled by the ostream manipulators. For instance in Figure 9-5 we
had:
out << left << setw( 15 ) << name << " "
<< right << fixed
<< setprecision( 2 ) << setw( 12 ) << salary;

the equivalent call to printf is

printf( "%-15s %12.2f", name, salary );

The most important thing to know about printf is that it is not type-safe. If the additional parameters do not match the conversion specifiers in the control string, you get gibberish.
Because C is not object-oriented, only the primitive types have conversion specifiers;
even in C++, you cannot define new specifiers for user-defined class types, making printf
vastily inferior to the C++ iostream library.
The C function that reads formatted input is scanf. The basic form is
int scanf( const char *control, void *obj1, void *obj2, ... );

The first parameter is the control string, as before, except that doubles should use %lf
instead of %f. Also, field width specifiers become maximums, instead of minimums (this is useful for strings, because you want to make sure you dont read more characters than you have
room for). The return value is the number of conversion specifiers that are actually matched, so
if this is less than the number of conversion specifiers, something has gone wrong. As an example,
int x;
double y;
char name[ 100 ];
scanf( "%d %lf %99s", &x, &y, name );

reads an int, double, and primitive string, putting them in the objects specified, and hopefully returns 3. Note that name is already an address, so we dont need the &.

c++book.mif Page 237 Tuesday, April 29, 2003 2:13 PM

C Library Routines

237

Like printf, scanf is not typesafe, and errors are likely to result in an abnormal program termination, because pointer variables are involved. The most common mistake with
scanf is to forget to pass an address, as in
scanf( "%d", x );

This attempts to put an integer in the memory location given by the integer stored in x. It is
unlikely that this is a valid location; it certainly is not xs location.
scanf is a dangerous routine that has little use in a C++ program. If possible you should
replace existing calls to print and scanf with the iostream equivalents. You should avoid
having a program use both libraries, since both libraries buffer, and I/O can be interwoven unexpectedly.
Use of printf and scanf, and all file routines requires the standard header cstdio
(or stdio.h).
12.4.2 File I/O
In C, file I/O is supported with the FILE* type and a set of functions. FILE is a typedef for a
struct, and because in C structs have historically been passed and returned by pointers, all
the I/O routines work with FILE* as the basic stream parameter. The FILE* type is used for
both reading and writing. Three standard streams, stdin, stdout, and stderr are already
predefined as standard input, standard output, and standard error.
In order to use a file, we must obtain a stream for it by using fopen. When we are done,
we close the file with fclose. The parameters to fopen are a primitive string representing the
filename, and a primitive string representing the mode. Typical modes are "r" for reading, "w"
for writing, and "a" for appending. Also "rb", "wb", and "ab" are used for binary files, on
systems that distinguish between text and binary files. fopen returns a FILE*, or NULL if
there is an error. Thus the declarations for fopen and fclose are:
FILE *fopen( const char *fileName, const char *mode );
int fclose( FILE *stream );

The calls getc and putc provide single-character input and output for files. Their declarations are:
int getc( FILE *stream );
int putc( int ch, FILE *stream );

getc returns EOF if the end-of-file is seen. The return type of getc is an int, for the
same reason that method java.io.Reader.read returns an int representing the char
that has been read: a char is not sufficient to store all possible characters and the EOF symbol.
Unfortunately, in C and C++, if the return value of getc is directly assigned to a char, the
code will probably not work on files with full character sets (e.g binary files). putc returns EOF
if the write fails, otherwise it returns ch. Attempts to write on a stream open for reading, or viceversa, yields a bad return value, instead of the compile-time error that you get in C++.
Like printf and scanf, fprintf and fscanf can be used for formatted input and
output. The first parameter is a FILE*, and the remaining parameters are as described earlier.

c++book.mif Page 238 Tuesday, April 29, 2003 2:13 PM

238

Chapter 12 C-Style C++

Figure 12-6 shows a routine that copies from one file to another. We write it in C-style,
using primitive strings. Line 4 declares charCounted, and line 5 declares ch which will represent a character read by fgetc. Note that ch must have type int. If it is char, and this program is applied to large binary files, the copy will most likely be short. At line 6, observe that
both sfp and dfp must BOTH be declared as pointer variables. The first * applies only to sfp.
Line 8 is a simple alias test. The error messages at line 10 illustrate the use of the stderr
stream. Lines 13 and 18 open the files as binary files. If the second open fails, we must remem1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37

// Copy files; return number of chars copied

int copy( const char *destFile, const char *sourceFile )
{
int charsCounted = 0
int ch;
FILE *sfp, *dfp;
if( strcmp( sourceFile, destFile ) == 0 )
{
fprintf( stderr, "Cannot copy to self\n" );
return -1;
}
if( ( sfp = fopen( sourceFile, "rb" ) ) == NULL )
{
fprintf( stderr, "Bad input file %s\n", sourceFile );
return -1;
}
if( ( dfp = fopen( destFile, "wb" ) ) == NULL )
{
fprintf( stderr, "Bad output file %s\n", destFile );
fclose( sfp );
return -1;
}
while( ( ch = getc( sfp ) ) != EOF )
if( putc( ch, dfp ) == EOF )
{
fprintf( stderr, "Unexpected write error.\n" );
break;
}
else
charsCounted++;
fclose( sfp );
fclose( dfp );
return charsCounted;
}

Figure 12-6

Copy files using getc and putc; return number of characters copied

c++book.mif Page 239 Tuesday, April 29, 2003 2:13 PM

C Library Routines

239

ber to close the first file. The copying of files is performed at lines 25 to 32, and then the files are
closed at lines 34 and 35. Failing to close the output file could result in some buffered data not
being written out, especially if the program terminates abruptly. fflush can be used to force
buffered data to be written out, prior to closing the file.
Two similar looking functions are fgetc and fputc. Because getc and putc are preprocessor macros, they can be dangerous to use if the arguments involve side effects. In this
extremely rare case, that is almost certainly poor programming practice, fgetc and fputc,
which are guaranteed to be functions and not macros, can be safely used.
Another routine that is provided is ungetc, which allows the putting back of a character
onto the input stream:
int ungetc( int ch, FILE *stream );

To read and write lines at a time, we can use fgets and fputs:
char *fgets( char *str, int howMany, FILE *stream );
int fputs( const char *str, FILE *stream );

fputs outputs a string to an output stream. It does not supply a newline character unless
one is already present. fgets reads characters from an input stream until one of three events
occurs:
1. EOF is encountered.
2. A newline is encountered.
3. howMany-1 characters are seen, before event 1 or 2 occurs.
After the characters are read, a null terminator is appended. A newline is stored only if it
was encountered. str is returned on success; if no characters were read because of an EOF or
any other error, a NULL pointer is returned.
As an example, suppose we want to read a large file one line at a time, using normal C++:
void processFile( string fileName )
{
ifstream fin( fileName.c_str( ) );
string oneLine;
while( getline( fin, oneLine ) )
...
}

On one of our systems, processing a file of about 78 million bytes containing about 2 million lines takes about 7.5 seconds simply to do the I/O. The slowness seems to have to do with
the fact that this form of getline deals with strings. However, here is an method in
istream called getline, that works somewhat like fgets, except that it does not retain the
newline character.

c++book.mif Page 240 Tuesday, April 29, 2003 2:13 PM

240

Chapter 12 C-Style C++

1
2
3
4
5
6
7
8
9
10
11
12

bool getlineFast( istream & in, string & oneLine )

{
static const int MAX_LINE_LEN = 10000;
char str[ MAX_LINE_LEN + 2 ];
in.getline( str, MAX_LINE_LEN );
if( in.gcount( ) == 0 )
return false;
oneLine = str;
return true;
}

Figure 12-7

getlineFast:: takes half the time of getline, but can fail

Figure 12-7 shows routine called getlineFast that uses this getline routine, presuming that no line is longer than MAX_LINE_LEN. At line 7, the call to gcount returns the
number of characters read by getline. Presumably, if this count is 0, getlineFast should
return false; otherwise we copy the primitive string into oneLine. If we use getlineFast
instead of getline, 7.5 seconds becomes 2.8 seconds! However, getlineFast only works
if lines have less than MAX_LINE_LEN characters. If not, very long lines will be split into sepa1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

bool getline( FILE *fin, string & oneLine )

{
static const int MAX_LINE_LEN = 100;
char str[ MAX_LINE_LEN + 2 ];
for( oneLine.erase( ); ; oneLine += str )
{
char *result = fgets( str, MAX_LINE_LEN, fin );
if( result == NULL )
return oneLine.length( ) != 0;
int len = strlen( str );
if( str[ len - 1 ] == '\n' )
{
str[ len - 1 ] = '\0';
break;
}
}
oneLine += str;
return true;
}

Figure 12-8

Replacement for getline; even faster than getlineFast

c++book.mif Page 241 Tuesday, April 29, 2003 2:13 PM

C Library Routines

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20

241

void lastChars( const string & fileName, int howMany )

{
FILE *fin = fopen( fileName.c_str( ), "rb" );
if( fin == NULL )
throw io_exception( );
// made this one up
else if( howMany <= 0 )
throw invalid_argument( "howMany is negative" );
fseek( 0, SEEK_END, fin );
int fileSize = ftell( fin );
if( fileSize < howMany )
howMany = fileSize;
fseek( -howMany, SEEK_END, fin );
int ch;
while( ( ch = getc( fin ) ) != EOF )
putc( ch, stdout );
if( !feof( fin ) )
throw io_exception( );
}

