The generic implementation is simplified by removing the 'optimization' for !_IEEE_FP_INEXACT (which does not handle inexact neither some values). Checked on alpha-linux-gnu.
