[Bug tree-optimization/56790] New: VEC_COND_EXPR not constant folded

["glisse at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56790

             Bug #: 56790
           Summary: VEC_COND_EXPR not constant folded
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: enhancement
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: glisse@gcc.gnu.org
            Target: x86_64-linux-gnu


#include <x86intrin.h>
__m128d f(){
  __m128d a={5,7};
  __m128d b={11,13};
  __m128i m={-1,0};
  return m?a:b;
}

With -msse3 or less, this is lowered to scalar operations and thus to a
constant. However, with -msse4 or higher, nothing folds it to a constant,
either at tree level or rtl level, and we end up with:

    pxor    %xmm0, %xmm0
    pcmpgtq    .LC0(%rip), %xmm0
    movapd    .LC2(%rip), %xmm1
    blendvpd    %xmm0, .LC1(%rip), %xmm1
    movapd    %xmm1, %xmm0

The most natural seems to be to add something to fold-const.c, which currently
only handles the case where the first argument is all 1s or all 0s. It would be
good if the optimization worked when the second and third arguments can be
either VECTOR_CST or CONSTRUCTOR, similarly to VEC_PERM_EXPR.