public inbox for gcc-bugs@sourceware.org
help / color / mirror / Atom feed
* [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
@ 2024-05-13 13:05 colin.king at intel dot com
2024-05-13 13:07 ` [Bug target/115069] " colin.king at intel dot com
` (23 more replies)
0 siblings, 24 replies; 25+ messages in thread
From: colin.king at intel dot com @ 2024-05-13 13:05 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
Bug ID: 115069
Summary: 8 bit integer vector performance regression, x86,
between gcc-14 and gcc-13 using avx2 target clones on
skylake platform
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: colin.king at intel dot com
Target Milestone: ---
Created attachment 58188
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=58188&action=edit
reproducer source code
I'm seeing a ~12-14% performance regression in gcc-14 compared to gcc-13, using
gcc on Ubuntu 24.04:
Versions:
gcc version 13.2.0 (Ubuntu 13.2.0-23ubuntu4)
gcc version 14.0.1 20240412 (experimental) [master r14-9935-g67e1433a94f]
(Ubuntu 14-20240412-0ubuntu1)
cking@skylake:~$ gcc-13 reproducer-vecmath.c -O2
cking@skylake:~$ ./a.out
13540.16 vec8 ops per sec, duration = 14.77 secs
cking@skylake:~$ gcc-14 reproducer-vecmath.c -O2
cking@skylake:~$ ./a.out
11720.25 vec8 ops per sec, duration = 17.06 secs
The original issue appeared when regression testing stress-ng vecmath stressor
[1]. I've managed to extract the attached reproducer from the original code
(see attached).
Salient point to focus on:
1. The issue is also dependant on the TARGET_CLONES macro being defined as
__attribute__((target_clones("mmx,avx,avx2,default"))) - the avx2 target
clones seems to be an issue in reproducing this problem, remove it for gcc-14
and the performance regression is reduced.
Attached are the reproducer C source and disassembled object code.
References: [1]
https://github.com/ColinIanKing/stress-ng/blob/master/stress-vecmath.c
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
@ 2024-05-13 13:07 ` colin.king at intel dot com
2024-05-13 13:07 ` colin.king at intel dot com
` (22 subsequent siblings)
23 siblings, 0 replies; 25+ messages in thread
From: colin.king at intel dot com @ 2024-05-13 13:07 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
--- Comment #1 from Colin Ian King <colin.king at intel dot com> ---
Created attachment 58189
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=58189&action=edit
gcc-13 disassembly
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
2024-05-13 13:07 ` [Bug target/115069] " colin.king at intel dot com
@ 2024-05-13 13:07 ` colin.king at intel dot com
2024-05-15 8:51 ` haochen.jiang at intel dot com
` (21 subsequent siblings)
23 siblings, 0 replies; 25+ messages in thread
From: colin.king at intel dot com @ 2024-05-13 13:07 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
--- Comment #2 from Colin Ian King <colin.king at intel dot com> ---
Created attachment 58190
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=58190&action=edit
gcc-14 disassembly
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
2024-05-13 13:07 ` [Bug target/115069] " colin.king at intel dot com
2024-05-13 13:07 ` colin.king at intel dot com
@ 2024-05-15 8:51 ` haochen.jiang at intel dot com
2024-05-16 1:49 ` [Bug target/115069] [14/15 regression] " sjames at gcc dot gnu.org
` (20 subsequent siblings)
23 siblings, 0 replies; 25+ messages in thread
From: haochen.jiang at intel dot com @ 2024-05-15 8:51 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
Haochen Jiang <haochen.jiang at intel dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |haochen.jiang at intel dot com
--- Comment #3 from Haochen Jiang <haochen.jiang at intel dot com> ---
From my investigation, GCC14 generates some perm instructions, which caused
data dependency.
Going to bisect which commit caused this issue.
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] [14/15 regression] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
` (2 preceding siblings ...)
2024-05-15 8:51 ` haochen.jiang at intel dot com
@ 2024-05-16 1:49 ` sjames at gcc dot gnu.org
2024-05-17 6:59 ` kkanas at fastmail dot com
` (19 subsequent siblings)
23 siblings, 0 replies; 25+ messages in thread
From: sjames at gcc dot gnu.org @ 2024-05-16 1:49 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
Sam James <sjames at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Target Milestone|--- |14.2
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] [14/15 regression] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
` (3 preceding siblings ...)
