This comment is tangential at best, but coming from someone who has never written a compiler (i.e. I'm a Dunning-Kruger peak), I'm curious to know if ML approaches could drive more optimizations. As I understand it, compiler optimizations are NP-hard so they use heuristics and clever ways to constrain the search space for optimizations, which I would think could be potentially guided by ML?

I think that there would be a lot of cases where engineers would accept 10x longer compile times if it gave 10% more application performance.