bpo-36974: separate vectorcall functions for each calling convention (GH-13781)

Author: jdemeyer

Include/descrobject.h

@@ -91,9 +91,6 @@ PyAPI_FUNC(PyObject *) PyDescr_NewMember(PyTypeObject *,
 PyAPI_FUNC(PyObject *) PyDescr_NewGetSet(PyTypeObject *,
                                                struct PyGetSetDef *);
 #ifndef Py_LIMITED_API
-
-PyAPI_FUNC(PyObject *) _PyMethodDescr_Vectorcall(
-        PyObject *descrobj, PyObject *const *args, size_t nargsf, PyObject *kwnames);
 PyAPI_FUNC(PyObject *) PyDescr_NewWrapper(PyTypeObject *,
                                                 struct wrapperbase *, void *);
 #define PyDescr_IsData(d) (Py_TYPE(d)->tp_descr_set != NULL)

Include/methodobject.h

@@ -41,13 +41,6 @@ PyAPI_FUNC(int) PyCFunction_GetFlags(PyObject *);
 #endif
 PyAPI_FUNC(PyObject *) PyCFunction_Call(PyObject *, PyObject *, PyObject *);
 
-#ifndef Py_LIMITED_API
-PyAPI_FUNC(PyObject *) _PyCFunction_Vectorcall(PyObject *func,
-    PyObject *const *stack,
-    size_t nargsf,
-    PyObject *kwnames);
-#endif
-
 struct PyMethodDef {
     const char  *ml_name;   /* The name of the built-in function/method */
     PyCFunction ml_meth;    /* The C function that implements it */

Lib/test/test_call.py

@@ -586,6 +586,8 @@ def __call__(self, *args):
                 return super().__call__(*args)
 
         calls += [
+            (dict.update, ({},), {"key":True}, None),
+            ({}.update, ({},), {"key":True}, None),
             (MethodDescriptorHeap(), (0,), {}, True),
             (MethodDescriptorOverridden(), (0,), {}, 'new'),
             (MethodDescriptorSuper(), (0,), {}, True),

Lib/test/test_gdb.py

@@ -850,10 +850,10 @@ def test_pycfunction(self):
         # called, so test a variety of calling conventions.
         for py_name, py_args, c_name, expected_frame_number in (
             ('gmtime', '', 'time_gmtime', 1),  # METH_VARARGS
-            ('len', '[]', 'builtin_len', 2),  # METH_O
-            ('locals', '', 'builtin_locals', 2),  # METH_NOARGS
-            ('iter', '[]', 'builtin_iter', 2),  # METH_FASTCALL
-            ('sorted', '[]', 'builtin_sorted', 2),  # METH_FASTCALL|METH_KEYWORDS
+            ('len', '[]', 'builtin_len', 1),  # METH_O
+            ('locals', '', 'builtin_locals', 1),  # METH_NOARGS
+            ('iter', '[]', 'builtin_iter', 1),  # METH_FASTCALL
+            ('sorted', '[]', 'builtin_sorted', 1),  # METH_FASTCALL|METH_KEYWORDS
         ):
             with self.subTest(c_name):
                 cmd = ('from time import gmtime\n'  # (not always needed)

Misc/NEWS.d/next/Core and Builtins/2019-06-11-12-59-38.bpo-36974.bVYmSA.rst

@@ -0,0 +1,2 @@
+Implemented separate vectorcall functions for every calling convention of
+builtin functions and methods. This improves performance for calls.

Objects/call.c

@@ -216,7 +216,7 @@ PyVectorcall_Call(PyObject *callable, PyObject *tuple, PyObject *kwargs)
     PyObject *result = func(callable, args,
                             nargs | PY_VECTORCALL_ARGUMENTS_OFFSET, kwnames);
     _PyStack_UnpackDict_Free(args, nargs, kwnames);
-    return result;
+    return _Py_CheckFunctionResult(callable, result, NULL);
 }
 
 
@@ -625,26 +625,6 @@ _PyMethodDef_RawFastCallKeywords(PyMethodDef *method, PyObject *self,
     return result;
 }
 
