I know it is irrational - as compilers already perform optimizations that are non intuitive - but my lizard brain really doesn’t like the idea of ML possibly changing semantics at a layer that is unobservable.
1. Decisions that have no semantic change in the program but affect performance. This is things like code layout or register allocation that will be more cache friendly. This is what I expect to show up in industrial optimizing compilers more and more over the next bunch of years.
2. Decisions where you can prove that the new code sequence is semantically identical to the original code sequence via some formal technique. This is basically the same as various superoptimization techniques but with a new search strategy and therefore no new concerns about correctness. I'd expect this to remain niche and only used offline to generate particularly optimized sequences in extremely hot and small code paths.
Yeah the problem is that there are not a whole lot of those in PL.
That's quite a tall order. Godel and all that.