[Bug tree-optimization/57537] New: [4.8/4.9] gcc.dg/vect/slp-widen-mult-half.c generating wrong code on PowerPC64

[Bug tree-optimization/57537] New: [4.8/4.9] gcc.dg/vect/slp-widen-mult-half.c generating wrong code on PowerPC64

[brooks at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57537

            Bug ID: 57537
           Summary: [4.8/4.9] gcc.dg/vect/slp-widen-mult-half.c generating
                    wrong code on PowerPC64
           Product: gcc
           Version: 4.8.1
            Status: UNCONFIRMED
          Severity: major
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: brooks at gcc dot gnu.org

The gcc.dg/vect/slp-widen-mult-half.c test tests autovectorization of this
inner loop:

  for (i = 0; i < N/2; i++)
    {
      out[2*i] = in[2*i] * COEF;
      out2[2*i] = in[2*i] + COEF2;
      out[2*i+1] = in[2*i+1] * COEF;
      out2[2*i+1] = in[2*i+1] + COEF2;
    }


The execution portion of this test is failing on AltiVec-enabled powerpc64
targets on both the 4.8 branch (which is essentially at the point of the 4.8.1
release at the point I tested it) and on trunk.

The symptoms appear to strongly indicate that this is a simple matter of wrong
code generation: the out[] array is zeroed out rather than having the correct
values assigned.  The out2[] array is correct.

In case it helps someone, here is the line of instrumentation code that I added
to the test to investigate the computed values:

 for (i = 0; i < N; i++)
   printf("at %d: %d == %d, %d == %d\n", i, out[i], in[i] * COEF,
          out2[i], in[i] + COEF2);

I also tried putting some nonzero initial values in out[] before the vectorized
loop; the output remains zero, so this is being actively zeroed out.

[Bug tree-optimization/57537] [4.8/4.9 Regression] gcc.dg/vect/slp-widen-mult-half.c generating wrong code on PowerPC64

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57537

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |wrong-code
   Target Milestone|---                         |4.8.2
            Summary|[4.8/4.9]                   |[4.8/4.9 Regression]
                   |gcc.dg/vect/slp-widen-mult- |gcc.dg/vect/slp-widen-mult-
                   |half.c generating wrong     |half.c generating wrong
                   |code on PowerPC64           |code on PowerPC64

[Bug tree-optimization/57537] [4.8/4.9 Regression] gcc.dg/vect/slp-widen-mult-half.c generating wrong code on PowerPC64

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57537

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |ASSIGNED
   Last reconfirmed|                            |2013-06-12
                 CC|                            |jakub at gcc dot gnu.org
           Assignee|unassigned at gcc dot gnu.org      |jakub at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 30298
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=30298&action=edit
gcc49-pr57537.patch

Untested fix.

[Bug tree-optimization/57537] [4.8/4.9 Regression] gcc.dg/vect/slp-widen-mult-half.c generating wrong code on PowerPC64

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57537

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

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

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Author: jakub
Date: Wed Jun 12 13:01:49 2013
New Revision: 200000

URL: http://gcc.gnu.org/viewcvs?rev=200000&root=gcc&view=rev
Log:
    PR tree-optimization/57537
    * tree-vect-patterns.c (vect_recog_widen_mult_pattern): If
    vect_handle_widen_op_by_const, convert oprnd1 to half_type1.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/tree-vect-patterns.c

Author: jakub
Date: Wed Jun 12 13:01:29 2013
New Revision: 199999

URL: http://gcc.gnu.org/viewcvs?rev=199999&root=gcc&view=rev
Log:
    PR tree-optimization/57537
    * tree-vect-patterns.c (vect_recog_widen_mult_pattern): If
    vect_handle_widen_op_by_const, convert oprnd1 to half_type1.

Modified:
    branches/gcc-4_8-branch/gcc/ChangeLog
    branches/gcc-4_8-branch/gcc/tree-vect-patterns.c

[Bug tree-optimization/57537] [4.8/4.9 Regression] gcc.dg/vect/slp-widen-mult-half.c generating wrong code on PowerPC64

[brooks at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57537

--- Comment #3 from Brooks Moses <brooks at gcc dot gnu.org> ---
Thanks for the quick fix, Jakub!  And congratulations on the auspicious commit
number, as well.  :)