1 @c Copyright (C) 1988,89,92,93,94,96,98,99,2000 Free Software Foundation, Inc.
2 @c This is part of the GCC manual.
3 @c For copying conditions, see the file gcc.texi.
6 @chapter Extensions to the C Language Family
7 @cindex extensions, C language
8 @cindex C language extensions
10 GNU C provides several language features not found in ANSI standard C.
11 (The @samp{-pedantic} option directs GNU CC to print a warning message if
12 any of these features is used.) To test for the availability of these
13 features in conditional compilation, check for a predefined macro
14 @code{__GNUC__}, which is always defined under GNU CC.
16 These extensions are available in C and Objective C. Most of them are
17 also available in C++. @xref{C++ Extensions,,Extensions to the
18 C++ Language}, for extensions that apply @emph{only} to C++.
20 @c The only difference between the two versions of this menu is that the
21 @c version for clear INTERNALS has an extra node, "Constraints" (which
22 @c appears in a separate chapter in the other version of the manual).
25 * Statement Exprs:: Putting statements and declarations inside expressions.
26 * Local Labels:: Labels local to a statement-expression.
27 * Labels as Values:: Getting pointers to labels, and computed gotos.
28 * Nested Functions:: As in Algol and Pascal, lexical scoping of functions.
29 * Constructing Calls:: Dispatching a call to another function.
30 * Naming Types:: Giving a name to the type of some expression.
31 * Typeof:: @code{typeof}: referring to the type of an expression.
32 * Lvalues:: Using @samp{?:}, @samp{,} and casts in lvalues.
33 * Conditionals:: Omitting the middle operand of a @samp{?:} expression.
34 * Long Long:: Double-word integers---@code{long long int}.
35 * Complex:: Data types for complex numbers.
36 * Hex Floats:: Hexadecimal floating-point constants.
37 * Zero Length:: Zero-length arrays.
38 * Variable Length:: Arrays whose length is computed at run time.
39 * Macro Varargs:: Macros with variable number of arguments.
40 * Subscripting:: Any array can be subscripted, even if not an lvalue.
41 * Pointer Arith:: Arithmetic on @code{void}-pointers and function pointers.
42 * Initializers:: Non-constant initializers.
43 * Constructors:: Constructor expressions give structures, unions
45 * Labeled Elements:: Labeling elements of initializers.
46 * Cast to Union:: Casting to union type from any member of the union.
47 * Case Ranges:: `case 1 ... 9' and such.
48 * Function Attributes:: Declaring that functions have no side effects,
49 or that they can never return.
50 * Function Prototypes:: Prototype declarations and old-style definitions.
51 * C++ Comments:: C++ comments are recognized.
52 * Dollar Signs:: Dollar sign is allowed in identifiers.
53 * Character Escapes:: @samp{\e} stands for the character @key{ESC}.
54 * Variable Attributes:: Specifying attributes of variables.
55 * Type Attributes:: Specifying attributes of types.
56 * Alignment:: Inquiring about the alignment of a type or variable.
57 * Inline:: Defining inline functions (as fast as macros).
58 * Extended Asm:: Assembler instructions with C expressions as operands.
59 (With them you can define ``built-in'' functions.)
60 * Asm Labels:: Specifying the assembler name to use for a C symbol.
61 * Explicit Reg Vars:: Defining variables residing in specified registers.
62 * Alternate Keywords:: @code{__const__}, @code{__asm__}, etc., for header files.
63 * Incomplete Enums:: @code{enum foo;}, with details to follow.
64 * Function Names:: Printable strings which are the name of the current
66 * Return Address:: Getting the return or frame address of a function.
67 * Other Builtins:: Other built-in functions.
68 * Deprecated Features:: Things might disappear from g++.
69 * Backwards Compatibility:: Compatibilities with earlier definitions of C++.
74 * Statement Exprs:: Putting statements and declarations inside expressions.
75 * Local Labels:: Labels local to a statement-expression.
76 * Labels as Values:: Getting pointers to labels, and computed gotos.
77 * Nested Functions:: As in Algol and Pascal, lexical scoping of functions.
78 * Constructing Calls:: Dispatching a call to another function.
79 * Naming Types:: Giving a name to the type of some expression.
80 * Typeof:: @code{typeof}: referring to the type of an expression.
81 * Lvalues:: Using @samp{?:}, @samp{,} and casts in lvalues.
82 * Conditionals:: Omitting the middle operand of a @samp{?:} expression.
83 * Long Long:: Double-word integers---@code{long long int}.
84 * Complex:: Data types for complex numbers.
85 * Hex Floats:: Hexadecimal floating-point constants.
86 * Zero Length:: Zero-length arrays.
87 * Variable Length:: Arrays whose length is computed at run time.
88 * Macro Varargs:: Macros with variable number of arguments.
89 * Subscripting:: Any array can be subscripted, even if not an lvalue.
90 * Pointer Arith:: Arithmetic on @code{void}-pointers and function pointers.
91 * Initializers:: Non-constant initializers.
92 * Constructors:: Constructor expressions give structures, unions
94 * Labeled Elements:: Labeling elements of initializers.
95 * Cast to Union:: Casting to union type from any member of the union.
96 * Case Ranges:: `case 1 ... 9' and such.
97 * Function Attributes:: Declaring that functions have no side effects,
98 or that they can never return.
99 * Function Prototypes:: Prototype declarations and old-style definitions.
100 * C++ Comments:: C++ comments are recognized.
101 * Dollar Signs:: Dollar sign is allowed in identifiers.
102 * Character Escapes:: @samp{\e} stands for the character @key{ESC}.
103 * Variable Attributes:: Specifying attributes of variables.
104 * Type Attributes:: Specifying attributes of types.
105 * Alignment:: Inquiring about the alignment of a type or variable.
106 * Inline:: Defining inline functions (as fast as macros).
107 * Extended Asm:: Assembler instructions with C expressions as operands.
108 (With them you can define ``built-in'' functions.)
109 * Constraints:: Constraints for asm operands
110 * Asm Labels:: Specifying the assembler name to use for a C symbol.
111 * Explicit Reg Vars:: Defining variables residing in specified registers.
112 * Alternate Keywords:: @code{__const__}, @code{__asm__}, etc., for header files.
113 * Incomplete Enums:: @code{enum foo;}, with details to follow.
114 * Function Names:: Printable strings which are the name of the current
116 * Return Address:: Getting the return or frame address of a function.
117 * Deprecated Features:: Things might disappear from g++.
118 * Backwards Compatibility:: Compatibilities with earlier definitions of C++.
119 * Other Builtins:: Other built-in functions.
123 @node Statement Exprs
124 @section Statements and Declarations in Expressions
125 @cindex statements inside expressions
126 @cindex declarations inside expressions
127 @cindex expressions containing statements
128 @cindex macros, statements in expressions
130 @c the above section title wrapped and causes an underfull hbox.. i
131 @c changed it from "within" to "in". --mew 4feb93
133 A compound statement enclosed in parentheses may appear as an expression
134 in GNU C. This allows you to use loops, switches, and local variables
135 within an expression.
137 Recall that a compound statement is a sequence of statements surrounded
138 by braces; in this construct, parentheses go around the braces. For
142 (@{ int y = foo (); int z;
149 is a valid (though slightly more complex than necessary) expression
150 for the absolute value of @code{foo ()}.
152 The last thing in the compound statement should be an expression
153 followed by a semicolon; the value of this subexpression serves as the
154 value of the entire construct. (If you use some other kind of statement
155 last within the braces, the construct has type @code{void}, and thus
156 effectively no value.)
158 This feature is especially useful in making macro definitions ``safe'' (so
159 that they evaluate each operand exactly once). For example, the
160 ``maximum'' function is commonly defined as a macro in standard C as
164 #define max(a,b) ((a) > (b) ? (a) : (b))
168 @cindex side effects, macro argument
169 But this definition computes either @var{a} or @var{b} twice, with bad
170 results if the operand has side effects. In GNU C, if you know the
171 type of the operands (here let's assume @code{int}), you can define
172 the macro safely as follows:
175 #define maxint(a,b) \
176 (@{int _a = (a), _b = (b); _a > _b ? _a : _b; @})
179 Embedded statements are not allowed in constant expressions, such as
180 the value of an enumeration constant, the width of a bit field, or
181 the initial value of a static variable.
183 If you don't know the type of the operand, you can still do this, but you
184 must use @code{typeof} (@pxref{Typeof}) or type naming (@pxref{Naming
187 Statement expressions are not supported fully in G++, and their fate
188 there is unclear. (It is possible that they will become fully supported
189 at some point, or that they will be deprecated, or that the bugs that
190 are present will continue to exist indefinitely.) Presently, statement
191 expressions do not work well as default arguments.
193 In addition, there are semantic issues with statement-expressions in
194 C++. If you try to use statement-expressions instead of inline
195 functions in C++, you may be surprised at the way object destruction is
196 handled. For example:
199 #define foo(a) (@{int b = (a); b + 3; @})
203 does not work the same way as:
206 inline int foo(int a) @{ int b = a; return b + 3; @}
210 In particular, if the expression passed into @code{foo} involves the
211 creation of temporaries, the destructors for those temporaries will be
212 run earlier in the case of the macro than in the case of the function.
214 These considerations mean that it is probably a bad idea to use
215 statement-expressions of this form in header files that are designed to
216 work with C++. (Note that some versions of the GNU C Library contained
217 header files using statement-expression that lead to precisely this
221 @section Locally Declared Labels
223 @cindex macros, local labels
225 Each statement expression is a scope in which @dfn{local labels} can be
226 declared. A local label is simply an identifier; you can jump to it
227 with an ordinary @code{goto} statement, but only from within the
228 statement expression it belongs to.
230 A local label declaration looks like this:
233 __label__ @var{label};
240 __label__ @var{label1}, @var{label2}, @dots{};
243 Local label declarations must come at the beginning of the statement
244 expression, right after the @samp{(@{}, before any ordinary
247 The label declaration defines the label @emph{name}, but does not define
248 the label itself. You must do this in the usual way, with
249 @code{@var{label}:}, within the statements of the statement expression.
251 The local label feature is useful because statement expressions are
252 often used in macros. If the macro contains nested loops, a @code{goto}
253 can be useful for breaking out of them. However, an ordinary label
254 whose scope is the whole function cannot be used: if the macro can be
255 expanded several times in one function, the label will be multiply
256 defined in that function. A local label avoids this problem. For
260 #define SEARCH(array, target) \
263 typeof (target) _SEARCH_target = (target); \
264 typeof (*(array)) *_SEARCH_array = (array); \
267 for (i = 0; i < max; i++) \
268 for (j = 0; j < max; j++) \
269 if (_SEARCH_array[i][j] == _SEARCH_target) \
270 @{ value = i; goto found; @} \
277 @node Labels as Values
278 @section Labels as Values
279 @cindex labels as values
280 @cindex computed gotos
281 @cindex goto with computed label
282 @cindex address of a label
284 You can get the address of a label defined in the current function
285 (or a containing function) with the unary operator @samp{&&}. The
286 value has type @code{void *}. This value is a constant and can be used
287 wherever a constant of that type is valid. For example:
295 To use these values, you need to be able to jump to one. This is done
296 with the computed goto statement@footnote{The analogous feature in
297 Fortran is called an assigned goto, but that name seems inappropriate in
298 C, where one can do more than simply store label addresses in label
299 variables.}, @code{goto *@var{exp};}. For example,
306 Any expression of type @code{void *} is allowed.
308 One way of using these constants is in initializing a static array that
309 will serve as a jump table:
312 static void *array[] = @{ &&foo, &&bar, &&hack @};
315 Then you can select a label with indexing, like this:
322 Note that this does not check whether the subscript is in bounds---array
323 indexing in C never does that.
325 Such an array of label values serves a purpose much like that of the
326 @code{switch} statement. The @code{switch} statement is cleaner, so
327 use that rather than an array unless the problem does not fit a
328 @code{switch} statement very well.
330 Another use of label values is in an interpreter for threaded code.
331 The labels within the interpreter function can be stored in the
332 threaded code for super-fast dispatching.
334 You may not use this mechanism to jump to code in a different function.
335 If you do that, totally unpredictable things will happen. The best way to
336 avoid this is to store the label address only in automatic variables and
337 never pass it as an argument.
339 An alternate way to write the above example is
342 static const int array[] = @{ &&foo - &&foo, &&bar - &&foo, &&hack - &&foo @};
343 goto *(&&foo + array[i]);
347 This is more friendly to code living in shared libraries, as it reduces
348 the number of dynamic relocations that are needed, and by consequence,
349 allows the data to be read-only.
351 @node Nested Functions
352 @section Nested Functions
353 @cindex nested functions
354 @cindex downward funargs
357 A @dfn{nested function} is a function defined inside another function.
358 (Nested functions are not supported for GNU C++.) The nested function's
359 name is local to the block where it is defined. For example, here we
360 define a nested function named @code{square}, and call it twice:
364 foo (double a, double b)
366 double square (double z) @{ return z * z; @}
368 return square (a) + square (b);
373 The nested function can access all the variables of the containing
374 function that are visible at the point of its definition. This is
375 called @dfn{lexical scoping}. For example, here we show a nested
376 function which uses an inherited variable named @code{offset}:
379 bar (int *array, int offset, int size)
381 int access (int *array, int index)
382 @{ return array[index + offset]; @}
385 for (i = 0; i < size; i++)
386 @dots{} access (array, i) @dots{}
390 Nested function definitions are permitted within functions in the places
391 where variable definitions are allowed; that is, in any block, before
392 the first statement in the block.
394 It is possible to call the nested function from outside the scope of its
395 name by storing its address or passing the address to another function:
398 hack (int *array, int size)
400 void store (int index, int value)
401 @{ array[index] = value; @}
403 intermediate (store, size);
407 Here, the function @code{intermediate} receives the address of
408 @code{store} as an argument. If @code{intermediate} calls @code{store},
409 the arguments given to @code{store} are used to store into @code{array}.
410 But this technique works only so long as the containing function
411 (@code{hack}, in this example) does not exit.
413 If you try to call the nested function through its address after the
414 containing function has exited, all hell will break loose. If you try
415 to call it after a containing scope level has exited, and if it refers
416 to some of the variables that are no longer in scope, you may be lucky,
417 but it's not wise to take the risk. If, however, the nested function
418 does not refer to anything that has gone out of scope, you should be
421 GNU CC implements taking the address of a nested function using a
422 technique called @dfn{trampolines}. A paper describing them is
423 available as @uref{http://master.debian.org/~karlheg/Usenix88-lexic.pdf}.
425 A nested function can jump to a label inherited from a containing
426 function, provided the label was explicitly declared in the containing
427 function (@pxref{Local Labels}). Such a jump returns instantly to the
428 containing function, exiting the nested function which did the
429 @code{goto} and any intermediate functions as well. Here is an example:
433 bar (int *array, int offset, int size)
436 int access (int *array, int index)
440 return array[index + offset];
444 for (i = 0; i < size; i++)
445 @dots{} access (array, i) @dots{}
449 /* @r{Control comes here from @code{access}
450 if it detects an error.} */
457 A nested function always has internal linkage. Declaring one with
458 @code{extern} is erroneous. If you need to declare the nested function
459 before its definition, use @code{auto} (which is otherwise meaningless
460 for function declarations).
463 bar (int *array, int offset, int size)
466 auto int access (int *, int);
468 int access (int *array, int index)
472 return array[index + offset];
478 @node Constructing Calls
479 @section Constructing Function Calls
480 @cindex constructing calls
481 @cindex forwarding calls
483 Using the built-in functions described below, you can record
484 the arguments a function received, and call another function
485 with the same arguments, without knowing the number or types
488 You can also record the return value of that function call,
489 and later return that value, without knowing what data type
490 the function tried to return (as long as your caller expects
494 @findex __builtin_apply_args
495 @item __builtin_apply_args ()
496 This built-in function returns a pointer of type @code{void *} to data
497 describing how to perform a call with the same arguments as were passed
498 to the current function.
500 The function saves the arg pointer register, structure value address,
501 and all registers that might be used to pass arguments to a function
502 into a block of memory allocated on the stack. Then it returns the
503 address of that block.
505 @findex __builtin_apply
506 @item __builtin_apply (@var{function}, @var{arguments}, @var{size})
507 This built-in function invokes @var{function} (type @code{void (*)()})
508 with a copy of the parameters described by @var{arguments} (type
509 @code{void *}) and @var{size} (type @code{int}).
511 The value of @var{arguments} should be the value returned by
512 @code{__builtin_apply_args}. The argument @var{size} specifies the size
513 of the stack argument data, in bytes.
515 This function returns a pointer of type @code{void *} to data describing
516 how to return whatever value was returned by @var{function}. The data
517 is saved in a block of memory allocated on the stack.
519 It is not always simple to compute the proper value for @var{size}. The
520 value is used by @code{__builtin_apply} to compute the amount of data
521 that should be pushed on the stack and copied from the incoming argument
524 @findex __builtin_return
525 @item __builtin_return (@var{result})
526 This built-in function returns the value described by @var{result} from
527 the containing function. You should specify, for @var{result}, a value
528 returned by @code{__builtin_apply}.
