[Bug middle-end/45720] New: [4.6 regression] Revision 164367 miscompiled SPEC CPU 2K

[Bug middle-end/45720] New: [4.6 regression] Revision 164367 miscompiled SPEC CPU 2K

[hjl dot tools at gmail dot com]

Revision 164367:

http://gcc.gnu.org/ml/gcc-cvs/2010-09/msg00661.html

miscompiled SPEC CPU 2K at -O3.

On Linux/x86-64, I got

  Running 186.crafty ref peak lnx32e-gcc default
*** Miscompare of crafty.out, see
/export/gnu/import/svn/gcc-test/spec/2000/x86_
64/spec/benchspec/CINT2000/186.crafty/run/00000004/crafty.out.mis
...
  Running 200.sixtrack ref peak lnx32e-gcc default
*** Miscompare of inp.out, see
/export/gnu/import/svn/gcc-test/spec/2000/x86_64/
spec/benchspec/CFP2000/200.sixtrack/run/00000004/inp.out.mis

with -O3 -funroll-loops -ffast-math.

On Linux/ia32, I got

  Running 254.gap ref peak lnx32-gcc default
*** Miscompare of ref.out, see
/export/gnu/import/svn/gcc-test/spec/2000/i686/sp
ec/benchspec/CINT2000/254.gap/run/00000004/ref.out.mis

with -O3 -funroll-loops -msse2 -mfpmath=sse -ffast-math.


-- 
           Summary: [4.6 regression] Revision 164367 miscompiled SPEC CPU 2K
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: hjl dot tools at gmail dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45720

[Bug middle-end/45720] [4.6 regression] Revision 164367 miscompiled SPEC CPU 2K

[hjl dot tools at gmail dot com]

-- 

hjl dot tools at gmail dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |4.6.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45720

[Bug middle-end/45720] [4.6 regression] Revision 164367 miscompiled SPEC CPU 2K

[rguenth at gcc dot gnu dot org]

------- Comment #1 from rguenth at gcc dot gnu dot org  2010-09-18 18:45 -------
I also see 450.soplex and 481.wrf fail.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45720

[Bug middle-end/45720] [4.6 regression] Revision 164367 miscompiled SPEC CPU 2K

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45720

H.J. Lu <hjl.tools at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|                            |FIXED

--- Comment #13 from H.J. Lu <hjl.tools at gmail dot com> 2010-10-23 02:08:57 UTC ---
Fixed as of revision 165840.

[Bug middle-end/45720] [4.6 regression] Revision 164367 miscompiled SPEC CPU 2K

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45720

--- Comment #12 from Richard Guenther <rguenth at gcc dot gnu.org> 2010-10-22 14:44:53 UTC ---
Author: rguenth
Date: Fri Oct 22 14:44:48 2010
New Revision: 165832

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=165832
Log:
2010-10-22  Richard Guenther  <rguenther@suse.de>

    PR tree-optimization/45720
    * tree-vect-data-refs.c (vect_update_misalignment_for_peel):
    Handle negative step.
    (vect_enhance_data_refs_alignment): Likewise.
    * tree-vect-loop-manip.c (vect_gen_niters_for_prolog_loop): Likewise.
    (vect_create_cond_for_align_checks): Likewise.
    (vect_create_cond_for_alias_checks): Likewise.

    * gcc.dg/torture/pr45720.c: New testcase.

Added:
    trunk/gcc/testsuite/gcc.dg/torture/pr45720.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/tree-vect-data-refs.c
    trunk/gcc/tree-vect-loop-manip.c

[Bug middle-end/45720] [4.6 regression] Revision 164367 miscompiled SPEC CPU 2K

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45720

--- Comment #11 from Richard Guenther <rguenth at gcc dot gnu.org> 2010-10-22 13:22:18 UTC ---
Created attachment 22116
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22116
untested patch

Patch which fixes the testcase (and some more).

