After was fixed, a test was created to reflect the fix (Lib/test/pickletester.py @ AbstractPickleModuleTests.test_bad_input). The test makes sure that an UnpicklingError is raised whenever pickled data attempts to pop on an empty stack. Although tests on AbstractPickleModuleTests normally only run on the C implementation of pickle (on which UnpicklingError is thrown), they may also run on the Python implementation if the C implementation of pickle is not available (which throws IndexError). I'm not sure if it's still reasonable to assume that it's possible for the C implementation of pickle not to exist (the python module does), as I've discovered this by forcefully disabling the C implementation. >>> pickle.Unpickler(io.BytesIO(b'0')).load() _pickle.UnpicklingError: unpickling stack underflow >>> pickle._Unpickler(io.BytesIO(b'0')).load() IndexError: list assignment index out of range The current fix involves adding IndexError to the list of exceptions that may be raised and moving the testcase to AbstractPickleTests (where the tests are ran against both the C and Python impl). As I'm working on improving pickle in protocol v4, it might be worth considering adding more consistency to the exceptions that are being raised by the Python unpickler.
It would be best to have the pickle tests always run against both the C and python code. We do this for other modules that have C versions of some or all of the Python code. And yes, making the two consistent is also good. Since pickle is generally not used without the C code, making the Python consistent with the C would seem to be a low-impact change.
