[Bug rtl-optimization/55845] New: 454.calculix miscompares with -march=btver2 -O3 -ffastmath -fschedule-insns -mvzeroupper for test data run

[Bug rtl-optimization/55845] New: 454.calculix miscompares with -march=btver2 -O3 -ffastmath -fschedule-insns -mvzeroupper for test data run

[venkataramanan.kumar at amd dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55845

             Bug #: 55845
           Summary: 454.calculix miscompares with -march=btver2 -O3
                    -ffastmath -fschedule-insns -mvzeroupper for test data
                    run
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: venkataramanan.kumar@amd.com
              Host: x86_64-unknown-linux-gnu
            Target: x86_64-unknown-linux-gnu
             Build: x86_64-unknown-linux-gnu


GCC revision used: 194705
Flags used:  -march=btver2 -O3 -ffastmath -fschedule-insns -mvzeroupper

Observation:
(Snip)
*** Miscompare of beampic.dat; for details see
   
/local/home/vekumar/specruns/benchspec/CPU2006/454.calculix/run/run_base_test_gcc480_bt_sched.exe_nofma.0000/beampic.dat.mis
Error: 1x454.calculix
Producing Raw Reports
(Snip)

On triaging further the file SPOOLES/Utilities/src/Utilities_DV.Utilities_DV.c
when compiled with -march=btver2 -O3 -ffastmath -fschedule-insns -mvzeroupper
seems to schedule vzeroupper incorrectly.

vzeroupper will clobber all ymm registers, so should not be moved arround.

Also using -march=corei7-avx seems to have problem.

I am drilling down further.

[Bug rtl-optimization/55845] 454.calculix miscompares with -march=btver2 -O3 -ffastmath -fschedule-insns -mvzeroupper for test data run

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55845

Uros Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |vbyakovl23 at gmail dot com

--- Comment #1 from Uros Bizjak <ubizjak at gmail dot com> 2013-01-02 10:10:01 UTC ---
Adding CC.

[Bug rtl-optimization/55845] 454.calculix miscompares with -march=btver2 -O3 -ffastmath -fschedule-insns -mvzeroupper for test data run

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55845

--- Comment #2 from Uros Bizjak <ubizjak at gmail dot com> 2013-01-02 10:15:16 UTC ---
It looks to me that scheduler is doing something with unspec_volatile
instructions, since vzeroupper insertion will be cancelled if there are live
call-saved SSE registers at the insertion point. Please see
ix86_avx_emit_vzeroupper function in i386.c

[Bug rtl-optimization/55845] 454.calculix miscompares with -march=btver2 -O3 -ffastmath -fschedule-insns -mvzeroupper for test data run

[vbyakovl23 at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55845

--- Comment #3 from Vladimir Yakovlev <vbyakovl23 at gmail dot com> 2013-01-04 06:40:44 UTC ---
The test fails corei7-avx also. I build a simple reproducer.

-------------------------------------------------------------
#include <stdio.h>

#define N 100

double
foo ( 
   int      size, 
   double   y[], 
   double   x[] 
)
{
  double sum = 0.0 ;
  int i ;
  for (i = 0, sum = 0. ; i < size ; i++)
    sum += y[i] * x[i] ; 
  return(sum);
}

int main ()
{
  double x[N];
  double y[N];
  double s;
  int i;

  for (i = 0; i < N; i++)
    {
      x[i] = i;
      y[i] = i;
    }

  s = foo (N, y, x);

  printf("%s\n", s == 328350 ? "pass" : "fail");
}
------------------------------------------------------

$ gcc -mavx -g -static  -o t  -O3 -ffast-math -march=corei7 t.c -fno-inline &&
./t
pass
$ gcc  -fschedule-insns -mavx -g -static  -o t  -O3 -ffast-math -march=corei7
t.c -fno-inline && ./t                                                          
fail

Responsible phase is jump2. To switch off the phase I did changes

diff --git a/gcc/cfgcleanup.c b/gcc/cfgcleanup.c
index 5d142e9..be04c5d 100644
--- a/gcc/cfgcleanup.c
+++ b/gcc/cfgcleanup.c
@@ -3070,6 +3070,7 @@ struct rtl_opt_pass pass_jump =
 static unsigned int
 execute_jump2 (void)
 {
+  if (!getenv("NOJMP2"))
   cleanup_cfg (flag_crossjumping ? CLEANUP_CROSSJUMP : 0);
   return 0;
 }

env NOJMP2=1 gcc  -fschedule-insns -mavx -g -static  -o t  -O3 -ffast-math
-march=corei7 t.c -fno-inline && ./t
pass

