public inbox for gcc-bugs@sourceware.org
help / color / mirror / Atom feed
* [Bug target/102591] New: Failure to optimize search for value in vector-sized area to use SIMD
@ 2021-10-04 12:12 gabravier at gmail dot com
2021-10-05 6:44 ` [Bug target/102591] " rguenth at gcc dot gnu.org
` (2 more replies)
0 siblings, 3 replies; 4+ messages in thread
From: gabravier at gmail dot com @ 2021-10-04 12:12 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102591
Bug ID: 102591
Summary: Failure to optimize search for value in vector-sized
area to use SIMD
Product: gcc
Version: 12.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: gabravier at gmail dot com
Target Milestone: ---
bool match8(char *tpl)
{
int found = 0;
for (int at = 0; at < 16; at++)
if (tpl[at] == 0)
found = 1;
return found;
}
This function can be greatly optimized by using SIMD. It can be optimized to
something like this:
typedef char v16i8 __attribute__((vector_size(16)));
bool match8v2(char *tpl)
{
v16i8 values;
__builtin_memcpy(&values, tpl, 16);
v16i8 compared = (values == 0);
return _mm_movemask_epi8((__m128i)compared) != 0;
}
This optimization is done by LLVM, but not by GCC.
PS: I've marked this as an x86 bug, but only because I could not find a
portable way of expressing `_mm_movemask_epi8((__m128i)compared)`, I would
assume other architectures have similar ways of expressing the same thing
cheaply.
(For example, Altivec should be able to implement that operation with a
`vec_extract(vec_vbpermq((__vector unsigned char)compared, perm), 1)` with
`perm` looking like this: `{120, 112, 104, 96, 88, 80, 72, 64, 56, 48, 40, 32,
24, 16, 8, 0}` and the 1 replaced with 14 on big-endian)
^ permalink raw reply [flat|nested] 4+ messages in thread
* [Bug target/102591] Failure to optimize search for value in vector-sized area to use SIMD
2021-10-04 12:12 [Bug target/102591] New: Failure to optimize search for value in vector-sized area to use SIMD gabravier at gmail dot com
@ 2021-10-05 6:44 ` rguenth at gcc dot gnu.org
2021-10-05 9:46 ` gabravier at gmail dot com
2021-10-05 10:19 ` [Bug tree-optimization/102591] " rguenth at gcc dot gnu.org
2 siblings, 0 replies; 4+ messages in thread
From: rguenth at gcc dot gnu.org @ 2021-10-05 6:44 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102591
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |rguenth at gcc dot gnu.org
Ever confirmed|0 |1
Last reconfirmed| |2021-10-05
Status|UNCONFIRMED |WAITING
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Hmm, but
__builtin_memcpy(&values, tpl, 16);
could trap since 'tpl' is not aligned to 16 bytes? So LLVM creates wrong code
here?
^ permalink raw reply [flat|nested] 4+ messages in thread
* [Bug target/102591] Failure to optimize search for value in vector-sized area to use SIMD
2021-10-04 12:12 [Bug target/102591] New: Failure to optimize search for value in vector-sized area to use SIMD gabravier at gmail dot com
2021-10-05 6:44 ` [Bug target/102591] " rguenth at gcc dot gnu.org
@ 2021-10-05 9:46 ` gabravier at gmail dot com
2021-10-05 10:19 ` [Bug tree-optimization/102591] " rguenth at gcc dot gnu.org
2 siblings, 0 replies; 4+ messages in thread
From: gabravier at gmail dot com @ 2021-10-05 9:46 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102591
--- Comment #2 from Gabriel Ravier <gabravier at gmail dot com> ---
memcpy can fail on unaligned memory ??? I used it specifically to avoid this
problem !
(also, LLVM's code, I am pretty sure, does not have any issue with alignment,
as it uses either AVX instructions which care not for it, or specifically does
a movdqu (i.e. unaligned load) of the memory)
^ permalink raw reply [flat|nested] 4+ messages in thread
* [Bug tree-optimization/102591] Failure to optimize search for value in vector-sized area to use SIMD
2021-10-04 12:12 [Bug target/102591] New: Failure to optimize search for value in vector-sized area to use SIMD gabravier at gmail dot com
2021-10-05 6:44 ` [Bug target/102591] " rguenth at gcc dot gnu.org
2021-10-05 9:46 ` gabravier at gmail dot com
@ 2021-10-05 10:19 ` rguenth at gcc dot gnu.org
2 siblings, 0 replies; 4+ messages in thread
From: rguenth at gcc dot gnu.org @ 2021-10-05 10:19 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102591
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|WAITING |NEW
Component|target |tree-optimization
Blocks| |53947
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Gabriel Ravier from comment #2)
> memcpy can fail on unaligned memory ??? I used it specifically to avoid this
> problem !
>
> (also, LLVM's code, I am pretty sure, does not have any issue with
> alignment, as it uses either AVX instructions which care not for it, or
> specifically does a movdqu (i.e. unaligned load) of the memory)
Ah, sorry - I was reading the loop as
for (int at = 0; at < 16; at++)
if (tpl[at] == 0)
{
found = 1;
break;
}
thus as if the suggested transform would eventually access storage that is
not accessed originally...
Btw, we vectorize
bool match8(char *tpl)
{
char found = 0;
for (int at = 0; at < 16; at++)
if (tpl[at] == 0)
found = 1;
return found;
}
but use
vector(16) char vect_found_4.8;
vect__3.7_29 = MEM <vector(16) char> [(char *)tpl_10(D)];
_32 = vect__3.7_29 != { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
vect_found_4.8_33 = VEC_COND_EXPR <_32, { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0 }, { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 }>;
_35 = .REDUC_MAX (vect_found_4.8_33);
_8 = (bool) _35;
return _8;
where we fail to apply "magic" to the .REDUC_MAX as we know the values
are all 0 or 1.
The conditional reduction support doesn't support producing 'int' from
char compares and we fail to narrow the reduction vector.
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53947
[Bug 53947] [meta-bug] vectorizer missed-optimizations
^ permalink raw reply [flat|nested] 4+ messages in thread
end of thread, other threads:[~2021-10-05 10:19 UTC | newest]
Thread overview: 4+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2021-10-04 12:12 [Bug target/102591] New: Failure to optimize search for value in vector-sized area to use SIMD gabravier at gmail dot com
2021-10-05 6:44 ` [Bug target/102591] " rguenth at gcc dot gnu.org
2021-10-05 9:46 ` gabravier at gmail dot com
2021-10-05 10:19 ` [Bug tree-optimization/102591] " rguenth at gcc dot gnu.org
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).