pythonGH-115776: Allow any fixed sized object to have inline values (p…

…ythonGH-123192)

Include/internal/pycore_object.h

@@ -803,10 +803,11 @@ _PyObject_GetManagedDict(PyObject *obj)
 static inline PyDictValues *
 _PyObject_InlineValues(PyObject *obj)
 {
+    PyTypeObject *tp = Py_TYPE(obj);
+    assert(tp->tp_basicsize > 0 && tp->tp_basicsize % sizeof(PyObject *) == 0);
     assert(Py_TYPE(obj)->tp_flags & Py_TPFLAGS_INLINE_VALUES);
     assert(Py_TYPE(obj)->tp_flags & Py_TPFLAGS_MANAGED_DICT);
-    assert(Py_TYPE(obj)->tp_basicsize == sizeof(PyObject));
-    return (PyDictValues *)((char *)obj + sizeof(PyObject));
+    return (PyDictValues *)((char *)obj + tp->tp_basicsize);
 }
 
 extern PyObject ** _PyObject_ComputedDictPointer(PyObject *);

Misc/NEWS.d/next/Core_and_Builtins/2024-08-21-08-53-00.gh-issue-115776.9A7Dv_.rst

@@ -0,0 +1,2 @@
+Enables inline values (Python's equivalent of hidden classes) on any class
+who's instances are of a fixed size.

Objects/object_layout.md

@@ -28,6 +28,10 @@ So the pre-header is these two fields:
 If the object has no physical dictionary, then the ``dict_pointer``
 is set to `NULL`.
 
+In 3.13 only objects with no additional data could have inline values.
+That is, instances of classes with `tp_basicsize == sizeof(PyObject)`.
+In 3.14, any object whose class has `tp_itemsize == 0` can have inline values.
+In both versions, the inline values starts `tp_basicsize` bytes after the object.
 
 <details>
 <summary> 3.12 </summary>

Objects/typeobject.c

@@ -8340,7 +8340,7 @@ type_ready_managed_dict(PyTypeObject *type)
             return -1;
         }
     }
-    if (type->tp_itemsize == 0 && type->tp_basicsize == sizeof(PyObject)) {
+    if (type->tp_itemsize == 0) {
         type->tp_flags |= Py_TPFLAGS_INLINE_VALUES;
     }
     return 0;

Python/bytecodes.c

@@ -2012,9 +2012,10 @@ dummy_func(
             DEOPT_IF(!_PyObject_InlineValues(owner_o)->valid);
         }
 
-        split op(_LOAD_ATTR_INSTANCE_VALUE, (index/1, owner -- attr, null if (oparg & 1))) {
+        split op(_LOAD_ATTR_INSTANCE_VALUE, (offset/1, owner -- attr, null if (oparg & 1))) {
             PyObject *owner_o = PyStackRef_AsPyObjectBorrow(owner);
-            PyObject *attr_o = _PyObject_InlineValues(owner_o)->values[index];
+            PyObject **value_ptr = (PyObject**)(((char *)owner_o) + offset);
+            PyObject *attr_o = *value_ptr;
             DEOPT_IF(attr_o == NULL);
             STAT_INC(LOAD_ATTR, hit);
             Py_INCREF(attr_o);
@@ -2196,16 +2197,17 @@ dummy_func(
             EXIT_IF(_PyObject_InlineValues(owner_o)->valid == 0);
         }
 
-        op(_STORE_ATTR_INSTANCE_VALUE, (index/1, value, owner --)) {
+        op(_STORE_ATTR_INSTANCE_VALUE, (offset/1, value, owner --)) {
             PyObject *owner_o = PyStackRef_AsPyObjectBorrow(owner);
 
