gh-134043: use stackrefs in vectorcalling methods (#134044)

Adds `_PyObject_GetMethodStackRef` which uses stackrefs and takes advantage of deferred reference counting in free-threading while calling method objects in vectorcall.

Include/internal/pycore_object.h

@@ -897,6 +897,9 @@ extern PyObject *_PyType_LookupRefAndVersion(PyTypeObject *, PyObject *,
 extern unsigned int
 _PyType_LookupStackRefAndVersion(PyTypeObject *type, PyObject *name, _PyStackRef *out);
 
+extern int _PyObject_GetMethodStackRef(PyThreadState *ts, PyObject *obj,
+                                       PyObject *name, _PyStackRef *method);
+
 // Cache the provided init method in the specialization cache of type if the
 // provided type version matches the current version of the type.
 //

Objects/call.c

@@ -834,12 +834,15 @@ PyObject_VectorcallMethod(PyObject *name, PyObject *const *args,
     assert(PyVectorcall_NARGS(nargsf) >= 1);
 
     PyThreadState *tstate = _PyThreadState_GET();
-    PyObject *callable = NULL;
+    _PyCStackRef method;
+    _PyThreadState_PushCStackRef(tstate, &method);
     /* Use args[0] as "self" argument */
-    int unbound = _PyObject_GetMethod(args[0], name, &callable);
-    if (callable == NULL) {
+    int unbound = _PyObject_GetMethodStackRef(tstate, args[0], name, &method.ref);
+    if (PyStackRef_IsNull(method.ref)) {
+        _PyThreadState_PopCStackRef(tstate, &method);
         return NULL;
     }
+    PyObject *callable = PyStackRef_AsPyObjectBorrow(method.ref);
 
     if (unbound) {
         /* We must remove PY_VECTORCALL_ARGUMENTS_OFFSET since
@@ -855,7 +858,7 @@ PyObject_VectorcallMethod(PyObject *name, PyObject *const *args,
     EVAL_CALL_STAT_INC_IF_FUNCTION(EVAL_CALL_METHOD, callable);
     PyObject *result = _PyObject_VectorcallTstate(tstate, callable,
                                                   args, nargsf, kwnames);
-    Py_DECREF(callable);
+    _PyThreadState_PopCStackRef(tstate, &method);
     return result;
 }
 
@@ -868,11 +871,14 @@ PyObject_CallMethodObjArgs(PyObject *obj, PyObject *name, ...)
         return null_error(tstate);
     }
 
-    PyObject *callable = NULL;
-    int is_method = _PyObject_GetMethod(obj, name, &callable);
-    if (callable == NULL) {
+    _PyCStackRef method;
+    _PyThreadState_PushCStackRef(tstate, &method);
+    int is_method = _PyObject_GetMethodStackRef(tstate, obj, name, &method.ref);
+    if (PyStackRef_IsNull(method.ref)) {
+        _PyThreadState_PopCStackRef(tstate, &method);
         return NULL;
     }
+    PyObject *callable = PyStackRef_AsPyObjectBorrow(method.ref);
     obj = is_method ? obj : NULL;
 
     va_list vargs;
@@ -880,7 +886,7 @@ PyObject_CallMethodObjArgs(PyObject *obj, PyObject *name, ...)
     PyObject *result = object_vacall(tstate, obj, callable, vargs);
     va_end(vargs);
 
-    Py_DECREF(callable);
+    _PyThreadState_PopCStackRef(tstate, &method);
     return result;
 }
 
@@ -897,12 +903,15 @@ _PyObject_CallMethodIdObjArgs(PyObject *obj, _Py_Identifier *name, ...)
     if (!oname) {
         return NULL;
     }
-
-    PyObject *callable = NULL;
-    int is_method = _PyObject_GetMethod(obj, oname, &callable);
-    if (callable == NULL) {
+    _PyCStackRef method;
+    _PyThreadState_PushCStackRef(tstate, &method);
+    int is_method = _PyObject_GetMethodStackRef(tstate, obj, oname, &method.ref);
+    if (PyStackRef_IsNull(method.ref)) {
+        _PyThreadState_PopCStackRef(tstate, &method);
         return NULL;
     }
+    PyObject *callable = PyStackRef_AsPyObjectBorrow(method.ref);
+
     obj = is_method ? obj : NULL;
 
     va_list vargs;
@@ -910,7 +919,7 @@ _PyObject_CallMethodIdObjArgs(PyObject *obj, _Py_Identifier *name, ...)
     PyObject *result = object_vacall(tstate, obj, callable, vargs);
     va_end(vargs);
 
-    Py_DECREF(callable);
+    _PyThreadState_PopCStackRef(tstate, &method);
     return result;
 }
 

Objects/object.c

@@ -1664,6 +1664,116 @@ _PyObject_GetMethod(PyObject *obj, PyObject *name, PyObject **method)
     return 0;
 }
 
