From C++ to Objective-C
version 2.1 en
Pierre Chatelier
e-mail: pierre.chatelier@club-internet.fr
Copyright c© 2005, 2006, 2007, 2008, 2009 Pierre Chatelier
English adaptation : Aaron Vegh
Document revisions available at :
http://pierre.chachatelier.fr/programmation/objective-c.php
This document is also available in french
Ce document est aussi disponible en français
With special thanks to: For their attentive reading and many helpful comments, I would like
to thank Pascal Bleuyard, Jérôme Cornet, François Delobel and Jean-Daniel Dupas, whose
help was important in making this work the best possible. Jack Nutting, Ben Rimmington and
Mattias Arrelid have also provided many feedback. Jonathon Mah has been particularly implied
in bringing a lot of very judicious corrections.
They are not responsible of any mistake I could add after their reviewing.
1
Contents
Table of contents
2
Introduction
5
1 Objective-C and Cocoa
6
1.1 A short history of Objective-C
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
1.2 Objective-C 2.0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
2 Syntax overview
7
2.1 Keywords . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
2.2 Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
2.3 Mixing up code and declarations . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
2.4 New types and values
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
2.4.1 BOOL, YES, NO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
2.4.2 nil, Nil and id
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
2.4.3 SEL
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
2.4.4 @encode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
2.5 Organization of source code: .h and .m files, inclusion . . . . . . . . . . . . . . . .
8
2.6 Class names: why NS? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
2.7 Differencing functions and methods . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
3 Classes and objects
10
3.1 Root class, type id, nil and Nil values . . . . . . . . . . . . . . . . . . . . . . . .
10
3.2 Class declaration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
3.2.1 Attributes and methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
3.2.2 Forward declarations: @class, @protocol . . . . . . . . . . . . . . . . . . .
11
3.2.3 public, private, protected
. . . . . . . . . . . . . . . . . . . . . . . . . .
12
3.2.4 static attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
3.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
3.3.1 Prototype and call, instance methods, class methods . . . . . . . . . . . . .
12
3.3.2 this, self and super
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
3.3.3 Accessing instance variables inside a method
. . . . . . . . . . . . . . . . .
14
3.3.4 Prototype id and signature, overloading . . . . . . . . . . . . . . . . . . . .
14
3.3.5 Pointer to member function: Selector . . . . . . . . . . . . . . . . . . . . . .
15
3.3.6 Default values of parameters
. . . . . . . . . . . . . . . . . . . . . . . . . .
18
3.3.7 Variable number of arguments
. . . . . . . . . . . . . . . . . . . . . . . . .
18
3.3.8 Anonymous arguments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
18
3.3.9 Prototype modifiers (const, static, virtual, “= 0“, friend, throw) . . .
18
3.4 Messages and transmission . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
3.4.1
Sending a message to nil
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
3.4.2 Delegating a message to an unknown object . . . . . . . . . . . . . . . . . .
19
3.4.3 Forwarding: handling an unknown message . . . . . . . . . . . . . . . . . .
19
3.4.4 Downcasting
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
20
4 Inheritance
21
4.1 Simple inheritance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
4.2 Multiple inheritance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
4.3 Virtuality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
4.3.1 Virtual methods
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
4.3.2
Silent redefinition of virtual methods . . . . . . . . . . . . . . . . . . . . . .
21
4.3.3 Virtual inheritance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
4.4 Protocols
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
22
4.4.1 Formal protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
22
4.4.2 Optional methods
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
23
4.4.3
Informal protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
23
2
4.4.4 Object of type Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
24
4.4.5 Message qualifiers for distant objects . . . . . . . . . . . . . . . . . . . . . .
24
4.5 Class categories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
25
4.6 Joint use of protocols, categories, subclassing: . . . . . . . . . . . . . . . . . . . . .
26
5 Instantiation
27
5.1 Constructors, initializers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
27
5.1.1 Distinction between allocation and initialization
. . . . . . . . . . . . . . .
27
5.1.2 Using alloc and init . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
27
5.1.3 Example of a correct initializer . . . . . . . . . . . . . . . . . . . . . . . . .
28
5.1.4 self = [super init...] . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
5.1.5
Initialization failure
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
30
5.1.6
“Splitting” construction into alloc+init . . . . . . . . . . . . . . . . . . .
31
5.1.7 Default constructor : designated initializer . . . . . . . . . . . . . . . . . . .
32
5.1.8 List of initialization and default value of instance data . . . . . . . . . . . .
34
5.1.9 Virtual constructor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
34
5.1.10 Class constructors
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
34
5.2 Destructors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
34
5.3 Copy operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
35
5.3.1 Classical cloning, copy, copyWithZone:, NSCopyObject() . . . . . . . . . .
35
5.3.2 NSCopyObject()
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
36
5.3.3 Dummy-cloning, mutability, mutableCopy and mutableCopyWithZone:
. .
37
6 Memory management
39
6.1 new and delete . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
6.2 Reference counting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
6.3 alloc, copy, mutableCopy, retain, release
. . . . . . . . . . . . . . . . . . . . .
39
6.4 autorelease
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
6.4.1 Precious autorelease . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
6.4.2 The autorelease pool . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
6.4.3 Using several autorelease pools . . . . . . . . . . . . . . . . . . . . . . . . .
41
6.4.4 Caution with autorelease
. . . . . . . . . . . . . . . . . . . . . . . . . . .
41
6.4.5 autorelease and retain
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
42
6.4.6 Convenience constructor, virtual constructor
. . . . . . . . . . . . . . . . .
42
6.4.7
Setter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
44
6.4.8 Getters
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
46
6.5 Retain cycles
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
48
6.6 Garbage collector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
48
6.6.1 finalize
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
48
6.6.2 weak, strong . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
48
6.6.3 NSMakeCollectable() . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
48
6.6.4 AutoZone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
48
7 Exceptions
49
8 Multithreading
51
8.1 Thread-safety . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
8.2 @synchronized . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
9 Strings in Objective-C
52
9.1 The only static objects in Objective-C . . . . . . . . . . . . . . . . . . . . . . . . .
52
9.2 NSString and encodings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
52
9.3 Description of an object, %@ format extension, NSString to C string . . . . . . . . .
52
3
10 C++ specific features
53
10.1 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
10.2 Inlining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
10.3 Templates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
10.4 Operators overloading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
10.5 Friends . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
10.6 const methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
10.7 List of initialization in the constructor . . . . . . . . . . . . . . . . . . . . . . . . .
53
11 STL and Cocoa
54
11.1 Containers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
11.2 Iterators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
11.2.1 Classical enumeration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
11.2.2 Fast enumeration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
11.3 Functors (function objects)
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
11.3.1 Using selectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
11.3.2 IMP caching
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
11.4 Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
12 Implicit code
56
12.1 Key-value coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
56
12.1.1 Principle
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
56
12.1.2 Interception . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
57
12.1.3 Prototypes
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
57
12.1.4 Advanced features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
57
12.2 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
58
12.2.1 Use of properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
58
12.2.2 Description of properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
58
12.2.3 Properties attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
12.2.4 Custom implementation of properties
. . . . . . . . . . . . . . . . . . . . .
60
12.2.5 Syntax to access properties
. . . . . . . . . . . . . . . . . . . . . . . . . . .
60
12.2.6 Advanced details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
61
13 Dynamism
62
13.1 RTTI (Run-Time Type Information) . . . . . . . . . . . . . . . . . . . . . . . . . .
62
13.1.1 class, superclass, isMemberOfClass, isKindOfClass
. . . . . . . . . . .
62
13.1.2 conformsToProtocol
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
62
13.1.3 respondsToSelector, instancesRespondToSelector
. . . . . . . . . . . .
62
13.1.4 Strong typing or weak typing with id
. . . . . . . . . . . . . . . . . . . . .
63
13.2 Manipulating Objective-C classes at run-time . . . . . . . . . . . . . . . . . . . . .
63
14 Objective-C++
64
15 The future of Objective-C
64
15.1 The blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
64
15.1.1 Support and use cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
64
15.1.2 Syntax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
65
15.1.3 Capturing the environment . . . . . . . . . . . . . . . . . . . . . . . . . . .
65
15.1.4 __block variables
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
65
Conclusion
67
References
67
Document revisions
68
Index
69
4
Introduction
This document is designed to act as a bridge between C++ and Objective-C. Many texts exist
to teach the object model through Objective-C, but to my knowledge, none target advanced C++
developers wanting to compare these concepts with their own knowledge. At first, the Objective-C
language seems to be an obstacle rather than a boost for Cocoa programming (cf. section 1 on the
following page): it was so different that I couldn’t get into it. It took me some time to appreciate its
challenges, and understand the many helpful concepts it supplies. This document is not a tutorial,
but a quick reference, to these concepts. I hope that it will be useful in preventing a developer
from either abandoning Objective-C or misusing its features, because of a misunderstanding of
the language. This document does not claim to be a full reference, but a quick one. For detailed
explanations of a concept, be sure to read a specialized Objective-C manual [4].
Making comparisons with C# would require another document, because that language is much
closer to Objective-C than C++ can be. Thus, a C# developer would certainly learn Objective-C
faster. According to me, C# is, despite a bunch of advanced concepts, far less interesting than
Objective-C, because it gives hard access to simple Objective-C features, and the Cocoa API quality
is miles ahead of .NET. This personal opinion is not the subject of the present document.
5
1 Objective-C and Cocoa
An initial distinction should be made : Objective-C is a language, while Cocoa is a set of classes
that contribute to native MacOS X programming. Theoretically, it is possible to use Objective-C
without Cocoa: there is a gcc front-end. But under MacOS X, both are almost inseparable, as
most of the classes supplied by the language are part of Cocoa.
More precisely, Cocoa is the implementation by Apple, for MacOS X, of the OpenStep standard,
originally published in 1994. It consists of a developer framework based upon Objective-C. The
GNUstep project [6] is another implementation, which is free. Its goal is to be as portable as
possible on most Unix systems, and is still under development.
1.1 A short history of Objective-C
It is hard to give a precise date of birth for a language, owing to the fact that there is some
time between first steps, improvements, standardisation and official announcement. However, a
rough history is given in Figure 1 to get a quick look at Objective-C amongst its ancestors and
“challengers”.
