Except in some rare edge cases, it’s mostly the latter, indirectly: in the average crate the vast majority of the time is spent in LLVM optimization passes and linking. Sometimes IR generation gets a pretty high score, but that’s somewhat inconsistent.