I used compiler
Target: x86_64-unknown-linux-gnu
Configured with: ../gcc/configure CFLAGS='-O0 -g3' CXXFLAGS='-O0 -g3'
--prefix=/gnumnt/msticlxl16_users/vbyakovl/workspaces/gcc/install
--enable-clocale=gnu --disable-bootstrap --with-system-zlib --enable-shared
--with-demangler-in-ld --with-fpmath=sse --with-arch=corei7-avx
--with-cpu=corei7-avx --enable-languages=c,c++,fortran,java,lto,objc
--no-create --no-recursion

[Bug rtl-optimization/55845] 454.calculix miscompares with -march=btver2 -O3 -ffastmath -fschedule-insns -mvzeroupper for test data run

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55845

Uros Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2013-01-04
     Ever Confirmed|0                           |1

--- Comment #4 from Uros Bizjak <ubizjak at gmail dot com> 2013-01-04 08:13:03 UTC ---
Please note how unspec_volatile vzeroupper (insn 225) moves above conditional
jump comparison, passing many instructions:

_.c.218r.csa:

  ...
  110: {r8:SI=r8:SI+r11:SI;clobber flags:CC;}
      REG_UNUSED flags:CC
  111: flags:CCZ=cmp(bx:SI,r11:SI)
      REG_DEAD r11:SI
      REG_DEAD bx:SI
  105: xmm2:V4DF=vec_select(vec_concat(xmm1:V4DF,xmm1:V4DF),parallel)
  106: xmm1:V4DF=xmm1:V4DF+xmm2:V4DF
      REG_DEAD xmm2:V4DF
  234: xmm1:V2DF=xmm1:V2DF
  235: xmm1:DF=xmm1:DF
  109: xmm0:DF=xmm0:DF+xmm1:DF
      REG_DEAD xmm1:DF
  112: pc={(flags:CCZ==0)?L230:pc}
      REG_DEAD flags:CCZ
      REG_BR_PROB 0xd06
  226: NOTE_INSN_BASIC_BLOCK 22
  225: unspec/v[0] 43
  113: L113:
  114: NOTE_INSN_BASIC_BLOCK 23
  ...

_.c.219r.jump2:

  ...
  110: {r8:SI=r8:SI+r11:SI;clobber flags:CC;}
      REG_UNUSED flags:CC
  225: unspec/v[0] 43
  111: flags:CCZ=cmp(bx:SI,r11:SI)
      REG_DEAD r11:SI
      REG_DEAD bx:SI
  105: xmm2:V4DF=vec_select(vec_concat(xmm1:V4DF,xmm1:V4DF),parallel)
  106: xmm1:V4DF=xmm1:V4DF+xmm2:V4DF
      REG_DEAD xmm2:V4DF
  109: xmm0:DF=xmm0:DF+xmm1:DF
      REG_DEAD xmm1:DF
  112: pc={(flags:CCZ==0)?L231:pc}
      REG_DEAD flags:CCZ
      REG_BR_PROB 0xd06
  113: L113:
  114: NOTE_INSN_BASIC_BLOCK 23
  ...

Confirmed as rtl-optimization infrastructure problem.

[Bug rtl-optimization/55845] 454.calculix miscompares with -march=btver2 -O3 -ffastmath -fschedule-insns -mvzeroupper for test data run

[venkataramanan.kumar at amd dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55845

--- Comment #5 from Venkataramanan <venkataramanan.kumar at amd dot com> 2013-01-04 09:12:09 UTC ---
Hi Uros.

Thank you for conforming.

I suspected this is a problem with jump optimization.  Uros what flag we need
to pass to bypass the .jump2 pass?

Regards,
Venkat.


-----Original Message-----
From: ubizjak at gmail dot com [mailto:gcc-bugzilla@gcc.gnu.org] 
Sent: Friday, January 04, 2013 1:43 PM
To: Kumar, Venkataramanan
Subject: [Bug rtl-optimization/55845] 454.calculix miscompares with
-march=btver2 -O3 -ffastmath -fschedule-insns -mvzeroupper for test data run


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55845

Uros Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2013-01-04
     Ever Confirmed|0                           |1

--- Comment #4 from Uros Bizjak <ubizjak at gmail dot com> 2013-01-04 08:13:03
UTC --- Please note how unspec_volatile vzeroupper (insn 225) moves above
conditional jump comparison, passing many instructions:

