Enable specialization of STORE_ATTR free-threaded builds.

Author: nascheme

Include/internal/pycore_opcode_metadata.h

@@ -129,6 +129,112 @@ def mutate():
         # with the cell binding being changed).
         do_race(access, mutate)
 
+    def test_racing_to_bool(self):
+
+        seq = [1]
+
+        class C:
+            def __bool__(self):
+                return False
+
+        def access():
+            if seq:
+                return 1
+            else:
+                return 2
+
+        def mutate():
+            nonlocal seq
+            seq = [1]
+            time.sleep(0)
+            seq = C()
+            time.sleep(0)
+
+        do_race(access, mutate)
+
+    def test_racing_store_attr_slot(self):
+        class C:
+            __slots__ = ['x', '__dict__']
+
+        c = C()
+
+        def set_slot():
+            for i in range(10):
+                c.x = i
+            time.sleep(0)
+
+        def change_type():
+            def set_x(self, x):
+                pass
+
+            def get_x(self):
+                pass
+
+            C.x = property(get_x, set_x)
+            time.sleep(0)
+            del C.x
+            time.sleep(0)
+
+        do_race(set_slot, change_type)
+
+        def set_getattribute():
+            C.__getattribute__ = lambda self, x: x
+            time.sleep(0)
+            del C.__getattribute__
+            time.sleep(0)
+
+        do_race(set_slot, set_getattribute)
+
+    def test_racing_store_attr_instance_value(self):
+        class C:
+            pass
+
+        c = C()
+
+        def set_value():
+            for i in range(100):
+                c.x = i
+
+        set_value()
+
+        def read():
+            x = c.x
+
+        def mutate():
+            # Adding a property for 'x' should unspecialize it.
+            C.x = property(lambda self: None, lambda self, x: None)
+            time.sleep(0)
+            del C.x
+            time.sleep(0)
+
+        do_race(read, mutate)
+
+    def test_racing_store_attr_with_hint(self):
+        class C:
+            pass
+
+        c = C()
+        for i in range(29):
+            setattr(c, f"_{i}", None)
+
+        def set_value():
+            for i in range(100):
+                c.x = i
+
+        set_value()
+
+        def read():
+            x = c.x
+
+        def mutate():
+            # Adding a property for 'x' should unspecialize it.
+            C.x = property(lambda self: None, lambda self, x: None)
+            time.sleep(0)
+            del C.x
+            time.sleep(0)
+
+        do_race(read, mutate)
+
 
 if __name__ == "__main__":
     unittest.main()

Include/internal/pycore_uop_metadata.h

@@ -1375,6 +1375,74 @@ def send_yield_from():
         self.assert_specialized(send_yield_from, "SEND_GEN")
         self.assert_no_opcode(send_yield_from, "SEND")
 
+    @cpython_only
+    @requires_specialization_ft
+    def test_store_attr_slot(self):
+        class C:
+            __slots__ = ['x']
+
+        def set_slot():
+            c = C()
+            for i in range(100):
+                c.x = i
+
+        set_slot()
+
+        #dis.dis(set_slot, adaptive=True)
+
+        self.assert_specialized(set_slot, "STORE_ATTR_SLOT")
+        self.assert_no_opcode(set_slot, "STORE_ATTR")
+
+        # Adding a property for 'x' should unspecialize it.
+        C.x = property(lambda self: None, lambda self, x: None)
+        set_slot()
+        self.assert_no_opcode(set_slot, "STORE_ATTR_SLOT")
+
+    @cpython_only
+    @requires_specialization_ft
+    def test_store_attr_instance_value(self):
+        class C:
+            pass
+
+        def set_value():
+            c = C()
+            for i in range(100):
+                c.x = i
+
+        set_value()
+
+        self.assert_specialized(set_value, "STORE_ATTR_INSTANCE_VALUE")
+        self.assert_no_opcode(set_value, "STORE_ATTR")
+
+        # Adding a property for 'x' should unspecialize it.
+        C.x = property(lambda self: None, lambda self, x: None)
+        set_value()
+        self.assert_no_opcode(set_value, "STORE_ATTR_INSTANCE_VALUE")
+
+    @cpython_only
+    @requires_specialization_ft
+    def test_store_attr_with_hint(self):
+        class C:
+            pass
+
+        c = C()
+        for i in range(29):
+            setattr(c, f"_{i}", None)
+
+        def set_value():
+            for i in range(100):
+                c.x = i
+
+        set_value()
+
+        self.assert_specialized(set_value, "STORE_ATTR_WITH_HINT")
+        self.assert_no_opcode(set_value, "STORE_ATTR")
+
+        # Adding a property for 'x' should unspecialize it.
+        C.x = property(lambda self: None, lambda self, x: None)
+        set_value()
+        self.assert_no_opcode(set_value, "STORE_ATTR_WITH_HINT")
+
     @cpython_only
     @requires_specialization_ft
     def test_to_bool(self):

