[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114 Ilya Enkovich changed: What|Removed |Added Status|NEW |RESOLVED Resolution|--- |FIXED --- Comment #23 from Ilya Enkovich --- Fixed
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114 --- Comment #22 from Ilya Enkovich --- Author: ienkovich Date: Tue May 17 09:28:15 2016 New Revision: 236315 URL: https://gcc.gnu.org/viewcvs?rev=236315=gcc=rev Log: gcc/ PR target/71114 * config/i386/i386.c (dimode_scalar_chain::convert_op): Fix insertion point for instructions generated by validize_mem. gcc/testsuite/ PR target/71114 * gcc.target/i386/pr70799-1.c: Fix scan for Darwin. Modified: trunk/gcc/ChangeLog trunk/gcc/config/i386/i386.c trunk/gcc/testsuite/ChangeLog trunk/gcc/testsuite/gcc.target/i386/pr70799-1.c
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114 Richard Biener changed: What|Removed |Added Target Milestone|--- |7.0
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114
--- Comment #21 from Dominique d'Humieres ---
> Probably just scan for movdqa then to get rid of addressing dependencies? Or
> use
>
> movdqa[ t]+.?LC0
>
> pattern if it works fine on Darwin.
The failure is indeed fixed with the following patch
--- ../_clean/gcc/testsuite/gcc.target/i386/pr70799-1.c 2016-05-10
18:46:37.0 +0200
+++ gcc/testsuite/gcc.target/i386/pr70799-1.c 2016-05-17 10:25:19.0
+0200
@@ -3,7 +3,7 @@
/* { dg-options "-O2 -march=slm" } */
/* { dg-final { scan-assembler "pxor" } } */
/* { dg-final { scan-assembler "pcmpeqd" } } */
-/* { dg-final { scan-assembler "movdqa\[ \\t\]+.LC0" } } */
+/* { dg-final { scan-assembler "movdqa\[ \\t\]+.?LC0" } } */
long long a, b, c;
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114 --- Comment #20 from Ilya Enkovich --- (In reply to Dominique d'Humieres from comment #18) > > smoke test passes, I'll leave it to Dominique's full-run to confirm. > > With the patch in comment 15 applied on top of revision r236286 the reported > failures are gone. > > From the fix, would it be possible to derive a test failing on linux? On Linux local objects are accessed in another way and additional instructions are not required. > > Also the test gcc.target/i386/pr70799-1.c fails > > FAIL: gcc.target/i386/pr70799-1.c scan-assembler movdqa[ t]+.LC0 > > In the assembly I see > > movdqa LC0-L3$pb(%eax), %xmm0 > > with nothing between \t and L. Probably just scan for movdqa then to get rid of addressing dependencies? Or use movdqa[ t]+.?LC0 pattern if it works fine on Darwin.
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114 --- Comment #19 from Dominique d'Humieres --- Results at https://gcc.gnu.org/ml/gcc-testresults/2016-05/msg01774.html.
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114 --- Comment #18 from Dominique d'Humieres --- > smoke test passes, I'll leave it to Dominique's full-run to confirm. With the patch in comment 15 applied on top of revision r236286 the reported failures are gone. From the fix, would it be possible to derive a test failing on linux? Also the test gcc.target/i386/pr70799-1.c fails FAIL: gcc.target/i386/pr70799-1.c scan-assembler movdqa[ t]+.LC0 In the assembly I see movdqa LC0-L3$pb(%eax), %xmm0 with nothing between \t and L.
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114 --- Comment #17 from Iain Sandoe --- smoke test passes, I'll leave it to Dominique's full-run to confirm.
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114 --- Comment #16 from Ilya Enkovich --- (In reply to Iain Sandoe from comment #15) > that code won't build - did you mean : > Sure. Thanks for noticing!
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114
--- Comment #15 from Iain Sandoe ---
(In reply to Dominique d'Humieres from comment #14)
> > Looks I found the problem. validize_mem generates new instructions which
> > are placed wrongly. This patch should help. Unfortunately I can't test
> > it properly on Darwin.
>
> Thanks for the feedback. I just started a clean regstrap, allow for ~6h+.
that code won't build - did you mean :
if (!standard_sse_constant_p (vec_cst, V2DImode))
- vec_cst = validize_mem (force_const_mem (V2DImode, vec_cst));
+ {
+ start_sequence ();
+ vec_cst = validize_mem (force_const_mem (V2DImode, vec_cst));
+ rtx_insn *seq = get_insns ();
+ end_sequence ();
+ emit_insn_before (seq, insn);
+ }
?
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114 --- Comment #14 from Dominique d'Humieres --- > Looks I found the problem. validize_mem generates new instructions which > are placed wrongly. This patch should help. Unfortunately I can't test > it properly on Darwin. Thanks for the feedback. I just started a clean regstrap, allow for ~6h+.
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114
--- Comment #13 from Ilya Enkovich ---
Looks I found the problem. validize_mem generates new instructions which are
placed wrongly. This patch should help. Unfortunately I can't test it
properly on Darwin.
diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index 046717b..5fdd12b 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -3688,7 +3688,13 @@ dimode_scalar_chain::convert_op (rtx *op, rtx_insn
*insn)
gen_rtvec (2, *op, const0_rtx));
if (!standard_sse_constant_p (vec_cst, V2DImode))
- vec_cst = validize_mem (force_const_mem (V2DImode, vec_cst));
+ {
+ start_sequence ();
+ vec_cst = validize_mem (force_const_mem (V2DImode, vec_cst));
+ rtx_insn *seq = get_insns ();
+ end_sequence ();
+ emit_insn_before (insns, insn)
+ }
emit_insn_before (gen_move_insn (tmp, vec_cst), insn);
*op = tmp;
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114 --- Comment #12 from Iain Sandoe --- (In reply to Iain Sandoe from comment #11) > I wonder if it's a stack/data alignment problem maybe not; comparing 6.1 and trunk with the function annotated to __attribute__((noinline)) it seems that the problem is introduced before the inlining; and it looks, as you say, to be related to the load of the constant value equivalent to data[0]. For 6.1 since the function is static the constant load is moved into variable_shift() and the value of i is passed in eax. This all shows up in a Linux X x86_64-apple-darwin10 configured like: ../gcc-mainline/configure --prefix=/somewhere --target=x86_64-apple-darwin10 --host=x86_64-linux-gnu --build=x86_64-linux-gnu --enable-languages=c --enable-targets=all
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114 --- Comment #11 from Iain Sandoe --- I wonder if it's a stack/data alignment problem
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114 --- Comment #10 from Iain Sandoe --- (In reply to Ilya Enkovich from comment #9) > (In reply to Iain Sandoe from comment #8) > > Created attachment 38495 [details] > > assembly for linux case > > > > Here's the asm for the case from my build, although I don't expect it'll > > help much if you can't reproduce it. > > But I can build and run this asm with no problem. What's wrong with it then? not sure, I'll see if I can analyse more.
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114 --- Comment #9 from Ilya Enkovich --- (In reply to Iain Sandoe from comment #8) > Created attachment 38495 [details] > assembly for linux case > > Here's the asm for the case from my build, although I don't expect it'll > help much if you can't reproduce it. But I can build and run this asm with no problem. What's wrong with it then?
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114 --- Comment #8 from Iain Sandoe --- Created attachment 38495 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=38495=edit assembly for linux case Here's the asm for the case from my build, although I don't expect it'll help much if you can't reproduce it.
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114 --- Comment #7 from Ilya Enkovich --- (In reply to Dominique d'Humieres from comment #2) > Created attachment 38486 [details] > Assembly for gcc.c-torture/execute/ashldi-1.c compiled with r236090 > > Assembly for gcc.c-torture/execute/ashldi-1.c compiled with r236090 and -Os > -m32. This part looks wrong: movd%xmm1, %edi movaps (%edi), %xmm0 Most probably it is an attempt to load a constant. Unfortunately I still don't have a way to reproduce it. Will try more.
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114
--- Comment #6 from Ilya Enkovich ---
(In reply to Iain Sandoe from comment #5)
> Also fails on Linux for the m32 multilib with -msse2avx :
> make -k check-gcc-c
> RUNTESTFLAGS="--target_board=unix/-msse/-msse2avx\{-m32,-m64\}
> execute.exp=ashldi-1.c"
Can't reproduce the problem on Linux (using r236170)
=== gcc Summary for unix/-msse/-msse2avx/-m32 ===
# of expected passes16
Running target unix/-msse/-msse2avx/-m64
=== gcc Summary for unix/-msse/-msse2avx/-m64 ===
# of expected passes16
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114
Iain Sandoe changed:
What|Removed |Added
Target|x86_64-apple-darwin*|x86_64-apple-darwin*,x86_64
||-pc-linux-gnu
Host|x86_64-apple-darwin*|x86_64-apple-darwin*,x86_64
||-pc-linux-gnu
Build|x86_64-apple-darwin*|x86_64-apple-darwin*,x86_64
||-pc-linux-gnu
--- Comment #5 from Iain Sandoe ---
Also fails on Linux for the m32 multilib with -msse2avx :
make -k check-gcc-c
RUNTESTFLAGS="--target_board=unix/-msse/-msse2avx\{-m32,-m64\}
execute.exp=ashldi-1.c"
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114
--- Comment #4 from Dominique d'Humieres ---
Reduced test
#include
extern void abort(void);
extern void exit(int);
#define BITS 8
static unsigned long long const data[8] = {
0x123456789abcdefULL,
0x2468acf13579bdeULL,
0x48d159e26af37bcULL,
0x91a2b3c4d5e6f78ULL,
0x123456789abcdef0ULL,
0x2468acf13579bde0ULL,
0x48d159e26af37bc0ULL,
0x91a2b3c4d5e6f780ULL
};
static unsigned long long
variable_shift(unsigned long long x, int i)
{
return x << i;
}
int
main()
{
int i;
for (i = 0; i < BITS; ++i)
{
unsigned long long y = variable_shift (data[0], i);
if (y != data[i])
abort ();
}
exit (0);
}
It succeeds if I replace '#define BITS 8' with '#define BITS 7'.
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114 Dominique d'Humieres changed: What|Removed |Added CC||ienkovich at gcc dot gnu.org --- Comment #3 from Dominique d'Humieres --- Also seen on x86_64-apple-darwin10 (see https://gcc.gnu.org/ml/gcc-testresults/2016-05/msg01499.html).
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114 Iain Sandoe changed: What|Removed |Added Status|UNCONFIRMED |NEW Last reconfirmed||2016-05-14 CC||iains at gcc dot gnu.org Ever confirmed|0 |1
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114 --- Comment #2 from Dominique d'Humieres --- Created attachment 38486 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=38486=edit Assembly for gcc.c-torture/execute/ashldi-1.c compiled with r236090 Assembly for gcc.c-torture/execute/ashldi-1.c compiled with r236090 and -Os -m32.
[Bug target/71114] [7 Regression] Several test suite failures on x86_64-apple-darwin* after revision r236090
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71114 --- Comment #1 from Dominique d'Humieres --- Created attachment 38485 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=38485=edit Assembly for gcc.c-torture/execute/ashldi-1.c for r236089 Assembly for gcc.c-torture/execute/ashldi-1.c compiled with r236089 and -Os -m32.
