gh-124153: Introduce PyType_GetBaseByToken function (PoC)

Doc/c-api/type.rst

@@ -264,6 +264,24 @@ Type Objects
 
    .. versionadded:: 3.11
 
+.. c:function:: int PyType_GetBaseByToken(PyTypeObject *type, void *token, PyTypeObject **result)
+
+   Find the first superclass in *type*'s :term:`method resolution order` whose
+   :c:macro:`Py_tp_token` token is equal to the given one.
+
+   * If found, set *\*result* to a new :term:`strong reference`
+     to it and return ``1``.
+   * If not found, set *\*result* to ``NULL`` and return ``0``.
+   * On error, set *\*result* to ``NULL`` and return ``-1`` with an
+     exception set.
+
+   The *result* argument may be ``NULL``, in which case *\*result* is not set.
+   Use this if you need only the return value.
+
+   The *token* argument may not be ``NULL``.
+
+   .. versionadded:: 3.14
+
 .. c:function:: int PyUnstable_Type_AssignVersionTag(PyTypeObject *type)
 
    Attempt to assign a version tag to the given type.
@@ -488,6 +506,11 @@ The following functions and structs are used to create
       * ``Py_nb_add`` to set :c:member:`PyNumberMethods.nb_add`
       * ``Py_sq_length`` to set :c:member:`PySequenceMethods.sq_length`
 
+      An additional slot is supported that does not correspond to a
+      :c:type:`!PyTypeObject` struct field:
+
+      * :c:data:`Py_tp_token`
+
       The following “offset” fields cannot be set using :c:type:`PyType_Slot`:
 
          * :c:member:`~PyTypeObject.tp_weaklistoffset`
@@ -538,4 +561,47 @@ The following functions and structs are used to create
       The desired value of the slot. In most cases, this is a pointer
       to a function.
 
-      Slots other than ``Py_tp_doc`` may not be ``NULL``.
+      *pfunc* values may not be ``NULL``, except for the following slots:
+
+      * ``Py_tp_doc``
+      * :c:data:`Py_tp_token` (for clarity, prefer :c:data:`Py_TP_USE_SPEC`
+        rather than ``NULL``)
+
+.. c:macro:: Py_tp_token
+
+   A :c:member:`~PyType_Slot.slot` that records a static memory layout ID
+   for a class.
+
+   If the :c:type:`PyType_Spec` of the class is statically
+   allocated, the token can be set to the spec using the special value
+   :c:data:`Py_TP_USE_SPEC`:
+
+   .. code-block:: c
+
+      static PyType_Slot foo_slots[] = {
+         {Py_tp_token, Py_TP_USE_SPEC},
+
+   It can also be set to an arbitrary pointer, but you must ensure that:
+
+   * The pointer outlives the class, so it's not reused for something else
+     while the class exists.
+   * It "belongs" to the extension module where the class lives, so it will not
+     clash with other extensions.
+
+   Use :c:func:`PyType_GetBaseByToken` to check if a class's superclass has
+   a given token -- that is, check whether the memory layout is compatible.
+
+   To get the token for a given class (without considering superclasses),
+   use :c:func:`PyType_GetSlot` with ``Py_tp_token``.
+
+   .. versionadded:: 3.14
+
+   .. c:namespace:: NULL
+
+   .. c:macro:: Py_TP_USE_SPEC
+
+      Used as a value with :c:data:`Py_tp_token` to set the token to the
+      class's :c:type:`PyType_Spec`.
+      Expands to ``NULL``.
+
+      .. versionadded:: 3.14

Doc/whatsnew/3.14.rst

@@ -554,6 +554,9 @@ New Features
 
   (Contributed by Victor Stinner in :gh:`107954`.)
 
+* Add :c:func:`PyType_GetBaseByToken` and :c:data:`Py_tp_token` slot for easier
+  superclass identification, which attempts to resolve the `type checking issue
+  <https://peps.python.org/pep-0630/#type-checking>`__ mentioned in :pep:`630`.
 
