On superscalars running xor trick as is would be significantly slower because it implies a data dependency where there isn't one. But all OOO x86's optimize it away internally.