2024-05-16 1:49 ` [Bug target/115069] [14/15 regression] " sjames at gcc dot gnu.org
@ 2024-05-17 6:59 ` kkanas at fastmail dot com
2024-05-17 7:06 ` liuhongt at gcc dot gnu.org
` (18 subsequent siblings)
23 siblings, 0 replies; 25+ messages in thread
From: kkanas at fastmail dot com @ 2024-05-17 6:59 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
Krzysztof Kanas <kkanas at fastmail dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |kkanas at fastmail dot com
--- Comment #4 from Krzysztof Kanas <kkanas at fastmail dot com> ---
I bisected the issue and it seems that commit
0368fc54bc11f15bfa0ed9913fd0017815dfaa5d introduces regression.
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] [14/15 regression] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
` (4 preceding siblings ...)
2024-05-17 6:59 ` kkanas at fastmail dot com
@ 2024-05-17 7:06 ` liuhongt at gcc dot gnu.org
2024-05-17 7:08 ` haochen.jiang at intel dot com
` (17 subsequent siblings)
23 siblings, 0 replies; 25+ messages in thread
From: liuhongt at gcc dot gnu.org @ 2024-05-17 7:06 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
--- Comment #5 from Hongtao Liu <liuhongt at gcc dot gnu.org> ---
(In reply to Krzysztof Kanas from comment #4)
> I bisected the issue and it seems that commit
> 0368fc54bc11f15bfa0ed9913fd0017815dfaa5d introduces regression.
I guess the real guilty commit is
commit 52ff3f7b863da1011b73c0ab3b11f6c78b6451c7
Author: Uros Bizjak <ubizjak@gmail.com>
Date: Thu May 25 19:40:26 2023 +0200
i386: Use 2x-wider modes when emulating QImode vector instructions
Rewrite ix86_expand_vecop_qihi2 to expand fo 2x-wider (e.g. V16QI ->
V16HImode)
instructions when available. Currently, the compiler generates following
assembly for V16QImode multiplication (-mavx2):
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] [14/15 regression] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
` (5 preceding siblings ...)
2024-05-17 7:06 ` liuhongt at gcc dot gnu.org
@ 2024-05-17 7:08 ` haochen.jiang at intel dot com
2024-05-17 8:26 ` ubizjak at gmail dot com
` (16 subsequent siblings)
23 siblings, 0 replies; 25+ messages in thread
From: haochen.jiang at intel dot com @ 2024-05-17 7:08 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
--- Comment #6 from Haochen Jiang <haochen.jiang at intel dot com> ---
(In reply to Hongtao Liu from comment #5)
> (In reply to Krzysztof Kanas from comment #4)
> > I bisected the issue and it seems that commit
> > 0368fc54bc11f15bfa0ed9913fd0017815dfaa5d introduces regression.
>
> I guess the real guilty commit is
>
> commit 52ff3f7b863da1011b73c0ab3b11f6c78b6451c7
> Author: Uros Bizjak <ubizjak@gmail.com>
> Date: Thu May 25 19:40:26 2023 +0200
>
> i386: Use 2x-wider modes when emulating QImode vector instructions
>
> Rewrite ix86_expand_vecop_qihi2 to expand fo 2x-wider (e.g. V16QI ->
> V16HImode)
> instructions when available. Currently, the compiler generates following
> assembly for V16QImode multiplication (-mavx2):
Yes, since 0368fc54bc11f15bfa0ed9913fd0017815dfaa5d only fixed a typo in that
patch.
Original thread: https://gcc.gnu.org/pipermail/gcc-patches/2023-May/619745.html
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] [14/15 regression] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
` (6 preceding siblings ...)
2024-05-17 7:08 ` haochen.jiang at intel dot com
@ 2024-05-17 8:26 ` ubizjak at gmail dot com
2024-05-17 8:41 ` ubizjak at gmail dot com
` (15 subsequent siblings)
23 siblings, 0 replies; 25+ messages in thread
From: ubizjak at gmail dot com @ 2024-05-17 8:26 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
--- Comment #7 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Hongtao Liu from comment #5)
> (In reply to Krzysztof Kanas from comment #4)
> > I bisected the issue and it seems that commit
> > 0368fc54bc11f15bfa0ed9913fd0017815dfaa5d introduces regression.
