gh-87729: add LOAD_SUPER_ATTR instruction for faster super() (#103497)

This speeds up `super()` (by around 85%, for a simple one-level
`super().meth()` microbenchmark) by avoiding allocation of a new
single-use `super()` object on each use.

Doc/library/dis.rst

@@ -1036,6 +1036,24 @@ iterations of the loop.
       pushed to the stack before the attribute or unbound method respectively.
 
 
+.. opcode:: LOAD_SUPER_ATTR (namei)
+
+   This opcode implements :func:`super` (e.g. ``super().method()`` and
+   ``super().attr``). It works the same as :opcode:`LOAD_ATTR`, except that
+   ``namei`` is shifted left by 2 bits instead of 1, and instead of expecting a
+   single receiver on the stack, it expects three objects (from top of stack
+   down): ``self`` (the first argument to the current method), ``cls`` (the
+   class within which the current method was defined), and the global ``super``.
+
+   The low bit of ``namei`` signals to attempt a method load, as with
+   :opcode:`LOAD_ATTR`.
+
+   The second-low bit of ``namei``, if set, means that this was a two-argument
+   call to :func:`super` (unset means zero-argument).
+
+   .. versionadded:: 3.12
+
+
 .. opcode:: COMPARE_OP (opname)
 
    Performs a Boolean operation.  The operation name can be found in

Include/internal/pycore_opcode.h

@@ -179,6 +179,7 @@ const uint8_t _PyOpcode_Deopt[256] = {
     [LOAD_GLOBAL_BUILTIN] = LOAD_GLOBAL,
     [LOAD_GLOBAL_MODULE] = LOAD_GLOBAL,
     [LOAD_NAME] = LOAD_NAME,
+    [LOAD_SUPER_ATTR] = LOAD_SUPER_ATTR,
     [MAKE_CELL] = MAKE_CELL,
     [MAKE_FUNCTION] = MAKE_FUNCTION,
     [MAP_ADD] = MAP_ADD,
@@ -236,7 +237,7 @@ const uint8_t _PyOpcode_Deopt[256] = {
 #endif   // NEED_OPCODE_TABLES
 
 #ifdef Py_DEBUG
-static const char *const _PyOpcode_OpName[263] = {
+static const char *const _PyOpcode_OpName[266] = {
     [CACHE] = "CACHE",
     [POP_TOP] = "POP_TOP",
     [PUSH_NULL] = "PUSH_NULL",
@@ -378,9 +379,9 @@ static const char *const _PyOpcode_OpName[263] = {
     [STORE_DEREF] = "STORE_DEREF",
     [DELETE_DEREF] = "DELETE_DEREF",
     [JUMP_BACKWARD] = "JUMP_BACKWARD",
-    [STORE_FAST__STORE_FAST] = "STORE_FAST__STORE_FAST",
+    [LOAD_SUPER_ATTR] = "LOAD_SUPER_ATTR",
     [CALL_FUNCTION_EX] = "CALL_FUNCTION_EX",
-    [STORE_SUBSCR_DICT] = "STORE_SUBSCR_DICT",
+    [STORE_FAST__STORE_FAST] = "STORE_FAST__STORE_FAST",
     [EXTENDED_ARG] = "EXTENDED_ARG",
     [LIST_APPEND] = "LIST_APPEND",
     [SET_ADD] = "SET_ADD",
@@ -390,15 +391,15 @@ static const char *const _PyOpcode_OpName[263] = {
     [YIELD_VALUE] = "YIELD_VALUE",
     [RESUME] = "RESUME",
     [MATCH_CLASS] = "MATCH_CLASS",
+    [STORE_SUBSCR_DICT] = "STORE_SUBSCR_DICT",
     [STORE_SUBSCR_LIST_INT] = "STORE_SUBSCR_LIST_INT",
-    [UNPACK_SEQUENCE_LIST] = "UNPACK_SEQUENCE_LIST",
     [FORMAT_VALUE] = "FORMAT_VALUE",
     [BUILD_CONST_KEY_MAP] = "BUILD_CONST_KEY_MAP",
     [BUILD_STRING] = "BUILD_STRING",
+    [UNPACK_SEQUENCE_LIST] = "UNPACK_SEQUENCE_LIST",
     [UNPACK_SEQUENCE_TUPLE] = "UNPACK_SEQUENCE_TUPLE",
     [UNPACK_SEQUENCE_TWO_TUPLE] = "UNPACK_SEQUENCE_TWO_TUPLE",
     [SEND_GEN] = "SEND_GEN",
-    [161] = "<161>",
     [LIST_EXTEND] = "LIST_EXTEND",
     [SET_UPDATE] = "SET_UPDATE",
     [DICT_MERGE] = "DICT_MERGE",
@@ -500,11 +501,13 @@ static const char *const _PyOpcode_OpName[263] = {
     [JUMP] = "JUMP",
     [JUMP_NO_INTERRUPT] = "JUMP_NO_INTERRUPT",
     [LOAD_METHOD] = "LOAD_METHOD",
+    [LOAD_SUPER_METHOD] = "LOAD_SUPER_METHOD",
+    [LOAD_ZERO_SUPER_METHOD] = "LOAD_ZERO_SUPER_METHOD",
+    [LOAD_ZERO_SUPER_ATTR] = "LOAD_ZERO_SUPER_ATTR",
 };
 #endif
 