Figure 12-9

Routine that prints last howMany characters from fileName

rate lines silently. Although there are some solutions available in the istream class, all of
them will take longer than 2.8 seconds. It turns out that this is a great place to use fgets.
We can write a getline that takes a FILE*, and then have the user create a FILE*
instead of an istream. With reasonable care, we can localize these changes, so that the rest of
the program doesnt have to change. Best of all, our routine will handle arbitrarily long lines and
be faster than getlineFast, taking about 2.4 seconds. The additional getline is shown in
Figure 12-8.
We begin by clearing out oneLine at line 6. Note that assignment of "" to oneLine
with = is significantly slower than calling erase, and since this is done on every line (2000000
times), this inefficiency is significant enough to be worth using erase. In the main loop, the
idea is to keep calling fgets, concatenating the result to oneLine (also line 6) until we read
characters that contain a terminating newline character (that test is performed at line 13). Those
characters are also appended (at line 20 outside of the main loop), after the newline character is
stripped out at line 15.
Other routines that are available include feof, the random access routines fseek, and
ftell that mimic the routines in istream, discussed in Section 9.6. (Actually the istream
routines mimic these.) For fseek, the constants that specify beginning, current, and end, are
SEEK_SET SEEK_CUR, and SEEK_END. The code in Figure 12-9 shows the direct correspondence between the FILE* and istream routines, by recoding Figure 9-7 line-for-line. (However, as we mention in Section 12.5, several aspects of this code that are legal in C++, such as
exceptions, are not legal C).

c++book.mif Page 242 Tuesday, April 29, 2003 2:13 PM

242

Chapter 12 C-Style C++

Finally, we mention sprintf and sscanf which allow printing and scanning from a
primitive string. These signatures are:
int sprintf( char *buffer, const char *control, val1, val2, ... );
int sscanf( const char *buffer, const char *control,
void *obj1, void *obj2, ... );

For instance, if we have

int x1 = 37, x2;
double y1 = 3.14, y2;
char oneLine[ 100 ];
sprintf( oneLine, "%d %f", x1, y1 );

oneLine will contain "37 3.14". A subsequent

sscanf( oneLine, "%d %lf", &x2, &y2 );

populates x2 and y2 with 37 and 3.14, respectively. Note that

sscanf( oneLine, "%d", &x2);
sscanf( oneLine, "%lf", &y2 );

sets x2 to 37 and y2 to 37.0, because sscanf does not maintain a notion of previous parsing of
the buffer string.
sprintf is dangerous because it is possible that buffer does not contain enough characters to hold the result. sscanf is dangerous, especially in the case where the target tokens are
strings. The functionality of these functions is provided in ostringstream and
istringstream, so there is little need to use them.
12.4.3 malloc, calloc, realloc, and free
In C, memory is allocated from the memory heap by calling a family of alloc functions,
malloc, calloc, and realloc. The memory is released back to the memory heap by calling
free. The declarations found in the standard header cstdlib and stdlib.h are:
void
void
void
void

*malloc( size_t howManyBytes );

*calloc( size_t objectSize, int numObjects );
*realloc( void *original, size_t newSize );
free( void *p );

The most important things to know about these routines are that they are not typesafe, and
they do not invoke constructors. Instead, they simply return a pointer to raw memory, that can be
cast to the appropriate type. In the case of malloc, the memory is not initialized. In the case of
calloc, memory is initialized to all bits zero, but that might not be suitable for pointer types or
doubles. realloc can be used to get more memory, which is initialized by using the original
memory (and then freeing the original, which needs to have been heap-allocated). free returns
the memory back to the heap, but does not invoke destructors.
It is important not to mix malloc and delete or new and free. Specifically, memory
that was allocated by malloc should be released by free and not by delete. Memory that

c++book.mif Page 243 Tuesday, April 29, 2003 2:13 PM

C Library Routines

243

was allocated by new should be released by delete, and not by free. Otherwise, havoc
results.
C-library routines that return heap-allocated memory get the memory from an alloc
function. Thus their memory is returned to the heap by calling free. One such example is
strdup.
12.4.4 atoi, atof, strtol, and strtod
atoi and atof are original C library routines that parse a primitive string, returning an int or
a double respectively. These routines handled errors poorly. Their replacements, strtol and
strtod, which yield long and double, respectively, and were added in a later version of C
do a better job. Although, these are not needed in C++ since istringstream can be used to
do the same thing, it is certainly possible that strtol and strtod could be more efficient
than using an istringstream.
12.4.5 system
The system function, in standard header cstdlib or stdlib.h, is used to invoke a command. It takes a primitive string representing the command; this string is passed to the operating
systems command processor and is run. How this is done is highly system dependent, and obviously non-portable, since few commands are available on all systems. For instance,
system( "dir" );

invokes the dir command (if one exists).

12.4.6 qsort
qsort is a generic sorting algorithm with declaration
void qsort( void *base, size_t numItems, size_t itemSize,
int cmp( const void *, const void * ) );

The parameters to qsort represent

1. base: the array being sorted
2. numItems: the number of items being sorted
3. itemSize: the size of the items being sorted
4. cmp: a comparison function that takes pointers to two items and returns a negative number, zero, or a positive number, if the first item is smaller than, equal to, or larger than the
second item.
Figure 12-10 shows the call to qsort to sort an array of integers, along with the definition of an
appropriate comparison function. The comparison function takes two generic pointers to the
items. At line 3, we convert lhs to a pointer to a const int, and then dereference it to get the
int that lhs is pointing at. In C++, we would sensibly use reinterpret_cast, rather than

c++book.mif Page 244 Tuesday, April 29, 2003 2:13 PM

244

Chapter 12 C-Style C++

1
2
3
4
5
6
7
8
9

int intCmp( const void *lhs, const void *rhs )

{
int lhint = *(const int *)lhs;
int rhint = *(const int *)rhs;
return lhint < rhint ? -1 : lhint > rhint;
}
...
int arr[ ] = { 3, 5, 1, 2, 6 };
qsort( arr, 5, sizeof( int ), intCmp );

Figure 12-10 Parameters to qsort, and a comparison function

the old style, but of course in C++ we would use the STL sort algorithm in the first place.
qsort is not required to be efficient, and although it is typically implemented as quicksort, older versions are known to have deficiencies that can cause quadratic behavior on degenerate inputs (such as arrays with only two unique items, repeatedly occurring randomly).
Furthermore, because qsort is already compiled, it is unreasonable to expect any inline optimizations to be performed. So applying qsort on an array of ints is going to be significantly
slower than using the STL sort algorithm, or even any quicksort that can be found in a standard textbook.
12.4.7 Variable Number of Arguments
Some routines, such as print and scanf, require a variable number of arguments. Allowing a
variable number of arguments is not a good thing for a type-safe language, which is why the feature is not part of Java. However, it has long been part of C and C++.
To declare an unspecified number of arguments, we use ellipses (...). They can be processed by a set of macros in standard header stdarg.h (or cstdarg). The macros are
va_start, va_end, and va_arg.
A function declares at least one fixed parameter, and then uses ellipses to specify an
unknown (possibly zero) number of additional parameters.
Macro va_start is used to initialize an object of type va_list with the last fixed
argument. Then, each call to va_arg returns the next argument. va_arg takes the va_list
object and the type of the parameter that is expected. This type cannot be one that widens when
passed as an argument; use double, int, or unsigned int in place of shorter alternative
such as float, and short. Prior to returning, the function should call va_end.
As an example, the code in Figure 12-11 illustrates a function printStrings that
prints an arbitrary number of primitive strings. At line 5, we see the ellipses used to specify
additional parameters after the first. The va_list object is declared at line 8, and initialized at
line 11. Then we repeatedly loop at lines 12 to 13 reading parameters of type const char *.
There needs to be some way to signal the end. Alternatives include encoding this information in
the first parameter (as is done in printf and scanf), or simply terminating the list with a
NULL pointer. We use the second alternative. Thus we see at line 20, that the last parameter to

c++book.mif Page 245 Tuesday, April 29, 2003 2:13 PM

C Programming

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

245

#include <iostream>
#include <cstdarg>
using namespace std;
void printStrings( const char *str1, ... )
{
const char *nextStr;
va_list argp;
cout << str1 << endl;
va_start( argp, str1 );
while( ( nextStr = va_arg( argp, const char * ) ) != NULL )
cout << nextStr << endl;
va_end( argp );
}
int main( )
{
printStrings( "This", "is", "a", "test", (const char *) NULL );
return 0;
}

Figure 12-11 Illustration of variable arguments

printStrings is NULL (we use a pointer for the technical reason that we want to ensure that
the type is a pointer type. This would be automatic if we had a fixed parameter list, but not with
variable parameter lists). At the end, we call va_end at line 15.

