https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115981
--- Comment #5 from GCC 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:df2b444a233e93b987adec76655ab89589b3fa10 commit r15-2473-gdf2b444a233e93b987adec76655ab89589b3fa10 Author: Jakub Jelinek <ja...@redhat.com> Date: Thu Aug 1 10:32:54 2024 +0200 i386: Fix up *<extract_type>_vinsert<shuffletype><extract_suf>_0 [PR115981] The r14-537 change started canonicalizing VEC_MERGE operands based on swap_commutative_operands_p or if they have the same precedence least significant bit of the third operand. The *<extract_type>_vinsert<shuffletype><extract_suf>_0 pattern was added for combine matching and no longer triggers after that change, as it used the reg_or_0_operand as the first operand and VEC_DUPLICATE as the second. Now, reg_or_0_operand could be a REG, SUBREG of object or CONST_VECTOR. REG has commutative_operand_precedence -1 or -2, SUBREG of object -3, CONST_VECTOR -4, while VEC_DUPLICATE has 0, so VEC_DUPLICATE will always go first and REG, SUBREG or CONST_VECTOR second. This patch swaps the operands so that it matches again. 2024-08-01 Jakub Jelinek <ja...@redhat.com> PR target/115981 * config/i386/sse.md (*<extract_type>_vinsert<shuffletype><extract_suf>_0): Swap the first two VEC_MERGE operands, renumber match_operands and test for 0xF or 0x3 rather than 0xFFF0 or 0xFC immediate. * gcc.target/i386/avx512dq-pr90991-1.c: Add tests for no separate zero extension instructions. * gcc.target/i386/avx512dq-pr90991-2.c: Likewise.
