https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90106
--- Comment #8 from JunMa <JunMa at linux dot alibaba.com> ---
(In reply to Alexander Monakov from comment #6)
> Reopening and confirming, GCC's code looks less efficient than possible for
> no good reason.
>
> CDCE does
>
> y = sqrt (x);
> ==>
> y = IFN_SQRT (x);
> if (__builtin_isless (x, 0))
> sqrt (x);
>
> but it could do
>
> y = IFN_SQRT (x);
> if (__builtin_isless (x, 0))
> y = sqrt (x);
>
> (note two assignments to y)
>
what is the difference between this and LLVM's approach ?
> or to mimic LLVM's approach:
>
> if (__builtin_isless (x, 0))
> y = sqrt (x);
> else
> y = IFN_SQRT (x);
I have finished a patch which do as same as LLVM in cdce pass, and test with
case below:
#include <math.h>
int main () {
float x = 1.0;
float y;
for (int i=0; i<100000000; i++) {
y += sqrtf (x+i);
}
return y;
}
And I've got, for x86-64 with O2:
# original asm of IFN_SQRT part
.L4:
pxor %xmm0, %xmm0
cvtsi2ssl %ebx, %xmm0
addss %xmm3, %xmm0
ucomiss %xmm0, %xmm4
movaps %xmm0, %xmm2
sqrtss %xmm2, %xmm2
ja .L7
and perf stat :
1,423,652,277 cycles # 2.180 GHz
(83.31%)
1,121,862,980 stalled-cycles-frontend # 78.80% frontend cycles
idle (83.31%)
634,957,413 stalled-cycles-backend # 44.60% backend cycles
idle (66.62%)
1,102,109,423 instructions # 0.77 insn per cycle
# 1.02 stalled cycles per
insn (83.31%)
200,400,940 branches # 306.873 M/sec
(83.44%)
7,734 branch-misses # 0.00% of all branches
(83.44%)
#transformed asm :
.L4:
pxor %xmm0, %xmm0
cvtsi2ssl %ebx, %xmm0
addss %xmm3, %xmm0
ucomiss %xmm0, %xmm2
ja .L8
sqrtss %xmm0, %xmm0
and perf stat:
1,418,560,722 cycles # 2.180 GHz
(83.25%)
1,116,732,674 stalled-cycles-frontend # 78.72% frontend cycles
idle (83.25%)
674,837,417 stalled-cycles-backend # 47.57% backend cycles
idle (66.81%)
1,003,067,037 instructions # 0.71 insn per cycle
# 1.11 stalled cycles per
insn (83.41%)
200,619,151 branches # 308.272 M/sec
(83.40%)
5,637 branch-misses # 0.00% of all branches
(83.28%)
The transformed case has less instructions and gets better performance which
looks good to me. However, one thing that I noticed is the original case gets
less 'stalled-cycles-backend', since its code has better ILP.
I'm not sure which approach is better.
Environment:
gcc version: gcc trunk@270488
OS: centos7.2
HW: Intel(R) Xeon(R) CPU E5-2430 0 @ 2.20GHz
