bpo-45340: Don't create object dictionaries unless actually needed

Include/cpython/object.h

@@ -270,6 +270,7 @@ struct _typeobject {
 
     destructor tp_finalize;
     vectorcallfunc tp_vectorcall;
+    Py_ssize_t tp_inline_values_offset;
 };
 
 /* The *real* layout of a type object when allocated on the heap */

Include/internal/pycore_dict.h

@@ -101,6 +101,8 @@ extern uint64_t _pydict_global_version;
 
 #define DICT_NEXT_VERSION() (++_pydict_global_version)
 
+PyObject *_PyObject_MakeDictFromInstanceAttributes(PyObject *obj, PyDictValues *values);
+
 #ifdef __cplusplus
 }
 #endif

Include/internal/pycore_object.h

@@ -181,6 +181,16 @@ extern int _Py_CheckSlotResult(
 extern PyObject* _PyType_AllocNoTrack(PyTypeObject *type, Py_ssize_t nitems);
 
 extern int _PyObject_InitializeDict(PyObject *obj);
+extern int _PyObject_StoreInstanceAttribute(PyObject *obj, PyDictValues *values,
+                                          PyObject *name, PyObject *value);
+PyObject * _PyObject_GetInstanceAttribute(PyObject *obj, PyDictValues *values,
+                                        PyObject *name);
+PyDictValues ** _PyObject_ValuesPointer(PyObject *);
+PyObject ** _PyObject_DictPointer(PyObject *);
+int _PyObject_VisitInstanceAttributes(PyObject *self, visitproc visit, void *arg);
+void _PyObject_ClearInstanceAttributes(PyObject *self);
+void _PyObject_FreeInstanceAttributes(PyObject *self);
+int _PyObject_IsInstanceDictEmpty(PyObject *);
 
 #ifdef __cplusplus
 }

Include/object.h

@@ -333,6 +333,7 @@ given type object has a specified feature.
 */
 
 #ifndef Py_LIMITED_API
+
 /* Set if instances of the type object are treated as sequences for pattern matching */
 #define Py_TPFLAGS_SEQUENCE (1 << 5)
 /* Set if instances of the type object are treated as mappings for pattern matching */

Lib/opcode.py

@@ -231,7 +231,7 @@ def jabs_op(name, op):
     "BINARY_SUBSCR_DICT",
     "JUMP_ABSOLUTE_QUICK",
     "LOAD_ATTR_ADAPTIVE",
-    "LOAD_ATTR_SPLIT_KEYS",
+    "LOAD_ATTR_INSTANCE_VALUE",
     "LOAD_ATTR_WITH_HINT",
     "LOAD_ATTR_SLOT",
     "LOAD_ATTR_MODULE",
@@ -242,8 +242,9 @@ def jabs_op(name, op):
     "LOAD_METHOD_CACHED",
     "LOAD_METHOD_CLASS",
     "LOAD_METHOD_MODULE",
+    "LOAD_METHOD_NO_DICT",
     "STORE_ATTR_ADAPTIVE",
-    "STORE_ATTR_SPLIT_KEYS",
+    "STORE_ATTR_INSTANCE_VALUE",
     "STORE_ATTR_SLOT",
     "STORE_ATTR_WITH_HINT",
     # Super instructions

