[Bug target/98167] [x86] Failure to optimize operation on indentically shuffled operands into a shuffle of the result of the operation

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98167

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #11 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Just would like to clarify, for some of the intrinsics like these that take an
immediate argument which encodes the permutation, the right thing is to fold
them in the gimple folder hook, because not all the immediate values might map
to __builtin_shuffle and even if they would, expressing it in C code would be
too large.
And then there are intrinsics that don't have any immediate arguments nor
vector arguments containing the permutation, they just take one or two vectors
and apply one specific permutations.  Those IMHO should be done in the headers
using __builtin_shuffle.