snehasish wrote: At Google we use debug information based AutoFDO. With this change to better resolve macro arguments, we measured a 0.3% improvement on a large internal workload. Thanks! cc: @amharc https://github.com/llvm/llvm-project/pull/163982