http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56935
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|ASSIGNED |WAITING
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> 2013-04-15
14:38:40 UTC ---
Reduced testcase:
typedef struct {
long int x;
long int y;
} S;
void foo (S *s)
{
s->x--;
s->y--;
}
Difference in cost model analysis:
before:
t.c:7: note: Cost model analysis:
Vector inside of basic block cost: 5
Vector prologue cost: 0
Vector epilogue cost: 0
Scalar cost of basic block: 6
after:
t.c:7: note: Cost model analysis:
Vector inside of basic block cost: 5
Vector prologue cost: 1
Vector epilogue cost: 0
Scalar cost of basic block: 6
after is more correct, as we need to synthesize the { 1, 1 } vector.
what isn't really optimal is the unchanged vector inside cost.
It's an unaligned load with cost 2, the vector operation with cost 1
and the unaligned store with cost 2.
Before we generated
pcmpeqd %xmm0, %xmm0
movdqu (%rdi), %xmm1
paddq %xmm1, %xmm0
movdqu %xmm0, (%rdi)
ret
and afterwards
subq $1, (%rdi)
subq $1, 8(%rdi)
I'd say it's obvious that the non-vectorized variant is better.
So, are you sure _this_ basic-block is really the issue?