532 @section Naming an Expression's Type
535 You can give a name to the type of an expression using a @code{typedef}
536 declaration with an initializer. Here is how to define @var{name} as a
537 type name for the type of @var{exp}:
540 typedef @var{name} = @var{exp};
543 This is useful in conjunction with the statements-within-expressions
544 feature. Here is how the two together can be used to define a safe
545 ``maximum'' macro that operates on any arithmetic type:
549 (@{typedef _ta = (a), _tb = (b); \
550 _ta _a = (a); _tb _b = (b); \
551 _a > _b ? _a : _b; @})
554 @cindex underscores in variables in macros
555 @cindex @samp{_} in variables in macros
556 @cindex local variables in macros
557 @cindex variables, local, in macros
558 @cindex macros, local variables in
560 The reason for using names that start with underscores for the local
561 variables is to avoid conflicts with variable names that occur within the
562 expressions that are substituted for @code{a} and @code{b}. Eventually we
563 hope to design a new form of declaration syntax that allows you to declare
564 variables whose scopes start only after their initializers; this will be a
565 more reliable way to prevent such conflicts.
568 @section Referring to a Type with @code{typeof}
571 @cindex macros, types of arguments
573 Another way to refer to the type of an expression is with @code{typeof}.
574 The syntax of using of this keyword looks like @code{sizeof}, but the
575 construct acts semantically like a type name defined with @code{typedef}.
577 There are two ways of writing the argument to @code{typeof}: with an
578 expression or with a type. Here is an example with an expression:
585 This assumes that @code{x} is an array of functions; the type described
586 is that of the values of the functions.
588 Here is an example with a typename as the argument:
595 Here the type described is that of pointers to @code{int}.
597 If you are writing a header file that must work when included in ANSI C
598 programs, write @code{__typeof__} instead of @code{typeof}.
599 @xref{Alternate Keywords}.
601 A @code{typeof}-construct can be used anywhere a typedef name could be
602 used. For example, you can use it in a declaration, in a cast, or inside
603 of @code{sizeof} or @code{typeof}.
607 This declares @code{y} with the type of what @code{x} points to.
614 This declares @code{y} as an array of such values.
621 This declares @code{y} as an array of pointers to characters:
624 typeof (typeof (char *)[4]) y;
628 It is equivalent to the following traditional C declaration:
634 To see the meaning of the declaration using @code{typeof}, and why it
635 might be a useful way to write, let's rewrite it with these macros:
638 #define pointer(T) typeof(T *)
639 #define array(T, N) typeof(T [N])
643 Now the declaration can be rewritten this way:
646 array (pointer (char), 4) y;
650 Thus, @code{array (pointer (char), 4)} is the type of arrays of 4
651 pointers to @code{char}.
655 @section Generalized Lvalues
656 @cindex compound expressions as lvalues
657 @cindex expressions, compound, as lvalues
658 @cindex conditional expressions as lvalues
659 @cindex expressions, conditional, as lvalues
660 @cindex casts as lvalues
661 @cindex generalized lvalues
662 @cindex lvalues, generalized
663 @cindex extensions, @code{?:}
664 @cindex @code{?:} extensions
665 Compound expressions, conditional expressions and casts are allowed as
666 lvalues provided their operands are lvalues. This means that you can take
667 their addresses or store values into them.
669 Standard C++ allows compound expressions and conditional expressions as
670 lvalues, and permits casts to reference type, so use of this extension
671 is deprecated for C++ code.
673 For example, a compound expression can be assigned, provided the last
674 expression in the sequence is an lvalue. These two expressions are
682 Similarly, the address of the compound expression can be taken. These two
683 expressions are equivalent:
690 A conditional expression is a valid lvalue if its type is not void and the
691 true and false branches are both valid lvalues. For example, these two
692 expressions are equivalent:
696 (a ? b = 5 : (c = 5))
699 A cast is a valid lvalue if its operand is an lvalue. A simple
700 assignment whose left-hand side is a cast works by converting the
701 right-hand side first to the specified type, then to the type of the
702 inner left-hand side expression. After this is stored, the value is
703 converted back to the specified type to become the value of the
704 assignment. Thus, if @code{a} has type @code{char *}, the following two
705 expressions are equivalent:
709 (int)(a = (char *)(int)5)
712 An assignment-with-arithmetic operation such as @samp{+=} applied to a cast
713 performs the arithmetic using the type resulting from the cast, and then
714 continues as in the previous case. Therefore, these two expressions are
719 (int)(a = (char *)(int) ((int)a + 5))
722 You cannot take the address of an lvalue cast, because the use of its
723 address would not work out coherently. Suppose that @code{&(int)f} were
724 permitted, where @code{f} has type @code{float}. Then the following
725 statement would try to store an integer bit-pattern where a floating
726 point number belongs:
732 This is quite different from what @code{(int)f = 1} would do---that
733 would convert 1 to floating point and store it. Rather than cause this
734 inconsistency, we think it is better to prohibit use of @samp{&} on a cast.
736 If you really do want an @code{int *} pointer with the address of
737 @code{f}, you can simply write @code{(int *)&f}.
740 @section Conditionals with Omitted Operands
741 @cindex conditional expressions, extensions
742 @cindex omitted middle-operands
743 @cindex middle-operands, omitted
744 @cindex extensions, @code{?:}
745 @cindex @code{?:} extensions
747 The middle operand in a conditional expression may be omitted. Then
748 if the first operand is nonzero, its value is the value of the conditional
751 Therefore, the expression
758 has the value of @code{x} if that is nonzero; otherwise, the value of
761 This example is perfectly equivalent to
767 @cindex side effect in ?:
768 @cindex ?: side effect
770 In this simple case, the ability to omit the middle operand is not
771 especially useful. When it becomes useful is when the first operand does,
772 or may (if it is a macro argument), contain a side effect. Then repeating
773 the operand in the middle would perform the side effect twice. Omitting
774 the middle operand uses the value already computed without the undesirable
775 effects of recomputing it.
778 @section Double-Word Integers
779 @cindex @code{long long} data types
780 @cindex double-word arithmetic
781 @cindex multiprecision arithmetic
783 GNU C supports data types for integers that are twice as long as
784 @code{int}. Simply write @code{long long int} for a signed integer, or
785 @code{unsigned long long int} for an unsigned integer. To make an
786 integer constant of type @code{long long int}, add the suffix @code{LL}
787 to the integer. To make an integer constant of type @code{unsigned long
788 long int}, add the suffix @code{ULL} to the integer.
790 You can use these types in arithmetic like any other integer types.
791 Addition, subtraction, and bitwise boolean operations on these types
792 are open-coded on all types of machines. Multiplication is open-coded
793 if the machine supports fullword-to-doubleword a widening multiply
794 instruction. Division and shifts are open-coded only on machines that
795 provide special support. The operations that are not open-coded use
796 special library routines that come with GNU CC.
798 There may be pitfalls when you use @code{long long} types for function
799 arguments, unless you declare function prototypes. If a function
800 expects type @code{int} for its argument, and you pass a value of type
801 @code{long long int}, confusion will result because the caller and the
802 subroutine will disagree about the number of bytes for the argument.
803 Likewise, if the function expects @code{long long int} and you pass
804 @code{int}. The best way to avoid such problems is to use prototypes.
807 @section Complex Numbers
808 @cindex complex numbers
810 GNU C supports complex data types. You can declare both complex integer
811 types and complex floating types, using the keyword @code{__complex__}.
813 For example, @samp{__complex__ double x;} declares @code{x} as a
814 variable whose real part and imaginary part are both of type
815 @code{double}. @samp{__complex__ short int y;} declares @code{y} to
816 have real and imaginary parts of type @code{short int}; this is not
817 likely to be useful, but it shows that the set of complex types is
820 To write a constant with a complex data type, use the suffix @samp{i} or
821 @samp{j} (either one; they are equivalent). For example, @code{2.5fi}
822 has type @code{__complex__ float} and @code{3i} has type
823 @code{__complex__ int}. Such a constant always has a pure imaginary
824 value, but you can form any complex value you like by adding one to a
827 To extract the real part of a complex-valued expression @var{exp}, write
828 @code{__real__ @var{exp}}. Likewise, use @code{__imag__} to
829 extract the imaginary part.
831 The operator @samp{~} performs complex conjugation when used on a value
834 GNU CC can allocate complex automatic variables in a noncontiguous
835 fashion; it's even possible for the real part to be in a register while
836 the imaginary part is on the stack (or vice-versa). None of the
837 supported debugging info formats has a way to represent noncontiguous
838 allocation like this, so GNU CC describes a noncontiguous complex
839 variable as if it were two separate variables of noncomplex type.
840 If the variable's actual name is @code{foo}, the two fictitious
841 variables are named @code{foo$real} and @code{foo$imag}. You can
842 examine and set these two fictitious variables with your debugger.
844 A future version of GDB will know how to recognize such pairs and treat
845 them as a single variable with a complex type.
851 GNU CC recognizes floating-point numbers written not only in the usual
852 decimal notation, such as @code{1.55e1}, but also numbers such as
853 @code{0x1.fp3} written in hexadecimal format. In that format the
854 @code{0x} hex introducer and the @code{p} or @code{P} exponent field are
855 mandatory. The exponent is a decimal number that indicates the power of
856 2 by which the significant part will be multiplied. Thus @code{0x1.f} is
857 1 15/16, @code{p3} multiplies it by 8, and the value of @code{0x1.fp3}
858 is the same as @code{1.55e1}.
860 Unlike for floating-point numbers in the decimal notation the exponent
861 is always required in the hexadecimal notation. Otherwise the compiler
862 would not be able to resolve the ambiguity of, e.g., @code{0x1.f}. This
863 could mean @code{1.0f} or @code{1.9375} since @code{f} is also the
864 extension for floating-point constants of type @code{float}.
867 @section Arrays of Length Zero
868 @cindex arrays of length zero
869 @cindex zero-length arrays
870 @cindex length-zero arrays
872 Zero-length arrays are allowed in GNU C. They are very useful as the last
873 element of a structure which is really a header for a variable-length
883 struct line *thisline = (struct line *)
884 malloc (sizeof (struct line) + this_length);
885 thisline->length = this_length;
889 In ISO C89, you would have to give @code{contents} a length of 1, which
890 means either you waste space or complicate the argument to @code{malloc}.
892 In ISO C99, you would use a @dfn{flexible array member}, which uses a
893 slightly different syntax: leave out the @code{0} and write
896 GCC allows static initialization of the zero-length array if
897 the structure is not nested inside another structure. I.e.
901 struct line x = @{ 4, @{ 'g', 'o', 'o', 'd' @} @};
906 @} y = @{ @{ 3, @{ 'b', 'a', 'd' @} @} @};
909 @node Variable Length
910 @section Arrays of Variable Length
911 @cindex variable-length arrays
912 @cindex arrays of variable length
914 Variable-length automatic arrays are allowed in GNU C. These arrays are
915 declared like any other automatic arrays, but with a length that is not
916 a constant expression. The storage is allocated at the point of
917 declaration and deallocated when the brace-level is exited. For
922 concat_fopen (char *s1, char *s2, char *mode)
924 char str[strlen (s1) + strlen (s2) + 1];
927 return fopen (str, mode);
931 @cindex scope of a variable length array
932 @cindex variable-length array scope
933 @cindex deallocating variable length arrays
934 Jumping or breaking out of the scope of the array name deallocates the
935 storage. Jumping into the scope is not allowed; you get an error
938 @cindex @code{alloca} vs variable-length arrays
939 You can use the function @code{alloca} to get an effect much like
940 variable-length arrays. The function @code{alloca} is available in
941 many other C implementations (but not in all). On the other hand,
942 variable-length arrays are more elegant.
944 There are other differences between these two methods. Space allocated
945 with @code{alloca} exists until the containing @emph{function} returns.
946 The space for a variable-length array is deallocated as soon as the array
947 name's scope ends. (If you use both variable-length arrays and
948 @code{alloca} in the same function, deallocation of a variable-length array
949 will also deallocate anything more recently allocated with @code{alloca}.)
951 You can also use variable-length arrays as arguments to functions:
955 tester (int len, char data[len][len])
961 The length of an array is computed once when the storage is allocated
962 and is remembered for the scope of the array in case you access it with
965 If you want to pass the array first and the length afterward, you can
966 use a forward declaration in the parameter list---another GNU extension.
970 tester (int len; char data[len][len], int len)
976 @cindex parameter forward declaration
977 The @samp{int len} before the semicolon is a @dfn{parameter forward
978 declaration}, and it serves the purpose of making the name @code{len}
979 known when the declaration of @code{data} is parsed.
981 You can write any number of such parameter forward declarations in the
982 parameter list. They can be separated by commas or semicolons, but the
983 last one must end with a semicolon, which is followed by the ``real''
984 parameter declarations. Each forward declaration must match a ``real''
985 declaration in parameter name and data type.
988 @section Macros with Variable Numbers of Arguments
989 @cindex variable number of arguments
990 @cindex macro with variable arguments
991 @cindex rest argument (in macro)
993 In GNU C, a macro can accept a variable number of arguments, much as a
994 function can. The syntax for defining the macro looks much like that
995 used for a function. Here is an example:
998 #define eprintf(format, args...) \
999 fprintf (stderr, format , ## args)
1002 Here @code{args} is a @dfn{rest argument}: it takes in zero or more
1003 arguments, as many as the call contains. All of them plus the commas
1004 between them form the value of @code{args}, which is substituted into
1005 the macro body where @code{args} is used. Thus, we have this expansion:
1008 eprintf ("%s:%d: ", input_file_name, line_number)
1010 fprintf (stderr, "%s:%d: " , input_file_name, line_number)
1014 Note that the comma after the string constant comes from the definition
1015 of @code{eprintf}, whereas the last comma comes from the value of
1018 The reason for using @samp{##} is to handle the case when @code{args}
1019 matches no arguments at all. In this case, @code{args} has an empty
1020 value. In this case, the second comma in the definition becomes an
1021 embarrassment: if it got through to the expansion of the macro, we would
1022 get something like this:
1025 fprintf (stderr, "success!\n" , )
1029 which is invalid C syntax. @samp{##} gets rid of the comma, so we get
1030 the following instead:
1033 fprintf (stderr, "success!\n")
1036 This is a special feature of the GNU C preprocessor: @samp{##} before a
1037 rest argument that is empty discards the preceding sequence of
1038 non-whitespace characters from the macro definition. (If another macro
1039 argument precedes, none of it is discarded.)
1041 It might be better to discard the last preprocessor token instead of the
1042 last preceding sequence of non-whitespace characters; in fact, we may
1043 someday change this feature to do so. We advise you to write the macro
1044 definition so that the preceding sequence of non-whitespace characters
1045 is just a single token, so that the meaning will not change if we change
1046 the definition of this feature.
1049 @section Non-Lvalue Arrays May Have Subscripts
1050 @cindex subscripting
1051 @cindex arrays, non-lvalue
1053 @cindex subscripting and function values
1054 Subscripting is allowed on arrays that are not lvalues, even though the
1055 unary @samp{&} operator is not. For example, this is valid in GNU C though
1056 not valid in other C dialects:
1060 struct foo @{int a[4];@};
1066 return f().a[index];
1072 @section Arithmetic on @code{void}- and Function-Pointers
1073 @cindex void pointers, arithmetic
1074 @cindex void, size of pointer to
1075 @cindex function pointers, arithmetic
1076 @cindex function, size of pointer to
1078 In GNU C, addition and subtraction operations are supported on pointers to
1079 @code{void} and on pointers to functions. This is done by treating the
1080 size of a @code{void} or of a function as 1.
1082 A consequence of this is that @code{sizeof} is also allowed on @code{void}
1083 and on function types, and returns 1.
1085 The option @samp{-Wpointer-arith} requests a warning if these extensions
1089 @section Non-Constant Initializers
1090 @cindex initializers, non-constant
1091 @cindex non-constant initializers
1093 As in standard C++, the elements of an aggregate initializer for an
1094 automatic variable are not required to be constant expressions in GNU C.
1095 Here is an example of an initializer with run-time varying elements:
1098 foo (float f, float g)
1100 float beat_freqs[2] = @{ f-g, f+g @};
1106 @section Constructor Expressions
1107 @cindex constructor expressions
1108 @cindex initializations in expressions
1109 @cindex structures, constructor expression
1110 @cindex expressions, constructor
1112 GNU C supports constructor expressions. A constructor looks like
1113 a cast containing an initializer. Its value is an object of the
1114 type specified in the cast, containing the elements specified in
1117 Usually, the specified type is a structure. Assume that
1118 @code{struct foo} and @code{structure} are declared as shown:
1121 struct foo @{int a; char b[2];@} structure;
1125 Here is an example of constructing a @code{struct foo} with a constructor:
1128 structure = ((struct foo) @{x + y, 'a', 0@});
1132 This is equivalent to writing the following:
1136 struct foo temp = @{x + y, 'a', 0@};
1141 You can also construct an array. If all the elements of the constructor
1142 are (made up of) simple constant expressions, suitable for use in
1143 initializers, then the constructor is an lvalue and can be coerced to a
1144 pointer to its first element, as shown here:
1147 char **foo = (char *[]) @{ "x", "y", "z" @};
1150 Array constructors whose elements are not simple constants are
1151 not very useful, because the constructor is not an lvalue. There
1152 are only two valid ways to use it: to subscript it, or initialize
1153 an array variable with it. The former is probably slower than a
1154 @code{switch} statement, while the latter does the same thing an
1155 ordinary C initializer would do. Here is an example of
1156 subscripting an array constructor:
1159 output = ((int[]) @{ 2, x, 28 @}) [input];
1162 Constructor expressions for scalar types and union types are is
1163 also allowed, but then the constructor expression is equivalent
1166 @node Labeled Elements
1167 @section Labeled Elements in Initializers
1168 @cindex initializers with labeled elements
1169 @cindex labeled elements in initializers
1170 @cindex case labels in initializers
1172 Standard C89 requires the elements of an initializer to appear in a fixed
1173 order, the same as the order of the elements in the array or structure
1176 In ISO C99 you can give the elements in any order, specifying the array
1177 indices or structure field names they apply to, and GNU C allows this as
1178 an extension in C89 mode as well. This extension is not
1179 implemented in GNU C++.