 #define EXTRA_CASES \
-    case 161: \
     case 166: \
     case 167: \
     case 168: \

Include/internal/pycore_typeobject.h

@@ -98,6 +98,9 @@ _Py_type_getattro(PyTypeObject *type, PyObject *name);
 PyObject *_Py_slot_tp_getattro(PyObject *self, PyObject *name);
 PyObject *_Py_slot_tp_getattr_hook(PyObject *self, PyObject *name);
 
+PyObject *
+_PySuper_Lookup(PyTypeObject *su_type, PyObject *su_obj, PyObject *name, int *meth_found);
+
 #ifdef __cplusplus
 }
 #endif

Include/opcode.h

@@ -95,6 +95,7 @@ extern "C" {
 #define STORE_DEREF                            138
 #define DELETE_DEREF                           139
 #define JUMP_BACKWARD                          140
+#define LOAD_SUPER_ATTR                        141
 #define CALL_FUNCTION_EX                       142
 #define EXTENDED_ARG                           144
 #define LIST_APPEND                            145
@@ -142,7 +143,10 @@ extern "C" {
 #define JUMP                                   260
 #define JUMP_NO_INTERRUPT                      261
 #define LOAD_METHOD                            262
-#define MAX_PSEUDO_OPCODE                      262
+#define LOAD_SUPER_METHOD                      263
+#define LOAD_ZERO_SUPER_METHOD                 264
+#define LOAD_ZERO_SUPER_ATTR                   265
+#define MAX_PSEUDO_OPCODE                      265
 #define BINARY_OP_ADD_FLOAT                      6
 #define BINARY_OP_ADD_INT                        7
 #define BINARY_OP_ADD_UNICODE                    8
@@ -198,18 +202,21 @@ extern "C" {
 #define STORE_ATTR_SLOT                        111
 #define STORE_ATTR_WITH_HINT                   112
 #define STORE_FAST__LOAD_FAST                  113
-#define STORE_FAST__STORE_FAST                 141
+#define STORE_FAST__STORE_FAST                 143
-#define STORE_SUBSCR_DICT                      143
+#define STORE_SUBSCR_DICT                      153
-#define STORE_SUBSCR_LIST_INT                  153
+#define STORE_SUBSCR_LIST_INT                  154
-#define UNPACK_SEQUENCE_LIST                   154
+#define UNPACK_SEQUENCE_LIST                   158
-#define UNPACK_SEQUENCE_TUPLE                  158
+#define UNPACK_SEQUENCE_TUPLE                  159
-#define UNPACK_SEQUENCE_TWO_TUPLE              159
+#define UNPACK_SEQUENCE_TWO_TUPLE              160
-#define SEND_GEN                               160
+#define SEND_GEN                               161
 
 #define HAS_ARG(op) ((((op) >= HAVE_ARGUMENT) && (!IS_PSEUDO_OPCODE(op)))\
     || ((op) == JUMP) \
     || ((op) == JUMP_NO_INTERRUPT) \
     || ((op) == LOAD_METHOD) \
+    || ((op) == LOAD_SUPER_METHOD) \
+    || ((op) == LOAD_ZERO_SUPER_METHOD) \
+    || ((op) == LOAD_ZERO_SUPER_ATTR) \
     )
 