1972
1978
1989
1980
1995
1998-99
1983
Smalltak-72
Smalltak-80
C
ANSI C
C 99
C++
Standard C++
C++0x draft
Objective-C
Objective-C 2.0
Objective-C++
Java
C♯
2001
C♯2 (C♯3 to come)
2005-07
Figure 1: Timeline of Java, C, C#, C++ and Objective-C
Smalltalk-80 is one the first “real” object languages. C++ and Objective-C are two different
branches that build a superset of the C language. Objective-C is very close to Smalltalk in terms
of syntax and dynamism, while C++ is much more static, with the goal of having better run-time
performance. Java targets a C++ audience, but is also very inspired by Smalltalk for its object
model. That’s why, despite this document’s title, many references are made to Java. The C#
language, developed by Microsoft, is a direct challenger to Objective-C.
Objective-C++ is a kind of merge between Objective-C and C++. It is already usable, but some
behaviours are still not perfect. The goal of Objective-C++ is to mix up the syntaxes of Objective-C
and C++ to benefit from the best features of both worlds (cf. section 14 on page 64).
1.2 Objective-C 2.0
The present document has been updated to take in account the new features of Objective-C 2.0,
which has been released alongside MacOS X10.5. Those features are deep technical improvements,
but the high-level modifications for the developers are easily enumerable. They can now use:
• a garbage-collector : cf. section 6.6 on page 48;
• properties : cf. section 12.2 on page 58;
• fast enumeration : cf. section 11.2.2 on page 55;
• new keywords @optional and @required for protocols : cf. section 4.4 on page 22;
• updated run-time Objective-C library features : cf. section 13.2 on page 63.
Each one is detailed in a specific section.
6
2 Syntax overview
2.1 Keywords
Objective-C is a superset of the C language. Like with C++, a well-written C program should be
compile-able as Objective-C, as long as it is not using some of the bad practices allowed by C.
Objective-C has only added some concepts and their associated keywords. To avoid conflicts, these
keywords begin with the @ (at) character. Here is the (short) exhaustive list: @class, @interface,
@implementation, @public, @private, @protected, @try, @catch, @throw, @finally, @end,
@protocol, @selector, @synchronized, @encode, @defs (no more documented in [4]). Objective-C
2.0 (cf. 1.2 on the preceding page) has added @optional, @required, @property, @dynamic,
@synthesize. Let us alo quote the values nil et Nil, the types id, SEL and BOOL, the boolean
values being YES et NO. At last, a few kewords are available in particular contexts, and are not
reserved outside: in, out, inout, bycopy, byref, oneway (they can be met when defining protocols
: cf. section 4.4.5 on page 24) and getter, setter, readwrite, readonly, assign, retain, copy,
nonatomic (they can be met when defining properties : cf. section 12.2 on page 58).
There is an easy confusion between the language keywords and some methods inherited from the
root class NSObject (the mother of every class, cf. section 3.1 on page 10). For instance, the similar-
looking “keywords” for memory managament, named alloc, retain, release and autorelease,
are in fact methods of NSObject. The words super and self (cf. section 3.3.1 on page 12), could
also be understood as keywords, but self is in fact a hidden parameter to each method, and super
an instruction asking the compiler to use self differently. However, the confusion between these
false keywords and the true ones will not likely prove problematic in normal use.
2.2 Comments
The comments /* . . . */ and // are allowed.
2.3 Mixing up code and declarations
Like in C++, it is possible to insert the declaration of a variable in the middle of a block of
instructions.
2.4 New types and values
2.4.1 BOOL, YES, NO
In C++, the boolean type is bool. In Objective-C, it is BOOL, which can be set to YES or NO.
2.4.2 nil, Nil and id
These three keywords are explained later in the document, but briefly:
• Every object is of type id. This is a tool for weak-typing;
• nil is the equivalent of NULL for a pointer to an object. nil and NULL should not be
interchangeable.
• Nil is the equivalent of nil for a class pointer. In Objective-C, a class is an object (it is the
instance of a meta-class).
7
2.4.3 SEL
The SEL type can store selectors values, which are method identifiers unrelated to any class instance
object. These values can be computed by a call to @selector. A selector can be used as a kind of
pointer to a method, even if it is not technically a real pointer to a function. See section 3.3.5 on
page 15 for more details.
2.4.4 @encode
For the purpose of interoperability, teh data types in Objective-C, even custom types, and proto-
types of functions or methods, can be ASCII-encoded, according to a documented format [4]. A
call to @encode(a type ) returns a C string (char*) representing the type.
2.5 Organization of source code: .h and .m files, inclusion
Like in C++, it is useful to split the code between interface and implementation for each class.
Objective-C uses .h files for headers, and .m files for the code; .mm files are used for Objective-C++
(see Section 14 on page 64). Objective-C introduces the #import directive to replace #include.
Indeed, every C header should use compilation guards to prevent multiple inclusions. This is auto-
matic when using #import. Below is a typical interface/implementation example. The Objective-C
syntax is explained later.
C++
//In file Foo.h
#ifndef __FOO_H__ //compilation guard
#define __FOO_H__ //
class Foo
{
...
};
#endif
//In file Foo.cpp
#include "Foo.h"
...
Objective-C
//In file Foo.h
//class declaration, different from
//the "interface" Java keyword
@interface Foo : NSObject
{
...
}
@end
//In file Foo.m
#import "Foo.h"
@implementation Foo
...
@end
2.6 Class names: why NS?
In this document, almost all class names begin with NS , like NSObject or NSString. The reason
is simple: they are Cocoa classes, and most Cocoa classes begin with NS since they were initiated
under NeXTStep.
It is a common practice to use a prefix to identify the origin of a class.
8
2.7 Differencing functions and methods
Objective-C is not a language with “function calls using square brackets”. This would be a legiti-
mate thought when observing code like that :
[object doSomething];
instead of
object.doSomething();
But in fact, Objective-C is a superset of C, so that functions match the same syntax and semantics
as C for declaration, implementation and call. On the contrary, methods, which do not exist in
C, have a special syntax, which includes square brackets. Moreover, the difference is not only in
the syntax, but also the meaning. This is detailed further in Section 3.2 on the next page: this
is not a method call, this is sending a message. This is not just a simple academic distinction;
it has implications on the mechanism of Objective-C. Even if it is the same regarding the source
code organization, this mechanism allows much more dynamism. For instance, it is compatible
with adding a method at run-time (cf. section 13.2 on page 63). The syntax is also more readable,
especially with nested calls (cf. section 3.3.1 on page 12).
9
3 Classes and objects
Objective-C is an object-oriented language: it manages classes and objects. Objective-C uses a
strict object model, unlike C++ which has many discrepancies against the ideal object model. For
instance, in Objective-C, classes are objects and can be dynamically managed: it is possible to add
classes at run-time, create instances based on the name of the class, ask a class for its methods,
and so on. This is much more powerful than C++ RTTI (cf. section 13.1 on page 62), which have
been added to a very “static” language. Discouraging RTTI use is common since the results may
depend on the compiler itself and lack portability.
3.1 Root class, type id, nil and Nil values
In an object-oriented language, each program makes use of a set of classes. Unlike C++, Objective-C
defines a root class. Every new class should be a descendant of the root class. In Cocoa, that class
is NSObject, and it provides a huge number of facilities for the run-time system. The root class
notion is not specific to Objective-C; it’s related to the object model. Smalltalk and Java make
use of a root class, while C++ does not.
Strictly speaking, every object should be of type NSObject, and every pointer to an object
could be declared as NSObject*. In fact, one can use the type id instead. This is a short and
handy way to declare a pointer to any object, and provides dynamic type-checking instead of static
type-checking. It is very useful for some weak typing on generic methods. Please note that a null
pointer to an object should be set to nil, not NULL. These values are not interchangeable. A normal
C pointer can be set to NULL, but nil was introduced in Objective-C for pointers to objects. In
Objective-C, classes are also objects (meta-class instances), and it is possible to declare a pointer
to a class. Their null value is Nil.
3.2 Class declaration
It is hard to show with a single example all the differences between Objective-C and C++ for class
declaration and implementation. Syntax and concepts are interleaved and require explanation. In
the following, the differences are exposed sequentially and specifically.
3.2.1 Attributes and methods
In Objective-C, attributes are called instance data, and member functions are called methods.
C++
Objective-C
class Foo
{
double x;
public:
int
f(int x);
float g(int x, int y);
};
int
Foo::f(int x) {...}
float Foo::g(int x, int y) {...}
@interface Foo : NSObject
{
double x;
}
-(int)
f:(int)x;
-(float) g:(int)x :(int)y;
@end
@implementation Foo
-(int)
f:(int)x {...}
-(float) g:(int)x :(int)y {...}
@end
In C++, attributes and methods are declared together inside the braces of the class. Method
implementation syntax is similar to C, with the addition of the scope resolution operator (Foo:: ).
10
In Objective-C, attributes and methods cannot be mixed. The attributes are declared in braces,
the methods follow. Their implementation lies in an @implementation block.
This is a major difference with C++, since some methods can be implemented without being
exposed in the interface. This is detailed later. Briefly, this is a way to clean up header files by
removing unnecessary declarations (“private” methods, and silently redefined virtual methods like
destructors). Please see Section 4.3.2 on page 21 for further explanations.
Instance methods are prefixed by the minus symbol “-”, and class methods by the plus symbol
“+”(cf. section 3.3.9 on page 18); this symbol has nothing to do with the UML notation and
the meaning public or private. The type of the parameters are enclosed in parenthesis, and the
parameters are separated by the symbol “:”. Please see Section 3.3.1 on the next page for further
explanations on the syntax of prototypes.
In Objective-C, there is no need for a semi-colon at the end of a class declaration. Also note
that the keyword to declare a class is @interface and not @class. The keyword @class is only
used in forward declarations (cf. section 3.2.2 on the current page). Finally, if there is no instance
data in a class, the braces, which would enclose nothing, can be ommitted.
3.2.2 Forward declarations: @class, @protocol
To avoid cyclic dependencies in header files, the C language supports the forward declaration,
that allows the coder to declare a class when the only required knowledge is its existence and
not its structure. In C++, the keyword class is used; in Objective-C, it is @class. The keyword
@protocol can also be used to anticipate the declaration of a protocol (cf. section 4.4 on page 22).
C++
//In file Foo.h
#ifndef __FOO_H__
#define __FOO_H__
class Bar; //forward declaration
class Foo
{
Bar* bar;
public:
void useBar(void);
};
#endif
//In file Foo.cpp
#include "Foo.h"
#include "Bar.h"
void Foo::useBar(void)
{
...
