[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498 Jeffrey A. Law changed: What|Removed |Added CC||law at redhat dot com Priority|P3 |P2
[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498 Richard Biener changed: What|Removed |Added Status|NEW |WAITING Blocks||26163 --- Comment #20 from Richard Biener --- The patch from comment#4 got installed so re-analysis is necessary I think. Referenced Bugs: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=26163 [Bug 26163] [meta-bug] missed optimization in SPEC (2k17, 2k and 2k6 and 95)
[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498 --- Comment #19 from Richard Biener --- Doesn't seem fixed.
[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498 --- Comment #18 from Richard Biener --- (In reply to Martin Liška from comment #17) > Can the issue be closed as resolved? I've fixed a -fno-common issue on the tester and will report back after the next run.
[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498 Martin Liška changed: What|Removed |Added Status|UNCONFIRMED |NEW Last reconfirmed||2020-01-27 CC||marxin at gcc dot gnu.org Ever confirmed|0 |1 --- Comment #17 from Martin Liška --- Can the issue be closed as resolved?
[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498 --- Comment #16 from ro at CeBiTec dot Uni-Bielefeld.DE --- > --- Comment #13 from Uroš Bizjak --- > (In reply to rguent...@suse.de from comment #10) >> > +FAIL: gcc.target/i386/minmax-7.c scan-assembler pminsd >> >> Not for me. I've used -march=haswell in all of these testcases to >> rule out ISA and cost issues. So I really don't know what's the issue >> with these FAILs. > > IIRC 32bit Solaris doesn't guarantee 16byte stack alignment, so STV is > disabled. Indeed: adding -mno-stackrealign (as is done in a couple of other STV tests already) fixes this failure and the ones reported in PR rtl-optimization/91154. Tested with the appropriate runtest invocation on i386-pc-solaris2.11 and x86_64-pc-linux-gnu (both multilibs each).
[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498 --- Comment #15 from Rainer Orth --- Created attachment 46737 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=46737&action=edit Patch for 32-bit Solaris/x86 failures
[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498 --- Comment #14 from Richard Biener --- Author: rguenth Date: Wed Aug 21 08:44:59 2019 New Revision: 274792 URL: https://gcc.gnu.org/viewcvs?rev=274792&root=gcc&view=rev Log: 2019-08-21 Richard Biener PR target/91498 PR target/91503 * config/i386/i386-features.c (general_scalar_chain::make_vector_copies): Copy stack temporary rtx when using it multiple times. (general_scalar_chain::convert_reg): Likewise. Modified: trunk/gcc/ChangeLog trunk/gcc/config/i386/i386-features.c
[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498 --- Comment #13 from Uroš Bizjak --- (In reply to rguent...@suse.de from comment #10) > > +FAIL: gcc.target/i386/minmax-7.c scan-assembler pminsd > > Not for me. I've used -march=haswell in all of these testcases to > rule out ISA and cost issues. So I really don't know what's the issue > with these FAILs. IIRC 32bit Solaris doesn't guarantee 16byte stack alignment, so STV is disabled.
[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498
Dmitry G. Dyachenko changed:
What|Removed |Added
CC||dimhen at gmail dot com
--- Comment #12 from Dmitry G. Dyachenko ---
(In reply to Arseny Solokha from comment #9)
> (In reply to H.J. Lu from comment #7)
> > Our SPEC CPU 2017 failed with
> >
> > 39 util.c:205:1: error: invalid rtl sharing found in the insn
> > 40 205 | }
> > 41 | ^
> > 42 (insn 29 28 3 2 (set (subreg:V2DI (reg:DI 91) 0)
> > 43 (vec_concat:V2DI (mem/c:DI (plus:DI (reg/f:DI 19 frame)
> > 44 (const_int -8 [0xfff8])) [0 S8 A64])
> > 45 (const_int 0 [0]))) "util.c":134:1 -1
> > 46 (nil))
> > 47 util.c:205:1: error: shared rtx
> > 48 (mem/c:DI (plus:DI (reg/f:DI 19 frame)
> > 49 (const_int -8 [0xfff8])) [0 S8 A64])
> > 50 during RTL pass: stv
> >
> > We have an invalid shared rtx.