_.c.218r.csa:

  ...
  110: {r8:SI=r8:SI+r11:SI;clobber flags:CC;}
      REG_UNUSED flags:CC
  111: flags:CCZ=cmp(bx:SI,r11:SI)
      REG_DEAD r11:SI
      REG_DEAD bx:SI
  105: xmm2:V4DF=vec_select(vec_concat(xmm1:V4DF,xmm1:V4DF),parallel)
  106: xmm1:V4DF=xmm1:V4DF+xmm2:V4DF
      REG_DEAD xmm2:V4DF
  234: xmm1:V2DF=xmm1:V2DF
  235: xmm1:DF=xmm1:DF
  109: xmm0:DF=xmm0:DF+xmm1:DF
      REG_DEAD xmm1:DF
  112: pc={(flags:CCZ==0)?L230:pc}
      REG_DEAD flags:CCZ
      REG_BR_PROB 0xd06
  226: NOTE_INSN_BASIC_BLOCK 22
  225: unspec/v[0] 43
  113: L113:
  114: NOTE_INSN_BASIC_BLOCK 23
  ...

_.c.219r.jump2:

  ...
  110: {r8:SI=r8:SI+r11:SI;clobber flags:CC;}
      REG_UNUSED flags:CC
  225: unspec/v[0] 43
  111: flags:CCZ=cmp(bx:SI,r11:SI)
      REG_DEAD r11:SI
      REG_DEAD bx:SI
  105: xmm2:V4DF=vec_select(vec_concat(xmm1:V4DF,xmm1:V4DF),parallel)
  106: xmm1:V4DF=xmm1:V4DF+xmm2:V4DF
      REG_DEAD xmm2:V4DF
  109: xmm0:DF=xmm0:DF+xmm1:DF
      REG_DEAD xmm1:DF
  112: pc={(flags:CCZ==0)?L231:pc}
      REG_DEAD flags:CCZ
      REG_BR_PROB 0xd06
  113: L113:
  114: NOTE_INSN_BASIC_BLOCK 23
  ...

Confirmed as rtl-optimization infrastructure problem.

--
Configure bugmail: http://gcc.gnu.org/bugzilla/userprefs.cgi?tab=email
------- You are receiving this mail because: ------- You reported the bug.

[Bug rtl-optimization/55845] 454.calculix miscompares with -march=btver2 -O3 -ffastmath -fschedule-insns -mvzeroupper for test data run

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55845

--- Comment #6 from Uros Bizjak <ubizjak at gmail dot com> 2013-01-04 09:20:16 UTC ---
(In reply to comment #5)

> I suspected this is a problem with jump optimization.  Uros what flag we need
> to pass to bypass the .jump2 pass?

The pass should be fixed to handle unspec_volatile insns correctly, not
disabled.

[Bug rtl-optimization/55845] 454.calculix miscompares with -march=btver2 -O3 -ffastmath -fschedule-insns -mvzeroupper for test data run

[venkataramanan.kumar at amd dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55845

--- Comment #7 from Venkataramanan <venkataramanan.kumar at amd dot com> 2013-01-04 09:24:47 UTC ---
Hi Uros, 

Sorry , my intention was just know if we have some flags to  disable this pass.

Not as a solution :)

Regards,
Venkat.



-----Original Message-----
From: ubizjak at gmail dot com [mailto:gcc-bugzilla@gcc.gnu.org] 
Sent: Friday, January 04, 2013 2:50 PM
To: Kumar, Venkataramanan
Subject: [Bug rtl-optimization/55845] 454.calculix miscompares with
-march=btver2 -O3 -ffastmath -fschedule-insns -mvzeroupper for test data run


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55845

--- Comment #6 from Uros Bizjak <ubizjak at gmail dot com> 2013-01-04 09:20:16
UTC --- (In reply to comment #5)

> I suspected this is a problem with jump optimization.  Uros what flag 
> we need to pass to bypass the .jump2 pass?

The pass should be fixed to handle unspec_volatile insns correctly, not
disabled.

--
Configure bugmail: http://gcc.gnu.org/bugzilla/userprefs.cgi?tab=email
------- You are receiving this mail because: ------- You reported the bug.