1181 To specify an array index, write
1182 @samp{[@var{index}] =} before the element value. For example,
1185 int a[6] = @{ [4] = 29, [2] = 15 @};
1192 int a[6] = @{ 0, 0, 15, 0, 29, 0 @};
1196 The index values must be constant expressions, even if the array being
1197 initialized is automatic.
1199 An alternative syntax for this which has been obsolete since GCC 2.5 but
1200 GCC still accepts is to write @samp{[@var{index}]} before the element
1201 value, with no @samp{=}.
1203 To initialize a range of elements to the same value, write
1204 @samp{[@var{first} ... @var{last}] = @var{value}}. This is a GNU
1205 extension. For example,
1208 int widths[] = @{ [0 ... 9] = 1, [10 ... 99] = 2, [100] = 3 @};
1212 Note that the length of the array is the highest value specified
1215 In a structure initializer, specify the name of a field to initialize
1216 with @samp{.@var{fieldname} =} before the element value. For example,
1217 given the following structure,
1220 struct point @{ int x, y; @};
1224 the following initialization
1227 struct point p = @{ .y = yvalue, .x = xvalue @};
1234 struct point p = @{ xvalue, yvalue @};
1237 Another syntax which has the same meaning, obsolete since GCC 2.5, is
1238 @samp{@var{fieldname}:}, as shown here:
1241 struct point p = @{ y: yvalue, x: xvalue @};
1244 You can also use an element label (with either the colon syntax or the
1245 period-equal syntax) when initializing a union, to specify which element
1246 of the union should be used. For example,
1249 union foo @{ int i; double d; @};
1251 union foo f = @{ .d = 4 @};
1255 will convert 4 to a @code{double} to store it in the union using
1256 the second element. By contrast, casting 4 to type @code{union foo}
1257 would store it into the union as the integer @code{i}, since it is
1258 an integer. (@xref{Cast to Union}.)
1260 You can combine this technique of naming elements with ordinary C
1261 initialization of successive elements. Each initializer element that
1262 does not have a label applies to the next consecutive element of the
1263 array or structure. For example,
1266 int a[6] = @{ [1] = v1, v2, [4] = v4 @};
1273 int a[6] = @{ 0, v1, v2, 0, v4, 0 @};
1276 Labeling the elements of an array initializer is especially useful
1277 when the indices are characters or belong to an @code{enum} type.
1282 = @{ [' '] = 1, ['\t'] = 1, ['\h'] = 1,
1283 ['\f'] = 1, ['\n'] = 1, ['\r'] = 1 @};
1286 You can also write a series of @samp{.@var{fieldname}} and
1287 @samp{[@var{index}]} element labels before an @samp{=} to specify a
1288 nested subobject to initialize; the list is taken relative to the
1289 subobject corresponding to the closest surrounding brace pair. For
1290 example, with the @samp{struct point} declaration above:
1293 struct point ptarray[10] = @{ [2].y = yv2, [2].x = xv2, [0].x = xv0 @};
1297 @section Case Ranges
1299 @cindex ranges in case statements
1301 You can specify a range of consecutive values in a single @code{case} label,
1305 case @var{low} ... @var{high}:
1309 This has the same effect as the proper number of individual @code{case}
1310 labels, one for each integer value from @var{low} to @var{high}, inclusive.
1312 This feature is especially useful for ranges of ASCII character codes:
1318 @strong{Be careful:} Write spaces around the @code{...}, for otherwise
1319 it may be parsed wrong when you use it with integer values. For example,
1334 @section Cast to a Union Type
1335 @cindex cast to a union
1336 @cindex union, casting to a
1338 A cast to union type is similar to other casts, except that the type
1339 specified is a union type. You can specify the type either with
1340 @code{union @var{tag}} or with a typedef name. A cast to union is actually
1341 a constructor though, not a cast, and hence does not yield an lvalue like
1342 normal casts. (@xref{Constructors}.)
1344 The types that may be cast to the union type are those of the members
1345 of the union. Thus, given the following union and variables:
1348 union foo @{ int i; double d; @};
1354 both @code{x} and @code{y} can be cast to type @code{union} foo.
1356 Using the cast as the right-hand side of an assignment to a variable of
1357 union type is equivalent to storing in a member of the union:
1362 u = (union foo) x @equiv{} u.i = x
1363 u = (union foo) y @equiv{} u.d = y
1366 You can also use the union cast as a function argument:
1369 void hack (union foo);
1371 hack ((union foo) x);
1374 @node Function Attributes
1375 @section Declaring Attributes of Functions
1376 @cindex function attributes
1377 @cindex declaring attributes of functions
1378 @cindex functions that never return
1379 @cindex functions that have no side effects
1380 @cindex functions in arbitrary sections
1381 @cindex functions that behave like malloc
1382 @cindex @code{volatile} applied to function
1383 @cindex @code{const} applied to function
1384 @cindex functions with @code{printf}, @code{scanf} or @code{strftime} style arguments
1385 @cindex functions that are passed arguments in registers on the 386
1386 @cindex functions that pop the argument stack on the 386
1387 @cindex functions that do not pop the argument stack on the 386
1389 In GNU C, you declare certain things about functions called in your program
1390 which help the compiler optimize function calls and check your code more
1393 The keyword @code{__attribute__} allows you to specify special
1394 attributes when making a declaration. This keyword is followed by an
1395 attribute specification inside double parentheses. Ten attributes,
1396 @code{noreturn}, @code{const}, @code{format},
1397 @code{no_instrument_function}, @code{section}, @code{constructor},
1398 @code{destructor}, @code{unused}, @code{weak} and @code{malloc} are
1399 currently defined for functions. Other attributes, including
1400 @code{section} are supported for variables declarations (@pxref{Variable
1401 Attributes}) and for types (@pxref{Type Attributes}).
1403 You may also specify attributes with @samp{__} preceding and following
1404 each keyword. This allows you to use them in header files without
1405 being concerned about a possible macro of the same name. For example,
1406 you may use @code{__noreturn__} instead of @code{noreturn}.
1409 @cindex @code{noreturn} function attribute
1411 A few standard library functions, such as @code{abort} and @code{exit},
1412 cannot return. GNU CC knows this automatically. Some programs define
1413 their own functions that never return. You can declare them
1414 @code{noreturn} to tell the compiler this fact. For example,
1417 void fatal () __attribute__ ((noreturn));
1422 @dots{} /* @r{Print error message.} */ @dots{}
1427 The @code{noreturn} keyword tells the compiler to assume that
1428 @code{fatal} cannot return. It can then optimize without regard to what
1429 would happen if @code{fatal} ever did return. This makes slightly
1430 better code. More importantly, it helps avoid spurious warnings of
1431 uninitialized variables.
1433 Do not assume that registers saved by the calling function are
1434 restored before calling the @code{noreturn} function.
1436 It does not make sense for a @code{noreturn} function to have a return
1437 type other than @code{void}.
1439 The attribute @code{noreturn} is not implemented in GNU C versions
1440 earlier than 2.5. An alternative way to declare that a function does
1441 not return, which works in the current version and in some older
1442 versions, is as follows:
1445 typedef void voidfn ();
1447 volatile voidfn fatal;
1450 @cindex @code{pure} function attribute
1452 Many functions have no effects except the return value and their
1453 return value depends only on the parameters and/or global variables.
1454 Such a function can be subject
1455 to common subexpression elimination and loop optimization just as an
1456 arithmetic operator would be. These functions should be declared
1457 with the attribute @code{pure}. For example,
1460 int square (int) __attribute__ ((pure));
1464 says that the hypothetical function @code{square} is safe to call
1465 fewer times than the program says.
1467 Some of common examples of pure functions are @code{strlen} or @code{memcmp}.
1468 Interesting non-pure functions are functions with infinite loops or those
1469 depending on volatile memory or other system resource, that may change between
1470 two consecutive calls (such as @code{feof} in a multithreading environment).
1472 The attribute @code{pure} is not implemented in GNU C versions earlier
1474 @cindex @code{const} function attribute
1476 Many functions do not examine any values except their arguments, and
1477 have no effects except the return value. Basically this is just slightly
1478 more strict class than the "pure" attribute above, since function is not
1479 allowed to read global memory.
1481 @cindex pointer arguments
1482 Note that a function that has pointer arguments and examines the data
1483 pointed to must @emph{not} be declared @code{const}. Likewise, a
1484 function that calls a non-@code{const} function usually must not be
1485 @code{const}. It does not make sense for a @code{const} function to
1488 The attribute @code{const} is not implemented in GNU C versions earlier
1489 than 2.5. An alternative way to declare that a function has no side
1490 effects, which works in the current version and in some older versions,
1494 typedef int intfn ();
1496 extern const intfn square;
1499 This approach does not work in GNU C++ from 2.6.0 on, since the language
1500 specifies that the @samp{const} must be attached to the return value.
1503 @item format (@var{archetype}, @var{string-index}, @var{first-to-check})
1504 @cindex @code{format} function attribute
1505 The @code{format} attribute specifies that a function takes @code{printf},
1506 @code{scanf}, or @code{strftime} style arguments which should be type-checked
1507 against a format string. For example, the declaration:
1511 my_printf (void *my_object, const char *my_format, ...)
1512 __attribute__ ((format (printf, 2, 3)));
1516 causes the compiler to check the arguments in calls to @code{my_printf}
1517 for consistency with the @code{printf} style format string argument
1520 The parameter @var{archetype} determines how the format string is
1521 interpreted, and should be either @code{printf}, @code{scanf}, or
1522 @code{strftime}. The
1523 parameter @var{string-index} specifies which argument is the format
1524 string argument (starting from 1), while @var{first-to-check} is the
1525 number of the first argument to check against the format string. For
1526 functions where the arguments are not available to be checked (such as
1527 @code{vprintf}), specify the third parameter as zero. In this case the
1528 compiler only checks the format string for consistency.
1530 In the example above, the format string (@code{my_format}) is the second
1531 argument of the function @code{my_print}, and the arguments to check
1532 start with the third argument, so the correct parameters for the format
1533 attribute are 2 and 3.
1535 The @code{format} attribute allows you to identify your own functions
1536 which take format strings as arguments, so that GNU CC can check the
1537 calls to these functions for errors. The compiler always checks formats
1538 for the ANSI library functions @code{printf}, @code{fprintf},
1539 @code{sprintf}, @code{scanf}, @code{fscanf}, @code{sscanf}, @code{strftime},
1540 @code{vprintf}, @code{vfprintf} and @code{vsprintf} whenever such
1541 warnings are requested (using @samp{-Wformat}), so there is no need to
1542 modify the header file @file{stdio.h}.
1544 @item format_arg (@var{string-index})
1545 @cindex @code{format_arg} function attribute
1546 The @code{format_arg} attribute specifies that a function takes
1547 @code{printf} or @code{scanf} style arguments, modifies it (for example,
1548 to translate it into another language), and passes it to a @code{printf}
1549 or @code{scanf} style function. For example, the declaration:
1553 my_dgettext (char *my_domain, const char *my_format)
1554 __attribute__ ((format_arg (2)));
1558 causes the compiler to check the arguments in calls to
1559 @code{my_dgettext} whose result is passed to a @code{printf},
1560 @code{scanf}, or @code{strftime} type function for consistency with the
1561 @code{printf} style format string argument @code{my_format}.
1563 The parameter @var{string-index} specifies which argument is the format
1564 string argument (starting from 1).
1566 The @code{format-arg} attribute allows you to identify your own
1567 functions which modify format strings, so that GNU CC can check the
1568 calls to @code{printf}, @code{scanf}, or @code{strftime} function whose
1569 operands are a call to one of your own function. The compiler always
1570 treats @code{gettext}, @code{dgettext}, and @code{dcgettext} in this
1573 @item no_instrument_function
1574 @cindex @code{no_instrument_function} function attribute
1575 If @samp{-finstrument-functions} is given, profiling function calls will
1576 be generated at entry and exit of most user-compiled functions.
1577 Functions with this attribute will not be so instrumented.
1579 @item section ("section-name")
1580 @cindex @code{section} function attribute
1581 Normally, the compiler places the code it generates in the @code{text} section.
1582 Sometimes, however, you need additional sections, or you need certain
1583 particular functions to appear in special sections. The @code{section}
1584 attribute specifies that a function lives in a particular section.
1585 For example, the declaration:
1588 extern void foobar (void) __attribute__ ((section ("bar")));
1592 puts the function @code{foobar} in the @code{bar} section.
1594 Some file formats do not support arbitrary sections so the @code{section}
1595 attribute is not available on all platforms.
1596 If you need to map the entire contents of a module to a particular
1597 section, consider using the facilities of the linker instead.
1601 @cindex @code{constructor} function attribute
1602 @cindex @code{destructor} function attribute
1603 The @code{constructor} attribute causes the function to be called
1604 automatically before execution enters @code{main ()}. Similarly, the
1605 @code{destructor} attribute causes the function to be called
1606 automatically after @code{main ()} has completed or @code{exit ()} has
1607 been called. Functions with these attributes are useful for
1608 initializing data that will be used implicitly during the execution of
1611 These attributes are not currently implemented for Objective C.
1614 This attribute, attached to a function, means that the function is meant
1615 to be possibly unused. GNU CC will not produce a warning for this
1616 function. GNU C++ does not currently support this attribute as
1617 definitions without parameters are valid in C++.
1620 @cindex @code{weak} attribute
1621 The @code{weak} attribute causes the declaration to be emitted as a weak
1622 symbol rather than a global. This is primarily useful in defining
1623 library functions which can be overridden in user code, though it can
1624 also be used with non-function declarations. Weak symbols are supported
1625 for ELF targets, and also for a.out targets when using the GNU assembler
1629 @cindex @code{malloc} attribute
1630 The @code{malloc} attribute is used to tell the compiler that a function
1631 may be treated as if it were the malloc function. The compiler assumes
1632 that calls to malloc result in a pointers that cannot alias anything.
1633 This will often improve optimization.
1635 @item alias ("target")
1636 @cindex @code{alias} attribute
1637 The @code{alias} attribute causes the declaration to be emitted as an
1638 alias for another symbol, which must be specified. For instance,
1641 void __f () @{ /* do something */; @}
1642 void f () __attribute__ ((weak, alias ("__f")));
1645 declares @samp{f} to be a weak alias for @samp{__f}. In C++, the
1646 mangled name for the target must be used.
1648 Not all target machines support this attribute.
1650 @item no_check_memory_usage
1651 @cindex @code{no_check_memory_usage} function attribute
1652 The @code{no_check_memory_usage} attribute causes GNU CC to omit checks
1653 of memory references when it generates code for that function. Normally
1654 if you specify @samp{-fcheck-memory-usage} (see @pxref{Code Gen
1655 Options}), GNU CC generates calls to support routines before most memory
1656 accesses to permit support code to record usage and detect uses of
1657 uninitialized or unallocated storage. Since GNU CC cannot handle
1658 @code{asm} statements properly they are not allowed in such functions.
1659 If you declare a function with this attribute, GNU CC will not generate
1660 memory checking code for that function, permitting the use of @code{asm}
1661 statements without having to compile that function with different
1662 options. This also allows you to write support routines of your own if
1663 you wish, without getting infinite recursion if they get compiled with
1664 @code{-fcheck-memory-usage}.
1666 @item regparm (@var{number})
1667 @cindex functions that are passed arguments in registers on the 386
1668 On the Intel 386, the @code{regparm} attribute causes the compiler to
1669 pass up to @var{number} integer arguments in registers @var{EAX},
1670 @var{EDX}, and @var{ECX} instead of on the stack. Functions that take a
1671 variable number of arguments will continue to be passed all of their
1672 arguments on the stack.
1675 @cindex functions that pop the argument stack on the 386
1676 On the Intel 386, the @code{stdcall} attribute causes the compiler to
1677 assume that the called function will pop off the stack space used to
1678 pass arguments, unless it takes a variable number of arguments.
