Specialize LOAD_ATTR_SLOT

- Use atomics and _Py_TryIncrefCompareStackRef in _LOAD_ATTR_SLOT
- Pass type version during specialization

Author: mpage

Lib/test/test_opcache.py

@@ -947,6 +947,33 @@ def write(items):
         opname = "LOAD_ATTR_PROPERTY"
         self.assert_races_do_not_crash(opname, get_items, read, write)
 
+    @requires_specialization_ft
+    def test_load_attr_slot(self):
+        def get_items():
+            class C:
+                __slots__ = ["a", "b"]
+
+            items = []
+            for i in range(self.ITEMS):
+                item = C()
+                item.a = i
+                item.b = i + self.ITEMS
+                items.append(item)
+            return items
+
+        def read(items):
+            for item in items:
+                item.a
+                item.b
+
+        def write(items):
+            for item in items:
+                item.a = 100
+                item.b = 200
+
+        opname = "LOAD_ATTR_SLOT"
+        self.assert_races_do_not_crash(opname, get_items, read, write)
+
     @requires_specialization_ft
     def test_load_attr_with_hint(self):
         def get_items():

Python/bytecodes.c

@@ -2255,12 +2255,20 @@ dummy_func(
         split op(_LOAD_ATTR_SLOT, (index/1, owner -- attr, null if (oparg & 1))) {
             PyObject *owner_o = PyStackRef_AsPyObjectBorrow(owner);
 
-            char *addr = (char *)owner_o + index;
-            PyObject *attr_o = *(PyObject **)addr;
+            PyObject **addr = (PyObject **)((char *)owner_o + index);
+            PyObject *attr_o = FT_ATOMIC_LOAD_PTR(*addr);
             DEOPT_IF(attr_o == NULL);
+            #ifdef Py_GIL_DISABLED
+            int increfed = _Py_TryIncrefCompareStackRef(addr, attr_o, &attr);
+            DEOPT_IF(!increfed);
+            #else
+            // XXX - Bug in cases generator
+            Py_INCREF(attr_o);
+            attr = PyStackRef_FromPyObjectSteal(attr_o);
+            #endif
             STAT_INC(LOAD_ATTR, hit);
+
             null = PyStackRef_NULL;
-            attr = PyStackRef_FromPyObjectNew(attr_o);
             DECREF_INPUTS();
         }
 

Python/executor_cases.c.h

@@ -1203,7 +1203,6 @@ do_specialize_instance_load_attr(PyObject* owner, _Py_CODEUNIT* instr, PyObject*
         }
         case OBJECT_SLOT:
         {
-            FT_UNIMPLEMENTED();
             PyMemberDescrObject *member = (PyMemberDescrObject *)descr;
             struct PyMemberDef *dmem = member->d_member;
             Py_ssize_t offset = dmem->offset;
@@ -1222,17 +1221,16 @@ do_specialize_instance_load_attr(PyObject* owner, _Py_CODEUNIT* instr, PyObject*
             assert(dmem->type == Py_T_OBJECT_EX || dmem->type == _Py_T_OBJECT);
             assert(offset > 0);
             cache->index = (uint16_t)offset;
-            write_u32(cache->version, type->tp_version_tag);
+            write_u32(cache->version, tp_version);
             specialize(instr, LOAD_ATTR_SLOT);
             return 0;
         }
         case DUNDER_CLASS:
         {
-            FT_UNIMPLEMENTED();
             Py_ssize_t offset = offsetof(PyObject, ob_type);
             assert(offset == (uint16_t)offset);
             cache->index = (uint16_t)offset;
-            write_u32(cache->version, type->tp_version_tag);
+            write_u32(cache->version, tp_version);
             specialize(instr, LOAD_ATTR_SLOT);
             return 0;
         }