https://llvm.org/bugs/show_bug.cgi?id=31866
Bug ID: 31866
Summary: Squaring a complex float gives inefficient code
Product: new-bugs
Version: trunk
Hardware: PC
OS: Linux
Status: NEW
Severity: normal
Priority: P
Component: new bugs
Assignee: unassignedb...@nondot.org
Reporter: drr...@gmail.com
CC: llvm-bugs@lists.llvm.org
Classification: Unclassified
Consider:
#include <complex.h>
complex float f(complex float x) {
return x*x;
}
clang trunk with -O3 -march=core-avx2 gives
f: # @f
vmovaps xmm2, xmm0
vmovshdup xmm1, xmm2 # xmm1 = xmm2[1,1,3,3]
vmulss xmm0, xmm2, xmm2
vmulss xmm3, xmm1, xmm1
vmulss xmm4, xmm2, xmm1
vsubss xmm0, xmm0, xmm3
vaddss xmm3, xmm4, xmm4
vucomiss xmm0, xmm0
jnp .LBB0_3
vucomiss xmm3, xmm3
jp .LBB0_2
.LBB0_3:
vinsertps xmm0, xmm0, xmm3, 16 # xmm0 = xmm0[0],xmm3[0],xmm0[2,3]
ret
.LBB0_2:
push rax
vmovaps xmm0, xmm2
vmovaps xmm3, xmm1
call __mulsc3
vmovshdup xmm3, xmm0 # xmm3 = xmm0[1,1,3,3]
add rsp, 8
vinsertps xmm0, xmm0, xmm3, 16 # xmm0 = xmm0[0],xmm3[0],xmm0[2,3]
ret
The Intel C compiler with -fp-model strict gives
f:
vmovsldup xmm1, xmm0 #3.12
vmovshdup xmm2, xmm0 #3.12
vshufps xmm3, xmm0, xmm0, 177 #3.12
vmulps xmm4, xmm1, xmm0 #3.12
vmulps xmm5, xmm2, xmm3 #3.12
vaddsubps xmm0, xmm4, xmm5 #3.12
ret
"strict" should be value safe, turn on floating point exception semantics and
also disable fuse multiply add. "precise" is the setting to use if you just
want it to be value safe. -fp-model precise -fp-model except gives
f:
vmovshdup xmm1, xmm0 #3.12
vshufps xmm2, xmm0, xmm0, 177 #3.12
vmulps xmm4, xmm1, xmm2 #3.12
vmovsldup xmm3, xmm0 #3.12
vfmaddsub213ps xmm0, xmm3, xmm4 #3.12
ret
gcc 7 gives code that is shorter than clang does but still call __mulsc3 .
f:
vmovq QWORD PTR [rsp-16], xmm0
vmovss xmm3, DWORD PTR [rsp-12]
vmovss xmm2, DWORD PTR [rsp-16]
vmovaps xmm1, xmm3
vmovaps xmm0, xmm2
jmp __mulsc3
If you enable -ffast-math in clang it is much better although not quite optimal
with:
f: # @f
vmovshdup xmm1, xmm0 # xmm1 = xmm0[1,1,3,3]
vaddss xmm2, xmm0, xmm0
vmulss xmm2, xmm1, xmm2
vmulss xmm1, xmm1, xmm1
vfmsub231ss xmm1, xmm0, xmm0
vinsertps xmm0, xmm1, xmm2, 16 # xmm0 = xmm1[0],xmm2[0],xmm1[2,3]
ret
>From my non-expert eyes it seems there are two questions:
1) In the "no fast-math" case is ICC actually meeting the C99 specs?
2) In the "fast-math" case can clang/llvm be persuaded/changed to use one call
to vmulps instead of two calls to vmulss?
--
You are receiving this mail because:
You are on the CC list for the bug.
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
