[Bug target/54566] New: __builtin_shuffle: use psrldq+pslldq+por for rotations

[glisse at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54566

             Bug #: 54566
           Summary: __builtin_shuffle: use psrldq+pslldq+por for rotations
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassig...@gcc.gnu.org
        ReportedBy: gli...@gcc.gnu.org
            Target: x86_64-linux-gnu


Hello,

this PR is based on those 2 emails:
http://gcc.gnu.org/ml/libstdc++/2012-09/msg00048.html
http://gcc.gnu.org/ml/libstdc++/2012-09/msg00050.html

which say that permutations that are rotations (like {1,2,3,4,5,6,7,0}) should
more often be based on the rotation instructions (as in _mm_srli_si128 and
_mm_slli_si128). Whether it is better than pshufb is not for me to say, but it
could at least help where pshufb is not available. (256 bit AVX2 versions are
also possible, although they require an additional lane swap)