Well, zero-cost abstractions are still abstractions. It’s not junk per-se, but things that will be optimized out if the IR has enough information to safely do so, so basically lots of extra metadata to actually prove to LLVM that these things are safe.