gh-144007: Eliminate redundant refcounting in the JIT for BINARY_OP

Lib/test/test_capi/test_opt.py

@@ -2897,6 +2897,27 @@ def testfunc(n):
         self.assertIn("_POP_TOP_NOP", uops)
         self.assertLessEqual(count_ops(ex, "_POP_TOP"), 2)
 
+    def test_binary_op_refcount_elimination(self):
+        class CustomAdder:
+            def __init__(self, val):
+                self.val = val
+            def __add__(self, other):
+                return CustomAdder(self.val + other.val)
+
+        def testfunc(n):
+            a = CustomAdder(1)
+            b = CustomAdder(2)
+            res = None
+            for _ in range(n):
+                res = a + b
+            return res.val if res else 0
+
+        res, ex = self._run_with_optimizer(testfunc, TIER2_THRESHOLD)
+        self.assertEqual(res, 3)
+        self.assertIsNotNone(ex)
+        uops = get_opnames(ex)
+        self.assertIn("_BINARY_OP", uops)
+
     def test_binary_op_extend_float_long_add_refcount_elimination(self):
         def testfunc(n):
             a = 1.5
@@ -2911,8 +2932,6 @@ def testfunc(n):
         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):

Misc/NEWS.d/next/Core_and_Builtins/2026-01-19-01-56-44.gh-issue-144007.1xjdBf.rst

@@ -0,0 +1 @@
+Eliminate redundant refcounting in the JIT for ``BINARY_OP``.

Modules/_testinternalcapi/test_cases.c.h

@@ -32,6 +32,9 @@
             _PyStackRef lhs;
             _PyStackRef rhs;
             _PyStackRef res;
+            _PyStackRef l;
+            _PyStackRef r;
+            _PyStackRef value;
             // _SPECIALIZE_BINARY_OP
             {
                 rhs = stack_pointer[-1];
@@ -65,18 +68,26 @@
                     JUMP_TO_LABEL(error);
                 }
                 res = PyStackRef_FromPyObjectSteal(res_o);
+                l = lhs;
+                r = rhs;
+            }
+            // _POP_TOP
+            {
+                value = r;
+                stack_pointer[-2] = res;
+                stack_pointer[-1] = l;
                 _PyFrame_SetStackPointer(frame, stack_pointer);
-                _PyStackRef tmp = lhs;
-                lhs = res;
-                stack_pointer[-2] = lhs;
-                PyStackRef_CLOSE(tmp);
-                tmp = rhs;
-                rhs = PyStackRef_NULL;
-                stack_pointer[-1] = rhs;
-                PyStackRef_CLOSE(tmp);
+                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);
             }
             DISPATCH();
         }

Python/bytecodes.c

@@ -5119,7 +5119,7 @@ dummy_func(
             assert(oparg <= NB_OPARG_LAST);
         }
 
-        op(_BINARY_OP, (lhs, rhs -- res)) {
+        op(_BINARY_OP, (lhs, rhs -- res, l, r)) {
             PyObject *lhs_o = PyStackRef_AsPyObjectBorrow(lhs);
             PyObject *rhs_o = PyStackRef_AsPyObjectBorrow(rhs);
 
@@ -5129,10 +5129,13 @@ dummy_func(
                 ERROR_NO_POP();
             }
             res = PyStackRef_FromPyObjectSteal(res_o);
-            DECREF_INPUTS();
+            l = lhs;
+            r = rhs;
+            DEAD(lhs);
+            DEAD(rhs);
         }
 
-        macro(BINARY_OP) = _SPECIALIZE_BINARY_OP + unused/4 + _BINARY_OP;
+        macro(BINARY_OP) = _SPECIALIZE_BINARY_OP + unused/4 + _BINARY_OP + POP_TOP + POP_TOP;
 
         pure replicate(2:4) inst(SWAP, (bottom, unused[oparg-2], top --
                     bottom, unused[oparg-2], top)) {

Python/optimizer_bytecodes.c

@@ -211,7 +211,9 @@ dummy_func(void) {
         sym_set_type(left, &PyFloat_Type);
     }
 
-    op(_BINARY_OP, (lhs, rhs -- res)) {
+    op(_BINARY_OP, (lhs, rhs -- res, l, r)) {
+        l = lhs;
+        r = rhs;
         REPLACE_OPCODE_IF_EVALUATES_PURE(lhs, rhs, res);
         bool lhs_int = sym_matches_type(lhs, &PyLong_Type);
         bool rhs_int = sym_matches_type(rhs, &PyLong_Type);