GH-128563: Generate `opcode = ...` in instructions that need `opcode` (GH-129608)

* Remove support for GO_TO_INSTRUCTION

1 files changed, 437 insertions, 29 deletions

diff --git a/Python/generated_cases.c.h b/Python/generated_cases.c.h
index 44a78c41048..7dd9d6528bb 100644
--- a/Python/generated_cases.c.h
+++ b/Python/generated_cases.c.h
@@ -966,6 +966,7 @@
             PREDICTED_CALL:;
             _Py_CODEUNIT* const this_instr = next_instr - 4;
             (void)this_instr;
+            opcode = CALL;
             _PyStackRef *callable;
             _PyStackRef *self_or_null;
             _PyStackRef *args;
@@ -1744,12 +1745,11 @@
         }
 
         TARGET(CALL_FUNCTION_EX) {
-            frame->instr_ptr = next_instr;
+            _Py_CODEUNIT* const this_instr = frame->instr_ptr = next_instr;
+            (void)this_instr;
             next_instr += 1;
             INSTRUCTION_STATS(CALL_FUNCTION_EX);
-            PREDICTED_CALL_FUNCTION_EX:;
-            _Py_CODEUNIT* const this_instr = next_instr - 1;
-            (void)this_instr;
+            opcode = CALL_FUNCTION_EX;
             _PyStackRef func;
             _PyStackRef callargs;
             _PyStackRef kwargs_in;
@@ -2030,6 +2030,7 @@
             PREDICTED_CALL_KW:;
             _Py_CODEUNIT* const this_instr = next_instr - 4;
             (void)this_instr;
+            opcode = CALL_KW;
             _PyStackRef *callable;
             _PyStackRef *self_or_null;
             _PyStackRef *args;
@@ -2297,6 +2298,7 @@
             frame->instr_ptr = next_instr;
             next_instr += 4;
             INSTRUCTION_STATS(CALL_KW_NON_PY);
+            opcode = CALL_KW_NON_PY;
             static_assert(INLINE_CACHE_ENTRIES_CALL_KW == 3, "incorrect cache size");
             _PyStackRef *callable;
             _PyStackRef kwnames;
@@ -2914,6 +2916,7 @@
             frame->instr_ptr = next_instr;
             next_instr += 4;
             INSTRUCTION_STATS(CALL_NON_PY_GENERAL);
+            opcode = CALL_NON_PY_GENERAL;
             static_assert(INLINE_CACHE_ENTRIES_CALL == 3, "incorrect cache size");
             _PyStackRef *callable;
             _PyStackRef *self_or_null;
@@ -4045,6 +4048,7 @@
             (void)this_instr;
             next_instr += 1;
             INSTRUCTION_STATS(ENTER_EXECUTOR);
+            opcode = ENTER_EXECUTOR;
             #ifdef _Py_TIER2
             PyCodeObject *code = _PyFrame_GetCode(frame);
             _PyExecutorObject *executor = code->co_executors->executors[oparg & 255];
@@ -4097,6 +4101,7 @@
             frame->instr_ptr = next_instr;
             next_instr += 1;
             INSTRUCTION_STATS(EXTENDED_ARG);
+            opcode = EXTENDED_ARG;
             assert(oparg);
             opcode = next_instr->op.code;
             oparg = oparg << 8 | next_instr->op.arg;
@@ -4646,6 +4651,7 @@
             (void)this_instr;
             next_instr += 4;
             INSTRUCTION_STATS(INSTRUMENTED_CALL);
+            opcode = INSTRUMENTED_CALL;
             _PyStackRef *callable;
             _PyStackRef *self_or_null;
             _PyStackRef *args;
@@ -4815,11 +4821,188 @@
         }
 
         TARGET(INSTRUMENTED_CALL_FUNCTION_EX) {
-            frame->instr_ptr = next_instr;
+            _Py_CODEUNIT* const this_instr = frame->instr_ptr = next_instr;
