pythongh-115999: Specialize LOAD_GLOBAL in free-threaded builds (python#126607)

Enable specialization of LOAD_GLOBAL in free-threaded builds.

Thread-safety of specialization in free-threaded builds is provided by the following:

A critical section is held on both the globals and builtins objects during specialization. This ensures we get an atomic view of both builtins and globals during specialization.
Generation of new keys versions is made atomic in free-threaded builds.
Existing helpers are used to atomically modify the opcode.
Thread-safety of specialized instructions in free-threaded builds is provided by the following:

Relaxed atomics are used when loading and storing dict keys versions. This avoids potential data races as the dict keys versions are read without holding the dictionary's per-object lock in version guards.
Dicts keys objects are passed from keys version guards to the downstream uops. This ensures that we are loading from the correct offset in the keys object. Once a unicode key has been stored in a keys object for a combined dictionary in free-threaded builds, the offset that it is stored in will never be reused for a different key. Once the version guard passes, we know that we are reading from the correct offset.
The dictionary read fast-path is used to read values from the dictionary once we know the correct offset.

Author: mpage

Include/internal/pycore_dict.h

@@ -90,6 +90,17 @@ extern PyObject *_PyDict_FromKeys(PyObject *, PyObject *, PyObject *);
 extern uint32_t _PyDictKeys_GetVersionForCurrentState(
         PyInterpreterState *interp, PyDictKeysObject *dictkeys);
 
+/* Gets a version number unique to the current state of the keys of dict, if possible.
+ *
+ * In free-threaded builds ensures that the dict can be used for lock-free
+ * reads if a version was assigned.
+ *
+ * The caller must hold the per-object lock on dict.
+ *
+ * Returns the version number, or zero if it was not possible to get a version number. */
+extern uint32_t _PyDict_GetKeysVersionForCurrentState(
+        PyInterpreterState *interp, PyDictObject *dict);
+
 extern size_t _PyDict_KeysSize(PyDictKeysObject *keys);
 
 extern void _PyDictKeys_DecRef(PyDictKeysObject *keys);

Include/internal/pycore_object.h

@@ -14,6 +14,7 @@ extern "C" {
 #include "pycore_interp.h"        // PyInterpreterState.gc
 #include "pycore_pyatomic_ft_wrappers.h"  // FT_ATOMIC_STORE_PTR_RELAXED
 #include "pycore_pystate.h"       // _PyInterpreterState_GET()
+#include "pycore_stackref.h"
 #include "pycore_uniqueid.h"      // _PyObject_ThreadIncrefSlow()
 
 // This value is added to `ob_ref_shared` for objects that use deferred
@@ -595,6 +596,20 @@ _Py_TryIncrefCompare(PyObject **src, PyObject *op)
     return 1;
 }
 
+static inline int
+_Py_TryIncrefCompareStackRef(PyObject **src, PyObject *op, _PyStackRef *out)
+{
+    if (_Py_IsImmortal(op) || _PyObject_HasDeferredRefcount(op)) {
+        *out = (_PyStackRef){ .bits = (intptr_t)op | Py_TAG_DEFERRED };
+        return 1;
+    }
+    if (_Py_TryIncrefCompare(src, op)) {
+        *out = PyStackRef_FromPyObjectSteal(op);
+        return 1;
+    }
+    return 0;
+}
+
 /* Loads and increfs an object from ptr, which may contain a NULL value.
    Safe with concurrent (atomic) updates to ptr.
    NOTE: The writer must set maybe-weakref on the stored object! */

Lib/test/test_opcache.py

@@ -546,7 +546,6 @@ def count_args(self, *args):
 
 
 @threading_helper.requires_working_threading()
-@requires_specialization
 class TestRacesDoNotCrash(TestBase):
     # Careful with these. Bigger numbers have a higher chance of catching bugs,
     # but you can also burn through a *ton* of type/dict/function versions:
@@ -588,6 +587,7 @@ def assert_races_do_not_crash(
             for writer in writers:
                 writer.join()
 
+    @requires_specialization
     def test_binary_subscr_getitem(self):
         def get_items():
             class C:
@@ -617,6 +617,7 @@ def write(items):
         opname = "BINARY_SUBSCR_GETITEM"
         self.assert_races_do_not_crash(opname, get_items, read, write)
 
+    @requires_specialization
     def test_binary_subscr_list_int(self):
         def get_items():
             items = []
@@ -640,6 +641,7 @@ def write(items):
         opname = "BINARY_SUBSCR_LIST_INT"
         self.assert_races_do_not_crash(opname, get_items, read, write)
 
+    @requires_specialization
     def test_for_iter_gen(self):
         def get_items():
             def g():
@@ -671,6 +673,7 @@ def write(items):
         opname = "FOR_ITER_GEN"
         self.assert_races_do_not_crash(opname, get_items, read, write)
 
