[Bug middle-end/52614] New: [4.8 Regression] Test failures in gcc.dg/vect: vectorizing unaligned access

[Bug middle-end/52614] New: [4.8 Regression] Test failures in gcc.dg/vect: vectorizing unaligned access

[dominiq at lps dot ens.fr]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52614

             Bug #: 52614
           Summary: [4.8 Regression] Test failures in gcc.dg/vect:
                    vectorizing unaligned access
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: dominiq@lps.ens.fr
                CC: Greta.Yorsh@arm.com, iains@gcc.gnu.org,
                    rguenth@gcc.gnu.org
            Target: arm*-*-* ppc*-*-* ia64-*-*


AFAICT pr52603 is not fixed after revision 185474 (see
http://gcc.gnu.org/ml/gcc-testresults/2012-03/msg02065.html ). In addition
similar errors occur on ppc*-*-* and ia64-*-* (see
http://gcc.gnu.org/ml/gcc-testresults/2012-03/msg02068.html ,
http://gcc.gnu.org/ml/gcc-testresults/2012-03/msg02104.html , and
http://gcc.gnu.org/ml/gcc-testresults/2012-03/msg02063.html ).

On powerpc-apple-darwin9, the tests pass with the following patch

--- ../../_gcc_clean/gcc/testsuite/gcc.dg/vect/vect.exp    2011-11-10
13:08:06.000000000 +0100
+++ ../../work/gcc/testsuite/gcc.dg/vect/vect.exp    2012-03-19
10:47:12.000000000 +0100
@@ -40,7 +40,7 @@ if ![check_vect_support_and_set_flags] {
 }

 # These flags are used for all targets.
-lappend DEFAULT_VECTCFLAGS "-ftree-vectorize" "-fno-vect-cost-model"
+lappend DEFAULT_VECTCFLAGS "-ftree-vectorize" "-fno-vect-cost-model"
"-fno-common"

 # Initialize `dg'.
 dg-init

        === gcc Summary for unix/-m64 ===

# of expected passes        3709
# of expected failures        127
# of unsupported tests        37

        === gcc Summary ===

# of expected passes        7443
# of expected failures        241
# of unsupported tests        64

I don't know if this related, but the failures reported in pr52491 disappeared
between revisions 185328 and 185486. Also on x86_64-apple-darwin10, the
polyhedron test gas_dyn.f90 runs 10% faster after revision 185454 than for
revision 185095.

[Bug testsuite/52614] [4.8 Regression] Test failures in gcc.dg/vect: vectorizing unaligned access

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52614

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|middle-end                  |testsuite
   Target Milestone|---                         |4.8.0

[Bug testsuite/52614] [4.8 Regression] Test failures in gcc.dg/vect: vectorizing unaligned access

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52614

William J. Schmidt <wschmidt at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |bergner at gcc dot gnu.org,
                   |                            |pthaugen at gcc dot
                   |                            |gnu.org, wschmidt at gcc
                   |                            |dot gnu.org

--- Comment #1 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2012-03-19 18:58:44 UTC ---
I was about to open a PR on this as well.  Adding some powerpc folks to the CC
list.

[Bug testsuite/52614] [4.8 Regression] Test failures in gcc.dg/vect: vectorizing unaligned access

[Greta.Yorsh at arm dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52614

--- Comment #2 from Greta Yorsh <Greta.Yorsh at arm dot com> 2012-03-20 11:24:41 UTC ---
The tests reported in PR52603 are still failing on arm (last tested on trunk
r185474).

[Bug testsuite/52614] [4.8 Regression] Test failures in gcc.dg/vect: vectorizing unaligned access

[dominiq at lps dot ens.fr]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52614

Dominique d'Humieres <dominiq at lps dot ens.fr> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2012-03-20
     Ever Confirmed|0                           |1

--- Comment #3 from Dominique d'Humieres <dominiq at lps dot ens.fr> 2012-03-20 15:16:45 UTC ---
> The tests reported in PR52603 are still failing on arm (last tested on trunk
> r185474).

When asked on IRC, Richard Guenther said he preferred a new PR rather than
reopening PR52603. He also asked me to test the addition of -fno-common to
DEFAULT_VECTCFLAGS in gcc/testsuite/gcc.dg/vect/vect.exp (patch in comment #0).
Could some of you test it on your favorite plaforms (it is enough to vect.exp
only)?

[Bug testsuite/52614] [4.8 Regression] Test failures in gcc.dg/vect: vectorizing unaligned access

[Greta.Yorsh at arm dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52614

--- Comment #4 from Greta Yorsh <Greta.Yorsh at arm dot com> 2012-03-20 16:01:02 UTC ---
These tests pass on arm-none-eabi with the option -fno-common. 

Thank you,
Greta

[Bug testsuite/52614] [4.8 Regression] Test failures in gcc.dg/vect: vectorizing unaligned access

[ro at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52614

--- Comment #5 from Rainer Orth <ro at gcc dot gnu.org> 2012-03-20 16:37:04 UTC ---
For the single testcase I've been using for the reghunt
(gcc.dg/vect/vect-109.c),
adding -fno-common fixes the failure.  On SPARC, one probably also has to add
-fno-common to gcc.target/sparc/fpmul-2.c.

I'll now run a full bootstrap with those changes.