>
> I guess the real guilty commit is
>
> commit 52ff3f7b863da1011b73c0ab3b11f6c78b6451c7
> Author: Uros Bizjak <ubizjak@gmail.com>
> Date: Thu May 25 19:40:26 2023 +0200
>
> i386: Use 2x-wider modes when emulating QImode vector instructions
>
> Rewrite ix86_expand_vecop_qihi2 to expand fo 2x-wider (e.g. V16QI ->
> V16HImode)
> instructions when available. Currently, the compiler generates following
> assembly for V16QImode multiplication (-mavx2):
The patch is at:
https://gcc.gnu.org/pipermail/gcc-patches/2023-May/619715.html
As mentioned in Comment #3, it looks that VPERMQ is a problematic insn. This
should be reflected in some cost function. Alternatively, we can simply change
the first line in:
+ if ((qimode == V16QImode && !TARGET_AVX2)
+ || (qimode == V32QImode && !TARGET_AVX512BW)
+ /* There are no V64HImode instructions. */
+ || qimode == V64QImode)
+ return false;
to check "qimode == V16QImode && !TARGET_AVX512VL" to avoid VPERMQ:
diff --git a/gcc/config/i386/i386-expand.cc b/gcc/config/i386/i386-expand.cc
index 4e16aedc5c1..450035ea9e6 100644
--- a/gcc/config/i386/i386-expand.cc
+++ b/gcc/config/i386/i386-expand.cc
@@ -24493,7 +24493,7 @@ ix86_expand_vecop_qihi2 (enum rtx_code code, rtx dest,
rtx op1, rtx op2)
bool op2vec = GET_MODE_CLASS (GET_MODE (op2)) == MODE_VECTOR_INT;
bool uns_p = code != ASHIFTRT;
- if ((qimode == V16QImode && !TARGET_AVX2)
+ if ((qimode == V16QImode && !TARGET_AVX512VL)
|| (qimode == V32QImode && (!TARGET_AVX512BW || !TARGET_EVEX512))
/* There are no V64HImode instructions. */
|| qimode == V64QImode)
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] [14/15 regression] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
` (7 preceding siblings ...)
2024-05-17 8:26 ` ubizjak at gmail dot com
@ 2024-05-17 8:41 ` ubizjak at gmail dot com
2024-05-17 8:48 ` ubizjak at gmail dot com
` (14 subsequent siblings)
23 siblings, 0 replies; 25+ messages in thread
From: ubizjak at gmail dot com @ 2024-05-17 8:41 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
Uroš Bizjak <ubizjak at gmail dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Ever confirmed|0 |1
Last reconfirmed| |2024-05-17
--- Comment #8 from Uroš Bizjak <ubizjak at gmail dot com> ---
A better patch:
--cut here--
diff --git a/gcc/config/i386/i386-expand.cc b/gcc/config/i386/i386-expand.cc
index 4e16aedc5c1..88bfc43201b 100644
--- a/gcc/config/i386/i386-expand.cc
+++ b/gcc/config/i386/i386-expand.cc
@@ -24493,6 +24493,10 @@ ix86_expand_vecop_qihi2 (enum rtx_code code, rtx dest,
rtx op1, rtx op2)
bool op2vec = GET_MODE_CLASS (GET_MODE (op2)) == MODE_VECTOR_INT;
bool uns_p = code != ASHIFTRT;
+ /* ??? VPERMQ is slow and VPMOWVB is only available under AVX512BW. */
+ if (!TARGET_AVX512BW)
+ return false;
+
if ((qimode == V16QImode && !TARGET_AVX2)
|| (qimode == V32QImode && (!TARGET_AVX512BW || !TARGET_EVEX512))
/* There are no V64HImode instructions. */
--cut here--
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] [14/15 regression] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
` (8 preceding siblings ...)
2024-05-17 8:41 ` ubizjak at gmail dot com
@ 2024-05-17 8:48 ` ubizjak at gmail dot com
2024-05-17 9:13 ` haochen.jiang at intel dot com
` (13 subsequent siblings)
23 siblings, 0 replies; 25+ messages in thread
From: ubizjak at gmail dot com @ 2024-05-17 8:48 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
--- Comment #9 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Uroš Bizjak from comment #8)
> A better patch:
The real issue is that the following permutation (truncation):
+ for (i = 0; i < d.nelt; ++i)
+ d.perm[i] = i * 2;
+
+ ok = ix86_expand_vec_perm_const_1 (&d);
results in a slow code involving VPERMQ. Ideally, ix86_expand_vec_perm_const_1
should emit faster code for truncation, because this will benefit other code as
well.