}
Objective-C
//In file Foo.h
@class Bar; //forward declaration
@interface Foo : NSObject
{
Bar* bar;
}
-(void) useBar;
@end
//In file Foo.m
#import "Foo.h"
#import "Bar.h"
@implementation Foo
-(void) useBar
{
...
}
@end
11
3.2.3 public, private, protected
One major feature of the object model is data encapsulation, which limits the visibility of data to
some parts of the code, in order to ensure its integrity.
C++
Objective-C
class Foo
{
public:
int x;
int apple();
protected:
int y;
int pear();
private:
int z;
int banana();
};
@interface Foo : NSObject
{
@public
int x;
@protected:
int y;
@private:
int z;
}
-(int) apple;
-(int) pear;
-(int) banana;
@end
In C++, attributes and methods can belong to a public, protected or private scope. The
default mode is private.
In Objective-C, only the instance data can be public, protected or private, and the default
visibility is protected. Methods can only be public. However, it is possible to mimic the private
mode, by implementing some methods in the @implementation, without declaring them in the
@interface, or using the notion of class category (cf. section 4.5 on page 25). It does not prevent
methods from being called, but they are less exposed. Implementing a method without a previ-
ous declaration is a special property of Objective-C and has a specific purpose, as explained in
Section 4.3.2 on page 21.
Inheritance cannot be tagged public, protected or private. The only way is public. Inher-
itance in Objective-C looks more like Java than C++ (section 4 on page 21).
3.2.4 static attributes
It’s not possible in Objective-C to declare a class data attribute (static as in C++). However, it
is possible to do the same thing in a different way: use a global variable in the implementation file
(optionally with the static C keyword to limit the scope). The class can then use accessors on
it (with class methods or normal methods), and its initialization can be done in the initialize
method of the class (cf. section 5.1.10 on page 34).
3.3 Methods
The syntax for a method in Objective-C is different from the syntax for common C functions. This
section aims to describe this syntax and adds some information on the underlying message sending
principle.
3.3.1 Prototype and call, instance methods, class methods
• A method is prefixed by “−” if it is an instance method (common case), or by “+” if it is
a class method (static in C++). This symbol is in no way linked to the UML notation for
public or private. Methods are always public in Objective-C;
12
• the types for the return value, or for the parameters, are enclosed by parentheses;
• the parameters are separated by a colon “:”;
• the parameters can be associated with a label, a name specified just before “:”; the label
is then part of the name of the method and modifies it. That makes calls to this
function particularly readable. In fact, the label usage should be systematic. Note that the
first parameter cannot have a label – indeed, its label is already the name of the method;
• a method name can be the same as an attribute name, without conflict. This is very useful
for getters (cf. section 6.4.8 on page 46).
C++
//prototype
void Array::insertObject(void *anObject, unsigned int atIndex)
//use with a "shelf" instance of the Array class and a "book" object
shelf.insertObject(book, 2);
Objective-C
Without label (direct translation from C++)
//prototype
//The method is named "insertObject::", the colon being used to separate
//the parameters (it is not a scope resolution operator like in C++)
-(void) insertObject:(id)anObject:(unsigned int)index
//use with a "shelf" instance of the Array class and a "book" object
[shelf insertObject:book:2];
With label
//prototype. The "index" parameter is now labeled "atIndex"
//The method is now named "insertObject:atIndex:"
//The call can easily be read like a sentence
-(void) insertObject:(id)anObject atIndex:(unsigned int)index
//use with a "shelf" instance of the Array class and a "book" object
[shelf insertObject:book:2];
//Error !
[shelf insertObject:book atIndex:2]; //OK
Please note that the syntax with square brackets should not be read as calling the method
insertObject of object “shelf” but rather sending the message insertObject to object “shelf”.
This is what makes Objective-C. One can send any message to any target. If it is not able to
handle the message, it can ignore it (an exception is raised but the program does not abort). If,
at the time it receives the message, a target is able to handle it, then the matching method will
be called. The compiler is able to raise a warning if a message is sent to a class for which it is
known that no matching method exists. This is not considered as an error thanks to the forwarding
facility (cf. section 3.4.3 on page 19). If a target is only known under the type id, no warning will
be made at compile time; a potential error would be raised at run-time.
3.3.2 this, self and super
There are two particular targets for a message: self and super. self is the current object (like
this in C++), super is the parent class. The keyword this does not exist in Objective-C. It is
replaced by self.
NB: self is not a real keyword, it is a hidden parameter that each method receives, the value
of which is the current object. Its value can be changed, unlike C++’s this keyword. However,
this is only useful in constructors (cf. section 5.1 on page 27).
13
3.3.3 Accessing instance variables inside a method
Like in C++, an Objective-C method can access the instance variables of its current object. The
optional this-> in C++ can be written self->.
C++
Objective-C
class Foo
{
int x;
int y;
void f(void);
};
void Foo::f(void)
{
x = 1;
int y; //making ambiguity with this->y
y = 2; //using local y
this->y = 3; //disambiguating
}
@interface Foo : NSObject
{
int x;
int y;
}
-(void) f;
@end
@implementation Foo
-(void) f
{
x = 1;
int y; //making ambiguity with this->y
y = 2; //using local y
self->y = 3; //disambiguating
}
@end
3.3.4 Prototype id and signature, overloading
A function is a part of code that can be referenced, for instance to use as function pointers, or
functors. Moreover, even if the name of the function is a good candidate for a unique id, one
must take care when using overloading. C++ and Objective-C use opposite ways to differentiate
prototypes. The first is based on parameters types, while the second is based on parameter labels.
In C++,
two functions can have the same name as long as their parameters have different types.
When using methods, the const option is also discriminative.
C++
int f(int);
int f(float); //OK, float is different from int
class Foo
{
public:
int g(int);
int g(float); //OK, float is different from int
int g(float) const; //OK, const is discriminative
};
class Bar
{
public:
int g(int); //OK, we are in Bar::, different from Foo::
}
14
In Objective-C, all functions are C functions: they cannot be overloaded (unless the compiler
can be told to use C99; like gcc does). However, methods use a different syntax, and can be
differentiated thanks to parameter labels.
Objective-C
int f(int);
int f(float); //Error : C functions cannot be overloaded
@interface Foo : NSObject
{
}
-(int) g:(int) x;
-(int) g:(float) x; //Error : this method is not different
// from the previous one (no label)
-(int) g:(int) x :(int) y;
//OK : two anonymous labels
-(int) g:(int) x :(float) y; //Error : not different from the
//previous one
-(int) g:(int) x andY:(int) y;
//OK : second label is "andY"
-(int) g:(int) x andY:(float) y; //Error : not different from the
//previous one
-(int) g:(int) x andAlsoY:(int) y; //OK : second label is
//"andAlsoY", different from "andY"
@end
The label-based identifying method is useful for expressing the exact “name” of the function,
as stated below.
@interface Foo : NSObject {}
//The method name is "g"
-(int) g;
//The method name is "g:"
-(int) g:(float) x;
//The method name is "g::"
-(int) g:(float) x :(float) y;
//The method name is "g:andY:"
-(int) g:(float) x andY:(float) y;
//The method name is "g:andZ:"
-(int) g:(float) x andZ:(float) z
@end
Clearly, two Objective-C methods are differentiated by labels instead of types. This is a way to
replace “pointers on member functions”, expressed by selectors, as explained in section 3.3.5 on
the current page.
3.3.5 Pointer to member function: Selector
In Objective-C, the methods have a particular syntax using parentheses and labels.
It is not
possible to declare functions with this syntax. The notion of pointer to functions is the same in C
and Objective-C. It is only for pointers on methods where a difference occurs.
15
In C++,
the syntax is hard but consistent with the C langage : the pointer focus is on types.
C++
class Foo
{
public:
int f(float x) {...}
};
Foo bar
int (Foo::*p_f)(float) = &Foo::f; //Pointer to Foo::f
(bar.*p_f)(1.2345); //calling bar.f(1.2345);
In Objective-C, a new type has been introduced. Such a “pointer to a method” is called
a selector . Its type is SEL and its value is computed using @selector on the exact name of the
method (with labels of parameters). Calling the method can be done using the class NSInvocation.
Most of the time, utility methods of the performSelector: family (inherited from NSObject) are
handier, but a little restrictive. The simplest three ones are:
-(id) performSelector:(SEL)aSelector;
-(id) performSelector:(SEL)aSelector withObject:(id)anObjectAsParameter;
-(id) performSelector:(SEL)aSelector withObject:(id)anObjectAsParameter
withObject:(id)anotherObjectAsParameter;
The returned value is the same as that of the called method. For methods which have parameters
that are not objects, one should generally use wrapper classes like NSNumber, that can provide
float, int, and so on. The NSInvocation class, which is more general and powerful, can also be
used (see the documentation).
According to what has been said, nothing prevents a method from being called on an object,
even if the object does not implement it. In fact, the method is effectively triggered if the message
is accepted. But an exception, that can be caught, is raised when the object does not know the
method; the application is not interrupted. Moreover, it is possible to check that an object can
trigger a method by calling respondsToSelector:.
Finally, the value of @selector() is computed at compile-time, which does not slow down the
code.
16
Objective-C
@interface Slave : NSObject {}
-(void) readDocumentation:(Document*)document;
@end
//Let us suppose an array "array[]" of 10 slaves,
//and a document "document"
//Normal method call
for(i=0 ; i<10 ; ++i)
[array[i] readDocumentation:document];
//Just for the example, try to use performSelector: instead
for(i=0 ; i<10 ; ++i)
[array[i] performSelector:@selector(readDocumentation:)
withObject:document];
//The type of a selector is SEL
//The following version is not more efficient than the previous one,
//since @selector() is evaluated at compile-time
SEL methodSelector = @selector(readDocumentation:);
for(i=0 ; i<10 ; ++i)
[slaves[i] performSelector:methodSelector withObject:document];
//for an object "foo" which type is not known (id)
//the test is not mandatory, but prevents an exception to be raised
//if the object has no readDocumentation: method.
if ([foo respondsToSelector:@selector(readDocumentation:)])
[foo performSelector:@selector(readDocumentation:) withObject:document];
A selector can thus be used as a very simple function parameter. Generic algorithms, like
sorting, may be easily specialized that way (cf. 11.3 on page 55).
A selector is not, strictly speaking, a pointer to a function; its real underlying type is a C
string, registered by the run-time as a method identifier. When a class is loaded, its methods
are automatically registered in a table, so that @selector() can work as expected. This way, for
instance, the equality of two selectors can be done on adresses with == rather than reyling on a
string comparison.
