| blob | 08ae86fa57a04a69b7991e9f237ee0bffe2f6164 |
1 /* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
2 /* vim: set ts=8 sts=2 et sw=2 tw=80: */
3 /* This Source Code Form is subject to the terms of the Mozilla Public
4 * License, v. 2.0. If a copy of the MPL was not distributed with this
5 * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
7 /* C++11-style, but C++98-usable, "move references" implementation. */
9 #ifndef mozilla_Move_h
10 #define mozilla_Move_h
16 /*
17 * "Move" References
18 *
19 * Some types can be copied much more efficiently if we know the original's
20 * value need not be preserved --- that is, if we are doing a "move", not a
21 * "copy". For example, if we have:
22 *
23 * Vector<T> u;
24 * Vector<T> v(u);
25 *
26 * the constructor for v must apply a copy constructor to each element of u ---
27 * taking time linear in the length of u. However, if we know we will not need u
28 * any more once v has been initialized, then we could initialize v very
29 * efficiently simply by stealing u's dynamically allocated buffer and giving it
30 * to v --- a constant-time operation, regardless of the size of u.
31 *
32 * Moves often appear in container implementations. For example, when we append
33 * to a vector, we may need to resize its buffer. This entails moving each of
34 * its extant elements from the old, smaller buffer to the new, larger buffer.
35 * But once the elements have been migrated, we're just going to throw away the
36 * old buffer; we don't care if they still have their values. So if the vector's
37 * element type can implement "move" more efficiently than "copy", the vector
38 * resizing should by all means use a "move" operation. Hash tables should also
39 * use moves when resizing their internal array as entries are added and
40 * removed.
41 *
42 * The details of the optimization, and whether it's worth applying, vary
43 * from one type to the next: copying an 'int' is as cheap as moving it, so
44 * there's no benefit in distinguishing 'int' moves from copies. And while
45 * some constructor calls for complex types are moves, many really have to
46 * be copies, and can't be optimized this way. So we need:
47 *
48 * 1) a way for a type (like Vector) to announce that it can be moved more
49 * efficiently than it can be copied, and provide an implementation of that
50 * move operation; and
51 *
52 * 2) a way for a particular invocation of a copy constructor to say that it's
53 * really a move, not a copy, and that the value of the original isn't
54 * important afterwards (although it must still be safe to destroy).
55 *
56 * If a constructor has a single argument of type 'T&&' (an 'rvalue reference
57 * to T'), that indicates that it is a 'move constructor'. That's 1). It should
58 * move, not copy, its argument into the object being constructed. It may leave
59 * the original in any safely-destructible state.
60 *
61 * If a constructor's argument is an rvalue, as in 'C(f(x))' or 'C(x + y)', as
62 * opposed to an lvalue, as in 'C(x)', then overload resolution will prefer the
63 * move constructor, if there is one. The 'mozilla::Move' function, defined in
64 * this file, is an identity function you can use in a constructor invocation to
65 * make any argument into an rvalue, like this: C(Move(x)). That's 2). (You
66 * could use any function that works, but 'Move' indicates your intention
67 * clearly.)
68 *
69 * Where we might define a copy constructor for a class C like this:
70 *
71 * C(const C& rhs) { ... copy rhs to this ... }
72 *
73 * we would declare a move constructor like this:
74 *
75 * C(C&& rhs) { .. move rhs to this ... }
76 *
77 * And where we might perform a copy like this:
78 *
79 * C c2(c1);
80 *
81 * we would perform a move like this:
82 *
83 * C c2(Move(c1));
84 *
85 * Note that 'T&&' implicitly converts to 'T&'. So you can pass a 'T&&' to an
86 * ordinary copy constructor for a type that doesn't support a special move
87 * constructor, and you'll just get a copy. This means that templates can use
88 * Move whenever they know they won't use the original value any more, even if
89 * they're not sure whether the type at hand has a specialized move constructor.
90 * If it doesn't, the 'T&&' will just convert to a 'T&', and the ordinary copy
91 * constructor will apply.
92 *
93 * A class with a move constructor can also provide a move assignment operator.
94 * A generic definition would run this's destructor, and then apply the move
95 * constructor to *this's memory. A typical definition:
96 *
97 * C& operator=(C&& rhs) {
98 * MOZ_ASSERT(&rhs != this, "self-moves are prohibited");
99 * this->~C();
100 * new(this) C(Move(rhs));
101 * return *this;
102 * }
103 *
104 * With that in place, one can write move assignments like this:
105 *
106 * c2 = Move(c1);
107 *
108 * This destroys c2, moves c1's value to c2, and leaves c1 in an undefined but
109 * destructible state.
110 *
111 * As we say, a move must leave the original in a "destructible" state. The
112 * original's destructor will still be called, so if a move doesn't
113 * actually steal all its resources, that's fine. We require only that the
114 * move destination must take on the original's value; and that destructing
115 * the original must not break the move destination.
