gh-134584: Eliminate redundant refcounting from _LOAD_ATTR_WITH_HINT (GH-143062)

Eliminate redundant refcounting from _LOAD_ATTR_WITH_HINT

Include/internal/pycore_opcode_metadata.h

@@ -1434,7 +1434,7 @@ _PyOpcode_macro_expansion[256] = {
     [LOAD_ATTR_NONDESCRIPTOR_WITH_VALUES] = { .nuops = 4, .uops = { { _GUARD_TYPE_VERSION, 2, 1 }, { _GUARD_DORV_VALUES_INST_ATTR_FROM_DICT, OPARG_SIMPLE, 3 }, { _GUARD_KEYS_VERSION, 2, 3 }, { _LOAD_ATTR_NONDESCRIPTOR_WITH_VALUES, 4, 5 } } },
     [LOAD_ATTR_PROPERTY] = { .nuops = 5, .uops = { { _CHECK_PEP_523, OPARG_SIMPLE, 1 }, { _GUARD_TYPE_VERSION, 2, 1 }, { _LOAD_ATTR_PROPERTY_FRAME, 4, 5 }, { _SAVE_RETURN_OFFSET, OPARG_SAVE_RETURN_OFFSET, 9 }, { _PUSH_FRAME, OPARG_SIMPLE, 9 } } },
     [LOAD_ATTR_SLOT] = { .nuops = 3, .uops = { { _GUARD_TYPE_VERSION, 2, 1 }, { _LOAD_ATTR_SLOT, 1, 3 }, { _PUSH_NULL_CONDITIONAL, OPARG_SIMPLE, 9 } } },
-    [LOAD_ATTR_WITH_HINT] = { .nuops = 3, .uops = { { _GUARD_TYPE_VERSION, 2, 1 }, { _LOAD_ATTR_WITH_HINT, 1, 3 }, { _PUSH_NULL_CONDITIONAL, OPARG_SIMPLE, 9 } } },
+    [LOAD_ATTR_WITH_HINT] = { .nuops = 4, .uops = { { _GUARD_TYPE_VERSION, 2, 1 }, { _LOAD_ATTR_WITH_HINT, 1, 3 }, { _POP_TOP, OPARG_SIMPLE, 4 }, { _PUSH_NULL_CONDITIONAL, OPARG_SIMPLE, 9 } } },
     [LOAD_BUILD_CLASS] = { .nuops = 1, .uops = { { _LOAD_BUILD_CLASS, OPARG_SIMPLE, 0 } } },
     [LOAD_COMMON_CONSTANT] = { .nuops = 1, .uops = { { _LOAD_COMMON_CONSTANT, OPARG_SIMPLE, 0 } } },
     [LOAD_CONST] = { .nuops = 1, .uops = { { _LOAD_CONST, OPARG_SIMPLE, 0 } } },

Include/internal/pycore_uop_ids.h

@@ -807,7 +807,7 @@ extern "C" {
 #define _LOAD_ATTR_NONDESCRIPTOR_WITH_VALUES_r11 1000
 #define _LOAD_ATTR_PROPERTY_FRAME_r11 1001
 #define _LOAD_ATTR_SLOT_r11 1002
-#define _LOAD_ATTR_WITH_HINT_r11 1003
+#define _LOAD_ATTR_WITH_HINT_r12 1003
 #define _LOAD_BUILD_CLASS_r01 1004
 #define _LOAD_BYTECODE_r00 1005
 #define _LOAD_COMMON_CONSTANT_r01 1006

