[Bug target/69252] [4.9/5/6 Regression] gcc.dg/vect/vect-iv-9.c FAILs with -Os -fmodulo-sched -fmodulo-sched-allow-regmoves -fsched-pressure
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69252 --- Comment #12 from Jakub Jelinek --- Thus, Roman, can you please post your patch to gcc-patches? Thanks.
[Bug target/69252] [4.9/5/6 Regression] gcc.dg/vect/vect-iv-9.c FAILs with -Os -fmodulo-sched -fmodulo-sched-allow-regmoves -fsched-pressure
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69252 --- Comment #13 from Roman Zhuykov --- (In reply to Jakub Jelinek from comment #12) > Thus, Roman, can you please post your patch to gcc-patches? Ok, in addition to comment 3 link, reposted it right now https://gcc.gnu.org/ml/gcc-patches/2016-01/msg01336.html, and ask Martin to help with creating new regression test, since I'm not ready to setup powerpc qemu vm at the moment.
[Bug target/69252] [4.9/5/6 Regression] gcc.dg/vect/vect-iv-9.c FAILs with -Os -fmodulo-sched -fmodulo-sched-allow-regmoves -fsched-pressure
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69252 --- Comment #9 from Martin Sebor --- (In reply to Jakub Jelinek from comment #5) > a) bootstrap/regtest it The patch passes all C, C++, and FORTRAN tests with no regressions WRT baxeline. I'll look at (b) next.
[Bug target/69252] [4.9/5/6 Regression] gcc.dg/vect/vect-iv-9.c FAILs with -Os -fmodulo-sched -fmodulo-sched-allow-regmoves -fsched-pressure
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69252 --- Comment #11 from Martin Sebor --- The following is a list of names of tests that fail in the baseline but not with the patch: c-c++-common/torture/pr53505.c gcc.c-torture/execute/pr54471.c gcc.c-torture/execute/pr61682.c gcc.dg/vect/costmodel/ppc/costmodel-vect-outer-fir.c gcc.dg/vect/vect-cond-6.c gcc.dg/vect/vect-outer-fir.c gcc.dg/vect/vect-outer-fir-lb-big-array.c gcc.dg/vect/vect-outer-fir-lb.c gcc.dg/vect/vect-reduc-dot-s8a.c gfortran.dg/inline_sum_2.f90 libgomp.c/autopar-1.c libgomp.c/autopar-2.c libgomp.c/autopar-3.c libgomp.c/autopar-4.c libgomp.c/autopar-5.c libgomp.c/autopar-6.c libgomp.c/autopar-7.c libgomp.c/autopar-8.c libgomp.c/simd-5.c libgomp.c++/simd-5.C libgomp.c++/simd-7.C libgomp.c++/target-2.C libgomp.fortran/jacobi.f libgomp.fortran/udr7.f90 tr1/5_numerical_facilities/special_functions/08_cyl_bessel_i/check_value.cc tr1/5_numerical_facilities/special_functions/09_cyl_bessel_j/check_value.cc tr1/5_numerical_facilities/special_functions/11_cyl_neumann/check_value.cc tr1/5_numerical_facilities/special_functions/21_sph_bessel/check_value.cc tr1/5_numerical_facilities/special_functions/23_sph_neumann/check_value.cc
[Bug target/69252] [4.9/5/6 Regression] gcc.dg/vect/vect-iv-9.c FAILs with -Os -fmodulo-sched -fmodulo-sched-allow-regmoves -fsched-pressure
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69252 --- Comment #10 from Martin Sebor --- (In reply to Jakub Jelinek from comment #5) > b) try make check -jN > RUNTESTFLAGS='--target_board=unix/fmodulo-sched/-fmodulo-sched-allow- > regmoves/-fsched-pressure' with unpatched vs. patched, see if it doesn't > regress any test and perhaps improves some other than this one The patch reduces the number of failed tests. Below, the first numeric column is the baseline (unpatched) result and the second one is patched. === gcc Summary === # of expected passes9990999925 # of unexpected failures278 262 # of unexpected successes 20 20 # of expected failures 315 315 # of unresolved testcases 11 # of unsupported tests 2027 2027 === gfortran Summary === # of expected passes4145841461 # of unexpected failures21 18 # of expected failures 89 89 # of unsupported tests 83 83 === g++ Summary === # of expected passes9724697247 # of unexpected failures13 12 # of expected failures 263 263 # of unsupported tests 3799 3799 === libgomp Summary === # of expected passes4596 4610 # of unexpected failures16 2 # of unresolved testcases 22 # of unsupported tests 302 302 === libstdc++ Summary === # of expected passes1096010965 # of unexpected failures46 41 # of expected failures 68 68 # of unsupported tests 232 233
[Bug target/69252] [4.9/5/6 Regression] gcc.dg/vect/vect-iv-9.c FAILs with -Os -fmodulo-sched -fmodulo-sched-allow-regmoves -fsched-pressure
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69252 --- Comment #8 from Jakub Jelinek --- (In reply to Roman Zhuykov from comment #7) > (In reply to Jakub Jelinek from comment #5) > > insufficient SMS testsuite coverage. > Not sure it's helpful, but 3 weeks ago I succesfully reg-strapped some bunch > of my SMS patches including this fix on x86-64 and aarch64 using trunk > 20151222. fmodulo-sched and fmodulo-sched-allow-regmoves options were > enabled by default, fsched-pressure left untouched. No new regressions, > excluding some scan-assembler-times failures because of loop versioning. > Maybe later on stage1 I'll send all my stuff to gcc-patches. Please do, if the patches aren't posted, others can't benefit from them.