12.5 C Programming
If you program in C instead of C++, you need to be aware that some C++ features are missing in
C. The partial list, based on ANSI C (note that items 11 and 12 will behave the same in recently
adopted C99 as in C++) includes:
1. C has no classes, and does not support even object-based programming.
2. C has no vector or string type. You have to use pointers.
3. C has weaker type checking than C++. The compiler will catch fewer problems.
4. C does not have reference variables or call-by-reference. You will have to pass pointers to
objects to simulate the effect.
5. C does not provide function overloading. Once you use a name, thats it.
6. C does not provide namespaces.
7. C does not provide operator overloading.
8. C does not have templates.
9. C does not have exceptions.

c++book.mif Page 246 Tuesday, April 29, 2003 2:13 PM

246

Chapter 12 C-Style C++

10. C does not allow default parameters to functions.

11. C requires that all local variables be declared together at the top of a function. You cannot
declare variables once you have a non-declaration statement, unless you create a block.
12. C does not have the // comment, although some compilers support it.
Basically, whats left? Not much. Its a long way down to C from the C++ or Java world.

12.6 Key Points

The standard C library is part of the C++ library. Header files such as stdio.h are available as both stdio.h and cstdio. The latter header file places all of the library in the
std namespace; the older header file leaves the library in the global namespace.
The preprocessor can be used to implement macros; however, macros can be dangerous if
invoked with a parameter that contains side effects.
C-style parameter passing is needed to interact with libraries that were written for C. To
simulate call-by-reference, or call-by-constant reference for large objects, we pass the
address of the object instead of the object. The formal parameter is declared as a pointer
(or a pointer to a constant).
printf and scanf are not typesafe and their functionality is completely contained in
ostream and istream.
Files can be accessed by using FILE* to represent both input and output streams.
Attempts to read from a file open for writing (or vice-versa) results in a bad return code
for the read, rather than a compile-time error. Most of the C-style I/O routines are cosmetically similar to the C++ counterparts.
Sometimes I/O that uses char* instead of string can be more efficient. Occasionally,
using fgets can help implement the more efficient I/O.
Never call free on an object created with new; never call delete on an object created
with malloc. Avoid using the alloc family because they allocate raw memory, without
calling constructors, and are not typesafe.
strtol and strtod can be used to extract a long or double (or smaller type) from a
primitive string. Although istringstream can accomplish the same thing, it is reasonable to expect that strtol and strtod can be more efficient for this task.
system can be used to invoke a system command.
C and C++ support variable-length argument lists, via the standard header cstdarg or
stdarg.h.
Programming exclusively in C requires abandoning lots of comforts of Java and C++.

12.7 Exercises
1. Describe the basic functionality of the preprocessor.
2. What does reinterpret_cast do?

c++book.mif Page 247 Tuesday, April 29, 2003 2:13 PM

Exercises

247

3. How do C programs simulate call-by-reference?

4. How do C programs pass large structures?
5. How can a function accept a variable number of arguments?
6. The majority function for booleans returns true if more than half of the parameters are
true, and false otherwise. Write the macro
#define MAJORITY( x, y, z )

which returns true if at least two of x, y, and z are true, and false otherwise.
7. Using one / and one % operator, implement the macro
#define DIVIDE( numerator, denomibator, quotient, remainder )

8. Write an iteration macro:

#define ITERATE( i, low, high )

that generates the clause that directs a for loop running from low to high, inclusive.
9. Write a macro, max4, to find the maximum of four numbers.
10. A safe macro can be implemented by declaring a block and copying the macro parameters
into block variables. Parameters thus become implicitly typed. Write the max4 macro for
ints as a safe macro.
11. Implement the macro
#define main( )

that prints the compilation date and time and then starts execution of the program.
12. Write a function that takes a month, day, and year, and returns the day of the week. Implement your function by using the routines in standard header file <ctime>.
13. Write a program that uses printf to print out the following numbers in decimal, octal
and hexadecimal: 37, 037, 0x37.
14. Reimplement Exercise 9.9 using I/O routines in standard header file <cstdio>.
15. Reimplement Exercise 9.10 using I/O routines in standard header file <cstdio>.
16. Reimplement Exercise 9.11 using I/O routines in standard header file <cstdio>.
17. Reimplement Exercise 9.12 using I/O routines in standard header file <cstdio>.
18. Implement a version of findMax that takes a variable number of int arguments.

c++book.mif Page 248 Tuesday, April 29, 2003 2:13 PM

248

Chapter 12 C-Style C++

c++book.mif Page 249 Tuesday, April 29, 2003 2:13 PM

H A P T E R

1 3

Using Java and C++: The

JNI

native method is a method that is implemented in

another language, such as C or C++, and run in the Java Virtual Machine. The JDK provides a
standard programming interface called the Java Native Interface (JNI), that in theory allows the
Virtual Machine to invoke C and C++ code somewhat portably.
In this chapter, we describe the Java Native Interface. We begin by explaining why one
might want to implement some Java methods in an alternate language. Then we describe the
basic layout of the JNI, provide an example that involves invoking a simple native method, and
then write native methods that access Java objects. Finally, we briefly discuss the use of Java
features, such as exceptions and object monitors in native code.

13.1 JNI Basics

A native method is a method that is specified in Java with the reserved word native, but is
implemented in another language. Typically the other language is C or C++. We will use C++ in
most of our examples. Using C involves slightly more syntax, but no additional concepts. We
show a C example in Section 13.7. There are several common reasons to use native methods:
1. You already have significantly large and tricky code written in another language and you
would rather not rewrite it in Java. Instead, you would like to use the existing codebase.
2. You need to access system devices, or perform some platform specific task that is beyond
the capability of Java. Many Java library routines eventually invoke private native methods for just this purpose. For instance the I/O, threading, and networking packages all contain private native methods.
3. You think that Java might be too slow for your application, and that performance can be
enhanced by implementing time-critical code in C++.

249

c++book.mif Page 250 Tuesday, April 29, 2003 2:13 PM

250

Chapter 13 Using Java and C++: The JNI

Although Java used to be painfully slow, a modern Java implementation has performance
that is comparable to C++ for many applications. Using JNI to achieve performance improvements is possible, but is no longer needed as much as it used to be.
Using JNI has significant downsides. First, you lose portability. A native implementation
must be supplied for each platform. Given the large Java library that already makes use of native
methods, if a new native method has identical C++ code that can be used on all platforms, then it
is likely that the code could have been implemented in Java in the first place. Second, you lose
safety. Native methods are not afforded the same protections as Java methods. Once you enter in
C++ code, all bets are off, and any C++ bug, such as indexing an array out-of-bounds, using a
stale pointer, and trashing memory can occur. Third, the implementation of the native method is
contained in a dynamic library (in a Windows environment, a .dll; in a Unix environment a
.so file). Any Java program that uses native methods must load dynamic libraries, and this is an
operation that the Java Security Manager might object to, because of the safety concerns listed
above. Lack of safety means lack of security, so for instance, by default user-defined native
methods generally cannot be invoked inside an applet. Fourth, the code is cumbersome, often
compiles and often fails to run because of silly typing errors such as poor capitalization or missing semicolons inside of string constants.
Once we decide to use JNI, the basic procedure is relatively straightforward.
1. A Java class declares that some methods have non-Java implementations by marking the
methods as native.
2. A C++ function is written that implements the native method, using the JNI protocols.
3. The C++ function is compiled in a dynamic library.
4. The Java Virtual Machine loads the dynamic library, and then calls to the native method
are handled by invoking the implementation in the dynamic library.
The devil, of course, is in the details, which are numerous, since the JNI is expected to work not
only for C++, but also for C and other languages that are not object-oriented. For instance,
1. How are fields and methods of a Java object used, given that C has no classes?
2. How are parameters passed from Java to C/C++?
3. How is a value returned from C/C++?
4. What about function overloading, since C does not allow it?
5. How do we differentiate between static and non-static members?
6. What about strings and arrays?
7. How can the C/C++ code throw an exception, and what happens if it invokes a Java
method that throws an exception?
We will begin our discussion by first implementing a single native method. To avoid all of
the above complications, our method will be static, with no parameters, and no return type, and
simply print a string. Still, this is tricky, since it is the first use of various incantations that will

c++book.mif Page 251 Tuesday, April 29, 2003 2:13 PM

JNI Basics

251

be part of all JNI implementations. Then we will access fields and invoke methods of an object,
discuss arrays and strings, have the native method return a value and throw an exception, and
then quickly examine some the JNI support for object monitors.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17

class HelloNative
{
native public static void hello( );
static
{
System.loadLibrary( "HelloNative" );
}
}
class HelloNativeTest
{
public static void main( String[ ] args )
{
HelloNative.hello( );
}
}

Figure 13-1

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20

Declaration of a native method

/* DO NOT EDIT THIS FILE - it is machine generated */

#include <jni.h>
/* Header for class HelloNative */
#ifndef _Included_HelloNative
#define _Included_HelloNative
#ifdef __cplusplus
extern "C" {
#endif
/*
* Class:
HelloNative
* Method:
hello
* Signature: ()V
*/
JNIEXPORT void JNICALL Java_HelloNative_hello( JNIEnv*, jclass );
#ifdef __cplusplus
}
#endif
#endif

