https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117562
--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> --- Created attachment 59662 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=59662&action=edit preprocessed source This is the mgau_eval from the non-LTO slowdown. When you build with -Ofast -g -fopt-info-vec -march=znver4 you'll see this vmovhlps + vmovaps "spill"/reload pair. I didn't yet attempt to reduce further (register pressure is needed for the spill).