[Bug target/69252] [4.9/5/6 Regression] gcc.dg/vect/vect-iv-9.c FAILs with -Os -fmodulo-sched -fmodulo-sched-allow-regmoves -fsched-pressure
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69252 --- Comment #5 from Jakub Jelinek --- (In reply to Martin Sebor from comment #4) > Thanks for your help! I can confirm that the first patch fixes the problem > in the test cases on powerpc64le. (I haven't done any other testing with > it.) Martin, do you think you could a) bootstrap/regtest it b) try make check -jN RUNTESTFLAGS='--target_board=unix/fmodulo-sched/-fmodulo-sched-allow-regmoves/-fsched-pressure' with unpatched vs. patched, see if it doesn't regress any test and perhaps improves some other than this one c) maybe try even bootstrap/regtest with both unpatched/patched compiler changed to use these 3 options by default and see if we don't regress there? Because a) is probably very insufficient testing here, given the insufficient SMS testsuite coverage.
[Bug target/69252] [4.9/5/6 Regression] gcc.dg/vect/vect-iv-9.c FAILs with -Os -fmodulo-sched -fmodulo-sched-allow-regmoves -fsched-pressure
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69252 --- Comment #7 from Roman Zhuykov --- (In reply to Jakub Jelinek from comment #5) > insufficient SMS testsuite coverage. Not sure it's helpful, but 3 weeks ago I succesfully reg-strapped some bunch of my SMS patches including this fix on x86-64 and aarch64 using trunk 20151222. fmodulo-sched and fmodulo-sched-allow-regmoves options were enabled by default, fsched-pressure left untouched. No new regressions, excluding some scan-assembler-times failures because of loop versioning. Maybe later on stage1 I'll send all my stuff to gcc-patches.
[Bug target/69252] [4.9/5/6 Regression] gcc.dg/vect/vect-iv-9.c FAILs with -Os -fmodulo-sched -fmodulo-sched-allow-regmoves -fsched-pressure
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69252 --- Comment #6 from Martin Sebor --- (In reply to Jakub Jelinek from comment #5) Sure, I was already planning to do some of it today.
[Bug target/69252] [4.9/5/6 Regression] gcc.dg/vect/vect-iv-9.c FAILs with -Os -fmodulo-sched -fmodulo-sched-allow-regmoves -fsched-pressure
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69252 Richard Biener changed: What|Removed |Added Priority|P3 |P2
[Bug target/69252] [4.9/5/6 Regression] gcc.dg/vect/vect-iv-9.c FAILs with -Os -fmodulo-sched -fmodulo-sched-allow-regmoves -fsched-pressure
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69252 --- Comment #3 from Roman Zhuykov --- I'll try to help. While working with expanding SMS functionality 4-5 years ago (https://gcc.gnu.org/ml/gcc-patches/2011-07/msg01807.html), I create several fixes not connected to my main non-doloop-support patch. Unfortunately only two of that fixes were approved and only one fix was committed. Could you try this patch first of all: https://gcc.gnu.org/ml/gcc-patches/2011-12/msg01800.html This patch suits current trunk. There are links to other fixes, I can provide newer versions for trunk if needed. https://gcc.gnu.org/ml/gcc-patches/2011-09/msg02049.html https://gcc.gnu.org/ml/gcc-patches/2011-07/msg01804.html https://gcc.gnu.org/ml/gcc-patches/2011-12/msg00505.html https://gcc.gnu.org/ml/gcc-patches/2011-12/msg00506.html
[Bug target/69252] [4.9/5/6 Regression] gcc.dg/vect/vect-iv-9.c FAILs with -Os -fmodulo-sched -fmodulo-sched-allow-regmoves -fsched-pressure
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69252 --- Comment #4 from Martin Sebor --- Thanks for your help! I can confirm that the first patch fixes the problem in the test cases on powerpc64le. (I haven't done any other testing with it.)
