This re-implements the shl_with_overflow operation from scratch,
making it a lot more robust and outputs the equal code to the LLVM backend.