https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88461
Jakub Jelinek <jakub at gcc dot gnu.org> changed: What |Removed |Added ---------------------------------------------------------------------------- CC| |jakub at gcc dot gnu.org --- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> --- What we can do easily, but not sure if it won't lead to worse code for non-mask code, would be to add either =k, km or =*k, *km alternatives to the various zeroextend?i?i2 patterns, all we have right now are the mov?i_internal patterns. Because even the mov?i_internal patterns use *, perhaps best to use just that. For this testcase that would still imply a kmov* instruction. If we want to go further, we'd need patterns for the instructions that set masks (comparisons etc.), which would contain zero_extend around the mask to wider modes. That implies two iterators and guess it wouldn't play nicely with define_subst, so we'd need to do that manually.