 #define HAS_CONST(op) (false\

Lib/dis.py

@@ -41,6 +41,7 @@ JUMP_BACKWARD = opmap['JUMP_BACKWARD']
 FOR_ITER = opmap['FOR_ITER']
 SEND = opmap['SEND']
 LOAD_ATTR = opmap['LOAD_ATTR']
+LOAD_SUPER_ATTR = opmap['LOAD_SUPER_ATTR']
 
 CACHE = opmap["CACHE"]
 
@@ -475,6 +476,10 @@ def _get_instructions_bytes(code, varname_from_oparg=None,
                     argval, argrepr = _get_name_info(arg//2, get_name)
                     if (arg & 1) and argrepr:
                         argrepr = "NULL|self + " + argrepr
+                elif deop == LOAD_SUPER_ATTR:
+                    argval, argrepr = _get_name_info(arg//4, get_name)
+                    if (arg & 1) and argrepr:
+                        argrepr = "NULL|self + " + argrepr
                 else:
                     argval, argrepr = _get_name_info(arg, get_name)
             elif deop in hasjabs:

Lib/importlib/_bootstrap_external.py

@@ -439,7 +439,8 @@ _code_type = type(_write_atomic.__code__)
 #     Python 3.12a7 3523 (Convert COMPARE_AND_BRANCH back to COMPARE_OP)
 #     Python 3.12a7 3524 (Shrink the BINARY_SUBSCR caches)
 #     Python 3.12b1 3525 (Shrink the CALL caches)
-#     Python 3.12a7 3526 (Add instrumentation support)
+#     Python 3.12b1 3526 (Add instrumentation support)
+#     Python 3.12b1 3527 (Optimize super() calls)
 
 #     Python 3.13 will start with 3550
 
@@ -456,7 +457,7 @@ _code_type = type(_write_atomic.__code__)
 # Whenever MAGIC_NUMBER is changed, the ranges in the magic_values array
 # in PC/launcher.c must also be updated.
 
-MAGIC_NUMBER = (3526).to_bytes(2, 'little') + b'\r\n'
+MAGIC_NUMBER = (3527).to_bytes(2, 'little') + b'\r\n'
 
 _RAW_MAGIC_NUMBER = int.from_bytes(MAGIC_NUMBER, 'little')  # For import.c
 

Lib/opcode.py

@@ -196,7 +196,7 @@ hasfree.append(138)
 def_op('DELETE_DEREF', 139)
 hasfree.append(139)
 jrel_op('JUMP_BACKWARD', 140)    # Number of words to skip (backwards)
-
+name_op('LOAD_SUPER_ATTR', 141)
 def_op('CALL_FUNCTION_EX', 142)  # Flags
 
 def_op('EXTENDED_ARG', 144)
@@ -264,6 +264,9 @@ pseudo_op('JUMP', 260, ['JUMP_FORWARD', 'JUMP_BACKWARD'])
 pseudo_op('JUMP_NO_INTERRUPT', 261, ['JUMP_FORWARD', 'JUMP_BACKWARD_NO_INTERRUPT'])
 
 pseudo_op('LOAD_METHOD', 262, ['LOAD_ATTR'])
+pseudo_op('LOAD_SUPER_METHOD', 263, ['LOAD_SUPER_ATTR'])
+pseudo_op('LOAD_ZERO_SUPER_METHOD', 264, ['LOAD_SUPER_ATTR'])
+pseudo_op('LOAD_ZERO_SUPER_ATTR', 265, ['LOAD_SUPER_ATTR'])
 
 MAX_PSEUDO_OPCODE = MIN_PSEUDO_OPCODE + len(_pseudo_ops) - 1
 

Lib/test/shadowed_super.py

@@ -0,0 +1,7 @@
+class super:
+    msg = "truly super"
+
+
+class C:
+    def method(self):
+        return super().msg

Lib/test/test_gdb.py

@@ -962,7 +962,7 @@ id(42)
         cmd = textwrap.dedent('''
             class MyList(list):
                 def __init__(self):
-                    super().__init__()   # wrapper_call()
+                    super(*[]).__init__()   # wrapper_call()
 
             id("first break point")
             l = MyList()

Lib/test/test_super.py

@@ -1,6 +1,8 @@
 """Unit tests for zero-argument super() & related machinery."""
 