Lib/test/test_free_threading/test_races.py

@@ -1467,7 +1467,7 @@ dummy_func(
         };
 
         specializing op(_SPECIALIZE_STORE_ATTR, (counter/1, owner -- owner)) {
-            #if ENABLE_SPECIALIZATION
+            #if ENABLE_SPECIALIZATION_FT
             if (ADAPTIVE_COUNTER_TRIGGERS(counter)) {
                 PyObject *name = GETITEM(FRAME_CO_NAMES, oparg);
                 next_instr = this_instr;
@@ -1476,7 +1476,7 @@ dummy_func(
             }
             OPCODE_DEFERRED_INC(STORE_ATTR);
             ADVANCE_ADAPTIVE_COUNTER(this_instr[1].counter);
-            #endif  /* ENABLE_SPECIALIZATION */
+            #endif  /* ENABLE_SPECIALIZATION_FT */
         }
 
         op(_STORE_ATTR, (v, owner --)) {
@@ -2129,7 +2129,7 @@ dummy_func(
         op(_GUARD_TYPE_VERSION, (type_version/2, owner -- owner)) {
             PyTypeObject *tp = Py_TYPE(PyStackRef_AsPyObjectBorrow(owner));
             assert(type_version != 0);
-            EXIT_IF(tp->tp_version_tag != type_version);
+            EXIT_IF(FT_ATOMIC_LOAD_UINT32_RELAXED(tp->tp_version_tag) != type_version);
         }
 
         op(_CHECK_MANAGED_OBJECT_HAS_VALUES, (owner -- owner)) {
@@ -2337,25 +2337,27 @@ dummy_func(
             assert(Py_TYPE(owner_o)->tp_dictoffset < 0);
             assert(Py_TYPE(owner_o)->tp_flags & Py_TPFLAGS_INLINE_VALUES);
             EXIT_IF(_PyObject_GetManagedDict(owner_o));
-            EXIT_IF(_PyObject_InlineValues(owner_o)->valid == 0);
+            EXIT_IF(FT_ATOMIC_LOAD_UINT8(_PyObject_InlineValues(owner_o)->valid) == 0);
         }
 
         op(_STORE_ATTR_INSTANCE_VALUE, (offset/1, value, owner --)) {
             PyObject *owner_o = PyStackRef_AsPyObjectBorrow(owner);
-
+            DEOPT_IF(!LOCK_OBJECT(owner_o));
+            if (_PyObject_GetManagedDict(owner_o) == NULL) {
+                UNLOCK_OBJECT(owner_o);
+                DEOPT_IF(true);
+            }
             STAT_INC(STORE_ATTR, hit);
-            assert(_PyObject_GetManagedDict(owner_o) == NULL);
             PyObject **value_ptr = (PyObject**)(((char *)owner_o) + offset);
             PyObject *old_value = *value_ptr;
-            *value_ptr = PyStackRef_AsPyObjectSteal(value);
+            FT_ATOMIC_STORE_PTR_RELEASE(*value_ptr, PyStackRef_AsPyObjectSteal(value));
             if (old_value == NULL) {
                 PyDictValues *values = _PyObject_InlineValues(owner_o);
                 Py_ssize_t index = value_ptr - values->values;
                 _PyDictValues_AddToInsertionOrder(values, index);
             }
-            else {
-                Py_DECREF(old_value);
-            }
+            UNLOCK_OBJECT(owner_o);
+            Py_XDECREF(old_value);
             PyStackRef_CLOSE(owner);
         }
 
@@ -2370,16 +2372,33 @@ dummy_func(
             assert(Py_TYPE(owner_o)->tp_flags & Py_TPFLAGS_MANAGED_DICT);
             PyDictObject *dict = _PyObject_GetManagedDict(owner_o);
             DEOPT_IF(dict == NULL);
+
+            DEOPT_IF(!LOCK_OBJECT(dict));
+            if (dict == NULL) {
+                UNLOCK_OBJECT(dict);
+                DEOPT_IF(true);
+            }
             assert(PyDict_CheckExact((PyObject *)dict));
             PyObject *name = GETITEM(FRAME_CO_NAMES, oparg);
-            DEOPT_IF(hint >= (size_t)dict->ma_keys->dk_nentries);
-            DEOPT_IF(!DK_IS_UNICODE(dict->ma_keys));
+            if (hint >= (size_t)dict->ma_keys->dk_nentries ||
+                    !DK_IS_UNICODE(dict->ma_keys)) {
+                UNLOCK_OBJECT(dict);
+                DEOPT_IF(true);
+            }
             PyDictUnicodeEntry *ep = DK_UNICODE_ENTRIES(dict->ma_keys) + hint;
-            DEOPT_IF(ep->me_key != name);
+            if (ep->me_key != name) {
+                UNLOCK_OBJECT(dict);
+                DEOPT_IF(true);
+            }
             PyObject *old_value = ep->me_value;
-            DEOPT_IF(old_value == NULL);
+            if (old_value == NULL) {
+                UNLOCK_OBJECT(dict);
+                DEOPT_IF(true);
+            }
             _PyDict_NotifyEvent(tstate->interp, PyDict_EVENT_MODIFIED, dict, name, PyStackRef_AsPyObjectBorrow(value));
-            ep->me_value = PyStackRef_AsPyObjectSteal(value);
+            FT_ATOMIC_STORE_PTR_RELEASE(ep->me_value, PyStackRef_AsPyObjectSteal(value));
+            UNLOCK_OBJECT(dict);
+
             // old_value should be DECREFed after GC track checking is done, if not, it could raise a segmentation fault,
             // when dict only holds the strong reference to value in ep->me_value.
             Py_XDECREF(old_value);
@@ -2395,10 +2414,12 @@ dummy_func(
         op(_STORE_ATTR_SLOT, (index/1, value, owner --)) {
             PyObject *owner_o = PyStackRef_AsPyObjectBorrow(owner);
 
+            DEOPT_IF(!LOCK_OBJECT(owner_o));
             char *addr = (char *)owner_o + index;
             STAT_INC(STORE_ATTR, hit);
             PyObject *old_value = *(PyObject **)addr;
-            *(PyObject **)addr = PyStackRef_AsPyObjectSteal(value);
+            FT_ATOMIC_STORE_PTR_RELEASE(*(PyObject **)addr, PyStackRef_AsPyObjectSteal(value));
+            UNLOCK_OBJECT(owner_o);
             Py_XDECREF(old_value);
             PyStackRef_CLOSE(owner);
         }