sebpop wrote: https://github.com/llvm/llvm-project/pull/170519 does batch delinearization for DA and caches the results. Delinearization results are more precise because delinearization happens with more context. https://github.com/llvm/llvm-project/pull/164379