[Bug rtl-optimization/55845] [4.8 Regression] 454.calculix miscompares with -march=btver2 -O3 -ffastmath -fschedule-insns -mvzeroupper for test data run

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55845

Uros Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |wrong-code
   Target Milestone|---                         |4.8.0
            Summary|454.calculix miscompares    |[4.8 Regression]
                   |with -march=btver2 -O3      |454.calculix miscompares
                   |-ffastmath -fschedule-insns |with -march=btver2 -O3
                   |-mvzeroupper for test data  |-ffastmath -fschedule-insns
                   |run                         |-mvzeroupper for test data
                   |                            |run

--- Comment #8 from Uros Bizjak <ubizjak at gmail dot com> 2013-01-04 09:35:12 UTC ---
(In reply to comment #7)

> Sorry , my intention was just know if we have some flags to  disable this pass.

It looks to me that -fno-crossjumping should do the trick.

[Bug rtl-optimization/55845] [4.8 Regression] 454.calculix miscompares with -march=btver2 -O3 -ffastmath -fschedule-insns -mvzeroupper for test data run

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55845

Uros Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |bonzini at gnu dot org

--- Comment #9 from Uros Bizjak <ubizjak at gmail dot com> 2013-01-06 13:23:54 UTC ---
It is try_head_merge_bb that moves unspec_volatile to the wrong place. However,
the core of the problem lies in can_move_insns_accross, which does not
special-case unspec_volatiles.

The untested df-problems.c patch that fixes the failing testcase:

--cut here--
Index: df-problems.c
===================================================================
--- df-problems.c       (revision 194945)
+++ df-problems.c       (working copy)
@@ -3916,6 +3916,10 @@ can_move_insns_across (rtx from, rtx to, rtx acros
        break;
       if (NONDEBUG_INSN_P (insn))
        {
+         /* Do not move unspec_volatile insns.  */
+         if (GET_CODE (PATTERN (insn)) == UNSPEC_VOLATILE)
+           break;
+
          if (may_trap_or_fault_p (PATTERN (insn))
              && (trapping_insns_in_across || other_branch_live != NULL))
            break;
--cut here--

Adding Paolo B. to CC.

[Bug rtl-optimization/55845] [4.8 Regression] 454.calculix miscompares with -march=btver2 -O3 -ffastmath -fschedule-insns -mvzeroupper for test data run

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55845

Uros Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                URL|                            |http://gcc.gnu.org/ml/gcc-p
                   |                            |atches/2013-01/msg00252.htm
                   |                            |l

--- Comment #10 from Uros Bizjak <ubizjak at gmail dot com> 2013-01-06 15:50:49 UTC ---
Patch at [1].

[1] http://gcc.gnu.org/ml/gcc-patches/2013-01/msg00252.html

[Bug rtl-optimization/55845] [4.8 Regression] 454.calculix miscompares with -march=btver2 -O3 -ffastmath -fschedule-insns -mvzeroupper for test data run

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55845

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P3                          |P1

[Bug rtl-optimization/55845] [4.8 Regression] 454.calculix miscompares with -march=btver2 -O3 -ffastmath -fschedule-insns -mvzeroupper for test data run

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55845

--- Comment #11 from Jakub Jelinek <jakub at gcc dot gnu.org> 2013-01-08 18:00:30 UTC ---
Author: jakub
Date: Tue Jan  8 18:00:10 2013
New Revision: 195028

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=195028
Log:
    PR rtl-optimization/55845
    * df-problems.c (can_move_insns_across): Stop scanning at
    volatile_insn_p source instruction or give up if
    across_from .. across_to range contains any volatile_insn_p
    instructions.

    * gcc.target/i386/pr55845.c: New test.

Added:
    trunk/gcc/testsuite/gcc.target/i386/pr55845.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/df-problems.c
    trunk/gcc/testsuite/ChangeLog

[Bug rtl-optimization/55845] [4.8 Regression] 454.calculix miscompares with -march=btver2 -O3 -ffastmath -fschedule-insns -mvzeroupper for test data run

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55845

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
                 CC|                            |jakub at gcc dot gnu.org
         Resolution|                            |FIXED

--- Comment #12 from Jakub Jelinek <jakub at gcc dot gnu.org> 2013-01-08 18:03:48 UTC ---
Fixed.

[Bug rtl-optimization/55845] [4.8 Regression] 454.calculix miscompares with -march=btver2 -O3 -ffastmath -fschedule-insns -mvzeroupper for test data run

[venkataramanan.kumar at amd dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55845

Venkataramanan <venkataramanan.kumar at amd dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|RESOLVED                    |VERIFIED

--- Comment #13 from Venkataramanan <venkataramanan.kumar at amd dot com> 2013-01-15 14:04:51 UTC ---
Veried the bugfix with gcc r195045.