https://gcc.gnu.org/bugzilla/show_bug.cgi?id=124271
--- Comment #13 from GCC Commits <cvs-commit at gcc dot gnu.org> --- The master branch has been updated by hongtao Liu <[email protected]>: https://gcc.gnu.org/g:8835caacb1b5d929b9df3ec873570b73e41a6847 commit r17-5837-g8835caacb1b5d929b9df3ec873570b73e41a6847 Author: liuhongt <[email protected]> Date: Thu Oct 29 19:44:35 2026 -0700 match.pd: Fold masked vector multiply to widening even/odd multiply [PR124271] (x & lowhalf_mask) * (y & lowhalf_mask) on an integer vector only multiplies the zero-extended low halves of each element. That is VEC_WIDEN_MULT_EVEN_EXPR of x and y viewed as vectors of half-width unsigned elements (VEC_WIDEN_MULT_ODD_EXPR on big-endian, where the low halves are the odd lanes). On x86 this replaces the V*DImode multiply emulation (3 pmuludq plus shifts and adds) with a single pmuludq, and replaces vpmullq when AVX512DQ is available. gcc/ChangeLog: PR target/124271 * match.pd ((x & lowhalf_mask) * (y & lowhalf_mask)): New simplification to VEC_WIDEN_MULT_{EVEN,ODD}_EXPR. gcc/testsuite/ChangeLog: PR target/124271 * gcc.dg/torture/pr124271.c: New test. * gcc.target/i386/pr124271-1.c: New test. * gcc.target/i386/pr124271-2.c: New test.