Figure 13-2

Result of javah for HelloNative (yields HelloNative.h)

c++book.mif Page 252 Tuesday, April 29, 2003 2:13 PM

252

Chapter 13 Using Java and C++: The JNI

13.2 Implementing A Simple Parameter-less, Return-less Method

We begin with the easy stuff. Figure 13-1 shows a class, HelloNative, that contains a single native method, hello. As we see at line 3, declaring a native method is trivial on the Java
side. In the test program, at line 15, we invoke the native method, as we would any Java method.
Because the method is native, the Java Virtual Machine will look for it amongst the
dynamic libraries that have been loaded prior to the call. Thus, at line 7, we load the library that
will contain the implementation of the native method. If the dynamic library that contains the
implementation of the native method has not been loaded prior to invoking the method, you will
get an UnsatisfiedLinkError. You will also get this exception if you provide the wrong
name for the library, or if the library is not located in the directories specified in the Java property java.library.path. Because it is important that the dynamic library be loaded prior
to invoking the native method, the library is typically loaded in the static initializer block of the
class that declares the native method.
The code in Figure 13-1 compiles. It generates an UnsatisfiedLinkError because
the dynamic link library does not yet exist.
Next, we have to write the C++ code. Although the name of the method is hello, we
cannot expect to implement a C++ function named hello. After all, there could be several
methods named hello in different classes or packages, or even an overloaded hello (and not
all languages support overloading). The JNI specifies a complex algorithm that generates a function name based on the complete class name, method name, and parameter types of the Java
native method. The complete class name includes the package name if appropriate, and the
parameter types are considered in the name only if there is function overloading in the same
class.
Fortunately, you do not have to know what this complicated algorithm is. Instead, the
javah utility can be used to generate a C++ header file from a Java class. Note that the Java
class must already be compiled. The header file will list all of the function declarations that you
must provide to implement the native methods declared in the class. In our example, we run
javah HelloNative

and the output is shown in Figure 13-2. We can see that the name of the method is
Java_HelloNative_hello, which reflects the (default) package, class, and method name.
1
2
3
4
5
6
7
8
9

#include <iostream>
using namespace std;
#include "HelloNative.h"
JNIEXPORT void JNICALL
Java_HelloNative_hello( JNIEnv *env, jclass cls )
{
cout << "Hello world" << endl;
}

Figure 13-3

Implementation of HelloNative.cpp

c++book.mif Page 253 Tuesday, April 29, 2003 2:13 PM

Implementing A Simple Parameter-less, Return-less Method

253

The C++ declaration lists two parameters. The first, is a pointer to a JNIEnv object, and
will be used extensively to access fields and methods of objects. The second parameter is a
jclass object, representing information about the HelloNative class. Java programmers
who are familiar with the Reflection API will recognize this as being the equivalent of a Class
object. A jclass object allows us to obtain information and use fields and methods for any
Java class. This second parameter is passed for static methods only. For instance methods, the
second parameter is a jobject object, representing the moral equivalent of the this reference. Given a jobject, one can always obtain the jclass object representing the objects
class type, and at that point, the objects fields can be manipulated, and methods can be invoked.
But more on that in Section 13.4.
Since the native method takes no parameters, the C++ declaration lists no additional
parameters after the first two. We can implement the method trivially, as shown in Figure 13-3.1
You should make sure to include the header file generated by javah, and to give names to the
formal parameters. It is standard to name the first parameter env; the second parameter is often
either cls or ths, depending on whether we are implementing a static or instance method. Do
not use class or this, since these are C++ reserved words (avoid them if you are using C,
too, in case you want to painlessly upgrade to C++ later on).
13.2.1 Compiling a Shared Library
All that is left to do is to compile the C++ code into a shared library. How this is done depends
on your environment. Observe first, that the header file has an include directive for jni.h.
Thus is not part of Standard C++, but instead is part of the JDK.
If the JDK is installed in directory JAVA_HOME, then the header file is in subdirectory
named include of JAVA_HOME. jni.h itself includes a second file, and this second file is
found in a subdirectory of include. This is important, because compiler options must be provided to search for header files in these nonstandard places.
Under Windows, the shared library must end in .dll, but the extension is not part of the
string passed to System.loadLibrary. Under Unix, the shared library must end with .so,
begin with lib, but neither the extension nor prefix is part of the string passed to
System.loadLibrary.
Using Visual Studio 6.0 or Visual Studio DotNet, we can compile the shared library into a
DLL either by using a DLL project (with the include search path augmented as specified above,
and an option to specify the target directory for the DLL file), or by using the command-line
tools. Visual Studio DotNet provides a command line window; in Visual Studio 6.0, after opening an MS/DOS window, find and run vcvars32.bat to set up command-line compilation.
Then command-line compilation can be used. The magic incantation for our example (all on one
line, of course), assuming that the JDK is in C:\jdk is
1.

Note that printing to the standard output from a native method is a bad plan, because writing to standard output
in both C++ and Javas System.out.println could yield intermixed output, due to buffering.

c++book.mif Page 254 Tuesday, April 29, 2003 2:13 PM

254

Chapter 13 Using Java and C++: The JNI

cl -GX /GR -IC:\jdk\include -IC:\jdk\include\win32

HelloNative.cpp -LD -FeHelloNative.dll

The first two options suppress complaints about exception handling not being enabled,
and are not needed to compile C++ programs.
In a Unix world, we can use g++. For g++, we first compile the file into a single .o, and
then create the shared library. Thus, we have two commands (of course the first two lines should
be typed as one):
g++ -c -fPIC -I$JAVA_HOME/include
-I$JAVA_HOME/include/linux HelloNative.cpp
g++ -shared -o libHelloNative.so HelloNative.o

Using g++ on Solaris, replace linux with solaris in the first command.
If you are using the standard Sun C++ compiler, one command compiles into the shared
library:
CC -G -I$JAVA_HOME/include -I$JAVA_HOME/include/solaris
HelloNative.cpp -o libHelloNative.so

For other platforms, consult local documentation. Once we have compiled the shared
library, the Java program should run as expected. If an UnsatisfiedLinkedError is still
being reported, verify that the dynamic library is correctly named, as specified above, and that it
is in a viewable directory. On Unix, make sure the dynamic library is located either in the directory that is current when the Virtual Machine is invoked, or in a directory listed in environment
variable LD_LIBRARY_PATH. On Windows, make sure the dynamic library is located either in
the directory that is current when the Virtual Machine is invoked, or in a directory listed in environment variable PATH.

13.3 JNI Types

The header file jni.h defines, a host of types, all of which start with the letter j.
Each of the eight Java primitive types has a corresponding C++ type. Thus, these types are
jbyte, jshort, jint, jlong, jfloat, jdouble, jchar, and jboolean. Constants
JNI_TRUE and JNI_FALSE are also defined. typedef statements are used to define these
types. For instance, on a 64 bit machine, a jlong might very well be typedefed as an int.
These types should be used in implementations of native methods instead of standard C++ types
such as int and double, to avoid writing code that depends on the particular representation of
C++ types. The JNI also defines the type jsize, which is the equivalent of C++s size_t,
representing the type that can be used to index an array.
Two Java class types also have a C++ type. These C++ types are jstring and
jobject, representing String and Object. The C++ programmer should not assume any
underlying implementation of jstring; specifically, routines are provided to create a char *
from a jstring object, and the char * can then be manipulated. Similarly, a jstring can
be created from a const char * by invoking a JNI routine. Array objects in Java are represented by corresponding C++ JNI types. These types include arrays of all types of primitives,

c++book.mif Page 255 Tuesday, April 29, 2003 2:13 PM

Accessing Java Objects

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

255

class Date
{
public Date( int m, int d, int y )
{ month = m; day = d; year = y; }
static
{
System.loadLibrary( "Date" );
}
native public void printDate( );
public int
{ return
public int
{ return
public int
{ return

getMonth( )
month; }
getDay( )
day; }
getYear( )
year; }

public String toString( )

{ return month + "/" + day + "/" + year; }
private int month;
private int day;
private int year;
}

Figure 13-4

Date class with a native method for printDate

such as jintArray, and jobjectArray. As was the case with strings, routines are provided
to create a jint * from a jintArray object. However, as we discuss in Section 13.5.2, we
cannot create a jobject * from a jobjectArray.
Finally, two objects are used to represent information about a method and a field. These
are jmethodId and jfieldID, which are the moral equivalents of Method and Field in
the Reflection API.

13.4 Accessing Java Objects

To see how Java objects can be manipulated by native code, we provide a Date class in
Figure 13-4 that is completely implemented except for a single method, printDate. Of
course, printDate could easily be implemented in Java too, but instead, we have declared it
as native.
Clearly an implementation of instance method printDate requires access to either the
instance data, or possibly an invocation of some of the Date instance methods. In this section
we provide three implementations of printDate. The first implementation accesses the data
members. The second implementation invokes the three accessors, getMonth, getDay, and

c++book.mif Page 256 Tuesday, April 29, 2003 2:13 PM

256

Chapter 13 Using Java and C++: The JNI

1
2
3
4
5
6
7
8

class TestDate
{
public static void main( String[ ] args )
{
Date d = new Date( 8, 23, 2003 );
d.printDate( );
}
}

Figure 13-5

Test program for Date class

getYear. The third implementation invokes the toString method, and illustrates how a Cstyle string is extracted from a jstring. The test program that we will use is shown in
Figure 13-5, for completeness.
13.4.1 Accessing Fields
Figure 13-6 shows the header file that is obtained as a result of running javah on class Date.
Observe that the second parameter at line 15 is of type jobject, signalling that we are implementing an instance method.
In order to access fields of a jobject we need to follow the following steps:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20

/* DO NOT EDIT THIS FILE - it is machine generated */

#include <jni.h>
/* Header for class Date */
#ifndef _Included_Date
#define _Included_Date
#ifdef __cplusplus
extern "C" {
#endif
/*
* Class:
Date
* Method:
printDate
* Signature: ()V
*/
JNIEXPORT void JNICALL Java_Date_printDate( JNIEnv *, jobject );
#ifdef __cplusplus
}
#endif
#endif

