https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125949
--- Comment #3 from vekumar at gcc dot gnu.org --- (In reply to Richard Biener from comment #2) > Not sure what this is about - I see lots of spilling (so missed > optimization?), > is the size cost of choosing xmm/ymm > 15? Is it wrong-code for avxvnniint8 > specifically? Zen6 supports avxvnniint8 which are vex variants as shown below. VPDPBSSD xmm1, xmm2, xmm3/m128 VPDPBSSD ymm1, ymm2, ymm3/m256 VPDPBSSDS xmm1, xmm2, xmm3/m128 VPDPBSSDS ymm1, ymm2, ymm3/m256 VPDPBSUD xmm1, xmm2, xmm3/m128 VPDPBSUD ymm1, ymm2, ymm3/m256 VPDPBSUDS xmm1, xmm2, xmm3/m128 VPDPBSUDS ymm1, ymm2, ymm3/m256 VPDPBUUD xmm1, xmm2, xmm3/m128 VPDPBUUD ymm1, ymm2, ymm3/m256 VPDPBUUDS xmm1, xmm2, xmm3/m128 VPDPBUUDS ymm1, ymm2, ymm3/m256 Evex variants will result in illegal instructions.