>
> And the testcase is as simple as:
>
> int
> ff (int jn)
> {
> return jn > 0 ? jn : 0;
> }
>
> % x86_64-unknown-linux-gnu-gcc-10.0.0-alpha20190818 -march=nano-3000 -Os -c
> f9272pqv.c
> f9272pqv.c: In function 'ff':
> f9272pqv.c:5:1: error: invalid rtl sharing found in the insn
> 5 | }
> | ^
> (insn 18 17 3 2 (set (subreg:V4SI (reg:SI 87) 0)
> (vec_merge:V4SI (vec_duplicate:V4SI (mem/c:SI (plus:DI (reg/f:DI 19
> frame)
> (const_int -4 [0xfffc])) [0 S4 A32]))
> (const_vector:V4SI [
> (const_int 0 [0]) repeated x4
> ])
> (const_int 1 [0x1]))) "f9272pqv.c":3:1 -1
> (nil))
> f9272pqv.c:5:1: error: shared rtx
> (mem/c:SI (plus:DI (reg/f:DI 19 frame)
> (const_int -4 [0xfffc])) [0 S4 A32])
> during RTL pass: stv
> f9272pqv.c:5:1: internal compiler error: internal consistency failure
> 0x9f3eca verify_rtx_sharing
>
> /var/tmp/portage/sys-devel/gcc-10.0.0_alpha20190818/work/gcc-10-20190818/
> gcc/emit-rtl.c:2927
> 0x9f3dba verify_rtx_sharing
>
> /var/tmp/portage/sys-devel/gcc-10.0.0_alpha20190818/work/gcc-10-20190818/
> gcc/emit-rtl.c:2942
> 0x9f3dba verify_rtx_sharing
>
> /var/tmp/portage/sys-devel/gcc-10.0.0_alpha20190818/work/gcc-10-20190818/
> gcc/emit-rtl.c:2942
> 0x9f3dba verify_rtx_sharing
>
> /var/tmp/portage/sys-devel/gcc-10.0.0_alpha20190818/work/gcc-10-20190818/
> gcc/emit-rtl.c:2942
> 0x9f4178 verify_insn_sharing
>
> /var/tmp/portage/sys-devel/gcc-10.0.0_alpha20190818/work/gcc-10-20190818/
> gcc/emit-rtl.c:3013
> 0x9f79ed verify_rtl_sharing()
>
> /var/tmp/portage/sys-devel/gcc-10.0.0_alpha20190818/work/gcc-10-20190818/
> gcc/emit-rtl.c:3036
> 0xc7a8f3 execute_function_todo
>
> /var/tmp/portage/sys-devel/gcc-10.0.0_alpha20190818/work/gcc-10-20190818/
> gcc/passes.c:2004
> 0xc7b536 execute_todo
>
> /var/tmp/portage/sys-devel/gcc-10.0.0_alpha20190818/work/gcc-10-20190818/
> gcc/passes.c:2037
sounds like PR91503
[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498
--- Comment #11 from rguenther at suse dot de ---
On Tue, 20 Aug 2019, asolokha at gmx dot com wrote:
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498
>
> Arseny Solokha changed:
>
>What|Removed |Added
>
> CC||asolokha at gmx dot com
>
> --- Comment #9 from Arseny Solokha ---
> (In reply to H.J. Lu from comment #7)
> > Our SPEC CPU 2017 failed with
> >
> > 39 util.c:205:1: error: invalid rtl sharing found in the insn
> > 40 205 | }
> > 41 | ^
> > 42 (insn 29 28 3 2 (set (subreg:V2DI (reg:DI 91) 0)
> > 43 (vec_concat:V2DI (mem/c:DI (plus:DI (reg/f:DI 19 frame)
> > 44 (const_int -8 [0xfff8])) [0 S8 A64])
> > 45 (const_int 0 [0]))) "util.c":134:1 -1
> > 46 (nil))
> > 47 util.c:205:1: error: shared rtx
> > 48 (mem/c:DI (plus:DI (reg/f:DI 19 frame)
> > 49 (const_int -8 [0xfff8])) [0 S8 A64])
> > 50 during RTL pass: stv
> >
> > We have an invalid shared rtx.
>
> And the testcase is as simple as:
>
> int
> ff (int jn)
> {
> return jn > 0 ? jn : 0;
> }
>
> % x86_64-unknown-linux-gnu-gcc-10.0.0-alpha20190818 -march=nano-3000 -Os -c
> f9272pqv.c
> f9272pqv.c: In function 'ff':
> f9272pqv.c:5:1: error: invalid rtl sharing found in the insn
> 5 | }
> | ^
> (insn 18 17 3 2 (set (subreg:V4SI (reg:SI 87) 0)
> (vec_merge:V4SI (vec_duplicate:V4SI (mem/c:SI (plus:DI (reg/f:DI 19
> frame)
> (const_int -4 [0xfffc])) [0 S4 A32]))
> (const_vector:V4SI [
> (const_int 0 [0]) repeated x4
> ])
> (const_int 1 [0x1]))) "f9272pqv.c":3:1 -1
> (nil))
I have a patch.
