http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55723
vincenzo Innocente <vincenzo.innocente at cern dot ch> changed:
What |Removed |Added
----------------------------------------------------------------------------
Summary|SLP vectorization vs loop: |loop vectorization
|SLP more efficient: loop |inefficient in presence of
|vectorization inefficient |multiple identical
|in presence of multiple |conditions
|"blends" |
--- Comment #2 from vincenzo Innocente <vincenzo.innocente at cern dot ch>
2012-12-20 15:39:13 UTC ---
It seems that in presence of identical conditions the vectorizer prefers to
compute two "full" branches
and do just one blend.
This is not always the most efficient choice as the benchmark in comment 1
demonstrates.
Another simple example:
for bar two rsqrtps and one blend
for foo one rsqrtps and two blends
#include<cmath>
float a[1024];
float b[1024];
void bar(){
for (int i=0;i!=1024;++i) {
auto z = a[i];
if (a[i] > 3.14f) z-=1.f;
b[i] = 1.f/std::sqrt(z);
if (a[i] > 3.14f) b[i]-=1.f;
}
}
void foo(){
for (int i=0;i!=1024;++i) {
auto z = a[i];
if (a[i] > 3.14f) z-=1.f;
b[i] = 1.f/std::sqrt(z);
if (a[i] > 1.f) b[i]-=1.f;
}
}
c++ -std=c++11 -Ofast -march=corei7 -S twoif.cc -ftree-vectorizer-verbose=1
-ftree-loop-if-convert-stores; cat twoif.s | c++filt
bar():
LFB221:
movaps LC0(%rip), %xmm6
leaq signed char(%rip), %rax
movaps LC1(%rip), %xmm5
leaq bool(%rip), %rdx
movaps LC2(%rip), %xmm4
leaq 4096+signed char(%rip), %rcx
movaps LC3(%rip), %xmm7
.align 4,0x90
L3:
movaps (%rax), %xmm0
addq $16, %rax
addq $16, %rdx
rsqrtps %xmm0, %xmm3
movaps %xmm0, %xmm2
subps %xmm6, %xmm2
rsqrtps %xmm2, %xmm1
mulps %xmm1, %xmm2
mulps %xmm1, %xmm2
mulps %xmm4, %xmm1
addps %xmm5, %xmm2
mulps %xmm1, %xmm2
movaps %xmm3, %xmm1
mulps %xmm0, %xmm1
subps %xmm6, %xmm2
mulps %xmm3, %xmm1
mulps %xmm4, %xmm3
addps %xmm5, %xmm1
mulps %xmm3, %xmm1
movaps %xmm7, %xmm3
cmpltps %xmm0, %xmm3
movaps %xmm3, %xmm0
blendvps %xmm0, %xmm2, %xmm1
movaps %xmm1, -16(%rdx)
cmpq %rcx, %rax
jne L3
rep; ret
LFE221:
.align 4,0x90
.globl foo()
foo():
LFB222:
movaps LC3(%rip), %xmm7
leaq signed char(%rip), %rax
movaps LC0(%rip), %xmm3
leaq bool(%rip), %rdx
movaps LC1(%rip), %xmm6
leaq 4096+signed char(%rip), %rcx
movaps LC2(%rip), %xmm5
.align 4,0x90
L7:
movaps (%rax), %xmm2
movaps %xmm7, %xmm0
addq $16, %rax
addq $16, %rdx
movaps %xmm2, %xmm1
cmpltps %xmm2, %xmm0
movaps %xmm2, %xmm4
subps %xmm3, %xmm1
blendvps %xmm0, %xmm1, %xmm4
rsqrtps %xmm4, %xmm0
movaps %xmm4, %xmm1
mulps %xmm0, %xmm1
mulps %xmm0, %xmm1
mulps %xmm5, %xmm0
addps %xmm6, %xmm1
mulps %xmm0, %xmm1
movaps %xmm3, %xmm0
cmpltps %xmm2, %xmm0
movaps %xmm1, %xmm4
subps %xmm3, %xmm4
blendvps %xmm0, %xmm4, %xmm1
movaps %xmm1, -16(%rdx)
cmpq %rcx, %rax
jne L7
rep; ret