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] [14/15 regression] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
` (9 preceding siblings ...)
2024-05-17 8:48 ` ubizjak at gmail dot com
@ 2024-05-17 9:13 ` haochen.jiang at intel dot com
2024-05-20 1:23 ` liuhongt at gcc dot gnu.org
` (12 subsequent siblings)
23 siblings, 0 replies; 25+ messages in thread
From: haochen.jiang at intel dot com @ 2024-05-17 9:13 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
--- Comment #10 from Haochen Jiang <haochen.jiang at intel dot com> ---
A patch like Comment 8 could definitely solve the problem. But I need to test
more benchmarks to see if there is surprise.
But, yes, as Uros said in Comment 9, maybe there is a chance we could do it
better.
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] [14/15 regression] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
` (10 preceding siblings ...)
2024-05-17 9:13 ` haochen.jiang at intel dot com
@ 2024-05-20 1:23 ` liuhongt at gcc dot gnu.org
2024-05-20 5:50 ` haochen.jiang at intel dot com
` (11 subsequent siblings)
23 siblings, 0 replies; 25+ messages in thread
From: liuhongt at gcc dot gnu.org @ 2024-05-20 1:23 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
--- Comment #11 from Hongtao Liu <liuhongt at gcc dot gnu.org> ---
(In reply to Haochen Jiang from comment #10)
> A patch like Comment 8 could definitely solve the problem. But I need to
> test more benchmarks to see if there is surprise.
>
> But, yes, as Uros said in Comment 9, maybe there is a chance we could do it
> better.
Could you add "arch=skylake-avx512" to target_clones and try disable whole
ix86_expand_vecop_qihi2 to see if there's any performance improvement?
For x86, cross-lane permutation(truncation) is not very efficient(3-4 cycles
for both vpermq and vpmovwb).
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] [14/15 regression] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
` (11 preceding siblings ...)
2024-05-20 1:23 ` liuhongt at gcc dot gnu.org
@ 2024-05-20 5:50 ` haochen.jiang at intel dot com
2024-05-20 6:27 ` ubizjak at gmail dot com
` (10 subsequent siblings)
23 siblings, 0 replies; 25+ messages in thread
From: haochen.jiang at intel dot com @ 2024-05-20 5:50 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
--- Comment #12 from Haochen Jiang <haochen.jiang at intel dot com> ---
(In reply to Hongtao Liu from comment #11)
> (In reply to Haochen Jiang from comment #10)
> > A patch like Comment 8 could definitely solve the problem. But I need to
> > test more benchmarks to see if there is surprise.
> >
> > But, yes, as Uros said in Comment 9, maybe there is a chance we could do it
> > better.
>
> Could you add "arch=skylake-avx512" to target_clones and try disable whole
> ix86_expand_vecop_qihi2 to see if there's any performance improvement?
> For x86, cross-lane permutation(truncation) is not very efficient(3-4 cycles
> for both vpermq and vpmovwb).
When I disable/enable ix86_expand_vecop_qihi2 with arch=skylake-avx512 on
trunk, there is no performance regression comparing to GCC13 + avx2.
It seems that the regression only happens when GCC14 + avx2.
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] [14/15 regression] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
` (12 preceding siblings ...)
2024-05-20 5:50 ` haochen.jiang at intel dot com
@ 2024-05-20 6:27 ` ubizjak at gmail dot com
2024-05-20 6:35 ` liuhongt at gcc dot gnu.org
` (9 subsequent siblings)
23 siblings, 0 replies; 25+ messages in thread
From: ubizjak at gmail dot com @ 2024-05-20 6:27 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
--- Comment #13 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Haochen Jiang from comment #12)
> (In reply to Hongtao Liu from comment #11)
> > (In reply to Haochen Jiang from comment #10)
> > > A patch like Comment 8 could definitely solve the problem. But I need to
> > > test more benchmarks to see if there is surprise.
> > >
> > > But, yes, as Uros said in Comment 9, maybe there is a chance we could do it
> > > better.
> >
> > Could you add "arch=skylake-avx512" to target_clones and try disable whole
> > ix86_expand_vecop_qihi2 to see if there's any performance improvement?
> > For x86, cross-lane permutation(truncation) is not very efficient(3-4 cycles
> > for both vpermq and vpmovwb).