+            (void)this_instr;
             next_instr += 1;
             INSTRUCTION_STATS(INSTRUMENTED_CALL_FUNCTION_EX);
-
-            goto PREDICTED_CALL_FUNCTION_EX;
+            opcode = INSTRUMENTED_CALL_FUNCTION_EX;
+            _PyStackRef func;
+            _PyStackRef callargs;
+            _PyStackRef kwargs_in;
+            _PyStackRef tuple;
+            _PyStackRef kwargs_out;
+            _PyStackRef func_st;
+            _PyStackRef null;
+            _PyStackRef callargs_st;
+            _PyStackRef kwargs_st;
+            _PyStackRef result;
+            // _MAKE_CALLARGS_A_TUPLE
+            {
+                kwargs_in = stack_pointer[-1];
+                callargs = stack_pointer[-2];
+                func = stack_pointer[-4];
+                PyObject *callargs_o = PyStackRef_AsPyObjectBorrow(callargs);
+                if (PyTuple_CheckExact(callargs_o)) {
+                    tuple = callargs;
+                    kwargs_out = kwargs_in;
+                }
+                else {
+                    _PyFrame_SetStackPointer(frame, stack_pointer);
+                    int err = _Py_Check_ArgsIterable(tstate, PyStackRef_AsPyObjectBorrow(func), callargs_o);
+                    stack_pointer = _PyFrame_GetStackPointer(frame);
+                    if (err < 0) {
+                        goto error;
+                    }
+                    _PyFrame_SetStackPointer(frame, stack_pointer);
+                    PyObject *tuple_o = PySequence_Tuple(callargs_o);
+                    stack_pointer = _PyFrame_GetStackPointer(frame);
+                    if (tuple_o == NULL) {
+                        goto error;
+                    }
+                    kwargs_out = kwargs_in;
+                    stack_pointer += -2;
+                    assert(WITHIN_STACK_BOUNDS());
+                    _PyFrame_SetStackPointer(frame, stack_pointer);
+                    PyStackRef_CLOSE(callargs);
+                    stack_pointer = _PyFrame_GetStackPointer(frame);
+                    tuple = PyStackRef_FromPyObjectSteal(tuple_o);
+                    stack_pointer += 2;
+                    assert(WITHIN_STACK_BOUNDS());
+                }
+            }
+            // _DO_CALL_FUNCTION_EX
+            {
+                kwargs_st = kwargs_out;
+                callargs_st = tuple;
+                null = stack_pointer[-3];
+                func_st = func;
+                (void)null;
+                PyObject *func = PyStackRef_AsPyObjectBorrow(func_st);
+                // DICT_MERGE is called before this opcode if there are kwargs.
+                // It converts all dict subtypes in kwargs into regular dicts.
+                EVAL_CALL_STAT_INC_IF_FUNCTION(EVAL_CALL_FUNCTION_EX, func);
+                PyObject *result_o;
+                assert(!_PyErr_Occurred(tstate));
+                if (opcode == INSTRUMENTED_CALL_FUNCTION_EX) {
+                    PyObject *callargs = PyStackRef_AsPyObjectBorrow(callargs_st);
+                    PyObject *kwargs = PyStackRef_AsPyObjectBorrow(kwargs_st);
+                    assert(kwargs == NULL || PyDict_CheckExact(kwargs));
+                    assert(PyTuple_CheckExact(callargs));
+                    PyObject *arg = PyTuple_GET_SIZE(callargs) > 0 ?
