================ ---------------- folkertdev wrote: Here the new approach is apparently worse in terms of instruction count. Based on some grepping a big difference is the number of spills. Anything we can/should do about that? https://github.com/llvm/llvm-project/pull/204802