https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93594
--- Comment #5 from CVS Commits <cvs-commit at gcc dot gnu.org> --- The master branch has been updated by Jakub Jelinek <ja...@gcc.gnu.org>: https://gcc.gnu.org/g:3f740c67dbb90177aa71d3c60ef9b0fd2f44dbd9 commit r10-6472-g3f740c67dbb90177aa71d3c60ef9b0fd2f44dbd9 Author: Jakub Jelinek <ja...@redhat.com> Date: Thu Feb 6 11:08:59 2020 +0100 i386: Improve avx* vector concatenation [PR93594] The following testcase shows that for _mm256_set*_m128i and similar intrinsics, we sometimes generate bad code. All 4 routines are expressing the same thing, a 128-bit vector zero padded to 256-bit vector, but only the 3rd one actually emits the desired vmovdqa %xmm0, %xmm0 insn, the others vpxor %xmm1, %xmm1, %xmm1; vinserti128 $0x1, %xmm1, %ymm0, %ymm0 The problem is that the cast builtins use UNSPEC_CAST which is after reload simplified using a splitter, but during combine it prevents optimizations. We do have avx_vec_concat* patterns that generate efficient code, both for this low part + zero concatenation special case and for other cases too, so the following define_insn_and_split just recognizes avx_vec_concat made of a low half of a cast and some other reg. 2020-02-06 Jakub Jelinek <ja...@redhat.com> PR target/93594 * config/i386/predicates.md (avx_identity_operand): New predicate. * config/i386/sse.md (*avx_vec_concat<mode>_1): New define_insn_and_split. * gcc.target/i386/avx2-pr93594.c: New test.