The real adress of a method, seen as a C function, can be obtained with a different notion,
and the type IMP, that is briefly presented in section 11.3.2 on page 55. It is rarely used but for
optimization. Virtual calls, for instance, are handled by selectors but not by IMP. The Objective-C
equivalent for C++ method pointers is definitely in selectors.
Finally, you may remember that self is in Objective-C, like this in C++, a hidden parameter
of every method that stores the current object. You can also note that there is a second hidden
parameter, _cmd, which is the current method .
17
Objective-C
@implementation Foo
-(void) f:(id)parameter //equivalent to a C function of type
//"void f(id self, SEL _cmd, id parameter)"
{
id currentObject = self;
SEL currentMethod = _cmd;
[currentObject performSelector:currentMethod
withObject:parameter]; //recursive call
[self performSelector:_cmd withObject:parameter]; //idem
}
@end
3.3.6 Default values of parameters
Objective-C does not allow a default value to be specified for the parameters of functions or meth-
ods. One should create as much functions as necessary when some parameters could be optional.
In the case of constructors, one should use the notion of designated initializer (section 5.1.7 on
page 32).
3.3.7 Variable number of arguments
Objective-C allows the use of methods with a variable number of arguments. Like in C, the syntax
relies on using “...” as the last argument. This is rarely useful, even if many Cocoa methods are
doing so. More details can be found in the Objective-C documentation.
3.3.8 Anonymous arguments
In C++, it is possible to give no name to a parameter in a prototype, since the type is sufficient to
characterize the signature of a function. This is not possible in Objective-C.
3.3.9 Prototype modifiers (const, static, virtual, “= 0“, friend, throw)
In C++, a few modifiers can be added to the prototype of a function. None of them exist in
Objective-C. Here is the list:
• const: a method cannot be set const. As a consequence, the mutable keyword cannot exist;
• static: making the difference between an instance method and a class method is done
through the use of “-” or “+” in front of the prototype;
• virtual: Objective-C methods are virtual, so the keyword is useless. Pure virtual methods
are implemented with a formal protocol (cf. section 4.4 on page 22);
• friend: there is no “friend” notion in Objective-C;
• throw:
in C++, one can restrict a method to transmit some exceptions only. This is not
possible in Objective-C.
18
3.4 Messages and transmission
3.4.1 Sending a message to nil
By default, it is legal to send a message (call a method) to nil. The message is just ignored. The
code can be greatly simplified by reducing the number of tests usually made with the null pointer.
GCC has an option to disable that handy behaviour, for extra optimizations.
3.4.2 Delegating a message to an unknown object
Delegation is common with user interface elements in Cocoa (tables, outlines. . . ), taking advantage
of the ability to send a message to an unknown object. An object can, for example, delegate some
tasks to an assistant.
//This is a function to define an assistant
-(void) setAssistant:(id)slave
{
[assistant autorelease]; //see the section about memory management
assistant = [slave retain];
}
//the method performHardWork can use delegation
-(void) performHardWork:(id)task
{
//the assistant is not known
//we check that it can handle the message
if ([assistant respondsToSelector:@selector(performHardWork:)])
[assistant performHardWork:task];
else
[self findAnotherAssistant];
}
3.4.3 Forwarding: handling an unknown message
In C++, code cannot be compiled if a method is called on an object that does not implement it.
In Objective-C, there’s a difference: one can always send a message to an object. If it can’t be
handled at run-time, it will be ignored (and raise an exception); moreover, instead of ignoring it,
it can forward the message to another object.
When the compiler is told the type of an object, it can detect if a message sending – method
call – will fail, and raise a warning. However, this is not an error, since in this case, an alternative is
available. This second chance is represented by a call to the forwardInvocation: method, which
can be redefined to redirect the message at the very last moment. This is obviously a method of
NSObject, which does nothing by default. Here is another way to manage assistant objects.
-(void) forwardInvocation:(NSInvocation*)anInvocation
{
//if we are here, that is because the object cannot handle
//the message of the invocation
//the bad selector can be obtained by sending "selector"
//to the object "anInvocation"
if ([anotherObject respondsToSelector:[anInvocation selector]])
[anInvocation invokeWithTarget:anotherObject];
else //do not forget to try with the superclass
[super forwardInvocation:anInvocation];
}
19
Even if a message can be handled, at last, in a forwardInvocation:, and only there, a check
based on respondsToSelector: will still return NO. Indeed, the respondsToSelector: mechanism
is not designed to guess whether forwardInvocation: will work or not.
Using the forward invocation can be thought as a bad practice, because it is triggering some
code while an error should occur. In fact , very good uses can be made of that mechanism, like in
the implementation of the Cocoa’s NSUndoManager. It allows an exceptionnaly pleasant syntax :
the Undo manager can records calls to methods, though it is not itself the target of those calls.
3.4.4 Downcasting
Downcasting is needed in C++ to call methods on a subclass, when only a parent class pointer is
known, this is made possible using dynamic_cast. In Objective-C, this practice is not necessary,
since a message can be sent to an object event if it seems that the object cannot handle it.
However, to avoid a compiler warning, one can simply cast the type of the object; there is no
explicit downcasting operator in Objective-C, the traditionnal cast syntax of the C language can
be used.
//NSMutableString is a subclass of NSString (string of characters)
//that allows mutating operations
//the "appendString:" method only exists in NSMutableString
NSMutableString* mutableString = ...initializing a mutable string...
NSString* string = mutableString;//storing in an NSString pointer
//those different calls are all valid
[string
appendString:@"foo"];//compiler warning
[(NSMutableString*)string appendString:@"foo"];//no warning
[(id)string
appendString:@"foo"];//no warning
20
4 Inheritance
4.1 Simple inheritance
Objective-C obviously implements the notion of inheritance, but does not support multiple in-
heritance. This limitation is addressed by other concepts (protocols, class categories) that are
explained later in this document (sections 4.4 on the following page, 4.5 on page 25).
C++
Objective-C
class Foo : public Bar,
protected Wiz
{
}
@interface Foo : Bar //single inheritance
//An alternative technique must be
//used to also "inherit" from Wiz
{
}
@end
In C++, a class can be derived from one or several other classes, using public, protected or
private mode. In the methods, one can reference a superclass using the scope resolution operator
:: (Bar::, Wiz::).
In Objective-C, one can derive from one class only, using publicmode. A method can reference
the superclass with the (false) keyword super, like in Java.
4.2 Multiple inheritance
Objective-C does not implement multiple inheritance, but brings other concepts, the protocols
(cf. 4.4 on the following page) and the categories (cf. 4.5 on page 25).
4.3 Virtuality
4.3.1 Virtual methods
In Objective-C, all methods are virtual. Hence, the virtual keyword does not exist and has no
equivalent.
4.3.2 Silent redefinition of virtual methods
It is possible in Objective-C to implement a method that has not been declared in the interface
section. This feature does not replace the @private notion for the methods (even if it can be used
to “hide” methods) : they can still be called; but it does lighten the interface declarations.
This is not a bad practice: the methods one uses with this technique are often “well-known”
methods of the super-classes. Many methods of the root-class NSObject are silently redefined. One
can quote the constructor init (cf. section 5.1 on page 27), the destructor dealloc (cf. section 5.2
on page 34), the drawing method drawRect: of views, and so on.
The interfaces are then lighter and easier to read, even if it is more difficult to see what can be
redefined, implying a regular reading of the super class documentation.
The notion of pure virtual methods (a method that must be redefined in sub-classes), is
addressed by the concept of formal protocols (cf. section 4.4.1 on the following page on the next
page).
4.3.3 Virtual inheritance
Virtual inheritance is not relevant in Objective-C, since inheritance cannot be multiple and has
none of the associated problems.
21
4.4 Protocols
Java and C# fix the lack of multiple inheritance by the notion of interface. In Objective-C, the
same notion is used, and is called a protocol. In C++, this would be an abstract class. A protocol
is not a real class: it can only declare methods and cannot hold any data. There are two types of
protocols : formal and informal.
4.4.1 Formal protocol
A formal protocol is a set of methods that must be implemented in any conforming class. This
can be seen as a certification regarding a class, ensuring that it is able to handle everything that
is necessary for a given service. A class can conform to an unlimited number of protocols.
C++
class MouseListener
{
public:
virtual bool mousePressed(void) = 0; //pure virtual method
virtual bool mouseClicked(void) = 0; //pure virtual method
};
class KeyboardListener
{
public:
virtual bool keyPressed(void) = 0; //pure virtual method
};
class Foo : public MouseListener, public KeyboardListener {...}
//Foo MUST implement mousePressed, mouseClicked and keyPressed
//It can then be used as an event listener for the mouse and the keyboard
Objective-C
@protocol MouseListener
-(BOOL) mousePressed;
-(BOOL) mouseClicked;
@end
@protocol KeyboardListener
-(BOOL) keyPressed;
@end
@interface Foo : NSObject
{
...
}
@end
//Foo MUST implement mousePressed, mouseClicked and keyPressed
//It can then be used as an event listener for the mouse and the keyboard
In C++, a protocol is implemented by an abstract class and pure virtual methods. The abstract
class in C++ is more powerful than the Objective-C protocol since it can contain data.
22
In Objective-C,
the protocol is a specific concept. The syntax using angular brackets <...>
is not linked to the C++ templates, a feature that does not exist in Objective-C.
A class can implement all the methods of a protocol without declaring its conformance. In this
case, the conformsToProtocol: method returns NO. For efficiency reasons, conformsToProtocol:
does not check, method-by-method, the conformance to a protocol, but is based on the explicit
declaration from the developer. However, the negative answer to conformsToProtocol: does not
prevent the program from behaving correctly if the methods of the protocol are called. Here is the
prototype of conformsToProtocol:
-(BOOL) conformsToProtocol:(Protocol*)protocol
//a Protocol object is returned by a call to @protocol(protocol name)
The type of an object that conforms to a formal protocol can be added the name of the protocol
itself, between angular braces. This is useful for assertions. For example :
//the following standard Cocoa method takes one parameter which is of
//any type (id), but must conform to the protocol NSDraggingInfo
-(NSDragOperation) draggingEntered:(id )sender;
4.4.2 Optional methods
It may be desirable that a class conforms to a protocol, to show that it can handle a particular
service, but without forcing it to conform to the whole protocol. For instance, in Cocoa, the notion
of a delegate object is widely used: an object can be given an assistant, to handle some tasks, but
not all of them.
