Ref: comment chain at https://github.com/python/cpython/pull/27008#discussion_r663450441.

setup_ga fails only if PyTuple_Pack fails, which usually happens when Python is out of memory. The cleanup code for setup_ga shouldn't use PyObject_GC_DEL as mentioned in Pablo's comment above.

I don't know how to add a test for out of memory, so I'll send a PR without test soon.

New changeset d33943a6c368c2184e238019c63ac7a267da5594 by Ken Jin in branch 'main':
bpo-44562: Remove invalid PyObject_GC_Del from error path of types.GenericAlias … (GH-27016)
https://github.com/python/cpython/commit/d33943a6c368c2184e238019c63ac7a267da5594

There is the same bug in ga_new(): tp_free is called directly.

And is is correct to call deallocator for not initialized GenericAlias object?

New changeset 68330b681a4b63cedad58fcfd1d9573209bad21d by Miss Islington (bot) in branch '3.10':
bpo-44562: Remove invalid PyObject_GC_Del from error path of types.GenericAlias … (GH-27016)
https://github.com/python/cpython/commit/68330b681a4b63cedad58fcfd1d9573209bad21d

New changeset 4684a34c8d2a2ffac7b779edb4ba57f043783478 by Miss Islington (bot) in branch '3.9':
bpo-44562: Remove invalid PyObject_GC_Del from error path of types.GenericAlias … (GH-27016) (GH-27018)
https://github.com/python/cpython/commit/4684a34c8d2a2ffac7b779edb4ba57f043783478

> And is is correct to call deallocator for not initialized GenericAlias object?

The correct thing to do is to call Py_DECREF so the object is properly unlinked and the other potential cleanups work. Calling the deallocator directly is almost always dangerous.

But is alias->origin initialized to safely call Py_XDECREF(alias->origin)?

> But is alias->origin initialized to safely call Py_XDECREF(alias->origin)?

Why would that be unsafe? PyType_GenericAlloc sets all fields to NULL so that is going to either be NULL or have an object, no? setup_ga only sets alias->origin if there has been no failures

AFAIK PyType_GenericAlloc is used indirectly in ga_new(), but not in Py_GenericAlias().

> AFAIK PyType_GenericAlloc is used indirectly in ga_new(), but not in Py_GenericAlias().

Oh, I see what you mean, that's a good point.

New changeset fe847a62852c3baaec6c97a5e2e7b2e66732bdb8 by Pablo Galindo in branch '3.9':
Revert "bpo-44562: Remove invalid PyObject_GC_Del from error path of types.GenericAlias … (GH-27016) (GH-27018)" (GH-27022)
https://github.com/python/cpython/commit/fe847a62852c3baaec6c97a5e2e7b2e66732bdb8

> Revert "bpo-44562: Remove invalid PyObject_GC_Del from error path of types.GenericAlias … (GH-27016) (GH-27018)" (GH-27022)

That's a good idea :-)

The reverted commit caused many segfaults on macOS:
https://buildbot.python.org/all/#/builders/108/builds/122

4 tests failed:
    test_cmd_line_script test_multiprocessing_main_handling
    test_tarfile test_trace

New changeset b324c4c5f763c5116a97db8591e6dcb94456570a by Pablo Galindo in branch 'main':
bpo-44562: Use PyType_GenericAlloc in Py_GenericAlias (GH-27021)
https://github.com/python/cpython/commit/b324c4c5f763c5116a97db8591e6dcb94456570a

New changeset d17cc1ff9fe82b17bbe589b83e440959c8f135d7 by Miss Islington (bot) in branch '3.10':
bpo-44562: Use PyType_GenericAlloc in Py_GenericAlias (GH-27021)
https://github.com/python/cpython/commit/d17cc1ff9fe82b17bbe589b83e440959c8f135d7

New changeset 51a29c42f10bd9368db9a21f2f63319be2e30b95 by Ken Jin in branch '3.9':
[3.9] bpo-44562: Remove invalid PyObject_GC_Del from error path of types.GenericAlias … (GH-27016) (GH-27028)
https://github.com/python/cpython/commit/51a29c42f10bd9368db9a21f2f63319be2e30b95