[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 Andrew Pinski changed: What|Removed |Added CC||thomas at monjalon dot net --- Comment #34 from Andrew Pinski --- *** Bug 88096 has been marked as a duplicate of this bug. ***
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #33 from janus at gcc dot gnu.org --- (In reply to janus from comment #30) > I have verified that the problem goes away when using the following as > version: > GNU assembler (GNU Binutils) 2.31.51.20181124 Same ǵoes for the 2.30 branch as of today. (I hope there will be a 2.30.1 release in the near future.) According to comment 22, binutils 2.29 also seems to work well.
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #32 from H.J. Lu --- *** Bug 86831 has been marked as a duplicate of this bug. ***
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #31 from H.J. Lu --- *** Bug 86828 has been marked as a duplicate of this bug. ***
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #30 from janus at gcc dot gnu.org --- I have verified that the problem goes away when using the following as version: GNU assembler (GNU Binutils) 2.31.51.20181124
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 H.J. Lu changed: What|Removed |Added Status|NEW |RESOLVED See Also||https://sourceware.org/bugz ||illa/show_bug.cgi?id=23465 Resolution|--- |MOVED --- Comment #29 from H.J. Lu --- This is a binutils bug: https://sourceware.org/bugzilla/show_bug.cgi?id=23465
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735
--- Comment #28 from Alexander Monakov ---
It seems like object file is not correctly assembled, note wrong offset to the
last instruction (movdqu):
.s:
vpgatherqd (%rax,%ymm3,4), %xmm4{%k4}
vpgatherqd (%rax,%ymm1,4), %xmm0{%k1}
vshufi32x4 $0, %ymm0, %ymm4, %ymm0
vmovdqu32 %ymm0, 32(,%rsi,4)
.o:
52a: 62 f2 7d 2c 91 24 98vpgatherqd (%rax,%ymm3,4),%xmm4{%k4}
531: 62 f2 7d 29 91 04 88vpgatherqd (%rax,%ymm1,4),%xmm0{%k1}
538: 62 f3 5d 28 43 c0 00vshufi32x4 $0x0,%ymm0,%ymm4,%ymm0
53f: 62 f1 7e 28 7f 04 b5vmovdqu32 %ymm0,0x1(,%rsi,4)
546: 01 00 00 00
On older Binutils I get vmovdqu32 %ymm0,0x20(,%rsi,4) as expected.
Probably relevant Binutils bugs:
https://sourceware.org/bugzilla/show_bug.cgi?id=23465
https://sourceware.org/bugzilla/show_bug.cgi?id=23314
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #27 from janus at gcc dot gnu.org --- Created attachment 45086 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=45086=edit object file
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #26 from janus at gcc dot gnu.org --- (In reply to janus from comment #25) > Created attachment 45085 [details] > assembler output In fact this seems to be identical to Jakub's assembly (except for filenames and versions strings), but while it works for him, it fails for me.
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #25 from janus at gcc dot gnu.org --- Created attachment 45085 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=45085=edit assembler output assembler output obtained from comment 0 via "gfortran-9 c0.f90 -march=skylake-avx512 -Ofast -S", exhibiting wrong results
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #24 from Alexander Monakov --- Janus, can you attach your .s and .o files please?
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #23 from Jakub Jelinek --- Created attachment 45083 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=45083=edit pr86735.s Assembly I get which works for me.
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #22 from Jakub Jelinek --- Fedora 27, kernel 4.16.4-200 Intel(R) Core(TM) i9-7960X CPU glibc 2.26-30, binutils, 2.29-13
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #21 from janus at gcc dot gnu.org --- (In reply to Jakub Jelinek from comment #20) > Tried again, this time on real hw (i9-7960X) and still can't reproduce, Thanks for the data point, Jakub. Could you please share your kernel and glibc version? H.J.'s comment 15 sounds like the kernel version might make a difference. (Unfortunately I am stuck with kernel 4.15 right now and cannot verify this.) In any case, I still see the problem with current trunk and this config: * Ubuntu 18.04, kernel 4.15.0 * Intel(R) Core(TM) i9-7980XE CPU * glibc 2.27, binutils 2.30 My trunk build has been configured with: configure --program-suffix=-9 --build=x86_64-linux-gnu --host=x86_64-linux-gnu --target=x86_64-linux-gnu --with-arch=skylake-avx512 --prefix=/ --enable-languages=c,c++,fortran --disable-multilib
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #20 from Jakub Jelinek --- Tried again, this time on real hw (i9-7960X) and still can't reproduce, tried current trunk, current 8 branch and r263070 (i.e. trunk from the time this PR has been filed), none of this reproduces it.
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #19 from janus at gcc dot gnu.org --- (In reply to janus from comment #17) > Huh, sounds possible. However, I see the buggy behavior also with sde64 (on > the same machine where I also see it in real HW, see comment 16). Another machine (Haswell CPU, Ubuntu 18.04) also shows the bug with sde64.
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #18 from janus at gcc dot gnu.org --- (In reply to janus from comment #17) > However, I see the buggy behavior also with sde64 (on > the same machine where I also see it in real HW, see comment 16). But then again, I don't see it with sde64 on a different machine: * Ubuntu 16.04 * kernel 4.15.0 * glibc 2.23 * binutils 2.26 Kernel version (and gcc trunk version) is the same. So what could it be that makes the difference? glibc? binutils? Something else?
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #17 from janus at gcc dot gnu.org --- (In reply to Jakub Jelinek from comment #14) > /usr/src/sde/sde64 -- ./pr86735 >1 2 3 4 5 6 > 7 8 9 10 11 12 13 > 14 15 16 >1 2 3 4 5 6 > 7 8 9 10 11 12 13 > 14 15 16 > Perhaps sde64 behaves differently from real HW? Huh, sounds possible. However, I see the buggy behavior also with sde64 (on the same machine where I also see it in real HW, see comment 16).
