[Bug middle-end/55027] New: simplify vector multiplication by 1

[Bug middle-end/55027] New: simplify vector multiplication by 1

[glisse at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55027

             Bug #: 55027
           Summary: simplify vector multiplication by 1
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: glisse@gcc.gnu.org


Hello,

vector multiplication by 1 is not simplified at the tree level.

#include <x86intrin.h>
__m128d f(__m128d x){
  __m128d y={1,1};
  return x*y;
}

Worse, gcc can actually create such multiplies. In this modified version of
PR55016:

#include <math.h>
float v0[1024];
float v1[1024];
void v() {
    for(int i=0; i!=1024; ++i)
          v0[i] = 1/sqrtf(v1[i]);
}

compiled on x86_64 with gcc -O3 -ffast-math e.c -S -std=gnu99

*.optimized still shows:
  vect_var_.9_23 = vect_var_.8_21 * { 1.0e+0, 1.0e+0, 1.0e+0, 1.0e+0 };

The back-end later removes it, but it should also be handled at tree level.

[Bug middle-end/55027] simplify vector multiplication by 1

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55027

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
           Keywords|                            |missed-optimization
   Last reconfirmed|                            |2012-10-23
                 CC|                            |rguenth at gcc dot gnu.org
     Ever Confirmed|0                           |1
           Severity|normal                      |enhancement

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> 2012-10-23 09:41:54 UTC ---
For integer values the generic folding with

      if (! FLOAT_TYPE_P (type))
        {
          if (integer_zerop (arg1))
            return omit_one_operand_loc (loc, type, arg1, arg0);
          if (integer_onep (arg1))
            return non_lvalue_loc (loc, fold_convert_loc (loc, type, arg0));

catches this.  For float values we have

          /* Maybe fold x * 0 to 0.  The expressions aren't the same
             when x is NaN, since x * 0 is also NaN.  Nor are they the
             same in modes with signed zeros, since multiplying a
             negative value by 0 gives -0, not +0.  */
          if (!HONOR_NANS (TYPE_MODE (TREE_TYPE (arg0)))
              && !HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (arg0)))
              && real_zerop (arg1))
            return omit_one_operand_loc (loc, type, arg1, arg0);
          /* In IEEE floating point, x*1 is not equivalent to x for snans.
             Likewise for complex arithmetic with signed zeros.  */
          if (!HONOR_SNANS (TYPE_MODE (TREE_TYPE (arg0)))
              && (!HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (arg0)))
                  || !COMPLEX_FLOAT_TYPE_P (TREE_TYPE (arg0)))
              && real_onep (arg1))
            return non_lvalue_loc (loc, fold_convert_loc (loc, type, arg0));

but neiter real_zerop or real_onep handle vector types (there are more
predicates like that).  I adjusted the integer predicates to also handle
vector types, so I suppose the same needs to be done for the real predicates.

[Bug middle-end/55027] simplify vector multiplication by 1

[glisse at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55027

--- Comment #2 from Marc Glisse <glisse at gcc dot gnu.org> 2012-10-24 19:35:03 UTC ---
Created attachment 28525
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=28525
patch

Surprisingly enough, the attached patch doesn't cause any regression in the
testsuite (default languages). I'll add a simple testcase and submit it.

Some remaining predicates that don't handle vectors are integer_nonzerop,
integer_pow2p, but some adaptations are necessary in the users before they can
be changed (at least for pow2p).

[Bug middle-end/55027] simplify vector multiplication by 1

[glisse at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55027

--- Comment #3 from Marc Glisse <glisse at gcc dot gnu.org> 2012-10-29 17:17:01 UTC ---
Author: glisse
Date: Mon Oct 29 17:16:51 2012
New Revision: 192954

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=192954
Log:
2012-10-29  Marc Glisse  <marc.glisse@inria.fr>

    PR middle-end/55027

gcc/
    * tree.c (real_zerop, real_onep, real_twop, real_minus_onep):
    Handle VECTOR_CST.

testsuite/
    * gcc.dg/pr55027.c: New testcase.


Added:
    trunk/gcc/testsuite/gcc.dg/pr55027.c   (with props)
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/tree.c

Propchange: trunk/gcc/testsuite/gcc.dg/pr55027.c
            ('svn:eol-style' added)

Propchange: trunk/gcc/testsuite/gcc.dg/pr55027.c
            ('svn:keywords' added)

[Bug middle-end/55027] simplify vector multiplication by 1

[glisse at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55027

Marc Glisse <glisse at gcc dot gnu.org> changed:

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

--- Comment #4 from Marc Glisse <glisse at gcc dot gnu.org> 2012-10-29 17:20:37 UTC ---
Fixed.
See this email for potential bugs introduced by this patch:
http://gcc.gnu.org/ml/gcc-patches/2012-10/msg02596.html