3 * Copyright (C) Igor Sysoev
4 * Copyright (C) Nginx, Inc.
9 #define NGX_SMP_LOCK "lock;"
27 * The "r" means the general register.
28 * The "=a" and "a" are the %eax register.
29 * Although we can return result in any register, we use "a" because it is
30 * used in cmpxchgl anyway. The result is actually in %al but not in %eax,
31 * however, as the code is inlined gcc can test %al as well as %eax,
32 * and icc adds "movzbl %al, %eax" by itself.
34 * The "cc" means that flags were changed.
37 static ngx_inline ngx_atomic_uint_t
38 ngx_atomic_cmp_set(ngx_atomic_t
*lock
, ngx_atomic_uint_t old
,
39 ngx_atomic_uint_t set
)
49 : "=a" (res
) : "m" (*lock
), "a" (old
), "r" (set
) : "cc", "memory");
63 * The "+r" means the general register.
64 * The "cc" means that flags were changed.
68 #if !(( __GNUC__ == 2 && __GNUC_MINOR__ <= 7 ) || ( __INTEL_COMPILER >= 800 ))
71 * icc 8.1 and 9.0 compile broken code with -march=pentium4 option:
72 * ngx_atomic_fetch_add() always return the input "add" value,
73 * so we use the gcc 2.7 version.
75 * icc 8.1 and 9.0 with -march=pentiumpro option or icc 7.1 compile
79 static ngx_inline ngx_atomic_int_t
80 ngx_atomic_fetch_add(ngx_atomic_t
*value
, ngx_atomic_int_t add
)
87 : "+r" (add
) : "m" (*value
) : "cc", "memory");
96 * gcc 2.7 does not support "+r", so we have to use the fixed
97 * %eax ("=a" and "a") and this adds two superfluous instructions in the end
98 * of code, something like this: "mov %eax, %edx / mov %edx, %eax".
101 static ngx_inline ngx_atomic_int_t
102 ngx_atomic_fetch_add(ngx_atomic_t
*value
, ngx_atomic_int_t add
)
104 ngx_atomic_uint_t old
;
111 : "=a" (old
) : "m" (*value
), "a" (add
) : "cc", "memory");
120 * on x86 the write operations go in a program order, so we need only
121 * to disable the gcc reorder optimizations
124 #define ngx_memory_barrier() __asm__ volatile ("" ::: "memory")
126 /* old "as" does not support "pause" opcode */
127 #define ngx_cpu_pause() __asm__ (".byte 0xf3, 0x90")