http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52975
Bug #: 52975
Summary: Ofast produces not optimized code for vectorized
"converted if"
Classification: Unclassified
Product: gcc
Version: 4.8.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: vincenzo.innoce...@cern.ch
this is a modified version of
gcc/testsuite/gcc.dg/torture/pr52969.c
notice
cmpps $0x1,%xmm2,%xmm1
cmpps $0x2,%xmm3,%xmm0
in case of Ofast
similar with -march=corei7 when blendv is generated
cat ifconv2.cc
int b;
float xsum[100];
float clus[100];
void bar2 ()
{
int j=0;
for (; j<100 ; ++j) {
xsum[j] = clus[j];
if (xsum[j] > 0)
xsum[j] = 0;
// xsum[j] = (clus[j] > 0.) ? 0. : clus[j];
}
if (xsum[0])
b = 0;
}
pb-d-128-141-131-26:bugs48 innocent$ c++ -O3 -c ifconv2.cc
-ftree-loop-if-convert-stores -ftree-vectorizer-verbose=2
Analyzing loop at ifconv2.cc:7
Vectorizing loop at ifconv2.cc:7
7: LOOP VECTORIZED.
ifconv2.cc:4: note: vectorized 1 loops in function.
pb-d-128-141-131-26:bugs48 innocent$ otool -t -X -v ifconv2.o
__Z4bar2v:
leaq 0x00000000(%rip),%rax
xorps %xmm2,%xmm2
leaq 0x00000000(%rip),%rdx
leaq 0x00000190(%rip),%rcx
nopl 0x00000000(%rax,%rax)
movaps (%rax),%xmm1
movaps %xmm2,%xmm0
addq $0x10,%rax
cmpps $0x1,%xmm1,%xmm0
andnps %xmm1,%xmm0
movaps %xmm0,(%rdx)
addq $0x10,%rdx
cmpq %rcx,%rax
jne 0x00000020
xorps %xmm0,%xmm0
ucomiss 0x00000000(%rip),%xmm0
jnp 0x00000054
movl $0x00000000,0xfffffffc(%rip)
ret
jne 0x00000049
repz/ret
pb-d-128-141-131-26:bugs48 innocent$ c++ -Ofast -c ifconv2.cc
-ftree-loop-if-convert-stores -ftree-vectorizer-verbose=2
Analyzing loop at ifconv2.cc:7
Vectorizing loop at ifconv2.cc:7
7: LOOP VECTORIZED.
ifconv2.cc:4: note: vectorized 1 loops in function.
pb-d-128-141-131-26:bugs48 innocent$ otool -t -X -v ifconv2.o
__Z4bar2v:
leaq 0x00000000(%rip),%rdx
xorps %xmm3,%xmm3
leaq 0x00000000(%rip),%rax
leaq 0x00000190(%rip),%rcx
nopl 0x00000000(%rax,%rax)
movaps (%rdx),%xmm2
movaps %xmm3,%xmm1
addq $0x10,%rdx
movaps %xmm2,%xmm0
cmpps $0x1,%xmm2,%xmm1
cmpps $0x2,%xmm3,%xmm0
andnps (%rax),%xmm1
andps %xmm0,%xmm2
andnps %xmm1,%xmm0
orps %xmm2,%xmm0
movaps %xmm0,(%rax)
addq $0x10,%rax
cmpq %rcx,%rdx
jne 0x00000020
xorps %xmm0,%xmm0
comiss 0x00000000(%rip),%xmm0
je 0x00000063
movl $0x00000000,0xfffffffc(%rip)
repz/ret
