https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80137
--- Comment #2 from Jonathan Wakely <redi at gcc dot gnu.org> ---
We need to handle targets without C99's nextafter, but maybe this would be OK:
if (__builtin_expect(__ret >= _RealType(1), 0))
{
#if _GLIBCXX_USE_C99_MATH_TR1
__ret = std::nextafter(_RealType(1), _RealType(0));
#else
__ret = _RealType(0);
#endif
}
If we reach 1.0 it's because we're rounding up or to nearest, and so returning
zero in that case would counteract the decreased probability of getting zero
otherwise.
