https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117839
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> --- It seems to me we fail to PRE the various rN:VnDI = const_vector [0] insns, one issue of course is that we have both V4DI and V2DI sets where with pseudos we don't know they are equivalent. While there's gcse after reload it's not really "PRE" AFAIK, and it might be as well confused by xmm0:V2DI=const_vector vs xmm0:V4DI=const_vector maybe we should try to canonicalize those after reload somehow?