 import unittest
+from unittest.mock import patch
+from test import shadowed_super
 
 
 class A:
@@ -283,17 +285,28 @@ class TestSuper(unittest.TestCase):
     def test_obscure_super_errors(self):
         def f():
             super()
-        self.assertRaises(RuntimeError, f)
+        with self.assertRaisesRegex(RuntimeError, r"no arguments"):
+            f()
+
+        class C:
+            def f():
+                super()
+        with self.assertRaisesRegex(RuntimeError, r"no arguments"):
+            C.f()
+
         def f(x):
             del x
             super()
-        self.assertRaises(RuntimeError, f, None)
+        with self.assertRaisesRegex(RuntimeError, r"arg\[0\] deleted"):
+            f(None)
+
         class X:
             def f(x):
                 nonlocal __class__
                 del __class__
                 super()
-        self.assertRaises(RuntimeError, X().f)
+        with self.assertRaisesRegex(RuntimeError, r"empty __class__ cell"):
+            X().f()
 
     def test_cell_as_self(self):
         class X:
@@ -325,6 +338,78 @@ class TestSuper(unittest.TestCase):
         with self.assertRaisesRegex(TypeError, "argument 1 must be a type"):
             super(1, int)
 
+    def test_shadowed_global(self):
+        self.assertEqual(shadowed_super.C().method(), "truly super")
+
+    def test_shadowed_local(self):
+        class super:
+            msg = "quite super"
+
+        class C:
+            def method(self):
+                return super().msg
+
+        self.assertEqual(C().method(), "quite super")
+
+    def test_shadowed_dynamic(self):
+        class MySuper:
+            msg = "super super"
+
+        class C:
+            def method(self):
+                return super().msg
+
+        with patch("test.test_super.super", MySuper) as m:
+            self.assertEqual(C().method(), "super super")
+
+    def test_shadowed_dynamic_two_arg(self):
+        call_args = []
+        class MySuper:
+            def __init__(self, *args):
+                call_args.append(args)
+            msg = "super super"
+
+        class C:
+            def method(self):
+                return super(1, 2).msg
+
+        with patch("test.test_super.super", MySuper) as m:
+            self.assertEqual(C().method(), "super super")
+            self.assertEqual(call_args, [(1, 2)])
+
+    def test_attribute_error(self):
+        class C:
+            def method(self):
+                return super().msg
+
+        with self.assertRaisesRegex(AttributeError, "'super' object has no attribute 'msg'"):
+            C().method()
+
+    def test_bad_first_arg(self):
+        class C:
+            def method(self):
+                return super(1, self).method()
+
+        with self.assertRaisesRegex(TypeError, "argument 1 must be a type"):
+            C().method()
+
+    def test_super___class__(self):
+        class C:
+            def method(self):
+                return super().__class__
+
+        self.assertEqual(C().method(), super)
+
+    def test_super_subclass___class__(self):
+        class mysuper(super):
+            pass
+
+        class C:
+            def method(self):
+                return mysuper(C, self).__class__
+
+        self.assertEqual(C().method(), mysuper)
+
 
 if __name__ == "__main__":
     unittest.main()

Misc/NEWS.d/next/Core and Builtins/2023-04-12-20-22-03.gh-issue-87729.99A7UO.rst

@@ -0,0 +1 @@
+Add :opcode:`LOAD_SUPER_ATTR` to speed up ``super().meth()`` and ``super().attr`` calls.

Objects/typeobject.c

@@ -9346,26 +9346,20 @@ super_repr(PyObject *self)
             su->type ? su->type->tp_name : "NULL");
 }
 
+// if `method` is non-NULL, we are looking for a method descriptor,
+// and setting `*method` to 1 means we found one.
 static PyObject *
-super_getattro(PyObject *self, PyObject *name)
+do_super_lookup(superobject *su, PyTypeObject *su_type, PyObject *su_obj,
+                PyTypeObject *su_obj_type, PyObject *name, int *method)
 {
-    superobject *su = (superobject *)self;
+    PyObject *mro, *res;
-    PyTypeObject *starttype;
-    PyObject *mro;
     Py_ssize_t i, n;
+    int temp_su = 0;
 
-    starttype = su->obj_type;
+    if (su_obj_type == NULL)
-    if (starttype == NULL)
         goto skip;
 
