Further optimizations: completely elide the lock, fully implement the descriptor protocol.

CHANGES.rst

@@ -54,8 +54,11 @@
 - Monkey-patching after the :mod:`ssl` module has been imported now
   prints a warning because this can produce ``RecursionError``.
 
-- :class:`gevent.local.local` objects are now about 3.4 times faster
-  reading simple attributes. See :issue:`1020`.
+- :class:`gevent.local.local` objects are now between 3 and 5 times faster
+  getting, setting and deleting attributes on CPython (the fastest
+  access occurs when ``local`` is not subclassed). This involved
+  implementing more of the attribute protocols directly. Please open
+  an issue if you have any compatibility problems. See :issue:`1020`.
 
 1.2.2 (2017-06-05)
 ==================

src/gevent/local.py

@@ -137,14 +137,24 @@ affects what we see:
    Use a weak-reference to clear the greenlet link we establish in case
    the local object dies before the greenlet does.
 
+.. versionchanged:: 1.3a1
+   Implement the methods for attribute access directly, handling
+   descriptors directly here. This allows removing the use of a lock
+   and facilitates greatly improved performance.
+
+.. versionchanged:: 1.3a1
+   The ``__init__`` method of subclasses of ``local`` is no longer
+   called with a lock held. CPython does not use such a lock in its
+   native implementation. This could potentially show as a difference
+   if code that uses multiple dependent attributes in ``__slots__``
+   (which are shared across all greenlets) switches during ``__init__``.
+
 """
 
 from copy import copy
 from weakref import ref
-from functools import partial
 from gevent.hub import getcurrent
 from gevent._compat import PYPY
-from gevent.lock import RLock
 
 __all__ = ["local"]
 
@@ -157,7 +167,7 @@ _oga = object.__getattribute__
 
 class _localimpl(object):
     """A class managing thread-local dicts"""
-    __slots__ = ('key', 'dicts', 'localargs', 'locallock', '__weakref__', 'dict_setter')
+    __slots__ = ('key', 'dicts', 'localargs', '__weakref__',)
 
     def __init__(self):
         # The key used in the Thread objects' attribute dicts.
@@ -225,20 +235,14 @@ class _localimpl(object):
 
 def _get_dict(self):
     impl = _oga(self, '_local__impl')
-    orig_dct = _oga(self, '__dict__')
-    lock = impl.locallock
     try:
         dct = impl.get_dict()
     except KeyError:
-        # it's OK to acquire the lock here and not earlier, because the above code won't switch out
-        # however, subclassed __init__ might switch, so we do need to acquire the lock here
         dct = impl.create_dict()
         args, kw = impl.localargs
-        with lock:
-            self.__init__(*args, **kw)
+        self.__init__(*args, **kw)
 
-    setter = impl.dict_setter
-    return dct, orig_dct, lock, setter
+    return dct
 
 
 _marker = object()
@@ -256,8 +260,6 @@ class local(object):
         self = object.__new__(cls)
         impl = _localimpl()
         impl.localargs = (args, kw)
-        impl.locallock = RLock()
-        impl.dict_setter = partial(_osa, self, '__dict__')
         _osa(self, '_local__impl', impl)
         # We need to create the thread dict in anticipation of
         # __init__ being called, to make sure we don't call it
@@ -265,10 +267,10 @@ class local(object):
         impl.create_dict()
         return self
 