-
-PyObject *
-_PyCFunction_Vectorcall(PyObject *func,
-                        PyObject *const *args, size_t nargsf,
-                        PyObject *kwnames)
-{
-    PyObject *result;
-
-    assert(func != NULL);
-    assert(PyCFunction_Check(func));
-    Py_ssize_t nargs = PyVectorcall_NARGS(nargsf);
-
-    result = _PyMethodDef_RawFastCallKeywords(((PyCFunctionObject*)func)->m_ml,
-                                              PyCFunction_GET_SELF(func),
-                                              args, nargs, kwnames);
-    result = _Py_CheckFunctionResult(func, result, NULL);
-    return result;
-}
-
-
 static PyObject *
 cfunction_call_varargs(PyObject *func, PyObject *args, PyObject *kwargs)
 {

Objects/descrobject.c

@@ -226,80 +226,199 @@ getset_set(PyGetSetDescrObject *descr, PyObject *obj, PyObject *value)
     return -1;
 }
 
-static PyObject *
-methoddescr_call(PyMethodDescrObject *descr, PyObject *args, PyObject *kwargs)
-{
-    Py_ssize_t nargs;
-    PyObject *self, *result;
 
-    /* Make sure that the first argument is acceptable as 'self' */
-    assert(PyTuple_Check(args));
-    nargs = PyTuple_GET_SIZE(args);
+/* Vectorcall functions for each of the PyMethodDescr calling conventions.
+ *
+ * First, common helpers
+ */
+static const char *
+get_name(PyObject *func) {
+    assert(PyObject_TypeCheck(func, &PyMethodDescr_Type));
+    return ((PyMethodDescrObject *)func)->d_method->ml_name;
+}
+
+typedef void (*funcptr)(void);
+
+static inline int
+method_check_args(PyObject *func, PyObject *const *args, Py_ssize_t nargs, PyObject *kwnames)
+{
+    assert(!PyErr_Occurred());
+    assert(PyObject_TypeCheck(func, &PyMethodDescr_Type));
     if (nargs < 1) {
         PyErr_Format(PyExc_TypeError,
-                     "descriptor '%V' of '%.100s' "
+                     "descriptor '%.200s' of '%.100s' "
                      "object needs an argument",
-                     descr_name((PyDescrObject *)descr), "?",
-                     PyDescr_TYPE(descr)->tp_name);
-        return NULL;
+                     get_name(func), PyDescr_TYPE(func)->tp_name);
+        return -1;
     }
-    self = PyTuple_GET_ITEM(args, 0);
+    PyObject *self = args[0];
     if (!_PyObject_RealIsSubclass((PyObject *)Py_TYPE(self),
-                                  (PyObject *)PyDescr_TYPE(descr))) {
+                                  (PyObject *)PyDescr_TYPE(func)))
+    {
         PyErr_Format(PyExc_TypeError,
-                     "descriptor '%V' for '%.100s' objects "
+                     "descriptor '%.200s' for '%.100s' objects "
                      "doesn't apply to a '%.100s' object",
-                     descr_name((PyDescrObject *)descr), "?",
-                     PyDescr_TYPE(descr)->tp_name,
-                     self->ob_type->tp_name);
+                     get_name(func), PyDescr_TYPE(func)->tp_name,
+                     Py_TYPE(self)->tp_name);
+        return -1;
+    }
+    if (kwnames && PyTuple_GET_SIZE(kwnames)) {
+        PyErr_Format(PyExc_TypeError,
+                     "%.200s() takes no keyword arguments", get_name(func));
+        return -1;
+    }
+    return 0;
+}
+
+static inline funcptr
+method_enter_call(PyObject *func)
+{
+    if (Py_EnterRecursiveCall(" while calling a Python object")) {
         return NULL;
     }
+    return (funcptr)((PyMethodDescrObject *)func)->d_method->ml_meth;
+}
 