  Rainer

[Bug testsuite/52614] [4.8 Regression] Test failures in gcc.dg/vect: vectorizing unaligned access

[pthaugen at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52614

--- Comment #6 from Pat Haugen <pthaugen at gcc dot gnu.org> 2012-03-20 17:21:21 UTC ---
Adding -fno-common fixes the failures on powerpc64 also.

[Bug testsuite/52614] [4.8 Regression] Test failures in gcc.dg/vect: vectorizing unaligned access

[dominiq at lps dot ens.fr]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52614

--- Comment #7 from Dominique d'Humieres <dominiq at lps dot ens.fr> 2012-03-21 12:57:54 UTC ---
I have posted the results for powerpc-apple-darwin9 with the patch in comment
#0 at http://gcc.gnu.org/ml/gcc-testresults/2012-03/msg02399.html . It shows
some remaining failures

FAIL: gcc.dg/vect/costmodel/ppc/costmodel-vect-outer-fir.c scan-tree-dump-times
vect "OUTER LOOP VECTORIZED" 2
FAIL: gcc.dg/vect/costmodel/ppc/costmodel-bb-slp-9a.c scan-tree-dump-times slp
"basic block vectorized using SLP" 1

They are fixed with the following patch

---
/opt/gcc/_gcc_clean/gcc/testsuite/gcc.dg/vect/costmodel/ppc/ppc-costmodel-vect.exp
   2011-01-05 08:39:45.000000000 +0100
+++
/opt/gcc/work/gcc/testsuite/gcc.dg/vect/costmodel/ppc/ppc-costmodel-vect.exp   
2012-03-21 10:55:27.000000000 +0100
@@ -34,7 +34,7 @@ if ![is-effective-target powerpc_altivec
 set DEFAULT_VECTCFLAGS ""

 # These flags are used for all targets.
-lappend DEFAULT_VECTCFLAGS "-O2" "-ftree-vectorize" "-fvect-cost-model"
+lappend DEFAULT_VECTCFLAGS "-O2" "-ftree-vectorize" "-fvect-cost-model"
"-fno-common"

 # If the target system supports vector instructions, the default action
 # for a test is 'run', otherwise it's 'compile'.  Save current default.

Note that I don't have write access to SVN, so someone will have to commit the
patches when approved.

Now I wonder if  '-fno-common' should not be the default when
'-ftree-vectorize' is used on the affected targets?

[Bug testsuite/52614] [4.8 Regression] Test failures in gcc.dg/vect: vectorizing unaligned access

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52614

--- Comment #8 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2012-03-21 18:34:55 UTC ---
I would be willing to commit the patches once they're approved.  Let me know.

[Bug testsuite/52614] [4.8 Regression] Test failures in gcc.dg/vect: vectorizing unaligned access

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52614

--- Comment #9 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2012-03-28 16:28:42 UTC ---
Dominique, can you please post your patches out on the gcc-patches mailing list
so they can be approved?  Again, I'm happy to commit them once they've met
approval by the maintainers, but I'm not permitted to post other people's
patches.

Thanks,
Bill

[Bug testsuite/52614] [4.8 Regression] Test failures in gcc.dg/vect: vectorizing unaligned access

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52614

--- Comment #10 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2012-04-05 12:11:58 UTC ---
Author: wschmidt
Date: Thu Apr  5 12:11:50 2012
New Revision: 186163

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=186163
Log:
gcc/testsuite:

2012-04-05  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>
        Dominique D'Humieures  <dominiq@lps.ens.fr>

    PR testsuite/52614
    * gcc.dg/vect/vect.exp: Use -fno-common on all targets.
    * gcc.dg/vect/costmodel/ppc/ppc-costmodel-vect.exp: Likewise.


Modified:
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/testsuite/gcc.dg/vect/costmodel/ppc/ppc-costmodel-vect.exp
    trunk/gcc/testsuite/gcc.dg/vect/vect.exp

[Bug testsuite/52614] [4.8 Regression] Test failures in gcc.dg/vect: vectorizing unaligned access

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52614

William J. Schmidt <wschmidt at gcc dot gnu.org> changed:

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

--- Comment #11 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2012-04-05 12:17:06 UTC ---
I am closing this as fixed, but the root cause (PR52571) remains open.  There
is still an open question whether -ftree-vectorize should now imply -fno-common
until the latter is no longer needed for the former to be useful.

[Bug testsuite/52614] [4.8 Regression] Test failures in gcc.dg/vect: vectorizing unaligned access

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52614

--- Comment #12 from William J. Schmidt <wschmidt at gcc dot gnu.org> 2012-04-05 12:46:12 UTC ---
Forget that last comment.  As Richard pointed out on gcc-patches:  "That's
probably more a C language question - you would get valid C rejected with
-fno-common.  But maybe -ftree-vectorize should suggest -fno-common when it
encounters a case it would like to promote alignment for.  Not sure if for
example Fortran would ever work with -fno-common though."

[Bug testsuite/52614] [4.8 Regression] Test failures in gcc.dg/vect: vectorizing unaligned access

[dominiq at lps dot ens.fr]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52614

--- Comment #13 from Dominique d'Humieres <dominiq at lps dot ens.fr> 2012-04-06 14:08:52 UTC ---
Thanks for the commit. I'll test the effect of -fno-common on gfortran.