There seems to be very little performance regression in doing so (+5ns) so might as well go for the more correct behaviour.