The other issue is instruction-level parallelism, as another poster in TFA pointed out. Even within a single loop iteration the "unoptimized" code is more likel

The other issue is instruction-level parallelism, as another poster in TFA pointed out. Even within a single loop iteration the "unoptimized" code is more likely to exploit multiple ALUs if they exist, regardless of vectorization instructions.