https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127773
--- Comment #8 from Hongtao Liu <liuhongt at gcc dot gnu.org> --- Created attachment 65813 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=65813&action=edit The patch fixed the testcase in the PR Could you please check if the patch also fix the issue in the real workload? The patch explicitly detect if there're sse registers live through call_insn, if not still set it as AVX_U128_CLEAN(insert vzeroupper before the call), otherwise use AVX_U128_ANY(insert vzeroupper after the call).
