[Bug target/57009] New: Select best typed instruction for scalar bitwise operations

[glisse at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57009



             Bug #: 57009

           Summary: Select best typed instruction for scalar bitwise

                    operations

    Classification: Unclassified

           Product: gcc

           Version: 4.9.0

            Status: UNCONFIRMED

          Keywords: missed-optimization

          Severity: enhancement

          Priority: P3

         Component: target

        AssignedTo: unassig...@gcc.gnu.org

        ReportedBy: gli...@gcc.gnu.org

            Target: x86_64-linux-gnu





Hello,



I purposedly took almost the same title as PR 54716, because this is the same

issue, but for scalars. Consider this code:



union A { double d; unsigned long long i; };



bool f(double a, double b, double c){

  A x, y, z;

  x.d = a * a;

  y.d = b * b;

  z.i = x.i & y.i;

  return z.d < c;

}



which compiles to:



    mulsd    %xmm0, %xmm0

    mulsd    %xmm1, %xmm1

    movq    %xmm0, %rax

    movq    %xmm1, %rdx

    andq    %rdx, %rax

    movq    %rax, %xmm0

    ucomisd    %xmm0, %xmm2

    seta    %al



when using andpd would save 3 movq. Note that for vectors, we get nicer code

thanks to Jakub's patch:



    mulpd    %xmm1, %xmm1

    mulpd    %xmm0, %xmm0

    andpd    %xmm1, %xmm0

    cmpltpd    %xmm2, %xmm0



It would be nice to extend that code to scalars (including the case where one

of the arguments is a constant).



I hit this while experimenting for PR 56944.