[Bug middle-end/45720] [4.6 regression] Revision 164367 miscompiled SPEC CPU 2K

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45720

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2010.10.22 12:25:43
     Ever Confirmed|0                           |1

--- Comment #10 from Richard Guenther <rguenth at gcc dot gnu.org> 2010-10-22 12:25:43 UTC ---
Peeling for alignment needs adjustment:

float val[256];
float x;
void __attribute__((noinline,noclone))
foo(int len, int beg)
{
  int i;
  for (i = len - 1; i >= beg; --i)
    x += val[i] * 2;
}
int main()
{
  foo(256-3, 0);
  return 0;
}

segfaults.  We align &val[len-1] instead of &val[len-4].

[Bug middle-end/45720] [4.6 regression] Revision 164367 miscompiled SPEC CPU 2K

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45720

--- Comment #9 from Uros Bizjak <ubizjak at gmail dot com> 2010-10-22 11:21:27 UTC ---
(In reply to comment #8)

> So I have no hypotheses about the reason of this exception.
> Maybe, it was indirect jump into the middle of instruction.
> I will continue attemts to create short test case after weekend.

Try to run the test under valgrind. This tool is much more powerful.

[Bug middle-end/45720] [4.6 regression] Revision 164367 miscompiled SPEC CPU 2K

[vladimir.a.kharchenko at intel dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45720

--- Comment #8 from Vladimir Kharchenko <vladimir.a.kharchenko at intel dot com> 2010-10-22 10:15:16 UTC ---
It can happen that this file is minimal test case. :(
The piece of source:
961              if ((l_maxabs = temp.s_max[rw]) < 0)
962              {
963                 l_maxabs = fabs(u.row.val[len]);
964
965                 for (i = len - 1; i >= beg; --i)
966                    if (l_maxabs < fabs(u.row.val[i]))  ----> Seg. fault
967                       l_maxabs = fabs(u.row.val[i]);
968
969                 temp.s_max[rw] = l_maxabs;               /* ##### */
970              }
is executed many times in this test before Segmentation fault.
GDB shows that all objects (l_maxabs, fabs(u.row.val[i]), len, i, beg) have
correct values after Seg exception. Instructions after disassembler:
0x000000000041385b <soplex::CLUFactor::selectPivots(soplex::Real)+2203>:       
and    $0x7,%r8d
0x000000000041385f <soplex::CLUFactor::selectPivots(soplex::Real)+2207>:       
movapd 0x0(%rbp),%xmm2     ----> Seg. fault
0x0000000000413864 <soplex::CLUFactor::selectPivots(soplex::Real)+2212>:       
cmp    %r11d,%edx
also looks correct. %rbp points to correct memory.

So I have no hypotheses about the reason of this exception.
Maybe, it was indirect jump into the middle of instruction.
I will continue attemts to create short test case after weekend.

[Bug middle-end/45720] [4.6 regression] Revision 164367 miscompiled SPEC CPU 2K

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45720

--- Comment #7 from Richard Guenther <rguenth at gcc dot gnu.org> 2010-10-22 08:59:56 UTC ---
We need some sort of a testcase.

[Bug middle-end/45720] [4.6 regression] Revision 164367 miscompiled SPEC CPU 2K

[vladimir.a.kharchenko at intel dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45720

--- Comment #6 from Vladimir Kharchenko <vladimir.a.kharchenko at intel dot com> 2010-10-22 04:28:46 UTC ---
Quick investigation of 450.soplex failure shows that Segmentation fault is in
line 966 (file factor.cc). When I recompiled this file without the option
"-ffast=math", test passed.

