https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95171
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> --- OK, so this is always_inline crossing a IL semantic boundary (-fnon-call-exceptions + -ffinite-math-only). In w9 NaN > 0.0 does not trap but when inlined into o7 it suddenly does and thus is invalid GIMPLE.
