total_cmp is precisely IEEE's separately specified total order for floats. It's just that the more common operators do something different, and that's perhaps better for most uses where NaN are inherently unexpected and generally indicate that some kind of error condition has occurred.