An immediate solution would be to split a formal protocol into multiple ones, and then make
the class conform to a subset of these protocols. This is not very handy. Cocoa brings a solution
in the notion of informal protocols. With Objective-C 1.0, the informal protocols coudl be used
(cf. section 4.4.3). With Objective-C 2.0, the new keywords @optional and required can make
the difference between optional and required methods.
@protocol Slave
@required //required part
-(void) makeCoffee;
-(void) duplicateDocument:(Document*)document count:(int)count;
@optional //optional part
-(void) sweep;
@required //you can split required/optional sections
-(void) bringCoffee;
@end
4.4.3
Informal protocol
The informal protocol i not really a “protocol” : it creates no constraint upon the code. But it is
“informal” by nature and targets it to the code auto-documentation.
An informal protocol lets a developer group methods by application field, so that it can organize
its classes consistently.
So, it is not so surprising that an informal protocol is not declared with the relaxation of a
formal protocol. Another concept is used : the class category (cf. section 4.5 on page 25).
Let us imagine a service called “document managing”. Assume there is a difference between
green, blue and red documents. Even if the class can only handle blue documents, a Slave
class is preferred to using three formal protocols: manageGreenDocuments, manageBlueDocuments
23
and manageRedDocuments. To the Slave class is rather added a category DocumentsManaging,
declaring the methods for the tasks it is able to accomplish. The name of the category is specified
in parenthesis (more explanations are given in Section 4.5 on the following page):
@interface Slave (DocumentsManaging)
-(void) manageBlueDocuments:(BlueDocument*)document;
-(void) trashBlueDocuments:(BlueDocument*)document;
@end
Any class can use the DocumentsManaging category, to declare methods that are related to
that service.
@interface PremiumSlave (DocumentsManaging)
-(void) manageBlueDocuments:(BlueDocument*)document;
-(void) manageRedDocuments:(RedDocument*)document;
@end
A developer can then browse the code and see the DocumentsManaging category. Hence, he can
suppose that the class is useful for some tasks, and he can check which ones exactly by consulting
the documentation. Even if he does not check the source code, a run-time request is still possible:
if ([mySlave respondsToSelector:@selector(manageBlueDocuments:)])
[mySlave manageBlueDocuments:document];
Strictly speaking, apart from the knowledge of the prototypes, the informal protocol is useless to
the compiler, it does not restrict the use of objects. However, it is vital for self-documenting code,
making APIs more readable.
4.4.4 Object of type Protocol
At run-time, a protocol is like a class represented by an object, and is typed Protocol*. Such an
object can be used as a parameter of a method like conformsToProtocol: (cf. section 13.1.2 on
page 62).
The keyword @protocol, that is used to declare protocols, is also used to build a Protocol*
object from its name:
Protocol* myProtocol = @protocol(protocol name ).
4.4.5 Message qualifiers for distant objects
Thanks to the dynamism of Objective-C, distant objects can communicate easily. They can belong
to distinct programs, on different machines, but can delegate some tasks and exchange some
information. Now, formal protocols are a perfect way to ensure that an object conforms to a given
service, wherever it comes from. The formal protocol concept has been given some extra keywords
to allow for more efficient communication between distant objects.
These keywords are in, out, inout, bycopy, byref and oneway.
It is only applicable to
distributed objects, and outside a protocol definition, they are not reserved keywords and can be
freely used.
These keywords are inserted inside the prototypes of the methods declared inside a formal
protocol, to add extra information regarding their behaviour. They can be used to specify which
parameters are input parameters, and which are ouput results; it is also possible to tell whether
they are to be used by copy or by reference; and the methods can be made synchronous or not.
Here are the different meanings:
• an in parameter is an input variable;
• an out parameter is an output variable;
• an inout parameter can be used in both ways (input and output);
24
• a bycopy parameter is transmitted by copy ;
• a byref parameter is transmitted by reference (without copy) ;
• a oneway method is asynchronous (the result is not immediately expected) - hence it must
return void.
For instance, this is an asynchronous method that returns an object:
-(oneway void) giveMeAnObjectWhenAvailable:(bycopy out id *)anObject;
By default, parameters are considered to be inout, except const pointers, that are supposed
to be in. Choosing in or out instead of inout is an optimization. The default mode to transmit
the parameters is byref, and the methods are synchronous by default (without oneway).
For parameters transmitted by value,
like non-pointers variables, out and inout make no
sense, only in is correct.
4.5 Class categories
Creating categories for a class is a way to split its implementation into several parts. Each category
is a part of the class. A class can use an arbitrary number of categories, but none can add instance
data. It brings the following benefits:
• For the meticulous developer, it is possible to make groups of methods. For a very rich class,
its different roles can be cleanly separated;
• It is possible to compile them separately, making collaborative work possible on the same
class;
• If the interface of a category and its implementation are present in some implementation
file (.m file), it is very easy to define private methods, which are only visible inside the file
(even if anybody who knows the prototype can use them, there is no calling restriction). An
appropriate name for such a category could be FooPrivateAPI ;
• A class can be extended differently in different applications, without duplicating the common
code. Any class can be extended, even existing Cocoa classes.
The last point is important: each developer would have their preference to extend a standard
class with the methods that are useful for them. This is not a real problem: inheritance is a way
to do that. However, in the context of simple inheritance, this can lead to a heavy subclassing
scheme. Moreover, it can seem a disproportional effort to make a subclass for a single method.
The class categories are an elegant solution to this problem.
C++
Objective-C
class MyString : public string
{
public:
//counts the vowels
int vowelCount(void);
};
int MyString::vowelCount(void)
{
...
}
@interface NSString (VowelsCounting)
//please note that no brackets {} are used
-(int) vowelCount; //counts the vowels
@end
@implementation NSString (VowelsCounting)
-(int) vowelCount
{
...
}
@end
25
In C++,
the new class is usable without restrictions.
In Objective-C,
the NSString class (a standard Cocoa class) is given an extension that is
usable inside the whole program. No new class is created. Every NSString object benefits from
the extension (even constant strings, cf. section 9.1 on page 52). But no instance variable can be
added inside a category, so there is no {. . . } block.
A category can even be anonymous, which is perfect for a “private” one.
@interface NSString ()
//please note that no brackets {} are used
-(int) myPrivateMethod;
@end
@implementation NSString ()
-(int) myPrivateMethod
{
...
}
@end
4.6 Joint use of protocols, categories, subclassing:
The only restriction in the joint specification of protocols, categories and derivation, is that a
subclass and a category cannot be declared concurrently; two steps are needed.
@interface Foo1 : SuperClass //ok
@end
@interface Foo2 (Category) //ok
@end
//below : compilation error
@interface Foo3 (Category) : SuperClass
@end
//a solution :
@interface Foo3 : SuperClass //step 1
@end
@interface Foo3 (Category) //step 2
@end
26
5 Instantiation
The instantiation of a class leads to two problems: how are the notions of a constructor/ destructor/
copy operator implemented, and how are they managed in memory ?
First, an important point: in the C and C++ languages, variables are said to be “automatic”
by default: unless they are declared static, they only exist inside their definition block. Only
dynamically allocated memory is usable beyond, until the matching free() or delete method is
called. Objects follow this rule in C++.
However, in Objective-C, all objects are dynamically allocated. This is rather logical, C++
being very static, and Objective-C being very dynamic. Objective-C’s dynamism wouldn’t be that
rich if objects weren’t created at run-time.
Please see the section 6 on page 39 for a detailed explanation about the way to retain or release
objects.
5.1 Constructors, initializers
5.1.1 Distinction between allocation and initialization
In C++, allocation and initialization of an object are mixed when calling the constructor.
In
Objective-C, they are two different methods.
Allocation is handled by the class method alloc, which also initializes all instance data.
Instance data are set to 0, except the isa (is-a) pointer of NSObject, which value describes the
exact type of the newly created object at run-time. Should instance data be set to a particular
value, depending on the parameters of construction, then the corresponding code is deported inside
an instance method. Its name usually begins with init. Therefore, the construction is clearly
divided into two steps : allocation and initialization. The alloc message is sent to the class, and
the init... message is sent to the object newly instantiated by alloc.
The initialization step is not optional, and alloc should always be followed by
init; with successive calls to superclass initializers, this must end in calling the init method of
NSObject, which performs important work.
In C++, the name of the constructor cannot be chosen. In Objective-C, the initializer is a
method like any other, and the init prefix is traditional but not mandatory. However, you are
strongly encouraged to follow a rule: the name of an initialization method must begin with
“init”.
5.1.2 Using alloc and init
A call to alloc returns a new object that must be sent init . The call to init also returns an
object. Most of the time, this will be the initial object. Sometimes, when one is using a singleton
(an object for which a single instance is allowed), init could substitute another return value. So,
the returned value of init should not be ignored. Usually, alloc and init are called on the same
line.
C++
Foo* foo = new Foo;
Objective-C
Foo* foo1 = [Foo alloc];
[foo1 init]; //bad : the returned value should be used
Foo* foo2 = [Foo alloc];
foo2 = [foo2 init]; //ok, but not handy
Foo* foo3 = [[Foo alloc] init]; //ok, usual way to do
To know whether an object was created or not, C++ requires either exception catching, or a
test against 0 (if new(nothrow) was used). With Objective-C, a test against nil is sufficient.
27
5.1.3 Example of a correct initializer
The constraints for a correct initializer are
• its name begins with init;
• it returns the object to use;
• it calls some init method of the superclass, so that init of NSObject will finally be called;
• it takes in account the value returned by [super init...];
• it handles correctly the construction errors, whether they are voluntary or inherited.
Here is an example of object instanciation, in C++ and in Objective-C.
C++
class Point2D
{
public:
Point2D(int x, int y);
private:
int x;
int y;
};
Point2D::Point2D(int anX, int anY) {x = anX; y = anY;}
...
Point2D p1(3,4);
Point2D* p2 = new Point2D(5, 6);
Objective-C
@interface Point2D : NSObject
{
int x;
int y;
}
//Note : id is somewhat similar to void* in Objective-C.
//(id) is the most "general" type for an object
-(id) initWithX:(int)anX andY:(int)anY;
@end
@implementation Point2D
-(id) initWithX:(int)anX andY:(int)anY
{
//an initializer of the superclass must be called
if (!(self = [super init])) //if the superclass is NSObject, this must be init
return nil; //in case of super-class failure, return nil to propagate
//in case of success, make additional initializations
self->x = anX;
self->y = anY;
return self; //and return the object itself
}
@end
...