Figure 13-6

Date.h, generated by javah

c++book.mif Page 257 Tuesday, April 29, 2003 2:13 PM

Accessing Java Objects

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

257

class Date extends java.lang.Object {

private int month;
/*
I
*/
private int day;
/*
I
*/
private int year;
/*
I
*/
static {};
/*
()V
*/
public Date(int,int,int);
/*
(III)V
*/
public int getDay();
/*
()I
*/
public int getMonth();
/*
()I
*/
public int getYear();
/*
()I
*/
public native void printDate();
/*
()V
*/
public java.lang.String toString();
/*
()Ljava/lang/String;
*/
}

Figure 13-7

Result of running javap -s -private on class Date

1. Obtain a jclass representing the class information for the object

2. Obtain a jfieldID representing field information for a particular field in the jclass.
To do this, we need to know that name of the field, and its type.
3. Invoke a get method, passing a jfieldID and the jobject to obtain the field for a particular object.
The trickiest step in the process is obtaining the jfieldID. The name of the field is easy
enough, but the type is difficult, because it can be any arbitrary type, such as array of some other
class type. Once again, JNI specifies a coding mechanism for types, and also for method signatures. Although the algorithm is relatively short, it can be tedious to do by hand. Fortunately,
Java provides a program to generate all the field types and method signatures for you. To do so,
invoke javap with the name of the class that you are interested. Once again, the class must
already have been compiled. javap requires two options to get the complete listing, so the
command for our example as
javap -s -private Date

The output of javap is shown in Figure 13-7. As we can see, int is represented as I. On
the other hand, String is Ljava/lang/String;, and omitting the L or the ; will give
incomprehensible runtime errors. Thus it is best to use javap.

c++book.mif Page 258 Tuesday, April 29, 2003 2:13 PM

258

Chapter 13 Using Java and C++: The JNI

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

#include "Date.h"
#include <iostream>
using namespace std;
JNIEXPORT void JNICALL
Java_Date_printDate( JNIEnv * env, jobject ths )
{
jclass cls = env->GetObjectClass( ths );
jfieldID monthID = env->GetFieldID( cls, "month", "I" );
jfieldID dayID =
env->GetFieldID( cls, "day", "I" );
jfieldID yearID = env->GetFieldID( cls, "year", "I" );
jint m = env->GetIntField( ths, monthID );
jint d = env->GetIntField( ths, dayID );
jint y = env->GetIntField( ths, yearID );
cout << m << "/" << d << "/" << y << endl;
}

Figure 13-8

Implementation #1 of printDate: access fields

The signature of a method is given by an open parenthesis, followed by a concatenated list

of parameter type encodings, followed by a close parenthesis, and an encoding for the return
type. Thus getYear is ()I, printDate is ()V (void return type), and toString is
()Ljava/lang/String;. Arrays of string are encoded as [Ljava/lang/String;.
With these preliminaries out of the way, Figure 13-8 illustrates the native implementation
of printDate that accesses the Date fields. The coding is straightforward, albeit torturous.
At line 8, we get the jclass entity representing ths. This contains information about
class Date. Throughout the code, we see invocation of member functions in JNIEnv. We call
these the environment member functions. Almost every function that is provided by the JNI is
invoked the same way, through the env pointer. Once we have a jclass, we can get
jfieldIDs for the three data fields at lines 10 to 12, by calling GetFieldID, passing the
class, field name, and field type. Then at lines 14 to 16, we can access the field of an object by
specifying the object and the jfieldID. Finally, we can print the date at line 18.
There are a host of GetXXXField methods, for each of the eight primitive types, and
also GetObjectField for Object. Strings are accessed by GetObjectField, and then
downcast to a jstring. At that point, the jstring is manipulated using the technique we
will discuss in Section 13.5. Fields (even final fields!) can also have their values changed, by
invoking SetXXXField methods.
Static fields are treated differently from instance fields. We use GetStaticFieldID
instead of GetFieldID, and we call GetStaticXXXField, passing a jclass entity cls
instead of the jobject entity ths. This is needed because some languages, such as C, do not
support overloading, so they cannot have identically named functions one of which accepts a

c++book.mif Page 259 Tuesday, April 29, 2003 2:13 PM

Accessing Java Objects

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20

259

#include "Date.h"
#include <iostream>
using namespace std;
JNIEXPORT void JNICALL
Java_Date_printDate( JNIEnv * env, jobject ths )
{
jclass cls = env->GetObjectClass( ths );
jmethodID getMonthID, getDayID, getYearID;
getMonthID = env->GetMethodID( cls, "getMonth", "()I" );
getDayID
= env->GetMethodID( cls, "getDay", "()I" );
getYearID = env->GetMethodID( cls, "getYear", "()I" );
jint m = env->CallIntMethod( ths, getMonthID );
jint d = env->CallIntMethod( ths, getDayID );
jint y = env->CallIntMethod( ths, getYearID );
cout << m << "/" << d << "/" << y << endl;
}

Figure 13-9

Implementation #2 of printDate: invoke instance methods

jclass, and the other of which accepts a jobject. Similarly, SetStaticXXXField

changes the value of a static field.
We remark that accessing String fields is done with GetObjectField, and the
resulting jobject must be typecast to a jstring. In modern C++, one would expect that this
could be done with dynamic_cast. However, to save space, in C++, jobject does not declare
any virtual methods; so dynamic_cast does not work. An alternate is to use
reinterpret_cast, but we will use the old-style cast, mostly to avoid excessive typing in
these examples.
13.4.2 Invoking Methods
Invoking a method is done in a manner that is similar to accessing (or setting) fields. First, we
need the jclass entity that represents the class that the method is part of. Then we get a
jmethodID, representing the method. To obtain a jmethodID, we need the jclass entity,
the name of the method, and the signature. Thus this step is logically the same as obtaining a
jfieldID. Finally, we can invoke the method. As in the case of the GetXXXField methods,
there are a host of different versions of functions to invoke methods:
XXX CallXXXMethod( jobject ths, jmethodID m, ... );
XXX CallStaticXXXMethod( jclass cls, jmethodID m, ... );

where ... represents the parameters to the method. In these calls, XXX represents the (Java)
return type of the method.

c++book.mif Page 260 Tuesday, April 29, 2003 2:13 PM

260

Chapter 13 Using Java and C++: The JNI

Figure 13-9 provides our second implementation of printDate and shows how instance
methods are invoked. As before, we get a jclass entity, and then at lines 11 to 13, we obtain
the jmethodIDs for each of the methods that we want to invoke (the jmethodID variable
declarations are all together on line 9 simply to avoid making lines 11 to 13 too long). Once we
have the jmethodIDs, we can use them to invoke the method, as shown at lines 15 to 17. As
with accessing instance fields, the first parameter is the object on which the method is to be
invoked. If this method took additional parameters, they would follow the jmethodIDs in the
parameter list.
Invocation of CallXXXMethod uses dynamic dispatch. An alternative is
XXX CallNonvirtualXXXMethod( jobject ths, jmethodID m, ... );

that does not use dynamic dispatch, but that is hardly ever something you would want to do.
13.4.3 Invoking Constructors
A constructor can be invoked by using environment method NewObject, and passing the
jclass, jmethodID, and parameters. If you already have a jclass entity, you can get a
jclass by invoking FindClass, passing the name. The jmethodID uses "<init>" as
the name of the method. As an example, to create a Date inside of a native method:
jclass dateClass = env->FindClass( "Date" );
jmethodID ctor = env->GetMethodID( dateClass, "<init>", "(III)V" );
jobject d1 = env->NewObject( dateClass, ctor, 8, 23, 2004 );

13.5 Strings and Arrays

The jstring and various array objects such as jintArray cannot be used directly by C++
code. Instead, environment functions must be used to create a primitive string or array from the
appropriate JNI type. Once this is done, we must remember to release the memory prior to
returning. Strings and arrays behave similarly, so we discuss strings first.
13.5.1 Strings
To obtain a primitive const char * from a jstring, we invoke environment function
GetStringUTFChars. The parameters to GetStringUTFChars are the jstring and
the address of a jboolean variable. If this address is not NULL, the Boolean variable will be
set to true if the char * is a copy of the original (which is to be expected, since Java uses 16bit Unicode, and the char * is an 8-bit UTF), and false if somehow a copy was not made.
Since it rarely matters whether or not a copy was made, usually NULL is passed as a second
parameter.
When we are done with the const
char
* that has been produced by
GetStringUTFChars by allocation from the memory heap, we must release them by invoking ReleaseStringUTFChars. Failure to do so leads to a memory leak.

c++book.mif Page 261 Tuesday, April 29, 2003 2:13 PM

Strings and Arrays

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

261

#include "Date.h"
#include <iostream>
using namespace std;
JNIEXPORT void JNICALL
Java_Date_printDate( JNIEnv * env, jobject ths )
{
jclass cls = env->GetObjectClass( ths );
jmethodID toStringID = env->GetMethodID( cls, "toString",
"()Ljava/lang/String;" );
jstring str = (jstring) env->CallObjectMethod( ths,
toStringID );
const char *c_ret = env->GetStringUTFChars( str, NULL );
cout << "(calling toString) " << c_ret << endl;
env->ReleaseStringUTFChars( str, c_ret );
}