 Porting to Python 3.14
 ----------------------

Include/cpython/object.h

@@ -269,6 +269,7 @@ typedef struct _heaptypeobject {
     struct _dictkeysobject *ht_cached_keys;
     PyObject *ht_module;
     char *_ht_tpname;  // Storage for "tp_name"; see PyType_FromModuleAndSpec
+    void *ht_token;  // Storage for the "Py_tp_token" slot
     struct _specialization_cache _spec_cache; // For use by the specializer.
 #ifdef Py_GIL_DISABLED
     Py_ssize_t unique_id;  // ID used for thread-local refcounting

Include/object.h

@@ -391,6 +391,10 @@ PyAPI_FUNC(PyObject *) PyType_FromMetaclass(PyTypeObject*, PyObject*, PyType_Spe
 PyAPI_FUNC(void *) PyObject_GetTypeData(PyObject *obj, PyTypeObject *cls);
 PyAPI_FUNC(Py_ssize_t) PyType_GetTypeDataSize(PyTypeObject *cls);
 #endif
+#if !defined(Py_LIMITED_API) || Py_LIMITED_API+0 >= 0x030E0000
+PyAPI_FUNC(int) PyType_GetBaseByToken(PyTypeObject *, void *, PyTypeObject **);
+#define Py_TP_USE_SPEC NULL
+#endif
 
 /* Generic type check */
 PyAPI_FUNC(int) PyType_IsSubtype(PyTypeObject *, PyTypeObject *);

Include/typeslots.h

@@ -90,3 +90,7 @@
 /* New in 3.14 */
 #define Py_tp_vectorcall 82
 #endif
+#if !defined(Py_LIMITED_API) || Py_LIMITED_API+0 >= 0x030E0000
+/* New in 3.14 */
+#define Py_tp_token 83
+#endif

Lib/test/test_capi/test_misc.py

@@ -1144,6 +1144,77 @@ class MyType:
         MyType.__module__ = 123
         self.assertEqual(get_type_fullyqualname(MyType), 'my_qualname')
 
+    def test_get_base_by_token(self):
+        def get_base_by_token(src, key, comparable=True):
+            def run(use_mro):
+                find_first = _testcapi.pytype_getbasebytoken
+                ret1, result = find_first(src, key, use_mro, True)
+                ret2, no_result = find_first(src, key, use_mro, False)
+                self.assertIn(ret1, (0, 1))
+                self.assertEqual(ret1, result is not None)
+                self.assertEqual(ret1, ret2)
+                self.assertIsNone(no_result)
+                return result
+
+            found_in_mro = run(True)
+            found_in_bases = run(False)
+            if comparable:
+                self.assertIs(found_in_mro, found_in_bases)
+                return found_in_mro
+            return found_in_mro, found_in_bases
+
+        create_type = _testcapi.create_type_with_token
+        get_token = _testcapi.get_tp_token
+
+        Py_TP_USE_SPEC = _testcapi.Py_TP_USE_SPEC
+        self.assertEqual(Py_TP_USE_SPEC, 0)
+
+        A1 = create_type('_testcapi.A1', Py_TP_USE_SPEC)
+        self.assertTrue(get_token(A1) != Py_TP_USE_SPEC)
+
+        B1 = create_type('_testcapi.B1', id(self))
+        self.assertTrue(get_token(B1) == id(self))
+
+        tokenA1 = get_token(A1)
+        # find A1 from A1
+        found = get_base_by_token(A1, tokenA1)
+        self.assertIs(found, A1)
+
+        # no token in static types
+        STATIC = type(1)
+        self.assertEqual(get_token(STATIC), 0)
+        found = get_base_by_token(STATIC, tokenA1)
+        self.assertIs(found, None)
+
+        # no token in pure subtypes
+        class A2(A1): pass
+        self.assertEqual(get_token(A2), 0)
+        # find A1
+        class Z(STATIC, B1, A2): pass
+        found = get_base_by_token(Z, tokenA1)
+        self.assertIs(found, A1)
+
+        # searching for NULL token is an error
+        with self.assertRaises(SystemError):
+            get_base_by_token(Z, 0)
+        with self.assertRaises(SystemError):
+            get_base_by_token(STATIC, 0)
+
+        # share the token with A1
+        C1 = create_type('_testcapi.C1', tokenA1)
+        self.assertTrue(get_token(C1) == tokenA1)
+
+        # find C1 first by shared token
+        class Z(C1, A2): pass
+        found = get_base_by_token(Z, tokenA1)
+        self.assertIs(found, C1)
+        # B1 not found
+        found = get_base_by_token(Z, get_token(B1))
+        self.assertIs(found, None)
+
+        with self.assertRaises(TypeError):
+            _testcapi.pytype_getbasebytoken(
+                'not a type', id(self), True, False)
 
