[Bug target/93594] Missed optimization with _mm256_set/setr_m128i intrinsics

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93594

--- Comment #10 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Jakub Jelinek <jakub@gcc.gnu.org>:

https://gcc.gnu.org/g:f82617f229b336d856c18313339b14657e05c129

commit r10-6499-gf82617f229b336d856c18313339b14657e05c129
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Fri Feb 7 09:28:39 2020 +0100

    i386: Better patch to improve avx* vector concatenation [PR93594]

    After thinking some more on this, we can do better; rather than having to
    add a new prereload splitter pattern to catch all other cases where it
might
    be beneficial to fold first part of an UNSPEC_CAST back to the unspec
    operand, this patch reverts the *.md changes I've made yesterday and
instead
    tweaks the patterns, so that simplify-rtx.c can optimize those on its own.
    Instead of the whole SET_SRC being an UNSPEC through which simplify-rtx.c
    obviously can't optimize anything, this represents those patterns through a
    VEC_CONCAT (or two nested ones for the 128-bit -> 512-bit casts) with the
    operand as the low part of it and UNSPEC representing just the high part of
    it (the undefined, to be ignored, bits).  While richi suggested using
    already in GIMPLE for those using a SSA_NAME default definition (i.e.
    clearly uninitialized use), I'd say that uninit pass would warn about
those,
    but more importantly, in RTL it would probably force zero initialization of
    that or use or an uninitialized pseudo, all of which is hard to match in an
    pattern, so I think an UNSPEC is better for that.

    2020-02-07  Jakub Jelinek  <jakub@redhat.com>

        PR target/93594
        * config/i386/predicates.md (avx_identity_operand): Remove.
        * config/i386/sse.md (*avx_vec_concat<mode>_1): Remove.
        (avx_<castmode><avxsizesuffix>_<castmode>,
        avx512f_<castmode><avxsizesuffix>_256<castmode>): Change patterns to
        a VEC_CONCAT of the operand and UNSPEC_CAST.
        (avx512f_<castmode><avxsizesuffix>_<castmode>): Change pattern to
        a VEC_CONCAT of VEC_CONCAT of the operand and UNSPEC_CAST with
        UNSPEC_CAST.