[Bug target/69252] [4.9/5/6 Regression] gcc.dg/vect/vect-iv-9.c FAILs with -Os -fmodulo-sched -fmodulo-sched-allow-regmoves -fsched-pressure
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69252
--- Comment #2 from Jakub Jelinek ---
Looks like a SMS pass bug.
In the original loop there is a a[i] memory load followed by addition of i to
that, for understandable reasons SMS wants to move the load as far as possible
from the addition, so it is done in another iteration.
For that purpose, schedule_reg_moves adds:
756 move->insn = gen_move_insn (move->new_reg, copy_rtx
(prev_reg));
(insn 64 38 37 4 (set (reg:DI 197)
(reg:DI 181 [ ivtmp.5 ])) 444 {*movdi_internal64}
(nil))
But, when emitting the prologue before the loop (unrolled two iterations, where
the first one does just the memory load and stuff before that and the second
one does the stuff from before the memory load from second iteration, then the
addition from first iteration and finally memory load from second iteration),
it emits a copy of insn 64 only in the second iteration (after the addition,
which is fine), but not in the first iteration.
Thus we end up with:
li 9,0
ld 5,.LC2@toc(2)
li 10,24
mtctr 10
sldi 6,9,2
lwzx 10,5,6
addi 9,9,1
sldi 6,9,2
add 7,10,8
mr 8,9
lwzx 10,5,6
add 3,7,3
addi 9,9,1
extsw 3,3
where another mr 8,9 instruction is missing from somewhere after li 9,0 until
before addi 9,9,1 instruction.
Do we have any SMS scheduling expert around?
[Bug target/69252] [4.9/5/6 Regression] gcc.dg/vect/vect-iv-9.c FAILs with -Os -fmodulo-sched -fmodulo-sched-allow-regmoves -fsched-pressure
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69252
Martin Sebor changed:
What|Removed |Added
Target|powerpc64-unknown-linux-gnu |powerpc64*-unknown-linux-gn
||u
Status|UNCONFIRMED |NEW
Last reconfirmed||2016-01-13
CC||msebor at gcc dot gnu.org
Host|x86_64-pc-linux-gnu |
Ever confirmed|0 |1
--- Comment #1 from Martin Sebor ---
Confirmed with today's top of trunk (6.0.0 20160113) also on powerpc764le with
-O1 and -Os.
Using a slightly simplified test case
int a[N] = { 0, 2, 4, 8 };
__attribute__ ((noinline, noclone)) int main1 (void)
{
int s = 0;
int i;
for (i = 0; i != sizeof a / sizeof *a; i++)
s += i + a[i];
return s;
}
and looking at the tree dumps, the first one that looks suspicious (with -O1)
is:
$ cat t.c.152t.ivopts
;; Function main1 (main1, funcdef_no=0, decl_uid=2350, cgraph_uid=0,
symbol_order=1)
__attribute__((noclone, noinline))
main1 ()
{
unsigned long ivtmp.10;
int i;
int s;
int _4;
int _5;
int * _14;
void * _15;
:
_14 = + 18446744073709551612;
ivtmp.10_13 = (unsigned long) _14;
:
# s_10 = PHI
# i_12 = PHI
# ivtmp.10_9 = PHI
ivtmp.10_8 = ivtmp.10_9 + 4;
_15 = (void *) ivtmp.10_8;
_4 = MEM[base: _15, offset: 0B];
_5 = _4 + i_12;
s_6 = _5 + s_10;
i_7 = i_12 + 1;
if (i_7 != 4)
goto ;
else
goto ;
:
goto ;
:
# s_11 = PHI
return s_11;
}
[Bug target/69252] [4.9/5/6 Regression] gcc.dg/vect/vect-iv-9.c FAILs with -Os -fmodulo-sched -fmodulo-sched-allow-regmoves -fsched-pressure
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69252 Andrew Pinski changed: What|Removed |Added Target Milestone|--- |4.9.4