[Bug middle-end/45720] [4.6 regression] Revision 164367 miscompiled SPEC CPU 2K

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45720

--- Comment #5 from H.J. Lu <hjl.tools at gmail dot com> 2010-10-22 00:28:51 UTC ---
With SPEC CPU 2006, revision 165771 gave me:

1. 64bit using-O3 -funroll-loops -ffast-math:

  Running 450.soplex ref peak lnx32e-gcc default

450.soplex: copy 0 non-zero return code (exit code=0, signal=11)

  Running 481.wrf ref peak lnx32e-gcc default

481.wrf: copy 0 non-zero return code (exit code=0, signal=11)

2. 32bit using -O3 -funroll-loops -msse2 -mfpmath=sse -ffast-math:  

  Running 481.wrf ref peak lnx32-gcc default

481.wrf: copy 0 non-zero return code (exit code=0, signal=11)

[Bug middle-end/45720] [4.6 regression] Revision 164367 miscompiled SPEC CPU 2K

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45720

--- Comment #4 from H.J. Lu <hjl.tools at gmail dot com> 2010-10-21 16:28:24 UTC ---
As of revision 165771, I still got

With runspec -c lnx-i686-gcc.cfg -n 1 -l -o asc -I all -T peak

*** Miscompare of ref.out, see
/export/gnu/import/svn/gcc-test/spec/2000/i686/spec/benchspec/CINT2000/254.gap/run/00000002/ref.out.mis

Error: 1x254.gap

With runspec -c lnx-x86_64-gcc.cfg -n 1 -l -o asc -I all -T peak

*** Miscompare of crafty.out, see
/export/gnu/import/svn/gcc-test/spec/2000/x86_64/spec/benchspec/CINT2000/186.crafty/run/00000002/crafty.out.mis

*** Miscompare of inp.out, see
/export/gnu/import/svn/gcc-test/spec/2000/x86_64/spec/benchspec/CFP2000/200.sixtrack/run/00000002/inp.out.mis

Error: 1x186.crafty 1x200.sixtrack

[Bug middle-end/45720] [4.6 regression] Revision 164367 miscompiled SPEC CPU 2K

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45720

--- Comment #2 from hjl at gcc dot gnu.org <hjl at gcc dot gnu.org> 2010-10-03 05:41:44 UTC ---
Author: hjl
Date: Sun Oct  3 05:39:32 2010
New Revision: 164914

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=164914
Log:
Disallow negative steps in vectorizer.

gcc/

2010-10-02  H.J. Lu  <hongjiu.lu@intel.com>

    PR tree-optimization/45720
    PR tree-optimization/45764
    * tree-vect-data-refs.c (vect_analyze_data_ref_access):
    Don't accept backwards consecutive accesses.
    (vect_create_data_ref_ptr): Disallow negative steps.

    * tree-vect-stmts.c (vectorizable_store): Allow negative steps.
    (perm_mask_for_reverse): Removed.
    (reverse_vec_elements): Likewise.
    (vectorizable_load): Don't hanle negative steps.

gcc/testsuite/

2010-10-02  H.J. Lu  <hongjiu.lu@intel.com>

    PR tree-optimization/45720
    PR tree-optimization/45764
    * g++.dg/torture/pr45764.C: New.

    * gcc.dg/vect/pr43432.c: Xfail.
    * gcc.dg/vect/vect-114.c: Likewise.
    * gcc.dg/vect/vect-15.c: Likewise.

Added:
    trunk/gcc/testsuite/g++.dg/torture/pr45764.C
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/testsuite/gcc.dg/vect/pr43432.c
    trunk/gcc/testsuite/gcc.dg/vect/vect-114.c
    trunk/gcc/testsuite/gcc.dg/vect/vect-15.c
    trunk/gcc/tree-vect-data-refs.c
    trunk/gcc/tree-vect-stmts.c

--- Comment #3 from Richard Guenther <rguenth at gcc dot gnu.org> 2010-10-21 10:03:49 UTC ---
Likely a dup of PR45764 which has all the analysis.

[Bug middle-end/45720] [4.6 regression] Revision 164367 miscompiled SPEC CPU 2K

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45720

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

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