Eliminate redundant refcounting for BINARY_OP_EXTEND

Author: aisk

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);
@@ -838,12 +838,18 @@ dummy_func(
             STAT_INC(BINARY_OP, hit);
 
             PyObject *res_o = d->action(left_o, right_o);
-            DECREF_INPUTS();
+            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)));