GH-91095: Specialize calls to normal Python classes. (GH-99331)

Include/cpython/object.h

@@ -246,6 +246,7 @@ struct _specialization_cache {
     //   *args nor **kwargs (as required by BINARY_SUBSCR_GETITEM):
     PyObject *getitem;
     uint32_t getitem_version;
+    PyObject *init;
 };
 
 /* The *real* layout of a type object when allocated on the heap */

Include/internal/pycore_frame.h

@@ -272,6 +272,30 @@ _PyFrame_PushUnchecked(PyThreadState *tstate, PyFunctionObject *func, int null_l
     return new_frame;
 }
 
+/* Pushes a trampoline frame without checking for space.
+ * Must be guarded by _PyThreadState_HasStackSpace() */
+static inline _PyInterpreterFrame *
+_PyFrame_PushTrampolineUnchecked(PyThreadState *tstate, PyCodeObject *code, int stackdepth, int prev_instr)
+{
+    CALL_STAT_INC(frames_pushed);
+    _PyInterpreterFrame *frame = (_PyInterpreterFrame *)tstate->datastack_top;
+    tstate->datastack_top += code->co_framesize;
+    assert(tstate->datastack_top < tstate->datastack_limit);
+    frame->f_funcobj = Py_None;
+    frame->f_executable = Py_NewRef(code);
+#ifdef Py_DEBUG
+    frame->f_builtins = NULL;
+    frame->f_globals = NULL;
+#endif
+    frame->f_locals = NULL;
+    frame->stacktop = code->co_nlocalsplus + stackdepth;
+    frame->frame_obj = NULL;
+    frame->prev_instr = _PyCode_CODE(code) + prev_instr;
+    frame->owner = FRAME_OWNED_BY_THREAD;
+    frame->return_offset = 0;
+    return frame;
+}
+
 static inline
 PyGenObject *_PyFrame_GetGenerator(_PyInterpreterFrame *frame)
 {

Include/internal/pycore_object.h

@@ -355,8 +355,10 @@ static inline int _PyType_SUPPORTS_WEAKREFS(PyTypeObject *type) {
 }
 
 extern PyObject* _PyType_AllocNoTrack(PyTypeObject *type, Py_ssize_t nitems);
+PyObject *_PyType_NewManagedObject(PyTypeObject *type);
 
 extern int _PyObject_InitializeDict(PyObject *obj);
+int _PyObject_InitInlineValues(PyObject *obj, PyTypeObject *tp);
 extern int _PyObject_StoreInstanceAttribute(PyObject *obj, PyDictValues *values,
                                           PyObject *name, PyObject *value);
 PyObject * _PyObject_GetInstanceAttribute(PyObject *obj, PyDictValues *values,

Lib/_opcode_metadata.py

@@ -85,6 +85,7 @@
         "CALL_METHOD_DESCRIPTOR_FAST_WITH_KEYWORDS",
         "CALL_NO_KW_METHOD_DESCRIPTOR_NOARGS",
         "CALL_NO_KW_METHOD_DESCRIPTOR_FAST",
+        "CALL_NO_KW_ALLOC_AND_ENTER_INIT",
     ],
 }
 

Lib/opcode.py

@@ -97,6 +97,7 @@ def pseudo_op(name, op, real_ops):
 def_op('UNARY_NOT', 12)
 
 def_op('UNARY_INVERT', 15)
+def_op('EXIT_INIT_CHECK', 16)
 
 # We reserve 17 as it is the initial value for the specializing counter
 # This helps us catch cases where we attempt to execute a cache.

Lib/test/test_sys.py

@@ -1557,7 +1557,7 @@ def delx(self): del self.__x
                   '10P'                 # PySequenceMethods
                   '2P'                  # PyBufferProcs
                   '6P'
-                  '1PI'                 # Specializer cache
+                  '1PIP'                 # Specializer cache
                   )
         class newstyleclass(object): pass
         # Separate block for PyDictKeysObject with 8 keys and 5 entries

Lib/test/test_sys_settrace.py

@@ -1614,8 +1614,30 @@ def func():
 
         self.run_and_compare(func, EXPECTED_EVENTS)
 
-    def test_settrace_error(self):
+    def test_correct_tracing_quickened_call_class_init(self):
+
+        class C:
+            def __init__(self):
+                self
+
+        def func():
+            C()
 
+        EXPECTED_EVENTS = [
+            (0, 'call'),
+            (1, 'line'),
+            (-3, 'call'),
+            (-2, 'line'),
+            (-2, 'return'),
+            (1, 'return')]
+
+        self.run_and_compare(func, EXPECTED_EVENTS)
+        # Quicken
+        for _ in range(100):
+            func()
+        self.run_and_compare(func, EXPECTED_EVENTS)
+
+    def test_settrace_error(self):
         raised = False
         def error_once(frame, event, arg):
             nonlocal raised