gh-128965: pickle load_build function checks if state is None, not False

[Legoclones]

Inside of the load_build() function for pickle's BUILD opcode, the C accelerator at one point checks if state is Py_None, while the Python version only checks if state.

cpython/Modules/_pickle.c

Line 6638 in 34ded1a

if (state != Py_None) {

cpython/Lib/pickle.py

Line 1765 in 34ded1a

if state:

This means if state is something like an empty dictionary or tuple, the code block under the if statement WILL be run in _pickle.c, but NOT in pickle.py.

As an example, the bytestream b']]b.' has the following disassembly:

    0: ]    EMPTY_LIST
    1: ]    EMPTY_LIST
    2: b    BUILD
    3: .    STOP
highest protocol among opcodes = 1

This will do nothing in pickle.py but error out in _pickle.c with the message state is not a dictionary. The easy solution is to change if state to if state != None, and it shouldn't break any existing functionality. I've attached a pull request.

• Issue: pickle load_build function checks if state is None, not False #128965

[picnixz]

Can we have tests that check this code path? Namely check that non-dict objects raise a TypeError and that falsey dicts do not raise exceptions? In addition, we need a NEWS entry for the bugfix, e.g:

Fix :meth:`!pickle._Pickler.load_build` for non-dictionary states.

(My NEWS entry is definitely poorly worded but I don't have a better suggestion for now. Maybe you can come up with a better formulation).

[Legoclones]

@picnixz I went ahead and added some tests and a NEWS entry. The problem here isn't that load_build is not detecting non-dictionary states. The problem is that if the inst argument of load_build is not a valid object (aka, doesn't have the __dict__ attribute), then a falsey state attribute won't cause it error out in Python, but the C accelerator does. The C accelerator checks if state != None, but Python only checks if it's falsey. By changing the check in Python to state is not None, then when a falsey state argument is provided, inst is actually checked to see if it's a valid object and will error now because it's not.

[Legoclones]

Okay, I'm not sure how to specify in the test that I want the Python version of the loads() function/Unpickler, not the C version. The loads() function of PyUnpicklerTests appears to use the Python version since it references pickle._Unpickler, but several of the tests were erroring out because they were using the C version (which throws an UnpicklingError instead of an AttributeError). Also, the C version was working fine, but the Python version is what's being changed here, so I want to make sure the tests are running on the Python version. What am I missing? I tried _loads but that doesn't exist in PyUnpicklerTests.