-    /* We want __class__ to return the class of the super object
+    mro = su_obj_type->tp_mro;
-       (i.e. super, or a subclass), not the class of su->obj. */
-    if (PyUnicode_Check(name) &&
-        PyUnicode_GET_LENGTH(name) == 9 &&
-        _PyUnicode_Equal(name, &_Py_ID(__class__)))
-        goto skip;
-
-    mro = starttype->tp_mro;
     if (mro == NULL)
         goto skip;
 
@@ -9374,14 +9368,14 @@ super_getattro(PyObject *self, PyObject *name)
 
     /* No need to check the last one: it's gonna be skipped anyway.  */
     for (i = 0; i+1 < n; i++) {
-        if ((PyObject *)(su->type) == PyTuple_GET_ITEM(mro, i))
+        if ((PyObject *)(su_type) == PyTuple_GET_ITEM(mro, i))
             break;
     }
     i++;  /* skip su->type (if any)  */
     if (i >= n)
         goto skip;
 
-    /* keep a strong reference to mro because starttype->tp_mro can be
+    /* keep a strong reference to mro because su_obj_type->tp_mro can be
        replaced during PyDict_GetItemWithError(dict, name)  */
     Py_INCREF(mro);
     do {
@@ -9389,19 +9383,23 @@ super_getattro(PyObject *self, PyObject *name)
         PyObject *dict = _PyType_CAST(obj)->tp_dict;
         assert(dict != NULL && PyDict_Check(dict));
 
-        PyObject *res = PyDict_GetItemWithError(dict, name);
+        res = PyDict_GetItemWithError(dict, name);
         if (res != NULL) {
             Py_INCREF(res);
-
+            if (method && _PyType_HasFeature(Py_TYPE(res), Py_TPFLAGS_METHOD_DESCRIPTOR)) {
-            descrgetfunc f = Py_TYPE(res)->tp_descr_get;
+                *method = 1;
-            if (f != NULL) {
+            }
-                PyObject *res2;
+            else {
-                res2 = f(res,
+                descrgetfunc f = Py_TYPE(res)->tp_descr_get;
-                    /* Only pass 'obj' param if this is instance-mode super
+                if (f != NULL) {
-                       (See SF ID #743627)  */
+                    PyObject *res2;
-                    (su->obj == (PyObject *)starttype) ? NULL : su->obj,
+                    res2 = f(res,
-                    (PyObject *)starttype);
+                        /* Only pass 'obj' param if this is instance-mode super
-                Py_SETREF(res, res2);
+                        (See SF ID #743627)  */
+                        (su_obj == (PyObject *)su_obj_type) ? NULL : su_obj,
+                        (PyObject *)su_obj_type);
+                    Py_SETREF(res, res2);
+                }
             }
 
             Py_DECREF(mro);
@@ -9417,7 +9415,34 @@ super_getattro(PyObject *self, PyObject *name)
     Py_DECREF(mro);
 
   skip:
-    return PyObject_GenericGetAttr(self, name);
+    if (su == NULL) {
+        PyObject *args[] = {(PyObject *)su_type, su_obj};
+        su = (superobject *)PyObject_Vectorcall((PyObject *)&PySuper_Type, args, 2, NULL);
+        if (su == NULL) {
+            return NULL;
+        }
+        temp_su = 1;
+    }
+    res = PyObject_GenericGetAttr((PyObject *)su, name);
+    if (temp_su) {
+        Py_DECREF(su);
+    }
+    return res;
+}
+
+static PyObject *
+super_getattro(PyObject *self, PyObject *name)
+{
+    superobject *su = (superobject *)self;
+
+    /* We want __class__ to return the class of the super object
+       (i.e. super, or a subclass), not the class of su->obj. */
+    if (PyUnicode_Check(name) &&
+        PyUnicode_GET_LENGTH(name) == 9 &&
+        _PyUnicode_Equal(name, &_Py_ID(__class__)))
+        return PyObject_GenericGetAttr(self, name);
+
+    return do_super_lookup(su, su->type, su->obj, su->obj_type, name, NULL);
 }
 
 static PyTypeObject *
@@ -9473,6 +9498,18 @@ supercheck(PyTypeObject *type, PyObject *obj)
     return NULL;
 }
 
+PyObject *
+_PySuper_Lookup(PyTypeObject *su_type, PyObject *su_obj, PyObject *name, int *method)
+{
+    PyTypeObject *su_obj_type = supercheck(su_type, su_obj);
+    if (su_obj_type == NULL) {
+        return NULL;
+    }
+    PyObject *res = do_super_lookup(NULL, su_type, su_obj, su_obj_type, name, method);
+    Py_DECREF(su_obj_type);
+    return res;
+}
+
 static PyObject *
 super_descr_get(PyObject *self, PyObject *obj, PyObject *type)
 {

