pythongh-87729: improve hit rate of LOAD_SUPER_ATTR specialization

Author: carljm

Include/internal/pycore_code.h

@@ -53,9 +53,6 @@ typedef struct {
 
 typedef struct {
     uint16_t counter;
-    uint16_t class_version[2];
-    uint16_t self_type_version[2];
-    uint16_t method[4];
 } _PySuperAttrCache;
 
 #define INLINE_CACHE_ENTRIES_LOAD_SUPER_ATTR CACHE_ENTRIES(_PySuperAttrCache)
@@ -226,8 +223,8 @@ extern int _PyLineTable_PreviousAddressRange(PyCodeAddressRange *range);
 
 /* Specialization functions */
 
-extern void _Py_Specialize_LoadSuperAttr(PyObject *global_super, PyObject *cls, PyObject *self,
-                                         _Py_CODEUNIT *instr, PyObject *name, int load_method);
+extern void _Py_Specialize_LoadSuperAttr(PyObject *global_super, PyObject *cls,
+                                         _Py_CODEUNIT *instr, int load_method);
 extern void _Py_Specialize_LoadAttr(PyObject *owner, _Py_CODEUNIT *instr,
                                     PyObject *name);
 extern void _Py_Specialize_StoreAttr(PyObject *owner, _Py_CODEUNIT *instr,

Include/internal/pycore_opcode.h

@@ -142,8 +142,6 @@ PyAPI_DATA(PyTypeObject) _PyBufferWrapper_Type;
 
 PyObject *
 _PySuper_Lookup(PyTypeObject *su_type, PyObject *su_obj, PyObject *name, int *meth_found);
-PyObject *
-_PySuper_LookupDescr(PyTypeObject *su_type, PyObject *su_obj, PyObject *name);
 
 #ifdef __cplusplus
 }

Include/internal/pycore_typeobject.h

@@ -442,6 +442,7 @@ def _write_atomic(path, data, mode=0o666):
 #     Python 3.12b1 3526 (Add instrumentation support)
 #     Python 3.12b1 3527 (Add LOAD_SUPER_ATTR)
 #     Python 3.12b1 3528 (Add LOAD_SUPER_ATTR_METHOD specialization)
+#     Python 3.12b1 3529 (Shrink the LOAD_SUPER_ATTR caches)
 
 #     Python 3.13 will start with 3550
 
@@ -458,7 +459,7 @@ 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 = (3528).to_bytes(2, 'little') + b'\r\n'
+MAGIC_NUMBER = (3529).to_bytes(2, 'little') + b'\r\n'
 
 _RAW_MAGIC_NUMBER = int.from_bytes(MAGIC_NUMBER, 'little')  # For import.c
 

Include/opcode.h

@@ -369,6 +369,7 @@ def pseudo_op(name, op, real_ops):
         "FOR_ITER_GEN",
     ],
     "LOAD_SUPER_ATTR": [
+        "LOAD_SUPER_ATTR_ATTR",
         "LOAD_SUPER_ATTR_METHOD",
     ],
     "LOAD_ATTR": [
@@ -446,9 +447,6 @@ def pseudo_op(name, op, real_ops):
     },
     "LOAD_SUPER_ATTR": {
         "counter": 1,
-        "class_version": 2,
-        "self_type_version": 2,
-        "method": 4,
     },
     "LOAD_ATTR": {
         "counter": 1,

Lib/importlib/_bootstrap_external.py

@@ -10140,18 +10140,6 @@ _PySuper_Lookup(PyTypeObject *su_type, PyObject *su_obj, PyObject *name, int *me
     return res;
 }
 
-PyObject *
-_PySuper_LookupDescr(PyTypeObject *su_type, PyObject *su_obj, PyObject *name)
-{
-    PyTypeObject *su_obj_type = supercheck(su_type, su_obj);
-    if (su_obj_type == NULL) {
-        return NULL;
-    }
-    PyObject *res = _super_lookup_descr(su_type, su_obj_type, name);
-    Py_DECREF(su_obj_type);
-    return res;
-}
-
 static PyObject *
 super_descr_get(PyObject *self, PyObject *obj, PyObject *type)
 {

