Eliminate redundant refcounting in the JIT for BINARY_OP

Author: aisk

Python/bytecodes.c

@@ -5113,7 +5113,7 @@ 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);
 
@@ -5123,10 +5123,13 @@ 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

@@ -210,8 +210,10 @@ dummy_func(void) {
         sym_set_type(left, &PyFloat_Type);
     }
 
-    op(_BINARY_OP, (lhs, rhs -- res)) {
+    op(_BINARY_OP, (lhs, rhs -- res, l, r)) {
         REPLACE_OPCODE_IF_EVALUATES_PURE(lhs, rhs, res);
+        l = lhs;
+        r = rhs;
         bool lhs_int = sym_matches_type(lhs, &PyLong_Type);
         bool rhs_int = sym_matches_type(rhs, &PyLong_Type);
         bool lhs_float = sym_matches_type(lhs, &PyFloat_Type);