Python/bytecodes.c

@@ -25,6 +25,7 @@
 #include "pycore_sliceobject.h"   // _PyBuildSlice_ConsumeRefs
 #include "pycore_sysmodule.h"     // _PySys_Audit()
 #include "pycore_tuple.h"         // _PyTuple_ITEMS()
+#include "pycore_typeobject.h"    // _PySuper_Lookup()
 #include "pycore_emscripten_signal.h"  // _Py_CHECK_EMSCRIPTEN_SIGNALS
 
 #include "pycore_dict.h"
@@ -1553,6 +1554,36 @@ dummy_func(
             PREDICT(JUMP_BACKWARD);
         }
 
+        inst(LOAD_SUPER_ATTR, (global_super, class, self -- res2 if (oparg & 1), res)) {
+            PyObject *name = GETITEM(frame->f_code->co_names, oparg >> 2);
+            if (global_super == (PyObject *)&PySuper_Type && PyType_Check(class)) {
+                int method = 0;
+                Py_DECREF(global_super);
+                res = _PySuper_Lookup((PyTypeObject *)class, self, name, oparg & 1 ? &method : NULL);
+                Py_DECREF(class);
+                if (res == NULL) {
+                    Py_DECREF(self);
+                    ERROR_IF(true, error);
+                }
+                // Works with CALL, pushes two values: either `meth | self` or `NULL | meth`.
+                if (method) {
+                    res2 = res;
+                    res = self;  // transfer ownership
+                } else {
+                    res2 = NULL;
+                    Py_DECREF(self);
+                }
+            } else {
+                PyObject *stack[] = {class, self};
+                PyObject *super = PyObject_Vectorcall(global_super, stack, oparg & 2, NULL);
+                DECREF_INPUTS();
+                ERROR_IF(super == NULL, error);
+                res = PyObject_GetAttr(super, name);
+                Py_DECREF(super);
+                ERROR_IF(res == NULL, error);
+            }
+        }
+
         family(load_attr, INLINE_CACHE_ENTRIES_LOAD_ATTR) = {
             LOAD_ATTR,
             LOAD_ATTR_INSTANCE_VALUE,

Python/ceval.c

@@ -21,6 +21,7 @@
 #include "pycore_sliceobject.h"   // _PyBuildSlice_ConsumeRefs
 #include "pycore_sysmodule.h"     // _PySys_Audit()
 #include "pycore_tuple.h"         // _PyTuple_ITEMS()
+#include "pycore_typeobject.h"    // _PySuper_Lookup()
 #include "pycore_emscripten_signal.h"  // _Py_CHECK_EMSCRIPTEN_SIGNALS
 
 #include "pycore_dict.h"

Python/compile.c

@@ -829,6 +829,10 @@ stack_effect(int opcode, int oparg, int jump)
 
         case LOAD_METHOD:
             return 1;
+        case LOAD_SUPER_METHOD:
+        case LOAD_ZERO_SUPER_METHOD:
+        case LOAD_ZERO_SUPER_ATTR:
+            return -1;
         default:
             return PY_INVALID_STACK_EFFECT;
     }
@@ -1047,6 +1051,24 @@ compiler_addop_name(struct compiler_unit *u, location loc,
         arg <<= 1;
         arg |= 1;
     }
+    if (opcode == LOAD_SUPER_ATTR) {
+        arg <<= 2;
+        arg |= 2;
+    }
+    if (opcode == LOAD_SUPER_METHOD) {
+        opcode = LOAD_SUPER_ATTR;
+        arg <<= 2;
+        arg |= 3;
+    }
+    if (opcode == LOAD_ZERO_SUPER_ATTR) {
+        opcode = LOAD_SUPER_ATTR;
+        arg <<= 2;
+    }
+    if (opcode == LOAD_ZERO_SUPER_METHOD) {
+        opcode = LOAD_SUPER_ATTR;
+        arg <<= 2;
+        arg |= 1;
+    }
     return codegen_addop_i(&u->u_instr_sequence, opcode, arg, loc);
 }
 
@@ -4230,6 +4252,89 @@ is_import_originated(struct compiler *c, expr_ty e)
     return flags & DEF_IMPORT;
 }
 
