I am writing some code with a SIMD kernel that uses two separate slices, and the autovectorizer is producing a good deal of unaligned load instructions which some profiling