The algorithm is the same, differing only on how we load data onto vector registers, so they can be merged, simplifying the code base. For all strings over 16 characters in length, we loop and then we perform a final overlapped comparison, if necessary. I've kept the 32- byte-per-loop solution even for pre-AVX2, because that should pipeline better. For any strings between 4 and 16 characters, we perform a pair of maybe-overlapped comparisons, of either 4 characters or of 8, so we consume the full string. That leaves a tail of at most 3 characters in scalar code. Change-Id: Ib42b3adc93bf4d43bd55fffd16c8c15d6d761af2 Reviewed-by: Lars Knoll <lars@knoll.priv.no>