gh-xxxxx: Add type-local type cache

Include/cpython/object.h

@@ -259,6 +259,9 @@ struct _specialization_cache {
     PyObject *getitem;
     uint32_t getitem_version;
     PyObject *init;
+#ifdef Py_GIL_DISABLED
+    struct local_type_cache *local_type_cache;
+#endif
 };
 
 /* The *real* layout of a type object when allocated on the heap */

Include/internal/pycore_typeobject.h

@@ -71,6 +71,31 @@ struct type_cache {
     struct type_cache_entry hashtable[1 << MCACHE_SIZE_EXP];
 };
 
+#ifdef Py_GIL_DISABLED
+
+// Type attribute lookup cache which is type-specific. Only used
+// for heap types where we store a small additional cache in free-threaded
+// builds which can be accessed without any locking.
+#define LOCAL_TYPE_CACHE_SIZE 64
+#define LOCAL_TYPE_CACHE_MAX_ENTRIES 48
+#define LOCAL_TYPE_CACHE_PROBE 3
+
+struct local_type_cache_entry {
+    PyObject *name;     // reference to exactly a str or NULL
+    PyObject *value;    // owned reference or NULL
+};
+
+struct local_type_cache {
+    unsigned int tp_version_tag;
+    unsigned int cache_count;
+    struct local_type_cache_entry entries[LOCAL_TYPE_CACHE_SIZE];
+#if 0
+    int hits[LOCAL_TYPE_CACHE_SIZE], probes[LOCAL_TYPE_CACHE_SIZE], miss[LOCAL_TYPE_CACHE_SIZE];
+#endif
+};
+
+#endif
+
 typedef struct {
     PyTypeObject *type;
     int isbuiltin;
@@ -85,6 +110,9 @@ typedef struct {
        are also some diagnostic uses for the list of weakrefs,
        so we still keep it. */
     PyObject *tp_weaklist;
+#ifdef Py_GIL_DISABLED
+    struct local_type_cache local_cache;
+#endif
 } managed_static_type_state;
 
 #define TYPE_VERSION_CACHE_SIZE (1<<12)  /* Must be a power of 2 */

Lib/test/test_sys.py

@@ -1731,6 +1731,7 @@ def delx(self): del self.__x
         s = vsize(fmt)
         check(int, s)
         typeid = 'n' if support.Py_GIL_DISABLED else ''
+        local_type_cache = 'P' if support.Py_GIL_DISABLED else ''
         # class
         s = vsize(fmt +                 # PyTypeObject
                   '4P'                  # PyAsyncMethods
@@ -1741,6 +1742,7 @@ def delx(self): del self.__x
                   '7P'
                   '1PIP'                # Specializer cache
                   + typeid              # heap type id (free-threaded only)
+                  + local_type_cache    # local type cache (free-threaded only)
                   )
         class newstyleclass(object): pass
         # Separate block for PyDictKeysObject with 8 keys and 5 entries