https://llvm.org/bugs/show_bug.cgi?id=31800
Bug ID: 31800
Summary: clang/llvm vectorize the sum of a complex array poorly
Product: libraries
Version: 3.9
Hardware: PC
OS: Linux
Status: NEW
Severity: normal
Priority: P
Component: Loop Optimizer
Assignee: unassignedb...@nondot.org
Reporter: drr...@gmail.com
CC: llvm-bugs@lists.llvm.org
Classification: Unclassified
Consider this code:
#include <complex.h>
complex float f(complex float x[]) {
complex float p = 1.0;
for (int i = 0; i < 32; i++)
p += x[i];
return p;
}
clang 3.9.1 with -O3 -march=core-avx2 -ffast-math gives
f: # @f
vmovq xmm0, qword ptr [rdi] # xmm0 = mem[0],zero
vmovss xmm1, dword ptr [rip + .LCPI0_0] # xmm1 = mem[0],zero,zero,zero
vaddps xmm0, xmm0, xmm1
vmovq xmm1, qword ptr [rdi + 8] # xmm1 = mem[0],zero
vmovq xmm2, qword ptr [rdi + 16] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vaddps xmm0, xmm0, xmm1
vmovq xmm1, qword ptr [rdi + 24] # xmm1 = mem[0],zero
vmovq xmm2, qword ptr [rdi + 32] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vmovq xmm2, qword ptr [rdi + 40] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vaddps xmm0, xmm0, xmm1
vmovq xmm1, qword ptr [rdi + 48] # xmm1 = mem[0],zero
vmovq xmm2, qword ptr [rdi + 56] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vmovq xmm2, qword ptr [rdi + 64] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vmovq xmm2, qword ptr [rdi + 72] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vaddps xmm0, xmm0, xmm1
vmovq xmm1, qword ptr [rdi + 80] # xmm1 = mem[0],zero
vmovq xmm2, qword ptr [rdi + 88] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vmovq xmm2, qword ptr [rdi + 96] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vmovq xmm2, qword ptr [rdi + 104] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vmovq xmm2, qword ptr [rdi + 112] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vaddps xmm0, xmm0, xmm1
vmovq xmm1, qword ptr [rdi + 120] # xmm1 = mem[0],zero
vmovq xmm2, qword ptr [rdi + 128] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vmovq xmm2, qword ptr [rdi + 136] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vmovq xmm2, qword ptr [rdi + 144] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vmovq xmm2, qword ptr [rdi + 152] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vmovq xmm2, qword ptr [rdi + 160] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vaddps xmm0, xmm0, xmm1
vmovq xmm1, qword ptr [rdi + 168] # xmm1 = mem[0],zero
vmovq xmm2, qword ptr [rdi + 176] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vmovq xmm2, qword ptr [rdi + 184] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vmovq xmm2, qword ptr [rdi + 192] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vmovq xmm2, qword ptr [rdi + 200] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vmovq xmm2, qword ptr [rdi + 208] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vmovq xmm2, qword ptr [rdi + 216] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vaddps xmm0, xmm0, xmm1
vmovq xmm1, qword ptr [rdi + 224] # xmm1 = mem[0],zero
vmovq xmm2, qword ptr [rdi + 232] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vmovq xmm2, qword ptr [rdi + 240] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vmovq xmm2, qword ptr [rdi + 248] # xmm2 = mem[0],zero
vaddps xmm1, xmm1, xmm2
vaddps xmm0, xmm0, xmm1
ret
The only vectorization is that the real and the imaginary parts are added in
parallel. The assembly also wastes half of the xmm register.
However in icc you get:
f:
vmovups ymm1, YMMWORD PTR [rdi] #5.10
vmovups ymm2, YMMWORD PTR [64+rdi] #5.10
vmovups ymm5, YMMWORD PTR [128+rdi] #5.10
vmovups ymm6, YMMWORD PTR [192+rdi] #5.10
vmovsd xmm0, QWORD PTR p.152.0.0.1[rip] #3.19
vaddps ymm3, ymm1, YMMWORD PTR [32+rdi] #3.19
vaddps ymm4, ymm2, YMMWORD PTR [96+rdi] #3.19
vaddps ymm7, ymm5, YMMWORD PTR [160+rdi] #3.19
vaddps ymm8, ymm6, YMMWORD PTR [224+rdi] #3.19
vaddps ymm9, ymm3, ymm4 #3.19
vaddps ymm10, ymm7, ymm8 #3.19
vaddps ymm11, ymm9, ymm10 #3.19
vextractf128 xmm12, ymm11, 1 #3.19
vaddps xmm13, xmm11, xmm12 #3.19
vmovhlps xmm14, xmm13, xmm13 #3.19
vaddps xmm15, xmm13, xmm14 #3.19
vaddps xmm0, xmm15, xmm0 #3.19
vzeroupper #6.10
ret
which is fully vectorized (and uses the wider ymm registers).
Another key difference seems to be that in the clang/llvm produced assembly
subsequent additions depend on each other. Whereas in the icc code the
additions work on subsequent items and so it benefits both from full
vectorization and superscalar parallelism.
(This report is related to https://llvm.org/bugs/show_bug.cgi?id=31677 where I
incorrectly stated at the end of the problem report that llvm could vectorise
this additive reduction loop.)
--
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