1680 The PowerPC compiler for Windows NT currently ignores the @code{stdcall}
1684 @cindex functions that do pop the argument stack on the 386
1685 On the Intel 386, the @code{cdecl} attribute causes the compiler to
1686 assume that the calling function will pop off the stack space used to
1687 pass arguments. This is
1688 useful to override the effects of the @samp{-mrtd} switch.
1690 The PowerPC compiler for Windows NT currently ignores the @code{cdecl}
1694 @cindex functions called via pointer on the RS/6000 and PowerPC
1695 On the RS/6000 and PowerPC, the @code{longcall} attribute causes the
1696 compiler to always call the function via a pointer, so that functions
1697 which reside further than 64 megabytes (67,108,864 bytes) from the
1698 current location can be called.
1700 @item long_call/short_call
1701 @cindex indirect calls on ARM
1702 This attribute allows to specify how to call a particular function on
1703 ARM. Both attributes override the @code{-mlong-calls} (@pxref{ARM Options})
1704 command line switch and @code{#pragma long_calls} settings. The
1705 @code{long_call} attribute causes the compiler to always call the
1706 function by first loading its address into a register and then using the
1707 contents of that register. The @code{short_call} attribute always places
1708 the offset to the function from the call site into the @samp{BL}
1709 instruction directly.
1712 @cindex functions which are imported from a dll on PowerPC Windows NT
1713 On the PowerPC running Windows NT, the @code{dllimport} attribute causes
1714 the compiler to call the function via a global pointer to the function
1715 pointer that is set up by the Windows NT dll library. The pointer name
1716 is formed by combining @code{__imp_} and the function name.
1719 @cindex functions which are exported from a dll on PowerPC Windows NT
1720 On the PowerPC running Windows NT, the @code{dllexport} attribute causes
1721 the compiler to provide a global pointer to the function pointer, so
1722 that it can be called with the @code{dllimport} attribute. The pointer
1723 name is formed by combining @code{__imp_} and the function name.
1725 @item exception (@var{except-func} [, @var{except-arg}])
1726 @cindex functions which specify exception handling on PowerPC Windows NT
1727 On the PowerPC running Windows NT, the @code{exception} attribute causes
1728 the compiler to modify the structured exception table entry it emits for
1729 the declared function. The string or identifier @var{except-func} is
1730 placed in the third entry of the structured exception table. It
1731 represents a function, which is called by the exception handling
1732 mechanism if an exception occurs. If it was specified, the string or
1733 identifier @var{except-arg} is placed in the fourth entry of the
1734 structured exception table.
1736 @item function_vector
1737 @cindex calling functions through the function vector on the H8/300 processors
1738 Use this option on the H8/300 and H8/300H to indicate that the specified
1739 function should be called through the function vector. Calling a
1740 function through the function vector will reduce code size, however;
1741 the function vector has a limited size (maximum 128 entries on the H8/300
1742 and 64 entries on the H8/300H) and shares space with the interrupt vector.
1744 You must use GAS and GLD from GNU binutils version 2.7 or later for
1745 this option to work correctly.
1747 @item interrupt_handler
1748 @cindex interrupt handler functions on the H8/300 processors
1749 Use this option on the H8/300 and H8/300H to indicate that the specified
1750 function is an interrupt handler. The compiler will generate function
1751 entry and exit sequences suitable for use in an interrupt handler when this
1752 attribute is present.
1755 @cindex eight bit data on the H8/300 and H8/300H
1756 Use this option on the H8/300 and H8/300H to indicate that the specified
1757 variable should be placed into the eight bit data section.
1758 The compiler will generate more efficient code for certain operations
1759 on data in the eight bit data area. Note the eight bit data area is limited to
1762 You must use GAS and GLD from GNU binutils version 2.7 or later for
1763 this option to work correctly.
1766 @cindex tiny data section on the H8/300H
1767 Use this option on the H8/300H to indicate that the specified
1768 variable should be placed into the tiny data section.
1769 The compiler will generate more efficient code for loads and stores
1770 on data in the tiny data section. Note the tiny data area is limited to
1771 slightly under 32kbytes of data.
1774 @cindex interrupt handlers on the M32R/D
1775 Use this option on the M32R/D to indicate that the specified
1776 function is an interrupt handler. The compiler will generate function
1777 entry and exit sequences suitable for use in an interrupt handler when this
1778 attribute is present.
1780 Interrupt handler functions on the AVR processors
1781 Use this option on the AVR to indicate that the specified
1782 function is an interrupt handler. The compiler will generate function
1783 entry and exit sequences suitable for use in an interrupt handler when this
1784 attribute is present. Interrupts will be enabled inside function.
1787 @cindex signal handler functions on the AVR processors
1788 Use this option on the AVR to indicate that the specified
1789 function is an signal handler. The compiler will generate function
1790 entry and exit sequences suitable for use in an signal handler when this
1791 attribute is present. Interrupts will be disabled inside function.
1794 @cindex function without a prologue/epilogue code on the AVR processors
1795 Use this option on the AVR to indicate that the specified
1796 function don't have a prologue/epilogue. The compiler don't generate
1797 function entry and exit sequences.
1799 @item model (@var{model-name})
1800 @cindex function addressability on the M32R/D
1801 Use this attribute on the M32R/D to set the addressability of an object,
1802 and the code generated for a function.
1803 The identifier @var{model-name} is one of @code{small}, @code{medium},
1804 or @code{large}, representing each of the code models.
1806 Small model objects live in the lower 16MB of memory (so that their
1807 addresses can be loaded with the @code{ld24} instruction), and are
1808 callable with the @code{bl} instruction.
1810 Medium model objects may live anywhere in the 32 bit address space (the
1811 compiler will generate @code{seth/add3} instructions to load their addresses),
1812 and are callable with the @code{bl} instruction.
1814 Large model objects may live anywhere in the 32 bit address space (the
1815 compiler will generate @code{seth/add3} instructions to load their addresses),
1816 and may not be reachable with the @code{bl} instruction (the compiler will
1817 generate the much slower @code{seth/add3/jl} instruction sequence).
1821 You can specify multiple attributes in a declaration by separating them
1822 by commas within the double parentheses or by immediately following an
1823 attribute declaration with another attribute declaration.
1825 @cindex @code{#pragma}, reason for not using
1826 @cindex pragma, reason for not using
1827 Some people object to the @code{__attribute__} feature, suggesting that ANSI C's
1828 @code{#pragma} should be used instead. There are two reasons for not
1833 It is impossible to generate @code{#pragma} commands from a macro.
1836 There is no telling what the same @code{#pragma} might mean in another
1840 These two reasons apply to almost any application that might be proposed
1841 for @code{#pragma}. It is basically a mistake to use @code{#pragma} for
1844 @node Function Prototypes
1845 @section Prototypes and Old-Style Function Definitions
1846 @cindex function prototype declarations
1847 @cindex old-style function definitions
1848 @cindex promotion of formal parameters
1850 GNU C extends ANSI C to allow a function prototype to override a later
1851 old-style non-prototype definition. Consider the following example:
1854 /* @r{Use prototypes unless the compiler is old-fashioned.} */
1861 /* @r{Prototype function declaration.} */
1862 int isroot P((uid_t));
1864 /* @r{Old-style function definition.} */
1866 isroot (x) /* ??? lossage here ??? */
1873 Suppose the type @code{uid_t} happens to be @code{short}. ANSI C does
1874 not allow this example, because subword arguments in old-style
1875 non-prototype definitions are promoted. Therefore in this example the
1876 function definition's argument is really an @code{int}, which does not
1877 match the prototype argument type of @code{short}.
1879 This restriction of ANSI C makes it hard to write code that is portable
1880 to traditional C compilers, because the programmer does not know
1881 whether the @code{uid_t} type is @code{short}, @code{int}, or
1882 @code{long}. Therefore, in cases like these GNU C allows a prototype
1883 to override a later old-style definition. More precisely, in GNU C, a
1884 function prototype argument type overrides the argument type specified
1885 by a later old-style definition if the former type is the same as the
1886 latter type before promotion. Thus in GNU C the above example is
1887 equivalent to the following:
1899 GNU C++ does not support old-style function definitions, so this
1900 extension is irrelevant.
1903 @section C++ Style Comments
1905 @cindex C++ comments
1906 @cindex comments, C++ style
1908 In GNU C, you may use C++ style comments, which start with @samp{//} and
1909 continue until the end of the line. Many other C implementations allow
1910 such comments, and they are likely to be in a future C standard.
1911 However, C++ style comments are not recognized if you specify
1912 @w{@samp{-ansi}} or @w{@samp{-traditional}}, since they are incompatible
1913 with traditional constructs like @code{dividend//*comment*/divisor}.
1916 @section Dollar Signs in Identifier Names
1918 @cindex dollar signs in identifier names
1919 @cindex identifier names, dollar signs in
1921 In GNU C, you may normally use dollar signs in identifier names.
1922 This is because many traditional C implementations allow such identifiers.
1923 However, dollar signs in identifiers are not supported on a few target
1924 machines, typically because the target assembler does not allow them.
1926 @node Character Escapes
1927 @section The Character @key{ESC} in Constants
1929 You can use the sequence @samp{\e} in a string or character constant to
1930 stand for the ASCII character @key{ESC}.
1933 @section Inquiring on Alignment of Types or Variables
1935 @cindex type alignment
1936 @cindex variable alignment
1938 The keyword @code{__alignof__} allows you to inquire about how an object
1939 is aligned, or the minimum alignment usually required by a type. Its
1940 syntax is just like @code{sizeof}.
1942 For example, if the target machine requires a @code{double} value to be
1943 aligned on an 8-byte boundary, then @code{__alignof__ (double)} is 8.
1944 This is true on many RISC machines. On more traditional machine
1945 designs, @code{__alignof__ (double)} is 4 or even 2.
1947 Some machines never actually require alignment; they allow reference to any
1948 data type even at an odd addresses. For these machines, @code{__alignof__}
1949 reports the @emph{recommended} alignment of a type.
1951 When the operand of @code{__alignof__} is an lvalue rather than a type, the
1952 value is the largest alignment that the lvalue is known to have. It may
1953 have this alignment as a result of its data type, or because it is part of
1954 a structure and inherits alignment from that structure. For example, after
1958 struct foo @{ int x; char y; @} foo1;
1962 the value of @code{__alignof__ (foo1.y)} is probably 2 or 4, the same as
1963 @code{__alignof__ (int)}, even though the data type of @code{foo1.y}
1964 does not itself demand any alignment.@refill
1966 It is an error to ask for the alignment of an incomplete type.
1968 A related feature which lets you specify the alignment of an object is
1969 @code{__attribute__ ((aligned (@var{alignment})))}; see the following
1972 @node Variable Attributes
1973 @section Specifying Attributes of Variables
1974 @cindex attribute of variables
1975 @cindex variable attributes
1977 The keyword @code{__attribute__} allows you to specify special
1978 attributes of variables or structure fields. This keyword is followed
1979 by an attribute specification inside double parentheses. Eight
1980 attributes are currently defined for variables: @code{aligned},
1981 @code{mode}, @code{nocommon}, @code{packed}, @code{section},
1982 @code{transparent_union}, @code{unused}, and @code{weak}. Other
1983 attributes are available for functions (@pxref{Function Attributes}) and
1984 for types (@pxref{Type Attributes}).
1986 You may also specify attributes with @samp{__} preceding and following
1987 each keyword. This allows you to use them in header files without
1988 being concerned about a possible macro of the same name. For example,
1989 you may use @code{__aligned__} instead of @code{aligned}.
1992 @cindex @code{aligned} attribute
1993 @item aligned (@var{alignment})
1994 This attribute specifies a minimum alignment for the variable or
1995 structure field, measured in bytes. For example, the declaration:
1998 int x __attribute__ ((aligned (16))) = 0;
2002 causes the compiler to allocate the global variable @code{x} on a
2003 16-byte boundary. On a 68040, this could be used in conjunction with
2004 an @code{asm} expression to access the @code{move16} instruction which
2005 requires 16-byte aligned operands.
2007 You can also specify the alignment of structure fields. For example, to
2008 create a double-word aligned @code{int} pair, you could write:
2011 struct foo @{ int x[2] __attribute__ ((aligned (8))); @};
2015 This is an alternative to creating a union with a @code{double} member
2016 that forces the union to be double-word aligned.
2018 It is not possible to specify the alignment of functions; the alignment
2019 of functions is determined by the machine's requirements and cannot be
2020 changed. You cannot specify alignment for a typedef name because such a
2021 name is just an alias, not a distinct type.
2023 As in the preceding examples, you can explicitly specify the alignment
2024 (in bytes) that you wish the compiler to use for a given variable or
2025 structure field. Alternatively, you can leave out the alignment factor
2026 and just ask the compiler to align a variable or field to the maximum
2027 useful alignment for the target machine you are compiling for. For
2028 example, you could write:
2031 short array[3] __attribute__ ((aligned));
2034 Whenever you leave out the alignment factor in an @code{aligned} attribute
2035 specification, the compiler automatically sets the alignment for the declared
2036 variable or field to the largest alignment which is ever used for any data
2037 type on the target machine you are compiling for. Doing this can often make
2038 copy operations more efficient, because the compiler can use whatever
2039 instructions copy the biggest chunks of memory when performing copies to
2040 or from the variables or fields that you have aligned this way.
2042 The @code{aligned} attribute can only increase the alignment; but you
2043 can decrease it by specifying @code{packed} as well. See below.
2045 Note that the effectiveness of @code{aligned} attributes may be limited
2046 by inherent limitations in your linker. On many systems, the linker is
2047 only able to arrange for variables to be aligned up to a certain maximum
2048 alignment. (For some linkers, the maximum supported alignment may
2049 be very very small.) If your linker is only able to align variables
2050 up to a maximum of 8 byte alignment, then specifying @code{aligned(16)}
2051 in an @code{__attribute__} will still only provide you with 8 byte
2052 alignment. See your linker documentation for further information.
2054 @item mode (@var{mode})
2055 @cindex @code{mode} attribute
2056 This attribute specifies the data type for the declaration---whichever
2057 type corresponds to the mode @var{mode}. This in effect lets you
2058 request an integer or floating point type according to its width.
2060 You may also specify a mode of @samp{byte} or @samp{__byte__} to
2061 indicate the mode corresponding to a one-byte integer, @samp{word} or
2062 @samp{__word__} for the mode of a one-word integer, and @samp{pointer}
2063 or @samp{__pointer__} for the mode used to represent pointers.
2066 @cindex @code{nocommon} attribute
2067 This attribute specifies requests GNU CC not to place a variable
2068 ``common'' but instead to allocate space for it directly. If you
2069 specify the @samp{-fno-common} flag, GNU CC will do this for all
2072 Specifying the @code{nocommon} attribute for a variable provides an
2073 initialization of zeros. A variable may only be initialized in one
2077 @cindex @code{packed} attribute
2078 The @code{packed} attribute specifies that a variable or structure field
2079 should have the smallest possible alignment---one byte for a variable,
2080 and one bit for a field, unless you specify a larger value with the
2081 @code{aligned} attribute.
2083 Here is a structure in which the field @code{x} is packed, so that it
2084 immediately follows @code{a}:
2090 int x[2] __attribute__ ((packed));
2094 @item section ("section-name")
2095 @cindex @code{section} variable attribute
2096 Normally, the compiler places the objects it generates in sections like
2097 @code{data} and @code{bss}. Sometimes, however, you need additional sections,
2098 or you need certain particular variables to appear in special sections,
2099 for example to map to special hardware. The @code{section}
2100 attribute specifies that a variable (or function) lives in a particular
2101 section. For example, this small program uses several specific section names:
2104 struct duart a __attribute__ ((section ("DUART_A"))) = @{ 0 @};
2105 struct duart b __attribute__ ((section ("DUART_B"))) = @{ 0 @};
2106 char stack[10000] __attribute__ ((section ("STACK"))) = @{ 0 @};
2107 int init_data __attribute__ ((section ("INITDATA"))) = 0;
2111 /* Initialize stack pointer */
2112 init_sp (stack + sizeof (stack));
2114 /* Initialize initialized data */
2115 memcpy (&init_data, &data, &edata - &data);
2117 /* Turn on the serial ports */
2124 Use the @code{section} attribute with an @emph{initialized} definition
2125 of a @emph{global} variable, as shown in the example. GNU CC issues
2126 a warning and otherwise ignores the @code{section} attribute in
2127 uninitialized variable declarations.
2129 You may only use the @code{section} attribute with a fully initialized
2130 global definition because of the way linkers work. The linker requires
2131 each object be defined once, with the exception that uninitialized
2132 variables tentatively go in the @code{common} (or @code{bss}) section
2133 and can be multiply "defined". You can force a variable to be
2134 initialized with the @samp{-fno-common} flag or the @code{nocommon}
2137 Some file formats do not support arbitrary sections so the @code{section}
2138 attribute is not available on all platforms.
2139 If you need to map the entire contents of a module to a particular
2140 section, consider using the facilities of the linker instead.
2143 @cindex @code{shared} variable attribute
2144 On Windows NT, in addition to putting variable definitions in a named
2145 section, the section can also be shared among all running copies of an
2146 executable or DLL. For example, this small program defines shared data
2147 by putting it in a named section "shared" and marking the section
2151 int foo __attribute__((section ("shared"), shared)) = 0;
2156 /* Read and write foo. All running copies see the same value. */
2162 You may only use the @code{shared} attribute along with @code{section}
2163 attribute with a fully initialized global definition because of the way
2164 linkers work. See @code{section} attribute for more information.