Lib/test/test_descr.py

@@ -5500,17 +5500,19 @@ class A:
         class B(A):
             pass
 
+        #Shrink keys by repeatedly creating instances
+        [(A(), B()) for _ in range(20)]
+
         a, b = A(), B()
         self.assertEqual(sys.getsizeof(vars(a)), sys.getsizeof(vars(b)))
         self.assertLess(sys.getsizeof(vars(a)), sys.getsizeof({"a":1}))
-        # Initial hash table can contain at most 5 elements.
+        # Initial hash table can contain only one or two elements.
         # Set 6 attributes to cause internal resizing.
         a.x, a.y, a.z, a.w, a.v, a.u = range(6)
         self.assertNotEqual(sys.getsizeof(vars(a)), sys.getsizeof(vars(b)))
         a2 = A()
-        self.assertEqual(sys.getsizeof(vars(a)), sys.getsizeof(vars(a2)))
-        self.assertLess(sys.getsizeof(vars(a)), sys.getsizeof({"a":1}))
-        b.u, b.v, b.w, b.t, b.s, b.r = range(6)
+        self.assertGreater(sys.getsizeof(vars(a)), sys.getsizeof(vars(a2)))
+        self.assertLess(sys.getsizeof(vars(a2)), sys.getsizeof({"a":1}))
         self.assertLess(sys.getsizeof(vars(b)), sys.getsizeof({"a":1}))
 
 

Lib/test/test_dict.py

@@ -994,8 +994,8 @@ class C:
 
     @support.cpython_only
     def test_splittable_setdefault(self):
-        """split table must be combined when setdefault()
-        breaks insertion order"""
+        """split table must keep correct insertion
+        order when attributes are adding using setdefault()"""
         a, b = self.make_shared_key_dict(2)
 
         a['a'] = 1
@@ -1005,7 +1005,6 @@ def test_splittable_setdefault(self):
         size_b = sys.getsizeof(b)
         b['a'] = 1
 
-        self.assertGreater(size_b, size_a)
         self.assertEqual(list(a), ['x', 'y', 'z', 'a', 'b'])
         self.assertEqual(list(b), ['x', 'y', 'z', 'b', 'a'])
 

Lib/test/test_gc.py

@@ -444,7 +444,7 @@ def __getattr__(self, someattribute):
         # 0, thus mutating the trash graph as a side effect of merely asking
         # whether __del__ exists.  This used to (before 2.3b1) crash Python.
         # Now __getattr__ isn't called.
-        self.assertEqual(gc.collect(), 4)
+        self.assertEqual(gc.collect(), 2)
         self.assertEqual(len(gc.garbage), garbagelen)
 
     def test_boom2(self):
@@ -471,7 +471,7 @@ def __getattr__(self, someattribute):
         # there isn't a second time, so this simply cleans up the trash cycle.
         # We expect a, b, a.__dict__ and b.__dict__ (4 objects) to get
         # reclaimed this way.
-        self.assertEqual(gc.collect(), 4)
+        self.assertEqual(gc.collect(), 2)
         self.assertEqual(len(gc.garbage), garbagelen)
 
     def test_boom_new(self):
@@ -491,7 +491,7 @@ def __getattr__(self, someattribute):
         gc.collect()
         garbagelen = len(gc.garbage)
         del a, b
-        self.assertEqual(gc.collect(), 4)
+        self.assertEqual(gc.collect(), 2)
         self.assertEqual(len(gc.garbage), garbagelen)
 
     def test_boom2_new(self):
@@ -513,7 +513,7 @@ def __getattr__(self, someattribute):
         gc.collect()
         garbagelen = len(gc.garbage)
         del a, b
-        self.assertEqual(gc.collect(), 4)
+        self.assertEqual(gc.collect(), 2)
         self.assertEqual(len(gc.garbage), garbagelen)
 
     def test_get_referents(self):
@@ -943,8 +943,8 @@ def getstats():
             A()
         t = gc.collect()
         c, nc = getstats()
-        self.assertEqual(t, 2*N) # instance object & its dict
-        self.assertEqual(c - oldc, 2*N)
+        self.assertEqual(t, N) # instance objects
+        self.assertEqual(c - oldc, N)
         self.assertEqual(nc - oldnc, 0)
 
         # But Z() is not actually collected.
@@ -964,8 +964,8 @@ def getstats():
         Z()
         t = gc.collect()
         c, nc = getstats()
-        self.assertEqual(t, 2*N)
-        self.assertEqual(c - oldc, 2*N)
+        self.assertEqual(t, N)
+        self.assertEqual(c - oldc, N)
         self.assertEqual(nc - oldnc, 0)
 
