What? If I understand correctly, the LLVM IR -> binary stage is shared. The Rust -> LLVM IR or C++ -> LLVM IR stages are obviously not shared, and these optimizations can happen there. Specifically C++ guarantees copy elision in some cases.