The memory operand version tends to be as slow or slower than the manual implementation, so LLVM is right to avoid it.