-    def __getattribute__(self, name):
+    def __getattribute__(self, name): # pylint:disable=too-many-return-statements
         if name == '__class__':
             return _oga(self, name)
-        dct, orig_dct, lock, setter = _get_dict(self)
+        dct = _get_dict(self)
         if name == '__dict__':
             return dct
         # If there's no possible way we can switch, because this
@@ -281,54 +283,100 @@ class local(object):
         # long as we were in here anyway.
         # Similarly, a __getattr__ will still be called by _oga() if needed
         # if it's not in the dict.
+
+        type_self = type(self)
+        # Optimization: If we're not subclassed, then
+        # there can be no descriptors except for methods, which will
+        # never need to use __dict__.
+        if type_self is local:
+            return dct[name] if name in dct else _oga(self, name)
+
+        type_attr = getattr(type_self, name, _marker)
         if name in dct:
-            if not hasattr(type(self), name):
-                # We elide even setting the dict in the first place in
-                # this case.
+            if type_attr is _marker:
+                # If there is a dict value, and nothing in the type,
+                # it can't possibly be a descriptor, so it is just returned.
                 return dct[name]
-        else:
-            # It's not in the dict at all. Is it in the type, but
-            # not a descriptor (has no __get__)? Then it can't execute any
-            # code and so doesn't need the lock or dct swizzle
-            type_attr = getattr(type(self), name, _marker)
-            if type_attr is not _marker and not hasattr(type_attr, '__get__'):
+
+            # It's in the type *and* in the dict. If the type value is
+            # a data descriptor (defines __get__ *and* either __set__ or
+            # __delete__), then the type wins. If it's a non-data descriptor
+            # (defines just __get__), then the instance wins. If it's not a
+            # descriptor at all (doesn't have __get__), the instance wins.
+            # NOTE that the docs for descriptors say that these methods must be
+            # defined on the *class* of the object in the type.
+            type_type_attr = type(type_attr)
+            if not hasattr(type_type_attr, '__get__'):
+                # Entirely not a descriptor. Instance wins.
+                return dct[name]
+            if hasattr(type_type_attr, '__set__') or hasattr(type_type_attr, '__delete__'):
+                # A data descriptor.
+                # arbitrary code execution while these run. If they touch self again,
+                # they'll call back into us and we'll repeat the dance.
+                return type_type_attr.__get__(type_attr, self, type_self)
+            # Last case is a non-data descriptor. Instance wins.
+            return dct[name]
+        elif type_attr is not _marker:
+            # It's not in the dict at all. Is it in the type?
+            type_type_attr = type(type_attr)
+            if not hasattr(type_type_attr, '__get__'):
+                # Not a descriptor, can't execute code
                 return type_attr
+            return type_type_attr.__get__(type_attr, self, type_self)
 
-        lock.acquire()
-        setter(dct)
-        try:
-            return _oga(self, name)
-        finally:
-            setter(orig_dct)
-            lock.release()
+        # It wasn't in the dict and it wasn't in the type.
+        # So the next step is to invoke type(self)__getattr__, if it
+        # exists, otherwise raise an AttributeError.
+        # we will invoke type(self).__getattr__ or raise an attribute error.
+        if hasattr(type_self, '__getattr__'):
+            return type_self.__getattr__(self, name)
+        raise AttributeError("%r object has no attribute '%s'"
+                             % (type_self.__name__, name))
 
     def __setattr__(self, name, value):
         if name == '__dict__':
             raise AttributeError(
                 "%r object attribute '__dict__' is read-only"
                 % self.__class__.__name__)
-        dct, orig_dct, lock, setter = _get_dict(self)
-        lock.acquire()
-        setter(dct)
-        try:
-            return _osa(self, name, value)
-        finally:
-            setter(orig_dct)
-            lock.release()
+        dct = _get_dict(self)
+
+        type_self = type(self)
+        if type_self is local:
+            # Optimization: If we're not subclassed, we can't
+            # have data descriptors, so this goes right in the dict.
+            dct[name] = value
+            return
+
+        type_attr = getattr(type_self, name, _marker)
+        if type_attr is not _marker:
+            type_type_attr = type(type_attr)
+            if hasattr(type_type_attr, '__set__'):
+                # A data descriptor, like a property or a slot.
+                type_type_attr.__set__(type_attr, self, value)
+                return
+        # Otherwise it goes directly in the dict
+        dct[name] = value
 
     def __delattr__(self, name):
         if name == '__dict__':
             raise AttributeError(
                 "%r object attribute '__dict__' is read-only"
                 % self.__class__.__name__)
-        dct, orig_dct, lock, setter = _get_dict(self)
-        lock.acquire()
-        setter(dct)
+
+        dct = _get_dict(self)
+        type_self = type(self)
+        type_attr = getattr(type_self, name, _marker)
+        if type_attr is not _marker:
+            type_type_attr = type(type_attr)
+            if hasattr(type_type_attr, '__delete__'):
+                # A data descriptor, like a property or a slot.
+                type_type_attr.__delete__(type_attr, self)
+                return
+        # Otherwise it goes directly in the dict
         try:
-            return object.__delattr__(self, name)
-        finally:
-            setter(orig_dct)
-            lock.release()
+            del dct[name]
+        except KeyError:
+            raise AttributeError(name)
 
     def __copy__(self):
         impl = object.__getattribute__(self, '_local__impl')