Include/internal/pycore_uop_metadata.h

@@ -191,7 +191,7 @@ const uint32_t _PyUop_Flags[MAX_UOP_ID+1] = {
     [_CHECK_MANAGED_OBJECT_HAS_VALUES] = HAS_DEOPT_FLAG,
     [_LOAD_ATTR_INSTANCE_VALUE] = HAS_DEOPT_FLAG,
     [_LOAD_ATTR_MODULE] = HAS_DEOPT_FLAG | HAS_ESCAPES_FLAG,
-    [_LOAD_ATTR_WITH_HINT] = HAS_ARG_FLAG | HAS_NAME_FLAG | HAS_DEOPT_FLAG | HAS_ESCAPES_FLAG,
+    [_LOAD_ATTR_WITH_HINT] = HAS_ARG_FLAG | HAS_NAME_FLAG | HAS_DEOPT_FLAG,
     [_LOAD_ATTR_SLOT] = HAS_DEOPT_FLAG | HAS_ESCAPES_FLAG,
     [_CHECK_ATTR_CLASS] = HAS_EXIT_FLAG,
     [_LOAD_ATTR_CLASS] = HAS_ESCAPES_FLAG,
@@ -1774,7 +1774,7 @@ const _PyUopCachingInfo _PyUop_Caching[MAX_UOP_ID+1] = {
         .best = { 1, 1, 1, 1 },
         .entries = {
             { -1, -1, -1 },
-            { 1, 1, _LOAD_ATTR_WITH_HINT_r11 },
+            { 2, 1, _LOAD_ATTR_WITH_HINT_r12 },
             { -1, -1, -1 },
             { -1, -1, -1 },
         },
@@ -3542,7 +3542,7 @@ const uint16_t _PyUop_Uncached[MAX_UOP_REGS_ID+1] = {
     [_LOAD_ATTR_INSTANCE_VALUE_r12] = _LOAD_ATTR_INSTANCE_VALUE,
     [_LOAD_ATTR_INSTANCE_VALUE_r23] = _LOAD_ATTR_INSTANCE_VALUE,
     [_LOAD_ATTR_MODULE_r11] = _LOAD_ATTR_MODULE,
-    [_LOAD_ATTR_WITH_HINT_r11] = _LOAD_ATTR_WITH_HINT,
+    [_LOAD_ATTR_WITH_HINT_r12] = _LOAD_ATTR_WITH_HINT,
     [_LOAD_ATTR_SLOT_r11] = _LOAD_ATTR_SLOT,
     [_CHECK_ATTR_CLASS_r01] = _CHECK_ATTR_CLASS,
     [_CHECK_ATTR_CLASS_r11] = _CHECK_ATTR_CLASS,
@@ -4501,7 +4501,7 @@ const char *const _PyOpcode_uop_name[MAX_UOP_REGS_ID+1] = {
     [_LOAD_ATTR_SLOT] = "_LOAD_ATTR_SLOT",
     [_LOAD_ATTR_SLOT_r11] = "_LOAD_ATTR_SLOT_r11",
     [_LOAD_ATTR_WITH_HINT] = "_LOAD_ATTR_WITH_HINT",
-    [_LOAD_ATTR_WITH_HINT_r11] = "_LOAD_ATTR_WITH_HINT_r11",
+    [_LOAD_ATTR_WITH_HINT_r12] = "_LOAD_ATTR_WITH_HINT_r12",
     [_LOAD_BUILD_CLASS] = "_LOAD_BUILD_CLASS",
     [_LOAD_BUILD_CLASS_r01] = "_LOAD_BUILD_CLASS_r01",
     [_LOAD_COMMON_CONSTANT] = "_LOAD_COMMON_CONSTANT",

Lib/test/test_capi/test_opt.py

@@ -2512,6 +2512,27 @@ class TestUopsOptimization(unittest.TestCase):
         self.assertNotIn("_POP_TOP", uops)
         self.assertIn("_POP_TOP_NOP", uops)
 
+    def test_load_attr_with_hint(self):
+        def testfunc(n):
+            class C:
+                pass
+            c = C()
+            c.x = 42
+            for i in range(_testinternalcapi.SHARED_KEYS_MAX_SIZE - 1):
+                setattr(c, f"_{i}", None)
+            x = 0
+            for i in range(n):
+                x += c.x
+            return x
+        res, ex = self._run_with_optimizer(testfunc, TIER2_THRESHOLD)
+        self.assertEqual(res, 42 * TIER2_THRESHOLD)
+        self.assertIsNotNone(ex)
+        uops = get_opnames(ex)
+
+        self.assertIn("_LOAD_ATTR_WITH_HINT", uops)
+        self.assertNotIn("_POP_TOP", uops)
+        self.assertIn("_POP_TOP_NOP", uops)
+
     def test_int_add_op_refcount_elimination(self):
         def testfunc(n):
             c = 1

Python/bytecodes.c

@@ -2426,7 +2426,7 @@ dummy_func(
             unused/5 +
             _PUSH_NULL_CONDITIONAL;
 
