gh-134584: Eliminate redundant refcounting from _CALL_TYPE_1

Lib/test/test_capi/test_opt.py

@@ -1819,9 +1819,10 @@ def testfunc(n):
         self.assertIsNotNone(ex)
         uops = get_opnames(ex)
         # When the result of type(...) is known, _CALL_TYPE_1 is replaced with
-        # _POP_CALL_ONE_LOAD_CONST_INLINE_BORROW which is optimized away in
+        # _SWAP_CALL_ONE_LOAD_CONST_INLINE_BORROW which is optimized away in
         # remove_unneeded_uops.
         self.assertNotIn("_CALL_TYPE_1", uops)
+        self.assertNotIn("_SWAP_CALL_ONE_LOAD_CONST_INLINE_BORROW", uops)
         self.assertNotIn("_POP_CALL_ONE_LOAD_CONST_INLINE_BORROW", uops)
         self.assertNotIn("_POP_CALL_LOAD_CONST_INLINE_BORROW", uops)
         self.assertNotIn("_POP_TOP_LOAD_CONST_INLINE_BORROW", uops)
@@ -1841,6 +1842,21 @@ def testfunc(n):
         uops = get_opnames(ex)
         self.assertNotIn("_GUARD_IS_NOT_NONE_POP", uops)
 
+    def test_call_type_1_pop_top(self):
+        def testfunc(n):
+            x = 0
+            for _ in range(n):
+                foo = eval('42')
+                x += type(foo) is int
+            return x
+
+        res, ex = self._run_with_optimizer(testfunc, TIER2_THRESHOLD)
+        self.assertEqual(res, TIER2_THRESHOLD)
+        self.assertIsNotNone(ex)
+        uops = get_opnames(ex)
+        self.assertIn("_CALL_TYPE_1", uops)
+        self.assertIn("_POP_TOP_NOP", uops)
+
     def test_call_str_1(self):
         def testfunc(n):
             x = 0

Misc/NEWS.d/next/Core_and_Builtins/2025-06-23-20-54-15.gh-issue-134584.ZNcziF.rst

@@ -0,0 +1 @@
+Eliminate redundant refcounting from ``_CALL_TYPE_1``. Patch by Tomas Roun

Python/bytecodes.c

@@ -4030,25 +4030,23 @@ dummy_func(
             DEOPT_IF(callable_o != (PyObject *)&PyType_Type);
         }
 
-        op(_CALL_TYPE_1, (callable, null, arg -- res)) {
+        op(_CALL_TYPE_1, (callable, null, arg -- res, a)) {
             PyObject *arg_o = PyStackRef_AsPyObjectBorrow(arg);
 
             assert(oparg == 1);
-            DEAD(null);
-            DEAD(callable);
-            (void)callable; // Silence compiler warnings about unused variables
-            (void)null;
             STAT_INC(CALL, hit);
+            INPUTS_DEAD();
             res = PyStackRef_FromPyObjectNew(Py_TYPE(arg_o));
-            PyStackRef_CLOSE(arg);
+            a = arg;
         }
 
         macro(CALL_TYPE_1) =
             unused/1 +
             unused/2 +
             _GUARD_NOS_NULL +
             _GUARD_CALLABLE_TYPE_1 +
-            _CALL_TYPE_1;
+            _CALL_TYPE_1 +
+            POP_TOP;
 
         op(_GUARD_CALLABLE_STR_1, (callable, unused, unused -- callable, unused, unused)) {
             PyObject *callable_o = PyStackRef_AsPyObjectBorrow(callable);
@@ -5360,6 +5358,17 @@ dummy_func(
             value = PyStackRef_FromPyObjectBorrow(ptr);
         }
 
+        tier2 op(_SWAP_CALL_ONE_LOAD_CONST_INLINE_BORROW, (ptr/4, callable, null, arg -- value, a)) {
+            PyStackRef_CLOSE(arg);
+            (void)null; // Silence compiler warnings about unused variables
+            (void)callable;
+            DEAD(null);
+            DEAD(callable);
+            assert(_Py_IsImmortal(PyStackRef_AsPyObjectBorrow(callable)));
+            value = PyStackRef_FromPyObjectBorrow(ptr);
+            a = arg;
+        }
+
         tier2 op(_LOAD_CONST_UNDER_INLINE, (ptr/4, old -- value, new)) {
             new = old;
             DEAD(old);

Python/optimizer_analysis.c

@@ -566,7 +566,9 @@ const uint16_t op_without_push[MAX_UOP_ID + 1] = {
     [_POP_TOP_LOAD_CONST_INLINE] = _POP_TOP,
     [_POP_TOP_LOAD_CONST_INLINE_BORROW] = _POP_TOP,
     [_POP_TWO_LOAD_CONST_INLINE_BORROW] = _POP_TWO,
+    [_POP_CALL_ONE_LOAD_CONST_INLINE_BORROW] = _POP_CALL_ONE,
     [_POP_CALL_TWO_LOAD_CONST_INLINE_BORROW] = _POP_CALL_TWO,
+    [_SWAP_CALL_ONE_LOAD_CONST_INLINE_BORROW] = _POP_CALL_ONE_LOAD_CONST_INLINE_BORROW,
 };
 
 const bool op_skip[MAX_UOP_ID + 1] = {
@@ -578,6 +580,10 @@ const bool op_skip[MAX_UOP_ID + 1] = {
 
 const uint16_t op_without_pop[MAX_UOP_ID + 1] = {
     [_POP_TOP] = _NOP,
+    [_POP_TOP_NOP] = _NOP,
+    [_POP_TOP_INT] = _NOP,
+    [_POP_TOP_FLOAT] = _NOP,
+    [_POP_TOP_UNICODE] = _NOP,
     [_POP_TOP_LOAD_CONST_INLINE] = _LOAD_CONST_INLINE,
     [_POP_TOP_LOAD_CONST_INLINE_BORROW] = _LOAD_CONST_INLINE_BORROW,
     [_POP_TWO] = _POP_TOP,

Python/optimizer_bytecodes.c

@@ -561,6 +561,11 @@ dummy_func(void) {
         value = PyJitRef_Borrow(sym_new_const(ctx, ptr));
     }
 
+    op(_SWAP_CALL_ONE_LOAD_CONST_INLINE_BORROW, (ptr/4, unused, unused, arg -- value, a)) {
+        value = PyJitRef_Borrow(sym_new_const(ctx, ptr));
+        a = arg;
+    }
+
     op(_POP_TOP, (value -- )) {
         PyTypeObject *typ = sym_get_type(value);
         if (PyJitRef_IsBorrowed(value) ||
@@ -969,16 +974,17 @@ dummy_func(void) {
        next = sym_new_type(ctx, &PyLong_Type);
     }
 
-    op(_CALL_TYPE_1, (unused, unused, arg -- res)) {
+    op(_CALL_TYPE_1, (unused, unused, arg -- res, a)) {
         PyObject* type = (PyObject *)sym_get_type(arg);
         if (type) {
             res = sym_new_const(ctx, type);
-            REPLACE_OP(this_instr, _POP_CALL_ONE_LOAD_CONST_INLINE_BORROW, 0,
+            REPLACE_OP(this_instr, _SWAP_CALL_ONE_LOAD_CONST_INLINE_BORROW, 0,
                        (uintptr_t)type);
         }
         else {
             res = sym_new_not_null(ctx);
         }
+        a = arg;
     }
 
     op(_CALL_STR_1, (unused, unused, arg -- res)) {