Figure 13-10 Implementation #3 of printDate: invoke toString method

Figure 13-10 which is our third implementation of printDate, illustrates how we can
access a String. Here printDate invokes the toString method, extracts the primitive
string from the jstring that is returned by toString, and then prints it. Lines 10 and 11
gets the jmethodID for toString, and then line 13 invokes the method (again with the oldstyle cast, to save typing). Simply printing the jstring does not work. If we try it on a Win1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17

JNIEXPORT jstring JNICALL Java_StringAdd_add

( JNIEnv *env, jclass cls, jstring a, jstring b )
{
const char *a1 = env->GetStringUTFChars( a, NULL );
const char *b1 = env->GetStringUTFChars( b, NULL );
char *c = new char[ strlen( a1 ) + strlen( b1 ) + 1 ];
strcpy( c, a1 );
strcat( c, b1 );
jstring result = env->NewStringUTF( c );
env->ReleaseStringUTFChars( a, a1 );
env->ReleaseStringUTFChars( b, b1 );
delete [] c;
return result;
}

Figure 13-11 Using strings (primitive style) for static native method StringAdd.add

c++book.mif Page 262 Tuesday, April 29, 2003 2:13 PM

262

Chapter 13 Using Java and C++: The JNI

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

JNIEXPORT jstring JNICALL Java_StringAdd_add

( JNIEnv *env, jclass cls, jstring a, jstring b )
{
const char *a1 = env->GetStringUTFChars( a, NULL );
const char *b1 = env->GetStringUTFChars( b, NULL );
string c = a1;
c += b1;
jstring result = env->NewStringUTF( c.c_str( ) );
env->ReleaseStringUTFChars( a, a1 );
env->ReleaseStringUTFChars( b, b1 );
return result;
}

Figure 13-12 Using C++ string library for static native method StringAdd.add

dows platform, we get 008D9E3C; other nonsense is produced on Unix machines. In fact, we
are lucky not to crash the Virtual Machine. Instead, at line 16 we get a null-terminated primitive
string, print it, and then free the C-style string at line 18.
New jstring objects can be created (mostly for the purposes of returning one) by
invoking NewStringUTF. As an example, the code in Figure 13-11 shows a routine that
implements a string concatenation in native code. Once again, this is a silly example, but illustrates the syntax.
As we can see by lines 1 and 2, we are implementing method add in class StringAdd.
This is a static method, since the second parameter is a jclass. The two parameters are both
String, and the return type is String. At lines 4 and 5 we obtain the UTF string, and at line
6, we allocate an array that will store the result of the concatenation. The +1 is needed to provide
space for the null terminator. At lines 8 and 9, we compute the result of concatenation by first
copying a1 to c, and then appending b1 to c. Then we call NewStringUTF at line 10 to form
a jstring that can be returned at line 16. Prior to returning, we must clean up memory. Lines
12 and 13 show the calls to ReleaseStringUTFChars, and line 14 cleans up the call to
new[] with a matching delete[].
A cleaner alternate that avoids the calls to new[] and delete[] by using the C++
string library type is shown in Figure 13-12.
13.5.2 Arrays
The JNI defines eight array types for the primitives. A typical example is jintArray. Additionally, jobjectArray is used to represent an array of Object. The environment function
GetArrayLength can be used get the length of any array object, passed as a parameter. To
access individual items in the array, we have to use one strategy for primitives, and another for
objects.

c++book.mif Page 263 Tuesday, April 29, 2003 2:13 PM

Strings and Arrays

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

263

class NativeSumDemo
{
native public static double sum( double [ ] arr );
static
{
System.loadLibrary( "Sum" );
}
public static void main( String [ ] args )
{
double [ ] arr = { 3.0, 6.5, 7.5, 9.5 };
System.out.println( sum( arr ) );
}
}

Figure 13-13 Declaration of native method that accepts an array as parameter

To access an array of primitive types, we can obtain a primitive array using the same idioms as was seen for strings. As an example, Figure 13-13 declares a native method, sum, that
returns the sum of the elements in an array of double. It also contains a test program.
When we run javah, we obtain the header file shown in Figure 13-15. There we see that
the double[] parameter in the Java native declarations becomes a jdoubleArray. Implementation of sum is shown in Figure 13-14. The sum is declared as a jdouble (rather than a

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

#include "NativeSumDemo.h"
JNIEXPORT jdouble JNICALL Java_NativeSumDemo_sum
( JNIEnv *env, jclass cls, jdoubleArray arr )
{
jdouble sum = 0;
jsize len = env->GetArrayLength( arr );
// Get the elements; don't care to know if copied or not
jdouble *a = env->GetDoubleArrayElements( arr, NULL );
for( jsize i = 0; i < len; i++ )
sum += a[ i ];
// Release elements; no need to flush back
env->ReleaseDoubleArrayElements( arr, a, JNI_ABORT );
return sum;
}

Figure 13-14 Implementation of NativeSumDemo.sum native method

c++book.mif Page 264 Tuesday, April 29, 2003 2:13 PM

264

Chapter 13 Using Java and C++: The JNI

double) and initialized at line 6. At line 7, we obtain the length of the array by calling
GetArrayLength.
Line 10 gets a C-style array from the jdoubleArray. As was the case with strings, this
array could be a copy, or it could be the original, depending mostly on the implementation of the
Virtual Machine. With arrays, if double and jdouble have identical representations, and if
the Java array is stored contiguously (which is not guaranteed, since the garbage collector may
elect to move parts of the array to reduce fragmentation), then it is possible that the pointer variable is actually pointing to the memory that stores the original double[] inside the virtual
machine. In such a case, no copy is made. However, once this pointer is handed out, the garbage
collector could not safely move parts of the array without invalidating the pointer. Thus, if no
copy is made, the original array is pinned and cannot relocate until the pointer is released.
Once we have the C-style array, we can compute its sum. Prior to returning, we must
release the array, as shown at line 16. When we were working with strings, releasing the string
returned the memory back to the system. With arrays, there are two independent issues.
First, if the array is a copy, we must copy any changes back to the original; otherwise they
are not reflected. Second, if the array is a copy, we must have memory reclaimed. As a result, the
last parameter to ReleaseXXXArrayElements can be either of 0, JNI_COMMIT, or
JNI_ABORT. If the parameter is 0, we flush the contents back to the original, the reflecting all
changes, and then reclaim the memory if needed. JNI_COMMIT flushes the contents, but does
not reclaim the memory. This can be useful if changes need to be reflected immediately, but
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21

/* DO NOT EDIT THIS FILE - it is machine generated */

#include <jni.h>
/* Header for class NativeSumDemo */
#ifndef _Included_NativeSumDemo
#define _Included_NativeSumDemo
#ifdef __cplusplus
extern "C" {
#endif
/*
* Class:
NativeSumDemo
* Method:
sum
* Signature: ([D)D
*/
JNIEXPORT jdouble JNICALL Java_NativeSumDemo_sum
(JNIEnv *, jclass, jdoubleArray);
#ifdef __cplusplus
}
#endif
#endif

Figure 13-15 Result of javah for NativeSumDemo (yields NativeSumDemo.h)

c++book.mif Page 265 Tuesday, April 29, 2003 2:13 PM

Using Java Exceptions

265

would not be needed if no copy was made (hence the second parameter to
GetXXXArrayElements could be useful to decide if this call should be made). JNI_ABORT
does not flush the contents, but reclaims the memory if needed. JNI_ABORT would be useful if
no changes were made to the array, because then we would avoid the flushing that would be
done with a parameter of 0. In fact, this is exactly the case we have, so we call
ReleaseXXXArrayElements with JNI_ABORT as the parameter.
Accessing elements in a jobjectArray is more difficult because we cannot obtain the
C-style equivalent. Instead, we must call
GetObjectArrayElement( array, idx );
SetObjectArrayElement( array, idx, val );

through the env pointer. Clearly this makes accessing arrays of objects fairly slow.
New arrays can be created by using
NewObjectArray( jclass cls, int len, jobject default );
NewXXXArray( int len, XXX default );

through the env pointer; the return type is the appropriate jarray type.