+    @requires_specialization
     def test_for_iter_list(self):
         def get_items():
             items = []
@@ -692,6 +695,7 @@ def write(items):
         opname = "FOR_ITER_LIST"
         self.assert_races_do_not_crash(opname, get_items, read, write)
 
+    @requires_specialization
     def test_load_attr_class(self):
         def get_items():
             class C:
@@ -721,6 +725,7 @@ def write(items):
         opname = "LOAD_ATTR_CLASS"
         self.assert_races_do_not_crash(opname, get_items, read, write)
 
+    @requires_specialization
     def test_load_attr_getattribute_overridden(self):
         def get_items():
             class C:
@@ -750,6 +755,7 @@ def write(items):
         opname = "LOAD_ATTR_GETATTRIBUTE_OVERRIDDEN"
         self.assert_races_do_not_crash(opname, get_items, read, write)
 
+    @requires_specialization
     def test_load_attr_instance_value(self):
         def get_items():
             class C:
@@ -773,6 +779,7 @@ def write(items):
         opname = "LOAD_ATTR_INSTANCE_VALUE"
         self.assert_races_do_not_crash(opname, get_items, read, write)
 
+    @requires_specialization
     def test_load_attr_method_lazy_dict(self):
         def get_items():
             class C(Exception):
@@ -802,6 +809,7 @@ def write(items):
         opname = "LOAD_ATTR_METHOD_LAZY_DICT"
         self.assert_races_do_not_crash(opname, get_items, read, write)
 
+    @requires_specialization
     def test_load_attr_method_no_dict(self):
         def get_items():
             class C:
@@ -832,6 +840,7 @@ def write(items):
         opname = "LOAD_ATTR_METHOD_NO_DICT"
         self.assert_races_do_not_crash(opname, get_items, read, write)
 
+    @requires_specialization
     def test_load_attr_method_with_values(self):
         def get_items():
             class C:
@@ -861,6 +870,7 @@ def write(items):
         opname = "LOAD_ATTR_METHOD_WITH_VALUES"
         self.assert_races_do_not_crash(opname, get_items, read, write)
 
+    @requires_specialization
     def test_load_attr_module(self):
         def get_items():
             items = []
@@ -885,6 +895,7 @@ def write(items):
         opname = "LOAD_ATTR_MODULE"
         self.assert_races_do_not_crash(opname, get_items, read, write)
 
+    @requires_specialization
     def test_load_attr_property(self):
         def get_items():
             class C:
@@ -914,6 +925,7 @@ def write(items):
         opname = "LOAD_ATTR_PROPERTY"
         self.assert_races_do_not_crash(opname, get_items, read, write)
 
+    @requires_specialization
     def test_load_attr_with_hint(self):
         def get_items():
             class C:
@@ -940,6 +952,7 @@ def write(items):
         opname = "LOAD_ATTR_WITH_HINT"
         self.assert_races_do_not_crash(opname, get_items, read, write)
 
+    @requires_specialization_ft
     def test_load_global_module(self):
         def get_items():
             items = []
@@ -961,6 +974,7 @@ def write(items):
             opname, get_items, read, write, check_items=True
         )
 
+    @requires_specialization
     def test_store_attr_instance_value(self):
         def get_items():
             class C:
@@ -983,6 +997,7 @@ def write(items):
         opname = "STORE_ATTR_INSTANCE_VALUE"
         self.assert_races_do_not_crash(opname, get_items, read, write)
 
+    @requires_specialization
     def test_store_attr_with_hint(self):
         def get_items():
             class C:
@@ -1008,6 +1023,7 @@ def write(items):
         opname = "STORE_ATTR_WITH_HINT"
         self.assert_races_do_not_crash(opname, get_items, read, write)
 
+    @requires_specialization
     def test_store_subscr_list_int(self):
         def get_items():
             items = []
@@ -1031,6 +1047,7 @@ def write(items):
         opname = "STORE_SUBSCR_LIST_INT"
         self.assert_races_do_not_crash(opname, get_items, read, write)
 
+    @requires_specialization
     def test_unpack_sequence_list(self):
         def get_items():
             items = []

Objects/dictobject.c

@@ -1285,6 +1285,20 @@ ensure_shared_on_resize(PyDictObject *mp)
 #endif
 }
 
+static inline void
+ensure_shared_on_keys_version_assignment(PyDictObject *mp)
+{
+    ASSERT_DICT_LOCKED((PyObject *) mp);
+    #ifdef Py_GIL_DISABLED
+    if (!IS_DICT_SHARED(mp)) {
+        // This ensures that a concurrent resize operation will delay
+        // freeing the old keys or values using QSBR, which is necessary to
+        // safely allow concurrent reads without locking.
+        SET_DICT_SHARED(mp);
+    }
+    #endif
+}
+
 #ifdef Py_GIL_DISABLED
 
 static inline Py_ALWAYS_INLINE int
@@ -1644,7 +1658,7 @@ insert_combined_dict(PyInterpreterState *interp, PyDictObject *mp,
     }
 