     def test_gen_get_code(self):
         def genf(): yield

Lib/test/test_sys.py

@@ -1718,7 +1718,7 @@ def delx(self): del self.__x
                   '3P'                  # PyMappingMethods
                   '10P'                 # PySequenceMethods
                   '2P'                  # PyBufferProcs
-                  '6P'
+                  '7P'
                   '1PIP'                # Specializer cache
                   + typeid              # heap type id (free-threaded only)
                   )

Misc/stable_abi.toml

@@ -2527,4 +2527,10 @@
 [function.PyLong_AsUInt64]
     added = '3.14'
 [const.Py_tp_vectorcall]
-    added = '3.14'
+    added = '3.14'
+[function.PyType_GetBaseByToken]
+    added = '3.14'
+[const.Py_tp_token]
+    added = '3.14'
+[const.Py_TP_USE_SPEC]
+    added = '3.14'

Modules/_collectionsmodule.c

@@ -2179,6 +2179,8 @@ typedef struct {
     PyObject *default_factory;
 } defdictobject;
 
+static PyType_Spec defdict_spec;
+
 PyDoc_STRVAR(defdict_missing_doc,
 "__missing__(key) # Called by __getitem__ for missing key; pseudo-code:\n\
   if self.default_factory is None: raise KeyError((key,))\n\
@@ -2358,23 +2360,16 @@ defdict_or(PyObject* left, PyObject* right)
 {
     PyObject *self, *other;
 
-    // Find module state
-    PyTypeObject *tp = Py_TYPE(left);
-    PyObject *mod = PyType_GetModuleByDef(tp, &_collectionsmodule);
-    if (mod == NULL) {
-        PyErr_Clear();
-        tp = Py_TYPE(right);
-        mod = PyType_GetModuleByDef(tp, &_collectionsmodule);
+    int ret = PyType_GetBaseByToken(Py_TYPE(left), &defdict_spec, NULL);
+    if (ret < 0) {
+        return NULL;
     }
-    assert(mod != NULL);
-    collections_state *state = get_module_state(mod);
-
-    if (PyObject_TypeCheck(left, state->defdict_type)) {
+    if (ret) {
         self = left;
         other = right;
     }
     else {
-        assert(PyObject_TypeCheck(right, state->defdict_type));
+        assert(PyType_GetBaseByToken(Py_TYPE(right), &defdict_spec, NULL) == 1);
         self = right;
         other = left;
     }
@@ -2454,6 +2449,7 @@ passed to the dict constructor, including keyword arguments.\n\
 #define DEFERRED_ADDRESS(ADDR) 0
 
 static PyType_Slot defdict_slots[] = {
+    {Py_tp_token, Py_TP_USE_SPEC},
     {Py_tp_dealloc, defdict_dealloc},
     {Py_tp_repr, defdict_repr},
     {Py_nb_or, defdict_or},