116 *
117 * (Opinions differ on whether move assignment operators should deal with move
118 * assignment of an object onto itself. It seems wise to either handle that
119 * case, or assert that it does not occur.)
120 *
121 * Forwarding:
122 *
123 * Sometimes we want copy construction or assignment if we're passed an ordinary
124 * value, but move construction if passed an rvalue reference. For example, if
125 * our constructor takes two arguments and either could usefully be a move, it
126 * seems silly to write out all four combinations:
127 *
128 * C::C(X& x, Y& y) : x(x), y(y) { }
129 * C::C(X& x, Y&& y) : x(x), y(Move(y)) { }
130 * C::C(X&& x, Y& y) : x(Move(x)), y(y) { }
131 * C::C(X&& x, Y&& y) : x(Move(x)), y(Move(y)) { }
132 *
133 * To avoid this, C++11 has tweaks to make it possible to write what you mean.
134 * The four constructor overloads above can be written as one constructor
135 * template like so[0]:
136 *
137 * template <typename XArg, typename YArg>
138 * C::C(XArg&& x, YArg&& y) : x(Forward<XArg>(x)), y(Forward<YArg>(y)) { }
139 *
140 * ("'Don't Repeat Yourself'? What's that?")
141 *
142 * This takes advantage of two new rules in C++11:
143 *
144 * - First, when a function template takes an argument that is an rvalue
145 * reference to a template argument (like 'XArg&& x' and 'YArg&& y' above),
146 * then when the argument is applied to an lvalue, the template argument
147 * resolves to 'T&'; and when it is applied to an rvalue, the template
148 * argument resolves to 'T&&'. Thus, in a call to C::C like:
149 *
150 * X foo(int);
151 * Y yy;
152 *
153 * C(foo(5), yy)
154 *
155 * XArg would resolve to 'X&&', and YArg would resolve to 'Y&'.
156 *
157 * - Second, Whereas C++ used to forbid references to references, C++11 defines
158 * 'collapsing rules': 'T& &', 'T&& &', and 'T& &&' (that is, any combination
159 * involving an lvalue reference) now collapse to simply 'T&'; and 'T&& &&'
160 * collapses to 'T&&'.
161 *
162 * Thus, in the call above, 'XArg&&' is 'X&& &&', collapsing to 'X&&'; and
163 * 'YArg&&' is 'Y& &&', which collapses to 'Y &'. Because the arguments are
164 * declared as rvalue references to template arguments, the rvalue-ness
165 * "shines through" where present.
166 *
167 * Then, the 'Forward<T>' function --- you must invoke 'Forward' with its type
168 * argument --- returns an lvalue reference or an rvalue reference to its
169 * argument, depending on what T is. In our unified constructor definition, that
170 * means that we'll invoke either the copy or move constructors for x and y,
171 * depending on what we gave C's constructor. In our call, we'll move 'foo()'
172 * into 'x', but copy 'yy' into 'y'.
173 *
174 * This header file defines Move and Forward in the mozilla namespace. It's up
175 * to individual containers to annotate moves as such, by calling Move; and it's
176 * up to individual types to define move constructors and assignment operators
177 * when valuable.
178 *
179 * (C++11 says that the <utility> header file should define 'std::move' and
180 * 'std::forward', which are just like our 'Move' and 'Forward'; but those
181 * definitions aren't available in that header on all our platforms, so we
182 * define them ourselves here.)
183 *
184 * 0. This pattern is known as "perfect forwarding". Interestingly, it is not
185 * actually perfect, and it can't forward all possible argument expressions!
186 * There are two issues: one that's a C++11 issue, and one that's a legacy
187 * compiler issue.
188 *
189 * The C++11 issue is that you can't form a reference to a bit-field. As a
190 * workaround, assign the bit-field to a local variable and use that:
191 *
192 * // C is as above
193 * struct S { int x : 1; } s;
194 * C(s.x, 0); // BAD: s.x is a reference to a bit-field, can't form those
195 * int tmp = s.x;
196 * C(tmp, 0); // OK: tmp not a bit-field
197 *
198 * The legacy issue is that when we don't have true nullptr and must emulate
199 * it (gcc 4.4/4.5), forwarding |nullptr| results in an |int| or |long|
200 * forwarded reference. But such a reference, even if its value is a null
201 * pointer constant expression, is not itself a null pointer constant
202 * expression. This causes -Werror=conversion-null errors and pointer-to-
203 * integer comparison errors. Until we always have true nullptr, users of
204 * forwarding methods must not pass |nullptr| to them.
205 */
207 /**
208 * Identical to std::Move(); this is necessary until our stlport supports
209 * std::move().
210 */
214 {
216 }
218 /**
219 * These two overloads are identical to std::forward(); they are necessary until
220 * our stlport supports std::forward().
221 */
225 {
227 }
232 {
236 }
238 /** Swap |aX| and |aY| using move-construction if possible. */
242 {
246 }