here's a new paper about bridging the gap between discovering a missing compiler optimization (both humans and superoptimizers are good at this) and obtaining a correct and general implementation of that optimization in a production compiler (this is very hard, even for experts)
https://users.cs.utah.edu/~regehr/generalization-oopsla24.pdf
this is my student Manasij's thesis work. I'll just throw modesty to the wind and say that this is a really important paper for how we'll build compilers in the future!