2166 The @code{shared} attribute is only available on Windows NT.
2168 @item transparent_union
2169 This attribute, attached to a function parameter which is a union, means
2170 that the corresponding argument may have the type of any union member,
2171 but the argument is passed as if its type were that of the first union
2172 member. For more details see @xref{Type Attributes}. You can also use
2173 this attribute on a @code{typedef} for a union data type; then it
2174 applies to all function parameters with that type.
2177 This attribute, attached to a variable, means that the variable is meant
2178 to be possibly unused. GNU CC will not produce a warning for this
2182 The @code{weak} attribute is described in @xref{Function Attributes}.
2184 @item model (@var{model-name})
2185 @cindex variable addressability on the M32R/D
2186 Use this attribute on the M32R/D to set the addressability of an object.
2187 The identifier @var{model-name} is one of @code{small}, @code{medium},
2188 or @code{large}, representing each of the code models.
2190 Small model objects live in the lower 16MB of memory (so that their
2191 addresses can be loaded with the @code{ld24} instruction).
2193 Medium and large model objects may live anywhere in the 32 bit address space
2194 (the compiler will generate @code{seth/add3} instructions to load their
2199 To specify multiple attributes, separate them by commas within the
2200 double parentheses: for example, @samp{__attribute__ ((aligned (16),
2203 @node Type Attributes
2204 @section Specifying Attributes of Types
2205 @cindex attribute of types
2206 @cindex type attributes
2208 The keyword @code{__attribute__} allows you to specify special
2209 attributes of @code{struct} and @code{union} types when you define such
2210 types. This keyword is followed by an attribute specification inside
2211 double parentheses. Three attributes are currently defined for types:
2212 @code{aligned}, @code{packed}, and @code{transparent_union}. Other
2213 attributes are defined for functions (@pxref{Function Attributes}) and
2214 for variables (@pxref{Variable Attributes}).
2216 You may also specify any one of these attributes with @samp{__}
2217 preceding and following its keyword. This allows you to use these
2218 attributes in header files without being concerned about a possible
2219 macro of the same name. For example, you may use @code{__aligned__}
2220 instead of @code{aligned}.
2222 You may specify the @code{aligned} and @code{transparent_union}
2223 attributes either in a @code{typedef} declaration or just past the
2224 closing curly brace of a complete enum, struct or union type
2225 @emph{definition} and the @code{packed} attribute only past the closing
2226 brace of a definition.
2228 You may also specify attributes between the enum, struct or union
2229 tag and the name of the type rather than after the closing brace.
2232 @cindex @code{aligned} attribute
2233 @item aligned (@var{alignment})
2234 This attribute specifies a minimum alignment (in bytes) for variables
2235 of the specified type. For example, the declarations:
2238 struct S @{ short f[3]; @} __attribute__ ((aligned (8)));
2239 typedef int more_aligned_int __attribute__ ((aligned (8)));
2243 force the compiler to insure (as far as it can) that each variable whose
2244 type is @code{struct S} or @code{more_aligned_int} will be allocated and
2245 aligned @emph{at least} on a 8-byte boundary. On a Sparc, having all
2246 variables of type @code{struct S} aligned to 8-byte boundaries allows
2247 the compiler to use the @code{ldd} and @code{std} (doubleword load and
2248 store) instructions when copying one variable of type @code{struct S} to
2249 another, thus improving run-time efficiency.
2251 Note that the alignment of any given @code{struct} or @code{union} type
2252 is required by the ANSI C standard to be at least a perfect multiple of
2253 the lowest common multiple of the alignments of all of the members of
2254 the @code{struct} or @code{union} in question. This means that you @emph{can}
2255 effectively adjust the alignment of a @code{struct} or @code{union}
2256 type by attaching an @code{aligned} attribute to any one of the members
2257 of such a type, but the notation illustrated in the example above is a
2258 more obvious, intuitive, and readable way to request the compiler to
2259 adjust the alignment of an entire @code{struct} or @code{union} type.
2261 As in the preceding example, you can explicitly specify the alignment
2262 (in bytes) that you wish the compiler to use for a given @code{struct}
2263 or @code{union} type. Alternatively, you can leave out the alignment factor
2264 and just ask the compiler to align a type to the maximum
2265 useful alignment for the target machine you are compiling for. For
2266 example, you could write:
2269 struct S @{ short f[3]; @} __attribute__ ((aligned));
2272 Whenever you leave out the alignment factor in an @code{aligned}
2273 attribute specification, the compiler automatically sets the alignment
2274 for the type to the largest alignment which is ever used for any data
2275 type on the target machine you are compiling for. Doing this can often
2276 make copy operations more efficient, because the compiler can use
2277 whatever instructions copy the biggest chunks of memory when performing
2278 copies to or from the variables which have types that you have aligned
2281 In the example above, if the size of each @code{short} is 2 bytes, then
2282 the size of the entire @code{struct S} type is 6 bytes. The smallest
2283 power of two which is greater than or equal to that is 8, so the
2284 compiler sets the alignment for the entire @code{struct S} type to 8
2287 Note that although you can ask the compiler to select a time-efficient
2288 alignment for a given type and then declare only individual stand-alone
2289 objects of that type, the compiler's ability to select a time-efficient
2290 alignment is primarily useful only when you plan to create arrays of
2291 variables having the relevant (efficiently aligned) type. If you
2292 declare or use arrays of variables of an efficiently-aligned type, then
2293 it is likely that your program will also be doing pointer arithmetic (or
2294 subscripting, which amounts to the same thing) on pointers to the
2295 relevant type, and the code that the compiler generates for these
2296 pointer arithmetic operations will often be more efficient for
2297 efficiently-aligned types than for other types.
2299 The @code{aligned} attribute can only increase the alignment; but you
2300 can decrease it by specifying @code{packed} as well. See below.
2302 Note that the effectiveness of @code{aligned} attributes may be limited
2303 by inherent limitations in your linker. On many systems, the linker is
2304 only able to arrange for variables to be aligned up to a certain maximum
2305 alignment. (For some linkers, the maximum supported alignment may
2306 be very very small.) If your linker is only able to align variables
2307 up to a maximum of 8 byte alignment, then specifying @code{aligned(16)}
2308 in an @code{__attribute__} will still only provide you with 8 byte
2309 alignment. See your linker documentation for further information.
2312 This attribute, attached to an @code{enum}, @code{struct}, or
2313 @code{union} type definition, specified that the minimum required memory
2314 be used to represent the type.
2316 Specifying this attribute for @code{struct} and @code{union} types is
2317 equivalent to specifying the @code{packed} attribute on each of the
2318 structure or union members. Specifying the @samp{-fshort-enums}
2319 flag on the line is equivalent to specifying the @code{packed}
2320 attribute on all @code{enum} definitions.
2322 You may only specify this attribute after a closing curly brace on an
2323 @code{enum} definition, not in a @code{typedef} declaration, unless that
2324 declaration also contains the definition of the @code{enum}.
2326 @item transparent_union
2327 This attribute, attached to a @code{union} type definition, indicates
2328 that any function parameter having that union type causes calls to that
2329 function to be treated in a special way.
2331 First, the argument corresponding to a transparent union type can be of
2332 any type in the union; no cast is required. Also, if the union contains
2333 a pointer type, the corresponding argument can be a null pointer
2334 constant or a void pointer expression; and if the union contains a void
2335 pointer type, the corresponding argument can be any pointer expression.
2336 If the union member type is a pointer, qualifiers like @code{const} on
2337 the referenced type must be respected, just as with normal pointer
2340 Second, the argument is passed to the function using the calling
2341 conventions of first member of the transparent union, not the calling
2342 conventions of the union itself. All members of the union must have the
2343 same machine representation; this is necessary for this argument passing
2346 Transparent unions are designed for library functions that have multiple
2347 interfaces for compatibility reasons. For example, suppose the
2348 @code{wait} function must accept either a value of type @code{int *} to
2349 comply with Posix, or a value of type @code{union wait *} to comply with
2350 the 4.1BSD interface. If @code{wait}'s parameter were @code{void *},
2351 @code{wait} would accept both kinds of arguments, but it would also
2352 accept any other pointer type and this would make argument type checking
2353 less useful. Instead, @code{<sys/wait.h>} might define the interface
2361 @} wait_status_ptr_t __attribute__ ((__transparent_union__));
2363 pid_t wait (wait_status_ptr_t);
2366 This interface allows either @code{int *} or @code{union wait *}
2367 arguments to be passed, using the @code{int *} calling convention.
2368 The program can call @code{wait} with arguments of either type:
2371 int w1 () @{ int w; return wait (&w); @}
2372 int w2 () @{ union wait w; return wait (&w); @}
2375 With this interface, @code{wait}'s implementation might look like this:
2378 pid_t wait (wait_status_ptr_t p)
2380 return waitpid (-1, p.__ip, 0);
2385 When attached to a type (including a @code{union} or a @code{struct}),
2386 this attribute means that variables of that type are meant to appear
2387 possibly unused. GNU CC will not produce a warning for any variables of
2388 that type, even if the variable appears to do nothing. This is often
2389 the case with lock or thread classes, which are usually defined and then
2390 not referenced, but contain constructors and destructors that have
2391 nontrivial bookkeeping functions.
2395 To specify multiple attributes, separate them by commas within the
2396 double parentheses: for example, @samp{__attribute__ ((aligned (16),
2400 @section An Inline Function is As Fast As a Macro
2401 @cindex inline functions
2402 @cindex integrating function code
2404 @cindex macros, inline alternative
2406 By declaring a function @code{inline}, you can direct GNU CC to
2407 integrate that function's code into the code for its callers. This
2408 makes execution faster by eliminating the function-call overhead; in
2409 addition, if any of the actual argument values are constant, their known
2410 values may permit simplifications at compile time so that not all of the
2411 inline function's code needs to be included. The effect on code size is
2412 less predictable; object code may be larger or smaller with function
2413 inlining, depending on the particular case. Inlining of functions is an
2414 optimization and it really ``works'' only in optimizing compilation. If
2415 you don't use @samp{-O}, no function is really inline.
2417 To declare a function inline, use the @code{inline} keyword in its
2418 declaration, like this:
2428 (If you are writing a header file to be included in ANSI C programs, write
2429 @code{__inline__} instead of @code{inline}. @xref{Alternate Keywords}.)
2430 You can also make all ``simple enough'' functions inline with the option
2431 @samp{-finline-functions}.
2433 Note that certain usages in a function definition can make it unsuitable
2434 for inline substitution. Among these usages are: use of varargs, use of
2435 alloca, use of variable sized data types (@pxref{Variable Length}),
2436 use of computed goto (@pxref{Labels as Values}), use of nonlocal goto,
2437 and nested functions (@pxref{Nested Functions}). Using @samp{-Winline}
2438 will warn when a function marked @code{inline} could not be substituted,
2439 and will give the reason for the failure.
2441 Note that in C and Objective C, unlike C++, the @code{inline} keyword
2442 does not affect the linkage of the function.
2444 @cindex automatic @code{inline} for C++ member fns
2445 @cindex @code{inline} automatic for C++ member fns
2446 @cindex member fns, automatically @code{inline}
2447 @cindex C++ member fns, automatically @code{inline}
2448 GNU CC automatically inlines member functions defined within the class
2449 body of C++ programs even if they are not explicitly declared
2450 @code{inline}. (You can override this with @samp{-fno-default-inline};
2451 @pxref{C++ Dialect Options,,Options Controlling C++ Dialect}.)
2453 @cindex inline functions, omission of
2454 When a function is both inline and @code{static}, if all calls to the
2455 function are integrated into the caller, and the function's address is
2456 never used, then the function's own assembler code is never referenced.
2457 In this case, GNU CC does not actually output assembler code for the
2458 function, unless you specify the option @samp{-fkeep-inline-functions}.
2459 Some calls cannot be integrated for various reasons (in particular,
2460 calls that precede the function's definition cannot be integrated, and
2461 neither can recursive calls within the definition). If there is a
2462 nonintegrated call, then the function is compiled to assembler code as
2463 usual. The function must also be compiled as usual if the program
2464 refers to its address, because that can't be inlined.
2466 @cindex non-static inline function
2467 When an inline function is not @code{static}, then the compiler must assume
2468 that there may be calls from other source files; since a global symbol can
2469 be defined only once in any program, the function must not be defined in
2470 the other source files, so the calls therein cannot be integrated.
2471 Therefore, a non-@code{static} inline function is always compiled on its
2472 own in the usual fashion.
2474 If you specify both @code{inline} and @code{extern} in the function
2475 definition, then the definition is used only for inlining. In no case
2476 is the function compiled on its own, not even if you refer to its
2477 address explicitly. Such an address becomes an external reference, as
2478 if you had only declared the function, and had not defined it.
2480 This combination of @code{inline} and @code{extern} has almost the
2481 effect of a macro. The way to use it is to put a function definition in
2482 a header file with these keywords, and put another copy of the
2483 definition (lacking @code{inline} and @code{extern}) in a library file.
2484 The definition in the header file will cause most calls to the function
2485 to be inlined. If any uses of the function remain, they will refer to
2486 the single copy in the library.
2488 GNU C does not inline any functions when not optimizing. It is not
2489 clear whether it is better to inline or not, in this case, but we found
2490 that a correct implementation when not optimizing was difficult. So we
2491 did the easy thing, and turned it off.
2494 @section Assembler Instructions with C Expression Operands
2495 @cindex extended @code{asm}
2496 @cindex @code{asm} expressions
2497 @cindex assembler instructions
2500 In an assembler instruction using @code{asm}, you can specify the
2501 operands of the instruction using C expressions. This means you need not
2502 guess which registers or memory locations will contain the data you want
2505 You must specify an assembler instruction template much like what
2506 appears in a machine description, plus an operand constraint string for
2509 For example, here is how to use the 68881's @code{fsinx} instruction:
2512 asm ("fsinx %1,%0" : "=f" (result) : "f" (angle));
2516 Here @code{angle} is the C expression for the input operand while
2517 @code{result} is that of the output operand. Each has @samp{"f"} as its
2518 operand constraint, saying that a floating point register is required.
2519 The @samp{=} in @samp{=f} indicates that the operand is an output; all
2520 output operands' constraints must use @samp{=}. The constraints use the
2521 same language used in the machine description (@pxref{Constraints}).
2523 Each operand is described by an operand-constraint string followed by
2524 the C expression in parentheses. A colon separates the assembler
2525 template from the first output operand and another separates the last
2526 output operand from the first input, if any. Commas separate the
2527 operands within each group. The total number of operands is limited to
2528 ten or to the maximum number of operands in any instruction pattern in
2529 the machine description, whichever is greater.
2531 If there are no output operands but there are input operands, you must
2532 place two consecutive colons surrounding the place where the output
2535 Output operand expressions must be lvalues; the compiler can check this.
2536 The input operands need not be lvalues. The compiler cannot check
2537 whether the operands have data types that are reasonable for the
2538 instruction being executed. It does not parse the assembler instruction
2539 template and does not know what it means or even whether it is valid
2540 assembler input. The extended @code{asm} feature is most often used for
2541 machine instructions the compiler itself does not know exist. If
2542 the output expression cannot be directly addressed (for example, it is a
2543 bit field), your constraint must allow a register. In that case, GNU CC
2544 will use the register as the output of the @code{asm}, and then store
2545 that register into the output.
2547 The ordinary output operands must be write-only; GNU CC will assume that
2548 the values in these operands before the instruction are dead and need
2549 not be generated. Extended asm supports input-output or read-write
2550 operands. Use the constraint character @samp{+} to indicate such an
2551 operand and list it with the output operands.
2553 When the constraints for the read-write operand (or the operand in which
2554 only some of the bits are to be changed) allows a register, you may, as
2555 an alternative, logically split its function into two separate operands,
2556 one input operand and one write-only output operand. The connection
2557 between them is expressed by constraints which say they need to be in
2558 the same location when the instruction executes. You can use the same C
2559 expression for both operands, or different expressions. For example,
2560 here we write the (fictitious) @samp{combine} instruction with
2561 @code{bar} as its read-only source operand and @code{foo} as its
2562 read-write destination:
2565 asm ("combine %2,%0" : "=r" (foo) : "0" (foo), "g" (bar));
2569 The constraint @samp{"0"} for operand 1 says that it must occupy the
2570 same location as operand 0. A digit in constraint is allowed only in an
2571 input operand and it must refer to an output operand.
2573 Only a digit in the constraint can guarantee that one operand will be in
2574 the same place as another. The mere fact that @code{foo} is the value
2575 of both operands is not enough to guarantee that they will be in the
2576 same place in the generated assembler code. The following would not
2580 asm ("combine %2,%0" : "=r" (foo) : "r" (foo), "g" (bar));
2583 Various optimizations or reloading could cause operands 0 and 1 to be in
2584 different registers; GNU CC knows no reason not to do so. For example, the
2585 compiler might find a copy of the value of @code{foo} in one register and
2586 use it for operand 1, but generate the output operand 0 in a different
2587 register (copying it afterward to @code{foo}'s own address). Of course,
2588 since the register for operand 1 is not even mentioned in the assembler
2589 code, the result will not work, but GNU CC can't tell that.