+static int
+can_optimize_super_call(struct compiler *c, expr_ty attr)
+{
+    expr_ty e = attr->v.Attribute.value;
+    if (e->kind != Call_kind ||
+        e->v.Call.func->kind != Name_kind ||
+        !_PyUnicode_EqualToASCIIString(e->v.Call.func->v.Name.id, "super") ||
+        _PyUnicode_EqualToASCIIString(attr->v.Attribute.attr, "__class__") ||
+        asdl_seq_LEN(e->v.Call.keywords) != 0) {
+        return 0;
+    }
+    Py_ssize_t num_args = asdl_seq_LEN(e->v.Call.args);
+
+    PyObject *super_name = e->v.Call.func->v.Name.id;
+    // detect statically-visible shadowing of 'super' name
+    int scope = _PyST_GetScope(c->u->u_ste, super_name);
+    if (scope != GLOBAL_IMPLICIT) {
+        return 0;
+    }
+    scope = _PyST_GetScope(c->c_st->st_top, super_name);
+    if (scope != 0) {
+        return 0;
+    }
+
+    if (num_args == 2) {
+        for (Py_ssize_t i = 0; i < num_args; i++) {
+            expr_ty elt = asdl_seq_GET(e->v.Call.args, i);
+            if (elt->kind == Starred_kind) {
+                return 0;
+            }
+        }
+        // exactly two non-starred args; we can just load
+        // the provided args
+        return 1;
+    }
+
+    if (num_args != 0) {
+        return 0;
+    }
+    // we need the following for zero-arg super():
+
+    // enclosing function should have at least one argument
+    if (c->u->u_metadata.u_argcount == 0 &&
+        c->u->u_metadata.u_posonlyargcount == 0) {
+        return 0;
+    }
+    // __class__ cell should be available
+    if (get_ref_type(c, &_Py_ID(__class__)) == FREE) {
+        return 1;
+    }
+    return 0;
+}
+
+static int
+load_args_for_super(struct compiler *c, expr_ty e) {
+    location loc = LOC(e);
+
+    // load super() global
+    PyObject *super_name = e->v.Call.func->v.Name.id;
+    RETURN_IF_ERROR(compiler_nameop(c, loc, super_name, Load));
+
+    if (asdl_seq_LEN(e->v.Call.args) == 2) {
+        VISIT(c, expr, asdl_seq_GET(e->v.Call.args, 0));
+        VISIT(c, expr, asdl_seq_GET(e->v.Call.args, 1));
+        return SUCCESS;
+    }
+
+    // load __class__ cell
+    PyObject *name = &_Py_ID(__class__);
+    assert(get_ref_type(c, name) == FREE);
+    RETURN_IF_ERROR(compiler_nameop(c, loc, name, Load));
+
+    // load self (first argument)
+    Py_ssize_t i = 0;
+    PyObject *key, *value;
+    if (!PyDict_Next(c->u->u_metadata.u_varnames, &i, &key, &value)) {
+        return ERROR;
+    }
+    RETURN_IF_ERROR(compiler_nameop(c, loc, key, Load));
+
+    return SUCCESS;
+}
+
 // If an attribute access spans multiple lines, update the current start
 // location to point to the attribute name.
 static location
@@ -4297,11 +4402,21 @@ maybe_optimize_method_call(struct compiler *c, expr_ty e)
             return 0;
         }
     }
+
     /* Alright, we can optimize the code. */
-    VISIT(c, expr, meth->v.Attribute.value);
     location loc = LOC(meth);
-    loc = update_start_location_to_match_attr(c, loc, meth);
+
-    ADDOP_NAME(c, loc, LOAD_METHOD, meth->v.Attribute.attr, names);
+    if (can_optimize_super_call(c, meth)) {
+        RETURN_IF_ERROR(load_args_for_super(c, meth->v.Attribute.value));
+        int opcode = asdl_seq_LEN(meth->v.Attribute.value->v.Call.args) ?
+            LOAD_SUPER_METHOD : LOAD_ZERO_SUPER_METHOD;
+        ADDOP_NAME(c, loc, opcode, meth->v.Attribute.attr, names);
+    } else {
+        VISIT(c, expr, meth->v.Attribute.value);
+        loc = update_start_location_to_match_attr(c, loc, meth);
+        ADDOP_NAME(c, loc, LOAD_METHOD, meth->v.Attribute.attr, names);
+    }
+
     VISIT_SEQ(c, expr, e->v.Call.args);
 