[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498 --- Comment #10 from rguenther at suse dot de --- On Tue, 20 Aug 2019, ro at gcc dot gnu.org wrote: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498 > > Rainer Orth changed: > >What|Removed |Added > > CC||ro at gcc dot gnu.org > > --- Comment #8 from Rainer Orth --- > The new testcase FAILs on both i386-pc-solaris2.11 and x86_64-pc-linux-gnu > with > -m32: > > +FAIL: gcc.target/i386/minmax-7.c scan-assembler pminsd Not for me. I've used -march=haswell in all of these testcases to rule out ISA and cost issues. So I really don't know what's the issue with these FAILs.
[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498
Arseny Solokha changed:
What|Removed |Added
CC||asolokha at gmx dot com
--- Comment #9 from Arseny Solokha ---
(In reply to H.J. Lu from comment #7)
> Our SPEC CPU 2017 failed with
>
> 39 util.c:205:1: error: invalid rtl sharing found in the insn
> 40 205 | }
> 41 | ^
> 42 (insn 29 28 3 2 (set (subreg:V2DI (reg:DI 91) 0)
> 43 (vec_concat:V2DI (mem/c:DI (plus:DI (reg/f:DI 19 frame)
> 44 (const_int -8 [0xfff8])) [0 S8 A64])
> 45 (const_int 0 [0]))) "util.c":134:1 -1
> 46 (nil))
> 47 util.c:205:1: error: shared rtx
> 48 (mem/c:DI (plus:DI (reg/f:DI 19 frame)
> 49 (const_int -8 [0xfff8])) [0 S8 A64])
> 50 during RTL pass: stv
>
> We have an invalid shared rtx.
And the testcase is as simple as:
int
ff (int jn)
{
return jn > 0 ? jn : 0;
}
% x86_64-unknown-linux-gnu-gcc-10.0.0-alpha20190818 -march=nano-3000 -Os -c
f9272pqv.c
f9272pqv.c: In function 'ff':
f9272pqv.c:5:1: error: invalid rtl sharing found in the insn
5 | }
| ^
(insn 18 17 3 2 (set (subreg:V4SI (reg:SI 87) 0)
(vec_merge:V4SI (vec_duplicate:V4SI (mem/c:SI (plus:DI (reg/f:DI 19
frame)
(const_int -4 [0xfffc])) [0 S4 A32]))
(const_vector:V4SI [
(const_int 0 [0]) repeated x4
])
(const_int 1 [0x1]))) "f9272pqv.c":3:1 -1
(nil))
f9272pqv.c:5:1: error: shared rtx
(mem/c:SI (plus:DI (reg/f:DI 19 frame)
(const_int -4 [0xfffc])) [0 S4 A32])
during RTL pass: stv
f9272pqv.c:5:1: internal compiler error: internal consistency failure
0x9f3eca verify_rtx_sharing
/var/tmp/portage/sys-devel/gcc-10.0.0_alpha20190818/work/gcc-10-20190818/gcc/emit-rtl.c:2927
0x9f3dba verify_rtx_sharing
/var/tmp/portage/sys-devel/gcc-10.0.0_alpha20190818/work/gcc-10-20190818/gcc/emit-rtl.c:2942
0x9f3dba verify_rtx_sharing
/var/tmp/portage/sys-devel/gcc-10.0.0_alpha20190818/work/gcc-10-20190818/gcc/emit-rtl.c:2942
0x9f3dba verify_rtx_sharing
/var/tmp/portage/sys-devel/gcc-10.0.0_alpha20190818/work/gcc-10-20190818/gcc/emit-rtl.c:2942
0x9f4178 verify_insn_sharing
/var/tmp/portage/sys-devel/gcc-10.0.0_alpha20190818/work/gcc-10-20190818/gcc/emit-rtl.c:3013
0x9f79ed verify_rtl_sharing()
/var/tmp/portage/sys-devel/gcc-10.0.0_alpha20190818/work/gcc-10-20190818/gcc/emit-rtl.c:3036
0xc7a8f3 execute_function_todo
/var/tmp/portage/sys-devel/gcc-10.0.0_alpha20190818/work/gcc-10-20190818/gcc/passes.c:2004
0xc7b536 execute_todo
/var/tmp/portage/sys-devel/gcc-10.0.0_alpha20190818/work/gcc-10-20190818/gcc/passes.c:2037
[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498 Rainer Orth changed: What|Removed |Added CC||ro at gcc dot gnu.org --- Comment #8 from Rainer Orth --- The new testcase FAILs on both i386-pc-solaris2.11 and x86_64-pc-linux-gnu with -m32: +FAIL: gcc.target/i386/minmax-7.c scan-assembler pminsd
[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498 --- Comment #7 from H.J. Lu --- Our SPEC CPU 2017 failed with 39 util.c:205:1: error: invalid rtl sharing found in the insn 40 205 | } 41 | ^ 42 (insn 29 28 3 2 (set (subreg:V2DI (reg:DI 91) 0) 43 (vec_concat:V2DI (mem/c:DI (plus:DI (reg/f:DI 19 frame) 44 (const_int -8 [0xfff8])) [0 S8 A64]) 45 (const_int 0 [0]))) "util.c":134:1 -1 46 (nil)) 47 util.c:205:1: error: shared rtx 48 (mem/c:DI (plus:DI (reg/f:DI 19 frame) 49 (const_int -8 [0xfff8])) [0 S8 A64]) 50 during RTL pass: stv We have an invalid shared rtx.
