EricWF wrote: What effect does this have on the benchmarks? If we're getting the optimization, I would expect this to be observable. Also, was the lines of assembly in a debug or optimized build configurations? https://github.com/llvm/llvm-project/pull/97201