-    result = _PyMethodDef_RawFastCallDict(descr->d_method, self,
-                                          &_PyTuple_ITEMS(args)[1], nargs - 1,
-                                          kwargs);
-    result = _Py_CheckFunctionResult((PyObject *)descr, result, NULL);
+/* Now the actual vectorcall functions */
+static PyObject *
+method_vectorcall_VARARGS(
+    PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames)
+{
+    Py_ssize_t nargs = PyVectorcall_NARGS(nargsf);
+    if (method_check_args(func, args, nargs, kwnames)) {
+        return NULL;
+    }
+    PyObject *argstuple = _PyTuple_FromArray(args+1, nargs-1);
+    if (argstuple == NULL) {
+        return NULL;
+    }
+    PyCFunction meth = (PyCFunction)method_enter_call(func);
+    if (meth == NULL) {
+        Py_DECREF(argstuple);
+        return NULL;
+    }
+    PyObject *result = meth(args[0], argstuple);
+    Py_DECREF(argstuple);
+    Py_LeaveRecursiveCall();
     return result;
 }
 
-// same to methoddescr_call(), but use FASTCALL convention.
-PyObject *
-_PyMethodDescr_Vectorcall(PyObject *descrobj,
-                          PyObject *const *args, size_t nargsf,
-                          PyObject *kwnames)
+static PyObject *
+method_vectorcall_VARARGS_KEYWORDS(
+    PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames)
 {
-    assert(Py_TYPE(descrobj) == &PyMethodDescr_Type);
-    PyMethodDescrObject *descr = (PyMethodDescrObject *)descrobj;
-    PyObject *self, *result;
+    Py_ssize_t nargs = PyVectorcall_NARGS(nargsf);
+    if (method_check_args(func, args, nargs, NULL)) {
+        return NULL;
+    }
+    PyObject *argstuple = _PyTuple_FromArray(args+1, nargs-1);
+    if (argstuple == NULL) {
+        return NULL;
+    }
+    PyObject *result = NULL;
+    /* Create a temporary dict for keyword arguments */
+    PyObject *kwdict = NULL;
+    if (kwnames != NULL && PyTuple_GET_SIZE(kwnames) > 0) {
+        kwdict = _PyStack_AsDict(args + nargs, kwnames);
+        if (kwdict == NULL) {
+            goto exit;
+        }
+    }
+    PyCFunctionWithKeywords meth = (PyCFunctionWithKeywords)
+                                   method_enter_call(func);
+    if (meth == NULL) {
+        goto exit;
+    }
+    result = meth(args[0], argstuple, kwdict);
+    Py_LeaveRecursiveCall();
+exit:
+    Py_DECREF(argstuple);
+    Py_XDECREF(kwdict);
+    return result;
+}
+
+static PyObject *
+method_vectorcall_FASTCALL(
+    PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames)
+{
+    Py_ssize_t nargs = PyVectorcall_NARGS(nargsf);
+    if (method_check_args(func, args, nargs, kwnames)) {
+        return NULL;
+    }
+    _PyCFunctionFast meth = (_PyCFunctionFast)
+                            method_enter_call(func);
+    if (meth == NULL) {
+        return NULL;
+    }
+    PyObject *result = meth(args[0], args+1, nargs-1);
+    Py_LeaveRecursiveCall();
+    return result;
+}
 
+static PyObject *
+method_vectorcall_FASTCALL_KEYWORDS(
+    PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames)
+{
     Py_ssize_t nargs = PyVectorcall_NARGS(nargsf);
-    /* Make sure that the first argument is acceptable as 'self' */
-    if (nargs < 1) {
-        PyErr_Format(PyExc_TypeError,
-                     "descriptor '%V' of '%.100s' "
-                     "object needs an argument",
-                     descr_name((PyDescrObject *)descr), "?",
-                     PyDescr_TYPE(descr)->tp_name);
+    if (method_check_args(func, args, nargs, NULL)) {
         return NULL;
     }
-    self = args[0];
-    if (!_PyObject_RealIsSubclass((PyObject *)Py_TYPE(self),
-                                  (PyObject *)PyDescr_TYPE(descr))) {
+    _PyCFunctionFastWithKeywords meth = (_PyCFunctionFastWithKeywords)
+                                        method_enter_call(func);
+    if (meth == NULL) {
+        return NULL;
+    }
+    PyObject *result = meth(args[0], args+1, nargs-1, kwnames);
+    Py_LeaveRecursiveCall();
+    return result;
+}
+
+static PyObject *
+method_vectorcall_NOARGS(
+    PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames)
+{
+    Py_ssize_t nargs = PyVectorcall_NARGS(nargsf);
+    if (method_check_args(func, args, nargs, kwnames)) {
+        return NULL;
+    }
+    if (nargs != 1) {
         PyErr_Format(PyExc_TypeError,
-                     "descriptor '%V' for '%.100s' objects "
-                     "doesn't apply to a '%.100s' object",
-                     descr_name((PyDescrObject *)descr), "?",
-                     PyDescr_TYPE(descr)->tp_name,
-                     self->ob_type->tp_name);
+            "%.200s() takes no arguments (%zd given)", get_name(func), nargs-1);
+        return NULL;
+    }
+    PyCFunction meth = (PyCFunction)method_enter_call(func);
+    if (meth == NULL) {
         return NULL;
     }
+    PyObject *result = meth(args[0], NULL);
+    Py_LeaveRecursiveCall();
+    return result;
+}
 
