We motivate using ML policies to replace fine-grained compiler optimization decisions, otherwise made by heuristics, and discuss:

• the current infrastructure available in the main repository of an open source compiler, Clang/LLVM;
• lessons learned from the practical application of ML on two optimizations over the last few years: a register allocation policy, and a size-optimizing inliner policy, both trained with an open source solution, on both internal and open-source projects, such as Chrome and Fuchsia;
• what we discovered as obstacles to the broader application of ML optimizations, and to broader community adoption.

We'll conclude with current status and investigation directions.