-        op(_LOAD_ATTR_WITH_HINT, (hint/1, owner -- attr)) {
+        op(_LOAD_ATTR_WITH_HINT, (hint/1, owner -- attr, o)) {
             PyObject *owner_o = PyStackRef_AsPyObjectBorrow(owner);
             assert(Py_TYPE(owner_o)->tp_flags & Py_TPFLAGS_MANAGED_DICT);
             PyDictObject *dict = _PyObject_GetManagedDict(owner_o);
@@ -2462,13 +2462,15 @@ dummy_func(
 #else
             attr = PyStackRef_FromPyObjectNew(attr_o);
 #endif
-            PyStackRef_CLOSE(owner);
+            o = owner;
+            DEAD(owner);
         }
 
         macro(LOAD_ATTR_WITH_HINT) =
             unused/1 +
             _GUARD_TYPE_VERSION +
             _LOAD_ATTR_WITH_HINT +
+            POP_TOP +
             unused/5 +
             _PUSH_NULL_CONDITIONAL;
 

Python/executor_cases.c.h

@@ -8261,11 +8261,12 @@
             break;
         }
 
-        case _LOAD_ATTR_WITH_HINT_r11: {
+        case _LOAD_ATTR_WITH_HINT_r12: {
             CHECK_CURRENT_CACHED_VALUES(1);
             assert(WITHIN_STACK_BOUNDS_IGNORING_CACHE());
             _PyStackRef owner;
             _PyStackRef attr;
+            _PyStackRef o;
             _PyStackRef _stack_item_0 = _tos_cache0;
             oparg = CURRENT_OPARG();
             owner = _stack_item_0;
@@ -8339,18 +8340,11 @@
             #else
             attr = PyStackRef_FromPyObjectNew(attr_o);
             #endif
-            stack_pointer[0] = attr;
-            stack_pointer += 1;
-            ASSERT_WITHIN_STACK_BOUNDS(__FILE__, __LINE__);
-            _PyFrame_SetStackPointer(frame, stack_pointer);
-            PyStackRef_CLOSE(owner);
-            stack_pointer = _PyFrame_GetStackPointer(frame);
+            o = owner;
+            _tos_cache1 = o;
             _tos_cache0 = attr;
-            _tos_cache1 = PyStackRef_ZERO_BITS;
             _tos_cache2 = PyStackRef_ZERO_BITS;
-            SET_CURRENT_CACHED_VALUES(1);
-            stack_pointer += -1;
-            ASSERT_WITHIN_STACK_BOUNDS(__FILE__, __LINE__);
+            SET_CURRENT_CACHED_VALUES(2);
             assert(WITHIN_STACK_BOUNDS_IGNORING_CACHE());
             break;
         }

Python/generated_cases.c.h

@@ -8280,6 +8280,8 @@
             static_assert(INLINE_CACHE_ENTRIES_LOAD_ATTR == 9, "incorrect cache size");
             _PyStackRef owner;
             _PyStackRef attr;
+            _PyStackRef o;
+            _PyStackRef value;
             _PyStackRef *null;
             /* Skip 1 cache entry */
             // _GUARD_TYPE_VERSION
@@ -8359,9 +8361,14 @@
                 #else
                 attr = PyStackRef_FromPyObjectNew(attr_o);
                 #endif
+                o = owner;
+            }
+            // _POP_TOP
+            {
+                value = o;
                 stack_pointer[-1] = attr;
                 _PyFrame_SetStackPointer(frame, stack_pointer);
-                PyStackRef_CLOSE(owner);
+                PyStackRef_XCLOSE(value);
                 stack_pointer = _PyFrame_GetStackPointer(frame);
             }
             /* Skip 5 cache entries */

Python/optimizer_bytecodes.c

@@ -634,9 +634,10 @@ dummy_func(void) {
         }
     }
 
-    op(_LOAD_ATTR_WITH_HINT, (hint/1, owner -- attr)) {
+    op(_LOAD_ATTR_WITH_HINT, (hint/1, owner -- attr, o)) {
         attr = sym_new_not_null(ctx);
         (void)hint;
+        o = owner;
     }
 
     op(_LOAD_ATTR_SLOT, (index/1, owner -- attr)) {

Python/optimizer_cases.c.h

@@ -1658,11 +1658,19 @@
         }
 
         case _LOAD_ATTR_WITH_HINT: {
+            JitOptRef owner;
             JitOptRef attr;
+            JitOptRef o;
+            owner = stack_pointer[-1];
             uint16_t hint = (uint16_t)this_instr->operand0;
             attr = sym_new_not_null(ctx);
             (void)hint;
+            o = owner;
+            CHECK_STACK_BOUNDS(1);
             stack_pointer[-1] = attr;
+            stack_pointer[0] = o;
+            stack_pointer += 1;
+            ASSERT_WITHIN_STACK_BOUNDS(__FILE__, __LINE__);
             break;
         }