     _PyDict_NotifyEvent(interp, PyDict_EVENT_ADDED, mp, key, value);
-    mp->ma_keys->dk_version = 0;
+    FT_ATOMIC_STORE_UINT32_RELAXED(mp->ma_keys->dk_version, 0);
 
     Py_ssize_t hashpos = find_empty_slot(mp->ma_keys, hash);
     dictkeys_set_index(mp->ma_keys, hashpos, mp->ma_keys->dk_nentries);
@@ -1686,7 +1700,7 @@ insert_split_key(PyDictKeysObject *keys, PyObject *key, Py_hash_t hash)
     ix = unicodekeys_lookup_unicode(keys, key, hash);
     if (ix == DKIX_EMPTY && keys->dk_usable > 0) {
         // Insert into new slot
-        keys->dk_version = 0;
+        FT_ATOMIC_STORE_UINT32_RELAXED(keys->dk_version, 0);
         Py_ssize_t hashpos = find_empty_slot(keys, hash);
         ix = keys->dk_nentries;
         dictkeys_set_index(keys, hashpos, ix);
@@ -2617,7 +2631,7 @@ delitem_common(PyDictObject *mp, Py_hash_t hash, Py_ssize_t ix,
         ASSERT_CONSISTENT(mp);
     }
     else {
-        mp->ma_keys->dk_version = 0;
+        FT_ATOMIC_STORE_UINT32_RELAXED(mp->ma_keys->dk_version, 0);
         dictkeys_set_index(mp->ma_keys, hashpos, DKIX_DUMMY);
         if (DK_IS_UNICODE(mp->ma_keys)) {
             PyDictUnicodeEntry *ep = &DK_UNICODE_ENTRIES(mp->ma_keys)[ix];
@@ -4429,7 +4443,7 @@ dict_popitem_impl(PyDictObject *self)
             return NULL;
         }
     }
-    self->ma_keys->dk_version = 0;
+    FT_ATOMIC_STORE_UINT32_RELAXED(self->ma_keys->dk_version, 0);
 
     /* Pop last item */
     PyObject *key, *value;
@@ -7417,20 +7431,54 @@ _PyDictKeys_DecRef(PyDictKeysObject *keys)
     dictkeys_decref(interp, keys, false);
 }
 
-uint32_t _PyDictKeys_GetVersionForCurrentState(PyInterpreterState *interp,
-                                               PyDictKeysObject *dictkeys)
+static inline uint32_t
+get_next_dict_keys_version(PyInterpreterState *interp)
 {
-    if (dictkeys->dk_version != 0) {
-        return dictkeys->dk_version;
-    }
+#ifdef Py_GIL_DISABLED
+    uint32_t v;
+    do {
+        v = _Py_atomic_load_uint32_relaxed(
+            &interp->dict_state.next_keys_version);
+        if (v == 0) {
+            return 0;
+        }
+    } while (!_Py_atomic_compare_exchange_uint32(
+        &interp->dict_state.next_keys_version, &v, v + 1));
+#else
     if (interp->dict_state.next_keys_version == 0) {
         return 0;
     }
     uint32_t v = interp->dict_state.next_keys_version++;
-    dictkeys->dk_version = v;
+#endif
     return v;
 }
 
+// In free-threaded builds the caller must ensure that the keys object is not
+// being mutated concurrently by another thread.
+uint32_t
+_PyDictKeys_GetVersionForCurrentState(PyInterpreterState *interp,
+                                      PyDictKeysObject *dictkeys)
+{
+    uint32_t dk_version = FT_ATOMIC_LOAD_UINT32_RELAXED(dictkeys->dk_version);
+    if (dk_version != 0) {
+        return dk_version;
+    }
+    dk_version = get_next_dict_keys_version(interp);
+    FT_ATOMIC_STORE_UINT32_RELAXED(dictkeys->dk_version, dk_version);
+    return dk_version;
+}
+
+uint32_t
+_PyDict_GetKeysVersionForCurrentState(PyInterpreterState *interp,
+                                      PyDictObject *dict)
+{
+    ASSERT_DICT_LOCKED((PyObject *) dict);
+    uint32_t dk_version =
+        _PyDictKeys_GetVersionForCurrentState(interp, dict->ma_keys);
+    ensure_shared_on_keys_version_assignment(dict);
+    return dk_version;
+}
+
 static inline int
 validate_watcher_id(PyInterpreterState *interp, int watcher_id)
 {

Objects/funcobject.c

@@ -289,12 +289,14 @@ functions is running.
 
 */
 
+#ifndef Py_GIL_DISABLED
 static inline struct _func_version_cache_item *
 get_cache_item(PyInterpreterState *interp, uint32_t version)
 {
     return interp->func_state.func_version_cache +
            (version % FUNC_VERSION_CACHE_SIZE);
 }
+#endif
 
 void
 _PyFunction_SetVersion(PyFunctionObject *func, uint32_t version)