https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71768
Bug ID: 71768
Summary: Missed trivial rematerialiation oppurtunity
Product: gcc
Version: 7.0
Status: UNCONFIRMED
Severity: enhancement
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: hubicka at gcc dot gnu.org
Target Milestone: ---
Compiling
#define vector __attribute__ ((vector_size (16)))
const vector int cst={10,10,10,10};
int t()
{
vector int val = cst;
asm("#%0"::"x"(val));
e();
asm("#%0"::"x"(val));
}
Results in:
t:
.LFB0:
.cfi_startproc
subq $24, %rsp
.cfi_def_cfa_offset 32
vmovaps .LC0(%rip), %xmm0
#APP
# 6 "t.c" 1
#%xmm0
# 0 "" 2
#NO_APP
xorl %eax, %eax
vmovaps %xmm0, (%rsp)
call e
vmovaps (%rsp), %xmm0
#APP
# 8 "t.c" 1
#%xmm0
# 0 "" 2
#NO_APP
addq $24, %rsp
.cfi_def_cfa_offset 8
ret
Which is clearly suboptimal, because xmm0 can be rematerialized again from LC0.
This hits pretty badly the exchange2 benchmark with -O3 -march=bdver2 where the
function is self recursive and we end up having many constants cached in XMM
registers.