Point2D* p1 = [[Point2D alloc] initWithX:3 andY:4];
28
5.1.4 self = [super init...]
The most surprising syntax in the construction is self = [super init...]. Let us recall that
self is a hidden argument given to every method, representing the current object. Thus, it is
a local variable; so why do we have to change its value ? Will Shipley[9] tried to show in a
very interesting document that it was useless. Its arguments were relying on hypothesis on the
Objective-C run-time that turned out to be false; indeed, self must be modified, but explanations
are welcome.
It can happen that [super init] returns another object than the current one. A singleton
would be such a case, but it is a counter-argument of [9] : il is illogic to call twice init for a
singleton; being there reveals a conception failure on top of that.
However, an API can substitute, to a newly-allocated object, another object. Core Data1 is
doing that, to handle a special treatment of instance data, which are then linked to the fields of
a database. When sub-classing the NSManagedObject class provided by Cocoa, it is mandatory to
take care of that substitution.
In such a case, self will hold successively two different values : the first is the one returned
by alloc, the second is the one returned by [super init...]. Modifying the value of self has
border effects : every access to a data member is using it implicitely, as shown in the code below.
@interface B : A
{
int i;
}
@end
@implementation B
-(id) init
{
//at that step, the value of self is the one returned by alloc
//let us suppose that A is performing subsitution, returning a different "self"
id newSelf = [super init];
NSLog(@"%d", i);//prints the value of self->i
self = nouveauSelf; //one could think "i" is untouched, but...
NSLog(@"%d", i);//prints the value of self->i, therefore newSelf->i,
//not necessarily the same as previously
return self;
}
@end
...
B* b = [[B alloc] init];
The concise form self = [super init] is the most simple to avoid introducing bugs after-
wards. However, it is legible to worry about the destiny ot the object pointed by the “old” self :
indeed, it must be freed.
The first rule is simple : handling the old self must be done by the one who substituted it
(if it happened); here, it is [super init...] who must perform the deallocation. For instance,
if you subclassNSManagedObject (a Cocoa class which performs substitution), you do not have to
care about the old self. On the contrary, the developer of NSManagedObject had to handle it
properly.
1Core Data is a Cocoa API provided by Apple
29
Thus, if you happen to develop class that performs substitution, you should know how to
deallocate an object during its initialization. This problem is the same as handling errors : what
should we do if we want to make the construction fail (invalid parameters, unavailable resources...)
? This is answered in Section 5.1.5.
5.1.5
Initialization failure
When constructing an object (initializing it fact), an error may happen, or being triggered, in three
different places :
• 1: before the call to [super init...] : if the construction parameters are considered invalid,
the initialization must be stopped as soon as possible;
• 2: with the call to [super init...] : if the super-class fails, we should abort the current
one;
• 3: after the call to [super init...] : if an additional resource allocation fails, for instance.
In every case, nil must be returned, and deallocating the current object must be done by the one
who triggered the error. Here, we are responsible of cases 1 and 3, but not of case 2. To deallocate
the current object, one just have to call [self release], which is very natural (cf. section 6 on
page 39 about memory management, wich explains the release).
The destruction of the object will end up by calling dealloc (cf section 5.2 on page 34 on
destructors); thus the implementation of that method must be compatible with partially-initialized
objects. The fact that all instance data are initialized to 0 by alloc is rather useful for that.
30
@interface A : NSObject {
unsigned int n;
}
-(id) initWithN:(unsigned int)value;
@end
@implementation A
-(id) initWithN:(unsigned int)value
{
//case #1 (is the construction legible ?)
if (value == 0) //here, we want a strictly positive value
{
[self release];
return nil;
}
//case #2 (is the super-class ok ?)
if (!(self = [super init])) //even is self is substitued, it’s the super-class
return nil; //who is reponsible, if an error occurs, for releasing "self"
//case #3 (can the initialization be complete ?)
n = (int)log(value);
void* p = malloc(n); //trying to allocate a resource
if (!p) //in case of failure, we want it to be an error
{
[self release];
return nil;
}
}
@end
5.1.6 “Splitting” construction into alloc+init
The successive application of alloc and init can sound laborious in some cases. Fortunately, it
can be shortened as a convenience constructor . The exact knowledge of such a constructor relies
on notions of memory management in Objective-C. So, the exact explanations are given further in
this document in Section 6.4.6 on page 42. Briefly, such a constructor, the name of which should
be prefixed by the name of the class, behaves like an init method, but it performs the alloc itself.
However, the object is registered in the autorelease pool (cf. section 6.4 on page 39) and will have
a limited life cycle if it is not sent a retain message. Here is an example :
//laborious
NSNumber* tmp1 = [[NSNumber alloc] initWithFloat:0.0f];
...
[tmp1 release];
//handier
NSNumber* tmp2 = [NSNumber numberWithFloat:0.0f];
...
//no need to release
31
5.1.7 Default constructor : designated initializer
The notion of default constructor has no real meaning in Objective-C. Because all objects are
allocated dynamically, their construction is always explicit. However, a preferred constructor can
be used to factorize some code. Indeed, a correct initializer is most of the time similar to:
if (!(self = [super init])) //"init" or another appropriated
return nil;
//initializer of the superclass
//in case of success, add some code...
return self;
Since code redundancy is a bad practice, it seems heavy-handed to repeat that scheme in each
possible initializer. The best solution is to polish up the most essential initializer with this code.
Then the other initializers will call this “preferred” initializer, known as the designated initializer.
Logically, the designated initializer is the one with the greatest number of parameters, since it is
not possible in Objective-C to give default values to these parameters.
-(id) initWithX:(int)x
{
return [self initWithX:x andY:0 andZ:0];
}
-(id) initWithX:(int)x andY:(int)y
{
return [self initWithX:x andY:y andZ:0];
}
//designated initializer
-(id) initWithX:(int)x andY:(int)y andZ:(int)z
{
if (!(self = [super init]))
return nil;
self->x = x;
self->y = y;
self->z = z;
return self;
}
If the designated initializer is not the one with the greatest number of parameters, then it’s not
very useful:
32
//the following code is not handy
-(id) initWithX:(int)x //designated initializer
{
if (!(self = [super init]))
return nil;
self->x = x;
return self;
}
-(id) initWithX:(int)x andY:(int)y
{
if (![self initWithX:x])
return nil;
self->y = y;
return self;
}
-(id) initWithX:(int)x andY:(int)y andZ:(int)z
{
if (![self initWithX:x])
return nil;
self->y = y;
self->z = z;
return self;
}
33
5.1.8 List of initialization and default value of instance data
The idea of list of initialization for C++ constructors does not exist in Objective-C. However, it
is remarkable that, unlike the C++ behaviour, the alloc method of Objective-C initializes all the
bits of instance data to 0, so the pointers are set to nil. This would be a problem in C++ for
objects as attributes, but in Objective-C, the objects are always represented by pointers.
5.1.9 Virtual constructor
It is possible in Objective-C to get real virtual constructors. For more details, please see Sec-
tion 6.4.6 on page 42, after the introduction to the memory management (Section 6 on page 39).
5.1.10 Class constructors
In Objective-C, since the classes are themselves objects, they supply a constructor that can be rede-
fined. it is obvioulsy a class method, inherited from NSObject; its prototype is +(void) initialize;
This method is automatically called when using the class or one of its sub-classes for the
first time. However, it is not true that this method is called only once for a given class; in-
deed, if a sub-class does not redefine +(void) initialize, the Objective-C mechanism calls
+(void) initialize from the mother class.
5.2 Destructors
In C++, the destructor, like the constructor, is a specific method that can be redefined.
In
Objective-C, it is an instance method named dealloc.
In C++, the destructor is called automatically when an object is freed; it is the same with
Objective-C; only the way to release the object is different (cf. Section 6 on page 39).
The destructor should never be called explicitly. In fact, in C++ there is one case where it can
be : when the developer himself is managing the memory pool used for the allocation. But in
Objective-C, no case justifies an explicit call to dealloc. One can use custom memory zones in
Cocoa, but their use does not influence common allocation/deallocation practices (cf. Section 5.3
on the following page).
C++
class Point2D
{
public:
~Point2D();
};
Point2D::~Point2D() {}
Objective-C
@interface Point2D : NSObject
-(void) dealloc; //this method can be redefined
@end
@implementation Point2D
//in this example the redefinition was not necessary
-(void) dealloc
{
[super dealloc]; //do not forget to transmit to the superclass
}
@end
34
5.3 Copy operators
5.3.1 Classical cloning, copy, copyWithZone:, NSCopyObject()
In C++, it is important to define a coherent implementation of the copy constructor and the
affectation operator. In Objective-C, operator overloading is impossible; one must only ensure
that the cloning method is correct.
Cloning in Cocoa is associated with a protocol (cf. Section 4.4 on page 22) named NSCopying,
requesting the implementation of the method
-(id) copyWithZone:(NSZone*)zone;
Its argument is a memory zone in which the clone should be allocated. Cocoa allows the use of
different custom zones: some methods take such a zone as an argument. Most of the times, the
default zone is perfect, and there is no need to specify it each time. Fortunately, NSObject supplies
a method
-(id) copy;
which encapsulates a call to copyWithZone: with the default zone as parameter. But it is the
copyWithZone: version which is required by NSCopying. Finally, the utility function NSCopyObject()
provides a slightly different approach, that can be simpler but also requires caution. First, the
code is presented without taking NSCopyObject(...)
into account, for which explanations are
given in section 5.3.2 on the following page.
The implementation of copyWithZone: for some Foo class would look like:
//if the superclass does not implement copyWithZone:, and that NSCopyObject()
//is not used
-(id) copyWithZone:(NSZone*)zone
{
//a new object must be created
Foo* clone = [[Foo allocWithZone:zone] init];
//instance data must be manually copied
clone->integer = self->integer; //"integer" is here of type "int"
//trigger the same mechanism for sub-objects to clone
clone->objectToClone = [self->objectToClone copyWithZone:zone];
//some sub-objects may not be cloned but shared
clone->objectToShare = [self->objectToShare retain]; //cf. memory management
//if there is a mutator, it can be used in both cases
[clone setObject:self->object];
return clone;
}
Note the use of allocWithZone:, instead of alloc, to handle the zone parameter. alloc
encapsulates a call to allocWithZone: with the default zone. To learn more about Cocoa’s
custom zone management, refer to the Cocoa documentation.