+                    PyTuple_GET_ITEM(callargs, 0) : &_PyInstrumentation_MISSING;
+                    stack_pointer[-2] = callargs_st;
+                    stack_pointer[-1] = kwargs_st;
+                    _PyFrame_SetStackPointer(frame, stack_pointer);
+                    int err = _Py_call_instrumentation_2args(
+                        tstate, PY_MONITORING_EVENT_CALL,
+                        frame, this_instr, func, arg);
+                    stack_pointer = _PyFrame_GetStackPointer(frame);
+                    if (err) {
+                        goto error;
+                    }
+                    _PyFrame_SetStackPointer(frame, stack_pointer);
+                    result_o = PyObject_Call(func, callargs, kwargs);
+                    stack_pointer = _PyFrame_GetStackPointer(frame);
+                    if (!PyFunction_Check(func) && !PyMethod_Check(func)) {
+                        if (result_o == NULL) {
+                            _PyFrame_SetStackPointer(frame, stack_pointer);
+                            _Py_call_instrumentation_exc2(
+                                tstate, PY_MONITORING_EVENT_C_RAISE,
+                                frame, this_instr, func, arg);
+                            stack_pointer = _PyFrame_GetStackPointer(frame);
+                        }
+                        else {
+                            _PyFrame_SetStackPointer(frame, stack_pointer);
+                            int err = _Py_call_instrumentation_2args(
+                                tstate, PY_MONITORING_EVENT_C_RETURN,
+                                frame, this_instr, func, arg);
+                            stack_pointer = _PyFrame_GetStackPointer(frame);
+                            if (err < 0) {
+                                Py_CLEAR(result_o);
+                            }
+                        }
+                    }
+                }
+                else {
+                    if (Py_TYPE(func) == &PyFunction_Type &&
+                        tstate->interp->eval_frame == NULL &&
+                        ((PyFunctionObject *)func)->vectorcall == _PyFunction_Vectorcall) {
+                        PyObject *callargs = PyStackRef_AsPyObjectSteal(callargs_st);
+                        assert(PyTuple_CheckExact(callargs));
+                        PyObject *kwargs = PyStackRef_IsNull(kwargs_st) ? NULL : PyStackRef_AsPyObjectSteal(kwargs_st);
+                        assert(kwargs == NULL || PyDict_CheckExact(kwargs));
+                        Py_ssize_t nargs = PyTuple_GET_SIZE(callargs);
+                        int code_flags = ((PyCodeObject *)PyFunction_GET_CODE(func))->co_flags;
+                        PyObject *locals = code_flags & CO_OPTIMIZED ? NULL : Py_NewRef(PyFunction_GET_GLOBALS(func));
+                        stack_pointer += -2;
+                        assert(WITHIN_STACK_BOUNDS());
+                        _PyFrame_SetStackPointer(frame, stack_pointer);
+                        _PyInterpreterFrame *new_frame = _PyEvalFramePushAndInit_Ex(
+                            tstate, func_st, locals,
+                            nargs, callargs, kwargs, frame);
+                        stack_pointer = _PyFrame_GetStackPointer(frame);
+                        // Need to sync the stack since we exit with DISPATCH_INLINED.
+                        stack_pointer += -2;
+                        assert(WITHIN_STACK_BOUNDS());
+                        if (new_frame == NULL) {
+                            goto error;
+                        }
+                        assert( 1 == 1);
+                        frame->return_offset = 1;
+                        DISPATCH_INLINED(new_frame);
+                    }
+                    PyObject *callargs = PyStackRef_AsPyObjectBorrow(callargs_st);
+                    assert(PyTuple_CheckExact(callargs));
+                    PyObject *kwargs = PyStackRef_AsPyObjectBorrow(kwargs_st);
+                    assert(kwargs == NULL || PyDict_CheckExact(kwargs));
+                    stack_pointer[-2] = callargs_st;
+                    stack_pointer[-1] = kwargs_st;
