khaki3 wrote: I think the regression is build noise. The erase_if only runs for a value already holding 1024 entries, which CTMark never reaches. Could you re-run the benchmark with 65db5ba672d4? https://github.com/llvm/llvm-project/pull/218438