[Bug tree-optimization/57601] New: Vector lowering could use larger modes

[glisse at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57601

            Bug ID: 57601
           Summary: Vector lowering could use larger modes
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: glisse at gcc dot gnu.org
            Target: x86_64-linux-gnu

typedef int vec __attribute__((vector_size(2*sizeof(int))));
vec f(vec a, vec b){
  return a-b;
}

    vmovq    %xmm0, %rcx
    vmovq    %xmm1, %rdx
    movl    %ecx, %eax
    shrq    $32, %rcx
    subl    %edx, %eax
    shrq    $32, %rdx
    subl    %edx, %ecx
    vmovd    %eax, %xmm2
    vpinsrd    $1, %ecx, %xmm2, %xmm0

(with -Ofast -mavx2) whereas if I change the size to 4, I get:

    vpsubd    %xmm1, %xmm0, %xmm0

which seems valid to me even for size 2. It is not clear to me how to model
that at tree level, maybe it would be easier to just implement V2SI operations
in the backend?