https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128081
--- Comment #2 from Roger Sayle <roger at nextmovesoftware dot com> --- It's impossible to completely trust google search any more, but it would like me to believe that llceilf of qNaN is undefined behaviour, and llceilf on +Inf and -Inf invokes FE_INVALID and returns an implementation defined value, e.g. LLONG_MAX, LLONG_MIN or zero. But I agree that if ever fold evaluates these at compile-time, they should produce the same value that you'd see at run-time. Apparently llceilf(NaN) produces a different value on x86_64 than on arm/aarch64. Still investigating...
