最佳答案
在编写优化的 ftol
函数时,我发现 GCC 4.6.1
中有一些非常奇怪的行为。让我先向您展示代码(为了清楚起见,我标记了不同之处) :
Fast _ trunc _ one,C:
int fast_trunc_one(int i) {
int mantissa, exponent, sign, r;
mantissa = (i & 0x07fffff) | 0x800000;
exponent = 150 - ((i >> 23) & 0xff);
sign = i & 0x80000000;
if (exponent < 0) {
r = mantissa << -exponent; /* diff */
} else {
r = mantissa >> exponent; /* diff */
}
return (r ^ -sign) + sign; /* diff */
}
Fast _ trunc _ two,C:
int fast_trunc_two(int i) {
int mantissa, exponent, sign, r;
mantissa = (i & 0x07fffff) | 0x800000;
exponent = 150 - ((i >> 23) & 0xff);
sign = i & 0x80000000;
if (exponent < 0) {
r = (mantissa << -exponent) ^ -sign; /* diff */
} else {
r = (mantissa >> exponent) ^ -sign; /* diff */
}
return r + sign; /* diff */
}
看起来是一样的吧? GCC 不同意。在用 gcc -O3 -S -Wall -o test.s test.c
编译之后,这是汇编输出:
Fast _ trunc _ one,生成:
_fast_trunc_one:
LFB0:
.cfi_startproc
movl 4(%esp), %eax
movl $150, %ecx
movl %eax, %edx
andl $8388607, %edx
sarl $23, %eax
orl $8388608, %edx
andl $255, %eax
subl %eax, %ecx
movl %edx, %eax
sarl %cl, %eax
testl %ecx, %ecx
js L5
rep
ret
.p2align 4,,7
L5:
negl %ecx
movl %edx, %eax
sall %cl, %eax
ret
.cfi_endproc
Fast _ trunc _ two,生成:
_fast_trunc_two:
LFB1:
.cfi_startproc
pushl %ebx
.cfi_def_cfa_offset 8
.cfi_offset 3, -8
movl 8(%esp), %eax
movl $150, %ecx
movl %eax, %ebx
movl %eax, %edx
sarl $23, %ebx
andl $8388607, %edx
andl $255, %ebx
orl $8388608, %edx
andl $-2147483648, %eax
subl %ebx, %ecx
js L9
sarl %cl, %edx
movl %eax, %ecx
negl %ecx
xorl %ecx, %edx
addl %edx, %eax
popl %ebx
.cfi_remember_state
.cfi_def_cfa_offset 4
.cfi_restore 3
ret
.p2align 4,,7
L9:
.cfi_restore_state
negl %ecx
sall %cl, %edx
movl %eax, %ecx
negl %ecx
xorl %ecx, %edx
addl %edx, %eax
popl %ebx
.cfi_restore 3
.cfi_def_cfa_offset 4
ret
.cfi_endproc
这是 极端的区别。这实际上也显示在配置文件中,fast_trunc_one
比 fast_trunc_two
快30% 左右。现在我的问题是,是什么导致了这一切?