         # The A() trash should have been reclaimed already but the
@@ -974,8 +974,8 @@ def getstats():
         zs.clear()
         t = gc.collect()
         c, nc = getstats()
-        self.assertEqual(t, 4)
-        self.assertEqual(c - oldc, 4)
+        self.assertEqual(t, 2)
+        self.assertEqual(c - oldc, 2)
         self.assertEqual(nc - oldnc, 0)
 
         gc.enable()
@@ -1128,8 +1128,7 @@ def test_collect_generation(self):
     @cpython_only
     def test_collect_garbage(self):
         self.preclean()
-        # Each of these cause four objects to be garbage: Two
-        # Uncollectables and their instance dicts.
+        # Each of these cause two objects to be garbage:
         Uncollectable()
         Uncollectable()
         C1055820(666)
@@ -1138,8 +1137,8 @@ def test_collect_garbage(self):
             if v[1] != "stop":
                 continue
             info = v[2]
-            self.assertEqual(info["collected"], 2)
-            self.assertEqual(info["uncollectable"], 8)
+            self.assertEqual(info["collected"], 1)
+            self.assertEqual(info["uncollectable"], 4)
 
         # We should now have the Uncollectables in gc.garbage
         self.assertEqual(len(gc.garbage), 4)
@@ -1156,7 +1155,7 @@ def test_collect_garbage(self):
                 continue
             info = v[2]
             self.assertEqual(info["collected"], 0)
-            self.assertEqual(info["uncollectable"], 4)
+            self.assertEqual(info["uncollectable"], 2)
 
         # Uncollectables should be gone
         self.assertEqual(len(gc.garbage), 0)

Lib/test/test_sys.py

@@ -1409,7 +1409,7 @@ def delx(self): del self.__x
         check((1,2,3), vsize('') + 3*self.P)
         # type
         # static type: PyTypeObject
-        fmt = 'P2nPI13Pl4Pn9Pn11PIPP'
+        fmt = 'P2nPI13Pl4Pn9Pn12PIPP'
         s = vsize(fmt)
         check(int, s)
         # class
@@ -1422,15 +1422,15 @@ def delx(self): del self.__x
                   '5P')
         class newstyleclass(object): pass
         # Separate block for PyDictKeysObject with 8 keys and 5 entries
-        check(newstyleclass, s + calcsize(DICT_KEY_STRUCT_FORMAT) + 8 + 5*calcsize("n2P"))
+        check(newstyleclass, s + calcsize(DICT_KEY_STRUCT_FORMAT) + 32 + 21*calcsize("n2P"))
         # dict with shared keys
-        check(newstyleclass().__dict__, size('nQ2P') + 5*self.P)
+        check(newstyleclass().__dict__, size('nQ2P') + 15*self.P)
         o = newstyleclass()
         o.a = o.b = o.c = o.d = o.e = o.f = o.g = o.h = 1
         # Separate block for PyDictKeysObject with 16 keys and 10 entries
-        check(newstyleclass, s + calcsize(DICT_KEY_STRUCT_FORMAT) + 16 + 10*calcsize("n2P"))
+        check(newstyleclass, s + calcsize(DICT_KEY_STRUCT_FORMAT) + 32 + 21*calcsize("n2P"))
         # dict with shared keys
-        check(newstyleclass().__dict__, size('nQ2P') + 10*self.P)
+        check(newstyleclass().__dict__, size('nQ2P') + 13*self.P)
         # unicode
         # each tuple contains a string and its expected character size
         # don't put any static strings here, as they may contain

Misc/NEWS.d/next/Core and Builtins/2021-10-08-09-47-38.bpo-45340.ukHgDb.rst

@@ -0,0 +1,3 @@
+Object attributes are held in an array instead of a dictionary. An object's
+dictionary are created lazily, only when needed. Reduces the memory
+consumption of a typical Python object by about 30%. Patch by Mark Shannon.