https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127055
--- Comment #27 from Matthias Kretz (Vir) <mkretz at gcc dot gnu.org> --- (In reply to Roger Sayle from comment #26) > I've just seen the proposed patch posted, which explains these instructions > should only trap in the presence of signaling NaNs. So I believe the > correct test sould be !(flag_trapping_math && flag_signaling_nans). > Unfortunately, this still means you need to use -Ofast (or equivalent) to > generate these instructions. I don't understand. 1. https://compiler-explorer.com/z/9WTE6T3ad: A single __builtin_floorf compiles to vroundss. Why should four __builtin_floorf then not be compiled to vroundps? 2. Intel Docs for VROUNDPS: "Invalid (signaled only if SRC = SNaN)." which matches the C and ISO 60559 spec, no?