However, one must take care of the possible implementation of copyWithZone: in the super-
class.
35
//if the superclass implements copyWithZone:, and that NSCopyObject() is not used
-(id) copyWithZone:(NSZone*)zone
{
Foo* clone = [super copyWithZone:zone]; //creates the new object
//you must clone instance data specific to the current sub-class
clone->integer = self->integer; //"integer" is here of type "int"
//trigger the same mechanism for sub-objects to clone
clone->objectToClone = [self->objectToClone copyWithZone:zone];
//some sub-objects may not be cloned but shared
clone->objectToShare = [self->objectToShare retain]; //cf. memory management
//if there is a mutator, it can be used in both cases
[clone setObject:self->object];
return clone;
}
5.3.2 NSCopyObject()
The NSObject class does not implement the NSCopying protocol, that is why a direct sub-class
cannot benefit from a call to [super copy...], and must use a standard initialization based on
the [[... alloc] init] scheme.
The utility function NSCopyObject() can be used to write simpler code, but requires some
caution with pointer members (including objects). This function creates a binary copy of an
object and its prototype is :
//the extra bytes is usually 0, but can be used for the extra space of
//indexed instance data
id NSCopyObject(id anObject,
unsigned int extraBytes, NSZone *zone)
The binary copy can automate the copy of instance data that are non-pointers; but for a pointer
member (including objects), one must keep in mind that it silently creates an additional reference
to the pointed data. The usual practice is to reset the pointers afterwards with values coming from
correct clonings.
//if the superclass does not implement copyWithZone:
-(id) copyWithZone:(NSZone*)zone
{
Foo* clone = NSCopyObject(self, 0, zone);//binary copy of data
//clone->integer = self->integer; //useless : binary data already copied
//a sub-object to clone must be really cloned
clone->objectToClone = [self->objectToClone copyWithZone:zone];
//a sub-object to share must only register the new reference
[clone->objectToShare retain]; //cf. memory management
//The mutator is likely to release clone->object. This is undesirable,
//because of the binary copy of the pointer value
//Therefore, before using the mutator, the pointer is reset
clone->object = nil;
[clone setObject:self->object];
return clone;
}
36
//if the superclass implements copyWithZone:
-(id) copyWithZone:(NSZone*)zone
{
Foo* clone = [super copyWithZone:zone];
//does a superclass implement NSCopyObject() ? This is important
//to know what remains to be done
clone->integer = self->integer;//only if NSCopyObject() has not been used
//In case of doubt, it can be done systematically
//a sub-object to clone must still be really cloned
clone->objectToClone = [self->objectToClone copyWithZone:zone];
//NSCopyObject() or not, a retain must be done (cf. memory management)
clone->objectToShare = [self->objectToShare retain];
clone->object = nil; //in case of doubt, it’s better to reset
[clone setObject:self->object];
return clone;
}
5.3.3 Dummy-cloning, mutability, mutableCopy and mutableCopyWithZone:
When cloning an object that cannot change, a fundamental optimization is to pretend that it is
cloned; instead of duplicating it, a reference to it can be returned. Starting from that, we can
distinguish the notion of an immutable and mutable object.
An immutable object cannot have any of its instance data changed; only the initializer gives
it a valid state. In this case, it can safely be “pseudo-cloned” by returning only a reference to
itself. Since neither it nor its clone can be modified, none of them can be undesirably affected by
a modification of the other. A very efficient copyWithZone: implementation can be proposed in
that case:
-(id) copyWithZone:(NSZone*)zone
{
//the object returns itself, counting one more reference
return [self retain]; //see the section about memory management
}
The use of retain comes from the memory management in Objective-C (cf. Section 6 on
page 39). The reference counter is incremented by 1 to make the existence of the clone “official”
so that deleting that clone will not destroy the original object.
“Pseudo-cloning” is not a marginal optimization. Creating an object requests a memory alloca-
tion, which is a “long” process that should be avoided when possible. That is why it is interesting
to identify two kind of objects: immutable ones, for which cloning can be fictive, and the others.
Making the distinction is as simple as creating “immutable” classes and optionally subclassing
them into “mutable” versions, adding methods to change their data instance. For instance, in
Cocoa, NSMutableString is a subclass of NSString, NSMutableArray is a subclass of NSArray,
NSMutableData is a subclass of NSData, and so on.
However, with the techniques presented here, it seems impossible to get a real clone, safely
mutable, from an immutable object which would only know how to “pseudo-clone” itself. Such a
limitation would significantly decrease the usefulness of immutable objects, isolating them from
the “external world”.
In addition to the NSCopying protocol, there is another protocol (cf. Section 4.4 on page 22)
named NSMutableCopying, requesting the implementation of
37
-(id) mutableCopyWithZone:(NSZone*)zone;
The mutableCopyWithZone: method must return a mutable clone, where modifications would not
apply to the original object. Similarly to the copy method, there is a mutableCopy method which
automatically calls mutableCopyWithZone: by giving it the default zone. The implementation of
mutableCopyWithZone: looks like the following, similar to the classical copy previously presented:
//if the superclass does not implement mutableCopyWithZone:
-(id) mutableCopyWithZone:(NSZone*)zone
{
Foo* clone = [[Foo allocWithZone:zone] init]; //or NSCopyObject() if possible
clone->integer = self->integer;
//Like with copyWithZone:, some sub-objects can be cloned, some others shared
//A mutable sub-object can be cloned with a call to mutableCopyWithZone:
//...
return clone;
}
Do not forget to use the possible mutableCopyWithZone: of the superclass:
//if the superclass implements mutableCopyWithZone:
-(id) mutableCopyWithZone:(NSZone*)zone
{
Foo* clone = [super mutableCopyWithZone:zone];
//...
return clone;
}
38
6 Memory management
6.1 new and delete
The C++ keywords new and delete do not exist in Objective-C (new exists as a method, but it
is just a deprecated shortcut for alloc+init). They are respectively replaced by calls to alloc
(cf. Section 5.1 on page 27) and release (cf. Section 6.2 on the current page).
6.2 Reference counting
Memory management in Objective-C is one of the most important parts of the language. In C or
C++, a memory area is allocated once and freed once. It can be referenced with as many pointers
as desired, but only one pointer will be given the delete call.
On the other hand, Objective-C implements a reference counting scheme. An object knows
how many times it is referenced. This can be explained by the analogy of dogs and leashes (an
analogy directly taken from Cocoa Programming for MacOS X [7]). If an object is a dog, everyone
can ask for a leash to hold it. If someones does not care about the dog anymore, it can drop its
leash. While the dog has at least one leash, it must stay there. But as soon as the number of
leashes falls to 0, the dog is free!
More technically, the reference counter of a newly created object is set to 1. If a part of the code
needs to reference that object, it can send it a retain message, which will increase the counter
by one. When a part of the code does not need the object any more, it can send it a release
message, that will decrease the counter by 1.
An object can receive as many retain and release messages as needed, as long as the reference
counter has a positive value. As soon as it falls to 0, the destructor dealloc is automatically called.
Sending release again to the address of the object, which is now invalid, triggers a memory fault.
This technique is not equivalent to the auto_ptr from the C++ STL. On the contrary, the
Boost library [5] supplies an encapsulation of pointers into a shared_ptr class, which implements
the reference counting scheme. But it is not part of the standard library.
6.3 alloc, copy, mutableCopy, retain, release
The understanding of memory management does not fully explain how it is used. The goal of this
section is to give some rules. The keyword autorelease is left aside for now, since it is more
difficult to understand.
The basic rule to apply is Everything that increases the reference counter with alloc,
[mutable]copy[WithZone:] or retain is in charge of the corresponding [auto]release.
Indeed, these are the three ways to increment the reference counter. It also means that you should
take care of releasing an object in only a limited number of cases:
• when you explicitly instantiate an object with alloc;
• when you explicitly clone the object with copy[WithZone:] or mutableCopy[WithZone:]
(whatever the copy is : a real or a pseudo-clone. This should not have any importance,
cf. section 5.3.3 on page 37);
• when you explicitly use retain.
Please remember that by default, it is legal to send a message (like release) to nil, without
any consequence (see Section 3.4.1 on page 19).
6.4 autorelease
6.4.1 Precious autorelease
The rule stated in the previous section is so important that it bears repeating: Everything that
increases the reference counter with alloc, [mutable]copy[WithZone:] or retain is in
charge of the corresponding [auto]release.
39
In fact, with the mere alloc, retain and release, this rule could not be applied. Indeed,
there are some methods that are not constructors, but that are designed to create objects : for
instance a binary addition operator in C++ (obj3 operator+(obj1, obj2)). In C++, the returned
object would be allocated on the stack and automatically destroyed when leaving its scope. But
in Objective-C, such objects do not exist. The function has necessarily used alloc, but cannot
release the object before returning it on the stack ! Here are given some illustrations of what is
going wrong:
-(Point2D*) add:(Point2D*)p1 and:(Point2D*)p2
{
Point2D* result = [[Point2D alloc] initWithX:([p1 getX] + [p2 getX])
andY:([p1 getY] + [p2 getY])];
return result;
}
//ERROR : the function performs "alloc", so, it is creating
//an object with a reference counter of 1. According
//to the rule, it should destroy the object.
//This can lead to a memory leak when summing three points :
[calculator add:[calculator add:p1 and:p2] and:p3];
//The result of the first addition is anonymous
//and nobody can release it. It is a memory leak.
-(Point2D*) add:(Point2D*)p1 and:(Point2D*)p2
{
return [[Point2D alloc] initWithX:([p1 getX] + [p2 getX])
andY:([p1 getY] + [p2 getY])];
}
//ERROR : This is exactly the same code as above. The fact that
//no intermediate variable is used does not change anything.
-(Point2D*) add:(Point2D*)p1 and:(Point2D*)p2
{
Point2D* result = [[Point2D alloc] initWithX:([p1 getX] + [p2 getX])
andY:([p1 getY] + [p2 getY])];
[result release];
return result;
}
//ERROR : obviously, it is nonsense to destroy the object after creating it
The problem seems intractable. It would be, if autorelease was not there. To simplify, let us
say that sending autorelease to an object means that it is sent a release that will be performed
“later”. But “later” does not mean “at any time”; this is detailed in Section 6.4.2 on the following
page. For a first approach, here is the only possible solution:
-(Point2D*) add:(Point2D*)p1 and:(Point2D*)p2
{
Point2D* result = [[Point2D alloc] initWithX:([p1 getX] + [p2 getX])
andY:([p1 getY] + [p2 getY])];
[result autorelease];
return result;
//a shorter writing is "return [result autorelease]"
}
//CORRECT : "result" will be automatically released later,
//after being used in the calling code
40
6.4.2 The autorelease pool
In the previous section, autorelease has been presented as a kind of magical release that is
automatically applied at the right moment. But it would make no sense to let the compiler guess
what the right moment is. In this case, a garbage collector would be more useful. To explain how
it works, more details must be given about autorelease.
