> Newer instructions (e.g. AVX2) would simplify this slightly, but the compiler is actually smart about replacing some of my intrinsics with better replacements

> Newer instructions (e.g. AVX2) would simplify this slightly, but the compiler is actually smart about replacing some of my intrinsics with better replacements. I've seen this story many times before. What the author should be doing is reorganizing his code to be branch-free, clear about aliasing, and aggressive with vectorizing-friendly pragmas. Notice that when he wrote vectorized code in the 25-year-old SSE instructions, the compiler was able to optimize into better, more contemporary SIMD instructions? It's not because he used SSE intrinsics. It's because SSE intrinsics forced him to write branch-free code. If he took the same steps with scalar code, it would now be clearer and easier for the compiler to truly optimize his code, and it would be stable throughout time as new architectures come out.