[Bug target/91560] Try harder for AVX non-AVX2 cross-lane permutations

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91560

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |ASSIGNED
   Last reconfirmed|                            |2019-08-28
           Assignee|unassigned at gcc dot gnu.org      |jakub at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 46770
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=46770&action=edit
gcc10-pr91560.patch

Untested fix.
With this, we emit
        vpermilps       $144, %ymm1, %ymm2
        vpermilps       .LC0(%rip), %ymm1, %ymm1
        vblendps        $238, %ymm2, %ymm0, %ymm0
        vperm2f128      $1, %ymm1, %ymm1, %ymm1
        vblendps        $16, %ymm1, %ymm0, %ymm0
for f1 and
        vshufps $68, %ymm1, %ymm0, %ymm0
        vpermilps       .LC1(%rip), %ymm1, %ymm1
        vperm2f128      $1, %ymm1, %ymm1, %ymm1
        vblendps        $48, %ymm1, %ymm0, %ymm0
for f2.