Each time an object receives autorelease, it is only registered into an “autorelease pool”.
When the pool is destroyed, the object receives a real release. The problem has moved : how is
this pool handled ?
There is not a single answer : if you use Cocoa for an application with a graphical interface,
most of the time there is nothing to do. Otherwise, you would have to create and destroy the pool
yourself.
An application with a graphical interface generally uses an event loop. Such a loop waits
for actions from the user, then wakes up the program to perform the action, then goes back to
sleep until the next event. When you create a graphical application with Cocoa, an autorelease
pool is automatically created at the beginning of the loop and destroyed at the end. This is
logical: generally, a user’s action triggers a cascade of tasks. Temporary objects are created, then
destroyed, since they do not have to be kept for the next event. If some of them must be persistent,
the developer must use retain as necessary.
On the other hand, when there is no graphical interface, you have to create an autorelease pool
around the code that is needing it. When an object receives autorelease, it knows how to find
the closest autorelease pool. Then, when it is time to empty the pool, you can just destroy it with
a simple release. Typically, a command-line Cocoa program contains the following code:
int main(int argc, char* argv[])
{
NSAutoreleasePool* pool = [[NSAutoreleasePool alloc] init];
//...
[pool release];
return 0;
}
Please note that MacOS X10.5 has added the drain method to the class NSAutoreleasePool.
This metid is equivalent to release if the garbage collector is enabled, and triggers a run of garbage
collection otherwise (cf. 6.6 on page 48). This is useful to write code which behaves the same in
both cases.
6.4.3 Using several autorelease pools
It is possible, and sometimes useful, to have more than one autorelease pool in the program. An
object that receives autorelease will register in the closest pool. Therefore, if a function creates
and uses a large number of temporary objects, increased performance can be obtained by creating
a local autorelease pool. That way, the crowd of temporary objects will be destroyed as soon as
possible and will not clutter the memory after the function has returned.
6.4.4 Caution with autorelease
The fact that autorelease is handy should not lead to its misuse.
• First, sending more autorelease calls than necessary is similar to sending too many release
calls: it triggers a memory fault when emptying the pool;
• Then, even if it is true that any release message can be replaced by autorelease, this
would be a performance hit, since the autorelease pool is more work than a normal release.
Moreover, delaying all deallocations can lead to useless and irrelevant peaks of memory
consumption.
41
6.4.5 autorelease and retain
Thanks to autorelease, a method can create an object that can plan its own releasing. However,
it is common that the object must be kept longer. In this case, one has to send a retain on it,
and plan for a release. Then there are two points of view regarding the lifetime of this object :
• from the point of view of the developer of the function, the object is created and its release
is planned;
• from the point of view of the caller of the function, the lifetime is increased with the retain
(the autorelease planned by the function will not decrease the reference counter down to
0), but since there was a 1 increment, te caller is now responsible for releasing later.
6.4.6 Convenience constructor, virtual constructor
The successive applications of alloc and init can sound laborious in some cases. Fortunately,
it can be shortened by the notion of convenience constructor . Such a constructor, the name of
which should be prefixed by the name of the class, behaves like an init method, but it performs
the alloc itself. However, the returned object is registered inside an autorelease pool, and will be
temporary if it is not sent any retain. For example:
//laborious
NSNumber* zero_a = [[NSNumber alloc] initWithFloat:0.0f];
...
[zero_a release];
...
//handier
NSNumber* zero_b = [NSNumber numberWithFloat:0.0f];
...
//no need of release
With help from the section about memory management (Section 6 on page 39), it’s obvious that
such a constructor relies on autorelease. The underlying code is not that obvious anyway, since
it requires the correct use of self. Indeed, a convenience constructor is a class method, so that
self refers to an object of type Class, which is a meta-class instance. In an initializer , which is
an instance method, self is an instance of the class, and refers to a “normal” object.
It is easy to write a bad convenience constructor. Let us assume a class Vehicle holding a
color, and supplying a convenience constructor.
//The Vehicle class
@interface Vehicle : NSObject
{
NSColor* color;
}
-(void) setColor:(NSColor*)color;
//convenience constructor
+(id) vehicleWithColor:(NSColor*)color;
@end
The implementation of the convenience constructor is somewhat subtle.
42
//bad convenience constructor
+(Vehicle*) vehicleWithColor:(NSColor*)color
{
//the value of "self" should not change here
self = [[self alloc] init]; // ERROR !
[self setColor:color];
return [self autorelease];
}
self in this class method refers to the class. It cannot be set to be the instance.
//Almost perfect constructor
+(id) vehicleWithColor:(NSColor*)color
{
id newInstance = [[Vehicle alloc] init]; // OK, but ignores potential
// sub-classes
[newInstance setColor:color];
return [newInstance autorelease];
}
We can still improve this. In Objective-C one can get the behaviour of a virtual constructor. The
constructor just needs to perform an introspection to know which is the real class of the object
performing the method. Then it can directly produce an object of the right subclass. The false
keyword class can be used; this is a method of NSObject that returns the current object’s class
object (a meta-class instance).
@implementation Vehicle
+(id) vehicleWithColor:(NSColor*)color
{
id newInstance = [[[self class] alloc] init]; // PERFECT, the class is
// dynamically identified
[newInstance setColor:color];
return [newInstance autorelease];
}
@end
@interface Car : Vehicle {...}
@end
...
//produces a (red) car
id car = [Car vehicleWithColor:[NSColor redColor]];
Similar to the rule about the init prefix for initializers, you are strongly encouraged to prefix a
convenience constructor with the name of the class. There are very few cases where this is not the
case, like [NSColor redColor] in the previous code, which should have been written [NSColor
colorRed].
Finally, let us repeat the rule: Everything that increases the reference counter with
alloc, [mutable]copy[WithZone:] or retain is in charge of the corresponding [auto]release.
When calling a convenience constructor, you are not explicitly calling alloc, so you are not in
charge of the release. However, when creating such a constructor, you are writing alloc, and
you must not forget the autorelease.
43
6.4.7 Setter
A setter (mutating accessor) is a typical example of something that is difficult to write without
the knowledge of memory management in Objective-C. Let us assume a class encapsulating an
NSString named title, and let us suppose that we want to change the value of this string. This
very simple example raises the main problem related to setters : how is the parameter supposd to
be used ? Unlike C++, only one prototype is legible (an object can only be used through pointers),
but several implementations can be found. It can be an assignation, an assignation with retain,
or a copy. Each one has a specific meaning regarding the data model chosen by the developer.
Moreover, in each case, the old resources must be released first to avoid a memory leak.
assignation (incomplete code)
The outer object is just referenced weakly, without any retain. If the outer object is modified,
it is visible from the current class. If the outer object happened to be deallocated without the
present reference being set to nil before, it would be an invalid reference.
-(void) setString:(NSString*)newString
{
... memory management to be detailed later
self->string = newString; //assignation
}
assignation with retain (incomplete code)
The outer object is referenced, and the reference counter is increased by 1 thanks to a retain.
If the outer object is modified, this is visible from the current class. The outer object cannot be
deallocated as long as the current reference is not released.
-(void) setString:(NSString*)newString
{
... memory management to be detailed later
self-> string = [newString retain]; //assignation with retain
}
copy (incomplete code)
The outer object is not references : a clone is created instead. If the oter object is modified, this
is not visible on the clone. Logically, the clone is handled by the current owner object, and its
lifetime should not be greater.
-(void) setString:(NSString*)newString
{
... memory management to be detailed later
self->string = [newString copy]; //cloning;
//the NSCopying protocol is used
}
To complete the code,
the previous state of the object should be considered : in each case,
the mutator would have to release the old reference (if any) before setting a new one. This part of
the code is tricky.
44
assignation (complete code)
The simplest case. The old reference can be overwritten.
-(void) setString:(NSString*)newString
{
//no strong link : the old reference can be overwritten
self->string = newString; //assignation
}
assignation with retain (complete code)
In this case, the old reference should be released, unless it is the same as the new one.
//Bad codes
-(void) setString:(NSString*)newString
{
self->string = [newString retain];
//ERROR : memoy leak : the old "string" is no more referenced
}
-(void) setString:(NSString*)newString
{
[self->string release];
self->string = [newString retain];
//ERROR : if newString == string, (it can happen)
//and that the reference counter of newString was 1,
//the it is invalid to use newString (string) after
//[self->string release], because it has been dealocated at this point
}
-(void) setString:(NSString*)newString
{
if (self->string != newString)
[self->string release]; //ok: it is safe to send release even to nil
self->string = [newString retain]; //ERROR : should be in the "if";
//because if string == newString,
//the counter should not be incremented
}
45
//Correct codes
//Practice "Check before change"
//the most intuitive for a C++ developer
-(void) setString:(NSString*)newString
{
//avoid degenerated case where there is nothing to do
if (self->string != newString)
{
[self->string release]; //release the old one
self->string = [newString retain]; //retain the new one
}
}
//Practice "Autorelease the old value"
-(void) setString:(NSString*)newString
{
[self->string autorelease]; //even if string == newString,
//it’s correct, because release is delayed
self->string = [newString retain];
//... and thus this retain happens before
}
//Practice "retain, then release"
-(void) setString:(NSString*)newString
{
[self->newString retain]; //the reference counter is increased by 1 (except on nil)
[self->string release]; //...so that is does not reach 0 here
self->string = newString; //but no "retain" is added here !
}
copy (complete code)
Regarding the typical errors or the good solutions, this case is almost identical to the assignation
with retain, where retain is replaced by copy.
pseudo-clonage
Please note that the copy can be a “dummy-cloning” under the hood (cf. section 5.3.3 on page 37),
without any consequences.
6.4.8 Getters
With the Objective-C language, all objects are dynamically allocated. They are referenced and
encapsulated as pointers. Typically, getters are only returning the pointer’s values, and do not
copy the object on the fly. The name of a getter is usually the same as the data member, this is
possible in Objective-C and does not create a conflict. In the case of a boolean value, the name
might begin with an is, in order to be read as a predicate.
46