>
> When I disable/enable ix86_expand_vecop_qihi2 with arch=skylake-avx512 on
> trunk, there is no performance regression comparing to GCC13 + avx2.
>
> It seems that the regression only happens when GCC14 + avx2.
This is what the patch in Comment #8 prevents. skylake-avx512 enables
TARGET_AVX512BW, so VPMOVB is emitted instead of problematic VPERMQ.
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] [14/15 regression] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
` (13 preceding siblings ...)
2024-05-20 6:27 ` ubizjak at gmail dot com
@ 2024-05-20 6:35 ` liuhongt at gcc dot gnu.org
2024-05-20 6:38 ` haochen.jiang at intel dot com
` (8 subsequent siblings)
23 siblings, 0 replies; 25+ messages in thread
From: liuhongt at gcc dot gnu.org @ 2024-05-20 6:35 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
--- Comment #14 from Hongtao Liu <liuhongt at gcc dot gnu.org> ---
(In reply to Uroš Bizjak from comment #13)
> (In reply to Haochen Jiang from comment #12)
> > (In reply to Hongtao Liu from comment #11)
> > > (In reply to Haochen Jiang from comment #10)
> > > > A patch like Comment 8 could definitely solve the problem. But I need to
> > > > test more benchmarks to see if there is surprise.
> > > >
> > > > But, yes, as Uros said in Comment 9, maybe there is a chance we could do it
> > > > better.
> > >
> > > Could you add "arch=skylake-avx512" to target_clones and try disable whole
> > > ix86_expand_vecop_qihi2 to see if there's any performance improvement?
> > > For x86, cross-lane permutation(truncation) is not very efficient(3-4 cycles
> > > for both vpermq and vpmovwb).
> >
> > When I disable/enable ix86_expand_vecop_qihi2 with arch=skylake-avx512 on
> > trunk, there is no performance regression comparing to GCC13 + avx2.
> >
> > It seems that the regression only happens when GCC14 + avx2.
>
> This is what the patch in Comment #8 prevents. skylake-avx512 enables
> TARGET_AVX512BW, so VPMOVB is emitted instead of problematic VPERMQ.
Yes, the patch looks good to me.
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] [14/15 regression] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
` (14 preceding siblings ...)
2024-05-20 6:35 ` liuhongt at gcc dot gnu.org
@ 2024-05-20 6:38 ` haochen.jiang at intel dot com
2024-05-20 6:40 ` liuhongt at gcc dot gnu.org
` (7 subsequent siblings)
23 siblings, 0 replies; 25+ messages in thread
From: haochen.jiang at intel dot com @ 2024-05-20 6:38 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
--- Comment #15 from Haochen Jiang <haochen.jiang at intel dot com> ---
I am doing like this way. Suppose should be same as Comment 8.
diff --git a/gcc/config/i386/i386-expand.cc b/gcc/config/i386/i386-expand.cc
index a6132911e6a..1e8334877d6 100644
--- a/gcc/config/i386/i386-expand.cc
+++ b/gcc/config/i386/i386-expand.cc
@@ -24323,8 +24323,8 @@ ix86_expand_vecop_qihi2 (enum rtx_code code, rtx dest,
rtx op1, rtx op2)
bool op2vec = GET_MODE_CLASS (GET_MODE (op2)) == MODE_VECTOR_INT;
bool uns_p = code != ASHIFTRT;
- if ((qimode == V16QImode && !TARGET_AVX2)
- || (qimode == V32QImode && (!TARGET_AVX512BW || !TARGET_EVEX512))
+ if (!TARGET_AVX512BW
+ || (qimode == V32QImode && !TARGET_EVEX512)
/* There are no V64HImode instructions. */
|| qimode == V64QImode)
Should we also run a SPEC on with -O2 -mtune=generic -march=x86-64-v3 to see if
there is any surprise?
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] [14/15 regression] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
` (15 preceding siblings ...)
2024-05-20 6:38 ` haochen.jiang at intel dot com
@ 2024-05-20 6:40 ` liuhongt at gcc dot gnu.org
2024-05-20 6:47 ` ubizjak at gmail dot com
` (6 subsequent siblings)
23 siblings, 0 replies; 25+ messages in thread
From: liuhongt at gcc dot gnu.org @ 2024-05-20 6:40 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
--- Comment #16 from Hongtao Liu <liuhongt at gcc dot gnu.org> ---
> Should we also run a SPEC on with -O2 -mtune=generic -march=x86-64-v3 to see
> if there is any surprise?