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #16 from janus at gcc dot gnu.org --- (In reply to H.J. Lu from comment #15) > I can no longer reproduce it under kernel 4.17.11. But five days ago you could, right? What changed in the meantime? I still see the problem with 8.2.0 and today's trunk, in the following setup: * Ubuntu 18.04, kernel 4.15.0 * Intel(R) Core(TM) i9-7980XE CPU * glibc 2.27, binutils 2.30
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #15 from H.J. Lu --- I can no longer reproduce it under kernel 4.17.11.
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735
Jakub Jelinek changed:
What|Removed |Added
CC||jakub at gcc dot gnu.org
--- Comment #14 from Jakub Jelinek ---
Can't reproduce, either with gcc 8.2.0, or current 8.2.1, or latest trunk:
/d/gcc-8.2.0/objdir/gcc/gfortran -B /d/gcc-8.2.0/objdir/gcc/
-march=skylake-avx512 -Ofast -o pr86735{,.f90} -B
/d/gcc-8.2.0/objdir/x86_64-pc-linux-gnu/libgfortran/.libs/ -B
/d/gcc-8.2.0/objdir/x86_64-pc-linux-gnu/libgfortran/ -B
/d/gcc-8.2.0/objdir/x86_64-pc-linux-gnu/libquadmath/.libs/
-Wl,-rpath,../x86_64-pc-linux-gnu/libgfortran/.libs/; /usr/src/sde/sde64 --
./pr86735
1 2 3 4 5 6
7 8 9 10 11 12 13
14 15 16
1 2 3 4 5 6
7 8 9 10 11 12 13
14 15 16
Perhaps sde64 behaves differently from real HW? Tried also -O3 -fstack-arrays,
-mprefer-vector-width={128,256,512}, but everything works.
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #13 from janus at gcc dot gnu.org --- gfortran.fortran-torture/execute/in-pack.f90 in the testsuite exhibits the same problem as comment 0, I think. It also contains Fortran array operations and fails with "-march=skylake-avx512 -Ofast". See PR 86831 comment 5.
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #12 from janus at gcc dot gnu.org --- (In reply to Dominique d'Humieres from comment #11) > Thanks for the explanation. Note that > > % gfcp pr86735.f90 -Ofast -march=skylake -Wall -Wextra -fcheck=all That probably just means that -fcheck prevents the optimization passes from applying AVX512 instructions. Please cut the noise!
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #11 from Dominique d'Humieres --- > If you build for haswell or skylake architecture, that executable will > not run on ivybridge, of course. Thanks for the explanation. Note that % gfcp pr86735.f90 -Ofast -march=skylake -Wall -Wextra -fcheck=all % ./a.out 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #10 from janus at gcc dot gnu.org --- (In reply to Dominique d'Humieres from comment #8) > On x86_64-apple-darwin17 without access to avx512, I see > >[..] > > Is this the same bug or should I file a new one? That is not a bug at all. You apparently have an ivybridge processor, so with that flag you get the expected output. If you build for haswell or skylake architecture, that executable will not run on ivybridge, of course.
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #9 from H.J. Lu --- (In reply to Dominique d'Humieres from comment #8) > On x86_64-apple-darwin17 without access to avx512, I see > > Is this the same bug or should I file a new one? Please file a new one.
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #8 from Dominique d'Humieres --- On x86_64-apple-darwin17 without access to avx512, I see % gfcp pr86735.f90 -Ofast -march=skylake % ./a.out 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Program received signal SIGILL: Illegal instruction. and % gfcp pr86735.f90 -Ofast -march=haswell % ./a.out 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Program received signal SIGILL: Illegal instruction. but % gfcp pr86735.f90 -Ofast -march=ivybridge % ./a.out 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Is this the same bug or should I file a new one?
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735
--- Comment #7 from H.J. Lu ---
diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index fc3d6f0aebc..a8709972e9c 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -40402,6 +40402,10 @@ ix86_multiplication_cost (const struct processor_costs
*cost,
? cost->mulsd : cost->mulss, true);
else if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT)
{
+ /* vpmullq is used in this case. No emulation is needed. */
+ if (TARGET_AVX512DQ)
+ return ix86_vec_cost (mode, cost->mulss, true);
+
/* V*QImode is emulated with 7-13 insns. */
if (mode == V16QImode || mode == V32QImode)
{
triggered gather coedgen, which may lead to a latent bug.
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 H.J. Lu changed: What|Removed |Added Status|UNCONFIRMED |NEW Last reconfirmed||2018-07-30 CC||hjl.tools at gmail dot com Ever confirmed|0 |1 --- Comment #6 from H.J. Lu --- This is caused by r257505.
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 --- Comment #5 from janus at gcc dot gnu.org --- (In reply to Alexander Monakov from comment #3) > An observation in case it might help to narrow down the problem: gcc-8 > changed to prefer 256-bit-wide vectors on skylake-avx512, so adding > -mprefer-vector-width=512 option for gcc-8/trunk should make vectorization > behave a bit closer to gcc-7. Indeed adding the flag "-mprefer-vector-width=512" makes the problem disappear. Thanks for the comment.
[Bug target/86735] [8/9 Regression] Bad wrong-code bug with "-march=skylake-avx512 -Ofast"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86735 Richard Biener changed: What|Removed |Added Target||x86_64-*-* CC||kyukhin at gcc dot gnu.org Component|tree-optimization |target Target Milestone|--- |8.3 --- Comment #4 from Richard Biener --- Looks more like a target issue to me. AVX2 uses very similar gather stuff.
