https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115693
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Blocks| |53947
Last reconfirmed| |2024-06-28
Target| |x86_64-*-*
Ever confirmed|0 |1
CC| |crazylht at gmail dot com
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Xi Ruoyao from comment #1)
> I'm transferring it to tree-optimization as the following cases are compiled
> to stupid code:
>
> char a[8], b[8];
>
> int test()
> {
> for (int i = 0; i < 8; i++)
> if (a[i] != b[i])
> return 0;
>
> return 1;
> }
>
> int test1()
> {
> int ret = 0;
> for (int i = 0; i < 8; i++)
> ret = ret || a[i] != b[i];
>
> return ret;
> }
>
> So it makes more sense to fix this in the optimization passes, instead of
> ad-hoc hack in libstdc++.
>
> But I'm not sure if there already exists a dup.
Let's keep this bug for the above testcase(s). For test() the issue is
that even with SSE4.1 we don't seem to support ptest for V8QImode?
For test1 cost modeling makes vectorization worthwhile, though with just SSE2
we get
test1:
.LFB1:
.cfi_startproc
movq a(%rip), %xmm1
pxor %xmm2, %xmm2
movq b(%rip), %xmm0
pcmpeqb %xmm1, %xmm0
movq .LC0(%rip), %xmm1
pandn %xmm1, %xmm0
movdqa %xmm0, %xmm1
punpcklbw %xmm2, %xmm0
punpcklbw %xmm2, %xmm1
pshufd $78, %xmm0, %xmm0
pxor %xmm2, %xmm2
movdqa %xmm0, %xmm3
punpcklwd %xmm2, %xmm0
punpcklwd %xmm2, %xmm3
pshufd $78, %xmm0, %xmm0
por %xmm3, %xmm0
movdqa %xmm1, %xmm3
punpcklwd %xmm2, %xmm1
punpcklwd %xmm2, %xmm3
pshufd $78, %xmm1, %xmm1
por %xmm3, %xmm1
por %xmm1, %xmm0
movdqa %xmm0, %xmm1
psrlq $32, %xmm1
por %xmm1, %xmm0
movd %xmm0, %eax
ret
with SSE4.2 it's a bit better and "just"
test1:
.LFB1:
.cfi_startproc
movq a(%rip), %xmm1
movq b(%rip), %xmm0
pcmpeqb %xmm1, %xmm0
movq .LC0(%rip), %xmm1
pandn %xmm1, %xmm0
pmovzxbw %xmm0, %xmm2
psrlq $32, %xmm0
pmovzxbw %xmm0, %xmm0
pmovzxwd %xmm0, %xmm1
psrlq $32, %xmm0
pmovzxwd %xmm0, %xmm0
por %xmm1, %xmm0
pmovzxwd %xmm2, %xmm1
psrlq $32, %xmm2
pmovzxwd %xmm2, %xmm2
por %xmm2, %xmm1
por %xmm1, %xmm0
movdqa %xmm0, %xmm1
psrlq $32, %xmm1
por %xmm1, %xmm0
movd %xmm0, %eax
ret
but we fail to realize that the bitwise-OR reduction could be narrowed to
char:
<bb 3> [local count: 954449106]:
# ret_12 = PHI <iftmp.0_5(7), 0(15)>
# i_14 = PHI <i_7(7), 0(15)>
# ivtmp_4 = PHI <ivtmp_3(7), 8(15)>
_1 = a[i_14];
_2 = b[i_14];
_8 = _1 != _2;
_9 = (int) _8;
iftmp.0_5 = _9 | ret_12;
i_7 = i_14 + 1;
ivtmp_3 = ivtmp_4 - 1;
if (ivtmp_3 != 0)
goto <bb 7>; [87.50%]
instead we keep 4 V2SImode "accumulators" and widen the compare results.
The best would be if scalar opts would make this a bool reduction though
IIRC we have a PR for that being not handled.
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53947
[Bug 53947] [meta-bug] vectorizer missed-optimizations
