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.

Author: carljm

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

@@ -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/internal/pycore_typeobject.h

@@ -41,6 +41,7 @@
 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:

Include/opcode.h

@@ -439,7 +439,8 @@ def _write_atomic(path, data, mode=0o666):
 #     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 @@ def _write_atomic(path, data, mode=0o666):
 # 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/dis.py

@@ -196,7 +196,7 @@ def pseudo_op(name, op, real_ops):
 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 @@ def pseudo_op(name, op, real_ops):
 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/importlib/_bootstrap_external.py

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

Lib/opcode.py

@@ -962,7 +962,7 @@ def test_wrapper_call(self):
         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/shadowed_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 @@ def f(self):
     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 @@ def test_super_argtype(self):
         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()

Lib/test/test_gdb.py

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