imdj wrote: Based on initial results [from benchmarks](https://github.com/llvm/llvm-project/pull/130387) this implementation doesn't (yet) offer better performance. I will try to profile it and tweak few things. https://github.com/llvm/llvm-project/pull/129565