https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103559
--- Comment #11 from Jakub Jelinek <jakub at gcc dot gnu.org> --- Actually, I was just misreading the tree dumps, we use hw insn for x u>= 0.0, which is UNGE_EXPR, so it is true if x is NAN or GE_EXPR, or as described in the tree-call-cdce.cc comment: y = sqrt (x); ==> if (__builtin_isless (x, 0)) y = sqrt (x); else y = IFN_SQRT (x); So we don't need to inspect anything on the libc sides, we already use the hw insn even for NANs. All we need to do is to use the ranger in tree-call-cdce.cc or perhaps later, and fold the x UNGE_EXPR 0.0 into true whenever frange says so (i.e. that the range is a subset of [-0.0, +INF] NAN).