+                    _PyFrame_SetStackPointer(frame, stack_pointer);
+                    result_o = PyObject_Call(func, callargs, kwargs);
+                    stack_pointer = _PyFrame_GetStackPointer(frame);
+                }
+                stack_pointer += -1;
+                assert(WITHIN_STACK_BOUNDS());
+                _PyFrame_SetStackPointer(frame, stack_pointer);
+                PyStackRef_XCLOSE(kwargs_st);
+                stack_pointer = _PyFrame_GetStackPointer(frame);
+                stack_pointer += -1;
+                assert(WITHIN_STACK_BOUNDS());
+                _PyFrame_SetStackPointer(frame, stack_pointer);
+                PyStackRef_CLOSE(callargs_st);
+                stack_pointer = _PyFrame_GetStackPointer(frame);
+                stack_pointer += -2;
+                assert(WITHIN_STACK_BOUNDS());
+                _PyFrame_SetStackPointer(frame, stack_pointer);
+                PyStackRef_CLOSE(func_st);
+                stack_pointer = _PyFrame_GetStackPointer(frame);
+                if (result_o == NULL) {
+                    goto error;
+                }
+                result = PyStackRef_FromPyObjectSteal(result_o);
+            }
+            // _CHECK_PERIODIC
+            {
+                _Py_CHECK_EMSCRIPTEN_SIGNALS_PERIODICALLY();
+                QSBR_QUIESCENT_STATE(tstate);
+                if (_Py_atomic_load_uintptr_relaxed(&tstate->eval_breaker) & _PY_EVAL_EVENTS_MASK) {
+                    stack_pointer[0] = result;
+                    stack_pointer += 1;
+                    assert(WITHIN_STACK_BOUNDS());
+                    _PyFrame_SetStackPointer(frame, stack_pointer);
+                    int err = _Py_HandlePending(tstate);
+                    stack_pointer = _PyFrame_GetStackPointer(frame);
+                    if (err != 0) {
+                        goto error;
+                    }
+                    stack_pointer += -1;
+                    assert(WITHIN_STACK_BOUNDS());
+                }
+            }
+            stack_pointer[0] = result;
+            stack_pointer += 1;
+            assert(WITHIN_STACK_BOUNDS());
+            DISPATCH();
         }
 
         TARGET(INSTRUMENTED_CALL_KW) {
@@ -4827,25 +5010,170 @@
             (void)this_instr;
             next_instr += 4;
             INSTRUCTION_STATS(INSTRUMENTED_CALL_KW);
-            uint16_t counter = read_u16(&this_instr[1].cache);
-            (void)counter;
-            uint32_t version = read_u32(&this_instr[2].cache);
-            (void)version;
-            int is_meth = !PyStackRef_IsNull(PEEK(oparg + 2));
-            int total_args = oparg + is_meth;
-            PyObject *function = PyStackRef_AsPyObjectBorrow(PEEK(oparg + 3));
-            PyObject *arg = total_args == 0 ? &_PyInstrumentation_MISSING
-        : PyStackRef_AsPyObjectBorrow(PEEK(total_args + 1));
-            _PyFrame_SetStackPointer(frame, stack_pointer);
-            int err = _Py_call_instrumentation_2args(
-                tstate, PY_MONITORING_EVENT_CALL,
-                frame, this_instr, function, arg);
-            stack_pointer = _PyFrame_GetStackPointer(frame);
-            if (err) {
-                goto error;
+            opcode = INSTRUMENTED_CALL_KW;
+            _PyStackRef *callable;
+            _PyStackRef *self_or_null;
+            _PyStackRef *args;
+            _PyStackRef kwnames;
+            _PyStackRef kwnames_in;
+            _PyStackRef *func;
+            _PyStackRef *maybe_self;
+            _PyStackRef kwnames_out;
+            _PyStackRef res;
+            /* Skip 1 cache entry */
+            /* Skip 2 cache entries */
+            // _MONITOR_CALL_KW
+            {
+                args = &stack_pointer[-1 - oparg];
+                self_or_null = &stack_pointer[-2 - oparg];
+                callable = &stack_pointer[-3 - oparg];
+                int is_meth = !PyStackRef_IsNull(self_or_null[0]);
+                PyObject *arg;
+                if (is_meth) {
+                    arg = PyStackRef_AsPyObjectBorrow(self_or_null[0]);
+                }
+                else {
+                    if (args) {
