MaxGraey wrote: The benchmarks have shown some benefits. The question is whether it’s worth making the code a bit more complex. WDYT? PR is ready for review. https://github.com/llvm/llvm-project/pull/191919