-    result = _PyMethodDef_RawFastCallKeywords(descr->d_method, self,
-                                              args+1, nargs-1, kwnames);
-    result = _Py_CheckFunctionResult((PyObject *)descr, result, NULL);
+static PyObject *
+method_vectorcall_O(
+    PyObject *func, PyObject *const *args, size_t nargsf, PyObject *kwnames)
+{
+    Py_ssize_t nargs = PyVectorcall_NARGS(nargsf);
+    if (method_check_args(func, args, nargs, kwnames)) {
+        return NULL;
+    }
+    if (nargs != 2) {
+        PyErr_Format(PyExc_TypeError,
+            "%.200s() takes exactly one argument (%zd given)",
+            get_name(func), nargs-1);
+        return NULL;
+    }
+    PyCFunction meth = (PyCFunction)method_enter_call(func);
+    if (meth == NULL) {
+        return NULL;
+    }
+    PyObject *result = meth(args[0], args[1]);
+    Py_LeaveRecursiveCall();
     return result;
 }
 
+
 /* Instances of classmethod_descriptor are unlikely to be called directly.
    For one, the analogous class "classmethod" (for Python classes) is not
    callable. Second, users are not likely to access a classmethod_descriptor
@@ -540,7 +659,7 @@ PyTypeObject PyMethodDescr_Type = {
     0,                                          /* tp_as_sequence */
     0,                                          /* tp_as_mapping */
     0,                                          /* tp_hash */
-    (ternaryfunc)methoddescr_call,              /* tp_call */
+    PyVectorcall_Call,                          /* tp_call */
     0,                                          /* tp_str */
     PyObject_GenericGetAttr,                    /* tp_getattro */
     0,                                          /* tp_setattro */
@@ -738,13 +857,40 @@ descr_new(PyTypeObject *descrtype, PyTypeObject *type, const char *name)
 PyObject *
 PyDescr_NewMethod(PyTypeObject *type, PyMethodDef *method)
 {
+    /* Figure out correct vectorcall function to use */
+    vectorcallfunc vectorcall;
+    switch (method->ml_flags & (METH_VARARGS | METH_FASTCALL | METH_NOARGS | METH_O | METH_KEYWORDS))
+    {
+        case METH_VARARGS:
+            vectorcall = method_vectorcall_VARARGS;
+            break;
+        case METH_VARARGS | METH_KEYWORDS:
+            vectorcall = method_vectorcall_VARARGS_KEYWORDS;
+            break;
+        case METH_FASTCALL:
+            vectorcall = method_vectorcall_FASTCALL;
+            break;
+        case METH_FASTCALL | METH_KEYWORDS:
+            vectorcall = method_vectorcall_FASTCALL_KEYWORDS;
+            break;
+        case METH_NOARGS:
+            vectorcall = method_vectorcall_NOARGS;
+            break;
+        case METH_O:
+            vectorcall = method_vectorcall_O;
+            break;
+        default:
+            PyErr_SetString(PyExc_SystemError, "bad call flags");
+            return NULL;
+    }
+
     PyMethodDescrObject *descr;
 
     descr = (PyMethodDescrObject *)descr_new(&PyMethodDescr_Type,
                                              type, method->ml_name);
     if (descr != NULL) {
         descr->d_method = method;
-        descr->vectorcall = _PyMethodDescr_Vectorcall;
+        descr->vectorcall = vectorcall;
     }
     return (PyObject *)descr;
 }