13.6 Using Java Exceptions

Java exceptions come into play in two basic ways.
First, the native method might have a throws list, so that the native code can throw an
exception if things go wrong. So how does a native method throw an exception?
Second, the native method might call a Java method, and the Java method throws an
exception. Can the native method determine that an exception has occurred and either handle it,
or let it propagate cleanly?
Both problems have relatively simple solutions in the JNI. Keep in mind that we cannot
assume that the target language has support for exceptions. In fact C does not, and although C++
has exceptions, the exception handling mechanism in C++ is not currently compatible with Java
exceptions.
Instead, an exception is signalled in native code simply by setting a flag. If a native
method calls a Java method and the Java method throws an exception, then when the Java
method returns back to the native method, by invoking an environment function
ExceptionOccurred, the native method can determine if an exception is pending. If so, it
can either return, in which case, the Virtual Machine will propagate the exception, or it can call
ExceptionClear to clear the exception, if it has a reasonable way of handling it.
The first case is most typical, resulting in code such as:
...
jthrowable e = env->ExceptionOccurred( );
if( e != NULL )
return;
// give up
...

c++book.mif Page 266 Tuesday, April 29, 2003 2:13 PM

266

Chapter 13 Using Java and C++: The JNI

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27

class NativeSumDemo
{
native public static double sum( double [] arr )
throws Exception;
static
{
System.loadLibrary( "Sum" );
}
public static void main( String[] args )
{
double [ ] arr1 = { 3.0, 6.5, 7.5, 9.5 };
double [ ] arr2 = { };
try
{
System.out.println( sum( arr1 ) );
System.out.println( sum( arr2 ) );
}
catch( Exception e )
{
System.out.println( "Caught the exception!" );
e.printStackTrace( );
}
}
}

Figure 13-16 Same main that illustrates exception being thrown by native call

If the native code does neither, and instead continues executing native code, even though
an exception is pending, then the behavior when other environment functions are called is undefined and dangerous.
If the native method needs to throw an exception on its own, then it can do so using the
environment functions Throw or ThrowNew. The result of calling either function is that an
exception is now pending; however, as before, the pending exception does not terminate the
native method. Instead, a return statement should immediately follow, and certainly other environment functions should not be called. So if it is important to release array elements, do so
before invoking Throw or ThrowNew.
ThrowNew is much easier to use than Throw because you can simply give the complete
name of the exception class and the parameter that is passed to its constructor. We illustrate how
a native method can throw an exception in Figure 13-16, which provides a main and,
Figure 13-17, which implements the native method itself.
In Figure 13-16, we see that method sum might throw an Exception; the exception is
thrown if the array has length 0. (This is terrible style, and a better exception should be used, in

c++book.mif Page 267 Tuesday, April 29, 2003 2:13 PM

Using Java Exceptions

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30

267

#include <iostream>
using namespace std;
#include "NativeSumDemo.h"
JNIEXPORT jdouble JNICALL Java_NativeSumDemo_sum
( JNIEnv *env, jclass cls, jdoubleArray arr)
{
jdouble sum = 0;
jsize len = env->GetArrayLength( arr );
if( len == 0 )
{
env->ThrowNew( env->FindClass( "java/lang/Exception" ),
"Empty array" );
cout << "Throwing an exception, but should exit" << endl;
return 0.0;
}
// Get the elements; don't care to know if copied or not
jdouble *a = env->GetDoubleArrayElements( arr, NULL );
for( jsize i = 0; i < len; i++ )
sum += a[ i ];
// Release elements; no need to flush back
env->ReleaseDoubleArrayElements( arr, a, JNI_ABORT );
return sum;
}

Figure 13-17 Implementation of NativeSumDemo.sum native method that throws exception

general; however using Exception illustrates a point as we will see when the native method is
implemented). Note also that the library loaded at line 8 is Sum, rather than NativeSumDemo.
When you compile the example, make sure the appropriate link library is created. The rest of the
code is standard fare, and we expect that the second call to sum triggers the catch block.
Figure 13-17 Illustrates the implementation of the native method. First, note that the
throws list declared in Java is not part of the native function name at line 6. The additional code
at lines 12 to 18 shows the test for the case of a zero-length array. When this test succeeds, a new
exception is created, and marked as pending at lines 14 and 15 by the call to ThrowNew. Note
that the name of the exception must reflect the complete class name, including package name.
However, ThrowNew does not cause an immediate return, so the print statement at line 16 will
be executed.

c++book.mif Page 268 Tuesday, April 29, 2003 2:13 PM

268

Chapter 13 Using Java and C++: The JNI

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

JNIEXPORT jstring JNICALL Java_StringAdd_add

( JNIEnv *env, jclass cls, jstring a, jstring b )
{
const char *a1 = (*env)->GetStringUTFChars( env, a, NULL );
const char *b1 = (*env)->GetStringUTFChars( env, b, NULL );
char *c = (char *) malloc( strlen(a1) + strlen(b1) + 1 );
jstring result;
strcpy( c, a1 );
strcat( c, b1 );
result = (*env)->NewStringUTF( env, c );
(*env)->ReleaseStringUTFChars( env, a, a1 );
(*env)->ReleaseStringUTFChars( env, b, b1 );
free( c );
return result;
}

Figure 13-18 Using C to implement a native method

The return value at line 17 is required to avoid warnings from the C++ compiler, but is
never used by the caller, because the return immediately causes the Virtual Machine to throw a
Java exception.

13.7 Using C Instead of C++

All of our examples use C++, but C can be used with only minor cosmetic changes. There are
two main differences. Probably most importantly, you must code in C. See Section 12.5 for reasons why this is a big deal; in part, you will have a harder time dealing with memory, and you
must declare all your local variables up front, instead of at first use.
The JNI difference is that all environment functions of the form
env->EnvFunction( ... );

are rewritten as
(*env)->EnvFunction( env, ... );

which does get ugly after a while, but is a relatively minor difference. Finally, make sure you
place your source code in a .c file, since many C compilers also compile C++ code, and make
their decisions based on the suffix of the source file name.
Figure 13-18 shows a C implementation of the add method in class StringAdd that
was previously written in Figure 13-11. The difference is mostly cosmetic.

13.8 JNI References

Data types such as jobject, jclass, jstring, jmethodID, and jclassID, and the

c++book.mif Page 269 Tuesday, April 29, 2003 2:13 PM

JNI References

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21

269

JNIEXPORT jint JNICALL Java_StringStuff_totalChars

( JNIEnv *env, jclass cls, jobjectArray arr)
{
jint totalChars = 0;
jsize len = env->GetArrayLength( arr );
for( jint i = 0; i < len; i++ )
{
jstring jstr = (jstring)
env->GetObjectArrayElement( arr, i );
if( jstr == NULL )
continue;
const char *cstr = env->GetStringUTFChars( jstr, NULL );
totalChars += strlen( cstr );
env->ReleaseStringUTFChars( jstr, cstr );
env->DeleteLocalRef( jstr );
}
return totalChars;
}

Figure 13-19 Illustration of DeleteLocalRef

array types all represent references to Java objects (whereas jint, jdouble, etc. do not).
When these references are created they are called local references and are valid:
1. only in the thread in which the native method is invoked
2. only for the duration of the native method
This means that you cannot cache local references, nor pass them to other threads. If you
do so, and try to use them later on, you may find that the garbage collector has reclaimed the
objects.
A global reference is a wrapper around a local reference. Unlike local references, global
reference are valid across multiple threads and multiple native calls, until a call to
DeleteGlobalRef. If there is no call to DeleteGlobalRef, the global reference is valid
for the entire duration of the Virtual Machine, which could be a problem for general objects and
arrays, but is probably reasonable behavior for jclass references.
From this discussion, we also know that local references are valid for the duration of the
native method call. However, if the native method makes a time-consuming function call, it
might be worth releasing local references prior to making the function call. Alternatively, if the
native method call creates many local references to large objects, it might be prudent to release
some of the references when they are no longer needed. As an example, if we are iterating
through a jobjectArray, each call to GetObjectArrayElement creates a jobject. It

c++book.mif Page 270 Tuesday, April 29, 2003 2:13 PM

270

Chapter 13 Using Java and C++: The JNI

may be worth reclaiming it as we advance to the next array element. This is done with
DeleteLocalRef. Figure 13-19 illustrates this with an example in a native routine that
counts the total string length in an array of strings. We use DeleteLocalRef after accessing
each string, prior to proceeding to the next string. Depending on the underlying implementation
of the JNI, the number of strings, and the size of the string objects, this could help performance,
or simply have no noticeable effect.

13.9 Java Monitors

If a native method is declared as synchronized then the usual monitor (this for instance
method, getClass() for static methods) is obtained prior to its invocation and released on its
return. If the method is not synchronized, or insufficiently synchronized, you may need to obtain
a monitor. The environment functions MonitorEnter and MonitorExit, which take a
jobject can be used to delimit a synchronized block in native code. Thus
env->MonitorEnter( obj );
/* synchronized block */
env->MonitorExit( obj );

is equivalent to
synchronized( obj )
{
/* synchronized block */
}

Dont forget to invoke MonitorExit. Although there are no environment functions to invoke
wait and notifyAll, these can be called by obtaining a jmethodID, and calling the appropriate function using the normal JNI mechanism.

