https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115161
--- Comment #9 from Jakub Jelinek <jakub at gcc dot gnu.org> --- In that case we should separate *.md patterns which are used for C conversions from the patterns used by the intrinsics, keep using what we are right now for the former and either use UNSPEC (the quickest but not best code generating variant) or something more complex. If UNSPEC is used, we could get some constant folding back by adding gimple_fold handling for those and the like of __builtin_ia32_psubusb128 or __builtin_ia32_pminub128. For __builtin_ia32_cvttps2dq etc. obviously the folding should either punt folding if some argument is out of range or fold those to what the hw does.
