gh-152235: Defer GC tracking during the lock-free fill in set.__init__

[corona10]

• Issue: Need to defer GC tracking of frozenset/set construction #152235

[corona10]

@vstinner@methane

Final PR of the set: list already has TSan-safe initialization, and this brings set.init up to the same level of protection.

cpython/Objects/listobject.c

Lines 3523 to 3544 in fe9c721

	list___init___impl(PyListObject*self, PyObject*iterable)
	/*[clinic end generated code: output=0f3c21379d01de48 input=b3f3fe7206af8f6b]*/
	{
	/* Verify list invariants established by PyType_GenericAlloc() */
	assert(0 <= Py_SIZE(self));
	assert(Py_SIZE(self) <= self->allocated||self->allocated==-1);
	assert(self->ob_item!=NULL||
	self->allocated==0||self->allocated==-1);
	/* Empty previous contents */
	if (self->ob_item!=NULL) {
	Py_BEGIN_CRITICAL_SECTION(self);
	list_clear(self);
	Py_END_CRITICAL_SECTION();
	}
	if (iterable!=NULL) {
	if (_list_extend(self, iterable) <0) {
	return-1;
	}
	}
	return0;
	}

cpython/Objects/listobject.c

Lines 1457 to 1511 in fe9c721

	_list_extend(PyListObject*self, PyObject*iterable)
	{
	// Special case:
	// lists and tuples which can use PySequence_Fast ops
	intres=-1;
	if ((PyObject*)self==iterable) {
	Py_BEGIN_CRITICAL_SECTION(self);
	res=list_inplace_repeat_lock_held(self, 2);
	Py_END_CRITICAL_SECTION();
	}
	elseif (PyList_CheckExact(iterable)) {
	Py_BEGIN_CRITICAL_SECTION2(self, iterable);
	res=list_extend_lock_held(self, iterable);
	Py_END_CRITICAL_SECTION2();
	}
	elseif (PyTuple_CheckExact(iterable)) {
	Py_BEGIN_CRITICAL_SECTION(self);
	res=list_extend_lock_held(self, iterable);
	Py_END_CRITICAL_SECTION();
	}
	elseif (PyAnySet_CheckExact(iterable)) {
	Py_BEGIN_CRITICAL_SECTION2(self, iterable);
	res=list_extend_set(self, (PySetObject*)iterable);
	Py_END_CRITICAL_SECTION2();
	}
	elseif (PyDict_CheckExact(iterable)) {
	Py_BEGIN_CRITICAL_SECTION2(self, iterable);
	res=list_extend_dict(self, (PyDictObject*)iterable, 0/*keys*/);
	Py_END_CRITICAL_SECTION2();
	}
	elseif (Py_IS_TYPE(iterable, &PyDictKeys_Type)) {
	PyDictObject*dict= ((_PyDictViewObject*)iterable)->dv_dict;
	Py_BEGIN_CRITICAL_SECTION2(self, dict);
	res=list_extend_dict(self, dict, 0/*keys*/);
	Py_END_CRITICAL_SECTION2();
	}
	elseif (Py_IS_TYPE(iterable, &PyDictValues_Type)) {
	PyDictObject*dict= ((_PyDictViewObject*)iterable)->dv_dict;
	Py_BEGIN_CRITICAL_SECTION2(self, dict);
	res=list_extend_dict(self, dict, 1/*values*/);
	Py_END_CRITICAL_SECTION2();
	}
	elseif (Py_IS_TYPE(iterable, &PyDictItems_Type)) {
	PyDictObject*dict= ((_PyDictViewObject*)iterable)->dv_dict;
	Py_BEGIN_CRITICAL_SECTION2(self, dict);
	res=list_extend_dictitems(self, dict);
	Py_END_CRITICAL_SECTION2();
	}
	else {
	Py_BEGIN_CRITICAL_SECTION(self);
	res=list_extend_iter_lock_held(self, iterable);
	Py_END_CRITICAL_SECTION();
	}
	returnres;
	}

[vstinner]

The issue title is "Need to defer GC tracking of frozenset/set construction", but in fact, the problem only arises for frozenset.

PyFrozenSet_Type defines tp_new (frozenset_new) but has no tp_init function.

PySet_Type defines tp_new (set_new) and tp_init (set_init).

So changing set_init() only impacts the mutable set. I don't see which problem you are trying to solve. It's fine to mutate a set object.

[corona10]

Yeah we don't need to handle