+                        arg = PyStackRef_AsPyObjectBorrow(args[0]);
+                    }
+                    else {
+                        arg = &_PyInstrumentation_MISSING;
+                    }
+                }
+                PyObject *function = PyStackRef_AsPyObjectBorrow(callable[0]);
+                _PyFrame_SetStackPointer(frame, stack_pointer);
+                int err = _Py_call_instrumentation_2args(
+                    tstate, PY_MONITORING_EVENT_CALL,
+                    frame, this_instr, function, arg);
+                stack_pointer = _PyFrame_GetStackPointer(frame);
+                if (err) {
+                    goto error;
+                }
             }
-            PAUSE_ADAPTIVE_COUNTER(this_instr[1].counter);
-            goto PREDICTED_CALL_KW;
+            // _MAYBE_EXPAND_METHOD_KW
+            {
+                kwnames_in = stack_pointer[-1];
+                func = &stack_pointer[-3 - oparg];
+                maybe_self = &stack_pointer[-2 - oparg];
+                if (PyStackRef_TYPE(callable[0]) == &PyMethod_Type && PyStackRef_IsNull(self_or_null[0])) {
+                    PyObject *callable_o = PyStackRef_AsPyObjectBorrow(callable[0]);
+                    PyObject *self = ((PyMethodObject *)callable_o)->im_self;
+                    maybe_self[0] = PyStackRef_FromPyObjectNew(self);
+                    PyObject *method = ((PyMethodObject *)callable_o)->im_func;
+                    _PyStackRef temp = callable[0];
+                    func[0] = PyStackRef_FromPyObjectNew(method);
+                    _PyFrame_SetStackPointer(frame, stack_pointer);
+                    PyStackRef_CLOSE(temp);
+                    stack_pointer = _PyFrame_GetStackPointer(frame);
+                }
+                kwnames_out = kwnames_in;
+            }
+            // _DO_CALL_KW
+            {
+                kwnames = kwnames_out;
+                args = &stack_pointer[-1 - oparg];
+                self_or_null = &stack_pointer[-2 - oparg];
+                callable = &stack_pointer[-3 - oparg];
+                PyObject *callable_o = PyStackRef_AsPyObjectBorrow(callable[0]);
+                PyObject *kwnames_o = PyStackRef_AsPyObjectBorrow(kwnames);
+                // oparg counts all of the args, but *not* self:
+                int total_args = oparg;
+                _PyStackRef *arguments = args;
+                if (!PyStackRef_IsNull(self_or_null[0])) {
+                    arguments--;
+                    total_args++;
+                }
+                int positional_args = total_args - (int)PyTuple_GET_SIZE(kwnames_o);
+                // Check if the call can be inlined or not
+                if (Py_TYPE(callable_o) == &PyFunction_Type &&
+                    tstate->interp->eval_frame == NULL &&
+                    ((PyFunctionObject *)callable_o)->vectorcall == _PyFunction_Vectorcall)
+                {
+                    int code_flags = ((PyCodeObject*)PyFunction_GET_CODE(callable_o))->co_flags;
+                    PyObject *locals = code_flags & CO_OPTIMIZED ? NULL : Py_NewRef(PyFunction_GET_GLOBALS(callable_o));
+                    stack_pointer[-1] = kwnames;
+                    _PyFrame_SetStackPointer(frame, stack_pointer);
+                    _PyInterpreterFrame *new_frame = _PyEvalFramePushAndInit(
+                        tstate, callable[0], locals,
+                        arguments, positional_args, kwnames_o, frame
+                    );
+                    stack_pointer = _PyFrame_GetStackPointer(frame);
+                    stack_pointer += -1;
+                    assert(WITHIN_STACK_BOUNDS());
+                    _PyFrame_SetStackPointer(frame, stack_pointer);
+                    PyStackRef_CLOSE(kwnames);
+                    stack_pointer = _PyFrame_GetStackPointer(frame);
+                    // Sync stack explicitly since we leave using DISPATCH_INLINED().
+                    stack_pointer += -2 - oparg;
+                    assert(WITHIN_STACK_BOUNDS());
+                    // The frame has stolen all the arguments from the stack,
+                    // so there is no need to clean them up.