[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498 --- Comment #6 from H.J. Lu --- (In reply to Uroš Bizjak from comment #5) > (In reply to Uroš Bizjak from comment #4) > > Following patch > > HJ, can you please put the patch through some benchmarks? (I have no access > to SPEC). Sure. We will measure it. Thanks.
[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498 --- Comment #5 from Uroš Bizjak --- (In reply to Uroš Bizjak from comment #4) > Following patch HJ, can you please put the patch through some benchmarks? (I have no access to SPEC).
[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498 --- Comment #4 from Uroš Bizjak --- Following patch --cut here-- diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c index 49ab50ea41bf..11c75be113e0 100644 --- a/gcc/config/i386/i386.c +++ b/gcc/config/i386/i386.c @@ -18601,9 +18601,9 @@ ix86_register_move_cost (machine_mode mode, reg_class_t class1_i, where integer modes in SSE registers are not tieable because of missing QImode and HImode moves to, from or between MMX/SSE registers. */ -return MAX (8, SSE_CLASS_P (class1) - ? ix86_cost->hard_register.sse_to_integer - : ix86_cost->hard_register.integer_to_sse); +return (SSE_CLASS_P (class1) + ? ix86_cost->hard_register.sse_to_integer + : ix86_cost->hard_register.integer_to_sse); if (MAYBE_FLOAT_CLASS_P (class1)) return ix86_cost->hard_register.fp_move; --cut here-- fixes the regression. (BTW: QI/HImodes are not "missing" at all. We can use %k modifier to move QImode and HImode values using MOVD instruction between register sets.)
[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498 --- Comment #3 from Richard Biener --- Author: rguenth Date: Tue Aug 20 08:45:56 2019 New Revision: 274694 URL: https://gcc.gnu.org/viewcvs?rev=274694&root=gcc&view=rev Log: 2019-08-20 Richard Biener PR target/91498 * config/i386/i386-features.c (general_scalar_chain::convert_op): Use (vec_merge (vec_duplicate..)) style vector from scalar move. (convert_scalars_to_vector): Add timode_p parameter and use it to guard TImode-only operation. (pass_stv::gate): Adjust so STV runs twice for TARGET_64BIT. (pass_stv::execute): Pass down timode_p. * gcc.target/i386/minmax-7.c: New testcase. Added: trunk/gcc/testsuite/gcc.target/i386/minmax-7.c Modified: trunk/gcc/ChangeLog trunk/gcc/config/i386/i386-features.c trunk/gcc/testsuite/ChangeLog
[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498 --- Comment #2 from Uroš Bizjak --- (In reply to Richard Biener from comment #1) > Patch that causes gcc.target/i386/minmax-6.c to spill to the stack: > https://gcc.gnu.org/ml/gcc-patches/2019-08/msg01283.html > > This makes SI/DImode chains handled by STV2 only (after combine). We can > take it from there if we think that's what we need to do anyways. In the > end it's STV/RA interaction, I guess if we can take combine out of the > equation it might simplify things. Yes, this is the intended way SI/DImode STV should be used.
[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498 --- Comment #1 from Richard Biener --- Patch that causes gcc.target/i386/minmax-6.c to spill to the stack: https://gcc.gnu.org/ml/gcc-patches/2019-08/msg01283.html This makes SI/DImode chains handled by STV2 only (after combine). We can take it from there if we think that's what we need to do anyways. In the end it's STV/RA interaction, I guess if we can take combine out of the equation it might simplify things.
[Bug target/91498] [10 Regression] STV change in r274481 causes 300.twolf regression on Haswell
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91498 Richard Biener changed: What|Removed |Added Keywords||missed-optimization, ra Target||x86_64-*-* i?86-*-* CC||hjl.tools at gmail dot com, ||uros at gcc dot gnu.org Target Milestone|--- |10.0
