bpo-42808: Add PyType_Type.tp_vectorcall for type(obj) performance (GH-24058)

sweeneyde · web-flow · commit b19855bb6ffd · 2021-02-22T11:59:16.000+09:00
diff --git a/Misc/NEWS.d/next/Core and Builtins/2021-01-02-05-10-58.bpo-42808.AOxgxl.rst b/Misc/NEWS.d/next/Core and Builtins/2021-01-02-05-10-58.bpo-42808.AOxgxl.rst
@@ -0,0 +1,2 @@
+Simple calls to ``type(object)`` are now faster due to the
+``vectorcall`` calling convention. Patch by Dennis Sweeney.
diff --git a/Objects/typeobject.c b/Objects/typeobject.c
@@ -2888,6 +2888,23 @@ type_new(PyTypeObject *metatype, PyObject *args, PyObject *kwds)
     return NULL;
 }
 
+static PyObject *
+type_vectorcall(PyObject *metatype, PyObject *const *args,
+                 size_t nargsf, PyObject *kwnames)
+{
+    Py_ssize_t nargs = PyVectorcall_NARGS(nargsf);
+    if (nargs == 1 && metatype == (PyObject *)&PyType_Type){
+        if (!_PyArg_NoKwnames("type", kwnames)) {
+            return NULL;
+        }
+        return Py_NewRef(Py_TYPE(args[0]));
+    }
+    /* In other (much less common) cases, fall back to
+       more flexible calling conventions. */
+    PyThreadState *tstate = PyThreadState_GET();
+    return _PyObject_MakeTpCall(tstate, metatype, args, nargs, kwnames);
+}
+
 /* An array of type slot offsets corresponding to Py_tp_* constants,
   * for use in e.g. PyType_Spec and PyType_GetSlot.
   * Each entry has two offsets: "slot_offset" and "subslot_offset".
@@ -3896,6 +3913,7 @@ PyTypeObject PyType_Type = {
     type_new,                                   /* tp_new */
     PyObject_GC_Del,                            /* tp_free */
     (inquiry)type_is_gc,                        /* tp_is_gc */
+    .tp_vectorcall = type_vectorcall,
 };
 
 

Original file line number	Diff line number	Diff line change
`@@ -0,0 +1,2 @@`
	`1`	+Simple calls to ``type(object)`` are now faster due to the
	`2`	+``vectorcall`` calling convention. Patch by Dennis Sweeney.