             STAT_INC(STORE_ATTR, hit);
             assert(_PyObject_GetManagedDict(owner_o) == NULL);
-            PyDictValues *values = _PyObject_InlineValues(owner_o);
-
-            PyObject *old_value = values->values[index];
-            values->values[index] = PyStackRef_AsPyObjectSteal(value);
+            PyObject **value_ptr = (PyObject**)(((char *)owner_o) + offset);
+            PyObject *old_value = *value_ptr;
+            *value_ptr = PyStackRef_AsPyObjectSteal(value);
             if (old_value == NULL) {
+                PyDictValues *values = _PyObject_InlineValues(owner_o);
+                int index = value_ptr - values->values;
                 _PyDictValues_AddToInsertionOrder(values, index);
             }
             else {

Python/gc.c

@@ -2055,6 +2055,9 @@ _PyObject_GC_New(PyTypeObject *tp)
         return NULL;
     }
     _PyObject_Init(op, tp);
+    if (tp->tp_flags & Py_TPFLAGS_INLINE_VALUES) {
+        _PyObject_InitInlineValues(op, tp);
+    }
     return op;
 }
 

Python/gc_free_threading.c

@@ -1810,6 +1810,9 @@ _PyObject_GC_New(PyTypeObject *tp)
         return NULL;
     }
     _PyObject_Init(op, tp);
+    if (tp->tp_flags & Py_TPFLAGS_INLINE_VALUES) {
+        _PyObject_InitInlineValues(op, tp);
+    }
     return op;
 }
 

Python/optimizer_bytecodes.c

@@ -452,10 +452,10 @@ dummy_func(void) {
         top, unused[oparg-2], bottom)) {
     }
 
-    op(_LOAD_ATTR_INSTANCE_VALUE, (index/1, owner -- attr, null if (oparg & 1))) {
+    op(_LOAD_ATTR_INSTANCE_VALUE, (offset/1, owner -- attr, null if (oparg & 1))) {
         attr = sym_new_not_null(ctx);
         null = sym_new_null(ctx);
-        (void)index;
+        (void)offset;
         (void)owner;
     }
 

Python/specialize.c

@@ -849,15 +849,19 @@ specialize_dict_access(
         assert(PyUnicode_CheckExact(name));
         Py_ssize_t index = _PyDictKeys_StringLookup(keys, name);
         assert (index != DKIX_ERROR);
-        if (index != (uint16_t)index) {
-            SPECIALIZATION_FAIL(base_op,
-                                index == DKIX_EMPTY ?
-                                SPEC_FAIL_ATTR_NOT_IN_KEYS :
-                                SPEC_FAIL_OUT_OF_RANGE);
+        if (index == DKIX_EMPTY) {
+            SPECIALIZATION_FAIL(base_op, SPEC_FAIL_ATTR_NOT_IN_KEYS);
+            return 0;
+        }
+        assert(index >= 0);
+        char *value_addr = (char *)&_PyObject_InlineValues(owner)->values[index];
+        Py_ssize_t offset = value_addr - (char *)owner;
+        if (offset != (uint16_t)offset) {
+            SPECIALIZATION_FAIL(base_op, SPEC_FAIL_OUT_OF_RANGE);
             return 0;
         }
         write_u32(cache->version, type->tp_version_tag);
-        cache->index = (uint16_t)index;
+        cache->index = (uint16_t)offset;
         instr->op.code = values_op;
     }
     else {

Tools/gdb/libpython.py

@@ -69,9 +69,6 @@ def _type_unsigned_int_ptr():
 def _sizeof_void_p():
     return gdb.lookup_type('void').pointer().sizeof
 
-def _sizeof_pyobject():
-    return gdb.lookup_type('PyObject').sizeof
-
 def _managed_dict_offset():
     # See pycore_object.h
     pyobj = gdb.lookup_type("PyObject")
@@ -505,7 +502,7 @@ def get_keys_values(self):
         dict_ptr = dict_ptr_ptr.cast(_type_char_ptr().pointer()).dereference()
         if int(dict_ptr):
             return None
-        char_ptr = obj_ptr + _sizeof_pyobject()
+        char_ptr = obj_ptr + typeobj.field('tp_basicsize')
         values_ptr = char_ptr.cast(gdb.lookup_type("PyDictValues").pointer())
         values = values_ptr['values']
         return PyKeysValuesPair(self.get_cached_keys(), values)