[Bug target/93395] AVX2 missed optimization : _mm256_permute_pd() is unfortunately translated into the more expensive VPERMPD instead of the cheap VPERMILPD

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93395

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |jakub at gcc dot gnu.org

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 47699
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=47699&action=edit
gcc10-pr93395.patch

We don't miss any patterns, just have them ordered incorrectly.  RTL-wise,
vpermilp* is a subset of vpermp*, but it was placed after vpermp* and thus
vpermp* matched always when -mavx2 was also enabled, thus vpermilp* got only
emitted with -mavx -mno-avx2.