2591 Some instructions clobber specific hard registers. To describe this,
2592 write a third colon after the input operands, followed by the names of
2593 the clobbered hard registers (given as strings). Here is a realistic
2594 example for the VAX:
2597 asm volatile ("movc3 %0,%1,%2"
2599 : "g" (from), "g" (to), "g" (count)
2600 : "r0", "r1", "r2", "r3", "r4", "r5");
2603 You may not write a clobber description in a way that overlaps with an
2604 input or output operand. For example, you may not have an operand
2605 describing a register class with one member if you mention that register
2606 in the clobber list. There is no way for you to specify that an input
2607 operand is modified without also specifying it as an output
2608 operand. Note that if all the output operands you specify are for this
2609 purpose (and hence unused), you will then also need to specify
2610 @code{volatile} for the @code{asm} construct, as described below, to
2611 prevent GNU CC from deleting the @code{asm} statement as unused.
2613 If you refer to a particular hardware register from the assembler code,
2614 you will probably have to list the register after the third colon to
2615 tell the compiler the register's value is modified. In some assemblers,
2616 the register names begin with @samp{%}; to produce one @samp{%} in the
2617 assembler code, you must write @samp{%%} in the input.
2619 If your assembler instruction can alter the condition code register, add
2620 @samp{cc} to the list of clobbered registers. GNU CC on some machines
2621 represents the condition codes as a specific hardware register;
2622 @samp{cc} serves to name this register. On other machines, the
2623 condition code is handled differently, and specifying @samp{cc} has no
2624 effect. But it is valid no matter what the machine.
2626 If your assembler instruction modifies memory in an unpredictable
2627 fashion, add @samp{memory} to the list of clobbered registers. This
2628 will cause GNU CC to not keep memory values cached in registers across
2629 the assembler instruction. You will also want to add the
2630 @code{volatile} keyword if the memory affected is not listed in the
2631 inputs or outputs of the @code{asm}, as the @samp{memory} clobber does
2632 not count as a side-effect of the @code{asm}.
2634 You can put multiple assembler instructions together in a single
2635 @code{asm} template, separated either with newlines (written as
2636 @samp{\n}) or with semicolons if the assembler allows such semicolons.
2637 The GNU assembler allows semicolons and most Unix assemblers seem to do
2638 so. The input operands are guaranteed not to use any of the clobbered
2639 registers, and neither will the output operands' addresses, so you can
2640 read and write the clobbered registers as many times as you like. Here
2641 is an example of multiple instructions in a template; it assumes the
2642 subroutine @code{_foo} accepts arguments in registers 9 and 10:
2645 asm ("movl %0,r9;movl %1,r10;call _foo"
2647 : "g" (from), "g" (to)
2651 Unless an output operand has the @samp{&} constraint modifier, GNU CC
2652 may allocate it in the same register as an unrelated input operand, on
2653 the assumption the inputs are consumed before the outputs are produced.
2654 This assumption may be false if the assembler code actually consists of
2655 more than one instruction. In such a case, use @samp{&} for each output
2656 operand that may not overlap an input. @xref{Modifiers}.
2658 If you want to test the condition code produced by an assembler
2659 instruction, you must include a branch and a label in the @code{asm}
2660 construct, as follows:
2663 asm ("clr %0;frob %1;beq 0f;mov #1,%0;0:"
2669 This assumes your assembler supports local labels, as the GNU assembler
2670 and most Unix assemblers do.
2672 Speaking of labels, jumps from one @code{asm} to another are not
2673 supported. The compiler's optimizers do not know about these jumps, and
2674 therefore they cannot take account of them when deciding how to
2677 @cindex macros containing @code{asm}
2678 Usually the most convenient way to use these @code{asm} instructions is to
2679 encapsulate them in macros that look like functions. For example,
2683 (@{ double __value, __arg = (x); \
2684 asm ("fsinx %1,%0": "=f" (__value): "f" (__arg)); \
2689 Here the variable @code{__arg} is used to make sure that the instruction
2690 operates on a proper @code{double} value, and to accept only those
2691 arguments @code{x} which can convert automatically to a @code{double}.
2693 Another way to make sure the instruction operates on the correct data
2694 type is to use a cast in the @code{asm}. This is different from using a
2695 variable @code{__arg} in that it converts more different types. For
2696 example, if the desired type were @code{int}, casting the argument to
2697 @code{int} would accept a pointer with no complaint, while assigning the
2698 argument to an @code{int} variable named @code{__arg} would warn about
2699 using a pointer unless the caller explicitly casts it.
2701 If an @code{asm} has output operands, GNU CC assumes for optimization
2702 purposes the instruction has no side effects except to change the output
2703 operands. This does not mean instructions with a side effect cannot be
2704 used, but you must be careful, because the compiler may eliminate them
2705 if the output operands aren't used, or move them out of loops, or
2706 replace two with one if they constitute a common subexpression. Also,
2707 if your instruction does have a side effect on a variable that otherwise
2708 appears not to change, the old value of the variable may be reused later
2709 if it happens to be found in a register.
2711 You can prevent an @code{asm} instruction from being deleted, moved
2712 significantly, or combined, by writing the keyword @code{volatile} after
2713 the @code{asm}. For example:
2716 #define get_and_set_priority(new) \
2718 asm volatile ("get_and_set_priority %0, %1": "=g" (__old) : "g" (new)); \
2723 If you write an @code{asm} instruction with no outputs, GNU CC will know
2724 the instruction has side-effects and will not delete the instruction or
2725 move it outside of loops. If the side-effects of your instruction are
2726 not purely external, but will affect variables in your program in ways
2727 other than reading the inputs and clobbering the specified registers or
2728 memory, you should write the @code{volatile} keyword to prevent future
2729 versions of GNU CC from moving the instruction around within a core
2732 An @code{asm} instruction without any operands or clobbers (and ``old
2733 style'' @code{asm}) will not be deleted or moved significantly,
2734 regardless, unless it is unreachable, the same way as if you had
2735 written a @code{volatile} keyword.
2737 Note that even a volatile @code{asm} instruction can be moved in ways
2738 that appear insignificant to the compiler, such as across jump
2739 instructions. You can't expect a sequence of volatile @code{asm}
2740 instructions to remain perfectly consecutive. If you want consecutive
2741 output, use a single @code{asm}.
2743 It is a natural idea to look for a way to give access to the condition
2744 code left by the assembler instruction. However, when we attempted to
2745 implement this, we found no way to make it work reliably. The problem
2746 is that output operands might need reloading, which would result in
2747 additional following ``store'' instructions. On most machines, these
2748 instructions would alter the condition code before there was time to
2749 test it. This problem doesn't arise for ordinary ``test'' and
2750 ``compare'' instructions because they don't have any output operands.
2752 For reasons similar to those described above, it is not possible to give
2753 an assembler instruction access to the condition code left by previous
2756 If you are writing a header file that should be includable in ANSI C
2757 programs, write @code{__asm__} instead of @code{asm}. @xref{Alternate
2760 @subsection i386 floating point asm operands
2762 There are several rules on the usage of stack-like regs in
2763 asm_operands insns. These rules apply only to the operands that are
2768 Given a set of input regs that die in an asm_operands, it is
2769 necessary to know which are implicitly popped by the asm, and
2770 which must be explicitly popped by gcc.
2772 An input reg that is implicitly popped by the asm must be
2773 explicitly clobbered, unless it is constrained to match an
2777 For any input reg that is implicitly popped by an asm, it is
2778 necessary to know how to adjust the stack to compensate for the pop.
2779 If any non-popped input is closer to the top of the reg-stack than
2780 the implicitly popped reg, it would not be possible to know what the
2781 stack looked like --- it's not clear how the rest of the stack ``slides
2784 All implicitly popped input regs must be closer to the top of
2785 the reg-stack than any input that is not implicitly popped.
2787 It is possible that if an input dies in an insn, reload might
2788 use the input reg for an output reload. Consider this example:
2791 asm ("foo" : "=t" (a) : "f" (b));
2794 This asm says that input B is not popped by the asm, and that
2795 the asm pushes a result onto the reg-stack, ie, the stack is one
2796 deeper after the asm than it was before. But, it is possible that
2797 reload will think that it can use the same reg for both the input and
2798 the output, if input B dies in this insn.
2800 If any input operand uses the @code{f} constraint, all output reg
2801 constraints must use the @code{&} earlyclobber.
2803 The asm above would be written as
2806 asm ("foo" : "=&t" (a) : "f" (b));
2810 Some operands need to be in particular places on the stack. All
2811 output operands fall in this category --- there is no other way to
2812 know which regs the outputs appear in unless the user indicates
2813 this in the constraints.
2815 Output operands must specifically indicate which reg an output
2816 appears in after an asm. @code{=f} is not allowed: the operand
2817 constraints must select a class with a single reg.
2820 Output operands may not be ``inserted'' between existing stack regs.
2821 Since no 387 opcode uses a read/write operand, all output operands
2822 are dead before the asm_operands, and are pushed by the asm_operands.
2823 It makes no sense to push anywhere but the top of the reg-stack.
2825 Output operands must start at the top of the reg-stack: output
2826 operands may not ``skip'' a reg.
2829 Some asm statements may need extra stack space for internal
2830 calculations. This can be guaranteed by clobbering stack registers
2831 unrelated to the inputs and outputs.
2835 Here are a couple of reasonable asms to want to write. This asm
2836 takes one input, which is internally popped, and produces two outputs.
2839 asm ("fsincos" : "=t" (cos), "=u" (sin) : "0" (inp));
2842 This asm takes two inputs, which are popped by the @code{fyl2xp1} opcode,
2843 and replaces them with one output. The user must code the @code{st(1)}
2844 clobber for reg-stack.c to know that @code{fyl2xp1} pops both inputs.
2847 asm ("fyl2xp1" : "=t" (result) : "0" (x), "u" (y) : "st(1)");
2851 @c Show the details on constraints if they do not appear elsewhere in
2857 @section Controlling Names Used in Assembler Code
2858 @cindex assembler names for identifiers
2859 @cindex names used in assembler code
2860 @cindex identifiers, names in assembler code
2862 You can specify the name to be used in the assembler code for a C
2863 function or variable by writing the @code{asm} (or @code{__asm__})
2864 keyword after the declarator as follows:
2867 int foo asm ("myfoo") = 2;
2871 This specifies that the name to be used for the variable @code{foo} in
2872 the assembler code should be @samp{myfoo} rather than the usual
2875 On systems where an underscore is normally prepended to the name of a C
2876 function or variable, this feature allows you to define names for the
2877 linker that do not start with an underscore.
2879 You cannot use @code{asm} in this way in a function @emph{definition}; but
2880 you can get the same effect by writing a declaration for the function
2881 before its definition and putting @code{asm} there, like this:
2884 extern func () asm ("FUNC");
2891 It is up to you to make sure that the assembler names you choose do not
2892 conflict with any other assembler symbols. Also, you must not use a
2893 register name; that would produce completely invalid assembler code. GNU
2894 CC does not as yet have the ability to store static variables in registers.
2895 Perhaps that will be added.
2897 @node Explicit Reg Vars
2898 @section Variables in Specified Registers
2899 @cindex explicit register variables
2900 @cindex variables in specified registers
2901 @cindex specified registers
2902 @cindex registers, global allocation
2904 GNU C allows you to put a few global variables into specified hardware
2905 registers. You can also specify the register in which an ordinary
2906 register variable should be allocated.
2910 Global register variables reserve registers throughout the program.
2911 This may be useful in programs such as programming language
2912 interpreters which have a couple of global variables that are accessed
2916 Local register variables in specific registers do not reserve the
2917 registers. The compiler's data flow analysis is capable of determining
2918 where the specified registers contain live values, and where they are
2919 available for other uses. Stores into local register variables may be deleted
2920 when they appear to be dead according to dataflow analysis. References
2921 to local register variables may be deleted or moved or simplified.
2923 These local variables are sometimes convenient for use with the extended
2924 @code{asm} feature (@pxref{Extended Asm}), if you want to write one
2925 output of the assembler instruction directly into a particular register.
2926 (This will work provided the register you specify fits the constraints
2927 specified for that operand in the @code{asm}.)
2935 @node Global Reg Vars
2936 @subsection Defining Global Register Variables
2937 @cindex global register variables
2938 @cindex registers, global variables in
2940 You can define a global register variable in GNU C like this:
2943 register int *foo asm ("a5");
2947 Here @code{a5} is the name of the register which should be used. Choose a
2948 register which is normally saved and restored by function calls on your
2949 machine, so that library routines will not clobber it.
2951 Naturally the register name is cpu-dependent, so you would need to
2952 conditionalize your program according to cpu type. The register
2953 @code{a5} would be a good choice on a 68000 for a variable of pointer
2954 type. On machines with register windows, be sure to choose a ``global''
2955 register that is not affected magically by the function call mechanism.
2957 In addition, operating systems on one type of cpu may differ in how they
2958 name the registers; then you would need additional conditionals. For
2959 example, some 68000 operating systems call this register @code{%a5}.
2961 Eventually there may be a way of asking the compiler to choose a register
2962 automatically, but first we need to figure out how it should choose and
2963 how to enable you to guide the choice. No solution is evident.
2965 Defining a global register variable in a certain register reserves that
2966 register entirely for this use, at least within the current compilation.
2967 The register will not be allocated for any other purpose in the functions
2968 in the current compilation. The register will not be saved and restored by
2969 these functions. Stores into this register are never deleted even if they
2970 would appear to be dead, but references may be deleted or moved or
2973 It is not safe to access the global register variables from signal
2974 handlers, or from more than one thread of control, because the system
2975 library routines may temporarily use the register for other things (unless
2976 you recompile them specially for the task at hand).
2978 @cindex @code{qsort}, and global register variables
2979 It is not safe for one function that uses a global register variable to
2980 call another such function @code{foo} by way of a third function
2981 @code{lose} that was compiled without knowledge of this variable (i.e. in a
2982 different source file in which the variable wasn't declared). This is
2983 because @code{lose} might save the register and put some other value there.
2984 For example, you can't expect a global register variable to be available in
2985 the comparison-function that you pass to @code{qsort}, since @code{qsort}
2986 might have put something else in that register. (If you are prepared to
2987 recompile @code{qsort} with the same global register variable, you can
2988 solve this problem.)
2990 If you want to recompile @code{qsort} or other source files which do not
2991 actually use your global register variable, so that they will not use that
2992 register for any other purpose, then it suffices to specify the compiler
2993 option @samp{-ffixed-@var{reg}}. You need not actually add a global
2994 register declaration to their source code.
2996 A function which can alter the value of a global register variable cannot
2997 safely be called from a function compiled without this variable, because it
2998 could clobber the value the caller expects to find there on return.
2999 Therefore, the function which is the entry point into the part of the
3000 program that uses the global register variable must explicitly save and
3001 restore the value which belongs to its caller.
3003 @cindex register variable after @code{longjmp}
3004 @cindex global register after @code{longjmp}
3005 @cindex value after @code{longjmp}
3008 On most machines, @code{longjmp} will restore to each global register
3009 variable the value it had at the time of the @code{setjmp}. On some
3010 machines, however, @code{longjmp} will not change the value of global
3011 register variables. To be portable, the function that called @code{setjmp}
3012 should make other arrangements to save the values of the global register
3013 variables, and to restore them in a @code{longjmp}. This way, the same
3014 thing will happen regardless of what @code{longjmp} does.
3016 All global register variable declarations must precede all function
3017 definitions. If such a declaration could appear after function
3018 definitions, the declaration would be too late to prevent the register from
3019 being used for other purposes in the preceding functions.
3021 Global register variables may not have initial values, because an
3022 executable file has no means to supply initial contents for a register.
3024 On the Sparc, there are reports that g3 @dots{} g7 are suitable
3025 registers, but certain library functions, such as @code{getwd}, as well
3026 as the subroutines for division and remainder, modify g3 and g4. g1 and
3027 g2 are local temporaries.
3029 On the 68000, a2 @dots{} a5 should be suitable, as should d2 @dots{} d7.
3030 Of course, it will not do to use more than a few of those.
3032 @node Local Reg Vars
3033 @subsection Specifying Registers for Local Variables
3034 @cindex local variables, specifying registers
3035 @cindex specifying registers for local variables
3036 @cindex registers for local variables
3038 You can define a local register variable with a specified register
3042 register int *foo asm ("a5");
3046 Here @code{a5} is the name of the register which should be used. Note
3047 that this is the same syntax used for defining global register
3048 variables, but for a local variable it would appear within a function.
3050 Naturally the register name is cpu-dependent, but this is not a
3051 problem, since specific registers are most often useful with explicit
3052 assembler instructions (@pxref{Extended Asm}). Both of these things
3053 generally require that you conditionalize your program according to
3056 In addition, operating systems on one type of cpu may differ in how they
3057 name the registers; then you would need additional conditionals. For
3058 example, some 68000 operating systems call this register @code{%a5}.