     if (kwdsl) {
@@ -5309,6 +5424,13 @@ compiler_visit_expr1(struct compiler *c, expr_ty e)
         return compiler_formatted_value(c, e);
     /* The following exprs can be assignment targets. */
     case Attribute_kind:
+        if (e->v.Attribute.ctx == Load && can_optimize_super_call(c, e)) {
+            RETURN_IF_ERROR(load_args_for_super(c, e->v.Attribute.value));
+            int opcode = asdl_seq_LEN(e->v.Attribute.value->v.Call.args) ?
+                LOAD_SUPER_ATTR : LOAD_ZERO_SUPER_ATTR;
+            ADDOP_NAME(c, loc, opcode, e->v.Attribute.attr, names);
+            return SUCCESS;
+        }
         VISIT(c, expr, e->v.Attribute.value);
         loc = LOC(e);
         loc = update_start_location_to_match_attr(c, loc, e);

Python/opcode_metadata.h

@@ -205,6 +205,8 @@ _PyOpcode_num_popped(int opcode, int oparg, bool jump) {
             return 1;
         case MAP_ADD:
             return 2;
+        case LOAD_SUPER_ATTR:
+            return 3;
         case LOAD_ATTR:
             return 1;
         case LOAD_ATTR_INSTANCE_VALUE:
@@ -589,6 +591,8 @@ _PyOpcode_num_pushed(int opcode, int oparg, bool jump) {
             return 0;
         case MAP_ADD:
             return 0;
+        case LOAD_SUPER_ATTR:
+            return ((oparg & 1) ? 1 : 0) + 1;
         case LOAD_ATTR:
             return ((oparg & 1) ? 1 : 0) + 1;
         case LOAD_ATTR_INSTANCE_VALUE:
@@ -879,6 +883,7 @@ const struct opcode_metadata _PyOpcode_opcode_metadata[256] = {
     [DICT_UPDATE] = { true, INSTR_FMT_IB },
     [DICT_MERGE] = { true, INSTR_FMT_IB },
     [MAP_ADD] = { true, INSTR_FMT_IB },
+    [LOAD_SUPER_ATTR] = { true, INSTR_FMT_IB },
     [LOAD_ATTR] = { true, INSTR_FMT_IBC00000000 },
     [LOAD_ATTR_INSTANCE_VALUE] = { true, INSTR_FMT_IBC00000000 },
     [LOAD_ATTR_MODULE] = { true, INSTR_FMT_IBC00000000 },

Python/opcode_targets.h

@@ -140,9 +140,9 @@ static void *opcode_targets[256] = {
     &&TARGET_STORE_DEREF,
     &&TARGET_DELETE_DEREF,
     &&TARGET_JUMP_BACKWARD,
-    &&TARGET_STORE_FAST__STORE_FAST,
+    &&TARGET_LOAD_SUPER_ATTR,
     &&TARGET_CALL_FUNCTION_EX,
-    &&TARGET_STORE_SUBSCR_DICT,
+    &&TARGET_STORE_FAST__STORE_FAST,
     &&TARGET_EXTENDED_ARG,
     &&TARGET_LIST_APPEND,
     &&TARGET_SET_ADD,
@@ -152,15 +152,15 @@ static void *opcode_targets[256] = {
     &&TARGET_YIELD_VALUE,
     &&TARGET_RESUME,
     &&TARGET_MATCH_CLASS,
+    &&TARGET_STORE_SUBSCR_DICT,
     &&TARGET_STORE_SUBSCR_LIST_INT,
-    &&TARGET_UNPACK_SEQUENCE_LIST,
     &&TARGET_FORMAT_VALUE,
     &&TARGET_BUILD_CONST_KEY_MAP,
     &&TARGET_BUILD_STRING,
+    &&TARGET_UNPACK_SEQUENCE_LIST,
     &&TARGET_UNPACK_SEQUENCE_TUPLE,
     &&TARGET_UNPACK_SEQUENCE_TWO_TUPLE,
     &&TARGET_SEND_GEN,
-    &&_unknown_opcode,
     &&TARGET_LIST_EXTEND,
     &&TARGET_SET_UPDATE,
     &&TARGET_DICT_MERGE,