python · brandtbucher · Oct 15, 2020 · Oct 9, 2020 · Oct 9, 2020 · Oct 13, 2020
diff --git a/Lib/test/test_finalization.py b/Lib/test/test_finalization.py
@@ -16,6 +16,15 @@ def __new__(cls, *args, **kwargs):
                 raise TypeError('requires _testcapi.with_tp_del')
         return C
 
+try:
+    from _testcapi import without_gc
+except ImportError:
+    def without_gc(cls):
+        class C:
+            def __new__(cls, *args, **kwargs):
+                raise TypeError('requires _testcapi.without_gc')
+        return C
+
 from test import support
 
 
@@ -94,9 +103,11 @@ def check_sanity(self):
         assert self.id_ == id(self)
 
 
+@without_gc
 class NonGC(NonGCSimpleBase):
     __slots__ = ()
 
+@without_gc
 class NonGCResurrector(NonGCSimpleBase):
     __slots__ = ()
 
@@ -109,8 +120,14 @@ def side_effect(self):
 class Simple(SimpleBase):
     pass
 
-class SimpleResurrector(NonGCResurrector, SimpleBase):
-    pass
+# Can't inherit from NonGCResurrector, in case importing without_gc fails.
+class SimpleResurrector(SimpleBase):
+
+    def side_effect(self):
+        """
+        Resurrect self by storing self in a class-wide list.
+        """
+        self.survivors.append(self)
 
 
 class TestBase:
@@ -178,6 +195,7 @@ def test_simple_resurrect(self):
             self.assert_survivors([])
         self.assertIs(wr(), None)
 
+    @support.cpython_only
     def test_non_gc(self):
         with SimpleBase.test():
             s = NonGC()
@@ -191,6 +209,7 @@ def test_non_gc(self):
             self.assert_del_calls(ids)
             self.assert_survivors([])
 
+    @support.cpython_only
     def test_non_gc_resurrect(self):
         with SimpleBase.test():
             s = NonGCResurrector()

diff --git a/Lib/test/test_gc.py b/Lib/test/test_gc.py
@@ -582,9 +582,9 @@ class UserIntSlots(int):
         self.assertTrue(gc.is_tracked(UserInt()))
         self.assertTrue(gc.is_tracked([]))
         self.assertTrue(gc.is_tracked(set()))
-        self.assertFalse(gc.is_tracked(UserClassSlots()))
-        self.assertFalse(gc.is_tracked(UserFloatSlots()))
-        self.assertFalse(gc.is_tracked(UserIntSlots()))
+        self.assertTrue(gc.is_tracked(UserClassSlots()))
+        self.assertTrue(gc.is_tracked(UserFloatSlots()))
+        self.assertTrue(gc.is_tracked(UserIntSlots()))
 
     def test_is_finalized(self):
         # Objects not tracked by the always gc return false

diff --git a/Misc/NEWS.d/next/Core and Builtins/2020-10-14-16-19-43.bpo-41984.SEtKMr.rst b/Misc/NEWS.d/next/Core and Builtins/2020-10-14-16-19-43.bpo-41984.SEtKMr.rst
@@ -0,0 +1,2 @@
+The garbage collector now tracks all user-defined classes. Patch by Brandt
+Bucher.
diff --git a/Modules/_testcapimodule.c b/Modules/_testcapimodule.c
@@ -3888,6 +3888,25 @@ with_tp_del(PyObject *self, PyObject *args)
     return obj;
 }
 
+static PyObject *
+without_gc(PyObject *Py_UNUSED(self), PyObject *obj)
+{
+    PyTypeObject *tp = (PyTypeObject*)obj;
+    if (!PyType_Check(obj) || !PyType_HasFeature(tp, Py_TPFLAGS_HEAPTYPE)) {
+        return PyErr_Format(PyExc_TypeError, "heap type expected, got %R", obj);
+    }
+    if (PyType_IS_GC(tp)) {
+        // Don't try this at home, kids:
+        tp->tp_flags -= Py_TPFLAGS_HAVE_GC;
+        tp->tp_free = PyObject_Del;
+        tp->tp_traverse = NULL;
+        tp->tp_clear = NULL;
+    }
+    assert(!PyType_IS_GC(tp));
+    Py_INCREF(obj);
+    return obj;
+}
+
 static PyMethodDef ml;
 
 static PyObject *
@@ -5805,6 +5824,7 @@ static PyMethodDef TestMethods[] = {
     {"meth_fastcall", (PyCFunction)(void(*)(void))meth_fastcall, METH_FASTCALL},
     {"meth_fastcall_keywords", (PyCFunction)(void(*)(void))meth_fastcall_keywords, METH_FASTCALL|METH_KEYWORDS},
     {"pynumber_tobase", pynumber_tobase, METH_VARARGS},
+    {"without_gc", without_gc, METH_O},
     {NULL, NULL} /* sentinel */
 };
 

diff --git a/Objects/typeobject.c b/Objects/typeobject.c
@@ -2612,10 +2612,10 @@ type_new(PyTypeObject *metatype, PyObject *args, PyObject *kwds)
     slots = NULL;
 
     /* Initialize tp_flags */
+    // All heap types need GC, since we can create a reference cycle by storing
+    // an instance on one of its parents:
     type->tp_flags = Py_TPFLAGS_DEFAULT | Py_TPFLAGS_HEAPTYPE |
-        Py_TPFLAGS_BASETYPE;
-    if (base->tp_flags & Py_TPFLAGS_HAVE_GC)
-        type->tp_flags |= Py_TPFLAGS_HAVE_GC;
+        Py_TPFLAGS_BASETYPE | Py_TPFLAGS_HAVE_GC;
 
     /* Initialize essential fields */
     type->tp_as_async = &et->as_async;
@@ -2815,21 +2815,11 @@ type_new(PyTypeObject *metatype, PyObject *args, PyObject *kwds)
     }
     type->tp_dealloc = subtype_dealloc;
 
-    /* Enable GC unless this class is not adding new instance variables and
-       the base class did not use GC. */
-    if ((base->tp_flags & Py_TPFLAGS_HAVE_GC) ||
-        type->tp_basicsize > base->tp_basicsize)
-        type->tp_flags |= Py_TPFLAGS_HAVE_GC;
-
     /* Always override allocation strategy to use regular heap */
     type->tp_alloc = PyType_GenericAlloc;
-    if (type->tp_flags & Py_TPFLAGS_HAVE_GC) {
-        type->tp_free = PyObject_GC_Del;
-        type->tp_traverse = subtype_traverse;
-        type->tp_clear = subtype_clear;
-    }
-    else
-        type->tp_free = PyObject_Del;
+    type->tp_free = PyObject_GC_Del;
+    type->tp_traverse = subtype_traverse;
+    type->tp_clear = subtype_clear;
 
     /* store type in class' cell if one is supplied */
     cell = _PyDict_GetItemIdWithError(dict, &PyId___classcell__);
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1,2 @@
		The garbage collector now tracks all user-defined classes. Patch by Brandt
Copy link Member pablogsal Oct 15, 2020 Choose a reason for hiding this comment The reason will be displayed to describe this comment to others. Learn more. I doubt that people are relying on the past behaviour (but I have been surprised in the past) so I am questioning if we should include a small sentence in the What's new of 3.10 in the "porting" section. What do you think? Copy link Member Author brandtbucher Oct 15, 2020 • edited Loading Choose a reason for hiding this comment The reason will be displayed to describe this comment to others. Learn more. Maybe? I don't really feel the need to, though, since the behavior wasn't documented before and we don't provide any "fixes" to actually port old code. pablogsal reacted with thumbs up emoji
		Bucher.