tblah wrote: Maybe we could just tweak the MLIR pipeline for the most obvious cases (LoopVersioning, Inlining) and be extremely conservative about anything with an uncertain payoff. https://github.com/llvm/llvm-project/pull/164707