gh-139103: Improve enum free-threaded scaling

Author: colesbury

Lib/enum.py

@@ -185,6 +185,10 @@ class property(DynamicClassAttribute):
     _attr_type = None
     _cls_type = None
 
+    def __init__(self, fget=None, fset=None, fdel=None, doc=None):
+        sys._enable_deferred_refcount(self)
+        super().__init__(fget, fset, fdel, doc)
+
     def __get__(self, instance, ownerclass=None):
         if instance is None:
             if self.member is not None:
@@ -1067,6 +1071,7 @@ def _find_new_(mcls, classdict, member_type, first_enum):
         return __new__, save_new, use_args
 
     def _add_member_(cls, name, member):
+        sys._enable_deferred_refcount(member)
         # _value_ structures are not updated
         if name in cls._member_map_:
             if cls._member_map_[name] is not member:

Misc/NEWS.d/next/Core_and_Builtins/2026-02-18-22-03-02.gh-issue-139103.ZOmb0m.rst

@@ -0,0 +1,2 @@
+Add :func:`sys._enable_deferred_refcount` and use it in :mod:`enum` to
+improve free-threaded scaling.

Objects/typeobject.c

@@ -10925,15 +10925,19 @@ static PyObject *
 slot_tp_descr_get(PyObject *self, PyObject *obj, PyObject *type)
 {
     PyTypeObject *tp = Py_TYPE(self);
-    PyObject *get;
+    PyThreadState *tstate = _PyThreadState_GET();
+    _PyCStackRef cref;
+    _PyThreadState_PushCStackRef(tstate, &cref);
 
-    get = _PyType_LookupRef(tp, &_Py_ID(__get__));
+    _PyType_LookupStackRefAndVersion(tp, &_Py_ID(__get__), &cref.ref);
+    PyObject *get = PyStackRef_AsPyObjectBorrow(cref.ref);
     if (get == NULL) {
 #ifndef Py_GIL_DISABLED
         /* Avoid further slowdowns */
         if (tp->tp_descr_get == slot_tp_descr_get)
             tp->tp_descr_get = NULL;
 #endif
+        _PyThreadState_PopCStackRef(tstate, &cref);
         return Py_NewRef(self);
     }
     if (obj == NULL)
@@ -10942,7 +10946,7 @@ slot_tp_descr_get(PyObject *self, PyObject *obj, PyObject *type)
         type = Py_None;
     PyObject *stack[3] = {self, obj, type};
     PyObject *res = PyObject_Vectorcall(get, stack, 3, NULL);
-    Py_DECREF(get);
+    _PyThreadState_PopCStackRef(tstate, &cref);
     return res;
 }
 

Python/clinic/sysmodule.c.h

@@ -1052,6 +1052,26 @@ sys__is_immortal_impl(PyObject *module, PyObject *op)
     return PyUnstable_IsImmortal(op);
 }
 
+/*[clinic input]
+sys._enable_deferred_refcount -> bool
+
+  op: object
+  /
+
+Enable deferred reference counting on the object.
+
+Return True if deferred reference counting was successfully enabled, and
+False otherwise. This is primarily useful for avoiding reference count
+contention on objects that are shared between multiple threads.
+[clinic start generated code]*/
+
+static int
+sys__enable_deferred_refcount_impl(PyObject *module, PyObject *op)
+/*[clinic end generated code: output=d19c0f74be9da2a8 input=92d197248dcfb1f7]*/
+{
+    return PyUnstable_Object_EnableDeferredRefcount(op);
+}
+
 /*
  * Cached interned string objects used for calling the profile and
  * trace functions.
@@ -2942,6 +2962,7 @@ static PyMethodDef sys_methods[] = {
     SYS_GETWINDOWSVERSION_METHODDEF
     SYS__ENABLELEGACYWINDOWSFSENCODING_METHODDEF
     SYS__IS_IMMORTAL_METHODDEF
+    SYS__ENABLE_DEFERRED_REFCOUNT_METHODDEF
     SYS_INTERN_METHODDEF
     SYS__IS_INTERNED_METHODDEF
     SYS_IS_FINALIZING_METHODDEF

Python/sysmodule.c

@@ -30,6 +30,7 @@
 import time
 from collections import namedtuple
 from dataclasses import dataclass
+from enum import Enum
 from operator import methodcaller
 from typing import NamedTuple
 
@@ -236,6 +237,25 @@ def instantiate_typing_namedtuple():
         obj = MyTypingNamedTuple(x=1, y=2, z=3)
 
 
+class MyEnum(Enum):
+    X = 1
+    Y = 2
+    Z = 3
+
+@register_benchmark
+def enum_attr():
+    for _ in range(1000 * WORK_SCALE):
+        MyEnum.X
+        MyEnum.Y
+        MyEnum.Z
+
+@register_benchmark
+def enum_value():
+    for _ in range(1000 * WORK_SCALE):
+        MyEnum.X.value
+        MyEnum.Y.value
+        MyEnum.Z.value
+
 @register_benchmark
 def deepcopy():
     x = {'list': [1, 2], 'tuple': (1, None)}