+                    if (new_frame == NULL) {
+                        goto error;
+                    }
+                    assert( 4 == 1 + INLINE_CACHE_ENTRIES_CALL_KW);
+                    frame->return_offset = 4 ;
+                    DISPATCH_INLINED(new_frame);
+                }
+                /* Callable is not a normal Python function */
+                STACKREFS_TO_PYOBJECTS(arguments, total_args, args_o);
+                if (CONVERSION_FAILED(args_o)) {
+                    PyStackRef_CLOSE(callable[0]);
+                    PyStackRef_XCLOSE(self_or_null[0]);
+                    for (int _i = oparg; --_i >= 0;) {
+                        PyStackRef_CLOSE(args[_i]);
+                    }
+                    PyStackRef_CLOSE(kwnames);
+                    stack_pointer += -3 - oparg;
+                    assert(WITHIN_STACK_BOUNDS());
+                    goto error;
+                }
+                stack_pointer[-1] = kwnames;
+                _PyFrame_SetStackPointer(frame, stack_pointer);
+                PyObject *res_o = PyObject_Vectorcall(
+                    callable_o, args_o,
+                    positional_args | PY_VECTORCALL_ARGUMENTS_OFFSET,
+                    kwnames_o);
+                stack_pointer = _PyFrame_GetStackPointer(frame);
+                STACKREFS_TO_PYOBJECTS_CLEANUP(args_o);
+                if (opcode == INSTRUMENTED_CALL_KW) {
+                    PyObject *arg = total_args == 0 ?
+                    &_PyInstrumentation_MISSING : PyStackRef_AsPyObjectBorrow(arguments[0]);
+                    if (res_o == NULL) {
+                        _PyFrame_SetStackPointer(frame, stack_pointer);
+                        _Py_call_instrumentation_exc2(
+                            tstate, PY_MONITORING_EVENT_C_RAISE,
+                            frame, this_instr, callable_o, arg);
+                        stack_pointer = _PyFrame_GetStackPointer(frame);
+                    }
+                    else {
+                        _PyFrame_SetStackPointer(frame, stack_pointer);
+                        int err = _Py_call_instrumentation_2args(
+                            tstate, PY_MONITORING_EVENT_C_RETURN,
+                            frame, this_instr, callable_o, arg);
+                        stack_pointer = _PyFrame_GetStackPointer(frame);
+                        if (err < 0) {
+                            Py_CLEAR(res_o);
+                        }
+                    }
+                }
+                PyStackRef_CLOSE(callable[0]);
+                PyStackRef_XCLOSE(self_or_null[0]);
+                for (int _i = oparg; --_i >= 0;) {
+                    PyStackRef_CLOSE(args[_i]);
+                }
+                PyStackRef_CLOSE(kwnames);
+                if (res_o == NULL) {
+                    stack_pointer += -3 - oparg;
+                    assert(WITHIN_STACK_BOUNDS());
+                    goto error;
+                }
+                res = PyStackRef_FromPyObjectSteal(res_o);
+            }
+            stack_pointer[-3 - oparg] = res;
+            stack_pointer += -2 - oparg;
+            assert(WITHIN_STACK_BOUNDS());
+            DISPATCH();
         }
 
         TARGET(INSTRUMENTED_END_FOR) {
@@ -4944,6 +5272,7 @@
             (void)this_instr;
             next_instr += 1;
             INSTRUCTION_STATS(INSTRUMENTED_INSTRUCTION);
+            opcode = INSTRUMENTED_INSTRUCTION;
             _PyFrame_SetStackPointer(frame, stack_pointer);
             int next_opcode = _Py_call_instrumentation_instruction(
                 tstate, frame, this_instr);
@@ -5001,6 +5330,7 @@
             (void)this_instr;
             next_instr += 1;
             INSTRUCTION_STATS(INSTRUMENTED_LINE);
+            opcode = INSTRUMENTED_LINE;
             int original_opcode = 0;
             if (tstate->tracing) {
                 PyCodeObject *code = _PyFrame_GetCode(frame);
@@ -5038,11 +5368,88 @@
             (void)this_instr;
             next_instr += 2;
             INSTRUCTION_STATS(INSTRUMENTED_LOAD_SUPER_ATTR);
+            opcode = INSTRUMENTED_LOAD_SUPER_ATTR;
+            _PyStackRef global_super_st;
+            _PyStackRef class_st;
+            _PyStackRef self_st;
+            _PyStackRef attr;
+            _PyStackRef null = PyStackRef_NULL;
             /* Skip 1 cache entry */
-            // cancel out the decrement that will happen in LOAD_SUPER_ATTR; we
