It depends. If your target is x86/x86_64 and the compiler has a LLVM or Intel backend, I've found that beating the compiler is almost impossible and largely futile. I took a long time to accept this since I know asm and the x86 instructions well.

That is my point of view as well, but then again the kind of code I write also does quite well with bounds checking enabled.

So I never had to deep dive into stuff like writing AVX by hand, just basing my remark on some comments that occasionally read about.

Compiler auto-vectorization is fairly easy to beat as it won’t even try when loop invariants are mildly complicated.