Sure, I guess no.
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] [14/15 regression] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
` (16 preceding siblings ...)
2024-05-20 6:40 ` liuhongt at gcc dot gnu.org
@ 2024-05-20 6:47 ` ubizjak at gmail dot com
2024-05-21 5:56 ` haochen.jiang at intel dot com
` (5 subsequent siblings)
23 siblings, 0 replies; 25+ messages in thread
From: ubizjak at gmail dot com @ 2024-05-20 6:47 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
--- Comment #17 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Haochen Jiang from comment #15)
> I am doing like this way. Suppose should be same as Comment 8.
Yes, but IMO the patch in Comment #8 better describes where the problem is.
Please note that without VPMOVWB we fall-back to the original
ix86_expand_vecop_qihi, where the expansion is implemented in a different way.
(BTW: If there is a better way to emulate VPMOVWB, it should be implemented in
vec-perm routines, it will universally benefit this permutation. In this case,
early exit, as introduced in the mentioned patch, could be removed.)
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] [14/15 regression] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
` (17 preceding siblings ...)
2024-05-20 6:47 ` ubizjak at gmail dot com
@ 2024-05-21 5:56 ` haochen.jiang at intel dot com
2024-05-21 5:58 ` haochen.jiang at intel dot com
` (4 subsequent siblings)
23 siblings, 0 replies; 25+ messages in thread
From: haochen.jiang at intel dot com @ 2024-05-21 5:56 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
--- Comment #18 from Haochen Jiang <haochen.jiang at intel dot com> ---
SPEC
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] [14/15 regression] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
` (18 preceding siblings ...)
2024-05-21 5:56 ` haochen.jiang at intel dot com
@ 2024-05-21 5:58 ` haochen.jiang at intel dot com
2024-05-22 2:25 ` cvs-commit at gcc dot gnu.org
` (3 subsequent siblings)
23 siblings, 0 replies; 25+ messages in thread
From: haochen.jiang at intel dot com @ 2024-05-21 5:58 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
--- Comment #19 from Haochen Jiang <haochen.jiang at intel dot com> ---
(In reply to Haochen Jiang from comment #18)
> SPEC
SPEC seems all same binary to me. So there is no surprise.
I suppose let's go with patch from Uros to just emphasize the problem.
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] [14/15 regression] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
` (19 preceding siblings ...)
2024-05-21 5:58 ` haochen.jiang at intel dot com
@ 2024-05-22 2:25 ` cvs-commit at gcc dot gnu.org
2024-05-22 2:27 ` cvs-commit at gcc dot gnu.org
` (2 subsequent siblings)
23 siblings, 0 replies; 25+ messages in thread
From: cvs-commit at gcc dot gnu.org @ 2024-05-22 2:25 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
--- Comment #20 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Haochen Jiang <jianghc@gcc.gnu.org>:
https://gcc.gnu.org/g:73a167cfa225d5ee7092d41596b9fea1719898ff
commit r15-764-g73a167cfa225d5ee7092d41596b9fea1719898ff
Author: Haochen Jiang <haochen.jiang@intel.com>
Date: Tue May 21 14:10:43 2024 +0800
i386: Disable ix86_expand_vecop_qihi2 when !TARGET_AVX512BW
Since vpermq is really slow, we should avoid using it for permutation
when vpmovwb is not available (needs AVX512BW) for ix86_expand_vecop_qihi2
and fall back to ix86_expand_vecop_qihi.
gcc/ChangeLog:
PR target/115069
* config/i386/i386-expand.cc (ix86_expand_vecop_qihi2):
Do not enable the optimization when AVX512BW is not enabled.
gcc/testsuite/ChangeLog:
PR target/115069
* gcc.target/i386/pr115069.c: New.
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] [14/15 regression] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
` (20 preceding siblings ...)
