gh-144005: Eliminate redundant refcounting in the JIT for BINARY_OP_EXTEND

Lib/test/test_capi/test_opt.py

@@ -2880,6 +2880,23 @@ def testfunc(n):
         self.assertIn("_POP_TOP_NOP", uops)
         self.assertLessEqual(count_ops(ex, "_POP_TOP"), 2)
 
+    def test_binary_op_extend_float_long_add_refcount_elimination(self):
+        def testfunc(n):
+            a = 1.5
+            b = 2
+            res = 0.0
+            for _ in range(n):
+                res = a + b
+            return res
+
+        res, ex = self._run_with_optimizer(testfunc, TIER2_THRESHOLD)
+        self.assertEqual(res, 3.5)
+        self.assertIsNotNone(ex)
+        uops = get_opnames(ex)
+        self.assertIn("_BINARY_OP_EXTEND", uops)
+        self.assertIn("_POP_TOP_NOP", uops)
+        self.assertLessEqual(count_ops(ex, "_POP_TOP"), 2)
+
     def test_remove_guard_for_slice_list(self):
         def f(n):
             for i in range(n):

Misc/NEWS.d/next/Core_and_Builtins/2026-01-19-01-26-12.gh-issue-144005.Z3O33m.rst

@@ -0,0 +1 @@
+Eliminate redundant refcounting from ``BINARY_OP_EXTEND``.

Modules/_testinternalcapi/test_cases.c.h

@@ -317,6 +317,9 @@
             _PyStackRef left;
             _PyStackRef right;
             _PyStackRef res;
+            _PyStackRef l;
+            _PyStackRef r;
+            _PyStackRef value;
             /* Skip 1 cache entry */
             // _GUARD_BINARY_OP_EXTEND
             {
@@ -348,25 +351,32 @@
                 STAT_INC(BINARY_OP, hit);
                 _PyFrame_SetStackPointer(frame, stack_pointer);
                 PyObject *res_o = d->action(left_o, right_o);
-                _PyStackRef tmp = right;
-                right = PyStackRef_NULL;
-                stack_pointer[-1] = right;
-                PyStackRef_CLOSE(tmp);
-                tmp = left;
-                left = PyStackRef_NULL;
-                stack_pointer[-2] = left;
-                PyStackRef_CLOSE(tmp);
                 stack_pointer = _PyFrame_GetStackPointer(frame);
-                stack_pointer += -2;
-                ASSERT_WITHIN_STACK_BOUNDS(__FILE__, __LINE__);
                 if (res_o == NULL) {
                     JUMP_TO_LABEL(error);
                 }
                 res = PyStackRef_FromPyObjectSteal(res_o);
+                l = left;
+                r = right;
+            }
+            // _POP_TOP
+            {
+                value = r;
+                stack_pointer[-2] = res;
+                stack_pointer[-1] = l;
+                _PyFrame_SetStackPointer(frame, stack_pointer);
+                PyStackRef_XCLOSE(value);
+                stack_pointer = _PyFrame_GetStackPointer(frame);
+            }
+            // _POP_TOP
+            {
+                value = l;
+                stack_pointer += -1;
+                ASSERT_WITHIN_STACK_BOUNDS(__FILE__, __LINE__);
+                _PyFrame_SetStackPointer(frame, stack_pointer);
+                PyStackRef_XCLOSE(value);
+                stack_pointer = _PyFrame_GetStackPointer(frame);
             }
-            stack_pointer[0] = res;
-            stack_pointer += 1;
-            ASSERT_WITHIN_STACK_BOUNDS(__FILE__, __LINE__);
             DISPATCH();
         }
 

Python/bytecodes.c

@@ -829,7 +829,7 @@ dummy_func(
             DEOPT_IF(!res);
         }
 
-       op(_BINARY_OP_EXTEND, (descr/4, left, right -- res)) {
+       op(_BINARY_OP_EXTEND, (descr/4, left, right -- res, l, r)) {
             PyObject *left_o = PyStackRef_AsPyObjectBorrow(left);
             PyObject *right_o = PyStackRef_AsPyObjectBorrow(right);
             assert(INLINE_CACHE_ENTRIES_BINARY_OP == 5);
@@ -838,13 +838,18 @@ dummy_func(
             STAT_INC(BINARY_OP, hit);
 
             PyObject *res_o = d->action(left_o, right_o);
-            DECREF_INPUTS();
-            ERROR_IF(res_o == NULL);
+            if (res_o == NULL) {
+                ERROR_NO_POP();
+            }
             res = PyStackRef_FromPyObjectSteal(res_o);
+            l = left;
+            r = right;
+            DEAD(left);
+            DEAD(right);
         }
 
         macro(BINARY_OP_EXTEND) =
-            unused/1 + _GUARD_BINARY_OP_EXTEND + rewind/-4 + _BINARY_OP_EXTEND;
+            unused/1 + _GUARD_BINARY_OP_EXTEND + rewind/-4 + _BINARY_OP_EXTEND + POP_TOP + POP_TOP;
 
         macro(BINARY_OP_INPLACE_ADD_UNICODE) =
             _GUARD_TOS_UNICODE + _GUARD_NOS_UNICODE + unused/5 + _BINARY_OP_INPLACE_ADD_UNICODE;

Python/optimizer_bytecodes.c

@@ -310,6 +310,12 @@ dummy_func(void) {
         r = right;
     }
 
+    op(_BINARY_OP_EXTEND, (left, right -- res, l, r)) {
+        res = sym_new_not_null(ctx);
+        l = left;
+        r = right;
+    }
+
     op(_BINARY_OP_INPLACE_ADD_UNICODE, (left, right -- res)) {
         if (sym_is_const(ctx, left) && sym_is_const(ctx, right)) {
             assert(PyUnicode_CheckExact(sym_get_const(ctx, left)));