gh-131586: Avoid refcount contention in some "special" calls (#131588)

In the free threaded build, the `_PyObject_LookupSpecial()` call can lead to
reference count contention on the returned function object becuase it
doesn't use stackrefs. Refactor some of the callers to use
`_PyObject_MaybeCallSpecialNoArgs`, which uses stackrefs internally.

This fixes the scaling bottleneck in the "lookup_special" microbenchmark
in `ftscalingbench.py`. However, the are still some uses of
`_PyObject_LookupSpecial()` that need to be addressed in future PRs.

Author: colesbury

Include/internal/pycore_global_objects_fini_generated.h

@@ -89,6 +89,7 @@ struct _Py_global_strings {
         STRUCT_FOR_ID(__bytes__)
         STRUCT_FOR_ID(__call__)
         STRUCT_FOR_ID(__cantrace__)
+        STRUCT_FOR_ID(__ceil__)
         STRUCT_FOR_ID(__class__)
         STRUCT_FOR_ID(__class_getitem__)
         STRUCT_FOR_ID(__classcell__)
@@ -113,6 +114,7 @@ struct _Py_global_strings {
         STRUCT_FOR_ID(__file__)
         STRUCT_FOR_ID(__firstlineno__)
         STRUCT_FOR_ID(__float__)
+        STRUCT_FOR_ID(__floor__)
         STRUCT_FOR_ID(__floordiv__)
         STRUCT_FOR_ID(__format__)
         STRUCT_FOR_ID(__fspath__)
@@ -218,6 +220,7 @@ struct _Py_global_strings {
         STRUCT_FOR_ID(__subclasscheck__)
         STRUCT_FOR_ID(__subclasshook__)
         STRUCT_FOR_ID(__truediv__)
+        STRUCT_FOR_ID(__trunc__)
         STRUCT_FOR_ID(__type_params__)
         STRUCT_FOR_ID(__typing_is_unpacked_typevartuple__)
         STRUCT_FOR_ID(__typing_prepare_subst__)

Include/internal/pycore_global_strings.h

@@ -891,6 +891,12 @@ extern bool _PyObject_TryGetInstanceAttribute(PyObject *obj, PyObject *name,
 extern PyObject *_PyType_LookupRefAndVersion(PyTypeObject *, PyObject *,
                                              unsigned int *);
 
+// Internal API to look for a name through the MRO.
+// This stores a stack reference in out and returns the value of
+// type->tp_version or zero if name is missing. It doesn't set an exception!
+extern unsigned int
+_PyType_LookupStackRefAndVersion(PyTypeObject *type, PyObject *name, _PyStackRef *out);
+
 // Cache the provided init method in the specialization cache of type if the
 // provided type version matches the current version of the type.
 //
@@ -946,6 +952,14 @@ extern int _PyObject_IsInstanceDictEmpty(PyObject *);
 PyAPI_FUNC(PyObject*) _PyObject_LookupSpecial(PyObject *, PyObject *);
 PyAPI_FUNC(PyObject*) _PyObject_LookupSpecialMethod(PyObject *self, PyObject *attr, PyObject **self_or_null);
 
+// Calls the method named `attr` on `self`, but does not set an exception if
+// the attribute does not exist.
+PyAPI_FUNC(PyObject *)
+_PyObject_MaybeCallSpecialNoArgs(PyObject *self, PyObject *attr);
+
+PyAPI_FUNC(PyObject *)
+_PyObject_MaybeCallSpecialOneArg(PyObject *self, PyObject *attr, PyObject *arg);
+
 extern int _PyObject_IsAbstract(PyObject *);
 
 PyAPI_FUNC(int) _PyObject_GetMethod(PyObject *obj, PyObject *name, PyObject **method);

Include/internal/pycore_object.h

@@ -592,7 +592,7 @@ PyStackRef_XCLOSE(_PyStackRef ref)
 
 // Note: this is a macro because MSVC (Windows) has trouble inlining it.
 