+int
+_PyObject_GetMethodStackRef(PyThreadState *ts, PyObject *obj,
+                            PyObject *name, _PyStackRef *method)
+{
+    int meth_found = 0;
+
+    assert(PyStackRef_IsNull(*method));
+
+    PyTypeObject *tp = Py_TYPE(obj);
+    if (!_PyType_IsReady(tp)) {
+        if (PyType_Ready(tp) < 0) {
+            return 0;
+        }
+    }
+
+    if (tp->tp_getattro != PyObject_GenericGetAttr || !PyUnicode_CheckExact(name)) {
+        PyObject *res = PyObject_GetAttr(obj, name);
+        if (res != NULL) {
+            *method = PyStackRef_FromPyObjectSteal(res);
+        }
+        return 0;
+    }
+
+    _PyType_LookupStackRefAndVersion(tp, name, method);
+    PyObject *descr = PyStackRef_AsPyObjectBorrow(*method);
+    descrgetfunc f = NULL;
+    if (descr != NULL) {
+        if (_PyType_HasFeature(Py_TYPE(descr), Py_TPFLAGS_METHOD_DESCRIPTOR)) {
+            meth_found = 1;
+        }
+        else {
+            f = Py_TYPE(descr)->tp_descr_get;
+            if (f != NULL && PyDescr_IsData(descr)) {
+                PyObject *value = f(descr, obj, (PyObject *)Py_TYPE(obj));
+                PyStackRef_CLEAR(*method);
+                if (value != NULL) {
+                    *method = PyStackRef_FromPyObjectSteal(value);
+                }
+                return 0;
+            }
+        }
+    }
+    PyObject *dict, *attr;
+    if ((tp->tp_flags & Py_TPFLAGS_INLINE_VALUES) &&
+         _PyObject_TryGetInstanceAttribute(obj, name, &attr)) {
+        if (attr != NULL) {
+           PyStackRef_CLEAR(*method);
+           *method = PyStackRef_FromPyObjectSteal(attr);
+           return 0;
+        }
+        dict = NULL;
+    }
+    else if ((tp->tp_flags & Py_TPFLAGS_MANAGED_DICT)) {
+        dict = (PyObject *)_PyObject_GetManagedDict(obj);
+    }
+    else {
+        PyObject **dictptr = _PyObject_ComputedDictPointer(obj);
+        if (dictptr != NULL) {
+            dict = FT_ATOMIC_LOAD_PTR_ACQUIRE(*dictptr);
+        }
+        else {
+            dict = NULL;
+        }
+    }
+    if (dict != NULL) {
+        // TODO: use _Py_dict_lookup_threadsafe_stackref
+        Py_INCREF(dict);
+        PyObject *value;
+        if (PyDict_GetItemRef(dict, name, &value) != 0) {
+            // found or error
+            Py_DECREF(dict);
+            PyStackRef_CLEAR(*method);
+            if (value != NULL) {
+                *method = PyStackRef_FromPyObjectSteal(value);
+            }
+            return 0;
+        }
+        // not found
+        Py_DECREF(dict);
+    }
+
+    if (meth_found) {
+        assert(!PyStackRef_IsNull(*method));
+        return 1;
+    }
+
+    if (f != NULL) {
+        PyObject *value = f(descr, obj, (PyObject *)Py_TYPE(obj));
+        PyStackRef_CLEAR(*method);
+        if (value) {
+            *method = PyStackRef_FromPyObjectSteal(value);
+        }
+        return 0;
+    }
+
+    if (descr != NULL) {
+        assert(!PyStackRef_IsNull(*method));
+        return 0;
+    }
+
+    PyErr_Format(PyExc_AttributeError,
+                 "'%.100s' object has no attribute '%U'",
+                 tp->tp_name, name);
+
+    _PyObject_SetAttributeErrorContext(obj, name);
+    assert(PyStackRef_IsNull(*method));
+    return 0;
+}
+
+
 /* Generic GetAttr functions - put these in your tp_[gs]etattro slot. */
 
 PyObject *

Tools/ftscalingbench/ftscalingbench.py

@@ -27,6 +27,7 @@ import queue
 import sys
 import threading
 import time
+from operator import methodcaller
 
 # The iterations in individual benchmarks are scaled by this factor.
 WORK_SCALE = 100
@@ -188,6 +189,18 @@ def thread_local_read():
         _ = tmp.x
         _ = tmp.x
 
+class MyClass:
+    __slots__ = ()
+
+    def func(self):
+        pass
+
+@register_benchmark
+def method_caller():
+    mc = methodcaller("func")
+    obj = MyClass()
+    for i in range(1000 * WORK_SCALE):
+        mc(obj)
 
 def bench_one_thread(func):
     t0 = time.perf_counter_ns()