3060 Defining such a register variable does not reserve the register; it
3061 remains available for other uses in places where flow control determines
3062 the variable's value is not live. However, these registers are made
3063 unavailable for use in the reload pass; excessive use of this feature
3064 leaves the compiler too few available registers to compile certain
3067 This option does not guarantee that GNU CC will generate code that has
3068 this variable in the register you specify at all times. You may not
3069 code an explicit reference to this register in an @code{asm} statement
3070 and assume it will always refer to this variable.
3072 Stores into local register variables may be deleted when they appear to be dead
3073 according to dataflow analysis. References to local register variables may
3074 be deleted or moved or simplified.
3076 @node Alternate Keywords
3077 @section Alternate Keywords
3078 @cindex alternate keywords
3079 @cindex keywords, alternate
3081 The option @samp{-traditional} disables certain keywords; @samp{-ansi}
3082 disables certain others. This causes trouble when you want to use GNU C
3083 extensions, or ANSI C features, in a general-purpose header file that
3084 should be usable by all programs, including ANSI C programs and traditional
3085 ones. The keywords @code{asm}, @code{typeof} and @code{inline} cannot be
3086 used since they won't work in a program compiled with @samp{-ansi}, while
3087 the keywords @code{const}, @code{volatile}, @code{signed}, @code{typeof}
3088 and @code{inline} won't work in a program compiled with
3089 @samp{-traditional}.@refill
3091 The way to solve these problems is to put @samp{__} at the beginning and
3092 end of each problematical keyword. For example, use @code{__asm__}
3093 instead of @code{asm}, @code{__const__} instead of @code{const}, and
3094 @code{__inline__} instead of @code{inline}.
3096 Other C compilers won't accept these alternative keywords; if you want to
3097 compile with another compiler, you can define the alternate keywords as
3098 macros to replace them with the customary keywords. It looks like this:
3106 @findex __extension__
3107 @samp{-pedantic} and other options cause warnings for many GNU C extensions.
3109 prevent such warnings within one expression by writing
3110 @code{__extension__} before the expression. @code{__extension__} has no
3111 effect aside from this.
3113 @node Incomplete Enums
3114 @section Incomplete @code{enum} Types
3116 You can define an @code{enum} tag without specifying its possible values.
3117 This results in an incomplete type, much like what you get if you write
3118 @code{struct foo} without describing the elements. A later declaration
3119 which does specify the possible values completes the type.
3121 You can't allocate variables or storage using the type while it is
3122 incomplete. However, you can work with pointers to that type.
3124 This extension may not be very useful, but it makes the handling of
3125 @code{enum} more consistent with the way @code{struct} and @code{union}
3128 This extension is not supported by GNU C++.
3130 @node Function Names
3131 @section Function Names as Strings
3133 GNU CC predefines two magic identifiers to hold the name of the current
3134 function. The identifier @code{__FUNCTION__} holds the name of the function
3135 as it appears in the source. The identifier @code{__PRETTY_FUNCTION__}
3136 holds the name of the function pretty printed in a language specific
3139 These names are always the same in a C function, but in a C++ function
3140 they may be different. For example, this program:
3144 extern int printf (char *, ...);
3151 printf ("__FUNCTION__ = %s\n", __FUNCTION__);
3152 printf ("__PRETTY_FUNCTION__ = %s\n", __PRETTY_FUNCTION__);
3170 __PRETTY_FUNCTION__ = int a::sub (int)
3173 The compiler automagically replaces the identifiers with a string
3174 literal containing the appropriate name. Thus, they are neither
3175 preprocessor macros, like @code{__FILE__} and @code{__LINE__}, nor
3176 variables. This means that they catenate with other string literals, and
3177 that they can be used to initialize char arrays. For example
3180 char here[] = "Function " __FUNCTION__ " in " __FILE__;
3183 On the other hand, @samp{#ifdef __FUNCTION__} does not have any special
3184 meaning inside a function, since the preprocessor does not do anything
3185 special with the identifier @code{__FUNCTION__}.
3187 GNU CC also supports the magic word @code{__func__}, defined by the
3191 The identifier @code{__func__} is implicitly declared by the translator
3192 as if, immediately following the opening brace of each function
3193 definition, the declaration
3196 static const char __func__[] = "function-name";
3199 appeared, where function-name is the name of the lexically-enclosing
3200 function. This name is the unadorned name of the function.
3203 By this definition, @code{__func__} is a variable, not a string literal.
3204 In particular, @code{__func__} does not catenate with other string
3207 In @code{C++}, @code{__FUNCTION__} and @code{__PRETTY_FUNCTION__} are
3208 variables, declared in the same way as @code{__func__}.
3210 @node Return Address
3211 @section Getting the Return or Frame Address of a Function
3213 These functions may be used to get information about the callers of a
3217 @findex __builtin_return_address
3218 @item __builtin_return_address (@var{level})
3219 This function returns the return address of the current function, or of
3220 one of its callers. The @var{level} argument is number of frames to
3221 scan up the call stack. A value of @code{0} yields the return address
3222 of the current function, a value of @code{1} yields the return address
3223 of the caller of the current function, and so forth.
3225 The @var{level} argument must be a constant integer.
3227 On some machines it may be impossible to determine the return address of
3228 any function other than the current one; in such cases, or when the top
3229 of the stack has been reached, this function will return @code{0}.
3231 This function should only be used with a non-zero argument for debugging
3234 @findex __builtin_frame_address
3235 @item __builtin_frame_address (@var{level})
3236 This function is similar to @code{__builtin_return_address}, but it
3237 returns the address of the function frame rather than the return address
3238 of the function. Calling @code{__builtin_frame_address} with a value of
3239 @code{0} yields the frame address of the current function, a value of
3240 @code{1} yields the frame address of the caller of the current function,
3243 The frame is the area on the stack which holds local variables and saved
3244 registers. The frame address is normally the address of the first word
3245 pushed on to the stack by the function. However, the exact definition
3246 depends upon the processor and the calling convention. If the processor
3247 has a dedicated frame pointer register, and the function has a frame,
3248 then @code{__builtin_frame_address} will return the value of the frame
3251 The caveats that apply to @code{__builtin_return_address} apply to this
3255 @node Other Builtins
3256 @section Other built-in functions provided by GNU CC
3257 @cindex builtin functions
3258 @findex __builtin_isgreater
3259 @findex __builtin_isgreaterequal
3260 @findex __builtin_isless
3261 @findex __builtin_islessequal
3262 @findex __builtin_islessgreater
3263 @findex __builtin_isunordered
3303 GNU CC provides a large number of built-in functions other than the ones
3304 mentioned above. Some of these are for internal use in the processing
3305 of exceptions or variable-length argument lists and will not be
3306 documented here because they may change from time to time; we do not
3307 recommend general use of these functions.
3309 The remaining functions are provided for optimization purposes.
3311 GNU CC includes builtin versions of many of the functions in the
3312 standard C library. The versions prefixed with @code{__builtin_} will
3313 always be treated as having the same meaning as the C library function
3314 even if you specify the @samp{-fno-builtin} (@pxref{C Dialect Options})
3315 option. Many of these functions are only optimized in certain cases; if
3316 not optimized in a particular case, a call to the library function will
3319 The functions @code{abort}, @code{exit}, and @code{_exit} are recognized
3320 and presumed not to return, but otherwise are not built in.
3321 @code{_exit} is not recognized in strict ISO C mode (@samp{-ansi},
3322 @samp{-std=c89} or @samp{-std=c99}).
3324 Outside strict ISO C mode, the functions @code{alloca}, @code{bcmp},
3325 @code{bzero}, @code{index}, @code{rindex} and @code{ffs} may be handled
3326 as builtins. Corresponding versions @code{__builtin_alloca},
3327 @code{__builtin_bcmp}, @code{__builtin_bzero}, @code{__builtin_index},
3328 @code{__builtin_rindex} and @code{__builtin_ffs} are also recognized in
3331 The ISO C99 function @code{llabs} is handled as a builtin except in
3332 strict ISO C89 mode. There are also builtin versions of the ISO C99
3333 functions @code{cosf}, @code{cosl}, @code{fabsf}, @code{fabsl},
3334 @code{sinf}, @code{sinl}, @code{sqrtf}, and @code{sqrtl}, that are
3335 recognized in any mode since ISO C89 reserves these names for the
3336 purpose to which ISO C99 puts them. All these functions have
3337 corresponding versions prefixed with @code{__builtin_}.
3339 The following ISO C89 functions are recognized as builtins unless
3340 @samp{-fno-builtin} is specified: @code{abs}, @code{cos}, @code{fabs},
3341 @code{fputs}, @code{labs}, @code{memcmp}, @code{memcpy}, @code{memset},
3342 @code{printf}, @code{sin}, @code{sqrt}, @code{strchr}, @code{strcmp},
3343 @code{strcpy}, @code{strlen}, @code{strncmp}, @code{strncpy},
3344 @code{strpbrk}, @code{strrchr}, and @code{strstr}. All of these
3345 functions have corresponding versions prefixed with @code{__builtin_},
3346 except that the version for @code{sqrt} is called
3347 @code{__builtin_fsqrt}.
3349 GNU CC provides builtin versions of the ISO C99 floating point
3350 comparison macros (that avoid raising exceptions for unordered
3351 operands): @code{__builtin_isgreater}, @code{__builtin_isgreaterequal},
3352 @code{__builtin_isless}, @code{__builtin_islessequal},
3353 @code{__builtin_islessgreater}, and @code{__builtin_isunordered}.
3357 @findex __builtin_constant_p
3358 @item __builtin_constant_p (@var{exp})
3359 You can use the builtin function @code{__builtin_constant_p} to
3360 determine if a value is known to be constant at compile-time and hence
3361 that GNU CC can perform constant-folding on expressions involving that
3362 value. The argument of the function is the value to test. The function
3363 returns the integer 1 if the argument is known to be a compile-time
3364 constant and 0 if it is not known to be a compile-time constant. A
3365 return of 0 does not indicate that the value is @emph{not} a constant,
3366 but merely that GNU CC cannot prove it is a constant with the specified
3367 value of the @samp{-O} option.
3369 You would typically use this function in an embedded application where
3370 memory was a critical resource. If you have some complex calculation,
3371 you may want it to be folded if it involves constants, but need to call
3372 a function if it does not. For example:
3375 #define Scale_Value(X) \
3376 (__builtin_constant_p (X) ? ((X) * SCALE + OFFSET) : Scale (X))
3379 You may use this builtin function in either a macro or an inline
3380 function. However, if you use it in an inlined function and pass an
3381 argument of the function as the argument to the builtin, GNU CC will
3382 never return 1 when you call the inline function with a string constant
3383 or constructor expression (@pxref{Constructors}) and will not return 1
3384 when you pass a constant numeric value to the inline function unless you
3385 specify the @samp{-O} option.
3387 @findex __builtin_expect
3388 @item __builtin_expect(@var{exp}, @var{c})
3389 You may use @code{__builtin_expect} to provide the compiler with
3390 branch prediction information. In general, you should prefer to
3391 use actual profile feedback for this (@samp{-fprofile-arcs}), as
3392 programmers are notoriously bad at predicting how their programs
3393 actually perform. However, there are applications in which this
3394 data is hard to collect.
3396 The return value is the value of @var{exp}, which should be an
3397 integral expression. The value of @var{c} must be a compile-time
3398 constant. The semantics of the builtin are that it is expected
3399 that @var{exp} == @var{c}. For example:
3402 if (__builtin_expect (x, 0))
3407 would indicate that we do not expect to call @code{foo}, since
3408 we expect @code{x} to be zero. Since you are limited to integral
3409 expressions for @var{exp}, you should use constructions such as
3412 if (__builtin_expect (ptr != NULL, 1))
3417 when testing pointer or floating-point values.
3420 @node Deprecated Features
3421 @section Deprecated Features
3423 In the past, the GNU C++ compiler was extended to experiment with new
3424 features, at a time when the C++ language was still evolving. Now that
3425 the C++ standard is complete, some of those features are superseded by
3426 superior alternatives. Using the old features might cause a warning in
3427 some cases that the feature will be dropped in the future. In other
3428 cases, the feature might be gone already.
3430 While the list below is not exhaustive, it documents some of the options
3431 that are now deprecated:
3434 @item -fexternal-templates
3435 @itemx -falt-external-templates
3436 These are two of the many ways for g++ to implement template
3437 instantiation. @xref{Template Instantiation}. The C++ standard clearly
3438 defines how template definitions have to be organized across
3439 implementation units. g++ has an implicit instantiation mechanism that
3440 should work just fine for standard-conforming code.
3442 @item -fstrict-prototype
3443 @itemx -fno-strict-prototype
3444 Previously it was possible to use an empty prototype parameter list to
3445 indicate an unspecified number of parameters (like C), rather than no
3446 parameters, as C++ demands. This feature has been removed, except where
3447 it is required for backwards compatibility @xref{Backwards Compatibility}.
3450 The named return value extension has been deprecated, and will be
3451 removed from g++ at some point.
3453 @node Backwards Compatibility
3454 @section Backwards Compatibility
3455 @cindex Backwards Compatibility
3458 Now that there is a definitive ISO standard C++, g++ has a specification
3459 to adhere to. The C++ language evolved over time, and features that used
3460 to be acceptable in previous drafts of the standard, such as the ARM, are
3461 no longer accepted. In order to allow compilation of C++ written to such
3462 drafts, g++ contains some backwards compatibilities. @emph{All such
3463 backwards compatibility features are liable to disappear in future
3464 versions of g++.} They should be considered deprecated @xref{Deprecated
3469 If a variable is declared at for scope, it used to remain in scope until
3470 the end of the scope which contained the for statement (rather than just
3471 within the for scope). g++ retains this, but issues a warning, if such a
3472 variable is accessed outside the for scope.
3474 @item implicit C language
3475 Old C system header files did not contain an @code{extern "C" @{...@}}
3476 scope to set the language. On such systems, all header files are
3477 implicitly scoped inside a C language scope. Also, an empty prototype
3478 @code{()} will be treated as an unspecified number of arguments, rather
3479 than no arguments, as C++ demands.
3482 @node C++ Extensions
3483 @chapter Extensions to the C++ Language
3484 @cindex extensions, C++ language
3485 @cindex C++ language extensions
3487 The GNU compiler provides these extensions to the C++ language (and you
3488 can also use most of the C language extensions in your C++ programs). If you
3489 want to write code that checks whether these features are available, you can
3490 test for the GNU compiler the same way as for C programs: check for a
3491 predefined macro @code{__GNUC__}. You can also use @code{__GNUG__} to
3492 test specifically for GNU C++ (@pxref{Standard Predefined,,Standard
3493 Predefined Macros,cpp.info,The C Preprocessor}).
3496 * Min and Max:: C++ Minimum and maximum operators.
3497 * Volatiles:: What constitutes an access to a volatile object.
3498 * Restricted Pointers:: C99 restricted pointers and references.
3499 * C++ Interface:: You can use a single C++ header file for both
3500 declarations and definitions.
3501 * Template Instantiation:: Methods for ensuring that exactly one copy of
3502 each needed template instantiation is emitted.
3503 * Bound member functions:: You can extract a function pointer to the
3504 method denoted by a @samp{->*} or @samp{.*} expression.
3508 @section Minimum and Maximum Operators in C++
3510 It is very convenient to have operators which return the ``minimum'' or the
3511 ``maximum'' of two arguments. In GNU C++ (but not in GNU C),
3514 @item @var{a} <? @var{b}
3516 @cindex minimum operator
3517 is the @dfn{minimum}, returning the smaller of the numeric values
3518 @var{a} and @var{b};
3520 @item @var{a} >? @var{b}
3522 @cindex maximum operator
3523 is the @dfn{maximum}, returning the larger of the numeric values @var{a}
3527 These operations are not primitive in ordinary C++, since you can
3528 use a macro to return the minimum of two things in C++, as in the
3532 #define MIN(X,Y) ((X) < (Y) ? : (X) : (Y))
3536 You might then use @w{@samp{int min = MIN (i, j);}} to set @var{min} to
3537 the minimum value of variables @var{i} and @var{j}.
3539 However, side effects in @code{X} or @code{Y} may cause unintended
3540 behavior. For example, @code{MIN (i++, j++)} will fail, incrementing
3541 the smaller counter twice. A GNU C extension allows you to write safe
3542 macros that avoid this kind of problem (@pxref{Naming Types,,Naming an
3543 Expression's Type}). However, writing @code{MIN} and @code{MAX} as
3544 macros also forces you to use function-call notation for a
3545 fundamental arithmetic operation. Using GNU C++ extensions, you can
3546 write @w{@samp{int min = i <? j;}} instead.
3548 Since @code{<?} and @code{>?} are built into the compiler, they properly
3549 handle expressions with side-effects; @w{@samp{int min = i++ <? j++;}}
3553 @section When is a Volatile Object Accessed?