2024-05-22 2:25 ` cvs-commit at gcc dot gnu.org
@ 2024-05-22 2:27 ` cvs-commit at gcc dot gnu.org
2024-05-22 2:27 ` haochen.jiang at intel dot com
2024-05-22 5:13 ` liuhongt at gcc dot gnu.org
23 siblings, 0 replies; 25+ messages in thread
From: cvs-commit at gcc dot gnu.org @ 2024-05-22 2:27 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
--- Comment #21 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-14 branch has been updated by Haochen Jiang
<jianghc@gcc.gnu.org>:
https://gcc.gnu.org/g:1ad5c9d524d8fa99773045e75da04ae958012085
commit r14-10229-g1ad5c9d524d8fa99773045e75da04ae958012085
Author: Haochen Jiang <haochen.jiang@intel.com>
Date: Tue May 21 14:10:43 2024 +0800
i386: Disable ix86_expand_vecop_qihi2 when !TARGET_AVX512BW
Since vpermq is really slow, we should avoid using it for permutation
when vpmovwb is not available (needs AVX512BW) for ix86_expand_vecop_qihi2
and fall back to ix86_expand_vecop_qihi.
gcc/ChangeLog:
PR target/115069
* config/i386/i386-expand.cc (ix86_expand_vecop_qihi2):
Do not enable the optimization when AVX512BW is not enabled.
gcc/testsuite/ChangeLog:
PR target/115069
* gcc.target/i386/pr115069.c: New.
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] [14/15 regression] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
` (21 preceding siblings ...)
2024-05-22 2:27 ` cvs-commit at gcc dot gnu.org
@ 2024-05-22 2:27 ` haochen.jiang at intel dot com
2024-05-22 5:13 ` liuhongt at gcc dot gnu.org
23 siblings, 0 replies; 25+ messages in thread
From: haochen.jiang at intel dot com @ 2024-05-22 2:27 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
--- Comment #22 from Haochen Jiang <haochen.jiang at intel dot com> ---
Fixed in GCC14 and GCC15
^ permalink raw reply [flat|nested] 25+ messages in thread
* [Bug target/115069] [14/15 regression] 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
` (22 preceding siblings ...)
2024-05-22 2:27 ` haochen.jiang at intel dot com
@ 2024-05-22 5:13 ` liuhongt at gcc dot gnu.org
23 siblings, 0 replies; 25+ messages in thread
From: liuhongt at gcc dot gnu.org @ 2024-05-22 5:13 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115069
Hongtao Liu <liuhongt at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|NEW |RESOLVED
Resolution|--- |FIXED
--- Comment #23 from Hongtao Liu <liuhongt at gcc dot gnu.org> ---
Fixed in GCC15 and GCC14.2
^ permalink raw reply [flat|nested] 25+ messages in thread
end of thread, other threads:[~2024-05-22 5:13 UTC | newest]
Thread overview: 25+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2024-05-13 13:05 [Bug target/115069] New: 8 bit integer vector performance regression, x86, between gcc-14 and gcc-13 using avx2 target clones on skylake platform colin.king at intel dot com
2024-05-13 13:07 ` [Bug target/115069] " colin.king at intel dot com
2024-05-13 13:07 ` colin.king at intel dot com
2024-05-15 8:51 ` haochen.jiang at intel dot com
2024-05-16 1:49 ` [Bug target/115069] [14/15 regression] " sjames at gcc dot gnu.org
2024-05-17 6:59 ` kkanas at fastmail dot com
2024-05-17 7:06 ` liuhongt at gcc dot gnu.org
2024-05-17 7:08 ` haochen.jiang at intel dot com
2024-05-17 8:26 ` ubizjak at gmail dot com
2024-05-17 8:41 ` ubizjak at gmail dot com
2024-05-17 8:48 ` ubizjak at gmail dot com
2024-05-17 9:13 ` haochen.jiang at intel dot com
2024-05-20 1:23 ` liuhongt at gcc dot gnu.org
2024-05-20 5:50 ` haochen.jiang at intel dot com
2024-05-20 6:27 ` ubizjak at gmail dot com
2024-05-20 6:35 ` liuhongt at gcc dot gnu.org
2024-05-20 6:38 ` haochen.jiang at intel dot com
2024-05-20 6:40 ` liuhongt at gcc dot gnu.org
2024-05-20 6:47 ` ubizjak at gmail dot com
2024-05-21 5:56 ` haochen.jiang at intel dot com
2024-05-21 5:58 ` haochen.jiang at intel dot com
2024-05-22 2:25 ` cvs-commit at gcc dot gnu.org
2024-05-22 2:27 ` cvs-commit at gcc dot gnu.org
2024-05-22 2:27 ` haochen.jiang at intel dot com
2024-05-22 5:13 ` liuhongt 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).