The presumed bug where the value computed without the sign overflows
even though the negative result can be represented is not actually a
problem, because this only happens with signed results and the temporary
value is computed as unsigned (thus with extra range).
