Still, a lot of problems seem to remain firmly in the "write it with intrinsics" region even if you do your best to help the compiler out. For any operation that involves filtering a list, the compiler will not generate the lookup table of shuffles that is used in the fastest solutions (maybe the story is different on AVX-512 where there are dedicated filtering instructions or on NEON where movemask costs more because there isn't such an instruction). For parsing and formatting strings, the compiler usually will not take branchy code and decide to compute both branches and vectorize it even if it is guaranteed chunks of the appropriate size. Some concrete examples of the how simple these problems can be while remaining out of reach are "Given a mutable list of u32, remove all multiples of 3" and "Given a list of newline-separated UUIDs and an equally large list of mutable u128, parse the hexadecimal bytes of each UUID into the corresponding u128".
Still, a lot of problems seem to remain firmly in the "write it with intrinsics" region even if you do your best to help the compiler out. For any operation that involves filtering a list, the compiler will not generate the lookup table of shuffles that is used in the fastest solutions (maybe the story is different on AVX-512 where there are dedicated filtering instructions or on NEON where movemask costs more because there isn't such an instruction). For parsing and formatting strings, the compiler usually will not take branchy code and decide to compute both branches and vectorize it even if it is guaranteed chunks of the appropriate size. Some concrete examples of the how simple these problems can be while remaining out of reach are "Given a mutable list of u32, remove all multiples of 3" and "Given a list of newline-separated UUIDs and an equally large list of mutable u128, parse the hexadecimal bytes of each UUID into the corresponding u128".