STORE_SUBSCR_LIST_INT and BINARY_OP_SUBSCR_LIST_INT

Author: Fidget-Spinner

Include/internal/pycore_uop_ids.h

@@ -949,6 +949,32 @@ dummy_func(
             DECREF_INPUTS();
         }
 
+        op(_BINARY_OP_SUBSCR_LIST_INT__NO_DECREF_INPUTS, (list_st, sub_st -- res)) {
+            PyObject *sub = PyStackRef_AsPyObjectBorrow(sub_st);
+            PyObject *list = PyStackRef_AsPyObjectBorrow(list_st);
+
+            assert(PyLong_CheckExact(sub));
+            assert(PyList_CheckExact(list));
+
+            // Deopt unless 0 <= sub < PyList_Size(list)
+            DEOPT_IF(!_PyLong_IsNonNegativeCompact((PyLongObject *)sub));
+            Py_ssize_t index = ((PyLongObject*)sub)->long_value.ob_digit[0];
+    #ifdef Py_GIL_DISABLED
+            PyObject *res_o = _PyList_GetItemRef((PyListObject*)list, index);
+                DEOPT_IF(res_o == NULL);
+                STAT_INC(BINARY_OP, hit);
+                res = PyStackRef_FromPyObjectSteal(res_o);
+    #else
+            DEOPT_IF(index >= PyList_GET_SIZE(list));
+            STAT_INC(BINARY_OP, hit);
+            PyObject *res_o = PyList_GET_ITEM(list, index);
+            assert(res_o != NULL);
+            res = PyStackRef_FromPyObjectNew(res_o);
+    #endif
+            STAT_INC(BINARY_OP, hit);
+            INPUTS_DEAD();
+        }
+
         macro(BINARY_OP_SUBSCR_LIST_SLICE) =
             _GUARD_TOS_SLICE + _GUARD_NOS_LIST + unused/5 + _BINARY_OP_SUBSCR_LIST_SLICE;
 
@@ -1152,6 +1178,34 @@ dummy_func(
             Py_DECREF(old_value);
         }
 
+        op(_STORE_SUBSCR_LIST_INT__NO_DECREF_INPUTS, (value, list_st, sub_st -- )) {
+            PyObject *sub = PyStackRef_AsPyObjectBorrow(sub_st);
+            PyObject *list = PyStackRef_AsPyObjectBorrow(list_st);
+
+            assert(PyLong_CheckExact(sub));
+            assert(PyList_CheckExact(list));
+
+            // Ensure nonnegative, zero-or-one-digit ints.
+            DEOPT_IF(!_PyLong_IsNonNegativeCompact((PyLongObject *)sub));
+            Py_ssize_t index = ((PyLongObject*)sub)->long_value.ob_digit[0];
+            DEOPT_IF(!LOCK_OBJECT(list));
+            // Ensure index < len(list)
+            if (index >= PyList_GET_SIZE(list)) {
+                UNLOCK_OBJECT(list);
+                DEOPT_IF(true);
+            }
+            STAT_INC(STORE_SUBSCR, hit);
+
+            PyObject *old_value = PyList_GET_ITEM(list, index);
+            FT_ATOMIC_STORE_PTR_RELEASE(_PyList_ITEMS(list)[index],
+                                        PyStackRef_AsPyObjectSteal(value));
+            assert(old_value != NULL);
+            UNLOCK_OBJECT(list);  // unlock before decrefs!
+            DEAD(sub_st);
+            DEAD(list_st);
+            Py_DECREF(old_value);
+        }
+
         macro(STORE_SUBSCR_DICT) =
             _GUARD_NOS_DICT + unused/1 + _STORE_SUBSCR_DICT;
 

Include/internal/pycore_uop_metadata.h

@@ -270,6 +270,19 @@ dummy_func(void) {
         }
     }
 
+    op(_STORE_SUBSCR_LIST_INT, (value, list_st, sub_st -- )) {
+        if (PyJitRef_IsBorrowed(list_st) && PyJitRef_IsBorrowed(sub_st)) {
+            REPLACE_OP(this_instr, _STORE_SUBSCR_LIST_INT__NO_DECREF_INPUTS, oparg, 0);
+        }
+    }
+
+    op(_BINARY_OP_SUBSCR_LIST_INT, (list_st, sub_st -- res)) {
+        if (PyJitRef_IsBorrowed(list_st) && PyJitRef_IsBorrowed(sub_st)) {
+            REPLACE_OP(this_instr, _BINARY_OP_SUBSCR_LIST_INT__NO_DECREF_INPUTS, oparg, 0);
+        }
+        res = sym_new_not_null(ctx);
+    }
+
     op(_BINARY_OP_SUBTRACT_FLOAT, (left, right -- res)) {
         if (sym_is_const(ctx, left) && sym_is_const(ctx, right)) {
             assert(PyFloat_CheckExact(sym_get_const(ctx, left)));

Python/bytecodes.c

@@ -635,6 +635,7 @@ def has_error_without_pop(op: parser.CodeDef) -> bool:
     "_PyLong_IsCompact",
     "_PyLong_IsNegative",
     "_PyLong_IsNonNegativeCompact",
+    "_PyList_GetItemRef",
     "_PyLong_IsZero",
     "_PyLong_BothAreCompact",
     "_PyCompactLong_Add",