Modern compilers are pretty good, but it's still quite possible for a good human coder to exceed their performance even for normal code. For really tight algor

Modern compilers are pretty good, but it's still quite possible for a good human coder to exceed their performance even for normal code. For really tight algorithms I've still seen humans beat compilers by a lot, though it takes a lot of skill and some domain knowledge of the CPU. For code that benefits a lot from vectorization, human coders still beat the crap out of compilers. I am not aware of any vectorizing compiler or JIT that can even approach what a human coder can do with SSE, AVX, or NEON (ARM's equivalent). I think these CPU extensions are just too complex for current generation compilers to effectively deal with. They require too much abstract understanding of what's actually happening in the code and the CPU to use really effectively. I'm a bit surprised that there's been so little attention paid to the opportunities for applying deep learning and other advanced techniques to compiler optimization. It seems like this area is ripe for a new wave of innovation, but it's not happening. I do believe that many companies would pay for an advanced compiler capable of generating code that was significantly faster than stock compilers, but the speedup would have to be more than a few percent to justify spending money on it.