Unless I am mistaken about the logic here. we should free the offset array at the end of the function.
Also, should we use an aligned allocation here? The memswap function documentation begs the user to use aligned addresses but that may not apply to things smaller than an AVX vector.