[Bug rtl-optimization/56876] New: Combine does not invent new moves

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

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56876

             Bug #: 56876
           Summary: Combine does not invent new moves
    Classification: Unclassified
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: enhancement
          Priority: P3
         Component: rtl-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: glisse@gcc.gnu.org
            Target: x86_64-linux-gnu


Hello,

I am looking at this testcase:

typedef unsigned long long vec __attribute__((vector_size(16)));
vec g;
vec f1(vec a, vec b){
  return ~a&b;
}
vec f2(vec a, vec b){
  return ~g&b;
}

which compiles to:

f1:
    pandn    %xmm1, %xmm0

f2:
    pcmpeqd    %xmm0, %xmm0
    pxor    g(%rip), %xmm0
    pand    %xmm1, %xmm0

whereas I would like to get, like I do with the _mm_andnot_si128 builtin:

    movdqa    g(%rip), %xmm0
    pandn    %xmm1, %xmm0

It seems that combine cannot match the pandn pattern because the first argument
is a memory load and not a register. In this case, it would be better if it
emitted a move to put it in a register so it can match, instead of giving up. I
don't know if there is a good way to characterize such situations where an
extra move is worth it.