Re moving the negation: it could be faster if x is available a few cycles before y in a superscalar architecture as you could save a cycle by negating x while in parallel computing y.