https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84844
--- Comment #4 from Uroš Bizjak <ubizjak at gmail dot com> --- (In reply to Jakub Jelinek from comment #3) > Or shall I got for ?*v instead, or ???*v, or !???*v ? Or we can revert PR78090 fix and use preferred_for_speed attribute again. This will allow the compiler to emit reg->xmm for cold parts. BTW: we shouldn't just grep assembly dumps for unoptimal insns, as was done and reported in PR78090. There is no harm to emit reg->xmm form of the insn in the cold parts. Based on this, I prefer the revert of PR78090 patch, perhaps with a run of 462.libquantum to assess the performance impact.