I wouldn't be surprised if the future compilers would optimize it away as well, since ffast-math basicly makes NaN's equivalent to undefined behaviour.