Lib/opcode.py

@@ -1556,17 +1556,18 @@ dummy_func(
 
         family(load_super_attr, INLINE_CACHE_ENTRIES_LOAD_SUPER_ATTR) = {
             LOAD_SUPER_ATTR,
+            LOAD_SUPER_ATTR_ATTR,
             LOAD_SUPER_ATTR_METHOD,
         };
 
-        inst(LOAD_SUPER_ATTR, (unused/9, global_super, class, self -- res2 if (oparg & 1), res)) {
+        inst(LOAD_SUPER_ATTR, (unused/1, global_super, class, self -- res2 if (oparg & 1), res)) {
             PyObject *name = GETITEM(frame->f_code->co_names, oparg >> 2);
             int load_method = oparg & 1;
             #if ENABLE_SPECIALIZATION
             _PySuperAttrCache *cache = (_PySuperAttrCache *)next_instr;
             if (ADAPTIVE_COUNTER_IS_ZERO(cache->counter)) {
                 next_instr--;
-                _Py_Specialize_LoadSuperAttr(global_super, class, self, next_instr, name, load_method);
+                _Py_Specialize_LoadSuperAttr(global_super, class, next_instr, load_method);
                 DISPATCH_SAME_OPARG();
             }
             STAT_INC(LOAD_SUPER_ATTR, deferred);
@@ -1584,17 +1585,38 @@ dummy_func(
             ERROR_IF(res == NULL, error);
         }
 
-        inst(LOAD_SUPER_ATTR_METHOD, (unused/1, class_version/2, self_type_version/2, method/4, global_super, class, self -- res2, res)) {
+        inst(LOAD_SUPER_ATTR_ATTR, (unused/1, global_super, class, self -- res2 if (oparg & 1), res)) {
+            assert(!(oparg & 1));
             DEOPT_IF(global_super != (PyObject *)&PySuper_Type, LOAD_SUPER_ATTR);
             DEOPT_IF(!PyType_Check(class), LOAD_SUPER_ATTR);
-            DEOPT_IF(((PyTypeObject *)class)->tp_version_tag != class_version, LOAD_SUPER_ATTR);
-            PyTypeObject *self_type = Py_TYPE(self);
-            DEOPT_IF(self_type->tp_version_tag != self_type_version, LOAD_SUPER_ATTR);
-            res2 = method;
-            res = self; // transfer ownership
-            Py_INCREF(res2);
+            STAT_INC(LOAD_SUPER_ATTR, hit);
+            PyObject *name = GETITEM(frame->f_code->co_names, oparg >> 2);
+            res = _PySuper_Lookup((PyTypeObject *)class, self, name, NULL);
+            ERROR_IF(res == NULL, error);
+            DECREF_INPUTS();
+        }
+
+        inst(LOAD_SUPER_ATTR_METHOD, (unused/1, global_super, class, self -- res2, res)) {
+            assert(oparg & 1);
+            DEOPT_IF(global_super != (PyObject *)&PySuper_Type, LOAD_SUPER_ATTR);
+            DEOPT_IF(!PyType_Check(class), LOAD_SUPER_ATTR);
+            STAT_INC(LOAD_SUPER_ATTR, hit);
+            PyObject *name = GETITEM(frame->f_code->co_names, oparg >> 2);
+            int method_found = 0;
+            res2 = _PySuper_Lookup((PyTypeObject *)class, self, name, &method_found);
             Py_DECREF(global_super);
             Py_DECREF(class);
+            if (res2 == NULL) {
+                Py_DECREF(self);
+                ERROR_IF(true, error);
+            }
+            if (method_found) {
+                res = self; // transfer ownership
+            } else {
+                Py_DECREF(self);
+                res = res2;
+                res2 = NULL;
+            }
         }
 
         family(load_attr, INLINE_CACHE_ENTRIES_LOAD_ATTR) = {