gh-131311: Fix memory leak in PyCStructUnionType_update_stginfo

[sergey-miryanov]

Fix memory leak in PyCStructUnionType_update_stginfo on fail path.

• Issue: Possible memory leak in _ctypes/PyCStructUnionType_update_stginfo on fail path #131311

[picnixz]

In order to locate the error label I needed to expand the entire file. This shows that this function should be probably splitted into multiple steps if possible, or that the if () at the end should be separate.

WDYT @encukou? any reason to keep a single large function?

[sergey-miryanov]

@picnixz Thanks for explanation.

I will try to explain further issues better. About this function - I fully agree that it has large complexity, and it is better to split it.

[encukou]

Thank you for the fix!

Since the function has the error label for cleanup, could you instead make layout_func use that?
(Declare it at the beginning with = NULL, XDECREF it in error, and Py_CLEAR instead of Py_DECREF in the main code path?)

(If you want to go further, you can do the same for type_block, and any other similar local.)

---

this function should be probably split into multiple steps if possible, or the if () at the end should be separate

Yes (though I don't think it's a priority).
We could do with a helper that combines PyObject_GetAttr + PyLong_AsSsize_t.
For format_spec, I think it might be worth a try to make StgInfo.format a Python string. (A bigger refactoring, but it seems PyUnicodeWriter would be a win over juggling strcats.)
The if can be split out of course.

[sergey-miryanov]

OK, will move cleanup to error section.

---

Would you like to me to open a new PR to split this function and maybe another one for StgInfo.format?

[sergey-miryanov]

I moved cleanup of locals to the error section except of type_block. I prefer to extract this

cpython/Modules/_ctypes/stgdict.c

Line 444 in a09c2b2

if (arrays_seen && (total_size <= MAX_STRUCT_SIZE)) {

to other function and do cleanup of type_block there. Should I do it in this PR or add new one?

[encukou]

Thank you for the fix!
Please open a new PR for the follow-ups.