13.10Invocation API
The Invocation API allows the C++ programmer to create a Virtual Machine from inside a C++
program. Once the Virtual Machine is created, the normal mechanism can be used to invoke the
main method of any class.
The code to do so is boilerplate and is shown in Figure 13-20. Clearly it can be generalized to allow any class, and to allow command-line arguments to main. The hard part of this
code is to do the compilation. In short, in addition to providing options to specify the include
directories, you must make sure that library jvm.lib (for Windows) or jvm (for Unix) is used
in the compilation. For Unix this involves using two options: -L to specify the search path for
libraries, and -l to specify the library itself. Using the Visual Studio products, the complete
library name is included with the compilation command, and the PATH environment variable is

c++book.mif Page 271 Tuesday, April 29, 2003 2:13 PM

Invocation API

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40

271

// Invoke the main method of Hello

#include <iostream>
#include <jni.h>
using namespace std;
const char *CLASS_NAME = "Hello";
int main( int argc, char *argv[] )
{
JavaVMInitArgs vm_args;
JavaVMOption options[ 1 ];
JavaVM *vm;
JNIEnv *env;
options[ 0 ].optionString = "-Djava.class.path=.";
vm_args.options = options;
vm_args.nOptions = 1;
vm_args.version = JNI_VERSION_1_2;
int res = JNI_CreateJavaVM( &vm, (void **)&env, &vm_args );
if( res < 0 )
{
cerr << "Failed to create VM (" << res << ")" << endl;
return -1;
}
jclass cls = env->FindClass( CLASS_NAME );
// Now try to call main
jmethodID mainMethodID = env->GetStaticMethodID( cls, "main",
"([Ljava/lang/String;)V" );
jclass classString = env->FindClass( "java.lang.String" );
jobjectArray argsToMain =
env->NewObjectArray(0, classString, NULL);
env->CallStaticVoidMethod( cls, mainMethodID, argsToMain );
vm->DestroyJavaVM( );
return 0;
}

Figure 13-20 Creating a Java VM from C++ main; invokes Hello with no parameters

set to include that hotspot compiler C:\jdk\jre\bin\hotspot. The online code contains
more specific compilation instructions.

c++book.mif Page 272 Tuesday, April 29, 2003 2:13 PM

272

Chapter 13 Using Java and C++: The JNI

13.11Key Points
Java native methods are specified with the native reserved word.
After the class is compiled, we can run javah to generate a C/C++ header file.
After the implementation is written in the native language, it must be compiled into a
dynamic library.
The library should be loaded by System.loadLibrary, typically invoked from the
static initializer of the native methods class.
The eight Java primitives have corresponding native types, such as jint.
JNI_TRUE and JNI_FALSE are defined to represent true and false, respectively.
jstring and jobject represent String and Object.
jXXXArray is used to represent the eight primitive Java arrays.
jobjectArray represents an array of Object.
jclass represents a Java class type.
jmethodID represents a method.
jfieldID represents a field.
The Java native method is implemented by a native function whose name incorporates the
package name, class name, and possibly signature of the Java native method.
The first parameter to a native function is the environment pointer.
The second parameter is a jclass, representing the class type for static methods, or a
jobject, representing this, for instance methods.
Additional parameters will be listed in the native function, as declared in the Java native
method.
javap is used to obtain a list of encoded field type signatures and method signatures.
An instance field of an object is accessed by getting a jfieldID from the class type,
field name, and (encoded) field type, and then invoking GetXXXField, with a jobject
and jfieldID. Static fields can be accessed by GetStaticFieldID and
GetStaticXXXField, with a jclass in place of a jobject. Fields can also be
changed with SetXXXField and SetStaticXXXField.
An instance method of an object is invoked using dynamic dispatch by getting a
jmethodID from the class type, method name, and (encoded) method signature, and
then invoking CallXXXMethod, with a jobject, jmethodID, and parameters to the
method. Static methods can be invoked by GetStaticMethodID and
CallStaticXXXMethod, with a jclass in place of a jobject. Methods can also
be invoked without dynamic dispatch by using CallNonvirtualXXXMethod.
A C-style string (const char *) can be extracted from a jstring by invoking the
environment function GetStringUTFChars. ReleaseStringUTFChars should
be called when the C-style string is no longer needed.
A C-style primitive array of primitive types, such as jint * can be extract from a
jarray (such as jintArray) by invoking the environment function

c++book.mif Page 273 Tuesday, April 29, 2003 2:13 PM

Exercises

273

GetXXXArrayElements. The corresponding ReleaseXXXArrayElements

should be called when the array is no longer needed, and also if flushing of data back to
the Java VM is required.
Elements in a jobjectArray can only be accessed individually through environment
functions GetObjectArrayElement and SetObjectArrayElement.
Environment function ThrowNew can be used to signal an exception from native code,
that can be sent back to the Virtual Machine. The exception is only marked in the native
code, and control does not automatically exit the native method.
Environment function ExceptionOccurred can be used to test if an exception is
active. If so, it is best to leave the native method as soon as possible.
C can be used instead of C++, with mostly cosmetic differences in coding required.
A local reference is valid only for the duration of the native method, and only in the thread
of the invoking native method.
A global reference is created by environment function NewGlobalRef, and wraps an
existing (local) reference. It is valid until a DeleteGlobalRef is invoked on it.
MonitorEnter and MonitorExit are environment functions that can be used to create a synchronized block in native code.
The Invocation API is boilerplate code that allows the creation of a Virtual Machine from
inside C++ code. Once the VM is running, the main method of a class can be invoked
using the usual techniques.

13.12Exercises
1. Why would it make sense to write a native method?
2. On the Java side, how is a native method declared?
3. On the Java side, how is the library that contains the native method implementation
loaded?
4. For a class that contains native methods, how is the corresponding C++ header file generated?
5. A C++ method is already implemented in a library and we want to be able to invoke it
from Java. What is the typical strategy that is used?
6. What are the first two parameters in the signature of a C++ native method implementation?
7. What is a jclass?
8. What is a jobject?
9. What are the JNI primitive types?
10. Explain how an instance field of a Java object is accessed in C++ code. How are static
fields accessed?
11. Explain how an instance method of a Java object is invoked in native code. How are static
methods invoked?

c++book.mif Page 274 Tuesday, April 29, 2003 2:13 PM

274

Chapter 13 Using Java and C++: The JNI

12. How are constructors for new Java objects invoked in C++ code?
13. Why cant a jstring be safely type-cast to a const char *? What is the correct way
to obtain a const char * from a jstring? What memory issues must be dealt with?
14. How is a jstring created from a const char *?
15. Explain how Java arrays are accessed in C++ code.
16. What does it mean for an array to be pinned? How can you tell if an array is pinned?
17. What does the last parameter to ReleaseXXXArrayElements do?
18. How can a C++ native method signal an exception?
19. How can a C++ native method tell if invoking a Java method caused an exception be
raised? Can a C++ native method handle an exception?
20. How does C native code differ from C++ native code?
21. What is a local reference, and how long are local references valid?
22. What is a global reference, and how long are global references valid?
23. How can a C++ native method obtain and release a monitor?
24. What is the invocation API?
25. Can a native method make changes to a final field? Find out by writing a program that
attempts to do so.
26. Can a native method invoke a private method of another class? Find out by writing a program that attempts to do so.
27. Class CIO, defined below is intented to allow the output of a single int, double, or
String, using C-style sprintf formatting. Each of the native methods returns a
String, in which the escape sequence in the control string is replaced by the second
parameter. The native methods are implemented by calling the C library routine sprintf. In
the call to sprintf, provide a large buffer for the first parameter (and hope for the best),
and then pass control and var as parameters to the C library sprintf.
class CIO
{
public native static String sprintf( String control, String var );
public native static String sprintf( String control, int var );
public native static String sprintf( String control, double var );
}

28. Implement the standard matrix multiplication algorithm using a native method, throwing
an exception if the matrices have incompatible sizes. The signature of your method is:
native public static double [][] multiply( double [][] a,
double [][] b );

c++book.mif Page 275 Tuesday, April 29, 2003 2:13 PM

Bibliography

he thinking behind C++ is described in [13]. The two

1,000-page gorrillas that describe C++ are [12] and [5] (with answers to the latter title provided
in [15]). The books [8] and [9] provide great tips for safe C++ programming. Advanced features
of the C++ language itself are discussed in [2], [4], [7], and [14]. A collection of answers to frequently asked C++ questions is provided in [1]. The C++ I/O library is described in great detail
in [6]. Good references for the STL include [10] and [11], as well as the 1000-page gorillas. The
classic reference for the C programming language is [3].
1. M. Cline, G. Lomow, and M. Girou, C++ FAQs, 2d ed., Addison-Wesley, Reading, Mass.,
1999.
2. J. O. Coplien, Advanced C++, Addison-Wesley, Reading, Mass., 1992.
3. B. W. Kernighan and D. M. Ritchie, The C Programming Language, 2d ed., Prentice-Hall,
Englewood Cliffs, N.J., 1997.
4. A. Koening and B. Moo, Ruminations on C++, Addison-Wesley, Reading, Mass., 1997.
5. J. Lajoie and S. Lippman, C++ Primer, 3d ed., Addison-Wesley, Reading, Mass., 1998.
6. A Langer and K. Kreft, Standard C++ IOStreams and Locales: Advanced Programmer's
Guide and Reference, Addison-Wesley, Reading, Mass., 2000.
7. S. Lippman, Essential C++, Addison-Wesley, Reading, Mass., 2000.
8. S. Meyers, Effective C++, 2d ed., Addison-Wesley, Reading, Mass., 1998.
9. S. Meyers, More Effective C++, Addison-Wesley, Reading, Mass., 1996.
10. S. Meyers, Effective STL, Addison-Wesley, Reading, Mass., 2001.
11. D. R. Musser and A. Saini, C++ Programming with the Standard Template Library, Addison-Wesley, Reading, Mass., 1996.

275

c++book.mif Page 276 Tuesday, April 29, 2003 2:13 PM

276

Bibliography

12. B. Stroustrop, The C++ Programming Language, 3d ed., Addison-Wesley, Reading,

Mass., 1997.
13. B. Stroustrop, The Design and Evolution of C++, Addison-Wesley, Reading, Mass., 1994.
14. H. Sutter, Exceptional C++, Addison-Wesley, Reading, Mass., 2000.
15. C. L. Tondo and B. P. Leung, C++ Primer Answer Book, Addison-Wesley, Reading,
Mass., 1999.

c++book.mif Page 277 Tuesday, April 29, 2003 2:13 PM

Index

277