gh-115999: Add partial free-thread specialzation for BINARY_SUBSCR

Author: corona10

Python/bytecodes.c

@@ -704,7 +704,7 @@ dummy_func(
         };
 
         specializing op(_SPECIALIZE_BINARY_SUBSCR, (counter/1, container, sub -- container, sub)) {
-            #if ENABLE_SPECIALIZATION
+            #if ENABLE_SPECIALIZATION_FT
             assert(frame->stackpointer == NULL);
             if (ADAPTIVE_COUNTER_TRIGGERS(counter)) {
                 next_instr = this_instr;
@@ -713,7 +713,7 @@ dummy_func(
             }
             OPCODE_DEFERRED_INC(BINARY_SUBSCR);
             ADVANCE_ADAPTIVE_COUNTER(this_instr[1].counter);
-            #endif  /* ENABLE_SPECIALIZATION */
+            #endif  /* ENABLE_SPECIALIZATION_FT */
         }
 
         op(_BINARY_SUBSCR, (container, sub -- res)) {
@@ -790,10 +790,15 @@ dummy_func(
             // Deopt unless 0 <= sub < PyList_Size(list)
             DEOPT_IF(!_PyLong_IsNonNegativeCompact((PyLongObject *)sub));
             Py_ssize_t index = ((PyLongObject*)sub)->long_value.ob_digit[0];
-            DEOPT_IF(index >= PyList_GET_SIZE(list));
+            DEOPT_IF(!LOCK_OBJECT(list));
+            if (index >= PyList_GET_SIZE(list)) {
+                UNLOCK_OBJECT(list);
+                DEOPT_IF(true);
+            }
             STAT_INC(BINARY_SUBSCR, hit);
             PyObject *res_o = PyList_GET_ITEM(list, index);
             assert(res_o != NULL);
+            UNLOCK_OBJECT(list);
             Py_INCREF(res_o);
             PyStackRef_CLOSE_SPECIALIZED(sub_st, (destructor)PyObject_Free);
             DEAD(sub_st);

Python/ceval_macros.h

@@ -295,6 +295,13 @@ GETITEM(PyObject *v, Py_ssize_t i) {
         dtrace_function_entry(frame); \
     }
 
+#ifdef Py_GIL_DISABLED
+#  define LOCK_OBJECT(op) PyMutex_LockFast(&(_PyObject_CAST(op))->ob_mutex._bits)
+#  define UNLOCK_OBJECT(op) PyMutex_Unlock(&(_PyObject_CAST(op))->ob_mutex)
+#else
+#  define LOCK_OBJECT(op) (1)
+#  define UNLOCK_OBJECT(op) ((void)0)
+#endif
 /* This takes a uint16_t instead of a _Py_BackoffCounter,
  * because it is used directly on the cache entry in generated code,
  * which is always an integral type. */

Python/executor_cases.c.h

@@ -1720,12 +1720,12 @@ _Py_Specialize_BinarySubscr(
     assert(ENABLE_SPECIALIZATION);
     assert(_PyOpcode_Caches[BINARY_SUBSCR] ==
            INLINE_CACHE_ENTRIES_BINARY_SUBSCR);
-    _PyBinarySubscrCache *cache = (_PyBinarySubscrCache *)(instr + 1);
     PyTypeObject *container_type = Py_TYPE(container);
+    uint8_t specialized_op;
     if (container_type == &PyList_Type) {
         if (PyLong_CheckExact(sub)) {
             if (_PyLong_IsNonNegativeCompact((PyLongObject *)sub)) {
-                instr->op.code = BINARY_SUBSCR_LIST_INT;
+                specialized_op = BINARY_SUBSCR_LIST_INT;
                 goto success;
             }
             SPECIALIZATION_FAIL(BINARY_SUBSCR, SPEC_FAIL_OUT_OF_RANGE);
@@ -1738,7 +1738,7 @@ _Py_Specialize_BinarySubscr(
     if (container_type == &PyTuple_Type) {
         if (PyLong_CheckExact(sub)) {
             if (_PyLong_IsNonNegativeCompact((PyLongObject *)sub)) {
-                instr->op.code = BINARY_SUBSCR_TUPLE_INT;
+                specialized_op = BINARY_SUBSCR_TUPLE_INT;
                 goto success;
             }
             SPECIALIZATION_FAIL(BINARY_SUBSCR, SPEC_FAIL_OUT_OF_RANGE);
@@ -1751,7 +1751,7 @@ _Py_Specialize_BinarySubscr(
     if (container_type == &PyUnicode_Type) {
         if (PyLong_CheckExact(sub)) {
             if (_PyLong_IsNonNegativeCompact((PyLongObject *)sub)) {
-                instr->op.code = BINARY_SUBSCR_STR_INT;
+                specialized_op = BINARY_SUBSCR_STR_INT;
                 goto success;
             }
             SPECIALIZATION_FAIL(BINARY_SUBSCR, SPEC_FAIL_OUT_OF_RANGE);
@@ -1762,9 +1762,10 @@ _Py_Specialize_BinarySubscr(
         goto fail;
     }
     if (container_type == &PyDict_Type) {
-        instr->op.code = BINARY_SUBSCR_DICT;
+        specialized_op = BINARY_SUBSCR_DICT;
         goto success;
     }
+#ifndef Py_GIL_DISABLED
     PyTypeObject *cls = Py_TYPE(container);
     PyObject *descriptor = _PyType_Lookup(cls, &_Py_ID(__getitem__));
     if (descriptor && Py_TYPE(descriptor) == &PyFunction_Type) {
@@ -1797,21 +1798,17 @@ _Py_Specialize_BinarySubscr(
         // struct _specialization_cache):
         ht->_spec_cache.getitem = descriptor;
         ht->_spec_cache.getitem_version = version;
-        instr->op.code = BINARY_SUBSCR_GETITEM;
+        specialized_op = BINARY_SUBSCR_GETITEM;
         goto success;
     }
+#endif   // Py_GIL_DISABLED
     SPECIALIZATION_FAIL(BINARY_SUBSCR,
                         binary_subscr_fail_kind(container_type, sub));
 fail:
-    STAT_INC(BINARY_SUBSCR, failure);
-    assert(!PyErr_Occurred());
-    instr->op.code = BINARY_SUBSCR;
-    cache->counter = adaptive_counter_backoff(cache->counter);
+    unspecialize(instr);
     return;
 success:
-    STAT_INC(BINARY_SUBSCR, success);
-    assert(!PyErr_Occurred());
-    cache->counter = adaptive_counter_cooldown();
+    specialize(instr, specialized_op);
 }
 
 void