https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128709

--- Comment #4 from Zoltan Hidvegi <zoltan at hidvegi dot com> ---
For AVX-512:

__v64qi gcc_shuffle8(__v64qi a, __v64qi b) {
    return __builtin_shuffle(a, b);
}

This generates something like 400 lines of assembly instead of vpermb! For
__v32hi it's better, but it's using vpermt2w which is a much more expensive
instruction than vpermw.

Reply via email to