[Bug target/63330] New: vector shuffle resembling vector shift not expanded optimally

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63330

            Bug ID: 63330
           Summary: vector shuffle resembling vector shift not expanded
                    optimally
           Product: gcc
           Version: 4.9.1
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: rguenth at gcc dot gnu.org
            Target: x86_64-*-*, i?86-*-*

typedef int v4si __attribute__((vector_size(16)));
v4si foo (v4si x)
{
  return __builtin_shuffle (x, (v4si){ 0, 0, 0, 0 },
                             (v4si){4, 3, 2, 1 });
}

and similar shuffles "shifting" a vector by whole-element amounts
left/right and inserting zeros are not expanded optimally (while
the target has at least a vec_shr optab which suggests sth would
be available).

With -mavx2 I get for the above

        vpxor   %xmm1, %xmm1, %xmm1
        vpalignr        $4, %xmm0, %xmm1, %xmm0
        vpshufd $27, %xmm0, %xmm0

while I expected sth like

        psrldq %xmm0, $4

__builtin_shuffle (x, (v4si) { -1, -1, -1, -1 }, ... )

Arbitrary constants "shifted in" could be handled the same by IORing
the shifted in value after the psrldq in the appropriate elements
for the cost of an extra vector constant.