https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123607
--- Comment #9 from Matthias Kretz (Vir) <mkretz at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #5)
> Looking at the Intel manuals, there are quite a few instruction forms which
> only accept {k1} without {z},
I would expect all stores to go without {z} even if one could define a meaning
for it. For COMPRESS there's no useful meaning for {z} even. VPTEST and
VFPCLASS are documented to unconditionally use {z} (i.e. there's never a
dependency on the dst register). For SCATTER the offsets under write-mask might
lead to invalid memory accesses (and also, it's a store). For GATHER, I have no
idea ...
