https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117734
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> --- (In reply to Hongtao Liu from comment #1) > But there's a saturation inside pmaddubsw, not a simple dot_prod pattern. Oh, I didn't see that - that wouldn't be correct then (saturation to signed int would match the source given overflow in int is UB, but the truncation is implementation-defined, not UB). So if there's no native dot_prod for qi->hi then let's close this as INVALID.
