https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115629
--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> --- I think we sort-of agreed to have .MASK_LOAD to have inactive lanes zeroed but we never got around to formalizing that. Note AVX512 supports zero-masking for .MASK_STORE IIRC, not just merge. Same for .MASK_LOAD where it supports merging as well. How would we ask the target whether it can do 'merge'? zero and undefined could be done via predicates but merge?