-#define PyStackRef_Is(a, b) (((a).bits & (~Py_TAG_REFCNT)) == ((b).bits & (~Py_TAG_REFCNT)))
+#define PyStackRef_Is(a, b) (((a).bits & (~Py_TAG_BITS)) == ((b).bits & (~Py_TAG_BITS)))
 
 #endif // !defined(Py_GIL_DISABLED) && defined(Py_STACKREF_DEBUG)
 
@@ -640,6 +640,28 @@ PyStackRef_FunctionCheck(_PyStackRef stackref)
     return PyFunction_Check(PyStackRef_AsPyObjectBorrow(stackref));
 }
 
+static inline void
+_PyThreadState_PushCStackRef(PyThreadState *tstate, _PyCStackRef *ref)
+{
+#ifdef Py_GIL_DISABLED
+    _PyThreadStateImpl *tstate_impl = (_PyThreadStateImpl *)tstate;
+    ref->next = tstate_impl->c_stack_refs;
+    tstate_impl->c_stack_refs = ref;
+#endif
+    ref->ref = PyStackRef_NULL;
+}
+
+static inline void
+_PyThreadState_PopCStackRef(PyThreadState *tstate, _PyCStackRef *ref)
+{
+#ifdef Py_GIL_DISABLED
+    _PyThreadStateImpl *tstate_impl = (_PyThreadStateImpl *)tstate;
+    assert(tstate_impl->c_stack_refs == ref);
+    tstate_impl->c_stack_refs = ref->next;
+#endif
+    PyStackRef_XCLOSE(ref->ref);
+}
+
 #ifdef Py_GIL_DISABLED
 
 static inline int
@@ -656,6 +678,17 @@ _Py_TryIncrefCompareStackRef(PyObject **src, PyObject *op, _PyStackRef *out)
     return 0;
 }
 
+static inline int
+_Py_TryXGetStackRef(PyObject **src, _PyStackRef *out)
+{
+    PyObject *op = _Py_atomic_load_ptr_relaxed(src);
+    if (op == NULL) {
+        *out = PyStackRef_NULL;
+        return 1;
+    }
+    return _Py_TryIncrefCompareStackRef(src, op, out);
+}
+
 #endif
 
 // Like Py_VISIT but for _PyStackRef fields

Include/internal/pycore_runtime_init_generated.h

@@ -65,6 +65,16 @@ typedef union _PyStackRef {
 #endif
 } _PyStackRef;
 
+// A stackref that can be stored in a regular C local variable and be visible
+// to the GC in the free threading build.
+// Used in combination with _PyThreadState_PushCStackRef().
+typedef struct _PyCStackRef {
+    _PyStackRef ref;
+#ifdef Py_GIL_DISABLED
+    struct _PyCStackRef *next;
+#endif
+} _PyCStackRef;
+
 
 #ifdef __cplusplus
 }

Include/internal/pycore_stackref.h

@@ -47,8 +47,9 @@ typedef struct _PyThreadStateImpl {
     struct _qsbr_thread_state *qsbr;  // only used by free-threaded build
     struct llist_node mem_free_queue; // delayed free queue
 
-
 #ifdef Py_GIL_DISABLED
+    // Stack references for the current thread that exist on the C stack
+    struct _PyCStackRef *c_stack_refs;
     struct _gc_thread_state gc;
     struct _mimalloc_thread_state mimalloc;
     struct _Py_freelists freelists;

Include/internal/pycore_structs.h

@@ -383,6 +383,10 @@ def __len__(self):
             __bool__ = None
         self.assertRaises(TypeError, bool, B())
 
+        class C:
+            __len__ = None
+        self.assertRaises(TypeError, bool, C())
+
     def test_real_and_imag(self):
         self.assertEqual(True.real, 1)
         self.assertEqual(True.imag, 0)

Include/internal/pycore_tstate.h

@@ -1746,6 +1746,11 @@ def test_repr(self):
         a[0] = a
         self.assertEqual(repr(a), '{0: {...}}')
 
+    def test_repr_blocked(self):
+        class C:
+            __repr__ = None
+        self.assertRaises(TypeError, repr, C())
+
     def test_round(self):
         self.assertEqual(round(0.0), 0.0)
         self.assertEqual(type(round(0.0)), int)

Include/internal/pycore_unicodeobject_generated.h

@@ -573,6 +573,8 @@ def testFloor(self):
         #self.assertEqual(math.ceil(NINF), NINF)
         #self.assertTrue(math.isnan(math.floor(NAN)))
 
+        class TestFloorIsNone(float):
+            __floor__ = None
         class TestFloor:
             def __floor__(self):
                 return 42
@@ -588,6 +590,7 @@ class TestBadFloor:
         self.assertEqual(math.floor(FloatLike(41.9)), 41)
         self.assertRaises(TypeError, math.floor, TestNoFloor())
         self.assertRaises(ValueError, math.floor, TestBadFloor())
+        self.assertRaises(TypeError, math.floor, TestFloorIsNone(3.5))
 
         t = TestNoFloor()
         t.__floor__ = lambda *args: args