3554 @cindex accessing volatiles
3555 @cindex volatile read
3556 @cindex volatile write
3557 @cindex volatile access
3559 Both the C and C++ standard have the concept of volatile objects. These
3560 are normally accessed by pointers and used for accessing hardware. The
3561 standards encourage compilers to refrain from optimizations
3562 concerning accesses to volatile objects that it might perform on
3563 non-volatile objects. The C standard leaves it implementation defined
3564 as to what constitutes a volatile access. The C++ standard omits to
3565 specify this, except to say that C++ should behave in a similar manner
3566 to C with respect to volatiles, where possible. The minimum either
3567 standard specifies is that at a sequence point all previous accesses to
3568 volatile objects have stabilized and no subsequent accesses have
3569 occurred. Thus an implementation is free to reorder and combine
3570 volatile accesses which occur between sequence points, but cannot do so
3571 for accesses across a sequence point. The use of volatiles does not
3572 allow you to violate the restriction on updating objects multiple times
3573 within a sequence point.
3575 In most expressions, it is intuitively obvious what is a read and what is
3576 a write. For instance
3579 volatile int *dst = <somevalue>;
3580 volatile int *src = <someothervalue>;
3585 will cause a read of the volatile object pointed to by @var{src} and stores the
3586 value into the volatile object pointed to by @var{dst}. There is no
3587 guarantee that these reads and writes are atomic, especially for objects
3588 larger than @code{int}.
3590 Less obvious expressions are where something which looks like an access
3591 is used in a void context. An example would be,
3594 volatile int *src = <somevalue>;
3598 With C, such expressions are rvalues, and as rvalues cause a read of
3599 the object, gcc interprets this as a read of the volatile being pointed
3600 to. The C++ standard specifies that such expressions do not undergo
3601 lvalue to rvalue conversion, and that the type of the dereferenced
3602 object may be incomplete. The C++ standard does not specify explicitly
3603 that it is this lvalue to rvalue conversion which is responsible for
3604 causing an access. However, there is reason to believe that it is,
3605 because otherwise certain simple expressions become undefined. However,
3606 because it would surprise most programmers, g++ treats dereferencing a
3607 pointer to volatile object of complete type in a void context as a read
3608 of the object. When the object has incomplete type, g++ issues a
3613 struct T @{int m;@};
3614 volatile S *ptr1 = <somevalue>;
3615 volatile T *ptr2 = <somevalue>;
3620 In this example, a warning is issued for @code{*ptr1}, and @code{*ptr2}
3621 causes a read of the object pointed to. If you wish to force an error on
3622 the first case, you must force a conversion to rvalue with, for instance
3623 a static cast, @code{static_cast<S>(*ptr1)}.
3625 When using a reference to volatile, g++ does not treat equivalent
3626 expressions as accesses to volatiles, but instead issues a warning that
3627 no volatile is accessed. The rationale for this is that otherwise it
3628 becomes difficult to determine where volatile access occur, and not
3629 possible to ignore the return value from functions returning volatile
3630 references. Again, if you wish to force a read, cast the reference to
3633 @node Restricted Pointers
3634 @section Restricting Pointer Aliasing
3635 @cindex restricted pointers
3636 @cindex restricted references
3637 @cindex restricted this pointer
3639 As with gcc, g++ understands the C99 feature of restricted pointers,
3640 specified with the @code{__restrict__}, or @code{__restrict} type
3641 qualifier. Because you cannot compile C++ by specifying the -std=c99
3642 language flag, @code{restrict} is not a keyword in C++.
3644 In addition to allowing restricted pointers, you can specify restricted
3645 references, which indicate that the reference is not aliased in the local
3649 void fn (int *__restrict__ rptr, int &__restrict__ rref)
3656 In the body of @code{fn}, @var{rptr} points to an unaliased integer and
3657 @var{rref} refers to a (different) unaliased integer.
3659 You may also specify whether a member function's @var{this} pointer is
3660 unaliased by using @code{__restrict__} as a member function qualifier.
3663 void T::fn () __restrict__
3670 Within the body of @code{T::fn}, @var{this} will have the effective
3671 definition @code{T *__restrict__ const this}. Notice that the
3672 interpretation of a @code{__restrict__} member function qualifier is
3673 different to that of @code{const} or @code{volatile} qualifier, in that it
3674 is applied to the pointer rather than the object. This is consistent with
3675 other compilers which implement restricted pointers.
3677 As with all outermost parameter qualifiers, @code{__restrict__} is
3678 ignored in function definition matching. This means you only need to
3679 specify @code{__restrict__} in a function definition, rather than
3680 in a function prototype as well.
3683 @section Declarations and Definitions in One Header
3685 @cindex interface and implementation headers, C++
3686 @cindex C++ interface and implementation headers
3687 C++ object definitions can be quite complex. In principle, your source
3688 code will need two kinds of things for each object that you use across
3689 more than one source file. First, you need an @dfn{interface}
3690 specification, describing its structure with type declarations and
3691 function prototypes. Second, you need the @dfn{implementation} itself.
3692 It can be tedious to maintain a separate interface description in a
3693 header file, in parallel to the actual implementation. It is also
3694 dangerous, since separate interface and implementation definitions may
3695 not remain parallel.
3697 @cindex pragmas, interface and implementation
3698 With GNU C++, you can use a single header file for both purposes.
3701 @emph{Warning:} The mechanism to specify this is in transition. For the
3702 nonce, you must use one of two @code{#pragma} commands; in a future
3703 release of GNU C++, an alternative mechanism will make these
3704 @code{#pragma} commands unnecessary.
3707 The header file contains the full definitions, but is marked with
3708 @samp{#pragma interface} in the source code. This allows the compiler
3709 to use the header file only as an interface specification when ordinary
3710 source files incorporate it with @code{#include}. In the single source
3711 file where the full implementation belongs, you can use either a naming
3712 convention or @samp{#pragma implementation} to indicate this alternate
3713 use of the header file.
3716 @item #pragma interface
3717 @itemx #pragma interface "@var{subdir}/@var{objects}.h"
3718 @kindex #pragma interface
3719 Use this directive in @emph{header files} that define object classes, to save
3720 space in most of the object files that use those classes. Normally,
3721 local copies of certain information (backup copies of inline member
3722 functions, debugging information, and the internal tables that implement
3723 virtual functions) must be kept in each object file that includes class
3724 definitions. You can use this pragma to avoid such duplication. When a
3725 header file containing @samp{#pragma interface} is included in a
3726 compilation, this auxiliary information will not be generated (unless
3727 the main input source file itself uses @samp{#pragma implementation}).
3728 Instead, the object files will contain references to be resolved at link
3731 The second form of this directive is useful for the case where you have
3732 multiple headers with the same name in different directories. If you
3733 use this form, you must specify the same string to @samp{#pragma
3736 @item #pragma implementation
3737 @itemx #pragma implementation "@var{objects}.h"
3738 @kindex #pragma implementation
3739 Use this pragma in a @emph{main input file}, when you want full output from
3740 included header files to be generated (and made globally visible). The
3741 included header file, in turn, should use @samp{#pragma interface}.
3742 Backup copies of inline member functions, debugging information, and the
3743 internal tables used to implement virtual functions are all generated in
3744 implementation files.
3746 @cindex implied @code{#pragma implementation}
3747 @cindex @code{#pragma implementation}, implied
3748 @cindex naming convention, implementation headers
3749 If you use @samp{#pragma implementation} with no argument, it applies to
3750 an include file with the same basename@footnote{A file's @dfn{basename}
3751 was the name stripped of all leading path information and of trailing
3752 suffixes, such as @samp{.h} or @samp{.C} or @samp{.cc}.} as your source
3753 file. For example, in @file{allclass.cc}, giving just
3754 @samp{#pragma implementation}
3755 by itself is equivalent to @samp{#pragma implementation "allclass.h"}.
3757 In versions of GNU C++ prior to 2.6.0 @file{allclass.h} was treated as
3758 an implementation file whenever you would include it from
3759 @file{allclass.cc} even if you never specified @samp{#pragma
3760 implementation}. This was deemed to be more trouble than it was worth,
3761 however, and disabled.
3763 If you use an explicit @samp{#pragma implementation}, it must appear in
3764 your source file @emph{before} you include the affected header files.
3766 Use the string argument if you want a single implementation file to
3767 include code from multiple header files. (You must also use
3768 @samp{#include} to include the header file; @samp{#pragma
3769 implementation} only specifies how to use the file---it doesn't actually
3772 There is no way to split up the contents of a single header file into
3773 multiple implementation files.
3776 @cindex inlining and C++ pragmas
3777 @cindex C++ pragmas, effect on inlining
3778 @cindex pragmas in C++, effect on inlining
3779 @samp{#pragma implementation} and @samp{#pragma interface} also have an
3780 effect on function inlining.
3782 If you define a class in a header file marked with @samp{#pragma
3783 interface}, the effect on a function defined in that class is similar to
3784 an explicit @code{extern} declaration---the compiler emits no code at
3785 all to define an independent version of the function. Its definition
3786 is used only for inlining with its callers.
3788 Conversely, when you include the same header file in a main source file
3789 that declares it as @samp{#pragma implementation}, the compiler emits
3790 code for the function itself; this defines a version of the function
3791 that can be found via pointers (or by callers compiled without
3792 inlining). If all calls to the function can be inlined, you can avoid
3793 emitting the function by compiling with @samp{-fno-implement-inlines}.
3794 If any calls were not inlined, you will get linker errors.
3796 @node Template Instantiation
3797 @section Where's the Template?
3799 @cindex template instantiation
3801 C++ templates are the first language feature to require more
3802 intelligence from the environment than one usually finds on a UNIX
3803 system. Somehow the compiler and linker have to make sure that each
3804 template instance occurs exactly once in the executable if it is needed,
3805 and not at all otherwise. There are two basic approaches to this
3806 problem, which I will refer to as the Borland model and the Cfront model.
3810 Borland C++ solved the template instantiation problem by adding the code
3811 equivalent of common blocks to their linker; the compiler emits template
3812 instances in each translation unit that uses them, and the linker
3813 collapses them together. The advantage of this model is that the linker
3814 only has to consider the object files themselves; there is no external
3815 complexity to worry about. This disadvantage is that compilation time
3816 is increased because the template code is being compiled repeatedly.
3817 Code written for this model tends to include definitions of all
3818 templates in the header file, since they must be seen to be
3822 The AT&T C++ translator, Cfront, solved the template instantiation
3823 problem by creating the notion of a template repository, an
3824 automatically maintained place where template instances are stored. A
3825 more modern version of the repository works as follows: As individual
3826 object files are built, the compiler places any template definitions and
3827 instantiations encountered in the repository. At link time, the link
3828 wrapper adds in the objects in the repository and compiles any needed
3829 instances that were not previously emitted. The advantages of this
3830 model are more optimal compilation speed and the ability to use the
3831 system linker; to implement the Borland model a compiler vendor also
3832 needs to replace the linker. The disadvantages are vastly increased
3833 complexity, and thus potential for error; for some code this can be
3834 just as transparent, but in practice it can been very difficult to build
3835 multiple programs in one directory and one program in multiple
3836 directories. Code written for this model tends to separate definitions
3837 of non-inline member templates into a separate file, which should be
3838 compiled separately.
3841 When used with GNU ld version 2.8 or later on an ELF system such as
3842 Linux/GNU or Solaris 2, or on Microsoft Windows, g++ supports the
3843 Borland model. On other systems, g++ implements neither automatic
3846 A future version of g++ will support a hybrid model whereby the compiler
3847 will emit any instantiations for which the template definition is
3848 included in the compile, and store template definitions and
3849 instantiation context information into the object file for the rest.
3850 The link wrapper will extract that information as necessary and invoke
3851 the compiler to produce the remaining instantiations. The linker will
3852 then combine duplicate instantiations.
3854 In the mean time, you have the following options for dealing with
3855 template instantiations:
3859 Compile your template-using code with @samp{-frepo}. The compiler will
3860 generate files with the extension @samp{.rpo} listing all of the
3861 template instantiations used in the corresponding object files which
3862 could be instantiated there; the link wrapper, @samp{collect2}, will
3863 then update the @samp{.rpo} files to tell the compiler where to place
3864 those instantiations and rebuild any affected object files. The
3865 link-time overhead is negligible after the first pass, as the compiler
3866 will continue to place the instantiations in the same files.
3868 This is your best option for application code written for the Borland
3869 model, as it will just work. Code written for the Cfront model will
3870 need to be modified so that the template definitions are available at
3871 one or more points of instantiation; usually this is as simple as adding
3872 @code{#include <tmethods.cc>} to the end of each template header.
3874 For library code, if you want the library to provide all of the template
3875 instantiations it needs, just try to link all of its object files
3876 together; the link will fail, but cause the instantiations to be
3877 generated as a side effect. Be warned, however, that this may cause
3878 conflicts if multiple libraries try to provide the same instantiations.
3879 For greater control, use explicit instantiation as described in the next
3883 Compile your code with @samp{-fno-implicit-templates} to disable the
3884 implicit generation of template instances, and explicitly instantiate
3885 all the ones you use. This approach requires more knowledge of exactly
3886 which instances you need than do the others, but it's less
3887 mysterious and allows greater control. You can scatter the explicit
3888 instantiations throughout your program, perhaps putting them in the
3889 translation units where the instances are used or the translation units
3890 that define the templates themselves; you can put all of the explicit
3891 instantiations you need into one big file; or you can create small files
3898 template class Foo<int>;
3899 template ostream& operator <<
3900 (ostream&, const Foo<int>&);
3903 for each of the instances you need, and create a template instantiation
3906 If you are using Cfront-model code, you can probably get away with not
3907 using @samp{-fno-implicit-templates} when compiling files that don't
3908 @samp{#include} the member template definitions.
3910 If you use one big file to do the instantiations, you may want to
3911 compile it without @samp{-fno-implicit-templates} so you get all of the
3912 instances required by your explicit instantiations (but not by any
3913 other files) without having to specify them as well.
3915 g++ has extended the template instantiation syntax outlined in the
3916 Working Paper to allow forward declaration of explicit instantiations
3917 (with @code{extern}), instantiation of the compiler support data for a
3918 template class (i.e. the vtable) without instantiating any of its
3919 members (with @code{inline}), and instantiation of only the static data
3920 members of a template class, without the support data or member
3921 functions (with (@code{static}):
3924 extern template int max (int, int);
3925 inline template class Foo<int>;
3926 static template class Foo<int>;
3930 Do nothing. Pretend g++ does implement automatic instantiation
3931 management. Code written for the Borland model will work fine, but
3932 each translation unit will contain instances of each of the templates it
3933 uses. In a large program, this can lead to an unacceptable amount of code
3937 Add @samp{#pragma interface} to all files containing template
3938 definitions. For each of these files, add @samp{#pragma implementation
3939 "@var{filename}"} to the top of some @samp{.C} file which
3940 @samp{#include}s it. Then compile everything with
3941 @samp{-fexternal-templates}. The templates will then only be expanded
3942 in the translation unit which implements them (i.e. has a @samp{#pragma
3943 implementation} line for the file where they live); all other files will
3944 use external references. If you're lucky, everything should work
3945 properly. If you get undefined symbol errors, you need to make sure
3946 that each template instance which is used in the program is used in the
3947 file which implements that template. If you don't have any use for a
3948 particular instance in that file, you can just instantiate it
3949 explicitly, using the syntax from the latest C++ working paper:
3952 template class A<int>;
3953 template ostream& operator << (ostream&, const A<int>&);
3956 This strategy will work with code written for either model. If you are
3957 using code written for the Cfront model, the file containing a class
3958 template and the file containing its member templates should be
3959 implemented in the same translation unit.
3961 A slight variation on this approach is to instead use the flag
3962 @samp{-falt-external-templates}; this flag causes template
3963 instances to be emitted in the translation unit that implements the
3964 header where they are first instantiated, rather than the one which
3965 implements the file where the templates are defined. This header must
3966 be the same in all translation units, or things are likely to break.
3968 @xref{C++ Interface,,Declarations and Definitions in One Header}, for
3969 more discussion of these pragmas.
3972 @node Bound member functions
3973 @section Extracting the function pointer from a bound pointer to member function
3976 @cindex pointer to member function
3977 @cindex bound pointer to member function
3979 In C++, pointer to member functions (PMFs) are implemented using a wide
3980 pointer of sorts to handle all the possible call mechanisms; the PMF
3981 needs to store information about how to adjust the @samp{this} pointer,
3982 and if the function pointed to is virtual, where to find the vtable, and
3983 where in the vtable to look for the member function. If you are using
3984 PMFs in an inner loop, you should really reconsider that decision. If
3985 that is not an option, you can extract the pointer to the function that
3986 would be called for a given object/PMF pair and call it directly inside
3987 the inner loop, to save a bit of time.
3989 Note that you will still be paying the penalty for the call through a
3990 function pointer; on most modern architectures, such a call defeats the
3991 branch prediction features of the CPU. This is also true of normal
3992 virtual function calls.
3994 The syntax for this extension is
3998 extern int (A::*fp)();
3999 typedef int (*fptr)(A *);
4001 fptr p = (fptr)(a.*fp);
4004 For PMF constants (i.e. expressions of the form @samp{&Klasse::Member}),
4005 no object is needed to obtain the address of the function. They can be
4006 converted to function pointers directly:
4009 fptr p1 = (fptr)(&A::foo);
4012 You must specify @samp{-Wno-pmf-conversions} to use this extension.