It seems like it should return the real bytecodes instead of the fast bytecodes that can be inserted by the interpreter.