-            // don't want to specialize instrumented instructions
-            PAUSE_ADAPTIVE_COUNTER(this_instr[1].counter);
-            goto PREDICTED_LOAD_SUPER_ATTR;
+            // _LOAD_SUPER_ATTR
+            {
+                self_st = stack_pointer[-1];
+                class_st = stack_pointer[-2];
+                global_super_st = stack_pointer[-3];
+                PyObject *global_super = PyStackRef_AsPyObjectBorrow(global_super_st);
+                PyObject *class = PyStackRef_AsPyObjectBorrow(class_st);
+                PyObject *self = PyStackRef_AsPyObjectBorrow(self_st);
+                if (opcode == INSTRUMENTED_LOAD_SUPER_ATTR) {
+                    PyObject *arg = oparg & 2 ? class : &_PyInstrumentation_MISSING;
+                    _PyFrame_SetStackPointer(frame, stack_pointer);
+                    int err = _Py_call_instrumentation_2args(
+                        tstate, PY_MONITORING_EVENT_CALL,
+                        frame, this_instr, global_super, arg);
+                    stack_pointer = _PyFrame_GetStackPointer(frame);
+                    if (err) {
+                        PyStackRef_CLOSE(global_super_st);
+                        PyStackRef_CLOSE(class_st);
+                        PyStackRef_CLOSE(self_st);
+                        goto pop_3_error;
+                    }
+                }
+                // we make no attempt to optimize here; specializations should
+                // handle any case whose performance we care about
+                PyObject *stack[] = {class, self};
+                _PyFrame_SetStackPointer(frame, stack_pointer);
+                PyObject *super = PyObject_Vectorcall(global_super, stack, oparg & 2, NULL);
+                stack_pointer = _PyFrame_GetStackPointer(frame);
+                if (opcode == INSTRUMENTED_LOAD_SUPER_ATTR) {
+                    PyObject *arg = oparg & 2 ? class : &_PyInstrumentation_MISSING;
+                    if (super == NULL) {
+                        _PyFrame_SetStackPointer(frame, stack_pointer);
+                        _Py_call_instrumentation_exc2(
+                            tstate, PY_MONITORING_EVENT_C_RAISE,
+                            frame, this_instr, global_super, arg);
+                        stack_pointer = _PyFrame_GetStackPointer(frame);
+                    }
+                    else {
+                        _PyFrame_SetStackPointer(frame, stack_pointer);
+                        int err = _Py_call_instrumentation_2args(
+                            tstate, PY_MONITORING_EVENT_C_RETURN,
+                            frame, this_instr, global_super, arg);
+                        stack_pointer = _PyFrame_GetStackPointer(frame);
+                        if (err < 0) {
+                            Py_CLEAR(super);
+                        }
+                    }
+                }
+                PyStackRef_CLOSE(global_super_st);
+                PyStackRef_CLOSE(class_st);
+                PyStackRef_CLOSE(self_st);
+                if (super == NULL) {
+                    goto pop_3_error;
+                }
+                PyObject *name = GETITEM(FRAME_CO_NAMES, oparg >> 2);
+                stack_pointer += -3;
+                assert(WITHIN_STACK_BOUNDS());
+                _PyFrame_SetStackPointer(frame, stack_pointer);
+                PyObject *attr_o = PyObject_GetAttr(super, name);
+                Py_DECREF(super);
+                stack_pointer = _PyFrame_GetStackPointer(frame);
+                if (attr_o == NULL) {
+                    goto error;
+                }
+                attr = PyStackRef_FromPyObjectSteal(attr_o);
+            }
+            // _PUSH_NULL_CONDITIONAL
+            {
+                null = PyStackRef_NULL;
+            }
+            stack_pointer[0] = attr;
+            if (oparg & 1) stack_pointer[1] = null;
+            stack_pointer += 1 + (oparg & 1);
+            assert(WITHIN_STACK_BOUNDS());
+            DISPATCH();
         }
 
         TARGET(INSTRUMENTED_NOT_TAKEN) {
@@ -6862,6 +7269,7 @@
             PREDICTED_LOAD_SUPER_ATTR:;
             _Py_CODEUNIT* const this_instr = next_instr - 2;
             (void)this_instr;
+            opcode = LOAD_SUPER_ATTR;
             _PyStackRef global_super_st;
             _PyStackRef class_st;
             _PyStackRef self_st;