gh-155503: Add more PyType C API tests

[vstinner]

Add tests on functions:

• PyType_ClearCache()
• PyType_GetFlags()
• PyType_IsSubtype()
• PyType_Ready()

Move PyType limited C API tests from _testcapi to _testlimitedcapi. Add a new Modules/_testlimitedcapi/type.c file.

• Issue: C API: Add more tests on PyType functions #155503

[vstinner]

Keep PyType_GetSlot() test in _testcapi since it does access directly PyTypeObject members (excluded from the limited C API).

[vstinner]

cc @serhiy-storchaka

[serhiy-storchaka]

Add tests for non-type arguments and NULL.

[serhiy-storchaka]

Absolutely not. You should not check the type of the argument. Use NULLABLE() instead.

If the tested API accepts non-type or NULL -- fine, you will add a test. If it crashes -- just add the commented out call to denote that such corner case was considered, but cannot be tested.

[serhiy-storchaka]

Do not enforce the type of arguments. Add tests for non-type arguments.

[vstinner]

@serhiy-storchaka: I addressed your review. Please review the updated PR.

If the tested API accepts non-type or NULL -- fine, you will add a test. If it crashes -- just add the commented out call to denote that such corner case was considered, but cannot be tested.

Ok, I modified the test wrappers to treat None as NULL. I added "CRASHES func(NULL)" comments: all tested PyType functions crash when NULL is passed.

Add tests for non-type arguments and NULL.

The PyType C API is special: the first argument type is PyTypeObject*, not PyObject*. If you pass a non-type, it will likely crash (or worse :-)), I don't see the purpose of testing that.

If you call a PyType C API with a PyObject*, the compiler fails with an error. Example with return PyType_GetName((PyObject*)type);:

./Modules/_testlimitedcapi/type.c: In function 'get_type_name':
./Modules/_testlimitedcapi/type.c:55:27: error: passing argument 1 of 'PyType_GetName' from incompatible pointer type [-Wincompatible-pointer-types]
   55 |     return PyType_GetName((PyObject*)type);

[serhiy-storchaka]

Remove type checks. If non-types are not supported, just do not pass non-types in tests. If they are supported, they could be tested. We should not accidentally test a TypeError raised by the wrapper.

For example, it looks to me that PyType_IsSubtype(a, b) works if b is a non-type or even NULL (the code only compares pointers).

[vstinner]

Remove type checks. If non-types are not supported, just do not pass non-types in tests. If they are supported, they could be tested. We should not accidentally test a TypeError raised by the wrapper.

I removed PyType_Check() checks.

For example, it looks to me that PyType_IsSubtype(a, b) works if b is a non-type or even NULL (the code only compares pointers).

I don't think that we should test this in test_capi. It looks like an implementation detail accident, not something that we want to test.

[serhiy-storchaka]

We should test implementation detail accidents in test_capi, this is one of purposes of the tests. There is a large chance that there is a code that depends on this, even if by accident. If we change that detail in future, we should be aware what we do and properly document the change. You can gate it as CPython implementation detail.

[vstinner]

We should test implementation detail accidents in test_capi, this is one of purposes of the tests. There is a large chance that there is a code that depends on this, even if by accident. If we change that detail in future, we should be aware what we do and properly document the change. You can gate it as CPython implementation detail.

I disagree. As I wrote previously, if you pass a PyObject* to PyType_Is_Subclass(), you get a compiler error. If a project pass (PyTypeObject*)b as the second argument of PyType_Is_Subclass() where b is a PyObject*, IMO it's a bug in the project, not something that we should support.

[vstinner]

I added "CRASHES ...: argument must be a type" comments.