Modules/_ctypes/_ctypes.c

@@ -500,7 +500,7 @@ CType_Type_dealloc(PyObject *self)
 {
     StgInfo *info = _PyStgInfo_FromType_NoState(self);
     if (!info) {
-        PyErr_WriteUnraisable(self);
+        PyErr_WriteUnraisable(NULL);  // NULL avoids segfault here
     }
     if (info) {
         PyMem_Free(info->ffi_type_pointer.elements);
@@ -560,6 +560,7 @@ static PyMethodDef ctype_methods[] = {
 };
 
 static PyType_Slot ctype_type_slots[] = {
+    {Py_tp_token, Py_TP_USE_SPEC},
     {Py_tp_traverse, CType_Type_traverse},
     {Py_tp_clear, CType_Type_clear},
     {Py_tp_dealloc, CType_Type_dealloc},
@@ -569,7 +570,7 @@ static PyType_Slot ctype_type_slots[] = {
     {0, NULL},
 };
 
-static PyType_Spec pyctype_type_spec = {
+PyType_Spec pyctype_type_spec = {
     .name = "_ctypes.CType_Type",
     .basicsize = -(Py_ssize_t)sizeof(StgInfo),
     .flags = (Py_TPFLAGS_DEFAULT | Py_TPFLAGS_IMMUTABLETYPE |

Modules/_ctypes/ctypes.h

@@ -108,6 +108,7 @@ get_module_state_by_def(PyTypeObject *cls)
 }
 
 
+extern PyType_Spec pyctype_type_spec;
 extern PyType_Spec carg_spec;
 extern PyType_Spec cfield_spec;
 extern PyType_Spec cthunk_spec;
@@ -488,18 +489,48 @@ PyStgInfo_FromAny(ctypes_state *state, PyObject *obj, StgInfo **result)
     return _stginfo_from_type(state, Py_TYPE(obj), result);
 }
 
+static inline void
+exercise_get_base_by_token(PyTypeObject *PyCType_Type)
+{
+    // CType_Type_traverse() almost alone trains PyType_GetBaseByToken() when
+    // running PGO tests (130,000 calls, whereas PyType_GetModuleByDef() has
+    // been trained with 2,700,000 calls), which is invoked by gc.collect().
+    // Its way of finding a superclass is currently biased, so the following
+    // should be complemented. Expects little perf impact on this module.
+    PyTypeObject *result;
+
+    // Find the type itself by token
+    if (PyType_GetBaseByToken(PyCType_Type, &pyctype_type_spec, &result) == 1) {
+        Py_DECREF(result);
+    }
+    assert(result == PyCType_Type);
+
+    // Reject a static type
+    if (PyType_GetBaseByToken(&PyBaseObject_Type, &PyBaseObject_Type, &result)) {
+        assert(0);
+    }
+}
+
 /* A variant of PyStgInfo_FromType that doesn't need the state,
  * so it can be called from finalization functions when the module
- * state is torn down. Does no checks; cannot fail.
- * This inlines the current implementation PyObject_GetTypeData,
- * so it might break in the future.
+ * state is torn down.
  */
 static inline StgInfo *
 _PyStgInfo_FromType_NoState(PyObject *type)
 {
-    size_t type_basicsize =_Py_SIZE_ROUND_UP(PyType_Type.tp_basicsize,
-                                             ALIGNOF_MAX_ALIGN_T);
-    return (StgInfo *)((char *)type + type_basicsize);
+    PyTypeObject *PyCType_Type;
+    if (PyType_GetBaseByToken(Py_TYPE(type), &pyctype_type_spec, &PyCType_Type) < 0) {
+        return NULL;
+    }
+    if (PyCType_Type == NULL) {
+        PyErr_Format(PyExc_TypeError, "expected a ctypes type, got '%N'", type);
+        return NULL;
+    }
+    exercise_get_base_by_token(PyCType_Type);
+
+    StgInfo *info = PyObject_GetTypeData(type, PyCType_Type);
+    Py_DECREF(PyCType_Type);
+    return info;
 }
 
 // Initialize StgInfo on a newly created type